Key words. a posteriori error estimators, quasi-orthogonality, adaptive mesh refinement, error and oscillation reduction estimates, optimal meshes.

Transcription

1 CONVERGENCE OF ADAPTIVE FINITE ELEMENT METHOD FOR GENERAL ECOND ORDER LINEAR ELLIPTIC PDE KHAMRON MEKCHAY AND RICARDO H. NOCHETTO Abstract. We prove convergence of adaptive finite element methods (AFEM) for general (nonsymmetric) second order linear elliptic PDE, thereby extending the result of Morin et al [6, 7]. The proof relies on quasi-orthogonality, which accounts for the bilinear form not being a scalar product, together with novel error and oscillation reduction estimates, which now do not decouple. We show that AFEM is a contraction for the sum of energy error plus oscillation. Numerical experiments, including oscillatory coefficients and convection-diffusion PDE, illustrate the theory and yield optimal meshes. Key words. a posteriori error estimators, quasi-orthogonality, adaptive mesh refinement, error and oscillation reduction estimates, optimal meshes. AM subject classifications. 65N12, 65N15, 65N30, 65N50, 65Y20 1. Introduction and Main Result. Let Ω be a polyhedral bounded domain in R d, (d = 2, 3). We consider a general second order elliptic Dirichlet boundary value problem Lu = (A u) + b u + c u = f in Ω, (1.1) u = 0 on Ω, (1.2) with the following assumptions: A : Ω R d d is Lipschitz and symmetric positive definite with smallest eigenvalue a and largest eigenvalue a +, i.e., a (x) ξ 2 A(x)ξ ξ a + (x) ξ 2, ξ R d, x Ω; (1.3) b [L (Ω)] d is divergence free ( b = 0 in Ω); c L (Ω) is nonnegative (c 0 in Ω); f L 2 (Ω). The purpose of this paper is to prove the following convergence results for adaptive finite element methods (AFEM) for ( ), and document their performance computationally. Theorem 1 (Convergence of AFEM). Let {u k } k N0 be a sequence of finite element solutions corresponding to a sequence of nested finite element spaces { V k} k N 0 produced by the AFEM of 3.5, which involves loops of the form OLVE ETIMATE MARK REFINE. There exist constants σ, γ > 0, and 0 < ξ < 1, depending solely on the shape regularity of meshes, the data, the parameters used by AFEM, and a number 0 < s 1 dictated by the angles of Ω, such that if the initial meshsize h 0 satisfies h s 0 b L < σ, then for any two consecutive iterates k and k + 1 we have u u k γ osc k+1 (Ω) 2 ξ 2 ( u u k 2 + γ osc k (Ω) 2). (1.4) Department of Mathematics, University of Maryland, College Park, MD 20742, UA (kxm@math.umd.edu). Partially supported by NF Grant DM Department of Mathematics and Institute for Physical cience and Technology, University of Maryland, College Park, MD 20742, UA (rhn@math.umd.edu). Partially supported by NF Grant DM

2 2 K. MEKCHAY AND R.H. NOCHETTO Therefore, AFEM converges with a linear rate ξ, namely u u k 2 + γ osc k (Ω) 2 C 0 ξ 2k, where C 0 := u u γ osc 0 (Ω) 2. Hereafter, denotes the energy norm induced by the operator L and osc(ω), the oscillation term, stands for information missed by the averaging process associated to FEM. This convergence result extends those of Morin et al. [6, 7] in several ways: We deal with a full second order linear elliptic PDE with variable coefficients A, b and c, whereas in [6, 7] A is assumed to be piecewise constant and b and c to vanish. The underlying bilinear form B is non-symmetric due to the first order term b u. ince B is no longer a scalar product as in [6, 7], the Pythagoras equality relating u, u k and u k+1 fails; we prove a quasi-orthogonality property instead. The oscillation terms depend on discrete solutions in addition to data. Therefore, oscillation and error cannot be reduced separately as in [6, 7]. The oscillation terms do not involve the oscillation of the jump residuals. This is achieved by exploiting positivity and continuity of A. ince error and oscillation are now coupled, in order to prove convergence we need to handle them together. This leads to a novel argument and result, the contraction property (1.4), according to which both error and oscillation decrease together. This paper is organized as follows. In section 2 we introduce the bilinear form, the energy norm, recall existence and uniqueness of solutions, and state the quasiorthogonality property. In section 3 we describe the procedures used in AFEM, namely, OLVE, ETIMATE, MARK, and REFINE, state new error and oscillation reduction estimates, present the adaptive algorithm AFEM and prove its convergence. In section 4 we prove the quasi-orthogonality property of section 2 and the error and oscillation reduction estimates of section 3. In section 5 we present three numerical experiments to illustrate properties of AFEM. We conclude in section 6 with extensions to A piecewise Lipschitz with discontinuities aligned with the initial mesh and non-coercive bilinear form B due to b Discrete olution and Quasi-Orthogonality. For an open set G R d we denote by H 1 (G) the usual obolev space of functions in L 2 (G) whose first derivatives are also in L 2 (G), endowed with the norm u H 1 (G) := ( u L 2 (G) + u L 2 (G)) 1/2 ; we use the symbols H 1 and L 2 when G = Ω. Moreover, we denote by H 1 0 (G) the space of functions in H 1 (G) that vanish on the boundary in the trace sense. A weak solution of (1.1) and (1.2) is a function u satisfying u H 1 0 (Ω) : B[u, v] = f, v v H 1 0 (Ω), (2.1) where u, v := Ω uv for any u, v L2 (Ω), and the bilinear form is defined on H 1 0 (Ω) H 1 0 (Ω) as B[u, v] := A u, v + b u + c u, v. (2.2) By Cauchy-chwarz inequality one can easily show the continuity of the bilinear form B[u, v] C B u H 1 v H 1,

3 Convergence of AFEM for General econd Order Linear Elliptic PDE 3 where C B depends only on the data. Combining Poincaré inequality with the divergence free condition b=0, one has coercivity B[v, v] a v 2 + cv 2 c B v 2 H, 1 Ω where c B depends only on the data. Existence and uniqueness of (2.1) thus follows from Lax-Milgram theorem [5]. We define the energy norm on H 1 0 (Ω) by v 2 := B[v, v], which is equivalent to H 1 0 (Ω)-norm H 1. In fact we have c B v 2 H 1 v 2 C B v 2 H 1 v H 1 0 (Ω). (2.3) 2.1. Discrete olutions on Nested Meshes. Let {T H } be a shape regular family of nested conforming meshes over Ω: that is there exists a constant γ such that H T γ T T H, (2.4) ρ T H where, for each T T H, H T is the diameter of T, and ρ T is the diameter of the biggest ball contained in T ; the global meshsize is h H := max T TH H T. Let {V H } be a corresponding family of nested finite element spaces consisting of continuous piecewise polynomials over T H of fixed degree n 1, that vanish on the boundary. Let u H be a discrete solution of (2.1) satisfying u H V H : B[u H, v H ] = f, v H v H V H. (2.5) Existence and uniqueness of this problem follows from Lax-Milgram theorem, since V H H 1 0 (Ω) Quasi-Orthogonality. Consider two consecutive nested meshes T H T h, i.e. T h is a refinement of T H. For the corresponding spaces V H V h H0 1 (Ω), let u h V h and u H V H be the discrete solutions. ince the bilinear form is nonsymmetric, it is not a scalar product and the orthogonality relation between u u H and u h u H, the so-called Pythagoras equality, fails to hold. We have instead a perturbation result referred to as quasi-orthogonality provided that the initial mesh is fine enough. This result is stated below and the proof is given in section 4. Lemma 2.1 (Quasi-orthogonality). Let f L 2 (Ω). There exist a constant C > 0, solely depending on the shape regularity constant γ and coercivity constant c B, and a number 0 < s 1 dictated only by the angles of Ω, such that if the meshsize h 0 of the initial mesh satisfies C h s 0 b L < 1, then u u h 2 Λ 0 u u H 2 u h u H 2, (2.6) where Λ 0 := (1 C h s 0 b L ) 1. The equality holds provided b = 0 in Ω. 3. Adaptive Algorithm. The Adaptive procedure consists of loops of the form OLVE ETIMATE MARK REFINE. The procedure OLVE solves (2.5) for the discrete solution u H. The procedure E- TIMATE determines the element indicators η H (T ) and oscillation osc H (T ) for all

