<%BANNER%>

Iterative Solvers for Hybridized Finite Element Methods

Permanent Link: http://ufdc.ufl.edu/UFE0024820/00001

Material Information

Title: Iterative Solvers for Hybridized Finite Element Methods
Physical Description: 1 online resource (100 p.)
Language: english
Creator: Tan, Shuguang
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: algorithm, condition, differential, discontinuous, discretization, element, equation, finite, galerkin, hybridized, iterative, lagrange, matrix, mesh, method, mixed, multigrid, numerical, partial, triangulation
Mathematics -- Dissertations, Academic -- UF
Genre: Mathematics thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: We consider the application of a variable V-cycle multigrid algorithm for the hybridized mixed method for second order elliptic boundary value problems. Our algorithm differs from previous works on multigrid for the mixed method in that it is targeted at efficiently solving the matrix system for the Lagrange multiplier of the method. Since the mixed method is best implemented by first solving for the Lagrange multiplier and recovering the remaining unknowns locally, our algorithm is more useful in practice. The critical ingredient in the algorithm is a suitable intergrid transfer operator. We design such an operator and prove mesh independent convergence of the variable V-cycle algorithm. We then extend this multigrid framework to the hybridized local discontinuous Galerkin method, and yield similar mesh independent convergence results. Numerical experiments are presented to indicate the asymptotically optimal performance of our algorithm, as well as the performance comparison among different hybridized finite element methods for targeted problems.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Shuguang Tan.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Gopalakrishnan, Jayadeep.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-02-28

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0024820:00001

Permanent Link: http://ufdc.ufl.edu/UFE0024820/00001

Material Information

Title: Iterative Solvers for Hybridized Finite Element Methods
Physical Description: 1 online resource (100 p.)
Language: english
Creator: Tan, Shuguang
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: algorithm, condition, differential, discontinuous, discretization, element, equation, finite, galerkin, hybridized, iterative, lagrange, matrix, mesh, method, mixed, multigrid, numerical, partial, triangulation
Mathematics -- Dissertations, Academic -- UF
Genre: Mathematics thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: We consider the application of a variable V-cycle multigrid algorithm for the hybridized mixed method for second order elliptic boundary value problems. Our algorithm differs from previous works on multigrid for the mixed method in that it is targeted at efficiently solving the matrix system for the Lagrange multiplier of the method. Since the mixed method is best implemented by first solving for the Lagrange multiplier and recovering the remaining unknowns locally, our algorithm is more useful in practice. The critical ingredient in the algorithm is a suitable intergrid transfer operator. We design such an operator and prove mesh independent convergence of the variable V-cycle algorithm. We then extend this multigrid framework to the hybridized local discontinuous Galerkin method, and yield similar mesh independent convergence results. Numerical experiments are presented to indicate the asymptotically optimal performance of our algorithm, as well as the performance comparison among different hybridized finite element methods for targeted problems.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Shuguang Tan.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Gopalakrishnan, Jayadeep.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-02-28

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0024820:00001


This item has the following downloads:


Full Text

PAGE 1

1

PAGE 2

2

PAGE 3

3

PAGE 4

IamgreatlythankfulforenormoushelpfromthefollowingpeoplewhoserveinmyPhDsupervisorycommittee,especiallymyadvisor: Prof.JaydeepGopalakrishnan,DepartmentofMathematics(advisor,committeechair) Prof.WilliamHager,DepartmentofMathematics Prof.SergeiPilyugin,DepartmentofMathematics Prof.TrevorPark,DepartmentofStatistics Prof.AlperUngor,DepartmentofComputerScience 4

PAGE 5

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 7 LISTOFFIGURES .................................... 8 ABSTRACT ........................................ 9 CHAPTER 1INTRODUCTION .................................. 10 2HYBRIDIZEDFINITEELEMENTMETHODS .................. 13 2.1AStandardFiniteElementMethod ...................... 14 2.2MixedMethod ................................. 16 2.3HybridizedMixedMethod ........................... 18 2.4HybridizedDiscontinuousGalerkinMethod .................. 23 2.4.1HDGMethod .............................. 24 2.4.2HIPMethod ............................... 26 3MULTIGRIDALGORITHMFORTHEHRTMETHOD ............. 29 3.1MultigridAlgorithm .............................. 29 3.2ProofoftheConvergenceResult ........................ 34 3.2.1VericationofCondition 3.4 ...................... 35 3.2.2VericationofCondition 3.5 ...................... 36 3.2.3VericationofCondition 3.6 ...................... 46 3.3NumericalExperiments ............................. 49 4ESTIMATESFORTHEHDGMETHOD ..................... 55 4.1EstimatesforHDG ............................... 55 4.1.1StabilityoftheHDGLocalSolvers .................. 57 4.1.2ErrorEstimatesfortheHDGMethod ................. 63 4.1.3ConditioningoftheHDGMethod ................... 65 4.2EstimatesforHIP ................................ 69 4.3NumericalExperiments ............................. 71 5MULTIGRIDFORHDG ............................... 78 5.1MultigridConvergenceAnalysis ........................ 80 5.1.1NormofProlongationOperator(Condition 3.4 ) ............ 81 5.1.2RegularityandApproximationProperty(Condition 3.5 ) ....... 84 5.2NumericalExperiments ............................. 93 5

PAGE 6

.................................... 95 REFERENCES ....................................... 97 BIOGRAPHICALSKETCH ................................ 100 6

PAGE 7

Table page 3-1Performancecomparisonbetweenconjugategradientandandmultigrid. ..... 50 3-2DiscretizationerrorsforthehybridizedmixedRaviart-Thomasmethod. .... 51 3-3Applicationofthemultigridalgorithmtoaproblemonthenon-convexdomain 52 3-4V-cyclewithconstantnumberofsmoothings. ................... 52 3-5Failureofcertainintergridoperators ........................ 54 4-1DiscretizationerrorkPuhkaofHDG(d=0)forvaryingandh. ....... 72 4-2DiscretizationerrorkuuhkL2ofHDG(d=0)forvaryingandh. ....... 72 4-3Discretizationerrork~q~qhkL2ofHDG(d=0)forvaryingandh. ....... 73 4-4DiscretizationerrorkPuhkaofHDG(d=1)forvaryingandh. ....... 73 4-5DiscretizationerrorkuuhkL2ofHDG(d=1)forvaryingandh. ....... 73 4-6Discretizationerrork~q~qhkL2ofHDG(d=1)forvaryingandh. ....... 73 5-1ThenumberofmultigriditerationsforHDG(d=0)withdierent 94 5-2ThenumberofmultigriditerationsforHDG(d=1)withdierent 94 7

PAGE 8

Figure page 2-1Illustrationofnotations ............................... 19 3-1Reningmesh:initialmeshontheleft,andrenedmeshontheright ...... 49 3-2Thenon-convexdomainusedinexperiments. ................... 51 3-3Arenedtriangleusedtodescribethefailedintergridtransferoperators .... 53 4-1ComparisonoflogkuuhkL2:CGvsHRTvsHDG ................ 75 4-2Comparisonoflogk~q~qhkL2:HRTvsHDG .................... 76 4-3ComparisonoflogkPuhka:HRTvsHDG ................... 77 8

PAGE 9

WeconsidertheapplicationofavariableV-cyclemultigridalgorithmforthehybridizedmixedmethodforsecondorderellipticboundaryvalueproblems.OuralgorithmdiersfrompreviousworksonmultigridforthemixedmethodinthatitistargetedatecientlysolvingthematrixsystemfortheLagrangemultiplierofthemethod.SincethemixedmethodisbestimplementedbyrstsolvingfortheLagrangemultiplierandrecoveringtheremainingunknownslocally,ouralgorithmismoreusefulinpractice.Thecriticalingredientinthealgorithmisasuitableintergridtransferoperator.WedesignsuchanoperatorandprovemeshindependentconvergenceofthevariableV-cyclealgorithm.WethenextendthismultigridframeworktothehybridizedlocaldiscontinuousGalerkinmethod,andyieldsimilarmeshindependentconvergenceresults.Numericalexperimentsarepresentedtoindicatetheasymptoticallyoptimalperformanceofouralgorithm,aswellastheperformancecomparisonamongdierenthybridizedniteelementmethodsfortargetedproblems. 9

PAGE 10

Theniteelementmethod(FEM)isoneofthemostoftenusedtoolsfornumericallysolvingpartialdierentialequations(PDE)oncomplicateddomains.Dependingontheimportanceofthepropertiesneededinanapproximatesolution,onechoosestherightkindofniteelementmethodforvariousapplications.HybridizedniteelementmethodshaverecentlyemergedasapowerfulsubclassofFEMwiththeabilitytoecientlyyieldapproximatesolutionswithinterestingproperties.Inthisstudy,weshallbeconcernedprimarilywithacceleratingthesolutionprocesswhenhybridizedniteelementsareused. Likeotherniteelementmethods,thehybridizedniteelementmethodsyieldmatrixsystemswithconditionnumberthatgrowsasmeshsizedecreases.Henceitisnecessarytousepreconditionediterativesolversorfastlinearsolverslikemultigrid(MG)algorithmstoobtainthesolutioneciently.However,theapplicabilityofsuchtechniquesforthehybridizedmethodshasnotbeeninvestigateduptonow.Inthisstudy,weshalldevelopmultigridalgorithmsforecientsolutionofthesystemsresultingfromdiscretizationthroughthehybridizedschemes. Whenlineariterativemethodsareappliedtosolvesuchsystems,wecanthinkthattheerrorconsistsoftwoparts:lowfrequencycomponentsandhighfrequencycomponents,correspondingtosmalleigenvaluesandlargeeigenvaluesrespectively.ClassicalmethodssuchastheGauss-Seideliteration,reducehighfrequencycomponentsquickly,butforthoselowfrequencycomponents,theconvergencebecomesreallyslowwhenthemeshsizedecreases.Amultigridmethod,ontheotherhand,usesadierentmechanismtoovercomethisdiculty.Generallyspeaking,itmaintainsasequenceofgridsstartingfromthecoarsesttothenest,withthelastonebeingthemeshonwhichtheproblemneedstobesolved.Amultigridmethodstillusesclassicaliterationstoeliminatehighfrequencycomponentsonnergrids(calledsmoothing),butforlowfrequencycomponents,ittransfersthemtocoarsergridstodothereduction(calledcorrection).Sincethenumber 10

PAGE 11

4ofthefollowingnermesh),therequiredamountofworkisgreatlyreduced.Thatiswhymultigridmethodhasoptimalperformanceoverallotherclassicaliterativemethods. Thedicultiesinadaptingmultigridtechniquestohybridizedniteelementschemesariseduetotwonon-standardfeaturesofhybridizedmethods: 1. Thespacesusedtondtheapproximatesolutionsconsistoffunctionsdenedontheedgesofa(twodimensional)meshoftriangles.Incontrast,otherniteelementmethodsusefunctionsdenedonelements(triangles)ofamesh. 2. Theapproximatesolutiongivenbyahybridizedmethodsatisesavariationalformulationinvolvingameshdependentbilinearform.IncontrastmanyofthestandardniteelementapproximationsarecharacterizedviaavariationalequationthatmakesenseinaSobolevspace(withoutreferringtoamesh). Becauseofthesefeatures,whenadaptingmultigridalgorithmstohybridizedniteelementmethods,wemustdesignthecomponentsofthealgorithmtoworkwithnon-nestedmultilevelniteelementspaces,andnon-inheritedbilinearforms.WeshalluseanabstracttheoryforthesocalledvariableV-cyclealgorithmwhichhasoftenprovedsuccessfulinanalyzingmultigridadaptationstootherproblemswithsimilardiculties. Thetwomainingredientsofanymultigridalgorithmaresmoothersandintergridtransferoperators.Fortheproblemsweshallconsider,smoothersdonotposenewchallenges.However,thedesignofproperintergridtransferoperatorstailoredtothehybridizedschemesarecriticalforthesuccessofthemultigridalgorithm.Theseareoperatorsusedwithinthealgorithmtomovedatafromacoarsergridtoanergrid,andviceversa.Sincehybridizedmethodsusefunctionsdenedonmeshedges,wemustdesignnon-trivialintergridtransferoperators.Asweshallsee,manyofthe\obvious"choicesdonotresultinecient,orevenconvergent,multigridalgorithms. Thenextchapterisdevotedtohybridizedniteelementmethods.Wegiveageneralintroductiontoniteelementtechniques,includingthe"standard"FEM,andwellknown 11

