An A Posteriori Error Estimate for Discontinuous Galerkin Methods

Transcription

1 An A Posteriori Error Estimate for Discontinuous Galerkin Methods Mats G Larson mgl@math.chalmers.se Chalmers Finite Element Center Mats G Larson Chalmers Finite Element Center p.1

2 Outline We present an a posteriori error estimate in a mesh dependent energy norm for a general class of dg methods for elliptic problems. Energy norm a posteriori error estimation for discontinuous Galerkin methods Comput. Methods Appl. Mech. Engrg., Volume 192, Issues 5-6, 2003, Pages Joint work with Roland Becker and Peter Hansbo. Mats G Larson Chalmers Finite Element Center p.2

3 A model problem Find u :Ω R such that σ(u) =f u = g D σ n (u) :=n σ(u) =g N in Ω on Γ D on Γ N where Ω denotes a bounded polygonal domain in R d, d =2or 3, with boundary Γ=Γ D Γ N, and σ(u) is defined by σ(u) = u Mats G Larson Chalmers Finite Element Center p.3

4 The dg method Find U Vsuch that a(u, v) =(f,v) for all v V Here the bilinear form is defined by a(v, w) = K ( v, w) K E ( n v, [w]) E + α E ([v], n w ) E + β E (h 1 E [v], [w]) E with α and β real parameters. Mats G Larson Chalmers Finite Element Center p.4

5 Conservation property Introducing the discrete normal flux σ n (U) βh 1 [U] on K \ Γ, Σ n (U) := σ n (U) βh 1 (U g D ) on K Γ D, g N on K Γ N we obtain the elementwise conservation law f + Σ n (U) =0 K K or Σ n (U) = K K σ n (u) Mats G Larson Chalmers Finite Element Center p.5

6 The energy norm We introduce the following mesh dependent energy norm: v 2 = v 2 K + v 2 K where v 2 K = K K(σ(v), v) K v 2 K = K K(h 1 [v], [v]) K\Γ /2+(h 1 v, v) K ΓD Mats G Larson Chalmers Finite Element Center p.6

7 The a posteriori error estimate ( e 2 c K K ρ 2 K with c independent of h and element indicator ) ρ 2 K = h 2 K f + σ(u) K 2 K + h K Σ n (U) σ n (U) 2 K\Γ D + h 1 K [U] 2 K\Γ N with Σ n (U) defined by Σ n (U) = { σn (U) βh 1 [U] on K \ Γ g N on K Γ N Mats G Larson Chalmers Finite Element Center p.7

8 Remarks Valid for several different dg methods including the symmetric and nonsymmetric formulations. Valid on nonconvex polyhedra in 2D and 3D Valid for general boundary conditions and variable coefficients. Mats G Larson Chalmers Finite Element Center p.8

9 Helmholtz decomposition of e There exists φ H 1 (Ω) and χ H(curl, Ω) such that e = φ + curl χ with φ =0on Γ D and n curl χ =0on Γ N and the stability estimate φ + curl χ c e K holds. Mats G Larson Chalmers Finite Element Center p.9

10 Remarks: Helmholtz decomp In 3D the stronger stability estimate χ [H1 (Ω)] d e K does not hold when Ω is a nonconvex polyhedron, while it holds in the convex case. Helmholtz decomposition used for derivation of a posteriori error estimates for nonconforming methods Dari-Duran-Padra-Vampa 1996, convex, Carstensen-Bartels-Jansche 2002, nonconvex 3D. Mats G Larson Chalmers Finite Element Center p.10

11 Basic idea of proof Using the Helmholtz decomposition of e we get: e 2 K =( e, e) = K K( e, φ) K +( e, curl χ) K residual(u) residual(u) e K ( ) φ + curl χ Thus e K residual(u) Mats G Larson Chalmers Finite Element Center p.11

12 Error representation: A φ) K = K K(σ(e), (σ(e), (φ π 0 φ)) K K K = K K(f + σ(u),φ π 0 φ) K +(σ n (u) σ n (U),φ π 0 φ) K\ΓD = K K(f + σ(u),φ π 0 φ) K +(Σ n (U) σ n (U),φ π 0 φ) K\ΓD π 0 φ is the piecewise constant L 2 -projection of φ. Mats G Larson Chalmers Finite Element Center p.12

