Metadata of the chapter that will be visualized online

Chapter Title On a Renewed Approach to A Posteriori Error Bounds for

Approximate Solutions of Reaction-Diffusion Equations
Copyright Year 2019
Copyright Holder Springer Nature Switzerland AG
Corresponding Author Family Name Korneev
Particle
Given Name Vadim G.
Suffix
Organization St. Petersburg State University
Address St. Petersburg, Russia
Email vad.korneev2011@yandex.ru
Abstract We discuss a new approach to obtaining the guaranteed, robust and
consistent a posteriori error bounds for approximate solutions of the
reaction-diffusion problems, modelled by the equation − Δu + σu = f
in Ω, u|∂Ω = 0, with an arbitrary constant or piece wise constant σ ≥
 0. The consistency of a posteriori error bounds for solutions by the
finite element methods assumes in this paper that their orders of
accuracy in respect to the mesh size h coincide with those in the
corresponding sharp a priori bounds. Additionally, it assumes that for
such a coincidence it is sufficient that the testing fluxes possess only the
standard approximation properties without resorting to the equilibration.
Under mild assumptions, with the use of a new technique, it is proved
that the coefficient before the L2-norm of the residual type term in the
a posteriori error bound is uniformly for all testing fluxes from
admissible set, which is the space H(Ω, div). As a consequence of
these facts, there is a wide range of computationally cheap and efficient
procedures for evaluating the test fluxes, making the obtained a posteriori
error bounds sharp. The technique of obtaining the consistent a posteriori
bounds was exposed in [arXiv:1711.02054v1 [math.NA] 6 Nov 2017]
and very briefly in [Doklady Mathematics, 96 (1), 2017, 380–383].
Chapter 12 1

On a Renewed Approach to A Posteriori 2

Error Bounds for Approximate Solutions 3

of Reaction-Diffusion Equations 4

Vadim G. Korneev 5

F
Abstract We discuss a new approach to obtaining the guaranteed, robust and 6

O
consistent a posteriori error bounds for approximate solutions of the reaction- 7

diffusion problems, modelled by the equation −u + σ u = f in Ω, u|∂Ω = 0,

O
8

with an arbitrary constant or piece wise constant σ ≥ 0. The consistency of a 9

PR
posteriori error bounds for solutions by the finite element methods assumes in 10

this paper that their orders of accuracy in respect to the mesh size h coincide 11

with those in the corresponding sharp a priori bounds. Additionally, it assumes 12

that for such a coincidence it is sufficient that the testing fluxes possess only the 13
D
standard approximation properties without resorting to the equilibration. Under 14

mild assumptions, with the use of a new technique, it is proved that the coefficient 15
TE

before the L2 -norm of the residual type term in the a posteriori error bound is O(h) 16

uniformly for all testing fluxes from admissible set, which is the space H(Ω, div). 17

As a consequence of these facts, there is a wide range of computationally cheap and 18

EC

efficient procedures for evaluating the test fluxes, making the obtained a posteriori 19

error bounds sharp. The technique of obtaining the consistent a posteriori bounds 20

was exposed in [arXiv:1711.02054v1 [math.NA] 6 Nov 2017] and very briefly in 21

R

[Doklady Mathematics, 96 (1), 2017, 380–383]. 22

R

12.1 Introduction
O

23
C

The paper is dedicated to the derivation of the guaranteed robust and computable 24

a posteriori error bounds possessing three properties. Firstly, the bound of the 25
N

error energy norm must provide the guaranteed accuracy equal in the order to the 26
U

accuracy of the unimprovable a priori error bound. Secondly, the first property must 27

be achieved with the use of the testing fluxes evaluated by pure approximation 28

instruments, in particular, without resorting to the equilibration of the testing fluxes. 29

V. G. Korneev ()
St. Petersburg State University, St. Petersburg, Russia

© Springer Nature Switzerland AG 2019

T. Apel et al. (eds.), Advanced Finite Element Methods with Applications,
Lecture Notes in Computational Science and Engineering 128,
https://doi.org/10.1007/978-3-030-14244-5_12
 V. G. Korneev

Thirdly, the evaluation of the a posteriori bound must be cheap. For brevity, we term 30

the bounds with the first two properties consistent. 31

A posteriori bounds, suggested in the paper, belong to the family of the error 32

energy norm bounds with two main terms in the right part called diffusion and 33

residual terms. These terms depend on the approximate solution of the boundary 34

value problem, the right part of the problem, and on the testing flux from some 35

admissible set. Effectiveness of such bounds is strongly influenced as by their 36

specific form, so the possibility to pick up a testing flux in the admissible set, 37

providing cheap computations and sufficient accuracy. In spite of the pointed out 38

common features, there were quite a number of suggestions in the literature with 39

F
differently defined specific forms, i. e., coefficients in front of the norms in the 40

diffusion and residual terms, and differently defined test fluxes. 41

O
The main model problem of the paper is 42

O

Lu ≡ − u + σ u = f (x), in Ω ⊂ Rm , m = 2, 3 , u∂Ω = 0 ,

PR
with the constant or piece wise (element wise) constant reaction coefficient σ ≥ 0. 43

If for simplicity we turn to the case of the constant reaction coefficient, the typical 44

residual term of the majorant1 looks as θ 1/2 fˆ − σ uh − div zL2 (Ω) , where uh is 45
D
the approximate solution, and z is the testing flux vector-function. Different authors 46

come to different expressions for θ and, in particular, to (a) θ = 1/σ for all σ > 0, 47
TE

(b) θ =const for σ ≡ 0, (c) θ = (ch)2 for σ ≤ (ch)−2 , and (d) on some subdomains 48

(finite elements) the residual equals to zero and, therefore, eliminated. Examples of 49

such majorants are given in the next section. 50

EC

Majorants related to the cases (a)–(c) behave differently in dependence on the 51

chosen test flux z. Most often, elliptic equations of second order are approximately 52

solved by the FEM’s (finite element methods) of the class C, whose solutions belong 53
R

to the finite dimensional subspaces of C(Ω) ∩ H 1 (Ω). The second derivatives of 54

these solutions, which are needed to calculate the residuals, are not defined. As 55
R

a consequence, the a posteriori majorants of the error are calculated with the use 56

of smoother testing fluxes, which are found from the FEM fluxes by means of 57
O

special flux recovery procedures. The need to improve the smoothness of numerical 58
C

fluxes without losses and even, under some conditions, with gaining in the accuracy 59

motivated the development of these procedures, which currently are numerous. Two 60
N

main classes of them can be distinguished, one aimed completely at achieving 61

highest accuracy without resorting to the equilibration and another which assumes 62
U

equilibration. For making the equilibration simpler and less dependent on the right 63

part f of the elliptic equation, it is commonly done not for f itself, but for some 64

its approximation fˆ. The latter is defined element wise, e.g., as the orthogonal L2 - 65

projection of f on the space of each finite element. Among great number of works 66

on the touched topics, we are able to refer to a few, in particular, to Zienkiewicz and 67

Zhu [39], Ainsworth and Oden [1], Babuska et al. [6, 7], Xu and Zhang [36], Ern 68

1 In this text, the right part of an a posteriori bound is termed majorant.

12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

et al. [19], Cheddadi et al. [14], Cai and Zhang [11], and Ainsworth and Vejchodsky 69

[3], where additional extensive bibliography can be found. 70

For the sequel it is worth to fix the following properties, additional to the 71

smoothing and paid attention at the creation of flux recovery procedures: 72

(α) preserving the orders of accuracy, e.g., in the energy and other norms, at least 73

the same as for the numerical fluxes determined by the approximate solution 74

of the problem; 75

(α∗ ) providing (α) and additionally superconvergence of the test fluxes z; 76

(β) obtaining the balanced or weakly balanced recovered fluxes; 77

(γ ) linear or almost linear computational cost. 78

F
If only α) or α∗ ) with or without (γ ) are satisfied, then it is easy to see that the 79

O
majorants in (a) and (b) can be larger in h−1 times, and even more in the case (a), 80

than the energy norm of the true error. Therefore, these bounds are inconsistent.