4 4 K. MEKCHAY AND R.H. NOCHETTO elements T T H. Depending on their relative size, these quantities are later used by the procedure MARK to mark elements T, and thereby create a subset T H of T H of elements to be refined. Finally, procedure REFINE partitions those elements in T H and a few more to maintain mesh conformity. These procedures are discussed more in detail below Procedure OLVE : Linear olver. We employ linear solvers, either direct or iterative methods, such as preconditioned GMRE, CG, and BICG, to solve linear system (2.5). In other words, given a mesh T k, an initial guest u k 1 for the solution, and the data A, b, c, f, OLVE computes the discrete solution u k := OLVE(T k, u k 1, A, b, c, f) 3.2. Procedure ETIMATE : A Posteriori Error Estimate. ubtracting (2.5) from (2.1), we have the Galerkin orthogonality B[u u H, v H ] = 0 v H V H. (3.1) In addition to T H, let H denote the set of interior faces (edges or sides) of the mesh (triangulation) T H. We consider the residual R(u H ) H 1 (Ω) defined by R(u H ) := f + (A u H ) b u H c u H, and its relation to the error L(u u H ) = R(u H ). It is then clear that to estimate u u H we can equivalently deal with R(u H ) H 1 (Ω). To this end, we integrate by parts elementwise the bilinear form B[u u H, v] to obtain the error representation formula B[u u H, v] = R T (u H )v + J (u H )v v H0 1 (Ω), (3.2) T T H T H where the element residual R T (u H ) and the jump residual J (u H ) are defined as R T (u H ) := f + (A u H ) b u H c u H in T T H, (3.3) J (u H ) := A u + H ν+ A u H ν := [[A u H ]] ν on H, (3.4) where is the common side of elements T + and T with unit outward normals ν + and ν, respectively, and ν = ν. Whenever convenient, we will use the abbreviations R T = R T (u H ) and J = J (u H ). by Upper Bound. For T T H, we define the local error indicator η H (T ) η H (T ) 2 := HT 2 R T (u H ) 2 L 2 (T ) + H J (u H ) 2 L 2 (). (3.5) T Given a subset ω Ω, we define the error estimator η H (ω) by η H (ω) 2 := η H (T ) 2. T T H, T ω Hence, η H (Ω) is the error estimator of Ω with respect to the mesh T H. Using (3.1),(3.2) and properties of Clément interpolation, as shown in [1, 3, 12], we obtain the upper bound of the error in terms of the estimator, u u H 2 C 1 η H (Ω) 2, (3.6) where the constant C 1 > 0 depends only on the shape regularity γ, coercivity constant c B and continuity constant C B of the bilinear form.

5 Convergence of AFEM for General econd Order Linear Elliptic PDE Lower Bound. Using the explicit construction of Verfürth [1, 12] via bubble functions and positivity and continuity of A, we can get a local lower bound of the error in terms of local indicators and oscillation. That is, there exist constants C 2, C 3 > 0, depending only on the shape regularity γ, C B, and c B, such that C 2 η H (T ) 2 C 3 R T R T 2 L 2 (T ) u u H 2 H 1 (ω T ), (3.7) T ω T H 2 T where the domain ω T consists of all elements sharing at least a side with T, and R T is a polynomial approximation of R T on T. We define the oscillation on the elements T T H by osc H (T ) 2 := HT 2 R T R T 2, (3.8) L 2 (T ) and for a subset ω Ω, we define osc H (ω) 2 := T T H, T ω osc H (T ) 2. Remark We see from (3.7) that if the oscillation osc H (ω T ) is small compared to the indicator η H (T ), then the size of the indicator η H (T ) will give reliable information about the size of the error u u H H 1 (ω T ). This explains why refining elements with large indicators usually tends to equi-distribute the errors, which is an ultimate goal of adaptivity. This idea is employed by the procedure MARK of 3.3. Remark The oscillation osc H (T ) does not involve oscillation of the jump residual J (u H ) as is customary [1, 12]. This result follows from the positivity and continuity of A, and is explained in 4.2. Remark The oscillation osc H (T ) depends on R T = R T (u H ) which in turn depends on the discrete solution u H. This is a fundamental difference with Morin et al. [6, 7], where the oscillation is purely data oscillation. It is not clear now that the oscillation will decrease when the mesh T H will be refined because u H will also change. Controlling the decay of osc H (T ) is thus a major challenge addressed in this work; see 3.3 and 3.4. It is not possible to show that oscillation will always decrease as the mesh gets refined as in [6, 7]. For a given mesh T H and discrete solution u H, along with input data A, b, c and f, the procedure ETIMATE computes indicators η H (T ) and oscillations osc H (T ) for all elements T T H according to (3.5) and (3.8): {η H (T ), osc H (T )} T TH = ETIMATE(T H, u H, A, b, c, f) 3.3. Procedure MARK. Our goal is to devise a marking procedure, namely to identify a subset T H of the mesh T H such that after refining, both error and oscillation will be reduced. We use two strategies for this: Marking trategy E deals with the error estimator, and Marking trategy O does so with the oscillation Marking trategy E : Error Reduction. This strategy was introduced by Dörfler [4] to enforce error reduction: Marking trategy E : Given a parameter 0 < θ < 1, construct a subset T H of T H such that T T H η H (T ) 2 θ 2 η H (Ω) 2, (3.9) and mark all elements in T H for refinement.

