# Generalized Prager-Synge Inequality and Equilibrated Error Estimators for Discontinuous Elements

The well-known Prager-Synge identity is valid in H^1(Ω) and serves as a foundation for developing equilibrated a posteriori error estimators for continuous elements. In this paper, we introduce a new inequality, that may be regarded as a generalization of the Prager-Synge identity, to be valid for piecewise H^1(Ω) functions for diffusion problems. The inequality is proved to be identity in two dimensions. For nonconforming finite element approximation of arbitrary odd order, we propose a fully explicit approach that recovers an equilibrated flux in H(div; Ω) through a local element-wise scheme and that recovers a gradient in H(curl;Ω) through a simple averaging technique over edges. The resulting error estimator is then proved to be globally reliable and locally efficient. Moreover, the reliability and efficiency constants are independent of the jump of the diffusion coefficient regardless of its distribution.

## Authors

• 7 publications
• 11 publications
• 30 publications
04/17/2020

### A polynomial-degree-robust a posteriori error estimator for Nédélec discretizations of magnetostatic problems

We present an equilibration-based a posteriori error estimator for Nédél...
07/13/2021

### Hybrid A Posteriori Error Estimators for Conforming Finite Element Approximations to Stationary Convection-Diffusion-Reaction equations

We consider the a posteriori error estimation for convection-diffusion-r...
10/17/2019

### Residual-based a posteriori error estimation for immersed finite element methods

In this paper we introduce and analyze the residual-based a posteriori e...
06/10/2020

### A virtual element-based flux recovery on quadtree

In this paper, we introduce a simple local flux recovery for 𝒬_k finite ...
05/01/2021

### A family of mixed finite elements for nearly incompressible strain gradient elastic models

We propose a family of mixed finite elements that are robust for the nea...
05/17/2021

### Quasi-monotonicity and Robust Localization with Continuous Piecewise Polynomials

We consider the energy norm arising from elliptic problems with disconti...
##### This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

## 1 Introduction

Equilibrated a posteriori error estimators have attracted much interest recently due to the guaranteed reliability bound with the reliability constant being one. This property implies that they are perfect for discretization error control on both coarse and fine meshes. Error control on coarse meshes is important but difficult for computationally challenging problems.

For the conforming finite element approximation, a mathematical foundation of equilibrated estimators is the Prager-Synge identity [31] that is valid in (see Section 3). Based on this identity, various equilibrated estimators have been studied recently by many researchers (see, e.g., [27, 22, 29, 20, 21, 6, 3, 33, 10, 12, 13, 34, 17, 14]). The key ingredient of the equilibrated estimators for the continuous elements is local recovery of an equilibrated (locally conservative) flux in the space through the numerical flux. By using a partition of unity, Ladevèze and Leguillon [27] initiated a local procedure to reduce the construction of an equilibrated flux to vertex patch based local calculations. For the continuous linear finite element approximation to the Poisson equation in two dimensions, an equilibrated flux in the lowest order Raviart-Thomas space was explicitly constructed in [10, 12]. This explicit approach does not lead to robust equilibrated estimator with respect to the coefficient jump without introducing a constraint minimization (see [17]). The constraint minimization on each vertex patch may be efficiently solved by first computing an equilibrated flux and then calculating a divergence free correction. For recent developments, see [14] and references therein.

The purpose of this paper is to develop and analyze equilibrated a posteriori error estimators for discontinuous elements including both nonconforming and discontinuous Galerkin elements. To do so, the first and the essential step is to extend the Prager-Synge identity to be valid for piecewise functions. This will be done by establishing a generalized Prager-Synge inequality (see Theorem 3.1) that contains an additional term measuring the distance between and piecewise . Moreover, by using a Helmholtz decomposition, we will be able to show that the inequality becomes an identity in two dimensions (see Lemma 3.4). A non-optimal inequality similar to ours was obtained earlier by Braess, Fraunholz, and Hoppe in [11] for the Poisson equation with pure Dirichlet boundary condition. Based on the generalized Prager-Synge inequality and an equivalent form (see Corollary 3.2), the construction of an equilibrated a posteriori error estimator for discontinuous finite element solutions is reduced to recover an equilibrated flux in and to recover either a potential function in

or a curl free vector-valued function in

.

