AN ANAL YSIS OF

THE FINITE ELEMENT METHOD

¡ "
1
 Prentice-Hall
Series in Automatic Computation
George Forsythe, editor

AHO, editor, Currents in the Theoty of Computing MARTIN ANO NORMAN, The Computerizt
AHO AND ULLMAN, Theory of Parsing, Translation, and Compi/ing, MATHISON AND WALKER, Computers an~
-Volume 1: Parsing; Volume 11: Compiling MCKEEMAN, et. al., A Compiler Gen~rat
(ANDREE) 3, Computer Programming: Techniques, Analysis, and Mathematics MEYERs, Time-Sharing Computation in l
ANSELONE, · Collectively Compact Operator Approximation Theory and App/ications to MINSKY, _Computation: Finite and lnfini
Integral Equations NIEVERGELT, et al., Computer Approacll
ARBm, Theories of Abstract Automata PLANE AND MCMILLAN, Discrete Optim¡
BATES AND DOUGLAS, Programming Language/One, 2nd ed. Analysis for Management Decisions
BLUMENTHAL, Management lnformation Systems PRITSKER ANO KIVIAT, Simu/ation with
BRENT, Algorithms for Minimization without Derivatives Simulation Language
BRINCH HANSEN, Operating Systems Principies PYLYSHYN, editor, Perspectives on the •
COFFMAN AND DENNING, Operating-Systems Theory RICH, Interna/ Sorting Methods: Illustr
CRESS, et. al., FORTRAN IV with WATFOR and WATFIV RUSTIN, editor, Algorithm Specification
DANIEL, The Approximate Minimization o/ Functionals RUSTIN, editor, Computer Networks
DESMONDE, A Conversational Graphic Dáta PriJcessing System RUSTIN, editor, Data Base Systems
DESMONOE, Computers and Their Uses, 2nd ed. RUSTIN, editor, Debugging Techniques
DESMONOE, Real-Time Data Processing Systems - RUSTIN, editor, Design and Optilflizati•
DRUMMONO, Evaluation and Measurement Techniques for Digital Computer Systems RUSTIN, editor, Formal Semantics of P
EVANS, et al., Simulation Using Digital Computers SACKMAN AND CITRENBAUM, editors, 0
FIKE, Coinputer Evaluation of Matheinatical Functions Creative Problem-Solving
FIKE, PL/1 for Scientific Programers SALTON, editor, The SMART Retrieva.
FORSYTHE AND MOLER, Computer Solution o/ Linear Algebraic Systems Automatic Document Processing
GAUTHIER ANO PONTO, Designing Systems·Programs SAMMET, Programming Languages: Hb
GEAR, Numerical Intitial Value Problema in Ordinary Differential Equations scHAEFER, A Mathematical Theory of 1
GOLDEN, FORTRAN IV Programming and Computing 'scHULTZ, Spline Analysis
GOLOEN AND LEICHUS, IBM/360 Programming and Computing SCHWARZ, et al., Numerical Analysis o
GORDON, System Simulation SHERMAN, Techniques in Computer Pro
HARTMANIS ANO STEARNS, Algebraic Structure Theory of Sequential Machines SIMON ANO SIKLOSSY, Representa/ion a
HULL, Introduction to Computing with Information Processing Systems
JACOBY, et al., Iterative Methods for Nonlinear Optimiza/ion Problems STERBENZ, Floating-Point Computation
JOHNSÓN, System Structure in Data, Programs and Computers STERLING ANO POLLACK, Jntroduction l
KANTER, The Computer and the Executive STOUTEMYER, PL/1 Programming jor 1
KIVIAT, et al., Tfre SJMSCRIPT JI Programming Language STRANG AND FIX, An Ana/ysis of the f
LORIN, Para/lelism in Hardware and Software: Real and Apparent Concurrency STROUO, Approximate CalcUlation of 1
LOUOEN AND LEDIN, Programming the IBM 1130, 2nd ed. TAVISS, editor, The Computer lmpact
MARTIN, Design of Man-Computer Dialogues TRAUB, Iterative Methods for the So_!u
MARTIN, Design of Real-Time Computer Systems UHR, Pattern Recognition, Learning, a
MARTIN, Future Developments in Telecommunications VAN TASSEL, Computer Security Mana
MARTIN, Programming Real-Time Computing Systems VARGA~ Matrix lterative Analysis
MARTIN, Systems Analysis /or Data Transmission wAITE, lmplementing Software for No
MARTIN, Telecommunications and the Computer WILKINSON, Rounding Errors in Algeb
MARTIN, Teleprocessing Network Organization WIRTH, Systematic Programming: An
ting MARTIN AND NORMAN, The Computerized Society
tion, and Compiling, MATHISON AND WALKER, Computers and Telecommunications: Issues in Public Policy
MCKEEMAN, et. at, A Compiler Genfirator
es, Analysis, and Mathematics MEYERS, Time-Sharing Computation in the Social Sciences
Jroximation Theory and Applications to MINSKY, _Computation: Finite and lrifinite Machines
NIEVERGELT, et al., Computer Approaches to Mathematical Problems
PLANE AND MCMILLAN, Discrete Optimization: Integer Prograínming and Network
One; 2nd ed. Analysis for, Management Decisions
ns PRITSKER AND KIVIAT, Simulation with GASP II: a FORTRAN-Based
)erivatives Simulation Language
PYLYSHYN, editor, Perspectives on the Computer Revolution
1eory RICH, Interna/ Sorting Methods: 11/ustrated with PL/1 Program
md WATFIV RUSTIN, editor, Algorithm Specification
zctionals RUSTIN, editor, Computer Networks
ocessing System RUSTIN, editor, Data Base Systems
RUSTIN, editor, Debugging Techniques in Large Systems
'S - RUSTIN, editor, Design and Optimization of Compilers
iniques for Digital Computer Systems RUSTIN, editor, Formal Semantics of Programming Languages
ters SACKMAN AND CITRENBAUM, editors, On-Line Planning: Towards
mctions Creative Problem-Svlving
SALTON, editor, The SMART Retrieval System: Experiments in
inear Algebraic Systems Automatic Document Processing
rrams SAMMET, Programming Languages: History andFundamentals
dinary Di./ferential Equations SCHAEFER, A Mathematical Theory of Global Program Optimization
nputing 'scHULTZ, Spline Analysís ·
and Computing SCHWARZ, et al., Numerical Analysis of Symmetric Matrices
SHERMAN, Techniques in Computer Programming
fheory of Sequential Machines SIMON AND SIKLOSSY, Representation and Meaning: Experíments
with Information Processing Systems
Optimization Problems STERBENZ, Flo.ating-Point Computation
and Computers STERLING AND POLLACK, lntroduction to Statistical Data Processing
STOUTEMYER, PL/1 Programming for Engineering and Science
1g Language STRANG AND FIX, An Analysis o/ the Finite Element Method
Real and Apparent Concurrency STROUD, Approximate Calculation of Multiple lntegrals
ro, 2nd ed. TAVISS, editor, The Computer Impact
TRAUB, lterative Methods for the Solution o/ Polynomial Equations
lS UHR, Paítern Recognition, Learning; and Thought
ations VAN TASSEL, Computer Security Management
vstems VARGA, Matrix lterative Analysis
m WAITE, lmplementing Software for Notj-Numeric Application
WILKINSON, Rounding Errors in Algehraic Processes
WIRTH, Systematic Programming: An 'Introduction
AN ANALYSI
FINITE ELEME

GILBERT _STRANG
Massachusetts Institute of Technolo,

GEORGE J. FIX
University of Maryland

PRENTICE-HALL, INC.

ENGLEWOOD CLIFFS, l
AN ANALYSIS OF THE
FINITE ELEMENT METHOD

GILBERT STRANG
Massachusetts Institute o/Technology

GEORGE J. FIX
University of Maryland

PRENTICE-HALL, INC.

