A Posteriori Error Analysis via Duality Theory

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY Advances in Mechanics and Mathematics Volume 8 Series Editors: David Y...

Author: Weimin Han

16 downloads 1065 Views 10MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Advances in Mechanics and Mathematics Volume 8

Series Editors: David Y. Gao Virginia Polytechnic Institute and State University, U.S.A.

Ray W. Ogden University of Glasgow, U.K.

Advisory Editors:

I. Ekeland University of British Columbia, Canada K.R. Rajagopal Texas A M University, U.S.A.

T . Ratiu Ecole Polytechnique, Switzerland

W. Yang Tsinghua University, P.R. China

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY With Applications in Modeling and Numerical Approximations

WEIMIN HAN Department of Mathematics University of Iowa Iowa City, IA 52242, U.S.A.

Library of Congress Cataloging-in-Publication Data

A C.I.P. Catalogue record for this book is available from the Library of Congress. ISBN 0-387-23536-1

e-ISBN 0-387-23537-X

Printed on acid-free paper.

O 2005 Springer Science+Business Media, Inc.

All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, Inc., 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now know or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks and similar terms, even if the are not identified as such, is not to be taken as an expression of opinion as t o whether or not they are subject to proprietary rights. Printed in the United States of America. 9 8 7 6 5 4 3 2 1

SPIN 11336112

Contents

List of Figures List of Tables Preface 1. PRELIMINARIES Introduction Some basic notions from functional analysis Function spaces Weak formulation of boundary value problems Best constants in some Sobolev inequalities Singularities of elliptic problems on planar nonsmooth domains An introduction of elliptic variational inequalities Finite element method, error estimates 2. ELEMENTS OF CONVEX ANALYSIS, DUALITY THEORY 2.1 Convex sets and convex functions 2.2 Hahn-Banach theorem and separation of convex sets 2.3 Continuity and differentiability 2.4 Convex optimization 2.5 Conjugate functionals 2.6 Duality theory 2.7 Applications of duality theory in a posteriori error analysis

3. A POSTERIORI ERROR ANALYSIS FOR IDEALIZATIONS IN LINEAR PROBLEMS 3.1 Coefficient idealization 3.2 Right-hand side idealization

vii xi xv 1 1 5 7 16 20 25 29 36

vi

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

3.3 3.4 3.5 3.6

Boundary condition idealizations Domain idealizations Error estimates for material idealization of torsion problems Simplifications in some heat conduction problems

4. A POSTERIORI ERROR ANALYSIS FOR LINEARIZATIONS 4.1 Linearization of a nonlinear boundary value problem 4.2 Linearization of a nonlinear elasticity problem 4.3 Linearizations in heat conduction problems 4.4 Nonlinear problems with small parameters 4.5 A quasilinear problem 4.6 Laminar stationary flow of a Bingham fluid 4.7 Linearization in an obstacle problem 5. A POSTERIORIERROR ANALYSIS FOR SOME NUMERICAL PROCEDURES 5.1 A posteriori error analysis for regularization methods 5.2 KaEanov method for nonlinear problems 5.3 KaEanov method for a stationary conservation law 5.4 KaEanov method for a quasi-Newtonian flow problem 5.5 Application in solving an elastoplasticity problem 6. ERROR ANALYSIS FOR VARIATIONAL INEQUALITIES OF THE SECOND KIND 6.1 Model problem and its finite element approximation 6.2 Dual formulation and a posteriori error estimation 6.3 Residual-based error estimates for the model problem 6.4 Gradient recovery-based error estimates for the model problem 6.5 Numerical example on the model problem 6.6 Application to a frictional contact problem REFERENCES

255 262 27 1 287

Index

301

List of Figures

Numerical analysis of a physical problem Smoothness of the boundary Smooth domains Lipschitz domains A crack domain A comer domain near the comer 0 A finite element mesh Continuity of a convex function Subdifferential of the absolute value function Choice of the auxiliary function near a comer domain: Subcase 2A Choice of the auxiliary function near a comer domain: Subcase 2B Domain idealization One-dimensional stress-strain relation Neighborhood around a corner The function for numerical examples Setting of a nonlinear elasticity problem The working problem Example 5.9, convergence of KaEanov iterates Example 5.10, convergence of KaEanov iterates Example 6.11, true solution Example 6.1 1, initial partition and adaptively refined partition after 5 iterations Example 6.11, l/u- u K ~ / / ( 0~ ) ,VS.~ h11211~- Xhllo,r( A )

viii

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

6.4 6.5 6.6 6.7 6.8

6.9

Example 6.1 1, plots of X h , 1 Example 6.11, plots of Example 6.11, results based on residual type estimator, IIu - ~ x ~ I 1 (01 1 , VS. ~ IIu - utd//1,n (A) Example 6.11, results based on gradient recovery type (D)VS. llu - utd (A) estimator, llu - uEn Example 6.11, results based on residual type estimator, IIu - ~ ; l l ~ l and l ~ ,q~~ on uniform mesh (o) vs. Ilu uidI 11 ,a and q~ on adapted mesh (A) Example 6.11, results based on gradient recovery type l ~q~, on~ uniform mesh (o) estimator, llu - ~ E ~ l and vs. llu and 7~ on adapted mesh (A) Example 6.11, performance comparison between the two error estimators Example 6.12, physical setting Example 6.12, initial mesh with 128 elements, 81 nodes Example6.12, ~ / U - U ; ~( /0/) ~~ sh. 1 / 2 1 1 ~ h T - ~ ~ ~ ~ ~ 0 , r c (A) Example 6.12, plot of X h , ~ Example 6.12, plot of Example 6.12, results based on residual type estimator, deformed configuration 5583 elements, 2921 nodes Example 6.12, results based on gradient recovery type estimator, deformed configuration with 5437 elements, 2832 nodes Example 6.12, results based on residual type estimator, Ilu - u y / / v(0) VS' 1/21 - u;~IIV (A) Example 6.12, results based on gradient recovery type estimator, iiu - u;lln (a) vs. iiu - uid/Iv (A) Example 6.12, results based on residual type estimator, I I U - U ; ~ . ~ I I ~ andqR onuniformmesh(o)vs. IIu-u~d11~ and q~ on adapted mesh ( A ) Example 6.12, results based on gradient recovery type estimator, /lu- u K ~ ( and ( ~ q~ on uniform mesh (o) vs. I 1 u - uid11 and q~ on adapted mesh (A) Example 6.13, physical setting Example 6.13, initial mesh with 160 elements, 105 nodes Example 6.13, results based on residual type estimator, deformed configuration with 5222 elements, 2755 nodes

~ t ~ / / ~ , ~

6.10 6.11 6.12 6.13 6.14 6.15 6.16 6.17

6.18 6.19 6.20

6.21

6.22 6.23 6.24

Itv

List of Figures

Example 6.13, results based on gradient recovery type estimator, deformed configuration with 4964 elements, 2624 nodes Example 6.13, plot of Ah,l Example 6.13, plot of Ah,2 Example 6.13, results based on residual type estimator, IIu - ~ ; ~ l(0) l vVS' IIu - u;~IIV (A) Example 6.13, results based on gradient recovery type estimator, Ilu - uKnIIV(o) VS. /Iu - uidlV(A) Example6.13, llu-uKnIIV (o)vs. h 1 / 2 / / ~ h T - ~ K ~ I I 0 , r c (A)

Example 6.13, results based on residual type estimator, I I U - U ; ~ I I ~ andqRonuniformmesh(o)vs. I I U - U ; ~ / ) ~ and q~ on adapted mesh (A) Example 6.13, results based on gradient recovery type estimator, Ilu - uKn11 and 7~ on uniform mesh (o) vs. /lu- uid11 and q~ on adapted mesh (A)

List of Tables

89 Example 3.8, effectivity of error bound 91 Example 3.10, effectivity of error bounds 98 Example 3.12, smallest roots for several angles 140 Example 4.1, true errors and error bounds, X = 0.1 140 Example 4.1, true errors and error bounds, X = 0.015 140 Example 4.1, true errors and error bounds, X = 0.0 101 Example 4.2, error bounds for problems on an L-shape domain 142 Singular exponents Example 4.11, error bounds for various parameters Example 4.12, error bound for energy difference Efficiency of error bounds on a one-dimensional obstacle problem Example 5.7, numerical results with a ( s ) = (2+s)/ (1+ s ) , Po = 718 Example 5.7,numerical results with a ( s ) = (7+s) / ( l + s ) , Po = 114 Example 5.7,numerical results with a ( s ) = (2+s2)/(1+ s2), Po = 7/16 Example 5.8, numerical results with K = 0.20488 Example 5.8, numerical results with K = 10-I Example 5.8, numerical results with K = Example 5.8, numerical results with K = l o w 3 Example 5.8, numerical results with K = Example 5.9, convergence of KaEanov iterates Example 5.10, convergence of KaEanov iterates

xii

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

5.11 5.12 6.1 6.2

Example 5.11, KaEanov iteration for a quasi-Newtonian flow problem, qo = 1,q, = 0.0001 Example 5-11,KaEanov iteration for a quasi-Newtonian flow problem, qo = 100,q, = 0.01 Nodes and weights of a7-point Gauss-Legendre quadrature formula over the reference triangle Example 6.11, numerical values of constants

226

226 263 268

Dedicated to

Preface This work provides a posteriori error analysis for mathematical idealizations in modeling boundary value problems, especially those arising in mechanical applications, and for numerical approximations of numerous nonlinear variational problems. An error estimate is called a posteriori if the computed solution is used in assessing its accuracy. A posteriori error estimation is central to measuring, controlling and minimizing errors in modeling and numerical approximations. In this book, the main mathematical tool for the developments of a posteriori error estimates is the duality theory of convex analysis, documented in the well-known book by Ekeland and Temam ([49]). The duality theory has been found useful in mathematical programming, mechanics, numerical analysis, etc. The book is divided into six chapters. The first chapter reviews some basic notions and results from functional analysis, boundary value problems, elliptic variational inequalities, and finite element approximations. The most relevant part of the duality theory and convex analysis is briefly reviewed in Chapter 2. This brief review is sufficient for the applications of the duality theory in all the following chapters. In mathematical modeling of differential equation problems, usually assumptions are made on various data. Qualitatively, for many problems, it is known that the solution depends continuously on the problem data. Frequently though, it is desirable also to estimate or bound quantitatively the effect on the solutions of the problems caused by the adoption of the assumptions on the data. In Chapter 3, a posteriori error estimates are derived for the effect on the solutions of mathematical idealizations on the data of elliptic linear boundary value problems. In Chapter 4, a posteriori error estimates are given for linearization in a number of nonlinear boundary value problems. The last two chapters are devoted to a posteriori error analysis of numerical solutions. In Chapter 5, the regularization method and the KaEanov method are considered, both being useful in handling certain types of nonlinearity. In Chapter 6, a posteriori error estimates are derived and studied for finite element solutions of some elliptic variational inequalities. This book is intended for researchers and graduate students in Applied and Computational Mathematics, and Engineering. Mathematical prerequisites include calculus, linear algebra, some exposures of differential equations, and concepts of normed spaces, Banach spaces and Hilbert spaces. In the theoretical development, some basic notions and results in functional analysis, duality thoery, weak formulations of boundary value problems, variational inequalities, and the finite element method are used. Brief reviews of these notions and

xvi

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

results in the first two chapters provide background materials for a reader who lacks knowledge in these areas. This work avoids giving the results in the most general, abstract form so that it is easier for the reader to understand more clearly the essential ideas involved. Many examples are included to show the usefulness of the derived error estimates. In preparing this book, I have benefited from many individuals. I am grateful to Professor Ivo BabuSka for introducing me the research topic and for providing valuable advice. Several of my collaborators (teachers, friends, and students) made contributions to various parts of the book, and I especially thank Dr. Viorel Bostan, Dr. Jiuhua Chen, Professor Hongci Huang, late Professor Sgren Jensen, Professor B.D. Reddy. I express my gratitude to Professor Kendall Atkinson and Professor Mircea Sofonea for their constant support. I thank Professor D.Y. Gao and Professor R.W. Ogden for inviting me to make the contribution in their Kluwer book series on Advances in Mechanics and Mathematics (AMMA). The supports of NSF under grant DMS-0106781 and the James Van Allen Fellowship of the University of Iowa are greatly appreciated.

Chapter 1

PRELIMINARIES

1.1.

INTRODUCTION

Numerical simulation/scientific computation is now playing a more and more important role, and has become one of the three basic tools in science and technology, in addition to experimentation and theory. Numerical simulation provides a relatively inexpensive and efficient way to help understanding the physical world and advancing the technology. A complete numerical analysis simulation session for a physical or engineering problem typically consists of several steps, described below. See Figure 1.1 for a description of the related flow chart, following [ 7 ] . First, the physical or engineering problem is brought to our attention. We want to predict and determine the response of the physical system to the external actions. To do this we need to establish a mathematical model for the problem. This is achieved by applying physical laws, material constitutive relations, and various experimental data such as the geometry of the system, densities of external forces. Most often, we obtain an initial-boundary or boundary value problem of differential equations or differential inequalities to describe the physical or engineering problem. We call this mathematical model the basic mathematical model, and identify it with the physical reality. It is a highly idealized assumption that we can have a mathematical problem which exactly describes the physical problem. The available data, which usually come from experiments, for the basic mathematical formulation can not be obtained as accurate as one wishes. As a consequence, we solve a simplified, or idealized mathematical problem instead. The idealized or simplified problem is the mathematical model we use to study the physical problem. The idealized mathematical model is usually still rather complicated and can be solved only by numerical methods. Popular methods to discretize initial-

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Physical Problem 1

Basic Mathematical Model Mathematical Idealization

-f

1

Error ~

n

a

m

4

Output Figure 1.1. Numerical analysis of a physical problem

boundary or boundary value problems include the finite difference method and the finite element method. As a result of the discretization, we obtain a discrete system. The discrete system is then solved by some numerical method. Once we have solved a discrete system, a natural question is: Can we use this solution? In other words, is this discrete solution sufficiently accurate for practical use? The reliability of a numerical solution of a physical or engineering problem depends on mathematical idealization of the physical problem and numerical treatment of the idealized mathematical problem. Various possibilities may arise and they demand our closer investigation. It is desirable to be able to estimate the errors associated with the steps described above. Standard topics in error analysis deal with the errors caused by discretization and solution of the discrete system. Fewer results are available for the estimation of errors in mathematical modelling. The error in mathematical modelling may be the most critical, however. If both the mathematical idealization and the numerical solution of the idealized problem are reliable, various information for the real problem is drawn based on the numerical solution of the idealized problem. If a discrete solution is found to be not accurate enough, we need to trace the sources of the inaccuracy, and decide whether we need to compute a more accurate numerical solution of the mathematical model, or refine and numerically solve a new mathematical model. There is a large amount of literature on numerical methods and their error analysis. Relatively few results are available in the literature for reliability analysis of mathematical idealizations, especially most desirably, certain good, practically useful quantitative assessments of the quality of solutions of idealized problems. Such quantitative assessments should be (hopefully) available once we have computed the solutions of idealized problems. One should not assume the quantitative knowledge of the solutions of basic mathematical problems for either exact descriptions of basic mathematical problems are usually not available in practice or, it is often too expensive to solve the basic mathematical prob-

Preliminaries

3

lems. Some papers devoted to modeling error analysis or error analysis of mathematical idealizations include [30, 71, 72, 116, 118, 119, 120, 121, 128, 1421. An error estimate is called a posteriori if the computed solution is used in assessing its accuracy. A posteriori error estimation is central to measuring, controlling and minimizing errors in modeling and numerical approximations. We now briefly describe the main features of the a posteriori error estimates to be derived and studied in this work. We let (P)stand for the basic mathematical model problem and use u for its solution, and let (Po)be an idealized mathematical problem with the solution uo, and ( ~ ta )numerical approximation of the idealized model with the solution u!. Here, h represents a discretization parameter. The basic mathematical model ( P ) is usually difficult to solve, even numerically, and the simpler problem (Po) is expected to be close to ( P ) . We want to use the solution uo to bound the error l u - uo 11:

where B ( u o ) is a quantity completely computable once uo and some information on the data of ( P ) are known. We also allow the case where the data for ( P ) are not completely given, and only some ranges of the data are available. In such a case, the data for ( P o ) can be obtained through certain averaging process on the data for (P). Of course, to be able to derive a posteriori error estimates, we need to make assumptions on the structure of the problem ( P ) . In this work, ( P ) is assumed to be a convex minimization problem; this allows the employment of the duality theory in convex analysis for deriving a posteriori error estimates. We will also use the numerical solution u! of the problem (P:) to bound the error 11 uo - u! 11. A posteriori estimation of the discretization error (uo - uk) has been a popular research topic since late 1970's (see the description in Chapter 6). Many of the a posteriori error estimates for the numerical solutions of differential equation problems can be derived via the duality theory. In Chapter 6, we focus on the a posteriori error analysis for finite element solutions of elliptic variational inequalities of the second kind. We will see there that the duality approach provides a general framework, leading to various a posteriori error estimators. In this work, errors are measured in terms of energy norms or energy-like norms. The salient features of the a posteriori error estimates presented in this work are: 1 The error estimates are rigorous in the sense that the error bounds are always satisfied. In this sense, when we say error estimates, most often we mean error bounds. We use the phrases "error estimates" and "error bounds" interchangeably.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

2 The error bounds are determined by the solutions of the idealized mathematical models or the numerical solutions of the given problems. There is no need to solve related dual problems.

3 All the error bounds, except those presented in Chapter 6 on the finite element approximations of variational inequalities, are completely computable in the sense that there are no unknown constants. In the literature, most a posteriori error estimates on numerical solutions of partial differential equations problems involve such theoretical unknown constants and their values are selected based on a few numerical examples (e.g., 1501).

4 Efficiency of the error estimates is demonstrated through numerous examples and theoretical analysis. Recently, goal-oriented or object-oriented error estimates have been developed to calculate error bounds of global or local quantities of interest, such as the error of stress or strain in a critical region, using a dual-weighted residual technique. The procedure of the technique can be described as follows. Consider a boundary value problem and suppose the purpose of the computation is the value of a functional of the solution. The boundary value problem is solved numerically, typically by the finite element method. The functional of the numerical solution is then computed. To bound the error involved in the functional value, a dual problem is introduced related to the functional and the boundary value problem, and is numerically solved. Then the error in the computed functional value is expressed in terms of the numerical solution of the dual problem together with some residual quantities, which is then localized and split into contributions related to the modeling error and discretization error. The survey papers [19, 621 provide detailed accounts of this technique. Note that the use of solutions of the dual problems may imply a possibly dramatic increase in the computational effort. In this work, dual problems will play a central role in the derivation of error estimators, but the error bounds do not involve solutions of the dual problems. The organization of the book is the following. In the remaining part of this first chapter, we will briefly review some basic notions and results from functional analysis, function spaces, weak formulation of boundary value problems, and the finite element method. In Chapter 2, we review some basic material from convex analysis and the duality theory, that plays the central role in this book for a posteriori error analysis. In Chapter 3, we employ the duality theory to derive a posteriori error estimates for mathematical idealizations in linear boundary value problems, paying particular attention to the situation with nonsmooth domains. The idealizations can occur in the coefficients of the differential equations, the right-hand sides, boundary value conditions, and the domain. In Chapter 4, we perform a posteriori error analysis for the effect of linearization in several nonlinear problems. In Chapter 5, we apply the duality theory to

Preliminaries

5

derive a posteriori error estimates for some numerical procedures in solving nonlinear boundary value problems, including the regularization method for problems involving non-differentiable terms, KaCanov iteration methods and linearizations. Finally, in Chapter 6, we derive a posteriori error estimates for finite element solutions of elliptic variational inequalities of the second kind. The error estimates that can be derived via the duality theory include some of the well-known a posteriori error estimates found in the finite element literature for solving elliptic differential equations.

1.2.

SOME BASIC NOTIONS FROM FUNCTIONAL ANALYSIS

We assume the reader is familiar with such basic notions as linear spaces, norms, inner products, Banach spaces, and Hilbert spaces. Details on these and the material to be reviewed in the following can be found in any standard textbook on functional analysis, e.g., [45, 481, or in a concise form, [6]. In this work, the general theory will be developed for domains in the space IRd of the d-dimensional vectors of the form x = (xl,. . . , x ~ )xi~ E, R. Recall that a domain R c Rd is an open, connected, bounded set in E X d . For p E [I,m], we have the following norms in IRd:

When Rd is viewed as a normed space, implicitly we understand the norm to be the Euclidean norm I . / = / . 12, unless otherwise stated. The Euclidean norm is induced by the canonical inner product in EXd:

The summation convention over a repeated index will be adopted. As an example, for the canonical inner product in IRd, we write (x,y) = xigi. The symbol sdstands for the space of second order symmetric tensors on Rd or, equivalently, the space of symmetric matrices of order d. The inner product and corresponding norm on S d are

For a normed space V ,we use V*to denote its dual space, i.e., the space of all the continuous linear functionals on V .The duality pairing between V* and V is usually denoted by e(v) or (v*,v) for e,v* E V*and v E V .We can then introduce different types of convergence.

6

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

DEFINITION1.1 Let V be a normed space, V* its dual space. A sequence {u,) C V converges or converges strongly to u E V, written u, + u as n+m,if lim llu - unll = 0. n+cc

The sequence {u,) converges weakly to u E V, written u,

--\

u as n + m, i f

We will use the following property of a weakly convergent sequence:

The dual space of a normed space is always complete, i.e. always a Banach space. Over a finite dimensional space, it is a well-known result that any bounded sequence contains a convergent subsequence. This property does not carry over to infinite dimensional spaces. For example, the sequence {sin j ~ x ) is ~ a>bounded ~ sequence in ~ ~ (I )0, but , none of its subsequences converges. I n many applications of the functional analytic approach, one needs the property that a bounded sequence contains a subsequence that converges in some sense. Reflexive Banach spaces enjoy this kind of desirable property. A space V is said to be rejlexive if (V*)*can be identified with V. A reflexive space must be complete and is hence a Banach space. We have the following important property of a reflexive space.

THEOREM 1.2 If V is a rejlexive Banach space, then any bounded sequence in V has a weakly convergent subsequence. We will see examples of reflexive Banach spaces in Section 1.3. For an inner product, there is an important property called the CauchySchwarz inequality:

with the equality holding iff u and v being linearly dependent. We recall an important property of a Hilbert space.

THEOREM 1 . 3 (Riesz representation theorem) Let V be a Hilbert space, 1 E V*. Then there is a unique u E V for which

In addition, lltll = Il.llv. Thus, the dual space of a Hilbert space can be identified with itself, and any Hilbert space is reflexive.

7

Preliminaries

1.3.

FUNCTION SPACES

We will use the multi-index notation for partial derivatives. An ordered collection of d non-negative integers, a = ( a l , . . , a d ) , is called a multiai is said to be the length of a. If v is an index. The quantity ( a (= m-times differentiable function, then for any a with la1 5 m,

c:=,

is the ath order partial derivative. For lower order partial derivatives, there are other notations in common use; e.g., the partial derivative dv/dxi is also written as dx,v, or div, or vjXi,or v,i.

1.3.1

CONTINUOUS FUNCTION SPACES

The notation ~ ( 2 is used ) for the space of functions continuous on 2. It is a Banach space with the norm

More generally, for a non-negative integer m, we define C m ( 2 ) = {v E ~ ( 2 : )D"v E C ( 2 ) for la1

m),

which is a Banach space with the norm

We also set ~ ~ (= 2n;==,crn(2) )

r {v E C ( 2 ) : v

E C m ( 2 ) 'dm = 0,1,. ..)

Given a function v on R, its support is defined to be suppv = {x E R : v ( x ) # 0). Here a bar over a set stands for the closure of the set. We say that v has a compact support if supp v is a proper subset of R: supp v CC R. Thus, if v has a compact support, then there is a strip about the boundary dR such that v is zero on the intersection of the strip and the domain. Later on, we will use the space C,CO(R) = {v E CCO(R): supp v cc 0).

Holder spaces. A function v defined on R is said to be Lipschitz continuous if for some constant c,

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

8

The smallest possible constant in the above inequality is called the Lipschitz constant of v , and is denoted by Lip(v). More generally, the function v is said to be Holder continuous with exponent /I E ( O , l ] if for some constant c,

(a)

(32) is defined to be the subspace of C which consists The Holder space CO)P of functions Holder continuous with the exponent P. With the norm

the space C O ) P ( ~ ) becomes a Banach space. For a nonnegative integer m and /I E (O,l],we similarly define the Holder space

C~)~= ( E{ V) E crn(a) : D f f vE C

O . ~ ( ~ for )

all u with la. = m } ;

this is a Banach space with the norm

1.3.2

LEBESGUE SPACES

In the study of Lebesgue spaces, we identify functions which are equal a.e. on R. For p E [ I , oo),LP(R) is the linear space of measurable functions v : R -+ R such that

The space L m ( R ) consists of all essentially bounded measurable functions v : R + R such that l l ~ l l l ~ r n= ( ~ )

inf

SUP

Iv(x)I

meas(nl)=O2 ~ n \ n f

< 00.

For a measurable function v defined on R, if v E Lp(Rf)for any R' C C R, then we say v is locally in LP(R) and write v E Lro,(R). We use meas(R)for the Lebesgue measure of R. For d = 3, meas(R) is the volume of R, and for d = 2, meas(R)is the area of R. Some basic properties of the LP spaces are summarized in the following theorem.

9

Preliminaries

THEOREM 1.4 Let R be an open bounded set in IRd. (a) For p E 11, oo],LP(R) is a ~ a n a c hspace with the norm dejined in (1.2)or (1.3). (b) For p E [ I , m ] ,every Cauchy sequence in LP(R) has a subsequence which converges pointwise a. e. on R. (c) If1 5 p 5 q oo, then L Q ( f l )c LP(R),

<

(d) I f 1 5 p 5 r 5 q O/p (1 - Q ) / q ,then

+

< oo and we choose O

E [O,1]such that l / r =

(e) For 1 5 p < m , (LP(R))*= LP*( f l ) . Hence for p E (1,m ) , the space Lp(R) is rejlexive.

1.3.3

SOBOLEV SPACES

Sobolev spaces are defined based on the concept of weak derivatives.

DEFINITION1.5 Let R be a nonempty open set in IRd, v , w E L;,,(R). Then w is called an ath weak derivative of v i f

It follows from the definition that a weak derivative, if it exists, is unique up to a set of measure zero. Another direct consequence of Definition 1.5 is that if a function v has a continuous ath derivative Dav in the classical sense, then Dav is also the athweak derivative of v. Thus we see that the notion of a weak derivative is indeed an extension of the classical derivative. For this reason, we will use the symbol Dav for the ath weak derivative of v. Let Ic be a nonnegative integer, p E [ I , oo].

DEFINITION 1.6 The Sobolev space wkJ'(R) is the set of all the functions v E L:,,(R) such that for each multi-index a with la1 5 Ic, the ath weak derivative Dav exists and D f f v E Lp(R). The norm in the space wkJ'(fl) is defined by the following:

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 1.2. Smoothness of the boundary

When p = 2, we write H k ( R )= w ~ ) ~ ( R ) . We will mainly use Sobolev spaces wk1p(fl) when f l is a domain in EXd. Usually we use simpler notation / ( . 11 k , p , R to denote the norm )I . 11 W h p ( R ) , and I . I k , n for the norm I . I H k ( n ) . The Sobolev space WkJ'(R) is a Banach space, and H k ((R is a Hilbert space with the inner product

Next we list several important properties of Sobolev spaces. Some properties require a certain degree of smoothness of the boundary J? of the domain R.

DEFINITION1.7 Let R be open bounded in iRd, and let V denote afunction space on JRdV1. We say d R is of class V iffor each point xo E X l , there exist an r > 0 and afunction g E V such that upon relabelling the coordinate axes ifnecessary, we have

Here, B ( x o ,r ) denotes the ball centered at xo with radius r. See Figure 1.2. In particular; when V consists of Lipschitz continuous functions, we say R is a Lipschitz domain. When V consists of Ck functions, we say R is a ck domain. Since d R is a compact set in IKd, we can actually find a finite number of points {xi):=, on the boundary so that for some positive numbers {ri):==,and functions {gi)l=l C V ,

upon relabelling the coordinate axes if necessary, and

Preliminaries

Figure 1.3. Smooth domains

Figure 1.4. Lipschitz domains

Figure 1.5. A crack domain

Throughout this work, we will assume R is Lipschitz continuous, unless stated explicitly otherwise. We observe that in engineering applications, most domains are Lipschitz continuous (Figures 1.3 and 1.4). A well-known nonLipschitz domain is one with cracks (Figure 1.5).

Approximation by smooth functions. Equalities and inequalities involving Sobolev functions are usually proved first for smooth functions followed by a density argument. A theoretical basis for this technique is results on density of smooth functions in Sobolev spaces.

T H E O R E1.8 M Assume R is a Lipschitz domain, 1 5 p < m. Then for any v E wk2p(R),there exists a sequence {v,) C P ( n )such that

Proof of this density theorem can be found, e.g., in [51].

12

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

DEFINITION1.9 We define w / ' P ( R )as the closure of C r ( S 1 )in wklp(0). When p = 2, we denote H$ (a) = (a).

~ i > ~

From Definition 1.9, we see that we always have the denseness of the smooth to be the space of function space Corn ( R ) in w;'P(R). We interpret w;'P(o) ) the "property" that all the functions v in w k > p ( Rwith

D a v ( z ) = 0 on d R , V a with ( a (5 k

-

1.

The meaning of this statement will be made clear later after the trace theorems are presented.

Traces. Notice that Sobolev spaces are defined through Lebesgue spaces. Hence Sobolev functions are uniquely defined only a.e, in R. Since the boundary dR has measure zero in Itd,it seems the boundary value of a Sobolev function is not well-defined. Nevertheless it is possible to define the trace of a Sobolev function on the boundary in such a way that for a Sobolev function continuous up to the boundary, its trace coincides with its boundary value.

THEOREM 1.10 Assume R is a Lipschitz domain in IRd, 1 5 p < m. Then there exists a linear operator y : wlJ'(R) -+ LP(dR) such that (a) yv = vIan $v E wlJ'(R) n C ( 2 ) . (b) For some constant c > 0, I l y ~ l l ~ p5( c~~~ w ) i , p b'v ( ~ E) W ' > ~ ( R ) . (c) The mapping y : wlJ'(R) -+ LP(dS2) is compact; i.e., for any bounded sequence {v,) in wlJ'(R), there is a subsequence {v,t) c {v,) such that {yv,t ) is convergent in LP(dR). In Theorem 1.10, the Lebesgue spaces on the boundary, LP(dR),are used. A precise definition of these spaces can be found in [101, Section 6.31. The operator y is called the trace operator, and with property (a) we can view yv as the the generalized boundary value of v. Property (b) states that the mapping y is continuous from w 1 ~ P ( Rto) LP(dR),whereas property (c) further states that actually the mapping y is compact from W1sp(R)to Lp(dS1). The trace operator is neither an injection nor a surjection from w 1 > p ( Rto) LP(dR). The range y (wlJ'(R)) is actually a space smaller than LP(dR),that an) example , of a fractional order Sobolev space. is denoted by ~ ~ - l l p , p ( d R This is a Banach space with the norm

. the future We will frequently use the space H1I2( d ~ =) y ( ~ ' ( 0 ) ) In to simplify the notation, for a function v E w13P((R),we will denote its trace on the boundary also by v. The trace operator is also continuous from

13

Preliminaries

wlJ'(R) to L Q ( R )for q in certain range. One such example is: for d = 2, y E C ( H 1 ( R )H, q ( d R ) )for any q E [I,oo), andforsomeconstantc = c ( R ,q), we have the inequality

With the notion of the trace of wlJ'(R) functions, we have

Here, the condition "v = 0 on dR" is understood as that the trace of v is a zero function on dR. This condition can be equivalently stated as "v = 0 a.e. on do". The particular case p = 2 leads to the Hilbert space

H; ( 0 )= {v E H 1( R ) : v = 0 on 8 0 ) . Its dual space is usually denoted as H-' ( 0 ) . In some occasions, we will need to use the first derivatives (e.g., the normal derivative) on the boundary. Let v = ( v l ,. . . , vd)T denote the outward unit normal to the boundary r of R. Since dR is Lipschitz continuous, v exists a.e. on dR. For a function v E H 2( R ) ,its first derivatives vxi E H 1 ( R ) . By Theorem 1.10, it is meaningful to write their traces vxi E L~( d R ). We can then define the normal derivative

and we have the integration by parts formula

This formula is first proved for u,v E cm(R),and is then extended to u E H 2 ( R )and v E H1( R )by applying the density result Theorem 1.8. For boundary value problems of higher-order partial differential equations, we need to use the traces of partial derivatives on the boundary. Such results can be found in, e.g., [67].

Sobolev embedding theorems. Sobolev embedding theorems are important, especially in analyzing the regularity of a weak solution of a boundary value problem.

DEFINITION 1.1 1 Let V and W be two Banach spaces with V C W . We say the space V is continuously embedded in W and write V v W ifthere exists a constant c > 0 such that

14

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

We say the space V is compactly embedded in W and write V 3- W , if ( 1 . 5 ) holds and each bounded sequence in V has a convergent subsequence in W. If V 9W , the functions in V are smoother than the rest of the functions in W . Proofs of most parts of the following two theorems can be found in [51]. The first theorem is on embedding of Sobolev spaces, and the second on compact embedding. These theorems are given in a form more general than what is needed later in this work. For any t E R, we denote [t]the largest integer less than or equal to t .

THEOREM 1 . 1 2 Let R c IRd be a Lipschih domain. Then the following statements are valid. (a) I f k < d l p , then wk2P(R) c, L q ( R ) for any q 5 p', where l l p ' = l / p - k/d. (b) I f k = d / p , then w k > p ( Rq ) L q ( R )for any q < cm. (c) I f k > d / p , then

where

4={

[d/pl+ 1- d / p any positive number

<1

if d l p # integer, i f d l p = integer.

THEOREM 1.13 Let 0 C IRd be a Lipschih domain. Then the following statements are valid. (a) I f k < d l p , then wkJ'(R) v* L 4 ( R ) for any q < p', where llpl = l l p - kld. (b) I f k = d l p , then wkJ'(R) qq L q ( R )for any q < oo. (c) I f k > d l p , then

where 4

E

+

[0,[ d l p ] 1 - d l p ) .

Equivalent norms. Associated with any vector space, one can define infinitely many different norms. A well-known result in analysis states that over a finite dimensional space, any two norms are equivalent, i.e., if I/ . and /I . 11(2) are two norms on a finite dimensional space V,then there exist two constants 0 < cl 5 c2 < oo such that

Thus, in a finite dimensional space, different norms lead to the same convergence property. For an infinite dimensional space, however, not any two norms are

15

Preliminaries

equivalent. For example, over the space C[O,11, the norm 11 . 11 c[o,ll is stronger than /I . I L l ( 0 , 1 ): A sequence converging with respect to the norm / //c[o,ll converges also with respect to the norm (1. ( I L 1 (0,1),whereas it is easy to construct examples of function sequences converging with respect to I/ . l i L l ( o , l ) but not with respect to I . I C I O , I l (cf. [6, Example 1.2.131). In some applications, it is convenient to use a different norm that is equivalent to the canonical norm of a given space. The next result can be used to generate various equivalent norms on Sobolev spaces. Over the Sobolev space wk>p(fl), I ~ l ~ , p ,is$ the seminorm defined bv

It can be shown that if f l is connected and IvlklPln= 0,then v is a polynomial of degree less than or equal to k - 1.

THEOREM 1.14 Let R c lRd be a Lipschitz domain, k 2 1, 1 ( p < m. Assume f j : wkJ'(fl) i R, 1 j ( J, are seminorms on wk>p(0) satisfying two conditions:

<

( H I )0 I f j ( v ) 5 c llvllrc,p,nv v E w k 2 ~ (1nI )j, 5 J . (Hz) If v is a polynomial of degree less than or equal to k f j ( v ) = 0, 1 ( j ( J , then v = 0.

-

1 and

Then, the quantity J

defines a norm on wkJ'(fl), which is equivalent to the norm

Il~l)~,~,$.

A proof of this result can be found in [79]. Many useful inequalities can be derived as consequences of the previous theorem. As an example, if rl is an open, non-empty subset of the boundary d f l , then there is a constant c > 0,depending only on f l such that

This inequality can be derived by applying Theorem 1.14 with k = 1, p = 2, J = 1 and r

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

16

where H&(R) = {v E H1(R) : v = 0a.e. o n r l ) . is a norm on H;~(R), equivalent to the usual Hence, the seminonn I . jl,' H1( 0 )-norm 1) 11 1,'. The inequality (1-6) is a PoincarC-type inequality. In and I - l , n will be used as the norm in Hhl (0). what follows, both I

1.4.

WEAK FORMULATION OF BOUNDARY VALUE PROBLEMS

Classical solutions of boundary value problems of partial differential equations may not exist even for smooth data (cf. [67, Section 1.11). The development of the theory of Sobolev spaces and weak formulations eliminates this problem and provides a general framework to derive powerful numerical methods. In this section, we briefly discuss weak formulations of boundary value problems. We start with the homogeneous Dirichlet boundary value problem of the Poisson equation

The standard weak formulation is

Here we assume f E ~ ~ ( 0 If f) .E HP1(R), then the right hand side of the equation is understood to be the duality pairing ( f , v ) between the spaces H-' (R) and (R). The weak formulation (1.9) is derived from (1.7)-(1.8) as follows: First, multiply the equation (1.7) by an arbitrary function v , suitably smooth so that the following calculations are justified, and assumed zero on the boundary. Next, integrate the resulting equation over R, and perform an integration by parts on the left hand side to obtain the integral identity

HA

S,

~ V U - V v d z =fvdz Finally, observe that both sides of the integral identity make sense even if we only require u , v E H1(R). Together with the zero boundary value condition, we thus require u , v E Hi ( a ) . Relations between the classical formulation (1.7)-(1.8) and the weak formulation (1.9) are:

17

Preliminaries

1 If u E C 2 ( R ) n C(2) is a classical solution of (1.7)-(1.8), i.e. (1.7)and (1.8)are satisfied at every point in R and on dR, respectively, then it is also a weak solution of (1.9). 2 If u E H; (R) is a weak solution of (1.9)and u E C 2 ( R )n C ( 2 ) , then it is also a classical solution of (1.7)-(1.8). Throughout this work, a boundary value problem written in the classical form will be understood in the sense of the corresponding weak formulation. In other words, we will understand the problem (1.7)-(1.8)to mean (1.9). Now consider a boundary value problem for the equation (1.7)with a nonhomogeneous Dirichlet boundary condition

Here we assume g is the trace of an H1(R) function, also denoted by g. We express this assumption as g E H1(R). Then the standard weak formulation for the boundary value problem (1.7)and (1.10)is

Setting u = uo + g,we can transform the problem (1.11)into one of seeking uo such that

In principle, non-homogeneous Dirichlet boundary conditions can always be rendered homogeneous in the way mentioned above. More general mixed boundary value conditions can also be considered. Assume d o is split into two parts dR = rl U with rl relatively closed, r2 relatively open, n r2 8 and meas(I'1) > 0. Let g E L ~ ( I ' ~be) given. We consider a boundary value problem for the equation (1.7)with the mixed boundary value condition

+

Denote

H&(0)= {v E H1(R) : v = 0 a.e. on r l ) . Then the standard weak formulation for the mixed boundary value problem (1.7)and (1.13)is

18

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

It is convenient to study the weak formulations of the boundary value problems within a general framework. Let V be a Hilbert space. We introduce the following definition.

DEFINITION 1.15 Let a : V x V -+ R . We say a(., .) is a bilinear form on V i f it is linear with respect to each argument. We say the bilinear form a ( . , .) is continuous or bounded ifthere exists a number M > 0 such that

The bilinear form a ( - ,.) is V-elliptic ifthere is a constant m

> 0 such that

and a ( . , .) is symmetric if

a ( u ,v ) = a ( v ,u ) V u , v E V. For a ( . , .) : V x V

+ R a bilinear form and t E V", consider the problem

We have the well-known Lax-Milgram lemma for the existence and uniqueness of a solution of the variational problem (1.17).

THEOREM 1-16 (Lax-Milgram Lemma) Let V be a Hilbert space. Assume a ( . , -) is a bounded, V-elliptic bilinear form on V, t! E V*. Then there is a unique solution to the problem (1.17). Note that the problems (1.9), (1.12) and (1.14) are all of the form (1.17). With the Lax-Milgram lemma, it is easy to show that these problems all admit a unique solution. The V-ellipticity of the associated bilinear form

is deduced from the PoincarC inequality (1.6). The Lax-Milgram lemma can be applied for an existence and uniqueness study of more general linear elliptic partial differential equations. Let R c Rd be a Lipschitz domain, its boundary dR being decomposed as dR = rDU with FD c dfl relatively closed, FN = dR\rD relatively open, and rDnrN= 0.Consider the boundary value problem

19

Preliminaries

Here v = (y, . . . , ud)T is the unit outward normal on rN. We assume the following conditions on the given functions aij, bi, c, f , and 9: aij, bi, c E L m ( 0 ) ; there exists a constant a0

(1.19)

> 0 such that

>

f

a i j t i t j a o 1 ~ 1 ~Y e = (&) E IRd, a.e. in 0; E L2(0), g E L 2 ( b ) .

(1.20) (1.21)

The weak formulation of the problem (1.18) is (1.17) with

We can apply Lax-Milgram Lemma to study the well-posedness of the boundary ~ is a Hilbert space, with the standard value problem. The space V = H ; (0) H 1(0)-norm. The assumptions (1.19)-(1.21) ensure that the bilinear form a ( . , .) of (1.22) is bounded on V, and the linear form is bounded on V. Various sufficient conditions for the V-ellipticity of the bilinear form can be obtained. As an example, denoting b = ( b l , . . . , bd)T, one can verify that any of the following three conditions is sufficient for the V-ellipticity: c or b. v

> co > 0, lbl < B a.e. in 0, and B~ < 4 ~ 0 ~ 0 ,

> 0 a.e. on rN,and c - -21 d i v b > co > 0 a.e, in 0,

or meas(rD) > 0, b = 0, and infc > -ao/c, n where c is the best constant of the PoincarC inequality

A constant E is said to be the best constant of the inequality (1.23) if the inequality holds with c = c and it does not hold for any c < E. The best constant of (1.23) can be computed by solving a linear elliptic eigenvalue problem:

20

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

with XI

> 0 the smallest eigenvalue of the eigenvalue problem

A special and important case is where b = 0.In this case, the bilinear form is symmetric, and V-ellipticity is assured if

BEST CONSTANTS IN SOME SOBOLEV INEQUALITIES Let R c IRd be a Lipschitz domain, rl a subset of dR with a positive

1.5.

measure. There are a lot of papers and monographs devoted to the study of Ilv 11 L 2 (rl, 11 V v 11 L 2 ( n ) , inequalities among norm quantities such as Ilv 11 L2 11 dvldvl)L2(r,),//Awl/L2(n)for functions v satisfying various smoothness and auxiliary conditions. For simplicity, in this work, we use IlVv 11 L z c n ) to mean IIVVll(L2(n))d, and we will use both notations. In [I391 one can find a list of 15 inequalities among these quantities for the case when r1= dR. Best possible constants in these inequalities are related to smallest positive eigenvalues of various linear elliptic eigenvalue problems. One example is (1.25) for the best constant (1.24) in the inequality (1.23). The reader is referred to [85, 86, 1031 and references therein for eigenvalue estimations and relations for eigenvalues of different eigenvalue problems. More generally, one can study inequalities between LP-norm of some quantity of v and LQ-normof another quantity of v , for functions v satisfying certain smoothness and auxiliary conditions. When either p or q # 2, the study of the best possible constant in such an inequality no longer leads to a linear eigenvalue problem. For certain specificp and q, we are led to linear elliptic boundary value problems. As an example, we consider the best possible constant in the following trace inequality:

Here R0 is a measurable open subset of R and meas(rl) # 0.This inequality follows directly from the inequality (1.6). We are interested in the computation of the best constant co of the inequality (1.26), characterized by the relation

Let us try to determine this best constant,

Preliminaries

First we have an existence result for the best constant.

L E M M A1.17 There exists 0 f u E H : ~( R )such that

Proof. We choose a sequence {u,) C H i l ( R )with the properties

and I / u n I I ~ l ( ~-o ))CO

as n

--?,

m.

Since {u,) is bounded in H& (O), we can apply Theorem 1.2 on the reflexive Banach space H : ~( R )and Theorem 1.13 for the compact embedding of H& ( 0 ) in L ' ( R ) to conclude the existence of a subsequence {un,} and u E H+l ( 0 ) such that

Thus we have (cf. (1.1))

Therefore,

.

By the definition of the constant co, the above inequality is an equality. Now we perform a formal calculation to derive a formula for the computation of co and then prove the formula rigorously. Let u E Hhl ( 0 )be the function specified in Lemma 1.17. For any v E Hhl ( R )the real function

is defined in a neighborhood of t = 0 and has its maximum at t = 0. We observe that if we keep the value of 11 u 11 L~ fixed, then the value 11 Qu I I L 2 ( n )

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

will be minimal when the function u does not change rapidly. So we expect an optimal function u should keep its sign. If this is true, then we will have

Hence, if we normalize u with the condition

then u E H& ( f l ) is the solution of the problem

This problem has a unique solution by the Lax-Milgram lemma (Theorem 1.16). So the best constant is expected to be

Let us show that these formulas indeed provide the best constant. For this purpose we need a preliminary result (see, e.g., [61]).