Recovery of equilibrated fluxes for discontinuous elements has been studied by many researchers. For discontinuous Garlerkin (DG) methods, equilibrated fluxes in Raviart-Thomas (RT) spaces were explicitly reconstructed in [2] for linear elements and in [23] for higher order elements. For nonconforming finite element methods, existing explicit equilibrated flux recoveries in RT spaces seem to be limited to the linear Crouzeix-Raviart (CR) and the quadratic Fortin-Soulie elements by Marini [28] (see [1] in the context of estimator) and Kim [26], respectively. For higher order nonconforming elements, a local reconstruction procedure was proposed by Ainsworth and Rankin in [4] through solving element-wise minimization problems. The recovered flux is not in the RT spaces. Nevertheless, the resulting estimator provides a guaranteed upper bound.

In this paper, we will introduce a fully explicit post-processing procedure for recovering an equilibrated flux in the RT space of index for the nonconforming elements of odd order of . Currently, we are not able to extend our recovery technique to even orders. This is because our recovery procedure heavily depends on the finite element formulation and the properties of the nonconforming finite element space; moreover, structure of the nonconforming finite element spaces of even and odd orders are fundamentally different.

Recovery of a potential function in for discontinuous elements was studied by some researchers (see, e.g., [4, 2, 11]). Local approaches for recovering equilibrated flux in [10, 12, 17, 13, 14] may be directly applied (at least in two dimensions) for computing an approximation to the gradient in the curl-free space. (As mentioned previously, this approach requires solutions of local constraint minimization problems over vertex patches.) The resulting a posteriori error estimator from either the potential or the gradient recoveries may be proved to be locally efficient. Nevertheless, to show independence of the efficiency constant on the jump, we have to assume that the distribution of the diffusion coefficient is quasi-monotone (see [30]).

In this paper, we will employ a simple averaging technique over edges to recover a gradient in . Due to the fact that the recovered gradient is not necessarily curl free, the reliability constant of the resulting estimator is no longer one. However, it turns out that the curl free constraint is not essential and, theoretically we are able to prove that the resulting estimator has the robust local reliability as well as the robust local efficiency without the quad-monotone assumption. This is compatible with our recent result in [15] on the residual error estimator for discontinuous elements.

This paper is organized as follows. The diffusion problem and the finite element mesh are introduced in Section 2. The generalized Prager-Synge inequality for piecewise functions are established in Section 3. Explicit recoveries of an equilibrated flux and a gradient and the resulting a posteriori error estimator for discontinuous elements are described in Section 4. Global reliability and local efficiency of the estimator are proved in Section 5. Finally, numerical results are presented in Section 6.

## 2 Model problem

Let be a bounded polygonal domain in , with Lipschitz boundary , where . For simplicity, assume that . Considering the diffusion problem:

 −∇⋅(A∇u)=finΩ, (2.1)

with boundary conditions

 u=0 on ΓDand−A∇u⋅n=g% on ΓN,

where and are the respective divergence and gradient operators; is the outward unit vector normal to the boundary; and are given scalar-valued functions; and the diffusion coefficient

is symmetric, positive definite, and piecewise constant full tensor with respect to the domain

. Here we assume that the subdomain, for , is open and polygonal.

We use the standard notations and definitions for the Sobolev spaces. Let

 H1D(Ω)={v∈H1(Ω):v=0 on ΓD}.

Then the corresponding variational problem of (2.1) is to find such that

 a(u,v):=(A∇u,∇v)=(f,v)−⟨g,v⟩ΓN,∀v∈H1D(Ω), (2.2)

where is the inner product on the domain . The subscript is omitted when .