O
81

The orders of smallness of a posteriori majorants, related to (c) and (d), can be 82

PR
equal to the orders in the corresponding a priori error bounds, but until recently 83

their examples assumed the equilibrated test fluxes. 84

Obviously, procuring the property (β) complicates flux recovery procedures, 85

makes them directly dependable upon the specific boundary value problem to be 86
D
solved and, therefore, much less universal. 87

In this paper, we develop the a posteriori error bounds, which for approximate 88
TE

solutions by the finite element method are consistent with the sharp a priori error 89

bounds on the wide range of testing fluxes satisfying only the property (α). They 90

have additional features, making them quite different from other bounds of the 91
EC

family under consideration. In particular, in their right parts, supposed for the 92

use as the error indicators, they have chf − σ uh − div zL2 (Ω) in the residual 93

term, independently of z ∈ H(Ω, div) := {y = (y1 , y2 , . . . , ym )

: yk ∈ 94

L2 (Ω), div y ∈ L2 (Ω)} and σ ∈ [0, (ch)−2 ]. As a consequence, the evaluation of

R

95

the test fluxes, providing sharp bounds, becomes a purely approximation problem 96
R

and the need to equilibrate fluxes disappears. These two features are meaningful 97

conceptually as establishing a proper correlation between a posteriori and a priori 98

O

error bounds, but as well they are important for the practice, because the range of 99
C

the numerical applicability of the sharp and cheap a posteriori error bounds greatly 100

widens. Indeed, a number of simple and cheap flux recovery techniques like local 101
N

weighted averaging, which have been developed and extensively studied in relation 102

with the residual based a posteriori error indicators, are applicable to the bounds of 103
U

this paper. They provide not only the standard approximation properties, needed for 104

a consistent a posteriori bound to became sharp, but, under some conditions, also 105

the property α∗ ) which implies superconvergence of the recovered fluxes, see, e.g., 106

[8, 36]. As is noted by some authors, these flux recovery procedures performed in 107

the practice “astonishingly well”, see, e.g., [1, 36]. 108

The approval of the new weights before the norms in the consistent a posteriori 109

error majorants required revision of the proofs. A key change is due to the fact that, 110

at the derivation of the majorants, we take into account the difference in the error 111
 V. G. Korneev

orders of the finite element solutions in L2 and H 1 norms. This allowed us to come 112

to θ of the optimal order under general conditions in two steps made in Sects. 12.4 113

and 12.5, respectively. Since the constants in the majorants is an important issue 114

for applications, we study two approaches to their derivation, which influence the 115

values of the constants and simplicity of their evaluation. One of them imposes 116

some restriction on the smoothness of the boundary, because it employs the Aubin- 117

Nitsche trick for a subsidiary problem, governed by the Poisson equation. The 118

constants resulting from another depend only upon properties of the local quasi- 119

interpolation operator H 1 (δ (r) ) → Vh (τr ), where Vh (τr ) is the space induced by the 120

finite element τr , δ (r) is the smallest patch of the finite elements neighboring τr . A 121

F
representative of such operators is the one of Scott and Zhang [34]. However, we do 122

not elaborate on the comparison of the effectiveness of different quasi-interpolation

O
123

and projection operators, see, e.g., in [9, 10, 30]. 124

O
The paper is organized as follows. Section 12.2 contains the formulation of the 125

boundary value problem of reaction-diffusion, examples of known error majorants, 126

PR
similar in structure to ones suggested in the paper, and a brief discussion of 127

their consistency. In Sect. 12.3, we derive new general a posteriori majorants for 128

the errors of approximations of the exact solution to the problem by arbitrary 129

sufficiently smooth functions v that satisfy the essential boundary conditions. The 130
D
cases of an arbitrary piece wise constant and a constant reaction coefficients are 131

considered separately. The majorants are well defined in both cases, if σ ≥ 0, and 132
TE

coincide with the Aubin’s [5] type majorant on those subdomains (finite elements), 133

where σ exceeds a certain critical value σ∗ . Several versions of the majorant are 134

discussed, related to different ways of defining σ∗ and, respectively, coefficients 135

EC

of the majorant. The easiest one corresponds to σ ≡ const and Galerkin method 136

with coordinate functions having additional smoothness, e.g., belonging to the space 137

H 2 (Ω). Such coordinate functions find implementation in the isogeometric , see 138
R

Cottrel etc. [17]. Majorants of Sect. 12.3 are more accurate at least for σ ≤ σ∗ than 139

other known general majorants valid for all z ∈ H(Ω, div) and allow sharpening 140
R

based upon estimating σ∗ for a specific numerical method. In this sense, they are the 141
O

base for the subsequent sections. 142

Consistent a posteriori error estimates for solutions by the finite element method, 143
C

which are the main results of the paper, are presented in Sects. 12.4 and 12.5. 144

Theorems 12.4 and 12.5, proved there on the basis of the results of Sect. 12.3, 145
N

suggest different approaches to the evaluation of constants in these estimates. For 146

instance, in Theorem 12.5 they are expressed through the constants in the estimates 147
U

of approximation produced by the quasi-interpolation projection operator of Scott 148

and Zhang [34]. In Sect. 12.5, some other important properties of the derived a 149

posteriori estimates are discussed, in particular, the approximation conditions for 150

test fluxes, sufficient for the a posteriori error estimators to be sharp, cheap flux 151

recovery procedures for computing such test fluxes. The conclusion about efficiency 152

of the majorants is supported also by the example of the inverse like bound. 153
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

Below, φH k (D) is the norm in the Sobolev space H k (D) on a domain D, 154

k  ∂lφ
φ2H k (D) = φ2L (D) + |φ|2H l (D) , |φ|2H l (D) = ( q q
2
q ) dx , 155
2
l=1 |q|=l D . . . ∂xmm
∂x1 1 ∂x2 2


where φ2 = D φ 2 dx and q = (q1 , q2 , . . . , qm ), qk ≥ 0, |q| = q1 + q2 + 156
L2 (D)
· · · + qm . If D = Ω, for  · L2 (Ω) ,  · H k (Ω) and | · |H k (Ω) will be also used 157

simpler symbols  · 0 ,  · k and | · |k , respectively. If on the entire boundary ∂Ω 158

or on its part ΓD the homogeneous Dirichlet boundary condition is specified, then 159

F
for the corresponding subspaces of H 1 (Ω) we use the notations H̊ 1 (Ω) := {v ∈ 160

H 1 (Ω) : v|∂Ω = 0} and H̊Γ1D (Ω) := {v ∈ H 1 (Ω) : v|ΓD = 0}. Besides, we will

O
161

need the spaces H̊ 1 (Ω, ) = {v ∈ H̊ 1 (Ω) : v ∈ L2 (Ω)} and H̊Γ1D (Ω, ) = 162

O
{v ∈ H̊Γ1D (Ω) : v ∈ L2 (Ω)}. 163

The finite element space will be denoted Vh (Ω) and by definition V˚h (Ω) = {v ∈

PR
164

Vh (Ω) : v|∂Ω = 0}. 165

Everywhere below it is assumed that the assemblage of compatible and, generally 166

speaking, curvilinear finite elements is given on Ω ⊂ Rm , m = 2, 3, with each 167

finite element occupying domain τr , r = 1, 2, . . . , R. Sometimes symbol R is also

D
168

used for the set of numbers of finite elements. The finite elements are defined by 169
TE

sufficiently smooth mappings x = X (r) (ξ ) : τ◦ → τr of the reference element, 170

defined on the standard triangle or tetrahedron τ◦ . The span of the coordinate 171

functions of the reference element is the space Pp of polynomials of degree p. 172

EC

If p > 1, we use additionally the notation Vh (Ω) = Vh,p (Ω). If other is not 173

mentioned, it is always assumed that the finite element assemblage satisfies the 174

generalized conditions of quasiuniformity with the mesh parameter h > 0, which 175
R

can be understood as the maximum of diameters of finite elements. These conditions 176

can be found, for instance, in Korneev and Langer [26, Section 3.2]. Occasionally 177
R

for more simplicity the following condition is assumed: 178

A) The domain Ω is a polygon in Rm , m = 2, 3, τr are compatible m- 179

O

dimensional simplices (with flat faces and, respectively, straight edges) forming the 180

triangulation Th of Ω, satisfying the conditions of shape regularity.

C

181

In the applications, as a rule Vh (Ω) ⊂ C(Ω) ∩ H 1 (Ω). At the same time, in the 182
N

isogeometric numerical analysis, the smoother spaces Vh (Ω) = Vh1 (Ω) ⊂ C 1(Ω)∩ 183

H 2 (Ω), see Cottrell et al. [17] and Langer et al. [28], are used for solving elliptic 184
U

equations of second order. Superscript l in the notations Vlh (Ω), Vlh,p (Ω) assumes 185

inclusions of these spaces in C l (Ω)∩H l+1(Ω). 186

Matrices and vectors are designated by bold capital and bold lowercase letters, 187

respectively. 188
 V. G. Korneev

12.2 Model Problem, Examples of A Posteriori Error 189

Majorants 190

We start from a very brief survey of a few well known a posteriori bounds which 191

illustrate some tendencies in the development of such error control instruments and 192

were successfully implemented in the numerical applications. We will concentrate 193

on the consideration of the model problem 194

Lu ≡ − u + σ u = f (x), x = (x1 , x2 , . . . , xm ) ∈ Ω ⊂ Rm ,

F
  (12.1)
uΓ = ψD , −∇ u · nΓ = ψN ,

O
D N

where ΓD , ΓN are disjoint simply connected parts of the boundary ∂Ω = Γ D ∪ 195

O
Γ N , mes ΓD > 0, n is the internal normal to ∂Ω. The reaction coefficient σ ≥ 0 is 196

assumed to be element wise constant, i.e., 197

PR
σ = σr = const, x ∈ τr , r = 1, 2, . . . R . (12.2)

The boundary of Ω and the right part f are always considered as sufficiently 198
D
smooth, in particular, f ∈ L2 (Ω), if the requirements on the smoothness are not 199
TE

formulated differently. 200

Our primal interest will be the error estimates in the energy norm 201

 1/2 
√
EC

|||v|||2 = |v|21 +  σ v2L2 (Ω) , |v|21 = ∇v · ∇v , (12.3)

Ω

with |||·|||τr denoting the restriction to τr . For vector-functions y = (y1 , y2 , . . . , ym )

, 202
R

we introduce also the spaces L2 (Ω) = (L2 (Ω))m , H(Ω,  div) = {y ∈ L (Ω) :
2 203

div y ∈ L2 (Ω)}, Wh,p (Ω, div) = {y ∈ H(Ω, div) : yk τ ∈ Pp , ∀r ∈ R, k =

R

204
r
1, . . . , m} and the respectivenorms. For the norm in L2 (Ω) we use additionally the 205
O

notation ]|y|[= yL2 (Ω) = ( Ω y · y)1/2. 206

The a posteriori error majorant of Aubin [5] is one of the earliest: 207
C

Theorem 12.1 Let f ∈ L2 (Ω), 0 < σ = const, ψD ∈ H 1 (Ω), ψN ∈ L2 (ΓN ), 208

N

v be any function of H 1 (Ω) that satisfies the boundary condition on ΓD . Then, for 209

any z ∈ H(Ω, div), satisfying on ΓN the boundary condition z · n = ψN , we have 210

U

1
|||v − u|||2 ≤]|∇ v + z|[2 + f − σ v − div z2L2 (Ω) . (12.4)
σ
Proof Estimate (12.4) is a special case of the results of Aubin [5], see Theorem 22
in Introduction and Theorems 1.2, 1.4, 1.6 in Chapter 10. 