PAGE 12

12

PAGE 13

Thepurposeofthischapteristointroduceaclassofmethodscalledhybridizedniteelementmethodsforboundaryvalueproblems,whichformsthemainsubjectofourresearch.Toputthismethodinproperperspective,webeginthischapterwithanintroductiontothemostbasicandstandardniteelementmethod,whichcanbetracedbacktoanearlypaperofCourant[ 21 ].Then,weshallpresenttheso-calledmixedniteelementmethod,whichwasintroducedasameanstoobtainbetteruxapproximations[ 12 31 ].Thesediscussionswillprovidethenecessaryleadintodiscussthehybridizedmethodsinthelatersections,beginningwiththehybridizedmixedmethod[ 2 17 ],andlaterthehybridizeddiscontinuousGalerkin(DG)methods[ 19 ]. Letusdenesomenotationshererst,whichwillbeusedthroughouttherestofthedissertation.LetbeapolygonaldomaininR2.A\subdivision"ofisanitecollectionofclosedsetscalled\elements"fKigsuchthattheirinteriorsaredisjointandtheirunionistheclosureof.Fortheniteelementmethodswehaveinmind,werequireasubdivisionwiththefurtherpropertythatallelementsKiaretrianglesandnovertexofanytriangleliesintheinteriorofanedgeofanothertriangle.Suchsubdivisionsarecalled\triangulations".WedenoteatriangulationofbyTh.Thesubscripthreferstoh=maxK2ThhK;wherehK=diam(K):

PAGE 14

whereisasabove,andf2L2().Thewellknownweakformulationofthisproblemistondu2H10()suchthata(u;v)=F(v);8v2H10() (2.2) wherea(u;v)=Z~ru~rvdx;F(v)=Zfvdx;H10()=fv2H1():vj@=0g;andH1()=fv2L2():@iv2L2();8ig: TheessentialideaofniteelementmethodsistoreplaceH10()withanite-dimensionalsubspaceVh,andthenseekanapproximationuhtotheexactsolutionufromthissubspace.Forexample,wecanchooseVh=fv2C():vj@=0;andvjKislinear;8K2Thg;

PAGE 15

Withthebasisabove,wecanwritetheapproximationasuh=nXi=1cii 2.2 ),wehaveZ~r(nXi=1cii)~rvdx=Zfvdx;8v2Vh;i.e.nXi=1ciZ~ri~rvdx=Zfvdx;8v2Vh: whereAi;j=R~ri~rjdx,Fi=Rfidx,andCisthevectorofcoecientsfcig.SincetheonlyunknownhereisC,bysolvingthematrixsystem,wecanndtheapproximationuhinVhtotheactualsolutionu.Moreover,thematrixAissymmetricandpositivedenite,allowingustosolveitconvenientlywithiterativetechniquesliketheConjugateGradientmethod[ 28 ]. Thefollowingwellknowntheorem[ 14 ]statesthetwopropertiesmostpertinenttothisstudy,namelytheconditioningofthesystem( 2.3 ),andanapriorierrorestimateforthediscreteapproximation.WeshallusekkHk(W)andjjHk(W)todenoteSobolevnormsandseminormsdenedonsomedomainWthroughoutthisstudy. 2.1 ,thenthespectralconditionnumberofthematrixAin( 2.3 )satises(A)Ch2;

PAGE 16

2.2 tellusthatifweusetheConjugateGradientmethodtosolve( 2.3 ),wecanexpectthenumberofiterationstogrowashdecreases.Specically,theconvergencerateoftheConjugateGradientmethodintheA-normis(p 2.3 )wemustresorteithertopreconditioning,ortotheconstructionofoptimallineariterativesolverslikemultigridmethods.Ouraimistostudytheapplicabilityofmultigridtechniquestosystemsanalogousto,butmorecomplexthan( 2.3 ),arisingfromhybridizedniteelementmethods(seeChapter 3 ). TheniteelementmethodintroducedaboveisanexampleofaGalerkinmethod.SinceitusesaniteelementspaceVhofcontinuousfunctions,itisaexampleofacontin-uousGalerkin(CG)method.MethodsthatusediscontinuousniteelementfunctionswillbecalleddiscontinuousGalerkin(DG)methods.Weshallseeexamplesofsuchmethodsinlatersections.Anengineeringapproachtotheconstructionofniteelementmethodscanfoundin[ 32 ]. 16

PAGE 17

Stillconsiderthemodelproblem( 2.1 ).Ifweintroduceanewvariable~q,denedas~q=~ru; 2.1 )canbereformulatedas Todistinguish,peopleoftenreferuastheprimalsolution,and~qastheuxsolution. TheRTmixedmethodseeksanapproximation(~qh;uh)totheexactsolution(~q;u)of( 2.4 ).LetPd(K)denotethespaceofpolynomialsonKofdegreeatmostd,andRd(K)bethecorrespondingRTspace,i.e.,Rd(K)=Pd(K)Pd(K)+~xPd(K): 17

PAGE 18

2.5 )givesrisetoamatrixsystemofblockform whereAi;j=R~vi~vjdx,Bi;j=Rwi~r~vjdx,Fi=Rfwidx,andQandUarethecoecientvectorsfor~qhanduhwithrespecttotheircorrespondingbasisf~vigandfwjg,respectively.Wecanthensolvethelinearsystemtogetanapproximationto(~q;u). Notethatincontrastto( 2.3 ),themixedmethodyieldsthesystem( 2.6 )thatisindenite.TheConjugateGradientmethodisnotappropriateforsuchsystems.WemustuseGMRES[ 33 ]orsimilarmethods.Whileitispossibletotailorpreconditioningstrategiesforthismethod,theyareconsiderablymoreinvolved,duetotheindenitenessofthesystem.Incontrast,thehybridizedversionofthemixedmethod,tobediscussedinthenextsection,resultsinasymmetricpositivedenitesystemthatgivestheexactlythesamesolutionasthemixedmethod.Apriorierrorestimateisavailableforthemixedmethod[ 18 23 ]. 2.6 ).Sincethesystemisnotpositivedenite,solvingforQandUisnotalwayseasy.AlthoughbyeliminationofQfromtheequations,wecanachieveapositivedenitesystem,thiswillrequiretheinversionofA,andtheinversematrixistypicallyafullmatrixofbigsize.Fortunately,anadvancedtechniquecalledhybridizationhelpsusovercomethisdiculty.Letusintroducethehybridizedmixedmethod,basedonthesamemodelproblem( 2.4 ). Themethodintroducesanadditionalunknownh,calledLagrangemultiplier,whichturnsouttobenothingbutanapproximationtothetraceofuoneachedgee2Eih.Aswewillshownext,itsintroductioncanactuallyeliminateboth~qhanduh,reducing 18

PAGE 19

Here[[]]iscalledthejumpoperator,denedasbelow. 2-1 ). Illustrationofnotations Itisawell-knownfact[ 2 17 ]thattheHRTmethodgivestheexactlysamesolutionsastheoriginalmixedmethod,asstatedinthefollowingtheorem. 19

PAGE 20

2.7 )isuniquelysolvable,and(~qh;uh)in( 2.7 )coincidewith(~qh;uh)in( 2.4 )fromthemixedmethod. 2.7 ),itmayseemthatweneedtosolveanevenmorecomplicatedequationsystemthantheoriginalmixedmethod.Butthatisreallynotthecase.Toexplainwhy,letusdenetwoso-calledliftingoperatorsasfollows. Giveneverym2L2(Eh),therstliftingpairoffunctions(~QRTm;URTm)2VhWhsatises andgiveneveryf2L2(),thesecondliftingpairoffunctions(~QRTf;URTf)2VhWhsatises forall(~v;w)2VhWh. OnethingtoknowisthedierencebetweentheFEMspacesfortheRTmethodandtheHRTmethod.IncaseoftheRTmethod,VhH(div;)impliesthatfunctionsinsideitneedtohavecontinuousnormalcomponentsacrossboundaries,whilefortheHRTmethod,thisrestrictionisliftedbysimplyrequiringVhL2()L2().Alsonotethatthesemappingsdenedaboveareuniquelydeterminedoneachelementbecauseofthesurjectivityofthemap(r):Vh7!Whrestrictedtoanelement.Moreimportantly,sincethefunctionsinVhWhhavenocontinuityconstraintsacrosselements,thecomputationoftheseliftingpairscanbedoneelementbyelementinadecoupledway.Suchcomputations,beinglocal,areinexpensive. 20

PAGE 21

17 ]. 2.7 ).Then

PAGE 22

2.11 )resultsinamatrixsystemARTC=BRT Byvirtueofthistheorem,insteadofsolving( 2.7 )directly,wecansolve( 2.11 )tondtheLagrangemultiplierh,andthenuse( 2.10 )torecoverthe~qhanduh.Thisrecoveryislocal,astheapplicationof~QRTandURTarelocaloperations,andconsequentlyofnegligiblecost,comparedtotheglobalinversionrequiredtondh. Thevariationalcharacterization( 2.11 )shouldbecomparedwith( 2.2 ).While( 2.2 )makessensenotonlyontheniteelementspace,butalsoonthewholeSobolevspaceH1(),theformulation( 2.11 )forthehybridizedRTmethodonlymakessenseonthespecicmeshunderconsideration.Thisisanimportantdierencethatwewillneedtokeepinmindduringmultigridanalysis. Notethatcomparedtothemixedmethod,thehybridizedversionhasthefollowingadvantages: 1. Thesizeofthematrixsystemcomingfrom( 2.11 )issmallerthantheonefromthemixedmethod( 2.4 )sincethespaceMhisdenedonEh,not. 2. Thematrixfrom( 2.11 )issymmetricpositivedenite,andhencewecansolveitwithfastiterativetechniquessuchastheConjugateGradientmethod. 22

PAGE 23

Oncehiscomputed,theothercomponentsofthesolutionpair,namely~QRTf,~QRTg,URTfandURTg,canbecomputedinexpensivelyinacompletelylocalfashion(elementbyelement),asseenfrom( 2.10 ). 4. 2 ]. 2.6 )canbediagonalized,whichmakesiteasiertocomputetheinverse.suchaniteelementmethodiscalleddiscontinuousGalerkin(DG)method.Similarasinthemixedmethod,byhybridizingDGmethods,wecangetevenbetterresults. Weconsideramoregeneralversionofthemodelproblem( 2.4 ): wheref2L2()andg2H1=2(@). TherearetwokindsofimportantDGmethods:localdiscontinuousGalerkinmethodandinteriorpenaltymethod.Basedonthegeneralframeworkofthehybridizedmixedmethod,wewillbrieyintroducethehybridizedversionsofthesetwomethods. Beforewecontinue,letusmakeafewconventions. 23