ENGLEWOOD CLIFFS, ·N.J.

Library of Congress Cataloging in Publication Data
STRANG, GJLBERT,
An analysis of the finite element method.
(Prentice-Hall series in automatic compuÚiÜon)
Bibliography: p.
1. Finite element method. l. Fix, Geor1;1e J.,
joint author. JI. Title.
TA335.S77 515'.624 72-12642
ISBN 0-13-032946-0

/
)
© 1973 by Prentice-Hall, Inc., Englewood Cliffs, N.J. To Jill an(

All rights reserved. No part of this book may be

reproduced in any form, by mimeograph or any other means,
without permission in writing from the publísher.

10 9 8 7 6 5 4

Printed in the United States of America.

PRENTICE-HALL INTERNATIONAL, INC., London

PRENTICE-HALL OF AUSTRALIA, PTY. LTD., Sydney
PRENTICE-HALL OF CANADA, LTD., Toronto
PRENTICE-HALL OF INDIA PRIVATE LIMITED, New Defhi
PRENTICE-HALL OF JAPAN, INC., Tokyo
glewood Cliffs, N.J. To Jill and Linda

; book may be
Jgraph or any other means,
1 the publisher.

1erica.

cherchez la fe. m .

. ,London
LTD., Sydney
1ronto
\iiTED, f'lew Delhi
vo
 PREFACE

The finite element method has been

to sol ve the complicated equations of e
for those problems it has essentially s
ences. Now other applications are ra1
in the geometry is important-and the
only to solve a system of equations, b
discrete approximation in the first place
thing to contribute.
From a mathematical point of vie
Rayleigh-Ritz-Galerkin technique. 1t
partial differential equations. The Ritz
directly with the differential equaÜon;
into an equivalent variational forrr:
assumed to be a combination 2: qirpJ
the method of weighted residuals, and
underlying variational principie. It is
puter actually solves.
So far the idea is an old one. What
in the finite element method they are
responsible for the method'~ success.
the domain, and enters the computati<
cular _node. In that neighborhood rp i i
low degree, and the computations are
that simultaneously, and .quite indep
become preeminent in the mathemati
·,_/' tions. Apparently it was the right id~
Because the mathematical foundat
stand why the method works. This is ti
 PREFACE

The finite element method has been an astonishing success. It was created
to sol ve the complicated equations of ehisticity and structural mechanics, and .
for those problems it has essentially superseded the method of finite differ-
ences. Now other applications are tapidly developing. Whenever flexibility
in the geometry is important-and the power of the computer is needed not
only to sol ve a system of equations, but also to formulate and assemble the
discrete approximation in the first place-the finite element method has some-
thing to contribute.
From a mathematical point of view, the method is an extension of the
Rayleigh-Ritz-Galerkin technique. It therefore applies to a wide class of
partial differential equations. The Ritz technique does not, however, operate
directly with the differential equation; instead, the contil1uous problem is put
into an equivalent vatiational form, and the approximate solution is
assumed to be a combination 2: q1 rp¡ of given trial functions rp¡(x). This is
the method of weighted residuals, and the weights q1 are computed from the
underlying variational principie. It is this discrete problem which the com-
puter actually solves.
So far the idea,-is an old one. What is new is the choice of trial functions:
in the finite element method they are piecewise polynomials. That choice is
responsible for the method'~ success. Each function rp1 is zero over most of ·
the domain, and enters the computation only in the neighborhood of aparti-
cular node. In that neighborhood rp 1 is pieced together from polynomials of
low degree, and the computations are as simple as possible. It is remarkable
that simultaneously, and quite independently, píecewise polynomials have
become preeminent in the mathematical theory of approximation of func-
tions. Apparently it was the right idea at the right time.
Because the mathematical foundations are sound, it is possible to under-
stand why the method works. This is the real reason for our book. It.s purpose

ix
X PREFACE

is to explain the effect ofe'!:ch bf the approximations that are essential for was to go out and program sorne fini1
the finite element technique to be computationally efficient. We list here sorne already available, students could be a~
of these approximations: ments with a theoretical seminar base:
Chapters 2 to 5 were also written b)
(1) interpolation of the original physical data w·ere drafted by the second author, ar
(2) choice of a finite number of polynomial trial functions the first. And the whole was typed b)
(3) sitÚplification of the geometry of the domain fully allowed us to believe that she er
(4) modification of the boundary conditions .
(5) numerica] integration of the underlying function~l in the variational
principie
(6) roundoff error in the solution of the discrete system.
Cambridge, Massachusetts
These questions are fundamentally mathematical, and so are the authors.
Nevertheless this book is absolutely not intended for the exclusive use of spe-
cialists in numerical análysis. On the contrary, we hope it may help establish
closer communication between the mathematical engineer and the mathe-
matical analyst. It seems to us that the finite element method provides a special
opportunity for this communication: the theory is attractive, the applications
are· growing, and best of all, the ·method is so new that tbe gap between
theory and application ought not yet to be insurmountable.
Of course we recognize that there are obstacles which cannot be made to
disa,ppear. One of them is the Janguage itself; we have kept the mathematical
notations to a mínimum, and indexed them (with definitions) at the end
of the book. We also know that, even after a norm has been interpreted as a
natural measure of strain energy, anda Hilbert space identified with the class
of admissible functions in a physically derived variational principie, there
still remains the hardest problem: to become comfortable with. these ideas,
and to make them one's own. This requires ·genuine patience and toleran ce
on both si des, as well as effort. Perhaps this book at least exhibits the kind of
proJ?lems which a mathematician is trained to solve, and those for which he
is useless.
In the last few years a great many numerical analysts have turned to finite
elements, and we are very much in their debt. This is acknowle,dged explicitly
throughout the book, and implicitly in the bibliography' even though we
have by no means attempted a formal history. Here, before the book begins,
we want to thank two others-engineers rathe'r than mathematicians-for
help that was the most important of aH. One is Isaac Fried, whose influence
led us to abandon an earlier (and completed) '''Fourier Analysis of the Finite
Element Method," and to study instead the real thing. The other is Bruce
lrons, whose remarkable intuitions are described (and proved correct, as
often as we can) in the book itself.
Chapter 1 is very much longer than the others, and was used by the first
author as the text in an introductory course at M.I.T. The only homewo~k
 PREFACE xi

:tpproximations that are essential for
utationally efficieni. We list here sorne
tysical data
ynomial trial functions
>f the domain
::onditions
derlying function~l in the variational
was to go out and program sorne finite elements. Where such programs are
already available, students could be asked to combine computational experi-
ments with a theoretical seminar based on the book.
Chapters 2 to 5 were al so written by the first author. The last three chapters
were drafted by the second author, and then revised and "homogenized" by
the first. And the whole was typed by Mrs. Ingrid Naaman, who has grace-
fully allowed us to believe that she enjoyed it; thank yo u.
GILBERT STRANG