Obviously, if σ → 0 the majorant of Aubin loses precision and for σ = 0 makes 211

no sense. 212
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

If σ ≡ 0, one can use the majorant of Repin and Frolov [32]. Let for simplicity, 213

ΓD = ∂Ω, ψD ≡ 0, and σ ≡ 0. Then for any  > 0 214

1
∇(v −u)2L2 (Ω) ≤ (1+)∇v +z2L2 (Ω) +cΩ (1+ )div z−f 2L2 (Ω) , (12.5)


where v and z are a function and an arbitrary vector-function from H̊ 1 (Ω) and 215

H(Ω, div), respectively, and cΩ is the constant from the Friedrichs inequality. 216

Correction of arbitrary vector-function z ∈ H(Ω, div) into the vector-function 217

τ , satisfying the balance/equilibrium equations, can be done by quite a few rather 218

F
simple and general techniques. Some of them can be found in Anufriev et al. [4] 219

and Korneev [22]. In particular, it is true for the correction of the finite element flux 220

O
vector-function ∇ufem into τ (ufem ). This leads to a specific family of a posteriori 221

error bounds. For simplicity, we turn to the same homogeneous Dirichlet problem 222

O
for the Poisson equation in a two-dimensional convex domain. Let Tk be the 223

PR
projection of the domain Ω on the axis x3−k and the equations of the left and lower 224

parts of the boundary be xk = ak (x3−k ), x3−k ∈ Tk . If βk are arbitrary bounded 225

functions and β1 + β2 ≡ 1, then according to [4, 22] 226

∇(v − u)L2 (Ω) ≤ ∇v + zL2 (Ω) +

D
  xk (12.6)
TE

k=1,2  ak (x3−k ) βk (f − div z)(ηk , x3−k ) dηk L2 (Ω) .

On the right of (12.6), we have integrals from the residual and this hopefully will 227
EC

make the majorant more accurate. Besides, there is an additional free function β1 228

or β2 and the right choice of it (for instance, with the use of the found approximate 229

solution v = ufem ) can accelerate the process of the minimization of the right part, 230
R

if such a process is implemented. If to estimate one-dimensional integrals under the 231

sign of the L2 -norm, then we come to the bound similar to (12.5). 232
R

There were attempts to modify the majorant of (12.4) with the aim of achieving 233

acceptable accuracy for all σ ≥ 0, see e.g., Repin and Sauter [33] and Churilova
O

234

[15]. The majorant of the latter, defined for ∀ σ = const ≥ 0, has the form 235
C

1
|||v − u|||2 ≤ (1 + )]|∇ v + z|[2 + f − σ v − div z2L2 (Ω) . (12.7)
N

σ+ 
cΩ (1+)
U

One of the efficient majorants for the finite element solutions was developed by 236

Ainsworth and Vejchodsky [2, 3]. For its record, we need additional notations: hr is 237
p
the diameter τr , Πr : L2 (τr ) → Pp (τr ) is the operator of orthogonal projection in 238

L2 (τr ). 239

In Theorems 12.2 and 12.3 below, for simplicity we assume ΓD = ∂Ω, ψD ≡ 0 240

and that the condition A) is fulfilled. 241

 V. G. Korneev

Theorem 12.2 Let u ∈ H̊ 1 (Ω) be the weak solution of the problem and ufem ∈ 242

V˚h (Ω) = V˚h,1 (Ω) be the solution by the finite element method. Then there exists
0
243

z ∈ Wh,2 (Ω, div) with the following properties: 244

(i) z is evaluated by the patch wise numerical procedure of linear numerical 245

complexity, √ 246

(ii) for all x ∈ τr and r ∈ R∗ = {r : σ r hr < 1} satisfies the equalities 247

Πr1 f − σr ufem + div z = 0 , (12.8)

(iii) for the error efem = u − ufem and the error indicator ητr (z), defined as

F
248

O
ητ2r (z) = z − ∇ufem 2L2 (τr ) , ∀r ∈ R∗ ,

O
ητ2r (z) = z − ∇ufem 2L2 (τr ) + 1
σr Πr1 f − σr ufem + div z2L2 (τr ) , ∀r ∈ R  R∗ ,

PR
(12.9)
there hold the bounds 249


2
|||efem |||2 ≤ ητr (z) + oscτr (f )
D
, (12.10)
τr ∈Th
TE

 
250
2
ηΩ (z) = ητ2r (z) ≤ C |||efem ||| + 2
osc2τr (f ) , (12.11)
r∈R r∈R
EC

 
hr √1
where oscτr (f ) = min π , σr f − Πr1 f L2 (τr ) . 251
R

Proof See Ainsworth and Vejchodsky [3] for the proof. In this work the
bounds (12.10), (12.11) in a more general form are derived under more general
R

conditions. In particular, ΓN = ∅, the bound (12.11) is proved in the local version,

i.e., with ητ2r (z) on the left and with the restriction of the right part to the patch
O

δ (r) . 

C

We present also the error majorant of Cheddadi et al. [14] for approximate 252
N

solutions of the reaction-diffusion problem by the method of vertex-centered finite 253

volumes. In this work the point wise equilibration is avoided and replaced by the 254
U

much more flexible equilibration in a weak sense. Let us introduce the notations: 255

Dh is the dual in respect to Th partition of Ω; Sh is the finer mesh, induced by the 256

partition Dh ; D is the polygon with the center in the vertex of triangulation Th and 257

containing all simplices of the finer mesh with this vertex, hD is its diameter; Dhint 258

is the set of all polygons D, for which ∂D ∩ ∂Ω = ∅. For additional information 259

about these entities, we refer to [14]. 260

12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

Theorem 12.3 Let uh be the solution by the method of vertex-centered finite 261

volumes, eh = u − uh , vector-function z ∈ H(Ω, div) satisfy the equalities 262

(f − div z − σ uh , 1)D = 0 , ∀D ∈ Dhint , (12.12)

and γD = min(CD h2D , σD−1 ), where CD is the constant from the Poincaré inequality 263

for the polygon D. Then 264

 √
2
|||eh |||2 ≤ ηΩ
2
(z) = ∇uh + zL2 (D) + γD f − σ uh − div zL2 (D) .
D∈Dh

F
(12.13)

O
Proof Theorem is one of the results of [14], see Theorem 4.5. 


O
Majorants in (12.5)–(12.7) have obvious merits, but are not consistent at the 265

application, e.g., to solutions by the finite element and other mesh methods. If v = 266

PR
ufem is the finite element solution to the problem (12.1) at ΓD = ∂Ω, ψD ≡ 0, 267

and σ = 0, then we can use 12.5). For our purpose, it is sufficient to consider the 268

approximate solutions from the space Vh (Ω) = V1h,p (Ω)⊂C 1 (Ω)∩H 2 (Ω). Under 269

the assumption u ∈ H l (Ω), we have a priori error bounds [16, 21]

D
270

u − vH k (Ω) ≤ chl−k uH l (Ω) , k = 1, 2, k ≤ l ≤ p +1.

