Introduction to Domain Decomposition Methods. Alfio Quarteroni

Size: px
Start display at page:

Download "Introduction to Domain Decomposition Methods. Alfio Quarteroni"

Transcription

1 . Aachen, EU Regional School Series February 18, 2013 Introduction to Domain Decomposition Methods Alfio Quarteroni Chair of Modeling and Scientific Computing Mathematics Institute of Computational Science and Engineering École Polytechnique Fédérale de Lausanne MOX Modeling and Scientific Computing Mathematics Department Politecnico di Milano 1

2 MOTIVATION DD can be used in the framework of any discretization method for PDEs (FEM, FV, FD, SEM) to make their algebraic solution more efficient on parallel computer platforms DDM allow the reformulation of a boundary-value problem on a partition of the computational domain into subdomains very convenient framework for the solution of heterogeneous or multiphysics problems, i.e. those that are governed by differential equations of different kinds in different subregions of the computational domain. 2

3 THE IDEA The computational domain Ω where the bvp is set is subdivided into two or more subdomains in which problems of smaller dimension are to be solved. Parallel solution algorithms may be used. There are two ways of subdividing the computational domain: with disjoint subdomains with overlapping subdomains Ω 2 Ω 2 Ω 1 Γ 2 Ω Ω 1 Γ Γ 1 Correspondingly, different DD algorithms will be set up. 3

4 AN HISTORICAL REMARK The classical logo of DDM is [ Why this symbol? What does it mean? We have to go back to a work of 1869 by Hermann Schwarz... 4

5 ...who wanted to establish the existence of harmonic functions with prescribed boundary values on regions with non-smooth boundaries: { u = f in Ω where u = g on Ω Ω Ω Schwarz s idea: split Ω into two elementary subdomains: Ω 2 Ω Ω 1 Γ 2 Γ 1 and set up a suitable iterative method passing information between the subproblems. 5

6 MODERN EXAMPLES OF SUBDIVISIONS 6

7 REFERENCES B.F. Smith, P.E. Bjørstad, W.D. Gropp (1996) Domain Decomposition Cambridge University Press, Cambridge. A. Quarteroni and A. Valli (1999) Domain Decomposition Methods for Partial Differential Equations Oxford Science Publications, Oxford. A. Toselli and O.B. Widlund (2005) Domain Decomposition Methods Algorithms and Theory Springer-Verlag, Berlin and Heidelberg. 7

8 CLASSICAL ITERATIVE DD METHODS 8

9 MODEL PROBLEM Consider the model problem: find u :Ω R s.t. { Lu = f in Ω u =0 on Ω L is a generic second order elliptic operator. The weak formulation reads: find u V = H0 1 (Ω) : a(u, v) =(f, v) v V where a(, ) isthebilinearformassociatedwithl. 9

10 SCHWARZ METHODS Consider a decomposition of Ω with overlap: Ω 2 Ω Ω 1 Γ 2 Γ 1 Given u (0) 2 on Γ 1,fork 1: solve solve Lu (k) 1 = f in Ω 1 u (k) 1 = u (k 1) 2 on Γ 1 u (k) 1 =0 on Ω 1 \ Γ 1 Lu (k) 2 = { f in Ω 2 u (k) 1 u (k 1) on Γ 2 1 u (k) 2 = u (k) 2 =0 on Ω 2 \ Γ 2. 10

11 Choice of the trace on Γ 2 : if u (k) 1 multiplicative Schwarz method if u (k 1) 1 additive Schwarz method We have two elliptic bvp with Dirichlet conditions in Ω 1 and Ω 2,andwe wish the two sequences {u (k) 1 } and {u(k) 2 } to converge to the restrictions of the solution u of the original problem: lim k u (k) 1 = u Ω1 and lim k u (k) 2 = u Ω2. The Schwarz method applied to the model problem always converges, with a rate that increases as the measure Γ 12 of the overlapping region Γ 12 increases. 11

12 Example Consider the model problem { u (x) =0 a < x < b, u(a) =u(b) =0, where Ω 2 a γ 2 γ 1 b Ω 1 The solution is u =0. Weshowafewiterationsofthemethod: u (1) 1 u (2) 1 γ 2 a b Clearly, the method converges with a rate that reduces as the length of the interval (γ 2,γ 1 )getssmaller. γ 1 u (0) 2 u (1) 2 u (2) 2 12

13 Remark The Schwarz method requires at each iteration the solution of two subproblems of the same kind as those of the original problem. The boundary conditions of the initial problem remain unchanged. Dirichlet-type boundary conditions are imposed across the interfaces. 13

14 NONOVERLAPPING DECOMPOSITION We partition now the domain Ω in two disjoint subdomains: Ω Ω 1 Ω 2 Γ The following equivalence result holds. Theorem The solution u of the model problem is such that u Ωi = u i for i =1, 2, where u i is the solution to the problem { Lui = f in Ω i u i =0 on Ω i \ Γ with interface conditions u 1 = u 2 and u 1 n L = u 2 n L on Γ / n L is the conormal derivative. 14

15 Example The problem { div (µ u) =f in Ω u =0 on Ω is equivalent to div (µ i u i )=f i u =0 u 1 = u 2 u 1 µ 1 n = µ u 2 2 n in Ω i on Ω i on Γ on Γ Remark The proof of the theorem can be done using the weak formulation ofthe problem. We refer to Lemma in [QV99]. 15

16 DIRICHLET-NEUMANN METHOD Given u (0) 2 on Γ, for k 1solvetheproblems: Lu (k) 1 = f in Ω 1 u (k) 1 = u (k 1) 2 on Γ u (k) 1 =0 on Ω 1 \ Γ Lu (k) 2 = f in Ω 2 u (k) 2 = u(k) 1 n L on Γ n L u (k) 2 =0 on Ω 2 \ Γ The equivalence theorem guarantees that when the two sequences {u (k) 1 } and {u(k) 2 } converge, their limit will be necessarily the solution to the exact problem. The DN algorithm is therefore consistent. However, its convergence is not always guaranteed. 16

17 Example Let Ω = (a, b), γ (a, b), L = d 2 /dx 2 and f =0. Ateveryk 1the DN algorithm generates the two subproblems: (u (k) 1 ) =0 a < x <γ u (k) 1 = u (k 1) 2 x = γ u (k) 1 =0 x = a (u (k) 2 ) =0 γ<x < b (u (k) 2 ) =(u (k) 1 ) x = γ u (k) 2 =0 x = b. The two sequences converge only if γ>(a + b)/2: u (0) 2 u (0) 2 2 a γ a+b b u (2) 2 a a+b 2 γ b u (1) 2 u (1) 2 17

18 AvariantoftheDNalgorithmcanbesetupbyreplacingthe Dirichlet condition in the first subdomain by u (k) 1 = θu (k 1) 2 +(1 θ)u (k 1) 1 on Γ using a relaxation parameter θ>0. In such a way it is always possible to reduce the error between two subsequent iterates. In the previous example, we can easily verify that, by choosing θ opt = u (k 1) 1 u (k 1) 2 u (k 1) 1 the algorithm converges to the exact solution in a single iteration. More in general, there exists a suitable value 0 <θ max < 1suchthat the DN algorithm converges for any possible choice of the relaxation parameter θ in the interval (0,θ max ). 18

19 NEUMANN-NEUMANN ALGORITHM Consider again a partition of Ω into two disjoint subdomains and denote by λ the (unknown) value of the solution u on their interface Γ: λ = u i on Γ (i =1, 2) Consider the following iterative algorithm: for any given λ (0) on Γ, for k 0andi =1, 2, solve the following problems: with Lu (k+1) i = f in Ω i u (k+1) i = λ (k) on Γ u (k+1) i =0 on Ω i \ Γ Lψ (k+1) i =0 inω i ψ (k+1) i = u(k+1) 1 u(k+1) 2 n n n on Γ i =0 on Ω i \ Γ ψ (k+1) λ (k+1) = λ (k) θ ( ) σ 1 ψ (k+1) 1 Γ σ 2 ψ (k+1) 2 Γ where θ is a positive acceleration parameter, while σ 1 and σ 2 are two positive coefficients. 19

20 ROBIN-ROBIN ALGORITHM For every k 0solvethefollowingproblems: then u (k+1) 1 = f in Ω 1 u (k+1) 1 =0 on Ω 1 \ Γ u (k+1) 1 n + γ 1 u (k+1) 1 = u(k) 2 n + γ 1u (k) 2 on Γ Lu (k+1) 2 = f in Ω 2 u (k+1) 2 =0 on Ω 2 \ Γ u (k+1) 2 n + γ 2 u (k+1) 2 = u(k+1) 1 n + γ 2 u (k+1) 1 on Γ where u (0) 2 is assigned and γ 1, γ 2 are non-negative acceleration parameters that satisfy γ 1 + γ 2 > 0. Aiming at parallelization, we could use u (k) 1 instead of u (k+1) 1. 20

21 THE STEKLOV-POINCARÉ INTERFACE EQUATION 21

22 MULTI-DOMAIN FORMULATION OF POISSON PROBLEM AND INTERFACE CONDITIONS We consider now the model problem: { u = f in Ω u =0 on Ω. For a domain partitioned into two disjoint subdomains, wecanwritethe equivalent multi-domain formulation (u i = u Ωi, i =1, 2): u 1 = f in Ω 1 u 1 =0 on Ω 1 \ Γ u 2 = f in Ω 2 u 2 =0 on Ω 2 \ Γ u 1 = u 2 on Γ u 1 n = u 2 n on Γ. 22

23 Remark On the interface Γ we have the normal unit vectors n 1 and n 2 : Ω1 Γ n 1 n 2 Ω 2 There holds: n 1 = n 2 on Γ. We denote n = n 1 so that n = n 1 = n 2 on Γ. 23

24 THE STEKLOV-POINCARÉ OPERATOR Let λ be the unknown value of the solution u on the interface Γ: λ = u Γ Should we know a priori the value λ on Γ, wecouldsolvethefollowing two independent boundary-value problems with Dirichlet condition on Γ (i =1, 2): w i = f in Ω i w i =0 on Ω i \ Γ w i = λ on Γ. 24

25 With the aim of obtaining the value λ on Γ, let us split w i as follows w i = w i + u 0 i, where wi and ui 0 represent the solutions of the following problems (i =1, 2): wi = f in Ω i wi =0 on Ω i Ω wi =0 on Γ and ui 0 = 0 in Ω i ui 0 =0 on Ω i Ω ui 0 = λ on Γ. 25

26 The functions w i depend solely on the source data f w i = G i f where G i is a linear continuous operator u 0 i depend solely on the value λ on Γ u 0 i = H i λ where H i is the so-called harmonic extension operator of λ on the domain Ω i. We have that u i = wi + ui 0 (i =1, 2) w 1 n = w 2 n on Γ n (w 1 + u0 1 )= n (w 2 + u0 2 ) on Γ n (G 1f + H 1 λ)= n (G 2f + H 2 λ) on Γ ( H1 n H ) ( 2 G2 λ = n n G ) 1 f on Γ. n 26

27 We have obtained the Steklov-Poincaré equation for the unknown λ on the interface Γ: Sλ = χ on Γ S is the Steklov-Poincaré pseudo-differential operator: Sµ = n H 1µ n H 2µ = χ is a linear functional which depends on f : The operator χ = n G 2f n G 1f = S i : µ S i µ = 2 i=1 2 i=1 n i H i µ n i G i f. (H i µ) n i i =1, 2, Γ is called local Steklov-Poincaré operator (Dirichlet-to-Neumann) which operates between the trace space and its dual Λ. Λ={µ : v V s.t. µ = v Γ } = H 1/2 00 (Γ) 27

28 Example To provide an example of the operator S, weconsiderasimple1d problem. Let Ω = (a, b) R as illustrated below. We split Ω in two nonoverlapping subdomains. In this case the interface Γreducestothepointγ (a, b) andthesteklov-poincaréoperators becomes ( dh1 Sλ = dx dh ) ( 2 1 λ = + 1 ) λ dx l 1 l 2 where l 1 = γ a et l 2 = b γ. λ H 1 λ H 2 λ a l 1 l 2 Ω 1 γ Ω 2 b 28