6 6 K. MEKCHAY AND R.H. NOCHETTO We will see later that Marking trategy E guarantees error reduction in the absence of oscillation terms. ince the latter account for information missed by the averaging process associated with the finite element method, we need a separate procedure to guarantee oscillation reduction Marking trategy O : Oscillation Reduction. This procedure was introduced by Morin et al. [6, 7] as a separate means for reducing oscillation: Marking trategy O : Given a parameter 0 < ˆθ < 1 and the subset T H T H produced by Marking trategy E, enlarge T H such that T T H osc H (T ) 2 ˆθ 2 osc H (Ω) 2, (3.10) and marked all elements in T H for refinement. Given a mesh T H and all information about the local error indicators η H (T ), and oscillation osc H (T ), together with user parameters θ and ˆθ, MARK generates a subset T H of T H TH = MARK(θ, ˆθ ; T H, {η H (T ), osc H (T )}T ) TH 3.4. Procedure REFINE. The following Interior Node Property, due to Morin et al [6, 7], is known to be necessary for error and oscillation reduction: Interior Node Property : Refine each marked element T T H to obtain a new mesh T h compatible with T H such that T and the d + 1 adjacent elements T T H of T, as well as their common sides, contain a node of the finer mesh T h in their interior. In addition to the Interior Node Property, we assume that the refinement is done in such a way that the new mesh T h is conforming, which guarantees that both T H and T h are nested. With this property, we have a reduction factor γ 0 < 1 of element size, i.e. if T T h is obtained by refining T T H, then h T γ 0 H T. For example, when d = 2 with triangular elements, to have Interior Node Property we can use 3 newest bisections for each single refinement step, whence γ 0 1/2. Given a mesh T H and a marked set T H, REFINE constructs the refinement T h satisfying the Interior Node Property: T h = REFINE(T H, T H ) Combining the marking strategies of 3.3 with the Interior Node Property, we obtain the following two crucial results whose proofs are given in 4. Lemma 3.1 (Error Reduction). There exist constants C 4 and C 5, only depending on the shape regularity constant γ and θ, such that η H (T ) 2 C 4 u h u H 2 H 1 (ω T ) + C 5osc H (ω T ) 2 T T H. (3.11) We realize that the local energy error between consecutive discrete solutions is bounded below by the local indicators for elements in the marked set T H, provided the oscillation term is relatively small.

7 Convergence of AFEM for General econd Order Linear Elliptic PDE 7 Lemma 3.2 (Oscillation Reduction). There exist constants 0 < ρ 1 < 1 and 0 < ρ 2, only depending on γ and ˆθ, such that osc h (Ω) 2 ρ 1 osc H (Ω) 2 + ρ 2 u h u H 2. (3.12) We have that oscillation reduces with a factor ρ 1 < 1 provided the energy error between consecutive discrete solutions is relatively small. Remark (Coupling of error and oscillation). Lemmas 3.1 and 3.2 seem to lead to conflicting demands on the relative sizes of error and oscillation. These two concepts are indeed coupled, which contrasts with [6, 7] where oscillation just depends on data and reduces separately from the error. This suggests that we must handle them together, this being the main contribution of this paper. We make this assertion explicit in Theorem 1 below Adaptive Algorithm AFEM. The adaptive algorithm consists of the loops of procedures OLVE, ETIMATE, MARK, and REFINE, consecutively, given that the parameters θ and ˆθ are chosen according to Marking trategies E and O. AFEM Choose parameters 0 < θ, ˆθ < Pick an initial mesh T 0, initial guest u 1 = 0, and set k = u k = OLVE(T k, u k 1, A, b, c, f). 3. {η k (T ), osc k (T )} T Tk = ETIMATE(T k, u k, A, b, c, f)). 4. T k = MARK(θ, ˆθ ; T k, {η k (T ), osc k (T )} T Tk ). 5. T k+1 = REFINE(T k, T k ). 6. et k = k + 1 and go to tep 2. Theorem 1 (Convergence of AFEM). Let {u k } k N0 be a sequence of finite element solutions corresponding to a sequence of nested finite element spaces { V k} k N 0 produced by AFEM. There exist constants σ, γ > 0, and 0 < ξ < 1, depending solely on the mesh regularity constant γ, data, parameters θ and ˆθ, and a number 0 < s 1 dictated by angles of Ω, such that if the initial meshsize h 0 satisfies h s 0 b L < σ, then for any two consecutive iterates k and k + 1, we have u u k γ osc k+1 (Ω) 2 ξ 2 ( u u k 2 + γ osc k (Ω) 2). (3.13) Therefore AFEM converges with a linear rate ξ, namely, u u k 2 + γ osc k (Ω) 2 C 0 ξ 2k, where C 0 := u u γ osc 0 (Ω) 2. Proof. We just prove the contraction property (3.13), which obviously implies the decay estimate. For convenience, we introduce the notation e k := u u k, ε k := u k+1 u k, osc k := osc k (Ω). The idea is to use the quasi-orthogonality (2.6) and replace the term u k+1 u k 2 using new results of error and oscillation reduction estimates (3.11) and (3.12). We proceed in three steps as follows.

8 8 K. MEKCHAY AND R.H. NOCHETTO 1. We first get a lower bound for ε k in terms of e k. To this end, we use Marking trategy E and the upper bound (3.6) to write θ 2 e 2 k C 1 θ 2 η k (Ω) 2 C 1 T T k η k (T ) 2. Adding (3.11) of Lemma 3.1 over all marked elements T T k, and observing that each element can be counted at most D := d + 2 times due to overlap of the sets ω T, together with v 2 H 1 v 2 for all v H0 1 (Ω), we arrive at If Λ 1 := c 1 B θ 2 e 2 k DC 1C 4 c B ε 2 k + DC 1 C 5 osc 2 k. θ2 c B DC 1C 4, Λ 2 := C 5c B C 4, then this implies the lower bound for ε 2 k, ε 2 k Λ 1 e 2 k Λ 2 osc 2 k. (3.14) 2. If h 0 is sufficiently small so that the quasi-orthogonality (2.6) of Lemma 2.1 holds, then e 2 k+1 Λ 0 e 2 k ε 2 k, where Λ 0 = (1 C h s 0 b L ) 1. Replacing the fraction βε 2 k of ε2 k via (3.14) we obtain e 2 k+1 (Λ 0 βλ 1 )e 2 k + βλ 2 osc 2 k (1 β)ε 2 k, where 0 < β < 1 is a constant to be chosen suitably. We now assert that it is possible to chose h 0 compatible with Lemma 2.1 and also that 0 < α := Λ 0 βλ 1 < 1. A simple calculation shows that this is the case provided i.e., h s 0 b L < σ with σ := h s 0 b L < βλ 1 C (1 + βλ 1 ) < 1 C, βλ 1 C (1+βΛ 1 ). Consequently e 2 k+1 αe 2 k + βλ 2 osc 2 k (1 β)ε 2 k. (3.15) 3. To remove the last term of (3.15) we resort to the oscillation reduction estimate of Lemma 3.2 osc 2 k+1 ρ 1 osc 2 k + ρ 2 ε 2 k. We multiply it by (1 β)/ρ 2 and add it to (3.15) to deduce e 2 k β ( osc 2 k+1 α e 2 k + βλ 2 + ρ ) 1 (1 β) osc 2 ρ 2 ρ k. 2 If γ := 1 β ρ 2, then we would like to choose β < 1 in such a way that βλ 2 + ρ 1 γ = µγ

