A Posteriori Estimates For Partial Differential Equations, Sergey Repin PDF

Radon Series on Computational and Applied Mathematics 4
Managing Editor
Heinz W. Engl (Linz/Vienna)
Editors
Hansjrg Albrecher (Linz)
Ronald H. W. Hoppe (Augsburg/Houston)
Karl Kunisch (Graz)
Ulrich Langer (Linz)
Harald Niederreiter (Singapore)
Christian Schmeiser (Linz/Vienna)
Sergey Repin
A Posteriori Estimates
for Partial Differential Equations
Walter de Gruyter Berlin New York

Author
Prof. Sergey I. Repin
V. A. Steklov Institute of Mathematics at St. Petersburg
Fontanka 27
191023 St. Petersburg
Russia
E-mail: repin@pdmi.ras.ru
Keywords
Partial differential equations, a posteriori error estimates, Poissons equation, diffusion problems,
elasticity, incompressible viscous fluids, nonlinear problems
Mathematics Subject Classification 2000
00-02, 35-02, 35J05, 35K05, 65M15, 65N15, 74Bxx, 76Dxx
Printed on acid-free paper which falls within the guidelines
of the ANSI to ensure permanence and durability.
ISBN 978-3-11-019153-0
Bibliographic information published by the Deutsche Nationalbibliothek
The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.
Copyright 2008 by Walter de Gruyter GmbH & Co. KG, 10785 Berlin, Germany.
All rights reserved, including those of translation into foreign languages. No part of this book
may be reproduced or transmitted in any form or by any means, electronic or mechanical,
including photocopy, recording, or any information storage or retrieval system, without permission
in writing from the publisher.
Printed in Germany
Cover design: Martin Zech, Bremen.
Typeset using the authors L
A
T
E
X files: Kay Dimler, Mncheberg.
Printing and binding: Hubert & Co. GmbH & Co. KG, Gttingen.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 A priori and a posteriori methods of error estimation . . . . . . . . . 1
1.2 Book structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 The error control problem . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Mathematical background and notation . . . . . . . . . . . . . . . . . 8
1.4.1 Vectors and tensors . . . . . . . . . . . . . . . . . . . . . . . 8
1.4.2 Spaces of functions . . . . . . . . . . . . . . . . . . . . . . . 11
1.4.3 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.4 Convex functionals . . . . . . . . . . . . . . . . . . . . . . . 17
2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1 Error indicator by Runge . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 PragerSynge estimate . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Mikhlin estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Ostrowski estimates for contractive mappings . . . . . . . . . . . . . 26
2.5 Error estimates based on monotonicity . . . . . . . . . . . . . . . . . 30
2.6 A posteriori error indicators for nite element approximations . . . . 31
2.6.1 Explicit residual methods . . . . . . . . . . . . . . . . . . . . 32
2.6.2 Implicit residual methods . . . . . . . . . . . . . . . . . . . . 35
2.6.3 A posteriori estimates based on post-processing of approxi-
mate solutions . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.6.4 A posteriori methods using adjoint problems . . . . . . . . . 42
3 Poissons equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1 The variational method . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 The method of integral identities . . . . . . . . . . . . . . . . . . . . 50
3.3 Properties of a posteriori estimates . . . . . . . . . . . . . . . . . . . 52
3.4 Two-sided bounds in combined norms . . . . . . . . . . . . . . . . . 57
3.5 Modications of estimates . . . . . . . . . . . . . . . . . . . . . . . 59
3.5.1 Galerkin approximations . . . . . . . . . . . . . . . . . . . . 59
3.5.2 Advanced forms of error bounds . . . . . . . . . . . . . . . . 60
3.5.3 Decomposition of the domain . . . . . . . . . . . . . . . . . 62
3.5.4 Estimates with partially equilibrated uxes . . . . . . . . . . 64
3.6 How can one use functional a posteriori estimates in practical compu-
tations? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.6.1 Post-processing of uxes . . . . . . . . . . . . . . . . . . . . 65
vi Contents
3.6.2 Runge type estimate . . . . . . . . . . . . . . . . . . . . . . 66
3.6.3 Minimization of the majorant . . . . . . . . . . . . . . . . . 66
3.6.4 Error indicators generated by error majorants . . . . . . . . . 70
4 Linear elliptic problems . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.1 Two-sided estimates for stationary diffusion problem . . . . . . . . . 75
4.1.1 Estimates for problems with mixed boundary conditions . . . 75
4.1.2 Modications of estimates . . . . . . . . . . . . . . . . . . . 78
4.1.3 Estimates for problems with Neumann boundary condition . . 80
4.2 The stationary reaction-diffusion problem . . . . . . . . . . . . . . . 81
4.3 Diffusion problems with convective term . . . . . . . . . . . . . . . . 87
4.3.1 The stationary convection-diffusion problem . . . . . . . . . 88
4.3.2 The reaction-convection-diffusion problem . . . . . . . . . . 92
4.3.3 Special cases and modications . . . . . . . . . . . . . . . . 95
4.3.4 Estimates for uxes . . . . . . . . . . . . . . . . . . . . . . . 99
4.4 Notes for the chapter . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5 Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.1 The linear elasticity problem . . . . . . . . . . . . . . . . . . . . . . 104
5.2 Estimates for displacements . . . . . . . . . . . . . . . . . . . . . . 107
5.3 Estimates for stresses . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.4 Isotropic linear elasticity . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4.1 3D problems . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.4.2 The plane stress problem . . . . . . . . . . . . . . . . . . . . 111
5.4.3 The plane strain problem . . . . . . . . . . . . . . . . . . . . 113
5.4.4 Error of the plane stress model . . . . . . . . . . . . . . . . . 114
6 Incompressible viscous uids . . . . . . . . . . . . . . . . . . . . . . . . 117
6.1 The Stokes problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.2 A posteriori estimates for the stationary Stokes problem . . . . . . . . 123
6.2.1 Estimates for the velocity eld . . . . . . . . . . . . . . . . . 123
6.2.2 Estimates for pressure . . . . . . . . . . . . . . . . . . . . . 127
6.2.3 Estimates for stresses . . . . . . . . . . . . . . . . . . . . . . 128
6.2.4 Estimates in combined norms . . . . . . . . . . . . . . . . . 128
6.2.5 Lower bounds of errors . . . . . . . . . . . . . . . . . . . . . 130
6.2.6 Mixed boundary conditions . . . . . . . . . . . . . . . . . . 131
6.2.7 Problems for almost incompressible uids . . . . . . . . . . . 137
6.2.8 Problems with the condition div u = . . . . . . . . . . . . 139
6.3 Generalized Stokes problem . . . . . . . . . . . . . . . . . . . . . . 140
6.3.1 Estimates for solenoidal approximations . . . . . . . . . . . . 141
6.3.2 Estimates for nonsolenoidal elds . . . . . . . . . . . . . . . 145
6.3.3 Estimates for the pressure eld . . . . . . . . . . . . . . . . . 146
Contents vii
6.3.4 Error minorant . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.5 Models with polymerization . . . . . . . . . . . . . . . . . . 148
6.3.6 Models with rotation . . . . . . . . . . . . . . . . . . . . . . 149
6.4 The Oseen problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.5 Stationary NavierStokes problem for J = 2 . . . . . . . . . . . . . 153
7 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.1 Linear elliptic problem . . . . . . . . . . . . . . . . . . . . . . . . . 158
7.1.1 The variational method . . . . . . . . . . . . . . . . . . . . . 159
7.1.2 The method of integral identities . . . . . . . . . . . . . . . . 164
7.1.3 Error estimates for the dual variable . . . . . . . . . . . . . . 168
7.1.4 Two-sided estimates for combined norms . . . . . . . . . . . 168
7.2 Elliptic problems with lower terms . . . . . . . . . . . . . . . . . . . 171
7.3 Problems with solutions dened in subspaces . . . . . . . . . . . . . 173
7.3.1 Abstract problem . . . . . . . . . . . . . . . . . . . . . . . . 173
7.3.2 Estimate for approximations lying in the subspace . . . . . . 173
7.3.3 Estimate for approximations lying in the energy space . . . . 174
7.4 Derivation of a posteriori estimates from saddle point relations . . . . 176
8 Nonlinear problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
8.1 Variational inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 178
8.1.1 Variational inequalities of the rst kind . . . . . . . . . . . . 179
8.1.2 Variational inequalities of the second kind . . . . . . . . . . . 185
8.2 General elliptic problem. Variational method. . . . . . . . . . . . . . 186
8.3 General elliptic problem. Nonvariational method . . . . . . . . . . . 191
8.4 A posteriori estimates for special classes of nonlinear elliptic problems 196
8.4.1 -Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
8.4.2 Problems with nonlinear boundary conditions . . . . . . . . . 201
8.4.3 Generalized Newtonian uids . . . . . . . . . . . . . . . . . 211
9 Other problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
9.1 Differential equations of higher order . . . . . . . . . . . . . . . . . 218
9.2 Equations with the operator curl . . . . . . . . . . . . . . . . . . . . 224
9.3 Evolutionary problems . . . . . . . . . . . . . . . . . . . . . . . . . 229
9.3.1 The linear evolutionary problem . . . . . . . . . . . . . . . . 229
9.3.2 First form of the error majorant . . . . . . . . . . . . . . . . 231
9.3.3 Second form of the error majorant . . . . . . . . . . . . . . . 235
9.3.4 Equivalence of the deviation and majorant . . . . . . . . . . . 238
9.3.5 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
9.4 A posteriori estimates for optimal control problems . . . . . . . . . . 242
9.4.1 Two-sided bounds for cost functionals . . . . . . . . . . . . . 243
viii Contents
9.4.2 Estimates for state and control functions . . . . . . . . . . . . 248
9.4.3 Estimate in a combined norm . . . . . . . . . . . . . . . . . 251
9.4.4 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . 252
9.4.5 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
9.5 Estimates for nonconforming approximations . . . . . . . . . . . . . 254
9.5.1 Estimates based on projecting to the energy space . . . . . . . 255
9.5.2 Estimates based on the Helmholtz decomposition . . . . . . . 257
9.5.3 Accuracy of approximations obtained by the Trefftz method . 263
9.5.4 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
9.6 Uncertain data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
9.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
9.6.2 Errors caused by indeterminacy in coefcients . . . . . . . . 270
9.6.3 Errors owing to uncertain C . . . . . . . . . . . . . . . . . . 276
9.6.4 Comments . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
9.7 Error estimates in terms of functionals and nonenergy norms . . . . . 279
9.7.1 General framework . . . . . . . . . . . . . . . . . . . . . . . 279
9.7.2 Estimates in local norms . . . . . . . . . . . . . . . . . . . . 280
9.7.3 Estimates in terms of linear functionals . . . . . . . . . . . . 281
9.7.4 Estimates based on the Poincar e inequality . . . . . . . . . . 284
9.7.5 Estimates based on multiplicative inequalities . . . . . . . . . 285
9.7.6 Estimates based on the maximum principle . . . . . . . . . . 285
9.7.7 Estimates in weighted norms . . . . . . . . . . . . . . . . . . 287
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Preface
Pure mathematicians sometimes are satised with show-
ing that the non-existence of a solution implies a logical
contradiction, while engineers might consider a numer-
ical result as the only reasonable goal. Such one sided
views seem to reect human limitations rather than ob-
jective values. In itself mathematics is an indivisible or-
ganism uniting theoretical contemplation and active ap-
plication. R. COURANT [112]
Partial differential equations (PDEs) were introduced as mathematical models of var-
ious physical phenomena. In the 20th century, the theory of differential equations was
mainly developed in the context of an a priori conception that can be expressed by
the triad: existence, regularity, and approximation. In it, the accent is made on a pri-
ori mathematical analysis, and numerical experiment is often regarded as the very last
(and in a sense technical) step, which is more related to practical applications than to
theory.
A certain revision of views has started 2030 years ago. It was stimulated by rapid
development of numerical methods for PDEs. The experience accumulated in this
area shows that a priori methods provide only one part of the information necessary
for a comprehensive analysis of models based on PDEs. If differential equations are
considered not as a self-contained branch of pure mathematics but as tools consigned
to serve natural sciences, then the imperfection of purely a priori analysis is easy to
observe. For example, almost all results of regularity theory and asymptotic analysis
have a qualitative meaning and are addressed to the whole class (or a subclass) of
boundary value problems considered. However, in the numerical experiment we al-
ways deal with an approximate solution of a particular problem the quality of which
must be certied by a certain quantitative criterion. The latter task calls for further
development of different mathematical methods focused on a posteriori analysis of
approximate solutions.
The need for new mathematical approaches to the analysis of PDEs is motivated
not only by routine arguments (such as getting accurate numerical approximations).
There is a more fundamental problem: validation of mathematical models. Certainly,
it can be solved only by joint efforts of mathematicians and specialists in a particular
x Preface
natural science and a posteriori error control methods able to guarantee the reliability
of mathematical experiments must play an essential role in such research.
This book can be viewed as an introduction to a posteriori error estimation theory
for PDEs, which is now in the process of formation and development. It includes
an extended version of the lecture course A posteriori estimates and adaptivity in
continuum mechanics that was prepared for the Special Radon Semester organized
in 2005 by the Radon Institute of Computational and Applied Mathematics in Linz.
That course was based on earlier lectures delivered in 20002001 for students of the
St. Petersburg Polytechnical University (Russia) and in 2003 for students and scientic
researchers of the University of Jyv askyl a (Finland) and the University of Houston
(USA). In 2006, I read a modied version of the course at the Helsinki University of
Technology (Finland) and in 2007 at the University of Valenciennes (France). The
work with lectures was also supported by the DAAD program of Germany and FIM
(Switzerland) during long-term visits to the University of Saarbr ucken (Germany) and
the Swiss Federal Institute of Tecnology (ETH, Zurich).
For these years, the content and structure of the text varied. However, the main
line of it remains the same: for each class of boundary value problems, a posteriori
estimates are derived by purely functional methods, which are used in the theory of
PDEs for analysis of the corresponding differential equations. In other words, the
method suggested for deriving a posteriori error estimates (as well as methods, which
study existence and regularity) exploits specic features of a particular mathematical
problem, but does not attract properties of approximate solutions, mesh, and numerical
method used (the latter information can be utilized later). As a result, we obtain esti-
mates that contain no mesh-dependent constants and are valid for any approximation
from the corresponding energy space. This new functional approach to the a posteriori
error estimation developed in the last decade is the main subject of the book.
A posteriori estimates of the functional type came about from two sources. The
variational statement of the problem (if it has variational form) generates the rst
derivation method, which is well exposed in many papers and in the book [244]. The
second method is based on transformations of integral identities that dene general-
ized solutions. Most results exposed in Chapters 39 are obtained with the help of this
nonvariational method. The main idea of it is briey as follows:
An integral identity (variational inequality) that denes a generalized solution is a
source of guaranteed and computable bounds of the difference between this solution
and any function from the corresponding energy space.
Chapter 2 contains a concise overview of various a posteriori error estimation meth-
ods developed in the 20th century, which are different from those considered in sub-
sequent chapters. Its purpose is to give only a general presentation and to discuss
several approaches to the error control problem. The reader interested in their detailed
investigation is referred to relevant literature.
In Chapter 3, the basic ideas of the functional approach are explained with the
paradigm of a simple elliptic problem. The next chapters are devoted to particular
Preface xi
classes of problems: diffusion, linear elasticity, variational inequalities, etc. Their goal
is not only to present new estimates, but to demonstrate the method used for deriving
them. The latter problem is even more important because the understanding of basic
ideas would allow the reader to derive a posteriori estimates for a concrete problem of
interest.
The book rests upon a moderate background in functional analysis and the theory
of PDEs. I hope that it will be useful for advanced specialists in the mathematics
of computations, as well as for students specialized in applied mathematics, and for
numerical analysts.
I wish to express my deep gratitude to Prof. O. A. Ladyzhenskaya for comments and
advice she gave me in 20002003. During the last decade I has had fruitful discussions
of problems related to the topic with many colleagues in Europe and USA; sincere
thanks to all of them. Some of these discussions have resulted in joint works presented
on the reference list.
I thank Dr. R. Plato and Dr. N. Lebedinskaya for their kind help in editing the book.
My special gratefulness is to the Radon Institute for Computational and Applied
Mathematics (RICAM) in Linz, where the idea of this book appeared and was actively
supported. Also, I am very grateful to the University of Jyv asyl a and the Academy of
Finland for a long-term support of my research.
Sergey Repin Saint Petersburg, 2008
1 Introduction
1.1 A priori and a posteriori methods of error estimation
Partial differential equations (PDEs) were divised as mathematical models of various
physical phenomena. Over the 20th century, mathematical models based on PDEs
have been intensively investigated as part of the new rapidly developing science Math-
ematical Modeling
1
.
Originally, the analysis of models generated by differential equations was performed
in the framework of an a priori method the main steps of which are as follows:
+
proving existence and uniqueness of a solution;
+
studying higher differentiability of a solution and deriving regularity estimates;
+
establishing rate convergence estimates for sequences of approximations.
Proving the fact that the problem considered possesses a (unique) solution is the rst
principal step. The second step is focused on qualitative properties of exact solutions
(the existence of higher derivatives, the continuity and smoothness of a solution, or
the behavior in the vicinity of special points, the type of possible singularities). This
a priori information can be further utilized in approximation methods. The third step
started receiving serious attention in the 5060s when rst systematic results in the er-
ror control theory for PDEs were obtained. Classical convergence theory is perfectly
presented, e.g., in P. Ciarlet [107] and G. Strang and G. Fix [344]. It is focused on
asymptotic estimates of approximation errors. The ultimate aim of this theory is to
prove that the difference between an exact solution u and an approximation u
k
found
in a nite-dimensional subspace of dimension k tends to zero as k o. In a sense,
proving such a convergence can be viewed as a formal justication of the approxima-
tion method used. More exact estimates show the convergence rate, i.e., they establish
the relation
[u u
k
[
V
_ c
_
1
k
_
n
. (1.1.1)
where V is an appropriate space and c and m are some positive real numbers inde-
pendent of k. The value of the constant c depends on the exact solution u and on the
type of approximations used. These estimates give the most general information on
the behavior of approximation errors and are often called a priori error estimates.
Usually, a priori error estimates establish the asymptotic behavior of an upper bound
of the approximation error for the whole set of solutions and their approximations of
a certain type. However, they are generally unable to evaluate efciently the error
1
In recent years, Mathematical Modeling has come under consideration as part of the Computational
Science that constitutes the third pillar of the scientic enterprize, peer alongside Theory and Physical
Experiment [267].
2 Chapter 1 Introduction
related to a particular approximate solution computed on a particular mesh. Moreover,
they possess other features that make their practical exploitation rather difcult. First,
they are valid only for Galerkin approximations. In practice, it is often difcult to
guarantee that an approximate solution computed by a numerical procedure is indeed
the exact solution of a respective nite-dimensional problem. Second, such estimates
require the extra regularity of exact solution. However, in many practically important
cases exact solutions do not have such regularity.
These drawbacks stimulated efforts focused on newerror estimation methods able to
characterize explicitly the accuracy of approximate solutions. In the late 70s and early
80s, it became clear that successful numerical methods for PDEs should be based
on the so-called mesh-adaptive procedures that modify nite-dimensional spaces (and
meshes) using the information comprised in the approximate solution computed at
the previous step
2
. In fact, at that time it was understood that a priori methods in
the theory of PDEs should be combined with a posteriori ones. In the a posteriori
method, getting an approximate solution is not the nal step but a starting point of
a new analysis, which is aimed at (a) estimating the accuracy of the solution and (b)
indicating the distribution of errors over the domain. Obviously, (a) gives a stopping
criterion and (b) provides information needed for a correction of the nite-dimensional
subspace.
Aposteriori error estimates, together with other related topics (such as mesh rene-
ment and adaptivity), have attracted much attention in the last decades. The above
problems create the basis for a new wave of investigations in numerical analysis united
by the common name reliable modeling. Nowadays, Fully Reliable Mathematical
Modeling based on advanced methods of computer simulation forms one of the most
challenging scientic directions the development of which is of great importance for
many applied sciences. Fully reliable modeling consists of
+
efcient computation of a sequence of approximate solutions that converges to the
desired exact solution and
+
reliable verication of the accuracy of the approximation obtained.
The second problem is discussed in the present book.
1.2 Book structure
Throughout Chapter 1, we discuss the error control problem in general terms and
present a short summary of mathematical notions and notation necessary for the un-
derstanding of the material.
2
An interesting discussion on the reliability of different computer simulation methods in nonlinear
mechanical problems is contained in [15]. It reects early observations of the fact that approximations
of strongly nonlinear problems (and not just them) may seriously depend on the structure of the nite-
dimensional space (mesh) used.
Section 1.2 Book structure 3
Chapter 2 contains a concise overview of a posteriori error estimation methods. We
discuss the very rst error indicators suggested by Runge, Prager and Synge, Mikhlin,
and Ostrowski. After that, we briey present another group of a posteriori error esti-
mates that follow from the monotonicity of operators. In the application to PDEs, the
methods of this group are close to the theory of positive solutions and the maximum
principle. Theoretically, they could give a posteriori error estimates of the strongest
(pointwise) type. A short discussion of them presented in the chapter is intended to
give preliminary knowledge (and to generate an interest in further study) rather than
to expose a complete theory.
The subsequent material is devoted to error indication methods created in the 70s
90s for nite element approximations. We consider the classical explicit residual
method, methods based on the post-processing of approximate solutions, and the dual-
weighted residual method, which have gained high popularity in the computational
community in the last decade. The purpose of this overview is to give only a general
idea of various a posteriori error indicators, which are usually used in adaptive numer-
ical methods. The reader interested in a detailed investigation of them is referred to
relevant literature.
In Chapter 3, we start the study of functional a posteriori estimates. To express the
main ideas in the most transparent form throughout the chapter we consider only one
problem: Poissons equation with Dirichlet boundary conditions. For this problem, we
explain both derivation methods (variational and nonvariational) and show that they
lead to exactly the same a posteriori error estimates. Further, we discuss properties,
modications, and practical implementation of the estimates.
Chapter 4 is devoted to diffusion problems. We begin with the linear stationary
problem, which can be viewed as the rst generalization of the problem treated in
Chapter 3. A posteriori estimates are now derived for mixed, Neumann, and Robin
boundary conditions. Further, we nd error estimates for reaction-diffusion and
convection-diffusion problems. In the last part of the chapter, we consider the case
where the stationary diffusion equation involves both reaction and convection terms.
A posteriori estimates for nonstationary diffusion equations are obtained in Chapter 9.
Chapter 5 deals with elliptic systems arising in linear elasticity theory. We obtain
error estimates in terms of displacements and stresses and especially discuss the case
of isotropic elasticity, which is typical of applications.
In Chapter 6, we present the error estimation method for models in the theory of
viscous uids. Among them we examine the stationary Stokes problem and its gener-
alizations, the Oseen problem, and the stationary NavierStokes equation (for a suf-
ciently small velocity). Models related to generalized Newtonian uids are discussed
in Chapter 8.
The experience gained by studying Chapters 36 will help the reader to look at the
theory from general positions. This is what Chapter 7 is intended for. Here, we evolve
the variational and nonvariational methods of a posteriori error estimation in operator
form. Using the material of this chapter it is easy to obtain a posteriori estimates for a
particular class of problems that ts the general conditions imposed on the operators.
In the last section of Chapter 7, we discuss mixed formulations of elliptic problems and
show that the a posteriori estimates can also be derived by transforming the relations
that dene the saddle point of the corresponding Lagrangian.
The book is mainly focused on linear problems. A concise excursion to the theory of
a posteriori error estimation for nonlinear problems is presented in Chapter 8. First, we
deal with elliptic variational inequalities (of the rst and second type) and show that
proper transformations of variational inequalities also lead to computable estimates
of approximation errors. It is of interest that the same estimates have earlier been
derived by the variational method, which we also discuss in the chapter. Further, we
consider generalizations of the variational and nonvariational methods to a wide class
of nonlinear elliptic problems. A posteriori estimates of the functional type for various
nonlinear problems have intensively been studying during the last decade. However, a
consequent exposition of the corresponding results is beyond the scope of this book.
Instead of this, in the last section of the chapter we give an overview of the results and
a deeper discussion of the estimates derived for the -Laplacian, for problems with
nonlinear boundary conditions, and for generalized Newtonian uids.
Chapter 9 is concerned with applications of error estimation theory to some other
problems. It starts with equations of higher order and a class of problems associated
with the operator curl. Certainly, the estimates presented follow from the general
theory of Chapter 7. However, it seems pertinent to give a more transparent discussion
of these (practically important) problems. A method that yields a posteriori functional
estimates for evolutionary problems is given in the third section. Section 4 is devoted
to a posteriori error estimates of the new type derived for optimal control problems
with the help of functional error majorants. In the next section, we examine two error
estimation methods for nonconforming approximations (they also follow from error
estimation theory presented in Chapters 34). Another important topic concerns the
inuence of errors caused by indeterminacy of data, which always exists in real-life
problems. In Section 6, we show that using new error estimation methods opens a
way for measuring the errors that arise owing to data uncertainty. Finally, we discuss
the problem of error estimation in terms of quantities other than the (global) energy
norm. We show that with the help of functional error majorants one can obtain errors in
local and weighted norms, in goal-oriented functionals, and in some other (nonenergy)
norms.
In conclusion, it is worth adding a comment on practical applications of the a pos-
teriori estimates presented in the book. In the last decade, serious efforts were concen-
trated on the study of this subject. The estimates were numerically tested for diffusion
(elliptic and parabolic), stationary Stokes and Oseen problems, linear elasticity and
thermoelasticity, variational inequalities, optimal control problems, and nonconform-
ing approximations of elliptic problems. A posteriori estimates of the functional type
were used together with standard software packages (such as MATLAB and ANSYS)
and compared with a posteriori error estimates of other types. Summarizing the ex-
Section 1.3 The error control problem 5
perience accumulated, we can say that new a posteriori estimates are more expensive
(e.g., if we compare them with error indicators based on gradient averaging). How-
ever, they have two serious advantages: the error bounds are guaranteed and are valid
for conforming approximations of all types. The latter property is especially impor-
tant in practical computations where it is difcult to guarantee that the true solution
is indeed regular, the approximate solution satises the Galerkin orthogonality condi-
tion, and it may be necessary to adapt quickly the error estimation method to approx-
imations of new types. In addition, functional a posteriori estimates do not contain
mesh-dependent constants and can be computed with the help of a unied numerical
technology (and one computer code, which is a checker independent of a solver).
The reader will nd references to publications including numerical results and their
discussion in respective sections of the book. However, a systematic exposition of all
these results would make the book unreasonably large. For this reason, it is concen-
trated on the explanation of key ideas, methods, and algorithms. Studying them would
allow the thoughtful reader to implement easily the estimates in his/her own code and
check how the error estimation technology works.
1.3 The error control problem
If we wish to use differential equations as mathematical models of real-life objects
and processes, then the main question is how to solve them reliably. By the word
solve we mean the ability to get detailed quantitative information on the exact solu-
tion. Regrettably, analytic methods yield solutions to a restricted amount of problems.
Therefore, in the vast majority of cases solving is reduced to a numerical proce-
dure resulting in an approximate solution. Thus, a mathematical experiment should
be regarded as the crucial step that provides information, which is either difcult or
impossible to get by other methods. The general principle of scientic objectivity
suggests that
the mathematical experiment must obey the same strict authenticity rules
as those commonly accepted in natural sciences.
In other words, mathematical modeling cannot be condent without answering the
question:
What is the guaranteed accuracy of an approximate solution?
A priori rate convergence estimates supply a rather philosophical answer to this
question, namely: If the dimension of subspaces increases in a certain proportion
(and if all subspaces are regular in some sense), then the upper bound of the error es-
tablished for the whole class of approximations decreases with a certain rate. Such an
answer cannot be accepted as a sufcient one. An a priori estimate may be unable to
give a realistic estimate for a particular approximate solution. Moreover, there are sev-
eral other questions that convincingly call for a different approach to the error control
problem. For example:
+
Which part of the difference between computed solutions and observable (physical)
data is related to approximations (numerical method) and which is generated by the
mathematical model used?
+
How accurately do we know the coefcients, domain, boundary and initial condi-
tions and how does the indeterminacy in this knowledge affects the solution?
+
How sensitive is the solution with respect to changes in the mathematical model
(e.g., with respect to adding (removing) lower terms to the equation, changing
boundary conditions or using other modications)?
These are difcult questions, and very often they are ignored in engineering and scien-
tic computations. However, if they are not answered, then computer simulation is but
a producer of more or less probable conjectures (which sometimes may lead to wrong
conclusions).
Reliable answers to these and other questions cannot be found unless the following
main error control problem is solved:
Given the data (coefcients,a domain, boundary conditions) of a bound-
ary value problem having the exact solution u and a function from the
corresponding (energy) space V , compute the radii r
1
and r
2
of two balls
B(. r
1
) and B(. r
2
) centered at such that
u , B(. r
1
) and u B(. r
2
). (1.3.1)
We say that a method used to solve the above problem is sharp if one can nd r
1
and
r
2
such that r
2
r
1
_ c for any given c > 0.
Obviously, this problem is solved if we have the estimate
M (. D) _ [u [
V
_ M (. D). V V. (1.3.2)
where [ [
V
stands for the norm of V , D denotes the set of known data and the func-
tionals M (error majorant) and M (error minorant) are directly computable. This es-
timate establishes two-sided guaranteed error bounds for conforming approximations
of all types.
Throughout the book, the functionals that give upper and lower bounds of approxi-
mation errors are denoted by M and M, respectively. Subscripts are used to specify
such a functional and relate it to a particular problem: for example, M
cT
is the majo-
rant for the Stokes problem.
Estimate (1.3.2) has a practical signicance if the functionals M and M satisfy an
additional consistency condition:
M (. D) 0. M (. D) 0 as uin V. (1.3.3)
which guarantees that the majorant (minorant) vanishes on any sequence converging
(in the energy space) to the exact solution. The majorants discussed in the book satisfy
this condition.
Section 1.3 The error control problem 7
Another important requirement imposed on M and M is that (1.3.2) must hold
without extra regularity assumptions on the exact solution u (which in a priori error
estimates are usually required). From the practical point of view, such a property of
an error estimate is very useful, because the regularity (smoothness) of exact solutions
is unstable with respect to small variations in data (coefcients of a PDE and the ge-
ometry of a domain). In real life problems, these data are never known exactly, so that
numerical analysts and engineers should use approximation and error control methods
that do not exploit (explicitly or implicitly) higher regularity of exact solutions and are
stable with respect to small variations in problem data.
In Chapters 39 we derive the functionals M and M for various classes of boun-
dary-value problems. In addition to , they also include the known data D and free
function(s) ,, which can be viewed as approximations of certain differential opera-
tor(s) applied to u. For example, in diffusion problems they are associated with Vu
or Vu (where is the diffusion matrix) and in linear elasticity with Lc(u) (where
L is the elasticity tensor and c(u) is the tensor of small strains). These differential
complexes have a physical meaning and are often called dual variables. We consider
them as arguments and rewrite (1.3.2) in the form
M (. ,. D) _ [u [
V
_ M (. ,. D). V V. , Y. (1.3.4)
where Y is a certain set (usually it is a space) that contains admissible ,.
Also we derive computable functionals that provide reliable bounds for the er-
rors estimated in terms of combined (primal-dual) norms (e.g., in terms of the norm
[(u )[
V
[( q)[
Y
, where denotes the exact dual function and q is an ap-
proximation of ).
Once the estimate (1.3.4) has been derived for a class of boundary value problems,
the computer simulation methods for this class are fully controllable. Therefore, such
estimates should be derived for each mathematical model used for the quantitative
analysis. The aim of this book is not only to present M and M for particular classes
of boundary value problems but to explain the basic principles that will help the reader
to derive similar estimates for a particular mathematical model he/she studies.
Finally, we note that (1.3.4) generates the variational statement
inf
:1,
,Y
M(. ,. D).
In principle, a sequence of approximations converging to the exact solution can be
computed by a certain minimization procedure applied to M(. ,. D). The values of
M indicate the quality of and , as approximations of the exact solutions u and ,
so that the efciency of an approximation process is explicitly controlled. However,
in practice such a minimization procedure may be rather expensive, and it is more
efcient to nd and , by other (e.g., mixed) methods.
1.4 Mathematical background and notation
It is assumed that the reader is familiar with basic facts in functional analysis and the
theory of differential equations and understands such notions as generalized deriva-
tive, weak solution, Lebesgue and Sobolev space. Those feel necessary to study the
mathematical background are advised to use, e.g., [151, 197, 234, 211]. The content
of these books is quite sufcient for the understanding of the material.
1.4.1 Vectors and tensors
By R
d
and M
dd
we denote the spaces of real J-dimensional vectors and J J
matrixes (tensors), respectively. The scalar product of vectors is denoted by , and for
the product of tensors we use the symbol :, i.e.,
u = u
i
i
. t : o = t
i}
o
i}
.
where summation (from 1 to J) over repeated indices is implied. The norms of vectors
and tensors are dened with the help of the products introduced above:
[a[ :=
_
a a. [o[ :=
_
o : o.
Henceforth, the symbol := means equals by denition. The multiplication of a
matrix M
dd
and a vector b R
d
is the vector, which we denote b. In the
book, matrixes are usually denoted by capital Latin letters or by Greek letters (e.g., o,
t, c).
T
and
-1
denote the transposed and inverse matrixes, respectively.
Any tensor t is decomposed into the deviatoric part t
D
and the trace tr t := t
i i
, so
that t := t
D
1
d
I tr t, where I is the unit tensor. It is easy to check that
t : I = tr t. t
D
: I = 0. (1.4.1)
[t[
2
= [t
D
[
2
1
J
(tr t)
2
. (1.4.2)
so that we have an orthogonal decomposition of t into two parts (which sometimes are
called deviatorical and spherical).
In the book, we use various inequalities for scalars, vectors, and tensors. First, we
recall the algebraic Youngs inequality
2ab _ a
2
b
2
. (1.4.3)
which is valid for any > 0. For a pair of vectors a and b we have a similar estimate
2a b _ [a[
2
[b[
2
. (1.4.4)
Section 1.4 Mathematical background and notation 9
which implies the inequalities
[a b[
2
_ (1 )[a[
2
[b[
2
. (1.4.5)
[a b[
2
_
1
1
[a[
2
[b[
2
. (1.4.6)
Similarly, for a pair of tensors o and t we have
2t : o _ [t[
2
[o[
2
. (1.4.7)
[t o[
2
_ (1 )[t[
2
[o[
2
. (1.4.8)
If H is a Hilbert space with scalar product (. ) and norm [ [ associated with the
product, then it is easy to extend (1.4.4)(1.4.6) to the elements of H.
The inequality (1.4.1) is a particular form of the more general Youngs inequality
ab _
1
(a)
]
t
_
b
_
]
/
.
1
t
= 1. (1.4.9)
Differential operations for vectors and tensors are introduced with the help of a sym-
bolic vector V = i
x
J
Jx
:
, where i
x
(s = 1. . . . . J) denote the init vectors of the Carte-
sian system. We recall that i
x
i
k
=
xk
, where
xk
= 0 if s ,= k and
kk
= 1.
If [ is a differentiable scalar-valued function then V[ is the vector
_
J)
Jx
1
.
J)
Jx
2
.
J)
Jx
3
_
,
which is the gradient of [. For a vector-valued function a = a
x
i
x
we have the follow-
ing differential operations:
V a = i
x
d
d.
x
a
k
i
k
=
da
x
d.
x
= a
x,x
:= div a.
V a = curl a := (a
3,2
a
2,3
: a
1,3
a
3,1
: a
2,1
a
1,2
) (for J = 3).
Va = V a =
da
k
d.
x
i
x
i
k
= a
k,x
i
x
i
k
.
Let [ and a be a differentiable function and vector-valued function, respectively. It is
easy to see that
V ([a) =
d
d.
x
i
x
([a
k
i
k
) =
d
d.
x
(
k
x
[a
k
) =
d[
d.
x
a
x

da
x
d.
x
[
and we arrive at the relation
div [a = a V[ [div a. (1.4.10)
Other differential relations, namely,
V([) = V[ [V.
V([ a) = V[ a [V a.
V ([ a) = [(curl a) V[ a.
V (a b) = b curl a a curl b.
V (a b) = b div a a Vb
are also often used in the analysis of boundary value problems.
Differential operations of the second order are not always dened. Some others lead
to a trivial result. We recall that
V (V[) = 0. V (V a) = 0.
div V[ = V V[ = (V V)[ = ^[.
and
V (V a) = V(V a) (V V)a = grad div a V
2
a.
A vector-valued function a is called solenoidal if it can be represented as curl b, where
b of another vector-valued function. In this case, div a = 0, so that solenoidal elds
are divergence-free.
Similar relations hold for tensors. For example, if t is a tensor-valued function with
differentiable components, then
Div (ta) = a Div t t
T
: Va. (1.4.11)
By Div t we denote the divergence of a tensor-valued function t, which is a vector-
valued function with components {t
i},}
].
In subsequent chapters, we use several known integral relations. The rst of them
is the Ostrogradski formula:
_
D
div a J. =
_
I
a nJs. (1.4.12)
where I is a closed surface of C.
By (1.4.10) and (1.4.12), we obtain
_
D
div ([a) J. =
_
D
([div a V[ a) J. =
_
I
[a nJs (1.4.13)
If [ vanishes on I, then from (1.4.13) it follows that
_
D
([div a V[ a) J. = 0. (1.4.14)
Another integral relation is
_
D
curl a J. =
_
D
a curl J.
_
I
(a n) Js. (1.4.15)
In the above relations, we assume that the functions are sufciently regular so that the
corresponding volume and surface integrals exist.
Finally, we recall the Helmholtzs theorem, which is known as the fundamental
theorem of vector calculus. It states that any sufciently smooth, rapidly decaying
vector-valued function can be resolved into curl-free and divergence-free (solenoidal)
components, i.e.,
q = V[ curl a. (1.4.16)
For a bounded domain C, Helmholtz decomposition states that q can be resolved into
a solenoidal vector-valued function q
0
and V[, where [

H
1
(C) (e.g., see [232]).
Another version (which is often used in mathematical hydrodynamics) resolves q into
q
0
and V[ such that div q
0
= 0, q
0
n = 0 on I, and [ is a scalar-valued function
having (generalized) derivatives of the rst order (e.g., see [210] or [348]).
1.4.2 Spaces of functions
We denote a bounded connected domain in R
d
by C and its boundary (which is as-
sumed to be Lipschitz continuous) by I. Usually, o stands for an open subset of C.
The closure of sets is denoted by a bar and the Lebesgue measure of a set o R
k
by meas
k
o. If o is a domain in R
d
and ; R
d-1
is its boundary, then we also use
simplied notation [o[ and [;[ for the corresponding J and J 1 measures.
By 1
]
(o) we denote the space of functions summable with power with norm
[n[
],o
:=
__
o
[n[
]
J.
_
1{]
.
The vector-valued functions with components that are square summable in C form
the Hilbert space 1
2
(C. R
d
). Analogously, 1
2
(C. M
dd
) is the Hilbert space of
tensor-valued functions (sometimes we use the special notation for this space). If
tensor-valued functions are assumed to be symmetric, then we write M
dd
x
(and
x
instead of 1
2
(C. M
dd
x
)).
For 1
2
(C. R
d
) and t 1
2
(C. M
dd
), the norms are dened by the relations
[[
2
:=
_
D
[[
2
J. and [o[
2
:=
_
D
[o[
2
J..
Since no confusion may arise, we denote the norm of 1
2
(C) and the norm of the space
1
2
(C. R
d
) by [ [. The space of measurable essentially bounded functions is denoted
by 1
o
(C). It is equipped with the norm
[u[
o,D
= ess sup
xD
[u(.)[.
By

C
o
(C) we denote the space of all innitely differentiable functions with compact
supports in C. The spaces of k-times differentiable scalar- and vector-valued func-
tions are denoted by C
k
(C) and C
k
(C. R
d
), respectively;

C
k
(C) is the subspace of
C
k
(C) that contains functions vanishing at the boundary; 1
k
(C) denotes the set of
polynomial functions dened in C R
d
, i.e., 1
k
(C) if
=
[[_n
a
. m _ k.
where := (
1
. . . . .
d
) is the so-called multi-index,
[[ =
1
2
. . .
d
. a
= a
1
,...,
o
.
and .
= .
1
.
2
. . . .
o
.
Solenoidal vector-valued functions the components of which are square summable
in C form the space S(C).
For partial derivatives we keep the standard notation and write
d
d.
i
or
,i
.
Usually, we understood them in a generalized sense: a function g =
i
is called the
generalized derivative of 1
1
(C) with respect to the .
i
if it satises the relation
_
D
n
,i
J. =
_
D
gnJ.. Vn C
1
0
(C). (1.4.17)
Generalized derivatives of higher orders are dened by similar integral relations (see
S. Sobolev [336]).
By fgg
S
we denote the mean value of a function g on S, i.e.,
fgg
S
:=
1
[S[
_
S
g J.
and g
S
:= g fgg
S
. The functions with zero mean form the space
1
2
(C) :=
_
q Q
fqg
D
= 0
_
.
The space H(C. div) is a subspace of 1
2
(C. R
d
) that contains vector-valued func-
tions with square-summable divergence, and H(C. Div) is a subspace of that con-
tains tensor-valued functions with square-summable divergence, i.e.,
H(C. div) := { 1
2
(C. R
d
) [ div 1
2
(C)].
H(C. Div) := {t 1
2
(C. M
dd
) [ Div t 1
2
(C. R
d
)].
Both spaces H(C. div) and H(C. Div) are Hilbert spaces endowed with scalar prod-
ucts
(u. )
div
:=
_
D
(u div udiv ) J.
and
(o. t)
Div
:=
_
D
(o : t Div o Div t) J..
respectively. The norms [ [
div
and [ [
Div
are associated with the above-dened
scalar products.
Similarly, H(C. curl ) is the Hilbert space of vector-valued functions having square-
summable curl or, i.e.,
H(C. curl ) := { 1
2
(C. R
d
) [ curl 1
2
(C)].
This space can be dened as the closure of smooth functions with respect to the norm
[n[
curl
:=
_
[n[
2
D
[curl n[
2
D
_
1{2
.
The Sobolev spaces W
n,]
(C) (where m and are positive integer numbers) con-
tain functions summable with power the generalized derivatives of which up to order
m belong to 1
]
. For a function W
n,]
(C), the norm is dened as usual:
[ [
n,],D
=
__
D
[[_n
[D
[
]
J.
_
1{]
.
Here = {
1
. . . . .
d
] is the multi index and
D
=
d
[[
d.
1
1
. . . d.
o
d
is the derivative of order [[. The Sobolev spaces with = 2 are denoted by the letter
H, i.e.,
H
n
(C) :=
1
2
(C) [ D
1
2
(C). Vm : [[ _ m
_
.
These spaces belong to the class of Hilbert spaces. A subset of H
n
(C) formed by the
functions vanishing on I is denoted by

H
n
(C).
The functions in Sobolev spaces have counterparts on I (and on other manifolds of
lower dimensions) that are associated with spaces of traces. Thus, there exist some
bounded operators mapping the functions dened in C to functions dened on the
boundary. For example, the operator ; : H
1
(C) 1
2
(I) is called the trace operator
if it satises the following conditions:
; = [
I
. V C
1
(C). (1.4.18)
[;[
2,I
_ c
TT
[[
1,2,D
. (1.4.19)
where c
TT
is a positive constant independent of . From these relations, we observe
that ; is a natural generalization of the trace dened for a continuous function (in the
pointwise sense). The image of ; is a subset of 1
2
(I), which is the space H
1{2
(I).
Thus, ; L
_
H
1
(C). H
1{2
(I)
_
and the space

H
1
(C) is the kernel of ;. The func-
tions from other Sobolev spaces are also known to have traces in Sobolev spaces with
fractional indices.
Also, for any H
1{2
(I), one can dene a continuation operator
L(H
1{2
(I). H
1
(C))
such that
= n. n H
1
(C). ;n = on I
and (e.g., see J.-L. Lions and E. Magenes [222])
[[
1
1/2
,I
_ c
;
[n[
1,2,D
. [n[
1,2,D
_ c
[[
1
1/2
,I
. (1.4.20)
Using the operator ;, we dene subspaces of functions vanishing on I or on some part
I
1
of I. Usually, such subspaces are marked by the zero subindex, e.g.,
V
0
:= { V [ ; = 0 a.e. on I
1
] .
Henceforth, we understand the boundary values of functions in the sense of traces, so
that the phrase u = on I means that the trace ;u of a function u dened in C
coincides with a given function dened on I (for the sake of simplicity, we usually
omit ;). If for two functions u and dened in C we say that u = on I, then we
mean that ;(u ) = 0 on I.
For 1
2
(C), the functional
(
,i
. ) :=
_
D
d
d.
i
J. (1.4.21)
is linear and continuous not only for functions in

C
o
(C) but also for all functions of
the space

H
1
(C) (this fact follows from the density of smooth functions in

H
1
(C)
and known theorems on the continuation of linear functionals). Such functionals can
be viewed as generalized derivatives of square summable functions. They form the
space H
-1
(C) dual to

H
1
(C). It is easy to see that the quantity
[[[ [[[
,i
[[[ [[[ := sup

!
1
(:)
,=0
[(
,i
. )[
[V[
D
(1.4.22)
is nonnegative and nite. It can be used to introduce the norm for H
-1
(C).
1.4.3 Inequalities
In the subsequent chapters, we use several inequalities well known in functional anal-
ysis (e.g., see [336, 214]). For convenience of the reader, we collect and discuss them
below.
First, we recall the inequality
[a b[ _
_
d
i=1
[a
i
[
_
1{
_
d
i=1
[b
i
[
/
_
1{
/
. (1.4.23)
where
1
/

1
= 1 and a. b R
d
. It is known as the discrete H older inequality. The
H older inequality in functional form is as follows:
_
D
u J. _ [u[
,D
[[
/
,D
. (1.4.24)
Let u and be two functions in 1
(C). Then
_
D
(u )
J. =
_
D
u(u )
-1
J.
_
D
(u )
-1
J.
_ [u[
,D
__
D
(u )
(-1)
/
J.
_
1{
/
[[
,D
__
D
(u )
(-1)
/
J.
_
1{
/
= ([u[
,D
[[
,D
)
__
D
(u )
(-1)
/
J.
_
(-1){
and we arrive at the Minkowski inequality
__
D
(u )
J.
_
1{
_ [u[
,D
[[
,D
. (1.4.25)
that states the triangle inequality for the norm. The H older and Minkowski inequalities
also hold for spaces of vector- and tensor-valued functions.
For the functions in

H
1
(C), we have the Friedrichs inequality
[n[
D
_ C
T D
[Vn[
D
. Vn

H
1
(C). (1.4.26)
where C
T D
is a positive constant independent of n. It is not difcult to observe that
the constant in (1.4.26) satises the relation
1
C
T D
= z
D
:= inf
u

!
1
(:)
u,=0
[Vn[
[n[
. (1.4.27)
Let C

C. For any n

H
1
(C), we can dene n = n in C and n(.) = 0 for any
.

C\ C. Obviously, n

H
1
(
C). Therefore,
z
D
_ inf
u

!
1
(:)
u,=0
[V n[
[ n[
= z
C
:
=
1
C
TD
and C
T D
_ C
TD
. Assume that
C := {. R
d
[ a
i
< . < b
i
. b
i
a
i
= l
i
].
Then,
z
D
_ z
=
_
i
1
l
2
i
.
and we obtain an explicit upper bound for C
T D
.
For n H
1
(C), the Friedrichs inequality has a more general form
[n[
2
D
_ c
2
T D
_
[Vn[
2
D
_
I
[n[
2
Js
_
. (1.4.28)
For n H
1
(C), the Poincar e inequality reads
[n[
2
D
_ C
2
1D
_
[Vn[
2
D
_
_
D
nJ.
_
2
_
. (1.4.29)
From (1.4.29) it follows that
[n[
D
_ C
1D
[Vn[
D
. Vn

1
2
(C). (1.4.30)
If
C =
I
:= {. R
d
[ .
i
(0. l
i
). l
i
> 0].
then the Poincar e inequality takes the form (e.g., see [217])
[n[
2
D
_
1
[
I
[
__
l
nJ.
_
2
J
2
_
l
d
i=1
l
2
i
n
2
,i
J.. (1.4.31)
In S. Mikhlin [234, 233], the reader will nd more information concerning the con-
stants in the Poincar e and Friedrichs inequalities.
In continuum mechanics, of importance is the following assertion known as the
Korns inequality. Let C be an open, bounded domain with Lipschitz continuous
boundary. Then
_
D
_
[n[
2
[c(n)[
2
_
J. _ C
1D
[n[
2
1,2,D
. Vn H
1
(C. R
d
). (1.4.32)
where C
1D
is a positive constant independent of n and c(n) denotes the symmetric
part of the tensor Vn, i.e.,
c
i}
(n) =
1
2
_
dn
i
d.
}
dn
}
d.
i
_
.
It is not difcult to verify that the left-hand side of (1.4.32) is bounded from above by
the H
1
-norm of n. Thus, it represents a norm equivalent to [ [
1,2,D
. The kernel of
c(n) is called the space of rigid deections and is denoted by R(C). If n R(C), then
it can be represented in the form n = n
0
o
0
., where n
0
is a vector independent of
. and o
0
is a skew-symmetric tensor with coefcients independent of .. It is easy to
understand that the dimension of R(C) is nite and equals J
d(d-1)
2
.
For the functions in

H
1
(C), the Korns inequality is easy to prove. Indeed,
[c(n)[
2
=
1
4
(n
i,}
n
},i
)(n
i,}
n
},i
)
=
1
4
(n
i,}
n
i,}
n
},i
n
},i
2n
i,}
n
},i
) =
1
2
([Vn[
2
n
i,}
n
},i
).
where the summation over repeated indices is implied. Therefore, for any n

C
2
(C)
we have
_
D
[c(n)[
2
J. =
1
2
_
D
_
[Vn[
2
n
i,}
n
},i
_
J. =
1
2
_
D
_
[Vn[
2
n
i
n
},i}
_
J.
=
1
2
_
D
_
[Vn[
2
n
i,i
n
},}
_
J. =
1
2
_
D
_
[Vn[
2
[n
i,i
[
2
_
J.
_
1
2
[Vn[
2
.
Hence,
[Vn[ _
_
2[c(n)[ Vn

C
2
(C). (1.4.33)
Since

C
2
(C) is dense in

H
1
(C), this inequality is also valid for functions in

H
1
(C).
The proofs of the Korns inequality (1.4.32) are much more complicated (e.g., see
[120]).
1.4.4 Convex functionals
Consider a Banach space V . A set 1 V is called convex if z
1
1
z
2
2
1 for
all
1
.
2
1 and all z
1
. z
2
R
such that z
1
z
2
= 1. Let 1 be a convex set. A
functional J : 1 R is said to be convex if
J(z
1
1
z
2
2
) _ z
1
J(
1
) z
2
J(
2
) (1.4.34)
for all
1
.
2
1 and all z
1
. z
2
R
such that z
1
z
2
= 1. A functional J is
called strictly convex if
J(z
1
1
z
2
2
) < z
1
J(
1
) z
2
J(
2
) (1.4.35)
for all
1
.
2
1 (such that
1
,=
2
) and z (0. 1). A functional J is called concave
(resp., strictly concave) if the functional (J) is convex (resp., strictly convex).
The functional
1
() =
_
0 if 1.
o if , 1
is called the characteristic functional of the set 1. It is clear that it is convex if and
only if the set 1 is convex.
If J
1
and J
2
are two convex functionals dened on a convex set 1 then the func-
tionals
1
J
1

2
J
2
(for
1
.
2
R
) and max{J
1
. J
2
] are also convex. It is worth
noting that the latter fact remains valid for any amount of convex functionals, i.e., the
upper bound taken over any set of convex functionals is a convex functional. There-
fore, convex functionals are often represented as upper bounds of afne functionals.
By denition, the space V
+
consists of all linear continuous functionals on V . It is
called topologically dual to V . The value of
+
V
+
on V is denoted by (
+
. ).
This product generates a duality pairing of the spaces V and V
+
. If V is a Banach
space, then V
+
can also be normed by setting
_
_
+
_
_
+
:= sup
V
(
+
. )
[[
. (1.4.36)
Henceforth, we assume that the supremum (or inmum) of a quotient is taken with
respect to all elements of V , except for the zero element 0
V
.
Any afne functional dened on elements of V has the form (
+
. ) , where
+
V
+
and R. A functional space is called reexive if it coincides with the
bidual space V
++
(i.e., if there exists a one-to-one mapping of V to V
++
and back that
preserves the metric). All Hilbert spaces are reexive. The same is true for the spaces
1
]
with 1 < < o.
The theorem of F. Riesz asserts that for Hilbert spaces, any functional
+
V
+
can
be written in the form of a scalar product introduced in such a space, i.e.,
(u. ) = (
+
. ). V V. (1.4.37)
where u is uniquely determined.
The functional J
+
: V
+
R dened by the relation
J
+
(
+
) = sup
V
{(
+
. ) J()] (1.4.38)
is said to be dual (or conjugate) to J.
Remark 1.1. If J is a smooth function that increases at innity faster than any linear
function, then J
+
is the Legendre transformof J. The dual functionals were studied by
Young, Fenchel, Moreau, and Rockafellar (e.g., see [121, 132, 324]). The functional
J
+
is also called polar to J.
The functional
J
++
() = sup
{(
+
. ) J
+
(
+
)] (1.4.39)
is called the second conjugate to J (or bipolar).
If J is a convex functional attaining nite values, then J coincides with J
++
.
To illustrate the denitions of conjugate functionals, consider functionals dened
on the Euclidean space 1
d
. In this case, V and V
+
consist of the same elements:
J-dimensional vectors (denoted by and
+
, respectively) and the quantity (
+
. ) is
given by the scalar product
+
.
Let = {a
i}
] be a positive denite matrix. We have the following pair of mutually
conjugate functionals:
J() =
1
2
and J
+
(
+
) =
1
2
-1
+

+
. (1.4.40)
Another example is given by the functionals
J() =
1
[[
and J
+
(
+
) =
1
t
[
+
[
/
. (1.4.41)
where
1
/
= 1. If is an odd convex function, then (([u[
V
))
+
=
+
([u
+
[
V
).
Let a functional J : V Rtakes a nite value at
0
V . The functional J is called
subdifferentiable at
0
if there exists an afne minorant l such that J(
0
) = l(
0
). A
minorant with this property is called the exact minorant at
0
. Obviously, any afne
minorant exact at
0
has the form
l() = (
+
.
0
) J(
0
). l() _ J(). V V. (1.4.42)
The element
+
is called a subgradient of J at
0
. The set of all subgradients of J at
0
forms a subdifferential, which is usually denoted by dJ(
0
). It may be empty, may
contain one element or innitely many elements.
An important property of convex functionals follows directly from the fact that they
have an exact afne minorant at any point (at which the functional attains a nite
value). Assume that J is a convex functional and
+
dJ(
0
). Then there exists an
afne minorant such that
(
+
. ) _ J(). V V.
and (
+
.
0
) = J(
0
). Hence, we obtain
J() J(
0
) _ (
+
.
0
). (1.4.43)
The inequality (1.4.43) represents the basic incremental relation for convex function-
als. For proper convex functionals, there exists a simple criterion that enables one
verify whether or not an element
+
belongs to the set dJ().
Proposition 1.2. The following two statements are equivalent:
J() J
+
(
+
) (
+
. ) = 0. (1.4.44)
+
dJ(). (1.4.45)
dJ
+
(
+
). (1.4.46)
Proof. Assume that
+
dJ(). In accordance with (1.4.43), we have
J(n) _ J() (
+
. n ). Vn V.
Hence,
(
+
. ) J() _ (
+
. n) J(n). Vn V
and, consequently,
(
+
. ) J() _ sup
uV
{(
+
. n) J(n)] = J
+
(
+
). (1.4.47)
However, by the denition of J
+
, we know that for any and
+
J
+
(
+
) _ (
+
. ) J(). (1.4.48)
We observe that (1.4.47) and (1.4.48) imply (1.4.44).
Assume that dJ
+
(
+
). Then J
+
(n
+
) _ J
+
(
+
) (n
+

+
. ). so that
(
+
. ) J
+
(
+
) _ sup
u
{(n
+
. ) J
+
(n
+
)] = J
++
().
On the other hand,
(
+
. ) J
+
(
+
) _ J
++
() = J().
and we again arrive at (1.4.44).
Assume that (1.4.44) holds. By the denition of J
+
, we obtain
0 = J() J
+
(
+
) (
+
. ) _ J() J(n) (
+
. n).
where n is an arbitrary element of V. Thus,
J(n) J() _ (
+
. n ). Vn V.
which means that J() (
+
. n) is an exact afne minorant of J (at ) and,
consequently, (1.4.45) holds.
The proof of (1.4.46) is quite similar.
Let J and J
+
be a pair of conjugate functionals. Then
D
J
(.
+
) := J() J
+
(
+
) (
+
. )
is called the compound functional. From Proposition 1.2 it follows that D
J
is non-
negative and vanishes only if the arguments satisfy (1.4.45) and (1.4.46), which are
also called the duality relations and very often represent the constitutive relations of a
physical model. Compound functionals play an important role in the a posteriori error
estimation of nonlinear problems (see Chapter 8). They serve as penalty functionals
that penalize errors caused by dissatisfaction of the duality relations. For this reason,
we denote the compound functionals by the letter D.
Note that the relation D
J
(.
+
) _ 0 generates inequalities that can be viewed as
generalizations of the Youngs inequality (cf. (1.4.3)(1.4.9)):
(
+
. ) _ J() J
+
(
+
). (1.4.49)
In particular, if V and V
+
coincide with R
d
and J() =
[[
, then J
+
(
+
) =
[
/
and (1.4.36) implies the estimate
+
_
[[
[
+
[
t
. V.
+
R
d
. (1.4.50)
Finally, we recall some basic notions related to the differentiation of convex func-
tionals. We say that J has a weak derivative J
t
(
0
) V
+
(at the point
0
) in the sense
of G ateaux if
lim
2-0
J(
0
zn) J(
0
)
z
= (J
t
(
0
). n). Vn V. (1.4.51)
Assume that J is differentiable in the above sense and
+
dJ(
0
). Then for any
V we know that J() J(
0
) _ (
+
.
0
). Set =
0
zn, where z > 0.
Now, we have J(
0
zn) J(
0
) _ z(
+
. n). Therefore,
(J
t
(
0
). n) = lim
2-0
J(
0
zn) J(
0
)
z
_ (
+
. n).
and (J
t
(
0
)
+
. n) _ 0 for any n V . This inequality means that, in such a case,
the G ateaux derivative coincides with
+
.
2 Overview
2.1 Error indicator by Runge
First attempts to formulate accuracy criteria for approximate solutions date back to
the very beginning of the 20th century when C. Runge suggested a heuristic error
indication rule. Originally, it was applied to quadrature formulas, but later the rule
was also adapted to integration procedures that he developed for ordinary differential
equations. In essence, Runges rule is based on comparing coarse and rened
solutions.
Assume that an approximate solution u
h
has been found on a mesh T
h
with mesh
size h and u
h
ref
is a solution on a rened mesh T
h
ref
. Let an asymptotic error estimate
u
h
u = h
n
O(h
n]
). m. > 0. (2.1.1)
be known, where is a certain (unknown) function. We have
u
ref
u = h
n
ref
O(h
n]
ref
).
and, consequently,
u
h
u
h
ref
(h
n
h
n
ref
) O(h
n]
) O(h
n]
ref
).
From here, we conclude that, up to higher order terms,
u
h
u h
n
u
h
u
h
ref
h
n
h
n
ref
. (2.1.2)
If h
ref
= kh with k (0. 1), then the above relation reads
u
h
u
1
1 k
n
(u
h
u
h
ref
). (2.1.3)
Certainly, the estimate adequately presents the error only if h is small enough (so
that the higher order terms are negligibly small) and if the asymptotic relation (2.1.1)
holds. In general, these (rather strong) a priori requirements cannot be guaranteed,
which means that (2.1.3) should be viewed as a conditional estimate.
Anyway, (2.1.3) partially justies the following heuristic rule often used in engi-
neering computations:
If u
h
u
h
ref
is small, then both approximations u
h
ref
and u
h
are probably
close to the exact solution u.
Section 2.2 PragerSynge estimate 23
Indeed, (2.1.3) shows that for sufciently small h, the error u
h
u is proportional to
u
h
u
h
ref
. Using the modern terminology, we can say that
E
T
(.) := u
h
(.) u
h
ref
(.)
is suggested as an error indicator and a certain measure of E
T
as a stopping criterion.
For this purpose, the quantities
max
xD
[E
T
(.)[ or [E
T
(.)[
],D
. _ 1
are often used. E
T
implies a simple error verication method. In many cases, it gives
quite acceptable (and even good) results. For these reasons, it would be logical to
compare any new error indicator suggested for a particular class of problems with E
T
.
However, E
T
is not fully reliable, because the fact that two subsequent elements of
an approximation sequence are close to each other cannot guarantee that they are close
to the exact solution. For example, E
T
may be small if a renement is improperly
performed (i.e., if new degrees of freedom appended to T
h
ref
do not correlate with the
error). Also, it is often important to have computable and guaranteed upper and lower
bounds for various norms (seminorms) of u
h
u, which cannot be derived by the
above heuristic arguments.
In subsequent chapters, we will present quantities computed by u
h
and u
h
ref
that do
provide guaranteed bounds of approximation errors (the simplest form is exposed in
the Section 3.6.2). However, their derivation is based on mathematical tools that did
not exist at the time of C. Runge.
2.2 PragerSynge estimate
In 1947, W. Prager and J. L. Synge [266] presented an a posteriori estimate valid for
approximations of linear elliptic problems. Originally, the proof was motivated by
an orthogonal decomposition of the energy space of a problem and purely geometric
arguments (this approach is often called the hypercircle method [347]).
Let us discuss the idea with the paradigm of the problem
û = 0. in C. (2.2.1)
u = 0. on I. (2.2.2)
By the Helmholtz decomposition, we know that q 1
2
(C. R
d
) is uniquely repre-
sented as q
0
V[, where [

H
1
(C), q
0
belongs to the set S(C) of vector-valued
solenoidal functions, and the orthogonality condition has the form
_
D
q
0
VnJ. = 0. Vn

H
1
(C).
24 Chapter 2 Overview
Figure 2.2.1: Two blind men and their hypercircle. Geometric interpretation from
[347]
Let
q Q
(
:=
_
j H(C. div)
_
D
j VnJ. =
_
D
nJ. Vn

H
1
(C)
_
.
The set Q
(
contains vector-valued functions that satisfy (in a generalized sense) the
relation div q = 0.
Since Vu q Q
0
= S(C), we have the orthogonality relation
_
D
V(u ) (Vu q) J. = 0.
which implies the estimate
[V(u )[
2
[Vu q[
2
= [V q[
2
. (2.2.3)
Figure 2.2.1 presents a geometric interpretation of the hypercircle formula: if two
blind men walking along two orthogonal roads are able to measure the distance be-
tween them, then they can construct a hypercircle and estimate the distance to the
crosspoint.
From (2.2.3) it also follows that
[V(u )[ = inf
qQ
[V q[. (2.2.4)
This relation and its analogs for more complicated problems generate various a pos-
teriori estimates that use equilibration of the dual variable (ux). We discuss some of
them in Section 2.6.3.
Section 2.3 Mikhlin estimate 25
2.3 Mikhlin estimate
S. Mikhlin [232] suggested to derive a posteriori estimates for linear elliptic problems
with the help of variational arguments. The main idea of this approach is easy to
explain with the paradigm of the problem (2.2.1)(2.2.2).
First, we note that
1
2
[V(u )[
2
= J() J(u). (2.3.1)
where
J(u) := inf

1
1
(D)
J(). J() :=
1
2
[V[
2
_
D
J..
This fact follows from the identity
J() J(u) =
1
2
[V(u )[
2
_
D
(Vu V( u) ( u)) J.
and the relation
_
D
Vu VnJ. =
_
D
nJ.. Vn

H
1
(C).
that holds for the minimizer u. Regrettably, the right-hand side of (2.3.1) is not com-
putable because the value of J(u) is not known. This difculty can be bypassed if a
computable lower bound of J(u) is known. A natural quantity that serves this task is
the value of the so-called dual variational functional
1
+
(q) :=
1
2
[q[
2
whose maximum taken over the set of equilibrated elds coincides with J(u), i.e.,
J(u) = 1() := sup
qQ
1(q).
Thus, we nd that for any q Q
(
1
2
[V(u )[
2
_ J()
1
2
[q[
2
=
1
2
[V[
2
_
D
J.
1
2
[q[
2
=
1
2
[V[
2
_
D
q V J.
1
2
[q[
2
=
1
2
[V q[
2
. (2.3.2)
Later (see, e.g, H. Gajewski, K. Gr oger, and K. Zacharias [145], S. Mikhlin [234],
P. Mosolov, and V. Myasnikov [239]), similar estimates were derived for some classes
of convex variational problems. Estimates based on complementary principles were
also obtained in A. M. Arthurth [14].
Fromthe practical point of viewestimates (2.2.3) and (2.3.2) have an essential draw-
back: they are valid only for q Q
(
. However, the set Q
(
is dened by the differ-
ential relation, which in general is difcult to exactly satisfy. In S. Mikhlin [232], a
strategy close to the known orthogonal projections method (see also M. Vishik [362],
H. Weil [369], S. Zaremba [375]) is discussed. It is based on the construction of an
approximating sequence for the dual problem. However, practical realization of this
approach within the framework of locally supported nite element approximations
may be faced with serious technical difculties. Certainly, for simple problems (e.g.,
homogeneous Laplace equation in a rectangular domain) it is possible to construct
such approximations but for problems with strongly nonhomogeneous coefcients in
complicated domains, vector-valued problems, problems with nonlinear and convec-
tive terms this task is much more difcult. For example, the so-called equilibrated
nite element approximations for problems in solid mechanics (e.g., see C. Johnson
and B. Mercier [188]) are constructed by means of macroelements.
Sometimes, numerical analysts use almost equilibrated approximations that may
not exactly satisfy the differential relation. Substitution of an almost equilibrated
function q may give a good error indicator but in this case the reliability of the upper
bound in (2.3.2) is not guaranteed.
Also, it is worth remarking that an upper bound constructed by some q Q
(
is
efcient only if q is close to the exact ux . Since the latter is unknown, the only
way to nd a suitable q is to minimize [V q[ over Q
(
(or over a certain subspace
of Q
(
). Note that
1
2
[V q[
2
=
1
2
[V[
2
_
D
J.
1
2
[q[
2
.
where the rst and second terms on the right-hand side do not depend on q. Therefore,
the problem is reduced to
inf
qQ
1
2
[q[
2
.
which is the dual variational problem. Hence, in the PragerSyngeMikhlin method,
getting sharp error bounds requires solving the dual problem with the help of conform-
ing approximations of the set Q
(
(which is dened by a differential relation). In the
next chapter, we will see that the requirement q Q
(
is superuous, and guaranteed
upper bounds can be found using a much wider set for q.
2.4 Ostrowski estimates for contractive mappings
Many problems admit a xed point formulation: Find .
in a complete metric space

(X. d) such that
.
= T.
. (2.4.1)
Section 2.4 Ostrowski estimates for contractive mappings 27
where T : X X is a continuous operator. Then approximations are usually con-
structed by the iteration procedure
.
i
= T.
i-1
. i = 1. 2. . . . . (2.4.2)
where .
0
X is a certain selected element. In this case, it is required
(a) to establish conditions that guarantee convergence of .
i
to .
and
(b) to nd computable estimates of the error e
i
= d(.
i
. .
).
Both problems can be effectively solved, provided that T possesses an additional
property.
Denition 2.1. An operator T : X X is called q-contractive on a set S X if there
exists a positive real number q such that the inequality
d(T.. T,) _ q d(.. ,) (2.4.3)
holds for any elements . and , of the set S.
The rst step in the analysis of (2.4.2) is given by the following well-known theo-
rem.
Theorem 2.2 (S. Banach). Let T be a q-contractive mapping of a closed nonempty set
S X to itself with q < 1. Then, T has a unique xed point in S and the sequence .
i
obtained by (2.4.2) converges to this point.
Proof. It is easy to see that
d(.
i1
. .
i
) = d(T.
i
. T.
i-1
)_ qd(.
i
. .
i-1
)_ . . . _ q
i
d(.
1
. .
0
).
Therefore, for any m > 1 we have
d(.
in
. .
i
) _ d(.
in
. .
in-1
) d(.
in-1
. .
in-2
) . . . d(.
i1
. .
i
)
_ q
i
(q
n-1
q
n-2
. . . 1)d(.
1
. .
0
). (2.4.4)
Since
n-1
k=0
q
k
_
1
1 q
.
(2.4.4) implies the estimate
d(.
in
. .
i
) _
q
i
1 q
d(.
1
. .
0
). (2.4.5)
If i o, then the right-hand side of (2.4.5) tends to zero, so that {.
i
] is a Cauchy
sequence. It has a limit , X. Then, d(.
i
. ,) 0 and
d(T.
i
. T,) _ qd(.
i
. ,) 0.
so that d(T.
i
. T,) 0 and T.
i
T,. Pass to the limit in (2.4.2) as i o. We
observe that T, = ,. Hence, any limit of such a sequence is a xed point.
Assume that two different xed points .
1
and .
2
exist. Since
d(.
1
. .
2
) = d(T.
1
. T.
2
) _ qd(.
1
. .
2
)
we arrive at a contradiction, which shows that .
is unique.
Theorem 2.2 also implies an a priori convergence estimate.
Let e
}
= d(.
}
. .
) denote the error at the th step. Then

e
}
= d(T.
}-1
. T.
) _ qe
}-1
_ q
}
e
0
and
e
}
_ q
}
e
0
.
This estimate gives a certain idea of how the error decreases. However, this a priori
upper bound may be rather coarse.
A posteriori error bounds for the iteration method are given by the following theo-
rem.
Theorem 2.3 (Ostrowski [257]). For any .
}
with _ 1, the following estimate holds:
M
}
_
:=
d(.
}1
. .
}
)
1 q
_ e
}
_ M
}
:=
q d(.
}
. .
}-1
)
1 q
. (2.4.6)
Proof. The right-hand side estimate in (2.4.6) follows from (2.4.5). For i = 1 it reads
d(.
1n
. .
1
) _
q
1 q
d(.
1
. .
0
).
Since .
1n
.
as m o, we pass to the limit with respect to m and obtain

d(.
. .
1
) _
q
1 q
d(.
1
. .
0
).
We may view .
}-1
as the starting point of the sequence. Then, in the above relation
.
0
= .
}-1
and .
1
= .
}
, and we arrive at the following upper bound of the error:
d(.
. .
}
) _
q
1 q
d(.
}
. .
}-1
).
The lower bound of the error follows from the relation
d(.
}
. .
}-1
) _ d(.
}
. .
) d(.
}-1
. .
) _ (1 q)d(.
}-1
. .
).
which shows that
d(.
}-1
. .
) _
1
1 q
d(.
}
. .
}-1
).
Section 2.4 Ostrowski estimates for contractive mappings 29
We observe that two-sided estimates of e
}
are easily computable, provided that q is
estimated from above by a number less than 1.
Remark 2.4. Note that
M
}
M
}
_
=
1 q
1 q
d(.
}
. .
}-1
)
d(.
}1
. .
}
)
q _
1 q
1 q
.
Therefore, the efciency of the upper and lower bounds in (2.4.6) crucially depends
on the value of 1 q.
Estimates (2.4.6) and their modications can be applied to linear and nonlinear al-
gebraic systems, integral equations, and other problems solved by iteration methods
(e.g., see E. Zeidler [376]). The major difculty in the application of the above a pos-
teriori estimates is that in practice it may be difcult to nd a sharp upper bound of q
and to establish that it is indeed less than 1.
A posteriori methods for various iteration schemes have been investigated by many
authors. Below we give some references that provide an idea of the results obtained
and will help to nd more pertinent information.
In G. Auchmuty [17], two sided -norm error bounds for solutions of linear sys-
tems are presented. O. Scherzer, H. W. Engl, and K. Kunisch [328] studied a posteriori
choice of the regularization parameter in a Tikhonov scheme for ill-posed problems
(see also H. W. Engl and O. Scherzer [123]). A further study of this strategy is pre-
sented in Jin Qi-nian and Hou Zong-yi [268]. A posteriori estimates for linear ill-posed
problems were investigated in A. Leonov [220]. More information on error estimates
for iteration Newton-type methods can be found, e.g., in K. Braune [75], Jin Qi-nian
[184], P. Meyer [231], F. Potra [263], K. Tsuruta and K. Ohmori [351], T. Yamamoto
[373].
Boundary value problems are often reformulated in terms of integral equations. A
posteriori error estimates for the respective numerical methods (e.g., boundary ele-
ment methods) are usually derived with the help of the techniques that exploits special
properties of the iteration operator. For the readers interested in iteration methods and
a posteriori estimates for integral type formulations of linear and nonlinear boundary
value problems we recommend the papers by C. Carstensen [89], Lin Qun and Shi
Jan [269], M. Schultz and O. Steinbach [330], S. Shaw and J. R. Whiteman [333],
and M. Schulz and W. L. Wendland [331], where they will also nd more references
related to the topic.
Finally, we note that the theory of contractive mappings can generate numerically
based existence theorems that establish existence of localized solutions to nonlinear
problems in the vicinity of a computed (constructed) approximate solution (see, e.g.,
J. G. Heywood, W. Nagata, and W. Xie [171], where such a method was applied to the
NavierStokes equation).
2.5 Error estimates based on monotonicity
The theory of monotone operators gives another way of constructing explicitly com-
putable error estimates. Such operators are dened on the so-called ordered (or par-
tially ordered) spaces that introduce the relation . _ , for all (or almost all) elements
.. , of the space. We recall that the operator T is called monotone if . _ , implies
T. _ T, and antitone if . _ , implies T. _ T,.
Consider an abstract xed point problem: Find .
in a complete ordered (partially

ordered) space X such that
.
= T.
. X. (2.5.1)
Assume that T = T
T
_
, where T
is monotone and T
_
is antitone. Let .
_0
, .
_1
,
.
0
, and .
1
be such that .
_0
_ .
_1
_ .
1
_ .
0
, and
.
_1
= T
.
_0
T
_
.
0
. .
1
= T
.
0
T
_
.
_0
.
Then, we observe that
.
_2
= T
.
_1
T
_
.
1
_ T
.
_0
T
_
.
0
= .
_1
.
.
2
= T
.
1
T
_
.
_1
_ T
.
0
T
_
.
_0
= .
1
.
By continuing the iterations we obtain elements such that
.
_k
_ .
_(k1)
_ .
(k1)
_ .
k
. (2.5.2)
If . T. maps a compact set D X to itself, then by the Schauder xed point
theorem .
exists and belongs to D. From (2.5.2) it follows that .
is bounded from
below and above by the sequences {.
_k
] and {.
k
], respectively.
Applications of this method are mainly oriented towards systems of linear simul-
taneous equations and integral equations (a detailed discussion of the monotonicity
methods is presented in L. Collatz [110]). As an example, we consider the system of
linear simultaneous equations
. = . .
which is supposed to have a unique solution .
. Assume that
=

_
.
_
= {a
_
i}
] M
dd
.
= {a
i}
] M
dd
. a
_
i}
_ 0. a
i}
_ 0.
We may partially order the space R
d
by saying that . _ , if and only if .
i
_ ,
i
for
i = 1. 2. . . . . n. Compute the vectors
.
_(k1)
=
.
_k

_
.
k
. .
(k1)
=
.
k

_
.
_k
.
Section 2.6 A posteriori error indicators for nite element approximations 31
If .
_0
_ .
_1
_ .
_ .
1
_ .
0
, then for the components of .
we obtain two-sided
estimates
.
(i)
_k
_ .
(i)
_(k1)
_ .
(i)
_ .
(i)
(k1)
_ .
(i)
k
. i = 1. 2. . . . . n. (2.5.3)
It should be noted that the convergence of .
(i)
_k
and .
(i)
k
to .
(and the convergence

rate) calls for a special investigation, which must use specic features of a particular
problem.
The reader interested in estimates based on monotonicity can nd more information
and relevant references in, e.g., E. Geisler, A. Tal, and D. Garg [150] (where they
are discussed in the context of ordinary nonlinear equations), J. Schr oder [329], and
M. Plum [262] (where monotonicity methods were used to perform computer-assisted
existence proofs based on Schauders xed point theorem).
2.6 A posteriori error indicators
for nite element approximations
It is commonly accepted that the mathematical concept of the nite element methods
brings its origin in the paper by R. Courant [112] where the method was introduced as a
variant of the B. Galerkin [147] and W. Ritz [323] methods using locally supported trial
functions. In 6070s, nite element methods have formed one of the main approaches
to the numerical analysis of PDEs. The foundations of the nite element method are
exposed in the books by Ph. Ciarlet [107], S. Brenner and R. L. Scott [76], C. Johnson
[185], J. T. Oden and J. N. Reddy [254], G. Strang and G. Fix. [344], O. C. Zienkiewicz
and K. Morgan [381], and many other publications. Achievements and some unsolved
problems of the nite element method are discussed, e.g., in the papers by I. Babu ska
[20] and O. C. Zienkiewicz [379] (see also I. Babu ska and J. E. Osborn [30]).
In the late 70s and early 80s, it became clear that the successive numerical methods
for PDEs should be based on the so-called mesh-adaptive procedures that modify
nite dimensional spaces with the help of the information contained in an approximate
solution computed at the previous step. This new concept generated an interest to
a posteriori error indicators that provide information for a proper improvement of
nite-dimensional spaces. Nowadays, mesh-adaptive methods based on a posteriori
estimates dominate in the numerical analysis of differential equations (e.g., see [261,
7, 8, 35, 40, 41, 100, 117, 124, 178, 187, 190, 322, 356, 383] and the papers cited
therein). The majority of a posteriori methods for FEM are based either on the analysis
of residuals or on special properties (e.g., additional regularity) of exact solutions.
A posteriori error estimation for nite element approximations is the main subject
of the books by M. Ainsworth and J. T. Oden [8], I. Babu ska and T. Stroboulis [35],
W. Bangerth and R. Rannacher [40], K. Eriksson, D. Estep, P. Hansbo and C. Johnson
[124], and R. Verf urth [356]. There, the reader will nd detailed expositions of various
approaches, results of numerical experiments, and a wide list of references. Below,
we shortly discuss several a posteriori error estimation methods developed for nite
element approximations. Certainly, the exposition is not complete. Its goal is to give a
view of the mathematical ideas underlying the methods.
2.6.1 Explicit residual methods
From the mathematical viewpoint, the classical residual method is a method for
nding an upper bound of the residual functional evaluated in the topology of the
image space of the respective operator. It leads to the so-called explicit residual
a posteriori estimate, which yields an error bound for a Galerkin approximation as a
sum of element-wise residuals and interelement jumps with weights given by constants
in the so-called Cl ements interpolation inequalities (see Ph. Cl ement [108]).
To present the main idea, we use the problem (2.2.1)(2.2.2), the generalized solu-
tion of which meets the integral identity
_
D
Vu VnJ. =
_
D
nJ.. n V
0
:=

H
1
(C). (2.6.1)
Let V
0
be an approximate solution of this problem. Then,
_
D
V(u ) VnJ. = F
(n). n V
0
. (2.6.2)
where
F
(n) :=
_
D
( n V Vn) J.
is a linear functional dened on V
0
. It is easy to see that this functional is equal to zero
if coincides with u. In all other cases, the norm of this functional, dened by the
relation
[[[ [[[ F
[[[ [[[ := sup

uV
0
,
u,=0
[ F
(n) [
[Vn[
. (2.6.3)
is positive. Therefore, it is natural to call F
the error functional. It is easy to show

that [[[ [[[ F
[[[ [[[ is indeed a measure of the deviation of from u. By (2.6.2) and (2.6.3), we
note that
_
D
[ V(u ) [
2
J. = F
(u ) _ [[[ [[[ F
[[[ [[[ [V(u )[.

