SIAM J. NUMER. ANAL. Vol. 43, No. 4, pp. 1481–1503 c 2005 Society for Industrial and Applied Mathematics A POSTERIORI ERROR ESTIMATIONS OF SOME CELL-CENTERED FINITE VOLUME METHODS∗ SERGE NICAISE† Abstract. This paper presents the natural framework to residual based a posteriori error estimation of some cell-centered finite volume methods for the Laplace equation in Rd , d = 2 or 3. For that purpose we associate with the finite volume solution a reconstructed approximation, which is a kind of Morley interpolant. The error is then the difference between the exact solution and this Morley interpolant. The residual error estimator is based on the jump of normal and tangential derivatives of the Morley interpolant. We then prove the equivalence between the discrete H 1 seminorm of the error and the residual error estimator. Numerical tests confirm our theoretical results. Key words. finite volume method, cell-centered method, a posteriori error estimates AMS subject classifications. 65N30, 65N15 DOI. 10.1137/S0036142903437787 1. Introduction. The finite volume method is a well-adapted method for the discretization of various partial differential equations and is very popular in the engineering community [24]. The mathematical community recently started to analyze it in detail. Presently, existence and uniqueness results as well as a priori error estimates are available for a quite large class of problems; we refer to [10] and the references cited there. Contrary to the finite element methods [26], a posteriori error estimates for finite volume methods are less developed, and until now only a few results have been obtained in that direction. See [14, 22, 1, 12, 13] for cell-centered finite volume methods, [17, 19, 25, 23] for vertex-centered methods, and [2, 3, 15, 16] for finite volume element methods. Since finite volume methods have some similarities with the finite element methods, we may hope that this gap will be filled soon. The goal of our paper is to present the natural framework to residual based a posteriori (efficient and reliable) error estimation of some cell-centered finite volume methods for linear elliptic equations. In a first attempt we restrict ourselves to the Laplace equation in Rd , d = 2 or 3. The case of diffusion–convection–reaction equations will be only sketched; for details, we refer to a forthcoming paper [20]. The key idea is the reconstruction of a piecewise polynomial approximation of the finite volume solution, its principal property being that the mean of its flux through any edge/face of the mesh is equal to the numerical flux through that edge/face (this interpolant is consequently smoother than the approximated solution). This reconstructed approximation is then a kind of Morley interpolant of the finite volume solution. In general a Morley interpolant is not in H 1 , and therefore the Morley interpolant may be considered as a nonconforming approximation of the exact solution. The second key idea is to use the Helmholtz decomposition of the error, the difference between the exact solution and this Morley interpolant, as was done in [7] for the a posteriori error analysis of a nonconforming finite element approximation of the Laplace equation. As in ∗ Received by the editors November 17, 2003; accepted for publication (in revised form) March 24, 2005; published electronically October 19, 2005. http://www.siam.org/journals/sinum/43-4/43778.html † Université de Valenciennes et du Hainaut Cambrésis, MACS, ISTV, F-59313 Valenciennes Cedex 9, France ([email protected]). 1481 1482 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS [7] the residual error estimator is then naturally based on the jump of normal and tangential derivatives of the Morley interpolant. We finally show the equivalence between the discrete H 1 seminorm of the error and the residual error estimator. The proof of the upper error bound uses the Helmholtz decomposition of the error and some quasiorthogonality relations obtained using the above-mentioned property of the Morley interpolant. The proof of the lower error bound is more standard and simply uses some Green’s formulas and inverse inequalities as for finite element methods [26]. Note that our purposes also require the introduction of new finite elements of Morley type on rectangles and tetrahedra. We further give explicitly the size of the constants appearing in the error estimates by estimating the constants involved in the interpolation error estimates (using some related eigenvalue problems and extension techniques) and in some inverse inequalities. In particular, we obtain constants in the upper error bound that are quite close to unity. The idea to interpolate the finite volume solution by a smoother function having the above-mentioned property on the flux was presented in [12] in an L1 framework for time-dependent nonlinear convection–diffusion equations in Rd × R+ . In that paper the authors obtain a reliable estimator in an L1 -norm, instead of the energy norm. Furthermore, their interpolant is a piecewise linear Lagrange interpolant on a dual mesh. As a consequence, to guarantee the property on the flux, the (primal) mesh has to be admissible in the sense of [10, Def. 9.1], a deep obstacle for adaptivity. To avoid this admissibility condition and use the energy norm framework, we need to use the natural degrees of freedom on the mesh, namely, the mean of the flux on the edges/faces, and consequently use higher order polynomials. The outline of the paper is as follows: In section 2 we describe the so-called cellcentered method for the Laplace equation on a mesh made of triangles, rectangles, or tetrahedra. Some standard inverse inequalities and interpolation error estimates are recalled in section 3, where some constants are specified as explicitly as possible. Section 4 is devoted to the introduction of some finite elements of Morley type. In section 5 we introduce the Morley interpolant of the approximated solution and prove its main properties. The upper and lower error bounds are then deduced in section 6. The upper error bound is based on the properties of the Morley interpolant and the use of the Helmholtz decomposition of the error, while the lower error bound is proved in a quite standard way. In section 7 we briefly describe how to extend our results to diffusion–convection–reaction equations. Finally, section 8 is devoted to numerical experiments that confirm our theoretical considerations. 