GEORGE J. FIX
of the discrete system.
Cambridge, Massachusetts
nathematical, and so are the authors.
in tended for the exclusive use of spe-
ntrary, we hope it may help establish
.thematical engineer and the mathe:-
tite element method provides a special
e theory·is attractive, the applications ·
wd is so new that the gap between
to be insurmountable.
e obstacles which cannot be made to
tself; we have kept the mathematical
them (with definitions) at the end
fter a norm has been interpreted as a
Hilbert space identified with the class
derived variátional principie, there
~come comfortable with ·these ideas,
uires genuine patience and tolerance
:his book at least exhibits the kind of
ned to solve, and those for which he

merical analysts have turned to finite

jebt. This is acknowledged explicitly
the bibliography, even though we
story. Here, before the book begins,
rs rather than mathematicians-for
One i~ Isaac Fried, whose influence
:ted) ~'Fourier Analysis of the Finite
l the real thing. The other is Bruce
described (and pro ved correct, as

he others, and was used by the first

urse at M.I.T. The only homework
 xiii CONTENTS

3 APPROXIMATION

3.1 Pointwise Approximation

3.2 Mean-square ApproximatiOI
3.3 Curved Elements and !sopar
3.4 Error Estimates

CONTENTS

4 VARIATIONAL CRIMES

4.1 Violations of the Rayleigh-l

4.2 Non-conforming Elements
4.3 Numerical Integratioa·
PREFACE ix 4.4 Approximation of Domain ~

1 AN INTRODUCTION TO THE THEORY 1

5 STABILITY
1.1 The Basic Ideas 1
1.2 A Two-point Boundary-value Problem 3 5.1 Independence of the Basis
1.3 The Variational Form ofthe Problem 8 5.2 The Condition Number
1.4 Finite Difference Approximations 16
1.5 The Ritz Method and Linear Elements 24
1.6 The Error with Linear Elements 39
1.7 The Finite Element Meth9d in One DJmension· 51
1.8
1.9
Boundary Value Problems in Two Dimensions
Triangular and Rectangular Elements
63
74
6 EIGENVALUE PROBLENIS
1.10 Element M~trices in Two-dimensional Problems 90
6.1 Variational Formulation an
6.2 Sorne Elementary Examples
6.3 Eigenvalue and Eigenfunctic
6.4 Computational Techniques

2 A ªUMMARY OF THE THEORY 101

2.1
2.2
Basis Functions for the Finite Element Spaces Sh
Rates of Cop.vergence
101
105 7 INITIAL-VALUE PROBLEMS
2.3 Galerkin's Method, Collocation, and the Mixed Method 116
2.4 Systems of Equations; Shell Problems; Variations on 7.1 The Galerkin-Crank-Nicol:
the Finite Element Method .126 Equation
7.2 Stability and Convergence i
7.3 Hyperbolic Equations

xii
 xiii 'CONTENTS

APPROXIMATION 136

3.1 Pointwise Approximation 136

3.2 Mean-square Approximation 143
3.3 Curved Elements and Isoparametric Transformations 156
3.4 Error Estiniates 165

4 VARIATIONAL CRIMES 172

4.1 Violations of the Rayleigh-Ritz Code 172

4.2 Non-conforming Elements 174
4.3 Numericat Integration 181
ix 4.4 Approximation of Domain and Boundary Conditions 192

:ORY 1
5 STABIUTY 205
1
3 5.1 Independence of the Basis '205
le m
8 5.2 The Condition Number 209
~m

16
!nts 24
39
·Dimension 51
Dimensions
1ts
63
74
6 EIGENVALUE PROBLEMS 216

nal Problems 90
6.1 Variational Formulation and the Min-max Principie 216
6.2 Sorne Elementary Examples 223
6.3 Eigenvalue and Eigenfunction Errors 228
6.4 Computational Techniques 236

101

tent Spaces Sh 101

105 7 INITIAL-VALUE PROBLEMS 241
1d the Mixed Method 116
ms; Variations on 7.1 The Galerkiri-Crank-Nicolson Method for the Heat
.126 Equation 241
7.2 Stability and Convergence in Parabotic Problems 245
7.3 Hyperbolic Equations 251
 CONTENTS XÍV

8 SINGULARITIES 257

8.1 Cornees and Interfaces 257

8.2 Singular Functions 263
8.~.. Errors in the Presence of Singularities 266
8.4. ' Experimental .R.esb.lts 268

BIBLIOGRAPHY 281 1 AN INTRODUCTI

INDEX OF NOTATIONS 297

INDEX 303
1.1. THE BASIC IDEAS

The finite element method car

the problem to be solved is in va
the function u which minimizes a
minimizing property leads to a dif
but normally an exact solution
necessary. The Rayleigh-Ritz-G:
trial functions f/Jp ••• , 'fJN, and <1
to find the one which is minimi
unknown weights qj are determi
a system of N discrete algebraic
The theoretical justification for 1
minimizing process automatically
u. Therefore, the goal is to choc
enough for the potential energy
same time general enough to ap
The real difficulty is the first e
·ity. In theory there always exists
their linear combinations fill the
and therefore the Ritz approxim
with them is another matter. Thi
The underlying idea is -simpl(
or the-region of physical interes
easy for the computer to record
tangles. Then within each piece
 CONTENTS Xiv

257

251
' 263
ties 266
268

281
1 AN INTRODUCTION TO THE THEORY

297

303
1.1. THE BASIC IDEAS

The finite element method can be described in a few words. Suppose that
the problem to be solved is in variational form-it may be required to find
the function u which minimizes a given expression of potential energy. This
minimizingproperty leads toa difterential equation for u (the Eulerequation),
but normally an exact solution is impossible and sorne approximation is
necessary. The Rayleigh~Ritz-Galerkin,idea is to choose a finite number of
trial functions rp 1, •.. , rpN, and among all their linear combinations :E qJrp1
to find· the one which is minimizing. This is the Ritz approximation. The
unknown weights qJ are determined, Iiot by a differential equation, but by
a system of N discrete algebraic equations which the computer can handle.
The theoretical justification for this ·method is simple, and compelling: The
minimizing process automatical/y seeks out the combination which is closest to
u. Therefore, the goal is to choosetrial functions rpJ which are coñvenient
enough for the potential energy to be computed and minimized, and at the
same time general enough to approximate closely the unknown solution u.
The real difficulty is the first one, to achieve convenience and computabil-
ity. Intheory there always exists a set of trial functions which is complete-
their linear combinations fill the space of all possible solutions as N oo,
and therefore the Ritz approximations converge-but to be able to compute
· with them is another matter. This is what finite elements have accomplished.
·The underlying idea is simple. It starts by a subdivision of the structure,
or the region of physical interest, into smaller pieces. These pieces must be
easy for the computer to record and identify; they may be triangles or rec-
tangles. Then within each piece the trial functions are given an extremely