Hence,
[V(u )[ _ [[[ [[[ F
[[[ [[[ . (2.6.4)

However, [ F
(n) [_ [V(u )[[Vn[, so that (2.6.3) implies the inequality op-

posite to (2.6.4). Thus, the norm of the norm of the deviation from the exact solution
coincides with the norm of F
. Here arises the problem of how to compute [[[ [[[ F
[[[ [[[ prac-

tically for a given . A straightforward computation of the norm based on (2.6.3) is
hardly possible. A more promising way is to nd computable upper bounds of [[[ [[[ F
[[[ [[[ .
In the papers of I. Babu` ska and W. C. Rheinboldt [31, 32] and some other publications
of them, a way was suggested for deriving such bounds, which was later called the
explicit residual method.
The explicit residual method is applicable if is a Galerkin approximation of the
exact solution computed on a nite-dimensional space V
h
V
0
, i.e., if = u
h
, where
_
D
Vu
h
Vn
h
J. =
_
D
n
h
J.. n
h
V
0h
. (2.6.5)
In this case,
_
D
V(u u
h
) Vn
h
J. = 0. n
h
V
0h
. (2.6.6)
i.e., the error V(u u
h
) is orthogonal to Vn
h
for any n
h
V
0h
, and we nd that
F
u
h
(n) =
_
D
V(u u
h
) VnJ. =
_
D
V(u u
h
) V(n
h
n) J..
where
h
: V
0
V
0h
is a continuous mapping (typically
h
is dened by the
Cl ements interpolation operator [108]). Let C be divided into a collection of sub-
domains C
k
, k = 1. 2. . . . . M, and u
h
be a smooth function in each subdomain. Then
F
u
h
(n) :=
k=1
_
D
k
^(u
h
u)(n
h
n) J.
k,I=1
_
I
kl
_
d(u u
h
)
dv
kI
_
I
kl
(n
h
n) Js.
where I
kI
is the common part of the boundaries of C
k
and C
I
, v
kI
is the unit normal
vector to this boundary, and
_
_
I
kl
denotes the jump of a quantity at the boundary
I
kI
. If the C
k
are simplexes, then for
h
the interpolation estimates are as follows:
[n
h
n[
D
k
_ ;
1k
diam(C
k
)[n[
1,2,o
1k
. (2.6.7)
[n
h
n[
I
kl
_ ;
2k
[I
kI
[
1{2
[n[
1,2,o
2k
. (2.6.8)
where o
1k
and o
2k
are certain domains (patches of neighbor elements) that contain
C
k
,
diam(C
k
) = sup
x
1
,x
2
D
k
[.
1
.
2
[.
and ;
1k
and ;
2k
are the interpolation constants (which depend not only on C
k
, but
also on the form of all elements in the patches o
1k
and o
2k
, respectively). By these
estimates we can represent [[[ [[[ F
u
h
[[[ [[[ as the sum of local quantities
j
2
k
:= diam(C
k
)
2
[û
h
[
2
D
k

1
2
I=1
[I
kI
[
_
_
_
_
_
_
du
h
dv
kI
_
I
kl
_
_
_
_
_
2
I
kl
.
which are related to the residual on C
k
and the values of the jumps in the normal
component of the gradient on the boundary. The respective estimate of the overall
error is as follows:
[V(u u
h
)[
2
_ [[[ [[[ F
u
h
[[[ [[[
2
_
k=1
c
k
j
2
k
. (2.6.9)
where the constants c
k
depend on ;
1k
and ;
2k
associated with the sampling con-
sidered. The quantities j
k
are used as error indicators. Comparing their values on
different elements, one can suggest an adequate mesh-adaptation procedure. In this
case, the estimate is used in the form
[V(u u
h
)[
2
_ C
k=1
j
2
k
. (2.6.10)
where the different c
k
are replaced by one common constant C. It should be noted that
often the values of two terms in j
k
are quite different. In this case, error indicators for
mesh-adaptation can be constructed with the help of only the dominant part of j
k
.
Difculties arise if a guaranteed and sharp upper bound of [V(u u
h
)[ is required.
Indeed, we must nd a large amount of local constants ;
1k
and ;
2k
in (2.6.7)(2.6.8).
In general, nding such a constant requires solving an innite-dimensional problem
on each patch. For example,
;
1k
= diam(C
k
)
-1
sup
u!
1
(o
1k
)
u,=0
[n
h
n[
D
k
[n[
1,2,o
1k
. (2.6.11)
In general, this problem is of the same level as (2.6.3). In the literature, it is sometimes
recommended to replace H
1
in (2.6.11) by a certain set of polynomial functions. This
provides a possibility of computing an approximate value of ;
1k
but, in such a case,
the reliability of the upper bound of the error may be lost. Another unpleasant feature
of the method is that ;
1k
and ;
2k
depend on T
h
and, consequently, all of them must be
recalculated if one sampling is replaced by another one.
Also, it is worth mentioning that the estimate (2.6.9) is derived by formal mathe-
matical transformations that clearly overestimate [[[ [[[ F
u
h
[[[ [[[ . Obviously, in (2.6.10) such
an overestimation may be much larger (especially for nonuniform meshes, where the
constants ;
1k
and ;
2k
are quite different). Thus, such estimates may lead to a con-
siderable overestimation of the total error. This fact was observed in some carefully
performed tests (e.g., see C. Carstensen and S. A. Funken [93]). Nevertheless, the ex-
plicit residual method is widely exploited in practice, and the quantities j
k
are often
used not for the computation of a guaranteed upper bound of the total error, but serve
as error indicators that give am idea of the distribution of local errors in C.
The amount of publications associated with residual based methods is huge. In ad-
dition to the books [8, 35, 124, 356], the reader will nd discussions of the method,
e.g., in M. Ainsworth and J. T. Oden [6, 7], C. Carstensen [87], C. Carstensen and
S. A. Funken [92, 93] (in these papers the authors consider ways of computing bounds
of interpolation constants in the residual based error estimator), C. Carstensen and
R. Verf urth [100] (the authors show that the edge component dominates in the resid-
ual error estimator), K. Eriksson and C. Johnson [125], and C. Johnson and P. Hansbo
[187].
In C. Carstensen and S. Sauter [99], a posteriori error estimates were derived for
elliptic PDEs on domains with complicated structures. Estimates for problems with
biharmonic operator are analyzed in A. Charbonneau, K. Dossou and R. Pierre [102].
Residual-type estimates for linear rst order systems of PDEs were obtained in
P. Houston, J. A. Mackenzie, E. Suli, and G. Warnecke [180].
A posteriori estimates taking into account the inuence of the non-discretized part
of the domain on the approximation error are considered in W. D orer and M. Rumpf
[117]. Estimates in the 1
o
-norm can be found, e.g., in S. W. Brady and A. R. Elcrat
[66]. Adaptive methods for convection-diffusion problems are considered in C. John-
son [186] and R. Verf urth [360]. Papers by R. Verf urth [359] and K. Eriksson and
C. Johnson [126] are devoted to parabolic type problems. A posteriori error esti-
mates for anisotropic meshes are presented in K. G. Siebert [334] and G. Kunert [203,
204]. Also, we recommend papers by I. Babu ska, R. Duran, and R. Rodriguez [25],
G. F. Carey and D. L. Humphrey [84], J. T. Oden, L. Demkowicz, W. Rachowicz, and
T. A. Westermann, [251], R. E. Ewing [129], D. W. Kelly, J. R. Gago, O. C. Zienkie-
wicz, and I. Babu ska [195], B. I. Wohlmuth and R. H. W. Hoppe [372], where the
reader will nd discussions of various a posteriori estimators and more references.
Finally, we conclude this overview by saying several words about residual a poste-
riori estimates for nonlinear boundary value problems. In an abstract form (for non-
linear mappings) residual type estimates were considered in J. Pousin and J. Rappaz
[264, 265] and R. Verf urth (e.g., see [355, 356]). The book [356] contains a system-
atic consideration of this question. Certain particular classes of nonlinear problems
were analyzed by many authors. For example, nonlinear diffusion-convection prob-
lems were considered in J. Medina, M. Picasso and J. Rappaz [229] and for nonlinear
parabolic problems in R. Verf urth [357]. A posteriori estimates for variational inequal-
ities related to problems with obstacles were analyzed in M. Ainsworth, J. T. Oden, and
C. Y. Lee [9], D. Braess [68], Z. Chen and R. H. Nochetto [103], R. H. W. Hoppe and
R. Kornhuber [178], R. Kornhuber [198], and A. Vesser [361].
2.6.2 Implicit residual methods
In the so-called implicit residual methods, the error is represented as the solution of
an auxiliary boundary value problem. For example, let u
h
V
0h
be a conforming
approximation of the problem (2.6.1). By (2.6.2) we have the relation
_
D
V(u u
h
) VnJ. = F
u
h
(n). Vn V
0
(2.6.12)
which shows that e = u u
h
is a solution of the problem
ê r(u
h
) = 0 in C. (2.6.13)
e = 0 on I. (2.6.14)
where r(u
h
) = div Vu
h
H
-1
(C) and (2.6.13) is understood in the sense of
distributions. Formally, this idea can be extended to a wide class of linear problems.
Indeed, if A : X Y is a linear operator for which we consider the problem
Au = 0 in C. (2.6.15)
u = 0 on I. (2.6.16)
then, for any X,
Ae r() = 0 in C. (2.6.17)
e = 0 on I. (2.6.18)
where r() = A is the residual.
We note that nding accurate approximate solutions of (2.6.13)(2.6.14) (and of
(2.6.17)(2.6.18)) may be a difcult task (because, in general, the error functional is
a distribution so that we have a boundary value problem with rather irregular right-
hand side). Moreover, the accuracy of an approximate solution obviously affects the
accuracy of error estimation, so that a new error estimation problem arises.
In practical applications, several methods are used in order to overcome the above
mentioned difculties. In particular, it is often suggested to split the residual functional
into a number of functionals dened as solutions of local subproblems. For example,
in R. E. Bank and A. Weiser [42]) error indicators are constructed with the help of
local boundary value problems with data dened by residuals and interelement jumps
(see also R. Duran and R. Rodriguez [119]). In the equilibrated residual method (e.g.,
see M. Ainsworth and J. T. Oden [8]), local boundary value problems are constructed
on each element, using the residuals and suitable Neumann conditions on boundaries
of the elements. Local Dirichlet and Neumann problems on patches are also used in
E. Stein and S. Ohnimus [339] and R. Verf urth [358, 360]. Applications to nonlinear
problems are discussed, e.g., in C. Carstensen, R. Klose, and A. Orlando [97].
Finally, we note that implicit type methods are often used for the indication of local
errors. Concerning a posteriori methods developed to evaluate local errors of FEM ap-
proximations, we address the reader to the books by M. Ainsworth and T. Oden [8] and
I. Babu ska and T. Strouboulis [35]. Also we recommend papers by I. Babu ska, F. Ih-
lenburg, A. Mathur, T. Strouboulis, S. K. Gangaraj, C. S. Upadhyay [27, 39, 39, 37],
E. Stein and S. Ohnimus [338], R. Verf urth [358, 360], and M. Ainsworth, J. T. Oden
and C. Y. Lee [9].
2.6.3 A posteriori estimates based on post-processing
of approximate solutions
Post-processing methods exploit certain a priori known properties of exact solutions.
In general terms, the situation that typically arises for nite element approximations
is as follows. Consider a conforming approximation u
h
(which belongs to a nite-
dimensional subspace V
h
) and the function u
h
, where is a certain linear operator
(e.g., the operator V). Usually, u
h
lies in a rather wide space U. For example, if
approximations of (3.2.1)(3.2.2) are constructed with the help of piece-wise afne
continuous functions then Vu
h
1
2
(C). However, in many cases a priori estimates
of the exact solution guarantee that
u U.
where U is a subset of U. In particular, if 1
2
(C), then the exact ux Vu of
the problem (3.2.1)(3.2.2) is a vector-valued function in U = H(C. div ). Moreover,
very often we know that Vu H
1
(C. R
d
) (globally or locally). Another option is to
set U = Q
(
. These observations suggest an idea to post-process Vu
h
and nd a close
vector-valued function that satises some of the above-menioned properties. Formally,
the principal scheme is as follows.
Assume that we have a continuous mapping G such that
(a) G(
h
) U,
(b) post-processing is a relatively inexpensive procedure,
(c) G(
h
) is much closer to u than
h
.
If G satises (a)(b), then the difference G(
h
)
h
generates an efcient indi-
cator of element-wise errors.
Regularization. Usually, elliptic problems with smooth coefcients have regular so-
lutions in internal subdomains, which suggests an idea to project uxes of approxi-
mate solutions to a set of more regular functions (e.g., see the paper by J. H. Bram-
ble and A. H. Schatz [73], which is one of the earliest publications in this area). If
the error caused by violations of a priori regularity properties dominates and a post-
processing operator efciently performs regularization of approximate solutions, then
one may hope that the difference between the approximate solution and its regularized
(smoothed) counterpart represents the major part of the error. Typically, regularization
is applied to uxes Vu
h
(or Vu
h
) and to stresses for problems related to continuum
mechanics.
Assume that G performs efcient regularization of uxes, i.e.,
[GVu
h
Vu[ _ [Vu
h
Vu[ . (2.6.19)
where < 1. In this case,
[Vu
h
Vu[ _ [GVu
h
Vu
h
[ [Vu
h
Vu[
and
[GVu
h
Vu
h
[ _ [Vu
h
Vu[ [Vu
h
Vu[ .
Hence,
(1 ) [Vu
h
Vu[ _ [GVu
h
Vu
h
[ _ (1 ) [Vu
h
Vu[ . (2.6.20)
If 1, then (2.6.20) shows that the image of Vu
h
computed by a post-processing
operator G is much closer to the exact solution than Vu
h
. Thus,
[Vu
h
Vu[ . [GVu
h
Vu
h
[ . (2.6.21)
and the function [GVu
h
Vu
h
[(.) (which is easily computable) can be used as an
error indicator. Certainly, the quality of such error indicator depends on the smallness
of .
The regularization of uxes (or other solution components) leads to a variety of
error indicators. Below we briey describe some of them.
Local post-processing. In most cases, post-processing is performed by elementwise
averaging procedures. Let O
i
be a patch of nite elements (see Fig 2.6.1), i.e.,
O
i
=
_
}=1,2,...,n
i
T
i}
.
O
T
T
T
i
i
i
i
1
2
3
Figure 2.6.1 Patch O
i
.
Dene g
i
as a vector-valued function in 1
k
(O
i
. R
d
) solving the minimization prob-
lem
inf
1
k
(O
i
,R
o
)
_
O
i
[g GVu
h
[
2
J.. (2.6.22)
The minimizer g
i
can be used to dene the values of an averaged ux at some points.
Further, these values are utilized by a prolongation procedure that denes an averaged
function GVu
h
: C R. Consider the simplest case. Let u
h
be a piecewise afne
continuous function. Then,
Vu
h
1
0
(T
i}
. R
d
) VT
i}
T
h
.
and Vu
h
, H
1
(C). Denote the values of Vu
h
on T
i}
by (Vu
h
)
i}
. Set k = 0 and nd
g
i
1
0
such that
_
O
i
[g
i
Vu
h
[
2
J. = inf
1
0
(O
i
)
_
O
i
[g Vu
h
[
2
J.
= inf
1
0
(O
i
)
_
_
_
[g[
2
[O
i
[ 2g
}=1
(Vu
h
)
i}
[T
i}
[
}=1
[(Vu
h
)
i}
[
2
[T
i}
[
_
_
_
. (2.6.23)
It is easy to see that g
i
is given by a weighted sum of (Vu
h
)
i}
, namely,
g
i
=
}=1
[T
i}
[
[O
i
[
(Vu
h
)
i}
. (2.6.24)
Now, we dene the value of GVu
h
(.
i
) as g
i
. Repeat this procedure for all nodes and
dene the vector-valued function GV(u
h
) by piecewise afne prolongation of these
values.
If the mesh is regular and all the quantities [T
i}
[ are equal, then (2.6.24) reads
g
i
=
}=1
1
M
i
(Vu
h
)
i}
. (2.6.25)
Various averaging formulas of this type are represented in the form
g
i
=
}=1
z
i}
(Vu
h
)
i}
.
}=1
z
i}
= 1. (2.6.26)
where the quantities z
i}
are weight factors. For internal nodes, they may be taken in
accordance with (2.6.24) or dened by the rule
z
i}
=
[;
i}
[
2
.
where [;
i}
[ is the radian measure of the angle of T
i}
associated with the node i . How-
ever, if a node belongs to the boundary, then it is better to choose special weights.
Their values depend on the mesh and on the type of the boundary. The reader can nd
a detailed consideration of this question in I. Hlav a` cek and M. K`ri` zek [176].
Another way of dening g
i
is to solve the problem
inf
P
k
(O
i
)
n
i
x=1
[g(.
x
) GVu
h
(.
x
)[
2
. (2.6.27)
where the points .
x

O
i
are the so-called superconvergent points. Evidently, in this
case, the integral type averaging is replaced by a discrete one.
Global averaging. Local minimization problems on patches can be replaced by the
following global problem: Find g
h
U
h
such that
[ g
h
u
h
[
2
D
= inf
h
U
h
[g
h
u
h
[
2
D
. (2.6.28)
where U
h
is a certain nite-dimensional subspace of U. The function g
h
can be viewed
as Gu
h
. Very often g
h
is a better image of Gu than the functions obtained by local
procedures. Moreover, mathematical justications of the methods based on global
averaging procedures can be performed under weaker assumptions what makes them
applicable to a wider class of problems (e.g., see C. Carstensen and S. A. Funken
[92] and C. Carstensen and S. Bartels [90] where it was shown that each averaging
procedure leads to a certain a posteriori estimate).
Estimates based upon global averaging of gradients were also considered in
B.-O. Heimsund, X.-C. Tai and J. Wang [170]. In J. Wang [367], it was suggested
the so-called least squares surface tting procedure that for problems with suf-
ciently smooth solutions lead to a recovered function with superconvergent properties.
The analysis is based on the presentation
u Q
r
u
h
= (u Q
r
u) Q
r
(u u
h
).
where u is the exact solution of a linear elliptic problem, u
h
is the Galerkin approx-
imation computed on a mesh T
h
and Q
r
is the 1
2
-projection operator on the nite-
dimensional space constructed on a mesh T
r
with the help of piecewise polynomial
functions of the order r _ 0. The key fact is that
[Q
r
u Q
r
u
h
[ _ Ch
x-1 min0,2-x
[u u
h
[
1
1. (2.6.29)
where (0. 1) is a parameter that relates h and t as t = h
, original problem is
assumed to be H
x
-regular with 1 _ s _ k 1, and k is the degree of polynomials
used in the Galerkin approximation. From (2.6.29) it follows that
[u Q
r
u
h
[ 6 C h
(h,,i,k)
_
_
_
u
(
_
_
i1,D
0
_
_
u
(
_
_
k1,D
_
. (2.6.30)
provided that u H
k1
(C) H
i1
(C
0
) V
0
. In (2.6.30), the rate depends on
t. h. r and k and is grater than 2 provided that u
(
is regular enough, and the space V
r
is
selected appropriately (i.e. it is sufciently rich). Concrete values of the convergence
rate for various k, r, and are presented in ([367]).
We conclude this short overview of the methods using regularization of approximate
solutions by several literature comments. One of the most well-known post-processing
methods originates from the works of O. C. Zienkiewicz and J. Z. Zhu [382, 383] (in
the literature it is often called ZZ error indicator). Simple and efcient error indicators
based upon gradient averaging were suggested by many authors. Here, we refer to,
e.g., M. Ainsworth, J. Z. Zhu, A. W. Craig and O. C. Zienkiewicz [10], I. Babu ska
and R. Rodriguez [33], B. Boroomand and O. C. Zienkiewicz [64], I. Babu ska and
R. Rodriguez [33], R. Duran, M. A. Muschietti and R. Rodriguez [118], R. Rodrigues
[326], O. C. Zienkiewicz and J. Z. Zhu [384, 378], O. C. Zienkiewicz, B. Boroomand
and J. Z. Zhu [380].
Mathematical justications of the condition 1 (which was used in (2.6.21)) are
based on the so-called superconvergence phenomenon, which states that certain com-
ponents of approximate solutions (Galerkin approximations) converge to the exact val-
ues with rates higher than the rate of the energy norm of the error. Probably, the earli-
est results on superconvergence were established in the papers by L. A. Oganesjan and
L. A. Ruchovets [256] and M. Zl amal [385, 386]. Error indicators and adaptive meth-
ods based upon superconvergence phenomenon are discussed in, e.g., to J. Brandts
[74], S.-S. Chow, G. F. Carey and R. D. Lazarov [106], R. E. Ewing, R. D. Lazarov
and J. Wang [130], I. Hlav a cek and M. Kri zek [176], M. Kr zek and P. Neittaanm aki
[202, 242], R. Lazarov [217], R. Verf urth [356], J. Wang [367], J. Wang and X. Ye
[368], N.-E. Wiberg, F. Abdulwahab and S. Ziukas [371], Z. Zhang and A. Naga [377],
and in many other publications. In L. Wahlbin [366] readers can nd a detailed expo-
sition of the subject and other references. Surveys on superconvergence are presented,
e.g., in M. Kr zek and P. Neittaanm aki [201], J. R. Whiteman and G. Goodsell [370],
and in the book [200].
Equilibration. Another group of methods exploits differential relations (usually they
follow from conservation laws), which for exact solutions must be exactly satised.
Consider again the problem (2.6.1)(2.6.2) and set U = Q
(
. If G
eq
is an equilibra-
tion operator that transforms Vu
h
into q
h
= G
eq
Vu
h
Q
(
, then we can apply the
estimate (2.2.4) and nd that
[V(u u
h
)[ _ [Vu
h
q
h
[.
If the equilibration is sharp (i.e., div q
h
= 0 ), then this method provides a guaran-
teed upper bound of the error. This property is lost if q
h
is equilibrated approximately
but usually Vu
h
q
h
serves as a good error indicator provided that q
h
is sufciently
close to the set of equilibrated elds.
Various approaches based on equilibration type post-processing of approximate so-
lutions are presented in, e.g, M. Ainsworth and J. T. Oden [8], D. W. Kelly [194],
P. Ladev eze and D. Leguillon [207] (the authors show that if the equilibrium equation
is exactly satised, then the estimate is reduced to the error in constitutive relations;
extensions of the method to some other other problems can be found in P. Ladev eze,
J.-P. Pelle and Ph. Rougeot [208], P. Ladev eze and Ph. Rougeot [209], L. Gallimard,
P. Ladev eze and J.-P. Pelle [148], P. Coorevits, J.-P. Dumeau and J.-P. Pelle [111]),
E. Stein and S. Ohnimus [338], E. Stein, F. J. Bartold, S. Ohnimus, and M. Schmidt
[337]).
Post-processing procedures based upon an equilibration of recovered stresses were
considered in B. Boroomand and O. C. Zienkiewicz [63]. They result in stress elds
that satisfy an equilibrium condition in a weak form. A method of equilibration is
presented in P. Destuynder and B. M etivet [114]. This paper also contain numerical
tests, in which a posteriori error estimates obtained by equilibration are compared with
other estimates.
Recently a new equilibration method was suggested in publications of D. Braess
and J. Sch oberl [67, 69]. In it uxes are projected to a subspace formed by Raviart
Thomas elements. After that a certain iteration procedure is performed on patches in
which normal components of local uxes are changed in order to satisfy the respective
integral (balance) relation on each element.
2.6.4 A posteriori methods using adjoint problems
In the recent decades, it was developed a new approach to a posteriori error estimation
based on the attraction of adjoint boundary value problems. Probably, rst results in
the error analysis for PDEs obtained with the help of adjoint problems are related to
the works of J.-P. Aubin [16] and J. A. Nitsche [249] where it was suggested a way of
deriving a priori rate convergence estimates in weaker norms. The idea to use adjoint
problems in order to establish upper bounds for linear functionals of approximation
errors is briey discussed in the book by S. Mikhlin [232] (with a reference to a paper
by M. Slobodyanskii [335]). Among early publications close to this approach, we also
mention the work of T. Kato [192].
Nowadays, error estimates using adjoint problems are widely used in computer sim-
ulation. Concerning this subject, we rst refer to the so-called dual-weighted residual
method. One of the advantages of this method is that the attraction of the adjoint prob-
lem allows one to avoid difculties with the evaluation of the interpolation constants.
Readers will nd a detailed exposition of the method and applications to various
problems in R. Rannacher [270] and in the book by W. Bangerth and R. Rannacher
[40]. Also, we recommend the papers by R. Becker and R. Rannacher [48], C. Johnson
and A. Szepessy [190], P. Houston, R. Rannacher, E. S uli [182], and R. Rannacher and
F. T. Suttmeier [272, 273].
Adjoint problems are used in the so-called goal-oriented a posteriori error estimates
that measure errors in terms of special goal-oriented quantities instead of global
energy norms (e.g., see W. Bangerth and R. Rannacher [40], J. T. Oden and S. Prud-
homme [253], J. Peraire and A. T. Patera [260], E. Stein, M. R uter, and S. Ohnimus
[340], and the literature cited in these publications). The idea of this approach can be
explained with the paradigm of the problem
_
D
(Vu Vn n) J. = 0 Vn V
0
:=

H
1
(C). (2.6.31)
where is a positive denite real matrix and and 1
2
(C) is a given function.
Let u
h
V
0
be an approximate solution computed on the mesh T
h
. Assume that
it is required to estimate the quantity (. u u
h
), where V
+
0
is a given linear
functional. Typically, is an integral type functional localized in a certain part of C.
Dene u
I
by the relation
_
D
?
Vu
I
VnJ. = (. n) Vn V
0
. (2.6.32)
in which
?
is the matrix adjoint to . From (2.6.31) and (2.6.32), it follows that
(. u u
h
) =
_
D
?
Vu
I
V(u u
h
) J.
=
_
D
( u
I
Vu
h
Vu
I
) J. = 1(u
I
. u
h
). (2.6.33)
Hence, (. (u )) can be easily estimated provided that u
I
is known. In practice, u
I
is replaced by an approximation u
Ir
computed on an adjoint mesh T
r
(which may not
coincide with T
h
). If u
Ir
is a sharp approximation of u
I
, then the quantity 1(u
Ir
. u
h
)
could be a good indicator of (. u u
h
). To obtain another indicator, we rewrite
(2.6.33) in the form
(. u u
h
) = 1
1
(u
h
. u
Ir
) 1
2
(u. u
h
. u
I
. u
Ir
). (2.6.34)
where
1
1
(u
h
. u
Ir
) :=
_
D
( u
Ir
Vu
h
Vu
Ir
) J.
is a directly computable functional and
1
2
(u. u
h
. u
I
. u
Ir
) :=
_
D
(Vu Vu
h
) (Vu
I
Vu
Ir
) J..
If T
r
and T
h
coincide, then 1
1
(u
h
. u
Ir
) = 0 (because u
h
is a Galerkin approximation).
In this case,
[(. u u
h
)[ _ 1
2
(u. u
h
. u
I
. u
Ir
). (2.6.35)
Estimate (2.6.34) serves as a source of various indicators. One of them is based
on the idea to replace unknown functions Vu and Vu
I
by averaged gradients G
h
Vu
h
and G
r
Vu
Ir
, where G
h
and G
r
are the respective averaging operators. It is proved
that under the standard assumptions that guarantee superconvergence of the primal
and adjoint approximations such a replacement leads to a higher order error (see [199,
244]). Then, the quantity
1
2
(u. u
h
. u
I
. u
Ir
) :=
_
D
(G
h
Vu
h
Vu
h
) (G
r
Vu
Ir
Vu
Ir
) J.
can be used instead of 1
2
.
Finally, we note that the quantity (. u ) cannot completely characterize the er-
ror because it vanishes if u is orthogonal to . Therefore, it is desirable to obtain
estimates for various functionals
x
, which amounts to solving several adjoint prob-
lems (e.g., see [244]). Recently, new forms of the above-discussed error indicators has
been derived and tested (see P. Neittaanm aki, S. Repin, and P. Turchin [246]). They
do not exploit superconvergence of the adjoint solution and, therefore, can be used if
adjoint meshes are not very regular.
3 Poissons equation
In this chapter, we begin studying estimates of the type (1.3.4). To present the main
ideas in the most transparent formwe, throughout the chapter consider only one elliptic
problem:
û = 0 in C. (3.0.1)
u = 0 on I. (3.0.2)
For this problem, we derive two-sided a posteriori estimates with the help of two dif-
ferent methods. The rst method uses variational arguments, the second one is based
on transformations of the corresponding integral identity. We discuss properties of
the estimates, their practical implementation, and relationships between them and a
posteriori estimates of other types.
3.1 The variational method
The variational method is based upon the variational statement of the problem (3.0.1):
Find u V
0
such that
J(u) = inf
V
0
J(). J() =
_
D
_
1
2
[V[
2

_
J..
Henceforth, we call it Problem P.
Note that
J() = sup
,Y
1(V. ,). 1(V. ,) =
_
D
_
V ,
1
2
[,[
2

_
J..
where Y = 1
2
(C. R
d
). Indeed, the value of the above supremum cannot exceed the
quantity that we obtain if, for almost all . C, the value of ,(.) is dened as the
maximizer of the problem
sup
R
o
_
V(.)
1
2
[[
2
_
.
It is easy to observe, that (at any . C) the solution of this problem is = V(.).
Since V(.) Y , we conclude that
sup
,Y
1(V. ,) = 1(V. V) = J().
46 Chapter 3 Poissons equation
Then, the original (or primal) problem takes the minimax form:
(P) inf
V
0
sup
,Y
1(V. ,). (3.1.1)
If the order of inf and sup is changed, then we arrive at the so-called dual problem
_
P
+
_
sup
,Y
inf
V
0
1(V. ,). (3.1.2)
Note that
inf
V
0
_
D
_
V ,
1
2
[,[
2

_
J. =
1
2
[,[
2
inf
V
0
_
D
(V , ) J.
=
_

1
2
[,[
2
if , Q
(
.
o if , , Q
(
.
where the set Q
(
is dened in Section 2.2. Hence, the dual problem has the form:
Find Q
(
such that
1
+
() = sup
,Q
1
+
(,). (3.1.3)
where
1
+
(q) =
1
2
[q[
2
.
How are the problems (P) and (P
+
) related to each other? To answer this question,
we rst establish a relation that holds regardless of the structure of 1(.. ,).
Lemma 3.1. Let 1(.. ,) be a functional dened on the elements of two nonempty sets
X and Y . Then
sup
,Y
inf
xA
1(.. ,) _ inf
xA
sup
,Y
1(.. ,). (3.1.4)
Proof. It is easy to see that
1(.. ,) _ inf
A
1(. ,). V. X. , Y.
Pass to the supremum over , Y . We obtain
sup
,Y
1(.. ,) _ sup
,Y
inf
A
1(. ,). V. X.
The left-hand side depends on ., whereas the right-hand side is a number. Thus, we
take the inmum over . X and conclude that
inf
xA
sup
,Y
1(.. ,) _ sup
,Y
inf
A
1(. ,).
Section 3.1 The variational method 47
Therefore, we always have sup P
+
_ inf P. However, in our case we have a
stronger relation, namely, sup P
+
= inf P. To prove this fact, we note that
_
D
Vu V J. =
_
D
J.. V V
0
.
Therefore Vu Q
(
and
1
+
() _ 1
+
(Vu) =
1
2
[Vu[
2
=
_
D
_
1
2
[Vu[
2
[Vu[
2
_
J.
=
_
D
_
1
2
[Vu[
2
u
_
J. = J(u).
Thus, we conclude that 1
+
() = J(u) and, consequently, u and = Vu are the
solutions of the primal and dual problems, respectively.
Recall the estimate by Mikhlin (2.3.1)
1
2
[V(u )[
2
= J() 1
+
() _ J() 1
+
(q). Vq Q
(
.
from which we obtain (see Section 2.3)
[V( u)[ _ [V q[. Vq Q
(
.
Take arbitrary , 1
2
(C). Then,
[V( u)[ _ [V ,[ inf
qQ
[, q[. (3.1.5)
Lemma 3.2. For any , 1
2
(C),
inf
qQ
[, q[ _ [[[ [[[ div , [[[ [[[ (3.1.6)

and for any , H(C. div),
inf
qQ
[, q[ _ C
D
[div , [. (3.1.7)
Proof. Consider an auxiliary problem
^n
(
div , = 0 in C
n
(
= 0 on I.
where , 1
2
(C) and, therefore, div , H
-1
. The corresponding solution n
(
exists, unique, and satises the relation
_
D
Vn
(
VnJ. =
_
D
( n , Vn) J.. (3.1.8)
From (3.1.8), we nd that
[Vn
(
[ _ [[[ [[[ div , [[[ [[[ := sup
u1
0
u,=0
_
D
(, Vn n) J.
[Vn[
. (3.1.9)
Also, (3.1.8) has the form
_
D
(Vn
(
,) VnJ. =
_
D
nJ.. Vn V
0
.
which means that q := Vn
(
, Q
(
. Therefore,
inf
qQ
[, q[ _ [, q[ = [Vn
(
[ _ [[[ [[[ div , [[[ [[[ .
If , has a square summable divergence, then
[[[ [[[ div , [[[ [[[ = sup
u1
0
u,=0
_
D
(div , )nJ.
[Vn[
_ sup
u1
0
u,=0
[div , [[n[
[Vn[
_ C
T D
[div , [. (3.1.10)
where C
T D
is the constant in the Friedrichs inequality for the domain C.
From (3.1.5) and (3.1.6) it follows that
[V( u)[ _ [V ,[ [[[ [[[ div , [[[ [[[ . V, 1
2
(C). (3.1.11)
However, the right-hand side of this estimate includes the norm [[[ [[[ [[[ [[[ , which is dened
as the supremum over a functional space and, therefore, is not explicitly computable.
A computable estimate follows from (3.1.10) and (3.1.11).
Theorem 3.3 ([276, 277, 282]). For any u
0
V
0
, the upper bound of the error is
given by the estimate
[V( u)[ _ [V ,[ C
T D
[div , [. (3.1.12)
where , is an arbitrary function in H(C. div).
Remark 3.4. From (3.1.12) it follows that the quantity
[V ,[ C[div , [
provides a guaranteed upper bound of the error for any constant C C
T D
. o). If
we set C = o, then we obtain the hypercircle estimate (2.2.4).
Section 3.1 The variational method 49
Henceforth, we denote the right-hand side of (3.1.12) by M
/
(. ,) and call it the
majorant of the deviation from exact solution or the error majorant. The majorant
M
/
(. ,) also depends on the external data D (represented by C and ) but for the
sake of simplicity we do not write them as explicit arguments.
Estimate (3.1.12) is the simplest one among the class of functional a posteriori esti-
mates. However, it possesses all principal features typical of all of them (see Section
3.3). A consequent exposition of the variational approach to a posteriori error estima-
tion is presented in the authors papers [276, 277, 278, 279, 282, 286] and in the book
by P. Neittaanm aki and S. Repin [244].
Remark 3.5. Let be a Galerkin approximation u
h
computed on a nite element
partition T
h
. Set , = Vu
h
. Then (3.1.11) implies the estimate
[V(u u
h
)[ _ [[[ [[[ û
h
[[[ [[[ .
If the right-hand side is estimated from above using the Galerkin orthogonality prop-
erty and the H
1
1
2
projection estimates on the patches, then we arrive at the
explicit residual estimate considered in Section 2.6.1.
A lower bound of the error is given in the theorem below.
Theorem 3.6. For any V
0
,
[V(u )[
2
_ M
2
/
(. n). (3.1.13)
where
M
2
/
(. n) := 2F
(n) [Vn[
2
. (3.1.14)
n is an arbitrary function in V
0
, and F
(n) is the residual functional (cf. 2.6.2).

Proof. From the relation
2(J() J(u)) = [V(u )[
2
.
it follows that
[V(u )[
2
_ 2(J() J( n)).
where n is an arbitrary function in V
0
. Therefore,
[V(u )[
2
_
_
D
_
[Vn[
2
2V Vn
_
J. 2
_
D
nJ..
and we arrive at (3.1.13).
3.2 The method of integral identities
The modern theory of partial differential equations considers integral relations as one
of the major mathematical objects. Integral identities dene generalized solutions of
differential equations and provide the basis for the analysis of their properties (e.g., see
[151, 217, 214, 222]). In this section, we show that two-sided guaranteed bounds of the
error can be derived by transformations of an integral identity. Originally, this modus
operandi was suggested in [287, 283]. It should come as no surprise that the corre-
sponding estimates coincide with those derived by the variational method. As before,
we explain the method with the paradigm of the problem (3.0.1) whose generalized
solution is dened by the integral identity
_
D
Vu VnJ. =
_
D
nJ.. Vn V
0
. (3.2.1)
Upper bound of the error. Let V
0
be a function viewed as an approximate
solution. Insert it in (3.2.1). We have
_
D
V(u ) VnJ. =
_
D
( n V Vn) J.. (3.2.2)
Note that for a vector-valued function , H(C. div) we have (cf. (1.4.14))
_
D
(ndiv , Vn ,) J. = 0.
In view of this relation,
_
D
V(u) VnJ.=
_
D
_
(,V) Vn(div , )n
_
J.. (3.2.3)
Set n = u , then we obtain
[V(u )[
2
_ [V ,[ [V(u )[ [ div ,[ [u [.
Hence,
[V(u )[ _ [V ,[ C
T D
[ div ,[
and the estimate (3.1.12) is derived by another method.
Remark 3.7. If ^ 1
2
(C), then from (3.2.2) we deduce the estimate
[V(u )[ _ C
T D
[^ [. (3.2.4)
Formally, it can be used for sufciently regular approximations. However, it violates
the consistency condition (1.3.3). If a sequence {
k
] of approximate solutions con-
verges to u in

H
1
(C), then the left-hand side of (3.2.4) tends to zero but this may be
not true for the right-hand one.
Section 3.2 The method of integral identities 51
Sometimes, it is required to nd approximate solutions of problems, the right-hand
sides of which are dened by linear functionals of a more general type (associated with
generalized derivatives of 1
2
-functions). Let u be dened by the integral identity
_
D
Vu VnJ. = (. n). Vn V
0
. (3.2.5)
where
(. n) =
_
D
( n t Vn) J.
and t 1
2
(C. R
d
) is a given vector-valued function. In this case,
_
D
V(u ) VnJ. =
_
D
(( div ,)n (, V) Vn t Vn) J..
This relation implies the estimate
[V(u )[ _ [, t V[ C
T D
[ div ,[. (3.2.6)
By (3.2.5) we know that , = Vu t Q
(
. It is easy to see that for , = , the
right-hand side of (3.2.6) is equal to the error.
Let t be represented as the sum of a divergence-free function and gradient of a
scalar-valued function, i.e., t = t
0
V0, where t
0
belongs to the space
S(C) :=
_
j 1
2
(C. R
d
) [ div j = 0. a.e. in C
_
.
and 0 V
0
. Then, the estimate (3.2.6) takes the form
[V(u )[ _ [V( 0) ,[ C
T D
[div , [. (3.2.7)
Remark 3.8. From the computational point of view, it is convenient to square both
parts of (3.1.12) and rewrite the upper bound in the form of a quadratic functional,
namely,
[V(u )[
2
_ M
2
,/
(. ,)
:= (1 )[V ,[
2
_
1
1
_
C
2
T D
[div , [
2
. (3.2.8)
Here, is a positive constant that comes from Youngs inequality (1.4.3).
Lower bound of the error. Lower bounds of the error can also be derived by non-
variational arguments. First, we note that
[V(u )[ = sup
uV
0
_
D
( n V Vn) J.
[Vn[
from which we conclude that a lower bound of the error is given by the quantity
M
/,(V
0h
)
() := sup
u
h
V
0h
_
D
( n
h
V Vn
h
) J.
[Vn
h
[
. (3.2.9)
where V
0h
is a nite-dimensional subspace of V
0
. Finding M
/,(V
0h
)
() requires solv-
ing a nite-dimensional maximization problem (so that it is indeed computable). The
more trial functions are contained in V
0h
the sharper estimate will be computed. How-
ever, (3.2.9) exploits a quotient type functional, which may lead to certain difculties
in maximization procedures.
Another lower bound is obtained with the help of a quadratic functional. We have
sup
uV
0
_
D
(V(u ) Vn) J.
1
2
[Vn[
2
_ sup
r1
2
__
D
(V(u ) t
1
2
[t[
2
) J.
_
=
1
2
[V(u )[
2
.
On the other hand
sup
uV
0
_
D
(V(u ) Vn
1
2
[Vn[
2
) J. _
1
2
[V(u )[
2
.
Thus, we conclude that
[V(u )[
2
= sup
uV
0
_
D
(2V(u ) Vn [Vn[
2
) J.
_ sup
uV
0
_
[Vn[
2
2
_
D
(V Vn n) J.
_
= sup
uV
0
M
2
/
(. n)
3.3 Properties of a posteriori estimates
Structure of the majorant First, we note that both terms on the right-hand side of
(3.1.11) and (3.1.12) have a clear meaning: they represent measures of the errors in
the basic relations
= Vu. (3.3.1)
div = 0. (3.3.2)
Section 3.3 Properties of a posteriori estimates 53
u
y
y
y
.
.
.
1
2
3
Q
Qf
v A
0
B
C
Figure 3.3.1 Geometrical interpretation of M
/
.
which jointly form the equation. In (3.1.11), the equation div = 0 is understood
in a weak sense, whereas in (3.1.12) the respective penalty is given in terms of the
1
2
-norm.
Ageometrical interpretation of the majorant is shown in Fig. 3.3.1. Here Q
0
denotes
the set of vector-valued functions representable as gradients of

H
1
-functions. It is
clear that Vu Q
0
Q
(
. Moreover, for any j Q
(
,
_
D
(j Vu) V J. = 0. V Q
0
.
so that the intersection is orthogonal. We observe that the error (associated with the
interval OA) is estimated from above by the length of two-chained curve ABC. The
length of AB is given by the rst term of M
/
, and the length of BC is estimated by
the second one. This estimate is valid for any ,, but the closer , lies to the exact ux
= Vu the sharper is the estimate. Indeed, it is easy to see that the length of the
two-chained curve associated with ,
2
provides a better approximation of [O[ than
[T[ [TC[. For the curve associated with ,
3
, the approximation is better than for
,
2
. Also, it is clear that the length of the curve cannot be smaller than [O[, so that
the estimate always yields a guaranteed upper bound of the error.
It should be outlined that in (3.1.11), the terms [V ,[ and [ div ,[ have
sharp multipliers. If the multiplier of the rst term is less than 1 and/or the multiplier
of the second one is less than C
T D
, then such a sum cannot be a guaranteed upper
bound of the error. Indeed, apply M
/
(. ,) to the case where = 0 and , = 0.
Then, we arrive at the energy estimate for the generalized solution
[Vu[ _ C
T D
[ [.
Therefore, no constant less than C
T D
can be used as a multiplier of the second term.
Next, set , = Vu. Then (3.1.12) holds as the equality, so that any constant less than 1
cannot be a multiplier of the rst term.
Asymptotic properties of two-sided bounds. Now, our goal is to show that M
/
and M
/
allow one to compute guaranteed two-sided bounds of the error with any
desired accuracy.
Denition 3.9. A sequence {X
k
]
o
k=1
of nite-dimensional subspaces of a Banach
space X is called limit dense in X if for any c > 0 and any X, one can nd
a natural number k
e
such that
inf
m
A
m
[
n
[
A
_ c. Vm > k
e
. (3.3.3)
Proposition 3.10. Let the spaces {Y
k
]
o
k=1
be limit dense in H(C. div). Then
lim
n-o
inf
,
m
Y
m
M
/
(. ,
n
) = [V( u)[ . (3.3.4)
lim
n-o
inf
,
m
Y
m
R
M
2
,/
(. ,
n
) = [V( u)[
2
. (3.3.5)
Proof. The proof of (3.3.4) is straightforward. Take an arbitrary small c > 0 and nd
a respective k
e
> 0 such that [
n
[
div
_ c for some
n
Y
n
if m > k
e
. Then,
M
/
(u. ,
n
) _ M
/
(u.
n
) _ [V(u )[ [
n
[ C
T D
[div (
n
)[
_ [V(u )[ max {1. C
T D
] c.
Analogously, for (3.2.8) we have
inf
,
m
Y
m
R
M
2
,/
(. ,
n
) _ M
2
=,/
(.
n
)
= (1 c)[V
n
[
2
_
1
1
c
_
C
2
T D
[ div
n
[
2
. (3.3.6)
Since
[V
n
[ _ [
n
[ [V( u)[ _ c [V( u)[
we obtain
[V
n
[
2
_ [V( u)[
2
j
1
c. j
1
= c 2[V( u)[.
Also,
[ div
n
[ _ [div (
n
)[ _ [
n
[
div
< c.
Section 3.3 Properties of a posteriori estimates 55
Therefore,
M
2
=,/
(.
n
) _ (1 c) [V( u)[
2
(j
1
C
2
T D
)(c c
2
)
= [V( u)[
2
o(c). (3.3.7)
Now, (3.3.6) and (3.3.7) imply (3.3.5).
For any > 0, the functional M
2
,/
(. ,) is a quadratic functional with respect to
,. Therefore, the quantities (approximate upper bounds of the error)
M
2
k
= inf
,
k
Y
k
R
M
2
,/
(. ,
k
) (3.3.8)
can be found by well-known methods. In view of Proposition 3.10, they form a se-
quence of computable upper bounds such that
M
2
k
[V( u)[
2
as k o. (3.3.9)
Lower bounds given by (3.1.13) possess similar properties.
Proposition 3.11. If the spaces {V
k
]
o
k=1
are limit dense in V
0
, then
lim
n-o
M
2
n
= [V( u)[
2
. (3.3.10)
where
M
2
n
:= sup
u
m
V
m
M
2
/
(. n
n
).
Proof. Take an arbitrary small c > 0 and nd k
e
such that [V(u n
n
)[ _ c for
m > k
e
. Then,
M
2
/
(. n
n
) = [Vn
n
[
2
2
_
D
(V Vn
n
n
n
) J.
= [Vn
n
[
2
2
_
D
V(u ) Vn
n
J.
= [V(u )[
2
[V(n
n
(u ))[
2
.
Hence,
[V(u )[
2
_ M
2
n
_ [V(u )[
2
c
2
. (3.3.11)
This estimate shows that
M
2
n
[V(u )[
2
as m o. (3.3.12)
Thus,
M
2
n
_ [V( u)[
2
_ M
2
k
. (3.3.13)
i.e., the error is bounded from below and above by two sequences of computable num-
bers. The relation (3.3.13) means that conforming approximations of the problem
(3.0.1) are fully controllable (i.e., in principle one can estimate the quality of any con-
forming approximation with any desirable accuracy).
Denition 3.12. Assume that M
k
and M
n
have been computed. Then we have the
quantity
1
(nk)
eff
:=
M
k
M
n
_ 1. (3.3.14)
which provides an idea of the quality of the error estimation. We call 1
(nk)
eff
the com-
putable efciency index.
We note that unlike the efciency indexes comparing error estimates with the norm
of the true error (which is known only in specially selected test problems), the quantity
1
(nk)
eff
is indeed computable. From (3.3.10) and (3.3.12) it follows that
1
(nk)
eff
1 as m. k o. (3.3.15)
The estimate (3.3.15) shows that the error bounds M
n
and M
k
are asymptotically
exact.
Properties of the minimizer. It is easy to prove that the exact lower bound of
M
/
(. ,) (and of M
,/
(. ,)) with respect to , is attained on a certain element of
H(C. div). Indeed, for any V
0
(and any > 0) the majorant is convex, continu-
ous, and coercive on H(C. div). By known results in the calculus of variations (e.g.,
see [121]), we conclude that a minimizer ,() exists. Since M
,/
(. ,) is a quadratic
functional, the corresponding minimizer ,(. ) is unique (in this case, it depends on
).
Lemma 3.13. Let , H(C. div) be such that
M
/
(. ,) = inf
,1(D,div)
M
/
(. ,). (3.3.16)
There exists n V
0
such that , = V n.
Proof. For any ,
0
S(C) we have
[V ,[ C
T D
[div , [ _ [V ,
0
,[ C
T D
[div , [.
Section 3.4 Two-sided bounds in combined norms 57
From the above we conclude that for any ,
0
,
_
D
, ,
0
J.
1
2
[,
0
[
2
_ 0.
This inequality holds if and only if
_
D
, ,
0
J. = 0. V,
0
S(C). (3.3.17)
Recall that , 1
2
(C. R
d
) admits the decomposition , = V n t
0
, where n V
0
and t
0
is a solenoidal eld. Set ,
0
= t
0
. From (3.3.17), it follows that [t
0
[ = 0.
Thus, , = V n.
The minimizer of M
,/
(. ,) has a similar property. We leave proving this fact for
the reader.
3.4 Two-sided bounds in combined norms
In the so-called mixed formulations, the solution of a boundary value problem is de-
ned as a pair of functions. For (3.0.1) it is the saddle point (u. ) of the Lagrangian
1 (cf. (3.1.1)(3.1.2)). The majorant M
/
(. q) also considers and q as independent
functions. Therefore, it is natural to measure the respective error in terms of combined
(primal-dual) norms of the product space
W := V
0
H(C. div).
for which we introduce the norm
[(. ,)[
W
:= [V[ [,[ [div ,[ = [V[ [,[
div
.
Two other equivalent norms are as follows:
[(. ,)[
(1)
W
:= [V[ [,[ C
T D
[div ,[.
[(. ,)[
(2)
W
:=
_
[V[
2
[,[
2
[div ,[
2
_
1{2
.
It is easy to see that
;
1
[(. ,)[
W
_ [(. ,)[
(1)
W
_ ;
2
[(. ,)[
W
. (3.4.1)
1
_
3
[(. ,)[
W
_ [(. ,)[
(2)
W
_ [(. ,)[
W
. (3.4.2)
where ;
1
= min{1. C
T D
] and ;
2
= max{1. C
T D
].
Let us show that the majorant M
/
(. ,) is equivalent to the error in the combined
norm [(u . ,)[
(1)
W
. Since
[ ,[ = [Vu ,[ _ [V(u )[ [V ,[
and [div ( ,)[ = [div , [, we nd that
[(u . ,)[
(1)
W
:= [V(u )[ [ ,[ C
T D
[div , [
_ 2[V(u )[ [V ,[ C
T D
[div , [
_ 3M
/
(. ,).
On the other hand,
M
/
(. ,) _ [V( u)[ [ ,[ C
T D
[div , [. (3.4.3)
Thus, we note that the following two-sided estimate holds:
M
/
(. ,) _ [(u . ,)[
(1)
W
_ 3M
/
(. ,). (3.4.4)
By (3.4.4) we conclude that M
/
is an efcient and reliable measure of the error in the
combined norm [(u . ,)[
(1)
W
.
In view of (3.4.1) and (3.4.2), the majorant is also equivalent to two other combined
norms, namely,
1
;
2
M
/
(. ,) _ [(u . ,)[
W
_
3
;
1
M
/
(. ,). (3.4.5)
1
_
3;
2
M
/
(. ,) _ [(u . ,)[
(2)
W
_
3
;
1
M
/
(. ,). (3.4.6)
Also, we can dene lower and upper bounds for the norm [(u . ,)[
W
with the
help of the functionals
M
Z
(. ,) = 3[V ,[ (1 2C
T D
) [div , [
and
M
Z
(. ,) := [V ,[ [div , [.
which consist of the same terms as those in M
/
but with different weights. We have
[(u . ,)[
W
:= [V(u )[ [ ,[ [div , [
_ 2[V(u )[ [V ,[ [div , [
_ 3[V ,[ (1 2C
T D
)[div , [ =: M
Z
(. ,).
Hence, we nd that
M
Z
(. ,) _ [(u . ,)[
W
_ M
Z
(. ,). (3.4.7)
Section 3.5 Modications of estimates 59
Similarly,
;
1
M
Z
(. ,) _ [(u . ,)[
(1)
W
_ ;
2
M
Z
(. ,). (3.4.8)
1
_
3
M
Z
(. ,) _ [(u . ,)[
(2)
W
_ M
Z
(. ,). (3.4.9)
Finally, we note that
M
/
(. ) = [V(u )[. (3.4.10)
M
/
(u. ,) = [, Vu[ C
T D
[div (, )[. (3.4.11)
Therefore,
[(u . ,)[
(1)
W
:= M
/
(. ) M
/
(u. ,). (3.4.12)
3.5 Modications of estimates
3.5.1 Galerkin approximations
Let V
0h
be a nite-dimensional subspace of V
0
. The Galerkin approximation u
h
satis-
es the orthogonality relation (cf. (2.6.6))
_
D
V(u u
h
) V
h
J. = 0. V
h
V
0h
.
Therefore,
[V(uu
h
)[
2
=
_
D
(Vu
h
,V
h
) V(uu
h
) J.
_
D
(div , )(uu
h
) J.
and we nd that
[V(u u
h
)[ _ [, Vu
h
V
h
[ C
T D
[ div ,[. (3.5.1)
Set
h
= u
h
n
h
, where n
h
is an arbitrary function in V
0h
. Then (3.5.1) has the form
[V(u u
h
)[ _ [, Vn
h
[ C
T D
[ div ,[. (3.5.2)
[V(u u
h
)[ _ C
T D
[ div ,[ inf
u
h
V
0h
[, Vn
h
[. (3.5.3)
Analogously, we obtain
[V(u u
h
)[ _ [[[ [[[ div , [[[ [[[ inf
u
h
V
0h
[, Vn
h
[. (3.5.4)
Note that the projection error estimate (C ea lemma) follows from (3.5.3) (and (3.5.4))
if we set , = Vu. In this case, the rst term vanishes and we arrive at the well-known
projection estimate
[V(u u
h
)[ _ inf
u
h
V
0h
[V(u n
h
)[.
3.5.2 Advanced forms of error bounds
In view of (3.2.3), we have
[V(u )[
2
=
_
D
(, V) V(u ) J.
_
D
r(,)(u ) J.. (3.5.5)
where
r(,) := div ,.
If , is properly selected (e.g., with the help of a post-processing procedure that gen-
erates a function similar to ), then , V ~ V(u ) and, therefore, the quantity
[, V[[V(u)[ does not essentially overestimate the rst term on the right-hand
side of (3.5.5). However, the quantity [r(,)[[u [ may essentially exceed the inte-
gral
_
D
r(,)(u ) J.. Thus, the second term of (3.1.12) may be larger than the rst
one.
One can try to improve the estimate as follows. Take a function 0 such that
0 V
0
:=
0 V
0
[ ^0 1
2
(C)
_
.
Since u vanishes at the boundary we observe that
_
D
( u)^0 J. =
_
D
V0 V(u ) J. =
_
D
( 0 V V0) J. = F
(0).
Note that F
(0) is easily computable and F

u
(0) = 0 for any 0 = 0. Now, we arrive
at the identity
[V(u )[
2
=
_
D
((r(,) ^0)(u ) J. (, V) V(u )) J. F
(0).
[V(u )[
2
_ M(. 0. ,)[V(u )[ F
(0). (3.5.6)
where M(. 0. ,) := C
T D
[r(,) ^0[ [, V[. Hence,
2[V(u )[ _ M(. 0. ,)
_
M
2
(. 0. ,) 4F
(0). (3.5.7)
By (1.4.3) and (3.5.6), we can obtain another upper bound
[V(u )[
2
_
;
2
M
2
(. 0. ,)
1
2;
[V(u )[
2
F
(0). (3.5.8)
which yields (for ; > 1,2) the estimate
[V(u )[
2
_
;
2
2; 1
M
2
(. 0. ,)
2;
2; 1
F
(0). (3.5.9)
Since min
;>1{2
;
2
2;-1
= 1, we observe that (3.5.9) converts to (3.1.12) if 0 = 0.
We note that (3.5.9) has an advantage with respect to (3.1.12) only if [r(y)[ is
sufciently large and ^0 compensates a considerable part of it. For this purpose, we
need to nd a proper function 0. A straightforward way is to take a collection of
linearly independent functions 0
i
V
0
, i = 1. 2. . . . . k, and nd
i
such that
[r(,)
k
i=1
i
^0
i
[ min .
Then, we set 0 =
k
i=1
i
0
i
. If [r(,) ^0[ is essentially smaller than [r(,)[, then
(3.5.7) with 0 = 0
k
supplies a sharper error bound than the basic estimate (3.1.12).
Certainly, the efciency of this method depends on the system of functions {0
i
].
Theoretically, the best choice of 0 is the function 0
,
that satises the equation
^0
,
r(,) = 0 with homogeneous boundary conditions. In this case, we obtain the
estimate
[V(u )[
2
_
;
2
2; 1
[, V[
2
2;
2; 1
F
(0
,
). (3.5.10)
In practice, instead of the unknown function 0
,
(which is the exact solution of a bound-
ary value problem) a certain approximation of it can be used (see 3.6.4).
Another form of the error majorant follows from the relation
[V(u)[
2
= F
(0)
_
D
(r(,)(u)(, V V0)V(u)) J.. (3.5.11)
where 0 is a function from the space V
0
(which is wider than V
0
and admits simpler
approximations). Then, we arrive at (3.5.7) with M(. 0. ,) replaced by
M(. 0. ,) := C
T D
[r(,)[ [, V0 V[.
If , Q
(
, then we have the estimate
2[V(u )[ _ [, V0 V[
_
[, V0 V[
2
4F
(0). (3.5.12)
which does not contain C
T D
. From (3.5.11) it also follows that
2[V(u )[ _

M(. t
0
. ,)
_

M
2
(. t
0
. ,) 4F
() 4F
(). (3.5.13)
where
M(. t
0
. ,) := C
T D
[r(,)[ [, V t
0
[
and and t
0
are arbitrary functions in V
0
and S(C), respectively.
By (3.5.11) and (1.4.3), we also obtain
[V(u )[
2
_ F
(0)

M(. 0. ,)[V(u )[
_ F
(0)
;
2
M
2
(. 0. ,)
1
2;
[V(u )[
2
.
Hence,
2; 1
2;
[V(u)[
2
_ F
(0)
;
2
M
2
(. 0. ,). (3.5.14)
Note that if 0 = u and , = Vu then

M(. 0. ,) = 0 and
F
(0) =
_
D
((u ) V V(u )) J. = [V(u )[
2
.
It is easy to see that in this case the right-hand side of (3.5.14) coincides with the
left-hand one if ; o.
Remark 3.14. In (3.5.12) and (3.5.14), the function 0 should be selected such that
[, V0 V[ is minimal. For example, one can take 0 = 0
h
/ , where
_
D
V0
h
/ Vn
h
/ J. =
_
D
(, V) Vn
h
/ J.. Vn
h
/ V
0h
/ .
where V
0h
/ is a certain nite dimensional subspace of V
0
. The function in (3.5.13)
can be taken as 0
h
/ .
3.5.3 Decomposition of the domain
Assume that C is decomposed into a set T of subdomains C
i
(in particular, C
i
may
coincide with nite elements) with Lipschitz continuous boundaries, i.e.,
C =
_
i=1,...,1
C
i
. and C
i
C
}
= 0 if i ,= .
It is not difcult to see that
[[[ [[[ r(,) [[[ [[[ := sup
uV
0
_
D
(, Vn n) J.
[Vn[
= sup
uV
0
1
i=1
_
D
i
(, Vn n) J.
[Vn[
= sup
uV
0
1
i=1
_
D
i
r(,) nJ.
[Vn[
_ sup
uV
0
1
i=1
_
D
i
(r(,) fr(,)g
D
i
)nJ.
[Vn[
sup
uV
0
1
i=1
fr(,)g
D
i
_
D
i
nJ.
[Vn[
. (3.5.15)
For the rst term on the right-hand side we have (cf. (1.4.23))
1
i=1
_
D
i
e
r(,)
D
i
nJ. _
1
i=1
[
e
r(,)
D
i
[
D
i
C
1D
i
[Vn[
D
i
_ [Vn[
_
1
i=1
[
e
r(,)
D
i
[
2
D
i
C
2
1D
i
(3.5.16)
and for the second one
1
i=1
fr(,)g
D
i
_
D
i
nJ. _
1
i=1
fr(,)g
D
i
[C
i
[
1{2
[n[
D
i
_
_
1
i=1
fr(,)g
2
D
i
[C
i
[ [n[
D
_ C
T D
[Vn[
_
1
i=1
fr(,)g
2
D
i
[C
i
[. (3.5.17)
From (3.5.15)(3.5.17) we deduce the estimate
[[[ [[[ r(,) [[[ [[[ _
_
1
i=1
[
e
r(,)
D
i
[
2
D
i
C
2
i1
C
T D
_
1
i=1
fr(,)g
2
D
i
[C
i
[. (3.5.18)
Let ,

Q(T ), where
Q(T ) := {, H(C. div) [ fdiv , g

D
i
= 0 VC
i
T ].
Then the second term of (3.5.18) vanishes and we obtain
[[[ [[[ r(,) [[[ [[[ _
_
1
i=1
[r(,)[
2
D
i
C
2
1D
i
. (3.5.19)
This relation infers the estimate
[V(u )[ _ [V ,[
_
1
i=1
[r(,)[
2
D
i
C
2
1D
i
. (3.5.20)
which, instead of C
T D
, involves constants in the Poincar e inequalities associated with
the subdomains C
i
.
Consider a special but important case, where T is a regular simplicial decomposi-
tion T
h
consisting of simplexes T
i
and
j
1
h _ diamT
i
_ j
2
h. Vi = 1. 2. . . . . N.
Let C
max1
= max
i
{C
1D
i
]. For regular triangulations the constant C
max1
is of the
same order as all other constants C
1D
i
, so that without big overestimation we can
replace all these constants by C
max1
. Then, we arrive at the upper bound
[V(u )[ _ [V ,[ C
max1
[div , [. (3.5.21)
where ,

Q(T ).
Remark 3.15. We note that C
max1
can be expressed throughout the constant

C
max1
for a similar simplex the diameter of which is equal to one and the mesh parameter h.
3.5.4 Estimates with partially equilibrated uxes
Assume that we have a vector-valued function ,

(
such that
div ,

(

= 0.
where

is close to in 1
2
-norm.
Set , = ,

(
t
0
, where t
0
S(C). Then div ,

= 0. We use (3.1.12) and
arrive at the estimate
[V(u )[ _ [V t
0
,

(
[ C
T D
[

[. (3.5.22)
In particular, we can set t
0
= curl j, where j is an arbitrary vector-valued function
in H(C. curl). If the value of [

[ is signicantly smaller than the tolerance
level accepted for approximations, then nding a sharp upper bound is reduced to the
problem
min
)1(D,curl)
[t curl j[.
where t = V ,

(
is the given vector-valued function. Using a suitable nite-
dimensional subspace for j (which is constructed with the help of conforming nite
element approximations of H(C. curl)), we nd an upper bound by solving a quadratic
minimization problem.
Partially equilibrated ux , can be used in (3.5.13). We have
2[V(u )[ _

M(. ,

(
)
_

M
2
(. ,

(
) 4F
() 4F
(). (3.5.23)
where
M(. ,

(
) := C
T D
[

[ [,

(
V[.
Here is any function in V
0
. This freedom can be used to minimize right-hand side of
(3.5.23).
Section 3.6 Practical applications 65
3.6 How can one use functional a posteriori estimates in
practical computations?
3.6.1 Post-processing of uxes
Let V
0h
V
0
be a nite-dimensional space. For example, V
0h
may contain piecewise
afne nite element approximations generated by the triangulation T
h
. Assume that
h
V
0h
is an approximate solution computed. In particular,
h
may coincide with the
Galerkin approximation u
h
dened by the relation (2.6.5). Also, it may be any other
approximation, which differs fromu
h
owing to the presence of a roundoff, integration,
or other errors. Using
h
, we nd a rough approximation of the ux
h
:= V
h
1
2
(C. R
d
). (3.6.1)
Generally,
h
does not belong to H(C. div) and we cannot directly substitute , =
h
in (3.1.12). For this reason, it is necessary to regularize
h
by a post-processing op-
erator G
h
: 1
2
(C. R
d
) H(C. div). After that, we obtain a vector-valued function
G
h
h
, which yields an easily computable estimate
[V(u u
h
)[ _ [Vu
h
G
h
h
[ C
T D
[div G
h
h
[ . (3.6.2)
The quality of the upper bound given by (3.6.2) depends on properties of the post-
processing operator used. In Chapter 2, we have discussed thr main classes of post-
processing (gradient averaging) operators. Any of them can be used in (3.6.2).
RaviartThomas elements of the lowest order (which are described and studied in,
e.g., F. Brezzi and M. Fortin [79] and J. E. Roberts and J.-M. Thomas [325]) suggest
one more post-processing operator, which we denote G
RT
. Consider a patch formed by
two elements having a common edge 1
In
(see Fig. 3.6.1). If u
h
is constructed by 1
1
-
approximations, then (Vu
h
)[
T
i
and (Vu
h
)[
T
are constant vectors. Dene the normal

ux on 1
In
as follows:
(, n
In
) [
T
ln
= k
In
(Vu
h
)[
T
i
(1 k
In
)(Vu
h
)[
T
.
where k
In
(0. 1). In the simplest case, we set k
In
= 1,2. Another option (which
takes into account sizes of elements) is
k
In
=
[T
i
[
[T
i
[ [T
}
[
.
For the boundary faces, we use the only one existing ux. Thus, we dene three normal
uxes on three sides of each element. The eld inside is obtained by the standard
1T
0
-extension of normal uxes. As a result, we have a function G
RT
h
H(C. div).
We note that E
RT
:= Vu
h
G
RT
h
is an error indicator generated by the procedure.
If the value of the term [div G
RT
h
[ is too large (in comparison with the term
[Vu
h
G
RT

h
[), then we can apply (3.5.7) or (3.5.9) in order to reduce it with the
k
m
T
i
T
j
l
n
k l
lm
mn
kn
E
ln
ln
Figure 3.6.1 Patch related to 1
In
.
help of 0. However, in general, substituting a post-processed gradient does not give a
very accurate upper bound. Numerical experiments have shown that if G is constructed
with the help of simple patch-averaging on the same mesh, then the upper bound given
by the right-hand side of (3.6.2) is rather coarse. More sophisticated post-processing
procedures usually lead to better estimates.
3.6.2 Runge type estimate
Let u
h
1
. u
h
2
. . . . . u
h
k
. . . . be a sequence of approximations on meshes T
h
k
. Compute
h
k
:= Vu
h
k
and average it by an averaging operator G
h
k
acting on T
h
k
. Then the
accuracy of the approximation u
h
k-1
can be measured by the estimate
[V(u u
h
k-1
)[ _ [Vu
h
k-1
G
h
k
h
k
[ C
T D
_
_
div G
h
k
h
k

_
_
. (3.6.3)
This estimate involves approximate solutions computed on two consequent meshes
T
h
k-1
and T
h
k
. Thus, it follows the same strategy as the Runge indicator. However,
the estimate (3.6.3) is mathematically justied and provides a guaranteed upper bound
for any pair of consequent meshes.
3.6.3 Minimization of the majorant
Minimization of the majorant with respect to y. Another strategy is to nd , by
minimizing the majorant on a certain subspace Y
r
H(C. div). In general, Y
r
may
be constructed using a mesh T
r
that differs from T
h
. Then
[V(u u
h
)[ _ inf
,
:
Y
:
{[Vu
h
,
r
[ C
T D
[div ,
r
[] .
The wider is Y
r
, the sharper upper bound is obtained. A detailed discussion of the
minimization methods and numerical results can be found in [134, 137, 244, 278, 303,
307] and some other publications cited therein.
If we intend to dene ,
r
by minimization of the majorant, then it is preferable to
represent the problem in the quadratic form:
min
>0
min
,
:
Y
:
M
2
,/
(. ,).
where
M
2
,/
(. ,) = (1 )[V ,[
2
_
1
1
_
C
2
T D
[div , [
2
.
Practical computations can be performed by the following minimization algorithm:
Step 1. Set k = 0, ,
r,0
= G
r
(Vu
h
), where G
r
is a post-processing operator dened
for T
r
. In particular, if T
h
= T
r
, then any standard gradient averaging operator (see
Section 2.6.3) on T
h
can be used.
Step 2. Find
k
such that
M
2
k
,/
(. ,
r,k
) = min
>0
M
2
,/
(. ,
r,k
).
Here, we have a simple minimization problem, which is solved analytically.
Step 3. Stop if the quantity M
k
= M
2
k
,/
(. ,
r,k
) is less than the desired accuracy
level (in this case, we guarantee that a sufciently accurate approximate solution
has already been constructed).
Otherwise go to Step 4.
Step 4. Dene ,
r,k1
by the relation
M
2
k
,/
(. ,
r,k1
) = min
,
:
Y
:
M
2
k
,/
(. ,
r
).
Step 5. Set k = k 1 and go to Step 2.
This algorithm generates a sequence M
0
. M
1
. . . . . M
k
of monotonically decreasing
upper bounds of the error. We terminate it if either the desired accuracy is conrmed
or if the difference M
k1
M
k
is considered insignicant (or if we have exceeded
the time limit). In any case, the value M
k
obtained at the very last step provides a
guaranteed upper bound of the error.
Remark 3.16. From Lemma 3.13, it follows that
[V(u )[ = inf
,Y
k
M
/
(. ,).
where
Y
k
:=
_
, H(C. div) [
_
D
,,
0i
J. = 0. for ,
01
. ,
02
. . . . . ,
0k
. ,
0i
S(C)
_
.
If (,) _ 0 is a penalty functional vanishing on Y
k
, then
inf
,1(D,div)
M
/
(. ,) = inf
,1(D,div)
_
M
/
(. ,) (,)
_
.
Analogously, the termcan be used with the squared majorant (3.2.8), which gives
the estimate
[V(u )[
2
_ (1 )[V ,[
2
_
1
1
_
C
2
T D
[div , [
2
(,).
If the majorant is numerically minimized with respect to ,, then such penalized forms
may have certain advantages because they include a stabilization term that penal-
izes deviations from the subspace, in which the exact minimizer lies.
On the construction of T
. One way to construct a nite element subspace of

H(C. div) is to use standard piecewise afne approximations of vector-valued func-
tions. It is well motivated if T
r
coincides with the mesh T
h
and is a nite element
approximation computed on this mesh.
Another natural class of conforming approximations of the space H(C. div) is rep-
resented by RT elements. In Section 3.6.1 we discussed the corresponding operator
G
RT
constructed by averaging of uxes on the edges. If T
r
and T
h
coincide, then an-
other regularization operator (which is used on the second step of the above-described
minimization algorithm) can be constructed as follows. First, we dene , = G
RT
Vu
h
.
Now, we describe a simple minimization procedure that can be used to obtain an
almost equilibrated ux without big computational expenditures. It operates with the
quantities ;
In
= , v
In
, which completely dene a piecewise afne vector-valued
function ,. By (1.4.12), we have
_
T
i
div , J. = ;
Ik
[1
Ik
[ ;
In
[1
In
[ ;
kn
[1
kn
[.
j
i
;
In
J
i
= (div ,)
T
i
. j
i
=
;
kI
[1
kI
[ ;
kn
[1
kn
[
[T
i
[
. J
i
=
[1
In
[
[T
i
[
.
Using analogous relation for T
}
, we obtain
j
}
;
In
J
}
= (div ,)
T
. j
}
=
;
In
[1
In
[ ;
nn
[1
nn
[
[T
}
[
. J
}
=
[1
In
[
[T
}
[
.
Our goal is to select ;
In
in such a way that
_
T
i
_
(div ,)
T
i

_
2
J.
_
T
_
(div ,)
T

_
2
J. min .
Since (div ,)
T
i
and (div ,)
T
are constant on T
i
and T
}
, respectively, we nd the
corresponding value of ;
In
by the relation
;
In
=
j
}
[T
i
[ j
i
[T
}
[ [T
i
[[T
}
[(f g
T
f g
T
i
)
[1
In
[([T
i
[ [T
}
[)
.
Using the same idea, we recompute normal uxes for all edges. At each step of this
procedure the value of [div , [
D
decreases and after several cycles of minimization
we obtain a vector-valued eld, which is equilibrated much better than the original one.
Minimization of the majorant with respect to the two variables v and y. Esti-
mate (3.1.12) implies a new variational statement of the problem (3.0.1):
M
/
(u. ) = inf
V
0
,
,1(D,div)
M
/
(. ,). (3.6.4)
which is generated by the error majorant M
/
.
Another variational statement follows from (3.2.4). Indeed, for any > 0 we have
M
2
,/
(u. ) = inf
V
0
,
,1(D,div)
M
2
,/
(. ,). (3.6.5)
It is easy to note that M
/
(. ,) (and M
,/
(. ,)) equals zero if and only if the
arguments coincide with the exact solution u and the exact ux, respectively. This
means that we have a new variational statements of the problem (3.0.1).
It is worth noting that there is a signicant difference between the primal variational
problem P and problems (3.6.4) and (3.6.5). The value of inf P is unknown an de-
pends on and C. The functional J() may be positive, as well as negative, and the
quantity J() does not indicate the accuracy of . In opposite, the exact lower bound
in (3.6.4) (and (3.6.5)) is known: it is equal to zero regardless of the problem data (
and C). Moreover, the functional M
/
(. ,) gives a guaranteed upper bound of the
error. It vanishes if and only if = u and , = . Thus, the value of M
/
(. ,)
supplies a measure of the quality for the approximations and ,.
In principle, one can use the above-discussed properties and solve the problem by
directly minimizing M
/
(. ,) with respect to both variables and , using two se-
quence of subspaces
{V
hk
] V
0
and {Y
hk
] H(C. div).
For this purpose one can use the methods developed in the theory of least square mixed
methods (e.g., see, J. H. Bramble, R. D. Lazarov, and J. E. Pasciak [72], G. F. Carey and
A. I. Pehlivanov [85], and the references therein). Assume that the nite dimensional
problem
inf
V
hk
,>0,
,Y
hk
,
M
(. ,. . C
T D
. ) = M
(
k
. ,
k
.
k
. C
T D
. ) := c
k
is solved. The quantity c
k
shows the accuracy achieved at the step k. If the subspaces
are limit dense in the respective functional spaces, then it is easy to prove that approx-
imate solutions (
k
. ,
k
) tend to (u. ) and the sequence of numbers c
k
tends to zero.
Sometimes, this method may be rather expensive and it may be more efcient to nd
rst
hk
(using Problem P) and after that ,
hk
. However, if
hk
and ,
hk
are dened
with the help of a mixed method, then the respective c
k
is directly computable by the
majorant.
3.6.4 Error indicators generated by error majorants
The theory considered in this chapter, was focused on getting guaranteed bounds of
approximation errors. In practice, it is also important to have easily computable func-
tions that furnish information on the overall error and adequately reproduce the error
function
[e(.)[ := [V(u )[.
Such functions are called error indicators. In Chapter 2, we discussed some of them
in the context of nite element approximations. Below, we introduce several error
indicators, which are generated by error majorants.
1. Let ,
r
be a vector-valued function found by minimization of M
/
(. ,) with re-
spect to , on a certain nite-dimensional space Y
r
. Then a simple indicator of the
squared error [e(.)[
2
is as follows:
E
1
(. ,
r
) = [j(.)[
2
. where j(.) := ,
r
V. (3.6.6)
Since
[e j[ = [V(u ) ,
r
V[ = [ ,
r
[. (3.6.7)
we see that the indicator E
1
(. ,
r
) is sharp (i.e., the computable function j is close to
[e[), if ,
r
is close to .
Let = u
h
, where u
h
is a nite element approximation computed on T
h
. Assume
that {,
r
k
] is a sequence of uxes computed by minimization of M
/
(. ,) on expand-
ing spaces {Y
r
k
], which are limit dense in H(C. div). By Proposition 3.10 we know
that
M
/
(. ,
r
k
) [V(u )[. (3.6.8)
Hence, the sequence {,
r
k
] is bounded in H(C. div) and a weak limit , of this se-
quence (or its subsequence) exists. Since M
/
(u
h
. ,) is convex and continuous with
respect to ,, we know that
[V(u u
h
)[ = lim
k-o
M
/
(u
h
. ,
r
k
) _ M
/
(u
h
. ,)
= [Vu
h
,[ C
T D
[div, [ _ [V(u u
h
)[. (3.6.9)
[Vu
h
,[ C
T D
[div, [ = [V(uu
h
)[
and, therefore, , minimizes the functional M
/
(u
h
. ,).
If Vu
h
, H(C. div) (which is typical of FEM approximations), then one can prove
that , = Vu. Indeed, by Lemma 3.13, we know that , = V u H(C. div), where
u V
0
. Then,
[V(u
h
u)[ C
T D
[^ u [ = [e[. (3.6.10)
where e = V(u u
h
). On the other hand,
[V(u u)[ _ [V u ,[ C
T D
[div , [
and, therefore,
C
T D
[^ u [ _ [V(u u)[. (3.6.11)
From (3.6.10) and (3.6.11) we conclude that
[e[ _ [V(u u)[ [V(u
h
u)[. (3.6.12)
By the triangle inequality,
[e[ _ [V(u u)[ [V(u
h
u)[. (3.6.13)
and, consequently, (3.6.12) and (3.6.13) result in the relation
[e[ = [V(u u)[ [V(u
h
u)[. (3.6.14)
which implies
_
D
V(u u) V( u u
h
) J. = [V(u u)[[V( u u
h
)[. (3.6.15)
Such a relation is true if (a) V(u u) = 0, (b) V(u
h
u) = 0, or (c)
V( u u
h
) = jV(u u) for some j R (j ,= 0). (3.6.16)
In view of the boundary conditions, the case (a) means that u = u holds. Since
Vu
h
, H(C. div), the case (b) is impossible. From (3.6.16), it follows that
Vu
h
= (1 j)V u jVu H(C. div).
so that if Vu
h
, H(C. div), then this relation does not hold and (c) cannot be true. It
remains to conclude that , = Vu.
Then,
[V(u u
h
)[ = lim
k-o
M
/
(u
h
. ,
r
k
) _ lim
k-o
[Vu
h
,
r
k
[
_ [Vu
h
,[ = [V(u u
h
)[.
so that
[Vu
h
,
r
k
[ [V(u
h
u)[ as k o.
From here, it follows that [,
r
k
[ [Vu[ and, consequently, ,
r
k
tends to Vu in
1
2
(C). Hence, [ ,
r
k
[ 0. By (3.6.7) we then conclude that the indicator
j
k
:= ,
r
k
Vu
h
tends to e as k o.
The indicator E
1
was veried in numerous tests not only for the Poissons equation
but also for diffusion, linear elasticity, Stokes, and Maxwells problems (where analogs
of this indicator were used). Experiments conrmed its efciency and stability with
respect to approximations of different types. Two other indicators discussed below
are less studied numerically, but we believe that they will be also useful in practical
computations.
2. Another error indicator follows from (3.5.10) if 0
,
is replaced by a sufciently
accurate approximation. For example, we can dene 0
,r
as a function in a nite-
dimensional space V
r
that satises the relation
_
D
V0
,r
Vn
r
J. =
_
D
(div , )n
r
J.. Vn
r
V
r
. (3.6.17)
Represent F
(0
,r
) in the form
F
(0
,
) =
_
D
( 0
,r
V V0
,r
) J.
_
D
(V V(0
,r
0
,
) (0
,
0
,r
)) J.
= F
(0
,r
)
_
D
V( u) V(0
,r
0
,
) J..
Assume that is the Galerkin solution u
h
computed on V
h
. Let G
h
and G
r
be averag-
ing operators on V
h
and V
r
, respectively. Then the quantity
E
2
(u
h
. 0
,r
) := min
;>1{2
_
;
2
2;-1
[, Vu
h
[
2
2;
2;-1
_
F
u
h
(0
,r
)
_
D
(Vu
h
G
h
Vu
h
) (V0
,r
G
r
V0
,r
)
_
J.
_
(3.6.18)
is an indicator of the energy norm of the error. If T
r
coincides with T
h
and ,
h
is
computed by a certain post-processing of Vu
h
on V
h
(e.g., by quasi-equilibration),
then the function 0
,h
can be found with the help of the same solver that was used for
nding u
h
. In this case, the expenditures of the error indication are approximately the
same as those required for getting u
h
.
A somewhat different estimate follows from the relation
_
D
V(u ) VnJ. =
_
D
(, V0
,
V) VnJ..
Then,
[V(u )[ _ [, V0
,
V[. (3.6.19)
Let = u
h
. From (3.6.19) we obtain a simple error indicator
[V(u u
h
)[