2. Discretization of the Laplace equation. Let Ω be an open subset of Rd , d = 2 or 3, with a polygonal (d = 2) or polyhedral (d = 3) boundary Γ. As usual, we denote by L2 (·) the Lebesgue spaces and by H s (·), s ≥ 0, the standard Sobolev spaces. If D is an open subset of Rd , d = 2 or 3, the usual norm and seminorm of H s (D) are denoted by · s,D and | · |s,D . For brevity the L2 (D)norm will be denoted by · D and in the case D = Ω, we will drop the index Ω. The space H01 (Ω) is defined, as usual, by H01 (Ω) := {v ∈ H 1 (Ω)/v = 0 on Γ}. In what follows the symbol | · | will denote either the Euclidean norm in Rd , d = 2 or 3, or the length of a line segment, or the area of a plane face, or finally the measure of a domain of Rd . We consider the standard elliptic problem: for f ∈ L2 (Ω) let u ∈ H01 (Ω) be the variational solution of (1) −Δu = f in Ω, SERGE NICAISE 1483 which means that u satisfies (2) ∇u · ∇v dx = f v dx ∀v ∈ H01 (Ω). Ω Ω To approximate this problem by a finite volume scheme we fix a family of meshes Th , h > 0, regular in Ciarlet’s sense [4, p. 124]. In two dimensions we assume that all elements of Th are either triangles or rectangles, while in three dimensions the mesh consists only of tetrahedra. For K ∈ Th we recall that hK is the diameter of K and h = maxK∈Th hK . For any edge/face E of K, we denote by hE,K its height in K, namely, hE,K = d|K| |E| if K is a triangle or a tetrahedron and hE,K = |K| |E| if K is a rectangle. For an edge/face E, its mean height is hE = 12 (hE,K + hE,L ), when E is the edge/face of K and L. The regularity of the mesh implies in particular that for any edge/face E of K one has σ1 hE,K ≤ hK ≤ σ2 hE,K , σ3 hE,K ≤ hE ≤ σ4 hE,K (3) (4) for some positive constants σi , i = 1, . . . , 4, depending on the aspect ratio of Th . Let us define Eh as the set of edges (d = 2) or faces (d = 3) of the triangulation and set Ehint = {E ∈ Eh /E ⊂ Ω} the set of interior edges/faces of Th , while Ehext = Eh \ Ehint is the set of exterior edges/faces of Th . For an edge E of a two-dimensional (2D) element K, introduce nK,E = (nx , ny ) the unit outward normal vector to K along E. Similarly for a face E of a tetrahedron K, set nK,E = (nx , ny , nz ) the unit outward normal vector to K on E. Furthermore, for each edge/face E, we fix one of the two normal vectors and denote it by nE . In two dimensions additionally introduce the tangent vector tK,E = n⊥ K,E := (−ny , nx ) such that it is oriented positively (with respect to K); similarly set tE := n⊥ . E The jump of some function v across an edge/face E at a point y ∈ E is defined as v(y) E := lim v(y + αnE ) − v(y − αnE ) ∀E ∈ Ehint , α→+0 v(y) E := v(y) ∀E ∈ Ehext . For any K ∈ Th or E ∈ Eh , we denote by MK χ and ME χ the mean of χ on K and E, respectively, i.e., 1 1 χ(x) dx ∀K ∈ Th , ME χ = χ(x) ds(x) ∀E ∈ Eh . MK χ = |K| K |E| E Finally, we will need local subdomains (also called patches). As usual, let ωK be the union of all elements having a common edge/face with K. Similarly let ωE be the union of both elements having E as edge/face. The finite volume approximation uh of u is piecewise constant on Th , i.e., uh := (uK )K∈Th (uK being the approximation of u(xK ) for K ∈ Th , xK being the “center” of the box K). To deduce the approximated equation satisfied by uh , we first integrate (1) on a control volume K and use the divergence formula to obtain − (5) ∇u · nK,E ds = f (x) dx ∀K ∈ Th , E∈EK E K 1484 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS where EK is the set of edges/faces of K. The diffusion flux E ∇u · nK,E is approximated by a numerical diffusion flux FK,E (uh ) obtained using quadrature rules and finite differences (see, e.g., [10]) and is consequently a linear combination of some values of uh around E [10, 5, 6]. For our further uses we do not need its exact form but the principle of conservation of flux is required: FK,E (uh ) = −FL,E (uh ) for E = K̄ ∩ L̄. These approximations lead to the following system. Find a solution uh := (uK )K∈Th of (6) FK,E (uh ) = f (x) dx ∀K ∈ Th . − K E∈EK K × xK nL,E E nK,E × xL L Fig. 1. The standard orthogonality condition. If the mesh Th is admissible in the sense of [10, Def. 9.1], i.e., satisfies standard orthogonality conditions (see Figure 1), then the numerical diffusion flux is defined by (7) |E|(uL − uK ) if E = K ∩ L, d(xK , xL ) |E|uK if E ⊂ K ∩ ∂Ω. FK,E (uh ) := − d(xK , Γ) FK,E (uh ) := For general meshes, a possible choice for FK,E (uh ) is proposed in [5, 6] using the diamond cell method. From now on we suppose that system (6) is well defined. This is the case if the mesh Th is admissible in the sense of [10] and if FK,E (uh ) is given by (7) (see, for instance, [10]); while for an arbitrary mesh and the choice of FK,E (uh ) from [5, 6], system (6) is well defined under some geometrical conditions on the mesh [5, 6]. 3. Some analytic tools. 3.1. Bubble functions, extension operator, and inverse inequalities. For our further analysis we require standard bubble functions and extension operators that satisfy certain properties recalled here for the sake of completeness. We need two types of bubble functions, namely, bK and bE associated with an element K and an edge/face E, respectively. For a triangle or a tetrahedron K, denoting by λaK , i = 1, . . . , d + 1, the barycentric coordinates of K and by aE i , i = 1, . . . , d, i 1485 SERGE NICAISE d+1 and the vertices of the edge/face E ⊂ ∂K, we recall that bK = (d + 1)d+1 i=1 λaK i d bE = dd i=1 λaE . i For a rectangle K we here enumerate its vertices in a clockwise sense. Denoting by λaK , i = 1, . . . , 4, the “barycentric” coordinates of K, namely, λaK is the unique i i K element in Q1 (K) such that λaK (a ) = δ , then we recall that b = 8λ λaK and i,j K aK j 1 3 i K K K K (λ + λ ) if the endpoints of the edge E are a and a . bE = 4λaK a2 a3 1 2 1 One recalls that bK = 0 on ∂K, bE = 0 on ∂ωE , bK ∞,K = bE ∞,ωE = 1. In two dimensions for an edge E ⊂ ∂K using temporarily the local coordinates system (x, y) such that E is included into the x-axis, then the extension Fext (vE ) of vE ∈ C(E) to K is defined by Fext (vE )(x, y) = vE (x). We proceed similarly in three dimensions. Now we may recall the so-called inverse inequalities, whose proof uses classical scaling techniques and the fact that all norms are equivalent in a finite-dimensional space [26]. Lemma 3.1 (inverse inequalities). Let K ∈ Th , E ∈ EK , vK ∈ Pk0 (K), and vE ∈ Pk1 (E) for some nonnegative integers k0 and k1 . Then there exist positive constants β0 , β1 (resp., α0 , α1 , and α2 ) depending on the form of K (triangle, rectangle, or tetrahedron), on the aspect ratio of the mesh Th , and on the polynomial degree k0 (resp., k1 ) such that (8) (9) (10) (11) (12) 1/2 1/2 vK bK 2K ≤ vK 2K ≤ β0 vK bK 2K , 2 ∇(vK bK )2K ≤ β1 h−2 K vK K , 1/2 1/2 vE bE 2E ≤ vE 2E ≤ α0 vE bE 2E , Fext (vE )bE 2K ≤ α1 hK vE 2E , 2 ∇(Fext (vE )bE )2K ≤ α2 h−1 K vE E . Remark 3.2. In the above lemma, if K is a square and k1 = 2, then α0 = √ √ 8(6+ 21) 8(56+ 881) ≈ 1.967, α1 = ≈ 0.269, α2 = ≈ 6.528. These numbers 315 105 are obtained by reducing