### 2.1 Triangulation

Let be a finite element partition of that is regular, and denote by the diameter of the element . Furthermore, assume that the interfaces,

 Γ={∂Ωi∩∂Ωj:i≠j and i,j=1,⋯,n},

do not cut through any element . Denote the set of all edges of the triangulation by

 E:=EI∪ED∪EN,

where is the set of interior element edges, and and are the sets of boundary edges belonging to the respective and . For each , denote by the length of and by a unit vector normal to . Let and be the two elements sharing the common edge such that the unit outward normal of coincides with . When , is the unit outward normal to and denote by the element having the edge .

## 3 Generalized Prager-Synge inequality

For the conforming finite element approximation, the foundation of the equilibrated a posteriori error estimator is the Prager-Synge identity [31]. That is, let be the solution of (2.1), then

 ∥A1/2∇(u−w)∥2+∥A−1/2\boldmathτ+A1/2∇u∥2=∥A−1/2\boldmathτ+A1/2∇w∥2

for all and for all , where is the so-called equilibrated flux space defined by

 Σf(Ω)={\boldmathτ∈H(div;Ω):∇⋅\boldmathτ=f in Ω and % \boldmathτ⋅n=gN}.

Here, denotes the space of all vector-valued functions whose divergence are in . The Prager-Synge identity immediately leads to

 ∥A1/2∇(u−w)∥2≤inf\boldmathτ∈Σf(Ω)∥A−1/2\boldmathτ+A1/2∇w∥2. (3.1)

Choosing to be the conforming finite element approximation, then (3.1) implies that

 ητ:=∥A−1/2\boldmathτ+A1/2∇w∥,∀\boldmathτ∈Σf(Ω) (3.2)

is a reliable estimator with the reliability constant being one.

We now proceed to establish a generalization of (3.1) for piecewise functions with applications to nonconforming and discontinuous Galerkin finite element approximations. To this end, denote the broken space with respect to by

 H1(T)={v∈L2(Ω):v|K∈H1(K),∀K∈T}.

Define be the discrete gradient operator on such that for any

 (∇hv)|K=∇(v|K),∀K∈T.
###### Theorem 3.1.

Let be the solution of (2.1). In both two and three dimensions, for all , we have

 ∥A1/2∇h(u−w)∥2≤inf\boldmathτ∈Σf(Ω)∥A−1/2\boldmathτ+A1/2∇hw∥2+infv∈H1D(Ω)∥A1/2∇h(v−w)∥2. (3.3)
###### Proof.

Let , for all and for all , it follows from integration by parts and the Cauchy-Schwarz and Young’s inequalities that

 2(∇h(u−w),A∇u+\boldmathτ) = 2(∇(u−v),A∇u+\boldmathτ)+2(∇h(v−w),A∇u+\boldmathτ) (3.4) = 2(∇h(v−w),A∇u+\boldmathτ) ≤ ∥A1/2∇h(v−w)∥2+∥A1/2∇u+A−1/2% \boldmathτ∥2.

It is easy to see that

 ∥A1/2∇hw+A−1/2\boldmathτ∥2=∥A1/2∇h(u−w)∥2+∥A1/2∇u+A−1/2\boldmathτ∥2−2(∇h(u−w),A∇u+\boldmathτ),

which, together with (3.4), implies

 ∥A1/2∇h(u−w)∥2 = ∥A1/2∇hw+A−1/2\boldmathτ∥2−∥A1/2∇u+A−1/2\boldmathτ∥2+2(∇h(u−w),A∇u+\boldmathτ) ≤ ∥A1/2∇hw+A−1/2\boldmathτ∥2+∥A1/2∇h(v−w)∥2

for all and all . This implies the validity of (3.3) and, hence, the theorem. ∎

A suboptimal result for the Poisson equation () with pure Dirichlet boundary condition is proved in [11] by Braess, Fraunholz, and Hoppe:

 ∥∇h(u−w)∥≤inf\boldmathτ∈Σf(Ω)∥∇w+\boldmathτ∥+2infv∈H10(Ω)∥∇h(v−w)∥.

Let be the space of all vector-valued functions whose curl are in , and denote its curl free subspace by

 ˚HD(curl;Ω)={\boldmathτ∈H(curl;Ω):∇×\boldmathτ=0 in Ω and \boldmathτ⋅t=0 on ΓD},

where denotes the tangent vector(s).