E
2
(u
h
. ,) = [, V0
,r
Vu
h
[. (3.6.20)
where 0
,r
is an approximation of 0
,
. Another version of the indicator arises if V0
,r
is replaced by a post-processed (e.g., averaged) vector-valued function
[V(u u
h
)[ [, G
r
V0
,r
Vu
h
[. (3.6.21)
Certainly the quality of indicators using approximations of 0
,
depends on ,. In prac-
tice, a suitable , can be found as follows. First, we post-process Vu
h
by a cheap
procedure and obtain , = G
h
Vu
h
H(C. div). If the values of div , are
large in some parts of the domain, then , should be modied to diminish them (exact
equilibration is not required). After that, we solve (3.6.17) and nd 0
,r
.
3. From (3.2.2) it follows that
[V(u)[
2
_ [,V[[V(u )[
,
(u ). (3.6.22)
where
,
(u ) :=
_
D
(div , )(u ) J..
By (3.6.22) we observe that
[V(u )[ _
1
2
[,V[
_
,
(u )
1
4
[, V[
2
. (3.6.23)
Since
,
(u ) =
_
D
(Vu ,) V(u ) J.
= [V(u )[
2
_
D
(, V) V(u ) J. _
1
4
[, V[
2
.
the determinant of (3.6.23) is nonnegative regardless of the sign of
,
. If = u
h
and an advanced approximation Q
r
u
h
is computed by one of the post-processing pro-
cedures that we discussed in Chapter 2 (cf. (2.6.30)), then (3.6.23) shows that the
quantity
E
3
(u
h
. ,) :=
1
2
[,Vu
h
[
_
,
(Q
r
u
h
u
h
)
1
4
[, Vu
h
[
2
(3.6.24)
may also serve as an error indicator. Its quality depends on the choice of , and the
efciency of post-processing provided by Q
r
.
In particular, if , = G
h
Vu
h
, then we arrive at the indicator
E
3
(u
h
. G
h
Vu
h
) := [G
h
Vu
h
Vu
h
[ E
h,r
. (3.6.25)
where
2E
h,r
=
_
_
D
_
4(div G
h
Vu
h
)(Q
r
u
h
u
h
) J. [G
h
Vu
h
Vu
h
[
2
_
J.
_
1{2
[G
h
Vu
h
Vu
h
[.
The rst term of E
3
(u
h
. G
h
Vu
h
) is the standard gradient averaging indicator. The
second term is obtained as a computable approximation of the term
,
(u u
h
).
It is easy to see that the error arising if the exact upper bound (3.6.22) is replaced
by the indicator (3.6.24) depends on the value of
_
D
(div , )(Q
r
u
h
u) J..
If Q
r
u
h
provides a good approximation (in 1
2
-sense) of u and div , is small (in
a weak integral sense), then (3.6.24) will give a correct representation of the error.
Finally, we note that (3.6.22) formally generates a simple indicator for the squared
error norm [V(uu
h
)[
2
:
[G
h
Vu
h
Vu
h
[
2
2
_
D
(div G
h
Vu
h
)(Q
r
u
h
u
h
) J..
However, it is clear that the efciency of such an indicator strongly depends on the
efciency of the averaging operators and deteriorates if G
h
Vu
h
does not properly re-
produce Vu and the residual div G
h
Vu
h
is not small enough.
4 Linear elliptic problems
4.1 Two-sided estimates for stationary diffusion problem
4.1.1 Estimates for problems with mixed boundary conditions
Stationary diffusion problem. First, we consider the problem
div Vu = 0 in C. (4.1.1)
u = u
0
on I
1
. (4.1.2)
n Vu = J on I
2
. (4.1.3)
where C R
d
is a bounded connected domain with Lipschitz continuous boundary
that consists of two measurable nonintersecting parts I
1
and I
2
. We assume that
meas
d-1
{I
1
] > 0, u
0
H
1
(C), and the matrix = {a
i}
] is symmetric and satises
the relation
c
2
1
[[
2
_ _ c
2
2
[[
2
. V R
d
. (4.1.4)
Also, we assume that 1
2
(C) and J 1
2
(I
2
). Let
u
0
V
0
:= {n = u
0
n
0
[ n
0
V
0
(C)].
where
V
0
:= {n H
1
(C) [ n = 0 on I
1
].
A generalized solution u of (4.1.1)(4.1.3) is a function in u
0
V
0
that meets the
integral identity
_
D
Vu VnJ. =
_
D
nJ.
_
I
2
JnJs. Vn V
0
(C). (4.1.5)
It is well known that a generalized solution dened by (4.1.5) exists and unique.
Upper estimates of the error norm. Now, we use (4.1.5) in order to obtain an es-
timate of the difference between u and an approximation u
0
V
0
in the energy
norm [[V(u )[[, where
[[,[[
2
:=
_
D
, , J..
Also, we use another norm
[[,[[
2
+
:=
_
D
-1
, , J..
76 Chapter 4 Linear elliptic problems
In view of (4.1.4), these norms are equivalent to the natural norm of the space
Y := 1
2
(C. R
d
).
Let , be a vector-valued function from the set
H
I
2
(C. div) :=
, H(C. div) [ , n 1
2
(I
2
)
_
.
Then,
_
D
((div ,)n Vn ,) J. =
_
I
2
(, n)nJs. Vn V
0
. (4.1.6)
By (4.1.5) and (4.1.6), we nd that
_
D
V(u ) VnJ. =
_
D
( div ,)nJ.
_
D
(, V) VnJ.
_
I
2
(J , n)nJs. (4.1.7)
By recalling the Friedrichs type inequality
[n[ _ C
T I
1
[Vn[. Vn V
0
. (4.1.8)
and the trace inequality
[n[
I
2
_ C
TI
2
[Vn[. Vn V
0
. (4.1.9)
we conclude that there exists a positive constant z
1
(C. I
2
) such that
z
2
1
(C. I
2
) = inf
uV
0
[[Vn[[
2
[n[
2
[n[
2
I
2
. (4.1.10)
Since
_
D
(V ,) VnJ. _ [[V ,[[
+
[[Vn[[
and
_
D
( div ,)nJ.
_
I
2
(J , n)nJs
_
_
[ div ,[
2
[J , n[
2
I
2
_
1{2
C[[Vn[[.
we arrive at the estimate
[[V(u )[[ _ [[V ,[[
+
C
_
[ div ,[
2
[J , n[
2
I
2
. (4.1.11)
Section 4.1 Two-sided estimates for stationary diffusion problem 77
In (4.1.11), C is any constant greater than z
-1
1
(C. I
2
). The right-hand side of (4.1.11)
represents a computable error majorant M
DF
(. ,) for the diffusion problem (4.1.1)
(4.1.3).
A somewhat different form of the error bound follows from the estimate
_
D
( div ,)nJ.
_
I
2
(J , n)nJs
_ C
T I
1
[ div ,[[Vn[ C
TI
2
[J , n[
I
2
[Vn[.
Applying it to (4.1.7), we obtain
[[V(u )[[ _ [[V ,[[
+
1
c
1
_
C
T I
1
[ div ,[
C
TI
2
[J , n[
I
2
_
. (4.1.12)
Lower estimates of the error norm. A lower bound of [[V(u )[[ can be derived
as follows. Note that
1
2
[[V(u )[[
2
= sup
,1
2
(D,R
o
)
_
D
_
V(u ) ,
1
2
, ,
_
J.
_ sup
uV
0
_
D
_
V(u ) Vn
1
2
Vn Vn
_
J.
_
_
D
_
V(u ) V(u )
1
2
V(u ) V(u )
_
J.
=
1
2
[[V(u )[[
2
and we conclude that
1
2
[[V(u )[[
2
= sup
uV
0
_
D
_
V(u ) Vn
1
2
Vn Vn
_
J.
= sup
uV
0
_
F
(n)
_
D
1
2
Vn VnJ.
_
.
where
F
(n) =
_
D
( n V Vn) J.
_
I
2
JnJs.
It is easy to see that the lower bound given by the left-hand side of the above estimate
is sharp (set n = u ). Thus, the minorant is dened by the relation
M
2
DF
(. n) := 2F
(n)
_
D
Vn VnJ.. (4.1.13)
4.1.2 Modications of estimates
Quadratic form of the error majorant. Square both parts of (4.1.11) and apply
Youngs inequality. We obtain
[[V(u )[[
2
_ M
2
DF
(. ,. )
:= (1 )[[V ,[[
2
+
C
2
_
[ div ,[
2
[J , n[
2
I
2
_
. (4.1.14)
where is an arbitrary positive number. For any > 0, the right-hand side of
(4.1.14) is a quadratic functional with respect to , the minimization of which on a
nite-dimensional subspace is equivalent to solving a system of linear simultaneous
equations. This estimate is exact in the sense that, by choosing proper and ,, one
can make the right-hand side as arbitrarily close to the left-hand side.
Galerkin approximations. Let = u
h
u
0
V
0h
, where
_
D
Vu
h
Vn
h
J. =
_
D
n
h
J.
_
I
2
Jn
h
Js. Vn
h
V
0h
V
0
.
In this case,
_
D
V(u
h
u) Vn
h
J. =
_
D
V(u
h
u) Vn
h
J. = 0. (4.1.15)
With the help of (4.1.15) we rewrite (4.1.7) in the form
_
D
V(u u
h
) V(u u
h
) J. =
_
D
( div ,)(u u
h
) J.
_
D
(, Vu
h
Vn
h
) V(u u
h
) J.
_
I
2
(J , n)(u u
h
) Js. (4.1.16)
From (4.1.16), we deduce the estimate
[[V(u u
h
)[[ _ [[, Vu
h
Vn
h
[[
+
C
_
[ div ,[
2
[J , n[
2
I
2
_
1{2
. (4.1.17)
Here, n
h
0h
, which can be used to reduce the value of the
rst term of the error majorant.
For , = Vu, the estimate (4.1.17) reads
[[V(u u
h
)[[ _ [[V(u u
h
n
h
)[[
+
= [[V(u u
h
n
h
)[[. (4.1.18)
Section 4.1 Two-sided estimates for stationary diffusion problem 79
Since n
h
0h
, we can take it as n
h
= u
h
n
h
, where
n
h
V
0h
. Then (4.1.18) implies the projection error estimate
[V(u u
h
)[ _
c
2
c
1
inf
u
h
V
0h
[V(u n
h
)[. (4.1.19)
An advanced form of the error majorant. In view of (4.1.7),
[[V(u )[[
2
=
_
D
(, V) V(u ) J.
_
D
r
:
(,)(u ) J.
_
I
2
r
T
2
(,)(u ) Js. (4.1.20)
where
r
:
(,) := div , and r
T
2
(,) = J , n.
Let 0 V
0
(C) be a function such that
V0 H
I
2
(C. div).
Since u vanishes on I
1
, we note that
_
D
div V0(u ) J. =
_
D
V0 V(u ) J.
_
I
2
(V0 n)(u ) Js
=
_
I
2
(V0 n)(u ) Js F
(0). (4.1.21)
where the functional F
is dened above.
By (4.1.21) we rewrite (4.1.20) in the form
[[V(u )[[
2
=
_
D
(, V) V(u ) J.
_
D
(r
:
(,) div V0)(u) J.
_
I
2
(r
T
2
(,) V0 n)(u) Js F
(0). (4.1.22)
Hence,
[[V(u )[[
2
_ M
2
11
(. ,. 0. ;). (4.1.23)
where ; is a positive constant and
M
2
11
(. ,. 0. ;) :=
2;
2; 1
F
(0)
;
2
2; 1
_
[[V ,[[
+
C
_
[r
:
(,) div V0[
2
[r
T
2
(,) V0 n[
2
I
2
_
1{2
_
2
represents an advanced form of the error majorant for the linear diffusion problem with
mixed boundary conditions.
Decomposition of . Assume that

C =
_
1
i=1

C
i
, where C
i
are nonintersecting
domains with Lipschitz continuous boundaries.
Assume that
fr
:
(,) div V0g
D
i
= 0. i = 1. 2. . . . . N.
r
T
2
(,) V0 n = 0. on I
2
.
In this case, the equation (4.1.22) yields an upper bound that contains constants C
1D
i
,
i = 1. 2. . . . . N, instead of C (we recall that C
1D
i
is the constant in the Poincar e
inequality for C
i
).
Indeed, we have
[[V(u )[[
2
=
_
D
(, V) V(u ) J.
i=1
C
1D
i
[r
:
(,) div V0[
D
i
[V(u )[
D
i
F
(0)
_ [[V ,[[
+
[[V(u )[[ F
(0)
_
1
i=1
C
2
1D
i
[r
:
(,) div V0[
2
D
i
[V(u )[.
Let and be positive numbers such that < 2. Then, we arrive at the estimate
(2 )[[V(u )[[
2
_
1
[[V ,[[
2
+
1
c
2
1
i=1
C
2
1D
i
[r
:
(,) div V0[
2
D
i
F
(0). (4.1.24)
4.1.3 Estimates for problems with Neumann boundary condition
Let I
1
= 0 and I
2
= I. In this case, the energy space is
V :=
_
n H
1
(C)
_
D
nJ. = 0
_
and the equilibrium condition
_
D
J.
_
I
J Js = 0 (4.1.25)
must be satised. The solution is dened by the integral identity
_
D
Vu VnJ.
_
D
nJ.
_
I
JnJs = 0. Vn

V . (4.1.26)
Section 4.2 The stationary reaction-diffusion problem 81
Let

V be an approximate solution and , H(C. div). A transformation of
(4.1.26) yields
_
D
V(u ) V(u ) J. =
_
D
( div ,)(u ) J.
_
D
(, V) V(u ) J.
_
I
(J , n)(u ) Js.
Since u

V , we nd that
_
D
( div ,)(u ) J. =
_
D
( div , f div ,g
D
)(u ) J.
=
_
D
E
( div ,)(u ) J.
_ [
D
div ,
D
[C
1D
[V(u )[.
Another term on the right-hand side is estimated by the trace inequality. We have
[[V(u )[[
2
_ [[, V[[
+
[[V(u )[[ (C
1D
[
D
div ,[
C
TI
[J , n[
I
)[V(u )[
_ c
-1
1
(C
1D
[
D
div ,[ C
TI
[J , n[
I
)[[V(u )[[ (4.1.27)
and arrive at the estimate
[[V(u )[[
_ [[V ,[[
+
c
-1
1
(C
1D
[
D
div ,[ C
TI
[J , n[
I
). (4.1.28)
4.2 The stationary reaction-diffusion problem
Diffusion problem with mixed DirichletNeumann boundary conditions. The
reaction-diffusion problem is represented by the system
div
2
u = in C. (4.2.1)
= Vu in C. (4.2.2)
u = u
0
on I. (4.2.3)
We assume that is a nonnegative function of . and the matrix satises (4.1.4).
Now V
0
=

H
1
(C) and the generalized solution u u
0
V
0
of (4.2.1)(4.2.3) is
dened by the integral identity
_
D
(Vu Vn
2
un) J. =
_
D
nJ.. n V
0
. (4.2.4)
It minimizes the functional
1(n) =
_
D
_
1
2
Vn Vn

2
2
[n[
2
n
_
J. (4.2.5)
on the set u
0
V
0
.
Let u
0
V
0
. Then (4.2.4) implies the relation
_
D
(V(u ) Vn
2
(u ) n) J. =
_
D
( n
2
n V Vn) J.. (4.2.6)
which holds for any n V
0
. Since n vanishes on the boundary, we rewrite (4.2.6) in
the form
_
D
(V(u ) Vn
2
(u ) n) J.
=
_
D
((
2
div ,)n (, V) Vn) J.. (4.2.7)
where , is a vector-valued function in the space H(C. div). The second term on the
right-hand side is estimated as in Section 4.1, but the rst one has two different upper
bounds:
_
D
(
2
div ,) nJ. _ [
1
(
2
div ,)[[n[. (4.2.8)
_
D
(
2
div ,) nJ. _ C[
2
div ,[[[Vn[[. (4.2.9)
where C is a constant in the inequality
[n[ _ C[[Vn[[. Vn V
0
. (4.2.10)
Note that C _ c
-1
1
C
T D
. By (4.2.7) and (4.2.8), we deduce the estimate
[u |[ _
_
[[V ,[[
2
+
[
1
r
:
(. ,)[
2
_
1{2
. (4.2.11)
where
[n|[
2
:= [[Vn[[
2
[ n[
2
is the energy norm related to the problem and
r
:
(. ,) :=
2
div ,.
From (4.2.7) and (4.2.9), we obtain another estimate
[[V(u )[[ _ [[V ,[[
+
C[r
:
(. ,)[. (4.2.12)
Let us denote the majorants in (4.2.11) and (4.2.12) by M
RD1
(. ,) and M
RD0
(. ,),
respectively (the motivation of this will be given below). Note that the majorant
M
RD1
(. ,) was earlier derived by the variational method (see [282]). It is easy to
show that
inf
,1(D,div)
M
RD1
(. ,) = [u |[.
This fact follows from the relation
M
2
RD1
(. ) = [[V( u)[[
2
+
[
1
r
:
(. )[
2
= [[V( u)[[
2
_
D
(u )
2
J. = [u |[
2
.
However, M
RD1
(. ,) has an essential drawback: if is small, then the second term
has a large multiplier that makes the whole estimate sensitive to the residual r
:
(. ,).
In the problems where is small (or zero) in one part of C and large in the other one,
the majorant M
RD1
(. ,) may lead to a considerable overestimation of the error. On
the contrary, M
RD0
(. ,) is robust with respect to small but it may have an inherent
gap between the left-hand and right-hand sides of (4.2.12).
An advanced form of the error majorant. In order to overcome the above difcul-
ties and to obtain an estimate that possesses positive features of the above estimates we
apply another modus operandi for the deviation of an upper bound of u suggested
in [305].
Represent the rst term on the right-hand side of (4.2.7) in the form
_
D
r
:
(. ,)nJ. =
_
D
r
:
(. ,)nJ.
_
D
(1 )r
:
(. ,)nJ..
where
1
o
0,1j
(C) := { 1
o
(C) [ 0 _ (.) _ 1].
Then, we have
_
D
r
:
(. ,)nJ.
_
_
_
_
_
r
:
(. ,)
_
_
_
_
[n[ C[(1 )r
:
(. ,)[ [[Vn[[.
Setting n = u , we arrive at the estimate
[u [|
2
_ (C[(1 )r
:
(. ,)[ [[V ,[[
+
)
2
_
_
_
_
r
:
(. ,)
_
_
_
_
2
=: M
2
DF
(. ,). (4.2.13)
It is easy to see that (4.2.11) and (4.2.12) are special cases of (4.2.13).
Also, we can represent (4.2.13) in the form
[u |[
2
_ C
2
(1 )[(1 )r
:
(. ,)[
2
[[V ,[[
2
+
[
r
:
(. ,)[
2
. (4.2.14)
where is an arbitrary positive number.
The minimization of the right-hand side of (4.2.14) with respect to is reduced to
the following auxiliary variational problem: Find 1
o
0,1j
(C) such that
g() = inf
1
c
0,1|
(D)
g(). g() :=
_
D
_
2
S(.) (1 )
2
T(.)
_
J.. (4.2.15)
where S and T are nonnegative integrable functions that do not vanish simultaneously.
It is easy to nd that for almost all .,
(.) =
T
S T
0. 1|. g() =
ST
S T
.
In our case, S = j
-2
r
2
:
(. ,) and T = C
2
(1 )r
2
:
(. ,). Therefore, we obtain
[u |[
2
_
_
D
C
2
(1 )
C
2
2
(1 ) 1
r
2
:
(. ,) J.
1
[[V ,[[
2
+
=: M
2
RD
(. ,. ). (4.2.16)
Since
M
2
RD
(. . ) =
_
D
_
C
2
(1 )
C
2
2
(1 ) 1
4
( u)
2
_
J.
1
[[V( u)[[
2
.
we nd that
inf
,1(D,div), >0
M
2
RD
(. ,. ) _ inf
>0
M
2
RD
(. . ) = [u |[
2
.
Therefore, (4.2.16) has no gap. At the same time the structure of the rst term of
(4.2.16) is such that it is not sensitive to small values of . Moreover, if = 0, then
[[V(u )[[
2
_ C
2
(1)[ div ,[
2
[[V ,[[
2
+
(4.2.17)
for the diffusion problem without convection.
If = const then the value of that minimizes the right-hand side of (4.2.16) can
be found analytically. At the end of Section 6.4, this question is discussed with the
paradigm of the generalized Stokes problem where a similar functional arises.
Diffusion problem with mixed DirichletRobin boundary conditions. Diffusion
problems are often considered with the Robin boundary condition
n Vu k(.)u = 0 on I
2
. (4.2.18)
where k(.) _ 0 and k(.) , 0. For this case, error estimates can be derived from the
integral identity
_
D
(Vu Vn
2
un) J.
_
I
2
kunJs =
_
D
nJ.. Vn V
0
. (4.2.19)
by the method discussed in the previous section.
Let u
0
V
0
. Then (4.1.6) and (4.2.19) imply the relation
_
D
(V(u ) Vn
2
(u ) n) J.
_
I
2
k(u )nJs
=
_
D
( n
2
n V Vn) J.
_
I
2
k Js
=
_
D
r
:
(. ,)nJ.
_
D
(, V) VnJ.
_
I
2
(, n k)nJs. (4.2.20)
Set n = u ; by (4.2.20) we have the estimate
[u |[ _
_
[[V ,[[
2
+
_
_
_
_
r
:
(. ,)
_
_
_
_
2
_
_
_
_
, n k
k
1{2
_
_
_
_
I
2
_
1{2
. (4.2.21)
where
[n|[
2
:= [[Vn[[
2
[n[
2
[k
1{2
n[
2
I
2
.
Estimate (4.2.21) is the simplest error majorant for the problem with Robin boundary
conditions. In Chapter 7, we deduce it by a different method, using general results of
the variational approach.
In view of the estimates
_
D
r
:
(. ,)nJ. _ C[r
:
(. ,)[[[Vn[[.
_
I
2
(, n k)nJs _ c
-1
1
C
TI
2
[, n k[[[Vn[[.
we obtain another upper bound:
[u|[ _ [[V,[[
+
C[r
:
(. ,)[
C
TT
2
c
1
[, nk[
I
2
. (4.2.22)
which is not sensitive to small values of and k.
By combining the methods used for the derivation of (4.2.21) and (4.2.22), one
can deduce a more general estimate (an analog of (4.2.13)) for the reaction-diffusion
problem with the Robin boundary condition.
Remark 4.1. For the problem û zu = 0 in C R
2
with boundary condition
Vu nku = 0 in which k does not depend on ., it is known that among all domains
with given area, the circle yields the lowest principle eigenvalue (e.g., see [65]) This
fact can be used for getting sharp estimates of the constant in the error majorant if
I = I
2
.
A problem generated by the SturmLiouville operator. Consider the boundary
value problem
(a(.)u
t
)
t
b(.)u = (.) . C := (
1
.
2
). (4.2.23)
u(
1
) = u
1
. u(
2
) = u
2
.
2
>
1
. (4.2.24)
generated by the SturmLiouville operator with bounded coefcients a and b. Also,
we assume that
a(.) _ a
0
> 0. b(.) _ 0. and 1
2
(C).
It can be viewed as the 1D form of (4.2.1)(4.2.3). In this case,
[[u [[
2
=
_

2
1
a(.)(u )
2
J. _ a
0
[u [
2
.
[[u [[
2
+
=
_

2
1
a
-1
(.)(u )
2
J..
If we set C =

2
-
1
o
0
t
, =
_
b, and r
:
(. ,) = ,
t
b, , then (4.2.14) provides an
upper bound of the error.
If the boundary conditions (4.2.24) are represented in a more general form, namely
u
t
c
i
u = u
i
at
i
. i = 1. 2. (4.2.25)
then the corresponding estimate is obtained by the same arguments we used for the
diffusion problem with Robin type boundary condition. We recommend the reader to
derive it as an exercise.
Decomposition of . Assume that

C =
_
1
i=1

C
i
, where C
i
are nonintersecting
domains with Lipschitz continuous boundaries and , H(C. div) is balanced in the
subdomains, i.e.,
fr
:
(. ,)g
D
i
= 0. i = 1. 2. . . . . N. (4.2.26)
, n k = 0. on I
2
. (4.2.27)
Section 4.3 Diffusion problems with convective term 87
Then instead of (4.2.22), we can use the estimate (which follows from (4.2.20))
[u|[ _ [[V,[[
+
1
c
1
_
1
i=1
C
2
1D
i
[r
:
(. ,)[
2
D
i
. (4.2.28)
where C
1D
i
are constants in the Poincar e inequalities associated with C
i
.
If (4.2.26) does not hold, then we can apply the same method as in Section 3.5.3.
We have
_
D
r
:
(. ,)nJ. =
1
i=1
_
D
i
n(
B
r
:
(. ,))
D
i
J.
1
i=1
fr
:
(. ,)g
D
i
_
D
i
nJ.
_
1
i=1
_
D
i
n(
B
r
:
(. ,))
D
i
J.
1
i=1
1
i
fr
:
(. ,)g
D
i
[C
i
[
1{2
[n[
D
i
.
where
i
= min
xD
i
(.). Note that
1
i=1
_
D
i
n(
B
r
:
(. ,))
D
i
J. _ 1
1
[[Vn[[.
1
i=1
1
i
fr
:
(. ,)g
D
i
[C
i
[
1{2
[n[
D
i
_ 1
2
[n[.
where
1
1
=
1
c
1
_
1
i=1
C
2
1D
i
[(
B
r
:
(. ,))
D
i
[
2
D
i
and 1
2
=
_
1
i=1
[C
i
[
2
i
fr
:
(. ,)g
2
D
i
.
We set n = u and use (4.2.20), which implies the estimate
[u |[
2
_ ([[V ,[[
+
1
1
)[[V(u )[[ 1
2
[(u )[.
Hence, we conclude that
[u |[
2
_ ([[V ,[[
+
1
1
)
2
1
2
2
. (4.2.29)
We outline that this estimate contains only Poincar e constants, so that if C can be
decomposed into a set of simple subdomains (for which C
1D
i
are known) then the
respective upper bound of the error is easily computable.
4.3 Diffusion problems with convective term
In this section, we analyze diffusion problems with convective term. First, we consider
the simplest convection-diffusion problem with homogeneous Dirichlet boundary con-
ditions. This problem is used to discuss transparently modications of the method,
which are due to the presence of the convective term. Subsequently, a more general
class of problems is analyzed.
4.3.1 The stationary convection-diffusion problem
Consider the simplest model involving the convective term.
div Vu a Vu = in C. (4.3.1)
u = 0 on I. (4.3.2)
Here a is a given vector-valued function satisfying the conditions
a 1
o
(C. R
d
). div a 1
o
(C). div a _ 0. (4.3.3)
The generalized solution u V
0
meets the integral identity
_
D
(Vu Vn (a Vu)n) J. =
_
D
nJ.. Vn V
0
. (4.3.4)
Again, an upper bound of the error is derived by transformations of the integral identity
that denes the solution. As before, we take V
0
and insert it into (4.3.4), which
yields the relation
_
D
(V(u ) Vn (a V(u ))n) J.
=
_
D
( n V Vn (a V)n) J.. (4.3.5)
Since (cf. (1.4.10))
div ((u ) a) = (u ) div a a V(u ) (4.3.6)
and n vanishes at the boundary, we have
_
D
na V(u ) J. =
_
D
_
div ((u ) a) (div a)(u )
_
nJ.
=
_
D
_
(u )a Vn (div a)(u ) n
_
J..
From here, we obtain
1
2
_
D
(div a)(u )
2
J. =
_
D
(u )a V(u ) J.. (4.3.7)
Set n = u and rearrange (4.3.5), using (4.3.7). We arrive at the relation
[[V(u )[[
2
1
2
_
D
(div a)(u )
2
J.
=
_
D
_
r
:
(. ,) (u ) (, V) V(u )
_
J.. (4.3.8)
where ,(.) is an arbitrary function in H(C. div) and
r
:
(. ,) = a V div ,.
Introduce the norm
[u |[
2
:= [[V(u )[[
2
[(u )[
2
.
where
2
=
1
2
div a _ 0.
If (.) > 0 for almost all . C, then, by the estimate
_
_
1
r
:
(. ,)
_
_
[(u )[ [[, V[[
+
[[V(u )[[
_
_
[[
1
r
:
(. ,)[[
2
[[, V[[
2
+
_
1{2
[u |[.
we deduce the rst error majorant:
[u |[
2
_
_
_
1
r
:
(. ,)
_
_
2
[[, V[[
2
+
=: M
2
CD1
(. ,). (4.3.9)
For small values of and large values of a this estimate may be coarse.
Another estimate that clearly follows from (4.3.8) is
[u |[ _ [[, V[[
+
C[r
:
(. ,)[ =: M
CD2
(. ,). (4.3.10)
If C is decomposed into a collection of subdomains C
i
and fr
:
(. ,)g
D
i
= 0, then
(4.3.8) implies the estimate similar to (4.2.26), namely
[u |[ _ [[, V[[
+
_
1
i
C
2
1D
i
c
2
1
[r
:
(. ,)[
2
D
i
. (4.3.11)
If fr
:
(. ,)g
D
i
,= 0, then repeating the arguments of the previous section we derive an
upper bound analogous to (4.2.29).
Consider a special but important case of the problem where the convection term is
dominant:
cû a Vu = in C. (4.3.12)
u = u
0
on I. (4.3.13)
Here c is a small positive number and (for the sake of simplicity) it is assumed that
I = I
T
. Also, we assume that is a positive constant and [a[ is of the order 1. In
this case,
C =
C
T D
_
c
. = c1.
-1
=
1
c
1.
[[,[[
2
=
_
D
c[,[
2
J. = c[,[
2
. [[,[[
2
+
=
_
D
1
c
[,[
2
J. =
1
c
[,[
2
.
[[, V[[
2
+
=
_
D
_
c[V[
2
c
-1
[,[
2
2, V
_
J. =
_
_
_
_
cV
1
_
c
,
_
_
_
2
.
We use the estimate (4.3.9) that does not include negative powers of c and obtain the
following upper bound of the error:
[u |[
2
:= c[V(u )[
2
2
[u [
2
_ M
2
CD1
(. ,)
=
_
_
_
1
( a V div,)
_
_
_
2
_
_
_
1
_
c
(, cV)
_
_
_
2
. (4.3.14)
Set , = cVu. Then,
div , = cû = a Vu.
and we nd that
E
2
1
() _
1
2
[a V(u )[
2
c[V(u )[
2
. (4.3.15)
where
E
1
() := inf
q1(D,div)
M
CD1
(. q).
By (4.3.14) and (4.3.15), we conclude that
0 _ E
2
1
() [u |[
2
_
1
2
[a V(u )[
2

2
[u [
2
.
Hence,
[u |[
2
_ E
2
1
() _ [u |[
2
[a[
2
2
[V(u )[
2
.
Since [V(u )[
2
_ c
-1
[u |[
2
, we deduce the estimate
[u |[ _ E
1
() _
_
1
[a[
2
2
c
[u |[. (4.3.16)
If = 0, then we use (4.3.10), which shows that
E
2
() := inf
q1(D,div)
M
CD2
(. q) _ C[a V(u )[ [u |[.
where [u |[ =
_
c[V(u )[. Therefore,
0 _ E
2
() [u |[ _ C[a V(u )[ _ C
[a[
_
c
[u |[.
and we nd that
[u |[ _ E
2
() _
_
1
[a[C
_
c
_
[u |[. (4.3.17)
It is not surprising that the quality of the upper bounds in (4.3.16) and (4.3.17) is
negatively affected by c. Moreover, properties of the exact solution u, as well as the
quality of approximations, usually deteriorates if c goes to zero. Such difculties are
predictable and typical for all problems with small parameter.
Assume that = u
h
where u
h
is computed on the nite-dimensional space V
h
and
the error majorant is computed on a nite-dimensional subspace Q
r
that satises the
condition
J
r
Q
r
. [
r
[ [div (
r
)[ _ j
r
. (4.3.18)
where j
r
is a small positive number. Then, we have
[u u
h
|[
2
_ E
2
1r
(u
h
) := inf
q
h
Q
:
M
2
CD1
(u
h
. q
h
) _ M
2
CD1
(u
h
.
r
)
_ 2
_
_
_
_
_
1
_
( a Vu
h
div)
_
_
_
_
2
_
_
_
_
1
_
div(
r
)
_
_
_
_
2
_
_
_
_
1
_
c
( cVu
h
)
_
_
_
_
2
1
c
[
r
[
2
_
_ 2
_
1
[a[
2
c
_
[u u
h
|[
2
2j
2
r
_
1
c
_
. (4.3.19)
If the subspaces V
h
and Q
r
are constructed such that the
r
approximates with the
same accuracy as u
h
approximates u in the energy norm (i.e., if (4.3.18) holds with
j
r
, which is of the same order as [u u
h
|[), then we conclude that the computable
quantity E
1r
(u
h
) is equivalent to the error and has the same rate with respect to c as
the quantity E
1
(u
h
) in (4.3.16).
A lower bound of the approximation error in terms of a different quantity can be
obtained by the method we used for the derivation of (4.1.13). Assume that div a = 0.
Then,
sup
uV
0
__
D
(cV(u ) Vn a V(u )n) J.
c
2
[Vn[
2
_
= sup
uV
0
__
D
(cV(u ) (u )a) VnJ.
c
2
[Vn[
2
_
_ sup
r1
2
_
D
(cV(u ) (u )a) t
c
2
[t[
2
) J.
=
1
2c
[cV(u) (u)a[
2
=
c
2
[V(u )[
2
1
2c
[a(u )[
2
_
D
V(u ) a(u ) J..
The last integral on the right-hand side is equal to zero. Since
_
D
(cVu Vn na Vu) J. =
_
D
nJ..
we conclude that
c
2
[V(u )[
2
1
2c
[a(u )[
2
_ sup
uV
0
_
c
2
[Vn[
2
_
D
(cV Vn na V n) J.
_
= sup
uV
0
M
2
CD
(. n). (4.3.20)
Note that
sup
uV
0
M
2
CD
(. n) _ M
2
CD
(. u ) =
c
2
[V(u )[
2
.
so that for small c the quality of the lower bound deteriorates no faster than [a[c
-1
.
4.3.2 The reaction-convection-diffusion problem
Now, we consider the most general statement of a linear elliptic problem related to
diffusion type models. It is the reaction-convection-diffusion problem
div Vu a Vu
2
u = in C. (4.3.21)
u = u
0
on I
1
. (4.3.22)
Vu n = J on I
2
. (4.3.23)
where the data satisfy the same assumptions as in the problem (4.1.1)(4.1.3) and
1
o
(C). _
. (4.3.24)
1
2
div a
2
:=
2
_
2
0
. (4.3.25)
One more assumption is that the function k(.) :=
1
2
(a n)(.) is dened at almost all
points of I and that the inow part of the boundary is a subset of I
1
, i.e.,
I
-
:= {. I [ k(.) < 0] I
1
. (4.3.26)
Now
V
0
:= {n H
1
(C) [ n = 0 on I
1
]
and u is dened as a function in u
0
V
0
that meets the integral identity
_
D
(Vu Vn (a Vu)n
2
n) J.
=
_
D
nJ.
_
I
2
JnJs. Vn V
0
(C). (4.3.27)
The existence of a generalized solution u follows from the known results in the theory
of partial differential equations (e.g., see [151, 217]).
The assertion below suggests a general form of a computable upper bound of the
error measured in terms of a natural energy type norm.
Theorem 4.2. Let the above made assumptions on the problem data be fullled. Then,
for any . 1
o
0,1j
, u
0
V
0
, and , H
I
2
(C. div) the following estimate
holds:
[u |[ _ M
RCD
(. . . ,) :=
_
_
_
_
r
:
(. ,)
_
_
2
D
[[, V[[
2
+
_
_
_

_
k
(J , n)
_
_
_
I
2
_
1{2
1
c
1
_
C
T D
[(1 )r
:
(. ,)[ C
TI
2
[(1 )(J , n)[
_
. (4.3.28)
where
r
:
(. ,) := a V
2
div ,.
[u |[
2
:= [[V(u )[[
2
_
D
2
(u )
2
J.
_
I
2
k(u )
2
Js.
C
:= {. C [ (.) ,= 0] . I
2
:= {. I
2
[ k(.) ,= 0] .
= 0 on C\ C
, = 0 on I
2
\ I
2
, and C
T I
1
and C
TI
2
are constants in (4.1.8)
and (4.1.9), respectively.
Proof. Rewrite (4.3.27) in the form
_
D
_
V(u ) V(u ) (a V(u ))(u )
2
(u )
2
_
J.
=
_
D
( (a V)
2
)(u ) J.
_
D
V V(u ) J.
_
I
2
J(u ) Js. (4.3.29)
Note that
1
2
_
D
(div a)(u )
2
J. =
1
2
_
D
a V(u )
2
J.
1
2
_
I
2
a n(u )
2
Js
=
_
D
(u )a V(u ) J.
_
I
2
k(u )
2
Js. (4.3.30)
Thus, the left-hand side of (4.3.29) is converted into the norm [u |[
2
, which is a
natural measure of the difference between u and .
Now, we have
[u |[
2
= m(. ,) :=
_
D
r
:
(. ,)(u ) J.
_
I
2
(J , n)(u ) Js
_
D
(, V) V(u ) J.. (4.3.31)
where ,(.) is an arbitrary function in H
I
2
(C. div).
Set (.) = 0 on C \ C
and represent the rst term on the right-hand side of

(4.3.31) in the form
_
D
r
:
(. ,)(u ) J. =
_
D
r
:
(. ,)(u ) J.
_
D
(1 )r
:
(. ,)(u ) J.
_
_
_
_
r
:
(. ,)
_
_
_
D
[(u )[
D
C
T D
c
-1
1
[(1 )r
:
(. ,)[[[V(u )[[.
Similarly,
_
I
2
(J , n)(u ) Js
_ [

_
k
(J , n)[
I
2
[
_
k(u )[
I
2
C
TI
2
c
-1
1
[(1 )(J , n)[
I
2
[[V(u )[[.
where 1
o
0,1j
and (.) = 0 on I
2
\ I
2
. From the above relations, we obtain
m(. ,) _
_
_
_
r
:
(. ,)
_
_
_
D
[(u )[ [[, V[[

+
[V(u )[
_
_
_
_
k
(J , n)
_
_
_
I
2
[
_
k(u )[
I
2

1
c
1
_
C
T D
[(1 )r
:
(. ,)[
C
TI
2
[(1 )(J , n)[
I
2
_
[[V(u )[[
_
__
_
_
r
:
(. ,)
_
_
_
2
D
[[, V[[
2
+
_
_
_
_
k
(J , n)
_
_
_
2
I
2
_
1{2
[u |[
1
c
1
_
C[(1 )r
:
(. ,)[
C
TI
2
[(1 )(J , n)[
I
2
_
[[V(u )[[. (4.3.32)
Estimate (4.3.28) follows from (4.3.31) and (4.3.32).
Remark 4.3. Estimate (4.3.28) shows that the upper bound is given by the norms of
residuals of three relations
, = V in C. (4.3.33)
div , a V
2
= 0 in C. (4.3.34)
, n = J on I
2
. (4.3.35)
which jointly dene the reaction-convection-diffusion problem. Therefore, the majo-
rant M
RCD
(. . . ,) vanishes if and only if u = and , = .
4.3.3 Special cases and modications
If = = 0, then we arrive at the estimate
[u |[ _ [[, V[[
+
C
T D
c
1
[r
:
(. ,)[
C
TI
2
c
1
[J , n[
I
2
. (4.3.36)
It includes two global constants C
T D
and C
TI
2
and does not contain and k in nega-
tive powers. Therefore, it is stable with respect to small values of parameters.
For = 1 and = 0 we obtain another estimate
[u |[ _
C
TI
2
c
1
[J , n[
I
2

__
_
_
1
r
:
(. ,)
_
_
_
2
[[, V[[
2
+
_
1{2
. (4.3.37)
which involves only the trace constant.
Set = = 1. Then we obtain such a form of the upper bound that contains
neither C
T D
nor C
TI
2
:
[u |[
2
_
_
_
_
r
:
(. ,)
_
_
_
2
_
_
_
J , n
_
k
_
_
_
2
I
2
[[, V[[
2
+
. (4.3.38)
Remark 4.4. If J is a simple (e.g., afne) function, then it is not difcult to satisfy the
condition I
2
exactly. For
, H
0
I
2
:= {, H
I
2
(C. div) [ , n J = 0 on I
2
].
the estimate (4.3.28) has a simplied form
[u |[ _
_
_
_
r
:
(. ,)
_
_
2
D
[[, V[[
2
+
_
1{2
C
T D
c
1
[(1 )r
:
(. ,)[. (4.3.39)
Remark 4.5. In (4.3.28), we can set
= min{. 1]. = min{k. 1]. (4.3.40)
This choice may be efcient if and k are small in some parts of C and I
2
, respec-
tively. However, the best upper bound is attained if (4.3.28) is minimized with respect
to , , and ,.
Set in (4.3.28) , = a t, where t H(C. div). Then, it has the form
[u |[ _ M
RCD
(. . . t)
:=
_
_
_
r
:
(. t)
_
_
2
D
[[t a V[[
2
+
_
_
_

_
k
(J t n (a n))
_
_
_
I
2
_
1{2
1
c
1
_
C
T D
[(1 ) r
:
(. t)[
C
TI
2
[(1 )(J t n (a n))[
_
. (4.3.41)
where r
:
(. t) :=
2
div t,
2
=
2
div a. This majorant is related to the
representation
= Vu au in C. (4.3.42)
div
2
= 0 in C. (4.3.43)
n = J (a n) u on I
2
. (4.3.44)
In particular, for = = 0 we obtain
[u |[ _ [[a t V[[
+
C
T D
c
-1
1
[
2
div t[
C
TI
2
c
-1
1
[J (t a) n[
I
2
. (4.3.45)
It is easy to see that the majorants dened by (4.3.41) and (4.3.45) possess the same
properties as (4.3.28) and (4.3.36); namely, they vanish at the exact solution and tend
to zero if and , tend to u and (or ).
We can obtain a better estimate, using the method based on introducing a function
0 V
0
(cf. (4.1.21)(4.1.23)) such that V0 H
I
2
(C. div). Since u vanishes
on I
1
, we note that
_
D
div
_
V0 0a
_
(u ) J.
=
_
D
_
V0 0a
_
V(u ) J.
_
I
2
_
V0 0a
_
n(u ) Js
=
_
D
_
(V0 0a) V
2
0 0
_
J.
_
I
2
J0 Js
_
I
2
_
V0 0a
_
n(u ) Js (4.3.46)
where we have used the relation
_
D
(Vu V0 (a Vu)0
2
0) J. =
_
D
0 J.
_
I
2
J0 Js.
By (4.3.46) we rearrange (4.3.31) as follows:
[u |[
2
=
_
D
r
:
(. ,. 0)(u ) J.
_
I
2
r
T
2
(,. 0)(u ) Js
_
D
(, V) V(u ) J. F
(0). (4.3.47)
where
r
:
(. ,. 0) := r
:
(. ,) div (V0 0a).
r
T
2
(,. 0) := J (, V0 0a) n.
and F
(0) =
_
D
_
0 (V0 0a) V
2
0
_
J.
_
I
2
J0 Js.
As before, let and be functions with values in 0. 1| such that (.) = 0 on
C\ C
and (.) = 0 on I
2
\ I
2
. Then,
_
D
r
:
(. ,. 0)(u ) J. _
_
_
_
r
:
(. ,. 0)
_
_
_
D
[(u )[
D
C
T D
c
-1
1
[(1 )r
:
(. ,. 0)[[[V(u )[[
and
_
I
2
r
T
2
(,. 0)(u ) Js
_
_
_
_
_
k
r
T
2
(,. 0)
_
_
_
I
2
[
_
k(u )[
I
2
C
TI
2
c
-1
1
[(1 )r
T
2
(,. 0)[
I
2
[[V(u )[[.
From the above relations, we obtain
[u |[
2
_
1
4j
1
_
_
_
r
:
(. ,. 0)
_
_
_
2
D
j
1
[(u )[
2
D
1
4j
2
C
2
T D
c
2
1
[(1 )r
:
(. ,. 0)[
2
j
2
[[V(u )[[
2
1
4j
3
_
_
_
_
_
k
r
T
2
(,. 0)
_
_
_
_
2
I
2
j
3
[
_
k(u )[
2
I
1
4j
4
C
2
TI
2
c
2
1
[(1 )r
T
2
(,. 0)[
2
I
2
j
4
[[V(u )[[
2
1
4j
5
[[, V[[
2
+
j
5
[[V(u )[[
2
F
(0). (4.3.48)
Here, j
1
. j
3
0. 1) and j
2
, j
4
, and j
5
are positive numbers such that
j
2
j
4
j
5
_ 1.
Now, we obtain an upper bound of the error evaluated in terms of a weighted norm:
(1j
2
j
4
j
5
)[[V(u)[[
2
(1j
1
)[(u)[
2
D
(1j
3
)[
_
k(u)[
2
I
2
_
1
4
1
_
_
_
r
:
(. ,. 0)
_
_
_
2
D

C
2
1:
4
2
c
2
1
[(1 )r
:
(. ,. 0)[
2
1
4
3
[

_
k
r
T
2
(,. 0)[
2
I
C
2
TT
2
4
4
c
2
1
[(1 )r
T
2
(,. 0)[
2
I
2
1
4
5
[[, V[[
2
+
F
(0). (4.3.49)
If and k are positive, then we can set = = 1, j
2
= j
4
= 0, and j
1
= j
3
=
1
2
.
In this case, the estimate (4.3.49) has a simplied form:
[u |[
2
_ M
2
RCD
(. ,. 0) :=
_
_
_
1
r
:
(. ,. 0)
_
_
_
2
D

_
_
_
1
_
k
r
T
2
(,. 0)
_
_
_
2
I
2
[[, V[[
2
+
F
(0). (4.3.50)
Set , = Vu and dene 0 as a solution of the problem
div
_
V0 0a
_
= a V(u )
2
(u ). (4.3.51)
0 = 0 on I
1
. (4.3.52)
(V0 0a) n = 0 on I
2
. (4.3.53)
Then,
r
:
(. ,) div
_
V0 0a
_
= 0.
[[, V[[
2
+
= [[V(u )[[
2
.
r
T
2
(,. 0) = J (Vu V0 0a) n = 0.
Since
_
D
_
V0 0a
_
V(u ) J. =
_
D
_
a V(u )(u )
2
(u )
2
_
J..
we have
F
(0) =
_
D
(V0 0a) V(u ) J.
=
_
D
(u )a V(u ) J.
_
D
2
(u )
2
J.
=
_
D
2
(u )
2
J.
_
I
2
k(u )
2
Js.
Hence, the right-hand side of (4.3.50) coincides with the left-hand side.
Remark 4.6. Other modications of the estimates (which are similar to those con-
sidered in Sections 3.5 and 4.2) can be also derived using the same arguments. In
particular, if C is decomposed into C
i
, then the estimates analogous to (4.2.28) and
(4.2.29) take place under the same assumptions on r
:
(. ,). We leave this task for the
reader as a good exercise.
Remark 4.7. In Section 8.5 we discuss another (rather general) way to obtain two-
sided estimates that have no gap between upper (lower) bounds and error measures.
For reaction-convection-diffusion problems it yields other modications of error esti-
mates.
4.3.4 Estimates for uxes
Theorem 4.2 allows us to derive computable bounds of errors measured in terms of the
dual variable (ux). Let , H
I
2
(C) be an approximation of the exact ux . For
any u
0
V
0
, we have
[[ ,[[
+
_ [[V(u )[[
+
[[V ,[[
+
(4.3.54)
= [[V(u )[[ [[V ,[[
+
_ M
RCD
(. . . ,) [[V ,[[
+
.
By (4.3.28) and (4.3.54), we obtain an upper bound for a combined primal-dual norm
[u |[ [[ ,[[
+
_ 2M
RCD
(. . . ,) [[V ,[[
+
. (4.3.55)
We can use (4.3.50) instead of (4.3.28). Then, in (4.3.55) M
RCD
(. . . ,) must be
replaced by M
RCD
(. ,. 0).
The norm
[,|[
+
:= [[,[[
+
[div ,[ [, n[
I
2
is another quantity, which is natural to use for uxes. For this norm, we have
[ ,|[
+
_ [[V(u )[[
+
[[V ,[[
+
[( ,) n[
I
2
[div ( ,)[
= [[V(u )[[ [[V ,[[
+
[, n J[
I
2
[div ( ,)[. (4.3.56)
By (4.3.28) and (4.3.56), we nd that
[ ,|[
+
_ M
RCD
(. . . ,) [[V ,[[
+
[, n J[
I
2
[div , a Vu
2
u [.
Assume that
0
> 0. Then,
[div , a Vu
2
u [
_ [r
:
(. ,)[ [a[[V(u )[
_
_
_
_
_
_[(u )[
_ [r
:
(. ,)[
_
_
[a[
2
c
2
1
_
_
_
_
_
_
_ _
[[V(u )[[
2
[(u )[
2
_
and we obtain
[ ,|[
+
_ M
RCD
(. . . ,) [[V ,[[
+
[r
:
(. ,)[
[, n J[
I
2

_
[a[
2
c
2
1
_
_
_
_
_
_[u |[.
which yields the estimate
[ ,|[
+
_ M
RCD
(. . . ,) [r
:
(. ,)[ [[V ,[[
+
[, n J[
I
2
. (4.3.57)
where = 1
_
|a|
2
c
2
1
_
_
_
_
_
_.
If , H
0
I
2
(C), then we have a simpler estimate
[ ,|[
+
_ M
RCD
(. . . ,)[r
:
(. ,)[ [[V ,[[
+
. (4.3.58)
Hence, for the pair (. ,) (u
0
V
0
) H
0
I
2
an upper bound in the full primal-dual
norm has the form
[u |[ [ ,|[
+
_ (1 )M
RCD
(. . . ,) [r
:
(. ,)[
[[V ,[[
+
. (4.3.59)
In particular, for = = 0 we have
[u |[ [ ,|[
+
_
_
1 (1 )
C
T D
c
1
_
[r
:
(. ,)[ (2 )[[V ,[[
+
. (4.3.60)
Equivalence to the error. By Theorem 4.2, we observe that the quantity
E() := inf
,1
T
2
(D)
M
RCD
(. . . ,)
yields an upper bound of the error. The proposition below shows that jE() also
furnishes a lower bound, where j does not depend on and is explicitly estimated by
the problem data.
Proposition 4.8. For any . 1
o
0,1j
and any u
0
V
0
,
E() _ c
(;)[u |[. (4.3.61)

where ; is an arbitrary positive number, and c
is dened by (4.3.62).
Proof. We have
inf
,1
T
2
(D)
M
RCD
(. . . ,) _ M
RCD
(. . . ).
Note that
[r
:
(. . . )[ = [ a V
2
div [
_ [a V(u )[ [
2
(u )[
_
_
2
[a[
2
c
2
1
[[V(u )[[
2
2
2
0
_
D
2
(u )
2
J..
Since [[ V[[
+
= [[V(u )[[ and n = J we nd that
1
(1 ;)
M
2
RCD
(. . . )
_
_
1
2
0
C
2
T D
;c
2
1
_
[r
:
(. ,)[
2
[[V(u )[[
2
_ 2
_
1
2
0
C
2
T D
;c
2
1
__
[a[
2
c
2
1
[[V(u )[[
2
2
0
_
D
2
(u )
2
J.
_
[[V(u )[[
2
=
_
1
_
1
2
0
C
2
T D
;c
2
1
_
2[a[
2
c
2
1
_
[[V(u )[[
2
_
1
2
0
C
2
T D
;c
2
1
_
2
2
2
0
_
D
2
(u )
2
J..
Dene
c
2
(;) := (1;) max

__
1
_
1
2
0
2C
2
T D
;c
2
1
_
2[a[
2
c
2
1
_
.
_
1
2
0
C
2
T D
;c
2
1
_
2
2
2
0
_
.
(4.3.62)
Then, M
RCD
(. . . ) _ c
(;)[u |[. and we obtain (4.3.61).

It is worth noting that, in the above consideration, we have overestimated the value
of M
RCD
(. . . ) at several places, so that the constant c
(;) dened by the rela-

tion (4.3.62) is rather coarse. Anyway it gives an idea how c
depends on [a[.
Also, we can establish the equivalence of M and the error computed in the full
primal-dual norm.
Proposition 4.9. For any u
0
V
0
and , H
0
I
2
(C), the following two-sided
estimate holds:
C
M
RCD
(. . . ,) _ [u |[ [ ,|[
+
_ C
M
RCD
(. . . ,). (4.3.63)
where = min{1. ], C
=
1
,
j = max
_
1.

C
T D
.
_
2
_
1

C
T D
[a[
c
1
_
.
_
2

C
T D
0
_
.

C
T D
:= 1
C
T D
c
1
.
and C
= 3
_
|a|
2
c
2
1
0
.
Section 4.4 Notes for the chapter 103
Proof. The right-hand side inequality follows from the structure of [u |[ and
(4.3.59).
To prove the left-hand side, we note that , _ 1. Therefore,
M
RCD
(. . . ,) _
_
[r
:
(. ,)[
2
[[, V[[
2
+
_
1{2
C
T D
c
1
[r
:
(. ,)[
_

C
T D
[r
:
(. ,)[ [[, V[[
+
.
We have
[r
:
(. ,)[ _ [div ( ,)[ [a V(u )[ [
2
(u )[
_ [div ( ,)[
|a|
c
1
[[V(u )[[

2
0
[(u )[.
[[, V[[
+
_ [[V(u )[[ [[ ,[[
+
.
Thus,
M
RCD
(. . . ,) _
_
1

C
T D
[a[
c
1
_
[[V(u )[[

C
T D
0
[(u )[
[[ ,[[
+

C
T D
[div ( ,)[
_
_
2 max
_
1

C
T D
[a[
c
1
.

C
T D
0
_
[u |[
max{1.

C
T D
][ ,|[
+
. (4.3.64)
Now (4.3.64) implies the left-hand side of (4.3.63).
4.4 Notes for the chapter
For linear and nonlinear diffusion problems, the rst a posteriori estimates of the func-
tional type were obtained in the context of the general variational method, basic princi-
ples of which are briey discussed at the beginning of this chapter (see [276][282] and
[244] for a consequent exposition). The nonvariational approach to the derivation of
a posteriori estimates for stationary diffusion problems (and other linear elliptic prob-
lems having the divergent form) was suggested in the authors paper [286]. Section 4.1
is based on these results. Later, a posteriori estimates for the convection-diffusion and
reaction-convection-diffusion problems were derived (see [293, 300]). A posteriori
estimates for the reaction-diffusion problem discussed in Section 4.2 were obtained in
the paper by S. Repin and S. Sauter [305]. Functional type a posteriori estimates for
the stationary reaction-convection-diffusion problem were analyzed in S. Nicaise and
S. Repin [248]. In [287], the nonvariational method was applied to the evolutionary
diffusion equation. The latter results are discussed in Chapter 9.
5 Elasticity
5.1 The linear elasticity problem
Consider an elastic body that occupies a domain C R
d
the boundary I of which
consists of two disjoint parts I
1
and I
2
([I
1
[ > 0). The classical statement of the
linear elasticity problem is as follows: Find a tensor-valued function o (stress) and a
vector-valued function u (displacement) that satisfy the system of equations
o = Lc(u) in C. (5.1.1)
Div o = 0 in C. (5.1.2)
u = u
0
on I
1
. (5.1.3)
on = J on I
2
. (5.1.4)
Here and J are given forces and L = {1
i}kn
] is the tensor of elasticity constants,
which is subject to the conditions
c
2
1
[c[
2
_ Lc : c _ c
2
2
[c[
2
. Vc M
dd
x
. (5.1.5)
Henceforth, we assume that the coefcients of the elasticity tensor are bounded and
possess natural symmetry, i.e.,
L
i}kn
= L
}ikn
= L
kni}
. i. . k. m = 1. . . . . J. (5.1.6)
L
i}kn
1
o
(C). (5.1.7)
Moreover, we assume that
1
2
(C. R
d
). J 1
2
(I
2
. R
d
). u
0
H
1
(C. R
d
). (5.1.8)
A function u u
0
V
0
is a generalized solution of (5.1.1)(5.1.5) if it satises the
integral relation
_
D
Lc(u) : c(n) J. = (. n) (5.1.9)
for all n V
0
. In (5.1.9), the right-hand side is determined by the volume and surface
loads
(. n) =
_
D
nJ.
_
I
2
J nJs.
and
V
0
:= {n H
1
(C. R
d
) [ n = 0 on I
1
].
Section 5.1 The linear elasticity problem 105
The linear elasticity problem has a variational statement: Find u u
0
V
0
such that
J(u) = inf
uu
0
V
0
J(). J() :=
1
2
_
D
Lc(n) : c(n) J. (. n). (5.1.10)
The existence and uniqueness of u are easy to prove if the coercivity of
[[c(n)[[ :=
__
D
Lc(n) : c(n) J.
_
1{2
on the space V
0
is established. This fact follows from Korns inequality (1.4.32), which
shows that [n[
2
1,2,D
is equivalent to
[n[|
2
1,2,D
:=
_
D
([c(n)[
2
[n[
2
) J..
Proposition 5.1. Let C be a Lipschitz domain in R
d
, and let I
1
satisfy the following
property:
J .
0
I
1
and c > 0 such that I T(.
0
. c) I
1
. (5.1.11)
Then
[n[
1,2,D
_ j
:T
1
[c(n)[. Vn V
0
. (5.1.12)
where j
:T
1
is a positive constant independent of n.
Proof. Assume that the assertion is wrong. Then, for any m N there exists
(n)
V
0
such that
[
(n)
[
1,2,D
> m
__
D
[c(
(n)
)[
2
J.
_
1{2
. (5.1.13)
Consider the sequence of normalized functions
n
:=

(m)
|
(m)
|
1,2,:
. Obviously,
[
n
[
1,2,D
= 1 and, by (5.1.13),
1
m
_
__
D
[c(
n
)[
2
J.
_
1{2
. (5.1.14)
Hence, we can extract a subsequence of {
n
] that converges to a vector-valued func-
tion V
0
weakly in H
1
(C. R
d
) and strongly in 1
2
(C. R
d
). Therefore, c(
n
)
weakly converges (in 1
2
) to c( ). By (5.1.13), we know that
[c(
n
)[ 0. (5.1.15)
Thus,
0 = liminf
n-o
[c(
n
)[ _ [c( )[.
106 Chapter 5 Elasticity
and, therefore, c( ) = 0. Then R(C) V
0
(cf. 1.4.3). In view of (5.1.11), the
intersection of these sets contains only the zero function, so that = 0. However, in
such a case we arrive at a contradiction. Indeed, on the one hand, by Korns inequality
we conclude that
[
(n)
[
1,2,D
_ C
1D
__
D
([c(
n
)[
2
[
n
[
2
) J.
_
1{2
n-o
0.
On the other hand, [
n
[
1,2,D
= 1 for any m and such a sequence cannot tend to zero.
It remains to conclude that (5.1.13) is wrong and the constant j
:T
1
exists.
Corollary 5.2. There exists a constant C such that
_
D
[n[
2
J.
_
I
2
[n[
2
Js _ C
2
[[c(n)[[
2
. Vn V
0
. (5.1.16)
Proof. The proof follows from Proposition 5.1 and trace inequalities. Indeed, by a
Friedrichs type inequality for the functions in V
0
and the trace inequality, we conclude
that
_
D
[n[
2
J.
_
I
2
[n[
2
Js _ C
2
D
[n[
2
1,2,D
. Vn V
0
. (5.1.17)
Now (5.1.16) follows from (5.1.5), (5.1.12), and (5.1.17).
Note that on V
0
the norms [Vn[ and [n[
1,2,D
are equivalent. Therefore, instead
(5.1.17) we can use the inequality
_
D
[n[
2
J.
_
I
2
[n[
2
Js _

C
2
D
[Vn[
2
. Vn V
0
. (5.1.18)
where

C
D
is a somewhat different constant (which also depends only on C and I
2
).
Also, (5.1.12) means that
[Vn[ _ c
DI
1
[c(n)[. Vn V
0
. (5.1.19)
with a constant c
DI
1
_ j
:T
1
.
Remark 5.3. In practice, the value of C can be estimated by minimizing the quotient
[[c(n)[[
2
_
D
[n[
2
J.
_
I
2
[n[
2
Js
over a sufciently representative nite-dimensional space V
0h
V
0
.
The estimation of constants arising in various versions of Korns inequality is a
more complicated task. This question was investigated by a number of authors (e.g.,
see C. Horgan [179]).
Section 5.2 Estimates for displacements 107
5.2 Estimates for displacements
Let u
0
V
0
. Represent (5.1.9) in the form
_
D
Lc(u ) : c(n) J. =
_
D
( n Lc() : c(n)) J.
_
I
2
J nJs. (5.2.1)
where n V
0
. Let (C) := 1
2
(C. M
dd
) and
t H
I
2
(C. Div) :=
t (C) [ Div t 1
2
(C. R
d
). t n 1
2
(I
2
. R
d
)
_
.
Since
Div (tn) = n Div t t
T
: Vn. (5.2.2)
we have (for t
x
(C))
_
I
2
t n nJs =
_
D
(n Div t t : c(n)) J.. (5.2.3)
By (5.2.1) and (5.2.3), we obtain
_
D
Lc(u ) : c(n) J. =
_
D
( Div t) nJ.
_
D
(t Lc()) : c(n) J.
_
I
2
(J t n) nJs. (5.2.4)
Set n = u and use (5.1.16). We arrive at the estimate
[[c(u )[[ _ M
EL
(. t)
:= [[t Lc()[[
+
C
_
[Div t [
2
[J t n[
2
I
2
_
1{2
. (5.2.5)
which is quite analogous to the estimate (4.1.11). Squaring both parts of (5.2.5), we
obtain an analog of (4.1.15):
[[c(u )[[
2
_ M
2
EL
(. t. ) := (1)
_
D
_
c()L
-1
t): (Lc()t
_
J.
C
2
[Div t [
2
[J t n[
2
I
2
_
. (5.2.6)
In (5.2.6), t is a symmetric tensor-valued function and is an arbitrary positive con-
stant. It is easy to note that the rst term is nonnegative and vanishes if and only
if
t = Lc().
It penalizes violations of the Hookes law. The meaning of the second and third terms
is obvious: they represent other two relations and penalize errors in the equilibrium
equation (5.1.2) and boundary condition (5.1.4). Thus, the majorant not only provides
a bound of the overall value of the error, but also shows its physically sensible parts.
Latter information suggests a correct way for nding a better approximation.
In (5.2.5) and (5.2.6), the error is measured in terms of the norm generated by the
tensor c. For this reason, the Korns constant does not explicitly occur in the estimates.
However, a tensor-valued function t in M
EL
(. t) must be symmetric. This require-
ment is easy to fulll if t is approximated with the help of nite element approxi-
mations having degrees of freedom at nodes (if the nodal tensors are symmetric then
an extension of them is also symmetric). For other approximations (e.g., for mixed
approximations degrees of freedom of which are associated with faces) the symmetry
condition may be rather burdensome. Below we discuss a way of avoiding it.
Take an arbitrary tensor-valued function t H
I
2
(C. Div) and rearrange (5.2.1)
with the help of (5.2.2) as follows:
_
D
Lc(u ) : c(n) J. =
_
D
_
(Divt ) n (t
T
Lc()) : Vn
_
J.
_
I
2
(J t
T
n) nJs. Vn V
0
. (5.2.7)
Decompose t into symmetric and skew-symmetric parts t
sm
andt
sk
. Then,
t
T
: Vn =t
sm
: c(n) t
sk
: Vn.
Note that Lc() is a symmetric tensor. Therefore, (5.2.7) implies the estimate
_
D
Lc(u ) : c(n) J. _ [Divt [[n[ [t
T
n J[
I
2
[n[
I
2
[[t
sm
Lc()[[
+
[[c(n)[[ [t
sk
[[Vn[. (5.2.8)
By (5.1.16), we nd that
[Divt [[n[ [t
T
n J[
I
2
[n[
I
2
_ C
_
[Divt [
2
[t
T
n J[
2
I
2
_
1{2
[[c(n)[[.
Set n = u and use (5.1.19). We arrive at the estimate
[[c(u )[[ _ C
_
[Divt [
2
[t
T
n J[
2
I
2
_
1{2
[[t
sm
Lc()[[
+
c
DI
1
c
1
[t
sk
[. (5.2.9)
This estimate includes an additional term which is positive if the tensor-valued func-
tion t is nonsymmetric. If t is a symmetric tensor (i.e., t
sk
= 0), then (5.2.9) is
equivalent to (5.2.5).
Section 5.3 Estimates for stresses 109
Remark 5.4. We note that the right-hand sides of (5.2.5) and (5.2.9) are equal to
[[c(u )[[ if t andt coincide with o.
By the same arguments as for the diffusion equation, one can prove that
_
D
Lc(u) : c(u)J. _ M
2
EL
(. n) := 2F
(n) [[c(n)[[
2
. (5.2.10)
where n V
0
and
F
(n) :=
_
D
nJ.
_
I
2
J nJs
_
D
Lc() : c(n) J..
Also, we can show that
[[c( u)[[
2
= sup
uV
0
M
2
EL
(. n).
By the maximization of the functional M
EL
(. n) on a sequence of nite-dimensional
spaces V
0k
V
0
. we obtain a sequence of computable lower bounds
M
k
= sup
uV
0k
M
EL
(. n
k
).
If the spaces V
0k
satisfy the limit density condition stated, then the sequence {M
k
]
tends to [[c( u)[[
2
.
5.3 Estimates for stresses
Assume that t H
I
2
(C. Div) is a symmetric tensor-valued function that approxi-
mates o. By (5.1.1), for any u
0
V
0
we have
[[o t[[
+
= [[Lc(u) t[[
+
_ [[Lc(u )[[
+
[[Lc() t[[
+
= [[c(u )[[ [[Lc() t[[
+
_ M
EL
(. t) [[Lc() t[[
+
. (5.3.1)
By (5.2.5) and (5.3.1), we obtain an upper bound for a combined primal-dual norm
[[c(u )[[ [[o t[[
+
_ 2M
EL
(. t) [[Lc() t[[
+
. (5.3.2)
For another norm
[t|[
+
:= [[t[[
+
[Div t[ [t n[
I
2
.
we have
[o t|[
+
_ M
EL
(. t) [[Lc() t[[
+
[t n J[
I
2
[Div (o t)[
= M
EL
(. t) [[Lc() t[[
+
[t n J[
I
2
[Div t [. (5.3.3)
Hence,
[[c(u )[[ [o t|[
+
_ 2M
EL
(. t) [[Lc() t[[
+
[t n J[
I
2
[Div t [
_ c M
EL
(. t). (5.3.4)
where c = 3
_
2,C. On the other hand,
M
EL
(. t) _ [[Lc( u)[[
+
[[o t[[
+
C
_
[Div (t o)[
2
[(o t)n[
2
I
2
_
1{2
_ [[c(u )[[ [[o t[[
+
C
_
[Div (t o)[ [(o t)n[
I
2
_
_ c([[c(u )[[ [o t|[
+
). (5.3.5)
where c = max{1. C]. Thus, we note that M
EL
(. t) is equivalent to the error evalu-
ated in the combined stress-strain norm.
5.4 Isotropic linear elasticity
In the important case of isotropic elastic media, the components of L and L
-1
depend
on two material constants only. Below we present respective forms of a posteriori
estimates derived in the previous section.
5.4.1 3D problems
For J = 3, we can represent the elasticity tensor in the form
L c = 1
0
tr c I 2jc
D
. (5.4.1)
L
-1
t =
1
91
0
tr tI
1
2j
t
D
. (5.4.2)
In (5.4.1)(5.4.2), 1
0
and j are positive (elasticity) constants that depend on proper-
ties of media to resist compression and shear forces, respectively.
In this case, the norms are dened by the relations
[[c(n)[[
2
:=
_
D
_
1
0
tr c(n)
2
2j
c
D
(n)
2
_
J.. (5.4.3)
[[t[[
2
+
:=
_
D
_
1
91
0
tr(t)
2
1
2j
(t)
D
2
_
J.. (5.4.4)
Since
c() L
-1
t =
_
1
3
tr(c())
1
91
0
tr(t)
_
I
_
c
D
()
1
2j
tr t
_
. (5.4.5)
Lc() t =
_
1
0
tr(c())
1
3
tr(t)
_
I
_
2j(c)
D
() t
D
_
. (5.4.6)
Section 5.4 Isotropic linear elasticity 111
we nd that
_
c() L
-1
t
_
: (Lc() t)
= 1
0
_
div
1
31
0
tr t
_
2
2j
c
D
()
1
2j
tr t
2
. (5.4.7)
Hence, for isotropic media the estimate (5.1.6) reads as follows:
_
D
_
1
0
[div (u )[
2
2j
c
D
(u )
2
_
J.
_ (1 )
_
D
_
1
0
_
div
1
31
0
tr t
_
2
2j
c
D
()
1
2j
tr t
2
_
J.
C
2
[Div t [
2
[J t n[
2
I
2
_
. (5.4.8)
Instead of 1
0
and j, one cane use another pair of constants and represent (5.4.1) and
(5.4.2) in the form
L c = z tr(c) I 2jc. (5.4.9)
L
-1
t =
1
2j
_
t
z
3z 2j
tr(t) I
_
. (5.4.10)
Here z = 1
0
2
3
. In this case,
[[c()[[
2
:=
_
D
_
z[div [
2
2j[c()[
2
_
J.. (5.4.11)
[[t[[
2
+
:=
_
D
1
2j
_
[t[
2
z
3z 2j
tr(t)
2
_
J.. (5.4.12)
and
_
c() L
-1
t
_
: (Lc() t)
= z[div [
2
2j[c()[
2
1
2j
_
[t[
2
z
3z 2j
tr(t)
2
_
2c() : t.
5.4.2 The plane stress problem
The plane stress problem arises if a thin elastic plate is subject to an action of forces
lying in its plane. Let
C :=

C
_
h
2
.
h
2
_
.
where

C R
2
is the so-called middle surface I
0
. It is assumed that this surface
contains .
1
and .
2
axes of a Cartesian coordinate system and .
3
is perpendicular to
I
0
. Throughout this chapter the symbol is used to mark plane components of
vectors and tensors. In particular, we write . = (.
1
. .
2
) for the plane coordinate
vector, = (
1
.
2
) is a plane vector, andt := {t
xt
], s. t = 1. 2, is a plane tensor.
Assume that h is small with respect to the character size of

C. In general, it may
be difcult to exactly dene this notion, but for sufciently regular domains we can
compare h with the diameter diam
C. Let = {
1
(.).
2
(.). 0] and assume that the
surfaces
I
:=
_
. = ( n. .
3
) [ .