PAGE 24

)willalwaysbeexplicitlymentioned. 24

PAGE 25

Then,similarlyasintheHRTmethoddiscussedinx ,wedenetwoliftingoperatorslocallyoneachmeshelementK. Giveneverym2L2(Eh)andeveryK2Th,therstliftingpairoffunctions(~QDGm;UDGm)2VhWhrestrictedonKisinV(K)W(K),andsatises: (~QDGm;~v)K(UDGm;r~v)K=(m;~v~n)@K;forall~v2V(K);(~rw;~QDGm)K+(w;(~QDGm~n+K(UDGmm))@K=0;forallw2W(K):(2.13) Giveneveryf2L2()andeveryK2Th,thesecondliftingpairoffunctions(~QDGf;UDGf)2VhWhrestrictedonKisinV(K)W(K),andsatises: (~QDGf;~v)K(UDGf;r~v)K=0;forall~v2V(K);(~rw;~QDGf)K+(w;(~QDGf~n+K(UDGf))@K=(f;w)K;forallw2W(K):(2.14) TheHDGmethodcanbeviewedasseekingtheapproximation(~qh;uh;h)in(VhWhMh)satisfying (~qh;~v)(uh;r~v)+XK2Th(h;~v~n)@K=(g;~v~n)@forall~v2Vh;(~qh;~rw)+XK2Th(w;bqh~n)@K=(f;w)forallw2Wh;XK2Th(m;bqh~n)@K=0forallm2Mh;(2.15) wherebqh,( 2.4 forthejumpoperator[[]]).AsimilarcharacterizationtheoremasfortheHRTmethodisgivenasfollows,withadditionaltermsinvolvingg: 25

PAGE 26

2.15 ).Then 2.17 )resultsinamatrixsystemADGC=BDG

PAGE 27

,wedenetwoliftingoperatorslocallyoneachmeshelementK. Giveneverym2L2(Eh)andeveryK2Th,therstliftingpairoffunctions(~QIPm;UIPm)2VhWhrestrictedonKisinV(K)W(K),andsatises: (~QIPm;~v)K(UIPm;r~v)K=(m;~v~n)@K;forall~v2V(K);(~rw;~QIPm)K+(w;a~rUIPm~n+K(UIPmm))@K=0;forallw2W(K):(2.18) Giveneveryf2L2()andeveryK2Th,thesecondliftingpairoffunctions(~QIPf;UIPf)2VhWhrestrictedonKisinV(K)W(K),andsatises: (~QIPf;~v)K(UIPf;r~v)K=0;forall~v2V(K);(~rw;~QIPf)K+(w;a~rUIPf~n+KUIPf)@K=(f;w)K;forallw2W(K):(2.19) AsintheHDGmethod,isdouble-valuedoneachinteriormeshfaceofEh. TheHIPmethodcanbeviewedasseekingtheapproximation(~qh;uh;h)in(VhWhMh)satisfying (~qh;~v)(uh;r~v)+XK2Th(h;~v~n)@K=(g;~v~n)@forall~v2Vh;(~qh;~rw)+XK2Th(w;bqh~n)@K=(f;w)forallw2Wh;XK2Th(m;bqh~n)@K=0forallm2Mh;(2.20) wherebqh,( 2.6 )isgivenasfollows: 2.20 ).Then 27

PAGE 28

2.20 )resultsinamatrixsystemAIPC=BIP 28

PAGE 29

Inthischapter,wewilldevelopamultigridalgorithmforsolvingthelinearsystemarisingfromthehybridizedmixedmethod.AswehavealreadyseeninSection 2.3 ,allsolutioncomponentscanberecoveredoncewendhinMhsatisfyinga(h;)=b();forall2Mh: 24 ]. 2.1 holds.ThenthespectralconditionnumberofARTsatises(ART)Ch2: 1 ],theGauss-Seidelmethod[ 27 ],ortheun-preconditionedConjugateGradientmethod[ 28 ]willyieldincreasingiterationcountsashgoesto0.Therefore,weneediterativesolutionstrategiesthatdonotdeteriorateinperformancewhenthemeshsizedecreases.Inthissection,wegiveonesuchtechnique.Thealgorithmtsintotheabstractframeworkof[ 9 10 ]asoneoftheirabstractvariableV-cyclealgorithms.Theiralgorithm,withtheabstractcomponentsparticularizedtoourapplicationisgivenbelow,followingwhichwestateourmainresultontheconvergenceofthealgorithm. Theresultsofthischapterhaveappearedin[ 25 ]. 29

PAGE 30

EachMkisendowedwithtwobilinearforms,(;)k,andak(;).While(;)kisjustthestandardL2()-innerproductfork=0;:::;J1;atlevelJitisameshdependentL2-likeinnerproductdenedby (;)J=XK2TJjKj j@KjZ@K;(3.1) wherejjdenotesthemeasure.TheotherbilinearformonMkMkisdenedbyak(u;v)=8>><>>:Z~ru~rv;k=0;1;;J1;Z~QRTu~QRTv;k=J;

PAGE 31

Themaincomplicationinanon-nestedsettingisthenecessityofdesigningappropriateintergridtransferoperatorsformovingdatabackandforthbetweenthemultilevelgrids.WedenetheprolongationoperatorIk:Mk1!Mk,(k=2;:::;J)by whereMJ:L2(EJ)!MJistheL2(EJ)-orthogonalprojectionontoMJ.Itisimportanttonotethattherearemanynaivechoicesofintergridtransferoperatorsthatdoesnotworkinourapplication.InSection 3.3 ,weshallshownumericalexperimentswithcertain\obvious"transferoperatorsthatleadtoslowconvergenceofmultigrid.Thereversemovementofdata,fromnetocoarselevels,isachievedthroughtherestrictionoperatorQk1:Mk!Mk1,denedby(Qk1!;')k1=(!;Ik')k;forall';!2Mk1: 31

PAGE 32

Thealgorithmgivenbelowperformsmkpre-andpost-smoothingsatlevelk.Ourconvergenceresultisunderanassumptionthatthenumberofsmoothingsincreaseinaspecicway(detailedinTheorem 3.3 )asweproceedtothecoarserlevels. wherethemapMGk(;):MkMk7!Mkisdenedrecursivelyasfollows. 1. First,atthecoarsestlevel,setMG1(u;f)=A11f. 2. Next,fork2,deneMGk(u(i);f)bythefollowingsteps: (a) Setv(0)=u(i). (b) Setresidualrk=fAkv(mk): 32

PAGE 33

(f) Finally,denethenextiteratebysettingMGk(u(i);f)=w(2mk): 10 ]thatEJisalinearoperatoradmittingarecursiveexpression.Usingtheabstracttheoryof[ 9 10 ],weprovethatthisiterativeerrordecreasesgeometricallyatamesh-independentrate,asstatedinthenexttheorem. 3.2 satises0aJ(EJu;u)aJ(u;u);forallu2MJ: 33

PAGE 34

WealsobrieyexplainhowtouseAlgorithm 3.2 asapreconditioner.Sincethealgorithmdenesalineariteration,theoperatorBJ:MJ7!MJdenedbyBJg=MGJ(0;g);forallg2MJ 3.3 showsthat(BJAJ)isboundedaboveandbelowbymeshindependentconstants.HenceBJisanoptimalpreconditioner. 3.3 .Weshallusetheabstractmultigridtheoryof[ 9 10 ]whichallowstheuseofnon-inheritedformsandnon-nestedspaces.Accordingtothistheory,onceweverifythefollowingthreeconditions,theproofofTheorem 3.3 iscomplete.

PAGE 35

Theremainderofthissectionisdividedintothreesubsections,eachdevotedtothevericationofoneoftheaboveconditions. 3.4 foranyK2TJ.Then,bythedenitionofthelifting~QRT(),ZK~QRT(IJvJ1)~r=Z@K(IJvJ1)~r~n;8~r2R0d(K): 3.2 ),therighthandsideabovecanberewrittenasZ@K(IJvJ1)~r~n=Z@K(MJvJ1)~r~n=Z@KvJ1~r~n=ZK(~rvJ1)~r; 35

PAGE 36

3.2 ),wehaveZ@K(IJvJ1)~r~n=Z@KvJ1~r~n=ZK(~rvJ1)~r; ThevericationofCondition 3.4 isnowcompletedbyobservingthatbecauseofLemma 3.7 ,aJ(IJv;IJv)=(~QRT(IJv);~QRT(IJv))=(~rvJ1;~rvJ1)=aJ1(vJ1;vJ1): 3.1 ). 3.5 3.5 typicallyfollowasaconsequenceofsomeregularityresultsfortheunderlyingboundaryvalueproblem,combinedwiththeapproximationpropertiesoftheniteelementspaces.Itiswell-knownthatCondition 3.5 holdsfork=0;1:::;J1[ 6 ],whichispartofthestandardfullregularitybasedproofsofmultigridconvergenceforthecontinuousGalerkinmethod[ 3 5 10 ].SoweonlyneedtoverifyCondition 3.5 withk=J. Forthis,weneedanumberofintermediatelemmasthatestablishpropertiesofvariouslocaloperators.LetusbeginwiththelocalliftingoperatorURT()denedearlier. Proof. 2.8 ),wehaveZK~QRT(IJw)~rZKURT(IJw)r~r=Z@KIJw(~r~n);8~r2Rd(K):

PAGE 37

3.7 ,thisimpliesZK(URT(IJw)w)r~r=0;8~r2Rd(K): Next,weneedtodeneanewlocaloperatorthatmapsapairofinteriorandboundaryfunctionsintoonefunction.LetLd(K)=fp2Pd+3(K):pje2Pd+(e);8edgeeofKg,whered+=8><>:d+1;ifdiseven,d+2;ifdisodd. Supposewearegivenp2L2(K)and2L2(@K).Considerafunction(p;)2Ld(K)thatsatises forallthethreeedgeseofK.Thatsuchaisuniqueisprovednext.Asusual,whenperformingstandardscalingarguments,weobtainconstantsthatdependontheshaperegularityofthemesh,namelyonaxedconstantwhichisthemaximumofdiam(K)=KoverallelementsK,whereKdenotesthediameterofthelargestballinscribedinK. 3.5 ).Furthermore,thereareconstantsC1andC2dependingonlyontheshaperegularityconstantsuchthatC1k(p;)kL2(K)kpkL2(K)+j@Kj1=2kkL2(@K)C2k(p;)kL2(K)

PAGE 38

Proof. 3.5 )formsasquaresystemfor(p;).Indeed,thenumberofequationsinthesystem( 3.5 )equals dim(Pd(K))+3dim(Pd+1(e))=1 2(d+1)(d+2)+3d+:(3.6) Ontheotherhand,thenumberofdegreesoffreedomofLd(K)canbecountedbyaddingtogetherthedimensionofP1(K)(equaling3),thedimensionofthespaceofalledgebubblesofLd(K)(equaling3(d+1)),andthedimensionofinteriorbubblesofLd(K)(equaling(d+1)(d+2)=2).Thus,dim(Ld(K))=3+3(d+1)+1 2(d+1)(d+2); 3.6 ).Thus( 3.5 )isasquaresystem. Toprovethatthereisaunique(p;)satisfying( 3.5 ),itnowsucestoshowthatifpandvanish,theonlysolutionof( 3.5 )istrivial.Tothisend,considerainLd(K)satisfyingZKs=0;8s2Pd(K) (3.7)Ze=0;8e;82Pd+1(e): Thelastequation( 3.8 )impliesthatoneachedgee,jeisapolynomialonPd+(e)thatisorthogonaltoallPd+1(e).HencejemustbetheLegendrepolynomialofdegreed+.Nomatterwhatdis,d+isalwaysodd,hencejeisanoddfunctionontheedgee.Sincethisholdsforallthreeedges,andsincemustbecontinuouson@K,weconcludethatvanisheson@K. Since2Ldvanishesonallthethreeedges,itmusthavetheform=123pd;forsomepd2Pd(K) 38

