Lecture Notes in Computational Science 61 and Engineering Editors TimothyJ.Barth MichaelGriebel DavidE.Keyes RistoM.Nieminen DirkRoose TamarSchlick Tarek P. A. Mathew Domain Decomposition Methods for the Numerical Solution of Partial Differential Equations With40Figuresand1Table ABC TarekPoonitharaAbrahamMathew [email protected] ISBN978-3-540-77205-7 e-ISBN978-3-540-77209-5 LectureNotesinComputationalScienceandEngineeringISSN1439-7358 LibraryofCongressControlNumber:2008921994 MathematicsSubjectClassification(2000):65F10,65F15,65N22,65N30,65N55, 65M15,65M55,65K10 (cid:1)c 2008Springer-VerlagBerlinHeidelberg Thisworkissubjecttocopyright.Allrightsarereserved,whetherthewholeorpartofthematerialis concerned,specificallytherightsoftranslation,reprinting,reuseofillustrations,recitation,broadcasting, reproductiononmicrofilmorinanyotherway,andstorageindatabanks.Duplicationofthispublication orpartsthereofispermittedonlyundertheprovisionsoftheGermanCopyrightLawofSeptember9, 1965,initscurrentversion,andpermissionforusemustalwaysbeobtainedfromSpringer.Violationsare liableforprosecutionundertheGermanCopyrightLaw. Theuseofgeneraldescriptivenames,registerednames,trademarks,etc.inthispublicationdoesnotimply, evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfromtherelevantprotectivelaws andregulationsandthereforefreeforgeneraluse. Coverdesign:WMXDesignGmbH,Heidelberg Printedonacid-freepaper 9 8 7 6 5 4 3 2 1 spinger.com In loving dedication to my (late) dear mother, and to my dear father and brother Preface These notes serve as an introduction to a subject of study in computational mathematics referred to as domain decomposition methods. It concerns divide and conquermethodsforthenumericalsolutionandapproximationofpartial differential equations, primarily of elliptic or parabolic type. The methods in this family include iterative algorithms for the solution of partial differential equations, techniques for the discretization of partial differential equations on non-matching grids, and techniques for the heterogeneous approximation of partial differential equations of heterogeneous character. The divide and conquer methodology used is based on a decomposition of the domain of the partial differential equation into smaller subdomains, and by design is suited forimplementationonparallelcomputerarchitectures.However,evenonserial computers, these methods can provide flexibility in the treatment of complex geometry and heterogeneities in a partial differential equation. Interest in this family of computational methods for partial differential equations was spawned following the development of various high perfor- mance multiprocessor computer architectures in the early eighties. On such parallel computer architectures, the execution time of these algorithms, as well as the memory requirements per processor, scale reasonably well with the size of the problem and the number of processors. From a computational viewpoint, the divide and conquer methodology based on a decomposition of the domain of the partial differential equation, yields algorithms having coarse granularity, i.e., a significant portion of the computations can be im- plemented concurrently on different processors, while the remaining portion requires communication between the processors. As a consequence, these al- gorithms are well suited for implementation on MIMD (multiple instruction, multiple data) architectures. Currently, such parallel computer architectures canalternativelybesimulatedusingacluster ofworkstationsnetworkedwith highspeedconnectionsusingcommunicationprotocolssuchasMPI(Message Passing Interface) [GR15] or PVM (Parallel Virtual Machines) [GE2]. VIII Preface The mathematical roots of this subject trace back to the seminal work of H.A.Schwarz[SC5]inthenineteenthcentury.Schwarzproposedaniterative method, now referred to as the Schwarz alternating method, for constructing harmonic functions on regions of irregular shape which can be expressed as theunion of subregionsof regular shape (such asrectangles and spheres). His motivation was primarily theoretical, to establish the existence of harmonic functions on irregular regions, and his method was not used in computations until recently [SO, MO2, BA2, MI, MA37, DR11, LI6, LI7, BR18]. A general development of domain decomposition methodology for par- tial differential equations occurred only subsequent to the development of parallel computer architectures, though divide and conquer methods such as Kron’s method for electrical circuits [KR] and the substructuring method [PR4] in structural engineering, pre-date domain decomposition methodol- ogy. Usage of the term “domain decomposition” seems to have originated around the mid-eighties [GL2] when interest in these methods gained mo- mentum. The first international symposium on this subject was held in Paris in 1987, and since then there have been yearly international conferences on this subject, attracting interdisciplinary interest from communities of engi- neers, applied scientists and computational mathematicians from around the globe. Early literature on domain decomposition methods focused primarily on