TE

(12.14)

In particular, if f ∈ L2 (Ω) and consequently l = 2, see [25], then according 271

to (12.14) the left part of (12.5) is estimated as O(h2 ). At the same time at the 272
EC

choice z = −∇v the first term in the right part of (12.5) vanishes, but cΩ (1 + 273

 )div z−f L2 (Ω) at any  > 0 can be bounded from above only by a constant. The
1 2 274

bounds (12.14) are not improvable in the order. For u ∈ H 2 (Ω), Ω ∈ R2 , Oganesian 275
R

and Ruhovets [31] gave the proof by estimating the corresponding Kolmogorov’s 276
R

width. From their results and results on the regularity of solutions of (12.1), it 277

follows existence of such f ∈ L2 (Ω) that u ∈ H 2 (Ω) and the second summand on 278
O

the right of (12.5) is estimated by the constant from below. Therefore, the orders of 279

smallness of the left and the right parts of the a posteriori bound are different, and 280
C

the value O(h2 ) on the left is estimated by the right part only with the order of unity. 281

If l > 2, the left and the right parts are estimated with not equal orders O(h2(l−k) ) 282
N

and O(h2(l−k−1)). 283

Inconsistency of the majorant (12.7) at σ ≤ ch−α , 0 ≤ α < 2, c = const, is

U

284

established in a similar way. Majorant (12.6) is consistent in 1-dimensional case, in 285

general it is also inconsistent. The inconsistency of the above mentioned majorants, 286

obviously, is retained, if finite elements of the class C are used and the test flux is 287

found by some recovery procedure, satisfying only the requirements α), γ ). 288

The equalities of the orders of smallness of the left and right parts of the 289

majorants (12.10) and (12.13) are well provided, as follows, e.g., from (12.11) 290

and similar bound, proved in [14]. However, it is achieved only for the test 291

fluxes, satisfying the additional conditions reflecting the requirement (β), see (12.8) 292

and (12.12). 293

 V. G. Korneev

12.3 A Posteriori Error Majorant Robust for Piece Wise 294

Constant and Constant Reaction Coefficients 295

In this section, we present a general a posteriori error bound for the approximations 296

of the exact solution to the reaction-diffusion equation with the piecewise constant 297

reaction coefficient. These bounds are not influenced by the origins of the approx- 298

imations and are robust in the respect to the reaction coefficient. In particular, they 299

are well defined in the vicinity of zero. 300

F
12.3.1 Reaction-Diffusion Problem with Piece Wise Constant

O
301

Reaction Coefficient 302

O
Let for some σ∗ ≡ const > 0, the piece wise constant function σ† be

PR
303

⎧
⎨ σ∗ , ∀ σ ∈ Ω∗ ,
σ† =
⎩
∀ σ ∈ Ωσ ,
D
σ,
TE

where Ω∗ := {x : σ (x) < σ∗ } and Ωσ = Ω \ Ω∗ . The basic assumption for what 304

follows is that for some σe ≥ 0 the error e = u − v satisfies the inequality 305
EC

1
(σ† − σ )1/2 e2L2 (Ω) ≤ (σ† − σ )1/2 ∇e2L2 (Ω) , (12.15)
σe

 be replaced by Ω∗ . We use the notation HψN (Ω, div) =

and, obviously, Ω in it can 306
R


{y ∈ H(Ω, div), ∇y · n = ψN }. 307
R

∂Ω
Lemma 12.1 Let σ be an arbitrary nonnegative element wise constant function as 308
O

in 12.2), f ∈ L2 (Ω), ψD ∈ H 1 (Ω), ψN ∈ L2 (ΓN ). Let also σe satisfy (12.15) and 309

σ∗ > 0 be arbitrary. Then for any function v ∈ H̊ψ1D (Ω) and any vector-function 310
C

z ∈ HψN (Ω, div) we have 311

N

γ   2
γ (∇v + z)2L2 (Ω) +  √
|||v − u|||2 ≤  σ† f − σ v − div z L2 (Ω) (12.16)
U

with 312

⎧
⎨ (σe + σ∗ )/(σe + σ ) , σ ∈ [0, σ∗ ] ,
2 =
γ (12.17)
⎩
1, σ ≥ σ∗ .

Proof Two additional features to the proof of the Aubin’s bound are a more tricky 313

use of the Cauchy inequality |ab| ≤  −1 a 2 + b2 , ∀ > 0, and the use of the 314
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

inequality (12.15). For an arbitrary v and z from Lemma 12.1, after integration by 315

parts of the first summand in the square of the energy norm of the error e = v − u, 316

we have 317

     
||| e |||2 = Ω ∇e · ∇e + σ e2 = Ω (∇v + z) · ∇e − (z + ∇u) · ∇e + σ e2 =
  
= Ω (∇v + z) · ∇e + [div (z + ∇u) + σ e] e =
   
= r τr (∇v + z) · ∇e − [f − σ v − div z] e .
(12.18)

F
√ √

O
For convenience, we will use also the notations κ = σ , κ∗ = σ∗ , and κ† = 318
√
σ† . By means of the Cauchy inequality, for γ (x) > 0 belonging to L2 (Ω) we find 319

O
that 320

PR
||| e |||2 = ∇e 2L2 (Ω) + κ e 2L2 (Ω) ≤ γ (∇v + z)2L2 (Ω) +

1/2
1/2
κ
+ κγ† (f − σ v − div z)2L2 (Ω) ×  γ1 ∇e 2L2 (Ω) +  γ† e 2L2 (Ω) .
D
(12.19)
TE

 In the case under consideration, it is natural to take γ element wise constant, i.e., 321

γ  = γr = const, ∀r ∈ R. The summand in the second multiplier of the right part 322
EC

τr
of (12.19), corresponding to one finite element τr for r ∈ R∗ := {r : τr ∈ Ω∗ }, we 323

multiply by γr and with the use of some βr ∈ (0, 1] rearrange as follows: 324
R

∇e 2L2 (τ ) + σ∗  e 2L2 (τr ) = ||| e |||2τr + (σ∗ − σ ) e 2L2 (τr ) =

r
R

(12.20)
= ||| e |||2τr + βr (σ∗ − σ ) e 2L2 (τr ) + (1 − βr )(σ∗ − σ ) e 2L2 (τr ) .
O

Summation of (12.20) over r ∈ R yields 325

C

1/2
σ† R
N

 γ1 ∇e 2L2 (Ω) +  γ e 2L2 (Ω) = 1

r=1 γr2 ||| e |||2τr +

(12.21)
U

+βr (σ† − σ )1/2 e 2L2 (τr ) + (1 − βr )(σ† − σ )1/2 e 2L2 (τr ) .


Let β be the element wise constant function with the restrictions β τr = βr . On Ω∗ 326

it can be uniquely defined from the equality 327

σ 1 + (σ† − σ ) = (1 − β)(σ† − σ ) + σ , x ∈ Ω∗ , (12.22)

σe
 V. G. Korneev

as β = σe /(σe + σ ) and continued by the unity on Ωσ . Estimating the sum of the 328

second terms in the brackets by means of (12.15) and combining (12.20) – (12.22) 329

at γ 2 = β, we conclude that 330

κ
 γ1 ∇e 2L2 (Ω) +  γ† e 2L2 (Ω) ≤

1/2
1/2
≤  γ1 1 + (σ† − σ ) σβe ∇e 2L2 (Ω) +  γ1 (1 − β)(σ† − σ ) + σ e 2L2 (Ω) ≤

σ∗ +σe
≤ ∇e 2L2 (Ω) + κe 2L2 (Ω) .

F
σe
(12.23)

O



O
Combining (12.19) with (12.23), we come to the bound (12.16).
Let us make a few remarks in relation to the bound (12.16). Since σ∗ and σe are 331

PR
constants, one can set σ∗ = c∗ σe , c∗ = const, and, therefore, (σ∗ +σe )/σe = 1+c∗ . 332

Obviously, γ ∈ [1/2, 1] and, if σ∗ ≤ σe , then (1 + c∗ ) ≤ 2 . Therefore, for σ∗ ≤ σe 333

the bound (12.16) can be written in a ruder, but simpler form 334
D
1  

||| e |||2 ≤ 2 (∇v + z)2L2 (Ω) +  √ f − σ v − div z 2L2 (Ω) . (12.24)

TE

σ†

At σ ≡ 0 the bound, (12.16) differs from (12.5) with  = 1 only by the coefficient 335
EC

before the second term on the right. In Sect. 12.4, we show that for the finite element 336

solutions of the problem in the domains with the sufficiently smooth boundaries this 337

coefficient will have, indeed, the order of σe−1 = O(h2 ) instead of the constant cΩ 338

in (12.5). In general, e.g., at piece wise constant or constant σ , it is possible to use 339
R

also rough values σe = σ∗ = 1/cF in (12.16) and (12.24), where cF = cF (Ω, ΓD ) 340
R

is the constant from the Friedrichs type inequality 341

O

(σ† − σ )1/2 φ2L2 (Ω) ≤ cF (σ† − σ )1/2 ∇φ2L2 (Ω) , ∀φ ∈ H̊Γ1D (Ω) . (12.25)
C

However, in this case minimization of the right parts of the a posteriori bounds with 342

respect to z becomes often unavoidable. 343

N
U

12.3.2 Reaction-Diffusion Problem with any Constant σ ≥ 0 344

The case σ = const, considered in this subsection, has an independent significance 345

and, besides, will allow us simpler transfer to more specific a posteriori error bounds 346

for the approximate solutions by the finite element method. Below we reformulate 347

the basic assumption (12.15) in a two simpler forms and give the respective version 348

of the preceding Lemma. 349

12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

In general, when v ∈ H̊Γ1D (Ω) is any approximation for u, σ∗ can be defined 350

from the inequality 351

|u − v|21 /u − v20 ≥ σ∗ > 0 . (12.26)

If v = uG is the approximate solution by the Galerkin method in the subspace 352

V(Ω) ⊂ H̊Γ1D (Ω, ), then this condition can be relaxed and it is sufficient that σ∗ 353

satisfies 354

|u − v|21 /u − Qu20 ≥ σ∗ > 0 . (12.27)

F
where Q is the operator of orthogonal projection L2 (Ω) → V(Ω), i.e., such that