C. .
3
=
h
2
_
are free from loads. Also, we assume that the boundary conditions on
I
i
:= ;
i

_
h
2
.
h
2
_
are dened by the functions u
0
= {u
0
(.). 0] and J = {
J(.). 0]. Here the ;

i
denote
two parts of ;, which is the boundary of
C.
In the plane stress model, it is suggested to consider the stress tensor as a plane
tensor. In other words, we set
o
i3
= 0. i = 1. 2. 3. (5.4.13)
In addition, in the plane stress model it is assumed that o insignicantly depends on
.
3
, so that
o =o(.). (5.4.14)
Strictly speaking these assumptions are not true and violate the 3D relations of linear
elasticity problem. However, they lead to a 2D problem which is much simpler. The
error arising owing to this dimension reduction procedure is the modeling error. In the
next section, we derive an upper bound for this error. Now, our goal is to present the
error majorant for the plane stress problem.
In view of (5.4.13), c
13
(u) = c
23
(u) = 0. For the component c
33
(u) we have the
relation (which comes from the Hookes law)
0 = o
33
= 1
0
(c
11
(u) c
22
(u) c
33
(u)) 2j
2c
33
(u) c
11
(u) c
22
(u)
3
.
which implies
tr c =
6j
31
0
4j
trc.
Therefore, 2D form of the Hookes law is
o =

Lc =

1
0
trc
I 2j(c)
D
. (5.4.15)
c =

L
-1
o =
1
4
1
0
tr o
I
1
2j
o
T
. (5.4.16)
where

1
0
=
91
0
31
0
4
and
trc :=c
11
c
22
. The other relations that dene the solution
are as follows:
b
Divo
= 0 in

C. (5.4.17)
u =u
0
on ;
1
. (5.4.18)
o n =

J on ;
2
. (5.4.19)
We repeat the arguments adduced for deriving the error majorant in the 3D elasticity
model and arrive at an estimate similar to (5.4.8):
_
D
_
1
0
[
trc(u )[
2
2j
c
D
(u )
2
_
J.
_ (1 )
_
D
_
1
0
_
div
1
2
1
0
trt
_
2
2j
c
D
()
1
2j
trt
2
_
J.
C
2
_
[
b
Divt
[
2
D
[
J t n[
2
;
2
_
. (5.4.20)
5.4.3 The plane strain problem
The deformation is called plane if at any point the displacement vector is collinear to
a certain plane. Assume that this plane is O.
1
.
2
and
u
1
= u
1
(.). u
2
=u
2
(.). u
3
= 0.
In this case, only the plane part of the deformation tensor c(u) has nonzero compo-
nents. By the Hookes law, we conclude that o
13
= o
23
= 0 and o
33
is excluded by
the relation
o
33
=
z
2(z j)
tr(o).
which also follows from the Hookes law. In view of the above relations, the plane
parts of stresses and strains are connected by the relations
o = z
tr(c(u))
I 2jc(u). (5.4.21)
c(u) =
1
2j
_
o
z
2(z j)
tr(o)
I
_
. (5.4.22)
In view of (5.4.22), the term of the majorant related to the constitutive relations of the
plane strain problem has the form
_
D
_
z[
b
div[
2
2j[c()[
2
1
2j
[t[
2
z
4j(z j)
tr(t)
2
2c() : t
_
J..
5.4.4 Error of the plane stress model
Estimate (5.4.8) allows one to measure the error that arises if the 3D elasticity problem
is replaced by a simplied 2D one. Let u and o be the displacement vector and the
stress tensor satisfying (5.4.17)(5.4.19) and the relations
o
11
=

1
0
(u
1,1
u
2,2
) j(u
1,1
u
2,2
). (5.4.23)
o
22
=

1
0
(u
1,1
u
2,2
) j(u
2,2
u
1,1
). (5.4.24)
o
12
= j(u
1,2
u
2,1
). (5.4.25)
With the help of u and o we construct (u. o), which yields an approximate solution of
the original (3D) problem. For this purpose, we set
u = (u
1
. u
2
. ).
o
=o
. o
3
= 0.
where (C) := { H
1
(C) [ = 0 on I
1
].
Since u =u
0
(.) on ;
1
, the function u belongs to u
0
V
0
. Therefore, we can use
(5.4.8) with =u and t =o. Note that
on = 0 on I
and on = {o n: 0] = J on I
2
. (5.4.26)
Divo = {
b
Divo: 0] = {
: 0] = . (5.4.27)
Therefore, the last term of (5.4.8) vanishes and the whole error is contained in the rst
one. To evaluate it, we use the relations
tr(c(u)) =u
1,1
u
2,2
,3
. (5.4.28)
tr(o) =
tr(o) =o
11
o
22
= 2
1
0
(u
1,1
u
2,2
) (5.4.29)
and nd that
tr(c(u))
1
31
0
tr(o) = j (u
1,1
u
2,2
)
,3
.
where
j = 1
2
1
0
31
0
=
31
0
2j
31
0
4j
=
1
2
_
1
0
j
1
_
.
Let us estimate the quantity
1
2
o (c)
D
(u)
2
.
In view of (5.4.23)(5.4.25), we have
2o
11
o
22
3
=
1
0
3
(u
1,1
u
2,2
) j(u
1,1
u
2,2
).
2o
22
o
11
3
=
1
0
3
(u
1,1
u
2,2
) j(u
2,2
u
1,1
).
o
11
o
22
3
=
2
1
0
3
(u
1,1
u
2,2
) .
Therefore,
1
2j
o
D
=
_
_
_
_
_
1
0
6j
(u
1,1
u
2,2
)
u
1,1
u
2,2
2
u
1,2
u
2,1
2
u
1,2
u
2,1
2
1
0
6
(u
1,1
u
2,2
)
u
2,2
-u
1,1
2
0 0
0
0
1
0
3
(u
1,1
u
2,2
)
_
_
.
Next,
c
D
(u) =
_
_
_
u
1,1
u
2,2
6

u
1,1
-u
2,2
2

,3
3
u
1,2
u
2,1
2
u
1,2
u
2,1
2
u
1,1
u
2,2
6

u
2,2
-u
1,1
2

,3
3
,1
2
,2
2
,1
2
,2
2
2
,3
3

u
1,1
u
2,2
3
_
_
and we obtain
1
2j
o c
D
(u) =
_
_
_
_
_
_
_
_
p (u
1,1
u
2,2
)
,3
3
0
,1
2
0
p (u
1,1
u
2,2
)
,3
3

,2
2
,1
2

,2
2

2(p (u
1,1
u
2,2
)
,3
)
3
_
_
.
Hence,
1
2j
o (c)
D
(u)
2
=
2
3
(j(u
1,1
u
2,2
)
,3
)
2
1
2
(
2
,1
2
,2
) .
and we nd that
_
D
_
1
0
[div (u u)[
2
2j
c
D
(u u)
2
_
J.
_
_
D
_
1
0
_
divu
1
31
0
tr o
_
2
2j
c
D
(u)
1
2j
tr o
2
_
J.
_
_
1
0
4j
3
__
D
_
j (u
1,1
u
2,2
)
,3
_
2
J. j
_
D
_
2
,1
2
,2
_
J.. (5.4.30)
It is easy to see that the right-hand side of the above estimate is positive. Indeed, if
the second integral is equal to zero then = (.
3
). Then the rst integral is positive
(because u
1
and u
2
depend only on .). The only one exception is the case j = 0 or
(
b
divu = 0) Since j =

1-
we observe that j can be equal to zero only if the Poisson
coefcient v is equal to zero. In all other cases, the modelling error related to the plane
stress model is positive.
The reader can nd a more detailed discussion of a posteriori estimates for linear
elasticity problems authors papers [280, 282, 284] and in the book P. Neittaanm aki and
S. Repin [244]. Numerical tests and extensions of the above theory to thermoelastic
problems are presented in A. Muzalevskii and S. Repin [240, 241].
In the last section of this chapter, we have touched an important problem: estimation
of modeling errors. This problem deserves a special consideration, which is beyond
the framework of the book. In the context of functional error majorants, a posteriori
error estimates for modeling errors arising in dimension reduction models of diffusion
type problems were derived in the papers by S. Repin, S. Sauter, and A. Smolianski
[310, 310, 312] and for elasticity problems in [284, 285].
6 Incompressible viscous uids
6.1 The Stokes problem
Statement of the problem. The Stokes model is one of the simplest models in the
theory of viscous incompressible uids. It is represented by the relations
u
t
vû = Vp in C. (6.1.1)
div u = 0. (6.1.2)
u(.. 0) =u(.). (6.1.3)
u = u
0
on I. (6.1.4)
where u(.. t ) is the velocity eld, p(.. t ) is the pressure function, v > 0 is the viscos-
ity parameter (or a positive function), and u(.) and u
0
are solenoidal functions that
dene the initial and boundary conditions, respectively.
In the stationary case, u does not depend on t . Then, the problem is to nd u(.)
and p(.) such that
vû = Vp in C. (6.1.5)
div u = 0. (6.1.6)
u = u
0
on I. (6.1.7)
In the case of mixed boundary conditions, (6.1.7) holds only on a part I
1
I, and on
another part I
2
, the Neumann boundary condition
on = J on I
2
. (6.1.8)
(where o = vVu p) is considered. We assume that
1
2
(C. R
n
) and u
0

S
1
(C). (6.1.9)
where

S
1
(C) denotes the closure of smooth solenoidal functions with compact sup-
ports in Cwith respect to the normof H
1
(C. R
d
). Henceforth, we denote H
1
(C. R
d
)
by V and dene V
0
as the subspace of V containing the functions with zero traces on
I (for problems with mixed boundary conditions V
0
contains functions vanishing on
the Dirichlet part of the boundary). We recall that the (Friedrichs) inequality
[n[ _ C
T D
[Vn[ (6.1.10)
holds for n V
0
. The set u
0
S
1
(C) consists of functions u
0
n, where n

S
1
(C).
The space of square summable functions with zero mean is denoted by

1
2
(C).
118 Chapter 6 Incompressible viscous uids
A generalized solution of the stationary Stokes problem (6.1.5)(6.1.7) is dened
by the integral identity
_
D
vVu : VnJ. =
_
D
nJ.. Vn

S
1
(C). (6.1.11)
Remark 6.1. Stokes problem can be stated using the tensor of small strains
c(u) = {c
i}
(u)]. c
i}
(u) :=
1
2
_
u
i,}
u
},i
_
.
instead of Vu. In view of the relation Div c() =
1
2
(^ Vdiv ), for solenoidal
elds we can write (6.1.11) in terms of the operator c (instead of V). Such a statement
is equivalent to (6.1.11) if v is multiplied by 2.
A generalized solution can also be dened as a function u u
0

S
1
(C) satisfying
the relation
_
D
vVu : VnJ. =
_
D
( n p div n) J.. n V
0
. (6.1.12)
on a wider set V
0
. In this case, the respective p is to be dened with the help of (6.1.6).
Inf-Sup condition. First, we recall an important result in the theory of functions
related to the operator div .
Lemma 6.2. Let C be a bounded domain with Lipschitz continuous boundary. Then,
for any function

1
2
(C) one can nd a function n
(
V
0
such that div n
(
=
and
[Vn
(
[ _ c
:
[ [. (6.1.13)
where c
:
is a positive constant dependent only on C.
The reader can nd a proof in I. Babu` ska and A. K. Aziz [22] (for J = 2) and
O. Ladyzhenskaya and V. Solonnikov [213]. Also, Lemma 6.2 can be viewed as a
special case of the closed range lemma (e.g., see F. Brezzi and M. Fortin [79] and
K. Yosida [374]). Lemma 6.2 implies several important results.
First, it leads to the key condition in the mathematical theory of incompressible
materials known in the literature as the Inf-Sup (or LadyzhenskayaI. Babu` skaBrezzi
(LBB)) condition. The latter reads: there exists a positive constant C
LBB
such that
inf
q
1
2
(:)
q=0
sup
u1
0
u=0
_
D
q div nJ.
[q[ [Vn[
_ C
LBB
. (6.1.14)
Section 6.1 The Stokes problem 119
In fact, by Lemma 6.2 we knowthat for any q

1
2
(C) one can nd a function
q
V
0
satisfying the conditions
div
q
= q. [V
q
[ _ c
:
[q[. (6.1.15)
In this case,
sup
:1
0
(:)
u,=0
_
D
qdiv J.
[V[ [q[
_
_
D
qdiv
q
J.
[V
q
[ [q[
=
[q[
[V
q
[
_
1
c
:
and, consequently, (6.1.14) holds with C
LBB
= (c
:
)
-1
.
The condition (6.1.14) and its discrete analogs are used for proving the stability and
convergence of numerical methods in various problems related to the theory of viscous
incompressible uids. In I. Babu` ska [21] and F. Brezzi [77], this condition was proved
and used to justify the convergence of mixed methods, in which a boundary value
problem is reduced to a saddle-point problem for a certain Lagrangian. It is worth
noting, that conditions similar to (6.1.14) arise in many other problems if they are
stated as saddle point problems.
Also, (6.1.14) follows from the Ne cas inequality
[q[ _ c
:
[Vq[
-1,D
:= sup
u1
0
u=0
_
D
q div nJ.
[Vn[
. Vq

1
2
(C). (6.1.16)
A simple proof of the Inf-Sup condition for domains with Lipschitz boundaries can
be found in J. Bramble [71]. Estimates of the value of C
LBB
for various domains are
discussed in, e.g., E. Chizhonkov and M. Olshanskii [105], M. Dobrowolski [116], and
L. Halpern [165].
Saddle-point statement, the existence of a solution, and energy estimates. With
the help of Lemma 6.2, it is not difcult to prove the existence of u, p, and o that
deliver a solution to the problem (6.1.5)(6.1.7). For this purpose, we use general
theorems in convex analysis concerning saddle-points of Lagrangians. Consider the
Lagrangian 1 : (u
0
V
0
)

1
2
(C) R of the form
1(. q) :=
_
D
_
v
2
[V[
2
q div
_
J.
_
D
J.
and the saddle point problem
1(u. q) _ 1(u. p) _ 1(. p). V u
0
V
0
. q

1
2
(C). (6.1.17)
It is not difcult to verify that a saddle point is formed by the velocity eld u and the
pressure function p satisfying (6.1.5)(6.1.7).
Indeed, the left-hand side of (6.1.16) reads
_
D
(q p)div uJ. = 0. Vq

1
2
(C). (6.1.18)
Hence, div u is orthogonal to any function with zero mean. However,
_
D
div uJ. =
_
I
u nJs =
_
I
u
0
nJs =
_
D
div u
0
J. = 0.
so that fdiv ug
D
= 0. By (6.1.18), we now conclude that div u = 0.
Set = u n, where > 0 and n V
0
. Then, the right-hand side of (6.1.17)
means that
_
D
(vVu : Vn div n n) J. _
v
2
_
D
[Vn[
2
J..
Since can be taken arbitrarily small the above inequalities imply
_
D
(vVu : Vn div n) J. =
_
D
nJ.. Vn V
0
. (6.1.19)
If n

S
1
(C), then (6.1.11) follows from (6.1.19).
The saddle point problem (6.1.17) is equivalent to two variational problems
(P
u
) inf
u
0
V
0
sup
q
1
2
(D)
1(. q) and (P
p
) sup
q
1
2
(D)
inf
u
0
V
0
1(. q).
Since fdiv g
D
= 0 for any u
0
V
0
and
sup
q
1
2
(D)
_
D
q div J. = sup
q1
2
(D)
_
D
q div J. =
_
0 if div = 0.
o if div ,= 0.
we note that
inf
u
0
V
0
sup
q
1
2
(D)
1(. q) = inf
u
0
S
1
(D)
1() = 1(u).
where
1(n) =
_
D
_
1
2
[Vn[
2
n
_
J.. (6.1.20)
The Problem P
p
determines the pressure eld, however the functional of this prob-
lem cannot be represented in explicit form.
The existence of u and p follow from Lemma 6.2 and known theorems in the theory
of saddle-points. Obviously, 1 is convex and continuous with respect to the rst vari-
able. Moreover, it is linear and continuous with respect to the second one. Therefore
(e.g., see I. Ekeland and R. Temam [121], Chapter 4, 2), it sufces to show that
Jq
+

1
2
(C) such that lim
||-o
1(. q
+
) = o (6.1.21)
Section 6.1 The Stokes problem 121
and
lim
|q|-o
inf
u
0
V
0
1(. q) = o. (6.1.22)
Set q
+
= 0, then (6.1.21) is satised.
To prove (6.1.22), we use Lemma 6.2 and select n
q
V
0
such that
div n
q
= q and [Vn
q
[ _ c
:
[q[. (6.1.23)
Then
inf
u
0
V
0
1(. q) _ 1(u
0
zn
q
. q)
=
_
D
_
v
2
z
2
[Vn
q
[
2
z[q[
2
_
J.
_
D
zvVn
q
: Vu
0
J.
z
_
D
n
q
J. C(u
0
).
where
C(u
0
) =
_
D
_
v
2
[V(u
0
)[
2
u
0
_
J..
By (6.1.23), we conclude that
_
D
n
q
J.
_ C
T D
c
:
[q[[ [
and
_
D
v
2
z
2
[Vn
q
[
2
J. _
_
D
vc
2
:
2
z
2
[q[
2
J..
Hence, we nd that
inf
u
0
V
0
1(. q) _
_
D
_
v
2
z
2
c
2
:
z
_
[q[
2
J.
zc
:
[q[(v[Vu
0
[ C
T D
[ [) C(u
0
).
Set z =
1
vc
2
D
. Then, we note that
inf
uV
0
1(n. q)_
[q[
2
2vc
2
:
C(u
0
) o as [q[ o.
Thus, (6.1.22) holds and the saddle point problem (6.1.16) has a solution.
Now we establish energy estimates for the velocity and pressure functions. Set
n = u u
0
in (6.1.11). We have the relation
v[Vu[
2
=
_
D
(vVu : Vu
0
(u u
0
)) J.
_ (v[Vu
0
[ C
T D
[ [)[Vu[ C
T D
[Vu
0
[[ [.
From here, it follows that the velocity norm is bounded by C([Vu
0
[ [ [), where
C depends only on the given data. If u
0
= 0, then the energy estimate has the simplest
form
v[Vu[ _ C
T D
[ [. (6.1.24)
It is not difcult to show that a similar estimate holds for the pressure eld p. Let
p
V
0
be the function dened as a counterpart of p in Lemma 6.2, i.e.,
div
p
= p and [V
p
[ _ c
:
[p[.
Then, by (6.1.19), we have
_
D
(vVu : V
p

p
) J. =
_
D
p div
p
J. = [p [
2
. (6.1.25)
Hence,
[p[
2
_ (v[Vu[ C
T D
[ [)[V
p
[.
and we obtain
[p[ _ c
:
(C
T D
[ [ v[Vu[). (6.1.26)
Since [Vu[ is bounded (by the norms of given data), we note that the same is true for
[p[, so that the saddle point problem is stable with respect to both components.
Estimates of the distance to the set
V
S
1
./. Approximations computed by a numer-
ical procedure may not belong to the space

S
1
(C). Lemma 6.2 allows us to estimate
the distance between such an approximation and the set of solenoidal elds. Subse-
quently, we will use such estimates in the derivation of a posteriori estimates valid for
nonsolenoidal approximations.
First, we note that an estimate of the distance in 1
2
-norm follows from Lemma 3.2.
However, in the case of ow problems we need an estimate in a stronger norm given
by the lemma below.
Lemma 6.3. For any function V
0
, there exists a function
0

S
1
(C) such that
[V(
0
)[ _ c
:
[div[. (6.1.27)
Section 6.2 A posteriori estimates for the stationary Stokes problem 123
Proof. Let = div. By Lemma 6.2, we nd a function n
(
V
0
such that
div n
(
=
and
[Vn
(
[ _ c
:
[div[.
Since div ( n
(
) = 0, the function
0
:= n
(
belongs to

S
1
(C). Moreover,
[V(
0
)[ = [Vn
(
[ _ c
:
[div[
and the estimate (6.1.27) follows.
In other words, the distance between V
0
and the set of solenoidal elds

S
1
(C)
is estimated from above by the quantity [div[ with the multiplier c
:
that comes from
Lemma 6.2.
Corollary 6.4. Let u
V be a vector-valued function such that

div u
= 1
2
(C).
From Lemma 6.3 it follows that for any function u
V
0
there exists a function

S
1
(C) satisfying the estimate
[V(
)[ _ c
:
[div [. (6.1.28)
Proof. Since u
V
0
, we can nd a function
0

S
1
(C) such that
[V( u

0
)[ _ c
:
[div ( u
)[ = c
:
[div [.
Hence, the function
= u

0
belongs to u

S
1
(C) and satises (6.1.28).
6.2 A posteriori estimates for the stationary Stokes problem
In this section, we derive functional a posteriori estimates for the stationary Stokes
problem. For the sake of simplicity, we assume that v is constant. Estimates for
problems with variable viscosity are considered in Section 6.3.
6.2.1 Estimates for the velocity eld
Let u
0

S
1
(C). Then (6.1.8) implies the relation
_
D
vV(u ) : VnJ. =
_
D
( n vV : Vn) J.. Vn

S
1
(C). (6.2.1)
For any tensor-valued function t (C) = 1
2
(C. M
dd
), the functional
F
r,(
(n) :=
_
D
( n t : Vn) J.
is linear and continuous on V
0
, and its norm can be characterized by the quantity
[[[ [[[ F
r,(
[[[ [[[ := sup
uV
0
, u,=0
_
D
( n t : Vn) J.
[Vn[
.
The set
Q
r,(
(C) :=
_
t (C)
_
D
t : VnJ. =
_
D
nJ.. Vn V
0
_
denes the kernel of F
r,(
. It contains the tensor-valued functions that satisfy (in a
generalized sense) the equilibrium equation Div t = 0.
Represent the integral identity (6.2.1) in the form
_
D
vV(u ) : VnJ. = F
r,(
(n)
_
D
(t vV) : VnJ..
Since [F
r,(
(n)[ _ [[[ [[[ F
r,(
[[[ [[[ [Vn[ and
_
D
(t vV) : VnJ. =
_
D
(t q I vV) : VnJ. _ [t q IvV[[Vn[.
we set n = u and arrive at the estimate
v[V(u )[ _ [[[ [[[ F
r,(
[[[ [[[ [t Iq vV[. (6.2.2)
where q is an arbitrary function in

1
2
(C) and t is an arbitrary tensor-valued function
in (C). If t H(C. Div)
x
, then
_
D
( n t : Vn) J. =
_
D
( Div t) nJ.
and we nd that
[[[ [[[ F
r,(
[[[ [[[ _ C
T D
[ Div t[.
where C
T D
is the constant in (6.1.10).
Now, the right-hand side (6.2.2) is presented by directly computable integrals,
namely,
v[V(u )[ _ M
ST1
(. t. q) := [t q I vV[ C
T D
[Div t [. (6.2.3)
If q H
1
(C), then a somewhat different form of the estimate follows by changing t
to j, where
t = j q I. (6.2.4)
which gives
v[V(u )[ _ M
ST2
(. j. q) := [j vV[ C
T D
[Div j Vq[. (6.2.5)
Estimates (6.2.3) and (6.2.5) have a clear meaning. Estimate (6.2.3) shows that the
upper bound of the error can be represented as the sum of two parts related to the
decomposition of the Stokes system as
o = p I vVu.
Div o = .
Its right-hand side vanishes if and only the above relations are exactly satised. Since
is a solenoidal eld satisfying the boundary condition, the right-hand side of the
majorant is zero if and only if = u. Similarly, (6.2.5) shows that the upper bound
of the error can be represented as the sum of two parts related to the decomposition of
the Stokes system as
o = vVu.
Div o = Vp.
We can also deduce other equivalent forms of (6.2.3) and (6.2.5). Squaring both parts
of (6.2.3), we obtain an estimate the right-hand side of which is given by a quadratic
functional:
v
2
[V(u )[
2
_ M
2
ST1
(. ,. q. )
:= (1 )[t q I vV[
2
(1 )C
2
T D
[Div t [
2
. (6.2.6)
where is an arbitrary positive number. Rearrange the rst term on the right-hand
side by (1.4.2)(1.4.2). Since div = tr V = 0, we have
[t q I vV[
2
= [
1
J
(tr t qJ) I t
D
vV[
2
=
_
D
_
1
J
(tr t Jq)
2
[t
D
vV[
2
_
J.. (6.2.7)
If t is selected in such a way that
ftrtg
D
= 0. (6.2.8)
then we can set q =
1
d
(trt) and obtain
v
2
[V(u )[
2
= (1 )[t
D
vV[
2
(1 )C
2
T D
[Div t [
2
. (6.2.9)
Note that the right-hand side of (6.2.9) does not contain q. It vanishes if
t
D
vV = 0.
Div t = 0.
Since div = 0, we have
tr(t q I v(V)) = tr t Jq = 0.
and, therefore, the constitutive relation in terms of traces is also satised. Thus, we
conclude that this majorant also vanishes if only if , t, and q coincide with the exact
solutions.
Now consider the case, in which an approximate solution may not exactly satisfy
the divergence-free condition. We mark such approximations by hats. Assume that
u
0
V
0
and div may be not equal to zero. In this case, the estimate of its
deviation from u can be obtained by the following arguments.
First, by Corollary 6.4, we know that for one can nd a function n u
S
1
(C)
such that
[V( n)[ _ c
:
[div[. (6.2.10)
Therefore, we have
v[V(u )[ _ v[V(u )[ v[V( )[.
Use (6.2.3) to estimate the rst norm on the right-hand side of this inequality. We
arrive at the estimate
v[V(u )[ _ [t q I vV[ C
T D
[Div t [ v[V( )[
_ [t q I vV[ C
T D
[Div t [ 2v[V( )[.
In view of (6.2.10), we nd that
v[V(u )[ _ [t q I vV[ C
T D
[Div t [ 2vc
:
[div[. (6.2.11)
where t H(C. Div) and q

1
2
(C).
If ftrtg
D
= 0. then we can set q =
1
d
(trt) and obtain
v [V(u)[ = [t
D
v (V)
D
[ C
T D
[Div t [ 2vc
:
[div[. (6.2.12)
Thus, if the constants C
T D
and c
:
are known (or we know suitable upper bounds
for them), then (6.2.11) provides a way for evaluating the deviation of from u. For
this purpose, we should select certain nite-dimensional subspaces
k
and Q
k
for
the functions t (or j) and q, respectively. The minimization of the right-hand side of
(6.2.11) with respect to t and q provides an estimate of the deviation, which will be
the sharper the greater is the dimension of the subspaces used.
6.2.2 Estimates for pressure
Estimates of [p q[ can also be derived with the help of Lemma 6.2. Since
(p q)

1
2
(C), we know that
div n = p q and [V n[ _ c
:
[p q[
for a certain vector-valued function n V
0
. Hence,
[p q[
2
=
_
D
div n(p q) J..
Recall that the exact solution u satises (6.1.18) and, therefore,
_
D
(p q)div nJ. =
_
D
(vVu : V n n q div n) J.
=
_
D
vV(u ) : V nJ.
_
D
(vV : V n n q div n) J..
We have
_
D
vV(u ) : V nJ. _ c
:
v[V(u )[ [p q[
and
_
D
(vV : V n n q div n) J.
=
_
D
(vV t q I) : V nJ.
_
D
(Div t ) nJ.
_
_
[vV t q I[ vC
T D
[Div t [
_
c
:
[p q[.
Therefore,
[p q[ _ c
:
_
v[V(u )[ [vV t q I[ C
T D
[Div t [)
_ 2c
:
_
[vV t q I[ C
T D
[Div t [ vc
:
[div[)
_
.
and we arrive at the estimate
1
2c
:
[p q[ _ [vV t q I[ C
T D
[Div t [ vc
:
[div[. (6.2.13)
It is easy to note that the right-hand side of (6.2.13) consists of the same terms as the
right-hand side of (6.2.3) and vanishes if and only if,
= u. t = o. and p = q.
However, in this case, the dependence of the penalty multipliers on the constant c
:
is
stronger.
6.2.3 Estimates for stresses
Let t (C) be an approximation of o. We have
[t o[ = [t pI vVu[
_ [t q I vV[ v[V( u)[
_
J[p q[. (6.2.14)
By (6.2.11) and (6.2.13) we conclude that
[t o[ _ 2(1
_
Jc
:
)[t q I vV[ C
T D
(1 2
_
Jc
:
)[Div t [
2vc
:
(1
_
Jc
:
)[div[. (6.2.15)
Since
[t o[
2
Div
= [t o[
2
[Div t [
2
.
it is not difcult to estimate the deviation t o in the norm of H(C. Div). However,
the estimate has a more symmetric form if the deviation is expressed in terms of the
norm
| j |
Div
:= [j[ C
T D
[Div j[.
In this case,
c |t o|
Div
_ [t q I vV[ C
T D
[Div t [ vc
:
[div[. (6.2.16)
where c =
1
2(1
_
Jc
:
)
.
6.2.4 Estimates in combined norms
We can measure errors in terms of combined norms of the product space
W := (u
0

S
1
(C)) H(C. Div)

1
2
(C).
for which we introduce two equivalent norms
[(. t. q)[
W
:= v[V[ [t[
Div
_
J[q[.
|(. t. q)|
W
:= v[V[ |t|
Div
g
_
J[q[.
;
1
[(. t. q)[
W
_ |(. t. q)|
W
_ ;
2
[(. t. q)[
W
. (6.2.17)
where ;
1
= min{1. C
T D
] and ;
2
= max{1. C
T D
].
We can show that the majorant M
ST1
(. t. q) is equivalent to the error in the com-
bined norm |(u . o t. p q)|
W
.
We have
|o t|
Div
= |vVu t p I|
Div
_ v[V(u )[ [vV t p I[ C
T D
[Div t [
_ v[V(u )[ [vV t q I[
_
J[p q[ C
T D
[Div t [.
Therefore,
|(u . o t. p q)|
W
:= v[V(u )[ |o t|
Div
_
J[p q[
_ 2v[V(u )[ [vV t q I[
2
_
J[p q[ C
T D
[Div t [.
Since u
0
S
1
(C), we can use (6.2.3) and (6.2.13) to estimate the terms [V(u)[
and [p q[. We obtain
|(u . o t. p q)|
W
_ C
_
[vV t q I[ C
T D
[Div t [
_
= C
M
ST1
(. t. q). (6.2.18)
where C
= 3 4
_
Jc
:
. On the other hand,
M
ST1
(. t. q) _ v[V( u)[ |o t|
Div
_
J[p q[. (6.2.19)
Thus, we nd that
M
ST1
(. t. q) _ |(u . o t. p q)|
W
_ C
M
ST1
(. t. q). (6.2.20)
In view of (6.2.17), the majorant M
ST1
(. t. q) is also equivalent to the combined norm
[(u . o t. p q)[
W
.
Remark 6.5. It is easy to show that the majorant M
ST2
(. j. q) is also equivalent to the
combined error norm. Moreover, one can prove that on a wider set
W := (u
0
V
0
) H(C. Div)

1
2
(C).
the majorant
M
cT1
(. t. q) := [t q IvV[ C
T D
[Div t [ 2vc
:
[div[
is also equivalent to the combined error norm |(u. ot. pq)|
W
.
6.2.5 Lower bounds of errors
A lower bound of [V(u )[ is derived by the same arguments that we have used in
Section 4.1. For u
0
V
0
, we have
v
2
[V(u )[
2
=
_
D
_
vV(u ) : V(u )
v
2
V(u ) : V(u )
_
J.
_ sup
uV
0
_
D
_
vV(u ) : V n
v
2
V n : V n
_
J.
_ sup
r(D)
_
D
_
vV(u ) : t
v
2
t : t
_
J. =
v
2
[V(u )[
2
.
Thus,
v[V(u )[
2
= sup
uV
0
_
D
(2vV(u ) : V n vV n : V n) J..
By (6.1.12), we reform the right-hand side of this relation and obtain
v[V(u )[
2
= sup
uV
0
{G(. n. p)].
where G is a quadratic functional dened by the relation
G(. n. p) =
_
D
_
2vV : V n v[V n[
2
2 n 2p div n
_
J..
Thus, for any n V
0
v[V(u )[
2
_ G(. n. q) 2[p q[[div n[.
We estimate the last term by (6.2.13) and arrive at the estimate
v[V(u )[
2
_ M
ST
(. n. t)
:= G(. n. q) 4c
:
_
[vV t q I[
C
T D
[Div t [ vc
:
[div[
_
[div n[. (6.2.21)
If the trial functions are taken from a narrower set

S
1
(C), then the last-mentioned term
in (6.2.21) vanishes, and we nd that
v[V(u )[
2
_ G(. n. q). Vn

S
1
(C). q

1
2
(C). (6.2.22)
Since
_
D
vVu : V(u ) J. =
_
D
( (u ) p div (u )) J.. (6.2.23)
we note that
G(. u . p)
=
_
D
_
2vV : V(u ) v[V(u )[
2
2 (u ) 2pdiv (u )
_
J.
= v[V( u)[
2
. (6.2.24)
Hence,
v[V(u )[
2
= sup
u
S
1
(:)
q
1
2
(:)
G(. n. q)
_
. (6.2.25)
6.2.6 Mixed boundary conditions
Now, we consider a more general statement of the Stokes problem (6.1.5) and (6.1.6)
with mixed DirichletNeumann boundary conditions
u = u
0
on I
1
. (6.2.26)
o n = Jm on I
2
. (6.2.27)
where I
1
and I
2
of I are two measurable nonintersecting parts of the boundary, u
0
is
a given function such that div u
0
= 0, and J 1
2
(I
2
. R
d
). In this section, we dene
the space V
0
as follows:
V
0
(C) :=
_
H
1
(C. R
d
) [ = 0 on I
1
_
.
Let S
1
0
(C) be a subspace of V
0
(C) formed by solenoidal functions. A generalized
solution is a function u S
1
0
(C) that satises the integral identity
_
D
vVu : VnJ. = (n). Vn S
1
0
(C). (6.2.28)
Here, : V
0
(C) R is the functional
(n) :=
_
D
n J.
_
I
2
J nJs.
[(n)[ _ C
I
[Vn[. Vn V
0
(C). (6.2.29)
Note that C
I
depends on C and I
2
and on the constants in the respective Friedrichs
and trace inequalities (for the functions vanishing on I
1
). For any t (C), the
functional
F
r,I
(n) := (n)
_
D
v t : VnJ.
is linear and continuous on V
0
(C). Its norm is dened by the relation
[[[ [[[ F
r,I
[[[ [[[ := sup
uV
0
(D)
[F
r,I
(n)[
[Vn[
_ C
I
v[t[. (6.2.30)
The set K
r,I
= Ker F
r,I
contains tensor-valued functions that satisfy (in a generalized
sense) the equilibrium equation
Div t = 0 in C (6.2.31)
and the boundary condition
t n = J on I
2
. (6.2.32)
Theorem 6.6. For any u
0
S
1
0
(C), q

1
2
(C), and t (C), the following
estimate holds:
v [V(u )[ _ [t q I vV[ [[[ [[[ F
r,I
[[[ [[[ . (6.2.33)
If
t H
T
2
(C. Div) :=
_
t [ Div t 1
2
(C. R
d
). t n 1
2
(I
2
. R
d
)
_
.
then the majorant for the Stokes problem with mixed boundary conditions is given by
the relation
v [V(u )[ _ M
STm
(. t. q)
:= [t q I vV[ C
T D
[Div t [
C
TI
2
[J t n[
I
2
. (6.2.34)
where C
T D
and C
TI
2
are the constants in the inequalities
[n[ _ C
T D
[Vn[. Vn V
0
(C). (6.2.35)
[n[
I
2
_ C
TI
2
[Vn[. (6.2.36)
Proof. From (6.2.28) we conclude that for any n S
1
0
(C)
_
D
vV(u ) : VnJ. = (n)
_
D
vV : VnJ..
Let t . Then,
_
D
vV(u ) : VnJ. =
_
D
(t vV) : VnJ. (n)
_
D
vt : VnJ.
_ ([t q I vV[ [[[ [[[ F
r,I
[[[ [[[ ) [Vn[ .
where q is an arbitrary function in

1
2
(C). By setting n = u we arrive at the
estimate (6.2.33).
Assume that t H
T
2
(C. Div). Then,
F
r,I
(n) =
_
D
(Div t ) nJ.
_
I
2
(J t n) nJs
_ [Div t [[n[ [J t n[[n[
I
2
_
_
C
T D
[Div t [ C
TI
2
[J t n[
I
2
_
[Vn[
and, therefore,
[[[ [[[ F
r,I
(n) [[[ [[[ _ C
T D
[Div t [ C
TI
2
[J t n[
I
2
. (6.2.37)
Now (6.2.34) follows from (6.2.33) and (6.2.37).
The functional M
STm
(. t. q) is directly computable, provided that the constants
C
T D
and C
TI
2
(or their upper bounds) are known. It vanishes if and only if
t = q I vV
and the relations Div t = 0 in C and t n = J on I
2
hold almost everywhere.
Since meets the Dirichlet boundary condition on I
1
and satises the relation
div = 0, we conclude that in such a case = u and t and q coincide with the
exact stress and pressure elds, respectively.
Remark 6.7. A modication of the above a posteriori estimate is obtained if ftrtg = 0
and we set q =
1
d
trt. Then (6.2.34) implies the estimate
v [V(u )[ _ [t
D
vV
D
[ C
T D
[Div t [ C
TI
2
[J t n[
I
2
. (6.2.38)
which that does not contain q. Note that if the right-hand side of (6.2.38) vanishes,
then t n = J on C and
Div (t
D
1
J
trtI) = 0 in C.
In addition, t
D
v(V)
D
= 0 and q =
1
d
tr t, so that
t = q I v(V)
D
.
Since u
0
S
1
0
(C), we conclude that t = qI vV satises the equilibrium
equation and, therefore, t coincides with o, with u, and q =
1
d
trt coincides with
p (up to a constant).
A lower bound of the error can be derived by arguments similar to those used in
Sections 3.2 and 4.1. It has the form
[[V(u )[[
2
_ 2(n)
_
D
_
v[Vn[
2
2vV : Vn
_
J.. Vn S
1
0
(C).
To derive estimates for approximations in u
0
V
0
(C) we rst prove an assertion
below, which can be viewed as a generalization of Lemma 6.3.
Lemma 6.8. Assume that

V
0
(C) := { V
0
(C) [ fdiv g
D
= 0].
Then, there exists
0
S
1
0
(C) such that
[V(
0
)[ _ c
:
[div [. (6.2.39)
Proof. For any a H
1{2
(I. R
d
) satisfying the condition
_
I
a nJs = 0 there exists
a solution n
o
of the Stokes problem
^n
o
Vp = 0.
div n
o
= 0 in C.
n
o
a = 0 on I.
Let a be the trace of

V
0
(C) on I. Then, n
o
V
0
. By Lemma 6.2 we know
that there exists n
0

S
1
(C) such that
[V(n
o
) Vn
0
[ _ c
:
[div (n
o
)[ = c
:
[div [.
This estimate means that
[V V(n
0
n
o
)[ _ c
:
[div [.
where the function
0
= n
0
n
o
is solenoidal and
0
= 0 on I
1
.
0
V
0
(C) such that
fdivg
D
= 0. (6.2.40)
q

1
2
(C), and t (C), the following estimate holds:
v [V(u )[ _ [t q I vV[ [[[ [[[ F
r,I
[[[ [[[ 2vc
:
[div[. (6.2.41)
If t H
T
2
(C. Div) then
v [V(u )[ _ [t q I vV()[ C
T D
[Div t [
C
TI
2
[t n J[
I
2
2vc
:
[div[. (6.2.42)
Proof. Let n := u
0
. This function belongs to V
0
(C). In view of (6.2.40)
fdiv ng
D
= fdiv div u
0
g
D
= 0.
so that n

V
0
(C). By Lemma 6.8, there exists a function
0
S
1
0
(C) such that
[V( n
0
)[ _ c
:
[div[. (6.2.43)
We have
v [V(u )[ = v [V(u n u
0
)[ (6.2.44)
_ v [V(u
0
u
0
)[ v [V( n
0
)[ .
Since div (
0
u
0
) = 0, we estimate the rst norm by (6.2.33) and nd that
v [V(u )[ _ [t q I vV(
0
u
0
)[ [[[ [[[ F
r,I
[[[ [[[ v [V( n
0
)[
_ [t q I vV[ [[[ [[[ F
r,I
[[[ [[[ 2v [V( n
0
)[ .
By (6.2.43), we obtain (6.2.41). Estimate (6.2.42) follows from (6.2.37) and (6.2.41).
Estimate (6.2.42) has the same principal structure as (6.2.4). The only difference is
that a new term C
TI
2
[t nJ[
I
2
arises. It serves as a penalty for a possible violation
of the Neumann boundary condition.
Remark 6.10. If ftrtg
D
= 0, then the pressure can be excluded, and we have the
estimate
v [V(u )[ _
_
_
t
D
v(V)
D
_
_
C
T D
[Div t [
C
TI
2
[t n J[
I
2
2vc
:
[div[. (6.2.45)
Now, our aim is to derive an upper bound of [p q[.
Let
|
V
0
(C) be a vector-valued function such that div
|
= 1 in C. We note that
there are many functions with such properties. Indeed, the nonhomogeneous Stokes
problem
^ V p = 0.
div = 1 in C.
= 0 on I
1
.
= a on I
2
.
_
I
2
a nJs = [C[
has a solution (e.g., see [348]). The latter can be taken as
|
.
Theorem 6.11. Let q

1
2
(C) be an approximation of the pressure eld p. Then
1
2c
|
:
[p q[ _ [vV t q I[ C
T D
[Div t [
C
TI
2
[t n J[
I
2
vc
:
[div[. (6.2.46)
where c
|
:
= c
:
[C[
-1{2
_
_
V
|
_
_
, and t are arbitrary functions in

V
0
(C) and
H
T
2
(C. Div), respectively.
Proof. Since
p q fp qg
D
= A p q

1
2
(C).
we apply Lemma 6.2 and conclude that there exists a function n
0
V
0
(C) (note that
n
0
vanishes on the boundary) such that
div n
0
= A p q (6.2.47)
and
[Vn
0
[ _ c
:
[A p q[. (6.2.48)
It is easy to see that n
|
0
:= n
0
fp qg
D
|
V
0
(C),
_
D
div n
|
0
(p q) J. =
_
D
(div n
0
(p q) div
|
fp qg
D
(p q)) J.
= [A p q[
2
[C[fp qg
2
= [p q[
2
. (6.2.49)
and
[Vn
|
0
[ _ [Vn
0
[ fp qg
D
[V
|
[ (6.2.50)
_ c
:
[A p q[ fp qg
D
[V
|
[ _ c
|
:
[p q[.
where c
|
:
= c
:
[C[
-1{2
[V
|
[.
Now, we use the integral identity
_
D
vVu : Vn
|
0
J. = (n
|
0
)
_
D
pdiv n
|
0
J. (6.2.51)
and rearrange (6.2.49) as follows:
[p q[
2
=
_
D
div n
|
0
(p q) J.
=
_
D
_
vVu : Vn
|
0
q div n
|
0
_
J. (n
|
0
)
=
_
D
vV(u ) : Vn
|
0
J.
_
D
_
vV : Vn
|
0
q div n
|
0
_
J. (n
|
0
).
In view of (6.2.50),
_
D
vV(u ) : Vn
|
0
J. _ c
|
:
v [V(u )[ [p q[
and
_
D
_
vV : Vn
|
0
q div n
|
0
_
J. (n
|
0
)
=
_
D
(vV t q I) : Vn
|
0
J.
_
D
(Div t ) n
|
0
J.
_
I
2
(t n J) n
|
0
Js
_
_
[vV t q I[ C
T D
[Div t [ C
TI
2
[J t n[
I
2
_
[Vn
|
0
[
_ c
|
:
_
[vV t q I[ C
T D
[Div t [ C
TI
2
[J t n[
I
2
_
[p q[.
Therefore,
[p q[ _ c
|
:
_
v [V(u )[ [vV t q I[
C
T D
[Div t [ C
TI
2
[J t n[
I
2
_
.
Now, we apply (6.2.42) and obtain the estimate
[p q[ _ 2c
|
:
_
[t q I vV()[ C
T D
[Div t [
C
TI
2
[t n J[
I
2
vc
:
[div[
_
.
which is equivalent to (6.2.46).
Remark 6.12. The constant c
|
:
contains the norm of a subsidiary function
|
, which
must satisfy the condition div
|
= 1 and
|
= 0 on I
1
. Usually, such a function is
not difcult to construct. For example, for polygonal domains
|
can be constructed
with the help of RaviartThomas elements of the lowest order. It is desirable to have a
function
|
such that [V
|
[ is as small as possible.
6.2.7 Problems for almost incompressible uids
In models of almost incompressible uids, the incompressibility condition is replaced
by a term that contains the divergence with a large multiplier. The respective pe-
nalized version of the stationary Stokes problem with Dirichlet boundary conditions
reads: Find u
u
0
V
0
(where V
0
=

H
1
(C. R
d
)) satisfying the integral identity
_
D
_
vVu
: Vn
1
div u
div n
_
J. =
_
D
nJ.. n V
0
(C). (6.2.52)
From (6.2.52) we nd that
v[Vu
[ _ C
T D
[ [. (6.2.53)
[div u
[
2
_ [ [[u
[ _ C
T D
[ [[Vu
[ _
C
2
T D
v
[ [
2
. (6.2.54)
We can deduce an estimate of the difference between u and u
. For this purpose, we

use (6.2.11) with = u
. Other functions in the right-hand side of (6.2.11), we dene

as follows:
t = t
:= vVu
div u
I and q =
:=
1
div u
.
_
D
t
: VnJ. =
_
D
nJ.. n V
0
(C). (6.2.55)
so that Div t
= 0 almost everywhere in C. Moreover,

t
I vVu
= 0.
Thus, by (6.2.11) we conclude that
1
2c
:
[V(u u
)[ _ [div u
[. (6.2.56)
By (6.2.13)
1
2vc
2
:
[
[ _ [div u
[. (6.2.57)
We note that the difference between the exact solutions of the Stokes problem and its
penalized counterpart is controlled by the 1
2
-norm of the divergence of the problem.
Moreover, from (6.2.54), (6.2.56), and (6.2.57) it follows that
[V(u u
)[ _ 2c
:
C
T D
_
v
[ [. (6.2.58)
[
[ _ 2c
2
:
C
T D
_
v[ [. (6.2.59)
These estimates show that u u
and
(this fact is known, see, e.g.,

R. Temam [348]). Moreover, we nd that the sequences converge with the rate no
less than
1{2
.
A similar estimate can be obtained for approximations constructed by means of the
Uzawa algorithm.
6.2.8 Problems with the condition div u D
In some cases, the following version of the Stokes problem is considered
Div o = Vp in C. (6.2.60)
o = vVu in C. (6.2.61)
u = u
0
on I. (6.2.62)
div u = in C. (6.2.63)
where is a given function in 1
2
and u
0
satises the compatibility condition
_
D
J. =
_
I
u
0
nJs. (6.2.64)
Let u
u
0
V
0
be a function satisfying (6.2.63). We set
u = u
u. u

S
1
(C).
Then (6.2.60)(6.2.63) is reduced to the Stokes problem with
= vû
H
-1
.
instead of . In other words, the problem is to nd u

S
1
(C) such that
_
D
vV u VnJ. =
_
D
( n t
: Vn p div n) J.. Vn V
0
. (6.2.65)
where t
= vVu
.
Let u be approximated by = u
, where

S
1
(C). By (6.2.65), we nd that
for any n

S
1
(C),
_
D
vV( u ) VnJ. =
_
D
( n t
: Vn vV : Vn q div n) J.
_
_
D
( Div t) nJ.
_
D
(t t
vV q I) : VnJ.. (6.2.66)
Since u

S
1
(C), we can set n = u . Then, (6.2.66) implies the estimate
v[V( u )[ _ [t q I vV t
[ C
T D
[Div t [. (6.2.67)
Since u = u , we have
v[V(u )[ _ [t q I vV[ C
T D
[Div t [. (6.2.68)
If = u
and V
0
(so that div ,= ), then (using the same arguments as
before) we deduce an estimate analogous to (6.2.11):
v[V(u )[ _ [t q I vV[ C
T D
[Div t [ 2vc
:
[div [. (6.2.69)
For = 0 this estimate coincides with (6.2.11).
6.3 Generalized Stokes problem
Various generalized statements of the Stokes problem are motivated by semidiscrete
formulations of evolutionary problems, in which u(.. t ) is replaced by a sequence
of approximations u
k
(.) representing u(.. t
k
), where t
k
, k = 0. 1. . . . . M are some
selected values of the time-variable t . For example, the scheme
_
_
_
u
k
u
k-1
k
vû
k
div (u
k-1
u
k-1
) = Vp
k
in C.
div u
k
= 0.
k
= t
k
t
k-1
.
(6.3.1)
leads to a stationary problem for u
k
. It differs from the Stokes problem by the presence
of the term ju (where j =
1
k
). In other cases, additional terms in the basic equation
(6.1.1) arise owing to some physical phenomena.
In this section, we rst consider the system
ju vû = Vp in C. (6.3.2)
div u = 0 in C. (6.3.3)
As before, the function u additionally satises the prescribed boundary condition
u = u
0
(u
0

S
1
(C)). In general, both parameters v > 0 and j _ 0 may be
large or small. Henceforth, we assume that j and v are positive functions such that
j j
. j
| and v v
. v
|.
Moreover, we assume that v
= 1 (this assumption does not lead to loss in generality,

because after a proper scaling (6.3.2)(6.3.3) can always be transformed to a system
that satises it).
For our purposes, it is convenient to state the generalized Stokes problem in the
so-called three-eld setting: Find u, p, and a tensor-valued function o (stress) such
that
Div o ju = in C. (6.3.4)
o = p I vVu in C. (6.3.5)
div u = 0 in C. (6.3.6)
u = u
0
on I. (6.3.7)
The velocity eld u u
0

S
1
(C) of (6.3.4)(6.3.5) is dened by the integral identity
_
D
(vVu : Vn ju n) J. =
_
D
nJ.. n

S
1
(C). (6.3.8)
or by the variational problem for the functional
inf
uu
0
S
1
(D)
1(n) = 1(u). (6.3.9)
Section 6.3 Generalized Stokes problem 141
where
1(n) =
_
D
_
v
2
[Vn[
2
j
2
[n[
2
n
_
J. (6.3.10)
is the corresponding energy functional. The existence and uniqueness of u is proved
by the same arguments as for the Stokes problem.
Our goal is to deduce two-sided bounds of the error evaluated in terms of the norm
[n[
2
:=
_
D
(v[Vn[
2
j[n[
2
) J..
which is the natural energy norm.
6.3.1 Estimates for solenoidal approximations
Let u
0

S
1
(C). Then (6.3.8) implies the relation
_
D
(vV (u ) : Vn j(u ) n) J.
=
_
D
( n j n vV : Vn) J.. (6.3.11)
which holds for any n

S
1
(C). Let t be a tensor-valued function in H(C. Div).
Since n vanishes on the boundary, we rewrite (6.3.11) as follows:
_
D
(vV (u ) : Vn j(u ) n) J. = F(n: . t. q). (6.3.12)
where F(n: . t. q) is a linear functional with respect to n dened by the relation
F(n: . t. q) :=
_
D
(( j Div t) n (t q I vV) : Vn) J.
and q

1
2
(C). Henceforth, we denote
r(. t) := j Div t and d(. t. q) := t q I vV.
Represent the rst term of F(n: . t. q) in the form
_
D
r(. t) nJ. =
_
D
_
j
r(. t)
_
jnJ.
_
D
(1 )r(. t) nJ..
where = (.) is a real function with values in 0. 1|. We have
_
D
r(. t) nJ.
_
_
_
_
_
_
j
r(. t)
_
_
_
_
_
_
_
jn
_
_
C
D
[(1 )r(. t)[ [
_
v Vn[.
where C
D
is a constant in the Friedrichs type inequality
[n[
2
_ C
2
D
_
D
v(.)[Vn[
2
J.. Vn V
0
(C). (6.3.13)
Since
[n[ _ C
T D
[Vn[ _
C
T D
_
v
[
_
vVn[.
we can set C
D
=
C
1:
_
_
(if c
= 1, then C
D
= C
T D
). Next,
_
D
d(. t. q) : VnJ.
_
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
[
_
v Vn[.
Thus, we obtain
[F(n: . t. q)[
_
_
C
D
[(1 )r(. t)[
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
_
[
_
vV n[
_
_
_
_
_
j
r(. t)
_
_
_
_
[
_
jn[
_
_
_
C
D
[(1)r(. t)[
_
_
_
_
d(. t. q)
_
v
_
_
_
_
_
2
_
_
_
_
r(. t)
_
j
_
_
_
_
2
_
1{2
[n[
. (6.3.14)
Set n = u . Then, we arrive at the following estimate for the generalized Stokes
problem:
[u [
2
_ M
GST
(. t. q) :=
_
_
_
_
_
j
r(. t)
_
_
_
_
2
_
C
D
[(1 )r(. t)[
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
_
2
. (6.3.15)
This estimate is valid for any 0. 1|, t H(C. Div), and q

1
2
(C). If j = 0
and v = const, then we set = 0 and arrive at (6.2.3).
Two particular forms of this estimate related to the choice = 0 and = 1 are as
follows:
[u [
_ C
D
[r(. t)[
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
=: M
0 GST
(. t. q) (6.3.16)
and
[u [
2
_
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
2
_
_
_
_
1
_
j
r(. t)
_
_
_
_
2
=: M
1 GST
(. t. q). (6.3.17)
By (6.3.4) and (6.3.5) we nd that
_
_
_
_
1
_
v
(o p I vV)
_
_
_
_
2
=
_
D
v[V(u )[
2
J.
and
_
D
1
j
[ j Div o[
2
J. =
_
D
j[u [
2
J..
Therefore,
M
1 GST
(. o. p) = [u [
and the majorant M

1 GST
(. t. q) can provide a sharp upper bound of the error if t and
q are properly chosen. However, it has an essential drawback: if j is small, then the
second term involves a large multiplier, which makes the whole estimate sensitive to
the residual r(. t). In practice, this may lead to a considerable overestimation of the
error.
The majorant M
0 GST
(. t. q) does not contain large parameters, but we cannot prove
that
inf
r,q
M
0 GST
(. t. q) = [u [
.
In other words, there may be an inherent gap between the left- and right-hand sides of
(6.3.17).
In order to avoid the above difculties and to obtain an estimate, which possesses
positive features of the above estimates without bad ones, we derive another upper
bound for the deviation u . For this purpose, we apply the same method as was
used in Chapter 4 for the reaction-diffusion problem.
Let us estimate the rst term on the right-hand side of (6.3.14) and rewrite it in the
form
[u [
2
_ C
2
D
(1 ) [(1 )r(. t)[
2
_
_
_
_
_
j
r(. t)
_
_
_
_
2
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
2
.
where is an arbitrary positive number. The minimum of the right-hand side with
respect is attained if
= H(. C
D
. j) :=
C
2
D
j(1 )
C
2
D
j(1 ) 1
0. 1).
which leads to another upper bound of the error for the generalized Stokes problem:
[u [
2
_
_
D
H(. C
D
. j)
j
r
2
(. t) J.
1
_
_
_
_
1
_
v
d(. t. q)
_
_
_
_
2
=: M
2
GST
(. t. q). (6.3.18)
Since
r(. o) = j(u ). d(. o. p) = o p I vV = vV(u ).
we note that
M
2
GST
(. o. p) =
_
D
_
C
2
D
(1 )
C
2
D
j(1 ) 1
j
2
( u)
2
v[V( u)[
2
_
J.
and
M
2
GST
(. o. p) [u [
2
as o.
Therefore, (6.3.18) has no gap between its left- and right-hand sides. At the same time,
the structure of the rst term of (6.3.18) is such that it is not sensitive to small values
of j.
If j = 0 and v = const, then (6.3.18) implies the estimate
v[V(u )[
2
_ C
2
D
(1 ) [ Div t[
2
_
_
_
_
d(. t. q)
_
v
_
_
_
_
2
. (6.3.19)
which gives an upper bound of the error for the generalized Stokes problem
(cf. (6.2.3)).
If j = const then we minimize M
2
GST
(. t. q) with respect to and set
=
_
_
D
C
D
jC
2
D
1
R jC
D
D
if R > jC
D
D.
o if R _ jC
D
D.
(6.3.20)
where
D :=
_
_
_
_
d(. t. q)
_
v
_
_
_
_
and R := [r(. t)[ .
Then, we arrive at the estimate
[u [
2
_
_
_
1
1 jC
2
D
(C
D
R D)
2
if R > jC
D
D.
1
j
R
2
D
2
if R _ jC
D
D.
(6.3.21)
Note that the second branch of (6.3.21) is bounded by the quantity D(C
D
RD) and,
therefore, (6.3.21) does not blow up if j 0.
6.3.2 Estimates for nonsolenoidal elds
Let u
0
V
0
be such that div ,= 0. To derive an upper bound of [u [
, we
use the same arguments as for the Stokes problem.
By Lemma 6.2, we know that there exists
0
u
0

S
1
(C) such that
[V(
0
)[ _ c
:
[div[.
Therefore,
[
0
[
2
_ k
2
D
[div[
2
. k
2
D
:= c
2
:
(v
C
2
T D
). (6.3.22)
and
[u [
_ [u [
[ [
. V u
0

S
1
(C). (6.3.23)
To estimate the rst norm on the right-hand side of (6.3.23), we apply (6.3.18) (assum-
ing for the sake of simplicity that j is a constant) represented in the form
[u [
2
_
H(. C
D
. j)
j
[r(. t)[
2
_
_
_
_
d(. t. q)
_
v
_
_
_
_
2
. (6.3.24)
Let
2
() = max
_
H(. C
D
. j).
1
_
.
Since
[r(. t)[
2
_ [r(. t)[
2
2j[r(. t)[[ [ j
2
[ [
2
.
_
_
_
_
d(. t. q)
_
v
_
_
_
_
2
_
_
_
_
_
d(. t. q)
_
v
_
_
_
_
2
2
_
_
_
_
d(. t. q)
_
v
_
_
_
_
[
_
vV( )[ [
_
vV( )[
2
.
and
H(. C
D
. j)[r(. t)[[ [
1
_
_
_
_
d(. t. q)
_
v
_
_
_
_
[
_
vV()[
_
_
H(. C
D
. j)
j
[r(. t)[
_
H(. C
D
. j)[
_
j( )[
_
_
_
_
d(. t. q)
_
v
_
_
_
_
[
_
vV()[
_ () M
GST
(. t. q)[ [
.
we nd that
[u [
_
_
M
2
GST
(. t. q) 2()M
GST
(. t. q)[ [
2
()[ [
2
_
1{2
[ [
. (6.3.25)
Set =
0
and apply (6.3.22). We obtain
[u [
_
_
M
2
GST
(. t. q) 2()k
D
M
GST
(. t. q)[div[
2
()k
2
D
[div[
2
_
1{2
k
D
[div[. (6.3.26)
With the help of positive parameters ; and , we represent (6.3.26) in quadratic form.
Set =
0
and apply (6.3.22). We obtain
[u [
2
_ (1 ;)(1 )M
2
GST
(. t. q)
(1 ;)
_
2
()

2
()
1
;
_
k
2
D
[div[
2
. (6.3.27)
If u
0

S
1
(C), then div = 0. In this case, we set ; = = 0 and see that
(6.3.27) is converted to (6.3.24).
6.3.3 Estimates for the pressure eld
Let q

1
2
(C) be an approximation of the pressure eld p. Then, there exists a
function n V
0
such that
div n = p q. and [V n[ _ c
:
[p q[. (6.3.28)
Therefore,
[ n[
_ k
D
[p q[. (6.3.29)
In view of (6.3.28), we note that
[p q[
2
=
_
D
div n(p q) J..
By the relation
_
D
div np J. =
_
D
(vVu : V n ju n n) J..
we obtain
[p q[
2
=
_
D
(vV(u ) : V n j(u ) n) J.
_
D
(vV : V n j n q div n n) J.
=
_
D
(vV(u ) : V n j(u ) n) J. F( n: . t. q). (6.3.30)
where is an arbitrary function in u
0

S
1
(C).
We have
_
D
(vV(u ) : V n j(u ) n) J. _ [u [
[ n[
.
By (6.3.14),
[F( n: . t. q)[
_
_
_
C
D
[(1 )r(. t)[
_
_
_
_
d(. t. q)
_
v
_
_
_
_
_
2
_
_
_
_
r(. t)
_
j
_
_
_
_
2
_
1{2
[ n[
.
Therefore, from (6.3.30) we nd that
[p q[ _ k
D
[u [
k
D
_
_
C
D
[(1)r(. t)[
_
_
_
_
d(. t. q)
_
v
_
_
_
_
_
2
_
_
_
_
r(. t)
_
j
_
_
_
_
2
.
In view of (6.3.15), we obtain
1
4k
2
D
[p q[
2
_
_
C
D
[(1 )r(. t)[
_
_
_
_
d(. t. q)
_
v
_
_
_
_
_
2
_
_
_
_
r(. t)
_
j
_
_
_
_
2
. (6.3.31)
In other words, the upper bound is given by the same expression as in (6.3.15) but with
multiplier 2k
D
. Since (.) is in our disposal, we can select it, e.g., as
(.) = min{j(.). 1], which makes the estimate robust with respect to small values
of j.
If we set (.) = 1, then (6.3.31) reads
1
4k
2
D
[p q[
2
_
_
_
_
_
d(. t. q)
_
v
_
_
_
_
2
_
_
_
_
r(. t)
_
j
_
_
_
_
2
. (6.3.32)
If (.) = 0, then we have another estimate:
1
2k
D
[p q[ _ C
D
[r(. t)[
_
_
_
_
d(. t. q)
_
v
_
_
_
_
. (6.3.33)
Let j = 0 and v = const. In this case, k
D
=
_
vc
:
and (6.3.33) is converted into
1
2
_
vc
:
[p q[ _
_
_
_
_
d(. t. q)
_
v
_
_
_
_
C
D
[r(. t. q)[. (6.3.34)
Since C
D
=
C
1:
_
and div = 0, this estimate coincides with (6.2.13) derived for the
Stokes problem.
6.3.4 Error minorant
In the case considered, lower bounds can be derived by variational arguments based
on (6.1.28). Assume that u
0

S
1
(C). For any n

S
1
(C), we have
1
2
[u [
2
= 1() 1(u) _ 1() 1( n)

=
_
D
_
v
2
[Vn[
2
j
2
[n[
2
vV : Vn j n n
_
J..
1
2
[u [
2
= sup
u
S
1
(D)
_
D
_
v
2
[Vn[
2
j
2
[n[
2
vV : Vn j n n
_
J.. (6.3.35)
To prove this, it sufces to take n = u and use the relation
_
D
v(Vu : V (u ) ju (u )) J. =
_
D
(u ) J..
Therefore, the maximization of the right-hand side of (6.3.35) with respect to a certain
nite-dimensional subspace of

S
1
(C) gives a computable lower bound of the error
norm.
6.3.5 Models with polymerization
Another version of the generalized Stokes problem is related to models of uids with
polymerization (e.g., see J. Bonvin, M. Picasso, and R. Stenberg [62]). It can be
represented in the form
Div o = 0 in C. (6.3.36)
div u = 0 in C. (6.3.37)
o = o
0
p I vVu in C. (6.3.38)
u = 0 on I. (6.3.39)
where o
0
is a given tensor-valued function such that tr o
0
= 0. Note that (6.3.37)
decomposes o into spherical and deviatoric parts, respectively.
The generalized solution u of the system (6.3.36)(6.3.39) is a function in

S
1
(C)
satisfying the integral identity
_
D
(vVu : Vn o
0
: Vn) J. =
_
D
nJ.. n

S
1
(C). (6.3.40)
For an approximation

S
1
(C), we have
_
D
vV (u ) : VnJ. =
_
D
( n o
0
: Vn vV : Vn) J.
=
_
D
_
( Div t) n (t q I o
0
vV) : Vn
_
J.. (6.3.41)
where t H(C. Div) and q

1
2
(C).
From (6.3.41) it follows that (if v is a constant)
v[V(u )[ _ [t q I o
0
vV[ C
T D
[ Div t[. (6.3.42)
If an approximation belongs to a wider set V
0
, then we apply the same arguments as
for the Stokes problem and deduce the estimate
v[V(u)[ _ [t q I o
0
vV[ C
T D
[ Div t[ 2vc
:
[div[. (6.3.43)
6.3.6 Models with rotation
In certain models, the NavierStokes problem is considered in a rotating coordinate
system. Then, additional terms arise in the equation of motion and we write the whole
system in the form (see J. P. Vanyo [353])
d
t
u div (u u) Div o = 2$ $ ($ r). (6.3.44)
o = p I vVu. (6.3.45)
div u = 0. (6.3.46)
u = u
0
on I. (6.3.47)
In (6.3.44), the term 2$ is due to the Coriolis force and the term $ ($ r)
is related to the centrifugal force (the latter term is usually appended to the source
function and disappears from the equation). The vector $ is oriented along the axis
.
3
, and its value depends on the rotation velocity. Mathematical properties of such
models were studied by a number of authors (e.g., see A. Babin, A. Mahalov, and
B. Nicolaenko [18, 19] and the literature cited in those papers).
A linearized version of (6.3.44)(6.3.47) can be viewed as a certain generalization
of the Stokes problem. It is dened by the relations
Div o ju = $ u. (6.3.48)
o = p I vVu. (6.3.49)
div u = 0. (6.3.50)
u = u
0
on I. (6.3.51)
This system of equations arises if the problem (6.3.44)(6.3.47) is solved by semi-
discrete approximations (then j > 0 comes from an approximation of the term d
t
u).
A generalized solution of the problem (6.3.48)(6.3.51) is dened by the integral
relation
_
D
(vVu : Vn ju n ($ u) n) J. =
_
D
nJ.. (6.3.52)
which holds for any n

S
1
(C). As in the models considered before, this integral
relation generates an estimate of the difference between u and any u
0

S
1
(C).
We have
_
D
(vV(u ) : Vn j(u ) n ($ (u )) n) J.
=
_
D
( n vV : Vn j n ($ ) n) J.
=
_
D
( j ($ )) nJ.
_
D
(q I vV) : VnJ..
where q

1
2
(C). Set n = u and note that
($ n) n = 0. (6.3.53)
Let t H(C. Div) and
r(. t) := j $ Div t.
d(. t. q) := t q I vV.
By arguments used at the beginning of this section, we obtain the estimate
[u [
2
_ M
$
(. t. q)
:=
__
_
_
_
d(. t. q)
_
v
_
_
_
_
C
D
[(1 )r(. t)[
_
2
_
_
_
_
r(. t)
_
j
_
_
_
_
2
. (6.3.54)
which is valid for any 0. 1|. Estimates for nonsolenoidal velocity elds and for
approximations of the pressure function can also be derived quite analogously (see
[159] for details).
Section 6.4 The Oseen problem 151
6.4 The Oseen problem
The Oseen problem is often regarded as a linearization of the NavierStokes problem
at a neighborhood of a constant velocity eld a

S
1
(C), which leads to the system
u
t
vû div (a u) = Vp in C. (6.4.1)
u(.. 0) =u(.). (6.4.2)
div u = 0. (6.4.3)
u = u
0
on I. (6.4.4)
In the stationary case, we formulate the problem as follows: Find u, p, and o such that
Div o Div (a u) = Vp in C. (6.4.5)
o = vVu. (6.4.6)
u = u
0
on I. (6.4.7)
div u = 0. (6.4.8)
where it is assumed that div u
0
= 0. A generalized solution of the above system is
dened as a function u u
0

S
1
(C) that satises the integral identity
_
D
(vVu : Vn (a u) : Vn) J. =
_
D
nJ.. n

S
1
(C). (6.4.9)
We assume that the problem data are such that u exists and is unique.
Estimates for the velocity eld. (6.4.9) implies computable bounds of errors for
solenoidal approximations of u. Let

S
1
(C). We rearrange (6.4.9) into the form
_
D
(vV(u ) : Vn (a (u )) : Vn) J.
=
_
D
( n vV : Vn (a ) : Vn) J.. n

S
1
(C). (6.4.10)
Note that
_
D
(a n) : VnJ. =
_
D
Div (a n) nJ. =
_
D
(a Vn) nJ.
=
_
D
a ((Vn)n) J. =
1
2
_
D
a V([n[
2
) J. = 0.
Take a symmetric tensor-valued function t H(C. Div) and rewrite (6.4.10) as fol-
lows:
_
D
vV(u ) : VnJ.
=
_
D
_
( Div t) n (t Iq a vV) : Vn
_
J.. (6.4.11)
where q

1
2
(C). By setting n = u , we note that (6.4.11) leads to the inequality
v[V(u )[ _ M
OS
(. t. q)
:= [t q I a vV[ C
T D
[ Div t[. (6.4.12)
Remark 6.13. If u is approximated by the function u
0
V
0
, then the correspond-
ing estimate for the deviation norm can be derived in exactly the same way as for the
Stokes problem. In this case, the majorant includes an additional term that penalizes
possible violations of the incompressibility condition. This estimate has the following
form:
v[V(u)[ _ [t q I a vV[ C
T D
[ Div t[ c
D
[div[. (6.4.13)
where c
D
depends on c
:
, C
T D
, [a[, and v.
Remark 6.14. We note that
M
OS
(. o. p) _ v[V(u )[ [a[[u [ _ (v [a[C
T D
)[V(u )[.
Therefore, the minimization of M
OS
(. t. q) with respect to t and q gives an upper
bound, which is equivalent to the energy error norm.
Estimates for the pressure eld. Assume that q

1
2
(C) is an approximation of
the pressure eld p. We take n V
0
as in Section 6.2.2, i.e., div n = p q and
[V n[ _ c
:
[p q[. Since
_
D
(p q)div nJ. =
_
D
(vVu : V n n (a u) : V n q div n) J.
_
_
D
(vV(u ) : V n (a (u )) : V n) J.
_
D
(vV : V n n (a ) : V n q div n) J..
we nd that
[p q[
2
_ (v C
T D
[a[)[V(u )[[V n[
[Div t [C
T D
[V n[ [t q I a vV[[V n[.
This relation implies the estimate
[p q[ _ c
:
_
(v C
T D
[a[)[V(u )[ C
T D
[Div t [
[t q I a vV[
_
. (6.4.14)
where [V(u )[ is estimated by (6.4.12).
Section 6.5 Stationary NavierStokes problem for J = 2 153
Remark 6.15. For the product space
W := (u
0

S
1
(C)) H(C. Div)

1
2
(C).
we introduce two equivalent norms
[(. t. q)[
W
:= v[V[ [t[
Div
_
J[q[.
|(. t. q)|
W
:= v[V[ |t|
Div
g
_
J[q[.
By the same arguments as in Section 6.2.4 one can showthat the majorant M
OS
(. t. q)
is equivalent to the error in the above-dened combined norms.
Generalized form of the Oseen problem. As in the case of the Stokes problem, a
generalized form of the Oseen problem arises if semidiscrete approximations of the
NavierStokes problem are used. For example, the scheme
_
u
k
-u
k-1
k
vû
k
div (u
k-1
u
k
) = Vp
k
in C.
div u
k
= 0.
(6.4.15)
leads to a stationary Oseen problem (for u
k
): Find u u
0

S
1
(C) that satises the
integral identity
_
D
(vVu : Vnju n(a u) : Vn) J. =
_
D
nJ.. n

S
1
(C). (6.4.16)
Estimates for such problems can be obtained without any serious difculties by re-
peating arguments we used in Section 6.3 for the generalized Stokes problems. In
this case, the estimates have the same form as, e.g., (6.3.15) and (6.3.31) but with
d(. t. q) := t q I a vV.
6.5 Stationary NavierStokes problem for d D 2
The NavierStokes equation
u
t
vû Div (u u) = Vp (6.5.1)
is the most known model in the theory of viscous incompressible uids. From the
mathematical point of view, the NavierStokes equation has yet to be completely un-
derstood.
1
It is known that for sufciently regular initial data it has a rather weak
LerayHopf solution (e.g, see G. Galdi [146], where the reader will nd a consequent
1
The problem to prove or counter the existence of a unique smooth solution in (0. T | R
3
is formu-
lated as one of the Millennium Prize Problems stated by the Clay Mathematical Institute.
exposition of the mathematical theory related to NavierStokes equation). So far, all
existence and uniqueness results of a stronger type have been conditional. In view of
these difculties, it is not surprising that at present we have no reliable a posteriori
estimates for this class of problems. Below, we consider one special case generated
by the stationary NavierStokes equation in a bounded Lipschitz domain C R
2
for
which existence of a unique solution is established (see O. Ladyzhenskaya [210]). We
consider the problem
Div (vVu) Div (u u) = Vp in C. (6.5.2)
div u = 0. (6.5.3)
u = u
0
on I. (6.5.4)
A generalized solution u u
0

S
1
(C) is dened by the integral identity
_
D
(vVu : Vn (u u) : Vn) J. =
_
D
nJ. n

S
1
(C). (6.5.5)
For any solenoidal u and n

S
1
(C), we have
_
D
(u n) : VnJ. =
1
2
_
D
u V([n[
2
) J. = 0. (6.5.6)
In particular,
_
D
(u u) : VuJ. = 0. and the relation
_
D
(vVu : Vu (u u) : Vu) J. =
_
D
uJ.
furnishes the energy estimate v[Vu[ _ C
T D
[ [.
Now, we use (6.5.5), for the derivation of an a posteriori estimate. Assume that

S
1
(C) is an approximation of u. Then,
_
D
(vV(u ) : Vn (u (u )) : Vn) J.
=
_
D
( n vV : Vn (u ) : Vn) J.. Vn

S
1
(C). (6.5.7)
Set n = u . In view of (6.5.6), the second term in the left-hand side of (6.5.7)
vanishes. Also, we note that
(u ) : V(u ) = ( ) : V(u ) ((u ) ) : V(u )
and
_
D
((u ) ) : V(u ) J. =
_
D
Div ((u ) ) (u ) J.
=
_
D
((u ) V) (u ) J.
=
_
D
V : ((u ) (u )) J.. (6.5.8)
Section 6.5 Stationary NavierStokes problem for J = 2 155
From (6.5.8), it follows that
v[V(u )[
2
_
D
V :
_
(u ) (u )
_
J.
_
_
D
( (u ) vV : V(u ) ( ) : V(u )) J.. (6.5.9)
For a symmetric t H(C. Div), we represent the right-hand side as follows:
_
D
( Div t Div ( )) (u ) J.
_
D
(t q I vV) : V(u ) J..
Note that
_
D
V : ((u ) (u )) J. _
_
D
[V[[(u )[
2
J. _ [V[[(u )[
2
4,D
_ j[V[[V(u )[
2
. (6.5.10)
where j is a constant in the inequality [(u )[
4,D
_ j[V(u )[ (which holds in
view of embedding theorems).
A computable upper bound of [V(u )[ follows from (6.5.10) if we assume that
v = v j[V[ > 0. (6.5.11)
It should be remarked that this assumption is very demanding (for approximate so-
lutions of problems with high velocities it does not hold). By (6.5.9)(6.5.11), we
conclude that
v[V(u )[ _ [t q I vV[ C
T D
[ Div t Div ( )[ (6.5.12)
has the same structure as the estimates derived for the Stokes and Oseen problems.
However, this is a conditional estimate valid only for sufciently slow ows.
Finally, we note that in view of the relation
_
D
V(u ) :
_
(u ) (u )
_
J. = 0.
(6.5.9) can be represented in the form
v[V(u )[
2
_
D
Vu :
_
(u ) (u )
_
J.
=
_
D
( Div t Div ( )) (u ) J.
_
D
(t q I vV) : V(u ) J.. (6.5.13)
If
v = v j[Vu[ > 0. (6.5.14)
then we nd that
v[V(u )[ _ [t q I vV[ C
T D
[ Div t Div ( )[. (6.5.15)
which implies the uniqueness of u. Indeed, assume that u and is a pair of solutions
satisfying (6.5.2)(6.5.4) that differs from u and p. In this case, we set = u, q = p,
and t = p I vV u. By (6.5.15), we conclude that
v[V(u u)[ = 0
and, consequently, u = u. In view of the energy estimate, (6.5.14) is satised if
j[Vu[ _
jC
T D
v
[ [ _ v. (6.5.16)
Hence, if is sufciently small (namely, if [ [ _ v
2
j
-1
C
-1
T D
), then a solution u
to the stationary NavierStokes problem is unique (this fact is well known, see, e.g.,
[210]).
Numerical methods for viscous ow problems are represented in many publications.
The reader will nd a systematic discussion of the topic in, e.g., M. Feistauer [131],
V. Girault and P. A. Raviart [152], R. Glowinski and O. Pironneau [157], R. Rannacher
[271], R. Rannacher and S. Turek [274], R. Temam [348], and S. Turek [352].
In the last decades, adaptive methods and a posteriori error indicators for approxi-
mate solutions of viscous ow problems attracted serious attention of many researches
who used methods different from those considered in this chapter. We cannot give
here a consequent overview of these results and conne ourselves to a short discus-
sion of some publications that set out the main approaches. The reader will nd more
literature references in the papers cited.
Residual type a posteriori methods for viscous ow problems are considered in the
book by R. Verf urth [356]. A posteriori analysis of approximations computed with
the help of the backward Euler scheme is given in C. Bernardi and R. Verf urth [55].
A posteriori error estimators for nite-element approximations of the Stokes problem
were obtained in R. E. Bank and B. D. Welfert [43] and R. Verf urth [354]. Error in-
dicators for the NavierStokes equations in stream function and vorticity statement
are discussed in M. Amara, M. Ben Younes, and C. Bernardi [13]. In D. Kay and
D. Silvester [193], various a posteriori estimators are investigated for stabilized mixed
approximations of the Stokes problem. Adaptive methods and a posteriori estimates
in computational uid dynamics are exposed, e.g., in M. Ainsworth and J. T. Oden
[8], J. G. Heywood and R. Rannacher [172], C. Johnson, R. Rannacher, and M. Bo-
man [189], J. T. Oden, W. Wu and M. Ainsworth [255], J. T. Oden, L. Demkowicz,
T. Strouboulis and P. Devloo [252], R. Rannacher [271], T. Strouboulis and J. T. Oden
[345]. A posteriori error estimators for some quasi-Newtonian uids were considered
in C. Padra [258] and in A. Berm udez, R. Dur an and R. Rodrguez [53] for combined
uid-solid systems. J. Wang and X. Ye [368] investigated error indicators based on
the superconvergence of nite-element approximations for Stokes and NavierStokes
equations.
For the stationary Stokes problem with Dirichlet boundary conditions, a posteriori
estimates of the functional type were originally derived by the author with the help of
a variational method in [288]. In [293], it was shown that for the Stokes problem the
same estimates follow from a pertinent integral identity. The material exposed in Sec-
tion 6.2 follow the lines of this paper. Later the variational technique was applied to
some classes of generalized Newtonian uids (see M. Bildhauer, M. Fuchs, and S. Re-
pin [57], M. Fuchs and S. Repin [139], and in the authors papers [289, 292, 293]).
Estimates of the same type were derived (by a nonvariational method) for generaliza-
tions of the Stokes problem in S. Repin and R. Stenberg [316]. Estimates for ow mod-
els with polymerization discussed in Section 6.3.5 were derived in [315]. Estimates
for the Oseen problem considered in Section 6.4 were obtained in [293]. Within the
framework of a posteriori analysis, problems with rotation were considered in E. Gor-
shkova, A. Mahalov, P. Neittaanm aki, and S. Repin [159] and in the PhD thesis of
E. Gorshkova [158], which also contains results of numerical tests and a comparative
study of different error indication methods for approximate solution of viscous ow
problems.
7 Generalizations
7.1 Linear elliptic problem
First, we consider the following general form of a linear elliptic problem: Find
u u
0
V
0
such that
(Au. n) (. n) = 0. Vn V
0
. (7.1.1)
Here V
0
is a closed subspace of a reexive Banach space V , is a linear bounded
operator acting from V to a Hilbert space U with scalar product (. ), V
+
0
. and
A L(U. U) is a self-adjoint operator. In this section, [ [ stands for the norm in U.
We assume that the operators and Asatisfy the relations
c
2
1
[,[
2
_ (A,. ,) _ c
2
2
[,[
2
. V, U. (7.1.2)
and
[n[ _ c
3
[n[
V
. Vn V
0
. (7.1.3)
with positive constants c
1
. c
2
, and c
3
. In most applied problems, the functional is
representable in the form (. n) = (. n)
V
(g. n), where V, g U, and V
is a Hilbert space such that V V V
+
0
. Such a functional is well dened and nite
on the elements of V . Henceforth, we assume that satises this condition.
For our analysis, it is convenient to introduce two additional spaces. The quantity
(A,. ,)
1{2
determines a new norm [[,[[, which is equivalent to the original norm
[,[ = (,. ,)
1{2
. Another equivalent norm is
[[,[[
+
= (A
-1
,. ,)
1{2
.
where A
-1
is the operator inverse to A. The spaces Y and Y
+
contain elements of U
equipped with the norms [[ [[ and [[ [[
+
, respectively.
Here and later on,
+
: U V
+
0
denotes the operator conjugate to in the sense
that
(,. n) = (
+
,. n). n V
0
. (7.1.4)
and (n
+
. n) is the value of the functional n
+
V
+
0
at n V
0
. Below we derive
functional a posteriori estimates for the problem (7.1.1) with the help of two different
methods.
Section 7.1 Linear elliptic problem 159
7.1.1 The variational method
In the variational method, we rely on the variational statement of the problem: Find
u u
0
V
0
such that
J(u) = inf
uu
0
V
0
J(n). where J(n) :=
1
2
[[n[[
2
(. n).
We call this problem primal or Problem P.
It is easy to show that it is coercive on u
0
V
0
. For any n V
0
, we have
J(n) _
c
2
1
2
[(n u
0
)[
2
(. u
0
) [[[ [[[ [[[ [[[ [[(n)[[
_
c
2
1
2
[n[
2
c
2
1
2
[u
0
[
2
(. u
0
) c
2
[[[ [[[ [[[ [[[ [n[ c
2
1
(n. u
0
)
1
.
where [[[ [[[ [[[ [[[ < ois a positive quantity dened by the relation
[[[ [[[ [[[ [[[ := sup
uV
0
(. n)
[[n[[
.
In view of (7.1.3), this norm is equivalent to the standard norm sup
uV
0
(I,u)
|u|
1
. It is
clear that J(n) o as [n[ o, which together with (7.1.3) proves the
coercivity of J on u
0
V
0
. Since J is also continuous (on V ) and strictly convex, we
conclude that the minimizer u exists and is unique.
The variational method of deriving an a posteriori error estimate attracts the so-
called dual problem, which we introduce below.
Let (u
0
V
0
) Y
+
R be the Lagrangian
1(. ,) = (,. )
1
2
[[,[[
2
+
(. ).
sup
,U
1(. ,) = J().
Dene the functional
1
+
(,) = inf
u
0
V
0
1(. ,) =
_
(,. u
0
)
1
2
[[,[[
2
+
(. u
0
). , Q
I
.
o. , Q
I
.
where
Q
I
:=
, U
(,. n) (. n) = 0. Vn V
0
_
.
The functional 1
+
generates the following (dual) Problem P
+
: Find Q
I
such that
1
+
() = sup
,Q
I
1
+
(,) := sup P
+
.
160 Chapter 7 Generalizations
The functionals J and (1
+
) are convex and coercive on u
0
V
0
and Y
+
, respectively.
The sets u
0
V
0
and Q
I
are closed afne manifolds. Therefore (e.g., see [121]), there
exist u u
0
V
0
and Q
I
such that
J(u) = inf P. 1
+
() = sup P
+
. (7.1.5)
The minimizer u satises (7.1.1). For any c R the maximizer satises the relation
( cj. u
0
)
1
2
[[ cj[[
2
+
_ (. u
0
)
1
2
[[[[
2
+
. Vj Q
+
0
.
which implies the necessary condition
(u
0
A
-1
. j) = 0. Vj Q
+
0
. (7.1.6)
where
Q
+
0
:=
, U
(,. n) = 0. Vn V
0
_
.
Note that Au Q
I
. so that we obtain
1
+
(Au) = (Au. u
0
)
1
2
[[Au[[
2
+
(. u
0
) _ sup P
+
. (7.1.7)
Since
[[Au[[
2
+
= (A
-1
Au. Au) = [[u[[
2
and
(Au. (u u
0
)) (. (u u
0
)) = 0.
we nd that
sup P
+
_ 1
+
(Au) = J(u) = inf P. (7.1.8)
On the other hand, we know that inf P _ sup P
+
. Therefore,
sup P
+
= inf P. (7.1.9)
(. u)
1
2
[[[[
2
+
(. u) =
1
2
[[u[[
2
(. u).
which is equivalent to the relation
1
2
[[u[[
2
1
2
[[[[
2
+
(. u) =
1
2
[[u A
-1
[[
2
= 0 (7.1.10)
or simply to
= Au. (7.1.11)
This is the duality relation for the pair (u. ).
Let u
0
V
0
and , U be some approximations of u and , respectively. First,
we establish the following basic result.
0
V
0
and q Q
I
, we have
[[( u)[[
2
[[q [[
2
+
= 2 (J() 1
+
(q)). (7.1.12)
[[( u)[[
2
[[q [[
2
+
= 2 D(. q). (7.1.13)
where
D(. q) :=
1
2
[[[[
2
1
2
[[q[[
2
+
(q. ) =
1
2
[[A q[[
2
+
.
Proof. In view of (7.1.1) and (7.1.6), we have
1
2
[[( u)[[
2
= J() J(u) (Au. ( u)) (. u)
= J() J(u).
and
1
2
[[q [[
2
+
= 1
+
() 1
+
(q) (u
0
A
-1
. q)
= 1
+
() 1
+
(q).
Since J(u) = 1
+
(), these relations imply (7.1.12). For q Q
I
we have
J() 1
+
(q) =
1
2
[[[[
2
(. u
0
) (q. u
0
)
1
2
[[q[[
2
+
=
1
2
[[[[
2
1
2
[[q[[
2
+
(q. )
= D(. q).
so that (7.1.13) follows from (7.1.12).
Remark 7.2. We note that (7.1.12) and (7.1.13) can be viewed as generalizations of
the Mikhlin and PragerSynge estimates (2.3.1) and (2.2.3), respectively.
[[( u)[[ _ [[A q[[
+
. Vq Q
I
.
Let , U, then
[[( u)[[ _ [[A ,[[
+
inf
qQ
I
[[q ,[[
+
. (7.1.14)
Thus, we need a computable estimate of the distance between , and the set Q
I
. For
this purpose, we introduce the space
Q
+
:= {, U [
+
, V ]
endowed with the norm
[,[
Q
:= [[,[[
+
[
+
,[
V
.
If , Q
+
, then
(,. n) = (
+
,. n)
V
. Vn V
0
.
We estimate the distance to the set Q
I
by the following lemma.
Lemma 7.3. For any , U,
inf
qQ
I
[[, q[[
+
_ [[[ [[[
+
, [[[ [[[ := sup
uV
0
(
+
,. n)
[[n[[
. (7.1.15)
If (. n) = (. n)
V
, then for any , Q
+
,
inf
qQ
I
[[, q[[
+
_ C[
+
, [
V
. (7.1.16)
where C =
c
c
1
and c is the constant in the inequality
[n[
V
_ c [n[. Vn V
0
. (7.1.17)
Proof. Consider the problem
(An
I
. n) =