###### Corollary 3.2.

Let be the solution of (2.1). In both two and three dimensions, for all , we have

 ∥A1/2∇h(u−w)∥2≤inf\boldmathτ∈Σf(Ω)∥A−1/2\boldmathτ+A1/2∇hw∥2+inf% \boldmathγ∈˚HD(curl;Ω)∥A1/2(% \boldmathγ−∇hw)∥2. (3.5)
###### Proof.

The result of (3.5) is an immediate consequence of (3.3) and the fact that . ∎

In the remaining section, we prove that, in two dimensions, the inequality (3.3) in Theorem 3.1 is indeed an equality. For each , in two dimensions, assume that , then denote by the unit vector tangent to and by and the start and end points of , respectively, such that . Let

 H={v∈H1(Ω):∫Ωvdx=0 and ∂v∂t=0 on ΓN}.

For a vector-valued function , define the curl operator by

 ∇×\boldmathτ=∂τ2∂x−∂τ1∂y.

For a scalar-valued function , define the formal adjoint operator of the curl by

 ∇⊥v=(∂v∂y,−∂v∂x).

For a fixed , there exist unique and for the following Helmholtz decomposition (see, e.g., [4]) such that

 A∇h(u−w)=A∇ϕ+∇⊥ψ, (3.6)

and and satisfy

 (A∇ϕ,∇v)=(A∇h(u−w),∇v)∀v∈H1D(Ω),

and

 (A−1∇⊥ψ,∇⊥w)=(∇h(u−w),∇⊥w)∀w∈H,

respectively. It is easy to see that and are orthogonal with respect to the inner product, which yields

 ∥A1/2∇h(u−w)∥2=∥A1/2∇ϕ∥2+∥A−1/2∇⊥ψ∥2. (3.7)
###### Lemma 3.3.

Let be a fixed function in and and be the corresponding Helmholtz decomposition of given in (3.6). We have

 inf\boldmathτ∈Σf(Ω)∥A−1/2% \boldmathτ+A1/2∇hw∥=∥A1/2∇ϕ∥ and infv∈H1D(Ω)∥A1/2∇h(v−w)∥=∥A−1/2∇⊥ψ∥. (3.8)
###### Proof.

For any , (3.6) and integration by parts give

 ∥A1/2∇ϕ∥2=(A∇h(u−w),∇ϕ)=(A∇u+% \boldmathτ,∇ϕ)−(\boldmathτ+A∇hw,∇ϕ)=−(\boldmathτ+A∇hw,∇ϕ),

which, together with the Cauchy-Schwarz inequality and the choice , yields the first equality in (3.8) as follows:

 ∥A1/2∇ϕ∥≤inf\boldmathτ∈Σf(Ω)∥A−1/2\boldmathτ+A1/2∇hw∥≤∥A1/2∇h(u−w)−A−1/2∇⊥ψ∥=∥A1/2∇ϕ∥.

Now we proceed to prove the second equality in (3.8). For any , by (3.6) and integration by parts, we have

 ∥A−1/2∇⊥ψ∥2=(∇h(u−w),∇⊥ψ)=(∇h(v−w),∇⊥ψ).

The second equality in (3.8) is then a consequence of the Cauchy-Schwartz inequality and the choice of :

 ∥A−1/2∇⊥ψ∥≤infv∈H1D(Ω)∥A1/2∇h(v−w)∥≤∥A1/2∇h(u−ϕ−w∥=∥A−1/2∇⊥ψ∥.

This completes the proof of the lemma. ∎

###### Lemma 3.4.

Let be the solution of (2.1). In two dimensions, for all , we have

 ∥A1/2∇h(u−w)∥2=inf\boldmathτ∈Σf(Ω)∥A−1/2\boldmathτ+A1/2∇hw∥2+infv∈H1D(Ω)∥A1/2∇h(v−w)∥2. (3.9)
###### Proof.