1
2 AN INTRODUCTION TO THE' THEORY CHAP. 1 SEC. 1.2. ATV

simple form-normally they .are polynomials, of at most the third or fifth We have attempted a fairly compl
degree. Boundary conditions are infinitely easier to impose locally, along the displacement method. A comparabl~ th
edge of a triangle or rectangle, than globally along a more complicated not yet exist, although it would cer
boundary. The accuracy of the approximation can be increased, if that is equations-in which the difficulties él
necessary, but not by the classical Rhz method of including more and more rriake a few preliminary comments or
compl~x trial funétions. Instead, the same polynomials are retained, and the an outstanding problem for the futu
subdivtlsion is refined. The computer follows a nearly identical set of instruc- method over the alternative variatio:
tions, a·nd just takes longer to finish. In fact, a large-scale finite element sys- 2, we ha ve opted to si de with the maj
tem can use the power of the computer, for the formulation of approximate version of the finite element method
equations as well as their solution, to a degree never before achieved.,.in would be the same for all formulatiot
complicated physical problems. throughout the whole subject makes
Unhappily none of the credit for this idea goes to numerical analysts. methods and force methods nearly au
The method was created by structural engineers, and it was not recognized Our goal in this chapter is to illust
at the start as an !nstance of the Rayleigh-Ritz principie. The subdivision method:
into simpler pieces, and the equations of equilibrium and compatibility
between the pieces, were initially constructed on the basis of physical reason- l. The variational formulation of
ing. The later development of more accurate elements happened in a similar 2. The construction of piecewise
way; it was recognized that increasing the degree of the polynomials would 3. The computation of the stiffm
greatly improve the accuracy, but· the unknowns qi computed in the discrete system.
approximation have always retained a physical significance. In this respect 4. The estimation of accuracy in
the computer output is ni.uch easier to interpret than the weights produced
by the classical method. We take the opportunity, when
The whole procedure became mathematically respectable at the moment insert sorne of the key mathematical
when the unknowns were identified as the coeffi.cients in a Ritz approximation Hilbert spaces xs and their norms, ti
u : : : : : 1:; q/Pr and the discrete equations were seen to be exactly the condi-
the data, and the energy inner produc
tions for minimizing the potential energy. Surely Argyris in Germany and specific problem. With these iools, tt
England, and Martín and Clough in America, were among those responsible; proved even for a very complicated g
we dare not guess who was first. The effect was instantly to provide a sound tional arguments permits an analysi
theoretical basis for the methoci. As the techniques of constructing more re- been achieved for finite differences.
fined elements have matured, the underlying theory has also begun to take
shape. 1.2. A TWO-POINT BOUNDARY-V•
The fundamental problem is to dlscover how closely piecewise polyno-
mials can approximate an unknown solution u. In other words, we must Our plan is to introduce the finit1
determine how well finite elements-which were developed on the basis of which Iies behind it, in terms ofa spec
computational simplicity-satisfy the second requirement of good trial func- to choose a one-dimensional probler
tions, to be effective in approximation. Intuitively, any reasonable function ments shall be simple and natural, él
u can be approached to arbitrary accuracy by piecewise linear functions. inanipulations shall be straightforwa
The mathematical task is to estímate the error as closely as possible and to than sorne general Green's formula.
determine how rapidly the error decreases as the number of pieces (or the tion
degree ofthe polynomial within .each piece) is increased. Of course, the finite
element method can proceed without the support .of precise mathematical (1)
theorems; it got on pretty well for more than 10 years. But we believe it will
be useful, especially in the future development of the method, to understand With suitable boundary conditions
and consolidate what has already been done. is a classical Sturm-Liouville probl
 CHAP.
,Jynomials, of at most the third or fifth
nitely easier to impose locally, along the
an globa1ly along a more complicated
)foximation can be increased, if that is
.itz method of inc1uding more and more
same polynomials are retained; and the
follows a nearly ideq¡t,ical set of instruc-
. In fact, a large-scale· finite element sys-
ter, for the formulation of approximate
to a degree never before achieved in
r this idea goes to numerical analysts.
al engineers, and it was not recognized
ayleigh-Ritz principie. The subdivision
[ons of equilibrium and compatibility
;tructed on the basis of physical reason-
Lccurate elements happened in a similar
tg the degree of the polynomials would
e unknowns q1 computed in the discrete
a physical significance. In this respect
to interpret than the weights produced
hematically respectable at the moment
the coefficients in a Ritz approximation
ms were seen to be exactly the condi-
tergy. Surely Argyris in Germany and
.merica, were among those responsible;
e:ffect was instantly to provide a sound
he techniques of constructing more re-
lerlying theory has also begun to take
iscover how c1osely piecewise polyno-
solution u. In other words, we must
.vhich were developed on the basis of
>econd requirement of good trial func-
l. Intuitively, any reasonable fQnction
curacy by piecewise linear functions.
~he error as c1osely as possible and to
!ases as the number of pieces (or the
•iece) is increased. Of course, the finite
the support .of precise mathematical
~e than lO years. But we beJieve it will
opment of the method, to understand
done.
1 SEC. 1.2. A TWO-POINT BOUNDARY-VALUE PROBLEM 3
We have attempted a fairly complete analysis of linear prob/ems and the
displacement method. A comparable theory for fuiiy nonlinear equations does
not yet exist, although it would certainly be possible to treat semilinear
equations-in which the difficulties are confined to Iower-order terms. We
make a few preliminary comments on nonlinear equations, but this remains
an outstanding problem for the future. In our choice of the displacement
method over the alternative variational formulations described in Chapter
2, we have opted to side withthe majority. This is the most commonly used
version of the finite eiement method. Of course; the approximation theory
would be the same for all formulations, and the duality which is so rampant
throughout the whole subject makes the conversion between displacement
methods and force methods nearly automatic.
Our goal in this chapter is to illustrate the basic steps in the finite el~ment
method:
l. The variational formulation of the problem.
2. The construction of piecewise polynomial trial functions.
3. The computation of the sti:ffness matrlx and solution of the discrete
system.
4. The estimation of accuracy in the final Ritz approximation.
We take the opportunity, when stating the problem variationally, to
insert sorne of the key mathematical ideas needed for a precise theory-the
Hilbert spaces xs and their norms, the estimates for the solution in terms of
the data, and the energy inner product which is naturally assodated with the
specific problem. With these iools, the cónvergence of finite elements can be
proved even for a very complicatedgeometry. In fact, the simplicity ofvaria-
tional arguments permits an analysis which already goes beyond what has
been achieved for finite differences.
1.2. A TWO-POINT BOUNDARY-VALUE PROBLEM
Our plan is to introduce the finite element method, and the mathematics
which lies behind it, in terms of a specific and familiar example. It makes sense
to choose a one-dimensional problem, in order that the construction of ele-
ments shall be simple and natural, and also in order that ·the mathematical
inanipulations shall be straightforward-requiring integration by parts rather
than sorne general Green's formula. Therefore, our choice falls on the equa-
tion
(1) - fx (p(x)~~) + q(x)u = f(x);
With suitable boundary conditions at the endpoints x = O and x = n, this
is a classical Sturm-Liouville problem. It represents a number of di:fferent
4 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.2. A

physical processes-the distribution of temperature along a rod, for example, sense "solved." Ofcourse there is sti
or thé displacement of a rotating string. Mathematically, the first point to which solution u corresponds to a gi
emphasize is that the equation and boundary conditions arise from a steady- book. But we believe it is worthwh
state problem, and not one which unfolds in time from initial conditions of first to get these function spaces rig:
disphicement and velocity. It would correspond in more space dimensions to the variational principies and theit
an ellipt{c boundarj-value problem, governed for example by Laplace's equa- space offunctions is admissible. (Tht
tion. n the implication that if u 1 and u2 art
In order to illustrate the treatment of different types of boundary''t:ondi- superposition is a natural property
tions, especially in the variational statement of the problem, we fix the left- We want to consider one specifi
hand end of the string and let the other be free. Thus at the end x = O there important to the theory, for the sp:
is an essential (or kinematic, or restrained, or geometric) boundary condition, admitted which have..finite energy. ·
in other words, one of Dirichlet type:

u(O) =O. (2) s: (/(>.

At the right-hand end x rc, the string is not constrained, and it assumes Any piecewise smooth function f is t
a natural (or dynamic, or stress) bound.ary condition, in other words, one of is not; we shall return later to this e
Neumann type: tions satisfying (2) is often denoted
u'(n) =O. cating by the superscript how many
energy (in this case it is only f itseU
We propose to consider this model problem from four different points For the simplest Sturm-Liouvi
ofview, in the following order: guess the corresponding space of so:
JCi-the subscript B refers to the
l. Pure mathematics. and the superscript 2 requires that th
2. Applied mathematics. The role of the pure mathematiciar
3. Numerical approximation by finite differences.
p(x) > Pmin > Oand q(x) > O, that
4. Numerical approximation by finite eléments. general equation -(pu')' + qu
j
in the following way:
1t is essential to recognize tne,common features of these four approaches
The operator L is a one-to-one ¡
to the same problem; the tools which are useful to the pure mathematician
for each f in 3C 0 the differential eq
in proving the existence and uniqueness of the solution, and to the applied
Furthermore, the solution depends CJ
mathematician in understanding its qualitative behavior, ought to be applied
sois u.
also to the study of the numerical algorithms.
The last sentence requires furth
We begin with the pure mathematician, who combines the differential
to measure the size off and u. The 1
equation and boundary conditions into a single equation,
space and solution space are differe
Lu f for the norms in terms ofthe energy