estimates (10)–(12) to an eigenvalue problem. Namely using the standard basis of P2 , estimate (12) is equivalent to (AX, X) ≤ α2 (BX, X) for all X ∈ R3 , where A and B are two explicit 3 × 3 matrices. Therefore, α2 is the largest eigenvalue of the matrix B −1/2 · A · B 1/2 , or equivalently the largest eigenvalue of the matrix B −1 · A, since B is invertible. A direct calculation yields the value of α2 . The other estimates are proved in the same manner. √ 10+ 30 4 3.2. Interpolation error estimates. Here we collect some standard interpolation error estimates but we specify as explicitly as possible the involved constants. As usual we start with the reference elements, which are the unit triangle K̂ of vertices (0, 0), (1, 0), (0, 1), the unit square K̂ of vertices (0, 0), (1, 0), (0, 1), (1, 1), or the unit tetrahedron K̂ of vertices (0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1). Lemma 3.3. Let Ê be the edge/face of K̂ included in the axis/plane xd = 0. Then there exist two positive constants μ and α such that for all v ∈ H 1 (K̂), the following estimates hold: (13) (14) v − MK̂ vK̂ ≤ μ∇vK̂ , v − MÊ vÊ ≤ α∇vK̂ . 1 ≈ 0.565244. If K̂ is If K̂ is the reference square, then μ = π1 and α = √π tanh π 1 1 2 the reference triangle, then μ = π and α = μ1 , where μ1 is the first positive root of 1486 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS the transcendental equation (15) sinh x + tan x = 0. √ 1/4 (2(11+4 6)) √ This means that α≈0.730276. √ If K̂ is the reference tetrahedron, then μ ≤ 3π (2(11+4 6))1/4 ≈ 0.466715 and α≤ √π tanh π ≈ 1.43549. Proof. The two estimates are Poincaré-like inequalities and follow from the Bramble–Hilbert lemma. But this argument does not give an estimate for μ and α. Therefore, we argue as follows. For the first estimate, denote by λ21 the first positive eigenvalue of the Laplace operator on K̂ with Neumann boundary2 conditions. Then ∇v by the min-max principle, we know that λ21 = min v∈H 1 (K̂) v 2K̂ . This identity is v=0,M v=0 K̂ K̂ equivalent to λ21 v2K̂ ≤ ∇v2K̂ ∀v ∈ H 1 (K̂) : MK̂ v = 0. We then obtain (13) with μ = λ−1 1 . If K̂ is the unit square, it is well known that λ21 = π 2 and consequently μ = π1 . If K̂ is the reference triangle, we use the following extension operator from K̂ to the unit square (0, 1)2 , temporarily denoted by Ŝ. Namely, for v ∈ H 1 (K̂), we define its extension Ev to Ŝ by Ev(y1 , y2 ) = v(y1 , y2 ) if (y1 , y2 ) ∈ K̂, Ev(y1 , y2 ) = v(1 − y2 , 1 − y1 ) if (y1 , y2 ) ∈ Ŝ \ K̂. Note that Ev ∈ H 1 (Ŝ) and from v2K̂ = v2Ŝ , ∇v2K̂ = ∇v2Ŝ , we easily get min 1 ∇v2K̂ v∈H (K̂) v=0,M v=0 K̂ v2K̂ ≥ min 1 v∈H (Ŝ) v=0,M v=0 Ŝ ∇v2Ŝ v2Ŝ = π2 . √ On the other hand, one readily checks that the function ψ(x, y) = 2(cos(πx) + cos(π(1 − y)) is an eigenvector of the eigenvalue π 2 of the Laplace operator with Neumann boundary conditions in K̂. Therefore, we actually have λ21 = π 2 . We use a similar argument for the reference tetrahedron. Namely we use an extension operator from K̂ to the standard reference pentahedron P̂ = Ê × (0, 1). For that purpose denote by K̂2 and K̂3 the tetrahedra of vertices (1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 0, 1) and (0, 1, 0), (0, 0, 1), (1, 0, 1), (0, 1, 1), respectively. We remark that P̂ = K̂ ∪ K̂2 ∪ K̂3 , that K̂ and K̂2 have a common face, and similarly that K̂2 and K̂3 have a common face. Note further that |K̂| = |K̂2 | = |K̂3 | = 16 . Therefore, as before there exists an affine transformation F1 which maps K̂ onto K̂2 and let their common face be invariant. Similarly denote by F2 the affine transformation which maps K̂2 onto K̂3 and let their common face be invariant. Denote by Ai , i = 1, 2, the 3 × 3 matrices and by bi ∈ R3 , i = 1, 2, such that Fi (x) = Ai x + bi for all x ∈ R3 . Now we are able to define the extension operator E: for v ∈ H 1 (K̂), we define Ev(y) = v(y), Ev(y) = Ev(y) = v(F1−1 (y)) if y ∈ K̂, v(F1−1 (F2−1 (y))) if y ∈ K̂2 , if y ∈ K̂3 . Using the above properties between the tetrahedra K̂, K̂2 , and K̂3 and somechanges of variables, we readily check that Ev ∈ H 1 (P̂ ) and satisfies P̂ Ev(y) dy = 3 K̂ v(x) dx, and |Ev(y)|2 dy = 3 |v(x)|2 dx, |∇Ev(y)|2 dy = ∇v(x) · T · ∇v(x) dx, P̂ K̂ P̂ K̂ 1487 SERGE NICAISE where the matrix T is given by ⎛ −1 −1 − T = Id + (A + A−1 A1 1 A1 ) 1 (A2 A2 ) −2 3 −1 5 = ⎝−2 0 ⎞ 0 −1⎠ . 5 √ An easy calculation yields T 2 = 2(11 + 4 6) ≈ 6.44949. These identities directly lead to ∇v2K̂ min 1 v∈H (K̂) v=0,M v=0 K̂ v2K̂ ≥ 3 T 2 min 1 ∇v2P̂ v∈H (P̂ ) v=0,M v=0 P̂ v2P̂ = 3π 2 . T 2 2 We then conclude that μ ≤ T 3π 2 . For the second estimate, we start with the following nonstandard eigenvalue problem in the unit square K̂ = (0, 1)2 . Find λ2 and v ∈ H 1 (K̂) solution of ∇v · ∇w = λ2 vw ∀w ∈ H 1 (K̂). (16) K̂ Ê For this eigenvalue problem, let us show that the min-max principle holds at least for the first positive eigenvalue λ̃21 . Namely λ̃21 is characterized by λ̃21 = (17) ∇v2K̂ min 1 v∈H (K̂) =0,M v=0 Ê Ê v v2Ê . Denote by m the above right-hand side. Consider a minimizing sequence vn of the above minimum, namely, for all n ∈ N, let vn ∈ H 1 (K̂) be such that MÊ vn = 0, vn Ê = 1, ∇vn 2K̂ → m as n → ∞. Since ∇vK̂ + vÊ is a norm on H 1 (K̂) equivalent to the standard norm, the sequence (vn )n is bounded in H 1 (K̂). Therefore, there exists a subsequence, still denoted by (vn )n , such that vn → v in H 1 (K̂) as n → ∞. From the above properties of the sequence (vn )n , we deduce that v satisfies MÊ v = 0, vÊ = 1, ∇v2K̂ = m. It remains to show that v is an eigenvector of problem (16) corresponding to the eigenvalue m. For that purpose let us fix z ∈ H 1 (K̂) such that MÊ z = 0 and (18) vz = 0. Ê Consider the mapping Φ:R→R:α→ ∇(v + αz)2K̂ 1 + α2 z2K̂ . 1488 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS From the above minimization problem and the properties of v, the mapping Φ hits its minimum at α = 0. Since Φ is smooth, we deduce that Φ (0) = 0, or (19) ∇v · ∇z = 0. K̂ Since any w ∈ H 1 (K̂) such that MÊ w = 0 may be written in the form w = βv + z, with β ∈ R and z ∈ H 1 (K̂) satisfying (18), we deduce that ∇v · ∇w = β ∇v · ∇v + ∇v · ∇z. K̂ K̂ K̂ By the properties of v and identity (19), we conclude that ∇v · ∇w = βm = m vw; K̂ Ê this last identity follows from (18). To find the first eigenvalue of problem (16), we remark that its strong form is ⎧ ⎪ ⎨−Δv = 0 in K̂, ∂v = λ2 v on Ê, ∂n ⎪ ⎩ ∂v = 0 on ∂ K̂ \ Ê. ∂n Using the standard argument of separation of variables, one finds a family of eigenvalues, its smallest one being π tanh π ≈ 3.12988. In order to be sure that this value is the smallest eigenvalue of problem (16), we penalize it by an integral term in K̂. Namely, for any > 0, we consider the problem 2 (20) ∇v · ∇w = λ vw + vw ∀w ∈ H 1 (K̂). K̂ Ê K̂ This problem is an eigenvalue problem related to a selfadjoint