+
,. n
. (7.1.18)
Represent the right-hand side in the form (. n) (,. n). Then, (7.1.18) implies
the relation
[[n
I
[[ _ sup
uV
0
[(,. n) (. n) [
[[n[[
= [[[ [[[
+
, [[[ [[[ .
Also, (7.1.18) has the form
((An
I
,). n) (. n) = 0. Vn V
0
.
which means that q
I
:= An
I
, Q
I
. Therefore,
inf
qQ
I
[[, q[[
+
_ [[, q
I
[[
+
= [[An
I
[[
+
= [[n
I
[[ _ [[[ [[[
+
, [[[ [[[ .
If (. n) = (. n)
V
and , Q
+
, then
[[[ [[[
+
, [[[ [[[ = sup
uV
0
(
+
, . n)
[[n[[
= sup
uV
0
(
+
, . n)
V
[[n[[
_ sup
uV
0
[
+
, [
V
[n[
V
[[n[[
_ [
+
, [
V
c
-1
1
sup
uV
0
[n[
V
[n[
.
Since V
0
is continuously embedded into V and (7.1.3) holds, we conclude that (7.1.17)
also holds with a constant c > 0 independent of n. Then
[[[ [[[
+
,[[[ [[[ _
c
c
1
[
+
, [
V
.
Now, (7.1.14) implies two estimates (see also [277, 282, 286]):
[[( u)[[ _ [[A ,[[
+
[[[ [[[
+
, [[[ [[[ . (7.1.19)
[[( u)[[ _ [[A ,[[
+
C[
+
,[
V
. (7.1.20)
We recall that in (7.1.19) , U and in (7.1.20) , Q
+
.
Henceforth, we denote the right-hand side of (7.1.20) by M
(. ,).
Remark 7.4. It is easy to see that
[[(u )[[ = inf
,Q
_
[[A,[[
+
C[
+
,[
_
.
To prove this fact, it sufces to set , = . Hence, the upper bound given by M
has
no gap.
From (7.1.20), it also follows that for any > 0,
[[( u)[[
2
_ (1 )[[ ,[[
2
+
_
1
1
_
C
2
[
+
,[
2
V
=: M
2
(. ,. ). (7.1.21)
This estimate provides a majorant having the form of a quadratic functional.
Error reduction property. Theorem 7.1 has a simple consequence related to the
so-called error reduction property. In practice, it is often of interest to predict how
signicantly an approximation error would decrease on account of a certain improve-
ment of a mesh. Assume that a coarse mesh T
h
is replaced by a rened one T
h
ref
. For
problems associated with quadratic functionals (similar to J), the value of the error
reduction is easy to compute. Indeed, for any u
0
V
0
,
1
2
[[(u )[[
2
= J() 1
+
() = J() J(u).
Therefore,
1
2
e
2
h
:=
1
2
[[(u u
h
)[[
2
= J(u
h
) J(u).
1
2
e
2
h
ref
:=
1
2
[[(u u
h
ref
)[[
2
= J(u
h
ref
) J(u).
From the above, a simple error reduction relation follows:
e
2
h
ref
= e
2
h
2
_
J(u
h
) J(u
h
ref
)
_
. (7.1.22)
We note that the reduction of the approximation error is equal to the difference of the
values of corresponding functionals. If an upper bound of the approximation error
related to the coarser mesh T
h
is known, i.e.,
[[(u u
h
)[[
2
_ M.
then for the rened mesh we have the estimate
[[(u u
h
ref
)[[
2
_ M 2
_
J(u
h
) J(u
h
ref
)
_
. (7.1.23)
If M is a sharp upper bound of the error on T
h
, then (7.1.23) gives a realistic value of
the error on T
h
ref
. Since the right-hand side of (7.1.23) is easily computable, getting
the upper bound can be performed with minimal expenditures.
From (7.1.23), it also follows that a mesh-adaptation strategy of rening the mesh
should minimize local contributions in the energy functional.
7.1.2 The method of integral identities
Upper bound. Let u
0
V
0
be an approximation of u. Then, the a posteriori
estimate can be derived directly from (7.1.1). Indeed,
(A(u ). n) =

+
,. n
(, A. n). (7.1.24)
Since
(A ,. n) _ [[A ,[[
+
[[n[[
and

+
,. n
_ [[[ [[[
+
, [[[ [[[ [[n[[.
we nd that
(A(u ). n) _
_
[[A ,[[
+
[[[ [[[
+
, [[[ [[[
_
[[n[[.
Setting n = u , we arrive at (7.1.19).
If (. n) = (. n)
V
and , Q
+
, then

+
,. u
= (
+
,. u )
V
_ [
+
,[
V
[u [
V
_ c[
+
,[
V
[(u )[ _ C[
+
,[
V
[[(u )[[
and (7.1.24) implies (7.1.20).
If (. n) = (. n)
V
(g. n) and , Q
+
, then

+
,. n
= (. n)
+
,. n
= (
+
,. n)
V
(g. n).
From (7.1.24), we obtain
(A(u ). n) = (
+
,. n)
V
(, A g. n). (7.1.25)
For n = u , (7.1.25) implies the estimate
[[(u )[[ _ [[g A ,[[
+
C[
+
,[
V
. (7.1.26)
If g is efciently approximated by g Q
+
, then another estimate can be used:
[[(u )[[ _ [[A ,[[
+
C[
+
(, g)[
V
[g g[. (7.1.27)
Lower bound. A lower bound of [[(u )[[ can be derived in the following way.
Note that
sup
uV
0
_
(A(u ). n)
1
2
(An. n)
_
_ sup
rY
_
(A(u ). t)
1
2
(At. t)
_
=
1
2
[[(u )[[
2
.
However,
sup
uV
0
_
(A(u ). n)
1
2
(An. n)
_
_ (A(u ). (u ))
1
2
(A(u ). (u )) =
1
2
[[(u )[[
2
.
1
2
[[(u )[[
2
= sup
uV
0
_
(A(u ). n)
1
2
(An. n)
_
_ sup
uV
0
_
1
2
(An. n) (A. n) (. n)
_
. (7.1.28)
It is easy to see that the lower bound given by the left-hand side of the above estimate
is sharp (set n = u ).
Modications of the estimates. Let 0 V
0
be such that
+
A0 V. Then, the
product (
+
A0. n) is represented by the scalar product of V and we have
(
+
A0. u)
V
= (A0. u) (A0. ) = (. 0) (A0. ).
Then, for (. n) = (. n)
V
and , Q
+
we express (7.1.24) in the form
[[(u )[[
2
= (
+
,
+
A0. u)
V
(, A. (u )) F
(0). (7.1.29)
where F
(0) := (. 0) (A0. )
V
. Hence, we obtain
[[(u )[[
2
_
2;
2; 1
F
(0)
;
2
2; 1
_
[[A ,[[
+
C[
+
(, A0)[
V
_
2
. (7.1.30)
If 0 solves the problem
(A0. n) = (
+
,. n)
V
. Vn V
0
.
then (7.1.29) implies that
[[(u )[[
2
_
2;
2; 1
F
(0)
;
2
2; 1
[[A ,[[
2
+
. (7.1.31)
Another modication is based on a generalization of the idea that was used in
Section 3.5.3. Consider the functional spaces U, V, V , and Y contain functions
dened in a domain C, which is decomposed into a collection of subdomains C
i
,
i = 1. 2. . . . . N. By V(C
i
) we denote the restriction of V associated with C
i
and
assume that the respective scalar product is additive with respect to the above decom-
position. Let
1
i
:= {n V(C
i
) [ n = 0 in C
i
]
and fng
D
i
1
i
denote the orthogonal projection of n, i.e.,
[n fng
D
i
[
V(D
i
)
= inf
u
0
1
i
[n n
0
[
V(D
i
)
.
In this case,
(n fng
D
i
. n
0
)
V(D
i
)
= 0. V n
0
1
i
. (7.1.32)
Assume that for any C
i
,
[n fng
D
i
[
V(D
i
)
_ C
D
i
[n[
U(D
i
)
. (7.1.33)
where C
D
i
is a positive constant depending only on C
i
(this estimate is a generaliza-
tion of the Poincar e inequality). Since fng
D
i
1
i
, we have
(r(,). n)
V(D
i
)
= (r(,) fr(,)g
D
i
. n)
V(D
i
)
(fr(,)g
D
i
. n)
V(D
i
)
= (r(,) fr(,)g
D
i
. n fng
D
i
)
V(D
i
)
(fr(,)g
D
i
. n)
V(D
i
)
. Vn V(C
i
).
where r(,) :=
+
,. Then
(r(,). n)
V
=
1
i=1
(r(,). n)
V(D
i
)
=
1
i=1
_
(r(,) fr(,)g
D
i
. n fng
D
i
)
V(D
i
)
(fr(,)g
D
i
. n)
V(D
i
)
_
_
_
1
i=1
C
2
D
i
[r(,) fr(,)g[
2
D
i
_
1{2
_
1
i=1
[n[
2
U(D
i
)
_
1{2
_
1
i=1
[fr(,)g
D
i
[
2
V(D
i
)
_
1{2
_
1
i=1
[n[
2
V(D
i
)
_
1{2
.
[(r(,). n)
V
[ _
_
1
i=1
C
2
D
i
[r(,) fr(,)g[
2
D
i
_
1{2
[n[
c
_
1
i=1
[fr(,)g
D
i
[
2
V(D
i
)
_
1{2
[n[. (7.1.34)
By (7.1.24), we have
(A(u ). n) _ [(r(,). n)
V
[ [[, A[[
+
[[n[[. (7.1.35)
Set here n = u and use (7.1.34). Then, we obtain the estimate
[[(u )[[ _ [[, A[[
+
1
c
1
_
1
i=1
C
2
D
i
[r(,) fr(,)g
D
i
[
2
D
i
_
1{2
C
_
1
i=1
[fr(,)g
D
i
[
2
V(D
i
)
_
1{2
. (7.1.36)
If the residuals are post-processed in such a way that fr(,)g
D
i
= 0 on any C
i
, then
(7.1.36) takes a simplied form:
[[(u )[[ _ [[, A[[
+
1
c
1
_
1
i=1
C
2
D
i
[r(,) fr(,)g
D
i
[
2
D
i
_
1{2
. (7.1.37)
which can be viewed as a generalization of (3.5.20).
Remark 7.5. If n is dened by the operator of small strains c(n), then fng
D
i
de-
notes the orthogonal projection to the space R(C
i
) of rigid deections.
7.1.3 Error estimates for the dual variable
Consider , as an approximation of (which is the exact solution of the dual problem).
To obtain an upper bound of [[ ,[[
+
, we use the relation
[[ ,[[
+
_ [[, A[[
+
[[( u)[[
_ 2[[, A[[
+
[[[ [[[
+
, [[[ [[[ . (7.1.38)
If = V and , Q
+
, then
[[ ,[[
+
_ [[, A[[
+
[[( u)[[
_ 2[[, A[[
+
C[
+
, [
V
. (7.1.39)
The norm [,[
Q
is another measure that characterizes the error ,. Since
[
+
( ,)[ = [
+
, [.
we note that
[ ,[
Q
_ 2[[, A[[
+
(C 1)[
+
, [
V
. (7.1.40)
7.1.4 Two-sided estimates for combined norms
If a pair (u. ) is viewed as a solution, then it is natural to measure the corresponding
error in terms of combined (primal-dual) norms of the product space W := V Q
+
,
for which we introduce the norm
[(. ,)[
W
:= [[[[ [,[
Q
.
We show that the majorant M
(. ,) is equivalent to the error in the combined norm

[(. ,)[
W
. We have
[(u . ,)[
W
:= [[(u )[[ [[ ,[[
Q
(7.1.41)
_ 3[[A ,[[ (2C 1)[
+
, [ _ c
(. ,).
Q
(
V
*
(
, )
p
p
y
y v
v
u
u
, )
Figure 7.1.1 Error in terms of the combined primal-dual norm.
where c
= max
3. 2
1
C
_
. On the other hand,
M
(. ,) _ [[( u)[[ [[ ,[[

+
C[
+
, [ (7.1.42)
_ max{1. C] [(u . ,)[
W
.
Thus, we note that the following two-sided estimate holds:
c
(. ,) _ [(u . ,)[
W
_ c
(. ,). (7.1.43)
where c
=
1
max1,C
.
Hence, the efciency index of the majorant (with respect to the combined error
norm) is estimated from above as follows:
1
eff
_
c
= max{1. C] max
3. 2
1
C
_
. (7.1.44)
Therefore, M
is an efcient and reliable measure of the error in the combined norm

[(u . ,)[
W
.
Remark 7.6. Since
M
(. ) = [(u )[ and M
(u. ,) = [[ ,[[
+
[
+
( ,)[.
we note that
[(u . ,)[
W
:= M
(. ) M
(u. ,). (7.1.45)

Remark 7.7. If the error is measured in a different (but equivalent) norm
[(. ,)[
(1)
W
:= [[[[ [[,[[
+
C[
+
,[.
then
M
(. ,) _ [(u . ,)[
(1)
W
_ 3 M
(. ,). (7.1.46)
Thus, the majorant is equivalent to such a norm and the corresponding efciency index
does not depend on c.
We can give another interpretation of the results discussed. Introduce the functionals
M
(. ,) = 3[[A ,[[
+
(1 2c) [
+
, [
and
M
(. ,) := [[A ,[[
+
[
+
, [.
which consist of the same terms as does the majorant but with different weights. We
have proved that
M
(. ,) _ [(u . ,)[
W
_ M
(. ,). (7.1.47)
In other words, if the terms [[A ,[[
+
and [
+
, [ are supplied with proper
weights, then their sum furnishes two-sided bounds of the error in the combined norm.
Comments. The functionals M
(. ,), M
(. ,) and M
(. ,) provide natural
error estimation tools for approximations computed with the help of the mixed nite
element method. Mixed approximations of boundary value problems are given by a
pair of functions (u
h
.
h
), which can be substituted in the corresponding majorant di-
rectly or after a certain post-processing procedure. A priori and a posteriori estimates
for mixed approximations were investigated by many authors. An elaborated theory
of mixed nite element methods can be found in the books by F. Brezzi and M. Fortin
[79], D. Braess [67], J. E. Roberts and J.-M. Thomas [325]. Various a posteriori error
estimators for mixed nite element methods were studied by many authors (e.g., see
A. Alonso [12], I. Babu` ska and G. N. Gatica [26], G. F. Carey and A. I. Pehlivanov
[85], C. Carstensen [86, 88], C. Carstensen and G. Dolzmann [91], C. Carstensen and
R. H. W. Hoppe [94], B. I. Wohlmuth and R. H. W. Hoppe [372]). In G. Gatica [149],
the reader will nd a proof of the efciency of residual-based a posteriori error es-
timators for mixed approximations. A posteriori estimates for approximations based
on RaviartThomas elements are represented by D. Braess and R. Verf urth [70], and
estimates based on superconvergence phenomenon are analyzed by J. H. Brandts [74].
A posteriori error estimators for mixed approximations of problems in linear elasticity
theory were investigated by M. Lonsing and R. Verf urth [223] and for mixed approxi-
mations of the Stokes problem by X. Cheng, W. Han, and H. Huang [104]. In B. Engel-
mann, R. H. W. Hoppe, B. Wohlmuth, Yu. Kuznetsov, Yu. Iliash, and Yu. Vasilevskii
[122], adaptive methods for hybrid nite element approximations are considered.
Section 7.2 Elliptic problems with lower terms 171
Functional a posteriori estimates for mixed approximations have been studied in
S. Repin, S. Sauter, and A. Smolianski [314, 313], where they were derived for linear
elliptic problems by variational techniques.
These estimates are applicable to mixed approximations of all types. In particular,
they applicable to the so-called dual mixed approximations, which use nonconform-
ing (with respect to the energy space V ) approximations of u and cell-equilibrated
approximations for (which integrally satisfy the relation
+
h
= 0 on each
cell/element). Approximations of such a type gained high popularity, because they pre-
serve relations coming from the physical (conservation) law and obtain a high exibil-
ity. The latter property was used to adapt them to approximations on highly distorted
meshes (e.g., see Yu. Kuznetsov and S. Repin [206, 205]). A posteriori estimates for
the dual-mixed approximations follow from the above-presented estimates, provided
that the function is post-processed (e.g., smoothed) in such a way that u
0
V
0
.
7.2 Elliptic problems with lower terms
Now we consider the problem: Find u u
0
V
0
such that
(Au. n) (T u. n)
1
(. n) = 0. n V
0
. (7.2.1)
where T : V V is a linear continuous operator satisfying the relation
j
2
1
[n[
2
V
_ (T n. n)
V
_ j
2
2
[n[
2
V
. (7.2.2)
The function u minimizes the functional
J(n) =
1
2
_
(An. n) (T n. n)
V
_
(. n)
on the set u
0
V
0
. Under the assumptions made, J(n) is coercive on u
0
V
0
and,
therefore, the minimizer u exists.
For a function u
0
V
0
, we have
(A(u ). n) (T(u ). n)
V
= (. n) (A. n) (T . n)
V
. (7.2.3)
Let , U and n = u . Then, (7.2.3) infers the relation
[u [|
2
= (
+
,. u ) (A ,. (u )) (T . u )
V
. (7.2.4)
Here,
[u [|
2
:= (A(u ). (u )) (T(u ). u )
V
is the energy error norm associated with the problem.
If is dened by an element V and , Q
+
, then (7.2.4) is rearranged as
follows:
[u |[
2
= (
+
, T . u)
V
(A ,. ( u)). (7.2.5)
We have
(
+
, T . u )
V
_ [r(. ,)[
V
[u [
V
_ (T
-1
r(. ,). r(. ,))
1{2
V
(T(u ). u )
1{2
V
.
(A ,. (u )) _ [[A ,[[
+
[[(u )[[.
where r(. ,) :=
+
, T With the help of these estimates and (7.2.5), we
obtain
[u |[
2
_ (T
-1
r(. ,). r(. ,))
V
[[A ,[[
2
+
. (7.2.6)
This estimate is sharp. Indeed, let , = Au. By (7.3.1), we nd that
(T u. n)
V
= (. n)
V
(,. n). n V
0
.
which means that
(T(u ). n)
V
= (
+
, T . n)
V
. n V
0
.
and, therefore, r(. ,) = T( u). Hence,
(T
-1
r(. ,). r(. ,))
V
= (T(u ). u )
V
.
Also,
[[A ,[[
+
= [[A( u)[[
+
= [[( u)[[
and the right-hand side of (7.2.6) is equal to the error norm [u |[
2
.
Remark 7.8. Since the problem has a variational statement, we can derive a posteriori
estimates using the variational method (see [276, 277, 282]), which gives the same
result.
Another estimate follows from (7.2.4) if we apply (7.1.17) and estimate the rst
term on the right-hand side of (7.2.5) as follows:
(
+
, T . u )
V
_ C[
+
, T [
V
[(u )[.
In this case, we arrive at the estimate
[[(u )[[ _ [A ,[
+
C[
+
, T [
V
. (7.2.7)
Remark 7.9. Also, we can derive a hybrid estimate if we introduce a function
(.) 0. 1| and split the rst term on the right-hand side of (7.2.5) (cf. Section
4.2). Then, we obtain the estimate
[(u )|[
2
_ ([A ,[
+
C[(1 )r(. ,)[
V
)
2
(T
-1
r(. ,). r(. ,))
V
. (7.2.8)
For = 0 and = 1, (7.2.8) implies the estimates (7.2.7) and (7.2.6), respectively.
Section 7.3 Problems with solutions dened in subspaces 173
7.3 Problems with solutions dened in subspaces
7.3.1 Abstract problem
The above-discussed method can be extended to problems the solutions of which be-
long to a certain subspace of the basic energy space.
Dene another pair of mutually conjugate linear operators T : V
0
H and
T
+
: H V
+
0
, where H is a Hilbert space endowed with scalar product (. )
1
.
The spaces and operators introduced are conveniently represented by the following
diagram:
H
T
V
0
U (Y. Y
+
)
H
T
+
V
+
0
U.
Consider the problem: Find H and u V
0,B
satisfying the relation
(Au. n) ( T
+
. n) = 0. n V
0
. (7.3.1)
where
V
0,B
:= { V
0
[ T = 0]
is a subspace dened by the kernel of T.
For q H, we have
(T
+
q. n) = (q. Tn)
1
= 0. Vn V
0,B
. (7.3.2)
Therefore, we can also dene u by the relation
(Au. n) (. n) = 0. n V
0,B
. (7.3.3)
The solvability of this problem requires special properties of the operator T, which
are analogous to the LBB condition (see Lemma 7.10). We assume that the necessary
conditions are satised and (7.3.1) has a solution u V
0,B
. Our goal is to present a
general scheme that delivers guaranteed estimates of the deviation from u.
7.3.2 Estimate for approximations lying in the subspace
Let V
0,B
. From (7.3.2), it follows that for any n V
0,B
and , U the identity
(A(u ). n) = (, A. n) (
+
,. n) (7.3.4)
holds. In view of (7.3.2), we have
sup
uV
0,!
[ (
+
,. n) [
[[n[[
= sup
uV
0,!
[ (
+
, T
+
q. n) [
[[n[[
_ sup
uV
0
[ (
+
, T
+
q. n) [
[[n[[
=: [[[ [[[
+
, T
+
q [[[ [[[ .
Thus (7.3.4) with n = u leads to the estimate
[[(u )[[ _ [[, A[[
+
[[[ [[[
+
, T
+
q [[[ [[[ . (7.3.5)
where q H. If is dened by V and
+
, and T
+
q also belong to V, then
(
+
, T
+
q. n) = (
+
, T
+
q. n)
V
_ [
+
, T
+
q[
V
[n[
V
.
and we nd that
sup
uV
0
[ (
+
, T
+
q. n) [
[[n[[
_ c sup
uV
0
[
+
, T
+
q[
V
[n[
[[n[[
.
In this case,
[[[ [[[
+
, T
+
q [[[ [[[ _ C[
+
, T
+
q[
V
.
[[(u )[[ _ [[, A[[
+
C[
+
, T
+
q[
V
. (7.3.6)
Another form of the estimate arises if there exists a linear continuous operator
T : H U such that
(T
+
q. n) = (Tq. n). Vn V
0
.
Then, (7.3.4) is rearranged as
(A(u ). n) = (, A Tq. n) (
+
,. n). (7.3.7)
and instead of (7.3.5) we obtain
[[(u )[[ _ [[, A Tq[[
+
[[[ [[[
+
, [[[ [[[ . (7.3.8)
If = and
+
, belongs to V, then (7.3.8) implies the estimate
[[(u )[[ _ [[, A Tq[[
+
C[
+
,[
V
. (7.3.9)
which is a counterpart of (7.3.5).
7.3.3 Estimate for approximations lying in the energy space
For functions that do not belong to the subspace V
0,B
, the deviations from u can be
estimated, provided that we can prove Lemma 7.10 (which is analogous to Lemma
6.2).
Section 7.3 Problems with solutions dened in subspaces 175
Lemma 7.10. For any g ImT, there exists
V
0
such that
T
= g and [
[
V
_ k[g[
1
. (7.3.10)
where k > 0 does not depend on g.
In general, this lemma can be proved with the help of closed range lemma (e.g.,
see K. Yosida [374]), which is valid for operators with closed range (we recall that the
operator T has a closed range if for any sequence {
k
] V such that T
k
converges
in H there exists V such that T
k
T).
Corollary 7.11. Set g = T, where is an element of V
0
. Then, we can nd an
element
V
0
satisfying the relation T(
) = 0 and such that

[
[
V
_ k[T[. (7.3.11)
Hence, the function n
0
= (
) belong to V
0,B
and satises the relation
[ n
0
[
V
_ k[T[.
Since is a bounded operator, we have the inequality
[( n
0
)[ _ k[T[. (7.3.12)
where the constant k does not depend on.
Let u be compared with a function V
0
. Take a function n
0
V
0,B
that satises
(7.3.12). Then,
[[(u )[[ _ [[(u n
0
)[[ [[( n
0
)[[
_ [[An
0
,[[
+
[[[ [[[
+
, T
+
q [[[ [[[ [[( n
0
)[[.
Insert into the rst term and use the triangle inequality. We have
[[(u )[[ _ [[A( n
0
)[[
+
[[A ,[[
+
[[[ [[[
+
, T
+
q [[[ [[[ [[( n
0
)[[
= 2[[( n
0
)[[ [[An
0
,[[
+
[[[ [[[
+
, T
+
q [[[ [[[
_ 2c
2
[( n
0
)[ [[An
0
,[[
+
[[[ [[[
+
, T
+
q [[[ [[[ . (7.3.13)
Hence,
[[(u )[[ _ 2c
2
k[T[ [[A ,[[
+
[[[ [[[
+
, T
+
q [[[ [[[ . (7.3.14)
If is dened by V and
+
, and T
+
q also belong to V, then (7.3.14) implies the
estimate
[[(u )[[ _ 2c
2
k[T[ [[A ,[[
+
C[
+
, T
+
q[
V
. (7.3.15)
Remark 7.12. We note that the problem considered can be expressed as a system
(
+
o T
+
. n) = 0. n V
0
. (7.3.16)
o = Au. (7.3.17)
T = 0. (7.3.18)
Each term on the right-hand side of (7.3.14) and (7.3.15) is a certain penalty for a
possible violation of one of the relations of this system.
Remark 7.13. For the Stokes problem, we have
= c(). T = div . and = vI.
It is easy to see that in this case, c
1
= c
2
= v,
[[A ,[[
+
=
1
_
v
[vc() ,[. and [[(u )[[ =
_
v[(u )[
and we nd that the estimate (7.3.14) coincides with (6.2.17).
7.4 Derivation of a posteriori estimates
from saddle point relations
In this section we again consider the problem (7.1.1) and show that a posteriori esti-
mates can be obtained by simple transformations of the saddle point relations.
First, we recall that the problem (7.1.1) can be stated as a minimax problem: Find
u u
0
V
0
and U such that
1(u. q) _ 1(u. ) _ 1(. ). V u
0
V
0
. q U. (7.4.1)
where
1(n. q) = (n. q)
1
2
(A
-1
q. q) (. n).
The corresponding system that denes the saddle point (u. ) follows from (7.4.1).
Indeed,
1(u. cq) _ 1(u. ). Vq U.
where c is an arbitrary positive number and
1(u. ) _ 1(u n. ). Vn V
0
.
Section 7.4 Derivation of a posteriori estimates from saddle point relations 177
Therefore, (u. ) must satisfy the relations
(u. q) (A
-1
. q) = 0. Vq U. (7.4.2)
(. n) (. n) = 0. Vn V
0
. (7.4.3)
or (equivalent) relations
(Au. q) (. q) = 0. Vq U. (7.4.4)
(
+
. n) (. n) = 0. Vn V
0
. (7.4.5)
Let the pair (. ,) (u
0
V
0
)U be an approximation of the saddle point. Estimates
of the deviations u and , directly follow from (7.4.4) and (7.4.5). Indeed, for
u and , we have
(A(u ) ( ,). q) = (, A. q). Vq U. (7.4.6)
(
+
( ,). n) = (n. ,) (. n). n V
0
. (7.4.7)
Here, we set q = (u ) and n = u . Then,
[[(u )[[
2
( ,. (u )) = (, A. (u )). (7.4.8)
(
+
( ,). u ) = (
+
,. u). (7.4.9)
We sum (7.4.8) and (7.4.9) and obtain
[[(u )[[
2
= (A ,. ( u)) (
+
,. u).
which leads to (7.1.19) and (7.1.20).
Another modus operandi applied to (7.4.2) and (7.4.3) implies estimates for the dual
variable. We put q = , and n = u, and we have
((u ). ,) ( A
-1
,. ,) = (A
-1
( ,). ,). (7.4.10)
( ,. ( u)) (,. ( u)) (. u) = 0. (7.4.11)
From (7.4.10) and (7.4.11) it follows that
[[, [[
2
+
= ( A
-1
,. , ) (
+
,. u)
_ [[, A[[
+
[[, [[
+
(, . ( u))
_ [[, A[[
+
[[, [[
+
[[( u)[[[[, [[
+
.
Hence, we arrive at the estimate
[[ ,[[
+
_ [[, A[[
+
[[( u)[[ (7.4.12)
and all estimates (see Section 7.1.3) that follow from it.
8 Nonlinear problems
In this chapter, we discuss a posteriori estimates for certain classes of nonlinear prob-
lems. As in previous chapters, we derive a posteriori estimates by two different meth-
ods. The rst method is based on variational techniques and the second one op-
erates with integral type relations that dene generalized solutions. Both methods
can be applied to nonlinear problems but we pay the main attention to the second
(nonvariational) method. Some results obtained with the help of variational tech-
niques are also discussed but without detailed proofs. The reader can nd them in
[61, 57, 58, 82, 139, 244, 276, 281, 277, 283, 319, 318, 320] and other publications.
8.1 Variational inequalities
Variational inequalities form an important class of nonlinear problems, which often
arise in mechanics and physics (e.g., see G. Duvaut and J.-L. Lions [120], A. Friedman
[133], and R. Glowinski [153]).
Let V be a reexive Banach space, a : V V R be a bilinear V -elliptic form,
and : V R be a given convex continuous functional. Consider the following
problem: Find u 1 such that the inequality
a(u. n u) (n) (u) _ (. n u) (8.1.1)
holds for any n 1, where 1 is a convex closed subset of V
0
and V
+
0
,
a(. n) :=
_
D
V VnJ..
C is a bounded domain in R
2
with Lipschitz continuous boundary, = {a
i}
] is a
symmetric matrix satisfying the conditions (4.1.4). It is well known (e.g., see [120,
153]) that (8.1.1) is equivalent to the variational problem: Find u 1 such that
J(u) = inf
u1
J(n). (8.1.2)
J(n) =
1
2
a(n. n) (n) (. n).
The existence of a minimizer to this problem follows from the coercivity of a on V .
In the literature, variational inequalities the nonlinear features of which arise owing to
the set 1 are often called inequalities of the rst kind. If a nonlinearity is caused
by the presence of a (nondifferentiable) functional , then the inequality is assigned to
the second kind.
Section 8.1 Variational inequalities 179
Exact solutions of variational inequalities may have a complicated structure. Typi-
cally, they contain unknown free boundaries the location and structure of which are a
priori unknown and investigation of their properties is an important part of the a priori
analysis of variational inequalities (e.g., see [133]).
In this section, we show that computable error bounds for elliptic variational in-
equalities of both kinds follow from the corresponding variational inequality.
8.1.1 Variational inequalities of the rst kind
We begin with a classical problem related to the variational inequalities of the rst
kind. We set 0 and dene 1 with the help of pointwise restrictions:
1 := { V
0
:=

H
1
(C) [ (.) _ (.) _ [(.) a.e. in C].
where and [ H
2
(C) are two given functions such that
(.) _ 0 and [(.) _ 0 on I.
(.) _ [(.) in C.
In this case, (8.1.1) reads: Find u 1 satisfying the inequality
a(u. n u) _ (. n u) (8.1.3)
Henceforth, we assume that is dened by an integrable function, i.e.,
(. n) =
_
D
nJ.
and C is a bounded domain with Lipschitz continuous boundary.
Problem (8.1.3) is the classical obstacle problem (e.g., see [133, 153]). Under the
assumptions made, the solution u exists and is unique. In general, C is divided into
three sets, where u is determined either by the differential equation or by obstacles.
They are as follows:
C
u
)
:= {. C [ u(.) = [(.)].
C
u
:= {. C [ u(.) = (.)].
C
u
0
:= {. C [ (.) < u(.) < [(.)].
The sets C
u
)
and C
u
are upper and lower coincidence sets and C

u
0
is an open set where
a solution satises the differential equation. Thus, we see that this problem involves
free boundaries, which are a priori unknown.
In [82, 283], the variational statement (8.1.3) was used to derive a guaranteed upper
bound of the error
[[V(u )[[ := (a(u . u ))
1{2
.
180 Chapter 8 Nonlinear problems
Below we derive the same estimate by transformations of (8.1.3). Assume that 1
is a function compared with u. Then,
a(u . u ) _
_
D
((u ) V V(u )) J.. (8.1.4)
For any vector-valued function , in the space H(C. div), we have
_
D
(ndiv , , Vn) J. = 0. Vn V
0
.
Therefore, (8.1.4) is transformed as follows:
a(u . u ) _
_
D
( div ,)(u ) J.
_
D
(, V) V(u ) J.
=
_
D
(, V) V(u ) J.
_
D
:
)
( div ,)(u ) J.
_
D
:
( div ,)(u ) J.
_
D
:
0
( div ,)(u ) J.. (8.1.5)
where
C
)
:= {. C [ (.) = [(.)].
C
:= {. C [ (.) = (.)].
C
0
:= {. C [ (.) < (.) < [(.)].
Introduce the function
div ,
:=
_
_
_
( div ,)
a.e. in C
)
.
div , a.e. in C
0
.
( div ,)
C
a.e. in C
.
where (g)
C
= max{0. g] and (g)
= min{0. g] denote the positive and negative parts

of g, respectively. Note that
div ,
1
2
(C).
Further analysis rests upon the relation
_
D
( div ,)(u ) J. _
_
D
div ,
(u ) J.. (8.1.6)
from which it follows that the rst term on the right-hand side of (8.1.5) is bounded
from above by the quantity [ div ,
[[ u[. Since
_
D
(V ,) VnJ. _ [[V ,[[
+
[[Vn[[.
we arrive at the estimate (cf. [283])
[[V(u )[[ _ M
OBS
(. ,) := [[V ,[[
+
C[ div ,
[. (8.1.7)
where C is the constant in the inequality [n[ _ C[[Vn[[.
We note that the error majorant M
OBS
(. ,) derived for the obstacle problem con-
sists of two terms. The rst term penalizes the error in the relation
V = ,. (8.1.8)
Another term of M
OBS
(. ,) penalizes the improper behavior of div , on the
sets C
0
, C
)
, and C
, respectively.
It is easy to see that the majorant is a nonnegative functional, which vanishes if and
only if (8.1.8) holds and
div ,
= 0 for a.e. . C. (8.1.9)

The latter relation means that
div ,(.) (.) _ 0 a.e. in C
. (8.1.10)
div ,(.) (.) = 0 a.e. in C
0
. (8.1.11)
div ,(.) (.) _ 0 a.e. in C
)
. (8.1.12)
If (8.1.8) and (8.1.10)(8.1.12) hold, then for any n 1, we have
_
D
V V(n ) J.
_
D
(n ) J. =
_
D
(div , )( n) J.
=
_
D
:
(div , )( n) J.
_
D
:
0
(div , )( n) J.
_
D
:
)
(div , )([ n) J. _ 0.
Thus, satises the variational inequality
a(. n ) _
_
D
(n ) J.. Vn 1.
which coincides with (8.1.3). We have proved the following assertion.
Theorem 8.1. For any V
0
and , H(C. div), the estimate (8.1.7) provides a
guaranteed upper bound of [[V(u )[[. Moreover, if M
OBS
(. ,) = 0, then = u
and , = Vu.
Remark 8.2. Theorem 8.1 shows that problem (8.1.7) admits a new variational state-
ment, namely: Find u 1 and H(C. div) such that
M
OBS
(u. ) = inf
:1
0
,
,!(:,div)
M
OBS
(. ,). (8.1.13)
Unlike the variational statement (8.1.2), the value of the exact lower bound for this
problem is known (it is zero).
A more sophisticated estimate can be derived if (8.1.5) is given in the form
a(u . u ) _
_
D
( div , z
1
z
2
)(u ) J.
_
D
(, V) V(u ) J.
_
D
(z
1
( u) z
2
(u )) J..
where z
i
(.), i = 1. 2, are arbitrary nonnegative functions. Since
_
D
z
1
( u) J. _
_
D
z
1
( ) J.
and
_
D
z
2
(u ) J. _
_
D
z
2
([ ) J..
from (8.1.5) it follows that
[[V(u )[[
2
_
_
C[ div , z
1
z
2
[ [[, V[[
+
_
[[V(u )[[
_
D
(z
1
( ) z
2
([ )) J.. (8.1.14)
By (8.1.14), we deduce an advanced error majorant:
[[V(u )[[ _ M(. ,. z
1
. z
2
)
_
_
D
(z
1
( ) z
2
([ )) J. M
2
(. ,. z
1
. z
2
). (8.1.15)
where
M(. ,. z
1
. z
2
) :=
1
2
_
[[, V[[
+
C[ div , z
1
z
2
[
_
.
Note that (8.1.7) is a special case of (8.1.15). Indeed, set z
1
=

z
1
and z
2
=

z
2
, where
z
1
= 0 on
_
C\ C
_
L
_
C
{. [ div ,(.) (.) _ 0]

_
.
z
1
(.) = [div ,(.) (.)[ on C
{. [ div ,(.) (.) < 0]

and
z
2
= 0 on
_
C\ C
)
_
L
_
C
)
() {. [ div ,(.) (.) _ 0]
_
.
z
2
(.) = [div ,(.) (.)[ on C
)
{. [ div ,(.) (.) > 0].
Then,
_
D
(
z
1
( )

z
2
([ )) J. = 0
and (8.1.15) takes the form
[[V(u )[[ _ 2M(. ,.
z
1
.
z
2
). (8.1.16)
Note that
[ div ,

z
1

z
2
[
2
=
_
D
:
[ div ,

z
1
[
2
J.
_
D
:
0
[ div ,[
2
J.
_
D
:
)
[ div ,

z
2
[
2
J.
=
_
D
:
( div ,)
2
C
J.
_
D
:
0
[ div ,[
2
J.
_
D
:
)
( div ,)
2
J.
= [ div ,
[
2
.
For this reason, 2M(. ,.
z
1
.
z
2
) coincides with M
OBS
(. ,).
From (8.1.14) it also follows that
_
1

2
_
[[V(u )[[
2
_
1
2
_
C[ div , z
1
z
2
[ [[, V[[
+
_
2
_
D
(z
1
( ) z
2
([ )) J.. (8.1.17)
where (0. 2). Using Youngs inequality, we rewrite (8.1.17) in the form
(2 ) [[V(u )[[
2
_
1
C
2
[ div , z
1
z
2
[
2
[[, V[[
2
+
2
_
D
(z
1
( ) z
2
([ )) J.. (8.1.18)
where is an arbitrary positive number. Let
c
:=
1
C
2
and r(,) := div , .
Now, we are aimed at selecting z
1
and z
2
in such a way that the value of the right-
hand side of (8.1.18) attains its minimal value. Unconstrained minimization leads to
the relations
_
D
_
r(,) z
1
z
2
_
J. = 0 (8.1.19)
and
_
D
_
r(,) z
1
z
2
[
c
_
J. = 0. (8.1.20)
which cannot be satised simultaneously. Therefore, either z
1
= 0 or z
2
= 0 (or both
of them are equal to zero). Assume that z
1
= 0. Then, we should set
z
2
= r(,)
[
c
. (8.1.21)
provided that z
2
> 0. Let
(r(,). ) := c
(r(,) z
1
z
2
)
2
2(z
1
( ) z
2
([ ))
denote the integrand of two terms that contain z
1
and z
2
. By (8.1.21), we obtain
(r(,). ) = c
(r(,) z
2
)
2
2z
2
([ )
= 2r(,)([ )
([ )
2
c
if r(,) >
[
c
. (8.1.22)
Another option is to take z
2
= 0 and
z
1
=

c
r(,). (8.1.23)
provided that z
1
> 0. Then,
(r(,). ) = c
(r(,) z
1
)
2
2z
1
( )
= 2r(,)( )
( )
2
c
if

c
> r(,). (8.1.24)

If both z
1
and z
2
are equal to zero, then
(r(y). ) = c
r
2
(,) if

c
< r(,) <

[
c
. (8.1.25)
Now, (8.1.18) takes the form
(2 ) [[V(u )[[
2
_
1
[[, V[[
2
+
_
D
(r(,). ) J.. (8.1.26)
where (r(,). ) is dened in accordance with (8.1.22), (8.1.24), and (8.1.25).
Remark 8.3. Estimates (8.1.7) and (8.1.26) were derived in [298] by the method,
which we have discussed above. In [283], the estimate (8.1.7) was derived by the
variational method. Set = 1. Then c
= c
:= (1 )C
2
, and we arrive at the
estimate
[[V(u )[[
2
_
1
[[, V[[
2
+
_
D
(r(,). ) J.. (8.1.27)
which was also derived in [283].
8.1.2 Variational inequalities of the second kind
Another group of variational inequalities is related to problems with nondifferentiable
functionals. In this case, a solution u is dened by the inequality
a(u. n u) (n) (u) _ (. n u). (8.1.28)
which holds for any n V
0
. As for the rst kind inequalities, the corresponding a
posteriori estimate can be derived directly from (8.1.28). Let be a function compared
with u. Set n = and rearrange (8.1.28) as follows:
a(u . u) () (u) _ (. u) a(. u). (8.1.29)
Since is a convex functional, we know that (cf. (1.4.43))
(u) () _ (u
+
. u ). Vu
+
d().
where d() V
+
0
denotes the subdifferential of at . Therefore, we rewrite (8.1.29)
in the form
a(u . u ) _ (u
+
. u) (. u ) a(. u).
Note that for , 1
2
(C. R
d
) (the divergence of which is an element of V
+
0
), we have
(div ,. n) (,. Vn) = 0. Vn V
0
.
By this identity, we obtain
a(u . u ) _ (u
+
div ,. u)
_
D
(V ,) V( u) J.. (8.1.30)
[[V(u )[[ _ [[[ [[[ u
+
div , [[[ [[[ [[V ,[[
+
. (8.1.31)
where [[[ [[[
+
[[[ [[[ := sup
uV
0
(
,u)
[|u[|
.
Assume that , H(C. div) and d() contains an element that can be identied
with an 1
2
(C)-function. Then, (8.1.31) is represented in the form
[[V(u )[[ _ C[u
+
div ,[ [[V ,[[
+
. (8.1.32)
where u
+
is such an element. In particular, if is G ateaux differentiable and
t
() 1
2
(C) denotes the G ateaux derivative, then we express (8.1.32) as
[[V(u )[[ _ C[
t
() div ,[ [[V ,[[
+
. (8.1.33)
In the next section, we show that the method applied can be extended to a considerably
wider class of problems.
8.2 General elliptic problem. Variational method.
A wide class of variational problems related to various physical models can be given
in the following abstract form:
inf
u
0
V
0
J(. ). J() := G() J(). V V. (8.2.1)
where G and J are convex continuous functionals, V is a reexive Banach space, and
is a linear continuous operator that maps V to another reexive Banach space Y . In
particular, if
G(,) =
1
2
(A,. ,). J() = (. ) . V
+
0
.
then (8.2.1) coincides with (7.1.1). Also, we assume that
c[n[
Y
_ [n[
V
. V V
0
. (8.2.2)
where c is a positive constant, and J is coercive on u
0
V
0
, i.e., for u
0
V
0
J(. ) o as [[
V
o. (8.2.3)
In this case, the problem (8.2.1) has a solution u (e.g., see [121]).
In ([276, 277, 282]), a posteriori error estimates for this class of problems were
derived by the variational method. In order to discuss them, we need to introduce
additional notation.
Section 8.2 General elliptic problem. Variational method. 187
As before, the product of V and
+
in the topologically dual space V
+
is
denoted by (
+
. ). We denote the space topologically dual to Y by Y
+
and the cor-
responding pairing by ((,
+
. ,)). We note that the spaces Y and Y
+
are essentially
different (unlike the case considered in Chapter 7). For this reason, throughout this
chapter we mark functions from Y
+
by stars. The operator
+
: V
+
Y
+
satisfying the relation
((,
+
. n)) =
+
,. n
. Vn V. (8.2.4)
is conjugate to .
By G
+
we denote the Fenchel conjugate of G (cf. (1.4.38)), which is dened by the
relation
G
+
(,
+
) = sup
,Y
_
((,
+
. ,)) G(,)
_
.
Denition 8.4. We say that G and G
+
are uniformly convex in the balls
T
:= T(0. ) Y and T
+
:= T(0.
+
) Y
+
.
respectively, if they satisfy the relations
G
_
,
1
,
2
2
_
_
,
1
,
2
2
_
_
1
2
_
G(,
1
) G(,
2
)
_
. (8.2.5)
G
+
_
,
+
1
,
+
2
2
_
_
,
+
1
,
+
2
2
_
_
1
2
_
G
+
(,
1
) G
+
(,
2
)
_
. (8.2.6)
where
: Y R
and
+
: Y
+
R
are certain nonnegative functionals

vanishing at the zero elements of Y and Y
+
, respectively.
It follows directly from (8.2.5) that any uniformly convex functional G is convex.
Moreover, the functional
reinforces the usual convexity inequality. For this reason,

sometimes it is called the forcing functional (e.g., see R. Glowinski [153]).
Let u
0
V be an approximation of u. The variational method of deriving
computable upper bounds for the quantity [( u)[
Y
is based on the inequality
_
( u)
2
_
_
1
2
_
J(. ) J(u. u)
_
. (8.2.7)
which can be viewed as a generalization of (2.3.1) for problems with uniformly convex
functionals. If G is uniformly convex, then it is easy to prove that (8.2.7) holds with
=
, provided that u and belong to T(0. ). Indeed,
_
( u)
2
_
_
1
2
(G() G(u)) G
_
(u )
2
_
.
0 _
1
2
(J() J(u)) J
_
u
2
_
.
Since
G
_
(u )
2
_
J
_
u
2
_
_ G(u) J(u).
we obtain (8.2.7).
We note that for problems with superquadratic growth, (8.2.7) holds for the whole
space, so that does not depend on .
Another key relation used in the derivation of a posteriori estimates is
inf P := J(u. u) = 1
+
(
+
.
+
+
) = sup P
+
.
where 1
+
is the functional of the so-called dual variational problem P
+
and
+
is the
corresponding solution that maximizes 1
+
on a set of admissible functions (cf. (7.1.9)).
Typically, this set consists of the functions that satisfy certain differential relations. We
will not give here a detailed exposition of the variational method (which can be found
in [282] and [244]) and pass to a discussion of the error majorant derived with the help
of it. This majorant provides an upper bound of the differences u (for the primal
problem) and ,
+

+
(for the dual one) evaluated in terms of the functionals and
+
.
Theorem 8.5 ([276, 282]). Let u be a minimizer of the problem, the functionals J and
G satisfy the above conditions, and u T
and
+
T
+
. Then for any u

0
V
0
and ,
+
Y
+
such that T
and ,
+
T
+
, the estimate
_
( u)
2
_
_
,
+

+
2
_
_ M
I
(. ,
+
)
=
1
2
(D
G
(. ,
+
) D
T
(.
+
,
+
)) (8.2.8)
holds, where
D
T
(.
+
,
+
) := J() J
+
(
+
,
+
)
+
,
+
.
.
D
G
(. ,
+
) := G() G
+
(,
+
) ((,
+
. ,)).
The functionals D
T
and D
G
are nonnegative (this fact follows from the denition
of a polar functional; cf. 1.4). They play an important role in a posteriori analysis
of various variational problems (e.g., see [57, 61, 139, 281, 282, 289, 293, 319, 320,
321]).
By Proposition 1.2, we know that the relation
J() J
+
(
+
,
+
)
+
,
+
.
= 0 (8.2.9)
is equivalent to
+
,
+
dJ() (8.2.10)
Section 8.2 General elliptic problem. Variational method. 189
and the relation
G() G
+
(,
+
) ((,
+
. )) = 0 (8.2.11)
is equivalent to
,
+
dG(). (8.2.12)
Note that (8.2.10) and (8.2.12) are duality relations, which hold if and only if and
,
+
coincide with u and
+
.
Remark 8.6. We observe that for nonlinear problems, it is natural to perform error
control with the help of special functionals . For linear problems, these functionals
coincide with the standard energy norms.
Example. As an example, we apply (8.2.8) to the reaction-diffusion equation with
mixed DirichletRobin boundary conditions. It can be represented as the variational
problem
inf
uV
0
J(. V)
for the functional
J(. V) =
_
D
_
1
2
[V[
2

2
[[
2
_
J.
_
I
2
_
2
[[
2
g
_
Js.
where and are positive real numbers and V
0
is a subspace of H
1
(C), which con-
tains functions vanishing at I
1
.
It is not difcult to show that the minimizer u of this variational problem satises
the relations
û u = 0 in C.
u = 0 on I
1
.
du
dn
u g = 0 on I
2
.
On I
2
the solution satises the so-called Robin boundary condition. An a posteriori
estimate for this problem follows from (8.2.8) if is associated with the operator V
and the functionals G and J are dened by the relations
G(n) =
_
D
1
2
[V[
2
J..
J() =
_
D
2
[[
2
J.
_
I
2
(
2
[[
2
g) Js.
In the case considered, Y and Y
+
are identied with 1
2
(C. R
d
),
,
+
. ,
:=
_
D
,
+
, J..
G
+
(,
+
) = sup
,
_
D
(,
+
,
1
2
[,[
2
) J. =
_
D
1
2
[,
+
[
2
J..
By the integration-by-parts formula
_
D
,
+
V J. =
_
D
div ,
+
J.
_
I
2
(,
+
n) Js. V V
0
.
we conclude that
+
,
+
should be understood as {div ,
+
[
D
. ,
+
n[
I
2
] and
+
,
+
.
=
_
D
div ,
+
J.
_
I
2
(,
+
n) Js.
provided that ,
+
possesses necessary regularity. By a direct substitution, we nd that
G() G
+
(,
+
)
,
+
.
=
_
D
_
1
2
[V[
2
1
2
[,
+
[
2
V ,
+
_
J.
and
J
+
(
+
,
+
) = sup
V
0
_
_
D
(div ,
+
) J.
_
I
2
(,
+
n) Js J()
_
= sup
V
0
_
_
D
(div ,
+
) J.
_
I
2
(,
+
n) Js
_
D
2
[[
2
J.
_
I
2
_
2
[[
2
g
_
Js
_
_ sup
1
2
(D)
_
D
_
(div ,
+
)

2
[[
2
_
J.
sup
)1
2
(I
2
)
_
I
2
_
(,
+
n)j

2
[j[
2
gj
_
Js
=
_
D
1
2
[div ,
+
[
2
J.
_
I
2
1
2
[,
+
n g[
2
Js.
From the above relations, we conclude that
J() J
+
(
+
,
+
)
+
,
+
.
_
_
D
1
2
(div ,
+
)
2
J.
_
I
2
1
2
[,
+
n g [
2
Js. (8.2.13)
Section 8.3 General elliptic problem. Nonvariational method 191
and
G() G
+
(,
+
)
,
+
.
=
_
D
1
2
[V ,
+
[
2
J.. (8.2.14)
which yields both terms of the error majorant. It is easy to see that they vanishes if and
only if
div ,
+
= 0 in C. (8.2.15)
,
+
n g = 0 on I
2
. (8.2.16)
,
+
= V in C. (8.2.17)
Since = 0 on I
1
, the relations (8.2.15)(8.2.17) mean that coincides with the
exact solution u and ,
+
coincides with
+
.
8.3 General elliptic problem. Nonvariational method
In this section, we derive an upper bound of the error from the variational inequality
associated with the problem (8.2.1). The method applied is a generalization of that
was used in Section 8.1.2 for variational inequalities with nondifferentiable terms.
For the sake of simplicity, we assume that G is G ateaux differentiable. Then, the
variational inequality follows from the variational statement with the help of well-
known arguments. Indeed, set n = u z( u) where u
0
V
0
and z is a
positive number. Then n u
0
V
0
and we have
J(n. n) J(u. u) = G(n) J(n) G(u) J(u) _ 0. (8.3.1)
Since
J(n) J(u) = J(u z( u)) J(u) _ z(J() J(u)).
we rewrite (8.3.1) in the form
1
z
(G ((u z( u))) G(u)) J() J(u) _ 0. V u
0
V
0
.
Let z 0, then we arrive at the inequality
((G
t
(u). ( u))) J() J(u) _ 0. (8.3.2)
((G
t
(u) G
t
(). ( u))) J() J(u) _ ((G
t
(). (u ))). (8.3.3)
which is equivalent to
(( u)) := ((G
t
() G
t
(u). ( u)))
_ J() J(u) ((G
t
(). ( u)))
_
u
+
. u
((G
t
(). ( u)))
_ ((G
t
() ,
+
. ( u)))
+
,
+
u
+
. u
. (8.3.4)
where u
+
is an element of the set dJ() and ,
+
is a function in Y
+
.
Since G is a convex functional, its derivative is a monotone operator. Therefore, the
quantity (( u)) := ((G
t
() G
t
(u). ( u))) is nonnegative and provides
a certain measure of the error.
We can deduce computable bounds of errors by one of the methods discussed below.
The rst method. Note that
((G
t
() ,
+
. ( u))) _ [G
t
() ,
+
[
Y
[( u)[
Y
(8.3.5)
and
+
,
+
u
+
. u
_ [
+
,
+
u
+
[
V
[ u[
V
. (8.3.6)
In view of (8.2.2),
[u [
V
_ c[( u)[
Y
. (8.3.7)
and we obtain the estimate
(( u)) _ [G
t
() ,
+
[
Y
c[
+
,
+
u
+
[
V
. (8.3.8)
where
(( u)) :=
(( u))
[( u)[
Y
is a nonnegative error functional. The second norm on the right-hand side of (8.3.8) is
a norm of the space topologically dual to V , which may be incomputable. Therefore,
it is desirable to represent the estimate in a somewhat different form. One can make
this if the second product in (8.3.4) is estimated by the inequality
+
,
+
u
+
. u
_ [
+
,
+
u
+
[
U
[ u[
U
. (8.3.9)
where U and U
+
is a pair of dual spaces the norms of which are dened by ex-
plicitly computable integrals (e.g., 1
]
and 1
]
/
with indices satisfying the condition
1, 1,
t
= 1), V is continuously embedded in U, and
[n[
U
_ C
I1
[n[
V
. Vn V. (8.3.10)
where C
I1
does not depend on n. Then, there exists a constant C (which is less than
C
I1
c) such that
[n[
U
_ C[n[
Y
. Vn V
0
. (8.3.11)
and (8.3.8) is replaced by
(( u)) _ M
II
(. ,
+
) := [G
t
() ,
+
[
Y
C[
+
,
+
u
+
[
U
. (8.3.12)
It is easy to note that the right-hand side of (8.3.12) vanishes if and only if
,
+
= G
t
() and
+
,
+
dJ().
which is equivalent to (8.2.10) and (8.2.12).
Assume that G is a differentiable functional. Then
G(,
1
,
2
,
1
2
) _ G(,
1
)
1
2
((G
t
(,
1
). ,
2
,
1
)).
1
2
((G
t
(,
2
). ,
2
,
1
)) _
1
2
G(,
2
)
1
2
G(,
1
).
and we nd that
1
2
((G
t
(,
2
) G
t
(,
1
). ,
2
,
1
)) _
1
2
G(,
2
)
1
2
G(,
1
) G
_
,
1
,
2
2
_
. (8.3.13)
In view of (8.2.5), we have
((G
t
(,
2
) G
t
(,
1
). ,
2
,
1
)) _ 2
_
,
2
,
1
2
_
(8.3.14)
for ,
1
. ,
2
T
. By (8.3.14), we conclude that

(( u)) _
2
_
(-u)
2
_
[( u)[
Y
. (8.3.15)
Thus, properties of the error functional in (8.3.12) are determined by properties of the
forcing functional.
Remark 8.7. If G does not have the G ateaux derivative at a certain point, then the
above relation holds, provided that it is replaced by an element of the corresponding
subdifferential set dG().
Remark 8.8. In addition to
and , errors can be estimated in terms of the quantity

D
G
(.
+
) = G() G
+
(
+
) ((
+
. )).
which is a certain measure of that how accurately reproduces
+
= u. Relations
between the error measures
(( u)), (( u)), and D

G
(.
+
) were stud-
ied in [244, 286, 298] (also, see comments and references in Section 8.4.1 related to
properties of D
G
(.
+
)). In particular, it is easy to see that
D
G
(.
+
) = G() G
+
(
+
) ((
+
. ))
_ ((G
t
(). ( u))) G(u) G
+
(
+
) ((G
t
(u). ))
= ((G
t
(). ( u))) ((G
t
(u). u)) ((G
t
(u). ))
= (( u)).
The second method. Let : R R be a nonnegative function and
+
be the
Fenchel conjugate to . We derive another estimate from (8.3.4), using the following
arguments. Take a number > 0 and apply Youngs inequality to the rst term on the
right-hand side of (8.3.4). We have
((G
t
() ,
+
. ( u)))
_ [( u)[
Y
_
_
_
_
G
t
() ,
+
_
_
_
_
Y
_ ([( u)[
Y
)
+
__
_
_
_
G
t
() ,
+
_
_
_
_
Y
_
. (8.3.16)
Analogously, let [ and [
+
be another pair of nonnegative functions, where [
+
is
conjugate to [. Take > 0 and apply Youngs inequality to the second term in
(8.3.4). With the help of (8.3.11) we nd that
+
,
+
u
+
. u
_
_
_
_
_
+
,
+
u
+
_
_
_
_
V
c[( u)[
Y
_ c[ ([( u)[
Y
) c[
+
__
_
_
_
+
,
+
u
+
_
_
_
_
V
_
. (8.3.17)
By (8.3.4), (8.3.16), and (8.3.17), we deduce the estimate
(( u))
_
+
__
_
_
_
G
t
() ,
+
_
_
_
_
Y
_
c [
+
__
_
_
_
+
,
+
u
+
_
_
_
_
V
_
. (8.3.18)
where
(( u)) := (( u)) ([( u)[

Y
) c[ ([( u)[
Y
) .
and it is assumed that , , , and [ are selected in such a way that
is a nonneg-
ative functional.
If (8.3.9) holds, then we obtain
(( u)) _
+
__
_
_
_
G
t
() ,
+
_
_
_
_
Y
_
C[
+
__
_
_
_
+
,
+
u
+
_
_
_
_
U
_
=: M
III
(. ,
+
: . ). (8.3.19)
Let the functional satisfy the relation
(( u)) _ 2 ([( u)[
Y
) . (8.3.20)
where is a nonnegative increasing function of the energy norm. Then,
([( u)[
Y
) _ (2 )([( u)[
Y
).
and it is natural to set = [ = . In this case, (8.3.19) has the form
(2 )([( u)[
Y
)
_
+
__
_
_
_
G
t
() ,
+
_
_
_
_
Y
_
C
+
__
_
_
_
+
,
+
u
+
_
_
_
_
U
_
. (8.3.21)
Particular case. We show that for the linear diffusion problem, the general estimates
given above lead to well-known estimates established in Chapter 4 for the problem:
Find u V
0
=

H
1
(C) such that
a(u. n) = (. n)
U
. Vn V
0
.
In this case,
= V. G() =
1
2
a(. ). and J() =
_
D
J..
where U = U
+
= 1
2
(C). Then
G
t
() = V. (( u)) = [[V(u )[[
2
.
+
,
+
= div ,
+
.
+
,
+
u
+
= div ,
+
.
and (8.3.8) gives the estimate
[[V(u )[[ _ [[,
+
V[[
+
C[div ,
+
[. (8.3.22)
Now, we derive this estimate with the help of (8.3.20). Note that (cf. (8.3.20))
(V( u)) = 2([V( u)[).
where (t ) =
1
2
t
2
. By (8.3.21), we have
+
__
_
_
_
G
t
() ,
+
_
_
_
_
Y
_
=
1
2
[[,
+
V[[
2
+
.
C
+
__
_
_
_
+
,
+
u
+
_
_
_
_
U
_
= C
1
2
_
_
div ,
+
_
_
2
.
and (8.3.21) implies the estimate
(2 )[[V( u)[[
2
_
1
[[,
+
V[[
2
+
_
_
div ,
+
_
_
2
. (8.3.23)
Set =
1
1;
and =
;
1;
, where ; > 0. Then = 1, and we arrive at the
estimate
[[V( u)[[
2
_ (1 ;)[[,
+
V[[
2
+
C
1 ;
;
_
_
div ,
+
_
_
2
. (8.3.24)
which is a special form of (4.1.14).
8.4 A posteriori estimates for special classes of nonlinear
elliptic problems
8.4.1 -Laplacian
Let be a real number greater than 1, and let
+
=

-1
be the corresponding conju-
gate number. Consider the problem: Find u V := W
1,
(C) such that u = u
0
on I
and
div
_
[Vu[
-2
Vu
_
= 0 in C. (8.4.1)
The weak statement of the problem is given by the integral identity
_
D
_
[Vu[
-2
Vu Vn n
_
J. = 0. V n V
0
. (8.4.2)
where is assumed to be of class 1
(C), V
0
is the subspace of V formed by the
functions vanishing at the boundary.
Variational method. By the variational method, a posteriori estimates were derived
in [278, 61]. In this analysis, we rest upon the minimization problem 1
inf
uu
0
V
0
J
(n). J
(n) :=
_
D
_
1
[Vn[
n
_
J.. (8.4.3)
Section 8.4 A posteriori estimates for special classes of nonlinear elliptic problems 197
As usual, by u
0
V
0
we denote the subspace of V containing functions n = u
0
n
0
,
where n
0
V
0
and u
0
is a given function in V . The existence and uniqueness of a
minimizer follow from the strict convexity of J
. In the variational method, we need to

consider the dual variational problem. For this purpose, we introduce the Lagrangian
1(. ,
+
) :=
_
D
_
V ,
+
+
[,
+
[

_
J.
and note that
J
() = sup
,
1(. ,).
The dual variational functional is dened by the relation
1
+
(,
+
) := inf
u
0
V
0
1(. ,
+
)
and has the form
1
+
(,
+
) :=
_
_
_
_
D
_
Vu
0
,
+
+
[,
+
[
u
0
_
J. if ,
+
Q
+
(
.
o if ,
+
Q
+
(
.
where
Q
+
(
:=
_
,
+
Y
+
[
_
D
,
+
VnJ. =
_
D
nJ.. Vn V
0
_
and Y
+
:= 1
(C: R
d
). The dual problem P
+
associated with the primal problem
(8.4.3) is as follows: Find
+
Y
+
such that 1
+
1
+
(
+
) = sup
,
1
+
(,
+
). (8.4.4)
Theorem 8.9. The problems P and P
+
have unique solutions u and
+
, respectively;
J
(u) = 1
(
+
). (8.4.5)
+
= [Vu[
-2
Vu a.e. in C. (8.4.6)
Vu = [
+
[
-2
+
a.e. in C. (8.4.7)
The proof of Theorem 8.9 follows from well-known results of convex analysis (e.g.,
see I. Ekeland and R. Themam [121]). If = 2, then we arrive at the quadratic
functional, which generates a linear elliptic equation. Properties of problems related
to the superquadratic ( > 2) and subquadratic ( (1. 2)) cases are rather different.
For this reason, we discuss them separately.
Estimates for problems with superquadratic growth. Variational problems with
superquadratic growth were among the rst studied in the context of the variational
approach to a posteriori error estimation (see [278]). For (1. o), computable
upper bounds of the error were derived in M. Bildhauer and S. Repin [61]. These
results are based on the uniform convexity of J
(see S. Sobolev [336] and P. Mosolov

and V. Myasnikov [239]). For _ 2, the uniform convexity follows from the rst
Clarksons inequality
_
D
_
,
1
,
2
2
,
1
,
2
2
_
J. _
1
2
[,
1
[
,D
1
2
[,
2
[
,D
. V,
1
. ,
2
Y.
(8.4.8)
0
V
0
, the following estimate holds:
[V( u)[
,D
_ 2
-1
_
1
() 1
+
(q
+
)
_
. Vq
+
Q
+
(
. (8.4.9)
Proof. We use (8.4.8). Setting ,
1
= Vu and ,
2
= V, we obtain
_
D
[V( u)[
J. _ 2
-1
__
D
[V[
J.
_
D
[Vu[
J. 2
_
D
V(u )
2
J.
_
= 2
-1
_
1
() 1
(u) 21
_
u
2
__
_ 2
-1
(1
() 1
(u)) .
1
(u) _ 1
+
(q
+
). Vq
+
Q
+
(
.
which leads to (8.4.9).
Remark 8.11. We note that (8.4.9) can be viewed as a generalized form of (2.3.1)
related to variational problems with power growth.
In [61], the difference 1
() 1
+
(q
+
) was analyzed and rearranged into a com-
putable form.
Theorem 8.12. Let _ 2 and u
0
V
0
. For any function ,
+
Y
+
that has
divergence summable with power
+
and for any real number > 0,
1
2
-1
[V( u)[
,D
_ M
(. ,
+
. ) := M
1
(. ,
+
. ) M
2
(,
+
. ). (8.4.10)
where
M
1
(V. ,
+
. ) = D
(V. ,
+
)

[[,
+
[
-2
,
+
V[
,D
.
M
2
(,
+
. ) = C
T
_
1
2
2-
(3
+
)
_
[ div ,
+
[
,D
.
C
T
is the constant in the Friedrichs type inequality
[n[
,D
_ C
T
[Vn[
,D
.
and the functional D
: Y Y
+
R
0
is dened by the relation
D
(,. ,
+
) :=
_
D
_
1
[,[
+
[,
+
[
, ,
+
_
J..
It is easy to see that the right-hand side of (8.4.10) vanishes if and only if the rela-
tions
[,
+
[
-2
,
+
= V. (8.4.11)
div ,
+
= 0 (8.4.12)
hold almost everywhere in C. Since the solution of the problem (8.4.3) is unique, the
relations (8.4.11) and (8.4.12) are equivalent to the fact that = u and ,
+
=
+
. The
functional D
can be viewed as a certain measure of the error in the duality relations

(8.4.6) and (8.4.7). Indeed, this functional is nonnegative and vanishes if and only if
(cf. Proposition 1.2)
, = [,
+
[
-2
,
+
and ,
+
= [,[
-2
,.
The second term of M
1
possesses the same properties. Therefore, M
1
is a measure of
the error in the duality relations. The term M
2
(,
+
. ) penalizes the violation of the
relation div,
+
= 0.
Comments. It should be noted that we cannot prove that there always exist ,
+
and
such that the right-hand side of (8.4.10) coincides with the left-hand one. How-
ever, such a property may hold for some special quantities introduced to character-
ize the accuracy of an approximate solution (this question is discussed in [61, 244,
282, 286, 298] and some other publications). For example, instead of the 1
-norm of
V( u) we can take a special measure of the error dened by the compound func-
tional D
G
(V.
+
), which in our case has the form
D
(V.
+
) =
_
D
_
1
[V[
+
[
+
[
V
+
_
J. = J
() 1
(
+
)
= J
() J
(u).
The majorant M
is deduced as an upper bound of J
() 1
(
+
) and, therefore,
provides an upper estimate for D
(V.
+
). It is easy to see that for ,
+
=
+
, the
term M
2
(,
+
. ) vanishes and M
1
(. ,
+
. ) coincides with D
(V.
+
). The latter
quantity is a nonnegative functional, which can be regarded as a certain measure of the
error. In terms of such a measure, the majorant M
is sharp. Similar results hold for

other nonlinear problems (see above-cited publications).
If = 2, then the quantity
J
() := D
2
(V.
+
)
coincides with
1
2
[V( u)[
2
2,D
and gives a natural energy norm of the error. If ,= 2,
then J
() is not a norm. Nevertheless,

J
() := D
(V.
+
) _ 0.
and J
() = 0 if and only if
V = [
+
[
-2
+
= Vu a.e. in C.
In some cases, J
() can be regarded as a certain weighted norm of the deviation from

the exact solution. For example, for = 3, we have
J
3
() _
1
3
_
D
([V[ 2[Vu[)[V( u)[
2
J..
However, in some cases it is important to get realistic error bounds in terms of the
energy norm. In general, this is a more complicated task. We shortly discuss it in
Section 8.5.
Estimates for problems with subquadratic growth. For < 2, it is more conve-
nient to derive a posteriori estimates within the framework of Problem (P)
+
. In this
case,
+
> 2 and the dual functional
_
1
+
_
(,
+
) =
_
D
_
Vu
0
,
+
+
[,
+
[
u
0
_
J.
is uniformly convex on Y
+
. Therefore, for any q
+
1
and q
+
2
Q
+
(
we apply Clarksons
inequality and obtain
_
1
+
_
_
q
+
1
q
+
2
2
_
_
D
_
q
+
1
q
+
2
2
_
J. _
1
2
_
_
1
+
_
(q
+
1
)
_
1
+
_
(q
+
2
)
_
.
which allows us to derive an upper bound of [
+
,
+
[
,D
, where ,
+
is an approxi-
mate one. The assertion below is an analog of Theorem 8.12 for the case
+
> 2.
Theorem 8.13. For any function u
0
V
0
, any vector-valued function ,
+
Y
+
such that div,
+
is summable with power
+
, and any > 0, the following estimate
holds:
[
+
,
+
[
,D
_ M
(. ,
+
. ) := M
1
(. ,
+
. ) M
2
(,
+
. ). (8.4.13)
Here M
1
(. ,
+
. ) = 2
+
j
2
_
D
(V. ,
+
)

2
[[,
+
[
-2
,
+
V[
2
,D
_
.
M
2
(,
+
. ) = 2
+
j
2
_
1
2
(r(,
+
))
2
(
+
1)kj
_
j[,
+
[
-2
,D
(r(,
+
))
-2
_
(r(y
+
))
2
_
2(r(y
+
))
ff
.
j = 2
-2
, and r(,
+
) is dened by the relation r(,
+
) := C
T
[ div ,
+
[
,D
.
The reader interested in a more detailed discussion of functional a posteriori esti-
mates for variational problems with power growth functionals is referred to [61, 57,
139, 244, 278].
8.4.2 Problems with nonlinear boundary conditions
Boundary conditions in general form. In many cases, the commonly used Dirich-
let or Neumann boundary conditions cannot properly describe the behavior of a model
and should be replaced by more sophisticated conditions that reect real physical situa-
tions. Typical examples are presented by problems with unilateral boundary conditions
and friction (e.g., see G. Duvaut and G.-L. Lions [120] and P. Panagiotopoulos [259]).
The corresponding boundary value problems are formulated as variational inequali-
ties and can be solved by known numerical methods (e.g., see R. Glowinski [153] and
R. Glowinski, J.-L. Lions, and R. Tr emolier` es [155]).
We study the case of nonlinear boundary conditions with the paradigm of the prob-
lem
div Vu = 0 in C. (8.4.14)
where satises (4.1.4) and
u(.) = u
0
(.). . I
1
. (8.4.15)
Vu n(.) d(u(.)). . I
2
. (8.4.16)
Henceforth, we assume that 1
2
(C), u
0
H
1
(C), and the boundary I consists
of two disjoint measurable parts I
1
and I
2
. Also, we assume that it is piecewise
smooth, so that one can uniquely dene the unit outward normal at almost all points
of I. By u
,n
we denote the normal derivative of u and : R
d
R is a convex
lower semicontinuous functional, which determines the so-called boundary dissipative
potential (e.g., see [259]). The relation (8.4.16) is the general form of a wide spectrum
of boundary conditions. Similar relations are often used in continuum mechanics if it
is necessary to model unilateral boundary contact or contact with friction. In this case,
boundary conditions are represented in the form
o
n
(.) d(u(.)). . I
2
. (8.4.17)
where o is the stress tensor and u is the displacement. Problem (8.4.12)(8.4.15) is a
simplied version of the elasticity model, in which u is a scalar-valued function and
(8.4.17) is replaced by (8.4.14). However, from the mathematical point of view these
two problems are quite similar.
We recall that any V = H
1
(C. R
d
) has a trace on I denoted by ; (cf.
(1.4.18) and (1.4.19)), where ; L
_
H
1
(C). H
1{2
(I)
_
is the trace operator. In the
preceding sections, the operator ; was omitted in the formulas associated with bound-
ary relations. In this section, we keep it in explicit form. Let V stand for H
1
(C)
and
V
0
:= { V [ ; = 0 a.e. on I
1
] .
which is a subspace of V . The set ;(V
0
) is denoted by T (it is a subspace of H
1{2
(I)).
The corresponding dual space T
+
contains traces (on I
2
) of the functions from the
space H(C. div). Indeed, for any smooth ,
+
and any V
0
, we have the relation
_
I
2
(,
+
n
) Js =
_
D
(,
+
V (div,
+
)) J.. (8.4.18)
where ,
+
n
:= ,
+
n. For any ,
+
H(C. div), the right-hand side of this identity is a
linear continuous functional
,
: V
0
R satisfying the relations
,
() = 0. V

H
1
(C). (8.4.19)
,
()
_ c
_
_
,
+
_
_
div
[;[
1
1/2
,I
. (8.4.20)
In fact,
,
is a linear continuous mapping dened on a factor space of V
0
two elements
of which are considered as different only if they have different traces on I. Indeed,
,
(
1
) =
,
(
2
) if
1
.
2
V
0
and ;
1
= ;
2
on I. For this reason,
,
can be
identied with a certain element in T
+
, which we denote by ;
,
+
n
and call the normal
trace of ,
+
on I
2
. The value of the functional
+
T
+
on T is denoted by
(
+
. )
I
2
. Then, (8.4.18) has the form
(;
,
+
n
. ;)
I
2
=
_
D
(,
+
V div ,
+
) J.. (8.4.21)
The norm of ;
,
+
n
is dened by the standard relation
_
_
;
,
+
n
_
_
T
:= sup
V
0
(;
,
+
n
. ;)
I
2
[;[
T
= sup
V
0
_
D
(,
+
V div ,
+
) J.
[;[
T
. (8.4.22)
_
_
;
,
+
n
_
_
T
_ c
_
_
,
+
_
_
div
. (8.4.23)
Conjugate functionals dened on spaces of traces. For any T we dene the
functional
J() :=
_
I
2
() Js.
We assume that the integrand : R
d
R
d
is a nonnegative, continuous, and convex
functional such that (0) = 0 and dom := { R
d
[ () < o] ,= 0. so that
belongs to the class of so-called proper convex functionals.
In this case, the functional J() is also nonnegative, convex, and lower semicontin-
uous on T. Since ; is a bounded linear operator, the functional J(;) also possesses
the above properties as a functional on V
0
.
By denition, dJ() contains elements
+
T such that
J( n) J() _ (
+
. n)
I
2
. Vn V
0
. (8.4.24)
and
J
+
(
+
) := sup
T
(
+
. )
I
2
J()
_
is the functional conjugate to the functional J.
By recalling (8.4.21), we nd that
J
+
(;
,
+
n
) = sup
uV
0
__
D
_
,
+
Vn div ,
+
n
_
J. J( ;n)
_
and dene the compound functional
D
T
2
(;. ;
,
+
n
) := J(;) J
+
(;,
+
n
) (;. ;,
+
n
)
I
2
generated by traces on I
2
. It is easy to see that
D
T
2
(;. ;
,
+
n
) := sup
uV
0
_
_
D
_
,
+
V(n ) div ,
+
(n )
_
J.
_
I
2
((;) (;n)) Js
_
_ 0.
Moreover, if D
T
2
(;. ;
,
+
n
) = 0. then (see Proposition 1.2) ;
,
+
n
dJ(;). If ,
+
n
is
sufciently regular (e.g., ;
,
+
n
1
2
(I
2
. R
d
)), then J
+
has an explicit form
J
+
(;
,
+
n
) =
_
I
2
+
(;
,
+
n
) Js.
where
+
(
+
) = sup
R
o {
+
()].
Variational inequality. Now we can state the problem as a variational inequality:
Find u u
0
V
0
such that
a(u. n u) J(n) J(u) _
_
D
(n u) J. (8.4.25)
holds for all n u
0
V
0
. This problem is equivalent to the variational problem: Find
u u
0
V
0
such that
J(u) = inf
uu
0
V
0
J(n). J(n) =
1
2
a(n. n) J(n)
_
D
nJ.. (8.4.26)
The functional J is strictly convex, continuous, and coercive on V , and the set u
0
V
0
is a convex closed subset of V . Therefore, the variational problem (8.4.25) is uniquely
solvable.
Assume that the function u is sufciently regular (e.g., has rst and second deriva-
tives in the classical sense). Then, from (8.4.24) it follows that
_
D
(div Vu )(n u) J. _
_
I
2
((n) (u)) Js
_
I
2
Vu n(n u) Js
for any n u
0
V
0
. Set n = u on I. Then, this relation implies (8.4.14). Hence,
we nd that
_
I
2
((n) (u) Vu n)(n u) Js _ 0. Vn u
0
V
0
.
Estimates of the difference between the exact and approximate solutions to (8.4.25)
can be obtained by any of the two methods (variational and nonvariational).
A posteriori estimates. The nonvariational method. Let u
0
V
0
be an ap-
proximation of u. We substitute it into (8.4.25) and rewrite in the form
a(u . u) J() J(u) _
_
D
( u) J. a(. u). (8.4.27)
By (8.4.24), we nd that
a(u . u ) _ (u
+
. u)
I
2

_
D
(u ) J. a(. u). (8.4.28)
where u
+
is an element of the set dJ(). In view of (8.4.21),
(;
,
+
n
. ;( u))
I
2
=
_
D
(,
+
V( u) div ,
+
( u)) J..
and we rearrange the right-hand side of (8.4.28) as follows:
a(u . u ) _ (u
+
;
,
+
n
. ;( u))
I
2