29 EQUIVALENCE BETWEEN THE DD SCHEMES AND CLASSICAL ITERATIVE METHODS 29

30 The DN method can be reinterpreted as a preconditioned Richardson method for the solution of the Steklov-Poincaré interface equation: P DN (λ (k) λ (k 1) )=θ(χ Sλ (k 1) ) The preconditioning operator is P DN = S 2 = (H 2 µ)/ n 2. The NN method can also be interpreted as a preconditioned Richardson algorithm P NN (λ (k) λ (k 1) )=θ(χ Sλ (k 1) ) with P NN =(σ 1 S σ 2 S 1 2 ) 1. 30

31 The Robin-Robin algorithm is equivalent to the following alternating direction (ADI) algorithm: (γ 1 I Λ + S 1 )µ (k) 1 = χ +(γ 1 I Λ + S 2 )µ (k 1) 2, (γ 2 I Λ + S 2 )µ (k) 2 = χ +(γ 2 I Λ + S 1 )µ (k 1) 1, where I Λ :Λ Λ here denotes the Riesz isomorphism between the Hilbert space Λ and its dual Λ. At convergence (for a convenient choice of γ 1 and γ 2 ), we have µ 1 = µ 2 = λ. 31

32 FINITE ELEMENT APPROXIMATION: MULTI-DOMAIN FORMULATION 32

33 Consider the Poisson problem: { u = f in Ω u =0 on Ω Its weak formulation reads find u V : a(u, v) =F (v) v V where V = H0 1(Ω), a(v, w) = v w v, w V and F (v) = fv v V. Ω Ω 33

34 Suppose that Ω is split into two nonoverlapping subdomains and consider a uniform triangulation T h of Ω, conforming on Γ: Ω 1 Γ Ω 2 34

35 The Galerkin finite element approximation of the Poisson problem reads: where find u h V h : a(u h, v h )=F (v h ) v h V h (1) V h = {v h C 0 (Ω) : v h K P r r 1, K T h, v h =0on Ω} is the space of finite element functions of degree r with basis {ϕ i } N h i=1. The Galerkin approximation (1) is equivalent to: find u h V h : a(u h,ϕ i )=F (ϕ i ) i =1,...,N h. (2) 35

36 We partition the nodes of the triangulation as follows: {x (1) j, 1 j N 1 } nodes in subdomain Ω 1 {x (2) j, 1 j N 2 } nodes in subdomain Ω 2 {x (Γ) j, 1 j N Γ } nodes on the interface Γ, and we split the basis functions accordingly: ϕ (1) j functions associated to the nodes x (1) j ϕ (2) j functions associated to the nodes x (2) j ϕ (Γ) j functions associated to the nodes x (Γ) j on the interface. 36

37 Example Ω 1 x (1) j Γ x (Γ) j Ω 2 x (2) j 37

38 Problem (2) can be equivalently rewritten as: find u h V h : a(u h,ϕ (1) i )=F (ϕ (1) i ) i =1,...,N 1 a(u h,ϕ (2) j )=F(ϕ (2) j ) j =1,...,N 2 a(u h,ϕ (Γ) k )=F(ϕ(Γ) k ) k =1,...,N Γ. We introduce the bilinear form on Ω i : a i (v, w) = v w v, w V, i =1, 2, Ω i the space V i h = {v H 1 (Ω i ):v =0on Ω i \ Γ} (i =1, 2) and let u (i) h = u h Ωi (i =1, 2). Then, problem (3) can be written equivalently as: find u (1) h V 1 h, u(2) h V 2 h s.t. a 1 (u (1) h,ϕ(1) i )=F 1 (ϕ (1) i ) i =1,...,N 1 a 2 (u (2) h,ϕ(2) j )=F 2 (ϕ (2) j ) j =1,...,N 2 a 1 (u (1) h,ϕ(γ) k Ω 1 )+a 2 (u (2) h,ϕ(γ) k Ω 2 ) = F 1 (ϕ (Γ) k Ω 1 )+F 2 (ϕ (Γ) k Ω 2 ) k =1,...,N Γ. (3) (4) 38

39 Remark Problem (4) corresponds to the finite element approximation of the multi-domain formulation of the Poisson problem: u 1 = f in Ω 1 u 1 =0 on Ω 1 \ Γ u 2 = f in Ω 2 u 2 =0 on Ω 2 \ Γ u 1 = u 2 on Γ u 1 n = u 2 on Γ. n The condition u 1 = u 2 on Γ is satisfied by definition of u (i) h. 39

40 We can write: u h (x) = N 1 j=1 u h (x (1) j )ϕ (1) j N Γ + u h (x (Γ) j j=1 N 2 (x)+ u h (x (2) )ϕ (Γ) j (x), j=1 j )ϕ (2) j (x) and we substitute this expression in (4) to obtain... 40

41 N 1 j=1 N 2 j=1 N Γ j=1 u h (x (1) j )a 1(ϕ (1) j,ϕ (1) u h (x (2) j )a 2(ϕ (2) j,ϕ (2) [ u h (x (Γ) j ) a 1(ϕ (Γ) j N Γ i )+ j=1 N Γ i )+,ϕ (Γ) i N 1 + u h (x (1) j )a 1(ϕ (1) j j=1 = F 1(ϕ (Γ) i j=1 u h (x (Γ) j )a 1(ϕ (Γ) j,ϕ (1) i )=F 1(ϕ (1) i ) i =1,...,N 1 u h (x (Γ) j )a 2(ϕ (Γ) j,ϕ (2) i )=F 2(ϕ (2) i ) i =1,...,N 2 ] )+a 2(ϕ (Γ) j,ϕ (Γ) i ) N 2 i )+ u h (x (2) j )a 2(ϕ (2) j,ϕ (Γ) i ),ϕ (Γ) j=1 Ω1 )+F 2(ϕ (Γ) i Ω2 ) i =1,...,N Γ. 41

42 A bit of algebra... N 1 j=1 N 2 j=1 N Γ j=1 N Γ u h (x (1) j )(A 11) ij + j=1 N Γ u h (x (2) j )(A 22) ij + [ u h (x (Γ) j ) j=1 j=1 (A (1) ΓΓ )ij + (A(2) u h (x (Γ) j )(A 1Γ ) ij = F 1(ϕ (1) i ) i =1,...,N 1 u h (x (Γ) j )(A 2Γ ) ij = F 2(ϕ (2) i ) i =1,...,N 2 ΓΓ )ij ] N 1 N 2 + u h (x (1) j )(A Γ1 ) ij + u h (x (2) j )(A Γ2 ) ij = F 1(ϕ (Γ) i j=1 Ω1 )+F 2(ϕ (Γ) i Ω2 ) i =1,...,N Γ. 42

43 A bit of algebra... [continued] N 1 N Γ u 1(A 11) ij + u Γ (A 1Γ ) ij = f 1 i =1,...,N 1 j=1 j=1 N 2 j=1 N Γ j=1 N Γ u 2(A 22) ij + u Γ (A 2Γ ) ij = f 2 i =1,...,N 2 u Γ [ j=1 ] (A (1) ΓΓ )ij +(A(2) ΓΓ )ij N 1 N 2 + u 1(A Γ1 ) ij + u 2(A Γ2 ) ij j=1 j=1 = f (Γ) 1 + f (Γ) 2 i =1,...,N Γ. 43

44 ... so that we obtain the algebraic form: A 11 u 1 + A 1Γ λ = f 1 A 22 u 2 + A 2Γ λ = f( 2 ) A Γ1 u 1 + A Γ2 u 2 + A (1) ΓΓ + A(2) ΓΓ λ = f (Γ) 1 + f (Γ) 2 (5) or A 11 0 A 1Γ 0 A 22 A 2Γ A Γ1 A Γ2 A ΓΓ where we denoted A ΓΓ = u 1 u 2 λ = f 1 f 2 f Γ ( ) A (1) ΓΓ + A(2) ΓΓ and f Γ = f (Γ) 1 + f (Γ) 2. (6) 44

45 THE SCHUR COMPLEMENT SYSTEM 45

46 Since λ represents the unknown value of u on Γ, its finite element correspondent is the vector λ of the values of u h at the interface nodes. By Gaussian elimination, we can obtain a new reduced system in the sole unknown λ: Matrices A 11 and A 22 are invertible since they are associated with two homogeneous Dirichlet boundary-value problems for the Laplace operator: u 1 = A 1 11 (f 1 A 1Γ λ) and u 2 = A 1 22 (f 2 A 2Γ λ). (7) From the third equation we obtain: [( ) ( )] A (1) ΓΓ A Γ1A 1 11 A 1Γ + A (2) ΓΓ A Γ2A 1 22 A 2Γ λ = f Γ A Γ1 A 1 11 f 1 A Γ2 A 1 22 f 2. 46

47 Setting and Σ=Σ 1 +Σ 2 with Σ i = A (i) ΓΓ A ΓiA 1 ii A iγ (i =1, 2) χ Γ = f Γ A Γ1 A 1 11 f 1 A Γ2 A 1 22 f 2 we obtain the Schur complement system Σλ = χ Γ Σ and χ Γ approximate S and χ. Σistheso-calledSchur complement of A with respect to u 1 and u 2. Σ i are the Schur complements related to the subdomains Ω i (i =1, 2). 47

48 Remark After solving the Schur complement system in λ, thanks to(7)we can compute u 1 and u 2. This is equivalent to solving two Poisson problems in the subdomains Ω 1 and Ω 2 with the Dirichlet boundary condition u (i) h Γ = λ h (i =1, 2) on the interface Γ. 48

49 PROPERTIES OF THE SCHUR COMPLEMENT Σ The Schur complement Σ inherits some of the properties of A: if A is singular, so is Σ; if A (respectively, A ii )issymmetric,thenσ(respectively,σ i )is symmetric too; if A is positive definite, so is Σ. Moreover, concerning the condition number, wehave κ(a) Ch 2 κ(σ) Ch 1 49

50 PRECONDITIONERS FOR THE SCHUR COMPLEMENT SYSTEM 50

51 The iterative methods that we have illustrated are equivalent to preconditioned Richardson methods for the Schur complement system with preconditioners: for the DN algorithm: P h =Σ 2 for the ND algorithm: P h =Σ 1 for the NN algorithm: P h =(σ 1 Σ σ 2 Σ 1 2 ) 1 for the RR algorithm: P h =(γ 1 + γ 2 ) 1 (γ 1 I +Σ 1 )(γ 2 I +Σ 2 ) 51

52 All these preconditioners are optimal in the sense of the following definition. Definition ApreconditionerPisoptimal for a matrix A R N N if the condition number of P 1 Aisboundeduniformly with respect to the dimension N of A. In particular, we have κ(σ 1 i Σ) = O(1) (i =1, 2) κ((σ 1 Σ σ 2 Σ 1 2 )Σ) = O(1) for all σ 1,σ 2 > 0 52

53 NONOVERLAPPING METHODS IN THE CASE OF MORE THAN TWO SUBDOMAINS 53

54 MULTI-DOMAIN FORMULATION FOR M > 2 SUBDOMAINS We generalize now the nonoverlapping methods to the case of a domain ΩsplitintoM > 2subdomains: 1 Ω i (i =1,...,M) suchthat Ω i = Ω Γ i = Ω i \ Ω Γ= Γ i. Plotmesh Ω Γ ik1 Γ ik2 Y Axis Γ ik4 Ω i Γ ik X Axis 54

55 At the differential level, we have the equivalent multi-domain formulation: { u = f in Ω u =0 on Ω u i = f in Ω i (i =1,...,M) u i = u k on Γ ik u i = u k n i n i on Γ ik u i =0 on Ω i Ω where Γ ik = Ω i Ω k. 55

56 FINITE ELEMENT APPROXIMATION Considering a conforming finite element approximation, weobtain the linear system: ( )( ) ( ) AII A I Γ ui fi = (8) A ΓI A ΓΓ λ f Γ where u I is the vector of unknowns in the internal nodes and λ is the vector of unknowns on Γ: λ = u Γ. The submatrix A II associated to the internal nodes is block-diagonal: A A II = A MM A I Γ is a banded matrix (interactions with local interfaces). 56

57 Remark On each subdomain Ω i the matrix ( ) Aii A A i = iγ A Γi A Γi Γ i represents the local stiffness matrix associated to a Neumann problem. 57