nonnegative operator. For that problem one can find all the eigenvalues by separation of variables. Since the eigenvalues of (20) depend continously on , the first positive eigenvalue λ̃21, tends to the first eigenvalue of problem (16). By direct calculations one shows that λ̃21, → π tanh π. We therefore conclude that λ̃21 = π tanh π. By the above “min-max” principle (17), we deduce that α = 1/λ̃1 . For the unit triangle, we start with the minization problem (17) (as before λ̃21 is the first positive eigenvalue of problem (16)). Using the extension operator E from K̂ to Ŝ, we deduce that λ̃21 = min 1 v∈H (K̂) =0,M v=0 Ê Ê v ∇v2K̂ v2Ê ≥ min 1 v∈H (Ŝ) +v =0,M v=0 Ê F̂ Ê∪F̂ v ∇v2Ŝ v2Ê + v2F̂ , where F̂ is the edge of Ŝ included into the line x1 = 1. The right-hand side is related to the eigenvalue problem: find λ2 and v ∈ H 1 (Ŝ) solution of (21) ∇v · ∇w = λ2 vw ∀w ∈ H 1 (Ŝ). Ŝ Ê∪F̂ 1489 SERGE NICAISE The same arguments as before give, as first positive eigenvalue μ21 , the first positive root of the transcendental equation (15). As for the first estimate we deduce that μ21 is the first positive eigenvalue of problem (16). Indeed, if w(x1 , x2 ) is the eigenvector of problem (21) associated with the eigenvalue μ21 , then one readily checks that v(x1 , x2 ) = w(x1 , x2 )−w(1−x2 , 1−x1 ) is an eigenvector of problem (16) associated with the eigenvalue μ21 . As before the situation is not so convenient for the unit tetrahedron Therefore, we first state the following estimate on the reference prism P̂ : v − MÊ vÊ ≤ √ 1 ∇vP̂ π tanh π ∀v ∈ H 1 (P̂ ), obtained as for the unit square. Now, using the extension operator E and this estimate, we may write v − MÊ vÊ = Ev − MÊ EvÊ ≤ √ 1 1 T 2 ∇vK̂ ; ∇EvP̂ ≤ √ π tanh π π tanh π this last estimate follows from the above properties of Ev. Remark 3.4. To our knowledge, the exact value of μ is not explicitly known for the unit tetrahedron. Numerical tests give for λ21 the approximated value λ21 ≈ 14.444208445. This gives for μ the approximated value μ ≈ 0.26312, which is relatively smaller than our theoretical upper bound. Similarly the exact value of α is not explicitly known for the unit tetrahedron; an approximated value is 0.340355, and therefore our theoretical upper bound is far from being optimal. In the above arguments, the main difference between the unit triangle and the unit tetrahedron concerns the extension operator. For the triangle, the extension operator uses an orthogonal transformation, which is impossible for the unit tetrahedron. That last case still requires more investigations. The above lemma and scaling arguments lead to the following lemma. Lemma 3.5. There exist two positive constants μ and α depending on K̂ such that for all K ∈ Th and v ∈ H 1 (K), the following estimates hold: (22) v − MK vK ≤ μρ̂−1 hK ∇vK , (23) v − ME vE ≤ αρ̂−1 hE,K hK ∇vK , −1/2 where E is an edge/face of K, and ρ̂ is the diameter of the inscribed ball of K̂. 4. Some finite elements of Morley type. As already mentioned the main idea of our a posteriori error analysis is to use an interpolant p satisfying ∂p ds = FK,E (uh ) ∀E ∈ EK . E ∂nK,E This means that we need to use finite elements having as degrees of freedom the mean of the normal derivative of p on each edge/face. The simplest element is the so-called Morley triangle [18, 4] usually used for the approximation of the plate problem. For our further uses we extend this kind of elements to rectangles and tetrahedra. We start by recalling the Morley triangle as well as a recent extension due to Nilssen, Tai, and Winther [21] and then introduce our new elements. 1490 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS 4.1. Triangles. Here K is a (nondegenerate) triangle with vertices aK i , i = 1, 2, Nf := 3. The standard Morley triangle is defined by the triple (K, PK , ΣK ) [18, 4], where PK = P2 (K) and (24) ΣK = {p(ai )}i=1,...,Nf ∪ E ∂p ds ∂nK,E . E∈EK Note that this element is not a C 0 -element; an extension which has this property was recently built in [21, sect. 4], where they take PK = P2 (K) ⊕ P1 (K)bK = {q + pbK : q ∈ P2 (K), p ∈ P1 (K)}, ∂p ΣK = {p(aK )} ∪ {p(m )} ∪ ds . i=1,2,3 E E∈E i K E ∂nK,E E∈EK 4.2. Rectangles. Here K is a (nondegenerate) rectangle of vertices aK i , i = 1, . . . , Nf := 4. The first element is defined by PK = P2 (K) ⊕ Span {x3 − 3xy 2 , y 3 − 3yx2 } with degrees of freedom ΣK defined by (24). We readily check that the triple (K, PK , ΣK ) is a finite element. The above choice is motivated by the fact that Δq ∈ R ∀q ∈ PK , since x3 − 3xy 2 and y 3 − 3yx2 are the unique homogeneous polynomials of degree 3 which are harmonic. The second example is to take PK = Q2 (K) and ΣK := {p(aK i )}i=1,...,5 ∪ K { E ∂n∂p ds} , where a is the center of gravity of K. E∈E 5 K K,E 4.3. Tetrahedra. Here K is a (nondegenerate) tetrahedron with vertices aK i , i = 1, 2, 3, Nf := 4. Inspired from the second triangular example from [21] we choose PK = P1 (K) ⊕ P1 (K)bK = {q + pbK : p, q ∈ P1 (K)}, and ΣK defined by (24). Similar to Lemma 4.1 of [21] (adapted to our setting) we can prove the following lemma. Lemma 4.1. The above triple (K, PK , ΣK ) is a C 0 -finite element. 5. The Morley interpolant. 5.1. Definition. For any vertex a of the triangulation we fix (wK (a))K∈Th :a∈K suitable weights of interpolation around K. Since our analysis below is independent of their choice, we do not describe them. They may be obtained using a discrete projection of piecewise constant functions over affine functions on ωa [5, 6], a standard technique to get a recovered gradient at the vertex a, leading further to the P1 exactness. Namely for any vertex a the weights wK (a) may be fixed such that w(a) = w (a)uK , where w ∈ P1 (ωa ) is the discrete projection of uh on P1 (ωa ), i.e., K K⊂ωa w ∈ P1 (ωa ) is the unique minimizer of |q(xK ) − uK |2 , q ∈ P1 (ωa ). K⊂ωa This choice implies that if uh were P1 (ωa ), then we would have w = uh in ωa . For instance, if ωa is made of four squares, then this choice yields wK (a) = 1/4. 1491 SERGE NICAISE We now introduce the Morley finite element space Vh := vh ∈ L2 (Ω) : vh|K ∈ PK ∀K ∈ Th , L K L vh|K (aK i ) = vh|L (aj ) ∀K, L ∈ Th , i, j ∈ {1, . . . , Nf } : ai = aj , K vh|K (aK i ) = 0 ∀K, L ∈ Th , i ∈ {1, . . . , Nf } : ai ∈ Γ, ∂vh|K ∂vh|L ds = ds ∀E ∈ Eh , K, L ∈ Th : E = K ∩ L . E ∂nE E ∂nE Since Vh is not necessarily included into H01 (Ω), the space Vh is equipped with the norm | · |1,h := ( K∈Th | · |21,K )1/2 . Notice that Vh is indeed included into H01 (Ω) for the second-triangular example and for our three-dimensional (3D) example. Definition 5.1. For uh = (uK )K∈Th , we define its Morley interpolant IM uh as the unique element vh in Vh satisfying (25) vh|K (aK wL (aK ∀K ∈ Th , i ∈ {1, . . . , Nf } : aK i )= i )uL i ∈ Ω, L∈Th :aK ∈L i (26) vh|K (aK i ) (27) E =0 ∀K ∈ Th , ∂vh|K ds = FK,E (uh ) ∂nK,E i ∈ {1, . . . , Nf } : aK i ∈ Γ, ∀E ∈ EK , K ∈ Th . For the second triangular element we have to add the conditions vh|K (mE ) = vh|L (mE ) = 12 (uK + uL ) ∀E ∈ Eh , K, L ∈ Th : E = K ∩ L, vh|K (mE ) = 0 ∀E ∈ Eh , K ∈ Th : E ⊂ K ∩ Γ. Similarly for the first-rectangular element we must add vh|K (aK 5 ) = uK for all K ∈ Th . 