O
355

for ∀φ ∈ L2 (Ω) we have (Qφ, ψ)Ω = (φ, ψ)Ω , ∀ ψ ∈ V(Ω) . 356

Let function fˆ(x), x ∈ Ω, be such that fˆ(x) = Πrl f for x ∈ τr , r =

O
357

1, 2, . . . , R and l is some nonnegative integer, e.g., related to the accuracy of the 358

PR
approximation v. In Theorem below domains τr can be understood as arbitrary 359

convex subdomains of some partition of the domain 360

R

D
Ω = interior{ τr} , τr ∪ τr  = ∅, r = r  , diam[τr ] = hr , 361

1
TE

for which the Poincaré inequalities hold, see, e.g., Nazarov and Poborchi [29], 362
EC

hr
inf φ − cL2 (τr ) ≤ |φ|H 1 (τr ) , φ ∈ H 1 (τr ) . 363
c∈R π

Lemma 12.2 Let f ∈ L2 (Ω), 0 < σ = const, σ∗ satisfy the inequality (12.26),
R

364

ΓD = ∂Ω, ψD ∈ H 1/2 (∂Ω). Then, for any function v ∈ H 1 (Ω) that satisfies the 365
R

boundary condition on ΓD and any z ∈ H(Ω, div) we have 366

O

||| e |||2 ≤ ΘM , M = M(σ, σ∗ , f, v, z) , (12.28)

C

where 367
N

M = ∇ v + z2L(Ω) + θ f − σ v − div z2L2 (Ω) , (12.29)

U

and 368

⎧ ⎫
⎨ 2/(1 + k) , ∀ σ ∈ [0, σ∗ ] ⎬
Θ=
⎩ ⎭
1, ∀ σ > σ∗
⎧ ⎫
⎨ 1/σ∗ , ∀ σ ∈ [0, σ∗ ] ⎬
,θ= . (12.30)
⎩ ⎭
1/σ , ∀ σ > σ∗
 V. G. Korneev

with k = σ/σ∗ . Besides, for σ ∈ [0, σ∗ ] and σ ≥ σ∗ , respectively, we have the 369

bounds 370

 h2r 
||| e |||2 ≤ Θ1 M(σ, σ∗ , fˆ, v, z) + r επ 2 τr (f − Πrl f )2 dx , ∀ε > 0,
 
||| e |||2 ≤ Θ2 M(σ, σ∗ , fˆ, v, z) + 1
r σ τr (f − Πrl f )2 dx ,
(12.31)

where 371

F
⎨ (2 + ε)/(1 + k) , 0 ≤ σ ≤ σ∗ /(1 + ε) ,
Θ1 =

O
⎩
1+ε, σ∗ /(1 + ε) ≤ σ ≤ σ∗ ,

O
Θ2 = 1 + 1/(1 + k−1 ) .

PR
If v = uG is the approximate solution by the method of Galerkin in the space 372

V(Ω) ⊂ H̊Γ1D (Ω, ), then the bound (12.28)–(12.30) takes place with σ∗ , satisfying 373

to the inequality (12.27) and z = zG := −∇ uG , i.e., 374

D
|||uG − u|||2 ≤ ΘMG ,
TE

(12.32)
MG = MG (σ, σ∗ , f, uG , zG ) = θ f − σ uG − div zG 2L2 (Ω) .
EC

Proof The direct proof of Lemma 12.2 is given in [25]. The proof of the
bound (12.28)–(12.30) follows also from (12.16) at σ∗ = σe . 

Remark 12.1 Other ways of obtaining majorants similar to (12.28)–(12.30) 375
R

and 12.31) are possible. One can consider the subsidiary problem 376
R


− u + λ u = fλ,σ , x ∈Ω, u∂Ω = 0 , (12.33)
O

with an arbitrary λ ≥ σ∗ and fλ,σ = f + (λ − σ )u, whose solution is the same as 377
C

for the problem (12.1) at the ΓN = ∅ and ψD ≡ 0. In the respective to (12.33) 378

Aubin’s majorant for the approximation v of the solution to the problem (12.33), 379
N

the approximation of the problem (12.1) can be used. The substitution of fλ,σ = 380

f +(λ−σ )u in this Aubin’s majorant, the application of the Cauchy inequality with
U

381

ε, the use of the inequality 12.26), and some manipulations produce the subsidiary 382

majorant, depending on λ, ε and β. At that, β plays similar role to the one at the 383

obtaining (12.16). In this way, we come to the set of majorants similar to those in 384

Lemma 12.2, and, obviously, by changing the choice of λ, β, and ε, we can change 385

the weights before the first and second norms in the majorants. If the ratio of the 386

weights is fixed, a suitable majorant can be obtained by the minimization of the 387

subsidiary majorant with respect to λ, β and ε. 388

12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

The bound (12.28)–(12.30) generalizes the Repin-Frolov’s bound (12.5) for 389

σ ≡ 0 upon interval σ ∈ [0, σ∗ ] with σ∗ ≤ 1/cΩ and the coefficients not greater 390

than in (12.5), i.e., Θ ≤ 2 and θ ≤ cΩ . At the same time (12.28)–(12.30) can be 391

considered as an updated Aubin’s bound (12.4), coinciding with it at σ ≥ σ∗ and 392

well defined for all σ ≥ 0. An additional unique feature of the bound (12.28)– 393

(12.30), if to compare with other attempts [15, 33, 35] to update the Aubin’s bound, 394

is that its accuracy can be significantly improved by the use of improved estimates 395

of σ∗ for concrete numerical methods. 396

F
12.4 Consistent A Posteriori Majorants for Finite Element 397

O
Method Errors 398

O
For specific classes of approximate solutions and, in particular, for solutions by the 399

PR
finite element method, the critical values σ∗ of the reaction coefficient in the derived 400

error majorants can be successfully estimated. In this section, we assume that σ ≥ 0 401

is a piece wise constant function (12.2) satisfying μ1 ≤ σ ≤ μ2 , ∀x ∈ Ω. 402

Lemma 12.3 Let ΓD = ∂Ω, σ ≥ 0 be the piece wise constant function satisfying 403
D
(12.2), ψD ≡ 0, f ∈ H −1 (Ω), the finite element assemblage generates the space 404

Vh,p (Ω), p ≥ 1, and efem = ufem − u. Then

TE

405


μ2
efem 0 ≤ c† h|efem |1 , c† = c cap , (12.34)
EC

μ1

with the constants c , cap , defined below, see (12.38), (12.39). 406
√
R

Proof Let us set κ = + σ and introduce notations u◦ , ufe and us for the functions 407

minimizing κ(u − φ)20 , |u − φ|21 , and h−2 κ(u − φ)20 + |u − φ|21 , respectively, 408
R

among all φ ∈ V˚h (Ω) and notations for the respective errors e◦ = u◦ − u, efe = 409

ufe − u and es = us − u. Since ufem minimizes |||u − φ|||2, φ ∈ V˚h (Ω), we conclude
O

410

that 411
C

|efem |21 + κ efem 20 ≤ |u − 

u|21 + κ (u − 
u)20 , (12.35)
N

where u can be any from functions  u = u◦ , ufe , us . If to take into attention the
U

412

inequalities efem 0 ≥ e◦ 0 and |efem |1 ≥ |efe |1 , following from the definitions of 413

functions u◦ and ufe , then (12.35) implies 414

κ efem 0 ≤ κ efe 0 ,
(12.36)
|efem |1 ≤ |e◦ |1 .
 V. G. Korneev

Let ϕ ∈ H̊ 1 (Ω) be the solution of the problem 415

aΩ (v, ϕ) = (v, efe )Ω , ∀ v ∈ H̊ 1 (Ω). (12.37)

Obviously, efe ∈ L2 (Ω), and, as a consequence, at sufficiently smooth domain ϕ ∈ 416

H 2 (Ω) and, see Ladyzhenskaya [27], 417

ϕ2 ≤ c efe 0 , c = c (Ω) . (12.38)

ϕ ∈ V˚h,p (Ω), for which

The function ϕ can be approximated by the function  418

F
|ϕ − 
ϕ |21 ≤ cap
2 2
h ϕ22 ≤ c cap h efe 20 .
2 2 2

O
(12.39)

O
Since in Lemma 12.3 only the constant cap but not the function  ϕ itself is used, one 419

can imply by  ϕ any approximation providing a good constant. Hence, in 2D  ϕ can 420

PR
be the interpolation and, in general, quasi-interpolation of Scott and Zhang [34], or 421

Bartels et al. [9], or one of the functions 

ϕ◦ , 
ϕfe , 
ϕs etc. 422

Estimating efe 0 by means of the Aubin-Nitsche trick [5] for the prob- 423

lem (12.37) and using the bound (12.39), we get: 424

D
efe 20 = aΩ (efe , ϕ) ≤ inf ˚ (Ω) |aΩ (efe , ϕ − w)| ≤
TE

w∈V h
(12.40)
|efe |1 inf ˚ (Ω) |ϕ − w|1 ≤ |efe |1 |ϕ − 
ϕ |1 ≤ c cap h|efe |1 efe 0 .
w∈V h
EC

This bound together with the inequality (12.36) and the definitions of functions
efe , efem results in the bound (12.34). 