9 Convergence of AFEM for General econd Order Linear Elliptic PDE 9 for some µ < 1. A simple calculation yields β = µ ρ 1 ρ 2 Λ 2 + µ ρ, 1 ρ 2 and shows that ρ 1 < µ < 1 guarantees that 0 < β < 1. Therefore, e 2 k+1 + γ osc 2 k+1 α e 2 k + µγ osc 2 k, and the asserted estimate (3.13) follows upon taking ξ = max(α, µ) < 1. Remark (Comparison with [6, 7]). In [6, 7] the oscillation is independent of discrete solutions, i.e. ρ 2 = 0, and is reduced by the factor ρ 1 < 1 in (3.12). Consequently, tep 3 above is avoided by setting β = 1 and the decay of e k and osc k is monitored separately. ince this is no longer possible, e k and osc k are now combined and decreased together. Remark (plitting of ε k ). The idea of splitting ε k is already used by Chen and Jia [2] in examining one time step for the heat equation. This is because a mass (zero order) term naturally occurs, which did not take place in [6, 7]. The elliptic operator is just the Laplacian in [2]. Remark (Effect of Convection). Assuming that h s 0 b L < σ implies that the local Peclet number is sufficiently small for the Galerkin method not to exhibit oscillations. This appears to be essential for u 0 to contain relevant information and guide correctly the adaptive process. This restriction is difficult to verify in practice because it involves unknown constants. However, starting from a coarse mesh does not seem to be a problem in practice (see numerical experiments in 5). Remark (Vanishing Convection). If b = 0, then Theorem 1 has no restriction on the initial mesh. This thus extends the convergent result of Morin et al. [6, 7] to variable diffusion coefficient and zero order terms. Remark (Optimal β). The choice of β can be optimized. In fact, we can easily see that α = Λ 0 βλ 1, µ = ρ 1 + β 1 β ρ 2Λ 2 yields a unique value 0 < β < 1 for which α = µ and the contraction constant ξ of Theorem 1 is minimal. This β depends on geometric constant Λ 0, Λ 1, Λ 2 as well on θ, ˆθ and h 0, but it is not computable. 4. Proofs of Lemmas. Let T H T H be a set of marked elements obtained from procedure MARK. Let T h be a refined mesh obtained from procedure REFINE, and let V H V h be nested spaces corresponding to compatible meshes T H and T h, respectively. For convenience, set e h := u u h, e H := u u H, ε H := u h u H Proof of Lemma 2.1: Quasi-Orthogonality. In view of Galerkin orthogonality (3.1), namely B[e h, v h ] = 0, v h V h, we have e H 2 = e h 2 + ε H 2 + B[ε H, e h ].

10 10 K. MEKCHAY AND R.H. NOCHETTO If b = 0, then B is symmetric and B[ε H, e h ] = B[e h, ε H ] = 0. For b 0, instead, B[ε H, e h ] 0, and we must account for this term. It is easy to see that b = 0 and integration by parts yield Hence B[ε H, e h ] = B[e h, ε H ] + b ε H, e h b e h, ε H = 2 b ε H, e h. e h 2 = e H 2 ε H 2 2 b ε H, e h. Using Cauchy-chwarz inequality and replacing the H 1 (Ω)-norm by the energy norm we have, for any δ > 0 to be chosen later, 2 b ε H, e h δ e h 2 L 2 + b 2 L δc B ε H 2. We then realize the need to relate L 2 (Ω) and energy norms to replace e h L 2 by e h. This requires a standard duality argument whose proof is reported in Ciarlet [3]. Lemma 4.1 (Duality). Let f L 2 (Ω) and u H 1+s (Ω) for some 0 < s 1 be the solution, where s depends on the angles of Ω (s = 1 if Ω is convex). Then, there exists a constant C 6, depending only on the shape regularity constant γ and the coercivity constant c B, such that e h L 2 C 6 h s e h. (4.1) Inserting this estimate in the preceding two bounds, and using h h 0, the meshsize of the initial mesh, we deduce ( ) ( 1 δc 2 6 h 2s ) 0 eh 2 e H 2 1 b 2 L ε H 2. δc B We now choose δ = b L C 6 cb h s 0 to equate both parenthesis, as well as the assumption that h 0 is sufficiently small for δc6h 2 2s 0 = C h s 0 b L < 1 with C := C 6 / c B. We end up with e h C b L h s e H 2 ε H 2. 0 This implies (2.6) and concludes the proof Proof of Lemma 3.1 : Error Reduction. Upon restricting the test function v in (3.2) to V h V H, we obtain the error representation B[ε H, v h ] = R T v h + (R T R T )v h + J v h v h V h, (4.2) T T H T T H where we use the abbreviations R T = R T (u H ), J = J (u H ), and R T = Π n 1 T R T denotes the L 2 -projection of R T onto the space of polynomials P n 1 (T ) over the element T T H. Except for avoiding the oscillation terms of the jump residual J, the proof goes back to [4, 6, 7]. We proceed in three steps.

11 Convergence of AFEM for General econd Order Linear Elliptic PDE Interior Residual. Let T T H, and let x T be an interior node of T generated by the procedure REFINE. Let ψ T V h be a bubble function which satisfies ψ T (x T ) = 1, vanishes on T, and 0 ψ T 1; hence supp (ψ T ) T. ince R T P n 1 (T ) and ψ T > 0 in a polyhedron of measure comparable with that of T, we have C R T 2 L 2 (T ) 2 ψ T R T = R T (ψ T R T ). T ince ψ T R T is a piecewise polynomial of degree n over T h, it is thus an admissible test function in (4.2) which vanishes outside T (and in particular on all H ). Therefore C R T 2 B[ε L 2 (T ) H, ψ T R T ] + (R T R T )ψ T R T T ( C H 1 T ε H H 1 (T ) + R T R L T 2 (T )) R L T 2 (T ), because of an inverse inequality for ψ T R T. This, together with the triangle inequality, yields the desired estimate for H 2 T R T 2 L 2 (T ) : H 2 T R T 2 L 2 (T ) C ( ε H 2 H 1 (T ) + H2 T T ) R T R T 2. (4.3) L 2 (T ) 2. Jump Residual. Let H be an interior side of T 1 T H, and let T 2 T H be the other element sharing. Let x be an interior node of created by the procedure REFINE. Let ψ V h be a bubble function in ω := T 1 T 2 such that ψ (x ) = 1, ψ vanishes on ω, and 0 ψ 1; hence supp (ψ ) ω. ince u H is continuous, then [[ u H ]] is parallel to ν, i.e. [[ u H ]] = j ν. Moreover, the coefficient matrix A(x) being continuous implies J = A(x) [[ u H ]] ν = j A(x)ν ν = a(x) j, where a(x) := A(x)ν ν satisfies 0 < a a(x) a with a, a the smallest and largest eigenvalues of A(x) on. Consequently, J 2 L 2 () a2 j 2 Ca 2 jψ 2 C a2 (j ψ )J, (4.4) a where the second inequality follows from j being a polynomial and ψ > 0 in a polygon of measure comparable with that of. We now extend j to ω by first mapping to the reference element, next extending constantly along the normal to Ŝ and finally mapping back to ω. The resulting extension E h (j ) is a piecewise polynomial of degree n 1 in ω so that ψ E h (j ) V h, and satisfies ψ E h (j ) L 2 (ω ) CH1/2 j L 2 (). ince v h = ψ E h (j ) is an admissible test function in (4.2) which vanishes on all sides of H but, we arrive at J (j ψ ) = B[ε H, v h ] C ( T 1 R T1 v h R T2 v h T 2 H 1/2 ε H H 1 (ω ) + H1/2 ) 2 (4.5) R Ti L 2 (T i ) j L2(). i=1