LEMMA1.18 I f v E H 1 ( f l ) ,then Ivl E ~ ' ( f l and ),

i

Vv

Vlvl =

0 -Vv

i f v > 0, i f v = 0, i f v < 0.

From this lemma, we immediately obtain the next result.

LEMMA1.19 I f v E H;l ( f l ) ,then J v JE Hhl ( f l ) and

THEOREM 1.20 Let u E H i 1( 0 )be the solution of the problem (1.27). Then the best constant co of the inequality (1.26) is given by

Proof. For any v

E

H;; (O), by Lemma 1-19,we have Iv 1 E Hhl ( f l ) and

Preliminaries

Therefore,

This completes the proof. Later in Subsection 3.5.2, we will use Theorem 1.20 in the derivation of an a posteriori error estimate in the study of material idealization for a torsion problem in linearized elasticity. We now examine some concrete examples.

EXAMPLE 1 . 2 1 Let d

> 2. Assume

is the d-dimensional open ball of radius Ro centered at 0, Ro = B!:) is the ro Ro, and rl= The d-dimensional ball of radius ro centered at 0, related problem (1.27) is the weak form of the boundary value problem:

<

a~gi.

together with the interface condition that both u ( x ) and d u ( x ) / d l x l are continuous across the sphere 1x1 = ro. For d = 2, the solution is

For d

> 2, the solution is "I2

--

u(x) =

+

2d rf d (d - 2)

3 2d

+

o'd

d ( d - 2)

(ri-d-~i-d) i f 0 5 1x1 < T O ,

~i-~)

( 1 ~ l 2 - ~

i f r o 5 1x15 RO

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

24 and so

Here r ( . ) denotes the gamma function (cf. e.g., [104]). Hence, we have the optimal inequality

and

In particular, when ro = Ro, we have the optimal inequality

for any dimension d 2 2.

rn

EXAMPLE 1.22 Assume R = Ro = n $ , ( 0 , a j ) , T1 = dR. The corresponding problem (1.27) is the weak form of the boundary value problem

The solution u ( x )of the boundary value problem is

Hence, the best constant co = I ~ V U I Iof~the ~~ corresponding ~) Sobolev inequality equals

25

Preliminaries

For a general domain, it is usually impossible to find a closed form for the best constant co. In this case, numerical methods can be used to determine the constant. An example of this kind can be found in [68]. Depending on the applications, best constants of other Sobolev inequalities may be useful. In [70], the following trace inequality is considered:

where Fl and r2 are disjoint subsets of d R with positive surface measures. Then the optimal constant is shown to be

where u E

HI!,

( 0 )satisfies the weak formulation

Discussion of the best constants in some more Sobolev inequalities can be found in [73].

1.6.

SINGULARITIES OF ELLIPTIC PROBLEMS ON PLANAR NONSMOOTH DOMAINS

The most significant property of elliptic boundary value problems on a smooth domain is the so-called "shift theorem". Roughly speaking, "shift theorem" states: For an elliptic boundary value problem on a smooth domain, the smoother the data (the coefficients, the right-hand side and the boundary value), the smoother the solution. As an example, consider the boundary value problem

Assume d R is smooth. Then for any integer k 2 0, f E H k ( R ) implies uE H~+~(R). On nonsmooth domains, however, such a nice property no longer exists. To explain this, consider the two dimensional model problem

We assume that d R is smooth everywhere except at the origin 0, in a neighborhood of which, R coincides with a cone

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 1.6. A comer domain near the comer 0

where w is the internal angle of R at 0; see Figure 1.6. Let a = ~ / w For . each positive integer k , define

r k f fsin k a 8 r k f f( l n r sin k a 8

+

if k a # integer, 8 cos k a 8 ) if k a = integer.

+

Then i f f E WmlP(f2) and m 2 - 2 / p is not an integer, we have the following smoothness property for the solution u (cf, citeGr):

for some constants c k , which are certain linear functionals of f . Hence, no matter how smooth the function f is, the smoothness of the solution u is determined by the smoothness of the singular term ul as long as cl = cl ( f ) # 0. We note that ul E w11p(R)if and only if 1 < a 2 / p . An early systematic study of singular behavior around a comer for the solution of an elliptic problem was done by Lehman, cf. [105]. The model problem considered in that paper is the Dirichlet problem for the equation

+

where R is an open planar domain and part of its boundary d R consists of two analytic arcs rl and r2,which meet at the origin and form there an internal

27

Preliminaries

angle w > 0. Assume ri(i = 1 , 2 ) is an analytic curve of which the origin 0 is a regular point. Let u E C 2 ( R )n C ( R )be a solution of the equation with boundary conditions

Assume A, B, C , f , cpl and cp2 are analytic in a neighborhood of the origin and w ( 0 )= cp2 (0) - Then

where the regular part U R together with its partial derivatives of all orders remain bounded when x + 0,and with a = T/W,

u s ( x )=

a) {W O(P r) ln

if a # integer, if a = integer.

These relations may be formally indefinitely differentiated. In particular, if f = 0, cpl = 0, cp2 = 0 in a neighborhood of 0 ,then for some constant k ,

u ( x ) = k r" sin a0

+ o(ra)

and this relation can be formally differentiated. Hence,

Later, Wigley [I561 extended Lehman's results to allow nonanalytic data and boundary conditions involving first order derivatives. In [96], Kondrat'ev developed a general method for studying singular behaviors of solutions of elliptic boundary value problems in the neighborhood of a singular point of the boundary. The influence of the paper on later researches on the topic is tremendous. Maz'ya and his co-researchers extended Kondrat'ev's results in several aspects, e.g., from L~ estimates to Lp estimates for any p E (1,m), from single equations to systems, etc. The interested reader is referred to [107]. More recent comprehensive references on the topic are [99, 1001. In the literature, one may find many other results on the topic. In [146], [157], etc., theoretical and computational aspects of singularities for elasticity systems are studied. The papers [92] and [I231 are devoted to a study of regularity of solutions of Stokes problems in a polygon. For boundary value problems of the biharmonic equation on comer domains, see [23], etc. For both edge and corner singularities in a three dimensional polyhedron, cf. [126], etc. There are also several papers devoted to singular behaviors of solutions of quasilinear elliptic problems on nonsmooth domains, cf. [log], [149], etc. The reader is referred to the bibliographies of [46], [67], 1971, [99] and [loo] for a relatively complete list of references on singularities caused by the nonsmoothness of domains.

28

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

For singular behaviors of solutions of elliptic problems on nonsmooth domains, the main reference used here is [67]. This reference also provides detailed discussions of Sobolev spaces over domains with piecewise smooth boundaries. The following result is Theorem 5.1.3.5 in [67]. Let R be a polygon, the boundary of which is the union of a finite number of linear segments 1 5 j N, listed in a counterclockwise order. Split the set { 1 , 2 , . . . , N) into two subsets N and V.Let u E H I ( 0 )be a weak solution of the boundary value problem

c,

<

where Y is the unit outward normal on the boundary and d u l d u is the normal derivative. Denote S j the common vertex of ( r j ,O j ) the polar coordinate and system with origin at S j , and O j = 0 on Fj+l, O j = wj on r j .Let

q+l,

When X j , , is not an integer, we let

and when X j , , is an integer, we let

where r2j is a smooth cutoff function that equals 1 in a neighborhood of S j and vanishes outside a larger neighborhood of S j . Then we have the following result.

T H E O R E1M . 2 3 Let f E W"P(R) and g j E ~ ~ + ~ - l l p > p (j I E ' ~N ) , be given, 1 < p < co. Denote q = p / ( p - 1 ) the conjugate exponent of p. Assume that (Qj+l - Q j - I c ~ j 2wjlq)/~ is not an integer for any j and that gj ( S j ) = gi+l ( S j ) whenever j and j 1 E V.Then, there exist numbers k j , , such that

+

29

Preliminaries

can be replaced by the In the theorem, the condition gj E ~~+~-llpJ'(r~) (rj). more easily checked condition gj E

ckS2

1.7.

AN INTRODUCTION OF ELLIPTIC VARIATIONAL INEQUALITIES

The theory of variational inequalities plays an important role in the study of both the qualitative and numerical analysis of some nonlinear boundary value problems arising in mechanics and other applications. Variational inequalities are problems involving either differential inequalities or inequality boundary conditions. Mechanics is a rich source of variational inequalities (cf. e.g. [125]), and some examples of problems that give rise to variational inequalities are obstacle and contact problems, plasticity and visco-plasticity problems, Stefan problems, unilateral problems of plates and shells, and non-Newtonian flows involving Bingham fluids. An early comprehensive reference on the topic is 1471, where many nonlinear boundary value problems in mechanics and physics are formulated and studied in the framework of variational inequalities. A concise introduction to the mathematical theory of some variational inequalities can be found in [95]. Numerical approximations of general variational inequalities are studied in detail in [65, 661. Numerical methods for some variational inequalities arising in mechanics are the subjects of [82, 831. Mathematical analysis and numerical approximations of variational inequalities arising in contact mechanics are presented in [94] (for elastic materials) and [81] (for viscoelastic and viscoplastic materials). In [78, 791, elastoplasticity problems are formulated and analyzed in the form of variational inequalities. Numerous variational inequalities in the modeling and control in solid mechanics are studied in [93]. Many free-boundary problems can be recast as variational inequalities, and vice versa. A mathematical study of free boundary problems is found in [57]. In [143], free boundary problems arising in mould filling are studied as variational inequalities, and both the finite element and finite volume methods are applied for their numerical solutions. In this section we review two basic results on the solvability of linear elliptic variational inequalities. More precisely, we recall the following theorems, following 165, 661. More general results on elliptic and time-dependent variational inequalities can be found in most of the references listed in the previous paragraph. A functional j : V -+ = R U {fco) from a normed space V to the extended real line is said to be proper if j(v) > -co for any v E V and j(v) $ co.The functional is said to be lower semicontinuous (1.s.c.) if

u, + u in V (cf. Definition 2.6 on page 49).

j(u)5 lim inf j(u,) n-+co

30

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

THEOREM 1.24 Let V be a Hilbert space and let K c V be a non-empty, convex, closed subset. Assume a : V x V + JR is a continuous, V-elliptic bilinear form and e : V -+ R is a linear continuous functional. Then there exists a unique solution to the elliptic variational inequality of the first kind

THEOREM 1.25 Let V be a Hilbert space. Assume a : V x V -+ R is a continuous, V-elliptic bilinear form, j : V + is propel; convex and 1.s.c. on V and e : V -+ IR is a linear continuousfunctional. Then there exists a unique solution to the elliptic variational inequality of the second kind

We remark that if we take j = I K ,

the indicator function of the set K, then the variational inequality (1.29) reduces to the variational inequality (1.28). Therefore, elliptic variational inequalities of the first kind can be viewed as special cases of elliptic variational inequalities of the second kind. However, variational inequalities of the first kind are usually studied separately from those of the second kind, since numerous variational inequalities important in applications are naturally of the form (1.28), and most of the time, it is convenient to study the variational inequality (1.29) where j is a real valued functional. In this connection, one can also study the following mixed type variational inequality

where the set K, the bilinear form a ( . , .), the functional j (.) and the linear form e ( . ) satisfy the conditions of Theorems 1-24and 1.25, and moreover, j (.) is real valued on K. This mixed variational inequality also has a unique solution. In the particular case of Theorem 1-24with K = V, or Theorem 1.25 with j = 0, we recover the well-known Lax-Milgram Lemma, Theorem 1.16. When the bilinear form a is symmetric (i.e. a ( u ,v ) = a ( v ,u ) for u , v E V ) , which is the case for the problems considered in this work, we can consider the problem min E ( v ) vEV

where E : V -+ is an energy functional. Then solving the variational inequality (1.29) is equivalent to solving the minimization problem (1.30) with the functional 1 E ( v ) = - a(v,v) j(v) - e(v), 2

+

31

Preliminaries

and solving the variational inequality (1.28) is equivalent to solving the minimization problem (1.30) with

We now look at two concrete examples.

EXAMPLE 1.26 ( A N OBSTACLE PROBLEM) A representative example of the elliptic variational inequality of the first kind is given by the obstacle problem. The problem is to determine the equilibrium position of an elastic membrane passing through the boundary of a planar domain, lying above an obstacle of height $, and being subject to the action of a vertical force of density T f , here 7 is the elastic tension of the membrane, and f is a given function. Denote by R the planar domain, and by r for its boundary. The unknown of the problem is the vertical displacement u of the membrane. Since the membrane is fixed along the boundary F,we have the boundary condition u = 0 on I?. We assume f E L 2 ( R ) ,$ E H 1( R ) , and the compatibility condition $ 5 0 on r. The set of admissible displacements is

K = { v E H ; ( R ) : v 2 $ a.e. in R). The set K is nonempty, because the function max(0, $1 belongs to K. The principle of minimal energy from mechanics asserts that the displacement u is a minimizer of the total energy,

uE K :

E ( u ) = inf E ( v ) , VE

K

(1.31)

where the energy functional is defined as

The minimizer u is characterized by the variational inequality

The equivalence between (1.31) and (1.32) is proved as follows. First, suppose u E K is a solution of the minimization problem (1.31). For any v E K andanyt E (0, I ) , u + t ( v - u ) E K andso

E(u) Easily,

< E ( u + t (v - u ) ) .

32

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Hence, for any t E (0, I ) ,

Cancel the factor t

> 0:

L e t t + O+ in this inequality to obtain (1.32). Conversely, suppose u E K satisfies (1.32). Then for any v E K,

= E(u)

+

/'

[Vu V(v - u)

-

f (v - u ) ] d x

i.e., u is a solution of (1.31). It is easy to verify the conditions stated in Theorem 1.24, and therefore, the obstacle problem has a unique solution. EXAMPLE1.27 (A FRICTIONAL CONTACT PROBLEM)This is an exxnple of an elliptic variational inequality of the second kind. In Section 6.6, we will solve the problem by an adaptive finite element method. Denote by R c IRd (d 5 3 in applications) an open connected and bounded set with Lipschitz boundary r. Since the boundary is Lipschitz continuous, the outward unit normal exists a.e. on r and is denoted by v . Recall S d denotes the space of second order symmetric tensors on IRd, and the canonical inner products and corresponding norms on TRd and S d are

We define the product spaces ~ ~ ( :=0 ( )~ ~ (and 0 H'(R) ) ) ~ := ( H ' ( R ) ) ~ equipped with the norms i v = J v i/:,n, k = 0 , l . The same notation v is used to denote the function and its trace on the boundary. For a vector v , we will use its normal component v, = v v and tangential component v, = v - v,v at a point on the boundary. Similarly for a tensor a E Sd, we define its normal component a, = a v . v and tangential component a, = a v - a,v. For a detailed treatment of traces for vector and tensor fields in contact problems and related spaces see [94] or [81].

~ t = ~

33

Preliminaries

The material we consider here is linearly elastic. We denote by C : R x Sd -+ S d the elasticity tensor of the material. We assume the fourth-order elasticity tensor C to be symmetric:

and pointwise stable: for some constant co

> 0,

We first describe the physical setting of the frictional contact problem. Details and other related problems can be found in [94, 811. Consider a linearly 5 3.The boundary F is parelastic body occupying the domain R inIRd, dtitioned into possibly three parts: F = rDU F N U rc with r D , rNand relatively open and mutually disjoint, and meas(FD) > 0. The subscripts "DM, "N" and "C" are intended as shorthand indications for Dirichlet, Neumann and contact boundary conditions. We assume that on rDthe body is clamped, on F N a surface traction of density f E (L2(I'N))d is applied and on Fc the body is in bilateral contact with a rigid foundation. The contact is frictional and is ) ~in R. modeled by Tresca's law. Volume forces of density f E ( L ~ ( R )act The classical formulation of the problem is the following: Find a displacement field u : R + IRd and a stress field a : S2 + S d such that

where the friction bound g > 0 on re.We observe that (1.33) is the constitutive relation of the linearized elasticity material, (1.34) defines the linearized strain and the displacement, (1.35) is the equilibrium equation. With a = (aij)dXd, D i v a : R + Ktd is defined by

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

34

where the summation convention is used. The classical displacement and traction boundary conditions are given in (1.36) and (1.37). Contact conditions are described in (1.38)-(1.41). The bilateral contact feature is reflected by (1.38). The relations (1-39)-(1.41) represent Tresca's friction law. In certain situations, the frictional contact problem stated above describes the material deformation quite accurately. In more complicated situations, such as when the contact zone is not prescribed a priori or when more realistic frictional contact laws are used, the frictional contact problem here can be viewed as an intermediate problem for a typical step in an iterative solution procedure for solving the more complicated contact problem. For a variational analysis of the problem, we need to introduce a function space and some functionals defined over the space. Let

Its inner product and norm are chosen to be

Since meas(FD) > 0, the Korn inequality (cf. e.g. [47] for its proof)

lv

is a norm over V which is equivalent to the canonical holds. Hence, llv norm 11v 111,fiOver the space V, we define

To derive the corresponding variational inequality, we temporarily assume the problem (1.33)-(1.41) has a solution u which is sufficiently smooth so that all the following calculations are valid. We multiply the differential equation (1.35) by (v - u)with an arbitrary v E V,

35

Preliminaries

We then perform an integration by parts for the left hand side, and use the boundary conditions (1.36) and (1.37) and the constitutive relation (1.33),

+ jflC E( u ). E ( V- u ) dx. Now consider the integral on r c . From the contact conditions (1.39)-(1.41), we have u , . u , = -gIuTl o n r c . This can be deduced by distinguishing two cases: la, / Hence, using the decomposition

<

g and lurl = g.

and recalling the boundary condition (1.38) for u, and v,, as well as the bound (1.39), we have

Therefore, u satisfies the inequality a ( u , v - U)

+ j(v) - j(u) 2t ( v - u ) ,

where the functionals are the ones defined in (1.43)-(1.45). In conclusion, the variational formulation of the problem (1.33)-(1.41) reads: Find u E V such that

Due to the assumptions on C and meas(rD) > 0, the bilinear continuous form a ( . , .) : V x V -+ R is V-elliptic, following the Korn inequality (1.42). It can be easily checked that t : V -+ R is a linear continuous functional and j : V -+ is a proper lower semicontinuous convex functional. Therefore, by Theorem 1.25, the variational inequality (1.46) has a unique solution u in V . Moreover, since the bilinear form a ( . , .) is symmetric and positive definite,

36

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

solving the variational inequality (1.46) is equivalent to minimizing the energy functional 1 E(v) = -a(v,v) - t(v) j(v) 2 over the space V. The equivalence between the variational inequality (1.46) and the minimization formulation can be shown by an argument similar to that used in Example 1.26. Note that the convexity of j (.) implies

+

f o r a n y u , ~E V andandt E [ O , l ] . Variational inequality formulations of many other contact problems can be found in [94, 811.

1.8.

FINITE ELEMENT METHOD, ERROR ESTIMATES

Weak formulations of boundary value problems are the basis for development of Galerkin methods, a general framework for approximation of variational problems, that include the finite element method as a special case. Consider the problem (1.17) (page 18) again. In general, it is impossible to find the exact solution of the problem (1.17) because the space V is infinite dimensional. A natural approach to construct an approximate solution is to solve a finite dimensional analog of the problem (1.17). Let VN c V be an N-dimensional subspace. We project the problem (1.17) onto VN,

Under the assumptions that the bilinear form a ( - ,.) is bounded and V-elliptic, and t E V', we can again apply the Lax-Milgram Lemma and conclude that the finite dimensional approximation problem (1.47) has a unique solution u ~ . The approximate solution U N is, in general, different from the exact solution u.To increase the accuracy, it is natural to seek the approximate solution u~ in a larger subspace VN. Thus, for a sequence of subspaces VN, c VN2 c - - - c V, we compute a corresponding sequence of approximate solutions u r ~E ~VNi from (1.47) for i = 1 , 2 , . . This solution procedure is called the Galerkin method. In the special case when the bilinear form a ( . , .) is also symmetric, a

a ( u , v) = a(v, u)

Vu, v E V,

the original problem (1.17) is equivalent to a minimization problem uEV,

E ( u ) = inf E ( v ) , vEV

(1.48)

Preliminaries

where the "energy functional" is defined as

Now with a finite dimensional subspace VN c V chosen, it is equally natural to develop a numerical method by minimizing the energy functional over the finite dimensional space VN, UN

E V N , E ( u N ) = inf E ( v ) . VEVN

Following the argument for the equivalence of (1.31) and (1-32)in Example 1.26, we can verify that the two approximate problems (1.47) and (1.50) are equivalent. The method based on minimizing the energy functional over finite dimensional subspaces is called the Ritz method. From the above discussion, we see that the Galerkin method is more general than the Ritz method, while when both methods are applicable, they are equivalent. Because of this, the Galerkin method is also called the Ritz-Galerkin method. Once a numerical method is formulated, an important issue is convergence and error estimation for the method. In this regard, for the Galerkin method, a key result is the following CCa's inequality.

PROPOSITION 1.28 Assume V is a Hilbert space, VN C V is a subspace, a ( . , is a bounded, V-elliptic bilinear form on V, and t? E V*. Let u E V be the solution of the problem (1.17),and U N E VN be the Galerkin approximation deJined in (1.47). Then there is a constant c such that a )

The proof of Cia's inequality follows from the error relation

obtained by subtraction of (1.47)from (1.17)with v E VN, and the boundedness and V-ellipticity of the bilinear form a(., .). CCa's inequality is a basis for convergence analysis and error estimations. As a simple consequence, we have the next convergence result.

COROLLARY 1.29 We make the assumptions stated in Proposition 1.28. Assume VN, c VN, c - . . is a sequence of subspaces of V with the property

Then the Galerkin method for (1.17) converges.

38

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

in V. Cia's In Corollary 1.29, Ui>lVN, stands for the closure of Ui>lVN, inequality also serves a s a basis for error estimates. When the finite dimensional space VN is constructed from piecewise (images of) polynomials, the Galerkin method leads to a finite element method. In other words, the finite element method (FEM) is a Galerkin method with the use of piecewise (images of) polynomials over a finite element partition of the domain 0. The finite element method today is the dominant numerical method for solving problems in structural mechanics and is popularly used in fluid mechanics. It is widely applied to both linear and nonlinear problems. General mathematical theory of finite element methods can be found in [8,41,43,117,144], among others. The textbook [89] offers an easily accessible mathematical introduction of finite element methods, whereas the two recent textbooks, [31, 321, provide deeper mathematical theory together with more recent and current research development such as the multigrid methods. Traditionally, convergence of finite element solutions is achieved through mesh refinement with the use of piecewise low degree polynomial. Since h is usually used to denote the mesh size, the traditional finite element method is also termed as the h-version finite element method. Convergence of the method can also be achieved by using piecewise increasingly higher degree polynomials over relatively coarse finite element meshes, leading to the p-version finite element method. Detailed discussion of the p-version finite element method can be found in [146]. The p-version approximation is more efficient in areas where the solution is smooth, so it is natural to combine the ideas of the p-version and the h-version to make the finite element method very efficient on many problems. A well-known result regarding the h-p-version finite element method is the exponential convergence rate for solving elliptic boundary value problems with corner singularities. Comprehensive mathematical theory of the p-version and h-p-version finite element methods with applications in solid and fluid mechanics can be found in [137]. Mixed and hybrid finite element methods are often used in solving boundary value problems with constraints and higher order differential equations. Mathematical theory of these methods can be found in [33,135]. Several monographs are available on the numerical solution of Navier-Stokes equations by the finite element method, see e.g. [63]. Theory of the finite element method for solving parabolic problems can be found in [I471 and more recently in [148]. Finally, we list a few representative engineering books on the finite element method, [20, 88, 163, 1641. The reader is referred to two historical notes [115, 1621 on the development of the finite element method. In this section, we will review some results of the finite element method. There are some basic aspects in the construction of finite element approximations. First we need a partition (or triangulation) of the domain of the differential equation into sub-domains called elements. Associated with the partition, we

Preliminaries

Figure 1.7. A finite element mesh

define a finite element space. Finally, we choose basis functions for the finite element space. The basis functions should have small supports. For simplicity, we assume R is a two-dimensional polygonal domain. The reader is referred to the aforementioned references for some detailed discussion on the use of the finite element method where R is not polygonal, or where the spatial dimension is higher than two.

Triangulation. A triangulation, partition, or mesh is a partition Ph = { K ) of the polygonal domain 2 into a finite number of subsets K , called elements, with the following properties: -

R = U K E P hK . 0

Each K is a triangle or quadrilateral with a nonempty interior K . For distinct K 1 ,K z E Ph, K 1 f l K 2 is either empty, or a common vertex, a common side or a common face of K l and K 2 . For any K E Ph, its diameter diam ( K ) 5 h. The assumption that the domain R is a polygon ensures that it can be partitioned into straight-sided triangles and quadrilaterals. See Figure 1.7 for a finite element mesh. For convenience in practical implementation as well as in theoretical analysis, it is assumed that there exist a finite number of fixed polyhedra, called reference elements or master elements, ambiguously represented by one symbol K ,such that for each element K , there is an invertible affine mapping function FK with K = FK ( K ). In the two-dimensional case, it is customary to use an equilateral or right isosceles triangle for triangular elements, and squares with side length 1 or 2 for quadrilateral elements.

40

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

For an arbitrary element K , we introduce its diameter

and let p~ be the diameter of the largest sphere inscribed in K. When dealing with the reference element K we denote the corresponding quantities by h and @. The quantity h K describes the size of K , whereas the ratio h K / p K is an indication whether the element is flat: the larger the ratio h K / p K ,the flatter the element K. For a triangulation P h , we denote h = max h K , KEP,

often called the mesh parameter. The quantity h is a measure of how refined the mesh is. The smaller h is, the finer the mesh. We say a family of triangulations { P h ) h is regular if 1 there exists a constant a 2 1 such that

2 the discretization parameter h

+ 0.

A condition equivalent to the first requirement is the minimal angle condition, i.e, the minimal angles of all the elements are bounded below away from 0.

Polynomial spaces on the reference elements. Function spaces over a general element can be constructed from those on the reference element. Thus we first introduce a polynomial space x on K. Although it is possible to choose other finite dimensional function spaces as X, the overwhelming choice of x for practical use is a polynomial space, due to the simplicity in constructing a polynomial and the good approximation property of polynomial spaces. Since the highest spatial derivatives are of order 2 for the mathematical problems studied later in this work, it is natural to use Lagrange finite elements, e.g., the parameters of the elements are function values at the nodes. Generally, let x be a polynomial space over K, dim^ = No, and we choose a set of nodal points { 8 i ) 2 1 , called the nodes, in K such that any function d E x is uniquely determined by its values at . Then we have the formula

{iii)2,

Nn

where the functions

{&)Fl form a basis for the space x with the property

41

Preliminaries

Affine-equivalent finite elements. We then define a function space over a general element K that is the image of the reference element K under an invertible affine mapping

The mapping FK is a bijection between K and K. Over the element K ,we define a finite dimensional function space X K by

Since FK is an invertible affine mapping, if x is a polynomial space of certain degree, then X K is a polynomial space of the same degree. Any function v defined on K is associated with a function 6 defined on K through .ir = v o FK. We see that v = 6 o Fil. Thus we have the relation

~ ( x=) 6 ( k ) V x E K ,2 E K ,w i t h x = FK(ii) Using the nodal points h i , 1 5 i 5 No, of K ,we can define the nodal points on K : aiK = F K ( h i ) , i = I , . . . , N O . (1.54) Recall that { & ) z , are the basis functions of the space x associated with the nodal points { h i ) F l with the property that

We define

#f=&oFil,

i=1,. . . , No.

Then the functions { @ ) Z 1 have the property that

Hence,

{@)zl

form a set of basis functions for the space X K

Finite element spaces. A global finite element function vh is defined piecewise by the formula v h l ~E X K V K E Ph. A natural question is whether such finite element functions can be directly used to solve a boundary value problem. For a linear second-order elliptic boundary value problem, the space V for the corresponding weak formulation is a subspace of H 1 ( R ) . Since the restriction of vh on each element K is a smooth function, a necessary and sufficient condition for vh E H 1 ( R ) is

42

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

vh E C ( 2 ) (cf. [41]). Thus we need to check if vh E ~ ( 2 holds. ) We then define a finite element space corresponding to the triangulation Ph, We observe that if x consists of polynomials, then a function from the space X h is a piecewise image of polynomials. In our special case of an affine family of finite elements, FK is an affine mapping, and vh 1 is a polynomial. We can use the finite element space Xh to approximate the space H 1( R ) . Some boundary value problems involve essential boundary conditions. Then we need finite element sets to approximate subsets of H1( R ) . As an example, suppose rl c r is a relatively closed subset of the boundary ?? consisting of a finite number of line segments. Consider the space We construct finite element partitions of R that are compatible to the subset r l , i.e., if K is a finite element, then either K n rl = 8 or K f l r c rl. Then we define the finite element space to be Vh = V n Xh. In other words,

For an affine family of finite elements, the mesh is completely determined by the reference element K and the family of affine mappings {FK : K E Ph). The finite element space Xh is completely described by the space x on K and the family of affine mappings.

Interpolation. On any element K , we introduce an interpolation operator l l K by

i=l We see that n K v E X K is uniquely determined by the interpolation conditions

For a function v E ~ ( 2 1 we,construct its global interpolant n h v in the finite element space X h by the formula

Let { a i ) z l c 2 be the set of the nodes collected from the nodes of all the elements K E Ph. We have the representation formula

43

Preliminaries

for the global finite element interpolant. Here $i, i = 1 , . . . , Nh, are the global basis functions that span X h . The basis function 41i is associated with the node ai, i.e., $i is a piecewise polynomial, $i l K E X K , and $i ( a j )= S i j . If the node ai is a vertex a r of the element K , then $i 1 = $ r . If ai is not a node of K , then $ilK = 0. Thus the functions $i are constructed from local basis functions .4:

EXAMPLE 1.30 We examine an example of linear elements. Assume R c IR2 is a poly onal domain, which is triangulated into triangles K , K E P h . Denote by {ai}$;' the set of the interior nodes, i.e. the vertices of the triangulation that lie in 12; and {ai}2Nint+l the set of the boundary nodes, i.e. the vertices of the triangulation that lie on 8 0 . From each vertex ai, construct K~ as the patch of the elements K which contain ai as a vertex. The basis function q!~~ associated with the node ai is a continuous function in 2,which is linear on each K and . corresponding piecewise linear function space is is non-zero only in K ~ The then X h = span{41i, 1 5 i 5 Nh}. Suppose we need to solve a linear elliptic boundary value problem with Neumann boundary condition. Then the function space is V = H ' ( R ) , and we choose the linear element space to be Vh = Xh. Now suppose the homogeneous Dirichlet boundary condition is specified. Then the function space is V = H: ( R ) ,whereas the linear element space is

In other words, we only use those basis functions associated with the interior nodes, so that again the finite element space is a subspace of the space for the boundary value problem. In case of a mixed boundary value problem, where we have a homogeneous Dirichlet boundary condition on a relatively closed part of the boundary rl and a Neumann boundary condition on the remaining part of the boundary r 2 , then the triangulation should be compatible with the splitting 8R = rl U r 2 , i.e., if an element K has one side on the boundary, then that side must belong entirely to either rl or E. The corresponding linear element space is then constructed from the basis functions associated with the interior nodes together with those boundary nodes located on PI. rn

Finite element error estimates. Let PI,( K )denote the space of the polynomials of degree less than or equal to k on K.A basic result concerning finite element interpolation errors is the following.

+

T H E O R E1.3 M 1 Let k and m be nonnegative integers with k > 0, k 1 2 m, and pk ( K ) C X . Let IIK be the operators defined in (1.55). Furthermore,

44

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

assume { P h ) is a regular family ofjnite element partitions. Then there is a constant c such that for m 5 k 1,

+

Since the global finite element interpolant is defined piecewise in terms of local finite element interpolants, we have the next result for global finite element interpolation error estimates.

T H E O R E1.32 M Assume that all the conditions of Theorem 1.31 hold. Then there exists a constant c independent of h such that for m = 0,1,

Notice that for the Lagrange finite elements, IIhv @ H 2 ( R ) .For 2 5 m 5

k

+ 1, (1.59) is replaced by

Cia's inequality and the interpolation error estimates together lead to convergence order error estimates for finite element solutions of the elliptic boundary value problem (1.17). As a sample result, consider the finite element approximation of the problem (1.9) (page 16). Assume R c Kt2 is a polygonal domain, f E L2( R ) . Let { P h ) be a regular family of triangulations of into triangles and let { V h ) be the corresponding family of linear element spaces given by (1.57). Under the solution regularity assumption u E H ~ ( R(valid ) if R is convex, cf. Section 1.6),then for the corresponding finite element solutions { u h ) ,we have the error estimate (1.60) 1 1 -~ ~ h l I 1 5, ~c 1 1 -~ n h ~ l l l , R5 c h iul2,R. This estimate is an example of an a priori error estimate , i.e., we have this estimate (under the stated assumptions) before we actually compute the solution uh. The estimate (1.60) expresses the fact that the convergence order of the linear element solutions is one, and if we refine the triangulation by connecting the three side mid-points of each element and compute the new finite element solution uh/2, then roughly speaking, we can expect the error ilu - uh/2 11 1 , would be about half of the error Ilu - uhiIlls2,at least when h is sufficiently small. Note, however, that the a priori error estimate (1.60) does not provide quantitative information on the size of the error 11 u - uh /I l , ~nor , does it provide information on the distribution of the error ( u - u h )throughout R. In contrast, in an a posteriori error estimate, computed solutions are used to provide such quantitative information on the solution errors.

~

Preliminaries

45

In a posteriori error analysis of finite element solutions (cf. Chapter 6), one needs finite element interpolations of functions with low degree smoothness, such as H 1(R) functions. Note that for d 2 2, an H' (R) function can have discontinuity and so the ordinary finite element interpolation based on the use of pointwise function values is not defined. In Section 6.1, we review the definition and some error estimates for one generalized finite element interpolation operator defined for H' (0)functions.

Chapter 2

ELEMENTS OF CONVEX ANALYSIS, DUALITY THEORY

The a posteriori error estimates presented in this work are derived based on the duality theory of convex analysis. The first research monograph specifically devoted to the topic of convex analysis is [136], emphasizing the finitedimensional case. Convex analysis and duality theory in general normed spaces, mostly infinite dimensional ones, are thoroughly discussed in the well-known reference [49]. Another comprehensive treatment of the topic is [159]. Duality theory has been also extended for nonconvex systems, see, e.g. [59, 601 where the mathematical theory is motived by duality in natural phenomena with particular emphasis on mechanics. In this chapter, we review some basic notions and results on convex sets, convex functions and their properties as well as the duality theory. Detailed discussions and proofs of the stated results can be found in [49] or [159]. In the theory of convex analysis, it is convenient to consider functions that take on values on the extended real line E. Recall that a functional f : V + is said to be proper if f ( v ) > -cc b'v E V and f ( u ) < cx for some u E V.

2.1.

CONVEX SETS AND CONVEX FUNCTIONS

Let V be a linear space.

DEFINITION 2 . 1 A subset K C V is said to be convex if it has the property U,V

( I - t ) u + t v E Kb't E [O,l].

EK

We see that if K is convex and u, v E K , then the line segment connecting u and v, i.e, the set ( ( 1 - t ) u t v : t E [0,I ] ) ,is contained in K. By an induction argument, for any ul , . . . ,u, E K and any nonnegative numbers tl , . . . , t , with C:='=,ti = 1, we have C:=l tiui E K . The expression C:=ltiui with

+

48

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

nonnegative numbers t l , . . . , tn satisfying Cy=2=1 ti = 1 is called a convex combination of the elements u l , . . . , u,.

DEFINITION 2 . 2 Afunction f : V

-+

is convex if

for any u, v E V and t E (0,I ) for which the right hand side is meaningful, i.e., f ( u )and f ( v )are not simultaneously infinite with opposite signs.

DEFINITION 2.3 Let K be a convex set in V avtd f : K

f ( v )=

+R. If

{ y; w

v E K, v

is convex, then we say f is convex on K. Thefunction f is strictly convex on K ifthe strict inequality in (2.1) holds for any u, v E K , u # v and t E ( 0 , l ) . In the future, for a function f defined on a subset K c V, we identify it with its extension f introduced in Definition 2.3. In other words, we will use the same symbol f for both the function defined on K and its extension by oo to the complement of K in the space V. Thus, we will say that a function is convex over a subset K c V to mean that the extension of the function is convex in the space V. By an induction argument, i f f is convex over a convex set K , then we have

for any u l , . . . , u, E K and any nonnegative numbers t l , . . . , t , with

The next result follows easily from the definition of a convex function.

PROPOSITION 2.4 Let V be a linear space, A be an index set. Assume g, t f ( t E f , g , fa ( a E A) : V + are convex. Then the functions f (0,oo)),sup{f , g ) and supaEAf a are all convex. Here we let f ( v )+ g ( v )= oo i f f ( v )= -g(v) = &oo.

+

In the study of convex functions, it is convenient to use the notions of the effective domain and the epigraph.

DEFINITION 2 . 5 Given afunction f : V

-+ E, we dejne its effective domain

dom ( f ) = {v E V : f ( v ) < oo}

Elements of convex analysis, duality theory

and its epigraph e p i ( f ) = { ( v , a )E V x R :f ( v ) 5 a ) . It is easy to show that for a convex function the effective domain is a convex set in V and the epigraph is a convex set in V x R. From now on, we assume V is a normed space.

DEFINITION 2.6 A function f : V -+ is said to be lower semicontinuous (1.s.c.) iffor any sequence {u,) C V with u, -+ u in V , f(

u ) < lim inf f (u,). - n+m

There is a useful characterization of the lower semicontinuity that provides an alternative definition of 1.s.c. in some references.

PROPOSITION 2.7 Thefinction f : V { v E V : f ( v ) 5 r ) is closed.

-+@ is 1.s.c. ifffor any r E R the set

Later on, we will also need the notion of weak 1.s.c.

-

DEFINITION 2.8 A function f : V -+ is said to be weakly lower semicontinuous (w.1.s. c.) i f for any sequence {u,) c V with u, u in V , f(

U)

< lim inf f - n-+m

(u,).

EXAMPLE2 . 9 Let K C V. The indicator function of the set K is defined by

Then it can be verified that K is closed iff IK is l.s.c., whereas K is a convex set iff IK is a convex function. 0

For a set K in the normed space V, we use int K = i n t ( K ) = K to denote its interior, i.e. the set of the points in K such that each point is contained in an open ball that in turn lies in K. Roughly speaking, int K is the subset of K excluding the boundary points. Some of the boundary points may not belong to K , unless K is a closed set. We use K to denote the closure of K , i.e., the union of the set K and its boundary. The following results are not difficult to prove.

PROPOSITION 2.10 Let f : V + E. Then (a) f is convex z f f epi (f ) is convex; (b) f is 1,s.c, ifSepi ( f ) is closed; (c) f is continuous at u and f ( u )# fcx ==+ int epi ( f ) # ( 4 f $ +W ===+ epi ( f ) # 0: (e) f is convex =+ dom (f) is convex.

0;

50

2.2.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

HAHN-BANACH THEOREM AND SEPARATION OF CONVEX SETS

The Hahn-Banach theorem and its corollaries are of central importance in functional analysis (cf. e.g. [48]). In this work, we only need Corollary 2.17 given at the end of the section. For completeness, we state some related results and show how they lead to a proof of Corollary 2.17.

DEFINITION 2 . 1 1 A function p : V -+ R is sublinear i f

We observe that p : V -+ R is a seminorm if it is sublinear and p ( t v ) = It/p(v) for any v E V and any t E R.The analytic form of a general HahnBanach Theorem is the following.

THEOREM 2.12 (Hahn-Banach Theorem) Let V be a real linear space, K C V a subspace. Assume f : K -+ R is linear and f (v) 5 p(v) for any v E K, with some sublinearfunctional p : V -+ R. Then f can be extended to a linear functional f : V -+ E% such that f (v) I p(v) for any v E V . Taking the functionp(.) to be a constant multiple of the norm, we immediately get the usual form of the Hahn-Banach Theorem.

COROLLARY2 . 1 3 Let V be a real Banach space, K Assume f : K -+ R is a linearfunctional satishing

c

V be a subspace.

Then f can be extended to a continuous linearfunctional on V with

There is a related geometric form of the Hahn-Banach theorem on separation of convex sets, Proposition 2.15. For this purpose, we introduce the following definition.

DEFINITION 2 . 1 4 Let V be a real normed space, A, B c V be non-empty. The sets A and B are separated ifthere exist k' E V * ,k' # 0 and a E R such that q ~I )a 5 e ( ~ ) v u E A, E B. The separation is strict if the inequalities can be replaced by strict inequalities

51

Elements of convex analysis, duality theory

PROPOSITION 2.15 (Separation of convex sets) Let V bea real normedspace, A , B c V be non-empty and convex. (a) Zfint(A)n B = 0 and int(A) # 0, then A and B can be separated; fiuthermore, f ( u ) < a V u E int( A ) . (b) I f A f l B = 8 and either A and B are open or A is closed and B is compact, then A and B can be strictly separated. LEMMA2.16 Let V be a real normed space, f : V Suppose -cc < a < f ( u )

-+

be convex and 1.s.c.

for some u E d o m ( f ) (hence it is possible f ( u ) = co). Then 3 ( u * ,a ) E V* x R such that In particulal; iff ( u ) # i m , then f ( v ) > a + ( u * , v - U ) V v E V, f ( v ) > -m. Proof. Every x* E (V x R)* has the form

+

( z * ,( v ,b ) ) = ( w * , v ) a*b V ( v ,b) E V x R, where w* E V * ,a* E R . Iff = co, then we choose u* = 0. Now assume f $ co. Then epi(f ) is convex, closed and non-empty. We have ( u ,a ) $! epi(f ), and the set { ( u ,a ) ) is convex and compact. By Proposition 2.15 (b), the sets { ( u ,a ) ) and epi(f) can be strictly separated in V x R.So 3 x* = ( w * ,a*) E (V x R ) * and p E R such that

+

+

( w * , ~ ) a*a > p > ( w * , v ) a*b V ( v ,b) E epi(f). Now for v E dom(f) with f ( v ) > - m , ( v ,f ( v ) )E epi(f). Hence Suppose a* 2 0. Then since u E dom(f ) , there is a sequence {v,) with v, + u as n + co. Then

+

( w * , u ) a*a > /? contradicting a

c dom(f )

> ( w * , v )+ a * f ( u ) ,

( u ) .Thus we must have a*

< 0 and therefore

Then the inequality is valid for any v E V with f ( v ) > -m. A consequence of Lemma 2.16 is the following.

rn

COROLLARY 2 . 1 7 Let V be a normed space, j : V -+ be propel; convex and 1.s.c. Assume u E dom ( j ) . Then there exist u* E V * and a E R such that

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 2.1. Continuity of a convex function

2.3.

CONTINUITY AND DIFFERENTIABILITY

The purpose of this section is to list a few results on the continuity and differentiability of convex functions so that readers with little background on convex analysis can get familiar with basic properties of convex functions.

Continuity. The basic result concerning the continuity of convex functions is the following.

PROPOSITION 2.18 Let V be a real normed space, f : V -i be convex. (a) Assume f (u)E R. Then f is continuous at u i f f is bounded from above in a neighborhood of u. (b) I f f ifJinite on an open set M C V and is continuous at some point of M , then f is continuous on M . Figure 2.1 shows a convex function f that is continuous in the interior of its effective domain, and is not continuous at b ( f (x) = cc for x 2 b), a boundary point of the effective domain. Note that f is not bounded from above to the right of b. The next two results can be deduced from Proposition 2.18.

COROLLARY 2.19 Let M C Rd be an open convex set. Then every convex function f : M -i R is continuous. COROLLARY 2.20 Let V be a real Banach space, M c V be closed and convex. Let f : M -+ R be convex and 1.s.c. Then f i f continuous on int ( M ) . Subdifferential. The notion of subdifferential is useful in describing various mechanical laws arising in contact problems, plasticity, etc. Although in later chapters, we do not explicitly use the notion of subdifferential in deriving a posteriori error estimates, it is an important concept in convex analysis. Now we introduce the definition of subdifferential and subgradient.

Elements of convex analysis, duality theory

Figure 2.2. Subdifferential of the absolute value function

DEFINITION2.21 Let V be a real normed space with the dual V*, and f : V -+ @. Let u E V be such that f ( u ) # +oo. Then the subdifferential off at u is defined to be the set

Any u* E df ( u )is called a subgradient off at u. We see that if d f ( u ) # 0, then f (v)

> -cc

for any v E V.

EXAMPLE2.22 For a real-valued real-variable function f : R -+ R, its subdifferential at u E R is the set of the slopes of straight lines passing through the point ( u ,f ( u ) )and lying below the curve of f . For example, the absolute value function f ( u ) = lul is not differentiable at u = 0, but is subdifferentiable there, and d f (0) = [- 1 , 11 (Figure 2.2). On the other hand, differentiable functions may not be subdifferentiable. For instance, the smooth function f ( u ) = u3 is not subdifferentiable at u = 0. The notion of the subdifferential is most suitable for convex functions. EXAMPLE2.23 (Support functional) Let V be a real normed space, K be a convex set. Consider the subdifferential of the indicator function 0 +cc

i f v ~ K , i f v $! K.