L is a linear operator, acting on a certain class of functions-those which in

sorne sense satisfy the boundary conditions and can be differenti~ted twice. 11 f llo [J (/(~)) 2 dx~
Mathematically, the fundamental question is precisely this: to match such
a space offunctions u with a class of inhoinogeneous terms J, in such a way that 11 u llz = [J {(u"(x))" -i
to ea eh f there corresponds one and on/y one solution u. Once this correspond-
ence betweenfand u has been established, the problem Lu isin anabstract tThese spaces are defined again in th<
 CHAP.
:emperature along a rod, for example,
g. Mathematically, the first point to
ndary conditions arise from a steady-
lds in time from initial conditions of
respond in more space dimensions to
~rned for example by Laplace's, equa-
1'\
)f different types of bbundary condi-
:ment of the problem, we fix the left-
be free. Thus at the end x =O there
d, or geometric) boundary condition,
:o.
g is not constrained, and it assumes
:ry condition, in other words, one of
=o.
problem from four different points
ite differences.
~e eletnents.
on features of these four approaches
re useful to the pure mathematician
; of the solution, and to the applied
itative behavior, ought to be applied
ithms.
cian, who combines the differential
a single equation,
f for the norms in terms of the energy, or rather its square root:
n class of functions-those which in
ons and can be differentiated twice.
ion is precisely ·this: to match such
1ogeneous terms f, in such a way that
me solution u. Once this correspond-
' the problem Lu is in an abstract
1 SEC. 1.2. A TWO-POINT BOUNDARY-VALUE PROBLEM 5
sense "sol ved." Of course there is still a little way to go in actually discovering
which solution u corresponds toa givenf That step is the real subject ofthis
book. But we believe it is worthwhile, and not just useless fussiness, to try
first to get these function spaces right. In fact, it is of special importance for
the variational principies and their 'approximation to know exactly which
space offunctions is admissible. (The references to "spaces" offunctions carry
the implication that if u 1 and u2 are admissible, then so is c 1u 1 + c2u2 ; this
superposition is a natural property in linear problems.)
We want to consider one specific choice, the one which is perhaps most
important to the theory, for the space of inhomogeneous data: Those f are
admitted which haveJinite energy. This means that
(2) J: (f(x)) 2 dx < oo.
Any piecewise smooth function f is thereby included, but the Dirac ó-function
is not; we shall return later to this case of a "point load." The space offunc-
tions satisfying (2) is often denoted by L 2 ; we prefer the notation 3C 0 , indi-
cating by the superscript how mat;ty derivatives off are required to have finite
energy (in this case it is only f itself).
For the simplest Sturm-Liouville equation -u" = f, it is not hard to
guess the corresponding space of solutions. This solution space is denoted by
3Cj-the subscript B refers to the boundary conditions u(O) il(n) =O,
and the superscript 2 requires that the second derivative ofu hasfiniteenergy.t
The role of the pure mathematician is th,en to show, under the assumptions
p(x) > Pmin > O and q(x) > O, that 3Ci is still the solution space for the more
general equation -(pu')' + qu f In fact, his final theorem can be stated
in the following way:
The operator L is a one-to-one transformation from 3Cj onto 3C 0 , so that
for each fin 3C 0 the differential equation (1) has a unique solution u in 3Ci.
Furthermore, the solution depends continuously on the data: Jff is small, then
sois u.
The last sentence requires further explanation; we need norms in which
to measure the size offand u. The two norms will be different, since the data
space and solution space are different. Fortunately, there is a natural choice
11 f llo [J (f(x)) dxJ12,
2
11 u ll2 [J ((u"(x))2 + (u'(x)) 2 + (u(x)) 2) dx J 12
~
tThese spaces are defined again in the index of notations at the end of the book.
6 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.2; A

With.·these definitions, the continuous dependen ce of the solution on the data energy, f f 2 < oo. The paradox is
can be expressed in, a quantitative form: There exists a constant C such that functions u" in 3C0 • Whether f sati
or not, its expansion is valid in tht
(3)

The 1uniqueness of the solution follows immediately from this estímate:

Iff = e); then necessarily u O. In fact, it is such esiimates which lie at the
very center of the modero theory of partial differential equations. A general The boundary conditions on f are 1
technique for proving (3), which applies also to boundary-value problems in N~ oo. Fi~ure 1.1 shows how a:
severa! space dimensions, has been created only in the last generation. In this could still converge to a function.)
book, we shall accept such estimates as pro ved: For elliptic equations of order The Sturm-Liouville differenti
2m, this means that solved: iff = 2: anun, then u has tl

(4). llullzm < Cll/llo· (7)

We move now to a more applied question, the actual construction of the

With this explicit const.ruction, the
solution. If the coefficients p and q are constant, then this can be carried out
of data space 3C0 with solution sp
in terms of an· infinite series. The key lies in knowing the eigenvalues and
The question of the boundar:
eigenfunctions of L:
further comment. We have seen al

(5) uix) = Jf sin(n !)x, An p(n - f) 2 + q.

in terms of the Un, which do SR
absolutely nothing to do with th
What is different about u? Why d1
It is immediate to check that Lun -pu~ +
qun = lnu,, that the functions answer is that the series expansio1
un satisfy the boundary conditions · and therefore lie in 3C~, and that they than the expansion for f: not onl)
á re orthonormal: square sense, but so do its first anc

Suppose the ·inhomogeneous term is expanded Üt a series of eigenfunctions:

"'
(6) f(x) = ~ an-y(ii
2 sin(n- f)x.

Then integrating formally, the orthqgonality of the un gives

11/lla J: /2
dx =~~a?;.
The functigns fin 3C 0 are exactly those which admita harmonic expansion
of the form (6),. with coefficients satisfying L: a?; < oo. Actually, this ought
to seem a little paradoxical, since apparently every f of the form (6) will
satisfy f(O) O, f'(n) =O, whereas no boundary conditions were meant x=O
to be imposed on f: The elements of 3C 0 are required only to ha ve finite Fig. 1.1 IN in aci a)
 CHAP. 1 SEC. 1.2. A TWO-POINT BOUNDARY-VALUE PROBLEM 7

pendence of the solution on the data energy, J f 2 < oo. The paradox is resolved by the completeness ofthe eigen-
There exists a constant e such that functions un in :JC 0 • Whether f satisfies these spurious boundary conditions
or not, its éxpansion is valid in the mean-square sense,
ll/llo·
ws immediately from this estima te:
it is such estimates \\:hich Jie at the
J: (f(x) ~ an¡{- sin(n -1)x )
2
dx-- O as N-- oo.

tial differential equaÚtins. A general
also to boundary-value problems in
!d only in the last generation. In this
·o ved: F or elliptic eq uations of order
The boundary conditions on f are thus unstabJe and disappear in the Iimit as
N-- oo. Figure 1.1 shows how a sequence of functions JN, all lying in Jei,
could still converge to a function f outside that space.
The Sturm-Liouville differential equation Lu = f is now ready to be
sol ved: iff = I: ar,un, then u has the expansion .