12 12 K. MEKCHAY AND R.H. NOCHETTO Therefore ( ) 2 H J 2 L 2 () C ε H 2 H 1 (ω ) + HT 2 i R Ti 2 L 2 (T i ). (4.6) 3. Final Estimate. To remove the interior residual from the right hand side of (4.6) we resort to (4.3) since T 1 and T 2 contain an interior node according to procedure REFINE. Hence ( ) 2 H J 2 L 2 () C ε H 2 H 1 (ω ) + HT 2 i R Ti R Ti 2. (4.7) i=1 i=1 L 2 (T i ) The asserted estimate for η H (T ) 2 is thus obtained by adding this bound to (4.3). The constant C depends on the shape regularity constant γ and the ratio a 2 /a of eigenvalues of A(x) on. Remark (Positivity). The use of A(x) being positive definite in (4.4) avoids having oscillation terms on. This comes at the expense of a constant depending on a 2 /a. If we were to proceed in the usual manner, as in [1, 8, 12], we would end up with oscillation of the form H 1/2 (A(x) A(x )) [[ u H ]] ν L 2 () = H1/2 (a(x) a(x ))j L 2 () CH 3/2 CH H 1/2 A W 1 () j L 2 (), J L 2 () where C > 0 also depends on the ratio a /a dictated by the variation of a(x) on. This oscillation can be absorbed into the term H 1/2 J L 2 () provided that the meshsize H is sufficiently small; see [8]. We do not need this assumption in our present discussion. Remark (Continuity of A). The continuity of A is instrumental in avoiding jump oscillation which in turn makes computations simpler. However, jump oscillation cannot be avoid when A exhibits discontinuities across inter-element boundaries of the initial mesh. We get instead of (4.7) 2 CH J 2 L 2 () ε H 2 H 1 (ω ) + HT 2 i R Ti R Ti 2 L 2 (T i ) +H J J 2 L 2 (), (4.8) i=1 where J is L 2 -projection of J onto P n 1 (). To obtain estimate (4.8) we proceed as follows. tarting from a polynomial J, we get an estimate similar to that of (4.4) C J 2 L 2 () 2 ψ J = J (ψ J ) + (J J )(ψ J ). (4.9) In contrast to (4.4), we see that the oscillation term (J J ) cannot be avoided when A has a discontinuity across. We estimate the first term on the right hand side of (4.9) exactly as we have argued with (4.5) and thereby arrive at ( J (J ψ ) C H 1/2 ε H H 1 (ω ) + H1/2 2 R Ti L 2 (T i )) J L. 2 () i=1

13 Convergence of AFEM for General econd Order Linear Elliptic PDE 13 This and a further estimate of the second term on the right hand side of (4.9), yield ( H J 2 2 C ε L 2 () H 2 H 1 (ω ) + i=1 H 2 T i R Ti 2 L 2 (T i ) + H ) J J 2, L 2 () whence the assertion (4.8) follows using triangle inequality for J L 2 (). Combining with (4.3), we deduce an estimate for η H (T ) similar to (3.11), namely, ( η H (T ) 2 C ε H 2 H 1 (ω T ) + osc H(ω T ) 2), with the new oscillation term involving jumps on interior sides osc H (T ) 2 := H 2 T R T R T 2 L 2 (T ) + T H J J 2 L 2 (). (4.10) In 6.1 we discuss the case of a discontinuous A. We show an oscillation reduction property of osc H (T ), defined by (4.10), similar to Lemma Proof of Lemma 3.2 : Oscillation Reduction. The proof hinges on the Marking trategy O and the Interior Node Property. We point out that if T T h is contained in T T H, then REFINE gives a reduction factor γ 0 < 1 of element size: h T γ 0 H T. (4.11) The proof proceeds in three steps as follows. 1. Relation between Oscillations. We would like to relate osc h (T ) and osc H (T ) for any T T H. To this end, we note that for all T T h contained in T, we can write where ε H = u h u H as before and R T (u h ) = R T (u H ) L T (ε H ) in T, L T (ε H ) := (A ε H ) + b ε H + c ε H in T. By Young s inequality, we have for all δ > 0 osc h (T ) 2 = h 2 T R T (u h ) R T (u h ) 2 L 2 (T ) (1+δ)h 2 T R T (u H ) R T (u H ) 2 L 2 (T )+(1+δ 1 )h 2 T L T (ε H ) L T (ε H ) 2 L 2 (T ), where R T (u h ), R T (u H ), and L T (ε H ) are L 2 -projections of R T (u h ), R T (u H ), and L T (ε H ) onto polynomials of degree n 1 on T. We next observe that L T (ε H ) L T (ε H ) L T (ε H ) L L 2 (T ) 2 (T ) and that, according to (4.11), h T γ T H T

14 14 K. MEKCHAY AND R.H. NOCHETTO provided γ T = γ 0 if T T H and γ T = 1 otherwise. Therefore, if T h (T ) denotes all T T h contained in T, osc h (T ) 2 = osc h (T ) 2 T T h (T ) (1 + δ)γ 2 T osc H(T ) 2 + (1 + δ 1 ) T T h (T ) h 2 T L T (ε H ) 2 L 2 (T ), (4.12) because R T (u H ) = R T (u H ) and R T (u H ) is the best approximation of R T (u H ) in T. 2. Estimate of L T (ε H ). In order to estimate L T (ε H ) L 2 (T ) in terms of ε H H 1 (T ), we first split it as follows L T (ε H ) L 2 (T ) (A ε H) L 2 (T ) + b ε H L 2 (T ) + c ε H L 2 (T ) and denote these terms N A, N B, and N C, respectively. ince N A ( A) ε H L 2 (T ) + A : H(ε H) L 2 (T ) where H(ε H ) is the Hessian of ε H in T, invoking the Lipschitz continuity of A together with an inverse estimate in T, we infer that ( ) N A C A ε H L 2 (T ) + h 1 T ε H L 2 (T ), where C A depends on A and the shape regularity constant γ. Besides, we readily have N B C B ε H L 2 (T ), N C C C ε H L 2 (T ), where C B, C C depend on b, c. Combining these estimates, we arrive at h 2 T L T (ε H ) 2 L 2 (T ) C ε H 2 H 1 (T ). (4.13) 3. Choice of δ. We insert (4.13) into (4.12) and add over T T H. Recalling the definition of γ T and utilizing (3.10), we deduce γt 2 osc H(T ) 2 = γ0 2 osc H (T ) 2 + osc H (T ) 2 T T H T T H T T H \ T H = osc H (Ω) 2 (1 γ0) 2 osc H (T ) 2 T T H (1 (1 γ 2 0)ˆθ 2 ))osc H (Ω) 2, where ˆθ is the user s parameter in (3.10). Moreover, since C ε H 2 H C 1 o ε H 2 with C o = C c 1 B in light of (2.3), we end up with osc h (Ω) 2 (1 + δ)(1 (1 γ 2 0)ˆθ 2 )osc H (Ω) 2 + (1 + δ 1 )C o ε H 2. To complete the proof, we finally choose δ sufficiently small so that ρ 1 = (1 + δ)(1 (1 γ 2 0)ˆθ 2 ) < 1, ρ 2 = (1 + δ 1 )C o.