If u $! K , then d I K ( u )= 0. Assume u E K. Then u* E d I K ( u )iff

Thus we have the characterization

cV

54

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Any subgradient u* E d I K (u) is called a support functional to K at u. We always have 0 E d I K ( u ) for u E K. It is easily seen that if u E int(K), then d I K ( u ) = (0). For a boundary point u E d K and the case int(K) # 0, by separating u and int(K), we can show the existence of a nonzero subgradient u* E d I K ( u ) If K is a subspace, then

which can be viewed as the orthogonal complement of K. The following important result plays a central role in the duality theory.

THEOREM 2.24 Assume V is a rejlexive Banach space, f : (-m, m ] is convexand1.s.c. Thenv* E d f ( v ) i r v € df*(v*). As is commented in Example 2.22, the notion of the subdifferential is mainly applied to convex functions. This is supported by the next result on the existence of subgradients.

THEOREM 2 . 2 5 Let V be a real normed space, f : V (a) For any v E V , d f (v) is convex and weak* closed. (b) Iff isjnite and continuous at v, then d f (v) # 0.

+

be convex.

Ordinary differentiation rules in calculus carry over to subdifferentials, either straightforwardly or with some additional assumptions. For example, it is easy to verify the following relations from the definition of the subdifferential.

A natural question is when the equality holds for the summation rule.

PROPOSITION 2.26 Let V be a real normed space, f i : V + ( - m , m ] be convex for i = 1 , . . . , n. Assume there is a uo E V such that fi(uo) E R, 1 5 i 5 n, and fi,1 5 i n - 1, are continuous at uo. Then

<

a(f1 + . . .

+ f n ) (v) = a(f1) (v) + . - . + d(fn) (v)

Vv E

V.

A proof of this result and that of the next Chain rule can be found in [49].

PROPOSITION 2 . 2 7 Let V and W be two real normed spaces, L : V -+ W be linear and continuous, f : W + R be convex and 1.s.c. Iff isjnite and continuous at some point, then

55

Elements of convex analysis, duality theory

Relationship between subgradient and Giiteaux derivative. First, we recall the definitions of the directional derivative and GPteaux derivative. DEFINITION 2.28 Let f : V -+

and f ( u ) E R. For a v E V , if

exists, we call it the directional derivative off at u in the direction v, and denote it by fl(u;v). I f there exists u* E V *such that

then f is said to be GPteaux differentiable at u. The element u* is called the Gdteaux derivative off at u and is denoted by f ' ( u ) . Higher order derivatives are defined recursively. For instance, the secondorder GPteaux derivative is defined to be the GPteaux derivative of the GPteaux derivative.

PROPOSITION 2.29 Let V be a real normed space, f : V -+ be convex. Assume f ( u ) E R. (a) I f f ' ( u )exists as a Gdteaux derivative, then df ( u ) = { f ' ( u ) ) . (b) I f f is continuous at u and df ( u )contains exactly one element, then f ' ( u ) exists as a Gdteaux derivative.

+

Proof. For any v E V, we define a function d ( t ) = f ( u t (v - u ) ) ,t E R. Then q5 is a convex function of the real variable t . By the Mean Value Theorem,

Since 4 is convex, @ ( O )

> $'(O).Then

f ( v ) - f ( u ) 2 ( u * , v - U ) 'dv E Let v = u

+ t h, h E V, and let t -+ O+

V.

to obtain

Therefore, f ' ( u )= u*. This completes a proof of part (a).

56

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

A proof of part (b) can be found in [49]. Characterization of convex functions. We can use the GBteaux derivative to characterize the convexity of a function. Let V be a normed space and f : V -+ R be GBteaux differentiable. Then the following three statements are equivalent. (a) f is convex. (b) 2 f(u)+ (fl(u),v- u) v u , v E V . (c) (fl(v)- fl(u),v-u)2 0 V u , v E V .

f(4

2.4.

CONVEX OPTIMIZATION

Given a space V, its subset K, and a functional f : K -+ R, we consider the problem inf f (v). vEK

When K is unbounded, we say the function f is coercive on K if

f (v)-+ cc as iivll -+ a,v

E

K.

We have a standard general result on the existence of a minimizer to the problem (2.2).

THEOREM 2.30 Assume V is a reflexive Banach space, K c V convex and closed, and f : K -+ E% is convex and k c . If either (a) K is bounded or (b) f is coercive on K , then the minimization problem (2.2)has a solution. Moreovel; iff is strictly convex on K , then a solution of the minimization problem (2.2)is unique. This theorem will be applied later to show the existence and uniqueness of weak solutions to some nonlinear boundary value problems that are equivalent certain convex minimization problems. From the definition of subdifferential, immediately we get an extremal principle.

PROPOSITION 2.31 Let V be a real normed space, f : V -+ Then u is a solution of infvEvf (v)if0 E d f (u).

be propel:

THEOREM 2.32 Suppose V is a real normed space, K C V is a non-empty convex set. Assume f : K -+ R is convex. Thenfor u E K to be a solution of the problem inf f (v) vEK

57

Elements of convex analysis, duality theory

a necessary and sufJicient condition is that u E V is a solution of the unconstrained optimization problem

or u E V satisfies the relation

3 u* E V * with (u*,v - u ) 2 0 'v'v E K , such that f ( v ) 2f ( u ) + ( u * , v - u ) Y v E V . The solution set is convex. I f f is 1.s.c. and K is closed, then the solution set is closed. Every local minimum o f f is also a global minimum. A minimizer o f f is unique i f f is strictly convex.

2.5.

CONJUGATE FUNCTIONALS

The idea of the duality theory can be described as follows: Let f be a given function on a normed space V . For a minimization problem

inf f ( v ),

vEV

we look for a maximization problem

such that inf f ( v )= supg(q).

VEV

~ E Q

The problem (2.3) is called the primal problem, and (2.4) is called the dual problem. Then we have the following two-sided bounds for the optimal value:

This is the basis for deriving most of the a posteriori error estimates in this book. The space Q and the function g in the dual problem (2.4) are to be constructed from the primal problem (2.3). In particular, the construction of g is related to the concept of conjugate functionals.

D E F I N I T I O2.33 N Assume V is a normed space, and let f conjugate functional f * : V * + is dejined by the formula f * ( v * ) = S U P [ ( V * ,v ) - f ( v ) ] vE V

v v* E

V*

:

V

+ E.

The

58

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Using the effective domain of the functional, we can also write f * ( v * )=

sup [ ( v *v, ) - f ( v ) ] Vv* E V * . v€dom(f)

It follows from the definition that the conjugate functional f * is convex and 1.s.c. on V * . We have the following generalized Young inequality:

f * ( v * )+ f (4 2 ( v * , v ) , f * ( v * ) +f ( v ) = ( v * , v ) iff v * ~ d f ( v )

+

for all v E V , v* E V * as long as the expression f * ( v * ) f ( v )is meaningful, i.e., not of the form cc - w . This inequality is a generalization of the usual Young inequality:

where p > 1 and p* is the conjugate exponent, defined through the equality llp* l l p = 1. We will frequently need to calculate the conjugate functional for a functional defined by an integral of the form

+

Before stating a theorem on how to calculate its conjugate function, we introduce the following notion.

DEFINITION 2.34 Let R be an open set of Rd, g : 0 x lR1 -+ R. We say g is a CarathCodory function if (a) 'i 6 E R1, x H g ( x ,5 ) is a measurable function; (b) for a.e. x E R, F+ g ( x ,6 ) is a continuous function.

<

Let there be given mi E (1,a ) i, = 1, . . . ,1. We have the following theorem which will be applied repeatedly in calculating conjugate functionals.

THEOREM 2.35 Assume g : R x R1 -+ R is a Carathe'odoryfunction. For ~~i ( R ) ,dejine any q E Q =

Thenfor the conjugate function of G, we have the formula

G*(q*)=

g* ( x ,q* ( x ) )d z

Vq* E V * ,

Elements of convex analysis, duality theory

EXAMPLE 2.36 Let Q be a domain in IRd, Q = ( L(Q)) ~ d . We equate the dual space Q* with Q. Let us compute the conjugate of the functional

By definition,

= sup QEQ

/

(q* . q

-

2I lq2) dx.

n

Applying Theorem 2.35, we have

In later chapters, we will follow the above procedure to compute conjugate functional~of similar kind, and we will not always state explicitly the application of Theorem 2.35.

2.6.

DUALITY THEORY

We now introduce some basic results in the duality theory; detailed discussion and proofs of these results can be found in [49]. Let V and Q be two normed spaces, V* and Q* denote their dual spaces. The duality pairings in both V, V* and Q, Q* will be denoted by (., .). Assume there exists a linear continuous operator A E C(V, Q). The transpose A* E C(Q*,V*) of the operator A is defined through the relation

(A*q*,v) = (q*, Av) V v E V, q* E Q*

E.

We consider the minimization

inf J ( v , Av).

(2.5)

Let J be a function mapping V x Q into problem (the primal problem) uEV

60

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Define its dual problem by

where J* : V * x Q* -+

is the conjugate function of J :

For the relation between problems (2.5) and (2.6), we have the following duality theorem.

THEOREM 2.37 Assume the following conditions: ( 1 ) V is a rejlexive Banach space and Q is a normed space; A E C(V,Q). ( 2 ) J : V x Q -+ IW is propel; lower semi-continuous and convex. (3) There exists uo E V such that J ( u o ,Auo) < cc and the mapping q e J ( u o ,q)from Q to IW is continuous at Auo. ( 4 ) J ( v ,Av) -+ +m as llvll -+ m, v E V . Then the problem (2.5) has a solution u E V , the problem ( 2 . 6 ) has a solution p* E Q*, and J ( u , Au) = - J* (A*p*,-p*). (2.8) Furthermore, i f J ( v ,Av) is strictly convex in its effective domain, then a solution u of the problem (2.5)is unique. It is possible to weaken the assumptions of Theorem 2.37, then a weaker conclusion holds.

THEOREM 2.38 Assume: ( 1 ) V and Q are normed spaces; A E L ( V ,Q). ( 2 ) J : V x Q -+ is convex. ( 3 ) There exists uo E V such that J ( u o ,Auo) < cc and the mapping q e J ( u o ,q)from Q to IW is continuous at Auo. (4) infvEvJ ( v ,Av) is$nite. Then the problem ( 2 . 6 ) has a solution p* E Q* and inf J ( v , Av) = - J* (A*p*,-p*).

vEV

(2.9)

Furthermore, i f J ( v ,Av) is strictly convex in its effective domain, then a solution u (ifit exists) of the problem (2.5) is unique. This theorem is of special interest where the primal minimization problem does not have a solution; one can study the primal problem through the dual problem. The two problems are connected by the equality (2.9) and note that

61

Elements of convex analysis, duality theory

the dual problem does have a solution. In the rest of this work, though, we do not need Theorem 2.38. We will often encounter the situation where the function J is of a separated form, i.e., (2.10) J ( v , q ) = F ( v ) + G M , v E V , q E Q. It is then usually more convenient to compute its conjugate as follows:

+

J* ( v * ,q*) = F* ( v * ) G*( q * ) , where F* and G* are the conjugate functions of F and G, respectively. This follows from the definition of the conjugate functional. Specializing Theorem 2.37 to this case. we obtain the next result.

THEOREM 2.39 Assume: ( 1 ) V is a rejkxive Banach space and Q is a normed space; A E L ( V ,Q ) . ( 2 ) F : V -+ E, G : Q Q E are propel; lower semi-continuous, convex functions. ( 3 ) There exists uo E V such that F ( u o ) < cc, G ( A u o ) < cc and the mapping q + G(q) is continuous at Auo. ( 4 ) F ( v ) G ( A v )-+ +cc as llvll -+ cc,v E V . Denote J ( v , q ) = F ( v ) G(q),then J*(v*,qi)= F * ( v * ) G*(q*).There is a solution u E V to the problem (2.5),a solution p* E Q* to the problem (2.6),and (2.8)holds. Moreovel; if J ( v ,Av) is strictly convex on its effective domain, then a solution u of the problem (2.5) is unique.

+

2.7.

+

+

APPLICATIONS OF DUALITY THEORY IN A POSTERIORI ERROR ANALYSIS

Let u E V be a solution of the minimization problem (2.5). For any v E V , we define the energy difference

E D ( u ,v) = J ( v ,Av) - J ( u ,Au).

(2.11)

Let v E V be any element with J ( v ,Av) < cc. If the directional derivative J 1 ( ( uAu); , ( v - u , Av - A u ) ) exists, then we further define the quantity

D ( u ,v ) = J ( v ,Av) - J ( u , Au) - J 1 ( ( uAu); , ( v - U , Av

-

A u ) ) . (2.12)

THEOREM 2.40 We make the same assumptions as in Theorem 2.37. Then E D ( u ,v ) 5 J ( v ,Av)

+ J*(A*q*,-q*)

V v E V , q* E Q*.

(2.13)

Assume the directional derivative J 1 ( ( uAu); , ( v - u , Av - A u ) ) exists, then

62

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

and

D ( u ,v )

< J ( v ,Av) + J*(A*q*,-q*)

Vq* E Q*.

(2.15)

Proof. The inequality (2.13) follows from the definition (2.1I), the equality (2.8), and the definition of the dual problem (2.6): with p* a solution of the dual problem, E D ( u ,v ) = J ( v ,Av) 5 J(v,Av)

+ J* (A*p*,-p*) + J*(A*q*,-q*)

for any q* E Q*. Now assume the directional derivative J 1 ( ( uAu); , ( v - u , Av - A u ) ) exists. What remains to be proved is the inequality (2.14). Since u is a solution of the minimization problem (2.5), for any v E V we have the inequality

Thus,

Taking the limit t

+ 0+, we obtain J 1 ( ( uAu); , ( v - u , Av

-

A u ) ) 2 0.

Hence, (2.14) holds. H In the case J ( . ,A*)is Giteaux-differentiable at u , we can replace (2.12) by

D ( u ,v ) = J ( v ,Av) - J ( u ,Au) - ( J 1 ( uAu), , ( v - u , Av

- Au)).

(2.16)

where .) is the duality pairing between ( V x Q)* and V x Q. In most of the applications of Theorem 2.40 later in this work, J ( v ,q) is of the separated form (2.10), F ( v ) is linear over its effective domain dom ( F ) c V , and G : Q ?r IR is real-valued Giteaux-differentiable over Q. Obviously, ( a ,

dom ( F ) = dom J ( . , A , ) . Suppose v E dom ( F ) . Since F ( . ) is linear over dom ( F ) ,it is easy to see that

where G' denotes the Giteaux derivative, and (., .) is the duality pairing between Q* and Q. We now consider further the quantity D ( u ,v ) , starting with the expression (2.17). For this purpose, we recall two elementary formulas from Taylor's

63

Elements of convex analysis, duality theory

theorem. Consider a real valued function f ( t ) ,continuously differentiable for t E [0,I ] . Then

f

(1) - f (0) =

/

1 fl@)

0

(2.18)

dt.

Iff ( t )is twice continuously differentiable, then we can apply the above formula to f l ( t ) :

and get rl

f ( I ) - f (0) - f ' ( 0 )= J ( I

-

t ) f " ( t )dt.

0

Assume G(q)is continuously or twice continuously Gdteaux differentiable. For p, q E Q , we apply (2.18) and (2.19) to the real variable function G(p+ t (q-p) ) to get rl

and

+

respectively. Here, G1'(p t (q - p ) )(q - p) is a mapping from Q to R , since the second order G2teaux derivative G1'(p t (q - p ) ) is a mapping from Q to Q*. Then we have the formulas

D ( u ,v ) =

i1 1

+

(G1(Au t (Av - Au)) - G 1 ( A u )Av ,

1

D ( u ,V ) =

+

-

Au) dt,

(2.20)

+

(1 - t ) (Gl1(Au t (Av - Au))(Au- Au), Au - Au) dt, (2.21)

for any u , v E dom J ( . , A,). We will mainly use the formula (2.21) later. In many situations, G is strongly convex in Q: for some constant a > 0,

64

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

In this case, from (2.21), we can conclude that a V u ,v E dom J ( - ,A.). D ( u ,v ) 2 2 ( ( A v- A U I ~ ;

(2.23)

EXAMPLE 2.41 We consider an example which will be useful for later chapters. Let R be a domain in P S ~Suppose . A ( X ) = ( a i j ( ~ is ) )symmetric, ~ ~ ~ bounded and uniformly positive definite:

for some constant a > 0. Suppose A is a linear continuous operator from H' ( R ) to Q = ( ~ ~ ( 0 In ) )such ~ . a situation, A = V will be the gradient operator in later applications. Define the functional

for q E ( ~ ~ ( 0In )the ) ~ following, . we will usually omit the variable x in integrands. Then for any p, q E ( L 2( 0 )d),

By (2.21), we have

Since the matrix A ( x ) is uniformly positive definite in R, we have

Therefore, the condition (2.22) is valid and we have the following lower bound from (2.23):

In particular, for

Elements of convex analysis, duality theory

and then

1 D ( u ,v ) = - ljhv - hull2 2 (L~(R))* for any u , v E dom J ( - ,A.).

EXAMPLE 2.42 In Section 4.5, we have the situation with

where p > 1, and f l is a domain in TRd. Using the chain rule, we can find the GBteaux derivatives:

By Cauchy-Schwarz inequality, ( p - q ) 2

< IP12jq12.

Hence,

And then, for u , v E dom J ( . ,As), a lower bound for D ( u ,v ) could be

Now if p 2 4, then

(l+lVu+tV(v-u)I

2 ~12-2

)

21

and a further lower bound for D ( u ,v ) could be

This lower bound is not of desirable form, since the natural space for the solution u of the corresponding weak formulation (cf. Section 4.5) is wlJ'(fl). For p < 4, the expression (2.24) does not lead to a convenient form for a lower

66

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

bound. In such a situation, we will use directly the energy difference (2.11) to measure the difference between u and v. Armed with Theorem 2.40, the procedure of deriving an estimate for the difference between u and v is decomposed into two steps:

STEP 1. Find a suitable lower bound for D(u,v) that measures the difference between u and v. Usually, this lower bound will be some quantity depending on llv - ull.We will use (2.23), (2.21) or (2.20) for this purpose. When it is not convenient to relate the lower bound with a norm-like quantity, we will directly use the energy difference ED(u,v) to measure the difference between u and v. STEP 2. Construct an appropriate dual variable q* so that the bound from the right hand side of (2.13) or (2.15) is as accurate as possible. If q* is chosen to be a solution p* of the dual problem, then the right-hand side of the estimate attains its minimum. However, usually it is not easy to find p*. So it is desirable to have a strategy on determining a q* that is easy to get and that produces a good bound for the right-hand side of the estimate. The function q* is called a dual variable since it is related to the dual problem; it will also be called an auxiliary function. To use Theorem 2.40 for an error estimate, we will take u to be the solution of the original problem, v = uo the solution of an idealized problem or an approximate problem. We will construct suitable auxiliary functions q* based on the information from the solution uo and the idealized or approximate problem to produce good estimates for the error (u- uo). As will be evident from the rest of the work, it is amazing that the above approach can be used to derive a posteriori error estimates in so many contexts.

Chapter 3

A POSTERIORI ERROR ANALYSIS FOR IDEALIZATIONS IN LINEAR PROBLEMS

This chapter is devoted to a posteriori error estimates for idealizations of linear elliptic problems on nonsmooth domains. The goal is to provide easily computable, efficient estimates once solutions of the idealized problems are found. The duality theory of convex analysis, reviewed in Chapter 2, is used to derive a posteriori error estimates for coefficient, boundary condition and domain idealizations. The estimates obtained through the duality technique involve auxiliary functions (i.e. dual variables) subjected to certain constraints. Selection of auxiliary functions influences the accuracy of an estimate dramatically. We discuss in length, especially for coefficient idealization, various selections of auxiliary functions. Numerical examples show that our selections lead to efficient error bounds. For quantitative error estimates of right-hand side idealization, only elementary calculus techniques are involved. Nevertheless, for the sake of completeness, a detailed derivation of error bounds for the right-hand side idealization is also given. In the first four sections, we study the effects on solutions of idealizations in coefficients, the right-hand side, the boundary condition and the domain for model second-order linear elliptic boundary value problems. The model problems discussed are taken to be the simple boundary value problems of the Poisson or Poisson-like equations. However, all the discussions extend easily to other linear elliptic boundary value problems, including the linearized elasticity systems. In the last two sections, we apply the results obtained for the derivation of a posteriori error estimates for material idealizations involved in linear torsion problems and for various idealizations in heat conduction problems. We comment that it is possible to derive a posteriori error estimates for combined effects of idealizations in several or all of the data. The resulting error

68

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

estimates will take rather complicated forms. To illustrate the error estimates more clearly, we present the results for idealizations in each type of data. For the numerical examples in this and later chapters, when exact values of quantities are not available, they are computed sufficiently accurately through numerical approximations, with the correct digits displayed. Boundary value problems without known solutions are solved by the finite element method with sufficient accuracy.

3.1.

COEFFICIENT IDEALIZATION

The coefficients of a differential equation describing a physical process characterize physical properties of the material. They are phenomenological and are obtained from physical experiments. Typically in applying the linearized elasticity theory for commercial materials, the coefficient uncertainty is in the range of 3 5%. It is important to assess quantitatively the influence of this uncertainty in the concrete environment of the problem being solved. We derive error estimates addressing this issue in this section. In Sections 3.5 and 3.6, the results are used to give a posteriori error estimates for material idealizations occurred in linear torsion problems and steady state heat transfer problems. Let R C IRd be a Lipschitz domain with its boundary decomposed as d R = rl U with I'l relatively closed, r2relatively open, rl n r2 = 0 and meas (rl)> 0. Since the boundary is Lipschitz continuous, the unit outward normal v = (vl, . . . , vd)T exists a.e, on dR. We consider linear elliptic boundary value problems allowing general mixed boundary conditions. (0), 1 5 i, j 5 d, and f be given functions in R , gl and g2 be given Let aij, aij functions on rl and r2.Consider the boundary value problem

-

a

--(aij*) dxj

dxi

=f

inn,

and its idealization

respectively. It is possible to consider more general equations, such as

d (aijg) dxj

--

+ cou = f,

69

A Posteriori Error Analysis for Idealizations in Linear Problems

>

where co( x ) 0 is a given function. Since the corresponding dual problem has a slightly different form at places where co ( x ) > 0 from that where co ( x ) = 0 , we choose to consider the model problem (3.1)-(3.3) only, for the sake of a more concise illustration on the usage of the duality technique. We assume

for a.e. x E f l and any J E IRd, where the bounding coefficient functions are bounded and positive: inf c l ( x ) , inf cl( 0 )( x ) > 0 ,

XER

XER

sup c 2 ( x ) sup , c2( 0 ) ( x ) < co.

x€R

xER

(3.9)

Implicit in the inequalities (3.7) and (3.8) are the assumptions that c 2 ( x )and c?) (I)are the best possible (or nearly best possible) upper bound functions for these inequalities, whereas cl ( x ) and cl( 0 )( x ) are the best possible (or nearly best possible) lower bound functions. We also assume

Regarding the assumption g l E H 1 ( a ) ,it is more natural to assume g l to belong to some Sobolev space over the part of the boundary rl and then apply trace theorems to extend g l to a function in f l with certain degree of smoothness. Here we assume from the outset that the function gl itself belongs to H 1( 0 ) and we use its trace on rl in the Dirichlet conditions (3.2) and (3.5). We will do the same in the rest of the work. We use the notation g l Hhl ( f l ) to denote the set

+

The weak formulation of the boundary value problem (3.1)-(3.3) is to find u E gl H ; (~R ) such that

+

and that of the boundary value problem (3.4)-(3.6) is to find uo E g l such that ( 0 ) duo dv a.. - -dx 'J dxi d x j

S,

=

f vdx

+l

g2vds V v E

2

+ H& ( f l )

~ ; ~ ( f l ) .

(3.12)

70

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Applying the Lax-Milgram Lemma (Theorem 1.16) to the unknown variables ( u - g l ) and (uo- gl), we see that under the stated assumptions, both problems (3.11) and (3.12) have a unique solution. is not known exactly We consider the situation where A ( x ) = ( 0 ) and we view A o ( x ) = (aij ( x ) ) d X as d an idealization of A ( x ) . We are interested in using the solution uo of the idealized problem (3.4)-(3.6) and some given bound on ( A ( x )- A. ( x ) )to estimate the error ( u - u o ) .

3.1.1

A GENERAL ERROR ESTIMATE

With the symmetry assumption on the coefficient matrix A ( x ) ,we note that the weak solution u E g l H;, (G) of the problem (3.1)-(3.3) is a minimizer of the energy functional

+

+

over the set g l H& ( R ) .To apply the duality theory for an estimation of the error ( u - u o ) ,we set

Then V * = ( H 1 ( G ) ) *and we identify Q* with Q. For q E Q, we write q = ( q l ,. . . , q d ) T with qi E L ~ ( R )1 , 5 i 5 d. We use a similar notation for q* E Q*. We define the operator A by the formula

Av = V v , which is obviously linear and continuous from V to Q. Corresponding to the functional (3.13), we introduce the functional

J ( v ,Av) = F ( v )

+ G(Av)

over the space V with

Then u is the unique solution of the problem inf J ( v ,Av).

vEV

71

A Posteriori Error Analysis for Idealizations in Linear Problems

We see that dom F = dom J ( , ,A*)= gl From Example 2.41, we know that

;

D ( u ,u O )= -

+ H;l ( R ) .

V ( U- U

~ ) ~ A V (U u 0 )dx.

JR

To find an upper bound for D ( u ,uo ) ,let us determine the conjugate functional J*. We first compute the conjugate functions of F and G. By definition,

F*(A*q*)= sup{(Av,q * ) - F ( v ) ) , vEV

we have

For the inner product of two vectors q l , q2 E E l d , we use ql 42 and qTq2 interchangeably. Replacing v E gl H : ~( R )by v g l with v E H:l (R), we have

F*(A*q*)=

a

+

+

(q* . Vgi + f g l ) dx +

g2g1

ds if q* E Qz, otherwise.

Here, the constraint set

We notice that the constraint

is the weak form of the conditions

We then compute the conjugate of G. By definition,

= sup 4E Q

{-

R

(q*Tq

+ 21 q T A q ) dz}.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

72

Applying Theorem 2.35, we have

G* ( - q * ) = -

1

sup (q*Tq ,ER*

+ j1 q T A q ) d x .

For fixed q* = q* (x),the supremum in the integrand is obtained when

i.e. when q = -A-lq*. Then we see that G*(-q*) =

Sn

1 2q*T~-1q*dx.

Thus, we have the following formula for the conjugate functional:

+o otherwise. Hence, by using Theorem 2.40 and (3.14), we obtain

for any q* E QF. Note that we have the equality

from the weak formulation (3.12) for the solution uo. Therefore, we have proved the following estimate.

T H E O R E3.1 M For the difSerence between the weaksolutions of the two boundary value problems (3.1)-(3.3) and (3.4)-(3.6), there holds the estimate:

A Posteriori Error Analysis for Idealizations in Linear Problems for

any q* E

QE.

The error estimate involves an auxiliary function or dual variable q* in a constraint set. An auxiliary function or dual variable in the constraint set will be called an admissible auxiliary function or admissible dual variable. The quality of the error estimate depends on the choice of the auxiliary function, a topic for the next subsection.

3.1.2

SELECTION OF THE AUXILIARY FUNCTION

For definiteness, in this subsection, we consider Dirichlet problems only; selection of the auxiliary function for other boundary value problems can be made similarly. So the original problem is

and the idealized problem is -

8dxj

" )'xi (0)du

=f

inn,

The constraint set for the auxiliary function is then

For a proper choice of the auxiliary function q* in the estimate (3.15), we distinguish two cases. To avoid the scaling effect, the upper bound function in the inequality (3.7) is assumed to be of size 0 ( 1 ) , i.e., for two constants 0 << rnz 5 M2 < m w i t h ~ ; >> ~ 0, wehave

CASE 1 . The lower bound function in the inequality (3.7) stays well away from 0, cl(x) where the constant cl

> cl

Vx E 0,

> > 0.In this case, we may take

(3.20)

74

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

which belongs to the constraint set QE. Then from Theorem 3.1,

Using the condition (3.20), we find the norm estimate

Notice that if A. is close to A, then the error bound of the estimate (3.22) is guaranteed to be small. The error ( u - uo)in either the energy norm

or the H i (R)-norm

llv(u - ~ o ) l l ~ 2 ( n ) has an error bound proportional to the error (A - Ao)in the coefficients. This kind of linear dependence of an error bound on the error in the data is the most one can expect. The effectiveness of the estimate will be shown in some numerical examples later.

CASE2. The lower bound function in the assumption (3.7) is close to zero in some part of 0, i.e., for a small positive number 6, inf cl ( x ) R

< 6.

(3.23)

This assumption could correspond to material damage in a local region for mechanics with damage effect. In this case, the matrix A ( x ) may be nearly singular in some region of R and so the selection (3.21) generally produces a "poor" estimate owing to the appearance of A-l on the right-hand side of the estimate (3.22). Therefore, we need to look for a better selection. To simplify the presentation, let us restrict now the consideration to a planar corner domain and the idealized problem for uo involves a Poisson equation. We assume the coefficients degenerate only around 0, i.e., there is an 7-1 > 0 such that on

we have the inequality

A Posteriori Error Analysis for Idealizations in Linear Problems

Figure 3.1. Choice of the auxiliary function near a comer domain: Subcase 2A

with

cl(x)> cl >> 0 a.e. x E Q,, .

Denote

Ro,,l = {x E R

:

1x1 < TI}.

We assume

f = 0 in Qo,,, . The necessity of the assumption will be explained in Example 3.2 below. We consider two subcases depending on the strength of the singularity.

SUBCASE2A. First we consider the situation where on QQ1,,, cl(x)goes to 0 slowly as x + 0 ;more precisely, we assume there are constants c,,B > 0 such that c l ( 4 c lxlB, 1x1 < r1.

>

We introduce a function

with the parameter y to be chosen, where vl is the unit outward normal on

r,, = {X E 0 : 1x1= rl ). See Figure 3.1 for the geometry. Then we define

76

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

This is an admissible function (i.e. q* E QF) since for any v E H; (R),

Here, d l d r denotes the tangential derivative. Using (3.25) as the choice for the dual variable, we have, by Theorem 3.1,

A Posteriori Error Analysis for Idealizations in Linear Problems

Figure 3.2. Choice of the auxiliary function near a comer domain: Subcase 2B

Easily we see that the quantity (el 12)lV(u

-

IiZ(nrl is bounded by the

u0)

right-hand side of (3.26). To cancel the effect of unboundedness of EO,,,, we choose y so that

2y

A-l on

+ 2 > p.

SUBCASE2B.Now we consider the situation where on Eo,,, , cl (x)goes to 0 + 0. We choose an r2 E (0, rl), and denote

rapidly as x

Denote v2 the unit outward normal vector on geometry. We construct a function

r,,.

See Figure 3.2 for the

where 6 > 0 is a constant to be suitably chosen. Then we take

78

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

which is an admissible function since for any v E H; ( R ) ,

LTl

(Vy,-V,).vl~ds-

+

LT2

(Vy,-V,)-v2~ds

= L f u d x - L T l $vds

LTlz LT2 -v d s

+

=

S,

-

u ds

vdx.

Using (3.27) as the choice for the dual variable, we have, by Theorem 3.1,

+/

[-(V,,-I/,)Vg R o , q \R0,T2

1 + -(v,, -v,)A-'(I/,, 2

-v,)T]d x .

(3.28)

It is easy to see in particular that the quantity ( c 1 / 2 ) I V ( u - uo)/2 ( R T l ) is bounded by the right-hand side of (3.28).

79

A Posteriori Error Analysis for Idealizations in Linear Problems

We will examine the effectiveness of the estimates (3.22), (3.26) and (3.28). In particular, we will see that for some problems with degenerating coefficients, selections of the auxiliary function according to Case 2 produce much better error bounds than those according to Case 1. Let us now look at an example which shows that when there is no assumption (3.24), the exact error itself will blow up. EXAMPLE3 . 2 Let R = {(r, 8 ) : 0 5 r < 1) be the unit disk, rl 6 > 0 a small number. Use r = 1x 1. Let the coefficient matrix be

E

(0,I ) ,

We consider the Dirichlet problem with f (x) = -4, g(x) = 1. Then the solution of the problem is

whereas the solution of the idealized problem is

Hence.

i.e., the exact error blows up as S + O+.

1

REMARK3 . 3 Some error estimates derived by the duality approach in this work can be shown by more elementary argument. This is the case with the estimate (3.22). Let us prove it directly. Subtracting (3.12) from (3.11), we obtain the relation

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

80

In particular, with v = u

-

uo, we have

Consider the quantity

Using (3.29), we have

S,

V ( u- U O ) ~ A V ( U - uo)dx

V ( u- U ~ ) ~ A V ( U uo)dx

.

[l

I

li2

V U ~ ( A ~ A-- I' ) A (A-'Ao

112

-

I ) Vuodx]

,

which is (3.22). Although direct calculations, whenever possible, are more straightforward, we note that the duality approach is a more powerful, more general way to systematically derive error estimates. In the case of error analysis for idealizations of the coefficient functions here, it is not clear how to directly prove the estimate (3.28), which is preferable to (3.22) in certain situations (cf. the numerical examples later). On the other hand, an estimate such as (3.28) naturally follows with the duality approach. In the more complicated problem setting when nonlinear boundary value problems are involved, direct derivation of error estimates appears difficult. Nevertheless, the duality approach can still 1 be employed in a systematic way. See Chapter 4 for detail.

REMARK 3.4 In the finite element method for solving boundary value problems, a time-consuming part is the construction of the stiffness matrix, that is usually done with the use of sufficiently accurate numerical quadratures to compute integrals. Thus, the actual stiffness matrix used in the finite element system is only an approximation of the true stiffness matrix. The error estimates shown in this subsection can be extended to their discrete analogues in assessing the effect of the stiffness matrix approximation on accuracy of the 1 finite element solutions.

A Posteriori Error Analysis for Idealizations in Linear Problems

3.1.3

81

AN ITERATIVE PROCEDURE

+

We consider a situation where A ( x ) = A. (x) E ( x )is known with E ( x ) = small and such that the boundary value problem with the coefficient (eij matrix A. ( x ) can be solved efficiently (say, via a fast solution solver). In this situation, it is advantageous to solve the boundary value problem with the coefficient matrix A ( x ) by an iterative procedure solving a sequence of boundary value problems with the coefficient matrix A. ( x ). Another situation where such an iterative procedure is of interest is when we need to solve a family of boundary value problems for each of them the coefficient matrix A ( x ) is close to a fixed matrix A. ( x ) . Once more, let us focus on the Dirichlet problem. We write

We assume e i j ( x )is "close" to 0. By Theorem 3.1, we have an estimate for the difference ( u - uo). It may turn out that the estimated error is not small enough. In this case, let us note that w = u - uo satisfies

Thus, if we denote wo the solution of the problem

+

we may expect that (uo wo)is a better approximation to u. Applying Theorem 3.1 to these problems, we have the next result.

T H E O R E3.5 M There holds the estimate ~ ( V U -

+

+

V ( u o W O ) ) ~( VAU - V ( U O W O ) )dx

(VWTAVW~ - q * T ~ - l q * dx ) for any q* E QE,o,where the constraint set

82

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Selection of a good q* can be discussed as in Subsection 3.1.2. If the estimated error for u - ( u o w o ) is still not small enough, we may repeat the above procedure. We define a sequence { w k ) k-> lby

+

wl, = 0 For Ic

> 1, let

Then we have the following result.

T H E O R E3.6 M There holds the estimate

It is possible to provide sufficient conditions that guarantee the convergence of the iterative procedure. For this purpose, we introduce three differential operators A, A o , £ : V -+ V*, with V = H ; ( R ) and V* = H - ' ( a ) , by the formulas

+

Then, A = A. £, u = A-I k = 0 , 1 , . . . . Denote

f, uo

=

A, 1 f , and wk

= ( - A ~ ~ £ ) ~ + ~ U ~

where Ami,(Ao ( x ) ) > 0 is the smallest eigenvalue of A. ( x ) ,A m a X ( E ( x )is) the eigenvalue of E ( x ) with the largest magnitude.

83

A Posteriori Error Analysis for Idealizations in Linear Problems

Consider the operator A;'& : V -+ V, and let us bound its norm. Let w = &'&v for any v E V. We have Aow = &v, i.e.,

Taking x = w in the relation, we obtain collwll$ 5 eollwllvllvllv,

Now if E ( x ) is small in the sense that eo < co, then by the geometric series theorem (see, e.g. [6, page 46]), Ego(-A,'&) converges to (I A;'E)-~. Then, as k -+ oo,

+

converges to

Meanwhile, wk -+ 0 in V as k q* = -EVwk-1 gives:

-+ oo.

The error bound in Theorem 3.6 with

A simple error bound is obtained by dropping the second integrand from the integrand:

84

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

which is in turn bounded by

with An advantage of using the iterative procedure is that if the finite element method is used to approximate uo, wo,wl,. . . , then the stiffness matrices are the same for all the finite element systems. This provides substantial time saving in forming and solving the finite element systems. The above consideration works for other kind of boundary value problems as well.

3.1.4

NUMERICAL EXPERIMENTS

We present here several numerical examples to show the effectiveness of the error estimates derived in this section.

EXAMPLE 3.7 Consider the problem

and its idealization

Here, R is a two-dimensional domain, aij ( x ) = dij Kronecker delta symbol:

+ eij ( x ) ,and Sij is the

) symmetric (to ensure the symmetry of the coefficient Let E ( x ) = ( e i j ( x )be , we matrix A ( x ) ) . We assume some knowledge about the size of e i j ( x ) but do not assume these functions are known exactly. Suppose first that we know some bound on the eigenvalues X I ( x )and X2(x)of E ( x ) :

for some E E (0,1 ) . We use the selection (3.21) to get the estimate

V ( U- U

O ) ~ VA( U -

u O dx ) 5

S,

+

VU;(A A-'

-

21)Vu0dx.

A Posteriori Error Analysis for Idealizations in Linear Problems

85

To proceed further, we have

Then

and so

By using the bounds on the eigenvalues of E(x),we have

where the matrix norm

11 . 11 is the spectral norm. Thus, we have the estimate

For example, if the domain is the unit square S2 = (0,I ) ~ ,f (a) = 1,

g(x) = 0,then / / V U ~ / /=~ 0.1874679.. Z(~) ..

As another example, we choose S2 to be an L-shape domain,

f (x)= 1, g(x) = 0. Then

86

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

We note that the estimate (3.34) is optimal. For if we assume eij ( x ) = -edij, then u = uo/ (1 - E). Thus,

More often, for the perturbation matrix E ( x ) ,we are given a bound on its entries rather than that on its eigenvalues. We assume

where

E

E

(0,O.s).Then, the estimate is

EXAMPLE 3.8 We reconsider the boundary value problem (3.30)-(3.3 1) and its idealization (3.32)-(3.33). This time, the coefficient matrix is assumed to be

where r = 1x1. We assume the coefficient function has the property that a ( r ) -+ 1 as r -+ I , and a ( r ) -+ 0 as r -+ 0. This property is assumed so that the differential equation exhibits singularity near r = 0. As examples of the coefficient function, we may take

The domain is a circular comer region with an internal angle w :

0 = { ( r , 8 ): 0 < r < I , 0

< 8 < w).

Denote a = T / W . Let f and g be such that the solution of the idealized problem (3.32)-(3.33) is ug = ra sin a8. We seek the solution in the form

A Posteriori Error Analysis for Idealizations in Linear Problems

It turns out that R is a bounded solution of the problem

In particular, for

a ( r ) = r'ln, n > 1 ,

Then it can be shown that the exact energy norm error is

The estimated error bound is

where q* E ( L( ~R ) )satisfies ~ the condition

We select q* according to the formula (3.25). For our problem,

V ( r ,0 ) = a r y - y - l r y + l

li2

sin a0 d0

After a tedious calculation, we find an analytic expression of the error bound:

where we assumed 2 ( y

+ 1 ) > lln. Note that a2 min 2(y+l)>l/n2 ( y

+ ( y+ I ) ~ +1) - l/n

88

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

For this y,the second term of (3.35) is negative, so we take

In conclusion,

To see the efficiency of the error bound, we use the effectivity index

Some numerical results are given in Table 3.1 on the values of the true error, error bound, and the effectivity index, for several choices of the internal angle w and the parameter n. We observe that the efficiency of the error bound is rather stable with respect to the increasing strength of comer singularities, i.e., as the singularity strength increases due to an increase of the internal angle, the effectivity index does not grow out of control. We also observe that the efficiency of the error estimate improves as the parameter n increases.

EXAMPLE 3.9 We again consider the problem (3.30)-(3.31) and its idealization (3.32)-(3.33) with the assumptions stated in Example 3.7. Consider the iterative procedure described in Subsection 3.1.3. Applying Theorem 3.5 with q* = -Vwo, we get the estimate

+

(VU - V(UO

W O ) ) ~ A( V u -

+

V(UO

WO))

dx

We then obtain

We can bound IIVwo11 L 2 ( R ) in terms of IIVuo11 L2(n),using the definition of Wo :

A Posteriori Error Analysis for Idealizations in Linear Problems

Table 3.1. Example 3.8, effectivity of error bound

i.e.,

<

l l ~ w o l l ~ ~ ( nE I) I V U O I I L ~ ( ~ ) Combining (3.36) and (3.37), we then have

Let us iterate further. Using Theorem 3.6 with q* = -Vwk, we get the estimate

Arguing as above, we have

90

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

+ + + w k ) converges to

Therefore, we conclude that the series { u o wo - . . u in H; (R), and we have the quantitative estimate (3.38).

E X A M P L E3.10 We have seen in Example 3.2 that when the elliptic differential operator in (3.1) degenerates in some region, usually we have to require f to vanish there. Even when f is identically zero, the simple choice (3.21) will lead to an error estimate which blows up if the problem degenerates, as can be seen from the estimate (3.22). We show by an example that the auxiliary function discussed in Case 2 in Subsection 3.1.2 still provides us reasonable error bound. Consider the problems (3.16)-(3.17) and (3.18)-(3.19) with

R

= {(r,O) : 0

< r < 1, 0 < 0 < w ) ,

and

where r2 E ( 0 ' 1 ) . Let a = ~ / w We . consider the Dirichlet problems with f = 0 and g = sina0. We are interested in the case when > 0 is small. Analytic expressions of the exact solutions can be found:

uO( x ) = r a sin a0, (1 P)ra ( 1 - P)riar-" sina0, r > r z , (1 + P ) + ( 1 - P)r,2" u ( x )= 2 ra sin a e , r < r2. ( 1 P ) + ( 1 - P)r,2"

+

+

+

The exact error is

The estimated error bound from the simple selection (3.21) is

which is readily seen to go to oo as P + O+. The error bound from the selection (3.25) is

A Posteriori Error Analysis for Idealizations in Linear Problems

Table 3.2. Example 3.10, effectivity of error bounds

where rl E ( r 2 ,1) and S > 0 are to be chosen. We take r2 = 0.1, r1 = 2 7-2,S = 1. Table 3.2 lists the exact error

its upper bound Eestl from (3.22) and another upper bound Eest2from (3.26), for various domains (i.e., various angles w) and various parameters P. The effectivity index in this example is

Again, we observe the efficiency of our error estimates.

3.2.

RIGHT-HAND SIDE IDEALIZATION

The interpretation of the right hand side of a partial differential equation for many mechanical problems is the load. Physically, usually the "worst scenario"

92

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

loads are of the major concern. The sensitivity of solutions to changes of loads needs to be assessed, especially changes of the loads in a neighborhood of critical places, such as corners. In a finite element system for solving a partial differential equation problem, the load vector is usually computed through the use of a numerical quadrature. The effect of the right-hand side approximation on the solution accuracy should be assessed quantitatively, and not only qualitatively. In this section, we bound the influence of the right-hand side function near a comer on the solution of a linear elliptic boundary value problem. We consider a domain R as shown in Figure 1.6 (page 26). Let R, be a small region around the comer point 0. We denote uo,u,E H' ( 0 ) the solutions of

and

respectively. As before, dR = rl uE,rl relatively closed, r2relatively open, and rl n rz= 0. We assume f,(x) = f ( x ) f o r x E R\R, and

f, - f

E ~ ~ ( 0 ) .

We also assume R, c for some fixed domain R1.

3.2.1

ERROR ESTIMATES

The weak formulations of the boundary value problems are

(3.45)

A Posteriori Error Analysis for Idealizations in Linear Problems

Subtract the two relations to obtain

In particular, taking v = u,

Denote w = u,

- UO.