58 THE SCHUR COMPLEMENT SYSTEM Since A II is non-singular, from (8) we can derive By eliminating u I from (8), we get that is Denoting and u I = A 1 II (f I A I Γ u Γ ) (9) A ΓΓ λ = f Γ A ΓI A 1 II (f I A I Γ λ) ( AΓΓ A ΓI A 1 ) II A I Γ λ = fγ A ΓI A 1 II f I. (10) Σ=A ΓΓ A ΓI A 1 II A I Γ χ Γ = f Γ A ΓI A 1 II f I we obtain the Schur complement system in the multi-domain case Σλ = χ Γ (11) 58

59 Example 0 Γ 1 Γ 4 Γ 2 Γ Γ 2 Γ c Γ 1 Γ c Γ 3 Γ nz = 597 On the left, decomposition of Ω = (0, 1) (0, 1) into four square subdomains. On the right, sparsity pattern of the Schur complement matrix arisingfromthe decomposition depicted on the left. Remark The local Schur complements are defined as so that Σ i = A Γi Γ i A Γi ia 1 ii A iγi Σ=R T Γ 1 Σ 1 R Γ1 +...R T Γ M Σ M R ΓM where R Γi is the restriction from the skeleton Γ to the boundary of Ω i. 59

60 ASIMPLEALGORITHM To compute a FE approximation of the solution u of the Poisson problem { u = f in Ω u =0 on Ω we can do the following steps: 1 solve the Schur complement system Σλ = χ Γ to compute λ on the whole interface Γ; 2 solve A II u I = f I A I Γ λ i.e., M independent problems of reduced dimension A ii u i I = g i (i =1,...,M) possibly in parallel. 60

61 ESTIMATE OF THE CONDITION NUMBER The following estimate can be proved for the condition number of the Schur complement matrix Σ: there exists a constant C > 0, independent of h and H such that κ(σ) C H max hh 2 min where H max and H min are the maximal and minimal diameters of the subdomains, respectively. 61

62 NUMERICAL RESULTS We consider the model problem { u = f in Ω = (0, 1) (0, 1) u =0 on Ω We decompose Ω into M square regions Ω i of characteristic dimension H such that M i=1 Ω i = Ω: Γ 4 Γ 2 Γ 3 Γ c Γ 1 (Decomposition of Ω = (0, 1) (0, 1) into four square subdomains) 62

63 Condition number of Σ: κ(σ) H =1/2 H =1/4 H =1/8 h=1/ h=1/ h=1/ h=1/ Notice the dependence on 1 h and 1 H. 63

64 The Schur complement Σ is dense. If we use iterative solvers to solve the Schur complement system Σλ = χ Γ it is not convenient from the point of view of memory usage to explicitly compute the elements of Σ. We need only to compute the product Σx Γ for any given vector x Γ. 64

65 For example, in the conjugate gradient method: a) Choose a starting solution λ 0 b) Compute r 0 = χ Γ Σλ 0, p 0 = r 0 For k 0untilconvergence,Do c) α k = (p k, r k ) (Σp k, p k ) d) λ k+1 = λ k + α k p k e) r k+1 = r k α k Σp k f) β k = (Σp k, r k+1 ) (Σp k, p k ) g) p k+1 = r k+1 β k p k EndFor 65

66 SCHUR COMPLEMENT MULTIPLICATION BY A VECTOR We can define the restriction and prolongation operators R Γi and R T Γ i that act on the vector of the interface nodal values. Recalling that Σ=Σ Σ M with Σ i = A Γi Γ i A Γi ia 1 ii A iγi we can use the following fully parallel algorithm. Given x Γ,computey Γ =Σx Γ as follows: a) Set y Γ = 0 For i =1,...,M Do in parallel: b) x i = R Γi x Γ c) z i = A iγi x i d) z i A 1 ii z i e) sum to the global vector y Γi A Γi Γ i x Γ A Γi iz i f) sum to the global vector y Γ RΓ T i y Γi EndFor 66

67 Before using for the first time the Schur complement matrix, an off-line startup phase is required: For i =1,...,M Do in parallel: a) Built the matrix A i b) Reorder A i as ( ) Aii A A i = iγi A Γi i A Γi Γ i and extract the submatrices A ii, A iγi, A Γi i and A Γi Γ i c) Compute the LU or Cholesky factorization of A ii EndFor 67

68 SCALABILITY Definition A preconditioner P h of Σ is said to be scalable if the condition number of the preconditioned matrix P 1 h Σ is independent of the number of subdomains. Iterative methods using scalable preconditioners allow henceforth to achieve convergence rates independent of the subdomain number. Remark General strategy: P 1 h = M RΓ T i P 1 i,h R Γ i + RΓ T P 1 H i=1 where P i,h is a suitable preconditioner for the local Schur complement Σ i.thelasttermisthecoarse-scalecorrection. R Γ 68

69 NEUMANN-NEUMANN PRECONDITIONER The Neumann-Neumann preconditioner for more subdomains reads: where Σ i is either Σ 1 i (P NN h ) 1 = M RΓ T i D i Σ i D i R Γi i=1 or an approximation of Σ 1 i. D i is a diagonal matrix of positive weights d 1 D i =... d n d j is the number of subdomains that share the j-th node. We have the following estimate: κ ( (Ph NN ) 1 Σ ) ( CH 2 1+log H ) 2 h 69

70 Remark The presence of D i and R Γi only entails matrix-matrix multiplications. Applying Σ 1 i may be done using local inverses: if q is a vector whose components are the nodal values on Γ i,then Σ 1 i q =[0, I ]A 1 [0, I ] T q In particular, the matrix-vector multiplication internal nodes } boundary nodes {{ } A 1 i i q corresponds to the solution of a Neumann problem in Ω i : { wi =0 inω i w i n = q in Γ i 70

71 Example Condition number of (Ph NN ) 1 Σ: κ((ph NN ) 1 Σ) H =1/2 H =1/4 H =1/8 H =1/16 h =1/ h =1/ h =1/ h =1/ Convergence history of the conjugate gradient using Ph NN preconditioner and h = 1/32: as a 10 0 Convergence history using P NN scaled residual H=1/4 H=1/ H=1/ Conjugate gradient iterations 71

72 BALANCED NEUMANN-NEUMANN PRECONDITIONER The Neumann-Neumann preconditioner of the Schur complement system is not scalable. Asubstantialimprovementcanbeachievedbyaddingacoarsegrid correction: (Ph BNN ) 1 Σ=P 0 +(I P 0 ) ( (PH NN ) 1 Σ ) (I P 0 ) where P 0 = R T 0 Σ 1 0 R 0ΣandΣ 0 = R 0 ΣR T 0, R 0 being the restriction from Γ onto the coarse level skeleton. This is called balanced Neumann-Neumann preconditioner. We can prove that κ ( (Ph BNN ) 1 Σ ) ( C 1+log H ) 2 h 72

73 Example κ((ph BNN ) 1 Σ) H =1/2 H =1/4 H =1/8 H =1/16 h =1/ h =1/ h =1/ h =1/

74 CONCLUDING REMARKS From the numerical results that we have presented, we can conclude with the following remarks: even if better conditioned with respect to A, Σisill-conditioned,and therefore a suitable preconditioner must be applied the Neumann-Neumann preconditioner can be satisfactory applied using a moderate number of subdomains, while for larger M, κ((ph NN ) 1 Σ) >κ(σ) the balancing Neumann-Neumann preconditioner is almost optimally scalable and therefore recommended for partitions with a large number of subdomains 74

75 FETI (Finite Element Tearing & Interconnecting) METHODS 75

76 THE MODEL PROBLEM We consider the variable coefficient elliptic problem: { div(ρ u) = f in Ω, u =0 on Ω, (12) where ρ is piecewise constant, ρ = ρ i R + in Ω i. We denote by H i = diam(ω i )thesizeofthei-th subdomain Ω i for i =1,...,M. 76

77 Domain Ω, its boundary Ω, skeleton Γ, subdomain Ω i,anditsboundary Ω i. 77

78 THE FINITE ELEMENT SPACES We introduce: the space of traces of finite element functions on the boundaries Ω i,sayw i = W h ( Ω i ), the product space of the previous trace spaces, say W =Π M i=1 W i, a subspace of continuous traces across the skeleton Γ, say Ŵ W. We denote by Ω ih the nodes in Ω i,by Ω ih the nodes on Ω i,by Ω h the nodes on Ω, and by Γ h the nodes on the skeleton Γ. 78

79 Finite element degrees of freedom on the boundary ( Ω h ), on the skeleton (Γ h ), in the subdomain (Ω ih ), and on its boundary ( Ω ih ). 79

80 THE SCALING COUNTING FUNCTIONS Let us introduce the following scaling counting functions: 1 x Ω ih ( Ω h \ Γ h ), ( ) δ i (x) = ργ j Nx j (x) /ρ γ i (x) x Ω ih Γ h, 0 elsewhere, x Γ h Ω h,where:γ [1/2, + ) andn x is the set of indices of the subregions having x on their boundaries. Then we set: δ i (x) (= pseudo inverses) = { δ 1 i (x) if δ i (x) 0, 0 if δ i (x) =0. 80

81 THE FINITE ELEMENT APPROXIMATION Based on the finite element approximation of the model problem (12), let us consider the local Schur complements: for which: Σ i = A (i) ΓΓ A Γ,Ω i A 1 Ω i,ω i A Ωi,Γ, i =1,... M, Σ=R T Γ 1 Σ 1 R Γ R T Γ M Σ M R ΓM, where R Γi is a restriction operator, that is a rectangular matrix of zeros and ones that map values on Γ onto those on Γ i, i =1,...,M. The matrices Σ i are positive semi-definite. We indicate the interface nodal values on Ω i as u i,andweset u =(u 1,...,u M ), the local load vectors on Ω i as χ i and we set χ =(χ 1,...,χ M ). Finally, we set: ablockdiagonalmatrix. Σ = diag(σ 1,...,Σ M ), 81

82 THE REDUCED FEM PROBLEM The original FEM problem, when reduced to the interface Γ,reads: Find u W s.t. J(u) = 1 2 Σ u, u χ, u min, B Γ u = 0. (13) The reduced problem (13) admits a unique solution iff Ker{Σ } Ker{B Γ } =0,thatisifΣ is invertible on Ker(B Γ ). The matrix B Γ is not unique, sothatweshouldimposecontinuity when u belongs to more than one subdomain; B Γ is made of {0, 1, 1} since it enforces interfaces continuity constraints at interfaces nodes. Remark We are using the same notation W to denote the finite element space trace and that of their nodal values at points of Γ h. 82

83 THE CHOICE OF THE MATRIX B Γ In 2D, there is a little choice on how to write the constraint of continuity at a point sitting on an edge, there are many options for a vertex point. For the edge node we only need to choose the sign, whereas for a vertex node, e.g. one common to 4 subdomains, a minimum set of three constraints can be chosen in many different ways to assure continuity at the node in question. Continuity constraints enforced by 3(non-redundant)conditions on the left, by 6(redundant)conditionson the right. 83

84 THE REDUCED FEM PROBLEM WITH LAGRANGE MULTIPLIERS We reformulate the reduced problem (13) by introducing suitable Lagrange multipliers: { Find (u, λ) W U s.t. Σ u + B T Γ λ = χ, B Γ u = 0. (14) Because of the inf-sup (LBB) condition, thecomponentλ of the solution to (14) is unique up to an additive vector of Ker(BΓ T ), so we choose U = range(b Γ ). 84

85 Let the matrix R = diag ( R (1),...,R (M)) be made of null-space elements of Σ. (E.g. R (i) corresponds to the rigid body motions of Ω i,incaseof linear elasticity operator.) R is a full column rank matrix. The solution of the first equation of the reduced problem (14) exists iff χ B T Γ λ range(σ ), a limitation that will be resolved by introducing asuitableprojectionoperatorp. Then, u =Σ (χ BT Γ λ)+rα if χ BT Γ λ Ker(Σ ), where α is an arbitrary vector and Σ apseudoinverseofσ. 85

86 Substituting u into the second equation of the reduced problem (14) yields: B Γ Σ BT Γ λ = B ΓΣ χ + B Γ Rα. (15) Let us set F = B Γ Σ BT Γ and d = B ΓΣ χ. Then choose P T to be a suitable projection matrix, e.g. P T = I G(G T G) 1 G T,withG = B Γ R. We obtain: { P T F λ = P T d, G T λ = e (= R T χ ). 86