15 Convergence of AFEM for General econd Order Linear Elliptic PDE Numerical Experiments. We test performance of the adaptive algorithm AFEM with several examples. We are thus able to study how meshes adapt to various effects from lack of regularity of solutions and convexity of domains to data smoothness, boundary layers, changing boundary conditions, etc. For simplicity, we strict our experiments to the case of piecewise linear finite element solutions with polygonal domains in 2 dimensions. The implementation is done within the finite element toolbox ALBERT of chmidt and iebert [10, 11] which provides tools for adaptivity Implementation. We employ the four main procedures as given by Morin et al. [6, 7]: OLVE, ETIMATE, MARK, and REFINE. We slightly modified the builtin adaptive solver for elliptic problems of ALBERT toolbox [10] to make it work for the general PDE (1.1) and mixed boundary conditions, as follows: OLVE. We used built-in solvers provided by ALBERT toolbox, such as GMRE, BICG, or CG. ETIMATE. We modified ALBERT for computing the estimator so that it works for (1.1), and added procedures for computing oscillations which are not provided. MARK. We utilized the same algorithm introduced by Morin et al [6, 7] for finding a marked set T H. REFINE. We employed 3 newest bisections for each refinement step to enforce the Interior Node Property. For simplicity and convenience of presentation, we introduce the following notation: DOF := number of elements in a given mesh, which is comparable with number of degree of freedoms; EOC := log(e(k 1)/e(k)) log(dof(k)/dof(k 1)), experimental order of convergence, e(k):= u u k H 1; log(η k 1 /η k ) log(dof(k)/dof(k 1)) EOC(η) :=, experimental order of convergence of estimator, η k := η k (Ω); Z e := e(k)/e(k 1) and Z o := osc k /osc k 1, reduction factors of error and oscillation; Eff := η k /e k, effectivity index, i.e. the ratio between the estimator and error Experiment 1 : Oscillatory Coefficients and Nonconvex Domain. We consider the PDE (1.1) with Dirichlet boundary condition u = g on the nonconvex L-shape domain Ω := [ 1, 1] 2 \[0, 1] [ 1, 0]. We also take the exact solution u(r) = r 2 3 sin( 2 3 θ), where r 2 := x 2 + y 2 and θ := tan 1 (y/x) [0, 2π). We deal with variable coefficients A(x, y) = a(x, y)i, b(x, y), c(x, y) defined by 1 a(x, y) = 4 + P (sin( 2πx ɛ ) + sin( 2πy ɛ )), (5.1) b(x, y) = ( y, x). (5.2) c(x, y) = A c (cos 2 (k 1 x) + cos 2 (k 2 x)), (5.3) where P, ɛ, A c, k 1 and k 2 are parameters. The functions f in (1.1) and g are defined accordingly. To see how AFEM performs comparing to standard uniform refinement, results of AFEM and standard uniform refinement are provided in Tables 5.1 and 5.2. ome examples of adapted refined meshes from AFEM are also displayed in Figure 5.1. Observations and conclusions about AFEM performance are as follows:

16 16 K. MEKCHAY AND R.H. NOCHETTO DOF u u k H 1 EOC Z e Z o Eff e e e e e e e e e Table 5.1 Experiment 1 (Oscillatory coefficients and nonconvex domain): The parameters of AFEM are θ = ˆθ = 0.6, and those controlling the oscillatory coefficients are P = 1.8, ɛ = 0.4, A c = 1.0, k 1 = k 2 = 4.0, as described in (5.1)-(5.3). The experimental order of convergence EOC is close to the optimal value 0.5, which indicates quasi-optimal meshes. The oscillation reduction factor Z o is small than the error reduction factor, which confirms that oscillation decreases faster than error. DOF u u k H 1 EOC Z e Z o e e e e Table 5.2 Experiment 1 (Oscillatory coefficients and nonconvex domain): tandard uniform refinement is performed using the same values for parameters P, ɛ, A c, k 1, and k 2 as that of AFEM given by Table 5.1 above. EOC is now suboptimal and close to the expected value 1/3. Fig Experiment 1 (Oscillatory coefficients and nonconvex domain): Parameters of AFEM are θ = ˆθ = 0.6, those of oscillatory coefficient A are P = 1.8, ɛ = 0.4, and both b and c vanish. This mesh sequence shows that mesh refinement is dictated by geometric (corner) singularities as well as periodic variations of the diffusion coefficient.

17 Convergence of AFEM for General econd Order Linear Elliptic PDE 17 AFEM gives optimal rate order of convergence 0.5 while standard uniform refinement achieves the suboptimal rate 0.33 expected from theory; see Tables 5.1 and 5.2. AFEM performs with effectivity index close to 1.0 and reduction factors of error and oscillation close to 0.7 and 0.5 as DOF increases (see Table 5.1). The oscillation thus decreases faster than the error and becomes insignificant asymptotically for k large. Figure 5.1 depicts the effect of a corner singularity and periodic variation of diffusion in mesh grading; here both b and c vanish Experiment 2 : Convection Dominated-Diffusion. We consider the convection dominated-diffusion elliptic model problem (1.1) with Dirichlet boundary condition u = g on convex domain Ω := [0, 1] 2, with isotropic diffusion coefficient A = ɛi, ɛ = 10 3, convection velocity b = (y, 1 2 x), and without the zero order term c = 0. The Dirichlet boundary condition g(x, y) on Ω is the continuous piecewise linear function given by 1 {.2 + ξ x.5 ξ; y = 0}, g(x, y) = 0 Ω \ {.2 x.5; y = 0}, linear {(.2 x.2 + ξ) or (.5 ξ x.5); y = 0}, where ξ = The parameters of AFEM are θ = 0.65, ˆθ = 0.6. Results are reported in Tables 5.3, 5.4 and Figures 5.2, 5.3. Conclusions and observations follow: We observe from Tables 5.3, 5.4 that at about the same level of DOF, AFEM gives much smaller error estimators, which implies that AFEM reduces error estimator much faster than standard uniform refinement. ince the computational decay of estimator η(ω) is close to the optimal value 0.5, the resulting meshes are quasi-optimal. Figure 5.2 depicts graded meshes which capture the nature of transport of a pulse from the boundary inside the domain. Figure 5.3 displays solutions without oscillations even though AFEM is not stabilized. Mesh grading is more noticeable at the boundary and internal layers, whereas the rest of the mesh barely changes. DOF η(ω) EOC(η) Z o e e e e e e e e e Table 5.3 Experiment 2 (Convection Dominated-Diffusion): Parameters of AFEM are θ = 0.65, ˆθ = 0.6, and model parameters ɛ = and ξ = The optimal decay 0.5 of estimator η(ω) is computational evidence of optimal meshes Experiment 3 : Drift-Diffusion Model. We consider a model problem that comes from a mathematical model in semi-conductors and chemotaxis.