-

uo, we get the inequality

The error w E

HA

( 0 ) and satisfies

We denote yl the part of X I L corresponding to that on 80, with a Dirichlet condition, and 7 2 the remaining part of a n l . We distinguish two cases according to whether yl is empty. CASE 1,meas(yl) Denote

# 0.

Introduce a change of variables x = 6 y , and an auxiliary function ?J defined on fll by the formula ?J(y ) = w (E y ). Then,

Thus, from (3.48), we obtain

CASE 2 . meas(yl) = 0. In this case, we assume that the idealized data f o f f , satisfies S,f.dx=/ R fds,

94

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

i.e., the average value of f and that of f, are equal. This relation between f and f, is quite natural since f is viewed as an idealization off,. Denoting

the average value of v on R1.Let

From

we obtain once more the estimate (3.50). Next, we comment on the number X1 for both cases. For Case 1, from the definition (3.49), we see that X1 eigenvalue of the problem

For Case 2, from the definition (3.51), we see that XI eigenvalue of the problem

> 0 is the smallest

> 0 is the smallest

For general domains Rl or general decomposition of d R 1 into 71 and 72, we have to compute X1 numerically. However, for certain special situations, it is possible to find XI analytically. In the following examples, we consider Dirichlet problems.

EXAMPLE 3.11 Assume R1 = (0,a)x (0,b) is a rectangle. For convenience, in this example, we use (x,Y) instead of (xl, 2 2 ) to denote the coordinates of a

A Posteriori Error Analysis for Idealizations in Linear Problems

95

point in 2.We use the method of separation of variables to solve the eigenvalue problem (3.52)-(3.54). In our special case, yl = ( ( 0 ,a ) x ( 0 ) )U ( ( 0 ) x ( 0 , b ) ) , "/z = ( ( 0 ,a ) x { b ) ) U ( { a ) x ( 0 ,b ) ). Substituting the expression v ( x ,y ) = X ( x ) Y ( y ) into (3.52), we find two sub-eigenvalue problems

XI1 = - ( A - p ) X in ( 0 ,a ) , X ( 0 ) = 0 , X 1 ( a )= 0 ,

(3.58) (3.59)

and

Y" = - p Y in ( 0 , b ) , Y ( 0 ) = 0 , Y 1 ( b )= 0 . Solving (3.60)-(3.6 I ) , we get a sequence of eigenvalues

and a sequence of corresponding eigenfunctions:

For each pl, the eigenvalue problem (3.52)-(3.54) admits a sequence of eigen-

with a sequence of corresponding eigenfunctions:

Thus, for the eigenvalue problem (3.52)-(3.54), we obtain a sequence of eigenvalues:

with a sequence of corresponding eigenfunctions:

It is known that (cf. [141]),{Xk)F=oforms an orthogonal basis of the space {V E H 1 ( 0 ,a ) : v ( 0 ) = 0 ) , { ~ ) forms ~ oan orthogonal basis of the space

96

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

{ v E ~ ' ( 0b), : v ( 0 ) = 0 ) . Thus, the sequence (3.63) forms an orthogonal ) 0 on 7 1 ) . Therefore, all the eigenvalues basis of the space { v E ~ ~ ( 0 : v1 = of the problem (3.52)-(3.54) are X k , ~k, , 1 = 0 , 1 , . . . , with the corresponding eigenfunctions X k ( x ) x ( y ) ,k , 1 = 0 , 1 , . . . . In particular, the smallest eigenvalue is

The estimation in this special case is

E X A M P L E3.12 Assume Rl is a circular corner domain: 0 1 =

{ ( r , Q ): 0

< r < r0,O < Q < w ) .

In this case, yl = { ( r , Q ): O < r < r o , Q = O , w ) , yz = { ( r , ~: )r = ro,O < Q < w ) .

By expanding v into the series C R k ( r )sin k a Q , a = T / W ,we see that the eigenvalue problem (3.52)-(3.54) is reduced to the problem: For each positive integer k , find the eigenvalues X of

lTO < R: ( r )r dr

co.

On the other hand, by using (3.49), we obtain

X 1 = min Therefore, X 1

1

J; [ r ~ ' ( r+)~~ ~ R ( r ) dr ~ / r ] :J

~R(r)~dr

> 0 is the smallest number such that the problem

A Posteriori Error Analysis for Idealizations in Linear Problems

has a nontrivial solution R = R ( r ) . By making the change of variable

we conclude that XI

> 0 is the smallest number such that the problem

has a nontrivial solution R = R ( r ) . From the differential equation (3.64) and the condition (3.66), we find that

where J, is the Bessel function of the first kind of order a (cf. [104]),

An integral representation of the Bessel function is Ja(4 =

(~/2)"

1

Thus by (3.65), XI

+1

In

cos(r cos 0) sin2a do.

(3.67)

0

> 0 is the smallest root of the equation

Using the representation formula (3.67), we see that 6ro > 0 is the smallest root of the equation

la

sin2, B (or cos(r cos 0) - r cos 0 sin(r cos 0)) d0 = 0.

(3.68)

We use Newton's method to solve (3.68). The algorithm is the following. (a) Initialization. Find an r ( ~>) 0 close to the smallest root of the equation (3.68). (b) Iteration. For k = 0,1, . . .,

where, c(k) =

J:

J: sin2ff0 (a cos(ck(0)) - r ( k )cos 0 sin(ck(0))) d0 0 ((1 + a ) cos 0 sin ck(0) + dk)cos2 0 cos ck (0)) de

98

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 3.3. Example 3.12, smallest roots for several angles

and ck (8) = d k )cos 8. For the computation of the two integrals on the right hand side of the iteration formula, we divide the integration interval into m equal-length subintervals, and on each subinterval, we use the six point Gauss quadrature formula. Numerical H results for several values of the angle are listed in Table 3.3.

3.2.2

NUMERICAL EXPERIMENTS

Below, we present some numerical results to show the effectivity of the error estimate (3.50).

EXAMPLE 3 . 1 3 Assume R = (0, I ) ~E , = 0.2, R, = ( 0 , 0 . 2 ) ~and ,

whereas the error bound from the estimate (3.50) is

If, instead of (3.69), we assume

Then,

jlV(u,

-~

= 0.006141 .

) l l ~ 2 ( ~ )

whereas the error bound from the estimate (3.50) is

'

(3.72)

A Posteriori Error Analysis for Idealizations in Linear Problems

99

Comparing (3.70) with (3.71), and (3.72) with (3.73), we see that our estiW mates provide good error bounds. Let us examine another example.

EXAMPLE 3.14 Assume 0 is an L-shape domain:

and

We choose

E

= 0.3. Then

From Example 3.12, we have 6= 1.4013. - - (note that ro = I). Thus, the estimate (3.50) implies

If, instead of (3.74), we assume

then

IIV(u, - u ) I I ~ ~=( 0.01256 ~) -

' '

whereas the error bound from the estimate (3.50) is

Now assume in fl\fl,, Here as usual, we use r = 1 x 1. Then

whereas the error bound from the estimate (3.50) is

In all the cases, we see that our estimates provide good numerical results. W

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

100

3.3.

BOUNDARY CONDITION IDEALIZATIONS

The boundary condition idealization is one of the most delicate aspect of the reliability of the computational analysis. Let us mention a concrete example which will be considered in detail in Section 3.6. Consider a heat (or membrane) problem leading to the Poisson problem with zero boundary condition. In reality, such a boundary condition does not exist, because the temperature (or displacements) can not be completely controlled by the surrounding medium. A more accurate description of the boundary condition would be of the type c ( x ) d u l d u u = 0.The function c ( x )has small values and depends globally on many factors. In practice, usually a set of boundary conditions needs to be analyzed. For some discussions in the context of linearized elasticity, we refer to [7], where the influence of boundary conditions on the reliability of the results is addressed in a computational way. Let us mention that the uncertainty in the boundary condition is especially large in the neighborhood of corners of the domain, where the theoretical fluxes (or stresses) could be infinite. A typical example in the elasticity is the plate paradox for hard, soft supports which practically can not occur in reality, because such conditions can not be practically imposed. For detail, see e.g., [9]. Hence, estimates of boundary condition idealizations are of major importance. In this section we show how to employ the duality technique to derive a posteriori error estimates for boundary condition idealizations. We will consider two idealization problems. Slight extensions of the results will be used in quantitative error estimation of some idealization problems related to heat conduction in Subsections 3.6.2 and 3.6.3.

+

BOUNDARY CONDITION IDEALIZATION (I) Let fl c Titd be a Lipschitz domain with its boundary dfl = Fl U G, rl relatively closed, r2relatively open, rl n r2= 0,and meas(rs) > 0.Let there 3.3.1

be given f E L2( Q ) , g E H1( a ) , A E Lw ( r 2 ) such that for some constants 0 < A. < Al < m,

We will be interested in the case where A. is of large size. Consider the boundary value problems

A Posteriori Error Analysis for Idealizations in Linear Problems

101

and -

nuo = f uo = g

infl, ondfl.

Note that on F l , the two problems have the same Dirichlet boundary condition. On F2, since the coefficient function A ( z ) is assumed big, the condition (3.77) is close to the condition (3.79). The weak formulation for the boundary value problem (3.75)-(3.77) is to find u E g H& (0)such that

+

and that for the boundary value problem (3.78)-(3.79) is to find uo E g+ H i (fl) such that S,vuo.vvd~= f ~ VdU E ~H ~ ( R ) .

S,

Our purpose here is to bound the difference (u - uo). We introduce the folowing function spaces:

and define the operator A E C(V, Q ) by

Let J(v,Av)=F(v)+G(Av),

VEV,

with

Here for q E Q, we write q = ( q l , q2) with q l E (L2(0))d , q 2 E L2(F2). A similar notation is used for q* E Q*. Then u E V is the unique solution of the minimization problem inf J ( v , Av).

vEV

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

102

For a lower bound of D ( u ,u o )defined in (2.12), we note that

Hence, by (2.21),

D(u, u 0 ) =

lo(1

+ t(AuO- Au)) (AuO- Au),AuO - Au) dt

- t ) (Gtt(Au

We now compute the conjugate functions.

F* (A*q*)= sup{ (Av,q*) - ~ ( v ) ) v€V

Introduce the constraint set

Then we have

For the conjugate function of G , we have

A Posteriori Error Analysis for Idealizations in Linear Problems

Thus,

Hence, applying Theorem 2.40, noting uo = g on we have the next result.

r2,and rearranging terms,

T H E O R E3.15 M For the difference between the solutions of (3.75)-(3.77) and (3.78)-(3.79), there holds the estimate:

for any q* E QE. Assume d u o / d v E L 2 ( r 2 ) .Then from (3.78)-(3.79), we have

*

So a simple choice of the auxiliary function is q; = -Vuo, q2 = and we obtain from Theorem 3.15 the following estimate:

Using the bound A ( x )

auo a V

-

A g,

> Ao, we obtain from (3.80) that

In the derivation of the estimates (3.80) and (3.81), we assumed d u o / d v E L 2 ( r 2 )on the solution uo of the problem (3.78)-(3.79). This assumption is valid if the domain is smooth, or the domain is a comer domain without cracks, and g is the trace of an H~ (R) function.

104

3.3.2

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

BOUNDARY CONDITION IDEALIZATION (11)

Let there be given a function a on dR with where, a0 I a1 are two positive constants, supposed to be small. Also given . are f E L2( R ) ,g E H1( R ) ,and h E ~ ~ ( 8 0We) assume P

Consider the boundary value problems

and

respectively. A weak formulation of the boundary value problem for u is to find u E H 1 ( R )such that

The compatibility condition (3.83) is assumed to ensure the idealized problem (3.86)-(3.87) has a solution, which is unique up to an additive constant. We use the duality theory to estimate ( u - u o ) .Let

d

For q E Q, we write q = ( q l ,qz) with ql E ( L ~ ( R ) and ) q2 E L2( d R ) . A similar notation is used for q* E Q*. Define the energy functional

J ( v ,Av) = F ( v ) with P

F ( v ) = - Jn f v d x ,

+ G(Av)

A Posteriori Error Analysis for Idealizations in Linear Problems

Then u minimizes the functional J ( v , Av) in V. For the Gsteaux derivatives of G, we have

Thus, from (2.21), we can find

We compute the conjugate functions. Define a constraint set

The constraint condition is a weak form of

We then have

and

Thus, J*(A*q*,-a*) =

{

+

1

(ah q 2 an 2a +m otherwise.

+g -)

ifq* E

Q:,

Hence, applying Theorem 2.40, we have the following error estimate.

THEOREM 3.16 Forsolutionsof (3.84)-(3.85) and (3.86)-(3.87), there holds the error estimate:

106

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

for any q* E Q*.

In particular, by the weak formulation of the boundary value problem (3.86)(3.87), we see that q* = ( - V u o , g ) is admissible. Then we have the error estimate

Using the smallness assumption (3.82), we have

for any solution uo (which is unique up to a constant) of the problem (3.86)(3.87).

3.4.

DOMAIN IDEALIZATIONS

Domain idealization can be a source of many difficulties in applications. As an example, in linearized elasticity, domains with reentrant comers lead to infinite stresses and strains, violating basic assumptions in deriving the differential equations in the theory of linearized elasticity. Moreover, except for cracks, it is impossible to fabricate exactly a comer. Hence, a comer has to be understood as a mathematical idealization with possibly non-physical consequences for some data of interest. Also, the domain often comes from a process of optimal design. It is important to have an approach for quantitative estimation of the effect of domain idealizations, especially in the neighborhood of comers, edges, etc. In this section, we consider two types of domain idealizations, for one of which, the problem domain is included in the idealized domain, whereas for the other, the problem domain includes the idealized domain.

3.4.1

DOMAIN IDEALIZATION (I) Let R, c R be a subset of R around a comer 0. Assume f

E L ~ ( Rand ) g E H 1 ( R ) are given. We apply the duality theory to give an a posteriori error estimate between the solutions u, E H1(R\R,) and uo E H 1 ( R )of the boundary value problems

and

A Posteriori Error Analysis for Idealizations in Linear Problems

We set

v = H~(R\R,),

V* = ( H I (R\R,))*, Q = Q* = ( L 2 ( f l \ W 2 , A = V : the gradient operator, J ( v , Av) = F ( v ) G(Av),

+

where

f v dx if v E V, v = g on aR\dR,, otherwise,

Then u, E V is the unique solution of the minimization problem inf J ( v , Av).

vEV

From Example 2.4 1, we have

Let us compute the conjugate functions. Introduce the constraint set

The constraint condition is a weak form of the relations divq* = f in R\R,, Then *

*

* =

{

"a; +oo

q * . v = 0 on dR,\dR.

( q * . V g + f g ) d x i f q * E Qz, otherwise

A POSTERIORI ERROR ANALYSlS VIA DUALITY THEORY

108

Thus, we obtain the following general error estimate

for any q* E QE. Now we discuss how to choose a suitable q* in (3.88). We will use the same notations for Ro,,, , Rr,, Fr1, vl.and so on as in Subsection 3.1.2. We assume there exist two positive numbers rz < r l , such that R, c

Define

where 6 > 0 is a constant to be suitably chosen. Then we take

It can be verified that the above q* is an admissible function, i.e., q* E QE. Using this selection q * , the estimate (3.88) becomes

3.4.2

DOMAIN IDEALIZATION (It)

Let R c Ktd be a Lipschitz domain, and 0,a small-size Lipschitz domain neighboring R (Figure 3.3). Let RE be the combined domain such that EE = RU Assume dR = U with n r2 = 0, d o E = U with

n,.

rl n r 2 , =~ 0.

A Posteriori Error Analysis for Idealizations in Linear Problems

Figure 3.3. Domain idealization

Let uo E H 1( R ) ,u, E H 1 ( R E )solve

and

We make the following assumptions:

and

A POSTERIORI ERROR ANALYSIS VIA DUALlTY THEORY

110

Let us derive an estimate for the difference between u,and uo.Take

Let J ( v , Av) = F ( v )

+ G(Av),

where F(v) =

{ '.Efvdx-12,E g2q2

+m

if ?I E a f H;, (OE), otherwise,

Then u,minimizes J(v, Av) among all functions v E gl For the functional G, we have

+ H;,

(RE).

Let Go E H1(RE) be an extension of uo to QE. Then from (2.21) for the we have quantity D ( u ,Go),

Now we compute the conjugate functions. Define a constraint set

The constraint is a weak form of the relations divq* = f in OE,

-q* . Y = 92 on r 2 , ~ .

Then we have F*(A*q*)=

(4' . Vgl

+ f g i ) dx +

/

~ 2 ~ 1if dq*~E

Qf,

r2,~

otherwise

A Posteriori Error Analysis for Idealizations in Linear Problems

and

Thus,

/q*I2+ q f ' V g l J * ( A * q * ,- q * ) =

+cc

+ f gl

if q* E Q z , otherwise.

Using Theorem 2.40, we obtain the following estimate 1 2 -I W - C0)/IL2(ilE) 2

for all q* E Q z . Recall the assumption (3.95). We can choose q* =

{ ,vuo

in,, in R,.

Noting the relation

we then have from (3.96) the following estimate

I/~(U

-CO)IIL~(~ I ~llvfioll~2(n~)~ E)

E X A M P L E3.17 Assume R is a corner domain with an internal angle w < 2 7r at the corner 0.Assume uo = X r f f cos a 0 is the solution of the problem (3.89)-(3.91) with homogeneous Neumann boundary condition around 0, where a = 7rlw. Assume R,

c { ( r ,0 ) : w < 0 < 27r).

Then, a simple choice of Go in R, is

1 uo = -r a ( 2 0 - 27r - w ) . 277 - w For this choice of Go, the estimate is 1 l l -~ C O I L ~ ( ~<-, )2a IIvr"(20 - 2~ - w ) l l ~ 2 ( n ~ ) .

rn

112

3.5.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

ERROR ESTIMATES FOR MATERIAL IDEALIZATION OF TORSION PROBLEMS

This section and the next show applications of the a posteriori error estimates and the techniques used in deriving them, in the context of two concrete problems: torsion problems in linearized elasticity and heat conduction problems.

3.5.1

DESCRIPTION OF THE TORSION PROBLEM

Let 0 c TR2 be a simply connected planar domain, whose diameter is small compared to a positive number 1. We consider a prismatic b a r n x [ O , l ] . Assume one end 2 x ( 0 ) of the bar is fixed in the 0XlX2 plane, the other end x { I ) is subject to a couple whose moment M33is along the OX3 axis. The OX3 axis passes through the center of twist of each section, i.e., through the point about which each cross section will rotate. By Saint-Venant's assumptions, for any cross section at a distance 2 3 from the origin, the displacement components are

n

where $ is the warping function, a is the twist per unit length, i.e., the relative angular displacement of two cross sections a unit distance apart. With the displacement (3.97), the components of the strain tensor are

and the stress components are

where G is the shear modulus. In reality, the value of G depends on the position of the point in the bar. In most elasticity theory books, however, G is assumed to be a constant. This is partly due to the difficulty associated with finding the function G. Since the function G is usually close to some constant for a commonly seen material, it is taken for granted that the torsion problem has been described accurate enough if the function G is replaced by a suitable constant Go. In this section we derive a posteriori quantitative error estimates

A Posteriori Error Analysis for Idealizations in Linear Problems

113

for effects of the idealization on solutions of torsion problems. The essential feature of such quantitative estimates is as follows: Assuming we can only solve the idealized torsion problem with a constant shear modulus Go and we know certain upper bound for the difference ( G - Go), we then should be able to estimate the error of the solution of the torsion problem caused by the idealization. The necessity for having an a posteriori error estimate also comes from the fact that by doing physical experiments, it is possible only to obtain an approximate shear modulus. The derivation of equations for the idealized torsion problem with a constant shear modulus Go can be found in any standard book on elasticity theory (e.g., [I101 or [154]). For the sake of completeness, we derive, along the same line, equations for the torsion problem, for which the shear modulus is a function G(x1, ~ 2 ) . The equilibrium equations without body force

reduce to a single equation

Hence, there exists a function 4 such that

The function 4 is usually called the Prandtl's stress function. Solving the system for V$ in terms of V 4 , we get

Hence we find the equation for 4:

For the boundary conditions, we first note that the lateral side of the bar is free. Let s be a length parameter on d o . Since the unit outward normal on the

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

114

lateral side is (dx2/ds, -dxlIds,o ) ~we , have

Thus 4 is a constant on dR,this constant being taken to be 0, On the boundary surfaces 2 x (0) and 2 x (11, the distribution of the stress must have no resultant forces and be equivalent to the torsional couple. The resultant force in the OX1direction is

where the boundary condition (3.99) is used. Similarly, the resultant force in the OX2direction is

The resultant torsional moment on the end of the bar due to the assumed stress distribution must equal to

4

Now we introduce a normalized Prandtl's stress function by the relation Then from (3.98) and (3.99), we see that is the solution of the following elliptic boundary value problem with variable coefficient:

4

= 2a

4.

4

A Posteriori Error Analysis for Idealizations in Linear Problems

115

From (3.100), once we have the function 4 , the unit length twist a can be obtained through

Thus, the quantity

is of paramount importance for the torsion problem. As mentioned earlier, practically it is impossible to have the exact shear modulus function G. Instead, usually a suitable constant Go is used as the shear modulus of the bar. For the corresponding idealized torsion problem, we then need to solve the following constant coefficient elliptic boundary value problem:

The relation between

M33 and

a is

where

Jo =

40d x dx2. ~

(3.106)

The main purpose of this section is to derive quantitative error estimates for

(3- $ 0 ) and ( J - J o ) ,by using the solution $0 only. Note that from an estimate for (4 - 40), it is straightforward to have an estimate for the error in warping functions. In the next subsection, we apply the results from previous sections to derive a posteriori upper bounds on ( 4 - $ 0 ) and ( J - Jo) for the torsion problem and its idealization.

3.5.2

ERROR ESTIMATES FOR MATERIAL IDEALIZATION

We recall the following result from Subsections 3.1.2. Let there be given

We denote the trace of g on dR by the same letter g . Let A ( x ) = ( a i j ( x ) )E Itdxd be a matrix-valued function that is measurable, symmetric, uniformly

116

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

positive definite and uniformly bounded from the above, i.e.

for some constants co, cl with 0 < co 5 cl < oo. Under the above assumptions, the elliptic boundary value problem

admits a unique weak solution u E H' ( 0 ) . An idealization of the above boundary value problem is

We assume that the matrix A ( O ) (I) = (a!;) ( 2 ) ) E ad is also measurable, uniformly bounded and positive definite. We have the following inequality for an error bound on u - u(O)in terms of u(O)from (3.22):

We apply the above result to derive a posteriori error estimates for material idealization of torsion problems. More precisely, we use the solution $0 of the idealized problem (3.104)-(3.105) to give error bounds for 11 v ( $ 0 - $) 11 Ll ( a ) and I Jo - J J ,where, $ is the solution of the problem (3.101)-(3.102), Jo, J are defined by (3.106) and (3.103). Thus, assume the shear modulus G is close to a constant Go in the sense

for some small number E posteriori error estimates.

>

0. We have the following two theorems for a

THEOREM 3.18 For the solution $0 of the problem (3.104)-(3.105) and the solution $ of the problem (3.101)-(3.102), we have

A Posteriori Error Analysis for Idealizations in Linear Problems

117

Proof. Let us rewrite problems (3.101)-(3.102) and (3.104)-(3.105) in the following forms:

and

-

-A#, = G~ &=O

in

R,

ondfl.

Now we apply (3.107) for the solutions of the problems (3.110)-(3.111) and (3.112)-(3.113) by taking

where I is the identity matrix. We have

By using the assumption (3.108), we then get the estimate

i.e., (3.109) holds.

H

THEOREM 3.19 For the quantities J and Jo defined in (3.103) and (3.106), we have E (3.114) I J - Jol I 1 - G o t /lv4ollLcn).

Proof. Let u be the solution of

From Theorem 1.20, we have

118

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

and jOu1I L 2 ( n ) is the best possible constant in the above inequality. Obviously, u = $ o / ~ o . Noticing that $0 - E H;( a ) , we have

4

where the estimate (3.109) is used. If instead of (3.108), we assume JG(x)-Go/5S, x E R for some small number S > 0. Then by taking

and applying (3.107) to the problems (3.10 I)-@. 102) and (3.104)-(3.105) directly, we obtain [G(x) - Go12

/v$o12dx.

Hence,

Correspondingly,

Therefore, we have proved the next result.

THEOREM 3.20 Under the assumption (3.117), we have the aposteriori error estimates (3.118) and (3.119).

A Posteriori Error Analysis for Idealizations in Linear Problems

119

Let us examine two concrete examples. When 0 is a disk of radius r , from Example 1.21 we know that for the solution u of (3.115)-(3.116),

Noticing that

$0 =

Gou, we get, from (3.109) and (3.1l4), the following error

and

IJ-

Jol I

n Gg r 2 8 (1 - Gar) "

Assume now 2 is a rectangle [0,a] x [0,b]. From Example 1.22 we know ~ ~ )to that I ~ V U I Iis~equal

Therefore, we have the error estimates

and

We comment that for the estimates presented in this subsection, we have assumed (3.108) and (3.117), i.e., G ( x ) is close to Go throughout the entire domain R. We can allow the situation where in a small region, G(x) is not close to Go, and in this case, (3.107) will still provide a small error bound. On the other hand, if G ( x ) varies greatly in a large region, such as in a composite material, then for the idealized problem, the shear modulus Go( x ) should be a function of the location x itself. The inequality (3.107) will still lead to meaningful error bounds.

3.6.

SIMPLIFICATIONS IN SOME HEAT CONDUCTION PROBLEMS

The purpose of the section is to give quantitative justification of some simplifications in steady state linear heat conduction problems. For convenience, here we recall the derivation of a steady state linear heat conduction problem for a stationary, homogeneous, isotropic solid R.For detail, see e.g. [124]. Let

120

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

T = T ( x ) be the temperature, q = q ( x ) the heat-flux vector, x E Fourier law of heat conduction states:

R. The

where k is the thermal conductivity of the material of the solid. Assume there is a heat source within the solid, and denote f (x) the heat generated per unit volume per unit time. Then the energy balance on an arbitrary small-volume element results in the equation divq(x) = f (x).

(3.121)

Combining (3.120) with (3.121), we then obtain the differential equation for the temperature variable T:

Newton's law of cooling provides a realistic boundary condition. It states that the heat energy flowing out per unit time per unit surface area is proportional to the difference between the temperature on the surface T and the temperature outside the surface g, i.e.,

with a proportionality function h > 0, which is usually taken to be a constant. A complete mathematical formulation of the heat conduction problem now consists of the differential equation (3.122) and the boundary value condition (3.123). If the boundary (or part of the boundary) is insulated, i.e., there is no heat flow across that portion of the boundary, then the boundary condition there is

which may be viewed as a limiting case of (3.123) for h + 0. In this case, in order to maintain the energy balance, we need to assume that the total heat generated within the solid to be zero:

f ( x ) dx = 0. For certain problems, the temperature on the boundary (or part of the boundary) is assumed to be fixed, T = g, (3.126) which may be viewed as a limiting case of (3.123) for h + co.

A Posteriori Error Analysis for Idealizations in Linear Problems

121

We assume R is an open bounded Lipschitz domain in IEd, d = 2 , 3 in most applications. When the thermal conductivity k is a function of x only, we and infZEnk(x)> 0. For the heat source function f , we assume k E Lm(R), . boundary function g is assumed to be the trace of an assume f E ~ ~ ( f l )The H ~ ( Rfunction ) on the boundary. The proportionality function h in (3.123) is an Lw (R)function. Our results are derived and presented for isotropic solids [i.e., k in (3.120) is a scalar] only. The generalization of the results to anisotropic solids is straightforward, as long as the associated thermal conductivity matrix is symmetric, positive definite, and bounded. Also, for simplicity in writing, we assume only one kind of boundary condition is specified on the whole boundary.

3.6.1

COEFFICIENT IDEALIZATION

Usually, one assumes that the thermal conductivity k(x)= ko is a constant, whereas in reality, k(x)is only close to the constant ko:

We assume that ko > E . Let us check how much error is introduced in replacing k(x)by ko.More preciously, we want to bound the difference between T and To, the solutions of the following problems:

and

Let us apply the result from Section 3.1 to estimate (T- To) by To. To do this, we rewrite the problems (3.128)-(3.129) and (3.130)-(3.13 1) as

T=g

and

ondR

122

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

+

We notice that A = (1 elko) I, and le(x) Ilko 5 &/kg. Thus, by using (3.107) and (3.127), we find

Once the quantity / / VTo j l L 2 ( n ) is computed, one obtains an upper bound for the error. Notice that the estimate is optimal.

3.6.2

APPROXIMATE INSULATION BOUNDARY CONDITION

We discuss the idealization of the boundary condition (3.124) for an "insulated" boundary. Practically, one does not have a pure Neumann boundary condition (3.124), instead, the boundary condition is of the form (cf. (3.123))

where the function h is small. Let us examine the error caused by idealizing an "insulated" boundary. We compare T and To, solutions of the problems -div(k V T ) = f

in R,

(3.133)

and

Here we need to assume (3.125) to ensure the boundary value problem (3.135)(3.136) has solutions. Indeed, the problem has infinitely many solutions and they are of the form To c, To being any particular solution of the problem, and c being an arbitrary constant. The estimates derived in Section 3.3 are for the Poisson equation. The same derivation technique can be used here to bound the error (T - To).We apply Theorem 2.40 and first cast the problem (3.133)-(3.134) in the form of (2.5). We set

+

123

A Posteriori Error Analysis for Idealizations in Linear Problems

For q E Q, we write q = ( q l , qz), with q1 E ( ~ ~ ( 0and ) q2 ) ~E L 2 ( a 0 ) . We use a similar notation for q * E Q*. Define J(v, q ) = F ( v ) + G ( q ) ,

(3.137)

where F(v) = -

f v dx,

Then, the weak solution of the problem (3.133)-(3.134) is the minimizer of the functional J ( v , Av) on V:

T E V,

J(T,AT) = inf J ( v , A v ) . vEV

Now consider the difference D ( T , To)given by (2.21). We have 0

for p, q E Q. Therefore,

From the definition of the conjugate function, we can find lq;I2dx

J * ( h * q * ,-q*) =

+oo

+

/

1

- (hg -

as2 2h

q;)2ds, if q * E

otherwise,

where the constraint set is r-

By Theorem 2.40, for an upper bound of D ( T , To),we have

&:,

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

124

Thus, we get the following error estimate:

for any q* E QE. In particular, if we take qT = k VTo,q; = 0,which belongs to the constraint set QE by the problem defining To,then from (3.138) we obtain

If we assume

Ihl 5

E

in

R,

then from (3.139), we get

for any solution To (which is unique up to a constant) of (3.135)-(3.136).

3.6.3

APPROXIMATE TEMPERATURE BOUNDARY CONDITION

In this subsection, we consider the Dirichlet boundary condition. Instead of (3.126), a more realistic boundary condition should be (cf. (3.123)):

where h is large. Let us compare T with To,the solutions of the problems

and

We have seen in the last subsection that the problem (3.140)-(3.141) corresponds to the minimization problem

T E V, J(T,AT) = inf J(v,Av), vEV

A Posteriori Error Analysis for Idealizations in Linear Problems

125

where, J ( v , A v ) is defined in (3.137). We already have the general error estimate (3.138) for (T - T o ) In particular, assuming d T o / d u E L2(dC2),then the choice dTo q ; = k V T o , q$=k-du is admissible, i.e. ( q ; , q;) E QF, and we get the following simple a posteriori error estimate:

If we assume

1 h ( x )t ; ,

then the error estimate (3.142) reduces to

Chapter 4

A POSTERIORI ERROR ANALYSIS FOR LINEARIZATIONS

Linearization technique is used frequently in modeling physical phenomena and in numerical computations. In this chapter, we derive a posteriori error estimates for the effect of linearization on solutions of nonlinear physical problems. A posteriori error estimates for the linearization technique used in numerical approximations will be derived in the Chapter 5.

4.1.

LINEARIZATION OF A NONLINEAR BOUNDARY VALUE PROBLEM

Detailed quantitative error analysis is given for linearizations of certain nonlinear elliptic problems, whose linearizations are boundary value problems of Poisson's equations. An example of a nonlinear problem considered in this section is the following:

When R c IK2 is a planar Lipschitz domain, for suitable data a,f and g, the problem (4.1)-(4.2) describes a nonlinear torsion problem (cf. [109]). In particular, when the coefficient function a(() = 1, the problem (4.1)-(4.2) reduces to

which represents a linear torsion problem (cf. Section 3.5). Owing to the difficulty associated with determining the material property function a((),in most elasticity theory books, it is taken for granted that the torsion problem of a

128

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

real material has been described accurately enough by the linearized problem (4.3)-(4.4), since the coefficient function a(<)is usually close to some constant (taken as 1 here) for many commonly seen materials. In this section, we give a posteriori error estimates for the effect of linearization on solutions of elliptic problems, using solutions of the linearized problems only. In particular, we provide error bounds on (u - uo) for the solutions of (4. I)-(&!) and (4.3)-(4.4) in terms of the solution uo of the linearized problem (4.3)-(4.4). In practice, numerical methods are used to solve problems (4.1)-(4.2) and (4.3)-(4.4). Like the continuous problems, it is usually not easy to solve the problem (4.1)-(4.2) numerically, and we may first solve the simpler problem (4.3)-(4.4) numerically and check if the numerical solution of (4.3)-(4.4) is close to, and can be used as a good approximation of the solution of (4.1)-(4.2). Moreover, nonlinear problems are usually solved by some iterative procedures, and solutions of the linearized problems can be used in the initialization step of an iterative procedure. For finite element solutions of problems (4.1)-(4.2) and (4.3)-(4.4), we will provide a discrete version of a posteriori estimates for the error caused by the linearization, i.e., some computable error bounds for the difference between a finite element solution of the problem (4.1)-(4.2) and a finite element solution of the problem (4.3)-(4.4), using the finite element solution of the problem (4.3)-(4.4) only.

4.1.1

THE NONLINEAR PROBLEM AND ITS LINEARIZATION

In this subsection, we derive a posteriori error estimates for linearization of nonlinear problems of the form (4.1)-(4.2). We state our results for two dimensional problems only. It is straightforward to derive similar results for any finite dimensional problems. We consider a more general mixed boundary value problem for the partial with rl relatively closed, r2 differential equation (4.1). Let d!2 = rl U relatively open, and rl n r2= 0. We assume meas (rl)# 0. Let

f E~~(01 gl ,E H1(W

g2 E L2(F2). We will use the same notation gl to denote the trace of gl E H1(!2) on the boundary. The model nonlinear elliptic boundary value problem we study is

129

A Posteriori Error Analysis for Linearizations

To have a class of problems for which we can make specific computations, we introduce the following scalar function as the coefficient function

= for some to Denote

{+

ift I to, i f t > to,

(1 - P

> 0, p E ( 0 , l ) .

The weak formulation of the boundary value problem is: Find u E H;l such that

,gl

(0)

We then introduce the corresponding minimization problem. Let

v = H ~ ( R ) ,v * =( H ~ ( R ) ) * , d

Q = Q* = ( ~ ~ ( ,0 ) ) Av = Vv. We define the energy function on V: J ( v , Av) = F ( v )

+ G(Av),

where F(v) = otherwise,

Then the weak formulation (4.9) is equivalent to the minimization problem inf J ( v , Av).

vEV

We can apply Theorem 2.37 to show that the minization problem (4. lo), and hence the weak formulation (4.9), has a unique solution u. We call Q(u) = {x E 0 : [ V U ( X ) ~to)

<

130

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

the linear region of the nonlinear solution u, and

the nonlinear region of u. The sets R2( u )and Rn(u)are the regions where the differential operator of the partial differential equation is linear and nonlinear, respectively. Since the solution depends continuously on the input data, when f , gl and gz are small, the nonlinear region R, ( u )is small. Note that on Q ( u ) , a ( l V u J 2=) 1. Hence, when f , g l and g2 are small, we expect that the solution uo of the following linear problem is a good approximation of u:

However, if R is a plane corner domain, such as the one shown in Figure 1.6 (page 26), then, no matter how smooth f , g l and g2 are, the solution uo of the linearized problem (4.11)-(4.13) is singular at the comer points; see Section 1.6 for a brief review of this topic. Consider the comer 0 (taken as the origin of the coordinate system we are working on) with an internal angle w > T. Let a = T/W. When on both sides of the part of the boundary around 0,either a Dirichlet or a Neumann boundary condition is specified, we have

lVuo( x )1

N

O(ra-I) for r = Ix / close to 0.

When on one side a Dirichlet condition is specified, while on the other side a Neumann condition is specified, we have

IVuo( a )1

N

O ( T " / ~ - ~for ) r = Ix I close to 0.

Since a < 1, l V u o ( x ) /-+ oo as x -+ 0. Hence, it is doubtful whether the linerized problem (4.11)-(4.13) is in any sense close to the nonlinear problem (4.5)-(4.7). We will derive an estimate for the difference between u and uo. Why the coefficient function a ( ( ) is chosen in the form (4.8)? The onedimensional stress-strain relation corresponding to the nonlinear problem (4.5)(4.7) is a = ~ ( 1 ~ E1 . ~ For ) scalars a and E , we require a stress-strain relation as illustrated in Figure 4.1, where we omit the part of the graph for negative a and E, which can be obtained by reflecting the graph for positive a and E with respect to the origin. The coefficient function a ( < )corresponding to Figure 4.1 is of the form (4.8) with ,B = E T / E , to= E $ . Now suppose we have computed the solution uo of the linearized problem (4.11)-(4.13). We are interested in bounding the error ( u - uo)by using uo.

A Posteriori Error Analysis for Linearizations

Figure 4.1. One-dimensional stress-strain relation

Let us first compute the conjugate functions. F*(A*q*)= sup [(Av,q * ) - F ( v ) ] vEV

=

sup

[S,(q*."v+fv)dx+

V E H ~(a) ~

L r

dsI

Introduce the constraint set

The constraint condition is a weak form of the relations divq* = f i n R ,

-q*v=gzonr2.

Then F*(A*q*)=

(Q* . Vgi

+f

gi) dx

+

gzgl ds if q* E

QE,

otherwise.

132

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

We also have

=

/

f?

=

q -

sup

,ad

y

I

*t -

1lq"

l2$0

:a(() dt] dx

dt] dx

Here, t = b(/q*1 ) is the solution of the equation

We can determine that

In computing G*( - q * ) , we used Theorem 2.35, also we assumed q* Obviously, the expression for G*( -q*) holds for q* = 0 too. Thus,

# 0.

A Posteriori Error Analysis for Linearizations

133

Now we consider the quantity D ( u , uo) defined in (2.12). We use Theorem 2.40 to derive an upper bound:

Hence,

for all q* E QE. Note that we also have the bound for the energy difference E D ( u , uo) of (2.11) between u and uo:

for all q* E QF. Then we turn to derive a lower bound for D ( u , uO).From the definition of the functional G, we can find

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

134 Thus, by (2.20), 1

=I

( ~ ' ( h+u t A(uo - u ) ) - G 1 ( h u )A(uo , - u ) )dt

-

a ( V u 2 )V U ] . V ( u o- u ) dxdt

- V (uo- u ) dr dx dt.

+

<

If ( V u t T V (uO- U ) I 6 , then the integrand is t I V (uo - u )1 2 . When IVU t r V ( u O- U ) 1 > 6 , using the definition of the coefficient function a (.) , we see that the integrand is not less than P t I V (uo- u )1 2 . Now denote

+

for the linear and nonlinear regions of uo and of both u and uo. Then on f l l ( u ,uo), jVu+t.rV(uo - u ) l <

A.

Thus,

Combining (4.14) with (4.16), we obtain the following estimate:

for any q* E Qz. In the next subsection, we discuss a proper selection of the auxiliary function q* E QE so that the right-hand side of (4.17) provides a good error bound.

135

A Posteriori Error Analysis for Linearizations

4.1.2

SELECTION OF THE AUXILIARY FUNCTION

We state two kinds of selections for q* in the estimates (4.15) and (4.17), similar to Subsection 3.1.2 where effects of coefficient idealizations are discussed in the context of linear boundary value problems. SELECTION1. We simply take

It is an admissible function. Then we have the estimates

and

For ,8 not close to 0, numerical experiments given in Subsection 4.1.3 will show that the selection (4.18) provides a good error bound. However, if P is close to 0, then the factor 1//3in (4.19), (4.20) leads to useless error bounds. To overcome this difficulty, we introduce another selection. SELECTION2. We describe this selection for a problem on a comer domain as shown in the Figure 4.2. First consider the case where the boundary condition around the comer is of the Dirichlet type. We assume there is an ro > 0 such that in the region Ro = R n {r > ro), JVuol5 where w is the internal angle of the comer. Denote ro = fl n {r = ro), vo the unit normal vector on roin the direction of being away from 0. We further assume

do/ d m

f (x)= 0, x We define a function in R\Ro:

E

R\flo.

,

(4.21)

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 4.2. Neighborhood around a comer

Then we take

in Ro, in R\Ro.

This function is admissible, since for any v E - i q * .

Vvdx =

lo

V u o. V v d x

HA ( R ) ,we have

+ l , n o ( V y ,- V z ) V v d x

The selected function q* has the property that Iq* (x)1 5 do. This inequality is obvious for x E Ro. For x E R\Ro, we use the relation

to derive the bound

A Posteriori Error Analysis for Linearizations

Then we have the estimates -a(E) d<

+ (q* + Vuo) . Vgl

and

When the boundary condition around the corner is of the Neumann type, we assume there exists an ro > 0 such that in Ro = R n { r > r o ) , IVuoJ5 & / d m . Again, denote ro = R n { r = ro). Besides the assumption (4.21), we also assume that g2 vanishes around 0 , i.e.,

where, instead of (4.22),

From the linear problem for uo and the assumptions made on f and g2, we have the equality

< 6.

Hence, q* is an admissible function. Once more, /q*(x)1 Therefore, the estimates (4.24) and (4.25) hold. The situation where the boundary condition around the comer changes its type can be treated similarly.

4.1.3

NUMERICAL EXPERIMENTS

We present two numerical examples to show the effectiveness of the estimates (4.19), (4.20), (4.24) and (4.25). We will compare the estimates resulted from

138

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

the two different selections of the auxiliary function q*. For the first example, we know the exact solutions for both the nonlinear problem and the linear problem. Hence, we will see how effective our estimates are by comparing our error bounds with exact errors.

EXAMPLE 4 . 1 We take R to be a unit disk excluding a small hole

where r, > 0 is a small number. Let us consider a family of nonlinear problems with a small parameter X > 0: in R, for r = 1, for r = r*,

-div [ a ( V u 2 ) ~ u=] 0 u = X ln(l/r,) u =0

where a(.)is the function defined in (4.8). The true solution is d o r llnr

+ X In-r*1

forrl

I

r r1h--(l-P)(r-r,) r*

< r < 1,

forr* < r < r l ,

where rl is the unique positive solution of the equation

with ro = A/& The linearized problem

-nuo= 0

in R, for r = 1, for r = r*,

uo = X ln(l/r,) uo = 0

has the solution It is easily verified that rl

< ro. Therefore,

Thus, in some sense, the linearized problem is more "rigid" than the nonlinear problem.

A Posteriori Error Analysis for Linearizations

139

We take to = 1, r, = 0.01. For various values of /? and A, we have the following numerical results. Here

is the square root of the difference in energy,

is the error in an energy-like norm. We have

For Selection 1, q* = -Vuo,

we obtain an upper bound for E J ( u ,u o )and E ( u ,u o ) :

For Selection 2,

where 7-2 = rod-, E ( u ,uo):

we have another upper bound for E j ( u ,u o ) and

140

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 4.1. Example 4.1, true errors and error bounds, X = 0.1

Table 4.2. Example 4.1, true errors and error bounds, X = 0.015

Table 4.3. Example 4.1, true errors and error bounds, X = 0.0101

to be the effectivity index of our error estimates. Numerical results for various values of the parameters X and P are reported in Tables 4.1, 4.2, and 4.3. We see that min{Eestl, Eest2) is a very good error bound for any values of ,B and A. For ,B away from 0, Eestlprovides a better bound, whereas when ,f?is as our error bound. close to 0, we should use Eest2

EXAMPLE 4.2 Let R be a circular comer domain

A Posteriori Error Analysis for Linearizations

141

We use again the notation cu = T / W . Consider a particular case when the linear solution is (4.30) u O ( x )= X r f fsin ae, where X

> 0 is a parameter. Since IVuo I = X a r f f - I ,the linear region of uo is

where, the radius Xa

1/(1-~)

r0 =

Let us estimate the difference between uo and u, the solution of the nonlinear problem

We try to give an upper bound for both EJ ( u ,u o )and E ( u ,u o ) ,defined in Example 4.1. Denote

If we use Selection 1, q* = -Vuo,

then an upper bound for the error ( u - uo)is

If we use Selection 2,

where

and

V ( r ,6') = -ar?-lr cos cue in R1,

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 4.4. Example 4.2, error bounds for problems on an L-shape domain

then an upper bound for the error is

.

We take to= 1, and w = 3 ~ / 2(i.e., Q is an L-shape domain), the numerical results are given in Table 4.4. We see that indeed, EeStlprovides a better bound for P away from 0,whereas Eestzis better than Eestlwhen /3 is close to 0. We end the subsection with several remarks.