87 ONE-LEVEL FETI METHOD The original one-level FETI method is a CG method in the space V applied to: P T F λ = P T d, λ λ 0 + V, with an initial λ 0 s.t. G T λ 0 = e. Here V = {λ U : λ, B Γ z =0, z Ker(Σ )} is the so-called space of admissible increments, Ker(G T )=range(p) and: V = {µ U : µ, B Γ z Q =0, z Ker(Σ )} = range(p T ). 87

88 FETI PRECONDITIONERS The original, most basic FETI preconditioner is: P 1 h = B Γ Σ B T Γ = M i=1 B(i) Σ i B (i)t. It is called a Dirichlet preconditioner since its application to a given vector involves the solution of M independent Dirichlet problems, one in every subdomain. The coarse space in FETI consists of the nullspace on each substructure. 88

89 In case B Γ has full row rank (i.e. the constraints are linearly independent and there are no redundant Lagrange multipliers), a better preconditioner can be defined as follows: P 1 h =(B Γ D 1 B T Γ ) 1 B Γ D 1 Σ D 1 B T Γ (B ΓD 1 B T Γ ) 1, where D is ( a block diagonal matrix with positive entries D = diag D (1),...,D (M)) with D (i) adiagonalmatrixwhose elements are δ i (x) correspondingtothepointx Ω ih (recall the scaling counting functions). 89

90 Remark Since B Γ D 1 BΓ T is block-diagonal, itsinversecanbeeasily computed by inverting small blocks whose size is n x,thenumberof Lagrange multipliers used to enforce continuity at point x. The matrix D, thatoperatesonelementsoftheproductspacew, can be regarded as a scaling from the right of B Γ by D 1/2.With this choice: 1 K 2 (P P h PT F ) C(1 + log(h/h)) 2, where K 2 ( ) isthespectral condition number and C is a constant independent of h, H, γ and the values of the ρ i. 90

91 FETI-DP (Dual Primal FETI) METHODS 91

92 OVERVIEW OF THE FETI-DP METHODS The FETI-DP method is a domain decomposition method that enforces equality of the solution at subdomains interfaces by Lagrange multipliers except at subdomains corners, which remain primal variables. The first mathematical analysis of the method was provided by Mandel and Tezaur (2001). The method was further improved by enforcing the equality of averages across the edges or faces on subdomain interfaces. This is important for parallel scalability. 92

93 FINITE ELEMENT SPACES We consider a 2D case for simplicity. We introduce a space W s.t. Ŵ W W for which we have continuity of the primal variables at subdomain vertices. We write W as: where: W = ŴΠ W, Ŵ Π Ŵ is the space of continuous interface functions that vanish at all nodal points of Γ h except at the subdomain vertices. ŴΠ is given in terms of the vertex variables and the averages of the values over the individual edges of the set of interface nodes Γ h. W is the direct sum of local subspaces W,i : W = M i=1 W,i, where W,i W i consists of local functions on Ω i that vanish at the vertices of Ω i and have zero average on each individual edge. 93

94 FINITE ELEMENT SPACES According to this space splitting: the continuous degrees of freedom associated with the subdomain vertices and with the subspace ŴΠ are called primal (Π); those (that are potentially discontinuous across Γ) that are associated with the subspaces W,i and with the interior of the subdomain edges are called dual ( ). The subspace ŴΠ, togetherwiththeinteriorsubspace,definesthe subsystem which is fully assembled, factored, and stored in each iteration step. 94

95 THE SCHUR COMPLEMENT All unknowns of the subspace ŴΠ as well as the interior variables are eliminated to obtain a new Schur complement Σ. We consider the following procedure. Let à denote the stiffness matrix obtained by restricting diag (A 1,...,A M )from M i=1 W h (Ω i )to W h (Ω) (these spaces now refer to subdomains, not their boundaries). Then à is no longer block diagonal because of the coupling that now exists between subdomains that share a common vertex. 95

96 According to the previous space decomposition, Ã can be split as: A II A I Π A I Ã = A T I Π A ΠΠ A Π. A T I AT Π A The subscript I refers to the internal degrees of freedom of the subdomains, Π to those associated to the subdomains vertices, and to those of the interior of the subdomains edges. The matrices A II and A are block diagonal (one block per subdomain). Any non-zero entry of A I represents a coupling between degrees of freedom associated with the same subdomain. 96

97 Degrees of freedom of the space W for one-level FETI (left) and those of the space W for one-level FETI-DP (right) in the case of primal vertices only. 97

98 Upon eliminating the variables of the I and Π sets, a Schur complement associated with the variables of the sets (interior and edges) is obtained as follows: Σ =A [ A T I A T Π ] [ ] 1 [ ] A II A I Π AI A T. I Π A ΠΠ A Π Correspondingly we obtain a reduced right hand side χ. 98

99 THE REDUCED FEM PROBLEM By indicating with u W the vector of degrees of freedom associated with the edges, similarly to what done in (13) for FETI, the finite element problem can be reformulated as a reduced minimization problem with constraints given by the requirement of continuity across all of Γ: Find u W : J(u ) = 1 2 Σu, u χ, u min, B u = 0. The matrix B is made of {0, 1, 1} as it was for B Γ. Remark The constraints associated with the vertex nodes are dropped since they are assigned to the primal set. Remark Since all the constraints refer to edge points, no distinction needs to be made between redundant and non-redundant constraints and Lagrange multipliers. (16) 99

100 THE REDUCED FEM PROBLEM WITH LANGRANGE MULTIPLIERS A saddle point formulation of (16), similar to (14), can be obtained by introducing a set of Lagrange multipliers λ V = range(b ). Indeed, since à is s.p.d., so is Σ: by eliminating the subvectors u we obtain the reduced system: F λ = d, (17) where F = B Σ 1 B T and d = B Σ 1 χ. Remark Once λ is found, u = Σ 1 ( χ B T λ) W,whiletheinterior variables u I and the vertex variables u Π are obtained by back-solving the system associated with Ã. 100

101 AFETI-DPPRECONDITIONER A preconditioner for F is introduced as done for FETI (in case of non-redundant Lagrange multipliers): P 1 =(B D 1 BT ) 1 B D 1 S D 1 BT (B D 1 BT ) 1. D is a diagonal scaling matrix with blocks D (i) :eachoftheir diagonal elements corresponds to a Lagrange multiplier that enforces continuity between the nodal values of some w i W i and w j W j at some point x Γ h and it is given by δ j (x); S = diag (Σ 1,,...,Σ M, ) with Σ i, being the restriction of the local Schur complement Σ i to W,i W i. 101

102 Remark When using the conjugate gradient method for the preconditioned system: P 1 F λ = P 1 d, in contrast with one level FETI methods we can use an arbitrary initial guess λ 0. We have a condition number that scales polylogarithmically, that is: K 2 (P 1 F ) C(1 + log(h/h)) 2, where C is independent of h, H,γ and the values of the ρ i. 102

103 COMPARISON OF FETI AND FETI-DP METHODS FETI-DP algorithms do not require the characterization of the kernels of local Neumann problems (as required by one-level FETI methods), because the enforcement of the additional constraints in each iteration always makes the local problems nonsingular and at the same time provides an underlying coarse global problem. FETI-DP methods do not require the introduction of a scaling matrix Q, whichentersintheconstructionofacoarsesolverfor one-level FETI algorithms. One-level FETI methods are projected conjugate gradient algorithms that cannot start from an arbitrary initial guess. In contrast, FETI-DP methods are standard preconditioned conjugate algorithms and can therefore employ an arbitrary initial guess λ

104 BDDC (Balancing Domain Decomposition with Constraints) METHODS 104

105 AQUICKVIEWOFBDDCMETHODS The BDDC method was introduced by Dohrmann (2003) as a simpler primal alternative to the FETI-DP domain decomposition method. The name BDDC was coined by Mandel and Dohrmann because it can be understood as further development of the balancing domain decomposition method. BDDC is used as a preconditioner for the conjugate gradient method. 105

106 AspecificversionofBDDCischaracterizedbythechoiceofcoarse degrees of freedom, whichcanbevaluesatthecornersofthe subdomains, or averages over the edges of the interface between the subdomains. One application of the BDDC preconditioner combines the solution of local problems on each subdomain with the solution of a global coarse problem with the coarse degrees of freedom as the unknowns. The local problems on different subdomains are completely independent of each other, so the method is suitable for parallel computing. 106

107 GENERAL FORM OF BDDC PRECONDITIONERS ABDDCpreconditionerreads: P 1 BDDC = R T DΓ Σ 1 RDΓ, where R Γ : Ŵ W is a restriction matrix, R DΓ is a scaled variant of R Γ with scale factor δ i (featuring the same sparsity pattern of R Γ ). This scaling is chosen in such a way that R Γ R T DΓ is a projection (then it coincides with its square). 107

108 ALGEBRAIC FORM OF SCHWARZ ITERATIVE METHODS 108

109 ALGEBRAIC FORM OF SCHWARZ METHODS FOR FE DISCRETIZATION Let Au = f be the system associated to the finite element approximation of the Poisson problem. We still assume that Ω is decomposed in two overlapping subdomains Ω 1 and Ω 2. We denote by N h the total number of interior nodes of Ω, and by n 1 and n 2 the interior nodes of Ω 1 and Ω 2 : N h n 1 + n 2 The stiffness matrix A contains two submatrices, saya 1 and A 2, which correspond to the local stiffness matrices associated to the Dirichlet problems in Ω 1 and Ω 2 : n 1 A 1 A N h A 2 n 2 109

110 MULTIPLICATIVE SCHWARZ METHOD Consider the multiplicative method: givenu (0) 2 on Γ 1,fork 1: Lu (k) 1 = f in Ω 1 u (k) 1 = u (k 1) 2 on Γ 1 u (k) 1 =0 on Ω 1 \ Γ 1 ( ) ( ) u (k) AΩ1 A Ω 1 Γ1 u (k) = f 1 Γ 1 u (k) Γ 1 = u (k 1) Ω 2 Γ 1 Lu (k) 2 = f in Ω 2 u (k) 2 = u (k) 1 on Γ 2 u (k) 2 =0 on Ω 2 \ Γ 2 ( ) ( ) u (k) AΩ2 A Ω 2 Γ2 u (k) = f 2 Γ 2 u (k) Γ 2 = u (k) Ω 1 Γ 2 Ω 2 Ω 1 Γ 2 Γ 1 110

111 We find: A Ω1 u (k) Ω 1 = f 1 A Γ1 u (k 1) Γ 1 A Ω2 u (k) Ω 2 = f 2 A Γ2 u (k) Γ 2 or, equivalently, by adding and subtracting u (k 1) Ω i : u (k) Ω 1 u (k) Ω 2 = u (k 1) Ω 1 = u (k 1) Ω 2 + A 1 Ω 1 (f 1 A Ω1 u (k 1) Ω 1 A Γ1 u (k 1) Γ 1 ) + A 1 Ω 2 (f 2 A Ω2 u (k 1) Ω 2 A Γ2 u (k) Γ 2 ) Remark This method can be seen as a block Gauss-Seidel method to solve the system ( )( ) ( ) AΩ1 A Γ1 uω1 f1 = A Ω2 u Ω2 f 2 A Γ2 111

112 If there is no direct coupling between the matrices assembled on opposite sides of Γ 1 and Γ 2 : A Ω2\Ω 1 =0 and A Ω1\Ω 2 =0 Ω 2 Ω 1 Γ 2 Γ 1 then, formally, we can write so that A Γ1 = A (Ω\Ω1)\(Ω 2\Ω 1) = A (Ω\Ω 1) A Γ1 u (k 1) Γ 1 = A (Ω\Ω1)u (k 1) (Ω\Ω 1) Proceeding in a similar way also for the second equation, we finally obtain: u (k) Ω 1 u (k) Ω 2 = u (k 1) Ω 1 = u (k 1) Ω 2 + A 1 Ω 1 (f 1 A Ω1 u (k 1) Ω 1 A (Ω\Ω1)u (k 1) (Ω\Ω ) 1) + A 1 Ω 2 (f 2 A Ω2 u (k 1) Ω 2 A (Ω\Ω2)u (k) (Ω\Ω ) 2) 112