13 Error representation: A Note that since Σ n (U) is conservative we have (σ n (u),π 0 φ) K\ΓD =(Σ n (U),π 0 φ) K\ΓD for all K not intersecting Γ D. Mats G Larson Chalmers Finite Element Center p.13

14 Error representation: B curl χ) K = K K( e, (u U, n curl χ) K\ΓN K K = K K(v U, n curl χ) K\ΓN We replaced u by an arbitrary function v V gd = {v H 1 : v = g D on Γ D } Mats G Larson Chalmers Finite Element Center p.14

15 Error representation: A + B e 2 K = K K(f + σ(u),φ π 0 φ) K for all v V gd. = I + II + III +(Σ n (U) σ n (U),φ π 0 φ) K\ΓD +(v U, n curl χ) K\ΓN The a posteriori error estimate now follows from estimates of terms I III. Mats G Larson Chalmers Finite Element Center p.15

16 Estimate of III (v U, n curl χ) K\ΓN v U H 1/2 ( K\Γ N ) n curl χ H 1/2 ( K\Γ N ) c v U H 1/2 ( K\Γ N ) curl χ K c v U H 1/2 ( K\Γ N ) e K We employed the trace inequality n curl χ H 1/2 ( K) c curl χ K Mats G Larson Chalmers Finite Element Center p.16

17 Est. III: trace inequality Trace inequality n curl χ H 1/2 ( K) c curl χ K follows from ) n w H 1/2 ( K) c ( w K + h K w K see Girault-Raviart with w = curl χ. Note that curl χ =0. Mats G Larson Chalmers Finite Element Center p.17

18 Est. III: Technical lemma We finally employ the following technical lemma inf v U 2 v V H 1/2 ( K\Γ N ) c gd K K K K with constant c independent of h. see paper for proof. h 1 K [U] 2 K\Γ N Mats G Larson Chalmers Finite Element Center p.18

19 Sinusoidal hill Consider u = f in Ω, u = 0 on Γ Let Then { f =2π 2 sin(πx) sin(πy) Ω=(0, 1) (0, 1) u =sin(πx) sin(πy) Mats G Larson Chalmers Finite Element Center p.19

20 Sinusoidal hill: Solution Mats G Larson Chalmers Finite Element Center p.20

21 Sinusoidal hill: Final mesh Mats G Larson Chalmers Finite Element Center p.21

22 Sinusoidal hill: Effectivity index Symmetric method Unsymmetric method Effectivity index Refinement level Mats G Larson Chalmers Finite Element Center p.22

23 Peak function Consider u = f in Ω, u = 0 on Γ Chose f such that u = e10 x2 +10 y (1 x) 2 x 2 (1 y) 2 y on the domain Ω=(0, 1) (0, 1). Mats G Larson Chalmers Finite Element Center p.23

24 Peak function: Solution Mats G Larson Chalmers Finite Element Center p.24

25 Peak function: Final mesh Mats G Larson Chalmers Finite Element Center p.25

26 Peak function: Effectivity index Symmetric method Unsymmetric method Effectivity index Refinement level Mats G Larson Chalmers Finite Element Center p.26

27 Discontinuous Galerkin Methods on Overlapping Grids Mats G Larson mgl@math.chalmers.se Chalmers Finite Element Center Mats G Larson Chalmers Finite Element Center p.1

28 FEM on overlapping grids Want to use independent meshes in different regions of the domain for Construction of a global mesh for a complex geometry by using overlapping meshes of elementary parts. Coupling of unstructured and structured meshes. Coupling of boundary fitted meshes to structured or unstructured meshes. Joint work with A. Hansbo and P. Hansbo. Mats G Larson Chalmers Finite Element Center p.2

29 Overlapping grids Ω covered by triangulations T h,1 and T h,2. Γ an artificial internal interface composed of edges from the triangles in one of the meshes T h,1. Mats G Larson Chalmers Finite Element Center p.3