REMARK4 . 3 We have seen in the numerical examples above that our a posteriori error bounds for the continuous problems are effective. When P is not close to 0, the error bound from the simple Selection 1 is better than that from the Selection 2. On the other hand, when P is close to 0, the error bound from the Selection 2 is better than that from the Selection 1. By continuity, at least when finite element solutions are sufficient close to the exact solutions, we can claim that our a posteriori error bounds for the discrete problems are effective too. When P is not close to 0, the error bound (4.20) is better than (4.25), whereas when ,B is close to 0, the error bound (4.25) is better than (4.20).

.

REMARK4.4 In the above discussions, we assumed a concrete form for the nonlinear problem. This is only for the purpose of making specific computations and of showing the effectiveness of the error estimates. In fact, we do not depend on a concrete form for the nonlinear problem. To be able to provide error estimates for the effects of mathematical idealizations on solutions, we

A Posteriori Error Analysis for Linearizations

143

only need a description on the structure and range of the coefficient function. This is a crucial point which allows the useful applications of the suggested a posteriori error estimates on practical problems, where all the data contain certain degree of uncertainty. Of course, the more accurate the description of the nonlinear problem, the more accurate the error estimates. REMARK4.5 In this section, we consider the effect of the linearization. The mathematical idealizations of a practical problem can be more complicated than the linearization, and the general framework for deriving a posteriori error H estimates discussed earlier allows us to get similar estimates. REMARK4.6 The sample problem discussed above is an elliptic boundary value problem for a Laplacian-like differential operator. The idea for deriving a posteriori error estimates works for more complicated problems in applications. 1 One such example is given in the next section.

4.2.

LINEARIZATION OF A NONLINEAR ELASTICITY PROBLEM

The classical theory on linearized elasticity is based on the assumption that the size of the displacement gradient tensor is small. The linear Hooke's law in the linearized elasticity theory characterizes a large class of elastic materials occupying smooth domains subject to moderate size body forces and boundary tractions. For engineering applications, however, the ideal assumption is not always satisfied. For instance, consider an elasticity problem on a domain with reentrant comers. Domains with comers and edges are common in engineering applications. The solution of the linearized elasticity system exhibits singularities; in a neighborhood of the comers, the strain and stress tensors computed from the solution become arbitrarily large in sizes. Thus, the basic assumption of the linearized elasticity theory is grossly violated around the comers. On the other hand, the linearized elasticity model is used for many applications, as long as it can be reasonably expected that the violation of the basic assumption is limited to a small region of the domain. Qualitatively, when the input data are small, the expectation is validated by the continuous dependence of the deformation of an elastic material on input data. Obviously, it is more useful to have quantitative information on the error caused by using the solution of the linear problem to describe the response of the elastic material on a comer domain. The purpose of the section is to provide quantitative error analysis for the effect of linearization on solutions of nonlinear Hencky materials. We notice that in practice there are material models more realistic than the nonlinear Hencky materials considered here. However, we may view the analysis given in this section as a preliminary step towards a general a posteriori error analysis

144

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

for linearization of more realistic materials; mathematical theory of nonlinear elasticity can be found in [42, 106, 1221. A typical result has the feature that we solve the linearized elasticity problem and use its solution uo to bound the error ( u - u o ) , with u the solution of a nonlinear Hencky material problem. We emphasize here that to obtain an error bound, all we need to know is the solution of the linearized elasticity problem.

4.2.1

NONLINEAR HENCKY MATERIALS

For a detailed description of nonlinear Hencky materials, one is referred to [ I l l ] or [160]. Let R c EXd be a d-dimensional domain. For a displacement field v : R + EXd, we use the notation v = (vl , . . . ,vd)Tand introduce the infinitesimal strain tensor E(V): R + S d with components

Denote by a the stress tensor with components the stress and strain tensors are defined by

aij,

1

< i, j

5 d. Deviators of

where t r ( r ) = q i , t r ( ~ = ) ~ i i Recall . that for the dot product of two tensors = (E . E ) " ~ . ~ , EqS d, E . q = E i j q i j Also, For a linearly isotropic, homogeneous material, the stored energy function is 1 Mo(E)= I X ltr(E)I2 p

+

or written as

1 Mo(E)= 2 k tr(c)12

+ p 1cD12,

+

where k = X 2 p l d is the bulk modulus, X and p are the Lam6 coefficients. A nonlinear Hencky material is characterized by a stored energy function of the form 1 (4.31) M ( E ) = 5 k ltr(c)I2 p 4 ( / E D 1 2 )

+

for a non-negative, continuously differentiable function

a then The constitutive law a = d M ( ~ ) l d is

4 satisfying

A Posteriori Error Analysis for Linearizations

and we have uD = 2 p

$'(IE~/~)

E

~

.

For the special case when 4 ( [ ) = 5, we obtain a linearly elastic material, and the constitutive law becomes

and we have uD = 2 p s D .

Now we assume a nonlinear Hencky material, initially occupying the domain

R,is subject to the action of a body force of density f , a traction g2 on part of the boundary r2c dR.We assume on another part of the boundary rl c do,the displacement field gl is specified. We may specify other boundary conditions as well, e.g., normal or tangential components of tractions or displacements. One can easily extend the a posteriori error estimates to be presented below for the cases of other boundary conditions. Let us assume

and meas(rl)

> 0.

We denote the Sobolev spaces

Then the displacement field u : R + Rid of the nonlinear Hencky material minimizes the total potential energy:

where the total potential energy

We now study the minimization problem (4.35). We need the following result.

LEMMA 4.7If R is a Lipschitz-continuous domain or a crack domain, then E ( v ) + m as lvllv

+ m, v - g1 E Vrl.

Proof. We will use Korn's inequality (page 34) lk(V)12dX 2 co llvll2. a v E v r , .

146

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Usually, Korn's inequality is proved for a Lipschitz-continuous domain ([l 1I]). For a domain with cracks, we decompose R into R1 and R2 in such a way that both R1 and R2 are Lipschitz-continuous, meas(dRl n F l ) > 0 and meas(df12 n r l ) > 0. Then for a v E Vrl, Korn's inequality holds for both vial and via,. Thus, Korn's inequality also holds for v. For v E V ,v - gl E Vr,, using the assumption (4.32) and Korn's inequality (4.36) on v - g l , we have

THEOREM 4.8 Let R be a Lipschitz domain or a crack domain. Assume (4.33) and (4.34). If we further assume 4(J) is second-order continuously differentiable and satisfies, for some a > 0,

then the minimization problem (4.35) has a unique solution u E g l solves the nonlinear elasticity system

+ Vrlthat

Proof. First we prove that the stored energy function M ( E ) defined in (4.31) is strictly convex. The second-order GBteaux derivative of M at E along the direction q is

A Posteriori Error Analysis for Linearizations

Here, we used the assumption (4.32) and the relation

If

4 " ( l ~ ~<10, ~ we ) use the assumption (4.37) to obtain

Thus, M ( E )is strictly convex. Hence the total potential energy function E ( v ) is strictly convex. By the assumptions on the function 4, E ( v ) is continuous. By Lemma 4.7, E ( v )is coercive. Therefore, the minimization problem (4.35) has a unique solution u E V with u - gl E Vrl(cf. Theorem 2.30). It is easy to verify the solution u satisfies the equation (4.38). In practice, because of its simplicity, linearized elasticity theory is adopted wherever the solution of the linearized elasticity system can provide sufficiently accurate information of physical interest. As mentioned at the beginning of this section, if the material occupies a comer domain and if the region where the basic assumption of the linearized elasticity theory is violated is limited to small neighborhoods of comers, then for a large part of the body the governing function 4 is close to the constant 1 (assuming the material is isotropic and homogeneous over that large part of the body). So it is intuitively reasonable to use the linear elasticity theory. The solution uo of the linearized elasticity theory is the minimizer of the following potential energy function

The solution uo can also be equivalently characterized by the relation

148

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

The main purpose of the section is to derive computable bounds in terms of

uo on the error (u - u o ) .

4.2.2

A POSTERIORI ERROR ESTIMATES FOR THE LINEARIZATION

To apply Theorem 2.40, let us take

V = (H'(o))~,V * = (H'(R)*)~, Q = Q* = ~ ~ x( (0 ~)

~ ( n ) ) ~ ~ ~ ,

Av = ( t r s ( v )s, D( v ) ) , J ( v ,A v ) = F ( v ) G ( A v ) ,

+

where

For q E Q, we use the notation q = ( q l ,q 2 ) with ql E L 2 ( 0 ) and q2 E ( L( ~ a )d)X d . We use a similar notation for q* E Q*. Note that the second component of Av is a symmetric tensor. We can also use the following spaces

where dxd ( ~ ~ ( f l-) {r ) = ~ (~r i j~) d x d : T i j =

rji E L 2 ( 0 ) , 1 5 i , j 5 d ) .

We notice that when v - gl E Vr,, J ( v , A v ) = E ( v ) is the total potential energy function. The minimization problem (4.35)can be equivalently written as U E V , J ( u , A u ) = inf J ( v , A v ) . vEV

From the corresponding properties of the potential energy function E ( v ) proved in Subsection 4.2.1, it is easy to verify that J ( v , A v ) satisfies all the assumptions of Theorem 2.40. Therefore, we may apply the inequality in Theorem 2.40 with v = uo, the solution of the linearized elasticity problem (4.41), to obtain an a posteriori bound on the error ( u - u o )as well as the energy difference E D ( u , u o )= E ( u o )- E ( u ) .

A Posteriori Error Analysis for Linearizations

Let us compute the conjugate functions. We have

with

where

and

where 7

2

= -+(q;) solves

Before giving an estimate for the error ( u - u o ) ,we provide a detailed analysis of the equation (4.42). Taking the inner product of both sides of the equation (4.42) with itself, we obtain an equation for the unknown scale

1 = 1r2I2:

, 1

( Q ( E ) ) ~=c 4P

1q;l2

(4.43)

Let f (0 = [4'(()]2 [ . We have

by assumptions (4.32) and (4.37). Thus, f ( t )is a strictly increasing continuous function. Since f ( 0 ) = 0, f ( E ) + m as + oo,we conclude that f is a oneto-one mapping from [0, m) to [0, co).Therefore, (4.43) is uniquely solvable. Having the solution [ of the nonlinear scalar equation (4.43), we obtain the solution of the system (4.42) from

<

150

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Now we apply Theorem 2.40 to the quantity D ( u , u o ) defined by (2.12). For any q; E L~(Q), q; E ( L Z( 0 ) )d x d such that

we have the following upper bound for D ( u , uo):

Taking v = uo - gl in (4.41), we obtain

Using this relation in (4.44), we then have

This inequality holds for any auxiliary functions q* E Qz.In particular, from (4.411, q? = -k t r s ( u o ) , q; = -2p E D ( u O ) (4.45) is an admissible pair of auxiliary functions. With the choice (4.45), we derive the following inequality from (4.44),

A Posteriori Error Analysis for Linearizations

Next, we derive a lower bound for D ( u ,u o ) .Note that

+ 2 4'(lp2I2) ~ lq212]dx. We use the formula (2.21):

where

+ - u)j2) [ ( a D ( u+ ) t c D ( u 0- u ) ). cD(u0- u ) ] + 4' (1cD(u)+ t c D ( u 0- u)I2)c D ( u o- u)l2.

h ( t ,u ,u O = ) 2 4" ( l c D ( u ) t E

~ ( U O

2

by using the assumption (4.32). If

4'' ( c D( u )+ t c D ( u o- u )1 2 ) < 0,then

by using the assumption (4.37). In any case, we have

h ( t ,u ,~

0 2 ) min(a,

P ) IcD(uo- u)I2,

and then

In conclusion, we have proved a posteriori bounds for the error between the displacement field of a nonlinear Hencky material and that of a linearized elastic material.

152

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

THEOREM 4.9 Under the assumptions (4.32), (4.33), (4.34) and (4.37), we have, for the energy difference

and for the error in energy norm

and

4.2.3

NUMERICAL EXAMPLES

To make specific computations, we introduce a family of nonlinear Hencky materials whose governing function $ in the stored energy function M is of the form (cf. Figure 4.3)

A Posteriori Error Analysis for Linearizations

Figure 4.3. The function

4 for numerical examples

with parameters ,B E (0, I ) , 0 # y E ((0.5 - ,@/(I - P ) , 1). A material with the governing function (4.48) is linear on the region where the size of the strain deviator tensor is small (bounded by d o ) . On the region where the size of the strain deviator tensor is large, the slope of the governing function 4 is close to E ( 0 , l ) . When P approaches 0, the material on the nonlinear elasticity region becomes softer. The parameter y is introduced to ensure q5 E C1. We have

So the assumption (4.32) is satisfied. With the choice (4.48), 4(() is only piecewisely twice differentiable; it is twice continuously differentiable for > 0 except at l = t o . So the results of the previous subsection are not applicable. Nevertheless, we can slightly modify the arguments presented in the previous sebsection. For example, it can be verified that M ( E ) is strictly convex. Indeed, for ( e D1' > t o , from (4.39) and (4.48), we have

It can be bounded from below as follows:

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

154

Using the assumptions on the parameters

Recall that k = X

and y, we have

+ 2 pld.

where the coefficient

is a positive constant by the restrictions on ,f? and y. We can also show directly that the equation (4.43) is uniquely solvable. 2 Obviously, we only need to prove the function f (J) = ($I([)) J is strictly increasing when J > to.In that case,

and

If y

2 112, then

If y

< 112, then

Hence, f (J) is a strictly increasing function.

155

A Posteriori Error Analysis for Linearizations

Finally, we modify the inequality (4.47), which is inadequate when the parameter ,f? is close to zero. By (2.20), we have

where h ( t , u , u o= ) 4' (1cD(u) +tcD(uO- u)12)( c D ( u+ ) tcD(u0 - u)) cD(u0- u ) - 4' (IcD(u)12) c D ( u .)cD(u0- u ) . Define the set n l ( uu , 0 )= (3: E fl : ~ E U ( U ( X )I) ~Eo,~

Eo)

I E ~ ( u o ( ~ : )I )I~

and then, h ( t , u , u O=) t (cD(uo - u)I2. On the remaining part of 0 , it can be verified that h(t,u , uo) 2 t P l c D ( w- u)l2 if y >_ 112, and

if y

< 112. These are proved in a fashion similar to the proof of (4.50). Define

Then we have the following relation:

156

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

In conclusion, we have the following result.

T H E O R E4.10 M For a nonlinear Hencky material defined by the governing finction (4.48),the following a posteriori error estimates for the linearization of the material hold: For the energy diference,

+

- 2 [ c D ( u o ) 3 ( - 2 ~E ~ ( u o ) ) ] . E ~ ( u o ) } d x ,

(4.51)

and also

-2

+

[ c D ( u 0 ) + ( - 2 ~~ ~ ( u o .)s D ) ]( u 0 ) }dx. (4.52)

As for numerical experiments, we consider elasticity problems on a planar comer domain R (d = 2). Near the reentrant comer 0,the domain R coincides with a cone

C, = { ( r , ~: r) > 0 , -w/2 < 0 < w / 2 ) . Assume the body is subject to boundary tractions only and in the neighborhood of the comer it is traction free. We have the following singularity expansion for the solution uo of the linear elasticity problem ([146]):

Here, Go is a smooth remainder such that &(Go)is bounded when x + 0 , G is the modulus of rigidity, K I and K I I are Mode 1 and Mode 2 stress-intensity factors,

Q I ( X I+ 1 ) ) cos X I O - X I C O S ( X I - 2)Q + QI(X1+ 1 ) )sin X I O + X I sin(X1 - 2)O

(K (K

Q1l(e)=

(

(K -(

K

+

Q I I ( X I I 1 ) ) sin A110 - X I I sin(X11 - 2)O

+ Q I I ( X I+I 1))cos A110 - X I I cos(X11- 2)O

A Posteriori Error Analysis for Linearizations

Table 4.5. Singular exponents

correspond to the most singular Mode 1 and Mode 2 displacement components, where

cos ((XI - 1 ) w / 2 )

+

cos((X1 l ) w / 2 ) 1 -- - XI sin((X1 - 1 ) w / 2 ) 1 XI s i n ( ( ~ ~l ) w / 2 )

+

The parameter

K

+

depends only on Poisson's ratio v,

K

= 3 - 4v

K=-

3-v l+v

for plane strain, for plane stress.

Listed in Table 4.5 are values of singular exponents XI and XII for some typical angles. For the convenience of later use. let us denote

158

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

It is easy to verify following formulas:

Note that when the angle is 360" or 270°, both & ( u Iand ) & ( u Iare I ) unbounded around the comer, whereas when the angle is 240°, 225" or 210°, only the most singular Mode 1 strain components are unbounded around the corner.

EXAMPLE 4.11 We consider a nonlinear Hencky elasticity problem as shown in Figure 4.4, where homogeneous natural boundary conditions are not written explicitly. The part of the boundary represented by thick line segments is fixed. Assume the stored energy function is given by (4.31) and (4.48). Owing to the symmetry of the problem, we consider only a quarter problem, shown in Figure 4.5, where again homogeneous natural boundary conditions are not written explicitly. We are going to use estimates (4.51) and (4.52) to examine how much error we are making if we take the material as a linear one. Let a = 5, b = 4, c = 4, h = 0.5. By using sufficiently accurate finite element analysis, we obtain that for the L-shape singularity (approximately) where a = la,,1 . The strain energy is, approximately,

where E is the Young's modulus. is mainly determined by the first two terms Around the L-shape comer, c(uO) in the expansion (4.53), since strain components for uoare negligible. Hence,

A Posteriori Error Analysis for Linearizations

Figure 4.4. Setting of a nonlinear elasticity problem

Figure 4.5. The working problem

we use the first two terms to approxiamtely represent the solution uo of the linear problem for producing estimates for error in the (nonlinear) energy between the solution uo and the solution u of the nonlinear problem, the corresponding estimates from (4.51) andlor (4.52) being denoted by EeSt. In the following, we take

then the linear energy of the linear solution is 0.502041 x ( 2 0 ) ~ .With u = 50, the linear energy is 5.020 x lo2. We compute the error estimates for various /3 and y.See Table 4.6 for the numerical results.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 4.6. Example 4.11, error bounds for various parameters

4.3.

LINEARIZATIONS IN HEAT CONDUCTION PROBLEMS

Besides the boundary conditions considered in Section 3.6, nonlinear boundary conditions arise in certain situations, e.g., thermal radiation boundary conditions with heat transfer obeying the fourth-power temperature law, or the natural convection boundary condition with heat transfer proportional to the (514)-power of temperature difference. Provided that the temperature differences are not big the nonlinearities associated with these boundary conditions are avoided by approximating them with linear boundary conditions. We will present a posteriori estimates for the errors caused by such approximations. We will also consider the effect on solutions of solving a heat conduction problem with constant thermal conductivity while the actual thermal conductivity varies slightly within the solid, or may even depend on the temperature.

4.3.1

TEMPERATURE-DEPENDENTTHERMAL CONDUCTIVITY

In certain applications, the thermal property of a material depends on the temperature. Here we consider a case where the thermal conductivity depends on the gradient of the temperature, i.e., we consider the nonlinear differential equation problem

A Posteriori Error Analysis for Linearizations

where

f

E L~(R),g E ~ ~ ( 0 ) .

In practice, one usually replaces the nonlinear problem (4.54)-(4.55) by solving a linear problem (linearization) or a sequence of linear problems (see the discussion on the KaEanov method next chapter). We assume the thermal conductivity does not change dramatically with respect to the gradient of the temperature, i.e., k ( z , 1 ~ ~% 1~ o~ ( x) ) , (4.56) and approximate (4.54)-(4.55) by the linear problem

Let us impose some conditions on the coefficient function k for the existence and uniqueness of a weak solution of the problem (4.54)-(4.55), as well as for the a posteriori error estimate on (T - T o ) To this end, we assume k (2,J) is continuous on R x R+ (R+ = [0, cm)), inf{k(x,[) : x E R, J E R+) sup{k(x,J) : x E R,

5 E R+)

> 0, < cm;

(4.59) (4.60)

the function

is strictly convex for a.e, x E R, and there exists a positive-valued function k(x) > 0 such that Vx E R, V E , 7 E IRd, -

Wenote that if k ( x , z ) is differentiable with respect to z , then (4.61) is equivalent to the inequality

for a.e, x E R. Under the above assumptions, one can slightly modify the proof of Theorem 25.J in [I611 to conclude that there is a unique T E H1( 0 ) with T = g on dR such that

Due to the closeness of the functions k and ko by (4.56), from (4.59) and (4.60), it is natural to assume

162

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Then applying the Lax-Milgram Lemma, we conclude there is a unique weak solution of the boundary value problem (4.57)-(4.58): To E H ' ( R ) , To = g on dR such that

Let us find a bound for the error ( T - T o ) ,using the solution To of the linear problem (4.57)-(4.58). To do this, we first express the nonlinear problem (4.54)-(4.55) as an equivalent minimization problem. For this purpose, take

V = H ~ ( R )V, * = (H1(C2))*, Q = Q* = ( Av = V v , and where =

{

~~(n))~,

-' fvdx

+cc

ifv=gondR, otherwise.

It is easy to verify that the weak formulation (4.62) is equivalent to the minimization problem

T E V , J ( T , A T ) = inf J ( v , A v ) vEV

(4.63)

For the a posteriori error estimation, we need the conjugate function of J :

I

+cc

2 Jo otherwise,

'"\'"? $ 1

where the constraint set is

and q = -k* ( x ,q * ) is the unique solution of the equation

-q* = k ( x , lq12)q.

163

A Posteriori Error Analysis for Linearizations

Now we consider the quantity D(T,T o )defined in (2.12). An upper bound of the difference, by Theorem 2.40, is

J(To,ATo)

+ J*(A*q*,-q*) = /R [q*. (Vg + k * ( x q, * ) )- f

(To - g )

for any q* E Q:. For a lower bound, we first note that from the definition of G,

J k(x,

W P ) , q) =

R

I P ( X ) I ~PW )

. q ( x )dx.

Recall the assumption (4.61). Use the formula (2.20),

D(T,To) =

1'

[ k ( x ,IVT

R 0

+ t V ( T o- T ) J (VT ~ ) + t V ( T O- T ) )

Thus, we have the following error estimate:

Snle(x)IWTo

<

[q*. (Vg

- T)I2dx

+

+ k * ( x q, * ) )- f

(To - g )

J ' ~ k (~x ,[) ~d [ ] 'dx ~b'q* E Q;. lk'(~,g*)1~

(4.64) One can then proceed to discuss the selection of a suitable auxiliary function q* E Q: which allows the right side of (4.64) to be an efficient estimate of the true error. A particularly simple and usually good choice is

which is admissible since To is the weak solution of (4.57)-(4.58). Then, the general error estimate (4.64) reduces to

164

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Notice that when (4.56) holds, from the definition of k* ( x ,q*),we have

i.e.,

k*(x,q*) z ko(z)-'q*. Hence,

Then it is easy to see that the right-hand side of the error estimate (4.65) is close to zero, indicating the error caused by the linearization will be small.

4.3.2

A THERMAL RADIATION BOUNDARY CONDITION

In radiative heat transfer problems, boundary conditions are of the type (the fourth-power temperature law, cf. [140])

where y > 0 is associated with emittance of the solid. The resulting heat transfer problem is thus a nonlinear boundary value problem. When the variation of the temperature is not large, one may consider to approximate the nonlinear boundary condition by a linear one,

where T,is a suitable function close to the expected real temperature. Thus, let us consider the following problems

and

from the knowledge of To. In this subsection, and estimate the error (T- To) we confine our discussion to two-dimensional problems. Note that the weak are formulations for T E H' (R)and ToE H' (0)

165

A Posteriori Error Analysis for Linearizations

and

respectively. Take

Define

J ( v ,q ) = F ( v ) +

w,

where

Recalling the trace property (1.4) (page 13), we see that J ( v , Av) is welldefined on V. A weak solution of the problem (4.66)-(4.67) is a minimizer of the functional J ( v ,Av). Since J ( v , Av) is strictly convex, continuous and weakly coercive, it has a unique minimizer. Hence, the nonlinear bounary value problem (4.66)-(4.67) has a unique solution T E V. Introduce a constraint set

Then it can be shown that

+

J*(A*q*,-q*) =

+

/

q;2dx dR if q* E QE, otherwise.

4

y-1/ 4 1g

- q;l5I4ds (4.70)

Let us consider the difference D ( T ,T o )defined by (2.12). From Theorem 2.40, we have the following upper bound:

166

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

for any q* E Qz. To find a lower bound for D(T,T o ) we , notice that

and

In the computation of the derivatives of the boundary integral term, we write

and view it as the composite function of the (512)-power function and the square function and then apply the chain rule. Hence, using the formula (2.21), we have

D(T,To) 2

Sn S

IP(T - To)I2dx.

In conclusion, we get the following error bound

for any q* E QE. In particular, if we take the simple choice

which is admissible due to the boundary value problem satisfied by To,we get

5

2

L,

;

Y ( I T O+I 4~~

~

1 - 5 1' T c 3 T ~ O 2ds. )~

~

The right side of the above inequality is presumably small if the real temperature is close to Tc on the boundary.

~

167

A Posteriori Error Analysis for Linearizations

4.3.3

A NATURAL CONVECTION BOUNDARY CONDITION

Consider a natural convection boundary condition of the form (cf. [124])

where T, is the temperature of the surrounding. After a change of variable T, - T =+ T , we see that without loss of generality we may consider the nonlinear boundary value problem -div(k V T ) = f

in R,

(4.72)

When the temperature variation is small, (T, - T )in (4.71) is close to zero. In this case, we may try to approximate the nonlinear boundary condition in (4.73) by a linear one, e.g., we may consider the following linear approximation of (4.72)-(4.73):

Our purpose here is to bound the error (T -To) in terms of To.We notice that the weak formulations of (4.72)-(4.73) and (4.74)-(4.75) are to find T E H1( R ) and To E H1(R)such that

and

respectively. Let

and define the functional

168

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

where

F(v)=

f vdx,

When d 5 9, the functional J ( v ,Av) is well-defined on V . In this subsection, we assume d 5 9. For most applications we have d = 2 or 3. Then the nonlinear problem (4.76) is equivalent to the minimization problem

T E V , J ( T , A T ) = inf J ( v ,Av). vEV

Let

From the definition of the conjugate function, we can find out that

(

+m

otherwise.

Now consider the quantity D ( T ,T o )defined by (2.12). An upper bound of the difference, given by Theorem 2.40, is

for any q* = (97,q;) E QF. For a lower bound of the quantity D ( T ,T o ) ,we first compute the Giteaux derivatives of the functional G.

A Posteriori Error Analysis for Linearizations

Using the formula (2.21), we have

Combining (4.79) and (4.80), we then obtain the error estimate:

for any q* E QF. In particular, if we take qT = -k VTo,qz = -hTo, which is admissible by (4.77), we then have

Recall that To here actually represents the difference (T, - T o ) .We expect the right side of (4.81) is a small quantity.

4.4.

NONLINEAR PROBLEMS WITH SMALL PARAMETERS

Assume we have a nonlinear problem for which the nonlinear part is of lower order in differentiation and contains a small parameter. We will derive an estimate showing how much error is introduced if we discard the term with the small parameter. Let R c IRd be a Lipschitz domain, aR = rl u with rl relatively closed, r2relatively open, rl n r2= 8, and

Let a parameter cu satisifes - 1 < a 5 4 / ( d - 2 ) for d 13 and - 1 < a in case d = 2. From the Sobolev embedding theorem (Theorem 1.12),

is valid for p 5 2 d / ( d

-

2). Thus for any v E H1( R ) ,the integral 0

170

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

is well-defined. For any X 2 0 , let us consider the following nonlinear boundary value problem

For the data, we assume

The weak form of the boundary value problem is to find ux E H 1 (R), ux = gl on I'l such that

It can be verified that this weak formulation is equivalent to the minimization of the energy function

+

among all functions v E gl H;, (a).We can apply Theorem 2.30 to conclude that the minimization problem, and then the weak formulation of the nonlinear boundary value problem as well, has a unique solution ux E H 1 ( 0 ) . If X is small, it is natural to set X = 0 in the equation, and obtain an approximate linear problem

The corresponding weak formulation is to find u0 E

gl

+ H : ~(R) such that

has a unique solution, following an application of the Lax-Milgram Lemma.

171

A Posteriori Error Analysis for Linearizations

We are interested in giving an estimate for the difference (ux - uo). We set

v = HI (R), V* =

(HI

(R))* ,

Q = ( ~ ~ ( nx )~ )~~+ ~ ( n ) , Q* = ( ~ ~ ( 0 x ~1( )u +~~ ) / ( @( R + )l ,) Av = (Vv, v), and introduce the functional

where

+a

otherwise,

We compute the conjugate functions. First, we have F*(A*q*)=

{k

[d . Vg1 + (q; + f ) 911 dx +

where the constraint set Qf = {q* E Q* :

Hence,

9291 d~ if Q* E

QE,

otherwise

172

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Now consider the quantity D ( u x ,uo)defined by (2.12). Applying Theorem 2.40, we have

for any q* E Qz. If we use the following relation for uo:

then the bound can be rewritten as

for any q* E Qz. In particular, if we take

then we have the bound

<

X

/

D(uAu , O) a+2

l u ~ dx. ~ + ~

We now turn to a lower bound for D ( u x ,u o ) Note that

We can verify that

(WP) q, q ) t Using the formula (2.21), we have

Iq1 12dx.

A Posteriori Error Analysis for Linearizations

Hence, we have the following general error bound:

for any q* E

QE,

and also the particular error bound:

Note that X is a small parameter.

4.5.

A QUASILINEAR PROBLEM

Consider the quasilinear problem

where p > 1 and as in the previous section, let R c JR2 be a Lipschitz domain, dR = rl U with rl relatively closed, r2relatively open, rl n F2 = 0 and meas (rl)# 0. We assume

When the input data (i.e., f , gl and g2) are small, we expect we may ignore the - ~problem ) / ~ and obtain a term IVuI2 in the expression (1 I V U ~ ~ ) (of~the linear problem

+

Our task in this section is to derive an estimate for the error caused by replacing the quasilinear problem with the linear problem. For the spaces, we take

174

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

where p* is the conjugate exponent associated with p, defined by the relation l / p l l p * = 1. Denote

+

For the operator, we take

Av = V v . For the functionals, we take

where,

We can then verify that the weak solution of the nonlinear boundary value problem (4.82)-(4.84) is also the solution of the minimization problem inf J ( v , Av).

vEV

Now we compute the conjugate functions. Let

We then have

where q = + ( q * ) is the unique solution of the equation 2 ( ~ - - 2 ) / 2=~ q*

( 1 + 141 )

175

A Posteriori Error Analysis for Linearizations

Let us prove that the above equation has a unique solution, and indicate a way to compute it. First, applying the vector norm to both sides of the equation, we obtain I4*l2 = (1 + 41

2 P-2

)

/q2-

(4.88)

+

Since the function h ( t ) = (1 t ) ~ - is ~ tstrictly increasing for t 2 0 , there is a unique solution 141 from (4.88) for any Iq* 1. After computing / q 1, we obtain +b*)from

*

= (1 + I

2 1-P/2

)

q*.

Let I q*/ and / q 1 be related by (4.88), the conjugate function G* (-q* ) can be written as

We assume the solution of the linear problem (4.85)-(4.87) satisfies

For a comer domain problem, denoting w the maximal internal angle of the domain, if the boundary conditions around the corner are of the same type and the boundary value function is smooth, the condition (4.89) is satisfied if

If the boundary condition type changes around the corner and the boundary value function are smooth, then (4.89) is satisfied if

Applying Theorem 2.40, we then have the following bound on the energy difference ED (u, uO),defined by (2.1l),

for any q* E Qz. Noting the relation

176

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

we can write the estimate as

-

I ~ u o / ~ + ( +q V* u o ) . V g l

1

dx V q * E Qz.

In particular, if we use the simple selection

q* = -Vuo, then we get the following estimate

E X A M P L E4.12 Let us assume that R is a unit circular corner domain with w E ( T ,27rI: R = { ( r , Q ) : O < r < 1, 0 < Q <w ) . Denote a = T / W . Assume the input data are such that the solution of the linear problem is uo = X r" sin aQ. For the numerical calculations we take X = 0.01. Table 4.7 provides numerical values of the energy J ( u o ,n u o ) and error bound from (4.90) on the energy difference for various values of the power p and the internal angle w . rn

4.6.

LAMINAR STATIONARY FLOW OF A BINGHAM FLUID

Consider an incompressible laminar stationary flow in an infinitely long cylindrical pipe under the effect of a drop in pressure. Let the pipe be represented by R x R,whose generators being parallel to the Ox3 axis in an orthonormal system 0 x l x 2 x 3 .The cross section of the pipe $2 C R2is a planar Lipschitz domain. Assume the flow satisfies the adhesion condition that the velocity field vanishes on d R x R.Denote a positive constant pd the drop in pressure per unit length along the pipe. Further assume the flow obeys the constituent laws for Bingham fluids. Then, the velocity field of the flow is ( 0 ,0 , u,), and u, E H: ( R ) solves the variational inequality

A Posteriori Error Analysis for Linearizations

Table 4.7. Example 4.12, error bound for energy difference

where the positive constants g and p are the yield limit (threshold of plasticity) and the viscosity of the Bingham fluid, respectively, and a ( u , v) =

S,

V u . Vv dx,

We use the subscript g for the solution ug to emphasize its dependence on g. Since the bilinear form a ( . , .) : H;(Q) x H; ( 0 ) + R is continuous and (a)-elliptic, j ( - ) : H:( 0 ) + R is continuous and convex, the problem (4.91) is uniquely solvable (Theorem 1-25in Section 1.7). When g = 0, we recover from (4.91) the flow problem for classical Newtonian fluids: Find uo E (0)such that

HA

HA

ug is Lipschitz continuous from &I It is proved in [47] that the mapping: g to H; ( 0 ) . Therefore, for small g we can view the Bingham fluid as a model close to the classical Newtonian fluid. Since the Newtonian fluid flow problem

178

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

(4.92) is much easier to be solved than the Bingham fluid flow problem (4.91), it is natural in practice to approximate the problem (4.91)by the problem (4.92) when the yield limit g is small. In the next subsection, we derive an a posteriori error bound for ( u 9 - u ) in terms of u . Then, once we have solved the linear problem (4.92),our a posteriori error bound will help to assess the reliability of taking the solution u to describe the laminar stationary flow of a Bingham fluid.

4.6.1

A POSTERIORI ERROR ESTIMATE

Let d

Q = Q* = [ L 2 ( f l ) ], Av = V v . Define the energy functional

J ( v ,Av) = F ( v )

+ G(Av),

v E V,

where

F(v)= -

I n pdv dx,

The derivation below is valid for any pd E ~ ~ ( f lalthough ) , in the context of application in this section, pd is assumed to be a positive constant. Then, the variational inequality (4.91)is equivalent to the minimization problem: u E V, J ( u , A u ) = inf J ( v , A v ) . vEV

Let us compute the conjugate functions.

F* (A*q*)= sup [ ( A v q, * ) - F ( v ) ] uEV

=

sup V E H ;(a)

0 cc where the constraint set

(q* . V V

+ pdv) dx

ifq*~Qz, otherwise,

A Posteriori Error Analysis for Linearizations

The constraint condition q* E Q f is a weak for of the relation in C2 .

div q* = p d Next,

=

/

sup (-q* q a

q€Bd

;

q12 - 9 q l ) dx

sup tlq*l - E t 2 - g t ) dx 2

=Sntto(

where r+ = max{r,0 ) . Hence, J * * *

-,*I

=

{

1 -

1 2 P

-9

*

if q* E

Qf,

otherwise.

co

Note that the functional G ( q )is not Giteaux differentiable. So we use the energy difference

E D ( u g ,uo) = J(u0,Auo) - J b g , hug) to derive a posteriori error estimates. First, by definition,

- ~d ( U O -

Take v = uo in (4.91) to obtain

Hence,

I

ug)

180

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

On the other hand, by Theorem 2.40, we have the following general bound:

=S,

1 ' ~ V ~ O ~ + ~-Pduo+ ~ V U- (O ( q *I 1 - 9 ) : 2P

12

I

dx

for any q* E Qz. In particular, we can choose q* = - p V u o . Then,

When lVuo1 2 g l p , the integrand equals g 2 / ( 2p ) . When l V u o1 integrand equals (g 1 V u o1 - p I V u o/ 1 2 ) . Therefore,

'1

g2 ED(U,, u o ) 5 meas (R) - 2P Ivuol<9IP

< g l p , the 2

(

~

-

V

U

~

~

)

~

X

p

Combining this with (4.93), we obtain the a posteriori error estimate: I I V ( U-~" 0 )

4.6.2

g2 I I L ~ ( ~ )< - p2 meas(R) -

/

2

("l~ul) IVUI<~IP p

dx. (4.94)

EXAMPLES

We examine two examples to show the effectiveness of the error estimate (4.94). Note that in the derivation of (4.94), we did not use the assumption that R c IR2. Hence, the estimate (4.94) holds as long as R is an open bounded domain in Etd,for any positive integer d.

EXAMPLE 4.13 Assume R = (0, I ) , and g nonlinear problem is ([65]):

< p d / 2 . Then the solution of the

The solution of the corresponding linear problem is

.

A Posteriori Error Analysis for Linearizations

For the first order derivatives, we have

and

u ~ ( x=) -Pd (I 2P

-2 ~ ) ,

The true error is

The estimated error from (4.94) is

Comparing (4.95) and (4.96), we observe a good agreement between the exact error and the estimated error. rn

<

EXAMPLE 4.14 Assume 0 = {x E R~ : 1x1 < T O ) ,g p d r 0 / 2 . Denote rb = 2 g / p d . Then the solution of the nonlinear problem is ([65])

The solution of the corresponding linear problem is

182

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

We have

Thus, the true error is

The error bound from (4.94) is

Once again, we have a good agreement between the true error (4.97) and the error bound (4.98).

4.7.

LINEARIZATION IN AN OBSTACLE PROBLEM

In this section, we derive a posteriori error estimates for the linearization of an obstacle problem, and show the effectivity of the estimates through a one-dimensional example. In addition to discussion on the continuous level, we also show how a posteriori error estimates can be similarly derived for the linearization of the discretized obstacle problem.

4.7.1

AN OBSTACLE PROBLEM

An obstacle problem determines the equilibrium position of an elastic membrane, which passes through a curve on part of its boundary, lies above an obstacle, and is acted by a vertical force ( [ 5 ] ) .One particular obstacle problem has been introduced in Example 1.26. Here, we consider a general form of the obstacle problem as follows.

183

A Posteriori Error Analysis for Linearizations

Let R c IBd be a Lipschitz domain. The boundary r = dR of the domain R is decomposed to two measurable parts: r = rl U with rl relatively closed, r2relatively open, rl n F2 = 0, and meas(rl) # 0. Let

be given, with the assumption

Define

H;,,,, (R) = {u E H' ( 0 ) : u = gl on r l } and the admissible set

Then the obstacle problem is to find a solution u E K of the minimization problem E ( u ) = inf E ( v ) , (4.100) WEK

where the energy function

The variational inequality, equivalent to the minimization problem (4.100) is to find u E K such that

It follows from Theorem 1.24 that under the stated conditions on the data, the obstacle problem (4.101) admits a unique solution u E K. In the definition (4.99) of the admissible set K, the constraint "v $ in R" represents the requirement that the membrane lies over an "obstacle" described by the function $. We note that the region

>

i.e., where the membrane touches the obstacle is an unknown.

184

4.7.2

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

LINEARIZATION, A POSTERIORI ERROR ESTIMATE

It is advantageous to make simplifications in modeling a real problem wherever possible, and it is highly desirable to know quantitatively how much error is introduced through the simplifications, once the simplified mathematical problem is solved. The goal of the section is to provide such an a posteriori error estimate for linearization. Let us drop the constraint "v 2 $ in R" for an admissible function, and minimize the energy function E ( v )over all v E H 1( 0 )satisfying v = g l on r l . In other words, we introduce the simplified problem of finding uo E Hhl,gl( 0 ) such that

The corresponding variational equation problem is to find uo E that

L V u o . V v d x = L f v d e + L g2vds

~h~ ,gl ( 0 )such

VVEH:~(R).

2

This is the weak form of the linear elliptic boundary value problem

Our question now is whether Go = max{uo,$1 can be accepted as an approximate solution of the obstacle problem (4.101). The reason we choose Go instead of uo is that Go E K, whereas uo does not have this property. Let us derive an a posteriori error estimate for ( u - G o ) . For this purpose, we introduce the spaces:

the linear operator A E L (V,Q ):

and the energy functional

J ( v ,Av) = F ( v )

+ G(Av),

A Posteriori Error Analysis for Linearizations

where F(v) =

{ ;if

v E K, otherwise,

vdx9

Obviously, E ( v ) = J ( v , Av), v E K. Consider the quantity D ( u , v) given by the formula (2.21). We notice that for P, q E Q3

Then by the formula (2.2I), we have

Now we turn to an upper bound for D ( u , v). We calculate the conjugate functionals. F*(A*q*)= sup [(Av,q*) - F ( v ) ] vEV

= sup [/,(q?

n u

WE K

={& Thus,

+f

.)dx

+

l2

n w s ]

1

q ; I2dx

if qz = g2 on T2, otherwise.

186

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

In the following, we make the particular choice

For any w E H : ~

(R), from the boundary value problem for UO,we obtain

Thus,

and we can modify the formula for J*(A*q*,- q * ) accordingly. Therefore, applying Theorem 2.40, we get the following upper bound for D ( u ,Go):

From this inequality and (4.102), we have thus proved the following general (R), error estimate: for any w E H;,,,,

In particular, if we take w = uo, then we have the a posteriori error estimate:

A Posteriori Error Analysis for Linearizations

187

And if we take w = fro, then

We will see later through a detailed analysis for a one-dimensional problem, that the a posteriori error estimates (4.104) and (4.105) provide efficient error bounds.

4.7.3

DISCRETE A POSTERIORI ERROR ESTIMATE

For simplicity, we will describe the results for triangular linear elements for an obstacle problem on a polygonal domain a. The generalization of the results for higher order elements, rectangular elements and higher dimensional problems on general domains is straightforward. In particular, our discrete a posteriori error estimates hold independent of concrete structures of finite element spaces. Thus, we triangulate into finite number triangles satisfying the standard condition that two distinct triangles from the triangulation either are disjoint, or have a common vertex or a common edge (cf. Section 1.8). We further assume and are vertices of the triangulation. Let h that the intersection points of be the maximum length of the triangle edges. Assume the pointwise values of gl and '$ exist. We let Nh denote the set of all the nodes (the triangle vertices Nh,rlthe set of those nodes lying on rl, and Nh,o= Nh\Nhlrlthe set in 2), of free nodes. We use the finite element set

Vh = { v h E ~ ' ( 0 : vh) is linear on each triangle, v h ( a ) = g l ( a ) v a E Nh,rl)

to approximate

HI^ (a), and ,gl

to approximate K. Then the finite element solution of the obstacle problem (4.100) is uh E Kh such that

E(uh)= inf E(vh) ~h€Kh

which is equivalent to the constrained quadratic programming problem (cf. [ 6 5 ] )

where A = ( a i j ) E IRNxN is symmetric positive definite, v = ( v l ,. . . , v ~ ) ~ , N is the number of nodes in Nh,0,and C is the constraint set

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

188

<

with { & , 1 5 i N) the set of the nodal values of $ on Nh,o. Convergence and error estimates of the finite element approximation have been given by several authors, see, e.g., [ 5 2 ] . An over-relaxation method with projection can be used to solve the constrained quadratic programming problem (4.107) (cf. [65]). The method is an iteration process, consisting of two parts.

Initialization, u0 E C. Iteration. For n 2 0,compute

In the above algorithm, w is a relaxation parameter, Pi is a projection operator defined through Pi ( t )= max{t,gi) It is proved in [65] that if 0 < w < 2, then for any initial guess u0 E C , the un = U , with u being the unique solution iteration method converges, limn,, of (4.107). Although theoretically the convergence of the method is guaranteed, as long as 0 < w < 2, in practice, the efficiency of the method depends on the initial guess and the choice of the relaxation parameter w. In general, a relaxation parameter w that is good for the SOR method in solving the linear system Au = b is also a good choice for the above relaxation iteration method in solving the constrained problem (4.107). Below, we focus on a discussion of providing a good initial guess from the corresponding finite element solution of a linearized problem. Our strategy is the following. First, we solve the finite element system of the linear problem, and construct an approximation of the finite element solution of the obstacle problem. Then, we employ a discrete a posteriori error estimate to bound the difference between the finite element solution of the obstacle problem and the approximation solution. It may happen that for certain given data and certain prescribed error tolerance, the approximation solution can be taken as the finite element solution of the obstacle problem. In this case, we stop our computation, and obviously, much energy has been saved. Otherwise, the approximation solution is not accurate enough to be taken as the finite element solution of the obstacle problem, we need to solve the finite element system (4.106) (or equivalently, (4.107)). Let us use the over-relaxation method with projection stated above, and let us use the obtained approximation solution as the initial guess. There are two advantages associated with this initial guess. First, to certain degree, the approximation solution is close to the finite element

A Posteriori Error Analysis for Linearizations

189

solution of the obstacle problem. Thus, we will need only a few iterations to get an acceptable solution. Second, more important from the point of view of practical computations, we have a computable bound on the error ( u - u O ) . For the iteration error, we have estimate of the form 1 u - u n 11 5 c kn 11 u - u011, where, the constants c > 0 and 0 < k < 1 can be computed from the problem setting or estimated from several iterates. Therefore, with the knowledge on the size of the initial error, it is easy to find the number of iterations needed for achieving required accuracy. As in the case for continuous problems, we introduce the finite element solution of a linearized problem. We drop the constraint in the finite element set K h , and use Vh to define a finite element solution of the linearized problem, - u ~ , hE Vh such that E(uO,h)= inf E ( v h ) . ~hcvh

In practice, it is easier to solve (4.108) than (4.106). Especially when the given data are of certain special type, fast solution method exists for solving (4.108). E Kh After we get the solution uo,h, we define an approximation solution by f i o , h ( a ) = max{uo,h(a), '$(a)) v a E N h , o . (4.109) We can then derive the following discrete versions of a posteriori error estimates (4.104) and (4.105):

Now, a modified over-relaxation method with projection for solving the finite element system (4.108) consists of three steps. In the first step, we solve the linear problem (4.108), and define an approximation through (4.109). In the second step, we apply one of the above two a posteriori estimates to see how big is the error //V(fio,h- uh) 1 1 L 2 ( n ) . If the error bound is smaller than the error tolerance, we accept as the finite element solution of the obstacle problem, uh, and stop computation. Otherwise, we go to the next step. In the third step, we take as the initial guess, and apply the over-relaxation method with projection to solve (4.106). Here, we also use the computed bound on IIV(fio,h- uh) 11 L 2 ( n ) to decide how many iterations we need to get an accurate solution uh.