The identity (3.9) is a direct consequence of (3.7) and Lemma 3.3. ∎

###### Remark 3.5.

It is easy to see that if in Lemma 3.4, i.e., is conforming, the second part on the right of (3.9) vanishes. It is thus natural to refer or as the conforming error and or as the nonconforming error.

For each , denote by and

the maximal and minimal eigenvalues of

, respectively. For each , let , , and if and if . To this end, let

 ΛT=maxK∈TΛKandλT=minK∈TλK.

Assume that each local matrix

is similar to the identity matrix in the sense that its maximal and minimal eigenvalues are almost of the same size. More precisely, there exists a moderate size constant

such that

 ΛKλK≤κ,∀K∈T.

Nevertheless, the ratio of global maximal and minimal eigenvalues, , is allowed to be very large.

For a function , denote its traces on by and and the jump of across the edge by

 [[w]]|F={w|−F−w|+F,∀F∈EI,w|−F,∀F∈ED∪EN.

In the following lemma, we show the relationship between the nonconforming error and the residual based error of solution jump on edges. It is noted that the constant is robust with respect to the coefficient jump.

###### Lemma 3.6.

Let be a fixed function in . In two dimensions, there exists a constant that is independent of the jump of the coefficient such that

 inf\boldmathτ∈˚HD(curl;Ω)∥A1/2(\boldmathτ−∇hw)∥≤Cr⎛⎝∑F∈EI∪EDλFh−1F∥[[w]]∥20,F⎞⎠1/2. (3.10)
###### Proof.

Let be given in the Helmholtz decomposition in (3.6), then integration by parts gives

 ∥A−1/2∇⊥ψ∥2=(∇h(u−w),∇⊥ψ)=−∑F∈EI∪ED∫F[[w]](∇⊥ψ⋅nF)ds.

Without loss of generality, assume that for each . It follows from Lemma 2.4 in [15] and the Cauchy-Schwarz inequality that

 ∑F∈EI∪ED∫F[[w]](∇⊥ψ⋅nF)ds ≤ C∑F∈EI∪EDh−1/2F∥[[w]]∥0,F∥∇⊥ψ∥0,K−F ≤ C⎛⎝∑F∈EI∪EDλFh−1F∥[[w]]∥20,F⎞⎠1/2∥A−1/2∇⊥ψ∥,

which, together with the above equality, yields

 ∥A−1/2∇⊥ψ∥≤C⎛⎝∑F∈EI∪EDλFh−1F∥[[w]]∥20,F⎞⎠1/2.

This completes the proof of the lemma. ∎

## 4 Error estimators and indicators

### 4.1 NC finite element approximation

For the convenience of readers, in this subsection we introduce the nonconforming finite element space and its properties.

Let and be the spaces of polynomials of degree less than or equal to on the element and , respectively. Define the nonconforming finite element space of order on the triangulation by

 Uk(T)={v∈L2(Ω):v|K∈Pk(K),∀K∈Tand∫F[[v]]pds=0,∀p∈Pk−1(F),∀F∈EI} (4.1)

and its subspace by

 UkD(T)={v∈Uk(T):∫Fvpds=0,∀p∈Pk−1(F)and∀F∈ED}.

The spaces defined above are exactly the same as those defined in [19] for , [24] for , [18] for and , [4] for general odd order, and [32, 5] for general order. Then the nonconforming finite element approximation of order is to find such that

 ah(uT,v):=(A∇huT,∇hv)=(f,v)−⟨g,v⟩ΓN,∀v∈UkD(T). (4.2)

Below we describe basis functions of and their properties. To this end, for each , let for and for . Denote by the set of all interior Lagrange points in with respect to the space and by the nodal basis function corresponding to , i.e.,

 Pj,K(xi)=δij for i=1,⋯,mk,

where is the Kronecker delta function. For each , let be the th order Gauss-Legendre polynomial on such that . Note that is an odd or even function when is odd or even. Hence, for odd and for even .