5.2. Some useful properties. We first prove a basic property of the Morley interpolant. Lemma 5.2. If uh is solution of (6), then IM uh satisfies (28) Δ(IM uh ) dx = − f (x) dx ∀K ∈ Th . K K Proof. By Green’s formula and property (27) satisfied by IM uh , we have ∂(IM uh ) Δ(IM uh ) dx = ds = FK,E (uh ), K E ∂nK,E E∈EK E∈EK and we conclude by (6). Now we prove some quasi-orthogonality relations that will be used for the upper error bound. We first define the gradient jump of IM uh in the normal and tangential direction by JE,n (uh ) = ∂n∂E (IM uh ) E ∀E ∈ Ehint , ∂ ∀E ∈ Eh for nonconforming 2D cases, ∂tE (IM uh ) E JE,t (uh ) = 0 ∀E ∈ Eh for conforming cases. Lemma 5.3. If u is a solution of (2) and uh is a solution of (6), then (29) ∇(u − IM uh ) · ∇χ dx = (f + ΔIM uh )(χ − MK χ) dx K K∈Th K K∈T h − JE,n (uh )(χ − ME χ) ds ∀χ ∈ H01 (Ω). int E∈Eh E 1492 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS Proof. For brevity denote the left-hand side of (29) by I1 (χ). By (2) and Green’s formula on each triangle K, and recalling that χ ∈ H01 (Ω), we may write ∂(IM uh ) I1 (χ) = f χ dx + Δ(IM uh )χ dx − χ ds ∂nK Ω K∈Th K K∈Th ∂K = (f + Δ(IM uh ))χ dx − JE,n (uh )χ ds. K K∈Th int E∈Eh E Using identity (28), we arrive at I1 (χ) = (f + Δ(IM uh ))(χ − MK χ) dx − JE,n (uh )χ ds. K∈Th K int E∈Eh E The conclusion now follows from the fact that E JE,n (uh ) ds = 0, for all E ∈ Ehint , due to (27) and the principle of conservation of flux, FK,E (uh ) = −FL,E (uh ), if E = K ∩ L, K, L ∈ Th . Corollary 5.4. Under the assumptions of Lemma 5.3 the next estimate holds (30) |I1 (χ)| ≤ √ μ2 2 2 2 hK f + Δ(IM uh )2K ρ̂ K∈Th α Nf 4ρ̂2 2 + 2 2 h−1 E,K hK JE,n (uh )E 1/2 |χ|1,Ω . K∈Th E∈E int ∩EK h Proof. Identity (29) and Cauchy–Schwarz’s inequality yield f + Δ(IM uh )K χ − MK χK + JE,n (uh )E χ − ME χE |I1 (χ)| ≤ int E∈Eh K∈Th ≤ f + Δ(IM uh )K χ − MK χK K∈Th + 1 2 JE,n (uh )E χ − ME χE . K∈Th E∈E int ∩EK h By the interpolation error estimates (22) and (23), we obtain μ α −1/2 |I1 (χ)| ≤ hK hE,K JE,n (uh )E |χ|1,K . f + Δ(IM uh )K + ρ̂ 2ρ̂ int E∈Eh ∩EK K∈Th We the discrete Cauchy–Schwarz’s inequality and the well-known estimate lconclude by l ( i=1 ai )2 ≤ l i=1 a2i , valid for l = 2, 3, 4 and all real numbers ai . Lemma 5.5. Assume that d = 2. If u is the solution of (2) and uh is the solution of (6), then (31) ∇(u−IM uh )·curl g dx = JE,t (uh )(g−ME g) ds ∀g ∈ H 1 (Ω), K∈Th K E∈Eh E where curl g = (∂2 g, −∂1 g) is the vectorial curl of g. 1493 SERGE NICAISE Proof. Denote the left-hand side of (31) by I2 (g). Green’s formula on each element K leads to (see Theorem I.2.11 of [11]) ∂ (u − IM uh )g ds = JE,t (uh )g ds, I2 (g) = − ∂K ∂t E K∈Th E∈Eh since u ∈ H01 (Ω) and g ∈ H 1 (Ω). The conclusion follows from the property (32) JE,t (uh ) ds = 0. E Indeed, if aiE , i = 1, 2, are the two extremities of E, we have E JE,t (uh ) ds = uh E (a1E ) − uh E (a2E ). Using properties (25) and (26), we have uh E (aiE ) = 0, i = 1, 2, and therefore (32) holds. Corollary 5.6. Under the assumptions of Lemma 5.5 the following estimate holds: 1/2 α −1 2 2 |I2 (g)| ≤ (33) Nf hE,K hK JE,t (uh )E |g|1,Ω . ρ̂ K∈Th E∈EK Remark 5.7. The above fundamental properties are only based on the definition of the scheme (6), the continuity of the interpolant at the interior nodes, the property (26), and the interpolation property (27). Therefore, our further analysis works for any finite element (K, PK , ΣK ) such that the associated interpolant satisfies these properties. But the finite element and the definition of the interpolant should be well chosen in order to guarantee the convergence of IM uh to the exact solution u. That is the reason of the introduction of the weights wK (a) in (25) since it was shown in [5, 6] that for a triangulation made of rectangles, the choice of the weights described at the beginning of section 5.1 guarantees the convergence of uh to u. Convergence analysis for arbitrary triangulations and appropriate weights is still to be done, but it is outside the scope of this paper. 6. Error estimators. 6.1. Residual error estimators. The exact element residual is defined by RK := f + ΔIM uh on K. As usual this exact residual is replaced by some finitedimensional approximation called approximate element residual rK ∈ Pk (K). A realistic choice is to take rK = MK f + ΔIM uh since in the case ΔIM uh ∈ R we have (thanks to Lemma 5.2) rK = 0. Definition 6.1 (residual error estimator). The local and global residual error estimators and approximation terms are defined by −1 −1 2 2 2 2 2 hE,K JE,n (uh )E + hE,K JE,t (uh )E , ηK := hK rK K + η 2 := int E∈EK ∩Eh E∈EK 2 ηK , K∈Th 2 ζK := K ⊂ωK h2K RK − rK 2K , ζ 2 := K∈Th 2 ζK . 1494 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS 6.2. Upper error bound. Theorem 6.2. Let u be a solution of (2) and let uh be a solution of (6). Then the error e := u − IM uh is bounded by |e|1,h ≤ cup (η + ζ), (34) 1/2 where cup = ρ̂1 max{2μ, αNf }. Proof. Denote by ∇h e the brokent gradient of e, namely, (∇h e)|K = ∇e|K on K ∀K ∈ Th . As in Theorem 3.1 of [7] we use its Helmholtz decomposition ∇h e = ∇χ + curl ψ, (35) with χ ∈ H01 (Ω) and ψ ∈ H 1 (Ω) if d = 2 and χ = e and ψ = 0 if d = 3 such that |χ|21,Ω + |ψ|21,Ω ≤ |e|21,h . (36) This estimate is direct in three dimensions, while in two dimensions it directly follows from the identity Ω ∇χ · curl ψ = 0, a consequence of Green’s formula (recalling that χ = 0 on the boundary). Owing to identity (35) we may write (with the notation from Lemmas 5.3 and 5.5) 2 2 |∇h e| dx = ∇h e · (∇χ + curl ψ) dx = I1 (χ) + I2 (ψ). |e|1,h = Ω Ω Using estimates (30) and (33), we obtain 1/2 1/2 √ μ2 αNf α 2 Nf 2 ηt |ψ|1,Ω , η |χ| + |e|21,h ≤ 2 2 Ξ2 + 1,Ω ρ̂ 4ρ̂2 n ρ̂ where for brevity we set Ξ2 = K∈Th h2K f + Δ(IM uh )2K and 2 2 2 2 2 h−1 h J (u ) , η = h−1 ηn2 = E,n h K E t E,K E,K hK JE,t (uh )E . K∈Th E∈E int ∩EK h K∈Th E∈EK By the discrete Cauchy–Schwarz inequality and estimate (36), we obtain |e|21,h ≤ 2μ2 2 α2 Nf 2 α2 Nf 2 4μ2 2 Ξ + η + η ≤ ζ n t ρ̂2 2ρ̂2 ρ̂2 ρ̂2 4μ2 2 α 2 Nf 2 + 2 hK rK 2K + (ηn + ηt2 ); ρ̂ ρ̂2 K∈Th this last estimate follows from the well-known estimate (a + b)2 ≤ 2a2 + 2b2 , valid for all real numbers a, b. By the definition of cup and η the above estimate implies that |e|1,h ≤ cup (ξ 2 + η 2 )1/2 ≤ cup (η + ξ). Remark 6.3. Thanks to Lemma 3.3, we can estimate the constant cup appearing in the above upper bound. For a triangulation made of rectangles, then cup = 2 max{ π1 , α} = 2α ≈ 1.13049. For a triangulation made of triangles, then √ √ cup = ρ̂1 max{ π2 , 3α} = ρ̂3α ≈ 2.15928. Finally for a mesh made of tetrahedra, one √ 1/4 6)) √ has cup ≤ 2 (2(11+4 ≈ 9.27912. For that last case, by Remark 3.4, a numerical π tanh π upper bound for cup is 2.20009. In both cases, the exact value, or the numerical upper bound for the tetrahedral case, of cup is quite close to unity. 