R

Theorem 12.4 Let ΓD = ∂Ω, ψD ≡ 0, and u ∈ H̊ 1 (Ω, ). Let also the finite 425

element assemblage generate the space V˚h,p 0

(Ω) ⊂ H̊ 1 (Ω), p ≥ 1, and ufem be 426
R

the solution by the finite element method. Then for σ satisfying 0 ≤ σ ≤ σ∗ = 427

1/(c† h)2 with c† , as in (12.34), and any z ∈ H(Ω, div) we have

O

428

(1)
C

|||efem |||2 ≤ 2
Mfem (σ, f, z) ,
1+c2† h2 σ

(1)
Mfem (σ, f, z) = |∇ ufem + z|2L2 (Ω) + c2† h2 f − σ ufem − div z2L2 (Ω) .
U

(12.41)

Under the condition A ), for σ ≤ σ∗ /(1 + ε) it holds also the bound 429

 h2 
2+ε (1) ˆ, z) + r p
|||efem ||| ≤
2
M (σ, f (f − Πr f )2 dx , ∀ε > 0.
1 + c2† h2 σ fem r
επ 2
τr
(12.42)
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

(1)
Proof Since Mfem (σ, f, z) = M(σ, σ∗ , f, v, z) with σ∗ defined according
to (12.34), Theorem is a direct consequence of Lemmas 12.2 and 12.3. 

Similar a posteriori error bounds can be derived in a quite similar way for the 430

finite element method solutions to the 2nth -order elliptic equations, n ≥ 1, see 431

Korneev [24]. The most complicated for their practical applications is the evaluation 432

of the constant c . However, in many cases such estimates are well known. For 433

instance, if the domain is convex, then c = 1, see Ladyzhenskaya [27, (6.5) in ch. 434

II]. 435

Existence of the constant c puts some restrictions on the smoothness of the 436

F
boundary and coefficients of the equation (if they are not constant). At the same 437

time, there is a possibility to avoid the mentioned additional restrictions, except for 438

O
those related to the suitable approximation operator. If there exists some interpola- 439

tion type or other approximation operator with locally defined approximations for 440

O
functions from H 1 (Ω), then constants in the a posteriori bounds may depend only 441

on the local approximation properties of the finite element space. A good example of

PR
442

such an operator is the quasi-interpolation operator of Scott and Zhang [34], which 443

will be used below to illustrate the statement. We start from the description of its 444

properties. 445

Let Ω ⊂ Rm , m ≥ 2, be a bounded Lipschitz domain, which is the domain of

D
446

the quasiuniform triangulation Th with vertices x (i) , i = 1, 2, . . . , I, and simplices 447

TE

τr of diameters not greater h. For simplicity it is assumed that faces of simplices are 448

plain and that the following quasiuniformity conditions are fulfilled: 449
EC

0 < c ≤ ρ r /hr , α̂ (1) h ≤ hr ≤ h , (12.43)

where ρ r and hr are the radius of the largest inscribed sphere and the diameter of 450

simplex τr , respectively. To each vertex x (i) , we relate (m − 1)-dimensional simplex

R

451
(m−1)
τi , which is the face of one of the simplices τr and has x (i) for the vertex. For 452
R

(m−1) (i)
m vertices of the simplex τi we will use also notations zl , l = 1, 2, . . . , m, 453
O

(i) (m−1)
assuming for definiteness that z1 = x (i) . Clearly the choice of the face τi is 454
(m−1)
not unique, but for x ∈ ∂Ω we always take one of the faces τi
(i) ⊂ ∂Ω. We 455
C

will formulate the result of Scott and Zhang using the simpler notations V (Ω), 456

Vtr (∂Ω), and V˚ (Ω) for the space of continuous piece wise linear functions
N

457

Vh,1
0
(Ω), its trace on the boundary, and its subspace of functions, vanishing on the 458
U

boundary, respectively. 459

We define functions θi ∈ P1 (τi(m−1) ), satisfying equations 460


θi λ(i)
l dx = δ1,l , l = 1, 2, .., m ,
(m−1)
τi

where λ(i)
l are the barycentric coordinates in τi
(m−1)
, corresponding to the vertices 461
(i)
zl , and δi,l is the Kronecker’s symbol. If φi ∈ V (Ω) are the basis functions in
 V. G. Korneev

V (Ω), defined by the equalities φi (xj ) = δi,j , i, j = 1, 2, .., I, then for any 462

v ∈ H 1 (Ω) the quasi-interpolation Ih v is the function 463

 

I
Ih v = θi v dx φi (x) .
(m−1)
i=1 τi

Lemma 12.4 The quasi-interpolation operator Ih : H 1 (Ω) → V (Ω) is a 464

projection and has the following properties: 465

(a) Ih v : H 1 (Ω) → V (Ω) and, if v ∈ V (Ω), then Ih v = v, 466

F
(b) (v − Ih v) ∈ H̊1 (Ω), if v|∂Ω ∈ Vtr (∂Ω), 467

v − Ih vt,Ω ≤ csz (t, s)hs−t vs,Ω for t = 0, 1, s = 1, 2, and ∀v ∈ H s (Ω) ,

O
(c) 468

(d) |Ih v|1,Ω ≤ c̆sz | v |1,Ω and Ih v1,Ω ≤ ĉsz  v 1,Ω , ∀v ∈ H 1 (Ω), 469

O
where csz (s, t), c̆sz , and ĉ are positive constants, depending on c. 470

PR
Proof Scott and Zhang [34] proved a much more general result. In the given form
the Lemma was formulated and proved by Xu and Zou [37]. 

Theorem 12.5 Let ΓD = ∂Ω, ψD ≡ 0, u ∈ H̊ 1 (Ω, ). Let also the finite element 471

assemblage satisfy the condition A) and generate the space V˚ (Ω) ⊂ H̊ 1 (Ω) and
D
472

z ∈ H(Ω, div). Then at any σ ≡ const ∈ [0, 1/(csz (0, 1)h)2 ] there holds the bound 473
TE

(2) (2) −1
|||efem |||2 ≤ Θsz Mfem (σ, f, z) , Mfem (σ, f, z) = M(σ, θsz , f, ufem , z) ,
(12.44)
EC

where 474

1 + 2 (1, 1)
csz
R

Θsz = , θsz = csz (0, 1)2 h2 , (12.45)

1 + csz (0, 1)h2 σ
2
R

and 
csz (1, 1) is the constant, depending only upon c and α̂ (1) found in (12.43). 475
O

Proof For any w ∈ V˚h,1

0 (Ω) we have the equality 476
C

  
|||efem |||2 = ∇(efem ) · ∇(efem ) + σ efem efem =
N

Ω
 
U

= Ω (∇ufem + z) · ∇(efem + w) − (z + ∇u) · ∇(efem + w)+ (12.46)


+σ (ufem − u)(efem + w) .
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

Integration by parts of the second summand in square brackets of the right part and 477

application of the Cauchy inequality with  > 0 result in the inequality 478

 
|||efem |||2 = Ω (∇ufem + z) · ∇(efem + w)+
  
div z + u + σ (ufem − u) (efem + w) ≤
 1/2 (12.47)
∇ufem + z20 + 1 f − σ ufem − div z20 ×

 1/2

F
∇(efem + w)20 + efem + w20

O
According to Lemma 12.4 and the definition of the operator Q of L2 -projection

O
479

upon V˚h,1
0
(Ω), for φ − Qφ with any φ ∈ H 1 (Ω), there are valid the bounds 480

PR
φ − Qφ0 ≤ φ0 ,

φ − Qφ0 ≤ csz (0, 1)h∇φ0 , (12.48)

D
∇(φ − Qφ)0 ≤ 
csz (1, 1)∇φ0 ,
TE

in which the constant 

csz (1, 1) depends only on c and α̂ (1) . The proof is needed only 481
EC

for the last bound, and it follows by the relations 482

∇(φ − Qφ)0 ≤ ∇(φ − Ih φ)0 + ∇(Ih φ − Qφ)0 ≤ c̆sz ∇φ0 +

R

c1,0 h−1 Ih φ − Qφ0 ≤ c̆sz ∇φ0 + c1,0 h−1 Ih φ − φ0 + φ − Qφ0 ≤ 483
R

 
O

c̆sz + 2c1,0csz (0, 1) ∇φ0 = 

csz (1, 1)∇φ0 ,
C

where c1,0 is the constant in the inverse inequality ∇(Ih φ − Qφ)0 ≤ 484

c1,0 h−1 Ih φ − Qφ0 . Therefore, 

N

csz (1, 1) = c̆sz + 2c1,0csz (0, 1). 485

It is worth noting, that the third inequality (12.48), indicating stability in 486
U

H 1 (Ω) of L2 -projection, was proved by Bramble and Xu [10] differently with the 487

differently defined constant 

csz (1, 1). 488
 V. G. Korneev

For the reason that w = Qefem ∈ V˚h,1 0 (Ω), it can be adopted w = Qe

fem . 489
−2
Combining with (12.48) and setting  = σsz := (csz (0, 1)h) lead to the bound 490

∇(efem + w)20 + σsz efem + w20 =

∇(efem + w)20 + βσ efem + w20 + +(σsz − βσ )efem + w20 ≤ (12.49)

σsz −βσ
2 (1, 1)∇e
csz fem 0 + βσ efem 0 +
2 2
σsz ∇efem 0 .
2

F
On the basis of (12.49) we conclude that 491

O
1 +2 (1, 1) 
csz 
∇(efem) + w20 + σsz efem + w20 ≤ ∇efem 20 + σ efem 20
1+k

O
(12.50)

PR
with k = σ/σsz . Now from (12.47) and (12.50) the Theorem follows . 