190

4.7.4

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

A ONE-DIMENSIONAL EXAMPLE

Let us consider a one-dimensional obstacle problem. We take R = ( 0 ,I ) ,

rl = F = d R , r2= 0,$ = 0, and the admissible set

K = { v E H ; ( R ) : v 2 Oin ( 0 , l ) ) . Given f E ~ ~ I( ) ,0the, one-dimensional obstacle problem is to find u E K such that E ( u ) = inf E ( v ) , (4.110) vEK

where the energy function is

Equivalently, u E K is the unique solution of the variational inequality:

To make specific computations we use the following form of the force function 2 (4.111 ) f(x)=--J-X, 3 where the parameter J E [O, 2/31, We first derive an analytic expression for the solution u of the obstacle problem (4.110)when f is given by (4.111). Formally, the solution u of the obstacle problem (4.110) solves the following boundary value problem:

- u l ' > f , ~ 2 0 ,( - u t l - f ) u = O i n n , u = O o n r , u = O and u l = O on F*, where r*is the "interface" between the sets { x E R : u ( x ) = 0 ) and { x E R : u ( x ) > 0 ) . We can obtain the following formula for the (unique) solution of the obstacle problem:

As for the solution uo of the corresponding linear problem for (4.1l o ) , we have 1 u o ( x )= - x 3 (4.112) 6 3

A Posteriori Error Analysis for Linearizations

Hence, our linear approximation Go = max{uo,0) to the solution u is

From the error estimate (4.104), we have the following a posteriori error estimate: 1

1

1

( ~ ~ - u I )5 ~ d x [Gf -uf - 2 0

(:

- J - L )

(Go - u o ) ] dx. (4.113)

We have a similar error estimate from (4.105). We note that the size of the parameter J determines the closeness of the linear problem to the obstacle problem. In particular, when J = 0, the two problems coincide, and the linearization causes no error. When f > 0 is close to 0, we expect the solution of the linear problem can be used as an approximation of the obstacle problem. The estimate (4.113) offers an upper bound for the error (Go - u), employing the solution uo of the linear problem only. We will see how accurate is our a posteriori error estimate (4.113) as compared to the true error. Since we have analytic expressions for u, Go and uo, we can compute both sides of the estimate (4.113) exactly. Thus, for the model one-dimensional obstacle problem (4.1 10) [with f given by (4.11I)], we can compare our a posteriori error bound with the exact error. Now let us compute the true error

The error upper bound from (4.113) is

We observe that as J -+ 0+, the approximation error 11 Gb - ulII L 2 ( n ) -+ 0, and the ratio between our error bound for / / f i b - ~ ' l l ~ 2 (and ~ , the exact error 11 Gb - u111 L2 cn)tends to

Thus, our a posteriori error estimate provides a quite accurate quantity for the exact error 11 Gb - ull/L2(n), when f is close to 0.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 4.8. Efficiency of error bounds on a one-dimensional obstacle problem

Table 4.8 shows some numerical results. Numerical results from the error estimate (4.105) are similar, and are omitted here. Although the above example is for the effectiveness of our a posteriori error estimates for continuous problems, it also indicates the effectiveness of the a posteriori error estimates for finite element solutions, for the discrete solutions will be sufficiently close to the continuous solutions when the finite element mesh is sufficiently refined.

Chapter 5

A POSTERIORI ERROR ANALYSIS FOR SOME NUMERICAL PROCEDURES

In this chapter, we apply the duality technique to derive a posteriori error estimates for some numerical procedures, including the regularization method and the KaEanov iteration method, in solving nonlinear boundary value problems. The regularization method is usually employed in numerical treatment of problems involving non-smooth terms. One family of such problems is the variational inequalities of the second kind. The KaEanov iteration method provides a sequence of linear problems to approximate a nonlinear problem, and can be quite efficient when the nonlinearity is not strong. For practical implementation of the regularization method or the KaEanov method, it is highly desirable to have some a posteriori error estimate so that once a regularization iterate or KaEanov iterate is computed, we can compute a (presumably efficient) error bound for the approximate solution with ease. Such an error estimate can provide us some information about the reliability of numerical solutions, and can also be used as a convenient stopping criterion for iterations. Our presentation of the a posteriori error analysis here is focused on the continuous problems, and can be easily adapted to treat the discretized problems, as is demonstrated in Section 5.4.

5.1.

A POSTERIORI ERROR ANALYSIS FOR REGULARIZATION METHODS

The purpose of the section is to give an a posteriori error analysis of regularization methods for solving non-differentiable minimization problems. The model problem to be discussed is a generalized version of an obstacle problem considered in [132]. Let R C IRd be a Lipschitz domain. In the study of obstacle problems for applications, d = 2. However, all the arguments 1 be ~ below are valid for any dimension d. Let f E L~( 0 )and g E ~ ' (OR)

194

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

given non-negative functions. Denote the energy functional

and the admissible set

K = { v E H ~ ( R: )u 2 0a.e. i n n ) , where

H$)

=

{v E ~

l ( n : v)= g

on an).

Then the obstacle problem is

E ( u ) = inf E ( v ) .

u E K,

WEK

(5.2)

The problem is equivalent to an elliptic variational inequality of the first kind:

Such an obstacle problem describes the vertical displacement of a thin elastic membrane that has a prescribed height g on its boundary, lies above a planar obstacle, and is subject to the action of a vertical force with a scaled density - f . By the non-negative assumption on f , the vertical force acts downward. Existence of a unique solution of the problem follows from the standard result on the unique solvability of variational inequalities of the first kind (Theorem 1.24, page 29). To develop the regularization method, we write the obstacle problem in another form.

T H E O R E5.1 M Denote

b ( v )=

1 (i R

j0v12

+f

v I ) dx.

Then the problem (5.2) is equivalent to the problem ofjnding u E H: ( 0 )such that E ( U ) = i n f { ~ ( v: )v E H ~ ( R ) } . (5.4)

Proof. We first show that the problem (5.4) has a unique solution. Since g E H112(dfl), it is the trace of an H 1 ( R )function, which is denoted by the same symbol g. Let us introduce a change of variable

vo = v

-g

for v E ~ ' ( 0 ) .

A Posteriori Error Analysis for Some Numerical Procedures

Then a solution of the problem (5.4) is

where uo E

HA

(R) minimizes the functional

HA

(0).This problem is equivalent to an elliptic among all the functions vo E (0)such that variational inequality of the second kind: Find uo E

HA

where

It is easy to verify that a(., -) is a continuous, H: (a)-elliptic bilinear form, j ( . ) : H; (R) + R is proper, convex and continuous, and E is a continuous linear form on ( a ) . So the variational inequality (5.6), and thus the problem (5.4), has a unique solution (Theorem 1.24 in Section 1.7). To prove the equivalence of the problems (5.2) and (5.4), we use the following result from Lemma 1.19 (page 22): If v E H1(R), then 1 v 1 E H' (0), and

HA

For the solution u of the problem (5.4), we have lul E H 1 ( R ) , lul = g on dR, and by (5.7), E(lul) L ~ ( 4 . By the uniqueness of a solution of the problem (5.4), we thus have u = lul 2 0 in R. Hence, the solution of (5.4) is also the unique solution of the problem (5.2). rn Recalling the relation ( 5 3 , we see that the solution u E H ~ ( R )of the problem (5.4) satisfies the variational inequality

f ( I V- I U ) dx 2 o 'dv E H;(R).

(5.8)

196

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

A major difficulty in solving the problem (5.8) numerically is the treatment of the non-differentiable term Ja f dx. In practice, there are several approaches to circumvent the difficulty. One approach is to introduce a Lagrange multiplier for the non-differentiable term, and the problem (5.8) (and its discretization) is solved by an iterative procedure, for detail, see, e.g., [65]. In this section, we will give a detailed analysis of another approach, namely, the regularization method. The idea of the regularization method is to approximate the non-differentiable term by a sequence of differentiable ones. The regularization method has been widely used in applications (cf. [65], [66], [94], [131]). An approximating differentiable sequence in the regularization method depends on a small parameter E > 0. The convergence is obtained when E goes to 0. However, as E + 0+, the conditioning of a regularized problem deteriorates. So, there is a tradeoff in the selection of the regularization parameter. Theoretically, to get more accurate approximations, we need to use smaller E . On the other hand, if E is too small, the numerical solution of the regularized problem cannot be computed accurately. Thus, it is highly desirable to have a posteriori error estimates which can give us computable error bounds once we have solutions of regularized problems. We can use the a posteriori error estimates in devising a stopping criterion in actual computations: If the estimated error is within the given error tolerance, we accept the solution of the regularized problem as the true solution; and if the estimated error bound is large, then we need to use a smaller value as the regularization parameter E . An adaptive algorithm can be developed based on the a posteriori error estimate.

VI

5.1.1

THE REGULARIZATION METHOD

In a regularization method, we approximate the non-differentiable term j (vo) by a sequence of differentiable ones,

The regularized problem is to find uo,,E H; (R)such that

or u, E H: (R)such that

+

The relation between the solutions of the two problems is u, = uo,, g. We expect under certain conditions, uo,,-+ uo (equivalently, u, + u)as E + 0. See the discussion below.

A Posteriori Error Analysis for Some Numerical Procedures

197

For a given non-differentiable term, there are many choices for a sequence of differentiable approximations. Let us list five natural choices of a regularizing sequence for the obstacle problem (the first two choices are taken from [94]).

Choice 1. j, (vo)=

Choice 2. j, (vo)=

Choice 3. j, (vo)=

Choice 5. j, (vo)=

f

I

$2 (vo+ g) dx with

f $:(vo

+ g) dx with

+

f

(vo g) dx with

S, 4:

(vo g) dx with

/a

f

=

+

{ (-

1 t2

2

E +&)

i i t

< r,

Following [66], we present a general convergence result on the regularization method.

L E M M A5.2 Let V be a Hilbert space, a : V x V -+ R bilineal; continuous and V-elliptic, j : V + R propel; convex and I.s.c., l : V -+ R linear continuous. For varepsilon > 0, assume j, : V + R ispropel; convex and 1.s.c. Assumefirther that for some constant cl, allowed to be negative, jE(v) > c1 'dv E V,

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

198

For relations between j

( a )

and j,

(e),

we assume

v, -+ v weakly in V

j(v) I lim inf j,(v,). &+O

(5.12)

Let u , u, E V be the solutions of the variational inequalities

and

respectively. Then, u,

-+

u in V, as E -+ 0.

Proof. In the variational inequality (5.14), we take v = 0 to get

>

cl. Since j,(O) -+ j(O), we see that {j,(O)) By the assumptions, j,(u,) is bounded for E small. Therefore, by the V-ellipticity of a(.,-), we conclude from the above inequality that {IIu,/lv) is bounded for E small. Then we have a subsequence, still denoted as {u,), and u E V, such that

Let us take the limit E -+ O+ in (5.14). By assumptions, V v E V,

Finally, from

we obtain

liminfa(u,,u,) e+O

> a(u,u).

Using all these relations, we see that the limit u satisfies the variational inequal1 ity (5.13). For the obstacle problem and its regularizations presented above, we can verify that all the conditions of Lemma 5.2 are satisfied. In particular, let us

A Posteriori Error Analysis for Some Numerical Procedures

199

show the conditions (5.11) and (5.12) and do these first for the choices 2 and 3. We notice that by the compact embedding theorem (Theorem 1.13, page 14), the space V is compactly embedded in Lp(R) for some p = p(d) > 2; for d = 2, p > 2 can be arbitrary, and for d 2 3, 2 < p < 2d/(d - 2). Let EO = (p - 2 ) / 2 > 0 and vo E Hi( 0 ) .Then for E < EO,we have

By the Holder inequality, the right hand side function is integrable:

Also, for both choices 2 and 3,

Applying the Lebesgue dominated convergence theorem (see, e.g., [18, page 2461 or most other textbooks on real analysis), we obtain the limiting relation v in V implies the existence of a (5.11). To show (5.12), we note that v, subsequence, still denoted by {v,), such that

-

v,

-+vo

in L P ( R )and a.e. in R.

We can then similarly apply the Lebesgue dominated convergence theorem to obtain (5.12), where the inequality is actually an equality:

j ( v o ) = lim j&(v,).

&to

The verification of (5.11) and (5.12) for the choices 1, 4 and 5 can be done similarly, and is actually simpler. It is not difficult to show that for any t E&!FI

Then for the choices 1 and 5,

and for the choice 4,

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

200

We have seen that for some choices of the regularization function, there is an inequality of the form

for some constant co. Under this additional assumption, we can actually have an error bound on (u- u,)for solutions of the variational inequality (5.13) and its regularization (5.14). We take v = u,in (5.13) and v = u in (5.14) to obtain

Add these two inequalities,

.) (denoting a the ellipticity constant) and the Using the V-ellipticity of a(., assumption (5.15), we get

5.1.2

A POSTERIORI ERROR ESTIMATES

As far as practical computation is concerned, a convergence result and an a priori error estimate are not enough for a complete numerical analysis with the regularization method. A posteriori error estimates are more desirable that will provide a quantitative error bound once a solution of the regularized problem is computed. In this regard, when the condition (5.15) holds, we can use (5.16) for an error bound. In this subsection, we will use the duality theory to derive some other a posteriori error estimates. We will first give an analysis in the framework of a general regularization method, and then discuss the applications to the regularization method with various choices of the regularization sequence. For the obstacle problem (5.4) considered in this section, we use the function spaces

the operator

Av and the functional

=

(VV~V)~

J(v,Av) = F(v)+ G(Av),

A Posteriori Error Analysis for Some Numerical Procedures

where F(v) =

0 oo

ifv = g o n d R , otherwise,

d

As before, we use the notation q = (q,, q2) for q E Q, with q, E ( ~ ~ ( 6 2 ) ) and 92 E L~( 0 ) . A similar notation is used for q * E Q*. With the above notations, the obstacle problem (5.4) can be rewritten in the form of (2.5). To apply Theorem 2.40 for a bound on (u, - u), we first compute the conjugate of the functional J. We have F * ( A * q * ) = sup / ( q ; - V u + v q j ) d z VEH;(R) R

and

-

lq212

dx if 1qz1

< f a.e. in 0,

otherwise. Hence, denoting the constraint set QE = {q* E Q* : L ( q ; V V a

+ qjv) dx = 0 Vv E H ~ ( R ) ,

we have

\

otherwise.

Since G(q) is not Giiteaux differentiable, we consider the energy difference

202

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Using (5.8) with v = u,,we find that

On the other hand, applying Theorem 2.40, we have

for any q* E Qz. Hence, we have the general inequality for a posteriori error analysis:

Let us then turn to the regularized problem (5.10). Since 4,is differentiable, the variational inequality (5.10) is equivalent to the equality

Thus u,is the weak solution of the elliptic boundary value problem

+

-nu, f (4,)'(u,) = 0 in R , u,= g o n d R . From (5.19), we observe that if the regularizing function 4 satisfies the inequality ' I 5 1 vt E 4 (5.20) then a natural selection of an auxiliary field q* in the basic inequality (5.18) is

From (5.18), the following a posteriori error estimate is obtained:

Taking v = u,- g E HA (R) in (5.19), we find that

Sn

[V% . V(%

- g)

+ f (4E)'b.)

(U&- g)] dx = 0.

Therefore, we can write the a posteriori error estimate in the form of

A Posteriori Error Analysis for Some Numerical Procedures

203

The regularizing functions in the choices 1, 4, and 5 satisfy the inequality (5.20). Hence, we have the following a posteriori error estimates. For the choice 1, we have

Thus the a posteriori error estimate is

For the choice 4, we have

Thus the a posteriori error estimate is

For the choice 5, we have

So we have the same form of a posteriori error estimate as that given in (5.23). For the choices 2 and 3, the regularizing functions do not satisfy the inequality (5.20). It is still possible to construct an admissible field q* from u,to produce a good error bound, but the procedure will be more involved.

REMARK5.3 The material of this section is based on the reference [87]. An earlier work on a posteriori error analysis of regularization methods for a holonomic elastic-plastic problem is given in [69]. Similar a posteriori error estimates have been derived for regularization methods in the more complicated situation of elastoplasticity problems. See [75, 77, 78, 79, 801.

5.2.

K A ~ A N O VMETHOD FOR NONLINEAR PROBLEMS

The KaEanov method is an iteration method for solving nonlinear problems, via linearization. An early reference on the method is [91], where the method is

204

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

applied to compute a stationary magnetic field in nonlinear media. Convergence of the method is proved in the context of the particular application there, though the technique of the proof is rather general. A numerical example in [91] shows that the method is quite effective for certain problems. See also [58]. Later, the method is applied to a nonlinear elasticity problem in [I 111. For applications of the method in solving variational inequalities for transonic flows in gas dynamics, see [53, 641. A general convergence result of the method is presented in [I121 (see also [161]) for solving a nonlinear variational inequality of the first kind. In [74], convergence of the method is proved for solving an even more general nonlinear variational inequality of the mixed kind (i.e., it is an inequality both because the problem is posed over a non-empty closed convex subset and because the problem contains non-differentiable terms), and the result is used in solving a nonlinear variational inequality of the second kind arising in elastoplasticity. The main purpose of the section is to introduce the KaEanov method and present a rather general convergence result. In the following sections of the chapter, we will derive a posteriori error estimates for KaEanov iterates in the context of concrete application problems. Whenever we discuss the KaEanov method, no summation is implied over the repeated index k for the KaEanov iterates. As an introductory example, let us consider the problem

where R c &idis a Lipschitz domain with its boundary dR = rl U E, Fl relatively closed, F2 relatively open, and rl r7 F2 = 0.A problem of the form (5.25) was considered in Subsection 4.1.1. The boundary value problem (5.25) is a stationary conservation law type problem. We will impose conditions on the data later for the unique solvability of the problem. The KaEanov method for solving (5.25) is the following. Given an initial guess uo with uo = g on F l , we define a sequence of iterates {uk)k>l - by the boundary value problems:

for k = 1 , 2 , . . . . Notice that in (5.26), we solve a sequence of linear problems instead of the nonlinear problem (5.25).

A Posteriori Error Analysis for Some Numerical Procedures

205

We now provide a rather general convergence result on the KaEanov method. The presentation here follows [39]. Let V be a Hilbert space, K c V a nonempty, closed, convex subset. Let A : K + R be Giteaux differentiable, j : K + R be non-negative, nondifferentiable, convex and continuous, and t E V* a continuous linear functional on V. Define the energy functional

Then we consider the minimization problem:

U E K , E ( u ) = inf E ( v ) . VEK

(5.27)

Since A is Giteaux differentiable and j is convex, a solution of (5.27) satisfies the variational inequality

When A(v) is not quadratic in v, the first term ( A 1 ( u v) ,) in (5.28) is non-linear in u. Assume there is a functional B : K x V x V + R such that

and for fixed u E K , B ( u ;v , w ) is bilinear with respect to v and w. The problem (5.28) can be rewritten as

Suppose { K i ) z ois a sequence of nonempty, closed and convex subsets of K with properties that Ki C Ki+l, i = 0,1,. . ., and U E oKi is dense in K . Then the KaEanov method for (5.30) is defined as follows: Choose uo E KOas an initial guess. For k = 1 , 2 , . . ., we find uk E Kk by solving the problem

In the conventional KaEanov method, KO= K1 = . . - = K . Here we have a more general setting by allowing the use of a sequence of nested subsets.

T H E O R E5.4 M We keep the above assumptions on the given data V , K , {Kk), A, j, !and B. (a) Assume for each u E K , the bilinearform ( v ,w ) H B ( u ;v , w ) is symmetric from V x V to R.Also, assume there are constants S1 > 0 and So > 0 such that (5.32) lB(u;v , w)l I 61 llull llwll V U E K , V v ,w E v,

206

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

and

Then the problem (5.31) has a unique solution uk E K k , which is also the unique solution of the minimization problem, 1 - B ( u k W lv; , v ) 2

+ j(v) - t ( v )-+

inf,

v E Kk.

(5.34)

(b) Further assume A' : K -+ V * is continuous and strongly monotone, i.e., there exists a constant po > 0 such that

Also, assume the following key inequality:

A(v) - A(u) 5

1 ( B ( u ;v , v ) - B ( u ;u, u ) ) 'v'u,v E K. 2

-

(5.36)

Then (5.27)has a unique solution u E K, which is also the unique solution of the variational inequality problem (5.28). The KaCanov method defined in (5.31 ) converges, i.e.,

Proof. The equivalence between the problems (5.31) and (5.34), or between the problems (5.27) and (5.28), can be established in a standard way; cf. Examples 1.26 and 1.27. From the stated assumptions, we can verify that for each k 2 1, the functional

is strictly convex, 1.s.c.and weakly coercive on Kk (i.e., Ek ( v ) -+ oo as llvli -+ oo, v E K k ) . Thus, by Theorem 2.37, the minimization problem (5.34) has a unique solution uk E K k . Similarly, the problem (5.27) can be shown to have a unique solution. Now we prove the convergence of the KaEanov method defined by (5.31). We consider the energy sequence { E ( U ~-) ) Using ~ > ~ (5.36) . and (5.31), we

A Posteriori Error Analysis for Some Numerical Procedures

have

Therefore, the sequence { E ( u k ) )is decreasing. Since { E ( u k ) )is bounded below by the minimum value of E ( v ) over K , we have lim [ E ( U ~ - ~E )( u k ) ]= 0.

k+co

Then, from (5.37) we have lim lluk - u ~ - ~= (0./

k+w

Using (5.35) and (5.28), we get, for any v E Kk+l,

From (5.31),

Hence,

u ) ) ~ < I I +inf Iz(v), ~€Kk+l

(5.39)

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

208 where

and 12(v) = B ( u k ;uk+l, v

- U)

+ j ( v ) - j ( u ) - t(v - u).

Il = 0. By (5.38) and the fact that {uk) is bounded in V, we see that limk,, Using (5.32) and the continuity of t , we know that there exists a constant c such that inf I12(v)l 5 c inf (llv - uII lj(v) - j(u)I). vEKk+l

Since limk,,

uEVk+1

+

Kk = UkKkis dense in K , we see that lim I z ( v ) = 0.

k+,

Therefore, limk,, Iluk - ull = 0 by (5.39). Theorem 5.4 extends various convergence results found in the literature as special cases. For example, if we let j ( . ) = 0 and take Ki = K for each i, then we recover the convergence result presented in [161]. Theorem 5.4 also generalizes the convergence result proved in [74]. We introduce the sequence {Ki) for more flexibility in applications of the method. We may, for instance, progressively enrich the finite element subsets while we perform the KaEanov iteration. More precisely, we can combine the finite element approximation and the KaEanov iteration in solving the problem (5.30). Let {Ki) be a sequence of finite element subsets such that Ki c Ki+l and Ui Ki is dense in K . Then (5.31) offers a numerical algorithm where for each k, we solve a finite dimensional linear variational inequality for uk, and as k + oo, we have the convergence uk + U. Although this coupling of the finite element method and the KaEanov method is not elaborated further in this work, it is worth exploring the possible efficiency gain offered by such a coupling. In the following, we focus on the use of the standard KaEanov method, i.e., where KO= Kl = - - . = K . For (5.28), the method is defined as follows. Let uo be an initial guess chosen from K. For k = 1 , 2 , . . . , we find uk E K such that

As a corollary of Theorem 5.4, we have the next result concerning the convergence of the standard KaEanov method.

T H E O R E5.5 M We keep the stated assumptions on the given data V, K , A, j, t and B.

209

A Posteriori Error Analysis for Some Numerical Procedures

(a) Assume for each u E K , the bilinearform ( v ,w ) t, B ( u ;v , w ) is symmetric from V x V to R . Assume there are constants S1 > 0 and So > 0, such that

and

>S~IIV-W~~~

B(u;v-w,v-w)

Vu,u,wEK.

(5.42)

Then the problem (5.40)has a unique solution uk E K , which is also the unique solution of the minimization problem 1 2

- B ( u k P 1v; , v )

(b)Further assume A' : K for a constant po > 0,

+ j ( v ) - t ( v ) + inf,

v E K.

+ V * is continuous and strongly monotone, i.e.,

Also assume the following key inequality: 1

A ( v ) - A ( u ) 5 I[ B ( u ;v , v ) - B ( u ;u , u ) ] V u ,v E K . 2

(5.44)

Then (5.27) has a unique solution u E K , which is also the unique solution of the variational inequality problem (5.28). The KaEanov method deJined in (5.40) converges, i.e.,

uk

5.3.

+u

in V , a s k

+ co.

K A ~ A N O VMETHOD FOR A STATIONARY CONSERVATION LAW

In this section, we consider the KaEanov meoth (5.26) for the stationary conservation law (5.25), derive an a posteriori error estimate for the KaEanov iterates, and present numerical examples to show the effectivity of the KaEanov method and the error estimate. Let us introduce the set

where and the functionals

v =H~(R),

210

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

where

We have with

P

B ( u ;v , w ) =

/a

a(lVu12)vv V w dx.

Then, the problems (5.25)for u E K and (5.26)for uk E K can be rewritten as

A ( u ) - t ( u ) = inf { A ( v )- t ( v ) }, UE

1 2

-B

(5.45)

K

( U ~ -uk, ~ ;u k ) - t ( u k )= inf

B ( u k - ~v;, v ) - t ( v )

The variational equation for uk E K is

We introduce the following assumptions.

( A l ) R c IRd bounded and Lipschitz continuous, H ' ( R ) , and h E H 1 I 2 ( r 2 ) * . ( A 2 ) The function a : R+ and Po, such that

rl #

a), f E H 1 ( R ) * ,g E

+ R+ is C1, and there are positive constants a l , a2

Here in ( A l ) , we assume that the given Dirichlet datum g is the trace of a gE ( R ) on rl. We note that ( A 2 )is a standard assumption in applications such as nonlinear magnetostatic field problem ( [ g l j )or elastoplasticity problem (11 121)The following result is derived from Theorem 5.5 (see also [I611).

T H E O R E5.6 M Assume ( A l ) and ( A 2 ) Then, (5.45) has a unique solution u.Let uo E K be given. Then for k = 1,2, . . . , (5.46) has a unique solution uk, and uk + u in H 1 ( R ) ,a s k + m. We now turn to an a posteriori error analysis for the KaEanov iterates. We note that a different a posteriori error estimate is given in [155],also via duality

21 1

A Posteriori Error Analysis for Some Numerical Procedures

theory, but in a different manner. That estimator is seemingly more involved to compute as it requires the computation of another linear problem roughly of the same size as the one required to produce the most recent term uk in the KaEanov sequence. There, no numerical results are given. Let

Q=( ~ Av = Vv,

~(n))~"~,

and identify Q* with Q. Define a functional from V x Q to E: J(v, 9 ) = F ( v ) + G ( 4 , where, F(v) =

{

- f v d x - ~ h v d s ifvtK, otherwise,

+m

Introduce the constraint set

Then, from the definition of a conjugate function, we can find Q* . Vg

+ f g + Iq*l b(Iq*l)

J*(A*q*,-q*)=

ghds otherwise,

+m

where t = b(lq*I) is the unique solution of the equation / q * /= a ( t 2 )t. From the definition of G ( q ) , we have, for p, q E Q, P

ifq*EQz,

212

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

<

Since a l ( s ) 0, we have

by assumption ( A 2 ) .Hence, for the quantity D(u,v ) defined by (2.12), we can use the formula (2.21) to find

For an upper bound of D(u,u k ) ,applying Theorem 2.40, we have, for any QE

QEl

D(u,uk) 5 J(uk Auk) + J* (A*Q*, -q*)

From (5.47) with v = uk

- g,

we obtain

Then,

+ I(?*\ b(\q;))- ~ ( l v u k - 1 1 2\ )v

I

~ k /

Combining the above lower and upper bounds of D(u,u k ) ,we have

A Posteriori Error Analysis for Some Numerical Procedures

213

In particular, by (5.47), q* = - a ( l V ~ ~ V -u ~ k/ E~ QE, ) and we obtain the estimate

When a is a positive constant, it is easy to verify that the right side of the error estimate (5.48) vanishes, i.e., the error estimate is exact if the differential operator associated with the original boundary value problem is linear. By a continuity consideration, we see that the right-hand side of (5.48) is guaranteed to be small if the coefficient function a is close to a positive constant (and satisfies ( A 2 ) ) .

Numerical examples. We now present some numerical results to show the performance of the KaEanov method and the effectivity of the a posteriori error estimate (5.48). The presentation here follows [76]. Our model problem is -

[ a ( l ~ ' 1 ~ ) u=' ]f ' for x E ( 0 ' 1 ) ) 4 0 ) = 40,

4 1 ) = q1.

(5.49) ( 5 SO)

([o,

Choose a function uo E C2 11) satisfying uo(0) = qo and uo( 1 ) = ql. For example, uo can be chosen as the linear interpolant of the boundary conditions (5.50):

The KaEanov method for solving the boundary value problem (5.49)-(5.50) on the continuous level is to find {uk)r=lby recursively solving the boundary value problems

We can express uk analytically in terms of uk-1. First, integrate the differential equation from 0 to x to get

214

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

where ck is a constant to be determined by the boundary conditions. We then use the value uk(0) = go,

So the integration constant c k involved in the solution formula can be computed from the boundary condition at x = 1:

In our numerical examples below, we use the composite trapezoidal rule with 401 quadrature points to compute the integrals in (5.51). The a posteriori error estimate to be used is given by (5.48). We will also apply the KaEanov method to compute finite element approximations of the problem (5.49)-(5.50). For the discrete solutions, we have an a posteriori error estimate similar to (5.48). In the examples below, finite element approximations uh and { u ~ , ~ are ) ~constructed > ~ using linear elements on a uniform mesh with 21 nodes on the domain [ O , l ] .

+

EXAMPLE 5.7 Let us consider the case when a ( s ) = 1 + X / ( 1 s ) , 0 5 X < 8. The assumptions ( A 1 )and ( A 2 )are satisfied, with Po = 1 - X/8. In this example, we let f ( x ) = 1, qo = ql = 0 , and uo( x ) = 0. Tables 5.1 and 5.2 show the convergence of the KaEanov method and the effectivity of our a posteriori error estimates for the KaEanov iterates. In these tables, I% is the iteration number, eh)k = ~ u L , ~u ; l L l ( o , l ) , ~ h , kis the a posteriori error estimate for eh,k from the discrete analog of (5.48), and = ~ ~ , ~is / e ~ the effectivity index. The corresponding values for the KaEanov method on the continuous level are denoted by ek, EI, and Bk. The first columns of the tables are iteration numbers. We observe that the KaEanov method converges fast both for solving the continuous problem and for computing the finite element solutions. For the case when X = 1 (Table 5. I), the effectivity indices lie between 1.4 and 1.5, showing the efficiency of our a posteriori error estimates. When X = 6 (Table 5.2), the convergence of the KaEanov method is still fast, while the effectivity indices are not close to 1. This is due to the fact that in our a posteriori error estimate (5.48), the left sides contain a factor of Po. Our a posteriori error estimate always bounds the true error, and hence, we expect the error estimate will become less efficient when certain parameter in a problem tends to a critical value ( A + 8- in the current example).

A Posteriori Error Analysis for Some Numerical Procedures

Table 5.1. Example 5.7, numerical results with a ( s ) = (2

+ s ) / ( l + s ) , Po = 718

Table 5.2. Example 5.7, numerical results with a ( s ) = (7

+ s ) / ( l + s ) , Po = 114

+

+

We have similar numerical results when a ( s ) = 1 X / ( 1 s2), 0 5 X < 1619. In this case, Po = 1 - 9x116. Table 5.3 provide numerical results for the problem with X = 1. H

EXAMPLE 5.8 In this example, we apply the KaEanov method to solve the problem -

[ a ( l ~ ' 1 ~ ) ~ ' ] ' =f o0 r z E [ o , l ] , u(0) = 0,

u(1) = 1.

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 5.3. Example 5.7, numerical results with a ( s ) = ( 2

+ s 2 ) / ( 1 + s 2 ) , Po = 7/16

This problem has an analytic solution u(x) = x. We choose a family of a ( s ) by following: Let 1 3 3 - -t, a ( t ) = -t 2 2 and positive parameters S and K , for x < 1 - S, forx~[l-b,l+S], for x > 1 S.

+

Then for s 2 - 1 E [-S, S], a ( s )

For s 2 - 1

> - 1, al(s) > -312

and s 2 5 1

+ h so that

> S we have

As n + 0, a tends to 1 uniformly and we might expect = &k/ek to tend to 1. Let yk = ek/ek-l, k = 1 , 2 , . . . , be the error reduction factors. Again, we observe the fast convergence of the KaEanov method, and the efficiency of the a posteriori error estimates. Tables 5.4-5.8 show us results for S = 712 and different K . Note that theoretically, our a posteriori error estimates always produce upper bounds on the exact error. Some of the effectivity index values in the tables are slightly smaller than 1, this reflects the influence of numerical errors resulting from using quadratures in computing both the error estimates and the true error. As expected, for n = 0 we get the solution in one iteration.

A Posteriori Error Analysis for Some Numerical Procedures

Table 5.4. Example 5.8, numerical results with r; = 0.20488

Table 5.5. Example 5.8, numerical results with r; = lo-'

Table 5.6. Example 5.8, numerical results with r; = lo-'

Table 5.7. Example 5.8, numerical results with r; =

217

218

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 5.8. Example 5.8, numerical results with K =

Table 5.9. Example 5.9, convergence of KaEanov iterates

EXAMPLE 5.9 As a further numerical example, we use the KaEanov method to solve the problem -

[ ~ ( ( u ' ~ )= u '1]0' ~ 0 ~ ( 1 O xfor ) x E [O, 11, u ( 0 ) = 0 , u ( 1 ) = 1.

(5.52) (5.53)

Here, a ( s ) = 712 - arctan(s)/lO. We can verify that the assumptions ( A 1 ) and ( A q )are satisfied, and d = 712 - ( 1 ~ / 4 ) / 1 0For . the initial iterate, we take ex - 1

+

4 4 = p-1.

Figure 5.1 and Table 5.9 show the convergence of the derivative u; for the problem (5.52)-(5.53). In Table 5.9, k is the iteration number, ek = 11u$ u'llL2(0,1),~k is the error bound for ek from (5.48), Ok = Ek/ek is the effectivity index, and yk = e k / e k - l is the error reduction factor. H

A Posteriori Error Analysis for Some Numerical Procedures

Figure 5.1. Example 5.9, convergence of KaEanov iterates

EXAMPLE 5.10 The last example is also interesting. Consider the problem -

[ a ( l ~ ' ~ ) u '= ] ' O for x E [01Ill u(0) = 0, u(1) = 1.

+

Here, a ( s ) = arctan(s) 7r/2. We take u o ( z ) = x2. It is easy to see the condition a l ( s ) 5 0 in (Az) is not satisfied. The other conditions in (A2) and (A1) hold, and we can take d = 71.12. Despite the fact that a does not satisfy conditions (A:,), we still have the convergence of the KaEanov iteration method, and the a posteriori error estimate (5.48) is efficient (see the numerical results reported in Table 5.10 and Figure 5.2).

5.4.

K A ~ A N O VMETHOD FOR A QUASI-NEWTONIAN FLOW PROBLEM

In this section, we extend some of the earlier discussions to cover nonlinear boundary value problems with constraints. We will take as a sample problem that of a quasi-Newtonian flow where the constraint is the incompressibility of the fluid. The discussion here, however, does not depend on the particular form of the problem we solve and can be extended in a straightforward way for other nonlinear problems with constraints. We now describe the problem of a quasi-Newtonian flow obeying the Carreau law. Let R c IEd be a Lipschitz domain. The boundary value problem

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 5.2. Example 5.10, convergence of KaEanov iterates

Table 5.10. Example 5.10, convergence of KaEanov iterates

governing a quasi-Newtonian flow is ([14, 151)

S,

p d x = 0.

22 1

A Posteriori Error Analysis for Some Numerical Procedures

) the ) ~ velocity field and ~ ( u: R ) Here u ( x )= (ul( x ) ,. . . , u ~ ( x is the rate of linearized deformation tensor with entries

-+ sdis

The unknown variablep is interpreted as apressure. As in [14,15], for simplicity we assume the homogeneous Dirichlet boundary condition for the velocity u . We consider the case of quasi-Newtonian flow and assume that the viscosity of the fluid obeys the Carreau law

where qo is the zero-shear rate viscosity, qw is the infinite-shear rate viscosity, 70 > qw > 0.The parameter X is a positive constant. Moreover, we assume r E ( 1 , 2 ] . The value r = 2 leads to the Newtonian flow, and detailed discussion of its finite element approximation can be found in [63]. For r E ( 1 , 2 ) , the problem models a pseudo-plastic fluid. Other possible forms of q can be found in [22]. To present variational formulations of the boundary value problem (5.54)(5.57), we need to introduce some function spaces. Let

v = W,l(Wd, U = { v E V : div v = 0 a.e, in R),

We observe that from the Korn's inequality (1.42), the formula

defines a norm on V . Assume the applied body force f E ( L ~ ( R )Then ) ~ .a variational formulation for (5.54)-(5.57) is to find u E V and p E M such that

B ( u ; u , v )- (p,divv) = ( f , v ) 'v'v E V, 'v'qM ~ . (div u , q ) = 0

(5.59) (5.60)

Here, (., .) is the L2( R )or (L2( R ) ) dinner product, and for w , u , v E V , 0

B ( w ;u , v ) = 2 JR

r l ( l ~ ( w ) 1~2() u. e)( v )dz.

The problem (5.59)-(5.60) is called a mixed formulation, because both the primal variable u and the dual variable p are present. Alternatively, we may eliminate the variable p to get a formulation involving the variable u only:

u E U,

B ( u ; u , v )= ( f , v ) 'v'v E U.

(5.62)

222

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

This problem is called the direct formulation in [15], and is equivalent to a constrained minimization problem

with

We notice that

( A 1 ( u v) ,) = B ( u ;u , v ) 'v'u,v E V. The mixed formulation (5.59)-(5.60) is the basis for developing mixed finite element methods. Error analysis for mixed finite element methods is done in [14] for the case r E ( 1 , 2 ] ,and in [15] for any r > 1,where improvements are made on some error estimates proved in [14]. Because of the nonlinearity introduced by the function q ( - ) the , resulting finite element system is nonlinear and is difficult to solve. One possible approach is to apply the KaEanov method to reduce the nonlinear system to a sequence of linearized problems. We may analyze the KaEanov method on both the continuous level when it is applied to the original problem and the discrete level when it is applied to solving finite element systems. Here we describe the method on the discrete level. Let Vh C V and Mh c ~ i ( f lbe) finite dimensional spaces. Define Uh = { V h E Vh : (divvh,qh) = 0 'v'qh E Mh). We now consider the approximation of the problem (5.62):

The KaEanov method for solving the problem (5.65) is the following. Choose uo E Uh as an initial guess. Then for L = 1 , 2 , . . . , we compute uk E Uh from

Let us apply Theorem 5.5 to conclude the convergence of the KaEanov method (5.66) when it is applied to solve the problem (5.65). It is easy to show that B , as defined in (5.61), satisfies (5.41) and (5.42), where we may let Sl = 70and So = qm. From Lemma 3.1 of [15], the functional A defined in (5.64) satisfies (5.43), where we may take po = vm. In order to show the convergence of {uk) defined in (5.66), it suffices to verify the key inequality (5.44).

A Posteriori Error Analysis for Some Numerical Procedures

For q ( z ) defined in (5.58), we have

Hence, for any v , w E V ,

and the key inequality (5.36) is satisfied. Thus, the method (5.66) converges following Theorem 5.5.

A posteriori error estimate. Define the operator

where

Q

{ q = (qij) : qij = qji E L ~ ( R )1, 5 i, j We identify Q* with Q. Then we define the functional

< d)

with F

Obviously, E ( v ) = J ( v , A v ) . We now compute the conjugate functional J* (A*q*,-q* ) :

where t = t(lq*1) is determined by the algebraic equation

224

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Since the function of t defined by the left-hand side of (5.67) is continuous, strictly increasing and taking on any non-negative value, for and q* E sd,the equation (5.67)has a unique solution t = t(lq*1 ) . Write ko = ko( jq* I ) = t(lq*I)/lq*1. Then ko is the solution of the equation

and

2 Xr

+ X k i iq*2)r/2

- - (70 - qm) [(1

-

1 ] }dx.

(5.69)

Define the constraint set

Then from (5.69),we get

14*12[(1- 1 1 ~ko) 2

J* (A*q*,-q*) =

+o

+m

-0

+ (217-- 1) r1&1 ifq* E

Qf,

(5.70)

otherwise.

Consider the quantity D ( u h ,u k )given by (2.21). From the definition of G ( q ) ,we have

and

Since r

Thus.

> 1 and ( p - q ) 2 5 lp121q12,

we have

A Posteriori Error Analysis for Some Numerical Procedures

Applying Theorem 2.40, we then get

Comparing the relations (5.66) and (5.70), we see that a convenient choice is q* = -2 q ( l ~ ( ~ k - l ) I 2 ) & ( u k )With . such a choice we get from (5.71) that an upper bound for both q,lluk - uh1I2and D ( u h ,u k )is

where ko is uniquely determined from (5.68) with

Notice that the expression (5.72) is guaranteed to be an upper bound for the left hand side of (5.71). It is computable once two consecutive iterates u k - ~and u k are found. Notice that usually numerical integration is needed to compute the expression (5.72).

E X A M P L E5.11 Let d = 2, R = [O, 112,f (x)= ( x 2 ,1 - x ~ )X ~= 2, , and r = 1.2. We use the Q1-Po element (see [55] for details). The domain R is divided into 9 equal squares. For the finite element spaces, Mh consists of piecewise constant functions and Vh consists of continuous vector-valued functions whose components are piecewise bilinear. A Gauss-Legendre quadrature of high degree is used to produce accurate values of Eh,k in (5.72). We use the effectivity index

to measure the efficiency of the error estimate. Some numerical results are given in Tables 5.11 and 5.12. We observe that the method converges very fast and the effectivity index Ok is close to 1, indicating that our a posteriori error estimate is effective. I

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Table 5.11. Example 5.11, KaEanov iteration for a quasi-Newtonian flow problem, = 0.0001

770

= 1,

7700

Table 5.12. Example 5.11, Kacanov iteration for a quasi-Newtonian flow problem, 770 = 100, = 0.01

77-

5.5.

APPLICATION IN SOLVING AN ELASTOPLASTICITY PROBLEM

In this section, we apply the KaEanov method and the regularization method to solve a nonlinear variational inequality of the second kind arising in elastoplasticity, and derive a posteriori error estimates. The problem is discussed in detail in [130, 771. Here, we first briefly review one mathematical formulation of the problem. Consider the quasistatic behavior of an elastoplastic body which occupies a bounded domain R c IRd (d 5 3 in practice) with Lipschitz boundary. The plastic behavior of the material is described in terms of a dissipation function, and we assume the material undergoes nonlinear kinematic hardening, the nonlinear part of which takes the form of an exponential decay. The boundary value problem we are going to present arises in a typical time-step in approximating the rate of change of the plastic strain by a backward Euler difference (cf. [130]). We assume the material is subject to the action of a body force with density b. For simplicity in presentation, we assume the boundary of the body is fixed. The unknown variables of the problem are the displacement u and the plastic strain tensor p. We seek the displacement in the space

and the plastic strain in the space

Qo = {q E Q : trq =0},

A Posteriori Error Analysis for Some Numerical Procedures

where t r p = pii is the trace of the tensor p , and

We require the restriction t r p = 0 on the plastic strain p by the conventional assumption of no volume change accompanying the plastic deformation. Both V and Q are Hilbert spaces with inner products

and the corresponding norms 11 v 11 = ( v ,v )' I 2 , 11 q 11 Q = ( q ,q ) 1 / 2 . Moreover, Qo is a closed subspace of Q. = V x Q To formulate the problem, we need to use the product space which is a Hilbert space with the inner product

v

and norm (/Ellt; = (8,8 ) g 2for , 8 = (u,p),B = ( v ,q ) E V . We also define V o = V x Qo, a closed subspace of Define an operator A : V + by

-

v*

v.

where

and

h ( a ) = ho

+ hie-pa.