18 18 K. MEKCHAY AND R.H. NOCHETTO DOF η(ω) EOC(η) Z o e e e e e e Table 5.4 Experiment 2 (Convection Dominated-Diffusion): tandard uniform refinement is performed using the same values for parameters ɛ and ξ as that of AFEM given in Table 5.3. The resolution with about elements is comparable with adapted meshes of about elements. Fig Experiment 2: Adaptively refined meshes after 3, 5, and 7 iterations. AFEM defects the region of rapid variation (circular transport of a pulse) and boundary layer in the outflow. The rest of the mesh remains unchanged Fig Experiment 2: The corresponding plots of solutions after 3, 5, and 7 iterations. No oscillations are detected even though AFEM is not stabilized. Adaptivity incorporates a nonlinear stabilization mechanism. ( u + χu ψ) = 0 in Ω := [0, 1] 2, u = g on Γ Ω, ν u = 0 on Ω \ Γ, where χ is a constant. The radial function ψ is defined on Ω by 1 { x 2 + y 2 r 1 }, ψ(x, y) := α { x 2 + y 2 r 1 + α}, linear {r 1 < x 2 + y 2 < r 1 + α},

19 Convergence of AFEM for General econd Order Linear Elliptic PDE 19 Fig Experiment 3: Refined meshes after k = 6, 8, 10 iterations and corresponding solutions u k. Mesh grading is quite pronounced in the internal layer where ψ does not vanish, and at the midpoints of the boundary sides, where boundary conditions change. where α is a small parameter and r 1 < 1 is a constant. The Dirichlet boundary condition on Γ is assumed to be { 1 g(x, y) = 0 {x = 0; 0 y 0.5} {y = 0; 0 x 0.5}, {x = 1; 0.5 y 1} {y = 1; 0.5 x 1}. We resort to the following transformation (exponential fitting) to symmetrize the problem ρ := exp(χψ)u = (exp( χψ) ρ) = 0, which gives a simpler form of model problem with variable scalar coefficient a = exp( χψ). We can now apply AFEM to solve for ρ and next compute u = exp( χψ)ρ at the nodes. Again, the experiment is performed by comparing the performance of AFEM with standard uniform refinement using parameters χ = 10.0, r 1 = 0.75, and α = 0.04 for the model problem, and parameters θ = 0.6, ˆθ = 0.75 for AFEM. Results reported in Tables 5.5, 5.6 and Figure 5.4. Conclusions and observations follow: From Tables 5.5, 5.6 we see again that AFEM outperforms the standard uniform refinement. ince the decay of estimator η(ω) is optimal, we have computational evidence of optimal meshes. Figure 5.4 displays meshes and corresponding solutions u k for iterations k = 6, 8, 10. Meshes adapt well to lack of smoothness, namely refinement concentrates in the transition layer, where ψ does not vanish, and at the midpoints of boundary sides, where boundary conditions change. 6. Extensions. We extend the model problem (1.1) by considering now A with discontinuities aligned with the initial mesh and a velocity field b no longer divergence free. Note that if b 0, then the bilinear form B is non-coercive.

20 20 K. MEKCHAY AND R.H. NOCHETTO DOF η(ω) EOC(η) Z o Table 5.5 Experiment 3 (Drift-Diffusion Model): Parameters of AFEM are θ = 0.6, ˆθ = 0.75, and model parameters are χ = 10, r 1 = 0.75 and α = The optimal decay 0.5 of estimator η(ω) is computational evidence of quasi-optimal meshes. AFEM outperforms uniform refinement (compare with Table 5.6). DOF η(ω) EOC(η) Z o Table 5.6 Experiment 3 (Drift-Diffusion Model): tandard uniform refinement is performed using the same values for parameters χ, r 1 and α as for AFEM given in Table Discontinuous A. The assumption of continuity of A is used for obtaining error and oscillation reduction estimates, Lemma 3.1 and Lemma 3.2, in that the element oscillation osc H (T ) does not involve oscillation of jump residual on T. Remark shows that when A has discontinuities across element faces, we still obtain error reduction estimate (3.11) of Lemma 3.1, but this time the oscillation is defined by (4.10) and involves oscillation of jump residual. To prove convergence it suffices to show the oscillation reduction estimate (3.12), for the new concept of element oscillation, namely osc H (T ) 2 = osc R,H (T ) 2 + T osc J,H() 2 with osc R,H (T ) 2 := HT 2 R T (u H ) R T (u H ) osc J,H () 2 J := H (u H ) J (u H ) 2 L 2 (T ) 2 L 2 () T T H, H. We proceed in three steps as follows. 1. Oscillation of Interior Residual. Invoking the same arguments as in the proof of Lemma 3.2 in 4.3, we obtain an oscillation reduction estimate for interior residual osc R,h (T ) 2 (1 + δ)γ 2 T osc R,H(T ) 2 + C (1 + δ 1 ) ε H 2 H 1 (T ) T T H,

21 Convergence of AFEM for General econd Order Linear Elliptic PDE 21 where osc R,h (T ) is defined to be osc h (T ) in (4.12). 2. Oscillation of Jump Residual. To obtain estimate for osc J,h () we write J (u h ) = γ [[A u H ] ν + [[A ε H ] ν s = γ J (u H ) + J (ε H ), where γ = 1 if H and γ = 0 otherwise, since A u H is continuous on in the second case. Using Young s inequality, we have for all δ > 0 osc J,h () 2 J (1 + δ)γ h (u H ) J (u H ) + (1 + δ 1 J )h (ε H ) J (ε H ) 2 L 2 () 2 L 2 () where J (u H ) and J (ε H ) are L 2 -projections of J (u H ) and J (ε H ) onto P n 1 (). For the second term we observe that J (ε H ) J (ε H ) J (ε H ) L L 2 () 2 () = [A ε H ]] ν L 2 () A + ε + H ν L 2 () + A ε H ν L 2 () ( A L (ω ) ε + H + ) L ε 2 () H C A h 1/2 ε H H 1 (ω ),, L 2 () where C A depends on A and shape regularity constant γ. For simplicity, let h (T ) denote all h contained in T T H ; hence osc J,h (T ) 2 = osc J,h () 2 h (T ) (1 + δ) h (T ) J γ h (u H ) J (u H ) 2 + (1 + δ 1 )C A ε H 2 L 2 H () 1 (ω T ). In light of reduction factor of element size h γ T H, and definitions of γ and γ T, we obtain J γ h (u H ) J (u H ) 2 γ T H J (u H ) J (u H ) L 2 () h (T ) H (T ) = γ T osc J,H (T ) 2, 2 L 2 ( ) because for T, we have J (u H ) = J (u H ) and J (u H ) is the best L 2 estimate for J (u H ) on. Therefore osc J,h (T ) 2 (1 + δ)γ T osc J,H (T ) 2 + (1 + δ 1 )C A ε H 2 H 1 (ω T ) T T H. 3. Choice of δ. Combining results from steps 1 and 2 above using γ T 1, C = max {C, C A } and definition of osc h (T ), we arrive at osc h (T ) 2 (1 + δ)γ T osc H (T ) 2 + C (1 + δ 1 ) ε H 2 H 1 (ω T ). Proceeding as in step 3 of the proof of Lemma 3.2, this time with Marking trategy O performed according to the new definition of osc H (T ), we arrive at osc h (Ω) 2 (1 + δ)(1 (1 γ 0 )ˆθ 2 )osc H (Ω) 2 + C o (1 + δ 1 ) ε H 2, with C o = C c 1 B. The assertion thus follows by choosing δ sufficiently small so that ρ 1 := (1 + δ)(1 (1 γ 0 )ˆθ 2 ) < 1, ρ 2 := C o (1 + δ 1 ).