1495 SERGE NICAISE 6.3. Lower error bound. Theorem 6.4. For all elements K, the following local lower error bound holds (37) ηK ≤ clow (∇h eωK + ζK ), where c2low = max{2β02 β1 + 2Nf α02 σ1−1 σ22 σ3 (3α2 σ4−1 + 8α1 β02 β1 σ1−1 σ4 ), 2β02 + 8Nf α02 α1 (1 +2β02 )σ1−1 σ22 σ3 σ4 }. Proof. Element residual. For a fixed element K denote wK = rK bK which belongs to H01 (K). From the definition of RK and integration by parts, we may write rK wK = (rK − RK )wK − Δ(u − IM uh )wK K K K = (rK − RK )wK + ∇e · ∇wK . K K By Cauchy–Schwarz’s inequality and the inverse inequalities (8) and (9), we conclude that (38) 1/2 hK rK K ≤ β0 (β1 ∇eK + hK rK − RK K ). Normal jump. Fix an arbitrary E ∈ Ehint . Recall that JE,n (uh ) ∈ Pk (E) for some k ∈ N and set wE := Fext (JE,n (uh ))bE ∈ H01 (ωE ). By elementwise partial integration, we get ∂e JE,n (uh )wE = − wE = − (∇e · ∇wE + ΔewE ) dx ∂nK E K⊂ωE ∂K K⊂ωE K ΔeK wE K . ≤ ∇h eωE ∇wE ωE + K⊂ωE Inequalities (10)–(12) and properties (3) and (4) in the previous estimate lead to −1 2 2 −1 2 2 2 hK RK K . hE JE,n (uh )E ≤ 2α0 σ1 2α2 σ4 ∇h eωE + 2α1 σ4 K⊂ωE By estimate (38) we arrive at (39) hE JE,n (uh )2E ≤ 4α02 σ1−1 (α2 σ4−1 + 4α1 σ4 β02 β1 )∇h e2ωE h2K RK − rK 2K . + 8α02 α1 σ1−1 σ4 (1 + 2β02 ) K⊂ωE Tangential jump (in two dimensions). For a fixed edge E set wE := Fext (JE,t (uh )) bE ∈ H01 (ωE ). For u ∈ H 1 (ωE ) and wE ∈ H01 (ωE ), partial integration leads to ∂u 0= ∇u · curl wE . wE = − ∂ωE ∂t ωE For IM uh we integrate elementwise and obtain using the above identity ∂(IM uh ) wE = JE,t (uh )wE = − ∇(IM uh ) · curl wE ∂t E K⊂ωE ∂K K⊂ωE K =− ∇(u − IM uh ) · curl wE ≤ ∇h eωE ∇wE ωE . K⊂ωE K 1496 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS The inverse inequalities (10) and (12), as well as (3) and (4), lead to hE JE,t (uh )2E ≤ 2α02 α2 σ1−1 σ4−1 ∇h e2ωE . (40) The conclusion follows from estimates (38)–(40) and properties (3) and (4). Remark 6.5. In the above proof, we see that if rK = 0, then the constant clow reduces to c2low = 2Nf α02 σ1−1 σ22 σ3 max{3α2 σ4−1 , 2α1 σ4 }. Let us illustrate this constant clow in the particular case considered in the next section. Take a triangulation made of squares, build the Morley interpolant with the help of the first rectangular element from section 4, and choose rK = 0. Then by Remark 3.2, we have clow = 24α02 α2 ≈ 24.622. 7. Diffusion–convection–reaction equations. In this section we describe how to extend the above results to diffusion–convection–reaction equations; for details we refer the reader to [20]. Consider the linearized diffusion–convection–reaction problem: for f ∈ L2 (Ω) let u ∈ H01 (Ω) be the unique solution of (41) Au := div (−∇u + vu) + bu = f in Ω, where is a fixed positive constant, v ∈ Rd , and b is a nonnegative real number. Integrating (41) on a control volume K and using the divergence formula, we obtain (−∇u + vu) · nK,E ds + bu dx = f (x) dx ∀K ∈ Th . E∈EK E K K The continuous diffusion flux −∇u · nK,E is approximated as before, the convection flux vu · nK,E by a first order upwind scheme, and the reaction term K u by a simple quadrature formula (see [10]). These approximations lead to the following system. Find uh := (uK )K∈Th , the solution of R −FK,E (uh ) + vK,E FEC (uh ) + bFK (42) (uh ) = f (x) dx ∀K ∈ Th , K E∈EK where vK,E = v · nK,E , the quantity FK,E (uh ) is supposed to satisfy the principle R of conservation of flux, while the quantities FEC (uh ) and FK (uh ) are, respectively, defined by (43) FEC (uh ) := |E|uE,+ , where for E ∈ Ehint , uE,+ = uKE,+ , KE,+ being the upstream control volume, i.e., vKE,+ ,E ≥ 0; while for E ∈ K̄ ∩ Γ, uE,+ = uK if vK,E ≥ 0, and uE,+ = 0 else. R FK (uh ) = |K|uK . For a restricted admissible mesh in the sense of [10, Def. 9.4], if the numerical diffusion fluxes FK,E (uh ) are given by (7), then system (42) is well defined as proved in [9]. For a general mesh as here, we simply assume that system (42) has a unique solution. As for the Laplace operator, we associate with the finite volume solution uh its Morley interpolant IM uh . This interpolant is related to the quantities involved in (42), namely, the diffusion and convection fluxes, and the reaction term. For that purpose, for each element K, we build a C 0 -finite element (K, PK , ΣK ) having as 1497 SERGE NICAISE degrees of freedom the mean of the normal derivative and of the function on each edge, as well as the mean on K. For instance, if K is a triangle, we may take PK = {q + (p + αbK )bK : q ∈ P2 (K), p ∈ P1 (K), α ∈ R}, K ΣK = p(ai ) ∪ p ∪ p ds ∪ K i=1,2,3 E E∈EK E ∂p ds ∂nK,E . E∈EK Then for uh = (uK )K∈Th , we define its interpolant IM uh as the unique element vh in Vh ∩ H01 (Ω) satisfying (25)–(27) and R vh ds = FEC (uh ) ∀E ∈ Ehint , vh dx = FK (uh ) ∀K ∈ Th . E K The key point of our a posteriori analysis is the following basic property of the Morley interpolant, obtained using Green’s formula and the above properties of IM uh . Lemma 7.1. If uh is a solution of (42), then IM uh satisfies (A(IM uh ) − f ) dx = 0 ∀K ∈ Th : measd−1 (K ∩ Γ) = 0. K This property and similar arguments to those used before allow us to prove the following error bounds. Theorem 7.2. Let u be a solution of (41), and let uh be a solution of (42). Then the error is bounded by 2 |∇e| + (44) Ω 1/2 1 div v + b |e|2 ≤ c1 (η + ζ), 2 where c1 is a positive constant depending on the aspect ratio of the mesh and of the size of . For all elements K, the following local lower error bound holds: (45) ηK ≤ c2 |∇e|2 + ωK 1/2 1 div v + b |e|2 + ζK , 2 where c2 is a positive constant depending on the aspect ratio of the mesh and of the size of . Remark 7.3. 1. By modifying appropriately the estimator ηK , we may skip the dependence of c1 with respect to and give explicitly the dependence of c2 on this parameter; see [20]. 2. In the case of a large Peclet number P e ≡ −1 |v| and/or large number Γ ≡ −1 b, problem (41) is singularly perturbed and the solution may generate sharp boundary or interior layers, where the solution of the limit problem (corresponding to = 0) is not smooth or does not satisfy the Dirichlet boundary condition. In that case, the use of anisotropic meshes is recommended. This will be addressed in [20]. 