'11/llo·
stion, the actual construction of the
mstant, then this can be carried out
ies in knowing the eigenvalues and
u:' + qun = A.nun, that the functions
therefore líe in :re¡, and that they
(7) u.
With this expJicit construction, the estimate 11 u 11 2 < e 11 f 1lo and the matching
of data space JC 0 with solution_ space Jei can be verified directly.
The question of the boundary conditions is more subtle and deserves
further comment. We have seen already that even thoughfcan be expanded
in terms of the un, which do satisfy the boundary conditions, still f has
absolutely nothing to do with these conditions. Therefore the question is:
What is different about u? Why does u satisfy the boundary conditions? The
answer is that the series expansion for u converges in a much stronger sense
than the expansion for f: not only does I: anunf A.n converge to u in the mean-
square sense, but so do its first and seconq derivatives. More precisely,

llu as N-- oo.

anded in a series of eigenfunctions: f {7T) :1= o

1

f~{7T) =o

· sin(n -1)x.

llity of the un gives

""
t' :¿:a;.
n=l
f{O) =F O
which admit a harmonic expansion
ng E a; < oo. Actually, this ought
~ently every f of the form {6) • will
boundary conditions were meant x=O x= 7T
C0 are required only to have finite Fig. 1.1 IN in 3C~ approxirnating a general functionf.
8 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.3. TF

The point is that when even the second derivatives converge, the boundary strong movement toward the minimi
conditions are stable; the limit function u is compelled to satisfy the boundary fundamental problem to be approxi1
conditions. (Note thát in Figure 1.1, the second derivatives of !N did not The relationship between the line;
converge to those off; therefore the limitfdid not have to satisfy the boun- if L, v, and f are just real numbers
dary conditions, and was outside the space ac¡. This is what will not happen provided that Lis positive, its minimu
for u.) ;, '
The g~neral rule is this: boundary conditions which invob'e only derivati~'es di 1 - 2(
below order s will make sense in the acs norm; those involving derivatives of dv v=u-
order s or higher will be unstable and will not apply to functions in the space
:Jes. We shall see that this is the rule which distinguishes between essential If L were not positive, the problerr
boundary conditions, which ~tay, and natural boundary conditions, which either the mínimum is oo or, if j

go. The distinction becomes crucial in the variational problem, whose state- straight line.
ment is in terms of first derivatives, that is, the JC 1 norm. The finite element A more interesting case is when
approximations will be required to satisfy aH boundary conditions below ordef is a symmetric positive definite ma1
1-that means the condition u( O) = 0-but they will not be required to satisfy stands for tlie usual inner product (
the condition on the first derivative. This leniency at x n will not prevent
the finite element approxi~ations from converging in the JC 1 norm to a solu- I(v) = L;L.~c
i,k J

tion u which does satisfy u'(n) = O. This is the key to the following section,
which extends the "pure mathematics" standpoint to the equivalent varia- Applying the symmetry Lik = Lk1; t
tional problem.

1.3. THE VARIATIONAL FORM OF THE PROBLEM

These n simultaneous equations m~
The linear equation Lu = f is related to the quadratic functional mínimum value of I, attained at u

I(v) = (Lv, v) - 2(f, v)

in the following way: l(v) is minimized at v u only if its derivative (or Since L is positive definite, so is L - l
first variation) vanishes there, arÍd the condition for the vanishing of this whenf = 0). Geometrically l(v) is 1
derivative is exactly the Euler equation Lu f The problems of inverting boloid opening upward, when L is
L and minimizing I are equivalent; they produce the same solution u. There- A vector derivation of the minin
fore, such problems can be investigated either in an operationalform, in terms all n coinponents at once. lf 1 has a
of the linear operator L, or in l'ariational form, in terms of the quadratic
l. The goal in this section is tó find the exact variatiohal equivalent of our l(u) l(u + EV) l(u) +
two-point boundary-value problem.
This -equivalence of•differential equations with variational problems is Since E can be arbitrarily small and
basic also to the choice of a computational scheme. The differential equation (8) (Lu, v) (f,
may be approximated by a discrete system, using finite differences, or the
variational integral can be minimized over a discrete class of functions, as This forces Lu We call the equ~
in the finite element method. In many applications-particularly in steady- of the problem. It no longer requir
state rather than transient problems-the variational statement is the pri- metric, since it deals not necessaril
mary physical principie, and the differential equation only a secondary tionary point. In other words, it stE
consequence. Therefore, it is not surprising to find in such applications a In this forro the problem leads to G
 CHAP.
l derivatives conv~rge, the boundary
~~ is compelled to sátisfy the boundary
.he second · derivatives of IN did 11ot
ait/did not have to satisfy the boun-
:tce Xi. This is what wil! not happen
nditions which involve;¡.only derivatives
norm; those involving· derivatives of
ill not apply to functions in the space
thich distinguishes between essential
natural boundary conditions, which
:he variational problem, whose state-
t is, the 3<3 1 norm. The finite element
y all boundary conditions below order
Jut they will not be required to satisfy
is leniency at x ::::::= 1t will not prevent
converging in the 3<3 1 norm toa solu-
is is the key to the following seciion,
standpoint to the equivalent varia-
: PROBLEM
i to the quadratic functional
2(/, v)
d at v u only if its derivative (or
condition for the vanishing of this
Lu =f. The problems of inverting
produce the same solution u. There-
!Íther in an operational form, in terms
•na! form, in terms of the quadratic
! exact variational equivalent of our
tations with .variational problems is
nal scheme. The differential equation
>tem, using finite differences, or the
>ver a discrete class of functions, as
applications-particularly in steady;..
the variational statement is the pri-
~rential equation only a secondary
·ising to find in such applications a
1 SEC. 1.!. THE VARIATIONAL FORM OF THE PROBLEM 9
strong movement toward the minimization of the quadratic functional as the
fundamental problem to be approximated .
The relationship between the linear and quadratic problems is transparent
if L, v, and f are just real numbers. I(v) = Lv 2 - 2fv is a parabola and,
provided that Lis posith•e, its mínimum is attained at the point u where
'!fv lv=u 2(Lu - /) O.
If L were not positive, the problem of minimization would break down:
either the mínimum is - oo or, if L = O, the para bola degenerates into a
straight line.
A more interesting case is when v and f are n-dimensional vectors and L
is a symmetric positive definite matrix of order n. The notation ( , ) then
stands for the usual inner product (or dot product) of two vectors, so
Applying the symmetry Lik Lki' the Euler equation is
m 1, ... , n.
These n simultaneous equations make up the vector equation Lu =f. The
mínimum value of /, attained at u= L- 1¡, is
Since L is positive definite, so is L - 1 , and this mínimum is negative (or zero,
when f = 0). Geometrically l(v) is represented by a convex surface, a para-
boloid opening upward, when L is positive definite.
A vector derivation of the minimizing equation Lu f comes by varying
all n components at once. If 1 has a mínimum at u, then for all v and f,
l(u) I(u + fV) I(u) + 2f[(Lu, v) - (f,.v)] + é(Lv, v).
Since f can be arbitrarily small and of either sign, its coefficient must vanish:
(8) (Lu, v) = (/, v) for every v.
This forces Lu We call the equation (8) the weakform, or Galerkinform,
of the problem. It no longer requires L to be positive definite or even sym-
metric, since it deals not necessarily with a mínimum but only with a sta-
tionary point. In other words, it states only that the first variation vanishes.
In this form the problem leads to Galerkin's approximation process.
10 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.3. TP.

lt .is worth mentioning that (Lv, v) - 2(/, v) is not the only quadratic is struck by the fact that. the expre
whose mínimum occurs at the point where Lu f lt is obvious that the because of the. integration by parts:
least-squares functional Q(v) (Lv- J, Lv f) has its mínimum (zero) I(v) will be well defined if only the fi.J
at the same point. Th~re is one significant difference, however: the Euler is required to ha ve finite energy. Th
equation, which arises by equating the derivatives aQ¡avm to zero, is not the class of functions which are ad
Lu J, put LTLu =-:::; Lrf Tfie two are theoretically equivalent, assuming that a space bigger than 3Cj.
L is inv~rtible, but in practice the appearance of LTL is a disadvantage. Our guiding principie will be th
Consider now the differential equation of the previous section: vided it can be obtained as the Iimit
"Jimit" we mean that the quadratic
I(v) converge:
Lu = [ - fx(p(x~) + q(x)Ju f,

u(O) = u'(n) O. (10) J: p(v' v:V) 2 + q(v

We want to construct I(v) = (Lv, v) - 2(/, v). These inner products now Notice that such an enlargement of
involve functions defined on the intervalO< x < n rather than vectors with mum value of /; each new value /(t
n components, but their definition is complete] y analogous to the vector case: the mimimum of 1 was already att'

(f, v) = J: f(x)v(x) dx.