113 In compact form we can write ( ) u (k+ 1 2 A ) = u (k) 1 + Ω 1 0 (f ) Au (k) 0 0 ( ) 0 0 ( u (k+1) = u (k+ 1 2 ) + 0 A 1 f Au (k+ 1 Ω 2 Let R i be the (rectangular) restriction matrix that returns the vector of coefficients in the interior of Ω i : ( ) uωi u Ωi = R i u =(I 0) u Ω\Ω1 Then, we obtain: u (k+ 1 2 ) = u (k) + R T 1 (R 1AR T 1 ) 1 R 1 ( f Au (k) ) 2 ) ) u (k+1) = u (k+ 1 2 ) + R T 2 (R 2AR T 2 ) 1 R 2 ( f Au (k+ 1 2 ) ) 113

114 Define, for i =1, 2 Q i = R T i (R i AR T i ) 1 R i We get u (k+ 1 2 ) = u (k) + Q 1 ( f Au (k)) u (k+1) = u (k+ 1 2 ) + Q 2 ( f Au (k+ 1 2 )) and ( u (k+1) = u (k) +Q 1 f Au (k)) +Q 2 [f A (u (k) + Q 1 (f Au (k)))] or, equivalently, u (k+1) = u (k) +(Q 1 + Q 2 Q 2 AQ 1 ) ( f Au (k)) 114

115 ADDITIVE SCHWARZ METHOD Proceeding as before, we can prove that the additive Schwarz method can be seen as a block Jacobi method to solve the system ( )( ) ( ) AΩ1 A Γ1 uω1 f1 = A Ω2 u Ω2 f 2 A Γ2 One iteration of the additive Schwarz method can be written as u (k+1) = u (k) +(Q 1 + Q 2 )(f Au (k) ) In the case of a M 2overlappingsubdomains{Ω i } we have: ( M ) u (k+1) = u (k) + Q i (f Au (k) ) i=1 115

116 THE SCHWARZ METHOD AS A PRECONDITIONER Multiplicative Schwarz method: can be interpreted as a Richardson iterative procedure for Au = f with the multiplicative Schwarz preconditioner P 1 ms = Q 1 + Q 2 Q 2 AQ 1 P ms is not symmetric and can be used within GMRES or BiCGStab iterations. Additive Schwarz method: corresponds to Richardson iterations for Au = f with the additive Schwarz preconditioner ( M ) 1 P as = Q i i=1 116

117 The convergence rate for the additive Schwarz method is slower than for the multiplicative Schwarz method. In the case of two subdomains, there is approximately a factor 1 2. (This is similar to the classical convergence results for the Jacobi and Gauss-Seidel methods) P as is not optimal as the condition number blows up if the size of the subdomains reduces: κ(p 1 as A) C 1 δh where C is a constant independent of h, H, andδ, δ being a linear measure of the overlapping region(s) (h δ H). 117

118 ALGORITHMIC ASPECTS & NUMERICAL RESULTS Create overlapping subdomains a) Triangulate the computational domain Ω T h. b) Partition the mesh into M nonoverlapping subdomains {ˆΩ i } M i=1. c) Extend every subdomain ˆΩ i by adding all the strips of finite elements of T h within distance δ from ˆΩ i. ˆΩ 1 ˆΩ 4 ˆΩ 7 ˆΩ 2 ˆΩ 5 ˆΩ 8 Ω 5 ˆΩ 3 ˆΩ 6 ˆΩ 9 Start-up phase for the application of P as a) On each subdomain Ω i build R i and Ri T. b) Build the matrix A corresponding to the FE discretization on T h. c) On each Ω i form the local submatrices A Ωi = R i ARi T. d) On each Ω i prepare the code to solve a linear system with A Ωi : compute a factorization (LU, ILU, Cholesky...) of A Ωi. 118

119 One-level additive Schwarz preconditioner P as Given a vector r, computez = Pas 1 r as follows a) set z = 0; For i =1,...,M, doinparallel b) restrict the residual over Ω i \ Ω i : r i = R i r; c) solve A Ωi z i = r i ; d) add to the global residual: z z + Ri T z i. EndFor Example κ(pas 1 A) H =1/2 H =1/4 H =1/8 H =1/16 h =1/ h =1/ h =1/ h =1/

120 The main limitation of Schwarz methods is to propagate information only among neighboring subdomains, sinceonly local solves are involved by the application of (P as ) 1. We have to introduce a coarse global problem over the whole domain to guarantee a mechanism of global communication among all subdomains. 120

121 TWO-LEVEL SCHWARZ PRECONDITIONERS As for the Neumann-Neumann method, we can introduce a coarse-grid mechanism that allows for a sudden information diffusion on the whole domain Ω: consider the subdomains as macro-elements of a new coarse grid T H build a corresponding stiffness matrix A H The matrix Q 0 = RH T A 1 H R H represents the coarse level correction for the two-level preconditioner, with R H the restriction operator from the fine to the coarse grid. The two-level preconditioner P cas is defined as: P 1 cas = M i=0 Q i 121

122 We can prove that there exists a constant C > 0, independent of both h and H such that ( κ(pcas 1 A) C 1+ H ) δ If δ is a fraction of H (generous overlap), the preconditioner P cas is scalable and optimal. If δ h, thepreconditioner P cas is neither scalable or optimal. Iterations on the original finite element system using P cas converge with a rate independent of h and H (and therefore of the number of subdomains) Thanks to the additive structure of the preconditioner, the preconditioning step is fully parallel as it involves the solution of independent systems, one per each local matrix A Ωi. 122

123 ALGORITHMIC ASPECTS & NUMERICAL RESULTS Start-up phase to apply P cas a) Do the start-up phase for P as. b) Define a coarse space triangulation T 0 with elements of order H and n 0 dofs: c) Build the restriction operator R H R n0 N h, R H (i, j) =Φ i (x j ). Φ i is the basis function associated to the coarse grid node i and x j are the coordinates of the fine grid node j. d) Construct the coarse matrix A H,eitheras A H (i, j) = Φ i Φ j or A H = R H ARH T Ω 123

124 Two-levels additive Schwarz preconditioner P cas Given a vector r, computez = Pcas 1 r as follows a) set z = 0; For i =1,...,M, doinparallel b) restrict the residual over Ω i \ Ω i : r i = R i r; c) solve A Ωi z i = r i ; d) add to the global residual: z z + Ri T z i. EndFor e) Compute the coarse-grid contribution z H = A 1 H (R Hr); f) Add to the global residual z z + RH T z H. Example κ(pcas 1 A) H =1/4 H =1/8 H =1/16 h =1/ h =1/ h =1/ h =1/ If H/δ = constant, thistwo-levelpreconditioneriseitheroptimaland scalable. 124

125 Practical indications: For decompositions with a small number of subdomains, the single level Schwarz preconditioner P as is very efficient. When the number M of subdomains gets large, using two-level preconditioners becomes crucial. When generating a coarse grid is difficult, other algebraic techniques, likeaggregation, canbeadoptedtodefinea H,and ( κ(paggre 1 A) C 1+ H ) δ Paggre 1 A H =1/4 H =1/8 H =1/16 h =1/ h =1/ h =1/ h =1/

126 REFERENCES 126

127 References Barry Smith, Petter Bjorstad, William Gropp. Domain Decomposition, Parallel Multilevel Methods for Elliptic Partial Differential Equations. CambridgeUniversityPress,1996. Alfio Quarteroni, Alberto Valli. Domain Decomposition Methods for Partial Differential Equations. OxfordSciencePublications,1999. Barbara I. Wholmuth. Discretization Methods and Iterative Solvers Based on Domain Decomposition. Springer,2001. Andrea Toselli, Olof Widlund. Domain Decomposition Methods Algorithms and Theory. Springer,2004. Clemens Pechstein. Finite and Boundary Element Tearing and Interconnecting Solvers for Multiscale problems. Springer,2013. Alfio Quarteroni. Numerical Models for Differential Problems. Springer, Official page of Domain Decomposition Methods 127

Multispace and Multilevel BDDC. Jan Mandel University of Colorado at Denver and Health Sciences Center

Multispace and Multilevel BDDC. Jan Mandel University of Colorado at Denver and Health Sciences Center Multispace and Multilevel BDDC Jan Mandel University of Colorado at Denver and Health Sciences Center Based on joint work with Bedřich Sousedík, UCDHSC and Czech Technical University, and Clark R. Dohrmann,

More information

A Balancing Algorithm for Mortar Methods

A Balancing Algorithm for Mortar Methods A Balancing Algorithm for Mortar Methods Dan Stefanica Baruch College, City University of New York, NY 11, USA Dan Stefanica@baruch.cuny.edu Summary. The balancing methods are hybrid nonoverlapping Schwarz

More information

Short title: Total FETI. Corresponding author: Zdenek Dostal, VŠB-Technical University of Ostrava, 17 listopadu 15, CZ Ostrava, Czech Republic

Short title: Total FETI. Corresponding author: Zdenek Dostal, VŠB-Technical University of Ostrava, 17 listopadu 15, CZ Ostrava, Czech Republic Short title: Total FETI Corresponding author: Zdenek Dostal, VŠB-Technical University of Ostrava, 17 listopadu 15, CZ-70833 Ostrava, Czech Republic mail: zdenek.dostal@vsb.cz fax +420 596 919 597 phone

More information

Multilevel and Adaptive Iterative Substructuring Methods. Jan Mandel University of Colorado Denver

Multilevel and Adaptive Iterative Substructuring Methods. Jan Mandel University of Colorado Denver Multilevel and Adaptive Iterative Substructuring Methods Jan Mandel University of Colorado Denver The multilevel BDDC method is joint work with Bedřich Sousedík, Czech Technical University, and Clark Dohrmann,

More information

A Balancing Algorithm for Mortar Methods

A Balancing Algorithm for Mortar Methods A Balancing Algorithm for Mortar Methods Dan Stefanica Baruch College, City University of New York, NY, USA. Dan_Stefanica@baruch.cuny.edu Summary. The balancing methods are hybrid nonoverlapping Schwarz

More information

CONVERGENCE ANALYSIS OF A BALANCING DOMAIN DECOMPOSITION METHOD FOR SOLVING A CLASS OF INDEFINITE LINEAR SYSTEMS

CONVERGENCE ANALYSIS OF A BALANCING DOMAIN DECOMPOSITION METHOD FOR SOLVING A CLASS OF INDEFINITE LINEAR SYSTEMS CONVERGENCE ANALYSIS OF A BALANCING DOMAIN DECOMPOSITION METHOD FOR SOLVING A CLASS OF INDEFINITE LINEAR SYSTEMS JING LI AND XUEMIN TU Abstract A variant of balancing domain decomposition method by constraints

More information

ASM-BDDC Preconditioners with variable polynomial degree for CG- and DG-SEM

ASM-BDDC Preconditioners with variable polynomial degree for CG- and DG-SEM ASM-BDDC Preconditioners with variable polynomial degree for CG- and DG-SEM C. Canuto 1, L. F. Pavarino 2, and A. B. Pieri 3 1 Introduction Discontinuous Galerkin (DG) methods for partial differential

More information

Convergence analysis of a balancing domain decomposition method for solving a class of indefinite linear systems

Convergence analysis of a balancing domain decomposition method for solving a class of indefinite linear systems NUMERICAL LINEAR ALGEBRA WITH APPLICATIONS Numer. Linear Algebra Appl. 2000; 00:1 6 [Version: 2002/09/18 v1.02] Convergence analysis of a balancing domain decomposition method for solving a class of indefinite

More information

Adaptive Coarse Space Selection in BDDC and FETI-DP Iterative Substructuring Methods: Towards Fast and Robust Solvers

Adaptive Coarse Space Selection in BDDC and FETI-DP Iterative Substructuring Methods: Towards Fast and Robust Solvers Adaptive Coarse Space Selection in BDDC and FETI-DP Iterative Substructuring Methods: Towards Fast and Robust Solvers Jan Mandel University of Colorado at Denver Bedřich Sousedík Czech Technical University

More information

Dirichlet-Neumann and Neumann-Neumann Methods

Dirichlet-Neumann and Neumann-Neumann Methods Dirichlet-Neumann and Neumann-Neumann Methods Felix Kwok Hong Kong Baptist University Introductory Domain Decomposition Short Course DD25, Memorial University of Newfoundland July 22, 2018 Outline Methods

More information

Extending the theory for domain decomposition algorithms to less regular subdomains

Extending the theory for domain decomposition algorithms to less regular subdomains Extending the theory for domain decomposition algorithms to less regular subdomains Olof Widlund Courant Institute of Mathematical Sciences New York University http://www.cs.nyu.edu/cs/faculty/widlund/