8. Numerical results. In this section we present two numerical tests that illustrate the efficiency and reliability of our estimator. The second example further indicates that our estimator is appropriate for adaptivity. Additionally, for both examples we provide the order of convergence of the error |e|1,h ; both cases confirm that IM uh is a good approximation of u. 1498 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS 8.1. A smooth solution. The first example is for a smooth solution in the unit square ]0, 1[2 and quasi-uniform meshes made of squares. Namely, we consider problem (1) in ]0, 1[2 with the following prescribed exact solution u(x, y) = xy(1−x)(1−y). The meshes are uniform ones made of squares of size h = 1/n for n = 4, 8, . . . , 256 obtained by dividing each segment in n subintervals. Since the meshes are made of squares we use the scheme (6) with the numerical diffusion flux given by (7); furthermore, the Morley interpolant is built using the first-rectangular element from subsection 4.2 and the weights wK (a) = 1/4. We first investigate the order of convergence of the approximated solution uh as well as its interpolant IM uh to the exact solution u in different norms. Namely, we present in Figure 2 the following norms: ū − uh (where ū is piecewise constant on Th and is equal to u(xK ) on each K); |u − uh |1,Th (where the mesh-depending norm | · |1,Th is defined in [10, Def. 9.3]); u − IM uh ; and |u − IM uh |1,h . These quantities are illustrated in Figure 2 by lines 1–4, respectively, in a double logarithmic scale so that the slope of the curve corresponds to the order of convergence. Theorem 9.3 of [10] yields the order of convergence 1 for |u − uh |1,Th . Figure 2 even reveals a better order of convergence of about 1.5, probably due to the smoothness of u. For the L2 -norms, we remark a quadratic order of convergence, a usual phenomenon. On the other hand, for the discrete H 1 -norm of the reconstructed approximation, we also see a quadratic order of convergence. This seems to be a superconvergence effect, probably due to the smoothness of the solution and of the use of structured meshes. log(error) −3.15 1 −4.35 1 −5.54 −6.74 −7.94 −9.13 (2) −10.33 −11.52 2 (4) −12.72 1 (3) −13.91 log(n) (1) −15.11 0.39 1.00 1.62 2.23 2.85 3.47 4.08 4.70 5.31 5.93 6.55 Fig. 2. Illustration of different norms for test 1. Now we investigate the main theoretical results which are the upper and lower error bounds (34) and (37). In order to present them appropriately, we consider the ratios |u − IM uh |1,h as a function of | log n|, η+ξ ηK := max as a function of | log n|. K∈Th ∇(u − IM uh )ωK + ξK qup := qlow The first ratio qup is frequently referred to as the effectivity index. It measures the reliability of the estimator and is related to the global upper error bound. The second 1499 SERGE NICAISE ratio is related to the local lower error bound and measures the efficiency of the estimator. From our theoretical considerations, both ratios should be bounded from above which is confirmed experimentally as shown in Figure 3. Hence our estimator is reliable and efficient. In Figure 4 we compare the discrete H 1 seminorm |u − IM uh |1,h and the global error estimator η with respect to n. We remark that the orders of convergence are the same (namely, 2). This figure further confirms the equivalence between |u − IM uh |1,h and η. All related quantities are summarized in Table 1. 0.6 0.35 0.55 0.3 0.5 qup low 0.4 q 0.25 0.45 0.2 0.4 0.15 0.35 0.1 0.5 1 1.5 | Log(n) | 2 0.3 0.5 2.5 1 1.5 | Log(n) | 2 2.5 Fig. 3. qup (left) and qlow (right) in dependence of | log n| for test 1. log(error) −2.87 (2) −3.93 (1) −4.99 −6.06 −7.12 −8.18 −9.24 −10.31 2 −11.37 1 −12.43 −13.49 0.39 log(n) 1.00 1.62 2.23 2.85 3.47 4.08 4.70 5.31 5.93 6.55 Fig. 4. Comparison between − ln |u − IM uh |1,h (line (1)) and − ln η (line (2)) with respect to ln n for test 1. 8.2. A nonsmooth solution. For the second example we use the L-shaped domain Ω :=] − 1, 1[2 \ ]0,1[×] − 1, 0[ with the exact singular solution given in polar coordinates by u = r2/3 sin 2θ 3 considered as a solution of the Dirichlet problem with nonhomogeneous Dirichlet boundary conditions. As before the domain is discretized using uniform meshes made of squares of size h = 1/n for n = 4, 8, . . . , 256. Since u presents singular behavior near the origin and uniform meshes are used, we have a reduction of the order of convergence from 1 to 2/3 for the norm |u − uh |1,h (see Theorem 2.4 of [8] and Figure 5). From Figure 5 we notice the same phenomenon of reduction of the order of convergence for the other norms. 1500 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS Table 1 Data for test 1. n 4 8 16 32 64 128 256 ū − uh 0,Ω 2.96e−03 7.57e−04 1.91e−04 4.77e−05 1.19e−05 2.99e−06 7.47e−07 |u − uh |1,Th 1.57e−02 5.58e−03 1.99e−03 7.09e−04 2.51e−04 8.90e−05 3.15e−05 u − IM uh 0,Ω 2.84e−03 8.58e−04 2.25e−04 5.68e−05 1.42e−05 3.56e−06 8.90e−07 |u − IM uh |1,h 1.37e−02 3.72e−03 9.53e−04 2.40e−04 6.00e−05 1.50e−05 3.75e−06 η 2.09e−02 6.43e−03 1.74e−03 4.48e−04 1.14e−04 2.86e−05 7.17e−06 qlow 0.4174 0.4875 0.5357 0.5513 0.5569 0.5580 0.5584 qup 0.2753 0.2708 0.2665 0.2638 0.2636 0.2616 0.2612 qlow 2.4949 2.4661 2.4527 2.4471 2.4448 2.4438 2.4435 qup 0.2958 0.3003 0.3022 0.3030 0.3033 0.3035 0.3034 log(error) −1.07 1 −2.02 2/3 −2.96 −3.91 −4.86 (4) (2) −5.80 −6.75 −7.69 4/3 −8.64 1 (3) (1) −9.59 −10.53 0.39 1.00 1.62 2.23 2.85 3.47 4.08 4.70 5.31 log(n) 5.93 6.55 Fig. 5. Illustration of different norms for test 2. Table 2 Data for tests 2. n 4 8 16 32 64 128 256 ū − uh 0,Ω 1.63e−02 6.86e−03 2.81e−03 1.14e−03 4.56e−04 1.82e−04 7.25e−05 |u − uh |1,Th 7.02e−02 4.52e−02 2.87e−02 1.82e−02 1.15e−02 7.23e−03 4.55e−03 u − IM uh 0,Ω 1.66e−02 8.11e−03 3.59e−03 1.51e−03 6.21e−04 2.52e−04 1.01e−04 |u − IM uh |1,h 1.26e−01 8.11e−02 5.15e−02 3.26e−02 2.06e−02 1.30e−02 8.16e−03 η 4.26e−01 2.70e−01 1.71e−01 1.07e−01 6.78e−02 4.27e−02 2.69e−02 Again we have tested the rate of convergence of ū − uh , |u − uh |1,Th , u − IM uh , and |u − IM uh |1,h . Here we notice that both discrete H 1 -norms have a rate of convergence of 2/3 and that the rate of convergence of the L2 -norms is twice, namely, 4/3. As before we further check the boundedness of the ratios qup and qlow . These quantities are given in Table 2 and illustrated in Figures 6 and 7. From these figures we can draw the same conclusion as before, namely, the efficiency and reliability of our estimator. From Tables 1 and 2, we see that the experimental bounds for qup and qlow are quite smaller than the theoretical ones. This is quite realistic since the experimental 1501 SERGE NICAISE 0.38 2.5 2.49 0.36 2.48 2.47 0.34 qup low 2.46 q 0.32 2.45 2.44 0.3 2.43 2.42 0.28 2.41 0.26 0.5 1 1.5 | Log(n) | 2 2.4 0.5 2.5 1 1.5 | Log(n) | 2 2.5 Fig. 6. qup (left) and qlow (right) in dependence of | log n| for test 2. log(error) 0.15 −0.45 (2) −1.04 −1.64 (1) −2.24 −2.83 −3.43 −4.02 2/3 −4.62 1 −5.21 log(n) −5.81 0.39 1.00 1.62 2.23 2.85 3.47 4.08 4.70 5.31 5.93 6.55 Fig. 7. Comparison between − ln |u − IM uh |1,h (line (1)) and − ln η (line (2)) with respect to ln n for test 2. Z 0.046 0.023 0.000 −1 0 Y −1 0 1 X 1 Fig. 8. Distribution of the local estimator for test 2 and n = 64. values depend on the chosen solution, while the theoretical analysis always considers the worse case. Finally, in Figure 8 we give the distribution of the local residual error estimators for our second example with the mesh corresponding to n = 64. From this figure we 1502 A POSTERIORI ESTIMATES FOR FINITE VOLUME METHODS may conclude that our estimator is appropriate for adaptivity, since it detects the region of large errors, namely, the neighborhood of the origin. 