30 Overlapping grids: Example 1 Mats G Larson Chalmers Finite Element Center p.4

31 Overlapping grid: Example 2 Mats G Larson Chalmers Finite Element Center p.5

32 A simple model problem Find u :Ω R d R such that u = f in Ω, u = 0 on Ω. This problem has a unique weak solution u H0 1 for f H 1. Mats G Larson Chalmers Finite Element Center p.6

33 A discontinuous space Let V c,i be continuous piecewise polynomials on T h,i and Ω 1 = T Th,1, Ω 2 =Ω\ Ω 1 The discontinuous space is defined by V = {v : v Ω1 V c,i,v Ω2 = w Ω2,w V c,2 } Mats G Larson Chalmers Finite Element Center p.7

34 DG method Find U Vsuch that a(u, v) =(f,v) for all v V a(u, φ) =( U, φ) Ω1 Ω 2 ( n U, [φ]) Γ ([U], n φ ) Γ +(βh 1 [U], [φ]) Γ, with a one-sided approximation of the normal flux: n v = n v 1 on Γ. Mats G Larson Chalmers Finite Element Center p.8

35 Main results Stability if the penalty parameter β is large enough. Optimal order of convergence in energy and L 2 norm. Valid for higher order elements in 2D and 3D and arbitrary intersecions. Energy norm a posteriori error estimates. Extends to the nonsymmetric case. Mats G Larson Chalmers Finite Element Center p.9

36 Earlier results Extends Becker, Stenberg, and Hansbo, where nonmatching meshes with standard mean value flux was used: weaker mesh assumption allowing for composition of arbitrarily overlapping grids (arbitrarily small parts in Ω 2 ) no ad hoc ("saturation") assumptions for the a posteriori result Mats G Larson Chalmers Finite Element Center p.10

37 Example Let Ω=( 4, 4) ( 4, 4) and f =64 2 x 2 2 y 2, corresponding to u =(x 4) (x +4)(y 4) (y +4). Mats G Larson Chalmers Finite Element Center p.11

38 Example: Linears and Quadratics Mats G Larson Chalmers Finite Element Center p.12

39 Example: Convergence in L 2 Convergence in L 2 norm on a sequence of refined meshes. We obtain second and third order convergence for the linear and quadratic approximations, respectively. 1 0 log(l 2 -error) Linear approximation Quadratic approximation log( h) Mats G Larson Chalmers Finite Element Center p.13

40 Energy norm a posteriori estimate 2 e 2 0,Ω 1 Ω 2 + [e] 2 1/2,h,Γ C i=1 K Ti h The element error indicators ρ K,i are defined by ρ 2 K,i. ρ 2 K,i = h 2 K f + U 2 0,P K + h K [n PK U] 2 0, P K +h 1 K [U] 2 0, P K Γ + Γ j K [U] 2 1/2,Γ j, where P K = K Ω i for K T h i. Mats G Larson Chalmers Finite Element Center p.14

41 Computing the 1/2 norm If the interface is one-dimensional: [U] 2 1/2,Γ j := [U] 2 0,Γ j [U](ξ) [U](η) + Γj 2 dξ dη. ξ η 2 Γ j [U] 2 1/2,Γ j is difficult to compute in general. Mats G Larson Chalmers Finite Element Center p.15

42 A simplified estimate A less sharp but more implementation-friendly variant: θ 2 K,i = h 2 K f + U 2 0,K + h K [n K U] 2 0, K Ω i + ( h K [ n U] j 2 + 0, K j h 1 1 Γ K [U] j 2 0, K j 1 j:γ j K All terms are integrals of single polynomials over the original elements or its edges. Mats G Larson Chalmers Finite Element Center p.16

43 Example: Refined meshes Mats G Larson Chalmers Finite Element Center p.17

44 Example: Refined meshes The meshes in both cases has a tendency to refine more at the interface. This is because the local error is the largest there. The exact solution is u = e (x2 +y 2) (4 x)(4+x)(4 y)(4+y), Mats G Larson Chalmers Finite Element Center p.18