Remark 12.2 Quasi-interpolation operator Ih is defined in [37] on triangulations, 492

satisfying the conditions of the shape regularity 0 < c ≤ ρ r /hr , hr ≤ h, with 493
D
preserving the properties (a), (b) and the properties (c), (d) taking the local form 494
TE

c) v − Ih vt,τr ≤ csz (t, s)hs−t

r vs,δr , t = 0, 1, s = 1, 2, ∀v ∈ H 1 (δr ) ,

d) |Ih v|1,τr ≤ c̆sz | v |1,δr , and Ih v1,τr ≤ ĉsz  v 1,δr , ∀v ∈ H 1 (δr ) ,
EC

(12.51)

where csz (t, s) = const, δr = interior{∪ τ  : τ  ∩ τ r = ∅} and r = 1, 2, . . . , R.

R

495

More over, these authors designed also the quasi-interpolation operators Ih,p : 496
R

H 1 (Ω) → Vh,p (Ω), for which again the properties (a), (b) are preserved, but 497

in (12.51) s = 1, 2, . . . , p + 1. This quasi-interpolation operator allows to expand 498

O

the a posteriori error bounds (12.44)–(12.45) on the solutions by the finite element 499

methods from the spaces Vh,p (Ω), p > 1. 500

C

The bounds (12.41), (12.42) of Theorem 4 essentially use some global properties 501
N

of the finite element solutions, see Lemma 12.3. In the bound (12.44)–(12.45) of 502

Theorem 12.5, we see only the constants defined by local properties of the quasi- 503
U

interpolation operator. 504

12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

12.5 Consistency and Local Effectiveness 505

For the right parts of the a posteriori error bounds (12.41), (12.44)–(12.45), 506

and (12.42) we introduce the notations ηk , k = 1, 2, 507

2 (1) (2)
η12 = M (σ, f, z) , η22 = Θsz Mfem (σ, f, z) , 508
1 + c2† h2 σ fem

and η1,ε , respectively. The error estimators ηk , η1,ε are consistent, they are 509

computable with the use of the testing fluxes which make the estimators sharp and 510

F
which can be calculated by a number of simple flux recovery procedures of linear 511

O
complexity. These facts lie practically on the surface and are established similarly 512

for all ηk , η1,ε . Hence, they are discussed below briefly for η1 . In support of the 513

O
effectiveness of the error indicators, we present also the inverse like inequality at 514

some choice of the testing fluxes. 515

PR
If ufem ∈ Vh,p (Ω) ⊂ H 1 (Ω), Vh,p (Ω)  H 2 (Ω) and u ∈ H l (Ω), 2 ≤ l ≤ 516

p + 1, p ≥ 1, then it is natural to require that for the recovered flux z the estimates 517

of convergence 518
D
h−1 ∇u + z)L2 (Ω) , div(∇u + z)0 ≤ c̃l hl−2 ul,Ω , c̃l = const , (12.52)
TE

hold with the orders corresponding to the convergence estimates of efem k ≤ 519

ck,l hl−k ul , k = 0, 1. From these bounds, it easily follows that for σ ≤ c†2 h−2 520
EC

|||efem ||| ≤ cl hl−1 ul , η1 (efem ) ≤ cl hl−1 ul , cl , ĉ1 = const , (12.53)

see [23–25] for additional details and discussion.

R

521

Thus, according to (12.53) the a posteriori bound (12.41) is consistent for σ ≤ 522

c†2 h−2 , and as the a priori bound is unimprovable in the order, so the same is true for
R

523

the bound (12.41). In other words, in the class of solutions, for which the a priori 524
O

bound is unimprovable in the order, there exist such that 525

C

η1 (efem ) ≤ C ||| efem ||| . (12.54)

N

Indeed, let f ∈ L2 (Ω), d = 2, and the domain is such thatthe inequality v2 ≤ 526

U

c◦ F 0 holds for solutions of the problems −v = F, u = 0 uniformly in 527

∂Ω
F ∈ L2 (Ω), see Ladyzhenskaya [27]. Then the inequalities (12.53) with l = 2 are 528

fulfilled as well. At the same time, such f ∈ L2 (Ω) exists that 529

hu2 ≤ c2 infφ∈Vh (Ω) ∇(u − φ)0 = c2 ∇efe 0 ≤

(12.55)
c2 ∇efem 0 ≤ c2 ||| ∇efem ||| , c2 = const .
 V. G. Korneev

The first of these inequalities follows from the estimates of the N-width of the 530

compact of functions v ∈ H̊ 1 (Ω), v1 = 1, for N = h−2 , see in [31] 531

Ch. 4, Section 4.1. In turn, (12.54) follows from (12.53), (12.55), which together 532

with (12.41) yield the two sided bound 533

||| efem |||2 ≤ η1 (efem ) ≤ C ||| efem |||2 . (12.56)

The error indicators ηk , η1,ε are computable for a wide range of problems and 534

FEM’s, and there are numerous works on the flux recovery procedures, which 535

can be used for obtaining the testing fluxes z satisfying (12.52) and, therefore, 536

F
making this indicators sharp. Majority of them, following the renown SPR of 537

[39], were created in the relation with the development of the residual type error 538

O
estimators η = ||∇ufem + Gufem ||L2 (Ω) , where G is the flux recovery operator. 539

For this reason, these procedures were aimed at achieving as high as possible

O
540

accuracy, in particular, superconvergence or even ultraconvergence [36, 38]. Under 541

PR
some assumptions on the finite element meshes, smoothness of the exact solutions, 542

finite element discretization and the operator G, these properties were approved 543

by the analysis and by the numerical practice. As is noted by some authors, flux 544

recovery procedures perform “astonishingly well”, see, e.g., [1, 36]. Clearly, the 545
D
superconvergence means that even better than (12.52) in the order bounds hold. 546

For instance, the superconvergence of fluxes produced by the simple and low cost 547
TE

procedures of linear complexity such as 548

(i) weighted averaging, 549

(ii) local L2 -projection, and

EC

550

(iii) the local discrete least squares fitting 551

for the first order finite element methods was studied in [36]. Under the assumptions 552

u ∈ H̊ 1 (Ω)∩H 2(Ω)∩W∞
R

3 (Ω) and mild regularity of the mesh, Theorem 4.2 of this 553

work establishes that ||∇ufem + Gufem ||L2 (Ω) = O(h1+ρ ) with ρ > 0 depending 554
R

on the mesh. Superconverging averaging techniques for p Lagrange elements were 555

studied in [8]. There are many other papers devoted to the expansion of efficient 556
O

recovery techniques on the more irregular meshes, problems with the discontinuous 557

coefficients etc. It is worth to stress again that the inequalities (12.52) do not assume
C

558

any superconvergence, but only retention by the testing flux z of the orders of 559
N

accuracy of the finite element flux zfem = −∇ufem . In general, the proof of (12.52) 560

does not differ much from the proofs of similar bounds for approximations by 561
U

functions from the finite element space. In particular, according to Lemma 4.3 of 562

[1], see also Corollary 4.2 there, under conditions of Theorems 12.4 and 12.5, for 563

the flux recovery procedures (i)–(iii) the inequalities (12.52) are fulfilled. 564

Obviously also, that for the global L2 (Ω) orthogonal projection of zfem upon 565

Vh, (Ω) or Wh, (Ω, div) with  = p or  = p − 1, the proof of (12.52) is very 566

easy while the complexity of such projection is linear. 567

The bounds of local effectiveness of the a posteriori error majorants, leading as 568

a rule to two-sided error bounds, are paid much attention, see [2, 3, 11–13, 18, 19]. 569

In part, it is for the reason that they are used in the proofs of the convergence of 570
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

adaptive algorithms. Here we give an example of the local effectiveness bound for 571

the a posteriori majorant (12.41) applied to the first order finite element method. 572

Theorem 12.6 Under conditions of Theorem 12.4, for ufem ∈ Vh,1 (Ω) and the 573

vector-function z ∈ Wh,2 (Ω, div), defined as L2 -projection of the vector-function 574

zfem = −∇ ufem on the space Wh,2 (Ω, div), we have 575

 R 2 
 
hr
M(k)
fem (σ, f, z) ≤ C |||e fem ||| 2
+ 2
(f − Πr1 f )2 dx (12.57)
π τr
r=1

F
with the constant C = C(Ω, c ) and k = 1, 2. 576

O
Proof We refer for the proof to Korneev [25]. 