For odd

, the set of degrees of freedom of

(see Lemma 2.1 in [4]) can be given by

 ∫KvPj,Kdx,j=1,⋯,mk (4.3)

for all and

 ∫FvLj,Fds,j=0,⋯,k−1 (4.4)

for all . Define the basis function satisfying

 {∫K′ϕi,KPj,K′dx=δijδKK′,∀j=1,⋯,mk,∀K′∈T,∫Fϕi,KLj,Fds=0,∀j=0,⋯,k−1,∀F∈E, (4.5)

for and , and the basis function satisfying

 (4.6)

for and . Then the nonconforming finite element space is the space spanned by all these basis functions, i.e.,

 Uk(T)=span{ϕi,K:K∈T}mki=1⊕span{ϕi,F:F∈E}k−1i=0.
###### Lemma 4.1.

For all , the basis functions have support on and vanish on the boundary of , i.e.,

 ϕj,K≡0 on ∂K.
###### Proof.

Obviously, (4.5) implies that . To show that , considering each edge , the second equation of (4.5) indicates that there exists such that

 ϕj,K|F=aFLk,F.

Note that is an odd function on and that values of at two end-points of are and , respectively. Now the continuity of in implies that and, hence, on . ∎

For each , denote by the set of all edges of . For each , denote by the union of all elements that share the common edge ; and define a sign function on the set (when is a boundary edge, let ) such that

###### Lemma 4.2.

For all , the basis functions have support on , and their restrictions on has the following representation:

 (4.7)

when is odd, and

 ϕj,F=⎧⎪ ⎪ ⎪ ⎪ ⎪⎨⎪ ⎪ ⎪ ⎪ ⎪⎩1∥Lj,F∥20,FLj,F,onF,χF(F′)∥Lj,F∥20,FLk,F′,onF′∈EK+F∪EK−F∖{F} (4.8)

when is even.

###### Proof.

By (4.6), it is easy to see that support of is . Since , there exist constants such that

 ϕj,F|±F=k∑i=0a±i,FLi,F.

Using (4.6) and the orthogonality of , it is obvious that

 a±i,F={∥Lj,F∥−20,F,fori=j,0,for0≤i≤k−1 and i≠j

and, hence,

 ϕj,F|±F=1∥Lj,F∥20,FLj,F+a±k,FLk,F. (4.9)

By (4.6), it is also easy to see that there exists constant for each such that

 ϕj,F|F′=aj,F,F′Lk,F′. (4.10)

Since is an odd function for all and is continuous in and , (4.10) implies that

 ϕj,F|K(sF)=ϕj,F|K(eF),K∈{K+F,K−F}. (4.11)

Combining the facts that for odd and that for even , (4.9), and (4.11), we have

 a±k,F=⎧⎪ ⎪⎨⎪ ⎪⎩−1∥Lj,F∥20,F,for odd j,0,for even j,

which, together with (4.9), leads to the formulas of in (4.7) and (4.8). Finally, for each , in (4.10) can be directly computed based on the continuity of in and . This completes the proof of the lemma.

###### Remark 4.3.

As a consequence of Lemma 4.2, the basis function is continuous on the edge , i.e., for all ; moreover, vanishes at end points of , i.e., , for odd .

###### Lemma 4.4.

Let be an edge of . Assume that . Then we have that

 ∫∂Kpϕj,Fds=∫Fpϕj,Fds. (4.12)

Moreover, if for all , then on .

###### Proof.

Since are orthogonal polynomials on , Lemma 4.4 is a direct consequence of Lemma 4.2. ∎

### 4.2 Equilibrated flux recovery

In this subsection, we introduce a fully explicit post-processing procedure for recovering an equilibrated flux. To this end, define by

 fk−1|K=ΠK(f),∀K∈T,

where is the projection onto . For simplicity, assume that the Neumann data is a piecewise polynomial of degree less than or equal to , i.e., for all .

Denote the conforming Raviart-Thomas (RT) space of index with respect to by

 RTk−1(T)={\boldmathτ∈H(div;Ω):\boldmathτ|K∈RTk−1(K),∀K∈T},

where . Let

 Σk−1f(T)={\boldmathτ∈RTk−1:∇⋅\boldmathτ=fk−1inΩand\boldmathτ⋅n