PAGE 39

3.7 )impliesZK(123)pds=0;8s2Pd(K): 3.5 ). Thenormestimateofthelemmafollowsbecauseifpandareasinthestatementofthelemma,then(p;)=0ifandonlyifp=0and=0.Thus,onaxedreferenceelement^KwehaveC1k(p;)kL2(^K)kpk2L2(^K)+kk2L2(@^K))1=2C2k(p;)kL2(^K); UsingtheabovedenedelementspaceLd(K)oneachmeshelement,wecandeneanewliftingoffromtheelementboundariesintotheelementinteriorsby NotethatSisinH(div;)sinceitsnormalcomponentsarecontinuousbecauseof( 3.5 ).ThenextlemmaestablishesafewpropertiesofSthatweneed.Initsstatement,andintheremainder,kkadenotesthe\energy"-likenormonthenestlevel,i.e.,kk2a=aJ(;)forall2MJ; 3.1 ). 39

PAGE 40

Proof. 3.10 ),rstobservethatiftakesaconstantvalueontheboundaryofsomemeshelement@K,thenStakesthesameconstantvalueonK(thisfollowsfromLemma 3.8 ).Hence,forany,wehave~r(SK)=0;where=1 3.9 18 ,Lemma3.3]Cj@Kj1=2kkL2(@K): 24 ,Theorem2.2],wegetXK2TJk~r(S)k2L2(K)CXK2TJj@Kj1kk2L2(@K)Cah(;) whichproves( 3.10 ). Theproofof( 3.11 )isastraightforwardconsequenceofLemma 3.9 Theidentity( 3.12 )isobviousfrom( 3.5 )and( 3.9 ). WeneedonemoreintermediatemapbeforewecangiveourproofofCondition 3.5 .Todescribethismap,rstwedeneainMJforeveryinMJby (S;S)=(~QRT;~QRT);82MJ:(3.13) ThisequationisuniquelysolvableforinMJ,becauseiftheright-handsideiszero,thenS=0,so=0bytheestimate( 3.11 )ofLemma 3.10 .Next,letf=URT.ThemapweuseinthelaterproofisamapfromMJintoMJ,whichfornotationalsimplicity,we 40

PAGE 41

Thefollowinglemmarevealstherelationshipbetween;~and. 3.15 ),kSkL2()=sup2MJ(S;S) 3.13 )=sup2MJ(AJ;)J 3.11 )CkAJkJ: 3.16 ),notethataJ(;)=(S;S)aJ(~;)=(WhS;S); 3.14 ),therighthandsidefunctionisf=URT=Wh(S) (3.17) by( 3.12 ).Subtracting,andsetting=~,wegetk~k2a=((IWh)S;S(~))=((IWh)S;(IWh)S(~)):

PAGE 42

11 ]kuWhukL2(K)ChJjujH1(K),wegetk~k2aXK2TJCh2JjSj2H1(K)1=2XK2TJCh2JjS(~)j2H1(K)1=2XK2TJCkSk2L2(K)1=2XK2TJCh2JjS(~)j2H1(K)1=2CkAJkJXK2TJCh2JjS(~)j2H1(K)1=2 3.15 ).Finally,applyingtheestimate( 3.10 )ofLemma 3.10 onthelasttermabove,andcancelingthecommonfactor,weobtain( 3.16 ). Withthesepreparations,wecannownishthevericationofCondition 3.5 3.5 with=1. First,byLemma 3.7 Second,theexpressionontherighthandsideabovesatises (~QRT+~r(PJ1);~rvJ1)=0;8vJ12MJ1;(3.19) because(~r(PJ1);~rvJ1)=aJ1(PJ1;vJ1)=aJ(;IJvJ1)=(~QRT;~QRT(IJvJ1))=(~QRT;~rvJ1):

PAGE 43

anditsucestoestimate~QRT+~r(PJ1):Here,andintheremainder,weusekk(withoutanysubscripts)aswellaskkL2()todenotetheL2()-norm. Tobegintheestimation,wesplit~QRT+~r(PJ1)intomanyterms,labelingeachoneasfollows:~QRT+~r(PJ1)=~QRT(~)(termA) (3.21)+~QRT~~QRT(MJ~u)(termB)+~QRT(MJ~u)+~r~u(termC)+~rPJ1~~r~u(termD)+~rPJ1(~);(termE) where~isasdenedin( 3.14 )and~uistheuniquefunctioninH10()thatsolves(~r~u;~rv)=(f;v);8v2H10(): byawellknownregularitytheorem[ 26 ]. Thersttermcanbeestimatedbyk(termA)k=k~QRT(~)k=k~kaChJkAJkJ 3.16 )ofLemma 3.11 .Forthenextterm,rstobservethatduetothecharacterizationofLagrangemultipliersgivenbyTheorem 2.6 ,~isthehybridizedmixedmethodapproximationto~u.HencebyapreviouslyestablishedLagrangemultipliererror 43

PAGE 44

18 ,Theorem3.1], where~q=~r~uandR~qistheRaviart-Thomasinterpolantof~q.Bythestandarderrorestimatesforthisinterpolant,weimmediatelyndthatk(termB)k=k~QRT~~QRT(MJ~u)kChJj~ujH2()ChJkfkL2()by( 3.22 )=ChJkWh(S)kL2()ChJkSkL2()ChJkAJkJ; 3.15 )ofLemma 3.11 Weproceedtoanalyzethenextterm.Forthis,recallthedivergencefreesubspaceR0d(K)denedin( 3.4 ).Bythedenitionof~QRT(),ZK~QRT(MJ~u)~r=Z@K(MJ~u)~r~n=Z@K~u~r~n=ZK~r~u~r Now,suppose~uJ12MJ1istheexactsolutionof(~r~uJ1;~rv)=(f;v);8v2MJ1: 3.24 ),wehave~QRT(MJ~u)+~r~u;~QRT(MJ~u)+~ruJ1K=0;

PAGE 45

3.22 ),and( 3.15 )ofLemma 3.11 For(termD),wewillrstshowthatPJ1~coincideswiththe~uJ1denedabove.Indeed,forallwJ12MJ1,wehave(~rPJ1~;~rwJ1)=(~QRT~;~rwJ1)by( 3.19 )=(~QRT~;~QRT(IJwJ1))byLemma 3.7 =(f;URT(IJwJ1))by( 3.14 )andTheorem 2.6 =(f;wJ1)byLemma 3.8 and( 3.17 ): 3.25 ). Forthenalterm,werstnotethatifwechoose=andvJ1=PJ1in( 3.19 ),thenwehavek~rPJ1k2=(~QRT;~rPJ1);

PAGE 46

3.16 )ofLemma 3.11 Returningto( 3.20 )andcombiningtheestimatesforeachofthetermsabove,weobtainaJ((IIJPJ1);)=k~QRT((IIJPJ1))k2Ch2JkAJk2J 24 ,Theorem2.3],weknowthatJCh2J.HencetheaboveinequalityprovesCondition 3.5 with=1. 3.6 10 ].Forthehighestlevelk=J,theargumentsarealsofairlystandard.Nonetheless,wewillsketchtheproofforthiscasenow. BoththeJacobiandGauss-Seideliterations,whicharewellknownclassicaliterations,canberewrittenusingthemodern\subspacedecomposition"framework.TodisplaythisdecompositionforthenestlevelspaceMJ,letiJ,i=1;2;:::;NJdenotealocalbasisforMJwiththepropertythateachiJissupportedonlyononemeshedge.Further,letMJ;i=spanfiJg.ThenthesubspacedecompositionofMJisMJ=NJXi=1MJ;i

PAGE 47

TheGauss-SeideloperatoroneachMJisdenedas Ontheremaininglevels,smoothersJkandGkcanbewrittenoutusingthestandardsubspacedecompositionsoftheconformingniteelementspaces. WebeginwithasimplelemmaonthestabilityofthedecompositioninthemeshdependentL2-likenormonEJ. 47

PAGE 48

3.26 ),byawellknownlemmaonadditiveoperators[ 10 ],splittingv=PNJi=1ciiJ,(J1Jv;v)=NJXi=1aJ(ciiJ;ciiJ)=NJXi=1c2i(AJiJ;iJ)JNJXi=1c2iJ(iJ;iJ)J=JNJXi=1c2ikiJk2J!CJkNJXi=1ciiJk2J=CJkvk2J; 3.12 Fromtheabovelemma,thesmoothingconditionscanbeveriedbystandardarguments.Indeed,theonlyotheringredientneededisaninequalityoftheformNJXj=1NJXl=1jaJ(vj;wl)jNJXj=1aJ(vj;vj)!1=2NJXl=1aJ(wl;wl)!1=2 7 8 10 ]thefollowinglemma: 1. ifRJ=JJ,then(~R1Jv;v)J1(J1Jv;v)J;8v2MJ: ifRJ=GJ,then(~R1Jv;v)J(J1Jv;v)J;8v2MJ: 3.6 notethat!kvkk

PAGE 49

3.14 withLemma 3.13 .Fortheremainingk,theestimateisstandardforpointJacobiandGauss-Seideloperators.ThusCondition 3.6 isveriedforallk. Fortherstexperimentthatweshallnowdescribe,westartedwithacoarsemeshT1generatedbythepublicdomainmeshingsoftwareTriangle[ 35 ],andthenproducedasequenceofrenementsT2;T3;:::;TJbyconnectingthemidpointsofedges,asexplainedbefore.ThedomainandthersttwomeshesareshowninFigure 3-1 .Supposeweneedtosolvethemodelproblem( 2.4 )onthenestmeshlevelTJforvariouschoicesofJ.Theexactsolutionisu(x;y)=sin(x)ey=2.ThisproblemrequiresanonzeroDirichletboundaryconditionu=gon@,whichentailstheadditionofthetermR@g~v~nontherighthandsideoftherstequationof( 2.7 ).Butthemultigridalgorithmisunaected. Reningmesh:initialmeshontheleft,andrenedmeshontheright.Cornercoordinatorsintheinitialmeshare(0;0);(1;0);(0:8;0:7)and(0;0:5). 49

PAGE 50

3.2 withmk=2Jk.Allexperimentsaredonewiththelowestordermethod,i.e.,d=0.Westartwiththezerofunctionastheinitialiterateandstoptheiterationswhentheinitialerrorisreducedbyafactorof108.WelisttheresultsinTable 3-1 .AllexperimentsarerunonIntelCoreDuoprocessor(CPU@1:73GHz,512MbRAM).Aftersolvingthesystemfor,werecoverbothuand~qasdescribedinx .ThreedierentkindsofdiscretizationerrorsarereportedinTable 3-2 .Theyshowtheconvergenceofniteelementerrorsinaccordancewiththeknowntheoreticalresults[ 2 18 ],andassuchmaybeconsideredtobeavalidationofourcode. Table3-1. Performancecomparisonbetweenunpreconditionedconjugategradientmethodandthemultigridmethod.(Entriesmarked*indicatesunavailabledataduetoexcessivecomputationaltime.) Sizeconjugatemultigrid IterationscpusecsIterationscpusecs 74480.00200.013161010.02260.0213042070.19310.1052964181.64330.392134483317.66342.18856961658148.03349.6034342432831209.653440.681374976650310190.0034163.085502464**34668.6322014976**342607.65 Ascanbeseenfromthelastcolumnofthetable,whenthesizeofthematrixincreasesbyafactorofabout4(whichhappenswhenhishalved),thenumberofmultigriditerationsaswellasthecputimeinsecondsalsoincreasesbyafactorof4.ThisindicatesthatourmultigridalgorithmindeedgivesaniterativeprocesswiththeasymptoticallyoptimalO(N)cost,whereNisthenumberofunknowns.Atthesametime,thecostincreasesbyafactorofaround8fortheconjugategradientmethod,eachtimethemeshsizeishalved.Thisclearlydemonstratesthebenetsofthemultigrid 50