We assume the fourth-order elasticity tensor (C to be symmetric:

bounded:

CijklE L m ( R ) , 1 5 i , j , k , l 5 d , and pointwise stable: for some constant co > 0,

We also assume that the coefficient p the conditions

> 0 and the functions ho and hl

satisfy

228

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

The smallness assumption on hl (relative to ho) is needed in proving the unique solvability of the elastoplasticity problem (5.74) below (cf. [130]). Then, associated with the density of body force b E V* we define the linear functional

e : V -+ R, e ( q = and for some material parameter g

S, .

b v dx

> 0 we define the functional

The functional j ( . ) is known as the dissipation function in plasticity. The functionals e ( . ) and j ( . ) are easily shown to be bounded, j ( . ) is a convex, positively homogeneous (i.e., j(XZI) = X j ( G ) for any X > O), non-negative Lipschitz continuous functional. Note that, however, j is not differentiable. The variational form of the elastoplasticity problem is: Find 2i = ( u ,p ) E Vo such that

By applying Theorem 1.25, one can show that the problem has a unique solution (see also [130]).

The Karanov method and its convergence. The KaEanov method for solving the problem (5.74) is: Choose an initial guess Go E T o ; for k = 1 , 2 , . . . , find u k E VOsuch that

where

To apply Theorem 5.5 for a convergence analysis of the KaEanov method, we notice that the problem (5.74) is equivalent to the minimization problem of finding U = ( u ,p ) E Vo such that

where

E ( v )= E l @ )

+ j ( F ) - l(5)

A Posteriori Error Analysis for Some Numerical Procedures

229

and

A crucial step in applying Theorem 5.5 for the convergence of the method (5.75) is to prove the key inequality (5.36), which in the present case, is equivalent to (5.78) in the next lemma.

LEMMA5.12 The following inequality holds:

Proof. For a fixed s

> 0, we define a function o f t by

Then from the definitions of H (.) and h ( . ) ,

Since f ' ( t )= p2t ( e - p t - e - p S ) ,

>

>

we have f ' ( t ) 0 for t E [0, s] and f 1 ( t )5 0 for t s. We conclude that the function f ( t ) ,t 2 0, attains its maximum value at t = s. Thus,

f ( t )5 f(s) = o

'dt > 0.

Hence, the inequality (5.78) is proved. The other conditions of Theorem 5.5 are easily verified. Therefore, the KaEanov method (5.75) converges.

A posteriori error estimation for the KaEanov method. Now that the convergence of the KaEanov method (5.75) is shown for the problem (5.74), we turn to derive some a posteriori error estimate for the KaEanov iterates. To derive an a posteriori error estimate for the approximation error (uk- u), we apply Theorem 2.40. Since it is customary to use p for the plastic strain, we will use s for a generic dual variable, and S for the space of the dual variables. Define the operator A : V o-+ S, Av = ~ ( v ) ,

230

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

We identify S* with S . Now define the functional

Then, the problem (5.76) can be rewritten as

J ( E , A E ) = inf { J ( B , A B ):;is E T o ) .

E E V o:

(5.79)

Note that the functional J(W, s ) is not of separated form. Define a constraint set

S; = { s * E S *

[E(v)-s*+b.v]dx=O V v EV

We can find the conjugate functional

J*(A*s*,-s*) =

{

[ic-ls* OC,

S*

I

+ K ( S * I ) dx

ifs* t

s.,

otherwise.

Here, ~ s* - ( l l d )tr ( s * ) I ,I being with the deviatoric strain tensor defined as s * = the second-order identity tensor,

If I S * ~ 5 ~ g, then t ( l ~ * = ~ I0.) If solution of the equation

> g, then t ( l ~ * ~> I0)is the unique

The unique solvability is guaranteed by the assumption ho > 0. We will say an s* E S* is admissible if s* E S,*. Notice that from (5.80), the value of J* (A*s*,- s * ) is infinite if s* is not admissible. Since the functional J ( B , s ) is not Giteaux differentiable, we consider the energy difference

for some constant E

> 0 (for an expression to calculate E > 0, see [77]).

A Posteriori Error Analysis for Some Numerical Procedures

23 1

To have an upper bound for the difference, we need to choose a suitable admissible variable s*. We notice that the problem (5.75) is equivalent to

and

From (5.83), it follows that

s* = -C ( E ( u k )- p k ) is admissible. With this choice of s * , applying Theorem 2.40, we obtain the following upper bound on E D @ , T i k ):

Applying (5.83) with v = u k , we obtain

+ H ( l ~ k l )+ g

pk

+ T ( ~ ( ~ ~ * ~ I ) ) ] dx.

Combined with (5.82), this implies the following a posteriori error estimate ( I u k - ~112, + lpk -PI/;)

where s * is defined in (5.85). For the purpose of comparing with the a posteriori error estimate to be derived next for the combined effect of a regularization procedure and the KaEanov iteration, we derive another a posteriori error estimate for the KaEanov method (5.75) as a consequence of the estimate (5.86). We take q = 0 in (5.84) to obtain

232

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Applying this inequality to the right-hand side of (5.86), we get the estimate

Regularization, a posteriori error estimate. We observe that in each step of the KaEanov iteration, one has a (linear) variational inequality of the second kind, (5.75). The difficulty in solving (5.75) directly lies in the fact that term j is non-differentiable. One approach used in practice for overcoming this difficulty is to use a regularization technique. In a regularization technique, j is replaced by a differentiable function j,, such that j, + j as E + 0. As is discussed in Section 5.1, there are many choices for the regularization function j,. Here, for definiteness, we take

Once the regularization function j, is chosen, the variational inequality (5.75) is approximated by the following

Owing to the fact that j, is differentiable, the regularized problem can be rewritten as B ( T i k - l ; T i E , k , C ) + ( j ~ ( ~ , , k ) , ~ 'dvEV0, )=e(v) or, in detail,

The convergence of the regularization method can be established by a standard procedure. From (5.16), we have some constant c such that

+ l l p k - p E , k I ; 5 C&.

l l ~ k- ~ & , k l l $

We now derive a posteriori error estimates for the approximation error Ti). Again we consider the energy difference

-

233

A Posteriori Error Analysis for Some Numerical Procedures

It can be proved that

for the same constant Zi > 0 as in (5.82). By (5.881, S*

= - c ( & ( ~ & , k -) P E , ~ )

is admissible. With this choice of s * , we have

Combined with (5.89), this implies the following a posteriori error estimate

] dx.

(5.91)

To see the efficiency of the estimate, we have that in the linear case (i.e., hl = O), the estimate (5.91) implies

g l ~ e , 1k&2

'/ JW(JW+ dx.

IP&,~I)

It is easy to see that, at least formally, the error estimate (5.91) reduces to (5.87) when E + 0.

Chapter 6

ERROR ANALYSIS FOR VARIATIONAL INEQUALITIES OF THE SECOND KIND

The finite element method today is the dominant numerical method for solving most problems in structural and fluid mechanics. It is widely applied to both linear and nonlinear problems. For practical use of the method, one of the most important problems is the assessment of the reliability of a finite element solution. The reliability of the numerical solution hinges on our ability to estimate errors after the solution is computed; such an error analysis is called a posteriori error analysis. A posteriori error estimates provide quantitative information on the accuracy of the solution and are the basis for the development of automatic, adaptive procedures for engineering applications of the finite element method. The research on a posteriori error estimation and adaptive mesh refinement for the finite element method began in the late 1970's. The pioneering work on the topic was done in [l1, 121. Since then, a posteriori error analysis and adaptive computation in the finite element method have attracted many researchers, and a variety of different a posteriori error estimates have been proposed and analyzed. In a typical a posteriori error analysis, after a finite element solution is computed, the solution is used to compute element error indicators and an error estimator. The element error indicator represents the contribution of the element to the error in the computation of some quantity by the finite element solution, and is used to indicate if the element needs to be refined in the next adaptive step. The error estimator provides an estimate of the error in the computation of the quantity of the finite element solution, and thus can be used as a stopping criterion for the adaptive procedure. Often, the error estimator is computed as an aggregation of the element error indicators, and one usually only speaks of error estimators. Most error estimators can be classified into residual type, where various residual quantities (residual of the equation, residual from derivative discontinuity, residual of material constitutive laws, etc.) are used; and recovery type, where a recovery operator is applied to the (discontinuous)

236

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

gradient of the finite element solution and the difference of the two is used to assess the error. Error estimators have also been derived based on the use of hierarchic bases or equilibrated residual. Two desirable properties of an a posteriori error estimator are the reliability and efficiency. The reliability requires the actual error to be bounded by a constant multiple of the error estimator, up to perhaps a higher order term, so that the error estimator provides a reliable error bound. The efficiency requires the error estimator to be bounded by a constant multiple of the actual error, again perhaps up to a higher order term, so that the actual error is not over-estimated by the error estimator. The study and applications of a posteriori error analysis is a current active research area, and the related publications grow fast. Some comprehensive summary accounts can be found, in chronicle order, in [153], [2], and [13]. Initially, a posteriori error estimates were mainly developed for estimating the finite element error in the energy norm. In the recent years, error estimators have also been developed for goal-oriented adaptivity. The goal-oriented error estimators are derived to specifically estimate errors in quantities of interest, other than the energy norm errors. Chapter 8 of [2] is devoted to such error estimators. The latest development in this direction is depicted in [19, 621. Most of the work so far on a posteriori error analysis has been devoted to ordinary boundary value problems of partial differential equations. In applications, an important family of nonlinear boundary value and initial-boundary value problems is that associated with variational inequalities. Although several standard techniques have been developed to derive and analyze a posteriori error estimates for finite element solutions to problems in the form of variational equations, they do not work directly for a posteriori error analysis of numerical solutions to variational inequalities. Nevertheless, numerous papers can be found on a posteriori error estimation of finite element solutions of obstacle problems, e.g., [3], [40], [84], [98], [I 141, [I501 (these papers consider numerical solutions on convex subsets of finite element spaces), as well as [56], [90] (these papers use a penalty approach for discrete solutions). Obstacle problems are so-called variational inequalities of the first kind, that is, they are inequalities involving smooth functionals and are posed over convex subsets. We also note that a posteriori error estimation is discussed in [24, 25, 1451, though the arguments in these papers are arguable. In the context of elastoplasticity with hardening, computable a posteriori error estimates are derived in [4, 34, 361 for the primal problem, which is a variational inequality of the second kind; that is, the inequality arises as a result of the presence of a non-differentiable functional. These works deal extensively also with a priori estimates, and in the latter work a number of numerical examples are presented. Residual type error estimators were studied for an elliptic variational inequality of the second kind in 128, 291.

Error Analysis for Variational Inequalities of the Second Kind

237

In this chapter, we derive and study some a posteriori error estimates for finite element solutions of elliptic variational inequalities of the second kind. The basic mathematical tool we will use is the duality theory in convex analysis, although some of the a posteriori error estimates can be derived by more direct calculations. In [134,133], the technique of the duality theory was used to derive a posteriori error estimates of the finite element method in solving boundary value problems of some nonlinear equations. In these papers, the error bounds are shown to converge to zero in the limit; however, no efficiency analysis of the estimates is given. In Section 6.1 we introduce a model elliptic variational inequality of the second kind and its finite element approximation. We will provide detailed derivation and analysis of a posteriori error estimates of the finite element solutions for the model problem. In Section 6.2 we formulate a dual problem for the model, and use the dual problem to establish a general a posteriori error estimate for any approximation of the solution of the model problem. The general a posteriori error estimate is featured by the presence of a dual variable. Different a posteriori error estimates can be obtained with different choices of the dual variable. In Section 6.3, we make a particular choice of the dual variable that leads to a residual-based error estimate of the finite element solution of the model problem, and explore the efficiency of the error estimate. In Section 6.4, we make another choice of the dual variable and obtain a gradient recoverybased error estimate of the finite element solution of the model problem. We also study the efficiency of the error estimator. In Section 6.5 we present some numerical results to illustrate the effectivity of the estimates in adaptive solution of the model variational inequality. In the last section of the chapter, we extend the discussion to solving a frictional contact problem. We comment that the discussions made here can be extended in solving time dependent variational inequalities, cf. e.g. [27].

6.1.

MODEL PROBLEM AND ITS FINITE ELEMENT APPROXIMATION

We introduce a model elliptic variational inequality of the second kind in this section. We comment that the ideas and techniques presented for a posteriori error analysis in solving the model problem can be extended to other elliptic variational inequalities of the second kind; in particular, in Section 6.6 we provide a posteriori error analysis for the finite element solution of a frictional contact problem. Let R be a domain in E t d , d 2 1, with a Lipschitz boundary I?. Let Fl c r be a relatively closed subset of r , and denote F2 = r\rl for the remaining part of the boundary. We allow the extreme situations where rl = 0 (i.e, r2= r ) and rl = r (i.e. r2= 0). Since the boundary is Lipschitz continuous, the unit outward normal vector Y exists a.e. on r . We will use d l d u to denote

238

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

the outward normal differentiation operator, that exists a.e. on I?. f E L~(R) and g > 0 are given. Over the space

Assume

we define a bilinear form and two functionals: a ( u , v) =

( V u - Vv

S,

+ u v) dx,

In the space V, we use the H1(R)-norm. The model problem is the following elliptic variational inequality of the second kind: Find u E V such that

This model is a so-called simplified friction problem following [65] as it can be viewed as a simplified version of a frictional contact problem in linearized elasticity (cf. Section 6.6). Since the bilinear form a ( . , is continuous and V-elliptic, the linear functional e(.) is continuous, and the functional j ( - ) is proper, convex and continuous, by Theorem 1.25, the variational inequality (6.1) has a unique solution. Moreover, due to the symmetry of the bilinear form a ( . , .), the variational inequality (6.1) is equivalent to the minimization problem: Find u E V such that E ( u ) = inf E ( v ) , (6.2) a )

vEV

where E is the energy functional:

The minimization problem (6.2) also has a unique solution. In the analysis of a posteriori error estimators later, we will need the following characterization of the solution u of (6.1): There exists a unique X E LCO(rz)such that

The function X can be viewed as a Lagrange multiplier.

Error Analysis for Variational Inequalities of the Second Kind

239

A proof of this characterization in the case r2= I'can be found in [65]. The argument there can be extended straightforward to the more general situation considered here; see also the proof of Theorem 6.2. It follows from (6.4) that the solution u of (6.1) is the weak solution of the boundary value problem

We now turn to finite element approximations of the model problem. For simplicity, we suppose that R has a polyhedral boundary I'.In order to define the finite element method for (6.1) we introduce a family of finite element spaces Vh c V, which consist of continuous piecewise polynomials of degree larger than or equal to 1, corresponding to partitions Phof into triangular or tetrahedral elements (other kinds of elements, such as quadrilateral elements, or hexahedral or pentahedral elements, can be considered as well). The partitions Phare compatible with the decomposition of r into rl and r2.In other words, if an element side lies on the boundary, then it belongs to one of the sets I'l or r2.For every element K E P h , let hK be the diameter of K and p~ be the diameter of the largest ball inscribed in K . For a side y of the element K , we denote by h, the diameter of y. We shall assume that the family of partitions p h , h > 0, satisfies the shape regularity assumption, i.e. the ratio h K / p K is uniformly bounded over the whole family by a constant C. Note that the shape regularity assumption does not require that the elements be of comparable size and thus locally refined meshes are allowed. We will use Eh for the set of the element sides, Eh,r, Eh,rl, and Eh,rz for the subsets of the element sides lying on I?, r l , and E,respectively, and Eh,0 = E h \ E h ,for ~ the subset of the element sides that do not lie on r. Let Nh be the set of all nodes in Ph and Nh,0c Nh the set of free nodes, i.e. those nodes that do not lie on rl. For a given element K E P h , N(K) and & ( K ) denote the sets of the nodes of K and sides of K , respectively. The patch g associated with any element K from a partition Ph consists of all elements sharing at least one vertex with K , i.e, g = U{Kt E Ph : Kt n K # 8). Similarly, for any side y E Eh, the patch 7 consists of the elements sharing y as a common side. Note that in the case where the side y lies on the boundary r , the patch 7 consists of only one element. For a given element K E P h , U K denotes the unit outward normal vector to the sides of K. When a side y lies on the boundary I?, v , denotes the unit outward normal vector to I?. For a side y in the interior, v , is taken to be one of the two unit normal vectors. In what follows, for any piecewise continuous function cp and any interior side y E Eh,O,[cp], denotes the jump of cp across y in the direction

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

v,, i.e.

In the derivation of a posteriori error estimates we will use the so-called weighted ClCment-type interpolation operator. There are several variants of such operators (see e.g. [16], [35], [37], [38], [158]), which are all versions of the interpolation operator introduced by Clkment [44]. The main difference among these interpolants lies in the way the interpolation is performed near the boundary. In this paper we will follow the approach used in [37]. Corresponding to the partition P h , we denote N j C Nh to be the set of the element vertices, Nu,rlc N;, the subset of the element vertices lying on r l , = N;, n Nh,0the subset of the interior vertices. Given a E N u , let and 4, be the linear element nodal basis function associated with a. For each fixed vertex a E Nu,rl,choose ((a) E Nuloto be an interior vertex of an element containing a . Let ( ( a ) = a if a E NuloFor each node a E Nu,odefine the class I(a) = {a E Nu : ((a) = a). In this way, the set of all the vertices N,, is partitioned into card(Nulo)classes of equivalence. For each a E Nuloset

Notice that {ga : a E N u , o v sa partition of unity. Let Fa = supp($,) and ha = diam(&). The set Ka is connected and that $, # Pa implies that I'l n & has a positive surface measure. , For a given v E L ~ ( R )let

Then define the interpolation operator Uh : V

--+ Vh as follows:

The next result summarizes some basic estimates for Uh. Its proof can be found in [35].

Error Analysis for Variational Inequalities of the Second Kind

THEOREM 6.1 There exists an h-independent constant C all v E V and f E L2(0),

24 1

> 0 such that for

The finite element method for the variational inequality (6.1) reads: Find

uh E Vhsuch that

The discrete problem has a unique solution uh E Vh. We need the following characterization of the finite element solution, similar to that of the solution of the continuous problem.

THEOREM 6.2 The unique solution uh E Vhof the discrete problem (6.11) is characterized by the existence of Ah E Lw (r2)such that

Proof. Assuming (6.1I), let us prove (6.12) and (6.13). Taking first vh = 0 and then vh = 2uhin (6.1 I), we obtain

Together with (6.14) the relation (6.11) leads to

v:,

HA

where V; = Vhn (0) and : V is the orthogonal Write Vh = V; @ complement of Vhin H : ~(0). It follows from (6.15) that !(vh) - a(uh, vh) = 0 V vh E Notice that the trace operator from V; onto : V lrz c L1(rz)is

v;.

242

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

an isomorphism. Therefore, the mapping L ( v h ) = t ( G h ) - a(uh,Gh) can be viewed as a linear functional on v:/~, , where Zh is any element from the space Vh whose trace on F 2 is vh. It follows from (6.15) that

Thus, by the Hahn-Banach Theorem the functional L(vh) can be extended to L ( v ) on L 1 ( r 2 )and so there exists Ah E Lm(F2)such that

and lXhl 5 1 a.e. on F2, from which (6.12) follows. Taking now vh = uh in relation (6.12), we have

and using (6.14) we get

Since / A h I 5 1 a.e. on r 2 ,we must have luh I = Xhuh a.e. on r2.This completes the proof of (6.12) and (6.13). Conversely, assume (6.12) and (6.13) hold. It follows from relation (6.12) that

which can be rewritten as

Then, relation (6.13) implies that

Since Xhvh 5 )vh)a.e. on r 2 ,it follows immediately that uh is the solution of the discrete problem (6.11). A priori error estimates for the finite element method (6.11) can be found in the literature, e.g. [65, 661. Here, we focus on the derivation and analysis of a

Error Analysis for Variational Inequalities of the Second Kind

243

posteriori error estimators that can be used in adaptive finite element solution of variational inequalities. Our argument of the efficiency of the a posteriori error estimators is based on that due to Verfiirth [153], with special attention paid to the inequality feature of the problem. The argument makes use of the canonical bubble functions constructed for each element K E Phand each side y E & ( K ) . Denote by PK a polynomial space associated with the element K . The following two theorems provide some basic properties of the bubble functions used to derive lower bounds. For more details on bubble functions and proofs see [2].

THEOREM 6.3 Let K E Phand gK be its corresponding bubble function. Then there exists a constant C, independent of hK, such that for any v E PK, the following inequalities hold:

THEOREM 6.4 Let K E Phand y E & ( K )be one of its sides. Let gY be the side bubble function corresponding to y. Then there exists a constant C, independent of h K , such that for any v E PK,

6.2.

DUAL FORMULATION AND A POSTERIORI ERROR ESTIMATION

We now present a dual formulation for the model problem. The dual formulation will be used in the derivation of a posteriori error estimators for approximate solutions. Let Q = ( L ~ x L~(a)x L2(r2). Define a function J : V x Q + by the formula

where q = (ql, q 2 , q3) E Q. Introduce a linear bounded operator A : V by the relation b'~ E V. AV = ( V V v,vlr2) , Then obviously, E ( v ) = J ( v , A v ) b'v E V.

+Q

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

244

Therefore, the minimization problem (6.2) can be rewritten as: Find u E V such that (6.17) J(u,Au) = inf J(v,Av). vEV

Let V * and Q* = ( L(~R ) )x~L~( 0 )x L~(r2)be the duals of V and Q placed in duality by the pairings (-, .) and (., a ) Q ,respectively. Let us compute the conjugate function

where A* : Q* + V * is the adjoint of A. We have

j*(A*q*, -q*) = sup

v v + (q;+ f) v ] dx +

[ q .~

vEV q€Q

It can easily be verified that

0, if lq;/ 5 g a.e. on r2, oo,

otherwise.

Also, note that the term sup UEV

equals

CG, unless

{S,

[q;. Vv

q* E

+ (qi+ f ) vl dx +

S,,

q:v ds}

Q*satisfies

and under this assumption, the above term equals 0. Define the set of admissible dual functions

Error Analysis for Variational Inequalities of the Second Kind

The classical relations for the constraint q* = (q?,q;, qg) E Qj,g are

Then the conjugate function (6.18) is

J * ( A * ~ *-q*) , =

Jn

Qi,gj

(q;12+ 1$12)dx, i f q * E otherwise.

The dual problem of (6.17) can now be stated: Find p* E Q j,gsuch that - J*(A*p*,-p*) =

{ - J*(h*q*,- q * ) ) .

sup q*

(6.20)

EQj.,g

Note that the mapping q* t+ J*(A*q*,-q*) is strictly convex over Qj,g. By Theorem 2.39, the dual problem (6.20) has a unique solution p* E Q jig, and

where u E V is the unique solution of the model problem. Now let w E V be an (arbitrary) approximation of u E V, the unique solution of (6.1). In the rest of the section, we present a general framework for a posteriori estimates of the error ( u - w ) . The error bounds are computable from the (known) approximant w. Later on, w will be taken as the finite element solution of the variational inequality. By using (6.1) and (6.3) we obtain

On the other hand, let p* be the solution of the dual problem (6.20). Relation (6.21) implies

E ( u ) = J ( u , Au) = - J* (A*p*,-p*) Therefore,

> -J*(A*q*,-q*)

V q* E Qi,g.

246

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Introduce the space

Q; = ( L ~ ( Qx) ~ ) ~~ ( 0 ) . Then for any r* = ( r ? r;) , E Q;, we write

Since q* = ( q ; ,qz, qi) E

Q>,g,

from (6.19),

Using this relation in (6.23) and recalling (6.22), we find that

Thus, we have established the following result.

247

Error Analysis for Variational Inequalities of the Second Kind

T H E O R E6M . 5 Let u E V be the unique solution of ( 6 . 1 ) , and w E V an approximation of u. Then the following estimate holds for any r* E Q::

Let us now deal with the second term I1 on the right side of the estimate (6.24). First, from the definition (6.19) it follows that

+ lq; - r z 2 )dx +

= inf sup { L ( l q ;- r;I2 EQ* V E V 14; 159

q*

+

( g /wl q j w ) ds

Here and below, the condition ''/q;l 5 g" stands for ''/q;I 5 g a.e. on Substitute q ; - ry by qT, qz - rz by q; and regroup the terms to get

Define the residual

r2".

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

248 Then

+

= inf ( ~ ( 9 3r *, ) 2

Jr2

(9 I W

+ d w )d

~ )

14;/5g

This last estimate is combined with Theorem 6.5, leading to the next result.

T H E O R E6.6 M Let u E V be the unique solution of (6.1), and w E V an approximation of u. Then the following error bound

+

Iq;15g inf

+

{ ~ ( q ; , r * ) 1~2 ( g l w l+ q j w ) d s } (6.26)

'v'r*= ( r T , r $ )E Q: is valid, where the residual R ( q $ ,r * ) is defined by (6.25).

6.3.

RESIDUAL-BASED ERROR ESTIMATES FOR THE MODEL PROBLEM

With the special selection rT = - V w , rz = -w, (6.26) leads to the following error estimate: I,

\/;a(u--W,U-w)

< R,

(6.27)

where

Although it is possible to derive (6.27)-(6.28) through other approaches, we comment that Theorem 6.6 provides a general framework for various a posteriori error estimates with different choices of the auxiliary variable r*. In the limiting case g = 0, the problem (6.1) reduces to the variational equation U E V , a ( u , v ) = l ( v ) 'V'vEV. We observe that correspondingly, the estimate (6.27)-(6.28) reduces to the familiar form

/:

-a(u

-w,u -w)

J

5 sup v€V

I/vI/~

R

[(f

- W)V

- vw

. VVI dx,

Error Analysis for Variational Inequalities of the Second Kind

249

which is a starting point in deriving some a posteriori error estimators for Galerkin approximations of linear elliptic partial differential equations (cf. [2]). Now we focus on a posteriori analysis for the finite element solution error; that is, for the situation where the approximant w = uh is the finite element solution. By taking qi = -g Ah, and substituting v by vh - v in (6.28) we obtain

{L

1 R 5 sup V E VIIvllv

.

[ ~ u hV ( V

- vh) + (uh - f) ( v - vh)I dx

for any vh E Vh.Here we have used (6.12). For a given v E V, take vh = Dhv in (6.29), where Dhv is defined in (6.6). Decompose the integrals into local contributions from each element K E Ph and integrate by parts over K to obtain

Define interior residuals for each element K E Ph by

and side residuals for each side y E &h,0,r2

-- &h,O U Eh,J2 by

where the quantity

represents the jump discontinuity in the approximation to the normal derivative on the side y which separates the neighboring elements K and K t . By using

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

250

definitions (6.31) and (6.32), relation (6.30) reduces to

Using the estimates (6.9) and (6.10) in (6.33), and applying the CauchySchwarz inequality, we have

R 5 sup uEV llvllv

c lv11,n(

+

~ $ I I ~ K I I ~ , K hT1lR&Y) K E P ~ ?'EEh,o,rz

'

We summarize the above results in the form of a theorem.

THEOREM 6.7 Let u and respectively. Then the error

uh

be the unique solutions of (6.1) and (6.11), - uh satisjes the a posteriori estimate

eh = u

where r~ and R, are interior and side residuals, defined by (6.31) and (6.32), respectively. In practical computations, the terms on the right side of (6.35) are regrouped by writing

lehI:.

5 CV;,

~a

=

c +iK1

where the local error indicator V R , K on each element K, defined by

identifies contributions from each of the elements to the global error.

25 1

Error Analysis for Variational Inequalities of the Second Kind

In the last part of the section, we explore the efficiency of the error bound from (6.35) or (6.36). We derive an upper bound for the error estimator. Integrating by parts over each element and using (6.4) and Theorem 6.13 we have, for any v E V,

( A- uh + j)v dx + KEP,

YE&~,O

-

[g] Y

v ds

Thus,

where 7-K and R, are interior and side residuals defined for each element K E side y E &h,O,rzby (6.31) and (6.32), respectively. In order to simplify notation, we will omit the subscripts K and y. In the following, we will apply Theorems 6.3 and 6.4 choosing PK as the space of polynomials of degree less than or equal to 1, and 1 is any integer larger than or equal to the local polynomial degree of the finite element functions. Let T be a discontinuous piecewise polynomial approximation to the residual r , that is, TI E PK. Applying Theorem 6.3, we get

Phand each

Since the function v = $KT vanishes on the boundary aK,it can be extended to a function in V by 0 to the rest of the domain R. Inserting this extended function v in the residual equation (6.38), one obtains

Using this relation, we obtain

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

The terms on the right side of (6.40) are bounded by making use of the CauchySchwarz inequality and the second part of Theorem 6.3,

Combined with (6.39) we have

With the aid of the triangle inequality, finally we get

Consider now an interior side y E it follows that

IlRlI;,,

&h,O. From

IC

I

the first part of Theorem 6.4

$,R2ds.

(6.42)

Let 7denote the subdomain of fl consisting of the side y and the two neighbouring elements. The function v = $,R vanishes on 87 and as before it can be extended continuously to the whole domain R by 0 outside 7.With this choice of v the residual equation (6.38) reduces to

Therefore,

0

Applying the Cauchy-Schwarz inequality and the second part of Theorem 6.4 to the terms on the right side of (6.43), we obtain

which, combined with (6.42) and (6.41), implies that for every interior side

E &h,03 1lRlIo.i

1/2 -

( h ~ 1 i 2 ~ ~ e+hh7 ~ ~ \Iri , i-. r ~ ~ o ,. ? )

(6.44)

Finally, consider those sides y lying on E. Denote R E PK an approximation to the residual R = gXh on y, y E E (K). The first part of Theorem 6.4 implies

%+

Error Analysis for Variational Inequalities of the Second Kind

253

and let y be the element whose boundary contains Define the function v = the side y.Then = 0. Extend this function to the whole domain by zero value outside 7. The residual equation (6.38), with this choice of v , becomes

~

l

~

~

\

~

which leads to

As before, the first three terms on the right side of (6.46) can be bounded by applying Theorem 6.4 and the Cauchy-Schwarz inequality. Using (6.45), we then obtain for each side y E & h , r z ,

Multiplying this inequality by h7 and summing over all sides y E Ehlr2, we get

where

We can bound Rh,r2as follows,

254

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Use this bound in (6.47) and apply the Cauchy-Schwarz inequality to get

Combining (6.41) and (6.48), we finally conclude that

Summarizing (6.41), (6.44) and (6.49), we have

For our model problem, the element residual r K of (6.31) and the side residual Ry of (6.32), -nuh uh in K and d u h / d v on y are polynomials. Therefore, the terms llrK - - I O , K and 11 R, - %110, , in the right hand side of (6.50) can be replaced by 11 f - f K 1 1 0 , ~ and 11 Ah - Ah,? 1 0 , , with discontinuous piecewise polynomial approximations f K and Ah,y.

+

T H E O R E6.8 M Let 7~ be dejined as in (6.36). Then

with discontinuous piecewise polynomial approximations f K , Ah,y o f f , Ah.

Error Analysis for Variational inequalities of the Second Kind

255

Let us comment on the three summation terms in (6.51). As long as f has suitable degree of smoothness, the approximation error CKtPh hgll f f K / I f , , will be of higher order than 1 eh I & . Due to the inequality nature of the variational problem, in the efficiency bound (6.5 1) of the error estimator, there h, / X are extra terms involving X and Ah. A sharp bound of the term '&,h,r2

X h / & is currently an open problem. Nevertheless, in Section 6.5, we will present numerical results showing that the presence of this term in (6.51) does not have an effect on the efficiency of the error estimator. Similar numerical h-, l X h - X h j 7 1 f , , results can also be used as an evidence that the term CYeEh,r2 does not have an effect on the efficiency of the error estimator.

6.4.

GRADIENT RECOVERY-BASED ERROR ESTIMATES FOR THE MODEL PROBLEM

An important class of a posteriori error estimates is based on local or global averaging of the gradient, e.g, in the form of Zienkiewicz-Zhu gradient recovery technique, [165], [166], [167]. In [16], [37], Carstensen and Bartels proved that all averaging techniques provide us with a reliable a posteriori error control for the Laplace equation with mixed boundary conditions on unstructured grids as well. In the context of solving variational inequalities, gradient recovery type error estimates for elliptic obstacle problem have been derived recently in [17], [158]. In this section, we study a gradient recovery type error estimator for the finite element solution of the model problem, and we restrict our discussion to linear elements. In this case, the nodes are also the vertices. To formulate the error estimator we need a gradient recovery operator. There are many types of gradient recovery operators. In order to have a "good" approximation of the true gradient Vu, a set of conditions to be satisfied by the recovery operator were identified in [2]. These conditions lead to a more precise characterization of the form of the gradient recovery operator, summarized in Lemma 4.5 in [2]. In particular, the recovered gradient at a node a is a linear combination of the values of V u hin a patch surrounding a. We define the gradient recovery operator Gh: Vh + ( v ~ as follows: )~

Since linear elements are used,

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

256

where ( V v h )K i denotes the vector value of the gradient V v h On the element K;, Ka = K~,CX = ;I K ; I / K a , i = 1, . . . , Na. Recall from Section 6.3 that residual-type error estimates are derived by applying Theorem 6.6 with r* = - ( V u h , u h ) ,where uh is the finite element solution. In this section, we consider a different choice.

uF1

T H E O R E6.9 M Let u and uh be the unique solutions of (6.1) and (6.11), respectively. Then the error u - uh satisJies the a posteriori estimate

(6.54)

where

Proof. Let Ah E L m ( r 2 )be provided by Theorem 6.2. Apply Theorem 6.6 with w = uh and r* = -(Ghuh, u h )to obtain

where

1

{L

R = sup VEV

llullv

!Gnuh V r

+ (uh- f)u] dx +

/r2

ghhv ds}

.

(6.58)

Let Dh be the interpolation operator defined by (6.6). Use (6.12) with vh = nhv,

Therefore, we can rewrite (6.58) as

{ L [ G h ~.hV ( U nhv) + (uh - f )

1 R = sup V E V llullv

-

( U - nhu)l

dx

Error Analysis for Variational Inequalities of the Second Kind

[n (Ghuh

1 sup VEV

IIvIIv

-

V u h ). V n h v dx 5 C l(Vuh- Ghuhlo,~.

where

Integrate by parts over each element K E Ph to get

R1 = sup uEV

/ / " / /KvE p h

+

{k(-diV(~hUh)

Uh -

f ) (V

- n h ~dx )

Since Ghuh is continuous, the integrals on the interior sides y E Eh,O cancel each other. Write

-div(Ghuh)

+Uh - f

+

= div(Vuh - Ghuh) (-Auh

+

Uh -

and rearrange the terms in (6.60) to obtain

Rl= SUP -

div(Vuh - Ghuh)(V

-

DhV) dx

f)

25 8

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

+

where r K = -Auh uh - f = uh - f denotes the interior residual on element K E Fh. We will use r to denote the piecewise interior residual, i.e. r 1 = r K for K E Ph. To estimate the first summand on the right side of (6.61), we use an elementwise inverse inequality of the form

Apply the Cauchy-Schwarz inequality, the inverse inequality (6.62) and the estimate (6.9) to get

5

c

(z

u

-

G h h K )l

2

KEP,

(

)

h i l ( v - nhv)lli,K

KEPh

For the second summand, we apply the estimate (6.8) to obtain

Now

C

h i min llr ra ER

~ E N ~ , o

2

- rallo,za

C

52

hilluh

-

~ I l i , ~

~ E N ~ , o

+2

x

h i min \ I f - fall2 -

faER ~ E N ~ , o

O,Ka t

where %denotes the integral mean of uh over Fa. Use the Poincare inequality and an inverse inequality of the form (6.62) to get

Error Analysis for Variational Inequalities of the Second Kind

Therefore,

Finally, with the aid of the Cauchy-Schwarz inequality and the estimate (6.lo), the third summand on the right side of (6.61) can be bounded by

Inserting (6.63), (6.64) and (6.65) into (6.61), and using the Cauchy-Schwarz inequality and (6.59), we deduce that

Split the first term on the right side of estimate (6.57) into local contributions from each K E Ph and insert (6.66) to conclude the proof. We can write (6.54) as

where

We observe that usually the term Rh is of higher order compared to llu - uh 11 l,n which is of order O ( h ) in the nondegenerate situations. This observation is argued as follows: ~ i i s t it, is easy to show from the definition of the finite element solutions that there is a constant C such that IIuh 11 l,n C for any h. So the term

<

260

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

is bounded by 0 ( h 2 ) Next, . for f E L2(R),

and if f E

HI

(a), then

Thus, (6.67) illustrates the reliability of the error estimator 7 ~ . We now turn to the efficiency of the estimator. We will relate the gradient . that for recovery-based estimator 7~ to the residual-type estimator 7 ~Recall the residual type estimator,

with the interior residual

For the error estimator its efficiency.

LEMMA6.10 Let ~

TK

=

-nuh

+ uh - f

& = CKEPh &K,

G , K be

in K and the side residual

we have the inequality (6.51) for

dejined in (6.56). Then the following bound holds

where .& denotes the set of inner sides of the patch element K.

corresponding to the

Proof. It follows from definition of Gh that we have on each element K, I

Error Analysis for Variational Inequalities of the Second Kind

For any K t $nyxl

#

-

K , there is a sequence of inner edges yl, . . . ,,y, 0 a n d K c 5, K t c G. Hence,

c

26 1 such that

Since uh is continuous on 0, [ d u h / d t ] , = 0 for all y E &h,O,where d u h / d t is the tangential derivative of uh. Therefore, / [ V U h I y I = / [duh/dvIyI if y E &h,O. The estimates (6.70) and (6.71) together with the shape regularity of the partition Ph imply

-

~

(6.72) Let K E Ph be such that & ( K )n Ehtr2 = definitions of ~ ] G , Kand R, that

0. It follows from (6.72) and

Consider now the case when the element K has at least one side lying on the boundary r2.Let y E & ( K ) n &h,r2. Apply the triangle inequality to get

From an inverse inequality and (6.72) we have

Inserting (6.72) and (6.75) into (6.74) concludes the proof.

262

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

From Lemma 6.10 and the inequality (6.51) we obtain

The with discontinuous piecewise polynomial approximations fK and comments at the end of Section 6.3 apply to the three summation terms in (6.76).

6.5.

NUMERICAL EXAMPLE ON THE MODEL PROBLEM

We now present some numerical results on a two dimensional problem to show the effectiveness of the error estimators q~ of (6.36) and q~ of (6.55). Triangular partitioning and linear elements are used in the experiments. A seven-point Gauss-Legendre quadrature is used to compute the load vector on each triangle. Numerical integration over a general triangle is reduced that over a reference triangle through a linear transformation of the variables. On the reference triangle

the seven-point Gauss-Legendre quadrature formula is defined by

where the nodes {(ti, %))L1 and weights {wi)L1 are given in Table 6.1. The discretized solution is computed by solving the equivalent minimization problem using an over-relaxation method with a relative error tolerance, in the maximum norm, of (see [65], [66]). In order to show the effectiveness of the adaptive procedure we compare numerical convergence orders of the approximate solutions. We compute these orders by considering families of uniform and adaptively refined partitions. Consider a sequence of finite element solutions uEn based on uniform partitions of the domain 0. Starting with an initial coarse partition P I , we construct a family of nested meshes by subdividing each element into four congruent elements for two-dimensional case. The solution from the most refined mesh will be taken as the "exact" solution u that will be used to compute the errors of the approximate solutions obtained on the other meshes.

Error Analysis for Variational Inequalities of the Second Kind

263

Table 6.1. Nodes and weights of a 7-point Gauss-Legendre quadrature formula over the reference triangle

Adaptive finite element solutions are obtained by the following algorithm: 1 Start with the initial partition Ph and corresponding finite element subspace

Vh. 2 Compute the finite element solution utd E Vh. 3 For each element K E Ph compute the error estimator q ~defined , in (6.37) for the residual type and (6.56) for the gradient recovery type.

4 Let

7)

=

$ CK,?~ ?K

with N being the number of elements in partition

Ph. An element K is marked for refinement if q ~ > p q, where p is a prescribed threshold. In the example of this section, p = 0.5.

5 Perform refinement and obtain a new triangulation Ph. 6 Return to step 2. In the computation of the error indicator q ~ we make use of the multiplier Ah defined on r2c r. In what follows we describe how Ah can be (approximately) recovered from the solution uh using characterization (6.12). = l nodes of the partition Ph belonging to G, and let Denote by { a i ) z n the { 4 i ) z 1be the basis functions corresponding to the nodes { a i ) . We first determine a piecewise linear function

by requiring an analogue of (6.12):

264

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Denote X h , ~= ( x ; , ~. ,. . , the interval [ - I , I ] to get Xk.2 =

X T ~ )We~ then . project the components of X h j l onto 1 = ( x ~ ., . ~ . , X, T ~ ) ~ :

max{min{~;,,, l ) ,-I),

i

= 1 , . . . , m.

The piecewise linear approximation of the multiplier Ah on computed as

r2 c r can be

m

We briefly comment on the method for finding Ah,,. Let n = dim Vh.Denote by K the standard ( n x n) stiffness matrix and by 1 E Rn the standard load vector. Let u E Rn be the nodal value vector of the finite element solution uh. Then the algebraic representation of (6.77) becomes

where vc denotes the subvector of v containing the nodal values of vh at the nodes { a i ) ~c, F 2 , and M is a tridiagonal ( m x m ) matrix. We can write v = ( v r ,v : ) ~E RnPm x Rm by assuming that the components of v , are listed last. We similarly split 1 to li and 1,. This decomposition yields a block structure for K .

Then (6.79) is equivalent to the following two relations:

Once the approximate solution uh is computed, we can obtain from the second relation that Ah,, = M - ' ( z , - KCiui- K,,u,). In the numerical examples below, we use uKn for finite element solutions on uniform meshes, and utd for finite element solutions on adaptive meshes. Since adaptive solutions are involved, numerical solution errors will be plotted against the number of degrees of freedom, rather than the meshsize.

E X A M P L E6 . 1 1 Let R = [O, 11 x [ O , 1 ] and F2 = r. The problem solved is: such ) that Find u E ~ ' ( 0

Error Analysis for Variational Inequalities of the Second Kind

Figure 6.1. Example 6.11, true solution

where

andfori = 1 , 2 ,

wi(x)=

{

(

1 - I ) ) if ri < 1, otherwise

(4 2 ( x z - x2,0) ( 2 ) 2 112 with ri = [ ( X I - x,,,) ] / ~ i In the experiments we let g = 1, xjt! = 0.8, x$t! = 0.1,

+

x(2) 2,0 -

= 0.3,

0.1, E I = 0 . 2 5 , ~ z= 0.2. We start with a coarse uniform triangulation shown on the left plot in Figure 6.2. Here, the interval 10, I ] is divided into l l h equal parts with h = 114 which is successively halved. The numerical solution corresponding to h = 11256 is taken as the "true" solution u, shown in Figure 6.1. We use the regular refinement technique (red-blue-green refinement), where the triangle is divided into four triangles by joining the midpoints of edges and adjacent triangles are refined in order to avoid hanging nodes. For a detailed description of this and other refinement techniques currently used see e.g. [I531 and references therein. Also, in order to improve the quality of triangulation, a

266

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 6.2. Example 6.11, initial partition and adaptively refined partition after 5 iterations

smoothing procedure is used after each refinement. For each triangle K of the triangulation we compute the triangle quality measure defined by

where hi, i = 1 , 2 , 3 , are the side lengths of the triangle K. Note that Q ( K ) = 1 if hl = h2 = h3. A triangle is viewed to be of acceptable quality if Q > 0.6, otherwise we modify the mesh by moving the interior nodes toward the center of mass of the polygon formed by the adjacent triangles. The adaptively refined triangulation after 5 iterations is shown on the right plot of Figure 6.2. To have an idea of the convergence behavior of the discrete Lagrange multipliers, we analyze the errors 11 X - Ah Il o , r corresponding to the sequence of uniform refinements. Here, X is the Lagrange multiplier corresponding to the parameter h = 11256. Figure 6.3 provides a comparison between the errors (ju - uKn11 l,n and h1I2( ( A - Ah jjO,r. The numerical convergence order of h1I211 X - Ah is obviously higher than that of Ilu - uXn 11 l , ~indicating , that the second term in the efficiency bounds (6.51) and (6.76) is expected to be of Graphs of Xh,1 and higher order compared to the first term Jju - uh with h = 11256 are provided in Figures 6.4 and 6.5. We use an adaptive procedure based on both residual type and gradient recovery type estimates to obtain a sequence of approximate solutions uzd. The adaptive finite element mesh after 5 adaptive iterations is shown on the right plot in Figure 6.2. The value ( / u ( (is~about , ~ 1.3228. Figures 6.6 and 6.7 contain the error values 11 u - uXn 11 1,a and I / u - uidI 11,a. We observe a substantial improvement of the efficiency using adaptively refined meshes. Figures 6.8 and 6.9 provide the values of 71 = 7):,K)1/2,I E { R ,G), where ~ I , K are computed using either residual type estimator (I = R) or gradient recov-

(C,

Error Analysis for Variational Inequalities of the Second Kind

1o2 1o3 Number of degrees of freedom

Figure 6.3. Example 6.11, 1

1o4

1 -~ u;fnlli,n (0)VS. h 1 ' 2 / I ~- Xh/10,r(A)

Figure 6.4. Example 6.11, plots of

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY Side: ( l ) x [ O , l ]

Side [ O , l h ( O )

0 -0 5 -1 -1.5

0

0.2

0.4

06

0.8

1

-1.5

0

0.2

04

0.6

08

Y

x

0.3 Side, (O)x[O,l]

Figure 6.5. Example 6.11, plots of

Table 6.2. Example 6.11, numerical values of constants

ery type estimator (I = G) on both uniform and adapted meshes. Table 6.2 contains the values of CI computed for uniform and adaptive solutions:

It is seen from Table 6.2 that for this numerical example, the gradient recovery type error estimator provides a better prediction of the true error than the residual type error estimator, a phenomenon observed in numerous references. In general, we use the a posteriori error estimates only for the purpose of designing adaptive meshes, due to the presence of the unknown constants CR and CG

Error Analysis for Variational Inequalities of the Second Kind

1o

-~ 10'

1o3 Number of degrees of freedom

1o2

I O*

Figure 6.6. Example 6.11, results based on residual type estimator, Ilu - u;tnlli,n ( 0 ) vs. 11u - uahdIl1,n(A)

1o

-~ 10'

I02 1o3 Number of degrees of freedom

I

10'

Figure 6.7. Example 6.11, results based on gradient recovery type estimator, Ilu - ~:"lll,n (0)vs. llu - u"hl11,n(4

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

4adapted mesh error adapted mesh I),

J 10'

Io2 1o3 Number of degrees of freedom

1o4

Figure 6.8. Example 6.11, results based on residual type estimator, 1)u- ~ ; l ~ and J l V~R ,on~ ~ V R on adapted mesh (A) uniform mesh (o) vs. ilu - u i d11 1 , and

:

10'

4adapted mesh error

1

1O-Z 10'

1oZ 10' Number of degrees of freedom

1 o4

Figure 6.9. Example 6.11, results based on gradient recovery type estimator, ilu and VG on uniform mesh ( 0 ) vs. Ilu - u P I I ~and , ~76 on adapted mesh (A)

~ x ~ l l ~ , ~

Error Analysis for Variational Inequalities of the Second Kind 10' adapted mesh (residual) 4- adapted mesh (gradient recovery) 4

10-21 10'

102 10' Number of degrees of freedom

1o4

Figure 6.10. Example 6.1 1, performance comparison between the two error estimators

For comparison of the performance between the two error estimators, we show in Figure 6.10 the errors of the adaptive solutions corresponding to the two error estimators. We observe that the two error estimators lead to very similar solution accuracy for same amount of degrees of freedom. More numerical examples can be found in [26].