iterative procedures for the solution of partial differential equations. As the methodology evolved, however, techniques were also developed for coupling discretizations on subregions with non-matching grids, and for constructing heterogeneous approximations of complicated systems of partial differential equationshavingheterogeneouscharacter.Thelatterapproximationsarebuilt by solving local equations of different character. From a mathematical view- point, these diverse categories of numerical methods for partial differential equations may be derived within several frameworks. Each decomposition of a domain typically suggests a reformulation of the original partial differen- tial equation as an equivalent coupled system of partial differential equations posed on the subdomains with boundary conditions chosen to match solu- tions on adjacent subdomains. Such equivalent systems are referred to in these notes as hybrid formulations, and provide a framework for develop- ingnoveldomaindecompositionmethods.Divideandconqueralgorithmscan be obtained by numerical approximation of hybrid formulations. Four hybrid formulations are considered in these notes, suited for equations primarily of elliptic type: • The Schwarz formulation. • The Steklov-Poincar´e (substructuring or Schur complement) formulation. • The Lagrange multiplier formulation. • The Least squares-control formulation. Alternative hybrid formulations are also possible, see [CA7, AC5]. Preface IX The applicability and stability of each hybrid formulation depends on the underlying partial differential equation and subdomain decomposition. For instance, the Schwarz formulation requires an overlapping decomposi- tion, while the Steklov-Poincar´e and Lagrange multiplier formulations are based on a non-overlapping decomposition. The least squares-control method can be formulated given overlapping or non-overlapping decompositions. Within each framework, novel iterative methods, discretizations schemes on non-matching grids, and heterogeneous approximations of the original par- tial differential equation, can be developed based on the associated hybrid formulations. In writing these notes, the author has attempted to provide an accessible introduction to the important methodologies in this subject, emphasizing a matrix formulation of algorithms. However, as the literature on domain de- composition methods is vast, various topics have either been omitted or only touched upon. The methods described here apply primarily to equations of ellipticorparabolictype,andapplicationstohyperbolicequations[QU2],and spectral or p-version elements have been omitted [BA4, PA16, SE2, TO10]. Applications to the equations of elasticity and to Maxwell’s equations have also been omitted, see [TO10]. Parallel implementation is covered in greater depth in [GR12, GR10, FA18, FA9, GR16, GR17, HO4, SM5, BR39]. For additional domain decomposition theory, see [XU3, DR10, XU10, TO10]. A broader discussion on heterogeneous domain decomposition can be found in [QU6], and on FETI-DP and BDDC methods in [TO10, MA18, MA19]. For additional bibliography on domain decomposition, see http://www.ddm.org. Readers are assumed to be familiar with the basic properties of ellip- tic and parabolic partial differential equations [JO, SM7, EV] and tradi- tional methods for their discretization [RI, ST14, CI2, SO2, JO2, BR28, BR]. Familiarity is also assumed with basic numerical analysis [IS, ST10], com- putational linear algebra [GO4, SA2, AX, GR2, ME8], and elements of op- timization theory [CI4, DE7, LU3, GI2]. Selected background topics are re- viewed in various sections of these notes. Chap. 1 provides an overview of domain decomposition methodology in a context involving two subdomain decompositions.Fourdifferenthybridformulationsareillustratedforamodel coercive 2nd order elliptic equation. Chapters 2, 3 and 4 describe the ma- trix implementation of multisubdomain domain decomposition iterative al- gorithms for traditional discretizations of self adjoint and coercive elliptic problems. These chapters should ideally be read prior to the other chapters. Readers unfamiliar with constrained minimization problems and their saddle point formulation, may find it useful to review background in Chap. 10 or in [CI4], as saddle point methodology is employed in Chaps. 1.4 and 1.5 and in Chaps. 4 and 6. With a few exceptions, the remaining chapters may be read independently. X Preface The author expresses his deep gratitude to the anonymous referees who made numerous suggestions for revision and improvement of the manuscript. Deep gratitude is also expressed to Prof. Olof Widlund who introduced the authortothissubjectovertwentyyearsago,toProf.TonyChan,forhiskind encouragement to embark on writing a book extending our survey paper on this subject [CH11], and to Prof. Xiao-Chuan Cai, Prof. Marcus Sarkis and Prof.JunpingWangfortheirresearchcollaborationsandnumerousinsightful discussions over the years. The author deeply thanks Prof. Timothy Barth for his kind permission to use the figure on the cover of this book, and for use of Fig. 5.1. To former colleagues at the University