PAGE 51

DiscretizationerrorsforthehybridizedmixedRaviart-Thomasmethod. hkPuhkakuuhkL2k~q~qhkL2 algorithm.Alsonoticethatthenumberofmultigriditerationsseemsboundedevenasthematrixsizegetsverylarge.ThisisinaccordancewiththeconclusionofTheorem 3.3 .Inotherwords,theerrorreductionfactorseemstobeindependentofmeshsize,whichisinaccordancewiththeconclusionofTheorem 3.3 Next,wepresentanexampledesignedtocheckifthesucientconditionthatisconvex(inTheorem 3.3 )isnecessary.WerepeatedtheexperimentswiththedomainasshowninFigure 3-2 .Table 3-3 givestheexperimentaldata.Thenumbersofmultigriditerationstillseemstoremainbounded.Weconcludethatourmultigridalgorithmcanbeeectiveevenwhenisnotconvex. Thenon-convexdomainusedinexperiments. 51

PAGE 52

Applicationofthemultigridalgorithmtoaproblemonthenon-convexdomainofFigure 3-2 Size116496204883203353613465653964821606408646656 Iterations232730323233333333 Next,weinvestigateexibilityinregardtothenumberofsmoothingsmk.Theorem 3.3 assumesthatthenumberofsmoothingsincreasesgeometricallyaswedecreasetherenementlevelk.Wenowrepeattherstexperiment,butinsteadofsettingmk=2Jk,wenowxmktobeoneforallk.Table 3-4 indicatesthattheV-cyclealgorithmcontinuestoexhibitmeshindependentconvergence. Table3-4. V-cyclewithconstantnumberofsmoothings. Size7431613045296213448569634342413749765502464 Iterations212631343434353535 Finally,wegivenumericalresultswhenwereplaceourintergridtransferoperatorwithtwoseeminglyplausibleintergridtransferoperatorsinthevariableV-cycle.Theseoperatorsfail,asweshallsee,buttheyprovideinsightintowhatoneshouldavoidwhenconstructingagoodprolongation.ConsiderAlgorithm 3.2 withthenonnestedmultilevelspacesMk=f2L2(Eik):je2P0(e);foralle2Eikgandnon-inheritedformsateverylevelgivenbyak(u;v)=R~QRTu~QRTv,wherenowtheliftings~QRT()aredenedwithrespecttoTk.Theformsak(;)andabaseinnerproductasin( 3.1 )generalizedtoalllevelsk,denethemultileveloperatorsAkandQkinthealgorithm.Inotherwordsthelowestorderhybridizedmixedmethodisusedtodenethespacesandformsateveryrenementlevelinthealgorithm.ThenweconsidertwodierentintergridtransferoperatorsIk:Mk1!Mk,givenasfollows.ConsideratriangleofmeshTk1,forinstance,thetriangleT=ABCshowninFigure 3-3 .Lete= 52

PAGE 53

Arenedtrianglewiththenotationsusedtodescribethefailedintergridtransferoperators. alltheneredgesasfollows:I(1)kAB=8>>>><>>>>:URT(AB)onthe3newedges( DF; EF)1onthe2newedges( BD)0ontheother4newedges( CF; BE; CE) whereURT()istheliftingoperatordenedin( 2.8 ),butnowwithrespecttothemeshTk1.ThesecondprolongationcandidateweshallconsiderisI(2)k:Mk1!MkisdenedbyI(2)kAB=8>>>><>>>>:1=2onthe2newedges( DF)1onthe2newedges( BD)0ontheother5newedges( CF; BE; CE; EF) Eachoftheseoperatorsgivesadierentmultigridalgorithm.WereportontheperformanceoftheV-cyclealgorithmwiththesetwoprolongationcandidatesandaxednumberofsmoothingsmk=1inTable 3-5 .Clearly,theresultsaredismal. WebelievethatthefailureisduetothefactthatprolongationoperatorslikeI(1)kandI(2)kincreaseenergyuponcontinualtransferofacoarsegridfunctiontoincreasingly 53

PAGE 54

Failureofcertainintergridoperators.(Anentry*indicatesthattheiterationdiverged.) SizeMGiterationcounts withI(1)kwithI(2)k nerlevels.Incontrast,thesuccessfulprolongationIkthatweanalyzed,doesnotincreaseenergy,ascanbeseenfromCondition 3.4 ,whichweveriedforourIk. 54

PAGE 55

ThischapterisdevotedtoastudyoftheHDGmethodintroducedinSection 2.4.1 .ThiswillformpreliminarymaterialforthestudyofthemultigridmethodsforHDGmethodspresentedinthenextchapter. HDGmethodswerediscoveredin[ 19 ].Thersterroranalysisofsuchmethodswasprovidedin[ 15 ].Thisanalysiswasimprovedusingaspecialprojectiontailoredtothemethodin[ 20 ].Theerrorestimatesweshallpresentinthischapterareprovedusingthesameprojection.Theseestimatesareslightlydierentfromthosein[ 20 ](and[ 15 ])inthatoursholdunderlessregularityassumptions.Suchlowregularityestimatesarenecessaryforthemultigridanalysislater,wherewedonotplaceconvexityorothersuchassumptionsonthedomainthatguaranteesolutionsofhigherregularity. Apartfromerrorestimates,wealsoproveboundsonthespectrumoftheoperatorsarisingfromtheHDGmethod.ThisyieldsaconditionnumberestimateforthematrixsystemresultingfromtheHDGmethod,indicatinghowtheperformanceofclassicaliterativetechniques(likeGauss-Seidelorconjugategradientmethods)deterioratewithsystemsize.TheyserveasapreludetothenecessityoftechniqueslikemultigridtoacceleratesolutionofHDGmethods.OurtechniquesalsoeasilyyieldaconditionnumberboundontheHIPmethod,asweshow.WhileconditionnumberboundsfortheHRTmethodswereknownpreviously,nosuchboundswereestablishedfortheHDGorHIPmethodspreviously. .Itsdeningequationsystemis( 2.17 ).Infact,itwillbenotationallyecienttorewritethisformintermsofthefollowingoperators.DeneA:Vh7!Vh,B:Vh7!Wh,C:Vh7!Mh,R:Wh7!Wh, 55

PAGE 56

forsome~gh2Vhandfh2Wh,wherethesuperscript\t"denotestheadjointwithrespectto(;)horh;i@hasappropriate.Itiseasytoseethat( 2.15 )canberewrittenastheabovesystemwithfh=Whf,whereWhdenotestheL2()-orthogonalprojectionintoWh,and~ghsettotheuniquefunctioninVhsatisfying (~gh;~v)h=hg;~v~ni@forall~v2Vh:(4.2) Notethatinthelowestordercasek=0,theoperatorBiszero,butthesystemcontinuestobeuniquelysolvable. Theresultontheabovementionedeliminationcanbedescribedusingadditional\local"operators~QDGV:Vh7!Vh,~QDGW:Wh7!Vh,UDGV:Vh7!Wh,UDGW:Wh7!Wh,whoseactionisdenedbysolvingthefollowingsystems0B@ABtBR1CA0B@~QDGV~ghUDGV~gh1CA=0B@~gh01CA0B@ABtBR1CA0B@~QDGWfhUDGWfh1CA=0B@0fh1CA forall~gh2Vhandfh2Wh. Onecanthenreadilyverifythat~QDGm=~QDGV(Ctm)~QDGW(Stm)andUDGm=UDGV(Ctm)UDGW(Stm),where(~QDGm;UDGm)isdenedin( 2.13 ).Forexample,( 4.3 ) 56

PAGE 57

(~QDGm;~v)K(UDGm;r~v)K=hm;~v~ni@Kforall~v2Vh; whichisequivalentto( 2.13 ). Basedontheseoperatorsabove,wewillobtainestimatesonthestability,conditioning,anddiscretizationerrorsoftheHDGmethod.OurtechniqueconsistsofrstobtainingboundsforvariouslocalsolutionoperatorsoftheHDGmethod.Thelocalboundsthenimplyglobalbounds,suchasboundsforthediscretizationerrorsandthelengthofthespectrum. 4.3 ),withtheobviousmodicationoftherighthandside.ItisinstructivetocompareTheorem 4.1 withasimilarresultfortheseHRToperators,asprovedin[ 18 ,Lemma3.3].Forinstance,onepairofinequalitiesof[ 18 ,Lemma3.3]isk~QRTkKCh1Kkkh;K;kURTkKCkkh;K; 57

PAGE 58

4.5 ),whenisofunitsize.Moreinterestingly,cf.( 4.6 )withk~QRTWfkKChKkfkK;kURTWfkKCh2KkfkK; 18 ,Lemma3.3].Observethatifisofunitsize,theselocalHRToperatorsaremorestablethanthecorrespondingHDGones.Indeed,whileURTWfdampsperturbationsinfbyO(h2),thecorrespondingHDGoperator,namelyUDGWf,dampsitbyonlyO(h)because(dKh)2=O(1=h). WewillnowdevelopaseriesofintermediateresultstoproveTheorem 4.1 intheremainderofthissubsection. 2.8 )bypartsonameshelementKwithm=,wehave (~QRT+~rURT;~v)K=hURT;~v~ni@K:(4.9) Thereisan~vinRd(K)suchthat~v~n=URTon@Kand(~v;~pd1)K=0forall~pd1inPd1(K)n(thisisobviousfromthewell-knowndegreesoffreedomofthespaceRd(K)).Additionally,byascalingargumentitisimmediatethat Withthis~vin( 4.9 ),weobtainkURTk2@K=(~QRT;~v)K; 4.10 ).