_
D
(div ,
+
)(u ) J.
_
D
(V ,
+
) V(u ) J.. (8.4.29)
Now we recall that (cf. (8.4.22) and (4.1.9))
(u
+
;
,
+
n
. ;( u))
I
2
_
_
_
u
+
;
,
+
n
_
_
T
[;( u)[
T
_ C
TI
2
_
_
u
+
;
,
+
n
_
_
T
[V(u )[.
By this relation, we obtain
[[V(u)[[ _ [[V ,[[
+
1
c
1
_
C
T D
[div ,
+
[
_
_
u
+
;
,
+
n
_
_
T
_
. (8.4.30)
This estimate yields the general form of the upper bound.
Assume that the boundary potential is sufciently regular, so that we can write
J() J(u) =
_
I
2
(() (u)) Js _
_
I
2
;( u) u
+
Js. (8.4.31)
where u
+
(which is determined by d()) can be identied with a square integrable
function dened on I
2
(we assume that such an element exists). In addition, assume
that ;
,
+
n
can be identied with a square integrable function dened on I
2
. Then,
[[V(u )[[ _ [[V ,[[
+
1
c
1
_
C
T I
1
[div ,
+
[ C
TI
2
[u
+
;
,
+
n
[
I
2
_
. (8.4.32)
where C
T I
1
and C
TI
2
are dened by (4.1.8) and (4.1.9). This estimate is a gener-
alization of (4.1.12). It shows that the error bound is represented as the sum of three
terms that penalize the relations
div ,
+
= 0 in C. (8.4.33)
,
+
= V in C. (8.4.34)
;
,
+
n
d() on I
2
. (8.4.35)
A posteriori estimates. The variational method. Since the minimizer u to the
problem P satises (8.4.25), we nd that
J() J(u) =
1
2
a( u. u) a(u. u) (. u) J() J(u)
_
1
2
a( u. u). V u
0
V
0
. (8.4.36)
which implies the basic deviation estimate
1
2
[[V(u )[[
2
_ J() inf P. V u
0
V
0
. (8.4.37)
The right-hand side of (8.4.37) can be estimated from above by a method similar to
that discussed in Section 3.1.
Let be an arbitrary approximation in u
0
V
0
. Then, the rst upper bound of the
error for problems with nonlinear boundary conditions is given by the estimate
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
) M
2
(;.
+
)
1
M
3
(,
+
.
+
). (8.4.38)
where ,
+
,
+
, and are arbitrary elements of the sets 1
2
(C. M
dd
), T
+
, and R
,
respectively. The functionals M
1
, M
2
, and M
3
are dened by the relations
M
1
(. ,
+
) =
1
2
_
D
(V V
-1
,
+
,
+
2V ,
+
) J..
M
2
(;.
+
) = J(;) J
+
(
+
) (
+
. ;)
I
2
.
M
3
(,
+
.
+
) =
1
2
inf
)
[[j
+
,
+
[[
2
+
.
where
Q
+
I
:=
_
,
+
1
2
(C. M
dd
) [
_
D
,
+
VnJ. =
(n). Vn V
0
_
.
and
(n) :=
_
D
nJ. (
+
. ;n)
I
2
.
The proof of (8.4.35) can be found in S. Repin and J. Valdman [319].
It is clear that the quantities M
1
, M
2
, and M
3
are nonnegative. The quantity
M
1
(. ,
+
) vanishes if and only if and ,
+
satisfy the relation (8.4.34) and
M
2
(;.
+
) = 0 if and only if
+
dJ(;) on I
2
. Thus, M
2
is a measure of the
error in the boundary condition (8.4.14) computed on I
2
for the function
+
T
+
(which can be thought of as an image of the normal component of the ux) and the
trace of . The quantity M
3
(,
+
) vanishes if and only if ,
+
Q
+
I
, i.e., if
_
D
,
+
VnJ. =
_
D
nJ. (
+
. ;n)
I
2
. Vn V
0
.
Since
_
D
,
+
VnJ. = (;
,
+
n
. ;n)
I
2

_
D
div ,
+
nJ..
we conclude that M
3
(,
+
) vanishes if and only if the equilibrium equation (8.4.33) and
the relation ;
,
+
n
=
+
hold (in a generalized sense).
Estimate (8.4.35) can be represented in a form, which is more convenient from the
practical point of view. Assume that
+
1
2
(I
2
). ,
+
H(C. div). and ;
,
+
n
1
2
(I
2
. R
d
). (8.4.39)
Then, the term M
3
can be estimated from above by a directly computable quantity
(using the same arguments as in Lemmas 3.2 and 7.3) and we obtain
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
) M
2
(;.
+
)
1
2c
2
1
_
C
T I
1
[r
:
(,
+
)[
C
TI
2
_
I
2
(
+
;
,
+
n
)
2
Js
_
2
. (8.4.40)
Here > 0 and r
:
(,
+
) := div ,
+
. We see that (8.4.40) has the same principal
structure as (8.4.38). The difference is that the functions ,
+
and
+
are integrable func-
tions. As before, the majorant vanishes if and only if the relations (8.4.33)(8.4.35)
hold.
Particular forms of (8.4.40) Estimate (8.4.40) has particular forms, which deserve
special comments.
1. The rst form arises if we set
+
= ;
,
+
n
. (8.4.41)
i.e., if we dene
+
(which is an image of the true boundary ux), using a known
approximation of the dual variable. In this case, the last integral on the right-hand side
of (8.4.40) vanishes, and we arrive at the estimate
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
) M
2
(;. ;
,
+
n
)
1
2c
2
1
C
2
T I
1
[r
:
(,
+
)[
2
. (8.4.42)
2. Another option is to take
+
in accordance with the relation
+
d(). (8.4.43)
i.e., dene it using the approximation . In this case, the second integral on the right-
hand side of (8.4.40) vanishes and we obtain
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
)
1
2c
2
1
_
C
2
T I
1
[r
:
(,
+
)[
2
C
TI
2
_
I
2
(u
+
;
,
+
n
)
2
Js
_
2
. (8.4.44)
Note that M
1
(. ,
+
) =
1
2
[[V ,
+
[[
2
+
. Therefore, (8.4.44) is the squared form of
(8.4.32).
3. We can try to select
+
in the best possible way. For this purpose, we estimate
the last term of (8.4.40) by means of Youngs inequality. Then, we get the following
inequality, which involves a new positive constant :
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
) M
2
(;.
+
)
1
2
_
1
1
_
(1 )C
2
T I
1
[r
:
(,
+
)[
2
1
2
_
1
1
_
_
1
1
_
C
2
TI
2
_
I
2
(
+
;
,
+
n
)
2
Js. (8.4.45)
We gather the terms related to I
2
and denote them by
1
I
2
(;. ;
,
+
n
.
+
) =
_
I
2
_
(;)
+
(
+
)
+
;
0
2
,
+
n

+
2
_
Js. (8.4.46)
where 0 =
_
1
1
_
_
1
1
_
C
2
TI
2
.
To minimize the right-hand side of (8.4.46), we need to minimize 1
I
2
with respect
to
+
, i.e., to solve the problem
inf
1
2
(I
2
)
_
+
(
+
)
0
2
[
+
[
2

+
;
+
;
,
+
n
_
.
Under the assumptions made, the corresponding minimum has the form
M
I
2
(;. ;
,
+
n
. 0) :=
_
I
2
_
(;)
0
2
,
+
n
2
j(; 0;
,
+
n
)
_
Js.
where j : R
d
R is the function conjugate to
+
(
+
)
0
2
+
2
. Now, we nd that
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
) M
I
2
(;. ;
,
+
n
. 0)
1
2
_
1
1
_
(1 )C
2
T I
1
[r
:
(,
+
)[
2
. (8.4.47)
Neumann type boundary condition. The boundary condition of this type corre-
sponds to the case
J() := (j
+
. )
I
2
. (8.4.48)
where j
+
T
+
. In particular, if j
+
is associated with a function J 1
2
(I
2
), then
(;u) = J ;u. d(;u) = J.
and (8.4.48) is the Neumann boundary condition
;(n
+
) = J a. e. on I
2
. (8.4.49)
Since
J() =
_
I
2
J Js
and
J
+
(
+
) =
_
0 if
+
= J a.e. on I
2
.
o otherwise.
we nd that
1
I
2
(;. ;
,
+
n
.
+
) =
_
I
2
_
J ; 0 J ;
0
2
,
+
n
J
2
_
Js
=
0
2
_
I
2
,
+
n
J
2
Js.
Hence, (8.4.47) has the form
[[V(u )[[
2
_ (1 )2M
1
(. ,
+
)
_
1
1
_
(1 )C
2
T I
1
r
:
2
(,
+
)
_
1
1
_
_
1
1
_
C
2
TI
2
_
I
2
,
+
n
J
2
Js. (8.4.50)
The minimization of the right-hand side with respect to leads to the estimate
[[V(u )[[
2
_ (1 )[[V ,
+
[[
2
+
_
1
1
_
_
C
T I
1
[r
:
(,
+
)[ C
TI
2
[;
,
+
n
J[
I
2
_
2
. (8.4.51)
which is the squared form of (4.1.12).
Friction type boundary condition. In this case,
(;) = j[;[. J() =
_
I
2
j[[ Js. j > 0. (8.4.52)
and
+
(
+
) =
_
0 if [
+
[ _ j.
o otherwise.
(8.4.53)
d() =
_
_
_
j if > 0.
t j. j| if = 0.
j if < 0.
Therefore, if [
+
n[ < j, then (cf. (8.4.16)) implies ; = 0; ; can take nonzero
values are only if [
+
n[ attains limit values stated by the constant j.s
Set
+
= ;
,
+
n
and impose the condition [;
,
+
n
[ _ j. Then,
M
2
(;. ;
,
+
n
) =
_
I
2
_
j[;[ (;
,
+
n
)(;)
_
Js. (8.4.54)
provided that ;
,
+
n
is a square integrable function on I
2
. Now the estimate (8.4.45)
takes the form
1
2
[[V(u )[[
2
_ (1 )M
1
(. ,
+
)
_
I
2
_
j[;[ (;
,
+
n
)(;)
_
Js
1
2
_
1
1
_
(1 )C
2
T I
1
r
:
2
(,
+
). (8.4.55)
where [;
,
+
n
[ _ j.
Assume that the right-hand side of (8.4.55) is zero. Then
,
+
= V in C. (8.4.56)
div ,
+
= 0 in C. (8.4.57)
j[;[ (;
,
+
n
)(;) = 0 on I
2
. (8.4.58)
We note that the last relation models boundary conditions of the friction type. Indeed,
if [;
,
+
n
[ < j, then (8.4.58) means that ; = 0. If ;
,
+
n
= j, then from (8.4.58) it
follows that ; < 0. If ;
,
+
n
= j then (8.4.58) implies ; > 0.
Winkler type boundary condition. Set
() =
1
2
k [[
2
. (8.4.59)
where k and is a positive constant. This relation can be viewed as a simplied variant of
the Winklers boundary condition widely used in solid mechanics. In this condition, on
I
2
a body is connected with an elastic foundation, which provides a certain response
to the boundary deections (such a condition can be modeled by a large amount of
springs connected with I
2
). In view of (8.4.59), we have
;
+
n
= k(;u) on I
2
(8.4.60)
and
+
(
+
) = sup
R
o
_
+

1
2
k [[
2
_
=
1
2k
2
. (8.4.61)
Consider the quantity
1
I
2
=
1
2
_
I
2
(k [;[
2
1
k
2
2
+
; 0
,
+
n

+
2
) Js.
The minimization of this quantity over
+
leads to the condition
_
1
k
0
_
+
= ; 0;
,
+
n
=
+
=
k(; 0;
,
+
n
)
1 k0
.
This gives a simple expression for 1
I
2
:
1
I
2
=
1
2
_
I
2
0
1 k0
(k(;) ;
,
+
n
)
2
Js.
By (8.4.36), we obtain the estimate
[[V(u )[[
2
_ (1 )2M
1
(. ,
+
)
_
I
2
0
1 k0
(k(;) ;
,
+
n
)
2
Js
_
1
1
_
(1 )C
2
1
r
:
2
(,
+
). (8.4.62)
If k = 0 (i.e., if we consider homogeneous Neumann boundary condition), then the
second term on the right-hand side of (8.4.62) has the form
_
1
1
_
_
1
1
_
C
2
TI
2
_
I
2
(;
,
+
n
)
2
Js
and (8.4.62) gives the same estimate as (8.4.50) (with J = 0).
8.4.3 Generalized Newtonian uids
Now, we briey discuss error estimates for some nonlinear models of viscous uids.
In these models, the basic relations are as follows:
u
t
Div o Div (u u) = Vp in C. (8.4.63)
o d(c(u)). (8.4.64)
where div u = 0, c(u) is the symmetric part of Vu, and is the so-called dissipative
potential.
Many physically motivated dissipative potentials have the form
d(c) = H
t
([c[) c. (8.4.65)
where H(0) = 0 and the prime denotes the derivative of H with respect to the argu-
ment. In particular, the classical Newtonian uid relates to the case
H() =
1
2
v
2
. (c) =
1
2
v [c[
2
. (8.4.66)
Another well-known example is the Bingham uid, where
(c) =
1
2
v [c[
2
k
+
[c[. k
+
> 0. (8.4.67)
For slow stationary ows, the classical statement of this type problem is as follows:
Find a vector-valued function u (velocity), a scalar-valued function (pressure) and a
tensor-valued function o (stress deviator) such that
Div o = V in C. (8.4.68)
div u = 0 in C. (8.4.69)
o d(c(u)) in C. (8.4.70)
u = u
0
on I. (8.4.71)
where and u
0
are given functions satisfying the same conditions as in Chapter 6,
and is the dissipative potential that denes physical properties of a uid. Consider
the class of potentials
(c) =
v
2
[c[
2
[(c).
where [ : M
dd
R
is a convex nonnegative function such that

[(0) = 0. [(c) _ c
1
[c[
2
c
2
. c
1
> 0. (8.4.72)
Obviously, the cases [ 0 and [(c) = k
+
[c[ correspond to Newtonian and Bingham
models, respectively. Investigation of mathematical properties of such type models
was started in the 60s (e.g., see O. A. Ladyzhenskaya [212]). Variational methods for
this and other classes of nonlinear problems are studied in the book by M. Fuchs and
G. Seregin [140], where the reader will nd a consequent exposition of the regularity
theory and many references related to the subject.
For dissipative potentials of such a type, energy estimates of the distance between
u
0
S
1
(C) and the exact solution u were derived in the authors papers [289, 292,
293], in M. Fuchs and S. Repin [139], and in M. Bildhauer, M. Fuchs, and S. Repin
[57]. In those papers, the derivation of a posteriori estimates is based on the variational
method.
In general, estimates of such a type have the following form:
v
2
[c( u)[
2
_ (1 )D
1
(c(). t
1
) D
2
(c(). t
2
)
_
1
1
_
1
2
C
2
T D
[div (t
1
t
2
) Vq[
2
. (8.4.73)
where
D
1
(c(). t
1
) :=
_
D
_
2
[c()[
2
1
2
[t
1
[
2
c() : t
1
_
J. =
1
2
[vc() t
1
[
2
.
D
2
(c(). t
2
) :=
_
D
_
[(c()) [
+
(t
2
) c() : t
2
_
J..
[
+
is the functional conjugate to [, t
1
t
2
H(C. Div) holds, q is a function in
H
1
(C)
1
2
(C), > 0, and
C
T D
is a constant in the inequality [n[ _

C
T D
[c(n)[.
Assume that the right-hand side of (8.4.73) is equal to zero. Then
div (t
1
t
2
) = Vq
and, in addition,
t
1
= vc() a.e. in C.
t
2
d[(c()) a.e. in C.
Since u
0

S
1
(C), we conclude that in such a case, coincides with the exact
solution u, t
1
= o
1
, and t
2
= o
2
.
Also, computable error bounds can be derived directly from the respective integral
identity, which has the form
_
D
vc(u) : c(n) J.
_
D
[
t
([c(u)[)
c(u) : c(n)
[c(u)[
J.
=
_
D
nJ.. Vn u
0

S
1
(C). (8.4.74)
Let u
0

S
1
(C) be an approximation of u. We transform (8.4.74) as follows:
_
D
vc(u ) : c(n) J.
_
D
_
[
t
([c(u)[)
c(u)
[c(u)[
[
t
([c()[)
c()
[c()[
_
: c(n) J.
=
_
D
( n vc() : c(n)) J.
_
D
[
t
([c()[)
c() : c(n)
[c()[
J.. (8.4.75)
We reform the right-hand side of (8.4.75), introducing two symmetric tensor-valued
functions t
1
and t
2
in H(C. Div) and using the identities
_
D
(n Div t
i
c(n) : t
i
) J. = 0. Vn V
0
. i = 1. 2.
We obtain
_
D
vc(u ) : c(n) J.
_
D
_
[
t
([c(u)[)
c(u)
[c(u)[
[
t
([c()[)
c()
[c()[
_
: c(n) J.
=
_
D
( Div (t
1
t
2
)) nJ.
_
D
(t
1
qI vc()) : c(n) J.
_
D
_
t
2
[
t
([c()[)
c()
[c()[
_
: c(n) J.. (8.4.76)
where q is a square summable function. Set n = u and note that
_
D
( Div (t
1
t
2
)) (u ) J. _

C
T D
[ Div (t
1
t
2
)[[c(u )[.
With the help of Youngs inequality, we deduce the estimate
(1
1

2

3
)
_
D
v[c(u )[
2
J.
_
D
_
[
t
([c(u)[)
c(u)
[c(u)[
[
t
([c()[)
c()
[c()[
_
: c(n) J.
=
1
4
1
C
2
T D
[ Div (t
1
t
2
)[
2
1
4
2
[t
1
qI vc()[
2
1
4
3
_
_
_
_
t
2
[
t
([c()[)
c()
[c()[
_
_
_
_
. (8.4.77)
where
1
,
2
, and
3
are positive numbers such that
1
3
< 1.
Estimate (8.4.77) has the same principal structure as (8.4.73): its right-hand side
consists of penalties for the violation of the constitutive law relations and the equilib-
rium equation formed by the total stress t
1
t
2
. If the right-hand side of (8.4.77)
vanishes, then
= Div (t
1
t
2
).
t
1
= qI vc().
t
2
= [
t
([c()[)
c()
[c()[
.
Since is a solenoidal eld satisfying the prescribed boundary condition, we conclude
that coincides with u.
Remark 8.14. The second term on the left-hand side of (8.4.77) is nonnegative and,
therefore, can be removed. However, for certain [ it may be evaluated in terms of
c(u ), which would make the overall estimate stronger.
1. Estimates discussed in this chapter are valid for a wide class of nonlinear varia-
tional problems in continuum mechanics associated with the functional
J() =
_
D
g(c()) J.
_
D
J.
_
I
2
J Js. (8.5.1)
where g (internal energy function) has the form
g(c) = Lc : c (c). (8.5.2)
In (8.5.1), L satises (5.1.5) and is a nonnegative convex function. A minimizer
of the functional (8.5.1) satises the relations (5.1.2)(5.1.4), and (5.1.1) is replaced
by a nonlinear constitutive relation. If is a differentiable functional, then the latter
relation has the form
o = Lc(u)
t
(c()). (8.5.3)
One example of such relations is offered by deformation plasticity theory (e.g., see
A. Iljushin [183] or R. Temam [349]), which is based on the constitutive relation
o = 1
0
tr(c) I ;(
c
T
) c
T
. (8.5.4)
where
;(t ) =
_
2j if t _ t
0
= k
+
,
_
2j.
(2j )t
0
t
-1
if t > t
0
.
1
0
and j are positive (elasticity) constants, k
+
> 0 is a plasticity module, and > 0
is a hardening module. In this case, the integrand g has the form
g(c) =
1
2
a(c. c) j
1
(
c
T
).
where j
1
= 1 ,2j,
a(c
1
. c
2
) = 1
0
tr(c
1
) tr(c
2
) c
T
1
: c
T
2
. Vc
1
. c
2
M
dd
.
and
(t ) =
_
_
_
j t
2
if [ t [_ t
0
= k
+
,
_
2j.
k
+
_
_
2 t k
+
,2j
_
if [ t [> t
0
.
Functional type a posteriori error estimates for variational problems with functionals
of the type (8.5.1)(8.5.2) were derived in S. Repin and L. Xanthis [320, 321].
A posteriori estimates for lower semicontinuous relaxations of some nonconvex
variational problems (which are related to simple phase transitions models in the the-
ory of solids) were obtained in the authors paper [281]). Estimates for the Ramberg
Osgood model and for elasto-plastic torsion problem have been recently derived in
M. Bildhauer, M. Fuchs, and S. Repin [58] and [59], respectively. Estimates for an
incremental evolutionary plasticity model were obtained in S. Repin and J. Valdman
[318].
2. As we have seen, functional a posteriori estimates derived for different problems
have certain common features. We summarize this experience and put forward the
following conjecture:
A majorant (guaranteed upper bound) of the deviation u, where is an arbi-
trary function from the energy space (i.e., from the functional class that contains
the generalized solution u) consists of terms, which can be thought of as penal-
ties for unconformity in all basic relations. Relevant multipliers are dened by
constants in the embedding inequalities for spaces pertaining to the mathematical
statement of the problem.
Certainly, this conjecture is yet to be justied for many classes of boundary value
problems (e.g., for those that are related to differential equations of a nondivergent
type).
3. All the majorants that we discussed satisfy the following conditions:
(a) [[(u )[[ _ M(. ,
+
. D) for any ,
+
Y
+
and u
0
V
0
.
(b) inf
,
Y
M(u. ,
+
. D) = 0.
(c) M(
k
. ,
+
k
. D) M(. ,
+
. D).
(8.5.5)
where {
k
] and {,
+
k
] are arbitrary sequences in V and Y
+
, respectively, such that
k
in V and ,
+
k
,
+
in Y
+
. We note that the above-stated requirements
are quite natural. Indeed, (a) means that the upper bound is guaranteed, (b) says that
if coincides with the exact solution u, then Y
+
contains a counterpart function ,
+
(which is
+
) such that the majorant vanishes, and (c) is the continuity property. Ob-
viously, practically valuable error majorants must satisfy (a)(c). Quite similar condi-
tions should be imposed on error minorants.
However, the conditions (a)(c) do not guarantee that the majorant M(. ,
+
. D) is
equal to [[(u )[[ for some ,
+
and other parameters involved in the majorant. In
other words, we cannot guarantee that there is no irremovable gap between the left-
hand and right-hand sides of (8.5.5 a).
Nevertheless, we can show that if M satises (a)(c), then another majorant pos-
sesses such a property. For this purpose, we introduce an arbitrary function n V
0
and dene a new majorant
M(. n. ,
+
. D) := [[n[[ M( n. ,
+
. D).
M(. n. ,
+
. D) also contains only known functions and, therefore, is explicitly com-
putable. This new majorant satises the properties (a)(c). Indeed, for any element
u
0
V
0
,
[[(u )[[ _ [[n[[ [[(u n)[[ _ M(. n. ,
+
. D).
Thus, M(. n. ,
+
. D) yields a guaranteed upper bound of the error in terms of the
energy norm. It is easy to see that M(. 0. ,
+
. D) = M(. ,
+
. D). so that
inf
,
uV
0
M(u. n. ,
+
. D) = 0.
Moreover, the functional M(. n. ,
+
. D) is continuous (i.e., the condition (c) holds).
However, M(. n. ,
+
. D) possesses one more property. Since
inf
,
uV
0
M(. n. ,
+
. D) _ inf
,
M(. u . ,
+
. D)
= [[(u )[[ inf
,
M(u. ,
+
. D)
= [[(u )[[. (8.5.6)
we nd that, in principle, the majorant M(. n. ,
+
. D) can give an upper bound of
the error with any desired accuracy.
Similar arguments lead to sharp lower bounds of the error: for any n V
0
we have
[[n[[ _ [[(u )[[ [[(n u )[[ _ [[(u )[[ M( n. ,
+
. D).
Dene the minorant
M(. n. ,
+
. D) := [[n[[ M( n. ,
+
. D).
We have
sup
,
uV
0
M(. n. ,
+
. D) _ sup
,
M(. u . ,
+
. D)
= [[(u )[[ inf
,
M(u. ,
+
. D)
= [[(u )[[
and, consequently, the minorant also has no gap.
Using continuity property (c), it is not difcult to prove that the above-dened two-
sided bounds of the error (which converge to the exact error) can be constructed by
solving only nite-dimensional problems. Thus, we arrive at the following conclu-
sion: if for a boundary value problem a majorant with properties (a)(c) has been
constructed, then (in principle) errors in the energy norm can be evaluated with any
desirable accuracy. Certainly, this theoretical conclusion may have different value for
different problems. It is very probable that for strongly nonlinear problem the practical
computation of sharp estimates (in terms of the global error norm) may lead to high
computational costs.
9 A posteriori estimates for other problems
9.1 Differential equations of higher order
Fourth order elliptic equation. A posteriori error estimation methods discussed in
previous chapters can be applied to boundary value problems associated with higher
order differential equations. As an example, we consider the 4th order elliptic problem
div Div (TVVu) = in C. (9.1.1)
u =
du
dn
= 0 on I. (9.1.2)
Here
1
2
(C). T = {b
i}kI
]. b
i}kI
= b
}ikI
= b
kIi}
1
o
(C).
where the indices change from 1 to J, and
c
2
1
[j[
2
_ Tj : j _ c
2
2
[j[
2
. Vj M
dd
x
. (9.1.3)
Then, the inverse tensor T
-1
exists and for any tensor-valued function t with square
summable components we dene the norms
[[t[[
2
:=
_
D
Tt : t J. and [[t[[
2
+
:=
_
D
T
-1
t : t J..
A posteriori estimates for this problem can be obtained from the general estimates
considered in Chapter 7 if we dene as the Hessian operator and set
U = 1
2
(C. M
dd
x
). V = H
2
(C).
V
0
=
_
n V [ n =
dn
dn
= 0 on I
_
.
However, to make the exposition more transparent and to obtain a posteriori esti-
mates of a different form, we derive them below directly from the integral identity
_
D
TVVu : VVnJ. =
_
D
nJ.. Vn V
0
. (9.1.4)
that denes a generalized solution u V
0
.
Let V
0
be an approximation of u. By the identity
_
D
(t : VVn ndiv Div t) J. = 0.
Section 9.1 Differential equations of higher order 219
where n V
0
and t is an arbitrary tensor-valued function from the space
H(div Div. C) :=
j U [ div Div j 1
2
(C)
_
.
we transform (9.1.4) as follows:
_
D
TVV(u ) : VVnJ.
=
_
D
( div Div t)nJ.
_
D
(t TVV) : VVnJ.. (9.1.5)
Let C
1D
denote the constant in the inequality
[n[
D
_ C
1D
[[VVn[[. Vn V
0
. (9.1.6)
Set n = u . From (9.1.5) and (9.1.6), it follows that
[[VV( u)[[
2
_ (1 )[[TVV t[[
2
+
_
1
1
_
C
2
1D
[div Div t [
2
D
. (9.1.7)
where is a positive real number. This estimate is quite analogous to (3.2.8) and can be
obtained as a particular case of (7.1.20). It corresponds to the following decomposition
of (9.1.1):
div Div o = .
o = TVVu.
However, the condition
div Div t 1
2
(C)
is rather demanding (for example, if t is constructed with the help of piecewise afne
continuous approximations, then it does not satisfy this condition). To avoid arising
technical difculties, we introduce a new vector-valued function , H(C. div) and
put (9.1.5) in the form
_
D
TVV(u ) : VVnJ.
=
_
D
( div Div t)nJ.
_
D
(t TVV) : VVnJ.
_
D
(ndiv , , Vn) J.
=
_
D
( div ,)nJ.
_
D
(Div t ,) VnJ.
_
D
(t TVV) : VVnJ.. (9.1.8)
220 Chapter 9 Other problems
Let C
2D
denote the constant in the inequality
[Vn[
D
_ C
2D
[[TVVn[[. Vn V
0
. (9.1.9)
Set n = u and estimate the right-hand side of (9.1.8), using (9.1.6) and (9.1.9).
We obtain
[[VV( u)[[ _ [[TVV t[[
+
C
1D
[div , [
D
C
2D
[Div t ,[
D
. (9.1.10)
Square both parts of the above estimate and apply Youngs inequality. Then (9.1.10)
implies another estimate:
[[VV( u)[[
2
_ (1 )[[TVV t[[
2
+
_
C
1D
[div , [
D
C
2D
[Div t ,[
D
_
2
. (9.1.11)
Estimates (9.1.10) and (9.1.11) have two free functions , H(div . C) and
t H(Div . C), which can be viewed as images of the gradient and double gradi-
ent, respectively. They reect the representation of (9.1.1) in the form
div , = .
Div o = ,.
o = TVVu.
Remark 9.1. Note that C
1D
and C
2D
are estimated by c
-1
1
C
1D
and c
-1
1
C
2D
, re-
spectively, where C
1D
and C
2D
are constants in the inequalities

[n[
D
_ C
1D
[VVn[
D
. Vn V
0
. (9.1.12)
[Vn[
D
_ C
2D
[VVn[
D
(9.1.13)
and C
is a rectangular domain containing C.

Remark 9.2. A posteriori estimates for the equation (9.1.1) with other boundary con-
ditions can be derived by the same arguments as for the second order problems we
discussed in Chapters 4 and 5. Instead of C
1D
and C
2D
they involve constants in the
inequalities analogous to (9.1.6) and (9.1.9) in which V
0
contains functions vanish-
ing on the Dirichlet part of the boundary. Also, such estimates involve constants in
the trace inequalities on other parts of the boundary (e.g., on the part related to the
Neumann boundary condition). We leave this task to the reader as an exercise.
Example. Consider the application of the above a posteriori error estimates to a plate
bending problem. In this case, C R
2
is associated with the middle surface of a
plate. Let the deformation of a plate be described by the KirchhoffLove model. Then
u = u(.
1
. .
2
) is the bending function, k(u) = VVu is the curvature tensor, which is
connected with the bending moments t by a linear constitutive law
t = Tk. (9.1.14)
If the plate is made of an isotropic elastic material, then (9.1.14) has a simple form
t
11
= H(k
11
vk
22
).
t
22
= H(k
22
vk
11
).
t
12
= H(1 v)k
12
.
where H =
Th
3
12(1-
2
)
, 1 > 0, and v (0. 1). Here 1 an v are elasticity constants and
h = h(.
1
. .
2
) is the thickness parameter. We see that T has the following nonzero
components:
b
1111
= H. b
2222
= H. b
1122
= vH.
b
2211
= vH. b
1212
= b
2121
= (1 v)H.
The natural condition
h
1
_ h(.
1
. .
2
) _ h
2
. (.
1
. .
2
) C. (9.1.15)
guarantees that the estimate (9.1.3) holds with
c
2
1
=
1h
3
1
12(1 v)
and c
2
2
=
1h
3
2
12(1 v)
.
The nonzero components the tensor C = T
-1
are as follows:
c
1111
=
1
H(1 v
2
)
. c
2222
=
1
H(1 v
2
)
. c
1122
=
v
H(1 v
2
)
.
c
2211
=
v
H(1 v
2
)
. c
1212
= c
2121
=
1
H(1 v)
.
Hence,
[[Tk t[[
2
+
=
_
D
(Tk : k Ct : t 2k : t) J..
and the other parts of the error majorants (9.1.10)(9.1.11) are directly computable.
Biharmonic equation. A somewhat different a posteriori estimate can be derived
for the biharmonic problem
^û = in C
with the boundary conditions (9.1.2). In this case, the integral identity has the form
_
D
û^nJ. =
_
D
nJ.. Vn V
0
.
Introduce a function H
2
(C). We have
_
D
^(u )^nJ. =
_
D
( n ^^n) J.
=
_
D
(( ^)n ( ^)^nJ..
From here, we nd that
[^(u )[ _ C
3D
[ ^[ [ ^[. (9.1.16)
where C
3D
is the constant in the inequality
[n[ _ C
3D
[^n[. Vn V
0
.
Variational inequalities. Consider the problem (9.1.1)(9.1.2) with the condition
u 1 := {n H
2
(C) [ n _ (.) a.e. in C].
which arises if the solution must lie above the obstacle (.). For the sake of simplic-
ity, we consider the case with one obstacle and assume that is sufciently regular
(e.g., continuous and piecewise smooth). Then, the solution satises the variational
inequality
_
D
TVVu : VV(n u) J. _
_
D
(n u) J.. Vn 1. (9.1.17)
Assume that 1 is a function, which is an approximation of u. Then,
[[u [[
2
_
_
D
((u ) TVV : VV(u )) J.. (9.1.18)
As in the linear case, we introduce two auxiliary functions , H(C. div) and
t H(C. Div). We have
[[u [[
2
_
_
D
( div ,)(u ) J.
_
D
(Div t ,) V(u ) J.
_
D
(t TVV) : VV(u ) J.
=
_
D
:
( div ,)(u ) J.
_
D
:
0
( div ,)(u ) J.
_
D
(Div t ,) V(u ) J.
_
D
(t TVV) : VVnJ.. (9.1.19)
where
C
:= {. C [ (.) = (.)] and C
0
:= {. C [ > (.)].
Since
_
D
:
( div ,)(u ) J. _
_
D
:
( div ,)
C
(u ) J..
we estimate the right-hand side of (9.1.19) in the same way as (9.1.8) and deduce the
estimate
[[VV( u)[[ _ [[TVV t[[
+
C
1D
[r(. ,)[
D
C
2D
[Div t ,[
D
. (9.1.20)
where
r(. ,) :=
_
( div ,)
C
in C
.
div , in C
0
.
Remark 9.3. If u is subject to two obstacles (i.e., u _ [ as in Section 8.1), then the
estimate is obtained quite analogously. In this case, we append the third branch and
set r(. ,) = ( div ,)
on C
)
.
Comments. Estimates (9.1.10) and (9.1.11) were derived in P. Neittaanma aki and
S. Repin [243] by variational techniqus. In [300], it was shown that the estimates fol-
low from the corresponding integral identity. Numerical testing of these estimates was
performed in the PhD thesis of M. Frolov [134], in which the estimate (9.1.16) was de-
rived (see also [135]). In the context of the above-discussed a posteriori estimates, the
classical KirchhoffLove plate model was considered in P. Neittaanma aki and S. Re-
pin [243] and in the authors paper [285]. Estimates of modeling errors arising if the
KirchhoffLove model is used instead of the 3D elasticity model has recently been de-
rived in S. Repin and S. Sauter [306]. Estimates for the ReissnerMindlin model were
obtained by the variational and nonvariational methods in the papers by M. Frolov,
P. Neittaanm aki, and S. Repin [138, 304]. A posteriori estimates for the fourth order
elliptic equations with obstacles were obtained in M. Bildhauer, M. Fuchs, and S. Re-
pin [60] with the help of the variational method. In M. Bildhauer and M. Fuchs [56]
these results were extended to a wider class of nonlinear functionals.
9.2 Equations with the operator curl
Basic problem. The simplest version of the Maxwells problem is given by the equa-
tion
curl j
-1
curl u k
2
u = in C. (9.2.1)
where C is a bounded domain in R
d
, is a given current density, and j is the perme-
ability of a medium (may be a positive constant or a positive bounded function). The
case k = 0 corresponds to magnetostatics. Equation (9.2.1) with positive k arises in
semidiscrete approximations of the evolutionary Maxwells problem.
On I the condition
n u = 0 (9.2.2)
is stated. In this section, V(C) is the space H(C. curl),
V
0
(C) := {n H(C. curl) [ [ n n = 0 on I]
and
(n) :=
_
D
nJ..
First, we consider the case k > 0. Multiply (9.2.1) by a smooth vector-valued function
n that satises (9.2.2) and integrate over C. We have
_
D
(curl j
-1
curl u n k
2
u n) J. = (n).
By the relation
_
D
(curl ) nJ. =
_
D
(curl n) J.
_
I
( n) nJs. (9.2.3)
we express the integral identity as follows
_
D
(j
-1
curl u curl n k
2
u n) J.
_
I
(j
-1
curl u n) nJs = (n).
Section 9.2 Equations with the operator curl 225
Note that
n (curl u n) = curl u (n n).
Therefore, if n V
0
, then the boundary term vanishes. Since smooth functions are
dense in V , we conclude that the generalized solution can be dened as a vector-valued
function u V
0
such that integral identity
_
D
_
j
-1
curl u curl n k
2
u n
_
J. = (n) (9.2.4)
holds for n V
0
.
If k = 0, then the additional compatibility condition
div = 0 in C. n = 0 on I (9.2.5)
is necessary to have a well-posed problem. In this case, a respective generalized solu-
tion is dened by the integral identity
_
D
j
-1
curl u curl nJ. = (n). Vn V
0
. (9.2.6)
Since curl V[ = 0 (see Section 1.4), a solution should be understood as an element of
the factor space, in which the functions are equivalent if their difference is a gradient
eld. To ensure the uniqueness of a solution, the Coulomb gauge condition
u V
00
:=
_
V
0
[
_
D
V J. = 0. V

H
1
(C)
_
. (9.2.7)
is usually attracted.
Let V
0
(C) be an approximation of u. Rewrite (9.2.4) in the form
_
D
_
j
-1
curl (u ) (curl n) k
2
(u ) n
_
J.
=
_
D
_
n j
-1
(curl ) (curl n) k
2
n
_
J.. Vn V
0
(C). (9.2.8)
By (9.2.3), we obtain
_
D
_
j
-1
curl (u ) (curl n) k
2
(u ) n
_
J.
=
_
D
_
( k
2
curl ,) n (, j
-1
curl ) curl n
_
J.. (9.2.9)
where , V(C). Introduce the norm
[[n[[
2
:=
_
D
_
j
-1
[ curl n[
2
k
2
[n[
2
_
J..
Setting n = u in (9.2.9), we obtain the relation
[[u [[
2
=
_
D
( k
2
curl ,) (u ) J.
_
D
(, j
-1
curl ) curl (u ) J..
which leads to the estimate
[[u [[
2
_
_
_
_
_
1
k
( k
2
curl ,)
_
_
_
_
[k(u )[
[j
1{2
(, j
-1
curl )[[j
-1{2
curl (u )[.
Hence, we nd that
[[u [[
2
_ M
2
MAX
(. ,). (9.2.10)
where
M
2
MAX
(. ,) :=
_
_
_
1
k
( k
2
curl ,)
_
_
_
2
[j
1{2
(, j
-1
curl )[
2
is the majorant for this type of Maxwells problem. It is easy to see that
inf
V
0
,
,1(D,curl)
M
MAX
(. ,) = 0.
and the exact lower bound is attained if and only if
curl , k
2
= a.e. in C. (9.2.11)
, = j
-1
curl a.e. in C. (9.2.12)
Since n = 0 on I, (9.2.11) and (9.2.12) mean that coincides with the exact
solution u and , coincides with j
-1
curl u.
For any , V
0
, the quantity M
2
MAX
(. ,) gives an upper bound of the error. It is
easy to observe that
inf
,V
0
M
2
MAX
(. ,) _ M
2
MAX
(. j
-1
curl u)
=
_
_
_
1
k
( k
2
curl j
-1
curl u)
_
_
_
2
[j
-1{2
curl (u )[
2
= [k(u )[
2
[j
-1{2
curl (u )[
2
= [[u [[
2
.
Therefore, the estimate (9.2.10) has no gap between the left- and right-hand sides. A
practically computable estimate can be determined if M
2
MAX
(. ,) is minimized over
a nite-dimensional subspace V
n
V(C).
Section 9.2 Equations with the operator curl 227
A lower bound of [[u [[ follows from the relation
sup
uV
0
_
D
_
j
-1
curl (u ) curl n k
2
n (u )
1
2
(j
-1
curl n curl n k
2
n n
_
J.
_ sup
r1
2
(D,R
o
)
)1
2
(D,R
o
)
_
D
_
j
-1
curl (u ) t
1
2
j
-1
t t k
2
j (u )
1
2
k
2
j j
_
J.
=
1
2
[[u [[
2
=
_
D
_
j
-1
curl (u ) curl (u ) k
2
(u ) (u )
1
2
(j
-1
[curl (u )[
2
k
2
[u [
2
_
J.
_ sup
uV
0
_
D
_
j
-1
curl (u ) curl n k
2
n (u )
1
2
(j
-1
curl n curl n k
2
n n
_
J..
1
2
[[u [[
2
= sup
uV
0
_
D
_
j
-1
curl (u ) curl n k
2
n (u )
1
2
(j
-1
curl n curl n k
2
n n
_
J..
[[u [[
2
_ M
2
(. n)
:=
_
D
_
2 n j
-1
[curl n[
2
k
2
[n[
2
2j
-1
curl curl n 2k
2
n
_
J.. (9.2.13)
For any n V
0
the quantity M
2
(. n) provides a lower bound of the error. Certainly,

the sharpest bound is given by
sup
uV
0
M
2
(. n).
It is not difcult to prove that this quantity coincides with the squared error (to prove
that it sufces to set n = u ). A practically computable lower bound can be de-
termined if maximization is performed over a nite-dimensional subspace V
0n
V
0
,
dimV
0n
= m. Then, nding the quantity
sup
uV
0m
M
2
(. n)
requires solving a quadratic type maximization problem.
Now, we consider the problem (9.2.6). Since u is dened up to a gradient eld, we
consider only the divergent part of , which leads to the assumption V
00
. Our aim
is to estimate the quantity
[u |[ := [curl (u )[.
First, we recall a result in the theory of functions in H(C. curl) (e.g., see [152, 236,
327]).
Lemma 9.4. Let C be a Lipschitz simply connected bounded domain. There exists a
constant C
D
such that
[n[ _ C
D
[curl n[. Vn V
0
. (9.2.14)
provided that div n = 0 in C.
By (9.2.6) we observe that
_
D
curl (u ) curl nJ. =
_
D
( n curl curl n) J.. Vn V
0
. (9.2.15)
Since
_
D
(curl , n , curl n) J. = 0. Vn V
0
.
we rearrange (9.2.15) as follows:
_
D
curl (u ) curl nJ. =
_
D
_
( curl ,) n (curl ,) curl n
_
J..
Set n = u V
0
. We obtain
[u |[
2
_ [ curl ,[ [u [ [curl ,[ [curl (u )[.
In view of (9.2.14), this estimate leads to the following upper bound of the error:
[u |[ _ C
D
[ curl ,[ [curl ,[. (9.2.16)
Note that in (9.2.16) occurs only as the argument of the operator curl. Therefore,
it should be understood as an estimate of the factor-norm for functions in the factor
space equivalent up to a gradient eld.
Section 9.3 Evolutionary problems 229
Comments. Approximation methods for the Maxwells equation were investigated
by many authors (e.g., see P. Monk [236] and R. Hiptmair [174]). Parallel multigrid
solvers were studied in G. Haase, M. Kuhn, and U. Langer [164]). A posteriori esti-
mates were derived in R. Beck, R. Hiptmair, R. Hoppe, and B. Wohlmuth [45] in the
context of the residual approach and in D. Braess and J. Sch oberl [69] with the help of
the equilibrated approach. A posteriori estimates for nonconforming approximations
of H(curl) elliptic partial differential equations were studied in P. Houston, I. Perugia,
and D. Schotzau [181].
Estimates (9.2.10) and (9.2.16) were obtained in the authors papers [296, 297] (es-
timate (9.2.10) was independently derived by A. Hannukainen [168]). We note that
(9.2.10) can also be derived from the general a posteriori estimate given in [280] (see
also [244]) for convex variational problems related to the functional G() J()
(see Chapter 7). The problem is encompassed in the general framework if we set
= curl,
G(,) =
_
D
1
2j
[,[
2
J.. and J() =
_
D
_
k
2
2
[[
2

_
J..
Advanced forms of a posteriori estimates for the Maxwells problem has been recently
obtained in P. Neittaanm aki and S. Repin [245].
9.3 Evolutionary problems
A posteriori error control for evolutionary problems is a substantial topic that requires
a special consideration. The goal of this section is to give only an idea of how to extend
methods considered in previous chapters to evolutionary problems. Below we shortly
discuss a posteriori estimates that follow from the evolutionary integral identity. As
before, we do this with the paradigm of the linear diffusion problem.
9.3.1 The linear evolutionary problem
Consider the classical initial-boundary value problem for the heat equation: Find
u(.. t ) such that
u
t
û = in Q
T
. (9.3.1)
u(.. 0) = . . C. (9.3.2)
u(.. t ) = 0. (.. t ) S
T
. (9.3.3)
where Q
T
:= C(0. T ) is a space-time cylinder and S
T
:= I 0. T |. This problem
is one of the simplest evolutionary problems, which is often used to model various
diffusion type processes (e.g., heat transfer). With the paradigm of this problem, we
explain how to derive computable estimates (in terms of a space-time norm) of the
difference between u and any admissible approximation . Our analysis follow the
lines of [287]. In A. Gaevskaya and S. Repin [144], the method was applied to a more
general class of linear parabolic problems (this paper also contains results of numerical
tests).
We begin by introducing some spaces of functions dened on Q
T
. Spaces of func-
tions that map (0. T ) into a Banach space X are called Bochner spaces. For example, if
1. o) and [ [
A
denotes the norm of X, then the Bochner space 1
]
((0. T ). X)
is the Banach space of mappings g such that
[g[
1
)
((0,T),A)
:=
_
_
T
0
[g(. t )[
]
A
Jt
_
1{]
< o.
In particular, 1
2
( (0. T ): H
1
(C)) consists of H
1
-functions (with respect to spatial
variables) the norms of which are 1
2
-functions with respect to t (0. T )). Let
H
1
0
(Q
T
) be the subspace of H
1
(Q
T
) that contain functions with zero traces on S
T
.
By V(Q
T
), we dene the Banach space of functions from1
2
( (0. T ): H
1
(C)) having
nite norm
[n[
2
V
:= vrai max
t(0,T)
[n(. t )[
2
D
[Vn[
2
2,Q
T
.
The space V
1,0
(Q
T
) = C(0. T |: 1
2
(C)) 1
2
( (0. T ): H
1
(C)) is a subspace of
V(Q
T
). For all t 0. T |, elements of this space have traces (which are square
summable functions on cross-sections of Q
T
) that continuously change with respect
to t 0. T |. By

V
1,0
(Q
T
), we denote another subspace of V(Q
T
), which is the in-
tersection of V
1,0
(Q
T
) and 1
2
((0. T ):

H
1
(C)). For elements of the space V
1,0
(Q
T
),
we dene the quantity
[[n[[
2
(k,e)
:= k[n(. T )[
2
D
c[Vn[
2
Q
T
. j. c > 0.
which will be used as a measure of the difference between the exact solutions and
approximations.
Also, we use the space H
1,Z
0
(Q
T
) that consists of functions n 1
2
(Q
T
) having
nite norm
[n[
1,Z
2,0
:=
_
Q
T
(n
2
n
2
t
[ Vn [
2
(^n)
2
) J. Jt
and vanishing on S
T
.
In the context of the well-known theory for parabolic type problems, a function
u

V
1,0
(Q
T
) is called a (generalized) solution of (9.3.1)(9.3.3) if it satises the
following integral identity:
_
Q
T
Vu VnJ. Jt
_
Q
T
un
t
J. Jt
_
D
(u(.. T )n(.. T ) u(.. 0)n(.. 0)) J.
=
_
Q
T
nJ. Jt. Vn H
1
0
(Q
T
). (9.3.4)
We recall the classical solvability results (e.g., see Ladyzhenskaya [217]) for this prob-
lem.
Theorem9.5. Let Cbe a bounded connected domain with Lipschitz continuous bound-
ary I.
(i) Let 1
2
(Q
T
) and (.)

H
1
(C). Then the problem (9.3.1)(9.3.3) is uniquely
solvable in the space H
1,Z
0
(Q
T
).
(ii) If 1
1
((0. T ). 1
2
(C)) and 1
2
(C), then u belongs to the class

V
1,0
(Q
T
).
9.3.2 First form of the error majorant
Assume that
1
2
(Q
T
). H
1
(C). (9.3.5)
In this case, Theorem 9.5 guarantees the existence of a solution u that satises (9.3.4).
Let H
1
0
(Q
T
) be a given function. In particular, may be an approximation of
u obtained by a semidiscrete approximation of (9.3.1)(9.3.3). We are interested in
deriving an upper bound of the deviation u evaluated in the norm [[[ [[[ or in terms
of the quantity |
(;,)
. From (9.3.4) we obtain
_
Q
T
V(u ) VnJ. Jt
_
Q
T
(u )n
t
J. Jt
_
D
((u(.. T ) (.. T ))n(.. T ) (u(.. 0) (.. 0))n(.. 0)) J.
=
_
Q
T
( n V Vn
t
n) J. Jt. Vn H
1
0
(Q
T
).
Set n = u and note that
_
D
[ n(.. T ) [
2
J.
_
D
[ n(.. 0) [
2
J.
=
_
Q
T
nn
t
J. Jt
1
2
[n(. T )[
2
D
1
2
[n(. 0)[
2
D
.
This yields the integral relation
[V(u )[
2
2,Q
T
1
2
[u(. T ) (. T )[
2
D
=
_
Q
T
((u ) V V(u )
t
(u )) J.Jt
1
2
[u(. 0) (. 0)[
2
D
. (9.3.6)
which presents the energy balance in terms of deviations fromthe exact solution u. The
relation (9.3.6) can be regarded as a generalization of the well-known energy-balance
equation for the heat equation (see, e.g. [217]). It is easy to see that the classical
energy balance equation follows from (9.3.6) if we set 0. Subsequently, we
use (9.3.6) as a starting point of the analysis. Introduce a new vector-valued function
,(.. t ) Y(Q
T
), where
Y(Q
T
) :=
,(.. t ) = {,
i
(.. t )] [ ,
i
1
2
(Q
T
). 1 _ i _ J
_
.
and rearrange (9.3.6) as follows:
[V(u )[
2
Q
T
1
2
[u(. T ) (. T )[
2
D
1
2
[u(. 0) (. 0)[
2
D
=
_
Q
T
((u )
t
(u ) , V(u )) J. Jt
_
Q
T
(, V) V(u ) J. Jt. (9.3.7)
For almost all t 0. T |, we can dene a linear functional J
t
:

H
1
(C) R by the
relation
J
t
( n: . ,) :=
_
D
( n
t
n , V n) J..
The quantity
[[[ [[[ J
t
(. ,) [[[ [[[ = sup
u

1
1
(D)
u,=0
_
D
( n
t
n , V n) J.
[V n[
D
is nite (it is bounded by C
T D
([(. t )[
D
[
t
(. t )[
D
) [,(. t )[
D
) and can be
viewed as a norm of this functional (note that [[[ [[[ J
t
(. ,) [[[ [[[ is square integrable on
(0. T )).
Now, we put (9.3.7) in the form
[V(u )[
2
Q
T
1
2
[(u )(. T )[
2
D
1
2
[(u )(. 0)[
2
D
=
_
T
0
J
t
((u ): . ,) Jt
_
Q
T
(, V) V(u ) J. Jt. (9.3.8)
Let and j be two given constants such that
0 < _ 2. 0 < j < 1. (9.3.9)
Dene the set
1
o
(0. T ) := {(t ) 1
o
(0. T ) [ (t ) _ j for almost all t (0. T )].
Take two scalar-valued functions
1
(t ) and
2
(t ) such that
1
(t ) =
1
_
1
1
(t )
_
and
2
(t ) =
1
(1 (t )) . (9.3.10)
In view of the YoungFenchel inequality, we have
_
T
0
J
t
((u). . ,) Jt _
_
T
0
_
1
(t)
2
[[[ [[[ J
t
(. ,) [[[ [[[
2
1
2
1
(t)
[V(u)[
2
D
_
Jt.
_
Q
T
(V ,) V(u) J. Jt _
_
T
0
_
2
(t)
2
[V ,[
2
D
1
2
2
(t)
[V(u)[
2
D
_
Jt.
Note that
1
(t ) and
2
(t ) satisfy the relation
1
1
(t )
2
(t )
= .
Now, by (9.3.9) and the inequalities we deduce the estimate
(2 )[V(u )[
2
Q
T
[(u )(. T )[
2
D
_ [(. 0) [
2
D
_
T
0
(1 (t ))[, V[
2
D
Jt
_
T
0
_
1
1
(t)
_
[[[ [[[ J
t
(. ,) [[[ [[[
2
D
Jt. (9.3.11)
This estimate is valid for any (t ) 1
o
(0. T ) and (0. 2|. It is not difcult to

observe that the right-hand side of (9.3.11) vanishes if and only if
J
t
(n: . ,) = 0 for all n

H
1
(C) and almost all t (0. T ).
, = V a.e. in Q
T
.
(0. .) = (.) for a.e. . C.
These relations mean that H
1
(Q
T
) satises the initial and boundary conditions
and for almost all t (0. T ) satises the relation
_
D
( n
t
n V Vn) J. = 0. Vn

H
1
(C).
which shows that is a solution of the problem.
To make the estimates computable, we should replace the norm of J
t
by an explic-
itly computable quantity. For this purpose, we take , in a certain subspace of Y(Q
T
).
Namely, suppose
, Y
div
(Q
T
) := {, Y(Q
T
) [ div , 1
2
(C) for a.e. t (0. T )]. (9.3.12)
Then, for almost all t (0. T ),
_
D
n(.. t )div ,(.. t ) J. =
_
D
,(.. t ) Vn(.. t ) J..
and we have
[[[ [[[ J
t
(. ,) [[[ [[[ _ C
T D
[
t
div ,[
D
. (9.3.13)
Then, by (9.3.11), we arrive at the following result.
Theorem 9.6. Let the conditions (9.3.5) and (9.3.9) be satised. Then
[[u [[
2
(1,2-)
_ M
2
EV1
(. . . ,) (9.3.14)
where
M
2
EV1
(. . . ,) :=
_
D
[(.. 0) (.)[
2
J.
_
Q
T
_
(1 )[, V[
2
C
2
T D
_
1
1
_
[
t
div ,[
2
_
J. Jt
and , Y
div
(Q
T
) and 1
o
(0. T ).
Consider two particular forms of (9.3.14), which deserve a special discussion. As-
sume that satises the initial condition and set = 1. Then we obtain
[V(u )[
2
Q
T
[(u )(. T )[
2
D
_ M
2
EV1
(. 1. . ,)
=
_
T
0
(1 (t ))[, V[
2
D
Jt
C
T D
_
T
0
_
1
1
(t)
_
[
t
div ,[
2
D
Jt. (9.3.15)
This estimate can be viewed as an analog of (3.2.8) derived for the stationary diffusion
problem. Note that in (9.3.15), = (t ) is a positive function (in the stationary case,
is a positive constant).
If = 2, then we arrive at another estimate:
[(u )(. T )[
2
D
_ M
2
EV1
(. 2. . ,)
=
1
2
_
T
0
(1 (t ))[, V[
2
D
Jt
C
T D
2
_
T
0
_
1
1
(t)
_
[
t
div ,[
2
D
Jt. (9.3.16)
which yields an upper bound of the error on the top of the space-time cylinder. Since
the right-hand side of (9.3.16) is a monotone function with respect to T , the right-hand
side also gives an upper bound of the quantity
vrai max
t0,Tj
[(u )(. t )[
2
D
.
Proposition 9.7. For any (0. 2| and 1
o
(0. T ), the variational problem

inf
:!
1
0
({
T
)
,Y
div
(Q
T
)
M
2
EV1
(. . . ,) (9.3.17)
has a solution. The exact lower bound of this problem is equal to zero. It is attained if
and only if = u and , = Vu.
Proof. The existence of a pair (. ,) H
1
0
(Q
T
)Y
div
(Q
T
) minimizing the functional
M
2
EV1
(. . . ,) is proved straightforwardly. Indeed, set = u and , = Vu. Since
u H
1,Z
2,1
, we see that div Vu 1
2
(Q
T
) and, therefore, , Y
div
(Q
T
). In this case,
M
(. ,. ) = 0, so that the exact lower bound is attained.

Assume that M
2
EV1
(. . . ,) = 0. Then, the function (.. t ) satises the initial
and boundary conditions. In addition, for almost all (.. t ) Q
T
the relations
V = , Y
div
(Q
T
) (9.3.18)
and
div ,
t
= 0 (9.3.19)
hold. Hence, is the exact solution.
Corollary 9.8. By (9.3.15) and (9.3.16), we nd that
[u [
2
V
_ M
2
EV1
(. 1. . ,) M
2
EV1
(. 2. . ,). (9.3.20)
Remark 9.9. The majorant M
2
EV1
(. . . ,) is well dened for H
1
0
(Q
T
),
1
2
(Q
T
), (.. 0) 1
2
(C), and (.) 1
2
(C). Using arguments close to
those often used in the theory of PDEs (e.g., see [217] (Chapter 2, 2)) one can extend
the estimates (9.3.15), (9.3.16), and (9.3.20) to wider sets of functions.
9.3.3 Second form of the error majorant
Now we reform the right-hand side of (9.3.7) by other means and deduce an advanced
form of the error majorant. Present the right-hand side of this relation as sum of three
terms, which are
1
1
=
_
Q
T
((u )
t
(u ) 0
t
(u ) , V(u )) J. Jt.
1
2
=
_
Q
T
(, V V0) V(u ) J. Jt.
1
3
=
_
Q
T
(0
t
(u ) V0 V(u )) J. Jt.
Here, , Y(Q
T
) and 0 H
1
0
(Q
T
) are some arbitrary functions (later we discuss
how one can chose these functions in order to obtain optimal estimates).
For almost all t (0. T ), we dene a linear functional
J
t
( : . 0. ,) :

H
1
(C) R
by the relation
J
t
( n: . 0. ,) :=
_
D
( n
t
n 0
t
n , V n) J..
It is easy to observe that the quantity
[[[ [[[ J
t
(. 0. ,) [[[ [[[ := sup
u

1
1
(D),
u,=0
_
D
( n
t
n 0
t
n , V n) J.
[V n[
D
is nite. It denes a norm of this functional and generates the estimate
1
1
_
_
T
0
[[[ [[[ J
t
(. 0. ,) [[[ [[[ [V(u )[
D
Jt.
The term 1
2
is estimated by
_
T
0
[, V V0[
D
[V(u )[
D
Jt
and 1
3
is represented in the form
1
3
=
_
Q
T
(V0 V 0
t
) J. Jt
_
Q
T
(V0 Vu 0
t
u) J. Jt
=
_
Q
T
(V0 V 0
t
0) J. Jt
_
D
(u(.. T )0(.. T ) u(.. 0)0(.. 0)) J.
= F (. 0)
_
D
((u )(.. T )0(.. T ) (u )(.. 0)0(.. 0)) J..
where
F (. 0) :=
_
Q
T
(V V0
t
0 0) J..
Since
_
D
((u )(.. T )0(.. T ) (u )(.. 0)0(.. 0)) J.
_
1
2;
[(u )(. T )[
2
D
;
2
[0(. T )[
2
D
_
D
((.) (.. 0))0(.. 0) J..
we deduce the estimate
(2 )[V(u )[
2
Q
T
_
1
1
;
_
[(u )(. T )[
2
D
_ ;[0(. T )[
2
D
2F (. 0)
_
T
0
_
(1 )[, V V0[
2
D
_
1
1
_
[[[ [[[ J
t
(. 0. ,) [[[ [[[
2
D
_
Jt
2F (. 0)
_
D
_
[(.) (.. 0)[
2
20(.. 0)((.) (.. 0))
_
J..
Here 0(.. t ) H
1
0
(Q
T
), , Y(Q
T
), (t ) 1
o
(0. T ), ; _ 1, and (0. 2|.

As in the previous section, we nd a computable majorant of the error, provided that
, Y
div
(Q
T
). In this case,
[[[ [[[ J
t
(. 0. ,) [[[ [[[ _ C
T D
[
t
0
t
div ,[
D
.
[[u [[
2
(1-
1
)
,2-)
_ M
2
EV2
(. ;. . . 0. ,). (9.3.21)
where
M
2
EV2
(. ;. . . 0. ,) := ;[0(. T )[
2
D
_
D
[(.) (.. 0)[
2
J.
_
T
0
_
(1 )[, V V0[
2
D
C
2
:
(1)
[
t
0
t
div ,[
2
D
_
Jt
2F (. 0) 2
_
D
0(.. 0)((.) (.. 0)) J..
If (.. 0) = (.) then the majorant has a simplied form:
M
2
EV2
(. ;. . . 0. ,) := ;[0(. T )[
2
D
2F (. 0)
_
T
0
_
(1 )[, V V0[
2
D
C
2
:
(1)
[
t
0
t
div ,[
2
D
_
Jt. (9.3.22)
Remark 9.10. If 0 0 and ; o, then M
2
EV2
is reduced to M
J
EV
. In practice,
the estimate (9.3.21) (which contains a correction function 0) gives a sharper upper
bound than (9.3.14). Using 0, one can reduce the residual term
t
div ,, which
may be difcult to make small by operating only with , (such a situation arises if
t
signicantly differs from u
t
).
Remark 9.11. By setting = 1 and = 2, we obtain the estimates
[V(u )[
2
Q
T
_ M
2
EV2
(. ;. 1. . 0. ,). (9.3.23)
vrai max
t0,Tj
[(u )(. t )[
2
D
_
;
;-1
M
2
EV2
(. ;. 2. . 0. ,). (9.3.24)
The proposition below shows that the majorant M
2
EV2
generates a variational prob-
lem the exact lower bound of which is attained on the solution of our problem.
Proposition 9.12. For any (0. 2|, ; _ 1, and 1
o
(0. T ), the variational

problem
inf
1
1
0
(Q
T
)
u1
1
0
(Q
T
),
,Y
div
(Q
T
)
M
2
EV2
(. ;. . . 0. ,). (9.3.25)
has a solution. The exact lower bound of this problem is equal to zero and is attained
if = u, n = 0, and , = Vu.
Proof. Obviously, the inmum in (9.3.25) is majorated by M
2
EV2
(. ;. . . 0. ,),
which coincides with M
J
EV
(. . . ,). Therefore, the result follows from Propo-
sition 9.7.
9.3.4 Equivalence of the deviation and majorant
Now we focus on another property of the majorant M
2
EV2
. Suppose (.. 0) = (.)
and consider the quantity
M
2
(;. . ) := inf
1
c
J
(0,T)
1
1
0
(Q
T
)
,Y
div
(Q
T
)
M
2
EV2
(. ;. . . 0. ,).
which is an upper bound of the error (see (9.3.21)). We are aimed at showing that this
bound is realistic, i.e., it does not lead to a large overestimation of the actual value of
the norm of the true error. For this purpose, we estimate M
(;. . ) from above and

show that this estimate is equivalent to the error.
Since u H
1,Z
0
(Q
T
), one can put , = Vu Y
div
(Q
T
). Moreover, we set
0 = u . Then

t
0
t
div , = u
t
û = 0.
and we nd that
M
2
(;. . ) _ ;[(u )(. T )[

2
D
2F (. u )
4
_
T
0
(1 )[V(u )[
2
D
Jt.
Note that
F (. u ) =
_
Q
T
(V V(u )
t
(u ) (u )) J. Jt
=
_
Q
T
(Vu V(u ) u
t
(u ) (u )) J. Jt
_
Q
T
_
[V(u )[
2
(u )
t
(u )
_
J. Jt.
The rst integral on the right-hand side is equal to zero. Hence,
M
2
(;. . ) _
_
T
0
(4
1
2)[V(u )[
2
D
Jt
;[(u )(. T )[
2
D
2
_
Q
T
(u )
t
(u ) J. Jt
_
_
T
0
(4
1 j
2)[V(u )[
2
D
Jt (; 1)[(u )(. T )[
2
D
.
Set
t
= 2 . Then, we obtain
M
2
(;. . ) _
2
(
t
2j)[V(u )[
2
Q
T
(; 1)[(u )(. T )[
2
D
.
Recall (9.3.21). We observe that for any H
1
0
(Q
T
)
[[u [[
2
(;
/
,
/
)
_ M
2
(;. . ) _ [[u [[
2
(;
//
,
//
)
_ k[[u [[
2
(;
/
,
/
)
.
where ;
t
=
;-1
;
,
tt
=
2
(
t
2j), ;
tt
= ; 1, and k = max
;.
2
_
1
2
/
__
.
This relation means that the quantity M
(;. . ) is equivalent to a certain measure of

u .
9.3.5 Comments
Generalizations. The derivation method considered above is extendable to other
parabolic equations of the form
u
t

+
Au = in Q
T
. (9.3.26)
u(.. 0) = . . C. (9.3.27)
u(.. t ) = 0. (.. t ) S
T
. (9.3.28)
which are generated by the elliptic operator
+
A (cf. Section 7.1). Generalized
solutions of such problems are dened by the integral identitiy
_
T
0
(Au. n) Jt
_
Q
T
un
t
J. Jt
_
D
(u(.. T )n(.. T ) u(.. 0)n(.. 0)) J.
=
_
Q
T
nJ. Jt. Vn V(Q
T
). (9.3.29)
where ( . ) is the scalar product associated with the spatial part of the operator and
V(Q
T
) is a suitable space of trial functions. Then, an upper bound of u is derived
by a procedure close to that used for the diffusion equation. Below we give a sketch
of it. Let be an approximation of u (which satises the boundary conditions and
belongs to the corresponding energy space). We have
_
T
0
(A(u ). n) Jt
_
Q
T
(u )n
t
J. Jt
_
D
(u(.. T ) (.. T ))n(.. T ) J.
_
D
(u(.. 0) (.. 0))n(.. 0) J.
=
_
T
0
_
D
_
n
t
n
_
J. Jt
_
T
0
(A. n) Jt.
Set n = u and note that
_
Q
T
nn
t
J. Jt =
1
2
_
[n(. T )[
2
D
[n(. 0)[
2
D
_
.
We arrive at the relation
_
T
0
[[(u )[[
2
Jt
1
2
[u(. T ) (. T )[
2
D
=
_
T
0
_
D
_
n
t
n
_
J. Jt
_
T
0
(A. n) Jt
1
2
[u(. 0) (. 0)[
2
D
. (9.3.30)
Further transformations are based on spatial properties of the operator and its con-
jugate counterpart
+
. Suppose , is such that we can represent
+
, as an integrable
function and write
_
T
0
(,. n) Jt =
_
T
0
_
D
+
, nJ. Jt.
where n V . Then, we have
_
T
0
[[(u )[[
2
Jt
1
2
[u(. T ) (. T )[
2
D
=
_
T
0
(, A. (u )) Jt
_
Q
T
(
+
,
t
)(u ) J. Jt
1
2
[ (. 0)[
2
D
. (9.3.31)
where (as in Chapter 7) [[ [[ denotes the spatial energy norm generated by A. Note
that
_
Q
T
(
+
,
t
)(u ) J. Jt _ C
_
T
0
[
+
,
t
[
D
[[(u )[[ Jt.
where C is the constant in the inequality [n[
D
_ C[[n[[.
Dene
1
and
2
by (9.3.10) and use the estimates
_
Q
T
(
+
,
t
)(u ) J. Jt
_
_
T
0
1
(t )
2
C
2
[
+
,
t
[
2
D
Jt
_
T
0
1
2
1
(t )
[[(u )[[
2
Jt
and
_
T
0
(, A. (u ))
_
_
T
0
2
(t )
2
[[, A[[
2
+
Jt
_
T
0
1
2
2
(t )
[[(u )[[
2
Jt.
Then, we obtain an estimate analogous to (9.3.14) in which [, V[
2
D
is replaced
by [[, A[[
+
, C
T D
by C, and div , by
+
,. We note that the above-discussed
method can be applied to other evolutionary problems (see [162, 301]).
Practical applications. Finally, we briey comment on possible applications of
the estimates derived. As for elliptic problems, the simplest way for the practical
exploitation of the above estimates consists of using post-processed uxes of ap-
proximate solutions. Let be an admissible approximation. Set , = 1V, where
1 : Y(Q
T
) Y
div
(Q
T
) is a post-processing operator (which performs necessary
regularization of ,). Then the quantity M
2
EV1
(. . . 1V) yields a directly com-
putable bound of the respective error norm, provided that and lie in admissible
sets.
Usually, the numerical analysis of evolutionary problems is based on using a se-
quence of consequently rening meshes. In this case, a directly computable error
estimate can be obtained in the same way as for elliptic problems (cf. Section 3.6.2).
Suppose is an approximate solution of the heat equation computed on a coarse mesh
with mesh-size t for the time variable and h for spatial variables. Let
ref
be an-
other approximate solution computed on a ner mesh (t
ref
. h
ref
) and 1
ref
denote a post-
processing operator on the rened mesh. Then, the estimates (9.3.14) and (9.3.21)
provide guaranteed and easily computable bounds of the approximation errors related
to :
[[u [[
2
(1,
/
)
_ M
2
EV1
(. . 1
ref
(Vu
ref
)) (9.3.32)
and
[[u [[
2
(;
/
,
/
)
_ M
2
EV2
(. ;. . .
ref
. 1
ref
(Vu
ref
)). (9.3.33)
These estimates can be viewed as justied quantitative forms of the Runges rule.
Minimization with respect to (t ) 1
o
(0. T ) (which is not difcult to perform) will

make the above estimates sharper.
If a more accurate error bound is required, then the majorant M
2
EV1
(. . . ,) (or
M
2
EV2
(. ;. . . 0. ,)) should be minimized with respect to , ;, , , (and 0). Cer-
tainly, such a procedure needs additional computational efforts. A rational way con-
sists of using the majorants on each step of time integration (instead of the whole inter-
val (0. T |). For example, if = 2, then we rewrite (9.3.14) for the interval (t
k
. t
k1
|
and obtain
e
2
k1
= e
2
k

1
2
_
t
k1
t
k
(1 )
_
D
_
[, V[
2
C
2
T D
[
t
div ,[
2
_
J. Jt.
where e
k
denotes the error at t = t
k
. This formula gives a way for controling the
accumulation of errors on time-steps. If the minimization with respect to , and does
not reduce the estimate below an acceptable level, then the corresponding interval (or
even several neighboring intervals) should be diminished.
9.4 A posteriori estimates for optimal control problems
Functional a posteriori error estimates open new ways of error estimation for some
classes of optimal control problems. Let [ 1
o
(C), o
d
1
2
(C. R
d
), and
1
2
(C) be given functions, and
U := {v 1
2
(C) [ v _ [ a.e. in C].
Section 9.4 A posteriori estimates for optimal control problems 243
Our goal is to nd a control function u U and a state function j
u
1
dened by the
boundary value problem
^j
v
= v a.e. in C. (9.4.1)
v = 0 on I (9.4.2)
such that the cost functional
J
1
(j. v) :=
1
2
[Vj o
d
[
2
a
2
[v u
d
[
2
. a > 0.
attains its minimal value J(j
u
. u). Another version of such a problem is generated by
the functional
J
2
(j. v) :=
1
2
[j
v
j
d
[
2
a
2
[v u
d
[
2
.
where j
d
1
2
(C).
It is well known that under the above assumptions, the optimal control problem
(with the functional J
1
or J
2
) has a unique solution (e.g., see J.-L. Lions [221]). In this
section, we show that functional a posteriori estimates allow us to obtain guaranteed
and computable bounds for the cost functional and for errors of approximations of the
state and control functions. First, we deduce an upper bound for the cost functional,
which leads to an unconstrained minimization problem. In this problem, the differ-
ential equation (9.4.1) (whose presence is the major difculty of the above optimal
control problem) does not appear explicitly. This new minimization problem can be
solved by well-known methods (e.g., by means of direct minimization). We prove that
the sequence of the so-obtained upper bounds converges to the exact value of the cost
functional and that the associated states and controls converge to the exact state and
control, respectively. These results have been established in the papers A. Gaevskaya,
R. Hoppe, and S. Repin [142, 143], where the reader will also nd numerical results
illustrating the reliability and efciency of the approach.
9.4.1 Two-sided bounds for cost functionals
Upper bounds. Assume that v U is an admissible control function computed by
some numerical procedure and j
v
is the corresponding state function. Since the latter is
a solution of the state boundary value problem, we do not know it exactly and instead
must operate with a certain approximation j

H
1
(C) (which may not satisfy the
1
In the literature devoted to optimal control problems, the control function is traditionally denoted by
u. However, in other parts of the book this letter was used to denote the exact solution of a boundary
value problem. We try to follow the style accepted and, at the same time, to avoid a collision of notation.
For this reason, in this section, we use special fonts (u, v) for the denotation of control functions.
differential equation). By the triangle inequality, we obtain the following upper bound
for the cost functional:
J
1
(j
v
. v) _
1
2
_
[Vj o
d
[ [V(j
v
j)[
_
2
a
2
[v u
d
[
2
. (9.4.3)
Now, we estimate the term [V(j
v
j)[ by (4.1.13) and nd that
[V(j
v
j)[ _ [t Vj[ C
T D
[div t v [.
where t is an arbitrary function in H(C. div).
In view of (9.4.3) and (9.4.4), we have
J
1
(j
v
. v) _ J
1
(j. t. v) :=
a
2
[v u
d
[
2
1
2
_
[Vj o
d
[ [t Vj[
C
T D
[div t v [
_
2
. (9.4.4)
The functional J
1
(j. t. v) is directly computable and provides an upper bound of the
cost functional for any j V
0
:=

H
1
(C), t H(C. div), and v U. It is easy to
see that (9.4.4) has no gap between its left- and right-hand sides. Indeed, set v = u,
t = Vj
u
, and j = j
u
. Then,
J
1
(j. t. v) =
a
2
[u u
d
[
2
1
2
[Vj
u
o
d
[
2
.
i.e., J
1
(j. t. v) coincides with the value of the cost functional computed for the exact
solution.
Using Youngs inequality with positive parameters and , we can represent this
bound in terms of a quadratic functional:
J
1
(j
u
. u) _ J
1
(j
v
. v) _ J
1
(. : j. t. v). (9.4.5)
where
J
1
(. : j. t. v) :=
1
2
[Vj o
d
[
2
(1 )(1 )
2
[t Vj[
2
(1 )(1 )
2
C
2
T D
[div t v [
2
1
2
[v u
d
[
2
. (9.4.6)
Setting v = u, t = Vj
u
, and j = j
u
and letting go to zero, we nd that
J
1
(. 0: j. t. v) also tends to J
1
(j
u
. u).
J
1
(j
u
. u) = inf
)V
0
, vU,
r1(D,div), ,R
J
1
(. : j. t. v). (9.4.7)
The majorants J
1
(j. t. v) and J
1
(. : j. t. v) can be used for nding guaranteed
upper bounds for the cost functional. For example, we can take j and v as approximate
solutions computed by a certain optimization procedure and additionally minimize the
majorant with respect to t (and to the parameters and ). In the simplest case, we can
take t as a post-processed ux V and perform a simple minimization with respect to
and . The respective value J gives an upper bound of the cost functional. It should
be outlined that the quantity J
1
(v. j) may not provide such a guaranteed upper bound
because j is not the exact solution of (9.4.1).
For the functional J
2
, the majorant can be easily derived by applying the same
method. We have
J
2
(j. v) =
1
2
[j j
d
[
2
a
2
[v u
d
[
2
_ J
2
(. : j. t. v). (9.4.8)
where
J
2
(. : j. t. v) :=
1
2
[j j
d
[
2
(1 )(1 )
2
C
2
T D
[t Vj[
2
(1 )(1 )
2
C
4
T D
[div t v [
2
a
2
[v u
d
[
2
.
Finding the sharpest upper bound for cost functionals requires the minimization of
J
1
(or J
2
) over j, t, v, , and , where the variables are taken in the above-stated sets
and are formally independent. Below, we show (with the paradigm of the majorant
J
1
) that the amount of independent variables can be reduced.
It is easy to observe that the minimization of J
1
with respect to v is equivalent to
the problem
inf
vU
M(. : t. v) =

M(. : t).
where
M(. : t. v) :=
C
2
[div t v [
2
1
2
[v u
d
[
2
and C
= C
2
T D
(1)(1)
. This problem is reduced to the minimization of the

integrand of M at almost all . C. If no constraints are imposed on the control
function (i.e., U = 1
2
(C)), then the respective minimizer v is easy to nd. It satises
the relation
v(.) =
1
C
a
_
au
d
(.) C
(div t(.) (.))