minimizing function. This is exactl)
we now have the enormous advan
of being permitted to try functiom
Note that our data f and functions v are assumed to be reai, as in almost JCj. In practice, this means that we
all applications. This is a convenient assumption, and the modifications to ous but only piecewise linear-thc
be made in the complex case are well known; in the integrand abo ve, one of derivative has finite energy, but the:
the factors should be conjugated. Our problem is now to descrit
The computation of 1 involves an integration by parts (indeed the whole In other words, we want to discove·
theory of differential operators rests on this rule): which is the limit in the sense of (
functions vN which have two qeriv'
(Lv, v) = J: ["- -:( pv')' + qv]v dx
tions.
There are two properties to be de
(9)
= J: [p(v') 2 + qv 2] dx pv'v 1: ·
admissible functions and the bour
first is comparatively easy: since ti
gen ce of the fitst derivative, the lim
If v satisfies the boqndary conditions v(O)'= v'(n) O, then the integrated means that the norm
term vanishes and the quadratic functional is

I(v) = J: [p(x)(v'(x)) 2 + q(x)(v(x)) 2

2f(x)v(x)] dx.
must be finite.
This is the functional to be minimized. Again the boundary conditions
The solution of the differential problem Lu = f is expected to coincide sequence from JCj will satisfy vN(O)
with the function u that minimizes l. But within, what class of functions shall both these properties? lt turns óu1
we search for a mínimum? Since we are looking for a solution u in. JCj, then the other is lost. To see that the limi
certainly we ought to admitas candidates at least all members of JCj. This v' (n) = Ó, suppose for example tr
leads to a correct and equivalent variational problem; the mínimum of /(v) vN in Fig. 1.2. Since v- vN is zero
over all functions v in JCj occurs at the desired point v = u. However, one and there O v' v:V 1, the req
 CHAP.
- 2(f, v) is not tpe only quadratic
here Lu f. It is obvious that the
1: Lv- f) has its mínimum (zero)
:::ant difference, however: the Euler
derivatives aQ¡avm to zero, is not
~oretically equivalent, assuming that
)earance of LTL is ~\disadvantage.
on of the previous se~tion:
)+ q(x)Ju f,
~) o.
· 2(f, v). These inner products now
O x n rather than vectors with
tpletely analogous to the vector case:
:x)v(x) dx.
re assumed to be real, as in almost
;sumption, and the modifications to
1own; in the integrand above, one of
tegration by parts (indeed the whole
this rule):
+ qv]v dx
qv 2 ] dx pv' v l: ·
(0)'= v'(n) O, then the integra;ed
•nal is
:)(v(x)) 2 2f(x)v(x)J dx.
blem Lu f is expected to coincide
t within what class of functions shall
looking for a solution u in. 3Ci, then
tes at Ieast all members of Xi. This
ional problem; the mínimum of J(v)
~ desired point v = u. However, one
1 SEC. 1.~. THE VARIATIONAL FORM OF THE PROBLEM 1.1
js struck by the fact that the expression for 1 involves no second derivatives;
beca use of the integration by parts, it involves only v and v'. It follows that
1(v) will be well defined if only the first derivative of v, rather than the second,
is required to ha ve finite energy. Therefore, the possibility arises of enlarging
the class of functions which are admissible in the minimization problem to
a space bigger than 3C¡.
Our guiding principie will be this: Any function v will be admitted, pro-
vided it can be obtained as the limit of a .sequence vN in 3Ci, where by the word
"limit" we mean that the quadratic terms in the potential-energy functional
1(v) converge:
(10) as N - oo.
Notice that such an enlargement of our space cannot actual/y lower the mini-
mum value of 1; each new value 1(v) is the limit of old values 1(vN). Thus if
the mimimum of 1 was already attained for sorne u in ;re¡, that remains the
minimizing function. This is exactly what we know to be the case. However,
we now have the enormous advantage, while searching for this mínimum,
of being permitted to try functions v which were outside the original class
3Ci. In practice, this means that we can now try functions which are continu-
ous but only piecewise Iinear-they are easy to construct, and their first
derivative has finite energy, but they do not lie in Xi.
Our problem is now to describe this new and larger admissible space.
In other words, we want to discover the properties possessed by a function v
which is the limit in the sen se of (l 0)-this is effectively the 3C 1 norm-of
functions v'j.¡ which have two derivatives and satisfy all the boundary condi-
tions.
There are two properties to be determined, the smoothness required ofthe
admissibJe functions and the boundary conditions they must satisfy. The
first is comparatively easy: sin ce the requirement (lO) no tices only conver-
gence of the first derivative, the limiting function v need only lie in 3C 1 • This
means that the norm
]1/2
11 v !11 = [I n
0
(v 2 + (v') 2) dx
must be finite.
Again the boundary conditions pose a more subtle problem: Since any
sequence from Xi will satisfy vN(O) =O and v:V(n) =O, will its limit v inherit
both these properties? lt turns out that the jirst condition is preserved, and
the other is lost. To see that the limit need not satisfy the Neumann condition
v'(n) = O, suppose for example that v(x) = x and consider the sequence
vN in Fig. 1.2. Since v vN is zero except over the small interval at the end,
and there O v' - v:V < 1, the requirement (lO) is obviously satisfied. Thus
12 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.3.

Since this holds for f on both s

must vanish:
v'(1r) =1
v~(1T) =O o= s:pu'v' + q
(11)
J: [ -(pu')' .

If the minimizing u has two de

parts, this expression wiJJ be zerc
fin the interval, and the natural
Also u( O) O, beca use like every
boundary condition. This comp
x=7r-/¡ X 1 is equivalent to the solution
x'=O x=1r approached from either directio;
Geometrical1y, the process G
Fig. 1.2 Convergence in :Jet, with v:V.(rc) O but v'(rc) ::1= O. interpretation. The quadratic 1 il
loid in infinite dimensions. At 1
v(x) x lies in the limiting space of admissible functions, even though there were "boles" in this surfa,
v'(n) =F O. surface has changed neither its st
On the other hand, the condition v(O) = O continues to hold in the limit. pinpricks, corresponding to func
In fact, vN will converge at each individual point x, since by the Schwarz exist.
inequality To conclude this section, w

J:
lv~x)- ~M(x)l 2 = 1 (v~(y)- v~(y)) dyl 2
are of great irnportance in appli
the variational forrn which was

J: J: (v~- v~)Z
12 dy dy ~o.
l(v) is still to be minirnized over 1
function u is the desired so1utio
becomes more cornplex, special
By standard results in anal~sis, the limiting function v is continuous and the the solution is no longer in x¡. 1
convergence of vN to v is uniform in x. In particular, at the point x = O, this speciaJ behavior cannot be
v(O) = lim vN(O) O. On the other ha~d, the first derivatives converge only understood. From the viewpoin1
in mean square, giving no assurance that v'(n) =O. it is significant that the algorithn
The space of admissible functions v in the minimization is therefore X 1; on the singularity. Such inforrn
its members ha ve first derivatives with finite energy and satisfy the essential the convergence of the approxin
boundªry condition v(O) = O indicated by the subscript E. The natural the algorithrn wiH not break dov
boundary condition v'(n) Ois not imposed. Notice that if our mathematics
is consistent, then the function u in X1 which minimizes 1 should automati,. Remark l. The coefficient p(.
cal/y satisfy u'(n) =O. This is easy to confirm, since for any f and any v in the elastic property of the strin
X};, flow) changes abruptly. At such
the solution u will no Jonger ha~
/(u) l(u + fV) tion" at x 0 , we depend on the
= I(u) + 2f J: pu'v' + quv - fv +f 2
J: p(v') 2 + qv 2 • variation J pu'v' + quv- fv m
. Assuming no other points of dif
 CHAP. 1 SEC. 1.3. THE VARIATIONAL FORM OF THE PROBLEM 13

Since this holds for f on both sides of zero, the linear term (first variation)
'must vanish:
v'{7r) =1
v~(TT)= O
(11)
O J: pu'v' + quv fv

= J: [ -(pu')' + qu f]v + p(n)u'(n)v(n).