More information

20. A Dual-Primal FETI Method for solving Stokes/Navier-Stokes Equations

20. A Dual-Primal FETI Method for solving Stokes/Navier-Stokes Equations Fourteenth International Conference on Domain Decomposition Methods Editors: Ismael Herrera, David E. Keyes, Olof B. Widlund, Robert Yates c 23 DDM.org 2. A Dual-Primal FEI Method for solving Stokes/Navier-Stokes

More information

IsogEometric Tearing and Interconnecting

IsogEometric Tearing and Interconnecting IsogEometric Tearing and Interconnecting Christoph Hofer and Ludwig Mitter Johannes Kepler University, Linz 26.01.2017 Doctoral Program Computational Mathematics Numerical Analysis and Symbolic Computation

More information

Selecting Constraints in Dual-Primal FETI Methods for Elasticity in Three Dimensions

Selecting Constraints in Dual-Primal FETI Methods for Elasticity in Three Dimensions Selecting Constraints in Dual-Primal FETI Methods for Elasticity in Three Dimensions Axel Klawonn 1 and Olof B. Widlund 2 1 Universität Duisburg-Essen, Campus Essen, Fachbereich Mathematik, (http://www.uni-essen.de/ingmath/axel.klawonn/)

More information

Domain Decomposition Preconditioners for Spectral Nédélec Elements in Two and Three Dimensions

Domain Decomposition Preconditioners for Spectral Nédélec Elements in Two and Three Dimensions Domain Decomposition Preconditioners for Spectral Nédélec Elements in Two and Three Dimensions Bernhard Hientzsch Courant Institute of Mathematical Sciences, New York University, 51 Mercer Street, New

More information

On the Use of Inexact Subdomain Solvers for BDDC Algorithms

On the Use of Inexact Subdomain Solvers for BDDC Algorithms On the Use of Inexact Subdomain Solvers for BDDC Algorithms Jing Li a, and Olof B. Widlund b,1 a Department of Mathematical Sciences, Kent State University, Kent, OH, 44242-0001 b Courant Institute of

More information

Multispace and Multilevel BDDC

Multispace and Multilevel BDDC Multispace and Multilevel BDDC Jan Mandel Bedřich Sousedík Clark R. Dohrmann February 11, 2018 arxiv:0712.3977v2 [math.na] 21 Jan 2008 Abstract BDDC method is the most advanced method from the Balancing

More information

18. Balancing Neumann-Neumann for (In)Compressible Linear Elasticity and (Generalized) Stokes Parallel Implementation

18. Balancing Neumann-Neumann for (In)Compressible Linear Elasticity and (Generalized) Stokes Parallel Implementation Fourteenth nternational Conference on Domain Decomposition Methods Editors: smael Herrera, David E Keyes, Olof B Widlund, Robert Yates c 23 DDMorg 18 Balancing Neumann-Neumann for (n)compressible Linear

More information

Domain Decomposition Algorithms for an Indefinite Hypersingular Integral Equation in Three Dimensions

Domain Decomposition Algorithms for an Indefinite Hypersingular Integral Equation in Three Dimensions Domain Decomposition Algorithms for an Indefinite Hypersingular Integral Equation in Three Dimensions Ernst P. Stephan 1, Matthias Maischak 2, and Thanh Tran 3 1 Institut für Angewandte Mathematik, Leibniz

More information

A FETI-DP Method for Mortar Finite Element Discretization of a Fourth Order Problem

A FETI-DP Method for Mortar Finite Element Discretization of a Fourth Order Problem A FETI-DP Method for Mortar Finite Element Discretization of a Fourth Order Problem Leszek Marcinkowski 1 and Nina Dokeva 2 1 Department of Mathematics, Warsaw University, Banacha 2, 02 097 Warszawa, Poland,

More information

Convergence Behavior of a Two-Level Optimized Schwarz Preconditioner

Convergence Behavior of a Two-Level Optimized Schwarz Preconditioner Convergence Behavior of a Two-Level Optimized Schwarz Preconditioner Olivier Dubois 1 and Martin J. Gander 2 1 IMA, University of Minnesota, 207 Church St. SE, Minneapolis, MN 55455 dubois@ima.umn.edu

More information

Optimal Interface Conditions for an Arbitrary Decomposition into Subdomains

Optimal Interface Conditions for an Arbitrary Decomposition into Subdomains Optimal Interface Conditions for an Arbitrary Decomposition into Subdomains Martin J. Gander and Felix Kwok Section de mathématiques, Université de Genève, Geneva CH-1211, Switzerland, Martin.Gander@unige.ch;

More information

Overlapping Schwarz preconditioners for Fekete spectral elements

Overlapping Schwarz preconditioners for Fekete spectral elements Overlapping Schwarz preconditioners for Fekete spectral elements R. Pasquetti 1, L. F. Pavarino 2, F. Rapetti 1, and E. Zampieri 2 1 Laboratoire J.-A. Dieudonné, CNRS & Université de Nice et Sophia-Antipolis,

More information

TR THREE-LEVEL BDDC IN THREE DIMENSIONS

TR THREE-LEVEL BDDC IN THREE DIMENSIONS R2005-862 REE-LEVEL BDDC IN REE DIMENSIONS XUEMIN U Abstract BDDC methods are nonoverlapping iterative substructuring domain decomposition methods for the solution of large sparse linear algebraic systems

More information

The All-floating BETI Method: Numerical Results

The All-floating BETI Method: Numerical Results The All-floating BETI Method: Numerical Results Günther Of Institute of Computational Mathematics, Graz University of Technology, Steyrergasse 30, A-8010 Graz, Austria, of@tugraz.at Summary. The all-floating

More information

Dual-Primal Isogeometric Tearing and Interconnecting Solvers for Continuous and Discontinuous Galerkin IgA Equations

Dual-Primal Isogeometric Tearing and Interconnecting Solvers for Continuous and Discontinuous Galerkin IgA Equations Dual-Primal Isogeometric Tearing and Interconnecting Solvers for Continuous and Discontinuous Galerkin IgA Equations Christoph Hofer and Ulrich Langer Doctoral Program Computational Mathematics Numerical

More information

OVERLAPPING SCHWARZ ALGORITHMS FOR ALMOST INCOMPRESSIBLE LINEAR ELASTICITY TR

OVERLAPPING SCHWARZ ALGORITHMS FOR ALMOST INCOMPRESSIBLE LINEAR ELASTICITY TR OVERLAPPING SCHWARZ ALGORITHMS FOR ALMOST INCOMPRESSIBLE LINEAR ELASTICITY MINGCHAO CAI, LUCA F. PAVARINO, AND OLOF B. WIDLUND TR2014-969 Abstract. Low order finite element discretizations of the linear

More information

AN INTRODUCTION TO DOMAIN DECOMPOSITION METHODS. Gérard MEURANT CEA

AN INTRODUCTION TO DOMAIN DECOMPOSITION METHODS. Gérard MEURANT CEA Marrakech Jan 2003 AN INTRODUCTION TO DOMAIN DECOMPOSITION METHODS Gérard MEURANT CEA Introduction Domain decomposition is a divide and conquer technique Natural framework to introduce parallelism in the

More information

A Robust Preconditioner for the Hessian System in Elliptic Optimal Control Problems

A Robust Preconditioner for the Hessian System in Elliptic Optimal Control Problems A Robust Preconditioner for the Hessian System in Elliptic Optimal Control Problems Etereldes Gonçalves 1, Tarek P. Mathew 1, Markus Sarkis 1,2, and Christian E. Schaerer 1 1 Instituto de Matemática Pura

More information

Parallel scalability of a FETI DP mortar method for problems with discontinuous coefficients

Parallel scalability of a FETI DP mortar method for problems with discontinuous coefficients Parallel scalability of a FETI DP mortar method for problems with discontinuous coefficients Nina Dokeva and Wlodek Proskurowski University of Southern California, Department of Mathematics Los Angeles,

More information

arxiv: v1 [math.na] 28 Feb 2008

arxiv: v1 [math.na] 28 Feb 2008 BDDC by a frontal solver and the stress computation in a hip joint replacement arxiv:0802.4295v1 [math.na] 28 Feb 2008 Jakub Šístek a, Jaroslav Novotný b, Jan Mandel c, Marta Čertíková a, Pavel Burda a

More information

INSTITUTE OF MATHEMATICS THE CZECH ACADEMY OF SCIENCES. A virtual overlapping Schwarz method for scalar elliptic problems in two dimensions

INSTITUTE OF MATHEMATICS THE CZECH ACADEMY OF SCIENCES. A virtual overlapping Schwarz method for scalar elliptic problems in two dimensions INSTITUTE OF MATHEMATICS THE CZECH ACADEMY OF SCIENCES A virtual overlapping Schwarz method for scalar elliptic problems in two dimensions Juan Gabriel Calvo Preprint No. 25-2017 PRAHA 2017 A VIRTUAL

More information

Scientific Computing WS 2017/2018. Lecture 18. Jürgen Fuhrmann Lecture 18 Slide 1

Scientific Computing WS 2017/2018. Lecture 18. Jürgen Fuhrmann Lecture 18 Slide 1 Scientific Computing WS 2017/2018 Lecture 18 Jürgen Fuhrmann juergen.fuhrmann@wias-berlin.de Lecture 18 Slide 1 Lecture 18 Slide 2 Weak formulation of homogeneous Dirichlet problem Search u H0 1 (Ω) (here,

More information

OUTLINE ffl CFD: elliptic pde's! Ax = b ffl Basic iterative methods ffl Krylov subspace methods ffl Preconditioning techniques: Iterative methods ILU

OUTLINE ffl CFD: elliptic pde's! Ax = b ffl Basic iterative methods ffl Krylov subspace methods ffl Preconditioning techniques: Iterative methods ILU Preconditioning Techniques for Solving Large Sparse Linear Systems Arnold Reusken Institut für Geometrie und Praktische Mathematik RWTH-Aachen OUTLINE ffl CFD: elliptic pde's! Ax = b ffl Basic iterative

More information

Construction of a New Domain Decomposition Method for the Stokes Equations

Construction of a New Domain Decomposition Method for the Stokes Equations Construction of a New Domain Decomposition Method for the Stokes Equations Frédéric Nataf 1 and Gerd Rapin 2 1 CMAP, CNRS; UMR7641, Ecole Polytechnique, 91128 Palaiseau Cedex, France 2 Math. Dep., NAM,

More information

On Nonlinear Dirichlet Neumann Algorithms for Jumping Nonlinearities

On Nonlinear Dirichlet Neumann Algorithms for Jumping Nonlinearities On Nonlinear Dirichlet Neumann Algorithms for Jumping Nonlinearities Heiko Berninger, Ralf Kornhuber, and Oliver Sander FU Berlin, FB Mathematik und Informatik (http://www.math.fu-berlin.de/rd/we-02/numerik/)

More information

Simple Examples on Rectangular Domains

Simple Examples on Rectangular Domains 84 Chapter 5 Simple Examples on Rectangular Domains In this chapter we consider simple elliptic boundary value problems in rectangular domains in R 2 or R 3 ; our prototype example is the Poisson equation

More information

BETI for acoustic and electromagnetic scattering

BETI for acoustic and electromagnetic scattering BETI for acoustic and electromagnetic scattering O. Steinbach, M. Windisch Institut für Numerische Mathematik Technische Universität Graz Oberwolfach 18. Februar 2010 FWF-Project: Data-sparse Boundary

More information

Scalable BETI for Variational Inequalities

Scalable BETI for Variational Inequalities Scalable BETI for Variational Inequalities Jiří Bouchala, Zdeněk Dostál and Marie Sadowská Department of Applied Mathematics, Faculty of Electrical Engineering and Computer Science, VŠB-Technical University

More information

33 RASHO: A Restricted Additive Schwarz Preconditioner with Harmonic Overlap

33 RASHO: A Restricted Additive Schwarz Preconditioner with Harmonic Overlap Thirteenth International Conference on Domain Decomposition ethods Editors: N. Debit,.Garbey, R. Hoppe, J. Périaux, D. Keyes, Y. Kuznetsov c 001 DD.org 33 RASHO: A Restricted Additive Schwarz Preconditioner

More information

From Direct to Iterative Substructuring: some Parallel Experiences in 2 and 3D

From Direct to Iterative Substructuring: some Parallel Experiences in 2 and 3D From Direct to Iterative Substructuring: some Parallel Experiences in 2 and 3D Luc Giraud N7-IRIT, Toulouse MUMPS Day October 24, 2006, ENS-INRIA, Lyon, France Outline 1 General Framework 2 The direct