_
.
which implies
M(. : t) =
C
a
2(C
a)
[div t u
d
[
2
. (9.4.9)
If U contains a nite constraint [, then
v(.) =
_
v(.) if . C
0
.
[(.) if . C
)
.
where C
)
:= {. C[v(.) > [(.)] and C
0
:= C\ C
)
. In this case,
M(. : t. v) =
C
a
2(C
a)
[div t u
d
[
2
D
0
2
[div t [ [
2
D
)
a
2
[[ u
d
[
2
D
)
. (9.4.10)
Hence,
inf
vU
J
1
(. : j. t. v) =
J
1
(. : j. t). (9.4.11)
where
J
1
(. : j. t) =
1
2
[Vj o
d
[
2
(1 )(1 )
2
[t Vj[
2

M(. : t. v).
Then, we conclude that the problem of nding the sharpest upper bound for the cost
functional can be formally reduced to the following minimization problem:
inf
)V
0
,r1(D,div),
,R
J
1
(. : j. t). (9.4.12)
Lower bounds. Suppose o
d
= Vj
d
, where j
d
V
0
. Then J
1
has the form
J
1
(j. v) :=
1
2
[V(j j
d
)[
2
a
2
[v u
d
[
2
. (9.4.13)
We note that if o
d
does not have such a form, then the optimization problem can be
reduced to the above-considered case. Indeed, let j
d
be the projection of o
d
onto V
0
,
i.e.,
_
D
(Vj
d
o
d
) VnJ. = 0. Vn V
0
.
Then
[Vj o
d
[
2
= [Vj Vj
d
[
2
[Vj
d
o
d
[
2
and
J(j. u) =
1
2
[Vj Vj
d
[
2
a
2
[u u
d
[
2
c.
where c = [Vj
d
o
d
[
2
is the distance fromj
d
to the set V
0
. Thus, the cost functional
can be reduced to the form (9.4.13).
We derive a lower bound for the functional J
1
(j
v
. v). For any j V
0
, we have
J
1
(j
v
. v) :=
1
2
[V(j
v
j)[
2
1
2
[V(j j
d
)[
2
_
D
V(j
v
j) V(j j
d
) J.
a
2
[v u
d
[
2
=
1
2
[V(j
v
j)[
2
1
2
[V(j j
d
)[
2
_
D
( v)(j j
d
) J.
_
D
Vj V(j j
d
) J.
a
2
[v u
d
[
2
.
Hence,
J
1
(j
u
. u) = inf
vU
J
1
(j
v
. v)
_
1
2
[V(j j
d
)[
2
_
D
((j j
d
) Vj V(j j
d
)) J.
inf
vU
__
D
v(j j
d
) J.
a
2
[v u
d
[
2
_
. (9.4.14)
Note that
inf
vU
__
D
gv J.
a
2
[v u
d
[
2
_
=
_
D
H(a. u
d
. [. g) J..
where
H(a. u
d
. [. g) J. :=
_
_
u
d
g
1
2a
g
2
if v := u
d
g
a
_ [.
[g
a
2
([ u
d
)
2
if v > [.
Now we obtain the following lower bound for the cost functional, which involves only
known functions:
J
1
(j
u
. u) _ J
1
(j) :=
1
2
[V(j j
d
)[
2
_
D
((j j
d
) Vj V(j j
d
)) J.
_
D
H(a. u
d
. [. j j
d
) J.. (9.4.15)
Assume that U = 1
2
(C). Then, it is easy to show that the minorant is sharp. Indeed,
sup
)V
0
J
1
(j) _ J
1
(j
u
) =
1
2
[V(j
u
j
d
)[
2
_
D
((j
u
j
d
) Vj
u
V(j
u
j
d
)) J.
_
D
H(a. u
d
. [. j
u
j
d
) J..
where
H(a. u
d
. [. j
u
j
d
) = u
d
(j
u
j
d
)
1
2a
[j
u
j
d
[
2
.
It is easy to prove (e.g., see [221]) that the corresponding solution of the optimal
control problem satises the necessary condition
u = u
d
1
a
(j
d
j
u
). (9.4.16)
Therefore,
_
D
((j
u
j
d
) Vj
u
V(j
u
j
d
)) J.
_
D
H(a. u
d
. [. j
u
j
d
) J.
=
_
D
_
(u
d
u)(j
u
j
d
)
1
2a
(j
u
j
d
)
2
_
J. =
a
2
[u u
d
[
2
.
Thus, J
1
(j
u
. u) = J
1
(j
u
).
To nd a more accurate lower bound for the cost functional, we estimate the rst
term on the right-hand side of (9.4.13) with the help of (3.1.13), which reads
1
2
[V(j
v
j)[
2
_
_
D
_
1
2
[Vn[
2
Vn Vj (v )n
_
J.. Vn V
0
. (9.4.17)
This way results in a more complicated estimate:
J
1
(j
u
. u) _ J
1
(j. n) :=
1
2
[Vn[
2
1
2
[V(j j
d
)[
2
_
D
((n j j
d
) Vj V(n j j
d
)) J.
inf
vU
__
D
v(n j j
d
) J.
a
2
[v u
d
[
2
_
. (9.4.18)
where the last term is equal to
_
D
H(a. u
d
. [. n j j
d
) J..
Remark 9.13. It should be noted that this estimate contains an additional function
n, which makes computations of the lower more expensive with respect to (9.4.15).
Numerical experiments performed (in part they are cited in [142, 143]) have shown
that in most cases, specications computed with the help of n are not very essential.
9.4.2 Estimates for state and control functions
Now our goal is to derive guaranteed upper bounds for the errors of v and j measured
in terms of a combined norm
[u v|[
2
:=
1
2
[V(j
u
j
v
)[
2
a
2
[u v[
2
.
Our analysis is based upon the following result, which can be viewed as a generaliza-
tion of the Mikhlin estimate (2.3.1) for the class of optimal control problems that we
consider.
Theorem 9.14. Let U = 1
2
(C). For any control function v U,
[u v|[
2
= J
1
(j
v
. v) J
1
(j
u
. u). (9.4.19)
Proof. We have
J(j
v
. v) J(j
u
. u) =
1
2
[V(j
v
j
u
)[
2
a
2
[v u[
2
_
D
V(j
u
j
d
) V(j
v
j
u
) J.
a
_
D
(u u
d
)(v u) J.. (9.4.20)
Note that
_
D
V(j
v
j
u
) V(j
u
j
d
) J. =
_
D
(v u)(j
u
j
d
) J..
By (9.4.16), we know that
(j
u
j
d
) a(u u
d
) = 0. (9.4.21)
_
D
V(j
v
j
u
) V(j
u
j
d
) J. = a
_
D
(v u)(u u
d
) J.
and the last two terms in (9.4.20) vanish. Hence, we arrive at (9.4.19).
Corollary 9.15. From (9.4.6), (9.4.14), and (9.4.19), it follows that
[v u|[
2
_ M
opt
(. . j. t. v) := J
1
(. : j. t. v) J
1
(j). (9.4.22)
where v is an arbitrary control function in 1
2
(C), t H(C. div), and are arbi-
trary positive numbers, and
M
opt
(. . j. t. v) :=

2
[V(j j
d
)[
2
(1 )(1 )
2
[t Vj[
2
(1 )(1 )
2
C
2
T D
[div t v [
2
1
2
[v u
d
[
2
_
D
(( u
d
)(j j
d
) Vj V(j j
d
)) J.
1
2a
[j j
d
[
2
J..
Proposition 9.16. The majorant M
opt
(. . j. t. v) attains the exact lower bound on
the exact solution of the optimal control problem, i.e.,
inf
,>0
M
opt
(. . j
u
. Vj
u
. u) = 0.
Proof. In view of (9.4.21),
_
D
(( u
d
)(j
u
j
d
) Vj
u
V(j
u
j
d
)) J.
=
_
D
((u u
d
)(j
d
j
u
) =
1
a
[j
d
j
u
[
2
(9.4.23)
and
1
2
[u u
d
[
2
=
1
2a
[j
d
j
u
[
2
.
For this reason, the last three terms of M
opt
(. . j. t. v) vanish. The second term is
also equal to zero, as well as the third one. Thus,
M
opt
(. . j
u
. Vj
u
. u) =

2
[V(j
u
j
d
)[
2
and the result follows if we let go to zero.
Remark 9.17. Finally, we note that it may be useful to represent the basic problem in
another (but equivalent) form. For a function V
0
, consider the following problem:
Minimize
J
1,(
(j. u) :=
1
2
[V(j )[
2
a
2
[u u
d
()[
2
(9.4.24)
over (j. u) V
0
1
2
(C) such that
^j = u a.e. in C. (9.4.25)
where u
d
() = u
d
1
a
(j
d
).
In [143], it is shown that for any j V
0
,
J
1
(j. u) = J
1,(
(j. u) C
(
. (9.4.26)
where
C
(
:= C(: j
d
. . u
d
) =
1
2
[V Vj
d
[
2
a
2
[u
d
() u
d
[
2
__
D
V V(j
d
) J.
_
D
(u
d
() )(j
d
) J.
_
.
We outline that C
(
depends only on and known functions j
d
. , and u
d
. This
constant gives a quantity, which is contained in the upper bound of the cost functional,
as well as in the lower one. In practice, it is convenient to reformulate the problem in
such a way that this a priori known constant is sufciently large. Then the estimates
based on the comparison of upper and lower estimates of the cost functional would
become more efcient (see [143] for details and numerical experiments).
9.4.3 Estimate in a combined norm
Introduce the following combined norm:
[v: q|[
2
2
:= [v|[
2
1
2
[q[
2
z[div q[
2
.
where (v. q) 1
2
(C) H(C. div) and z (0. a). This norm can be regarded as a
full primal-dual norm associated with the problem under consideration.
Proposition 9.18.
[(v u): (t )|[
2
2
_ c
M
opt
(. . j. t. v). (9.4.27)
where := Vj
u
and
c
:= 3
2
(1)(1)
max
_
6.

C
2
1:
_
4C
2
T D
2o
o-2
_
_
.
Proof. By the obvious inequality
1
2
[t [
2
=
1
2
[t Vj
u
[
2
_ [V(j
v
j
u
)[
2
[Vj
v
t[
2
= 2[v u|[
2
a[v u[
2
[Vj
v
t[
2
(9.4.28)
and (9.4.22), we nd that
1
2
[t [
2
_ 2M
opt
(. . j. t. v) a[v u[
2
[Vj
v
t[
2
. (9.4.29)
For any positive z and j, we have
z[div t div [ = z[div t u [
2
_ z
_
(1 j)[div t v [
2
1 j
j
[v u[
2
_
. (9.4.30)
Setting z
1
= a, we nd that j =
2
o-2
. Now (9.4.29) and (9.4.30) imply the
estimate
1
2
[t [
2
z[div (t )[
2
_ 2M
opt
(. . j. t. v)
za
a z
[div t v [
2
[Vj
v
t[
2
.
which, together with (9.4.22), yields
[(u u): (t )|[
2
2
_ 3M
opt
(. . j. t. v)
za
a z
[div t v [
2
[Vj
v
t[
2
. (9.4.31)
For the last term, we have
[Vj
v
t[
2
_ 2[V(j
v
j)[
2
2[Vj t[
2
_ 2([Vj t[ C
T D
[div t v [)
2
2[Vj t[
2
_ 6[Vj t[
2
4C
2
T D
[div t v [
2
.
[(u u): (t )|[
2
2
_ 3M
opt
(. . j. t. v)
_
4C
2
T D
za
a z
_
[div t v [
2
6[Vj t[
2
.
Recalling the structure of M
opt
(. . j. t. v), we arrive at (9.4.27).
Remark 9.19. In the proof, we overestimated the right-hand side of inequalities sev-
eral times. Therefore, in reality the constant c
is essentially smaller than it is dened

by Proposition 9.18.
Remark 9.20. With the help of similar arguments, it can be shown that
[(v u): (t )|[
2
2
_ c
M
opt
(. . j. t. v).
where c
depends only on the data of the problem.

9.4.4 Generalizations
As we have seen, error majorants of the functional type allow to consider state equa-
tions of optimal control problems in the form of penalty functionals. This method
yields computable upper bounds for cost functionals of many other optimal control
problems (provided that the majorant for the state problem is known). Below we
briey consider several examples related to nonlinear problems.
1. Consider the problem with the cost functional J
1
in which the state function is
dened not by (9.4.1) and (9.4.2) but by the variational inequality (see 8.1.1)
_
D
Vj
v
V(j j
v
) J. _
_
D
(v )(j j
v
) J.. Vj 1. (9.4.32)
where
1 := {j V
0
:=

H
1
(C) [ (.) _ j(.) _ [(.) a.e. in C].
Now the problem is to minimize J
1
(j
v
. v) on (j
v
. v) 1 U. We use (9.4.3) and
estimate the term [V(j
v
j)[ by the majorant derived in Section 8.1. Instead of
(9.4.4), we obtain
J
1
(j
v
. v) _ J
1
(j. t. v) :=
a
2
[v u
d
[
2
1
2
_
[Vj o
d
[ [t Vj[
C
T D
[ div , v
)
[
_
2
. (9.4.33)
where , H(C. div). For any (j
v
. v) 1 U, the quantity J
1
(j
v
. v) is computable
and gives a guaranteed upper bound for the cost functional.
Also, we can derive a computable lower bound. Suppose o
d
=Vj
d
, where j
d
1.
Let v U be an approximation of u and j be an approximation of j
v
. Then,
J
1
(j
v
. v) :=
1
2
[V(j
v
j)[
2
1
2
[V(j j
d
)[
2
_
D
V(j
v
j) V(j j
d
) J.
a
2
[v u
d
[
2
_
1
2
[V(j j
d
)[
2
a
2
[v u
d
[
2
[V(j
v
j)[[V(j j
d
)[
_
1
2
[V(j j
d
)[
2
a
2
[v u
d
[
2
1
2
_
[t Vj[ C
T D
[ div t v
)
[
_
2
. (9.4.34)
The right-hand side of (9.4.34) can be maximized with respect to v U by the same
method, which we applied to the optimal control problem with linear state equation.
Then we obtain a computable lower bound for the cost functional. The quality of this
lower bound depends on the value of approximation error for the state function.
2. The method is extendable to optimal control problems with convex cost function-
als and also to those that have the control function in the main part of the differential
operator (in the latter case, the existence of u requires a special investigation). If the
cost functional J(j. ) is convex with respect to the rst variable for any admissible v,
them we have the relations
J(j
v
. v) _ zJ
_
j
z
. v
_
(1 z)J
_
j
v
j
1 z
. v
_
. (9.4.35)
J
_
j
v
z
. v
_
_
1
z
J(j. v)
(1 z)
z
J
_
j j
v
1 z
. v
_
. (9.4.36)
where z (0. 1). Assume that J satises the growth condition
J(j. v) _ C(v)[j[
V
. (9.4.37)
where C > 0 does not depend on j and V is the energy space of the boundary value
problem (which denes the state function)
(v)j
v
= (v) (9.4.38)
generated by an elliptic operator . If for the problem (9.4.38) we have a computable
upper bound of the deviation from j
v
[j
v
j[
V
_ M(j. (v). (v). D). (9.4.39)
then (9.4.35), (9.4.37), and (9.4.39) imply a computable upper bound for the cost func-
tional, namely,
J(j
v
. v) _ inf
2(0,1)
_
zJ
_
j
z
. v
_
(1 z)
1-
C(v)M
(j. (v). (v). D)

_
. (9.4.40)
A lower bound follows from (9.4.36). However, in general, getting computable lower
bounds for cost functionals is a more difcult task. Usually it requires a deeper analy-
sis, which attracts specic properties of the problem considered.
9.4.5 Comments
For the reader interested in other approaches to a posteriori estimation for optimal con-
trol problems (and in the adaptive numerical methods developed for such problems),
we recommend the papers by R. Becker and R. Rannacher [49], R. Becker, H. Kapp,
and R. Rannacher [47], M. Hinterm uller [173], R. H. W. Hoppe, Y. Iliash, C. Iyyunni,
and N. H. Sweilam [177], A. Gaevskaya, R. H. W. Hoppe, Y. Iliash, and M. Kieweg
[141], D. Meidner and B. Vexler [230], and R. Becker and B. Vexler [50]. These
papers also contain an overview of the results in the area and many references.
9.5 Estimates for nonconforming approximations
Let C be divided into a collection of subdomains C
i
, i = 1. 2. . . . . N, and
C =
_
i
C
i
.
We consider nonconforming approximations that may violate the continuity on the
boundaries of subdomains C
i
and the boundary conditions on I
1
. The corresponding
functions are marked by hats and form a broken Sobolev space
H
1
:=
n 1
2
(C)[ n H
1
(C
i
). i = 1. 2. . . . . N
_
.
the norm of which is dened by the relation
[n|[
2
=
i
[n[
2
1,2,D
i
.
Section 9.5 Estimates for nonconforming approximations 255
Also, we dene the following norms:
[[q[[
2
D
i
:=
_
D
i
q q J.. [[q[[
2
+,D
i
:=
_
D
i
-1
q q J..
[q|[
2
:=
i
[[q[[
2
D
i
. [q|[
2
+
:=
i
[[q[[
2
+,D
i
.
Guaranteed bounds of approximation errors generated by nonconforming approx-
imations can be derived by two methods. The rst method projects a nonconform-
ing approximation into the energy space (with the help of a suitable post-processing
procedure) and applies the functional error majorant to the post-processed approxima-
tion. Since the majorant is valid for any conforming approximation, this procedure
implies computable and guaranteed error bounds. The second method is based on the
Helmholtz decomposition of the error. Below we briey discuss both methods with
the paradigm of the problem (4.1.1)(4.1.3).
9.5.1 Estimates based on projecting to the energy space
Let

V be a nonconforming approximation of u u
0
V
0
(we recall that V
0
contains functions vanishing on I
1
). Probably the simplest modus operandi is to dene
:= P().
where P :

V u
0
V
0
is a projection operator. In particular, one can use the or-
thogonal projection to a certain nite-dimensional subspace of u
0
V
0
. If P produces
~
0
+
b a
~
P
P
0
u V
u
u
v
v
v
v
u
0
+V
0
V
^
V
^
^
^
Figure 9.5.1 Projection to u
0
V
0
.
almost orthogonal projection (see Figure 9.5.1 a), then is a better approximation of
u than. In this case, it is logical to consider as an approximate solution and use the
estimates
[[Vu V[[ _ M
DF
(. ,). , H(C. div).
[[Vu V[[ _ M
DF
(. n). n V
0
.
to evaluate the respective error. Similar estimates can easily be obtained for noncon-
forming approximations of other boundary value problems, where the corresponding
functional error majorants are known.
However, orthogonal projection is equivalent to solving of an auxiliary problem,
which may lead to essential expenditures. On the other hand, cheap post-processing
procedures may destroy approximation properties of , so that is less accurate than
(see Figure 9.5.1 b). In this case, we should estimate the error in terms of the broken
norm [
V Vu|[, in which
V is dened at almost all points of C by the relation
V(.) := V(.). . C
i
. i = 1. 2. . . . . N.
The simplest way to obtain such an estimate is as follows. For any u
0
V
0
, we
have the triangle inequality
[Vu
V|[ _ [V(u )|[ [
V V|[. V u
0
V
0
. (9.5.1)
Remark 9.21. It should be noted that [
V|[ is a seminorm on

V . One can view this
quantity as a factor norm dened for the class of functions that differ from each other
by a constant, so that in (9.5.1) we can write u
0
V
0
R.
Consider the rst term on the right-hand side of (9.5.1). Since u V
0
, the
broken norm [V(u )|[ coincides with the usual energy norm [[V(u )[[, which
is estimated from above by the error majorant. Moreover, the value of this norm does
not change if we replace by c (where c R). For this reason, we arrive at the
estimate
[Vu
V|[ _ inf
u
0
V
0
,1(D,div)
M
DF
(. ,) [
V V|[
_
. (9.5.2)
If = u and , = Vu, then M
DF
(. ,) = 0. Therefore, (9.5.2) holds as the equality.
Let := P(). Then, (9.5.2) is replaced by
[Vu
V|[ _ inf
,1(D,div)
M
DF
(. ,) [
V V|[. (9.5.3)
where , H(C. div) should be used to minimize the right-hand side of (9.5.2).
Lower bounds of approximation errors are derived analogously. By the obvious
inequality
[Vu
V|[ _ [V(u )|[ [
V V|[. V u
0
V
0
. (9.5.4)
we nd that
[Vu
V|[ _ sup
u
0
V
0
uV
0
_
M
DF
(. n) [
V V|[
_
(9.5.5)
and, therefore,
[Vu
V|[ _ sup
uV
0
M
DF
(. n) [
V V|[. (9.5.6)
If the term [
V V|[ (related to the nonconformity error) is large, then the

estimates (9.5.3) and (9.5.6) may be not sharp. In this case, estimates of a somewhat
different type can be helpful. They are obtained if instead of the triangle inequality we
use another representation of the error.
9.5.2 Estimates based on the Helmholtz decomposition
Consider the error function
:= Vu
V.
Since
V is dened at almost all points of C and V 1

2
(C
i
. R
d
) for i = 1. . . . . N,
we can regard as a vector-valued function in 1
2
(C. R
d
), for which the well-known
Helmholtz decomposition takes place. For our purposes, it is more convenient to use
a similar decomposition of , namely,
= Vu
)
t
)
. (9.5.7)
where u
)
is a function in V
0
and
t
)
S
I
2
:=
_
t 1
2
(C. R
d
) [
_
D
t VnJ. = 0. Vn V
0
_
.
If t S
I
2
then div t = 0 and t n = 0 on I
2
(these relations should be understood
in a generalized sense; they hold in the classical sense if t is a sufciently regular
function).
The decomposition (9.5.7) is motivated by the problem: Find u
)
V
0
such that
_
D
Vu
)
VnJ. =
_
D
VnJ.. Vn V
0
. (9.5.8)
which is uniquely solvable. From (9.5.8) it follows that
_
D
t
)
VnJ. = 0. Vn V
0
. where t
)
:= ( Vu
)
). (9.5.9)
which means that t
)
S
I
2
and we obtain (9.5.7).
By the denition of S
I
2
, t
)
is orthogonal to Vn for any n V
0
. Therefore, we
arrive at the important relation
[|[
2
=
i
_
D
i
J.
=
_
D
_
Vu
)
Vu
)

-1
t
)
t
)
2Vu
)
t
)
_
J.
= [[Vu
)
[[
2
[[t
)
[[
2
+
. (9.5.10)
A posteriori estimates based on the Helmholtz type decomposition of were studied
in E. Dari, R. Duran, C. Padra, and V. Vampa [113] and M. Ainsworth [4, 5]. In those
papers, t
)
is represented as curl and is dened by the relation
_
D
-1
curl curl J. =
_
D
curl J.. V H.
where H is the subspace of H
1
,R that consists of functions having zero tangential
derivatives on I
2
. Hence, the overall error is represented by (9.5.10), where u
)
and t
)
are dened as solutions of auxiliary boundary value problems.
Below we discuss a somewhat different modus operandi, which was suggested and
numerically tested in S. Repin and S. Tomar [317]. First, we estimate u
)
with the help
of the same method as we used for deriving a posteriori error estimates (for conforming
approximations) in previous chapters. This method provides computable two-sided
bounds for the term [[Vu
)
[[. After that, we show that the value of [[t
)
[[ is estimated
from above by the broken norm of the difference V

V (i.e., by a penalty for
nonconformity). The sum of these two estimates yields a directly computable bound
of the error, expressed in terms of the broken energy norm.
Upper bound of [|[. First we nd computable upper bounds for the norms
[[Vu
)
[[. Note that
_
D
Vu
)
VnJ. =
i
_
D
i
VnJ.
_
D
t
)
VnJ.
=
_
D
Vu VnJ.
i
_
D
i
V VnJ.
=
_
D
nJ.
_
I
2
JnJs
i
_
D
i
V VnJ.. Vn V
0
. (9.5.11)
Now, we rearrange the right-hand side of (9.5.11) by introducing a vector-valued func-
tion , H(C. div) and obtain
_
D
Vu
)
VnJ. =
_
D
(div , )nJ.
_
I
2
(J , n)nJs
i
_
D
i
(, V) VnJ.. (9.5.12)
[[Vu
)
[[ _ [,
V|[
+
C
_
[div , [
2
[, n [
2
I
2
_
1{2
. (9.5.13)
For the function t
)
, we have
_
D
-1
t
)
t
)
J. =
i
_
D
i
t
)
J. =
i
_
D
i
(Vu V) t
)
J.
=
i
_
D
i
(V V) t
)
J.. V u
0
V
0
. (9.5.14)
Hence,
[[t
)
[[
2
+
_
_
i
_
D
i
(V V) (V V) J.
_
1{2
[[t
)
[[
+
.
and we nd that
[[t
)
[[
+
_ [V
V|[. (9.5.15)
By (9.5.10), (9.5.13), and (9.5.15), we deduce the estimate
[Vu
V|[
2
_ inf
u
0
V
0
[V
V|[
2
_
[,
V|[
+
C
_
[div , [
2
[, n J[
2
I
2
.
_
1{2
_
2
. (9.5.16)
which provides an upper bound of the error in terms of the broken energy norm.
Remark 9.22. The right-hand side of (9.5.16) presents a natural decomposition of the
overall error into three terms: the error owing to nonconformity, the error in the duality
relation for uxes, and the error in the equilibrium equation and boundary condition
for uxes.
Remark 9.23. If , is subject to the boundary condition , n = J, then (9.5.16) takes
a simplied form
[|[
2
_ [V
V|[
2
_
[,
V|[ C[div , [
_
2
. (9.5.17)
where is an arbitrary function in u
0
V
0
. We set = P() (i.e., dene as a
projection onto u
0
V
0
). Then,
[|[
2
_
2
P
(1 )[,
V|[
2
C
2
_
1
1
_
[div , [
2
. (9.5.18)
where is an arbitrary positive number and
P
:= [VP()
V|[ is the projection

error (which is directly computable). Minimization with respect to , is now reduced
to a quadratic problem.
Finally, we consider a modication of (9.5.16). Let I
i}
= dC
i
dC
}
and n
i}
denote the unit normal vector to I
i}
external to C
i
if i < . In (9.5.16), we assume
that , belongs to H(C. div) (which means that , n
i}
is continuous on I
i}
). However,
V is a vector-valued function that may have jumps on I

i}
. For this reason, it may
be useful to have another form of the upper bound, which operates with , from a
wider set. To deduce such an estimate, we transform the right-hand side of (9.5.11) as
follows. Let
,

Y :=
_
, = ,
(i)
in C
i
. ,
(i)
H(C
i
. div). i = 1. . . . . N
_
.
Denote the broken divergence by
b
div (i.e.,
b
div ,
(i)
= div ,
(i)
in C
i
).
For any o H(C. div), we have
_
D
(n
b
div, Vn ,) J. =
i
_
D
i
(ndiv ,
(i)
Vn ,
(i)
) J.
=
_
D
(n
b
div (, o) Vn (, o)) J.
_
I
2
(o n)nJs.
Substitute this identity into (9.5.11). We have
_
D
Vu
)
VnJ. =
_
D
(
b
div,)nJ.
_
I
2
(J o n)nJs
i
_
D
i
(, V) VnJ.
_
D
(n
b
div (, o) Vn (, o)) J..
Since
_
D
(n
b
div (, o) Vn (, o)) J.
_
_
C
T I
1
[
b
div (, o)[ [, o[
_
[Vn[.
we nd that
[[Vu
)
[[ _ c
-1
1
_
C
T I
1
[
b
div,[ C
TI
2
[J o n[
I
2
C
T I
1
[
b
div (, o)[ [, o[
_
[,
V|[
+
.
Now, by (9.5.10) and (9.5.13) we obtain
[Vu
V|[
2
_ inf
u
0
V
0
[V
V|[
2
_
c
-1
1
_
C
T I
1
[
b
div,[ C
TI
2
[J o n[
I
2
C
T I
1
[
b
div (, o)[ [, o[
_
[,
V|[
+
_
2
. (9.5.19)
where , is any vector-valued function in
Y .
From (9.5.19), we deduce the estimate
[Vu
V|[
2
_ inf
u
0
V
0
[V
V|[
2
_
c
-1
1
C
T I
1
[
b
div,[ [,
V|[
+
c
-1
1
inf
c1(D,div)
cn=T on I
2
_
C
T I
1
[
b
div (, o)[ [, o[
__
2
. (9.5.20)
which operates with piecewise continuous approximations and uxes. We note that
the last term on the right-hand side of the above estimate can be viewed as a penalty
for nonconformity of ,.
Lower bound of [|[. To derive a lower bound of the error we again use (9.5.10).
Now, our goal is to nd lower bounds for the norms on the right-hand side of (9.5.10).
To estimate the rst term, we rewrite (9.5.11) in the form
_
D
Vu
)
VnJ. = (n). (9.5.21)
where : V
0
R is a linear functional dened as
(n) =
_
I
2
JnJs
_
D
nJ.
i
_
D
i
V VnJ.. (9.5.22)
Note that (9.5.21) is the Eulers equation of the variational problem
min
uV
0
J
I
(n). where J
I
(n) =
1
2
[[Vn[[
2
(n). (9.5.23)
[[Vu
)
[[
2
= (. u
)
). (9.5.24)
and, therefore,
J
I
(u
)
) =
1
2
[[Vu
)
[[
2
. (9.5.25)
From (9.5.24) and (9.5.25), we obtain
[[Vu
)
[[
2
= 2J
I
(u
)
) = 2 inf
uV
0
1
2
_
[[Vn[[
2
(n)
_
= sup
uV
0
_
2(n) [[Vn[[
2
_
.
Hence,
[[Vu
)
[[
2
_ 2(n) [[Vn[[
2
. Vn V
0
. (9.5.26)
For the second term in (9.5.10), we proceed analogously. In view of (9.5.7), t
)
meets the identity
_
D
-1
t
)
t
0
J. =
_
D
( Vu
)
) t
0
J.
=
_
D
(Vu
V Vu
)
) t
0
J.. Vt
0
S
I
2
. (9.5.27)
Introduce the functional j : S
I
2
R,
j(t
0
) :=
i
_
D
i
(Vu
0
V) t
0
J.. (9.5.28)
Then, t
)
is a minimizer of the variational problem
min
r
0
S
T
2
1
(t
0
). where 1
(t
0
) :=
_
1
2
[[t
0
[[
2
+
j(t
0
)
_
. (9.5.29)
By the arguments similar to those used before, we conclude that
[[t
)
[[
2
+
= 21
(t
)
) = sup
r
0
S
T
2
_
2j(t
0
) [[t
0
[[
2
+
_
_ 2j(t
0
) [[t
0
[[
2
+
. Vt
0
S
I
2
. (9.5.30)
Combining (9.5.26) and (9.5.30), we obtain the following lower estimate of the error
[|[
2
_ 2(n) 2j(t
0
) [[Vn[[
2
[[t
0
[[
2
+
. (9.5.31)
In (9.5.31), n and t
0
are arbitrary functions in V
0
and S
I
2
, respectively. Certainly,
getting a realistic estimate requires a proper selection of these functions.
Remark 9.24. If u
0
V
0
, then j(t
0
) 0 and we should take t
0
= 0 to make the
right-hand side of (9.5.31) maximal. In this case, (9.5.31) is transformed to the lower
bound that was derived for conforming approximations in Chapter 4.
9.5.3 Accuracy of approximations obtained by the Trefftz method
Approximate solutions to boundary value problems can be constructed as series
i

i
i
(.), where the
i
exactly satisfy the differential equation. The coefcients
i
are selected in order to approximate the Dirichlet boundary condition as accurately
as possible (in the sense of least squares). Numerical methods of this type originate
from the Trefftz method (e.g., see S. Mikhlin [232]). For the problem (4.1.1)(4.1.3),
the respective approximate solution u satises the equation
div Vu = 0 in C.
but violates the condition u = u
0
on I
1
. We can view such a function as a noncon-
forming approximation of u and apply (9.5.16). In this case,
V = H
1
(C) and [Vu
Vu|[ = [[V(u u)[[.

Set , = Vu. Then, (9.5.16) implies the estimate
[[Vu Vu[[
2
_ inf
u
0
V
0
[[V( u)[[
2
C
2
[Vu n [
2
I
2
. (9.5.32)
If I
2
= 0, then we arrive at a simple projection estimate
[[Vu Vu[[ _ inf
u
0
V
0
[[V( u)[[ _ [[V(Pu u)[[. (9.5.33)
Since
(n) =
_
I
2
JnJs
_
D
nJ.
_
D
Vu VnJ. =
_
I
2
(J Vu n)nJs
and
j(t
0
) =
_
D
(Vu
0
Vu) t
0
J..
we nd that
[[V(u u)[[
2
_ 2
_
I
2
(J Vu)nJs 2
_
I
1
(u u
0
)t
0
nJs
[[Vn[[
2
[[t
0
[[
2
+
. Vn V
0
. t
0
S
I
2
. (9.5.34)
If I
2
= 0, then the right-hand side of (9.5.34) does not have the corresponding bound-
ary integral. In this case, the best choice of n is the zero function. Then, the lower
bound is given by a simple relation:
[[V(u u)[[
2
_ 2
_
I
(u u
0
)t
0
nJs [[t
0
[[
2
+
. t
0
H
0
. (9.5.35)
The vector-valued functions in S
I
2
can be represented in the form t
0
, where is a
real number and [[ t
0
[[
+
= 1. Take
=
_
I
(u u
0
) t
0
nJs.
Then,
[[V(u u)[[ _ sup
r
0
1
0
,
[| r
0
[|
=1
_
I
(u u
0
) t
0
nJs. (9.5.36)
Remark 9.25. Consider the special case in which u = u c, c R. The left-hand
sides of (9.5.33) and (9.5.36) are equal to zero. If we take = u, then the right-hand
side of (9.5.33) is also zero. On I we haveuu
0
=uu = c. Since
_
I
t
0
nJs = 0.
we conclude that the right-hand side of (9.5.36) is also equal to zero.
Similar estimates are easily obtained for different generalizations of the method.
For example, the differential equation may be satised in subdomains C
i
only approx-
imately, i.e.,
div Vu = c
i
(.). . C
i
. (9.5.37)
where c
i
(.) is a small residual. For this case, we have the following upper bound of
the error
[|[ _ inf
V
0
u
0
[V
Vu|[
2
C
2
_
i
[c
i
(.)[
2
D
i
[Vu n [
2
I
2
_
. (9.5.38)
9.5.4 Comments
A rapidly developing group of methods is related to nonconforming nite element ap-
proximations. Here, a posteriori error estimation methods are much less developed
than for classical (conforming) nite element methods. Concerning a posteriori esti-
mates for Discontinuous Galerkin (DG) approximations of elliptic type equations, we
refer to R. Becker, P. Hansbo, and M. G. Larsson [46] and R. Bustinza, G. N. Gatica,
and B. Cockburn [83], where a modication of the residual based estimate for the en-
ergy norm of the error was suggested. In P. Castillo [101] a posteriori estimates in
the 1
2
-norm were derived for the so-called local DG method applied to an elliptic
boundary value problem. A posteriori error estimates for DG approximations were
also obtained for other classes of problems. In particular, in S. Sun and M. F. Wheeler
[346] time-dependent (transport) equations were considered and in P. Houston, I. Peru-
gia, and D. Schotzau [181] the authors investigated elliptic problems of the Maxwells
type. The paper by J. Ma and H. Brunner [224] deals with a posteriori error estimates
for DG approximations of integral equations. In A. Ern and J. Proft [127] a posteriori
estimates were obtained for DG approximations of the convection-diffusion equation.
In A. Ern, A. F. Stephensen, and M. Vohralik [128] advanced a posteriori estimates
using constants in Poincar e inequalities for elements were derived for DG approxima-
Section 9.6 Uncertain data 265
tions. Error reduction and convergence analysis of an adaptive nonconforming nite
element method was studied in C. Carstensen and R. H. W. Hoppe [94].
Finite Volume (FV) method is another nonconforming scheme, which is widely used
in modern numerical analysis. Various approaches to a posteriori error control of
FV approximations can be found in, e.g., Y. Achdou, C. Bernardi and F. Coquel [1],
A. Bergam, Z. Mghazli, and R. Verf urth [51], C. Carstensen, R. Lazarov, and S. Tomov
[98], R. Lazarov and S. Tomov [219], V. Jovanovic and C. Rohde [191], D. Kr oner and
M. Ohlberger [196], K. W. Morton and E. S uli [238], S. Nicaise [247], and M. Vohralk
[365].
A consequent discussion of an a posteriori error estimation method based on (9.5.3)
and (9.5.6) is presented in R. Lazarov, S. Repin and S. Tomar [218], where it is stud-
ied with the paradigm of the Discontinuous Galerkin method (also see [350], which
includes a discussion of numerical experiments and practical efciency of the error es-
timation method). In the context of the Finite Volume method, functional a posteriori
estimates were studied in S. Cochez-Dhondt, S. Nicaise, and S. Repin [109].
9.6 Uncertain data
In practice, the data of a problem are always dened with certain indeterminacy. This
fact should be taken into account in constructing approximation and error control
methods that must be stable with respect to small variations in the coefcients and
other data caused by indeterminacy. Also, it is necessary to compare the approxima-
tion errors and the errors caused by indeterminacy in the data. There are three major
sources of the latter errors related to indeterminacy in
(a) coefcients of a PDE;
(b) boundary (initial) conditions;
(c) the conguration of C.
In what follows, we discuss the cases (a)(c) and apply a posteriori estimates derived
in previous chapters to the analysis of such errors. As everywhere, in this chapter
we demonstrate new possibilities arising on this way, using linear diffusion problem
as a basic example. Generalizations to other problems are rather transparent and can
be done by the thoughtful reader without big difculties. However, we begin with a
concise introduction to general principles underlying error control theory in the case
of uncertain data.
9.6.1 Introduction
Take a boundary value problem in abstract form
Au = . (9.6.1)
where the operator Aand the functional are dened with indeterminacy. In this case,
instead of concrete Aand , it is only known that
A U
A
and U
I
. (9.6.2)
where U
A
and U
I
are certain (bounded) sets of possible data. In the simplest case,
we assume that all pairs
(A. ) U
A
U
I
are of equal probability. In order to guarantee the solvability of problems, we must
assume that for any A, the problem (9.6.1) is correct and has a unique solution. More-
over, we must assume that variations of A are sufciently small, so that all problems
belong to the same type (e.g., the operator remains elliptic and bounded).
Exact solutions of (9.6.1) with data satisfying (9.6.2) form a set in the respective
energy space V , which we call the set of possible solutions and denote S. Formally,
this set is dened as follows:
S :=
_
u V
Au = for some A U
A
and U
I
_
.
Let V be an approximation of the exact solution (which is an unknown element of
S). Since the data are indeterminate, the error estimation problem takes two different
forms. The rst form is dened by the quantity
e(. S) = inf
uS
[ u[
V
(9.6.3)
that measures the distance between and the set S in terms of a certain norm [ [
V
selected for this purpose. The quantity e(. S) is equal to zero if A = for some pair
(A. ) U
A
U
I
. This quantity provides the lowest possible bound of the true error
or the error in the best-case situation (cf. Figure 9.6.1). Another important quantity is
e(. S) = sup
uS
[ u[
V
. (9.6.4)
It compares with the most remote element of S and yields the error in the worst-case
situation. Obviously,
e(. S) _ e(. S) and e(. S) > 0. (9.6.5)
If the problem (9.6.1) is uniquely solvable and the data are exactly determined, then S
contains only one element, which is the exact solution u. In this case,
e(. S) = e(. S) = [ u[
V
.
and the accuracy of can be evaluated by the error norm. However, if the data are not
fully determined, then the error estimation problem is more complicated, and we need
at least two different quantities to have an idea of the quality of .
e
e
S
diam S
v
Figure 9.6.1 The set of solutions S and errors e and e.
dominates
error
Approximation
Indeterminacy
error
dominates
0 1
Figure 9.6.2 The indicator k(. S).
In addition, we introduce the quantity
k(. S) :=
e(. S) e(. S)
e(. S) e(. S)
0. 1|. (9.6.6)
which is easily computable provided that e(. S) and e(. S) (or bounds of them) are
known.
This quantity could be used as an indicator that predicts the efciency of expendi-
tures spent on decreasing approximation errors. Indeed, if k(. S) is close to zero, then
approximation errors provide the major part of the overall error and the impact of inde-
terminacy is not signicant. In the special case of fully determined data, k(. S) = 0,
and we have only the error of approximation. However, if k(. S) is close to 1, then
e(. S) ; e(. S).
which means that is quite close to S. If k(. S) = 1, then e(. S) = 0, i.e.,
S. In this case, further attempts at reducing approximation errors are obviously
meaningless. These observations are schematically depicted in Figure 9.6.2.
Hence, guaranteed and computable estimates of e(. S) and e(. S), are indeed
required in practice because they could provide important information for an efcient
organization of the computational process.
If a boundary value problem is analyzed with account of the data indeterminacy,
then one more important quantity should be determined. It is the diameter of S, which
henceforth is denoted by the symbol diam. The quantity diamS gives an insight into
the accuracy limit caused by the indeterminacy in the data. By denition,
diamS := sup
u
1
,u
2
S
[u
1
u
2
[
V
. (9.6.7)
We outline that diamS does not depend on . It characterizes a particular boundary
value problem under the indeterminacy conditions imposed. The generation of ap-
proximate solutions that have approximation errors less than diamS has no practical
sense.
In general, the exact values of e, e, and diamS cannot be found. However, func-
tional a posteriori estimates supply a method for nding their computable bounds,
which we shortly discuss below.
First, we note that a computable lower bound of diamS stems from the relation
_
_
u
i
u
}
_
_
V
_
_
_
i

}
_
_
V
[u
i

i
[
V

_
_
u
}

}
_
_
V
. (9.6.8)
where
i
and
}
are conforming approximations of two exact solutions u
i
and u
}
(which belong to the set S), respectively. The functions
i
, i = 1. . . . . m, can be
constructed by solving the boundary value problems with A
i
U
A
and
i
U
I
with
the help of some numerical method. Their accuracy is controlled by error majorants,
which furnish relevant error bounds (e.g., see (7.1.20))
[u
i

i
[
V
_ c
i
:= M
A
(
i
. ,).
where the functions ,
i
are properly selected to avoid signicant overestimation. Then
diamS _ sup
i,}=1,2,...,n
_
_
i

}
_
_
V
c
i
c
}
_
. (9.6.9)
The right hand side of (9.6.9) is computable, because it contains only approximate
solutions and known functions. The larger is m, the sharper lower bound of diamS is
given by (9.6.9) (provided that the c
i
are sufciently small).
However, such a straightforward procedure may be rather expensive. To simplify
computations, we can use another method that we discuss with the paradigm of prob-
lem (7.1.1)(7.1.3). Let A and V be certain (central) elements of the sets U
A
and U
I
, respectively. Consider the quantity
j(S)
2
:= sup
uS
[[(u u)[[
2
.
where u is the solution generated by A and (we assume that it is either known or
computed with a high accuracy conrmed by the the corresponding error majorant),
[[ [[ denotes the energy norm associated with A, and u is the exact solution of the
problem with

Aand
. It is easy to see that

diamS = sup
u
1
,u
2
S
[[(u
1
u
2
)[[
_ sup
u
1
,u
2
S
_
[[(u
1
u)[[ [[(u
2
u)[[
_
= 2j(S).
On the other hand, diamS _ j(S). Thus, the quantity j(S) characterizes the diameter
of S. We can estimate j(S) from below by (7.1.28), which reads as follows:
[[(u u)[[
2
A
_ sup
uV
0
_
(
An. n) 2(
Au. n) 2
_
. n
__
= sup
uV
0
_
(
An. n) 2((
A A)u. n) 2
_
. n
__
.
From this relation, it follows that
j(S)
2
= sup
AU
A
,
IU
I
sup
uV
0
[[(u u)[[
2
= sup
uV
0
sup
AU
A
,
IU
I
_
(
An. n)
2((
A A)u. n) 2
_
. n
_ _
. (9.6.10)
Computable lower bounds follow from (9.6.10) if V
0
, U
A
, and U
I
are replaced by
some nite-dimensional subsets.
Further simplications can be made if we assume that
zA U
A
for all z 1 j
1
. 1 j
2
|.
where j
1
and j
2
are some known nonnegative numbers, and there exists > 0 such
that
B(. ) :=
_
V [ [
[
V
_
_
U
I
.
Set

A =

A
2
:= zA. Then
[[(u u)[[
2
= z[[(u u)[[
2
.
and we nd that
j(S)
2
= sup
A,
I
[[(u u)[[
2
_
1
z
sup
I
sup
uV
0
[[(u u)[[
2
_ sup
uV
0
_
(An. n) 2
_
1
1
z
_
(Au. n)
2
z
sup
IU
I
_
. n
_ _
. (9.6.11)
Since
sup
IU
I
_
. n
_
_ sup
IB(I,)
(
. n) = [n[
V
.
we have
j(S)
2
_ sup
uV
0
_
(An. n) 2
_
1
1
z
_
(Au. n)
2
z
[n[
V
_
. (9.6.12)
where z 1 j
1
. 1 j
2
|. Taking the supremum with respect to n V
0h
V
0
and
z, we obtain a lower bound of j(S).
Finally, consider the special case in which u
0
= 0. Then we set n = u and obtain
j(S)
2
_
1
z
sup
R
_
(z
2
2z 2)[[u[[
2
2[[[u[
V
_
_
(1 z)
2
z
2
[[u[[
2
2
z
2
[1 z[[u[
V
. (9.6.13)
If = 0, then (9.6.13) implies the estimate
j(S) _ max
21-
1
,1
2
j
[1 z[
z
[[u[[. (9.6.14)
which (for this special case) can be derived in a simpler way.
9.6.2 Errors caused by indeterminacy in coefcients
Indeterminacy in coefcients of a boundary value problem that models a certain phys-
ical object is quite typical. For example, the coefcients of diffusion, elasticity con-
stants, and viscosity are usually known only approximately. Functional a posteriori
estimates provide a way for estimating these errors and comparing them with approx-
imation errors. In this section, we consider this question by using again the problem
(4.1.1)(4.1.3) as a basic example. Assume that the indeterminacy in the coefcients
of the differential equation (4.1.1) is described by xing some mean elements
0
1
o
(C: M
dd
x
). (
0
. J
0
) 1
2
(C) 1
2
(I
2
).
and by dening bounds of possible variations. In this case, the elements of the sets
U
A
and U
I
are represented by elements of the following two functional sets:
U
:=
_
1
o
(C: M
dd
x
) [ =
0
c1. 1 E
_
.
U
I
:=
_
(. J) [ =
0
1
. J = J
0
2
.
i
J
i
. i = 1. 2
_
.
where
E :=
_
1 1
o
(C: M
dd
x
)
[ [1[ [
o,D
_ 1
_
.
J
1
:=
1
1
2
(C)
[
1
[
2,D
_ 1
_
.
J
2
:=
2
1
2
(I
2
)
[
2
[
2,I
2
_ 1
_
.
Here, c and
i
are small parameters characterizing the range of indeterminacy. Hence-
forth, we assume that c < c
2
1
, where c
1
is the ellipticity constant for
0
:
c
2
1
[[
2
_
0
_ c
2
2
[[
2
. V R
d
. (9.6.15)
Since [1 [ _ [1[ [[
2
, we nd that
_
0
c[1[ [[
2
_ (c
2
1
c)[[
2
. (9.6.16)
_
0
c[1[ [[
2
_ (c
2
2
c)[[
2
(9.6.17)
and analogous estimates for the inverse matrix are
c
-2
2
[[
2
_
-1
0
_ c
-2
1
[[
2
. (9.6.18)
(c
2
2
c)
-1
[[
2
_
-1
_ (c
2
1
c)
-1
[[
2
. (9.6.19)
Let the energy norm [[V(u )[[
0
generated by the mean matrix
0
be selected as
the energy norm. Our goal is to nd estimates of the quantities
e(. S) = inf
uS
[[V( u)[[
0
and e(. S) = sup
uS
[[V( u)[[
0
.
where
S(U
. U
I
) :=
_
u V
_
D
Vu VnJ. =
_
D
nJ.
_
I
2
JnJs.
Vn V
0
. for U
and U
I
_
.
The idea behind the derivation of computable bounds of e(. S) and e(. S) is to
use the majorants and minorants discussed in Chapter 4 that explicitly depend on the
coefcients. Then, taking the supremum (inmum) over u S is reduced to a mini-
mization (maximization) problem for the coefcients of a majorant (minorant).
By (4.1.12), we know that
[[V(u )[[
2
0
_ (1 c
e
)[[V(u )[[
2
_
_
D
_
V V
-1
, , V ,
_
J.
1
c
1
_
C
T I
1
[ div ,[ C
TI
2
[J , n[
I
2
_
. (9.6.20)
where c
e
=
e
c
2
1
-e
. Analogously, by (4.1.13) we obtain
[[V(u )[[
2
0
_ (1 c
e
)[[V(u )[[
2
_ (1 c
e
)
__
D
(Vn Vn 2V Vn 2 n) J.
2
_
I
2
JnJs
_
. (9.6.21)
Below we use (9.6.20) and (9.6.21) to nd the desired error bounds.
Upper bound of e.v; S/. Let u
(A,(,T)
denote the solution of the problem generated
by , , and J. Then,
[[V( u
(,(,T)
)[[
0
_ [[V( u
(,(
0
,T
0
)
)[[
0
[[V(u
(,(
0
,T
0
)
u
(,(,T)
)[[
0
. (9.6.22)
Also, we have
_
D
Vu
(,(,T)
VnJ. =
_
D
nJ.
_
I
2
JnJs. Vn V
0
. (9.6.23)
_
D
Vu
(,(
0
,T
0
)
VnJ. =
_
D
0
nJ.
_
I
2
J
0
nJs. (9.6.24)
From (9.6.23) and (9.6.24), we conclude that
[[V(u
(,(
0
,T
0
)
u
(,(,T)
)[[
2
_
_
C
T I
1
[
0
[
D
C
TI
2
[J J
0
[
I
2
_
[V(u
(,(
0
,T
0
)
u
(,(,T)
)[. (9.6.25)
Hence,
[[V(u
(,(
0
,T
0
)
u
(,(,T)
)[[
0
_
_
1 c
e
[[V(u
(,(
0
,T
0
)
u
(,(,T)
)[[
_
c
1
c
2
1
c
_
C
T I
1
1
[
1
[
D
C
TI
2
2
[
2
[
I
2
_
. (9.6.26)
By (9.6.22) and (9.6.26), we obtain
[[V( u
(,(,T)
)[[
0
_ [[V( u
(,(
0
,T
0
)
)[[
c
1
c
2
1
c
_
C
T I
1
1
C
TI
2
2
_
. (9.6.27)
Here,
[[V( u
(,(
0
,T
0
)
)[[
0
_ (1 c
e
)[[V( u
(,(
0
,T
0
)
)[[
_ (1 c
e
)
_
[[V ,[[
-1 C
T I
1
[div ,
0
[
C
TI
2
[, n J
0
[
I
2
_
. (9.6.28)
We must estimate the term
[[V ,[[
2
-1
= (V. V) (
-1
,. ,) 2(V. ,).
Since U
A
, we have
-1
=
_
0
(I c
-1
0
1)
_
-1
= (I cT)
-1
-1
0
.
where T =
-1
0
1 and
c[T[ _ c[
-1
0
[ [1[ _ cc
-2
1
< 1.
Thus, the absolute values of all eigenvalues of the matrix cT are less than one and we
can use the representation
(I cT)
-1
= I
o
}=1
(1)
}
c
}
T
}
.
Hence,
-1
, , = (1 cT)
-1
-1
0
, ,
=
-1
0
, , cT
-1
0
, ,
o
}=2
(1)
}
c
}
T
}
-1
0
, ,.
Estimate the last term as follows:
T
}
-1
0
, , _ [T
}
-1
0
[ [,[
2
_ [
-1
0
[
}1
[1[
}
[,[
2
_ c
-2(}1)
1
[,[
2
and
_
D
-1
, , J. _
_
D
(
-1
0
, , cT
-1
0
, ,) J.
_
_
o
}=2
(1)
}
c
}
c
-2(}1)
1
_
_
[,[
2
.
In view of (9.6.1) and the relation
_
D
V V J. =
_
D
(
0
V V c1V V) J..
we nd that
[[V ,[[
2
-1
= [[
0
V ,[[
2
-1
0
c
_
D
(1V V T
-1
0
, ,) J.
1
c
2
1
[,[
2
o
}=2
_
c
c
2
1
_
}
. (9.6.29)
Since
0
is symmetric, we have
1V V
-1
0
1
-1
0
, , = 1(V
-1
0
,) (V
-1
0
,).
To nd an upper bound of e(. S), we note that
sup
TE
_
D
1 :
_
(V
-1
0
,) (V
-1
0
,)
_
J. _ 1
1;,e
.
where
1
1;,e
(. ,) := c
_
D
[V
-1
0
,[[V
-1
0
,[ J..
Dene
1
2;e
:=
_
c
c
2
1
_
2
1
(c c
2
1
)
[,[
2
.
Then,
[[V ,[[
-1 _ [[
0
V ,[[
-1
0
_
1
1;,e
1
2;e
. (9.6.30)
and by (9.6.28), we conclude that
[[V( u
(,(
0
,T
0
)
)[[
0
_
_
1 c
e
_
M
A
0
,f
0
,F
0
(. ,)
_
1
1;,e
1
2;e
_
. (9.6.31)
where
M
A
0
,f
0
,F
0
(. ,) := [[
0
V ,[[
-1
0
C
T I
1
[div ,
0
[ C
TI
2
[, n J
0
[
I
2
.
Now (9.6.27) and (9.6.31) imply the estimate
[[V( u
(,(,T)
)[[
0
_
c
1
_
c
2
1
c
M
A
0
,f
0
,F
0
(. ,) M
,e
. (9.6.32)
where
M
,e
:=
c
1
_
c
2
1
c
_
1
1;,e
1
2;e

1
_
c
2
1
c
_
C
T I
1
1
C
TI
2
2
_
.
We note that the right-hand side of (9.6.32) contains only known functions and para-
meters. It gives an upper bound of the error for any u S. Hence,
e(. S) _
c
1
_
c
2
1
c
M
A
0
,f
0
,F
0
(. ,) M
,e
. (9.6.33)
In (9.6.33), the term M
A
0
,f
0
,F
0
(. ,) is related to the approximation error of a solution
to the mean problem and M
,e
reects the impact of data indeterminacy.
Lower bound of e(. S). In view of (9.6.16), we have
[[V(u )[[
2
0
_ (1 c
e
)[[V(u )[[
2
_ (1 c
e
)
__
D
(Vn Vn 2V Vn 2 n) J. 2
_
I
2
JnJs
_
= (1 c
e
)M
2
(A
0
,f
0
,F
0
)
(. n) (1 c
e
)c
_
D
(1Vn Vn 21V Vn) J.
2(1 c
e
)
_
D
(
0
)nJ. 2(1 c
e
)
_
I
2
(J J
0
)nJs. (9.6.34)
where n is an arbitrary function in V
0
and
M
2
(A
0
,f
0
,F
0
)
(. n) :=
_
D
(
0
Vn Vn 2
0
V Vn 2
0
n) J. 2
_
I
2
J
0
nJs.
Therefore,
e
2
(. S) = inf
uS
[[V( u)[[
2
0
_ (1 c
e
) inf
1E,
1
1
1
,
2
1
2
sup
uV
0
_
M
2
(A
0
,f
0
,F
0
)
(. n)
c
_
D
(1Vn Vn 21V Vn) J.
2
_
D
1
nJ. 2
_
I
2
2
nJs
_
_ (1 c
e
) sup
uV
0
_
M
2
(A
0
,f
0
,F
0
)
(. n)
c inf
TE
_
D
(1Vn Vn 21V Vn) J. 2
1
[n[ 2
2
[n[
I
2
_
.
Here,
inf
TE
_
D
(1Vn Vn 21V Vn) J. =
_
D
(Vn 2V) Vn
J..
Now, we obtain
e
2
(. S) _ (1 c
e
) sup
uV
0
_
M
2
(A
0
,f
0
,F
0
)
(. n) S
e,
(n)
_
. (9.6.35)
where the term
S
e,
(n) := c
_
D
(Vn 2V) Vn
J. 2
1
[n[ 2
2
[n[
I
2
depends on the indeterminacy parameters.
To obtain a computable lower bound, we replace V
0
by a nite-dimensional sub-
space V
0k
and solve a nite-dimensional maximization problem
sup
uV
0k
_
M
2
(A
0
,f
0
,F
0
)
(. n) S
e,
(n)
_
.
9.6.3 Errors owing to uncertain
In practice, the domain C may be also not completely dened. For example, it may be
only known that
C
-
C C
. (9.6.36)
where C
-
and C
are given domains with Lipschitz continuous boundaries I

-
and
I
, respectively. Henceforth, for the sake of simplicity, we assume that C is a simply

connected domain.
In general, accounting of this type uncertainty imposes a more complicated task.
One way to solve it is to make a proper parametrization that maps C to an etalon do-
main
C. Then the uncertainty is transformed to the coefcients of a modied differen-

tial problem related to the new coordinate system. Then the uncertainty is transformed
to the coefcients of a modied differential problem related to the new coordinate
system. In this case, the inuence of uncertain geometry can be estimated by the
above-discussed methods.
Below we consider another (simpler) way the applicability of which is, however,
restricted to the case of Dirichlet boundary conditions.
Consider the problem
div Vu = 0 in C
with homogeneous Dirichlet boundary conditions. We assume that 1
2
(C
) and
the matrix is dened and positive denite in C
. Let us denote the exact solutions

of the boundary value problems in C
and C
-
by u
and u
-
, respectively.
A function

H
1
(C) can be extended to

H
1
(C
) if we set = 0 in C
\ C.
Similarly, u
-
can be viewed as a function in

H
1
(C
) or

H
1
(C) with zero values
outside C
-
.
By analogy with the estimate (2.3.1), we deduce the relation
1
2
[[V(u
u)[[
2
D
= J
D
(u) J
D
(u
). (9.6.37)
where
J
D
() :=
_
D
(V V ) J..
Since
J
D
(u) = J
D
(u) _ J
D
(u
-
)
we nd that
1
2
[[V(u
u)[[
2
D
_ J
D
(u
-
) J
D
(u
) = J
D
(u
-
) J
D
(u
)
=
1
2
[[V(u
-
u
)[[
2
D
.
On the other hand,
1
2
[[V(u
-
u)[[
2
D
= J
D
(u
-
) J
D
(u) = J
D
(u
-
) J
D
(u)
_ J
D
(u
-
) J
D
(u
) =
1
2
[[V(u
-
u
)[[
2
D
.
[[V(u
u)[[
D
_ [[V(u
u
-
)[[
D
. (9.6.38)
This estimate leads to an upper bound of
diamS := sup
u
1
,u
2
[[V(u
1
u
2
)[[
D
.
where u
1
and u
2
are the exact solutions related to C
1
and C
2
, respectively (it is
assumed that C
-
C
i
C
for i = 1. 2). Then,

[[V(u
1
u
2
)[[
D
_ [[V(u
1
u
-
)[[
D
[[V(u
-
u
)[[
D
[[V(u
u
2
)[[
D
_ 3[[V(u
-
u
)[[
D
and we nd that
[[V(u
-
u
)[[
D
_ diamS _ 3[[V(u
-
u
)[[
D
. (9.6.39)
Let
-

H
1
(C
-
) be an approximation of u
-
extended by zero to C
. Then
1
2
[[V(u
-
u
)[[
2
D
= J
D
(u
-
) J
D
(u
)
_ J
D
(
-
) J
D
(u
) =
1
2
[[V(
-
u
)[[
2
D
and by (9.6.39) we conclude that

diamS _ 3[[V(
-
u
)[[
D
. (9.6.40)
Apply the majorant to the right-hand side of (9.6.40). We have
[[V(u

-
)[[
D
_ [[V
-
,[[
+,D
C
T D
[div , [
D
=: M
:
(
-
. ,). (9.6.41)
where , H(C
. div).
Remark 9.26. By the same arguments, we can obtain computable estimates of the
error arising from the uncertainty of the Dirichlet part of the boundary in many other
problems, for which the principal relation (9.6.37) holds.
9.6.4 Comments
In the fully reliable mathematical modeling, the errors caused by uncertainties in the
problem data must be measured, as well as the approximation errors. However, this
question has only recently started receiving serious attention. For example, effects
associated with uncertainty in the boundary conditions was considered in the papers
by I. Babu ska and J. Chleboun [23, 24]. Also, we refer to the papers by I. Babu` ska,
F. Nobile, and R. Tempone [29] and J. T. Oden, I. Babu` ska, F. Nobile, Y. Feng, and
R. Tempone [250]. A study of mathematical models generated by uncertain input
data and the worst scenario method is the main subject of the book by I. Hlav a` cek,
J. Chleboun, and I. Babu` ska [175].
In [290] (see also P. Neittaanm aki and S. Repin [244]), a posteriori estimates dis-
cussed in Chapter 4 were used to evaluate errors induced by indeterminacy in coef-
cients of elliptic problems. Subsequent investigations of this problem and results
of numerical tests are presented in O. Mali and S. Repin [225]. Recently, estimates
of diamS for linear diffusion problems has been derived by a different method (see
O. Mali and S. Repin [226, 227]).
It should be noted that the inuence of various uncertainties (e.g., roundoff errors)
affecting the accuracy of computations can also be investigated in terms of the so-
called interval analysis. In it, the operations are performed for intervals instead of
numbers. This theory is beyond the scope of the present book. The reader inter-
ested in it is referred to, e.g., G. Alefeld and J. Herzberger [11], B. S. Dobronets and
V. V. Shadurov [115], Y. V. Matijasevich [228], R. E. Moore [237] and to the literature
cited therein.
Section 9.7 Error estimates in terms of functionals and nonenergy norms 279
9.7 Error estimates in terms of functionals and
nonenergy norms
A posteriori estimates are intended to present a computable measure of the difference
between an exact solution of a certain boundary value problem and an approximate
one obtained by some numerical technology. In most cases, a posteriori estimates are
derived in global (e.g., energy) norms and, therefore, justify the overall accuracy of an
approximation considered. However, such information may be not sufcient because
numerical analysts are often interested in local errors and in errors expressed in terms
of special problem-oriented functionals. In other words, it may be desirable to have
estimates of the type
(u ) _ M
(. ,. D).
where : V R
is a given functional and R
denotes the set of nonnegative real

numbers. In this section, we derive computable estimates in terms of local norms and
other quantities, which may be used to obtain a comprehensive presentation on the
error structure.
9.7.1 General framework
It is often required to measure the accuracy of approximate solutions in terms of func-
tionals other than the energy norm. If the basic error control problem stated at the
beginning of Section 4 is solved and we have certain M and M satisfying (8.5.5),
then on can to construct computable and sharp error estimates for any error functional
subject to the energy norm.
Assume that we are interested in the value of ( u), where : V R is a
given functional such that
[(
1
2
)[ _ j
1
([(
1
)[ [(
2
)[) . V
1
.
2
V. (9.7.1)
[(
1
)[ _ j
2
[[
1
[[. (9.7.2)
where j
1
_ 1 and j
2
> 0.
Set
1
= and
2
= u , where V
0
. From (9.7.1) it follows that
[(u )[ _ j
1
[()[ j
1
[(u )[
_ j
1
[()[ j
1
j
2
[[u [[. (9.7.3)
Consider u
0
V
0
as an approximation of u and apply the majorant to the
second term. We have
[(u )[ _ j
1
[()[ j
1
j
2
M( . D). (9.7.4)
where D stands for the problem data and free functions.
Set
1
= u and
1
2
= . Then,
[(u )[ _
1
j
1
() j
2
M( . D). (9.7.5)
Since M(u. D) = 0, we observe that, by choosing = u the right-hand side of
(9.7.4) is equal to [(u )[. If j
1
= 1, then the lower bound (9.7.5) is also sharp.
Also, from (9.7.4) we observe that an upper bound of [(u )[ can be obtained if
M is additionally minimized on the kernel of , i.e.,
[(u )[ _ j
1
j
2
inf
Ker
_
M( . D)
_
. (9.7.6)
9.7.2 Estimates in local norms
Consider the diffusion problem (4.1.1)(4.1.3). Let o be a subdomain of C with Lip-
schitz continuous boundary do. Set
(u ) = [[V(u )[[
o
:=
__
o
V(u ) (u ) J.
_
1{2
.
In this case, j
1
= j
2
= 1 and we have the following estimates:
[[V(u )[[
o
_ [[V[[
o
M
DF
( . ,) =: M
DFo
(. . ,). (9.7.7)
[[V(u )[[
o
_ [[V[[
o
M
DF
( . ,) =: M
DFo
(. . ,). (9.7.8)
where V
0
. It is easy to see that the estimates (9.7.7) and (9.7.8) have no gaps (if
, = Vu and = u , they hold as equalities).
In particular, from (9.7.7) it follows that an upper bound of [[V(u )[[
o
is ob-
tained if M
DF
( . ,) is minimized over functions such that V = 0 in o (see
[291, 295, 294]).
Square both parts of (9.7.7) and use (4.1.14). Then, we obtain
[[V(u )[[
2
o
_ (1 ;)[[V[[
2
o

(1 )(1 ;)
;
[[V( ) ,[[
2
+
(1 )(1 ;)
;
C
_
[div , [
2
[J , n[
2
_
.
(9.7.9)
where and ; are positive numbers, , H(C. div), and is an arbitrary function
from V
0o
. If ,, , and ; are dened, then optimization of the upper bound is reduced
to minimization of the quadratic functional
J
o
() := k[[V[[
2
[[V[[
2
o
2k
_
D
V (V ,) J. (9.7.10)
on the set V
0
. In (9.7.10), k =
1
;
. This problem can be approximately solved if J
o
is minimized on a sufciently reach subspace V
0h
V
0
.
It should be noted that the last term on the right-hand side of (9.7.9) does not depend
on , so that the quality of , states an accuracy limit for the local norm. Therefore,
an efcient evaluation of [[V(u )[[
2
o
requires not only nding but also needs a
vector-valued function , such that [div , [ is sufciently small. However, such
difculties are quite predictable and it is natural to await that getting guaranteed and
accurate estimates for local norms (which provide a more detailed information on the
quality of an approximate solution) should be more expensive than for the global en-
ergy norm.
Remark 9.27. If , is equilibrated, then we can set = n, where n is obtained by
maximization of the minorant M
DF
(. n). The same choice is valid for any error func-
tional that satises (9.7.1)(9.7.2). Thus, we conclude that if the global upper and
lower bounds computed by the majorant and minorant are sufciently close, then the
corresponding n provides a good image of u and, therefore, it can also be used for
estimation of local errors and other quantities of interest without noticeable increasing
in the computational cost. In fact, these observations mean that the expenditures spent
for a careful control of global errors can be utilized for other purposes.
9.7.3 Estimates in terms of linear functionals
In Section 2.6.4, we discussed some error estimation methods used to evaluate the
quantity
E
I
(u ) :=[ (. u ) [.
where V
+
0
is a given linear functional and (. ) denotes the duality pairing of
the spaces V
+
0
and V
0
. The functional is usually taken in such a way that its value
characterizes some especially important properties of a solution. It is necessary to note
that E
I
is only a seminorm on V , so that small values of this indicator do not guarantee
small values of the error. Indeed, if u ,= , but e := u Ker , then E
I
(u) = 0.
Thus, indicators of this type are meaningful only if e contains a signicant component
orthogonal to Ker . Despite of this fact, error indicators of such a type are often used
in error control methods in combination with global error estimates. Guaranteed upper
bounds of [(. u )[ can be derived in several different ways discussed below.
1. The simplest way is to use the inequality
[(. u )[ _ [[[ [[[ [[[ [[[ [[V(u )[[. (9.7.11)
[(. u )[ _ [[[ [[[ [[[ [[[ M
DF
(. ,). (9.7.12)
where
[[[ [[[ [[[ [[[ := sup
uV
0
[(. n)[
[[Vn[[
.
However, the right-hand side of (9.7.11) may essentially overestimate the left-hand
one and, in general, the upper bound given by (9.7.12) may be very coarse.
2. Another upper bound follows from (2.6.34). Since
1
2
(u. u
h
. u
I
. u
Ir
) _ [[V(u
h
u)[[[[V(u
Ir
u
I
)[[.
we nd that
[(. u u
h
)[ _
_
D
( u
Ir
Vu
h
Vu
Ir
) J.
M
f,DF
(u
h
. ,
1
) M
I,DF
(u
Ir
. ,
2
)
. (9.7.13)
where ,
1
and ,
2
are two (different) functions in H(C. div). By M
f,DF
and M
I,DF
we
denote the majorants related to the diffusion problems with source terms and , re-
spectively. This upper bound is sharper than in (9.7.12). Regrettably, its computation is
rather expensive, because, in addition to u
Ir
, it requires nding ,
1
and ,
2
sufciently
close to exact uxes of solutions to the original and adjoint problems, respectively.
3. Another method is based upon the relation
[(. u )[ = [(. u )[. V V
0I
(C). (9.7.14)
where
V
0I
(C) := { V
0
(C) [ (. ) = 0].
By (9.7.14) we obtain
[(. u )[ _ [[[ [[[ [[[ [[[ inf
V
0I
[[u [[. (9.7.15)
We show that (9.7.15) holds as equality. Indeed, let u
I
V
0
be a function such that
_
D
Vu
I
VnJ. = (. n). Vn V
0
. (9.7.16)
Since is a bounded linear functional (on V
0
), the solution u
I
exists, is unique, and
satises the relation [[Vu
I
[[ = [[[ [[[ [[[ [[[ .
Set
= u
(. u )
[[[ [[[ [[[ [[[
2
u
I
.
Then,
(.

) = (. u )
_
1
(. u
I
)
[[[ [[[ [[[ [[[
2
_
= 0.
so that

V
0I
. Hence,
inf
V
0I
[[u [[ _ [[u

[[
= [(. u )[
[[Vu
I
[[
[[[ [[[ [[[ [[[
2
=
1
[[[ [[[ [[[ [[[
[(. u )[ .
and we see that the left-hand side of (9.7.15) is equal to the right-hand one.
Estimate the right-hand side of (9.7.15) by the error majorant. Then, we nd that
[(. u )[ _ [[[ [[[ [[[ [[[ M
DF
( . ,). (9.7.17)
[(. u )[ _ [[[ [[[ [[[ [[[ inf
,!
T
2
(:,div)
1
0I
M
DF
( . ,) _ [[[ [[[ [[[ [[[ M
DF
(

. Vu)
= [[[ [[[ [[[ [[[ [[V(u

)[[
+
= [[[ [[[ [[[ [[[ [[u

[[ = [(. u )[.
Thus, additional minimization of the majorant M
DF
over the set V
0I
is sufcient to get
a sharp upper bound of [(. u )[.
4. An estimate of the quantity [(. u )[ follows from (9.7.4) if we set
(u ) = (u ).
In this case, j
1
= 1 and j
2
= [[[ [[[ [[[ [[[ . Therefore, we have the estimate
[(. u )[ _ [(. )[ [[[ [[[ [[[ [[[ M
DF
( . ,). (9.7.18)
where , and are arbitrary functions in H
I
2
(C. div) and V
0
, respectively. Usually,
the value of [[[ [[[ [[[ [[[ is not difcult to estimate. For example, if
(. u ) =
_
D
z(u ) J.. where z 1
2
(C). (9.7.19)
then [[[ [[[ [[[ [[[ _ [z[
C
1:
c
1
. If
(. u ) =
_
D
t V(u ) J.. (9.7.20)
where t 1
2
(C. R
d
) is a given vector-valued function, then
_
D
t VnJ.
_ [[t[[
+
[[Vn[[
and [[[ [[[ [[[ [[[ _ [[t[[
+
.
In another frequently encountered case, the functional is dened by the integral
(. u ) =
_
Jo
; (u ) Js. (9.7.21)
where ; 1
2
(do) is a given weight function. In this case, the value of [[[ [[[ [[[ [[[ is esti-
mated by the constant in the trace inequality associated with do.
9.7.4 Estimates based on the Poincar e inequality
In Sections 3.5.3, 4.1.3, and 4.2 we used the Poincar e inequality for deriving error
estimates in terms of energy norms. Below we shortly discuss one other application of
this inequality.
Let o be again a connected subset of C with a Lipschitz boundary. In view of the
Poincar e inequality, we have
[u [
2
2,o
_ C
1o
_
[V(u )[
2
2,o

__
o
(u ) J.
_
2
_
. (9.7.22)
where u is the exact solution of (4.1.1)(4.1.3) and u
0
V
0
is an approximation.
For some subdomains (simplexes, circles, squares, cubes, etc.), the constants C
1o
can
be evaluated analytically. For example, if
o =
I
:= {. [ 0 < .
i
< l
i
. i = 1. 2. . . . . J].
then (see (1.4.31))
[u [
2
l
_
1
[
I
[
__
l
(u ) J.
_
2
J
2
max
i
{l
2
i
] [V(u )[
2
l
. (9.7.23)
Thus, if the error in the local energy norm is evaluated, then adding the error in terms
of the linear functional
_
l
(u ) J. provides an upper bound of the local error in
1
2
. Also, (9.7.23) implies a special characteristic of the error. Introduce the quantity
Os
o
(u ) := inf
R
[u [
o
.
which characterizes the oscillatory part of the error u related to o. Obviously
Os
o
(u ) _ [u [
o
.
where = fu g
o
. Therefore, (9.7.23) implies the estimate
Os
l
(u ) _
_
J
2
max
i
{l
i
] [V(u )[
l
. (9.7.24)
Hence, by (9.7.7) we can also evaluate certain part of the local error.
9.7.5 Estimates based on multiplicative inequalities
To derive estimates in 1
]
norms, we can use well-known embedding inequalities
analogous to (9.7.22). However, these estimates involve new constants (in place of
C
1o
). The multiplicative inequality (e.g., see O. A. Ladyzhenskaya and N. N. Uralt-
seva [214]) opens a way for avoiding the necessity of computing new constants. It has
the form
[n[
],o
_ C
],
[Vn[
2,o
[n[
1-
2,o
. Vn

H
1
(o). > 2. (9.7.25)
where
=
_
1
2

1
]
_ _
1
2

d-2
2d
_
-1
.
If J = 2, then we can take 2. o), and the constant is dened by the relation
C
],
= max
]
2
. 2
_
. If J = 3, then 2. 6| and C
],
= 4
.
Let W
1,o
(o) be a function such that
0 _ (.) _ 1. supp o. and (.) = 1 in o
t
o.
[(u )[
],o
_ C
],
[V((u ))[
2,o
[(u )[
1-
2,o
.
Since
[V((u ))[
2
2,o
_ (1 ) [ [V(u )[[
2
2,o

1
[[V[ (u )[
2
2,o
.
where is an arbitrary positive number, we obtain
[u [
],o
/ _ [(u )[
],o
_ C
],
_
(1 ) [V(u )[
2
2,o
_
_
_[V[
2
_
_
_
o,o
[u [
2
2,o
_
{2
[u [
1-
2,o
. (9.7.26)
Thus, having estimates of the local errors in terms of 1
2
-norms, we obtain an upper
bound of the error in any 1
]
-norm without computing new global constants.
9.7.6 Estimates based on the maximum principle
Pointwise estimates of approximation errors can be derived with the help of known
estimates for partial differential equations, which follow from the maximum prin-
ciple (e.g., see D. Gilbarg and N. S. Trudinger [151] and O. A. Ladyzhenskaya and
N. N. Uraltseva [214]). In the simplest case, it reads as follows:
Theorem 9.28. Let A be a uniformly elliptic operator of the second order, which
is dened in a bounded domain C with Lipschitz boundary I. Assume that
u
C
2
(C)
_
C
0
(C) and
Au
_ 0. (9.7.27)
Then, the function u
attains its maximum on I, i.e.,

sup
D
u
= sup
I
u
. (9.7.28)
This principle holds for many elliptic operators. In particular, it holds for the opera-
tor := div VbVc provided that c _ 0, the coefcients are bounded, and
[b[ is small with respect to the ellipticity constant c
1
(so that the ellipticity condition
is satised).
The function u
in Theorem 9.28 is called a sub-solution associated with A. A

function u
-
C
2
(C)
_
C
0
(C) that satises the condition Au
-
_ 0 is called a
super-solution. If A is the operator ^, then sub- and super-solutions are presented by
sub- and super-harmonic functions, respectively.
Consider the problem u = in C with homogeneous Dirichlet boundary condi-
tions. Assume that we have an approximate solution u C(
C) C
2
(C) that satises
the condition
u _ in C.
Then (u u) _ 0 and by the maximum principle we conclude that
sup
D
(u u) _ sup
I
(u u) = sup
I
u. (9.7.29)
The estimate (9.7.29) shows that the (pointwise) error is bounded by the quantity
sup
I
u. If u satises the relation u _ , then we nd that inf
D
(u u) _ inf
I
u.
Also, one can apply a more sophisticated estimate that based upon the following
theorem (e.g., see [151]):
Theorem 9.29. Let u _ and u C(
C) C
2
(C). Then
sup
D
u _ sup
I
(u)
C
jsup
D
[ ( )
[,z. (9.7.30)
If u = , then
sup
D
[u[ _ sup
I
[ (u)
C
[ jsup
D
[ [,z. (9.7.31)
where z(.) is the lowest eigenvalue of (.) and the constant j depends only on
diamC and on the ratio = sup [b[,z. In particular, if C lies between two parallel
surfaces and J is the distance between them, then j = exp{( 1)J] 1.
Theorem 9.29 implies estimates of the difference between u and any approximation
u C(
C) C
2
(C). Set e = u u. We have
e = u u = u =:

.
sup
D
[e[ _ sup
I
[ (u u)
C
[ jsup
D
[
[,z. (9.7.32)
where we can estimate z by (4.1.4). If the residual

= is of constant sign
in C, then we can apply (9.7.30). The maximum principle and the estimate (9.7.30)
can be extended to a wider class of functions. However, these functions must have
second generalized derivatives summable in any subdomain of C, so that the regularity
conditions that we must impose on u are rather strong. This fact may lead to certain
technical difculties in practical applications.
9.7.7 Estimates in weighted norms
Estimates in weighted norms can be useful if the signicance of errors in different parts
of C is different. We consider a way of deriving such estimates with the paradigm of
the problem
û = 0 in C. (9.7.33)
u = u
0
on I. (9.7.34)
Let be a smooth (or piecewise smooth) positive weight function. We wish to measure
the error in terms of the norm
[V(u )[
2
j
:=
_
D
[V(u )[
2
J..
From the corresponding integral identity, we nd that
_
D
V(u ) V(n) J. =
_
D
( n V V(n)) J.. Vn V
0
. (9.7.35)
where V
0
=