O
Note that in comparison with (12.54), the inequality (12.57) is weaker in a 577

sense that they assume z ∈ Wh,k (Ω, div) for k = 1, 2, respectively. But this is 578

PR
a consequence of the way of the proof, based on indirect use of the equilibrated 579

fluxes. D 580

References 581
TE

1. Ainsworth, M., Oden, J.T.: A Posteriori Estimation in Finite Element Analysis. Wiley, New 582
York (2000) 583
EC

2. Ainsworth, M., Vejchodský, T.: Fully computable robust a posteriori error bounds for singularly 584
perturbed reaction-diffusion problems. Numer. Math. 119(2), 219–243 (2011) 585
3. Ainsworth, M., Vejchodský, T.: Robust error bounds for finite element approximation of 586
reaction-diffusion problems with non-constant reaction coefficient in arbitrary space dimen- 587
R

sion. Comput. Methods Appl. Mech. Eng. 281, 184–199 (2014) 588
4. Anufriev, I.E., Korneev, V.G., Kostylev, V.S.: Exactly equilibrated fields, can they be efficiently 589
R

used for a posteriori error estimation? Uchenye zapiski Kazanskogo universiteta, seria 590
Fiziko-matematicheskie nauki (Scientific notes of Kazan State University, series Physical- 591
O

Mathematical Sciences) 148(4), 94–143 (2006) 592

5. Aubin, J.-P.: Approximation of Elliptic Boundary-Value Problems. Wiley-Interscience, New 593
C

York (1972) 594

6. Babuska, I., Strouboulis, T.: Finite Element Method and Its Reliability. Oxford University 595
N

Press, New York (2001) 596

7. Babuska, I., Witeman, J.R., Strouboulis, T.: Finite elements. An introduction to the method and 597
U

error estimation. University Press, Oxford (2011) 598

8. Bank, R.E., Xu, J., Zheng, B.: Superconvergent Derivative Recovery for Lagrange Triangular 599
Elements of Degree p on unstructured grids. SIAM J. Numer. Anal. 45(5), 2032–2046 (2007) 600
9. Bartels, S., Nochetto, R.H., Salgado, A.J.: A total variation diminishing interpolation operator 601
and applications. Math. Comput. 84, 2569–2587 (2015). https://doi.org/10.1090/mcom/2942 602
10. Bramble, J.H., Xu, J.: Some estimates for a weighted L2 projection. Math. Comput. 56(194), 603
463–476 (1991) 604
11. Cai, Z., Zhang, S.: Flux recovery and a posteriori error estimators: conforming elements for 605
scalar elliptic equations. SIAM J. Numer. Anal. 48(2), 578–602 (2010) 606
12. Carey, V., Carey, G.F.: Flexible patch post-processing recovery strategies for solution enhance- 607
ment and adaptive mesh refinement. Int. J. Numer. Methods Eng. 87(1–5), 424–436 (2011) 608
 V. G. Korneev

13. Carstensen, C.E., Merdon, C.: Effective postprocessing for equilibration a posteriori error 609
estimators. Numer. Math. 123(3), 425–459 (2013) 610
14. Cheddadi, I., Fučík, R., Prieto, M.I., Vohralik, M.: Guaranteed and robust a posteriori error 611
estimates for singularly perturbed reaction-diffusion problems. ESAIM: Math. Model. Numer. 612
Anal. 43, 867–888 (2009) 613
15. Churilova, M.A.: Vychislitel’nye svoistva funktsional’nyh aposteriornyh otsenok dlia stat- 614
sionarnoi zadachi reaktsii-diffuzii (Numerical properties of functional a posteriori bounds 615
for stationary reaction-diffusion problem). Vestnik SPbSU, Seria 1: Matematika, Mehanika, 616
Astronomija 1(1), 68–78 (2014, in Russian) 617
16. Ciarlet, P.: The Finite Element Method for Elliptic Problems. North-Holland, Amsterdam 618
(1978) 619
17. Cottrell, J.A., Hughes, J.R., Bazilevs, Y.: Isogeometric Analysis. Toward Integration of CAD 620

F
and FEA. Wiley, Chichester (2009) 621
18. Creusé, E., Nicaise, S.: A posteriori error estimator based on gradient recovery by averaging 622

O
for discontinuous Galerkin methods. J. Comput. Appl. Math. 234, 2903–2915 (2010) 623
19. Ern, A., Stephansen, A., Vohralik, M.: Guaranteed and robust discontinuous Galerkin a 624

O
posteriori error estimates for convection-diffusion-reaction problems. J. Comput. Appl. Math. 625
234(1), 114–130 (2009) 626

PR
20. Karakashian, O.A., Pascal, F.: A posteriori error estimates for a discontinuous Galerkin 627
approximation of second-order problems. SIAM J. Numer. Anal. 41, 2374–2399 (2003) 628
AQ1
21. Korneev, V.: Shemy metoda konechnyh elementov vysokih poriadkov tochnosti (The Finite 629
Element Methods of High Order of Accuracy). Leningrad State University, Leningrad (1977, 630
in Russian). 631
D
22. Korneev, V.G.: Prostye algoritmy vychislenia aposteriornyh otsenok pogreshnosti chislennyh 632
reshenii ellipticheskih uravnenii (Simple algorithms for calculation of a posteriori error 633
TE

estimates for approximate solutions of elliptic equations). Uchenye zapiski Kazanskogo 634
universiteta, seria Fiziko-matematicheskie nauki (Scientific notes of Kazan State University, 635
series Physical-Mathematical Sciences). Kazan State University, Kazan 154(4), 11–27 (2011, 636
in Russian) 637
EC

23. Korneev, V.G.: Robastnye soglasovannye aposteriornye mazhoranty pogreshnosti dlia prib- 638
lizhennyh reshenii uravntnii reaktsii-diffuzii (Robust consistent a posteriori error majorants for 639
approximate solutions of diffusion-reaction equations). In: Materialy 11-oi mezhdunarodnoi 640
konferentsii “Setochnye metody dlia kraievyh zadach i prilozhenia” (Kazan State University, 641
R

Kazan 20–25 Oct. 2016), pp 182–187 (2016, in Russian) 642

AQ2
24. Korneev, V.G.: On the accuracy of a posteriori functional error majorants for approximate 643
R

solutions of elliptic equations. Dokl. Math. 96(1), 380–383 (2017) 644

25. Korneev, V.G.: On the error control at numerical solution of reaction-diffusion equations, 6 645
O

Nov 2017. arXiv:1711.02054v1 646

26. Korneev, V.G., Langer U.: Dirichlet-Dirichlet Domain Decomposition Methods for Elliptic 647
C

Problems, h and hp Finite Element Discretizations. World Scientific, New Jersey (2015) 648
27. Ladyzhenskaya, O.A.: The Boundary Value Problems of Mathematical Physics. Springer, New 649
N

York (1985) 650

28. Langer, U., Moore, S.E., Neumuller, M.: Space-time isogeometric analysis of parabolic 651
U

evolution problems. 7 Sep 2015. arXiv:1509.02008v1 [math.NA] 652

29. Nazarov, F.B., Poborchi, S.V.: Neravenstvo Puankare i ego prilozhenia (Poincare inequality and 653
its applications). Publishing House of St. Petersburg State University, St. Petersburg (2012, in 654
Russian) 655
30. Nochetto, R.H., Otarola, E., Salgado, A.J.: Piecewise polynomial interpolation in Mucken- 656
houpt weighted Sobolev spaces and applications. Numer. Math. 132, 85–130 (2016). https:// 657
doi.org/10.1007/s00211-015-0709-6 658
31. Oganesian, L.A., Ruhovets, L.A.: Variatsionno-raznostnyie metody reshenia ellipticheskih 659
uravnenii (Variational-difference methods for solution of elliptic equations). Publishing House 660
of Armenian Academy of Sciences of Armenian SSR, Yerevan (1979, in Russian) 661
12 On a Renewed Approach to A posteriori Error Bounds for Approximate. . .

32. Repin, S., Frolov, M.: Ob aposteriornyh otsenkah tochnosti priblizhennyh reshenii kraievyh 662
zadach (On a posteriori error bounds for approximate solutions of elliptic boundary value 663
problems). Zhurnal vychislitel’noi matematiki i matematicheskoi fiziki. 42(12), 1774–1787 664
(2002, in Russian) 665
33. Repin, S., Sauter, S.: Functional a posteriori estimates for the reaction-diffusion problem. C. 666
R. Math. Acad. Sci. Paris 343(5), 349–354 (2006) 667
34. Scott, L., Zhang, S.: Finite element interpolation of nonsmooth functions satisfying boundary 668
conditions. Math. Comput. 54, 483–493 (1990) 669
35. Vejchodsk’y, T.: Guaranteed and locally computable a posteriori error estimate. IMA J. Numer. 670
Anal. 26, 525–540 (2006). https://doi.org/10.1093/imanum/dri043 671
36. Xu, J., Zhang, Z.: Analysis of recovery type a posteriori error estimators for mildly structured 672
grids. Math. Comput. 73(247), 1139–1152 (2003) 673

F
37. Xu, J., Zou, J.: Some nonoverlapping domain decomposition methods. SIAM Rev. 40(4), 857– 674
914 (1998) 675

O
38. Zhang, Z.: Ultracovergence of the patch recovery technique. Math. Comput. 65(216), 1431– 676
1437 (1996) 677

O
39. Zienkiewicz, O.C., Zhu, J.Z.: The superconvergence patch recovery (SPR) and adaptive finite 678
element refinement. Comput. Methods Appl. Mech. Eng. 101, 207–224 (1992) 679

PR
D
TE
EC
R
R
O
C
N
U
 AUTHOR QUERIES

AQ1. Ref. [20] is not cited in the text. Please provide the citation or delete it from
the list.
AQ2. Please check the presentation of Refs. [22, 23].

F
O
O
PR
D
TE
EC
R
R
O
C
N
U