More information

Additive Average Schwarz Method for a Crouzeix-Raviart Finite Volume Element Discretization of Elliptic Problems

Additive Average Schwarz Method for a Crouzeix-Raviart Finite Volume Element Discretization of Elliptic Problems Additive Average Schwarz Method for a Crouzeix-Raviart Finite Volume Element Discretization of Elliptic Problems Atle Loneland 1, Leszek Marcinkowski 2, and Talal Rahman 3 1 Introduction In this paper

More information

Two new enriched multiscale coarse spaces for the Additive Average Schwarz method

Two new enriched multiscale coarse spaces for the Additive Average Schwarz method 346 Two new enriched multiscale coarse spaces for the Additive Average Schwarz method Leszek Marcinkowski 1 and Talal Rahman 2 1 Introduction We propose additive Schwarz methods with spectrally enriched

More information

Substructuring for multiscale problems

Substructuring for multiscale problems Substructuring for multiscale problems Clemens Pechstein Johannes Kepler University Linz (A) jointly with Rob Scheichl Marcus Sarkis Clark Dohrmann DD 21, Rennes, June 2012 Outline 1 Introduction 2 Weighted

More information

Schwarz Preconditioner for the Stochastic Finite Element Method

Schwarz Preconditioner for the Stochastic Finite Element Method Schwarz Preconditioner for the Stochastic Finite Element Method Waad Subber 1 and Sébastien Loisel 2 Preprint submitted to DD22 conference 1 Introduction The intrusive polynomial chaos approach for uncertainty

More information

cfl Jing Li All rights reserved, 22

cfl Jing Li All rights reserved, 22 NYU-CS-TR83 Dual-Primal FETI Methods for Stationary Stokes and Navier-Stokes Equations by Jing Li A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy

More information

Parallel Scalability of a FETI DP Mortar Method for Problems with Discontinuous Coefficients

Parallel Scalability of a FETI DP Mortar Method for Problems with Discontinuous Coefficients Parallel Scalability of a FETI DP Mortar Method for Problems with Discontinuous Coefficients Nina Dokeva and Wlodek Proskurowski Department of Mathematics, University of Southern California, Los Angeles,

More information

An additive average Schwarz method for the plate bending problem

An additive average Schwarz method for the plate bending problem J. Numer. Math., Vol. 10, No. 2, pp. 109 125 (2002) c VSP 2002 Prepared using jnm.sty [Version: 02.02.2002 v1.2] An additive average Schwarz method for the plate bending problem X. Feng and T. Rahman Abstract

More information

Stabilization and Acceleration of Algebraic Multigrid Method

Stabilization and Acceleration of Algebraic Multigrid Method Stabilization and Acceleration of Algebraic Multigrid Method Recursive Projection Algorithm A. Jemcov J.P. Maruszewski Fluent Inc. October 24, 2006 Outline 1 Need for Algorithm Stabilization and Acceleration

More information

CLASSICAL ITERATIVE METHODS

CLASSICAL ITERATIVE METHODS CLASSICAL ITERATIVE METHODS LONG CHEN In this notes we discuss classic iterative methods on solving the linear operator equation (1) Au = f, posed on a finite dimensional Hilbert space V = R N equipped

More information

ETNA Kent State University

ETNA Kent State University Electronic Transactions on Numerical Analysis. Volume 11, pp. 1-24, 2000. Copyright 2000,. ISSN 1068-9613. ETNA NEUMANN NEUMANN METHODS FOR VECTOR FIELD PROBLEMS ANDREA TOSELLI Abstract. In this paper,

More information

Scalable Domain Decomposition Preconditioners For Heterogeneous Elliptic Problems

Scalable Domain Decomposition Preconditioners For Heterogeneous Elliptic Problems Scalable Domain Decomposition Preconditioners For Heterogeneous Elliptic Problems Pierre Jolivet, F. Hecht, F. Nataf, C. Prud homme Laboratoire Jacques-Louis Lions Laboratoire Jean Kuntzmann INRIA Rocquencourt

More information

From the Boundary Element Domain Decomposition Methods to Local Trefftz Finite Element Methods on Polyhedral Meshes

From the Boundary Element Domain Decomposition Methods to Local Trefftz Finite Element Methods on Polyhedral Meshes From the Boundary Element Domain Decomposition Methods to Local Trefftz Finite Element Methods on Polyhedral Meshes Dylan Copeland 1, Ulrich Langer 2, and David Pusch 3 1 Institute of Computational Mathematics,

More information

Auxiliary space multigrid method for elliptic problems with highly varying coefficients

Auxiliary space multigrid method for elliptic problems with highly varying coefficients Auxiliary space multigrid method for elliptic problems with highly varying coefficients Johannes Kraus 1 and Maria Lymbery 2 1 Introduction The robust preconditioning of linear systems of algebraic equations

More information

Lecture 9 Approximations of Laplace s Equation, Finite Element Method. Mathématiques appliquées (MATH0504-1) B. Dewals, C.

Lecture 9 Approximations of Laplace s Equation, Finite Element Method. Mathématiques appliquées (MATH0504-1) B. Dewals, C. Lecture 9 Approximations of Laplace s Equation, Finite Element Method Mathématiques appliquées (MATH54-1) B. Dewals, C. Geuzaine V1.2 23/11/218 1 Learning objectives of this lecture Apply the finite difference

More information

An Iterative Substructuring Method for Mortar Nonconforming Discretization of a Fourth-Order Elliptic Problem in two dimensions

An Iterative Substructuring Method for Mortar Nonconforming Discretization of a Fourth-Order Elliptic Problem in two dimensions An Iterative Substructuring Method for Mortar Nonconforming Discretization of a Fourth-Order Elliptic Problem in two dimensions Leszek Marcinkowski Department of Mathematics, Warsaw University, Banacha

More information

Optimal multilevel preconditioning of strongly anisotropic problems.part II: non-conforming FEM. p. 1/36

Optimal multilevel preconditioning of strongly anisotropic problems.part II: non-conforming FEM. p. 1/36 Optimal multilevel preconditioning of strongly anisotropic problems. Part II: non-conforming FEM. Svetozar Margenov margenov@parallel.bas.bg Institute for Parallel Processing, Bulgarian Academy of Sciences,

More information

Preface to the Second Edition. Preface to the First Edition

Preface to the Second Edition. Preface to the First Edition n page v Preface to the Second Edition Preface to the First Edition xiii xvii 1 Background in Linear Algebra 1 1.1 Matrices................................. 1 1.2 Square Matrices and Eigenvalues....................

More information

Efficient smoothers for all-at-once multigrid methods for Poisson and Stokes control problems

Efficient smoothers for all-at-once multigrid methods for Poisson and Stokes control problems Efficient smoothers for all-at-once multigrid methods for Poisson and Stoes control problems Stefan Taacs stefan.taacs@numa.uni-linz.ac.at, WWW home page: http://www.numa.uni-linz.ac.at/~stefant/j3362/

More information

Adaptive algebraic multigrid methods in lattice computations

Adaptive algebraic multigrid methods in lattice computations Adaptive algebraic multigrid methods in lattice computations Karsten Kahl Bergische Universität Wuppertal January 8, 2009 Acknowledgements Matthias Bolten, University of Wuppertal Achi Brandt, Weizmann

More information

GRUPO DE GEOFÍSICA MATEMÁTICA Y COMPUTACIONAL MEMORIA Nº 8

GRUPO DE GEOFÍSICA MATEMÁTICA Y COMPUTACIONAL MEMORIA Nº 8 GRUPO DE GEOFÍSICA MATEMÁTICA Y COMPUTACIONAL MEMORIA Nº 8 MÉXICO 2014 MEMORIAS GGMC INSTITUTO DE GEOFÍSICA UNAM MIEMBROS DEL GGMC Dr. Ismael Herrera Revilla iherrerarevilla@gmail.com Dr. Luis Miguel de

More information

Indefinite and physics-based preconditioning

Indefinite and physics-based preconditioning Indefinite and physics-based preconditioning Jed Brown VAW, ETH Zürich 2009-01-29 Newton iteration Standard form of a nonlinear system F (u) 0 Iteration Solve: Update: J(ũ)u F (ũ) ũ + ũ + u Example (p-bratu)

More information

SOME PRACTICAL ASPECTS OF PARALLEL ADAPTIVE BDDC METHOD

SOME PRACTICAL ASPECTS OF PARALLEL ADAPTIVE BDDC METHOD Conference Applications of Mathematics 2012 in honor of the 60th birthday of Michal Křížek. Institute of Mathematics AS CR, Prague 2012 SOME PRACTICAL ASPECTS OF PARALLEL ADAPTIVE BDDC METHOD Jakub Šístek1,2,

More information

for three dimensional problems are often more complicated than the quite simple constructions that work well for problems in the plane; see [23] for a

for three dimensional problems are often more complicated than the quite simple constructions that work well for problems in the plane; see [23] for a DUAL-PRIMAL FETI METHODS FOR THREE-DIMENSIONAL ELLIPTIC PROBLEMS WITH HETEROGENEOUS COEFFICIENTS AEL KLAWONN Λ, OLOF B. WIDLUND y, AND MAKSYMILIAN DRYJA z Abstract. In this paper, certain iterative substructuring

More information

INSTITUTE OF MATHEMATICS THE CZECH ACADEMY OF SCIENCES

INSTITUTE OF MATHEMATICS THE CZECH ACADEMY OF SCIENCES INSTITUTE OF MATHEMATICS THE CZECH ACADEMY OF SCIENCES On the approximation of a virtual coarse space for domain decomposition methods in two dimensions Juan Gabriel Calvo Preprint No. 31-2017 PRAHA 2017

More information

Algebraic Coarse Spaces for Overlapping Schwarz Preconditioners

Algebraic Coarse Spaces for Overlapping Schwarz Preconditioners Algebraic Coarse Spaces for Overlapping Schwarz Preconditioners 17 th International Conference on Domain Decomposition Methods St. Wolfgang/Strobl, Austria July 3-7, 2006 Clark R. Dohrmann Sandia National

More information

Scientific Computing WS 2018/2019. Lecture 15. Jürgen Fuhrmann Lecture 15 Slide 1

Scientific Computing WS 2018/2019. Lecture 15. Jürgen Fuhrmann Lecture 15 Slide 1 Scientific Computing WS 2018/2019 Lecture 15 Jürgen Fuhrmann juergen.fuhrmann@wias-berlin.de Lecture 15 Slide 1 Lecture 15 Slide 2 Problems with strong formulation Writing the PDE with divergence and gradient

More information

arxiv: v1 [math.na] 11 Jul 2011

arxiv: v1 [math.na] 11 Jul 2011 Multigrid Preconditioner for Nonconforming Discretization of Elliptic Problems with Jump Coefficients arxiv:07.260v [math.na] Jul 20 Blanca Ayuso De Dios, Michael Holst 2, Yunrong Zhu 2, and Ludmil Zikatanov

More information

Inexact Data-Sparse BETI Methods by Ulrich Langer. (joint talk with G. Of, O. Steinbach and W. Zulehner)

Inexact Data-Sparse BETI Methods by Ulrich Langer. (joint talk with G. Of, O. Steinbach and W. Zulehner) Inexact Data-Sparse BETI Methods by Ulrich Langer (joint talk with G. Of, O. Steinbach and W. Zulehner) Radon Institute for Computational and Applied Mathematics Austrian Academy of Sciences http://www.ricam.oeaw.ac.at

More information

Preconditioning of Saddle Point Systems by Substructuring and a Penalty Approach

Preconditioning of Saddle Point Systems by Substructuring and a Penalty Approach Preconditioning of Saddle Point Systems by Substructuring and a Penalty Approach Clark R. Dohrmann 1 Sandia National Laboratories, crdohrm@sandia.gov. Sandia is a multiprogram laboratory operated by Sandia

More information

Hybrid (DG) Methods for the Helmholtz Equation

Hybrid (DG) Methods for the Helmholtz Equation Hybrid (DG) Methods for the Helmholtz Equation Joachim Schöberl Computational Mathematics in Engineering Institute for Analysis and Scientific Computing Vienna University of Technology Contributions by

More information

ISOGEOMETRIC BDDC PRECONDITIONERS WITH DELUXE SCALING TR

ISOGEOMETRIC BDDC PRECONDITIONERS WITH DELUXE SCALING TR ISOGOMTRIC BDDC PRCONDITIONRS WITH DLUX SCALING L. BIRÃO DA VIGA, L.. PAVARINO, S. SCACCHI, O. B. WIDLUND, AND S. ZAMPINI TR2013-955 Abstract. A BDDC (Balancing Domain Decomposition by Constraints) preconditioner

More information

Domain Decomposition Methods for Mortar Finite Elements

Domain Decomposition Methods for Mortar Finite Elements Domain Decomposition Methods for Mortar Finite Elements Dan Stefanica Courant Institute of Mathematical Sciences New York University September 1999 A dissertation in the Department of Mathematics Submitted

More information

Sharp condition number estimates for the symmetric 2-Lagrange multiplier method

Sharp condition number estimates for the symmetric 2-Lagrange multiplier method Sharp condition number estimates for the symmetric -Lagrange multiplier method Stephen W. Drury and Sébastien Loisel Abstract Domain decomposition methods are used to find the numerical solution of large

More information

1. Fast Solvers and Schwarz Preconditioners for Spectral Nédélec Elements for a Model Problem in H(curl)

1. Fast Solvers and Schwarz Preconditioners for Spectral Nédélec Elements for a Model Problem in H(curl) DDM Preprint Editors: editor1, editor2, editor3, editor4 c DDM.org 1. Fast Solvers and Schwarz Preconditioners for Spectral Nédélec Elements for a Model Problem in H(curl) Bernhard Hientzsch 1 1. Introduction.