of Wyoming, and to professors Myron Allen, Gastao Braga, Benito Chen, Duilio Conceic¸˜ao, Max Dryja,FredericoFurtado,JuanGalvis,EtereldesGonc¸alves,RaytchoLazarov, Mary Elizabeth Ong, Peter Polyakov, Giovanni Russo, Christian Schaerer, Shagi-Di Shih, Daniel Szyld, Panayot Vassilevski and Henrique Versieux, the authorexpresseshisdeepgratitude.Deepappreciationisalsoexpressedtothe editors of the LNCSE series, Dr. Martin Peters, Ms. Thanh-Ha LeThi, and Mr. Frank Holzwarth for their patience and kind help during the completion of this manuscript. Finally, deep appreciation is expressed to Mr. Elumalai Balamurugan for his kind assistance with reformatting the text. The author welcomescommentsandsuggestionsfromreaders,andhopestopostupdates at www.poonithara.org/publications/dd. January 2008 Tarek P. A. Mathew Contents 1 Decomposition Frameworks................................ 1 1.1 Hybrid Formulations..................................... 2 1.2 Schwarz Framework...................................... 9 1.3 Steklov-Poincar´e Framework .............................. 16 1.4 Lagrange Multiplier Framework ........................... 27 1.5 Least Squares-Control Framework ......................... 36 2 Schwarz Iterative Algorithms .............................. 47 2.1 Background............................................. 48 2.2 Projection Formulation of Schwarz Algorithms .............. 56 2.3 Matrix Form of Schwarz Subspace Algorithms .............. 66 2.4 Implementational Issues .................................. 72 2.5 Theoretical Results ...................................... 77 3 Schur Complement and Iterative Substructuring Algorithms ................................107 3.1 Background.............................................108 3.2 Schur Complement System ...............................110 3.3 FFT Based Direct Solvers ................................125 3.4 Two Subdomain Preconditioners ..........................140 3.5 Preconditioners in Two Dimensions ........................155 3.6 Preconditioners in Three Dimensions.......................162 3.7 Neumann-Neumann and Balancing Preconditioners ..........175 3.8 Implementational Issues ..................................185 3.9 Theoretical Results ......................................192 4 Lagrange Multiplier Based Substructuring: FETI Method .............................................231 4.1 Constrained Minimization Formulation.....................232 4.2 Lagrange Multiplier Formulation ..........................239 4.3 Projected Gradient Algorithm.............................241 4.4 FETI-DP and BDDC Methods ............................250 XII Contents 5 Computational Issues and Parallelization ..................263 5.1 Algorithms for Automated Partitioning of Domains ..........264 5.2 Parallelizability of Domain Decomposition Solvers ...........280 6 Least Squares-Control Theory: Iterative Algorithms ......295 6.1 Two Overlapping Subdomains.............................296 6.2 Two Non-Overlapping Subdomains ........................303 6.3 Extensions to Multiple Subdomains........................310 7 Multilevel and Local Grid Refinement Methods ...........313 7.1 Multilevel Iterative Algorithms............................314 7.2 Iterative Algorithms for Locally Refined Grids ..............321 8 Non-Self Adjoint Elliptic Equations: Iterative Methods ....333 8.1 Background.............................................334 8.2 Diffusion Dominated Case ................................340 8.3 Advection Dominated Case ...............................348 8.4 Time Stepping Applications ..............................364 8.5 Theoretical Results ......................................366 9 Parabolic Equations .......................................377 9.1 Background.............................................378 9.2 Iterative Algorithms .....................................381 9.3 Non-Iterative Algorithms .................................384 9.4 Parareal-Multiple Shooting Method........................401 9.5 Theoretical Results .....................................408 10 Saddle Point Problems ....................................417 10.1 Properties of Saddle Point Systems ........................418 10.2 Algorithms Based on Duality .............................426 10.3 Penalty and Regularization Methods.......................434 10.4 Projection Methods......................................437 10.5 Krylov Space and Block Matrix Methods ...................445 10.6 Applications to the Stokes and Navier-Stokes Equations ......456 10.7 Applications to Mixed Formulations of Elliptic Equations.....474 10.8 Applications to Optimal Control Problems..................489 11 Non-Matching Grid Discretizations ........................515 11.1 Multi-Subdomain Hybrid Formulations.....................516 11.2 Mortar Element Discretization: Saddle Point Approach.......523 11.3 Mortar Element Discretization: Nonconforming Approach.....551 11.4 Schwarz Discretizations on Overlapping Grids...............555 11.5 Alternative Nonmatching Grid Discretization Methods .......559 11.6 Applications to Parabolic Equations .......................564