PAGE 59

^Ck^wk^Ksup~v2Pd(^K)j(^w;r~v)^Kj k~vk^K+k^wk^F;8^w2Pd(K);(4.11) foranyface^Fof^K.Thisfollowsbyequivalenceofnorms.Thattherighthandsideindeeddenesanormcanbeseenasfollows:divergenceisasurjectivemapfromPd(K)ntoPd1(K).Henceifthesupremumiszero,then^wisorthogonaltoPd1(K),inwhichcase^wiszeroonceitvanishesonanyface^F(see[ 20 ,Lemma2.1]).Thelemmafollowsbymapping( 4.11 )toanysimplexKandusingstandardscalingarguments. (iv) 59

PAGE 60

2.8 )from( 2.13 )wehave ((~QDG~QRT);~v)K(UDGURT;r~v)K=0 (4.17a)(r(~QDG~QRT);w)K+h(UDGURT);wi@K=h(URT);wi@K forall~v2Pd(K)nandforallwinPd(K).Notethatsincer~QRT=0,thelifting~QRTisinfactinPd(K)n.Hencef~QDG~QRT;UDGURTgformstheuniquesolutionof( 4.17 ). First,letusprovetherstassertion (i) ofthelemma.Indeed,iftakesaconstantvalueon@K,thenitiswellknownthatURTequalsthesameconstant[ 17 ]and~QRT=0,sotherighthandsideof( 4.17 )vanishes.HenceURTUDGand~QRT~QDGalsovanish,thusproving (i) Thestatement (ii) isprovedbythesametechniqueas (i) .TheonlydierenceisthattheanalogousresultfortheRTcaseislesswellknown,soletusrstshowit,namelyURTj@K=j@Kwhend>0andj@Kequalsthetraceofsomev2P1(K).Inlightof( 3.7 ),equation( 2.8 )becomes(~rv;~v)(URT;r~v)=hv;~v~ni@K=(~rv;~v)(v;r~v): (i) ,thesolutionof( 4.17 )vanishesinthiscasealso,andwehaveprovenitem (ii) Next,letusprovetheestimates.Setting~v=~QDG~QRTandw=UDGURTin( 4.17 ),wehavek~QDG~QRTk2K+kUDGURTk2;@K=h(URT);UDGURTi@K: 60

PAGE 61

4.18 )implieskUDGURTk;@KkURTk;@K; 4.12 )followsbytriangleinequalityandLemma 4.2 ApplyingCauchy-Schwarzinequalitytotherighthandsideof( 4.18 ),k~QDG~QRTk2K+kUDGURTk2;@KCh1=2Kk~QRTkKmaxK1=2kUDGURTk;@K 4.2 .Clearly( 4.13 )immediatelyfollows. Itremainsonlytoprove( 4.14 ).LetFmaxdenoteafaceofKwhere=maxK:ThenmaxKkUDGURTk2Fmax=kUDGURTk2;FmaxkUDGURTk2;@KCmaxKhKk~QRTk2K; 4.3 ,weobtainkUDGURTkKChKkBt(UDGURT)kK+h1=2KkUDGURTkFmaxChKk~QDG~QRTkK+hKk~QRTkK; 4.14 )follows.(Thisappliesevenifd=0,inwhichcasetheterminvolvingBtisabsent.)Thuswehaveproveditem (iii) Forthenalitem (iv) ,(JK~QDG;JK~v)K=h;JK~v~ni@Kby( 2.13 );=(~QRT;JK~v)K;by( 2.8 ); 4.15 ),as~QRTisintherangeofJKby( 2.8 ).Theestimate( 4.16 )isthenobviousasorthogonalprojectorshaveunitnorm. 61

PAGE 62

4.1 4.13 )ofLemma 4.4 4.8 ): 4.13 )inplaceof( 4.13 ). Next,consider~QDGWf;UDGWf.Fromtheirdenitions,itiseasytoseethat (~QDGWf;~QDGWf)K+hUDGWf;UDGWfi@K=(UDGWf;f)K:(4.19) LetFmaxdenoteafaceofKwhere=maxK:Then,kUDGWfkKC(hKkBtUDGWfkK+h1=2KkUDGWfkFmax)byLemma 4.3 ,C(hKk~QDGWfkK+h1=2KkUDGWfkFmax)asA~QDGWf+BtUDGWf=0,ChK(k~QDGWfkK+(maxKhK)1=2kUDGWfk;@K)ChK((UDGWf;f)1=2K+(maxKhK)1=2(UDGWf;f)1=2K)by( 4.19 );ChK(1+(maxKhK)1=2)kUDGWfk1=2Kkfk1=2K; 4.19 ),weimmediatelygetthestatedboundfor~QDGWfaswell. Finally,toprove( 4.7 ),westartfromthefollowingeasyconsequenceofthedenitionsof~QDGV~g;UDGV~g: Sinceitisimmediatefromtheabovethatk~QDGV~gkKCk~gkK,itonlyremainstoprovetheboundforUDGV~g.ByLemma 4.3 ,kUDGV~gkKChKkBtUDGV~gkK+Ch1=2KkUDGV~gkFmaxChK(kA~QDGV~gkK+k~gkK)+ChK(maxKhK)1=2k~QDGV~gk;@K;

PAGE 63

4.20 )intheabove. 20 ].Here,asanapplicationoftheestimatesweprovedinx ,weprovetwonewerrorestimatesnotin[ 20 ]. Theproofisquickandeasyonceweusethespecialprojectionof[ 20 ].Theprojection,denotedbyh(~q;u),isintotheproductspaceVhWh,anditsdomainisasubspaceofH(div;)L2()consistingofsucientlyregularfunctions,e.g.,H(div;)\Hs()nHs()fors>1=2: 20 ]: forall1=2
PAGE 64

2.12 )be(~q;u),andthediscretesolutionsatisfying( 2.17 )be(~qh;uh;h).Then,thefollowingerrorestimateshold:k~q~qhk2k~qVh~qk; 20 ]: 20 ,Lemma3.1]. 4.5 20 ](andtheproofiseasy),soweonlyprovetheremainingtwo. Toprove( 4.23 ),weapplyTheorem 2.10 to( 4.25 )ofLemma 4.6 .Then,wendthat"hsatisesah("h;)=(~eh;~QDG)=((Vh~q~q);~QDG) forallinMh.Hence( 4.23 )followsbychoosing="handapplyingtheCauchy-Schwarzinequality. 64

PAGE 65

4.24 ),weapplythelocalrecoveryequation( 2.16 )ofTheorem 2.10 to( 4.25 ),whichgives"uh=UDG"h+UDGV~eh:Therefore,k"uhkKkUDG"hkK+kUDGV~ehkKcKhCk"hkh;K+CdKhhKk~ehkK; 4.1 .SincedKhhKCh1=2K(maxK)1=2,kuuhk2KC(cKh)2k"hk2h;K+C(maxK)1hKkVh~q~qk2K+CkuWhuk2K: 4.9 ,weobtainkuuhk2hb(1)Ck"hk2a+Cb(2)kVh~q~qk2h+CkuWhuk2h: 4.24 )usingthepreviousestimate( 4.23 )for"h. 20 ]underadditionalregularityassumptions.Theonlyregularityrequirementfortheestimates( 4.24 )and( 4.23 )toholdisthat(~q;u)isinthedomainofh,whereastheanalysisin[ 20 ]assumesinadditionthefullregularityconditionsuitableforanAubin-Nitschetypeargument. 3.1 ).Themainresultofthissubsectionisthefollowing.

PAGE 66

4.8 imply(2)hCh2: TheproofofTheorem 4.8 reliesonthetwolemmasbelow.Tostatethem,weneedanadditionalnormpreviouslydenedinNotation 2.9 .NotethatsincekkhisanL2-likenorm,jjjjjjhisanH1-likenorm,andsincefunctionsinMhcanbethoughtofashavingzeroboundaryconditionson@,itisnotsurprisingthatthefollowingPoincare-typeinequalityholds: 24 ,ProofofTheorem2.3]. 66

PAGE 67

24 ],theproofoflemmacanbecompletedinstantlybyjjjjjjh;KCk~QRTkKCk~QDGkK 4.16 )ofLemma 4.4 LetTKbetheaneisomorphismmappingthereferenceunitsimplex^Kone-oneontoK.IthastheformTK(^x)=MK^x+bforsomennmatrixMK.WewillalsoneedthePiolamapKmappingfunctionsonKto^K,denedbyK(~v)=(detMK)1M1K~vTK.Westartbyletting^=TKandrecallingthatthereisafunction~v^inPk(K)nsuchthatr~v^=0;in^K;~v^~n=^m^K(^);on@^K;andk~v^k^KCk^m^K(^)k^K: 13 ]appliedto^m^K(^);orevenbymoreelementaryobservations.Nextlet~v=1K(~v^).BythewellknownpropertiesofthePiolamap[ 13 ],weknowthatr~v=0andh;~v~ni@K=h^;~v^~ni@^K: 4.4a ),weget(~QDG;~v)K=h;~v~ni@K=h^;~v^~ni@^K=(^m^K(^);~v^~n)@^K=k^m^K(^)k2@^K:

PAGE 68

4.8 4.12 )and( 4.16 )ofLemma 4.4 toconcludethatkUDGk2;@KCmaxKhKk~QRTk2KCmaxKhKk~QDGk2K: 4.1 .Thus,theupperboundfollows. Forthelowerbound,wecombinetheestimatesofLemmas 4.9 and 4.10 toobtainkk2hC0jjjjjj2hC0Ck~QDGk2hCC0ah(;); 68

PAGE 69

Settingm=;~v=~QDG;w=UDGin( 2.13 ),wehaveaDGh(;)jK=k~QDGkK+kUDGk;@K=(;~QDG~n)@K(;(UDG))@K: wherethelastinequalitycomesfromascalingargument,andj(;(UDG))@Kj(maxK)1=2h1=2(kk@Kh1=2)k1=2(UDG)k@K:

PAGE 70

19 ]:aIPh(;)=XK(k~QDGk2Kk~QDG+~rUDGk2K+kUDGk2;@K) (4.30)=XK(k~rUDGk2K2(UDG;~rUDG~n)@K+kUDGk2;@K) (4.31)=XK((;~rUDG~n)@K+(;K(UDG))@K): Firstletusprovethelowerbound.ByLemma 4.9 and 4.10 ,wejustneedtoshowk~QDGk2KaIPh(;).From( 4.30 ),itisclearthatitsucestoprove Tothisend,startingwiththefollowingidentity[ 19 ,Lemma2.2] (~QDG+~rUDG;~v)K=(UDG;~v~n)@K;8~v2V(K);(4.34) andalocalscalingargument,wehavek~QDG+~rUDGk2KC(minKhK)1=2kUDGk;@Kk~QDG+~rUDGkK; 4.33 )followsprovidedminKhK1=C2. 70

PAGE 71

4.32 ),aIPh(;)=XK(;~rUDG~n)@K+(;K(UDG))@K)Ch1XK(kk@Kh1=2K)(k~rUDGk2K+maxKhKkUDGk2;@K)1=2Ch1XKkk2@KhK!1=2XK(k~rUDGk2K+kUDGk2;@K)!1=2; Forthisweuse( 4.31 )andj(UDG;~rUDG~n)@KjC(minKhK)1=2kUDGk;@Kk~rUDGkK^C(kUDGk2;@K+k~rUDGk2K); 4.31 ),weseethat( 4.35 )follows.Thustheproofiscomplete. 4.13 ,weputstrongrestrictionsonthechoicesoftofavorthearguments.Butwesuspectthatthisestimate( 4.29 )maystillholdforlessrestrictive.Thatconjecturewillbeleftforfutureinvestigation. .Wewillalsoinvestigatethedependenceoftheerrorsonthepenaltyparameter.WeimplementedtheHDGmethodandperformedexperimentswiththesamemeshesanddomainasinx .Theexacttestsolutionisagainu(x;y)=sin(x)ey=2. 71

PAGE 72

4-1 to 4-6 ,namely,kuuhkL2,k~q~qhkL2andkPuhka,wherePuistheL2orthogonalprojectionontoMh.Asseenfromthosetables,forxed,alldiscretizationerrorsconvergeinO(h)ford=0,andO(h2)ford=1.TheseagreewiththeknowntheoreticalestimatesonHDGerrors[ 20 ,Theorem2.1]aswellasthetheoremweprovedearlier,namelyTheorem 4.5 .Noticehowtherateofconvergenceslowsdownifissetto1=h.ThishasbeenatraditionalchoiceinmanyDGmethods.However,forHDGmethods,itisimportantthatthepenaltynotbesetto1=hasseenfromthelastrowofTables 4-1 through 4-6 Table4-1. DiscretizationerrorkPuhkaofHDG(d=0)forvaryingandh. 1=20.903880.005990.002840.001400.000690.0003410.870650.011860.005660.002790.001390.0006920.814770.023190.011190.005550.002770.0013940.732730.044490.021910.010980.005520.0027780.636610.070420.038530.020530.010670.00544 Table4-2. DiscretizationerrorkuuhkL2ofHDG(d=0)forvaryingandh. 1=20.128230.034820.017410.008710.004350.0021810.121150.034820.017410.008710.004360.0021820.110200.034840.017430.008720.004360.0021840.096260.034890.017470.008750.004370.0021980.083050.035060.017620.008850.004430.00222 NextwetestandcomparetheperformanceamongthecontinuousGalerkin(CG)method,theHRTmethodandtheHDGmethodwithdierent.Thecomparisonsarebasedonthreekindsofdierentdiscretizationerrors,namely,theprimalsolutionerrorkuuhkL2,theuxsolutionerrork~q~qhkL2,andthetracesolutionerrorkhka. 72