6.6.

APPLICATION TO A FRICTIONAL CONTACT PROBLEM

In this section, we take the frictional contact problem discussed in Example 1.27 (page 32) to show that similar a posteriori error estimates can be derived for more complicated variational inequalities. Recall that R c IRd ( d 5 3 in applications) a Lipschitz domain. The unit outward normal vector v exists a.e. on the boundary r . We will use the product space ~ ' ( 0 = ()H I equipped with the norm d I : , n . Similarly, for w C R or w C I?, we use the space I V Il:,n = ~ ~ (=w( L) ~ ( w equipped ))~ with the norm lvl$,, = /vil;,,. Recall that the material is linearly elastic the fourth-order elasticity tensor @ : R x Sd -+ S d is assumed to be bounded, symmetric and positive definite in - - R. The boundary I? is partitioned as I? = I?D U F N U I?c with r D ,rNand rc relatively open and mutually disjoint, and meas(rD) > 0. The body is clamped on I'D; on the boundary part rNsurface tractions of density f E ( L ~ ( I ? N ) ) ~

27 2

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

are applied and on rc the body is in bilateral contact with a rigid foundation. The contact is frictional and is modeled by Tresca's law. Volume forces of ) ) in ~ R. density f E ( L ~ ( R act The function space for the displacement variable is V = {V E H ' ( R ) : vlr, = 0, vvlrc = O), with the inner product and norm

Over the space V , we use the following functionals:

a ( u , v )=

S,

C E ( U .) ~ ( vdx, )

Then the variational inequality of the frictional contact problem is to find u E V such that

This problem has a unique solution u in V , which is the minimizer of the energy functional 1 over the space V . In the analysis of a posteriori error estimators later, we will need the following characterization of the solution u of (6.80) through the use of a Lagrange multiplier: The unique solution u E V of the problem (6.80) is characterized by the existence of A, E (L" ( I ' c ) ) dsuch that

Turn now to a finite element approximation of the problem. Assume R has a polyhedral boundary I?. Let { P h }be a family of partitions of the domain R into straight-sided elements, and let V h C V be the associated standard finite element spaces of continuous piecewise polynomials of certain degree.

Error Analysis for Variational Inequalities of the Second Kind

273

Corresponding to the partition P h , we use the symbols h K ,h,, E ( K ) , Eh, Eh,O as introduced in Section 6.1. In addition, we use &h,rN,EhJc with obvious meanings. The discrete formulation of the variational inequality (6.80) reads: Find u h E V h such that Like the continuous variational inequality (6.80), the discrete problem (6.83) has a unique solution. Analysis of the finite element approximation of such problems in the general context of variational inequalities is extensively discussed in [65], [66]. In the context of finite element approximations of a problem more general than the one considered in this paper, one can find in [81, Section 8.21 an optimal order a priori error estimate under additional solution regularity, and a convergence result without any additional solution regularity assumption. For derivation of a posteriori error estimates, we need the following characterization of the discrete solution through a Lagrange multiplier: The unique solution uh E V h of the problem (6.83) is characterized by the existence of ~ that ) ) ~ Xhr E ( ~ ~ ( rsuch

where vh, and u h , denote the tangential components of vh and uh, respectively. Then similar to the result in Section 6.3, we have the residual-type error estimator: Ilehll", c'&, where

r K and Ry are the interior and side residuals, respectively, defined by ( a h =

c.5( ~ h ) ) :

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 6.1I .

Example 6.12, physical setting

Moreover,

with discontinuous piecewise polynomial approximations f l , K , Xhr,r of residuals f l, Ah7, respectively. The results in Section 6.4 on gradient recovery-based error estimators can be similarly extended to the finite element solution of the frictional contact problem (6.80). Now we present numerical results on two two-dimensional problems. In all the examples no body forces are assumed to act and the body is in bilateral frictional contact with a rigid foundation on the part rc. The friction is modeled by the Tresca law with a given slip bound g. We use the adaptive algorithm stated at the beginning of Section 6.5, with p = 1.

EXAMPLE 6.12 We consider the physical setting shown in Figure 6.11. The domain R = (0,4) x (0,4) is the cross section of a three-dimensional linearly = (4) x elastic body and plane stress condition is assumed. On the part (0,4) the body is clamped. Oblique tractions act on the part (0) x (0,4) and the part (0,4) x (4) is traction free. Thus rrv= ((0) x (0,4))U ((0,4) x (4)). The contact part of the boundary is Fc = (0,4) x (0). The elasticity tensor C satisfies

rD

Error Analysis for Variational Inequalities of the Second Kind

Figure 6.12. Example 6.12, initial mesh with 128 elements, 81 nodes

where E is the Young's modulus, u is the Poisson's ratio of the material and bij is the Kronecker symbol. We use the following data (the unit daN/rnm2 stands for "decanewtons per square milimeter"):

We start with the initial triangulation Pl shown in Figure 6.12. Here the interval [O, 41 is divided into 4 / h equal parts with h = 112 which is successively halved. The numerical solution corresponding to h = 1/64 is taken as the "exact" solution u . To have an idea of the convergence behavior of the discrete Lagrange multipliers, we compute the errors 1 A, - Ah, corresponding to the sequence of uniform refinements. Here, A, is the Lagrange multiplier corresponding to the parameter h = 1/64. Figure 6.13 provides a comparison of the errors )/ u - uxn11 and h1I2)I X I - Xhr I ojrc. The numerical convergence order of h1I2/Ix, - Ahrllo,rc is obviously higher than that of Ilu - uinIlv, indicating that the second term in the efficiency bound is expected to be of higher order compared to the first term llu - u h II$. Note that in the two-dimensional case, A, and Ah, are scalar valued functions, to be also denoted by X and Ah. The (approximate) calculation of Ah follows the procedure described in Section with h = 1/64 are provided in Figures 6.14 and 6.5. Graphs of X h , ~and 6.15. We use an adaptive procedure based on both residual type and gradient recovery type estimates to obtain a sequence of approximate solutions u i d . The deformed configuration and the adaptive finite element mesh after 4 adaptive iterations are shown in Figures 6.16 (residual type estimator) and 6.17 (gradient recovery type estimator). Figures 6.18 and 6.19 contain the error values / I u - u$IJ and ) / u- uidlJ v . We observe a substantial improvement of the

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

1o

-~ 1o2

1o3 Number of degrees of freedom

Figure 6.14. Example 6.12, plot of

10'

Xh,1

efficiency using adaptively refined meshes. Figures 6.20 and 6.21 provide the values of 71 = ( C , 1):,,)1/2, I E { R , G I , where 7 1 , are ~ computed using either residual type estimator ( I = R) or a gradient recovery type estimator H ( I = G ) on both uniform and adapted meshes.

EXAMPLE 6.13 Consider the physical setting shown in Figure 6.22. The domain R = ( 0 , l O ) x ( 0 , 2 ) is the cross section of a three-dimensional linearly elastic body with plane stress condition assumed. On the part rD= ( 0 , l O ) x

Error Analysis for Variational Inequalities of the Second Kind

421

0

05

I

15

2

25

3

Figure 6.15. Example 6.12, plot of 45

1

35

4

Xh,2

I

Figure 6.16. Example 6.12, results based on residual type estimator, deformed configuration 5583 elements. 2921 nodes

(2) the body is clamped. Horizontal tractions act on the part (0) x (0,2) and oblique tractions on (10) x (0,2). Here rc = (0,lO) x (0). The following data are used:

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 6.17. Example 6.12, results based on gradient recovery type estimator, deformed configuration with 5437 elements, 2832 nodes

J

1o2

Io3 Number of degrees of freedom

1o4

Figure 6.18. Example 6.12, results based on residual type estimator, Iu - u;tnllv(o) vs. llu - uPIIv (A)

As before we start with a coarse uniform triangulation Ph consisting of 160 triangular elements and 105 nodes depicted in Figure 6.23. The interval [O, 101 is divided into 1 0 / h equal parts and interval [O,21 is divided into 4 / h parts with h = 112. The numerical solution corresponding to h = 1/64 is taken as the

Error Analysis for Variational Inequalities of the Second Kind

1o0 -A- adapted mesh

1o3 Number of degrees of freedom

Figure 6.19. Example 6.12, results based on gradient recovery type estimator, Ilu - uinliv ( 0 ) VS. IIu - uahdllv (A)

Number of degrees of freedom

Figure 6.20. Example 6.12, results based on residual type estimator, Ilu - uXn uniform mesh ( 0 ) vs. llu - uidI/ and 1 7 on ~ adapted mesh (A)

and 1 7 on ~

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

4adapted mesh error adapted mesh q,

1o2

1o3 Number of degrees of freedom

1o4

Figure 6.21. Example 6.12, results based on gradient recovery type estimator, llu - u;lnllV and 1 7 on ~ uniform mesh (o) vs. ilu - U? dlV and 176. on adapted mesh (A)

-1

rigid obstacle

Figure 6.22. Example 6.13, physical setting

"exact" solution u.Figures 6.24 (residual type) and 6.25 (gradient recovery type) show the approximate solution and refined mesh after 4 consecutive refinements. Graphs of Xh,l and with h = 1/64 are provided in Figures l u - uKnllv, 6.26 and 6.27. Again, we compute the errors I h 1 1 2 ) I~T X h 7 ) ) ~ ; rand C 711, I E {R, G), whose values are provided in Figures 6.28-6.32.

Iu

~ ~ ~ l l ~

Error Analysis for Variational Inequalities of the Second Kind

28 1

w Figure 6.23. Example 6.13, initial mesh with 160 elements, 105 nodes

Fkure 6.24. Example 6.13, results based on residual type estimator, deformed configuration with 5222 elements, 2755 nodes

Figure 6.25. Example 6.13, results based on gradient recovery type estimator, deformed configuration with 4964 elements, 2624 nodes

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

Figure 6.26. Example 6.13, plot of Ah,]

Figure 6.27. Example 6.13, plot of

Ah,2

Error Analysis for Variational Inequalities of the Second Kind

1o0

+ uniform mesh 4adapted mesh

10-21 1

o2

I o3

1o4

Number of degrees of freedom

Figure 6.28. Example 6.13, results based on residual type estimator, Ilu - u;tnilv ( 0 ) vs. llu - uahdIIv (A)

+ uniform mesh 4- adapted mesh

10-21 1o2

1o3

1o4

Number of degrees of freedom

Figure 6.29. Example 6.13, results based on gradient recovery type estimator, Ilu - u;tnllv ( 0 ) VS. Ilu - uahdIIv (A)

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

+ solution -A- lambda

- ~L

1o

1o2

1o3 Number of degrees of freedom

1o4

1o3 Number of degrees of freedom

10'

10.'

adapted mesh error

1o

-~ 102

Figure 6.31. Example 6.13, results based on residual type estimator, Jlu- uznllVand 1 7 on ~ uniform mesh ( 0 ) vs. 1 lu - uRd/I and 1 7 on ~ adapted mesh (A)

I

1

S9.0

% qsau umjun mia qsau uiojlun -s

Ei===

References

[I] R.A. Adams, Sobolev Spaces, Academic Press, New York, 1975. [2] M. Ainsworth and J.T. Oden, A Posteriori Error Estimation in Finite Element Analysis, John Wiley & Sons, Inc., New York, 2000. [3] M. Ainsworth, J.T. Oden, and C.Y. Lee, Local aposteriori error estimators for variational inequalities, Numer: Meth. PDE 9 (1993), 23-33. [4] J. Alberty, C. Carstensen and D. Zarrabi, Adaptive numerical analysis in primal elastoplasticity with hardening, Comp. Meths. Appl. Mech. Engng. 171 (1999), 175-204. [5] A.M. Arthurs, Complementary Variational Principles, 2nd Edition, Oxford Math. Monographs, Oxford, 1980.

[6] K. Atkinson and W. Han, Theoretical Numerical Analysis: A Functional Analysis Framework, Springer-Verlag, New York, TAM, Volume 39, 2001. [7] I. BabuSka, The problem of modeling the elastomechanics in engineering, Computer Methods in Applied Mechanics and Engineering 82 (1990), 155-182. [8] I. BabuSka and A.K. Aziz, Survey lectures on the mathematical foundations of the finite element method, in A.K. Aziz, ed., The Mathematical Foundations of the Finite Element Method with Applications to Partial DifSerential Equations, Academic Press, New York, 1972, 3-359. [9] I. BabuSka, Y. Li, and K.L. Jerina, Reliability of computational analysis of plasticity problems, Technical Note BN-1123, IPST, UMCP, March 1991.

288

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

[lo] I. BabuSka and J. Pitkaranta, The plate paradox for hard and soft simple support, SIAM J. Math. Anal. 21 (1990), 55 1-576. [11] I. BabuSka and W.C. Rheinboldt, Error estimates for adaptive finite element computations, SIAM J. Numec Anal. 15 (1978), 736-754. [12] I. BabuSka and W.C. Rheinboldt, Aposteriori error estimates for the finite element method, Intern. J. Numer: Methods Engrg. 12 (1978), 15971615. [13] I. BabuSka and T. Strouboulis, The Finite Element Method and its Reliability, Oxford University Press, 2001. [14] J. Baranger and K. Najib, Analyse numCrique des Ccoulements quasiNewtonians dont la viscositC obCit i la loi puissance ou la loi de Carreau, Numer: Math. 58 (1990), 35-49. [15] J.W. Barrett and W.B. Liu, Finite element error analysis of a quasiNewtonian flow obeying the Carreau or power law, Numer: Math. 64 (1993), 433453. [16] S. Bartels and C. Carstensen, Each averaging technique yields reliable a posteriori error control in FEM on unstructured grids. Part 11: Higher order FEM, Math. Comp. 71 (2002), 97 1-994. [17] S. Bartels and C. Carstensen, Averaging techniques yield reliable a posteriori finite element error control for obstacle problems. Preprint Nr. 212001. Publications of the Max-Planck-Institute for Mathematics in the Sciences, Leipzig, Germany. [I81 H.S. Bear, An Introduction to Mathematical Analysis, Academic Press, San Diego, 1997. [19] R. Becker and R. Rannacher, An optimal control approach to a posteriori error estimation in finite element methods, in Acta Numerica 2001, Cambridge University Press, 1-102. [20] T. Belytschko, W.K. Liu and B. Moran, Nonlinear Finite Elements for Continua and Structures, Wiley, Chichester, England, 2000. [21] C. Bernardi and V. Girault, A local regularization operator for triangular and quadrilateral finite elements, SIAM J. Numer: Anal. 35 (1998), 18931916. [22] R.B. Bird, R.L. Armstrong and 0. Hassager, Dynamics of Polymeric Liquids, Vol. 1, Fluid Mechanics, Wiley, 1977.

REFERENCES

289

[23] H. Blum and R. Rannacher, On the boundary value problem of the biharmonic operator on domains with angular comers, Math. Meth. in the Appl. Sci. 2 (1980), 556-581. [24] H. Blum and F.-T. Suttmeier, An adaptive finite element discretisation for a simplified Signorini problem, CALCOLO 37 (2000), 65-77. [25] H. Blum and F.-T. Suttmeier, Weighted error estimates for finite element solutions of variational inequalities, Computing 65 (2000), 119-134. [26] V. Bostan and W. Han, Recovery-based error estimation and adaptive solution of elliptic variational inequalities of the second kind, to appear in Communications in Mathematical Sciences. [27] V. Bostan and W. Han, On a posteriori error analysis of a quasistatic variational inequality, submitted. [28] V. Bostan, W. Han and B. D. Reddy, A posteriori error analysis for elliptic variational inequalities of the second kind, in Computational Fluid and Solid Mechanics 2003, Proceedings of Second MIT Conference on Computational Fluid and Solid Mechanics, June 17-20, 2003, ed. K.J. Bathe, 1867-1870. [29] V. Bostan, W. Han and B. D. Reddy, A posteriori error estimation and adaptive solution of elliptic variational inequalities of the second kind, submitted. [30] M. Braack and A. Em, A posteriori control of modeling errors and discretization errors, Multiscale Model. Simul. 1 (2003), 221-238. [31] D. Braess, Finite Elements: Theory, Fast Solvers, and Applications in Solid Mechanics, second edition, Cambridge University Press, Cambridge, 2001. [32] S.C. Brenner and L.R. Scott, The Mathematical Theory of Finite Element Methods, second edition, Springer-Verlag, New York, 2002. [33] F. Brezzi and M. Fortin, Mixed and Hybrid Finite Element Methods, Springer-Verlag, Berlin, 1991. [34] C. Carstensen, Numerical analysis of the primal problem of elastoplasticity with hardening, Nuinel: Math. 82 (1999), 577-597. [35] C. Carstensen, Quasi-interpolation and a posteriori analysis in finite element methods, RAIRO Math. Model. Num. 33 (1999), 1187-1202.

290

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

[36] C. Carstensen and J. Alberty, Averaging techniques for reliable a posteriori FE-error control in elastoplasticity with hardening, Comput. Methods Appl. Mech. Engrg. 192 (2003), 1435-1450. [37] C. Carstensen and S. Bartels, Each averaging technique yields reliable a posteriori error control in FEM on unstructured grids. Part I: Low order conforming, nonconforming, and mixed FEM, Math. Comp. 71 (2002), 945-969. [38] C. Carstensen and R. Verfiirth, Edge residual dominate a posteriori error estimates for low order finite element methods, SIAM J. Numec Anal. 36 (1999), 1571-1587. [39] J. Chen, W. Han, and H. Huang, On the KaEanov method for a quasiNewtonian flow problem, Numerical Functional Analysis and Optimization 19 (1998), 961-970. [40] Z. Chen and R.H. Nochetto, Residual type a posteriori error estimates for elliptic obstacle problems, Numec Math. 84 (2000), 527-548. [41] P.G. Ciarlet, The Finite Element Method for Elliptic Problems, North Holland, Amsterdam, 1978. [42] P.G. Ciarlet, Mathematical Elasticity, Vol. I: Three-Dimensional Elasticity, North-Holland, Amsterdam, 1988. [43] P.G. Ciarlet, Basic error estimates for elliptic problems, in P.G. Ciarlet and J.-L. Lions, eds., Handbook of Numerical Analysis, Vol. ]I, NorthHolland, Amsterdam, 1991, 17-35 1. [44] Ph. ClCment, Approximation by finite element functions using local regularization, RAIRO Numerical Analysis R-2 (1975), 77-84. [45] J. Conway, A Course in Functional Analysis, 2nd ed., Springer-Verlag, New York, 1990. [46] M. Dauge, Elliptic Boundary Value Problems in Corner DomainsSmoothness andAsymptotics of Solutions, Lecture Notes in Mathematics, No. 1341, Springer-Verlag, 1988. [47] G. Duvaut and J.-L. Lions, Inequalities in Mechanics and Physics, Springer-Verlag, Berlin, 1976. [48] R.E. Edwards, Functional Analysis: Theory and Applications, Dover Publications, Inc., New York, 1994.

REFERENCES

29 1

[49] I. Ekeland and R. Temam, Convex Analysis and Variational Problems, North-Holland, Amsterdam, 1976. Reprinted as SIAM's Classics in Applied Mathematics, Volume 28, 1999. [50] K. Eriksson, D. Estep, P. Hansbo, and C. Johnson, Introduction to adaptive methods for differential equations, in Acta Numerica 1995, Cambridge University Press, 105-1 58. [51] L.C. Evans, Partial Differential Equations, AMS, 1998. [52] R.S. Falk, Error estimates for the approximation of a class of variational inequalities, Math. Comp. 28 (1974), 963-97 1. [53] M. Feistauer, J. Mandel, and J. NeEas, Entropy regularization of the transonic flow problem, Comment. Math. Univ. Carolin. 25 (1984), 43 1443. [54] V.M. Filippov, Variational Principles for Nonpotential Operators, Translations of Mathematical Monographs, Vol. 77, AMS, 1989. [55] M. Fortin, Old and new finite elements for incompressible flows, Int. J. Numel: Meth. Fluids. 1 (1981), 347-364. 1-56] D. French, S. Larsson, and R.H. Nochetto, Pointwise a posteriori error analysis for an adaptive penalty finite element method for the obstacle problem, Comput. Methods Appl. Math. 1 (2001), 18-38. [57] A. Friedman, Variational Principles and Free-boundary Problems, John Wiley & Sons, New York, 1982. 1581 S. Furik, A. Kufner, and J. NeEas, KaEanov's method and its application, Rev. Roumaine de Math. Pures et Appl., 20 (1975), 907-9 16. [59] D.Y. Gao, Duality Principles in Nonconvex Systems, Kluwer Academic Publishers, Dordrecht, 1999. [60] D.Y. Gao, Nonconvex semi-linear problems and canonical duality solutions, Chapter 5 of Advances in Mechanics and Mathematics 11, 2003, eds. D.Y. Gao andR.W. Ogden, Kluwer Academic Publishers, Dordrecht, pp. 261-312. [61] D. Gilbarg and N. Trudinger, Elliptic Differential Equations of Second Order, Springer-Verlag, 1977. [62] M.B. Giles and E. Siili, Adjoint methods for PDEs: a posteriori error analysis and postprocessing by duality, in Acta Numerica 2002, Cambridge University Press, 145-236.

292

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

[63] V. Girault and P.-A. Raviart, Finite Element Methods for Navier-Stokes Equations, Theory and Algorithms, Springer, Berlin, 1986. [64] H. Gittel, Studies on transonic flow problems by nonlinear variational inequalities, Z. Anal. Anwendungen 6 (1987), 449-458. [65] R. Glowinski, Numerical Methods for Nonlinear Variational Problems, Springer-Verlag, New York, 1984. [66] R. Glowinski, J.-L. Lions, and R. Trkmolikes, Numerical Analysis of Variational Inequalities, North-Holland, Amsterdam, 1981. [67] P. Grisvard, Elliptic Problems in Nonsmooth Domains, Pitman, Boston, 1985. [68] W. Han, The best constant in a Sobolev inequality, Applicable Analysis An International Journal 41 (1991), 203-208. [69] W. Han, Finite element analysis of a holonomic elastic-plastic problem, Numel: Math. 60 (1992), 493-508. [70] W. Han, The best constant in a trace inequality, Journal of Mathematical Analysis and Applications 163 (1992), 512-520. [7 11 W. Han, A posteriori error analysis for linearizations of nonlinear elliptic problems and their discretizations, Math. Meths. Appl. Sci. 17 (1994), 487-508. [72] W. Han, Quantitative error estimates for idealizations in linear elliptic problems, Math. Meths. Appl. Sci. 17 (1994), 971-987. [73] W. Han, Notes on best constants in some Sobolev's inequalities, Analysis Mathematica 20 (1994), 3-10. [74] W. Han and S. Jensen, The KaEanov method for a nonlinear variational inequality of the second kind arising in elastoplasticity, Chinese Annals of Mathematics 17B (1996), 129-138. [75] W. Han, S. Jensen, and B.D. Reddy, Numerical approximations of internal variable problems in plasticity: error analysis and solution algorithms, Numerical Linear Algebra with Applications 4 (1997), 191-204. [76] W. Han, S. Jensen, and I. Shimansky, The KaEanov method for some nonlinear problems, Applied Numerical Analysis 24 (1997), 57-79. [77] W. Han and B.D. Reddy, On the finite element method for mixed variational inequalities arising in elastoplasticity, SIAM J. Numel: Anal. 32 (1995), 1778-1807.

REFERENCES

293

[78] W. Han and B.D. Reddy, Computational plasticity: the variational basis and numerical analysis, Computational Mechanics Advances 2 (1995), 283-400. [79] W. Han and B.D. Reddy, Plasticity: Mathematical Theory and Numerical Analysis, Springer-Verlag, 1999. [80] W. Han, B.D. Reddy, and G.C. Schroeder, Qualitative and numerical analysis of quasistatic problems in elastoplasticity, SIAM J. Numer: Anal. 34 (1997), 143-177. [811 W. Han and M. Sofonea, Quasistatic Contact Problems in Viscoelasticity and Viscoplasticity, American Mathematical Society and International Press, 2002. [82] J. Haslinger, I. HlaviEek and J. NeEas, Numerical methods for unilateral problems in solid mechanics, in P.G. Ciarlet and J.-L. Lions, eds., Handbook of Numerical Analysis, Vol. IV, North-Holland, Amsterdam, 1996, 313-485. [83] I. HlaviEek, J. Haslinger, J. NeEas and J. LoviSek, Solution of Variational Inequalities in Mechanics, Springer-Verlag, New York, 1988. [84] R.H.W. Hoppe and R. Kornhuber, Adaptive multilevel methods for obstacle problems, SIAM J. Numer: Anal. 31 (1994), 301-323. [85] C.O. Horgan, Eigenvalue estimates and the trace theorem, J. of Math. Anal. and Appl. 69 (1979), 231-242. [86] C.O. Horgan and L.E. Payne, Lower bounds for free membrane and related eigenvalues, Rendiconti di Matematica 10 (1990), 457491. [87] H. Huang, W. Han, and J. Zhou, The regularization method for an obstacle problem, Numer: Math. 69 (1994), 155-166. [88] T.J.R. Hughes, The Finite Element Method, Prentice Hall Inc., New Jersey, 1987. [89] C. Johnson, Numerical Solutions of Partial Differential Equations by the Finite Element Method, Cambridge University Press, Cambridge, 1987. [90] C. Johnson, Adaptive finite element methods for the obstacle problem, Math. Models Methods Appl. Sci. 2 (1992), 483487. [91] J. KaE6r, J. NeEas, J. Polik, and J. SouEek, Convergence of a method for solving the magnetostatic field in nonlinear media, Aplikace Maternatiky 13 (1968), 456465.

294

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

[92] R.B. Kellogg and J.E. Osborn, A regularity result for the Stokes problem in a convex polygon, J. Functional Anal. 21 (1976), 3 9 7 4 3 1. [93] A.M. Khludnev and J. Sokolowski, Modelling and Control in Solid Mechanics, Birkhauser Verlag, Basel, 1997. [94] N. Kikuchi and J.T. Oden, Contact Problems in Elasticity: A Study of Variational Inequalities and Finite Element Methods, SIAM, Philadelphia, 1988. [95] D. Kinderlehrer and G. Stampacchia, An Introduction to Variational Inequalities and their Applications, Academic Press, New York, 1980. [96] V.A. Kondrat'ev, Boundary problems for elliptic equations in domains with conical or angular points, Trans. Moscow Math. Soc. 16 (1967), 227-3 13. [97] V.A. Kondrat'ev and O.A. Oleinik, Boundary-value problems for partial differential equations in non-smooth domains, Russian Math. Surveys 38 (1983), 1-86. [98] R. Komhuber, A posteriori error estimates for elliptic variational inequalities, Comput. Math. Appl. 31 (1996), 49-60. [99] V.A. Kozlov, V.G. Maz'ya, and J. Rossmann, Elliptic Boundary Value Problems with Point Singularities, AMS, Provindence, 1997. [loo] V.A. Kozlov, V.G. Maz'ya, and J. Rossmann, Spectral Problems Associated with Corner Singularities of Solutions to Elliptic Equations, AMS, Provindence, 2001. [ l o l l A. Kufner, 0 . John, and S. FuEik, Function Spaces, Noordhoff International Publishing, A division of A.W. Sijthoff International Publishing Company B .V., Leyden, The Netherlands, 1977. [I021 A. Kufner and A. Sangig, SomeApplications of Weighted Sobolev Spaces, Teubner-Texte zur Mathematik, Band 100, Leipzig, 1987. [I031 J.R. Kuttler and V.G. Sigillito, Estimating Eigenvalues with A Posteriori / A Priori Inequalities, Pitman Research Notes in Mathematics Series, Volume 135, London, 1985. [I041 N.N. Lebedev, Special Functions & Their Applciations, Dover Publications, Inc., New York, 1972. [lo51 R. Lehmann, Developments at an analytic comer of solutions of elliptic partial differential equations, J. Math. Mech. 8 (1959), 727-760.

REFERENCES

295

[106] J.E. Marsden and T.J.R. Hughes, Mathematical Foundations of Elasticity, Prentice-Hall, Englewood Cliffs, New Jersey, 1983. [107] V.G. Maz'ya, N.F. Morozov, B.A. Plamenevskii, and L. Stupyalis, Elliptic Boundary Value Problems, AMS Translations, Series 2, Volume 123, 1984. [lo81 E. Miersemann, Asymptotic expansions of solutions of the Dirichlet problem for quasilinear elliptic equations of second order near a conical point, Math. Nachl: 135 (1988), 239-274. [I091 S.G. Mikhlin, Variational Methods in Mathematical Physics, Oxford, New York, Pergamon Press, 1964. [I101 N.I. Muskhelishvili, Some Basic Problems of the Mathematical Theory of Elasticity, P. Noordoff Ltd, Groningen-Holland, 1953. [I 111 J. NeEas and I. HlavBEek, Mathematical Theory of Elastic and ElasticoPlastic Bodies: An Introduction, Elsevier Scientific Publishing Company, Amsterdam, Oxford, New York, 1981. [I121 J. NeEas and I. HlavBEek, Solution of Signorini's contact problem in the deformation theory of plasticity by secant modules method, Aplikace Matematiky 28 (1983), 199-214. [I 131 B. Noble and M.J. Sewell, On dual extremum principles in applied mathematics, J. Inst. Math. Appl. 9 (1972), 123-193. [114] R.H. Nochetto, K.G. Siebert, and A. Veeser, Pointwise a posteriori error control for elliptic obstacle problems, Numel: Math. 95 (2003), 163-195. [115] J.T. Oden, Finite elements: an introduction, in: P.G. Ciarlet and J.L. Lions, eds., Handbook of Numerical Analysis, Vol. 11, North-Holland, Amsterdam, 1991, pp. 3-15. [116] J.T. Oden and S. Prudhomme, Estimation of modeling error in computational mechanics, J. Comput. Phys. 182 (2002), 496-515. [117] J.T. Oden and J.N. Reddy, An Introduction to the Mathematical Theory of Finite Elements, John Wiley, New York, 1976. [1181 J.T. Oden and K. Vemaganti, Estimation of local modeling error and goaloriented modeling of heterogeneous materials, Part I: Error estimates and adaptive algorithms, J. Comput. Phys. 164 (2000), 2 2 4 7 . [119] J.T. Oden and K. Vemaganti, Estimation of local modeling error and goal-oriented modeling of heterogeneous materials, Part 11: A computational environment for adaptive modeling of heterogeneous elastic solids, Comp. Methods Appl. Mech. Engrg. 190 (2001), 6663-6684.

296

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

[I201 J.T. Oden, K. Vemaganti, and N. Moes, Hierarchical modeling of heterogeneous solids, Comp. Methods Appl. Mech. Engrg. 172 (1999), 3-25. [I211 J.T. Oden and T.I. Zohdi, Analysis and adaptive modeling of highly heterogeneous elastic structures, Comp. Methods Appl. Mech. Engrg. 148 (1997), 367-391. [I221 R.W. Ogden, Non-linear Elastic Deformations, Ellis Harwood, Chichester, and John Wiley, New York, 1984. [I231 J.E. Osbom, Regularity of solutions of the Stokes problem in a polygonal domain, in Numerical Solution of Partial Differential Equations-U, Ed. B. Hubbard, Synspade, 1975. [I241 M.N. Ozigik, Boundary Value Problems of Heat Conduction, International Textbook Company, Scranton, Pennsylvania, 1968. [I251 P.D. Panagiotopoulos, Inequality Problems in Mechanics and Applications, Birkhauser, Boston, 1985. [I261 T.v. Petersdorff and E.P. Stephan, Decompositions in edge and comer singularities for the solution of the Dirichlet problem of the Laplacian in a polyhedron, Math. Nachl: 149 (1990), 71-104. [127] W. Prager and P.G. Hodge, Theory of Perfectly Plastic Solids, Dover Publications, New York, 1968. [I281 S. Prudhomme, J.T. Oden, T. Westermann, J. Bass, and M.E. Botkin, Practical methods for a posteriori error estimation in engineering applications, Int. J. Numel: Meth. Engng 56 (2003), 1193-1224. [I291 A. Quarteroni and A. Valli, Numerical Approximation of Partial Differential Equations, Springer-Verlag, Berlin, 1994. [I301 B.D. Reddy, Mixed variational inequalities arising in elastoplasticity, Nonlinear Analysis, TMA 19 (1992), 1071-1089. [I311 B.D. Reddy and T.B. Griffin, Variational principles and convergence of finte element approximations of a holonomic elastic-plastic problem. Numel: Math. 52 (1988), 101-1 17. [I321 J.C. Redondo, The penalized obstacle problem, Duke Mathematical Journal 69 (1993), 43-85. [I331 S.I. Repin, A posteriori error estimation for variational problems with uniformly convex functionals, Mathematics of Computation 69 (2000), 48 1-500.

REFERENCES

297

[134] S.I. Repin and L.S. Xanthis, A posteriori error estimation for elastoplastic problems based on duality theory, Computer Methods in Applied Mechanics and Engineering 138 (1996), 317-339. [I351 J.E. Roberts and J.-M. Thomas, Mixed and hybrid methods, in P.G. Ciarlet and J.-L. Lions, eds., Handbook of Numerical Analysis, Vol. II, NorthHolland, Amsterdam, 1991, 523-639. [136] R.T. Rockafellar, Convex Analysis, Princeton University Press, New Jersey, 1970. [I371 Ch. Schwab, p- and hp-Finite Element Methods, Oxford University Press, 1998. [138] M.J. Sewell, Dual approximation principles and optimization in continuum mechanics, Phil. Trans. Roy. Soc. (London)A 265 (1969), 319-351. [I391 V.G. Sigillito, Explicit A Priori Inequalities with Applications to Boundary Value Problems, Pitman Research Notes in Mathematics Series, Volume 13, London, 1977. [I401 E.M. Sparrow and R.D. Cess, Radiation Heat Transfer, McGraw-Hill, Washington, 1978. [I411 I. Stakgold, Green's Functions and Boundary Value Problems, John Wiley & Sons, Inc., 1979. [142] E. Stein and S. Ohnimus, Anisotropic discretization- and model-error estimation in solid mechanics by local Neumann problems, Comput. Methods Appl. Mech. Engrg. 176 (1999), 363-385. [I431 J. Steinbach, A Variational Inequality Approach to Free Boundary Problems with Applications in Mould Filling, Birkhauser Verlag, Basel, 2002. [I441 G. Strang and G. Fix, An Analysis of the Finite Element Method, PrenticeHall, Englewood Cliffs, NJ, 1973. [I451 F.-T. Suttmeier, General approach for a posteriori error estimates for finite element solutions of variational inequalities, Computational Mechanics 27 (2001), 3 17-323. 11461 B.A. Szabo and I. BabuSka, Computation of the amplitude of stress singular terms for cracks and reentrant corners, in Fracture Mechanics: Nineteenth Symposium, ASTM STP969, Ed. T.A. Cruse, 1988. [I471 V. Thomke, Galerkin Finite Element Methods for Parabolic Problems, Springer Lecture Notes in Mathematics, No. 1054, 1984.

298

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY

[148] V. ThomCe, Galerkin Finite Element Methods for Parabolic Problems, Springer, 1997. [I491 P. Tolksdorf, On the Dirichlet problem for quasilinear equations in domains with conical boundary points, Comm. in PDEs 8 (1983), 773-817. [I501 A. Veeser, Efficient and reliable a posteriori error estimators for elliptic obstacle problems, SIAM J. Numel: Anal. 39 (2001), 146-167. [I511 W. Velte, Complementary variational principles, in Finite Element and Boundary Element Techniques from Mathematical and Engineering Point of View, Ed. E. Stein and W. Wendland, Springer-Verlag, WienNew York, 1988. [I521 R. Verfiirth, A posteriori error estimates for nonlinear problems, finite element discretizations of elliptic equations, Mathematics of Computation 62 (1994), 445-475. [I531 R. Verfiirth, A Review of A Posteriori Error Estimation and Adaptive Mesh Refinement Techniques, Wiley and Teubner, 1996. [I541 C. Wang, Applied Elasticity, McGraw-Hill Book Company, Inc., 1953. [I551 J. Weisz, A posteriori error estimate of approximate solutions to a special nonlinear boundary value problem, Zeit. Angew. Math. Mech. 75 (1995) 1, 79-81. [I561 N.M. Wigley, Asymptotic expansions at a corner of solutions of mixed boundary value problems, J. Math. Mech. 13 (1964), 549-576. [I571 M.L. Williams, Stress singularities resulting from various boundary conditions in angular corners of plates in extension, J. Appl. Mech. 19 (1952), 526-528. [I581 N. Yan, A posteriori error estimators of gradient recovery type for elliptic obstacle problems, Adv. Comput. Math. 15 (2001), 333-362. [I591 E. Zeidler, Nonlinear Functional Analysis and its Applications. III: Variational Methods and Optimization, Springer-Verlag, New York, 1986. [I601 E. Zeidler, Nonlinear Functional Analysis and its Applications. IV Applications to Mathematical Physics, Springer-Verlag, New York, 1988. [1611 E. Zeidler, Nonlinear Functional Analysis and its Applications, II/B Nonlinear Monotone Operators, Springer-Verlag, New York, 1990. [I621 O.C. Zienkiewicz, Origins, milestones and directions of the finite element method-a personal view, in P.G. Ciarlet and J.-L. Lions, eds., Handbook

REFERENCES

299

of Numerical Analysis, Vol. 4, North-Holland, Amsterdam, 1996, pp. 3-67. [I631 O.C. Zienkiewicz and R.L. Taylor, The Finite Element Method, Vol. 1 (Basic Formulation and Linear Problems), McGraw-Hill, New York, 1989. 11641 O.C. Zienkiewicz and R.L. Taylor, The Finite Element Method, Vol. 2 (Solid and Fluid Mechanics, Dynamics and Nonlinearity), McGraw-Hill, New York, 1991. [I651 0. C. Zienkiewicz and J. Z. Zhu, A simple error estimator and adaptive procedure for practical engineering analysis, Int. J. Numel: Meth. Engrg. 24 (1987), 337-357. [I661 0. C. Zienkiewicz and J. Z. Zhu, The superconvergent patch recovery and a posteriori error estimates. Part 1: The recovery technique, Int. J. Nurnel: Meth. Engrg. 33 (1992), 1331-1364. 11671 0. C. Zienkiewicz and J. Z. Zhu, The superconvergent patch recovery and a posteriori error estimates. Part 2: Error estimates and adaptivity, Int. J. Numel: Meth. Engrg. 33 (1992), 1365-1382.

Index

C(Q, 7

cm(a),7 C? (n),7 C m ( 2 ) ,7

cO,a (Q,8 cm>"(n), 8 H 1 ( R ) ,13 H ! ~(n),16 H"(R, 10 inner product, 10 norm, 10 H t ( R ) >11 H-'(a),13 H ' / ~ ( ~ R12) , L m ( O ) ,8 L p ( R ) 8, L f O C ( R8) , w k , p ( R ) 9, norm, 9 seminorm, 15 w $ ' P ( 11 ~), Rd,5 inner product, 5 norm, 5 sd,5 inner product, 5 norm, 5 a posteriori error estimate, 3 a priori error estimate, 44 auxiliary function, 66 admissible, 73 best constant, 19 bilinear form, 18 bounded, 18 continuous, 18 elliptic, 18 symmetric, 18

CCa's inequality, 37 Carathtodory function, 58 Cauchy-Schwarz inequality, 6 ClCment-type interpolation, 240 coercive function, 56 compact embedding, 14 conjugate functional, 57 convergence, 6 strong, 6 weak, 6 convex function, 48 characterization, 56 continuity, 52 strictly, 48 strongly, 63 convex set, 47 density theorems, 11 directional derivative, 55 domain, 5 Lipschitz, 10 dual problem, 57 dual space, 5 dual variable, 66 admissible, 73 dual-weighted residual technique, 4 duality pairing, 5 duality theory, 59 effective domain, 48 elliptic variational inequality first kind, 30 second kind, 30 embedding, 13 energy difference, 61 epigraph, 49 equivalent norms, 14 error estimator, 235 gradient recovery-based, 255

A POSTERIORI ERROR ANALYSIS VIA DUALITY THEORY residual-based, 248, 273 error indicator. 235 finite element interpolant, 43 finite element method, 38 h-p-version, 38 h-version, 38 p-version, 38 finite element space, 41 finite elements affine-equivalent, 41 frictional contact problem, 32 Gdteaux derivative, 55 Galerkin method, 36 gradient recovery-based error estimate, 255 Holder continuity, 8 Hahn-Banach theorem, 50 heat conduction problem, 119 coefficient idealization, 121 insulation boundary condition, 122 linearization, 160 temperature boundary condition, 124 Hencky material, 144 idealizations in linear problems, 67 boundary condition, 100 coefficient, 68 domain, 106 right-hand side, 91 indicator function, 49,53 KaEanov method, 204 convergence, 205 elastoplasticity, 226 quasi-Newtonian flow problem, 219 stationary conservation law, 209 Kom inequality, 34 Lagrange multiplier, 238 Lax-Milgram Lemma, 18 linearization, 127 Bingham flow, 176 heat conduction problem, 160 nonlinear elasticity, 143 obstacle problem, 182 problem with small parameter, 169 quasilinear problem, 173 Lipschitz continuity, 7 Lipschitz domain, 10 lower semicontinuous (l.s.c.),49 mathematical model, 1 basic, 1 idealized, 1

meas (a), 8 mesh, 39 mesh parameter, 40 multi-index notation. 7 normal derivative, 13 obstacle problem, 3 1, 182, 193 partition, 39 regular, 40 PoincarC inequality, 19 primal problem, 57 proper functional, 29 reference element, 39 reflexive Banach space, 6 regularization method, 196 a posteriori error estimate, 200 convergence, 197 elastoplasticity, 226 obstacle problem, 193 residual-based error estimate, 248, 273 Riesz representation theorem, 6 Ritz method, 37 separation of convex sets, 50 singularity, 25 smoothness of boundary, 10 Sobolev space, 9 strain deviator, 144 stress deviator, 144 subdifferential, 53 subgradient, 53 sublinear functional, 50 summation convention, 5 support, 7 support functional, 53 torsion problem, 112 normalized Prandtl's stress function, 114 Prandtl's stress function, 113 warping function, 112 trace, 12 triangulation, 39 regular, 40 variational inequality, 29 weak derivative, 9 weak formulation, 16 weakly lower semicontinuous (w.l.s.c.), 49 Young inequality, 58 generalized, 58