H
1
(C) and u
0
V
0
is an approximate solution. Set n = u and
rewrite (9.7.35) in the form
_
D
[V(u )[
2
J.
1
2
_
D
V V((u )
2
) J.
=
_
D
(( V V)(u ) V V(u )) J.. (9.7.36)
By the identity
_
D
(, V(u ) (u ), V (u ) div ,) J. = 0.
we obtain
_
D
[V(u )[
2
J.
1
2
_
D
V V((u )
2
) J.
=
_
D
_
( div ,) V (, V))(u )
(, V) V(u )
_
J.. (9.7.37)
If is sufciently regular, then we integrate by parts in the second term and deduce
the estimate
_
D
[V(u )[
2
J.
1
2
_
D
^ (u )
2
J.
=
_
D
_
( div ,) V (, V))(u )
(, V) V(u )
_
J.. (9.7.38)
If is a harmonic (or superharmonic) function, then (9.7.38) implies the estimate
_
D
[V(u )[
2
J.
_
_
D
(( div ,) V (, V))(u ) (, V) V(u )) J..
It is easy to see that a constant C
D
exists such that
[n[ _ C
D
[Vn[
j
. Vn V
0
.
In particular, one can set C
D
= C
T D
(
0
)
-1{2
, where
0
= min
xD
{(.)]. Since
_
D
(( div ,) V (, V))(u ) J.
_ C
D
[( div ,) V (, V)[[V(u )[
j
.
we arrive at the estimate
[V(u)[
j
_ C
D
[( div ,) V (, V)[ [
1{2
(, V)[. (9.7.39)
Remark 9.30. Let [(.) C
2
(C) be a superharmonic function and b R
d
be a
vector independent of .. Set (.) = [(.) b . c, where
c > [ min
xD
{[(.) b .][.
Since ^ = ^[, the function is a nonnegative superharmonic function, which can
be used in (9.7.39).
We can rearrange (9.7.37) without assuming that is superharmonic. For this rea-
son, we shift the term with V((u )
2
to the right-hand side of (9.7.38). Let (^)
C
denote the positive part of ^. Note that
1
2
_
D
V V((u )
2
) J.
=
1
2
_
D
^ (u )
2
J.
_
1
2
_
D
(^)
C
(u )
2
J.
_ [ (^)
C
[
o
[u [
2
_ [ (^)
C
[
o
C
2
D
[V(u )[
j
.
Hence, (9.7.37) implies the estimate
_
1 C
2
D
[ (^)
C
[
o
_
[V(u )[
j
_ C
D
[( div ,) V (, V)[ [
1{2
(, V)[. (9.7.40)
which has a meaning only if [ (^)
C
[
o
is sufciently small.
Another estimate follows from the relation
_
D
[V(u )[
2
J. =
_
D
(( div ,) V (, V)) (u ) J.
_
D
(, V) V(u ) J.
1
2
_
D
(q V) V((u )
2
) J..
where q is an arbitrary function in 1
o
(C. R
d
) S(C). Since
1
2
_
D
(q V) V((u )
2
) J.
_ [q V[
o
_
D
[u [ [V(u )[ J.
_ [q V[
o
[V(u )[
j
__
D
-1
[u [
2
J.
_
1{2
_
C
T D
0
[q V[
o
[V(u )[
2
j
we nd that
_
1
C
T D
0
[q V[
o
_
[V(u )[
j
_ C
T D
[
1{2
( div ,) V (, V))[ [
1{2
(, V)[. (9.7.41)
Certainly, this estimate makes a sense only if the quantity in parentheses is positive. If
is a harmonic function, then its gradient is a divergence free vector-valued function
and the term [q V[
o
vanishes.
Finally, we note that a posteriori estimates in weighted norms can be derived for
other elliptic problems, using transformations of integral identities quite similar to
those we applied to the problem (9.7.33)(9.7.34) (see [302]).
Bibliography
[1] Y. Achdou, C. Bernardi, and F. Coquel, A priori and a posteriori analysis of nite volume dis-
cretizations of Darcys equations, Numer. Math. 96 (2003), pp. 1742.
[2] A. Agouzal, On the saturation assumption and hierarchical a posteriori error estimator, Com-
put. Meth. Appl. Math. 1 (2001), pp. 125131.
[3] M. Ainsworth, A posteriori error estimation for fully discrete hierarchic models of elliptic bound-
ary value problems in thin domains, Numer. Math. 80 (1998), pp. 325362.
[4] , Robust a posteriori error estimation for nonconforming nite element approximation,
SIAM J. Numer. Anal. 42 (2005), pp. 23202341.
[5] , A posteriori error estimation for discontinuous Galerkin nite element approximation,
SIAM J. Numer. Anal. 45 (2007), 4, pp. 17771798.
[6] M. Ainsworth and J. T. Oden, A posteriori error estimation in the nite element method, Nu-
mer. Math. 60 (1992), pp. 429463.
[7] , A unied approach to a posteriori error estimation using element residual methods, Nu-
mer. Math. 65 (1993), pp. 2350.
[8] , A posteriori error estimation in nite element analysis, Wiley and Sons, New York, 2000.
[9] M. Ainsworth, J. T. Oden, and C. Y. Lee, Local a posteriori error estimators for variational in-
equalities, Numer. Methods for PDE 9 (1993), pp. 2333.
[10] M. Ainsworth, J. Z. Zhu, A. W. Craig, and O. C. Zienkiewicz, Analysis of the ZienkiewiczZhu a
posteriori error estimator in the nite element method, Int. J. Numer. Methods Engrg. 28 (1989),
pp. 21612174.
[11] G. Alefeld and J. Herzberger, Introduction to interval computations, Academic Press, New York,
1983.
[12] A. Alonso, Error estimators for a mixed method, Numer. Math. 74 (1996), pp. 385395.
[13] M. Amara, M. Ben Younes, and C. Bernardi, Error indicators for the NavierStokes equations in
stream function and vorticity formulation, Numer. Math. 80 (1998), pp. 181206.
[14] A. M. Arthurth, Complementary variational principles, Clarendon Press, Oxford, 1980.
[15] S. N. Atluri (ed.), Computational methods in the mechanics of fracture, North-Holland, NewYork,
1986.
[16] J. P. Aubin, Behavior of the error of the approximate solutions of boundary-value problems for
linear elliptic operators by Galerkins and nite difference methods, Ann, Scuola Norm. Sup. Pisa
21 (1967), pp. 599637.
[17] G. Auchmuty, A posteriori error estimates for linear equations, Numer. Math. 61 (1992), pp. 16.
[18] A. Babin, A. Mahalov, and B. Nicolaenko, Global splitting, integrability and regularity of 3D
Euler and NavierStokes equations for uniformly rotating uids, European. J. Mech. 15 (1996),
pp. 291300.
[19] , Global regularity of 3D rotating NavierStokes equations for resonant domains, Indiana
Univ. Math. J. 48 (1999), pp. 11331176.
292 Bibliography
[20] I. Babu` ska, Courant element: before and after, Fifty Years of Courant Element, edited by
M. K`ri` zek, P. Neittaanm aki, and R. Stenberg, Marcel Dekker, 1994, pp. 3751.
[21] , The nite element method with Lagrangian multipliers, Numer. Math. 20 (1973), pp. 179
192.
[22] I. Babu` ska and A. K. Aziz, Surway lectures on the mathematical foundations of the nite element
method, The mathematical formulations of the nite element method with applications to partial
differential equations, Academic Press, New York, 1972, pp. 5359.
[23] I. Babu` ska and J. Chleboun, Effects of uncertainties in the domain on the solution of Neumann
boundary-value problems in two spatial dimensions, Math. Comput. 71 (2001), pp. 13391370.
[24] , Effects of uncertainties in the domain on the solution of Drichlet boundary-value prob-
lems, Numer. Math. 93 (2003), pp. 583610.
[25] I. Babu` ska, R. Duran and R. Rodriguez, Analysis of the efciency of a posteriori error estimator
for linear triangular elements, SIAM J. Numer. Anal. 29 (1992), pp. 947964.
[26] I. Babu` ska and G. N. Gatica, On the mixed nite element method with Lagrange multipliers, Nu-
mer. Meth. Partial Diff. Eq. 19 (2003), pp. 192210.
[27] I. Babu` ska, F. Ihlenburg, T. Strouboulis, and S. K. Gangaraj, A posteriori error estimation for nite
element solutions on Helmholtz equation Part II: estimation of the pollution error, Int. J. Nu-
mer. Meth. Engrg. 40 (1997), pp. 38833900.
[28] I. Babu` ska, I. Lee, and C. Schwab, On the a posteriori estimation of the modeling error for the
heat conduction in a plate and its use for adaptive hierarchical modeling, Appl. Numer. Math.
14 (1994), pp. 521.
[29] I. Babu` ska, F. Nobile, and R. Tempone, Worst case scenarion analysis for elliptic problems with
uncertainty, Numer. Math. 101 (2005), pp. 185219.
[30] I. Babu` ska and J. E. Osborn, Can a nite element method perform arbitrarily badly?, Math. Com-
put. 69 (2000), pp. 443462.
[31] I. Babu` ska and W. C. Rheinboldt, A-posteriori error estimates for the nite element method, In-
ternat. J. Numer. Meth. Engrg. 12 (1978), pp. 15971615.
[32] , Error estimates for adaptive nite element computations, SIAM J. Numer. Anal. 15
(1978), pp. 736754.
[33] I. Babu` ska and R. Rodriguez, The problem of the selection of an a posteriori error indicator based
on smoothing techniques, Internat. J. Numer. Meth. Engrg. 36 (1993), pp. 539567.
[34] I. Babu` ska and C. Schwab, A posteriori error estimation for hierarchic models of elliptic bound-
ary value problems on thin domains, SIAM J. Numer. Anal. 33 (1996), pp. 221246.
[35] I. Babu` ska and T. Strouboulis, The nite element method and its reliability, Oxford University
Press, New York, 2001.
[36] I. Babu` ska, T. Strouboulis, S. K. Gangaraj, and C. S. Upadhyay, Pollution-error in the h-version
of the FEM and the local quality of recovered derivatives, Comput. Methods Appl. Mech. Engrg.
140 (1997), pp. 137.
[37] I. Babu` ska, T. Strouboulis, A. Mathur, and C. S. Upadhyay, Pollution error in the h-version of
the FEM and the local quality of a posteriori error estimators, Finite Elem. Anal. and Des. 17
(1994), pp. 273321.
Bibliography 293
[38] I. Babu` ska, T. Strouboulis, and C. S. Upadhyay, A model study of the quality of a posteriori error
estimators for linear elliptic problems. Error estimation in the interior of patchwise uniform grids
of triangles, Comp. Meth. Appl. Mech. Engrg. 114 (1994), pp. 307378.
[39] I. Babu` ska, T. Strouboulis, C. S. Upadhyay, and S. K. Gangaraj, A posteriori estimation and adap-
tive control of the pollution-error in the h-version of the FEM, Int. J. Numer. Meth. Engrg. 38
(1995), pp. 42074235.
[40] W. Bangerth and R. Rannacher, Adaptive nite element methods for differential equations,
Birkh auser-Verlag, Berlin, 2003.
[41] R. E. Bank and R. K. Smith, Mesh smoothing using a posteriori error estimates, SIAM J. Nu-
mer. Anal. 34 (1997), pp. 979997.
[42] R. E. Bank and A. Weiser, Some a posteriori error estimators for elliptic partial differential equa-
tions, Math. Comput. 44 (1985), pp. 283301.
[43] R. E. Bank and B. D. Welfert, A posteriori error estimates for the Stokes problem, SIAM J. Nu-
mer. Anal. 28 (1991), pp. 591623.
[44] S. Bartels and C. Carstensen, Averaging techniques yield reliable a posteriori nite element error
control for obstacle problems, Numer. Math. 99 (2004), pp. 225249.
[45] R. Beck, R. Hiptmair, R. Hoppe, and B. Wohlmuth, Residual based a posteriori error estimators
for eddy current computation, RAIRO Math. Model. Numer. Anal. 34 (2000), 159182.
[46] R. Becker, P. Hansbo, and M. G. Larsson, Energy norm a posteriori error estimation for discon-
tinuous Galerkin methods, Comput. Methods Appl. Mech. Engrg. 192 (2003), pp. 723733.
[47] R. Becker, H. Kapp, and R. Rannacher, Adaptive nite element methods for optimal control of
partial differential equations: Basic concept, SIAM J. Control Optim. 39 (2000), pp. 113132.
[48] R. Becker and R. Rannacher, A feed-back approach to error control in nite element methods:
Basic approach and examples, EastWest J. Numer. Math. 4 (1996), pp. 237264.
[49] , An optimal control approach to error estimation and mesh adaptation, Acta Numerica
2000, edited by A. Iserles, Cambridge University Press, 2001, pp. 1102.
[50] R. Becker and B. Vexler, A posteriori error estimation for nite element discretization of param-
eter identication problems, Numer. Math. 96 (2004), pp. 435459.
[51] A. Bergam, Z. Mghazli, and R. Verf urth, A posteriori estimators for the nite volume discretiza-
tion of an elliptic problem, Numer. Math. 95 (2003), pp. 599624.
[52] M. Bergounioux, M. Haddou, M. Hinterm uller, and K. Kunisch, A comparison of a Moreau
Yosida-Based Active Set Strategy and Interior Point Methods for Constrained Optimal Control
Problems, SIAM J. Optim. 11 (2000), pp. 495521.
[53] A. Berm udez, R. Dur an, and R. Rodrguez, Finite element analysis of compressible and incom-
pressible uid-solid systems, Math. Comput. 67 (1998), pp. 111136.
[54] C. Bernardi and V. Girault, A local regularization operator for triangular and quadrilateral nite
elements, SIAM J. Numer. Anal. 35 (1998) pp. 18931916.
[55] C. Bernardi and R. Verf urth, A posteriori error analysis of the fully discretized time-dependent
Stokes equations, ECIAM:M2NL 38 (2004), pp. 437455.
[56] M. Bildhauer and M. Fuchs, Error estimates for obstacle problems of higher order, Zap. Nauchn,
Sem. S.-Peterburg, Otdel. Mat. Inst. Steklov. (POMI) 326 (2007), pp. 518.
294 Bibliography
[57] M. Bildhauer, M. Fuchs, and S. Repin, A posteriori error estimates for stationary slow ows of
power-law uids, J. Non-Newtonian Fluid Mech. 142 (2007), pp. 112122.
[58] , A functional type a posteriori error analysis for RambergOsgood Model,
Z. Angew. Math. Mech. (ZAMM) 87 (2007), pp. 860876.
[59] , The elasto-plastic torsion problem: a posteriori estimates for approximate solutions, Uni-
versit at des Saarlandes, 2007, preprint.
[60] , Duality based a posteriori error estimates for higher order variational inequalities with
power growth functionals, Annales Academi Scentiarum Fennic, Mathematics 33 (2008),
pp. 475490.
[61] M. Bildhauer and S. Repin, Estimates for the deviation from exact solutions of variational prob-
lems with power growth functionals, Zapiski Nauchn, Semin V. A. Steklov Mathematical Institute
in St.-Petersburg (POMI) 336 (2006), pp. 524.
[62] J. Bonvin, M. Picasso, and R. Stenberg, GLS and EVSS methods for a three-eld Stokes problem
arising from viscoelastic ows, Comput. Methods Appl. Mech. Engrg. 190 (2001), pp. 3893
3914.
[63] B. Boroomand and O. C. Zienkiewicz, Recovery by equilibrium in patches (REP), Int. J. Nu-
mer. Meth. Engrg. 40 (1997), pp. 137164.
[64] , An improved REP recovery and the effectivity robustness test, Int. J. Numer. Meth. Engrg.
40 (1997), pp. 32473277.
[65] M.-H. Bossel, Membranes elastiquement li ees: extension du th eor` eme de RayleighFaberKrahn
et de lin equalit e de Cheeger, C. R. Acad. Sci., Paris, S er.I Math. 302 (1986), pp. 47-50.
[66] S. W. Brady and A. R. Elcrat, Some results on a posteriori error estimation for approximate solu-
tion of second order elliptic problems, SIAM J. Numer. Anal. 16 (1979), pp. 877889.
[67] D. Braess, Finite elements, Cambridge University Press, Cambridge, 1997.
[68] , A posteriori error estimators for obstacle problems another look, Numer. Math. 101
(2005), pp. 415421.
[69] D. Braess and J. Sch oberl, Equilibrated residual error estimator for Maxwells equation, RICAM
Preprint, 2006.
[70] D. Braess and R. Verf urth, A posteriori error estimators for the RaviartThomas element, SIAM
J. Numer. Anal. 33 (1996), pp. 24312444.
[71] J. H. Bramble, A proof of the inf-sup condition for the Stokes equations on Lipschitz domains,
Math. Models Methods Appl. Sci 13 (2003), pp. 361371.
[72] J. H. Bramble, R. D. Lazarov, and J. E. Pasciak, Least-squares for second-order elliptic problems,
Comput. Methods Appl. Mech. Engrg. 152 (1998), pp. 195210.
[73] J. H. Bramble and A. H. Schatz, Higher order local accuracy by averaging in the nite element
method, Math. Comput. 31 (1977), pp. 94111.
[74] J. H. Brandts, Superconvergence and a posteriori error estimation for triangular mixed nite ele-
ments, Numer. Math. 68 (1994), pp. 311324.
[75] K. Braune, A posteriori error bounds in the computation of inverse standard functions with the
aid of the Newton method, Z. Angew. Math. Mech. 70 (1990), pp. T579T581 (in German).
[76] S. Brenner and R. L. Scott, The mathematical theory of nite element methods, Springer-Verlag,
Berlin, Heidelberg, New York, 1994.
Bibliography 295
[77] F. Brezzi, On the existence, uniqueness and approximation of saddle-point problems arising from
Lagrange multipliers, R.A.I.R.O., Annal. Numer. 8 (1974), R-2, pp. 129151.
[78] F. Brezzi, J. J. Douglas, R. Duran, and M. Fortin, Mixed nite elements for second order elliptic
problems in three variables, Numer. Math. 51 (1987), pp. 237250.
[79] F. Brezzi and M. Fortin, Mixed and hybrid nite element methods, Springer Series in Computa-
tional Mathematics, vol. 15, New York, 1991.
[80] F. Brezzi, M. Fortin, and R. Stenberg, Error analysis of mixed-interpolated elements for Reissner-
Midlin plates, Math. Models and Methods in Applied Sciences 1 (1991), pp. 125151.
[81] F. Brezzi, C. Johnson, and B. Mercier, Analysis of a mixed nite element method for elasto-plastic
plates, Mathematics of Computation 31 (1977), pp. 809817.
[82] H. Buss and S. Repin, A posteriori error estimates for boundary-value problems with obsta-
cles, Numerical mathematics and advanced applications (Jyv askyl a, 1999), World Sci. Publishing,
River Edge, New York, 2000, pp. 162170.
[83] R. Bustinza, G. N. Gatica, and B. Cockburn, An a posteriori error estimate for the local Discon-
tinuous Galerkin method applied to linear and nonlinear diffusion problems, Journal of Scientic
Computing 22/23 (2005), pp. 147185.
[84] G. F. Carey and D. L. Humphrey, Mesh renement and iterative solution methods for nite element
computations, Int. J. Numer. Meth. Engrg. 17 (1981), pp. 17171734.
[85] G. F. Carey and A. I. Pehlivanov, Local error estimation and adaptive remeshing scheme for least-
squares mixed nite elements, Comput. Methods Appl. Mech. Engrg. 150 (1997), pp. 125131.
[86] C. Carstensen, A posteriori estimate for the mixed nite element method, Math. Comput. 66
(1997), pp. 218, 465476.
[87] , Quasi-interpolation and a posteriori error analysis of nite element methods, Mathemati-
cal Modelling in Numerical Analysis 33 (1999), pp. 11871202.
[88] , A posteriori error estimate for the mixed nite element method, Math. Comput. 66 (1997),
pp. 465476.
[89] , An a posteriori error estimate for rst-kind integral equation, Math. Comput. 66 (1997),
pp. 139155.
[90] C. Carstensen and S. Bartels, Each averaging technique yields reliable a posteriori error con-
trol in FEM on unstructured grids. I: Low order conforming, nonconforming, and mixed FEM,
Math. Comput. 71 (2002), pp. 945969.
[91] C. Carstensen and G. Dolzmann, A posteriori error estimates for mixed FEM in elasticity, Nu-
mer. Math. 81 (1998), pp. 187209.
[92] C. Carstensen and S. A. Funken, Fully reliable localized error control in the FEM, SIAM
J. Sci. Comput. 21 (2000), pp. 14651484.
[93] , Constants in Cl ement-interpolation error and residual based a posteriori error estimates
in Finite Element Methods, EastWest J. Numer. Math. 8 (2000), pp. 153175.
[94] C. Carstensen and R. H. W. Hoppe, Error reduction and convergence for an adaptive mixed nite
element method, Math. Comput. 75 (2006), pp. 10331042
[95] , Convergence analysis of an adaptive nonconforming nite element method, Nu-
mer. Math. 103 (2006), pp. 251266.
296 Bibliography
[96] C. Carstensen, J. Hu, and A. Orlando, Framework for the A Posteriori Error Analysis of Noncon-
forming Finite Elements, SIAM J. Numer. Anal. 45 (2007), pp. 6882.
[97] C. Carstensen, R. Klose, and A. Orlando, Reliable and efcient equilibrated a posteriori nite el-
ement error control in elastoplasticity and elastoviscoplasticity with hardening, Comput.Methods
Appl. Mech. Engrg. 195 (2006), pp. 25742598.
[98] C. Carstensen, R. Lazarov, and S. Tomov, Explicit and averaging a posteriori error estimates for
adaptive nite volume methods, SIAM J. Numer. Anal. 42 (2005), pp. 24962521.
[99] C. Carstensen and S. Sauter, A posteriori error analysis for elliptic PDEs on domains with com-
plicated structures, Numer. Math. 96 (2004), pp. 691721.
[100] C. Carstensen and R. Verf urth, Edge residuals dominate a posteriori error estimates for low order
nite element methods, SIAM J. Numer. Anal. 36 (1999), pp. 15711587.
[101] P. Castillo, An a posteriori error estimate for the local Discontinuous Galerkin method, Journal
of Scientic Computing 22/23 (2005), pp. 1872004.
[102] A. Charbonneau, K. Dossou, and R. Pierre, A residual-based a posteriori error estimator for
CiarletRaviart formulation of the rst biharmonic problem, Numer. Meth. for PDEs 13 (1997),
pp. 93111.
[103] Z. Chen and R. H. Nochetto, Rezidual type a posteriori error estimates for elliptic obstacle prob-
lems, Numer. Math. 84 (2000), pp. 527548.
[104] X. Cheng, W. Han, and H. Huang, Analysis of some mixed elements for the Stokes problem,
Journal of Computational and Applied Mathematics 85 (1997), pp. 1935.
[105] E. Chizhonkov and M. Olshanskii, On the domain geometry dependence of the LBB condition,
Mathematical Modelling and Numerical Analys 34 (2000), pp. 935951.
[106] S.-S. Chow, G. F. Carey, and R. D. Lazarov, Natural and post-processed superconvergence in
semilinear problems, Numer. Meth. PDE 7 (1991), pp. 245259.
[107] P. G. Ciarlet, The nite element method for elliptic problems, North Holland, New York, Oxford,
1978.
[108] Ph. Cl ement, Approximations by nite element functions using local regularization, RAIRO
Anal. Num er. 9 (1975), pp. 7784.
[109] S. Cochez-Dhondt, S. Nicaise, and S. Repin, A posteriori error estimates for nite volume ap-
proximations, Technical report, Universitet Valenciennes, 2008.
[110] L. Collatz, Funktionanalysis und Numerische Mathematik, Springer-Verlag, Berlin, 1964.
[111] P. Coorevits, J.-P. Dumeau, and J.-P. Pelle, Error estimator and adaptivity for three-dimensional
nite element analysis, Advances in Adaptive Computational Methods in Mechanics, edited by
P. Ladev eze and J. T. Oden, Elsevier, New York, 1998, pp. 443457.
[112] R. Courant, Variational methods for some problems of equilibrium and vibration, Bulletin of
AMS 49 (1943), pp. 123.
[113] E. Dari, R. Duran, C. Padra, and V. Vampa, A posteriori error estimators for nonconforming nite
element methods, M2AN Math. Model. Numer. Anal. 30 (1996), pp. 385400.
[114] P. Destuynder and B. M etivet, Explicit error bounds of a conforming nite element method,
Math. Comput. 68 (1999), pp. 13791396.
[115] B. S. Dobronets and V. V. Shadurov, Two-sided numerical methods, Nauka Sibirsk. Otdel.,
Novosibirsk, 1990 (in Russian).
Bibliography 297
[116] M. Dobrowolski, On the LBB constant on stretched domains, Math. Nachr. 254255 (2003),
pp. 6467.
[117] W. D orer and M. Rumpf, An adaptive strategy for elliptic problems including a posteriori con-
trolled boundary approximation, Math. Comput. 67 (1998), 224, pp. 13611362.
[118] R. Duran, M. A. Muschietti, and R. Rodriguez, On the asymptotic exactness of error estimators
for linear triangle elements, Numer. Math. 59 (1991), pp. 107127.
[119] R. Duran and R. Rodriguez, On the asymptotic exactness of BankWeisers estimator, Nu-
mer. Math. 62 (1992), pp. 297303.
[120] G. Duvaut and J.-L. Lions, Les inequations en mecanique et en physique, Dunod, Paris, 1972.
[121] I. Ekeland and R. Temam, Convex analysis and variational problems, North-Holland, Amsterdam,
1976.
[122] B. Engelmann, R. H. W. Hoppe, B. Wohlmuth, Yu. Kuznetsov, Yu. Iliash, and Yu. Vasilevskii,
Adaptive macro-hybrid nite element methods, Proc. of the 2nd European Conference on Nu-
merical Mathematics and Advanced Applications, Heidelberg, 1997, World Sci., River Edge,
New York, 1998, pp. 294302.
[123] H. W. Engl and O. Scherzer, Convergence rates results for iterative methods for solving nonlinear
ill-posed problems, Surveys on solution methods for inverse problems, Springer-Verlag, Vienna,
2000, pp. 734.
[124] K. Eriksson, D. Estep, P. Hansbo, and C. Johnson, Computational differential equations, Cam-
bridge University Press, Cambridge, 1996.
[125] K. Eriksson and C. Johnson, An adaptive nite element method for linear elliptic problems,
Math. Comput. 50 (1988), pp. 361383.
[126] , Adaptive nite element methods for parabolic problems. I. A linear model problem, SIAM
J. Numer. Anal. 28 (1991), pp. 4377.
[127] A. Ern and J. Proft, A posteriori Discontinuous Galerkin error estimates for transient convection-
diffusion equations, Appl. Math. Lett. 18 (2005), pp. 833841.
[128] A. Ern, A. F. Stephansen, and M. Vohralik, Improved energy norm a posteriori error estimation
based on ux reconstruction for discontinuous Galerkin methods, manuscript, 2007.
[129] R. E. Ewing, A posteriori error estimation, Comput. Meth. Appl. Mech. Engrg. 82 (1990), pp. 59
72.
[130] R. E. Ewing, R. D. Lazarov, and J. Wang, Superconvergence of the velocity along the Gauss lines
in mixed nite element methods, SIAM J. Numer. Anal. 18 (1991), pp. 10151029.
[131] M. Feistauer, Mathematical methods in uid dynamics, Longman, Harlow, 1993.
[132] W. Fenchel, On the conjugate convex functions, Canad. J. Math. 1 (1949), pp. 7377.
[133] A. Friedman, Variational principles and free-boundary problems, Wiley and Sons, New York,
1982.
[134] M. Frolov, Reliable control over approximation errors by functional type a posteriori error esti-
mates, Phd thesis, Jyv askyl a Studies in Computing 44, University of Jyv askul a, 2004.
[135] , On the efciency of the dual majorant method for the quality estimation of approximate
solutions of fourth-order elliptic boundary-value problems, Russian J. Numer. Anal. Math. Mod-
elling 19 (2004), pp. 407418.
298 Bibliography
[136] M. Frolov, P. Neittaanm aki, and S. Repin, On the reliability, effectivity and robustness of a posteri-
ori error estimation methods, Numerical Methods for Scientic Computing, Variational problems
and applications, CIMNE, Barcelona, 2003, pp. 153175.
[137] , On computational properties of a posteriori error estimates based upon the method of
duality error majorants, Numerical mathematics and advanced applications, Springer-Verlag,
Berlin, 2004, pp. 346357.
[138] , Guaranteed functional error estimates for the Reissner-Mindlin plate problem,
J. Math. Sci (New York) 132 (2006), pp. 553561.
[139] M. Fuchs and S. Repin, Estimates for the deviation from the exact solutions of variational prob-
lems modeling certain classes of generalized Newtonian uids, Math. Meth. Appl. Sci. 29 (2006),
pp. 22252244.
[140] M. Fuchs and G. Seregin, Variational methods for problems from plasticity theory and for gener-
alized Newtonian uids, Lecture Notes in Math., vol. 1749, Springer-Verlag, Berlin, 2000.
[141] A. Gaevskaya, R. H. W. Hoppe, Y. Iliash, and M. Kieweg, Convergence analysis of an adaptive
nite element method for distributed control problems with control constraints, Control of coupled
partial differential equations, Internat. Ser. Numer. Math., 155 (2007), Birkh auser-Verlag, Basel,
pp. 4768.
[142] A. Gaevskaya, R. H. W. Hoppe, and S. Repin, A posteriori estimates for cost functionals of op-
timal control problems, Proceedings of the 6th European Conference on Numerical Mathemat-
ics and Advanced Applications, Santiago de Compostela, 2005, Springer-Verlag, Berlin, 2006,
pp. 291298.
[143] , A posteriori error estimation for elliptic optimal control problems with distributed control,
Journal of Mathematical Sciences 144 (2007), pp. 114.
[144] A. Gaevskaya and S. Repin, A posteriori error estimates for approximate solutions of linear
parabolic problems, Differential Equations, 41 (2005), 7, pp. 970983 (in Russian); translation
in Differ. Equ. 41 (2005), 7, pp. 970983.
[145] H. Gajewski, K. Gr oger, and K. Zacharias, Nichtlineare Operatorgleichungen and Operatordif-
ferentialgleichungen, Akademie-Verlag, Berlin, 1974.
[146] G. Galdi, An introduction to the mathematical theory of the NavierStokes equations, Springer-
Verlag, New York, 1994.
[147] B. G. Galerkin, Beams and plates. Series in some questions of elastic equilibrium of beams and
plates, Vestnik Ingenerov 19 (1915), pp. 897908 (in Russian).
[148] L. Gallimard, P. Ladev eze, and J.-P. Pelle, Error estimation and adaptivity in elastoplasticity,
Int. J. Numer. Meth. Engrg. 39 (1996), pp. 189217.
[149] G. Gatica, A note on the efciency of residual-based a-posteriori error estimators for some mixed
nite element methods, Electron.Trans.Numer.Anal. 17 (2004), pp. 218233.
[150] E. Geisler, A. Tal, and D. Garg, On a posteriori error bounds for the solution of ordinary nonlinear
differential equations, Computers and Mathematics with Applications, Pergamon, Oxford, 1976,
pp. 407416.
[151] D. Gilbarg and N. S. Trudinger, Elliptic partial differential equations of second order, Springer-
Verlag, Berlin, 1977.
[152] V. Girault and P. A. Raviart, Finite element approximation of the NavierStokes equations,
Springer-Verlag, Berlin, 1986.
Bibliography 299
[153] R. Glowinski, Numerical methods for nonlinear variational problems, Springer-Verlag, New-
York, 1982.
[154] , Finite element methods for incompressible viscous ow, Handbook of numerical analysis,
Vol. IX, 31176, North-Holland, Amsterdam, 2003.
[155] R. Glowinski, J.-L. Lions, R. Tr emolier` es, Analyse num erique des in equations variationnelles,
Dunod, Paris, 1976.
[156] R. Glowinski, T.-W. Pan, and J. Periaux, A ctitious domain method for Dirichlet problem and
applications, Comput. Methods Appl. Mech. Engrg. 111 (1994), pp. 283303.
[157] R. Glowinski and O. Pironneau, Numerical Methods for the First Biharmonic Equation and for
the Two-Dimensional Stokes Problem, SIAM Review 21 (1999), pp. 167212.
[158] E. Gorshkova, A posteriori error estimates and adaptive methods for incompressible viscous ow
problems, Phd thesis, Jyv askyl a Studies in Computing 86, University of Jyv askul a, 2007.
[159] E. Gorshkova, A. Mahalov, P. Neittaanmaki, and S. Repin, Functional a posteriori estimates for
viscous ow problems with rotation, J. Math. Sci. New York 142 (2007), pp. 927935.
[160] E. Gorshkova and P. Neittaanmaki. Mesh-adaptive methods for viscous ow problems with rota-
tion, In Adaptive Modeling and Simulation 2007, edited by K. Runesson and P. Diez, CIMNE,
Barcelona, 2007, pp. 109111.
[161] E. Gorshkova, P. Neittaanmaki, and S. Repin, Comparative study of a posteriori error estimates
for the Stokes problem, Numerical Mathematics and Advanced Applications (ENUMATH 2005),
Springer-Verlag, Berlin, 2006, pp. 255262.
[162] , A posteriori error estimates for semi-discrete approximations of the evolutionary Stokes
problem, Technical Report, University of Jyv askyl a (2007).
[163] E. Gorshkova and S. Repin, Error control of the approximate solution to the Stokes equation using
a posteriori error estimates of functional type, European Congress on Computational Methods
in Applied Sciences and Engineering, ECCOMAS 2004, Jyv askyl a, 2428 July, 2004, edited
by P. Neittaanm aki, T. Rossi, K. Majava, and O. Pironeau, O. Nevanlinna and R. Rannacher
(assoc. eds.), (electronic).
[164] G. Haase, M. Kuhn, and U. Langer, Parallel multigrid 3D Maxwell solvers, Parallel Computing
27 (2001), pp. 761775.
[165] L. Halpern, Spectral methods in polar coordinates for the Stokes problem, Application to compu-
tation in unbounded domains, Math. Comput. 65 (1996), pp. 507531.
[166] W. Han, Finite element analysis of a holonomic elasto-plastic problem, Numer. Math. 60 (1992),
pp. 493508.
[167] , A posteriori error analysis for linearization of nonlinear elliptic problems and their dis-
cretization, Math. Meth. Appl. Sci. 17 (1994), pp. 487508.
[168] A. Hanukainen, Functional type a posteriori error estimates for the Maxwells equations, ENU-
MATH 2007, September 1014 2007, book of abstracts, p. 114.
[169] Y. Hayashi, On a posteriori error estimation in the numerical solution of systems of ordinary
differential equations, Hiroshima Math. J 9 (1979), pp. 201243.
[170] B.-O. Heimsund, X.-C. Tai, and J. Wang, Superconvergence for the gradient of nite element
approximations by 1
2
projections, SIAM J. Numer. Anal. 40 (2002), pp. 12631280.
[171] J. G. Heywood, W. Nagata, and W. Xie, A numerically based existence theorem for the Navier
Stokes equations, J. Math. Fluid. Mech. 1 (1999), pp. 523.
300 Bibliography
[172] J. G. Heywood and R. Rannacher, Finite element approximations of the nonstationary Navier
Stokes problem, Part II: Stability of solutions and error estimates uniformin time, SIAMJ. Numer.
Anal. 23 (1986), pp. 750777.
[173] M. Hinterm uller, A primal-dual active set algorithm for bilaterally control constrained optimal
control problems, Quarterly of Applied Mathematics LXI (2003), pp. 131161.
[174] R. Hiptmair, Multigrid method for Maxwells equations, SIAM J. Numer. Anal. 36 (1999),
pp. 204225.
[175] I. Hlav a` cek, J. Chleboun, and I. Babu` ska, Uncertain input data problems and the worst scenario
method, Elsevier, Amsterdam, 2004.
[176] I. Hlav a` cek and M. K`ri` zek, On a superconvergence nite element scheme for elliptic systems. I.
Dirichlet boundary conditions, Aplikace Matematiky, 32 (1987), pp. 131154.
[177] R. H. W. Hoppe, Y. Iliash, C. Iyyunni, and N. H. Sweilam, A posteriori error estimates for adap-
tive nite element discretizations of boundary control problems, J. Numer. Math. 14 (2006),
pp. 5782
[178] R. H. W. Hoppe and R. Kornhuber, Adaptive multilevel methods for obstacle problems, SIAM
J. Numer. Anal. 31 (1994), pp. 301323.
[179] C. O. Horgan. Korns inequalities and their applications in continuum mechanics, SIAM Review
37 (1995), pp. 491511.
[180] P. Houston, J. A. Mackenzie, E. Suli, and G. Warnecke, A posteriori analysis for numerical ap-
proximations of Friederichs systems, Numer. Math. 82 (1999), pp. 433470.
[181] P. Houston, I. Perugia, and D. Schotzau, A posteriori error estimation for discontinuous Galerkin
discretizations of H(curl)elliptic partial differential equations, IMA Journal of Numerical Anal-
ysis (2005), pp. 125.
[182] P. Houston, R. Rannacher, E. S uli, A posteriori error analysis for stabilised nite element approx-
imations of transport problems, Comput. Methods Appl. Mech. Engrg. 190 (2000), pp. 1483
1508.
[183] A. A. Iljushin, Plasticite. (Deformations elastico-plastiques), Eyrolles, Paris, 1956.
[184] Jin Qi-nian, Error estimates of some Newton-type methods for solving nonlinear inverse problems
in Hilbert scales, Inverse Problems 16 (2000), pp. 187197.
[185] C. Johnson, Numerical Solution of Partial Differential Equations by the Finite Element Methods,
Cambridge University Press, Cambridge, 1987.
[186] , Adaptive nite element methods for diffusion and convection problems, Comput. Methods
[187] C. Johnson and P. Hansbo, Adaptive nite elements in computational mechanics, Comput. Meth-
ods Appl. Mech. Engrg. 101 (1992), pp. 143181.
[188] C. Johnson and B. Mercier, Some equilibrium nite element methods two-dimensional elasticity
problems, Numer. Math. 30 (1978), pp. 101116.
[189] C. Johnson, R. Rannacher, and M. Boman, Numerics in hydrodynamic stability. Towards error
control in CFD, SIAM J. Numer. Anal. 32 (1995), pp. 10581079.
[190] C. Johnson and A. Szepessy, Adaptive nite element methods for conservation laws based on a
posteriori error estimates, Commun, Pure and Appl. Math. XLVIII (1995), pp. 199234.
Bibliography 301
[191] V. Jovanovic and C. Rohde, Finite-volume schemes for Friedrichs systems in multiple space di-
mensions: a priori and a posteriori error estimates, Numer. Methods Partial Differ. Equations 21
(2005), pp. 104131.
[192] T. Kato, On some approximation methods concerning the operator T T
+
, Math. Ann. 126 (1953),
pp. 253262.
[193] D. Kay and D. Silvester, A-posteriori error estimation for stabilised mixed approximations of the
Stokes equations, SIAM J. Sci. Comput. 21 (2000), pp. 13211336.
[194] D. W. Kelly, The self equilibration of residuals and complementary error estimates in the nite
element method, Internat. J. Numer. Meth. Engrg. 20 (1984) pp. 14911506.
[195] D. W. Kelly, J. R. Gago, O. C. Zienkiewicz, and I. Babu` ska, A posteriori error analysis and adap-
tive processes in the nite element method. Part I error analysis, Internat. J. Numer. Meth. En-
grg. 19 (1983), pp. 15931619.
[196] D. Kr oner and M. Ohlberger, A posteriori error estimates for upwind nite volume schemes for
nonlinear conservation laws in multi dimensions, Math. Comp. 69 (1999), pp. 2539.
[197] A. N. Kolmogorov and S. V. Fomin, Introductory real analysis, Dover Publications, Inc., New
York, 1975.
[198] R. Kornhuber, A posteriori error estimates for elliptic variational inequalities, Com-
put. Math. Appl. 31 (1996), pp. 4960.
[199] S. Korotov, P. Neittaanmaki, and S. Repin, A posteriori error estimation of goal-oriented quanti-
ties by the superconvergence patch recovery, J. Numer. Math. 11 (2003), pp. 3359.
[200] M. K` r` zek, P. Neittaanm aki, and R. Stenberg (eds.), Finite Element Methods. Superconvergence,
Post-Processing and A Posteriori Error Estimates, Lecture Notes in Pure and Applied Mathemat-
ics, Vol. 196, Marcel Dekker, New York, 1998.
[201] M. K`r` zek and P. Neittaanm aki, Finite Element Approximations of Variational Problems and Ap-
plications. Wiley and Sons, New York, 1990.
[202] , Superconvergence phenomenon in the nite element method arising from averaging of
gradients, Numer. Math. 45 (1984), pp. 105116.
[203] G. Kunert, An a posteriori residual error estimator for the nite element method on anisotropic
tetrahedral meshes, Numer. Math. 86 (2000), pp.471490.
[204] , Towards anisotropic mesh construction and error estimation in the nite element method,
Numer. Meth. PDE 18 (2002), pp. 625648.
[205] Yu. Kuznetsov and S. Repin, New mixed nite element method on polygonal and polyhedral
meshes, Russian J. Numer. Anal. Math. Modeling 18 (2003), pp. 261278.
[206] , Convergence analysis and error estimates for mixed nite element method on distorted
meshes, Journal of Numerical Mathematics 13 (2005), pp. 2251.
[207] P. Ladev eze and D. Leguillon, Error estimate procedure in the nite element method and appli-
cations, SIAM J. Numer. Anal. 20 (1983), pp. 485509.
[208] P. Ladev eze, J.-P. Pelle, and Ph. Rougeot, Error estimation and mesh optimization for classical
nite elements, Engineering Computations 8 (1991), pp. 6980.
[209] P. Ladev eze and Ph. Rougeot, New advances on a posteriori error on constitutive relation in
f.e. analysis, Comput. Methods Appl. Mech. Engrg. 150 (1997), pp. 239249.
302 Bibliography
[210] O. A. Ladyzhenskaya, The mathematical theory of viscous incompressible ow, Gos. Izdat. Fiz.-
Mat. Moscow, 1961 (in Russian); English transl. Gordon & Breach, New York-London, 1963.
[211] , The boundary value problems of mathematical physics, Springer-Verlag, New York, 1985.
[212] , On mosied NavierStokes equations for large velocity gradients, Zapiski Nauchnych
Seminarov LOMI 7 (1968), pp. 126154.
[213] O. A. Ladyzenskaja and V. A. Solonnikov, Some problems of vector analysis, and general-
ized formulations of boundary value problems for the NavierStokes equation, Zap. Nauchn,
Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 59 (1976), pp. 81116 (in Russian).
[214] O. A. Ladyzhenskaya and N. N. Uraltseva, Linear and Quasilinear Elliptic equations, Academic
Press, New York, 1968.
[215] U. Langer and W. Queck, On the convergence factor of Uzawas algorithm, J. Com-
put. Appl. Math. 15 (1986), pp. 191202.
[216] M. G. Larson and A. M alquist, Adaptive variational multiscale methods based on a posteriori
error estimation: Energy norm estimates for elliptic problems, Comput. Methods. Appl. Math.
Engrg. 196 (2007), pp. 23132334.
[217] R. Lazarov, Superconvergence of the gradient for triangular and tetrahedral nite elements of a
solution of linear problems in elasticity theory, Computational Processes and Systems 6 (1988),
pp. 180191 (in Russian).
[218] R. Lazarov, S. Repin, and S. Tomar, Functional a posteriori error estimates for Discontinuous
Galerkin method, Radon Institute of Computational and Applied Mathematics, Preprint 2006-40,
Linz.
[219] R. Lazarov and S. Tomov, A posteriori error estimates for nite volume approximations of
convection-diffusion-reaction equations, Comput. Geosci. 6 (2002), pp. 483503.
[220] A. S. Leonov, Some a posteriori stopping rules for iterative methods for solving linear ill-posed
problems, Comput. Math. Math. Phys. 34 (1994), pp. 121126.
[221] J.-L. Lions, Optimal Control of Systems Governed by Partial Differential Equations, Springer-
Verlag, Berlin, Heidelberg, New York, 1971.
[222] J.-L. Lions and E. Magenes, Probl mes aux limites non homog enes et applications, Dunod, Paris,
1968.
[223] M. Lonsing and R. Verf urth, A posteriori error estimators for mixed nite element methods in
linear elasticity, Numer. Math. 97 (2004), pp. 757778.
[224] J. Ma and H. Brunner, A posteriori error estimates of Discontinuous Galerkin methods for non-
standard Volterra integro-differential equations, IMA J. Numer. Anal. 26 (2006), pp. 7895.
[225] O. Mali and S. Repin, Estimates of the indeterminacy set for elliptic boundary-value problems
with uncertain data, In Adaptive Modeling and Simulation 2007, edited by K. Runesson and
P. Diez, CIMNE, Barcelona, 2007, pp. 158161.
[226] , Estimates of the indeterminacy set for elliptic boundary-value problems with uncertain
data, J. Math. Sci. 150 (2008), pp. 18691874.
[227] , Computable estimates of the solution set for the reaction-diffusion problem with uncertain
data (to appear).
[228] Y. V. Matijasevich, A posteriori interval analysis, Lecture Notes in Computer Sciences, 204
(1985), pp. 328334.
Bibliography 303
[229] J. Medina, M. Picasso, and J. Rappaz, Error estimates and adaptive nite elements for nonlinear
diffusion-convection problems, Ecole Polytechnique Federale de Lausanne, Preprint CH-1015,
1995.
[230] D. Meidner and B. Vexler, Adaptive space-time nite element methods for parabolic optimization
problems. SIAM J. Control Optim. 46 (2007), pp. 116142.
[231] P. Meyer, A unifying theorem on Newtons method, Numer. Funct. Anal. Optim. 13 (1992),
pp. 463473.
[232] S. G. Mikhlin, Variational methods in mathematical physics, Pergamon, Oxford, 1964.
[233] , Constants in some inequalities of analysis, Wiley and Sons, ChiesterNew York, 1986.
[234] , Error Analysis in Numerical Processes, Wiley and Sons, ChiesterNew York, 1991.
[235] M. M oller and D. Kuzmin, Adaptive mesh renement for high-resolution nite element schemes,
Int. J. Numer. Meth. Fluids 52 (2006), pp. 545569.
[236] P. Monk, Finite Element Methods for Maxwells Equations, Clarendon Press, Oxford, 2003.
[237] R. E. Moore, Interval analysis, Prentice-Hall, Englewood Cliffs, New Jersey, 1966.
[238] K. W. Morton and E. S uli, A posteriori and a priori error analysis of nite volume methods, in
The Mathematics of Finite Elements and Applications, edited by J. R. Whiteman, Wiley, 1994,
pp. 267288.
[239] P. Mosolov and V. Myasnikov, Mechanics of rigid plastic bodies, Nauka, Moscow, 1981 (in Rus-
sian).
[240] A. Muzalevsky and S. Repin, On two-sided error estimates for approximate solutions of problems
in the linear theory of elasticity, Russian J. Numer. Anal. Math. Modelling 18 (2003), pp. 6585.
[241] , On error estimates for approximate solutions in problems of the linear theory of thermoe-
lasticity, Russian Math. (Iz. VUZ) 49 (2005), pp. 6068.
[242] P. Neittaanm aki and M. K`r` zek, On O(h
4
) superconvergence of piecewise bilinear FE-
approximations, Teubner-Texte Math., vol. 107, Teubner-Verlag, 1988, pp. 250255.
[243] P. Neittaanm aki and S. Repin, A posteriori error estimates for boundary-value problems related
to the biharmonic operator, East-West J. Numer. Math. 9 (2001), pp. 157178.
[244] , Reliable methods for computer simulation, Error control and a posteriori estimates, Else-
vier, New York, 2004.
[245] , Guaranteed estimates of approximation errors for numerical solutions of the Maxwell
equation (to appear).
[246] P. Neittaanm aki, S. Repin, and P. Turchyn, New A Posteriori Error Indicators in Terms of Linear
Functionals for Linear Elliptic Problems, Russian J. Numer. Anal. Math. Modelling 23 (2008),
pp. 7787.
[247] S. Nicaise, A posteriori error estimation of some cell centered nite volume methods for diffusion-
convection-reaction problems, SIAM J. Numer. Anal. 44 (2006), pp. 949978.
[248] S. Nicaise and S. Repin, Functional a posteriori error estimates for the reaction-convection-
diffusion problem, Zapiski Nauchn, Semin V. A. Steklov Mathematical Institute in St.-Petersburg
(POMI) 326 (2007), pp. 12.
[249] J. A. Nitsche, Ein Kriterium f ur die Quasi-Optimalit at des Ritzschen Verfahrens, Numer. Math.
11 (1968), pp. 346348.
304 Bibliography
[250] J. T. Oden, I. Babu` ska, F. Nobile, Y. Feng, and R. Tempone, Theory and methodology for esti-
mation and control of errors due to modeling, approximation, and uncertainty, Comput. Methods
Appl. Mech. Engrg. 194 (2005), pp 195204.
[251] J. T. Oden, L. Demkowicz, W. Rachowicz, and T. A. Westermann, Towards a universal h
adaptive nite element strategy. Part 2. A posteriori error estimation, Comput. Methods
[252] J. T. Oden, L. Demkowicz, T. Strouboulis, and P. Devloo, Adaptive methods for problems in solid
and uid mechanics, Accuracy Estimates and Adaptive Renements in Finite Element Compu-
tations (Lisbon, 1984), edited by I. Babu` ska, O. C. Zienkiewicz, J. Gago, and E. R. de Oliveira,
Wiley, Chichester, 1996, pp. 249280.
[253] J. T. Oden and S. Prudhomme, Goal-oriented error estimation and adaptivity for the nite element
method, Comput. Math. Appl. 41 (2001), pp. 735756.
[254] J. T. Oden and J. N. Reddy, Mathematical theory of nite elements, Wiley and Sons, 1976.
[255] J. T. Oden, W. Wu, and M. Ainsworth, An a posteriori error estimate for nite element approx-
imations of the NavierStokes equations, Comput. Methods. Appl. Mech. Engrg. 111 (1994),
pp. 185202.
[256] L. A. Oganesjan and L. A. Ruchovec, Study of the rate of convergence of variational difference
scheme for second order elliptic equations in two-dimensional region with a smooth boundary,
U.S.S.R. Comput. Math. and Math. Phy. 9 (1969), pp. 158183.
[257] A. Ostrowski, Les estimations des erreurs a posteriori dans les proc ed es it eratifs, C. R. Acad. Sci,
Paris, S er. AB 275 (1972), pp. A275A278.
[258] C. Padra, A posteriori error estimators for nonconforming approximation of some quasi-
newtonian ows, SIAM J. Numer. Anal. 34 (1997), pp. 16001615.
[259] P. Panagiotopoulos, Inequality problems in mechanics and applications, Birkh auser-Verlag,
Boston, 1985.
[260] J. Peraire and A. T. Patera, Bounds for linear-functional outputs of coercive partial differential
equations: Local indicators and adaptive renement, Advances in Adaptive Computational Meth-
ods in Mechanics, edited by P. Ladev eze and J. T. Oden, Elsevier, New York, 1998, pp. 199228.
[261] T. Plewa, T. Linde, and W. G. Weirs (eds.), Adaptive mesh renement theory and applications.
Proceedings of Chicago workshop on adaptive mesh renement methods, Chicago, IL, USA,
September 35, 2003, Springer-Verlag, Berlin, 2005.
[262] M. Plum, Computer-assisted enclosure methods for elliptic differential equations, Linear Algebra
and its Applications 324 (2001) pp. 147187.
[263] F. Potra, Sharp error bounds for a class of Newton-like methods, Libertas Math. 5 (1985), pp. 71
84.
[264] J. Pousin and J. Rappaz, Consistance, stabilit e, erreurs a priori et a posteriori pour des probl emes
non lin earies, C. R. Acad. Sci. Paris 312 (1991), pp. 699703.
[265] , Consistency, stability, a priori and a posteriori errors for PetrovGalerkin methods ap-
plied to nonlinear problems, Numer. Math. 69 (1994), pp. 213231.
[266] W. Prager and J. L. Synge, Approximation in elasticity based on the concept of function space,
Quart. Appl. Math. 5 (1947), pp. 241269.
[267] Presidents Information Technology Advisory Committee, Computational Sci-
ence: Ensuring Americas Competitiveness, Report to the President, June 2005,
http://www.nitrd.gov/pitac/reports.
Bibliography 305
[268] Jin Qi-nian and Hou Zong-yi, On a posteriori parameter choice strategy for Tikhonov regulariza-
tion of nonlinear ill-posed problems, Numer. Math. 83 (1999), pp 139159.
[269] Lin Qun and Sci Jan, Iterative correctins and a posteriori error estimate for integral equations,
Journal of Computational Math. 11 (1993), pp. 297300.
[270] R. Rannacher, The dual-weighted-residual method for error control and mesh adaptation in -
nite element methods, The mathematics of nite elements and applications X, MAFELAP 1999.
Proceedings of the 10th conference, Brunel Univ., Uxbridge, Middlesex, GB, June 2225, 1999,
edited by J. Whiteman, Elsevier, Amsterdam, 2000, pp. 97116.
[271] , Finite element methods for the incompressible NavierStokes equations, Fundamental di-
rections in mathematical uid mechanics, edited by G. P. Galdi, J. H. Heywood, and R. Rannacher,
Birkhauser-Verlag, Basel, 2000, pp. 191293.
[272] R. Rannacher and F. T. Suttmeier, A feed-back approach to error control in nite element methods:
application to linear elasticity, IWR, Preprint 9642 (SFB 359), Heidelberg 1996.
[273] , A posteriori error control and mesh adaptation for F.E. models in elasticity and elasto-
plasticity, Advances in Adaptive Computational Methods in Mechanics, edited by P. Ladev eze
and J. T. Oden, Elsevier, New York, 1998, pp. 275292.
[274] R. Rannacher and S. Turek, A simple nonconforming quadrilateral Stokes element, Nu-
mer. Meth. Part. Diff. Eq. 8 (1992), pp. 97111.
[275] P.-A. Raviart and J.-M. Thomas, A mixed nite element for second order elliptic problems, Math-
ematical Aspects of Finite Element Methods, edited by I. Galligani and E. Magenes, Springer-
Verlag, Berlin, 1977, pp. 292315.
[276] S. Repin, A posteriori estimates for approximate solutions of variational problems with strongly
convex functionals, Problems of Mathematical Analysis 17 (1997), pp. 199226. (In Russian,
translated in J. Math. Sci, (New York) 97 (1999), no. 4, pp. 43114328).
[277] , A posteriori error estimation for nonlinear variational problems by duality theory, Za-
piski Nauchn, Semin, V. A. Steklov Mathematical Institute in St.-Petersburg (POMI) 243 (1997),
pp. 201214.
[278] , A posteriori error estimation for variational problems with power growth functionals
based on duality theory, Zapiski Nauchn, Semin, V. A. Steklov Mathematical Institute in St.-
Petersburg (POMI) 249 (1997), pp. 244255.
[279] , A posteriori error estimates for approximate solutions of variational problems, Proc. 2nd
European Conference on Numerical Mathematics and Advanced Applications, (Heidelberg)
1997, World Sci, Publishing, River Edge, New York, 1998, pp. 524531.
[280] , A unied approach to a posteriori error estimation based on duality error majorants,
Mathematics and Computers in Simulation 50 (1999), pp. 313329.
[281] , A posteriori estimates of the accuracy of variational methods for problems with nonconvex
functionals, Algebra i Analiz, 11 (1999), pp. 151182 (in Russian, translated in St.-Petersburg
Mathematical Journal 11 (2000), pp. 651672).
[282] , A posteriori error estimates for variational problems with uniformly convex functionals,
Math. Comput. 69 (230) (2000), pp. 481500.
[283] , Estimates of deviations from exact solutions of elliptic variational inequalities, Za-
piski Nauchn. Semin. V. A. Steklov Mathematical Institute in St.-Petersburg (POMI) 271 (2000),
pp. 188203.
306 Bibliography
[284] , Estimates for errors in two-dimensional models of elasticity theory, J. Math. Sci, (New
York) 106 (2001), pp. 30273041.
[285] , A posteriori estimates of the accuracy of dimensional reduction models in 3-D elasticity
theory, Finite element methods (Jyv askyl a), 2000, GAKUTO Internat. Ser. Math. Sci, Appl., 15,
Gakkotosho, Tokyo, 2001, pp. 240253.
[286] , Two-sided estimates of deviation from exact solutions of uniformly elliptic equa-
tions, Proc. St. Petersburg Math. Society IX (2001), pp. 143171, translation in
Amer. Math. Soc. Transl. Ser. 2, 209, Amer. Math. Soc., Providence, RI, 2003.
[287] , Estimates of deviation from exact solutions of initial-boundary value problems for the heat
equation, Rend. Mat. Acc. Lincei 13 (2002), pp. 121133.
[288] , Aposteriori estimates for the Stokes problem, J. Math. Sci, (New York) 109 (2002),
pp. 19501964.
[289] , Estimates of deviations for generalized Newtonian uids, Zapiski
Nauchn. Semin. V. A. Steklov Mathematical Institute in St.-Petersburg (POMI) 288 (2002),
pp. 178203.
[290] , A posteriori error estimates taking into account indeterminacy of the problem data, Rus-
sian J. Numer. Anal. Math. Modelling 18 (2003), pp. 507519.
[291] , A posteriori estimates in local norms. Problems in mathematical analysis, J. Math. Sci,
(New York) 124 (2004), pp. 50265035.
[292] , Functional type a posteriori error estimates for approximate solutions of problems with
incompressibility condition, European Congress on Computational Methods in Applied Sciences
and Engineering, ECCOMAS 2004, Jyv askyl a, 2428 July, 2004, edited by P.Neittaanm aki,
T. Rossi, K. Majava, and O. Pironeau, O. Nevanlinna and R. Rannacher (assoc. eds.), (electronic).
[293] , Estimates of deviations from exact solutions for some boundary-value problems with in-
compressibility condition, Algebra and Analiz 16 (2004), pp. 124161.
[294] , Local a posteriori estimates for the Stokes problem, Zap. Nauchn, Sem. S.-Peterburg
Otdel. Mat. Inst. Steklov. (POMI) 318 (2004), pp. 233245.
[295] , Functional Approach to Locally Based A Posteriori Error Estimates for Elliptic and
Parabolic Problems, Proc. 6th European Conference on Numerical Mathematics and Advanced
Applications (Santiago de Compostela) 2005, Springer-Verlag, Berlin, 2006, pp. 133148.
[296] , On the derivation of functional a posteriori estimates from integral identities, Advances
in numerical mathematics. Proceedings of International Conference on the occasion of the 60th
birthday of Yu. A. Kuznetsov Moscow, September 1617, 2005, W. Fitzgibbon, R. Hoppe, J. Pe-
riaux, O. Pironneau, and Y. Vassilevski, Institute of Numerical Mathematics Russian Academy of
sciences and Department of Mathematics University of Houston, USA, Moscow 2006, pp. 217
242.
[297] , Functional a posteriori estimates for the Maxwells problem, J. Math. Sci. New York 142
(2007), pp. 18211827.
[298] , Estimates of deviations from exact solutions of elliptic variational inequalities,
Zap. Nauchn, Sem. S.-Peterburg, Otdel. Mat. Inst. Steklov. (POMI) 326 (2007), pp. 147164.
[299] , On a posteriori error estimates for the stationary NavierStokes problem, J. Math. Sci.
150 (2008), pp. 18851889.
Bibliography 307
[300] , A posteriori error estimation methods for partial differential equations, In Lectures on
Advanced Computational Methods in Mechanics, edited by J. Kraus and U. Langer, Walter de
Gruyter, Berlin, 2007, pp. 161226.
[301] , A posteriori error estimates for evolutionary reaction-convection-diffusion problem,
Zap. Nauchn, Sem. St. Peterburg, Otdel. Mat. Inst. Steklov. (POMI) (2008), to appear.
[302] , Advanced forms of functional a posteriori estimates for elliptic problems, Russian J. Nu-
mer. Anal. Math. Modelling, to appear.
[303] S. Repin and M. Frolov, A posteriori error estimates for approximate solutions of elliptic bound-
ary value problems, Zh. Vychisl. Mat. Mat. Fiz. 42 (2002), pp. 17741787 (in Russian).
[304] , An estimate for deviations from the exact solution of the ReissnerMindlin plate problem,
J. Math. Sci, (New York) 132 (2006), pp. 331338.
[305] S. Repin and S. Sauter, Functional A Posteriori Estimates for the Reaction-Diffusion Problem,
C. R. Acad. Sci. Paris, Ser. 1 343 (2006), pp. 349354.
[306] , Computable estimates of the modeling error related to KirchhoffLove plate model,
Preprint of the University of Z urich, 2008.
[307] S. Repin, S. Sauter, and A. Smolianski, A posteriori error estimation for the Dirichlet problem
with account of the error in the approximation of boundary conditions, Computing 70 (2003),
pp. 205233.
[308] , Duality Based A Posteriori Error estimator for the Dirichlet Problem, Proc.Appl. Math.
Mech. 2 (2003), pp. 513514.
[309] , A posteriori estimation of dimension reduction errors for elliptic problems in thin domains,
[310] , A Posteriori Control of Dimension Reduction Errors on Long Domains, Proceedings in
Applied Mathematics and Mechanics 4 (2004), pp. 714715.
[311] , A posteriori error estimation for the Poisson equation with mixed Dirichlet/Neumann
boundary conditions, J. Comput. Appl. Math. 164/165 (2004), pp. 601612.
[312] , A posteriori estimation of dimension reduction errors, Proc. 5th European Conference on
Numerical Mathematics and applications (Prague) 2004, Springer-Verlag, pp. 717725.
[313] , Two-Sided A Posteriori Error Estimates for Mixed Formulations of Elliptic Problems,
[314] S. Repin and A. Smolianski, Functional-type a posteriori error estimates for mixed nite element
methods, Russian J. Numer. Anal. Math. Modelling 20 (2005), pp. 365382.
[315] S. Repin and R. Stenberg, Two-sided a posteriori estimates for the generalized Stokes problem,
Helsinki University of Technology, 2006, Preprint A516.
[316] , A posteriori error estimates for the generalized Stokes problem, J. Math. Sci. New Jork
142 (2007), pp. 18281843.
[317] S. Repin and S. Tomar, A posteriori error estimates for nonconforming approximations of elliptic
problems based on Helmholtz decomposition of the error, Radon Institute of Computational and
Applied Mathematics, Preprint 2007-40.
[318] S. Repin and J. Valdman, Functional a posteriori estimates for incremental models in elasto-
plasticity, Technical Report, Radon Institute of Computational and Applied Mathematics (RI-
CAM), Preprint 2007-41.
308 Bibliography
[319] , A posteriori error estimates for problems with nonlinear boundary conditions, J. Nu-
mer. Math. 16 (2008), pp. 5181.
[320] S. Repin and L. S. Xanthis, A posteriori error estimation for elasto-plastic problems based on
duality theory, Comput. Methods Appl. Mech. Engrg. 138 (1996), pp. 317339.
[321] , A posteriori error estimation for nonlinear variational problems, Comptes Rendus de
lAcad emie des Sciences, Math ematique 324 (1997), pp. 11691174.
[322] W. C. Rheinboldt, On a theory of mesh-renement processes, SIAM J. Numer. Anal. 17 (1980),
pp. 766778.
[323] W. Ritz,

Uber eine neue Methode zur L osung gewisser Variationsprobleme der Mathematischen
Physik, J. Reine Angew. Math. 135 (1909), pp. 161.
[324] R. T. Rockafellar, Convex analysis, Princeton Univ. Press, 1970.
[325] J. E. Roberts and J.-M. Thomas, Mixed and hybrid methods, Handbook of numerical analysis,
Vol. II, Handb. Numer. Anal., II, North-Holland, Amsterdam, 1991, pp. 523639.
[326] R. Rodrigues, Some remarks on ZienkiewiczZhu estimator, Numer. Methods for PDE 10 (1994),
pp. 625635.
[327] J. Saranen, On an inequality of Friedrichs, Math. Scand. 51 (1982), pp. 310322.
[328] O. Scherzer, H. W. Engl, and K. Kunisch, Optimal a posteriori parameter choice for Tikhonov reg-
ularization for solving nonlinear ill-posed problems, SIAM J. Numer. Anal. 30 (1993), pp. 1796
1838.
[329] J. Schr oder, Operator Inequalities, Academic Press, New York, 1980.
[330] M. Schultz and O. Steinbach, A new a posteriori error estimator in adaptive direct boundary
element methods: the Dirichlet problem, Calcolo 37 (2000), pp. 7996.
[331] M. Schulz and W. L. Wendland, A general approach to a posteriori error estimates for strictly
monotone and Lipschitz continuous nonlinear operators illustrated in elasto-plasticity, Proc. 2nd
European Conference on Numerical Mathematics and Advanced Applications 1997 (Heidelberg),
World Sci, Publishing, River Edge, New York, 1998, pp. 572579.
[332] C. Schwab, A-posteriori modelling error estimation for hierarchic plate models, Numer. Math.
74 (1996), pp. 221259.
[333] S. Shaw and J. R. Whiteman, Negative norm error control for second-kind convolution Volterra
equations, Numer. Math. 85 (2000), pp. 329341.
[334] K. G. Siebert, An a posteriori error estimation for anisotropic renement, Numer. Math. 73
(1996), pp. 373398.
[335] M. G. Slobodyanskii, Estimates of the error of an approximate solution in linear problems reduc-
ing to variational ones, and their application to the determination of two-sided approximations
in static problems of the theory of elasticity, Akad. Nauk SSSR. Prikl. Mat. Meh. 16 (1952),
pp. 449464.
[336] S. L. Sobolev, Some Applications of Functional Analysis in Mathematical Physics, Izdt. Lenin-
grad. Gos. Univ., Leningrad, 1955 (in Russian, translated in Translation of Mathematical Mono-
graphs, Volume 90, American Mathematical Society, Providence, RI, 1991).
[337] E. Stein, F. J. Bartold, S. Ohnimus, and M. Schmidt, Adaptive nite elements in elastoplasticity
with mechanical error indicators and Neumann-type estimators, Advances in Adaptive Compu-
tational Methods in Mechanics, edited by P. Ladev eze and J. T. Oden, Elsevier, New York, 1998,
pp. 81100.
Bibliography 309
[338] E. Stein and S. Ohnimus, Coupled model- and solution-adaptivity in the nite element method,
Comput. Methods Appl. Mech. Engrg. 150 (1997), pp. 327350.
[339] , Coupled anisotropic discretization and modelerror estimation in solid mechanics by local
Neumann problems, Comput. Methods Appl. Mech. Engrg. 176 (1999), pp. 363385.
[340] E. Stein, M. R uter, and S. Ohnimus, Error-controlled adaptive goal-oriented modeling and -
nite element approximations in elasticity, Comput. Methods Appl. Mech. Engrg. 196 (2007),
pp. 35983613.
[341] R. Stenberg, Some new families of nite elements for the Stokes equations, Numer. Math. 56
(1990), pp. 827838.
[342] , Error analysis of some nite element methods for the Stokes problem, Math. Comput. 54
(1990), pp. 495508.
[343] G. Stoyan, Towards Discrete Velte Decompositions and Narrow Bownds for Inf-Sup Constants,
Computer and Mathematics with Applications, 38 (1999), pp. 243261.
[344] G. Strang and G. Fix, An analysis of the nite element method, Prentice Hall, Englewood Cliffs,
1973.
[345] T. Strouboulis and J. T. Oden, A posteriori error estimation of nite element approximations in
uid mechanics, Comput. Meth. Appl. Mech. Engrg. 78 (1990), pp. 201242.
[346] S. Sun and M. F. Wheeler, 1
2
(H
1
)-norm a posteriori error estimation for Discontinuous
Galerkin approximations of reactive transport problems, Journal of Scientic Computing 22/23
(2005), pp.501530.
[347] J. L. Synge, The hypercircle method, Studies in numerical analysis (papers in honour of Cornelius
Lanczos on the occasion of his 80th birthday), Academic Press, London, 1974, pp. 201217.
[348] R. Temam, NavierStokes equations. Theory and numerical analysis, Studies in Mathematics and
its Applications, vol. 2, North-Holland, Amsterdam, 1979.
[349] , Probl` emes math ematique en plasticit e, Bordas, Paris, 1983.
[350] S. Tomar and S. Repin, Efcient computable error bounds for dicontinuous Galerkin approxima-
tions of elliptic problems, Radon Institute of Computational and Applied Mathematics, Preprint
2007-39.
[351] K. Tsuruta and K. Ohmori, A posteriori error estimation for Volterra integro-differential equa-
tions, Mem. Numer. Math. 1976, pp. 3347.
[352] S. Turek, Efcient solvers for incompressible ow problems: An algorythmic and computational
approach, Springer-Verlag, Berlin, 1999.
[353] J. P. Vanyo, Rotating uids in engineering and science, Dover Publ. Inc., New York, 2001.
[354] R. Verf urth, A posteriori error estimators for the Stokes equations, Numer. Math. 55 (1989),
pp. 309326.
[355] , A posteriori error estimates for nonlinear problems. Finite element discretisations of el-
liptic equations, Math. Comput. 62 (1994), pp. 445475.
[356] , A review of a posteriori error estimation and adaptive mesh-renement techniques, Wiley
and Sons, Teubner-Verlag, New-York, 1996.
[357] , A posteriori error estimates for nonlinear problems. 1
i
(0. T : 1
]
(C))-error estimates for
nite element discretizations of parabolic equations, Math. Comput. 67 (1998), pp. 13351360.
310 Bibliography
[358] , A posteriori error estimators for convection-diffusion problems, Numer. Math. 80 (1998),
pp. 641663.
[359] , A posteriori error estimates for nite element discretizations of the heat equation, Calcolo
40 (2003), pp. 195212.
[360] , Robust a posteriori error estimates for sationary convection-diffusion equations, SIAM
J. Numer. Anal. 43 (2005), pp. 17661782.
[361] A. Vesser, Efcient and reliable a posteriori error estimators for elliptic obstacle problems,
[362] M. Vishik , La methode des projections orthogonales pour les equations differentielles conjuguees
a ellesmemes, Doklady Acad. Sci. URSS (N.S.) 56 (1947), pp. 115118.
[363] M. Vogelius and I. Babu` ska, On a dimensional reduction method I. The optimal selection of basis
functions, Math. Comput. 37 (1981), pp. 3146.
[364] , On a dimensional reduction method III. A posteriori error estimation and an adaptive
approach, Math. Comput. 37 (1981), pp. 361384.
[365] M. Vohralik, Equivalence between mixed nite element and multi-point nite volume methods,
C. R. Acad. Sci, Paris, Ser. I 339 (2004), pp. 525528.
[366] L. B. Wahlbin, Superconvergence in Galerkin Finite Element Methods, Lecture Notes in Mathe-
matics, vol. 1605, Springer-Verlag, Berlin, 1995.
[367] J. Wang, Superconvergence analysis of nite element solutions by the least-squares surface tting
on irregular meshes for smooth problems, J. Math. Study 33 (2000), pp. 229243.
[368] J. Wang and X. Ye, Superconvergence analysis for the NavierStokes equations, Applied Numeri-
cal Mathematics 41 (2002), pp. 515527.
[369] H. Weil, The method of orthogonal projections in potential theory, Duke Math. J. 7 (1940),
pp. 411444.
[370] J. R. Whiteman and G. Goodsell, A survey of gradient superconvergence for nite approximations
to second order elliptic problems on triangular and tetrahedral meshes, The Mathematics of
Finite Elements and Applications VII, edited by J. R. Whiteman, Academic Press, 1991, pp. 55
74.
[371] N.-E. Wiberg, F. Abdulwahab, and S. Ziukas, Enhanced superconvergent patch recovery incor-
porating equilibrium and boundary conditions, Int. J. Num. Meth. Engng. 37 (1994), pp. 3417
3440.
[372] B. I. Wohlmuth and R. H. W. Hoppe, A comparison of a posteriori error estimators for mixed nite
element discretizations by RaviartThomas elements, Math. Comput. 68 (228) (1999), pp. 1347
1378.
[373] T. Yamamoto, A note on a posteriori error bound of Zabrejko and Nguen for Zincenkos iteration,
Numer. Funct. Anal. Optim. 9 (1987), pp. 987994.
[374] K. Yosida, Functional Analysis, Springer-Verlag, New Tork, 1996.
[375] S. Zaremba, Sur un probleme toujours possible comprenant, a titre de cas particuliers, le prob-
leme de Dirichlet et celui de Neumann, J. math. pures et appl. 6 (1927), pp. 127163.
[376] E. Zeidler, Nonlinear functional analysis and its applications. I. Fixed-point theorems, Springer-
Verlag, New York, 1986.
Bibliography 311
[377] Z. Zhang and A. Naga, A new nite element gradient recovery method: superconvergence prop-
erty, SIAM J. Sci. Comput. 26 (2005), pp. 11921213.
[378] J. Z. Zhu, A posteriori error estimation the relationship between different procedures, Com-
put. Methods Appl. Mech. Engrg. 150 (1997), pp. 411422.
[379] O. C. Zienkiewicz, Achievements and some unsolved problems of the nite element method,
Int. J. Numer. Meth. Engrg. 47 (2000), pp. 928.
[380] O. C. Zienkiewicz, B. Boroomand, and J. Z. Zhu, Recovery procedures in error estimation and
adaptivity: Adaptivity in linear problems, Advances in Adaptive Computational Methods in Me-
chanics, edited by P. Ladev eze and J. T. Oden, Elsevier, New York, 1998, pp. 323.
[381] O. C. Zienkiewicz and K. Morgan, Finite elements and approximation, Wiley and Sons, New
York, 1983.
[382] O. C. Zienkiewicz and J. Z. Zhu, A simple error estimator and adaptive procedure for practical
engineering analysis, Internat. J. Numer. Meth. Engrg. 24 (1987), pp. 337357.
[383] , Adaptive techniques in the nite element method, Commun. Appl. Numer. Methods 4
(1988), pp. 197204.
[384] , The superconvergent patch recovery and a posteriori error estimates. Part 1: The recovery
technique, Int. J. Numer. Meth. Engrg. 33 (1992), pp. 13311364.
[385] M. Zl amal, Some superconvergence results in the nite element method. Mathematical aspects of
nite element methods, Lecture Notes in Math., Vol. 606, Springer-Verlag, Berlin, 1977, pp. 353
362.
[386] , Superconvergence and reduced integration in the nite element method, Math. Comput.
32 (1978), pp. 663685.
Notation
:= equals by denition
V for all
= implies
R
d
space of real J-vectors
M
dd
space of real J J matrixes
M
dd
x
space of symmetric real J J matrixes
C, o open sets in R
d
I, I
1
, I
2
boundary of C and its parts
n exterior unit normal
meas
d
{o] Lebesgue measure of a set o R
d
diamS diameter of the set S
a b scalar product of vectors
a b tensor product of vectors
a b vector product of vectors
o : c scalar product of tensors
tro trace of o
o
D
deviator of o; o
D
:= o
1
d
I tr o
fgg
o
mean value of g on o; fgg :=
1
[o[
_
o
g J.
g
o
g
o
:= g fgg
o
supp g support set of g
g|
I
jump of g on I
(g)
negative part of g
(g)
C
positive part of g
Ker T kernel of the operator T
,i
partial derivative with respect to i -th coordinate;
i
:=
J
Jx
i
V gradient of a scalar-valued function; V :=
_
,1
. . . . .
,d
_
div divergence of a vector-valued function; div :=
i=1,d

i,i
Div divergence of a tensor-valued function; (Div t)
}
:=
i=1,d
t
i},i
curl rotor of a vector-valued function;
curl := (
3,2

2,3
:
1,3

3,1
:
2,1

1,2
)
^ Laplace operator; ^ = div V
M error majorant
M error minorant
e error
E error indicator
1
eff
efciency index of an error estimate
B(.. r) ball of the radius r centered at .
Notation 313
set of uncertain data
S set of possible solutions
P projection operator
1
]
(o) space of functions integrable in o with power
1
2
(C) set of functions in 1
2
(C) with integral mean 0
[ [
],o
1
]
(o) norm
[ [. [ [
D
1
2
(C) norm
[ [
o,D
supremum norm
W
I,]
(C) Sobolev space of functions having generalized derivatives
up to order l integrable with power
[ [
I,],D
W
I,]
(C) norm
[[ [[ energy norm
[ |[ broken energy norm
[ [
j
weighted norm with weight
C
T D
constant in the Friedrichs inequality for the domain C
C
1D
constant in the Poincar e inequality for the domain C
(C) space 1
2
(C. M
dd
) of tensor-valued functions
with square summable components
x
(C) space 1
2
(C. M
dd
x
) of symmetric tensor-valued functions
with square summable components
S(C) subspace of 1
2
(C. R
d
) formed by divergence-free functions
S
1
(C) the closure of smooth solenoidal functions with
compact supports in C with respect to the norm of H
1
(C. R
d
)
H
-1
(C) space dual to

H
1
(C)
H(C. div) := { 1
2
(C) [ div 1
2
(C)]
[ [
div
norm of H(C. div)
H(C. Div) := {t 1
2
(C. M
dd
) [ Div t 1
2
(C. R
d
)]
[ [
Div
norm of H(C. Div)
L(X. Y ) space of linear bounded operators acting from X to Y
(
+
. ) duality pairing between
+
V
+
and V
J
+
polar functional (YoungFenchel conjugate of J),
J
+
(
+
) := sup
{(
+
. ) J()]
J
++
bipolar functional
J
++
() := sup
{(
+
. ) J
+
(
+
)]
D(.
+
) compound functional,
D(.
+
) := J() J
+
(
+
) (
+
. )
dJ() subdifferential of J at
Index
A
A posteriori estimates
explicit residual method, 32
implicit residual methods, 35
iteration methods, 28
optimal control problems, 254
A priori error estimates, 1
Adaptive methods, 31
Adjoint problem, 42
Almost incompressible uids, 137
B
Banach theorem, 27
Boundary condition
DirichletRobin, 189
friction, 209
mixed, 75
Neumann, 80, 208
nonlinear, 201
Robin, 85
Winkler, 210
C
C ea lemma, 59
Cl ements interpolation estimates, 33
Clarksons inequality, 198
Coincidence set, 179
Compound functional, 189
Contractive mapping, 27
Convex set, 17
Cost functional, 243
majorant, 245
minorant, 247
D
Discontinuous Galerkin method, 264
Dual-weighted residual method, 42
Duality pairing, 18
Duality relation, 160
E
Efciency index, 56, 169
Equilibration, 41
Error indicator, 22, 65, 70
Error majorant, 6, 49, 50, 128, 132, 152,
163, 168
decomposition of C, 62
modications, 59
Error minorant, 6, 49, 52, 148, 165
Error reduction property, 163
F
Finite volume method, 265
Fixed point, 27
Fixed point problem, 30
Friedrichs inequality, 16
Functional
afne, 18
bidual (bipolar), 19
characteristic, 18
compound, 21
conjugate, 187, 203
convex, 17
dual (polar), 18
forcing, 187
G ateaux differentiable, 21
nondifferentiable, 185
strictly convex, 17
uniformly convex, 187
Functional a posteriori estimate
biharmonic problem, 219
data indeterminacy errors, 270
diffusion problem, 77
domain decomposition, 62, 166
equilibration of uxes, 64
general form, 188
generalized Stokes problem
models with polymerization, 148
generalized Stokes problem, 141
models with rotation, 149
in terms of functionals, 279
linear elasticity
estimates for displacements, 107
estimates for stresses, 109
isotropic media, 111
modeling error, 114
Maxwells problem, 225
Index 315
minimization of the error majorant, 66
nonconforming approximations, 254
nonvariational method, 50
optimal control problems, 242
principal structure, 215
reaction-convection-diffusion problem,
96
reaction-diffusion problem, 83
regularization of uxes, 65
Runges type estimate, 66
stationary Oseen problem, 152
stationary Stokes problem, 123
estimates for combined norms, 128
estimates for pressure, 127
estimates for stress, 128
estimates for velocity, 125
lower bounds of errors, 130
mixed boundary conditions, 131
structure, 52
variational method, 45
G
Galerkin approximation, 59, 65, 78
Galerkin orthogonality, 33
Gradient averaging, 40
H
Helmholtz decomposition, 11, 23, 257
Hookes law, 108
Hypercircle method, 23
I
Inf-Sup condition, 118
Interpolation constants, 33
K
KirchhoffLove plate model, 221
Korns inequality, 16
L
LBB condition, 118
Limit density, 54
M
Main error control problem, 6
Maximum principle, 285, 286
Mikhlin estimate, 25
Minkovski inequality, 15
Mixed approximations, 170
Modeling error, 112
Monotone operators, 30
N
Nonconforming approximations, 254
O
Ostrogradski formula, 10
Ostrowski estimate, 26
P
Poincar e inequality, 16
Post-processing, 37
PragerSynge estimate, 23
Primal-dual error norm, 57, 168
Problem
biharmonic, 222
convection-diffusion, 88
dual, 159, 188
evolutionary, 229
general elliptic, 186, 191
generalized Stokes, 140
linear diffusion, 75
linear elasticity, 104
linear elliptic
fourth order, 218
general form, 158
on a subspace, 173
with lower terms, 171
Maxwell, 224
NavierStokes, 153
nonlinear elliptic, 196
optimal control, 242
Oseen, 151
plane strain, 113
plane stress, 111
primal, 159
reaction-convection-diffusion, 93
reaction-diffusion, 81
stationary Stokes, 123
Stokes, 117
saddle point statement, 119
with subquadratic growth, 200
with superquadratic growth, 198
Projection operator, 255
R
Regularization, 37
316 Index
Runge estimate, 22
S
Saddle point, 176
Sobolev space, 13
Solenoidal eld, 10
Space
bidual, 18
of rigid deections, 17
of solenoidal functions, 122
topologically dual, 18
SturmLiouville operator, 86
Sub-solution, 286
Subdifferential, 19
Subgradient, 19
Super-solution, 286
Superconvergence, 40
U
Uncertain data, 265, 278
V
Variational inequality, 178
rst kind, 179
fourth order problems, 222
second kind, 185
Y
Youngs inequality, 8, 194

A Posteriori Estimates For Partial Differential Equations, Sergey Repin PDF

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

A Posteriori Estimates For Partial Differential Equations, Sergey Repin PDF

Transféré par

Droits d'auteur :

Formats disponibles

Radon Series on Computational and Applied Mathematics 4

Walter de Gruyter Berlin New York

in a complete metric space

) denote the error at the th step. Then

as m o, we pass to the limit with respect to m and obtain

in a complete ordered (partially

exists and belongs to D. From (2.5.2) it follows that .

(and the convergence

[[[ [[[ := sup

the error functional. It is easy to show

[[[ [[[ [V(u )[.

[[[ [[[ . (2.6.4)

(n) [_ [V(u )[[Vn[, so that (2.6.3) implies the inequality op-

. Here arises the problem of how to compute [[[ [[[ F

[[[ [[[ prac-

[, q[ _ [[[ [[[ div , [[[ [[[ (3.1.6)

(n) is the residual functional (cf. 2.6.2).

(0) is easily computable and F

Q(T ) := {, H(C. div) [ fdiv , g

are constant vectors. Dene the normal

. One way to construct a nite element subspace of

and represent the rst term on the right-hand side of

[(u )[ [[, V[[

(;)[u |[. (4.3.61)

(;) := (1;) max

(;)[u |[. and we obtain (4.3.61).

(;) dened by the rela-

J(.). 0]. Here the ;

V be a vector-valued function such that

. For this purpose, we

. Other functions in the right-hand side of (6.2.11), we dene

= 0 almost everywhere in C. Moreover,

(this fact is known, see, e.g.,

= 1 (this assumption does not lead to loss in generality,

and the majorant M

= 1() 1(u) _ 1() 1( n)

(. ,) is equivalent to the error in the combined norm

(. ,) _ [[( u)[[ [[ ,[[

is an efcient and reliable measure of the error in the combined norm

(u. ,). (7.1.45)

) = 0 and such that

are upper and lower coincidence sets and C

= min{0. g] denote the positive and negative parts

= 0 for a.e. . C. (8.1.9)

{. [ div ,(.) (.) _ 0]

{. [ div ,(.) (.) < 0]

> r(,). (8.1.24)

< r(,) <

are certain nonnegative functionals

reinforces the usual convexity inequality. For this reason,

, provided that u and belong to T(0. ). Indeed,

. Then for any u

. By (8.3.14), we conclude that

and , errors can be estimated in terms of the quantity

(( u)), (( u)), and D

(( u)) := (( u)) ([( u)[

. In the variational method, we need to

(see S. Sobolev [336] and P. Mosolov

can be viewed as a certain measure of the error in the duality relations

is deduced as an upper bound of J

is sharp. Similar results hold for

() is not a norm. Nevertheless,