PAGE 73

Discretizationerrork~q~qhkL2ofHDG(d=0)forvaryingandh. 1=21.215710.005930.002830.001390.000690.0003511.127910.011590.005590.002780.001390.0006920.986720.022160.010940.005490.002760.0013940.792810.040780.020930.010730.005460.0027680.580040.070410.038540.020540.010670.00545 Table4-4. DiscretizationerrorkPuhkaofHDG(d=1)forvaryingandh. 1=20.001700.000470.000123.2e-058.0e-062.0e-0610.002140.000580.000153.8e-059.5e-062.4e-0620.003320.000870.000225.6e-051.4e-053.5e-0640.005880.001530.000399.8e-052.5e-056.2e-0680.010840.002890.000740.000194.7e-051.2e-05 Table4-5. DiscretizationerrorkuuhkL2ofHDG(d=1)forvaryingandh. 1=20.003200.000790.000194.9e-051.2e-053.1e-0610.002640.000660.000164.1e-051.0e-052.6e-0620.002470.000620.000163.9e-059.7e-062.4e-0640.002430.000610.000153.8e-059.5e-062.4e-0680.002460.000610.000153.8e-059.5e-062.4e-06 Table4-6. Discretizationerrork~q~qhkL2ofHDG(d=1)forvaryingandh. 1=20.004540.001150.000297.3e-051.8e-054.6e-0610.004720.001200.000307.7e-051.9e-054.8e-0620.005380.001390.000358.9e-052.2e-055.6e-0640.007250.001920.000490.000133.2e-057.9e-0680.011270.003150.000840.000225.5e-051.4e-05 73

PAGE 74

4-1 .xaxisrepresentsthelogarithmofthedegreesoffreedom,andyaxisrepresentslog(kuuhkL2).TheCGmethod(d=1)usestheniteelementspaceconsistingofcontinuouspiecewiselinearfunctionstoapproximatetheexactsolution.Incontrast,TheHDGmethod(d=1)usestheniteelementspaceofdiscontinuouspiecewiselinearfunctions.Welistthreecasescorrespondingdierent(1/10000,1and10000)forthelatter.TheresultsshowthattheCGmethodyieldsthebestapproximationsfortheprimalvariableu.However,theCGmethoddoesnotyieldgooduxes(whichisonetheprimarymotivationsforconstructingmixedandhybridDGmethodsbasedonthedualform). Next,wecomparek~q~qhkL2amongtheHRTmethod(d=0)andtheHDGmethod(d=1)inFigure 4-2 .Bothmethodsuselinearfunctionstoapproximatetheuxsolution~q,andthedierenceisthatthelatterusesthefullspaceoflinearpiecewisefunctions,whiletheformerusesonlyasubspaceofit(thenormalcomponentsneedtobecontinuousacrosstheinteriorfaces).WeseethatwhiletheperformanceoftheHDGmethodremainsrelativelyunalteredasgoesfromunitsizetozero,itdeterioratesifapproachesinnity. WeproceedtocomparekhkaamongtheHRTmethod(d=0)andtheHDGmethod(d=0)inFigure 4-3 .Bothmethodsusethespaceofpiecewiseconstantfunctionstoapproximatethetracesolution.TheHDGmethodsseemtohavelargererrorsoncoarsermeshes,buttheirperformancequicklybecomecomparabletotheHRTmethodwhengoingtonermeshes. Inallthesegraphs,itmayseemthatthesmalleris,thebettertheapproximationgets,whichconrmstheresultsasindicatedinTable 4-4 .Moreover,thebiggeris,theclosertheHDGcurvegetstotheCGcurve.Asamatteroffact,theCGmethodcanbeconsideredasalimitingcaseoftheHDGmethodbylettinggotoinnity[ 19 ]. 74

PAGE 75

ComparisonoflogkuuhkL2amongtheCGmethod(d=1),theHRTmethod(d=0)andtheHDGmethod(d=1)withdierent. 75

PAGE 76

Comparisonoflogk~q~qhkL2amongthetheHRTmethod(d=0)andtheHDGmethod(d=1)withdierent. 76

PAGE 77

ComparisonoflogkPuhkaamongthetheHRTmethod(d=0)andtheHDGmethod(d=0)withdierent. 77

PAGE 78

ThemaincontributionofthischapterisaprovablyecientmultigridalgorithmfortheHDGmethod.WhilethealgorithmitselfisastraightforwardgeneralizationofwhatwepresentedinChapter 3 fortheHRTmethod,itsanalysisismoreinvolvedduetocomplexities(suchas-dependence)intheestimatesfortheHDGmethod.Furthermore,werelaxtheregularityassumptionsontheoriginalboundaryvalueprobleminthischapter,resultinginamoregenerallyapplicableconvergencetheorem,butnecessitatingmoretechnicalproofs.Theresultsofthischapterwillbepublishedin[ 16 ]. WeusethesamealgorithmsettingasintheHRTcase,specicallywecontinuetouseAlgorithm 3.2 .OurmaingoalistoproveaconvergenceresultforthisalgorithmappliedtotheHDGcase.Again,weusetheabstractmultigridtheoryin[ 10 ],sothattheconvergenceproofreducestothevericationofCondition 3.4 throughCondition 3.6 .NotethattheproofofCondition 3.6 remainsthesameforHRTandHDG(smoothingsubspacesareidentical),soitleavesusonlytherst2conditionstoverify. RecallingthesettingofAlgorithm 3.2 ,notethattherearemanysimilaritieswhenitisadaptedtotheHDGsetting.Asinx ,werequireamultilevelhierarchyofmeshesandspacesandthatthemeshThinwhichthesolutionissought,isobtainedbysuccessiverenementsofacoarsemeshT1asinx .TheHDGmultilevelspacesMkareidenticaltothoseintheHRTsetting.Themaindierenceisthatthebilinearformak(;)andtheresultingmatrices(andoperatorsAk)arenowobtainedfromtheHDGmethod.ThesmoothersdependonAk,sotheymustalsobemodied.TheintergridtransferoperatorIkremainsthesame. Thefollowingisthemainresultofthischapter.Itisprovedunderthefollowingassumptions. 78

PAGE 79

Theboundaryvalueproblem( 2.12 )admitsthefollowingregularityestimateforitssolution: forsomenumber1=21=2.Notealsothatonce( 5.1 )holdswiths>1=2,wecanapplytheprojectionhto(~q;u). 3.2 modiedtoapplytotheHDGmethodasdescribedabove.SupposeAssumption 5.1 holds.Assumealsothatd1.Thenthereexistsapositive<1,independentofthemeshsizehJ,suchthattheerrorreducingoperatorEDGJofthealgorithmsatises0aJ(EDGJu;u)aJ(u;u);forallu2MJ: 5.2 maystillhold.Butthetheoreticalproofrequiressometediousmodications,whicharepostponedforfutureresearch.Throughouttherestofthischapter,wealwaysassumed>0. 79

PAGE 80

. Proof. 4.10 ,soitonlyremainstoprovetheupperbound.Forthis,notethatah(;)=k~QDGk2+kUDGk22k~QDG~QRTk2+2k~QRTk2+C(1+maxKh)k~QRTk2by( 4.12 )C(1)hk~QRTk2by( 3.24 ): 24 ]. 80

PAGE 81

4.4 (ii) .Ifd=0,itfollowsfromthedenitionof~QDG(IJw),namely( 4.4a ),whichreducesto(~QDG(IJw);~r)K0=hIJw;~r~ni@K=hw;~r~ni@K=(~rw;~r)K 5.2 )follows. Toprove( 5.3 ),weuse( 5.2 )inthedenitionofaJ(;)toget andobservethatthelasttermvanishesbecauseofLemma 4.4 (ii) Toprove( 5.4 ),weagainuse( 5.2 ):(~QDG+~rPJ1;~rw)=(~QDG;~rw)+(~rPJ1;~rw)=(~QDG;~QDG(IJw))+aJ1(PJ1;w)=(~QDG;~QDG(IJw))+aJ(;IJw)=h(UDG);UDG(IJw)wi: 4.4 (ii) andweget( 5.4 ). Finally,toprove( 5.5 ),aJ(IJPJ1;)=aJ(;)aJ1(PJ1;PJ1)=(~QDG;~QDG)(~rPJ1;~rPJ1)+kUDGk2=((~QDG+~rPJ1);(~QDG~rPJ1))+kUDGk2: 5.4 )wheneverd>0. 3.4 ) 81

PAGE 82

2.13 ),h(UDG(IJw)w);UDG(IJw)w0i@K=0: Applyingalocaltraceinequality, wherewehaveusedFriedrichsinequalityafterchoosingw0tobethemeanofwonK.Thisproves( 5.7 ). Inequality( 5.8 )followsfrom( 5.7 )bytriangleinequalityonceweprovethat SincethemeanofIJwwvanishesoneachfaceFof@K,applyingFriedrichsinequality,wehave 82

PAGE 83

5.13 )implies( 5.12 ). For( 5.9 ),weuseastandardlocalestimateforlinearfunctions,CkUDG(IJw)wkKhKk~r(UDG(IJw)w)kK+h1=2KkUDG(IJw)w)kF 5.10 ).Thenwenishtheproofasin( 5.11 ). 5.6 5.5 :aJ(IJv;IJv)=(~rv;~QDG(IJv))by( 5.3 )=(a~rv;~rv)by( 5.2 ): 5.2 )ofLemma 5.5 ,hencewecanput~QDG(IJv)=~rvinthedenitionofaj(;)togetaJ(IJv;IJv)=(~rv;~rv)+kUDG(IJv)IJvk2aJ1(v;v)+CmaxKhJk~rvk2; 5.8 )ofLemma 5.7 .Thisprovesthetheorem. 83

PAGE 84

3.5 ) 5.6 5.6 tothelastterm. Nextwedeneanewoperator,S,denedoneachedgeofthemesh.LetbetherestrictionofafunctioninMhon@KforsomemeshelementK.LetFidenotethefaceofKoppositetotheithvertexofK.ThendeneSKiinPk+1(K)by and,consideringallthen+1facesofK,dene(;)S=XK2Th1 5.14a )and( 5.14b )uniquelydeneaSKiinPk+1(K).Further-more,forallinMh,UDGjK=Wh(SKi);

PAGE 85

5.14 )formsasquaresystemforSi,toshowthatithasauniquesolution,itsucestoshowthattheonlysolutionwhentherighthandsidesarezeroisthetrivialsolution.Thatthisisindeedthecaseisanimmediateconsequenceof[ 20 ,Lemma2.1].Theidentity( 5.15 )isobviousfrom( 5.14b ).Letusprovetheremainingassertions. Weprove( 5.16 )byascalingargument.Tothisend,consideraxedreferencesimplex^K,withanarbitrarilychosenface^F,anddene^;^qbyh^;^q;i^F=h^;i^Fforall2Pk+1(^F);(^;^q;v)^K=(^q;v)^Kforallv2Pk(^K): SummingoverallfacesFi@K, wherewehaveusedTheorem 4.1 .Now,toobtain( 5.16 ),weneedonlysumoverallK. Toprove( 5.17 ),rstobservethatiftakesaconstantvalueontheboundaryofsomemeshelement@K,thenSKi.ThisisbecauseUDGbyLemma 4.4 (i) ,sothefunctionsatisesboththeequationsof( 5.14 ).Therefore,bytheuniquesolvabilityof( 5.14 ),SKi.Aconsequenceofthisfactisthatforany,wehave~r(SKi(mK())=0 85