22 22 K. MEKCHAY AND R.H. NOCHETTO 6.2. Non-coercive B. In this section we prove convergence of AFEM for noncoercive bilinear forms B. According to what we have so far, the assumption of B being coercive is used for proving quasi-orthogonality and for having equivalence between energy norm v 2 := B[v, v] and H 1 -norm as in (2.3). ince now B is no longer coercive, we cannot define energy norm in this manner. We instead define energy norm by v 2 := Ω A v v + c v2, and we have equivalence of norms c E v 2 H 1 (Ω) v 2 C E v 2 H 1 (Ω), (6.1) where constants c E and C E depend only on data A, c and Ω. The lack of coercivity is now replaced by Gårding s inequality v 2 γ G v 2 L 2 (Ω) B[v, v] v H1 0 (Ω), (6.2) where γ G = b /2. To see this we integrate by parts the middle term of B[v, v], b v v = 1 b (v 2 b ) = Ω 2 Ω Ω 2 v2 v H0 1 (Ω). The same calculation leads to the sharp upper bound for B[v, v]: B[v, v] v 2 + γ G v 2 L 2 (Ω) v H 1 0 (Ω). (6.3) Existence and uniqueness of weak solutions follows from the maximum principle for c 0 [5]. chatz showed in [9] that the discrete problem (2.5) has a unique solution if the meshsize h is sufficiently small, i.e. h h for some constant h depending on shape regularity and data but not computable; the results in [9] are valid also for graded meshes. Assuming h 0 h, to prove convergence of AFEM it thus suffices to prove quasi-orthogonality. We follow the steps of Lemma 2.1. Using the same notation as in 4 for e h, e H and ε H, expanding B[e H, e H ], and noticing that e H = e h + ε H and B[e h, ε H ] = 0, we arrive at B[e h, e h ] = B[e H, e H ] B[ε H, ε H ] B[ε H, e h ], (6.4) where this time integration by parts yields B[ε H, e h ] = B[e h, ε H ] + b ε H, e h b e h, ε H = 2 b ε H, e h + b e h, ε H. Consequently, using Cauchy-chwarz inequality and (6.1), we have for all δ > 0 B[ε H, e h ] (2 b ε H L 2 + b ε H L 2) e h L 2 C 2 b δ ε H 2 + δ 1 e h 2 L 2, where constant C b = max {2 b, b } c 1 E /2. Using (6.2) and (6.3) to estimate terms B[e h, e h ], B[e H, e H ], B[ε H, ε H ] in (6.4), and combining with the previous estimate, we infer that e h 2 (γ G + δ 1 ) e h 2 L 2 e H 2 + γ G e H 2 L 2 (1 C 2 b δ) ε H 2 + γ G ε H 2 L 2. ince ε H 2 L 2 2 e h 2 L e H 2 L 2, estimates for e h L 2 and e H L 2 of the form (4.1), obtained via duality, imply Λ h e h 2 Λ H e H 2 Λ ε ε H 2, (6.5)

23 Convergence of AFEM for General econd Order Linear Elliptic PDE 23 where Λ h = 1 C 2 6h 2s 0 (3γ G + δ 1 ), Λ H = 1 + 3γ G C 2 6h 2s 0, Λ ε = 1 C 2 b δ. Consequently, to get Λ h = Λ ε, it is easy to see that we need to choose δ depending on h 0, δ(h 0 ) = C Gh 2s 0 + CG 2 h4s 0 + 4C2 b C2 6 h2s 0 2Cb 2 > 0, where C G = 3γ G C6. 2 The assertion thus follows provided the meshsize h 0 is sufficiently small so that Cb 2δ(h 0) < 1. This can be achieved for h s 0 min { C 6 C b C 1 G, (3C 6C b ) 1} because Cb 2 δ(h 0 ) = C G 2 h2s 0 + C b C 6 h s h 2s 0 C2 G (4C2 b C2 6 ) 1 2C b C 6 h s 0( 1 + h s 0 C G (4C b C 6 ) 1) < 3C b C 6 h s 0 1. We conclude that if the meshsize h 0 of the initial mesh satisfies h s 0 min { C 6 C b C 1 G, (3C 6C b ) 1, (h ) s}, then quasi-orthogonality holds, i.e. for Λ 0 := Λ H /Λ h, e h 2 Λ 0 e H 2 ε H 2, (6.6) and Λ 0 can be made arbitrarily close to 1 by decreasing h 0. Convergence of AFEM finally follows as in Theorem 1. REFERENCE [1] M. Ainsworth and J.T. Oden, A Posteriori Error Estimation in Finite Element Analysis, John Wiley & ons, Inc., [2] Z. Chen and F. Jia, An adaptive finite element algorithm with reliable and efficient error control for linear parabolic problems, Math. Comp. (to appear) [3] Ph. Ciarlet, The Finite Element Method for Elliptic Problems., North-Holland, Amsterdam, [4] W. Dörfler. A convergent adaptive algorithm for poisson s equation, IAM J. Numer. Anal., 33(1996), pp [5] D. Gilbarg and N.. Trudinger, Elliptic Partial Differential Equations of econd Order, pringer- Verlag, Germany, [6] P. Morin, R.H. Nochetto, and K.G. iebert, Data oscillation and convergence of adaptive FEM, IAM J. Numer. Anal., 38, 2 (2000), pp [7] P. Morin, R.H. Nochetto, and K.G. iebert, Convergence of adaptive finite element methods, IAM Review, 44 (2002), pp [8] R.H. Nochetto, Removing the saturation assumption in a posteriori error analysis,istit. Lombardo ci. Lett. Rend. A, 127 (1993), [9] A.H. chatz, An observation concerning Ritz-Galerkin methods with indefinite bilinear forms, Math. Comp. 28 (1974), pp [10] A. chmidt and K.G. iebert, ALBERT: an adaptive hierarchical finite element toolbox, Documentation, Preprint 06/2000, Universität Freiburg. [11] A. chmidt and K.G. iebert, ALBERT - software for scientific computations and applications, Acta Mathematica Universitatis Comenianae 70 (2001), [12] R. Verfürth, A Review of A Posteriori Error Estimation and Adaptive Mesh-Refinement Technique, Wiley-Teubner, Chichester, 1996.