More information

Capacitance Matrix Method

Capacitance Matrix Method Capacitance Matrix Method Marcus Sarkis New England Numerical Analysis Day at WPI, 2019 Thanks to Maksymilian Dryja, University of Warsaw 1 Outline 1 Timeline 2 Capacitance Matrix Method-CMM 3 Applications

More information

On solving linear systems arising from Shishkin mesh discretizations

On solving linear systems arising from Shishkin mesh discretizations On solving linear systems arising from Shishkin mesh discretizations Petr Tichý Faculty of Mathematics and Physics, Charles University joint work with Carlos Echeverría, Jörg Liesen, and Daniel Szyld October

More information

Acceleration of a Domain Decomposition Method for Advection-Diffusion Problems

Acceleration of a Domain Decomposition Method for Advection-Diffusion Problems Acceleration of a Domain Decomposition Method for Advection-Diffusion Problems Gert Lube 1, Tobias Knopp 2, and Gerd Rapin 2 1 University of Göttingen, Institute of Numerical and Applied Mathematics (http://www.num.math.uni-goettingen.de/lube/)

More information

Preliminary/Qualifying Exam in Numerical Analysis (Math 502a) Spring 2012

Preliminary/Qualifying Exam in Numerical Analysis (Math 502a) Spring 2012 Instructions Preliminary/Qualifying Exam in Numerical Analysis (Math 502a) Spring 2012 The exam consists of four problems, each having multiple parts. You should attempt to solve all four problems. 1.

More information

Incomplete Cholesky preconditioners that exploit the low-rank property

Incomplete Cholesky preconditioners that exploit the low-rank property anapov@ulb.ac.be ; http://homepages.ulb.ac.be/ anapov/ 1 / 35 Incomplete Cholesky preconditioners that exploit the low-rank property (theory and practice) Artem Napov Service de Métrologie Nucléaire, Université

More information

From the Boundary Element DDM to local Trefftz Finite Element Methods on Polyhedral Meshes

From the Boundary Element DDM to local Trefftz Finite Element Methods on Polyhedral Meshes www.oeaw.ac.at From the Boundary Element DDM to local Trefftz Finite Element Methods on Polyhedral Meshes D. Copeland, U. Langer, D. Pusch RICAM-Report 2008-10 www.ricam.oeaw.ac.at From the Boundary Element

More information

Multipréconditionnement adaptatif pour les méthodes de décomposition de domaine. Nicole Spillane (CNRS, CMAP, École Polytechnique)

Multipréconditionnement adaptatif pour les méthodes de décomposition de domaine. Nicole Spillane (CNRS, CMAP, École Polytechnique) Multipréconditionnement adaptatif pour les méthodes de décomposition de domaine Nicole Spillane (CNRS, CMAP, École Polytechnique) C. Bovet (ONERA), P. Gosselet (ENS Cachan), A. Parret Fréaud (SafranTech),

More information

Application of Preconditioned Coupled FETI/BETI Solvers to 2D Magnetic Field Problems

Application of Preconditioned Coupled FETI/BETI Solvers to 2D Magnetic Field Problems Application of Preconditioned Coupled FETI/BETI Solvers to 2D Magnetic Field Problems U. Langer A. Pohoaţǎ O. Steinbach 27th September 2004 Institute of Computational Mathematics Johannes Kepler University

More information

PARTITION OF UNITY FOR THE STOKES PROBLEM ON NONMATCHING GRIDS

PARTITION OF UNITY FOR THE STOKES PROBLEM ON NONMATCHING GRIDS PARTITION OF UNITY FOR THE STOES PROBLEM ON NONMATCHING GRIDS CONSTANTIN BACUTA AND JINCHAO XU Abstract. We consider the Stokes Problem on a plane polygonal domain Ω R 2. We propose a finite element method

More information

ADDITIVE SCHWARZ FOR SCHUR COMPLEMENT 305 the parallel implementation of both preconditioners on distributed memory platforms, and compare their perfo

ADDITIVE SCHWARZ FOR SCHUR COMPLEMENT 305 the parallel implementation of both preconditioners on distributed memory platforms, and compare their perfo 35 Additive Schwarz for the Schur Complement Method Luiz M. Carvalho and Luc Giraud 1 Introduction Domain decomposition methods for solving elliptic boundary problems have been receiving increasing attention

More information

Coupled FETI/BETI for Nonlinear Potential Problems

Coupled FETI/BETI for Nonlinear Potential Problems Coupled FETI/BETI for Nonlinear Potential Problems U. Langer 1 C. Pechstein 1 A. Pohoaţǎ 1 1 Institute of Computational Mathematics Johannes Kepler University Linz {ulanger,pechstein,pohoata}@numa.uni-linz.ac.at

More information

Lecture 18 Classical Iterative Methods

Lecture 18 Classical Iterative Methods Lecture 18 Classical Iterative Methods MIT 18.335J / 6.337J Introduction to Numerical Methods Per-Olof Persson November 14, 2006 1 Iterative Methods for Linear Systems Direct methods for solving Ax = b,

More information

LECTURE # 0 BASIC NOTATIONS AND CONCEPTS IN THE THEORY OF PARTIAL DIFFERENTIAL EQUATIONS (PDES)

LECTURE # 0 BASIC NOTATIONS AND CONCEPTS IN THE THEORY OF PARTIAL DIFFERENTIAL EQUATIONS (PDES) LECTURE # 0 BASIC NOTATIONS AND CONCEPTS IN THE THEORY OF PARTIAL DIFFERENTIAL EQUATIONS (PDES) RAYTCHO LAZAROV 1 Notations and Basic Functional Spaces Scalar function in R d, d 1 will be denoted by u,

More information

FETI-DPH: A DUAL-PRIMAL DOMAIN DECOMPOSITION METHOD FOR ACOUSTIC SCATTERING

FETI-DPH: A DUAL-PRIMAL DOMAIN DECOMPOSITION METHOD FOR ACOUSTIC SCATTERING Journal of Computational Acoustics, c IMACS FETI-DPH: A DUAL-PRIMAL DOMAIN DECOMPOSITION METHOD FOR ACOUSTIC SCATTERING Charbel Farhat, Philip Avery and Radek Tezaur Department of Mechanical Engineering

More information

Two-scale Dirichlet-Neumann preconditioners for boundary refinements

Two-scale Dirichlet-Neumann preconditioners for boundary refinements Two-scale Dirichlet-Neumann preconditioners for boundary refinements Patrice Hauret 1 and Patrick Le Tallec 2 1 Graduate Aeronautical Laboratories, MS 25-45, California Institute of Technology Pasadena,

More information

Parallel Implementation of BDDC for Mixed-Hybrid Formulation of Flow in Porous Media

Parallel Implementation of BDDC for Mixed-Hybrid Formulation of Flow in Porous Media Parallel Implementation of BDDC for Mixed-Hybrid Formulation of Flow in Porous Media Jakub Šístek1 joint work with Jan Březina 2 and Bedřich Sousedík 3 1 Institute of Mathematics of the AS CR Nečas Center

More information

Adaptive Finite Element Methods Lecture Notes Winter Term 2017/18. R. Verfürth. Fakultät für Mathematik, Ruhr-Universität Bochum

Adaptive Finite Element Methods Lecture Notes Winter Term 2017/18. R. Verfürth. Fakultät für Mathematik, Ruhr-Universität Bochum Adaptive Finite Element Methods Lecture Notes Winter Term 2017/18 R. Verfürth Fakultät für Mathematik, Ruhr-Universität Bochum Contents Chapter I. Introduction 7 I.1. Motivation 7 I.2. Sobolev and finite

More information

Algebra C Numerical Linear Algebra Sample Exam Problems

Algebra C Numerical Linear Algebra Sample Exam Problems Algebra C Numerical Linear Algebra Sample Exam Problems Notation. Denote by V a finite-dimensional Hilbert space with inner product (, ) and corresponding norm. The abbreviation SPD is used for symmetric

More information

Parallel Scalable Iterative Substructuring: Robust Exact and Inexact FETI-DP Methods with Applications to Elasticity

Parallel Scalable Iterative Substructuring: Robust Exact and Inexact FETI-DP Methods with Applications to Elasticity Parallel Scalable Iterative Substructuring: Robust Exact and Inexact FETI-DP Methods with Applications to Elasticity Oliver Rheinbach geboren in Hilden Fachbereich Mathematik Universität Duisburg-Essen

More information

Deflated Krylov Iterations in Domain Decomposition Methods

Deflated Krylov Iterations in Domain Decomposition Methods 305 Deflated Krylov Iterations in Domain Decomposition Methods Y.L.Gurieva 1, V.P.Ilin 1,2, and D.V.Perevozkin 1 1 Introduction The goal of this research is an investigation of some advanced versions of

More information

Parallel Sums and Adaptive BDDC Deluxe

Parallel Sums and Adaptive BDDC Deluxe 249 Parallel Sums and Adaptive BDDC Deluxe Olof B. Widlund 1 and Juan G. Calvo 2 1 Introduction There has recently been a considerable activity in developing adaptive methods for the selection of primal

More information

Nonoverlapping Domain Decomposition Methods with Simplified Coarse Spaces for Solving Three-dimensional Elliptic Problems

Nonoverlapping Domain Decomposition Methods with Simplified Coarse Spaces for Solving Three-dimensional Elliptic Problems Nonoverlapping Domain Decomposition Methods with Simplified Coarse Spaces for Solving Three-dimensional Elliptic Problems Qiya Hu 1, Shi Shu 2 and Junxian Wang 3 Abstract In this paper we propose a substructuring

More information

Master Thesis Literature Study Presentation

Master Thesis Literature Study Presentation Master Thesis Literature Study Presentation Delft University of Technology The Faculty of Electrical Engineering, Mathematics and Computer Science January 29, 2010 Plaxis Introduction Plaxis Finite Element

More information

Optimal Left and Right Additive Schwarz Preconditioning for Minimal Residual Methods with Euclidean and Energy Norms

Optimal Left and Right Additive Schwarz Preconditioning for Minimal Residual Methods with Euclidean and Energy Norms Optimal Left and Right Additive Schwarz Preconditioning for Minimal Residual Methods with Euclidean and Energy Norms Marcus Sarkis Worcester Polytechnic Inst., Mass. and IMPA, Rio de Janeiro and Daniel

More information

EFFICIENT NUMERICAL TREATMENT OF HIGH-CONTRAST DIFFUSION PROBLEMS

EFFICIENT NUMERICAL TREATMENT OF HIGH-CONTRAST DIFFUSION PROBLEMS EFFICIENT NUMERICAL TREATMENT OF HIGH-CONTRAST DIFFUSION PROBLEMS A Dissertation Presented to the Faculty of the Department of Mathematics University of Houston In Partial Fulfillment of the Requirements

More information