PAGE 86

2.9 ).Therefore,k~r(SKi)kL2(K)=k~rSKi(mK())kL2(K);Ch1KkSKi(mK())kL2(K)(byaninverseinequality)Ch1K(1+(maxKhK)1=2)h1=2KkmK()kL2(@K)(by( 5.18 ))C(1=2)hjjjmK()jjjh;K; 5.17 )followsfromLemma 4.10 Next,wedeneamap7!~fromMJintoMJasfollows.First,giveninMJ,letbetheuniquefunctioninMJsatisfying (;)S=ah(;);82MJ:(5.19) ThisequationisuniquelysolvableforinMJ,becauseiftheright-handsideiszero,thenby( 5.16 )ofLemma 5.9 ,wehavethat=0.Next,letf=UDGanddene~2MJtobetheuniquesolutionoftheequation 5.21 )and( 5.22 )aresimilartotheproofof 5.10 .TheonlydierenceisthatwenowusetheestimatesofLemma 5.9 .Toprove( 5.21 ),rstobservethatf=WhSKi

PAGE 87

5.15 )ofLemma 5.9 .Therefore,kfk2=XK2Th1 5.21 ).Moreover,kk2S=sup2MJ(;)S 5.19 ),=sup2MJ(AJ;)J 5.16 )ofLemma 5.9 ,CkAJkJ; 5.21 ). Toprove( 5.22 ),letusrstnotethatwecanrewrite( 5.19 )and( 5.20 )asfollows:a(;)=XK2Th1 5.9 ,wherebyf=UDG=Wh(SKi)onanyelementK.Subtracting,andsetting=~,wegetk~k2a=XK2Th1 5.21 )and( 5.17 )ofLemma 5.9 .Cancelingthecommonfactorabove,weobtain( 5.22 ). 87

PAGE 88

5.23 ).Tothisend,givenanyinH10(),letJ1inMJ1denoteafunctionsatisfying Suchapproximationsarewellknowntoexist[ 34 ].Then,kfkH1()=sup2H10()(f;) 5.24 );(f;J1UDG(IJJ1))=0;byLemma 4.4 (ii) ,ifd>0,(f;UDG(IJJ1))=aJ(~;IJJ1)by( 5.20 );k~kaCdk~rJ1k;byTheorem 5.6 thetermsinthesupremumcanbeboundedaccordinglytogetthatkfkH1()Cdk~ka+ChJkfk: 5.21 )impliesthatkfkCkAJkJ,andsince( 5.22 )impliesk~kakka+ChJkAJkJ;

PAGE 89

4.8 .Hencewenishtheproof. Now,let~ubetheuniquefunctioninH10()thatsolves (~r~u;~rv)=(f;v);8v2H10();(5.25) andlet~uJ1betheuniquefunctioninMJ1satisfying (~r~uJ1;~rv)=(f;v);8v2MJ1:(5.26) Proof. by( 5.20 ).Now,ifd>0,byLemma 4.4 (ii) ,weknowthatUDG(IJw)w=0.Hencewehave(~rPJ1~;~rw)=(f;w)8w2MJ1; 89

PAGE 90

5.1 ,foranyinMJ,k~QDG+~rPJ1k2Ch2JkAJk2J+Ch2saJ(;)1skAJk2sJ: 5.10 ,kt2kk~PMh~ukak~qVh~qkcby( 4.23 )ofTheorem 4.5 ,Chs(j~qjHs()+maxKj~ujHs())by( 4.21 ), where~q=~r~u.Fort3,weuse( 4.5 )ofTheorem 4.1 togetthat 90

PAGE 91

Since~uJ1isastandardGalerkinapproximation[ 14 ]of~u,wehave Furthermore,astandarddualityargument[ 14 30 ]yields Summing( 5.29 )overallelementsandusing( 5.30 )and( 5.31 ),wecanestimatek~u~uJ1kh.Returningto( 5.28 )andusingthisbound,wehavekt3kC((1)h)1=2hsjujH1+s(): 5.2 )ofLemma 5.5 ,kt5kChJkAJkJ;ifd>0;kt6kCk~kaChkAJkJ;byLemmas 5.8 and 5.10 Combiningtheseestimatesforallti,weobtaink~QDG+~rPJ1k2Ch2JkAJk2J+C(1)hh2sJj~uj2H1+s()+Ch2s(j~qjHs()+maxKj~ujHs())2Ch2JkAJk2J+C(1+maxK2)h2skfk2H1+s() bytheregularityassumption( 5.1 ).SinceH1+s()isaninterpolationspace[ 4 ]inthescaleofintermediatespacesbetweenH1()andL2(),weknowthatkfk2H1+s()kfk2(1s)H1()kfk2s:

PAGE 92

5.23 )and( 5.21 )ofLemma 5.10 .Notethatd2(1s)C(1+(maxKh)1s).Returningto( 5.32 )andusingtheresultingbound,wethenobtaintheestimateofthelemma. 6 10 ]toholdforallk0: 5.12 ,letusrstinvestigatetheremainingterminvolvingUDG.Tothisend,thefollowinginequalitywillbehelpful:kUDG(IJPJ1)(IJPJ1)k;@KCp 4.12 )ofLemma 4.4 .ByLemma 3.7 ,wealsoknowthat~QDG(IJPJ1)=~rPJ1: wherethelastinequalitywasbecause~QDG+~r(PJ1)=JK(~QDG+~r(PJ1)):

PAGE 93

5.33 )togetak((IIkPk1)v;v)k~QDG+~rPJ1k2c+CmaxKhKk~QDG+~r(PJ1)k2KCk~QDG+~rPJ1k2c 5.12 withs=1. 3 .Weusethesamemeshandproblemsettingsasinx .Discretizationerrorsintheniteelementanalysisarepreviouslyreportedinx ,sointhissectionwewillonlyreportobservationsonthemultigridconvergence. Inordertocheckthattheiterationerrorinthemultigridcycleconvergesatarateindependentofthemeshsizeh,wedesigntherstexperimentinasuchwaythatweknowtheexactsolution,asshownbelow: 1. Wesetb=0,i.e.,trytosolvetheequationsystemAx=0withtheexactsolutionx=0. 2. Theinitialguessx0intheMGiterationoneachmulti-levelspaceMk;k=0;1;:::;J,issettobeIJIJ1I1v,wherevisthelinearcombinationoftheglobalbasisfunctionsinM0withthecoecients(1;:::;1)(i.e.,vtakesthevalue1forallinteriornodes,islinearonallmeshelements,iscontinuousacrosselements,anddecreasesto0ontheboundary). 3. WeuseGauss-Seideliterationasthesmoother.Thenumberofsmoothings,mk,equals2JkforHDG(d=0)and42JkforHDG(d=1),onmk. 4. Westopiterationseitherwhenkxixi1ka108kx0xka,orwhentheiterationcountreachesthelimit99. 93

PAGE 94

5-1 and 5-2 forvariousvaluesofillustratestheuniformconvergenceofthemultigriditerationweprovedinTheorem 5.2 .AswementionedinRemark 5.3 ,theexperimentresultsalsoindicatethatthemultigridV-cycleconvergesuniformlyeveninthecaseofd=0.Perhapssurprisinglythecase=1=halsoyieldedgoodmultigridconvergence(unlikethecaseofthediscretizationerrorconvergence{seex ). Table5-1. Thenumberofmultigriditerations(mk=2Jk)forHDG(d=0)withdierent.*indicatesdivergenceoraniterationcountexceedingthelimit. mesh012345 Table5-2. Thenumberofmultigriditerations(mk=42Jk)forHDG(d=1)withdierent. mesh012345 94

PAGE 95

Thismainproductofthisstudyisafastandecientiterativesolutiontechniqueforhybridizedniteelementmethods.Weadoptedthemultigridframeworkfordesigningsuchsolvers.Butattheinitialstages,wewerefacedwiththedicultiesarisingfromthefactthatthemultilevelspacesinallhybridizedmethodswerenon-nested.Infact,themultilevelfunctionsdidnotevensharethesamedomain.Wedevelopedamultilevelframeworkinwhichthisdicultycanbeovercome.Withinthisframework,weanalyzedtwospecicmethodsforwhichweareabletoprovepreciseresults.However,beyondtheregimeofproofs,theframeworkgivesonethekeyingredientstoformulateamultigridalgorithmforanyhybridizedmethod.Withtheinsightsfromournumericalexperiments,weputforththesetechniqueswithgoodcondence,evenfortheotherhybridizedmethods. Letussummarizethemaintheoreticalresults: 1. ConsideringthewellknownHRTmixedmethodrst,weprovedtheuniformconvergenceofamultigridV-cyclealgorithm.Althoughmultigridforthemixedmethodhasbeenstudiedpreviouslybymanyauthors,ouralgorithmisappliedtothecondensedhybridbilinearformandisthusthemostusefulpractically.Aproperlydesignedintergridtransferoperatorwasthekeyingredientthatopenedtheavenuetowardsasuccessfulalgorithm.Wealsoprovidednegativeresultsthatcertainseeminglyplausibleintergridoperatorsdonotwork. 2. Next,westudiedtherelativelyrecentclassofDGmethodscalledHDGmethods.Weprovednewerrorestimates.WeshowedthattheconditionnumberoftheHDGmatrixgrowslikeO(h2)asmeshsizehtendstozero.ThetechniquesdevelopedhereareanticipatedtohaveutilityinotherDGmethodsofthehybridtype. 3. Finally,weshowedthatamultigridalgorithmgaveanecientsolverforthematrixsystemsarisingfromtheHDGmethod.Werigorouslyprovedaconvergencerate 95

PAGE 96

Wehavesetapartforfutureresearchanumberofquestionsthataroseduringthisstudy.Forexample,althoughwehavefullproofsonthemultigridconvergencefortheHDGmethodwithd>0,thed=0partneedstobeinvestigatedfurther.Numericalexperimentshavestronglysuggestedthatouralgorithmalsoworkforthatcase.TheHDGd=0caseismorethanofacademicinterestasitusesspacessmallerthaneventhelowestordermixedmethod. ExtensionstoothermethodslikeHIParealsoofinterest.WehavealreadyprovedaconditionnumberestimatefortheHIPmethod(Theorem 4.13 ).However,errorestimatesandmultigridconvergenceboundsremainunknown. Finally,themodicationofadaptiveniteelementtechniquestohybridizedmethodsisaninterestingquestionofsomesignicance.AsHDGmethodsaregainingpopularity,severalgroupsaretryingthemoutforpracticalproblems.Treatmentofpracticalproblemsrequirenotonlyecientsolvers,butalsoadaptiverenementsandaposteriorierrorestimatorstoautomaticallyindicatewhichregionstorene.TheseareissuesthatcurrentlyremainopeninthearenaofHDGmethods. 96

PAGE 97

[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] 97

PAGE 98

[17] [18] ,Erroranalysisofvariabledegreemixedmethodsforellipticproblemsviahybridization,Math.Comp.,74(2005),pp.1653{1677(electronic). [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] 98

PAGE 99

[32] [33] [34] [35] 99

PAGE 100

ShuguangTanwasborninChina.HereceivedhisB.S.inInformationSciencefromthePekingUniversityin2001.HecametoU.S.A.in2003forcontinuinggraduatestudy,andreceivedhisM.SinStatisticsandPh.D.inMathematicsconcurrentlyfromtheUniversityofFloridainthesummerof2009. 100