9. Conclusions. We have proposed and rigorously analyzed a new a posteriori error estimator for a cell-centered finite volume method that is reliable and efficient. This estimator is based on the construction of an appropriate interpolant of Morley type and the use of a Helmholtz decomposition of the error. The size of the constants appearing in the error estimates has been given as explicitly as possible, as a function of the aspect ratio of the mesh and of the form of the elements (triangles, rectangles, or tetrahedra). Some numerical experiments confirm our theoretical predictions and show that our estimator is appropriate for adaptivity. The extension of our method to diffusion–convection–reaction equations is briefly described; the details are postponed to a forthcoming paper. Adaptive algorithms are not considered here since they require more investigations. They will be considered in forthcoming works. Acknowledgments. I am very grateful to Karim Djadel (University of Lille, France) who conducted the numerical experiments presented in section 8. I further thank Joachim Schöberl (University of Linz, Austria) and Philipp Frauenfelder (ETH Zürich, Switzerland) who gave me the numerical values of μ and α presented in Remark 3.4 and obtained using their hp-FEM code. REFERENCES [1] A. Agouzal and F. Oudin, A posteriori error estimator for finite volume methods, Appl. Math. Comput., 110 (2000), pp. 239–250. [2] A. Bergam and Z. Mghazli, Estimateurs a posteriori d’un schéma de volumes finis pour un problème non linéaire, C. R. Math Acad. Sci. Paris Sér. I Math., 331 (2000), pp. 475–478. [3] A. Bergam, Z. Mghazli, and R. Verfürth, A posteriori estimators for the finite volume discretization of an elliptic problem, Numer. Math., 95 (2003), pp. 599–624. [4] P. G. Ciarlet, The Finite Element Method for Elliptic Problems, North-Holland, Amsterdam, 1978. [5] Y. Coudière, J.-P. Villa, and P. Villedieu, Convergence rate of a finite volume scheme for a two dimensional convection-diffusion problem, M2AN Math. Model. Numer. Anal., 33 (1999), pp. 493–516. [6] Y. Coudière and P. Villedieu, Convergence rate of a finite volume scheme for the linear convection–diffusion equation on locally refined meshes, M2AN Math. Model. Numer. Anal., 34 (2000), pp. 1123–1149. [7] E. Dari, R. Durán, C. Padra, and V. Vampa, A posteriori error estimators for nonconforming finite element methods, M2AN Math. Model. Numer. Anal., 30 (1996), pp. 385–400. [8] K. Djadel, S. Nicaise, and J. Tabka, Some refined finite volume methods for elliptic problems with corner singularities, Int. J. Finite Volumes, (2003), http://averoes.math.univparis13.fr/IJFVDB, PAPERS/RevFv Nicaise.pdf. [9] J. Droniou, Error estimates for the convergence of a finite volume discretization of convection– diffusion equations, J. Numer. Math., 11 (2003), pp. 1–32. [10] R. Eymard, T. Gallouët, and R. Herbin, Finite volume methods, in Handbook of Numerical Analysis, Vol. 7, P. Ciarlet and J.-L. Lions, eds., North Holland, Amsterdam, 2000, pp. 723– 1020. [11] V. Girault and P.-A. Raviart, Finite Element Methods for Navier–Stokes Equations, Theory and Algorithms, Springer Series in Comput. Math. 5, Springer, New York, 1986. [12] R. Herbin and M. Ohlberger, A posteriori error estimate for finite volume approximation of convection–diffusion problems, in Finite Volume for Complex Applications, R. Herbin and D. Kröner, eds., Hermès, London, 2002, pp. 753–760. [13] N. Jullian, An error indicator for cell-centered finite volumes for linear convection–iffusion problems, in Finite Volume for Complex Applications, R. Herbin and D. Kröner, eds., Hermès, London, 2002, pp. 777–784. SERGE NICAISE 1503 [14] D. Kröner and M. Ohlberger, A posteriori error estimates for upwind finite volume schemes for nonlinear conservation laws in multi dimensions, Math. Comp., 69 (1999), pp. 25–39. [15] R. Lazarov and S. Tomov, Adaptive finite volume element method for convection-diffusionreaction problems in 3-d, in Scientific Computing and Application, Y. W. P. Minev and Y. Lin, eds., Nova Science Publishing House, Huntington, NY, 2001, pp. 91–106. [16] R. Lazarov and S. Tomov, A posteriori error estimates for finite volume approximations of convection-diffusion-reaction equations, Comput. Geosci., 6 (2002), pp. 483–503. [17] J. Mackenzie, T. Sonar, and G. Warnecke, A posteriori error estimates for the cell-vertex finite volume method, in Adaptive Methods: Algorithms, Theory and Applications, W. Hackbusch and G. Wittum, eds., Vieweg, Berlin, 1994, pp. 221–235. [18] L. Morley, The triangular equilibrium element in the solution of plate bending problems, Aero. Quarterly, 19 (1968), pp. 149–169. [19] K. W. Morton and E. Süli, A posteriori and a priori error analysis of finite volume methods, in The Mathematics of Finite Elements and Applications, J. R. Whiteman, ed., John Wiley and Sons, New York, 1994, pp. 267–288. [20] S. Nicaise, A posteriori error estimations of some cell-centered finite volume methods for diffusion-convection-reaction problems, SIAM J. Numer. Anal., submitted. [21] T. K. Nilssen, X.-C. Tai, and R. Winther, A robust nonconforming H 2 -element, Math. Comp., 70 (2000), pp. 489–505. [22] M. Ohlberger, A posteriori error estimate for finite volume approximations to singularly perturbed nonlinear convection-diffusion equations, Numer. Math., 87 (2001), pp. 737–761. [23] M. Ohlberger, A posteriori error estimates for vertex centered finite volume approximations of convection-diffusion-reaction equations, M2AN Math. Model. Numer. Anal., 35 (2001), pp. 355–387. [24] S. V. Patankar, Numerical Heat Transfer and Fluid Flow, Ser. Comput. Methods Mech. Thermal Sci., McGraw Hill, New York, 1980. [25] T. Sonar and E. Süli, A dual graph-norm refinement indicator for finite volume approximations of the Euler equations, Numer. Math., 78 (1998), pp. 619–658. [26] R. Verfürth, A Review of a Posteriori Error Estimation and Adaptive Mesh-Refinement Techniques, Wiley-Teubner, Chichester, 1996.
© Copyright 2026 Paperzz