X= 1T- k approached from either direction.
x=TT
vjy(n) = O but v'(n) =F- O.
:tdmissible functions, even though
1= O continues to hold in the limit.
lual point x, since by the Schwarz
y) - v:V(y)) dy 12
f: (v~ - v'M) 2 dy ~o.
ng function v is continuous and the
. In particular, at the point x O,
l, the first derivatives converge only
t v'(n) O.
¡ the minimization is therefore X 1;
nite energy and satisfy the essential
by the subscript E. The natural
)Sed. Notice that if our mathematics
which minimizes 1 should automati~
mfirm, since for any f and any v in
If the minimizing u has two derivatives, permitting the last integration by
parts, this expression will be zero for all v in X 1 only if both -(pu')' qu+
fin the interval, and the natural condition u'(n) =O holds at the boundary.
Also u(O) O, beca use like every other function in X 1, u satisfies the essential
boundary condition ..This completes the cycle: The minimizadon of 1 o ver
X 1 is equivalent to the solution of Lu = f, and the computation of u can be
Geometrically, the process of completion from Xi to X1 has a simple
interpretation. The quadratic 1 is represented by a convex surface, a para bo-
lo id in infinite dimensions. At first, when 1 was defined only for v in :re¡,
there were "boles" in this surface; aH we ha ve done is to fill them in. The
surface has ~hanged neither its shape nor its mínimum value; it isjust that the
pinpricks, corresponding to functions v lying in X 1 but not in Xi, no longer
exist.
To conclude this section, we remark on two singular problems which
are of great importance in applications. It is noteworthy that in both cases
the variational forro which was just established remains completely valid;
1(v) ~s still to be minimized over the admissible space X1, and the minimizing
funct'ion u is the desired solution. In contrast, the operational forro Lu = f
becomes more complex, special conditions enter at the singularity x 0 , and
the solution is no longer in Xi. From the viewpoint of applied mathematics,
this special behavior cannot be ignored; it is probably the very point to be
understood. From the viewpoint of finite element approximations, however,
it is significant that the algorithm can proceed without complete information
on the singularity. Such information will be extremely valuable in speeding
the convergence of the approximations, as in Chapter 8 on singularities, but
the algorithm will not break down without it.
Remark l. The coefficientp(x) may be discontinuous ata point Xo where
the elastic property of the string (or the diffusivity of the medium in heat
flow) changes abruptly. At such a point there appears an interna! houndary;
the solution u will no longer have two derivatives. To find the "jurrip coridi-
tion" at x 0 , we dépend on the variational forro of the problem; the first
vadation f pu'v' + quv- fv must still vanish for all v if u is minimizing.
Assuming no other points of difficulty, integration by parts o ver the separate
14 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.3.

intervals (0, x 0 ) and (x 0 ; n) yields But also for f = .5(x0), O < x 0 <
O=
f xo
0
[
1 1
-(pu) + qu - ]
f v + p_u_v_
1

I(v) J:
+ Jn.:ro [-(pu')' + qu f]v + p(n)u'(n)v(~)- P+u'+v:.
and the mínimum occurs at the raJ
The sub~hripts - and + indicate the limiting values as x approaches Xo the solution is not in 3Ci, and the
from the left and right, respectively. Recall that v_ v+ for any v in 3C1,
happens to fall at the bottom of t
since v is continuous, and in particular that u_ = u+. Varying v, it follows One more possibility: Suppos
that the differential equation holds in each interval, that u'(n) = O at the far that .5(n). Then the solution
leve/s off. In this case the so/utic
end,. and that
u'(n) O. Looking back at (11)
q O, and u(x) x, the first vari
This is the natural boundary condition at x 0 , and is a direct consequence
of the variational form: u' has a jump, but the combination pul remains
continuous.
J: (u'v' - fv) dx

Since there is a jump in u', the solution líes in 3C1 but not in 3C1. This is
for any v in 3C1. Thus the first vari
a case in which one of the holes in the surface l(v) was actually at the bottom.
minimizing. The only remarkable
The mínimum value of I(v) would have been the same over the original space
(11), and therefore the subsequent
3C 2 but within that space no function attained the mínimum. The surface
tion u'(n) = O, falls through. Thu
ca:e arbitrarily near to the hole, but it was filled in only by the function
there may no longer be a natural]
satisfying the jump condition. .
Once a J-function has been acc
The standard error estimates for finite elements, which rest on an assumed
the general question: Which class <
smoothness of u, will dégenerate at the discontinuity x 0 • These estimates can
More precisely, which space of dat:
be saved, however, if by placing a node at x 0 wé permit the trial fu?~tio~s
speaking, as long as /(v) remains f
in the approximation to copy the jump condition. Since the cond1tton IS
proceed. This rule accepts .5-functi
natural rather than essential, not every trial function has to satisfy it; as long
as the trial functions are not forced to have a continuous derivative, and
thereby violate the jump condition, the approximation will be good.

Remark 2. Up to now the inhomogeneous term f has been required to

come from 3C 0 , thereby excluding the .5-function. Physically this would be
ari interesting choice, representing a point load or point so urce, and mathe-
matically the corresponding u is the fundamental solutio.n. Therefore we try u(x):;:
harder t:o include it, by reconsidering the functional

I(v) = J: p(v') 2 + qv 2 - 2jv,

which is still to be minimized over 3C1.

x=·O
Suppose, for the sake of example, that p = 1 and q O. If f has finite
Fig. 1.3 Fundamental solut
energy, then the integral I is finite and its minimization is straightforward. 3C 0 and :JC2.
 CHAP.
p_u'_v_
+ p(n)u'(n)v(n) p+u'+v+.
imiting values as x ~pproaches X 0
;;all that v_ = v+ for any v in 3C1,
:hat u_ u+. Varying v, it follows
:h interval, that u'(n) = O at the far
at x 0 , and is a direct consequence
), but the combination pu' remains
::m lies in 3C1 but not in X l. This }s
face I(v) was actual/y at the bottoin.
1een the same o ver the original space
attained the mínimum. The surface
t was filled in only by the function
: elements, which rest on an assumed
tiscontinuity x 0 • These estimates can
at x 0 wé permit the trial functions
p condition. Since the condition is
~ial function has to satisfy it; as long
have a continuous derivative, and
tpproximation will be good.
eneous term f has been required to
)-funétion. Physically this would be
nt load or point so urce, and mathe-
zdamental solutio.n. Therefore we try
functional
f- qv 2 - 2fv,
hat p = 1 and q O. lf f has finite
its minimization is straightforward.
1 SEC. 1.3. THE VARIATIONAL FORM OF THE PROBLEM 15
But also for f = o(xo), o< Xo < 1' the integral is finite. In this case
I(v) = J: (v') 2 dx- 2v(x0 ),
and the mínimum occurs at the ramp function v = u shown in Fig. 1.3. Again
the solution is not in 3Cl, and the hole corresponding to this ramp function
happens to fall at the bottom of the infinite-dimensional paraboloid l(v).
One more possibility: Suppose the poi~t load is at the end x 0 = n, so
that f = o(n). Then the solution is u(x) x, arid the ramp function never
levels off. In this case the solution vio/ates the natural boundary condition
u'(n) = O. Looking back at (11) this is perfectly consistent; with p = 1,
q = O, and u(x) = x, the first variation is
s: (u'v' fv) dx = s: (v' o(n)v) dx . o
for any v in 3C.1. Thus the first variation vanishes at u(x) = x, which is indeed
minimizing. The only remarkable point is that the integration by parts in
(11), and therefore the subsequent derivation of the natural boundary condi-
tion u'(n) O, falls through. Thus if f is allowed to be singular at x n,
there may no longer be a natural boundary condition at that point.
Once a o-function has been accepted as a possible choice of J, there arises
the general question: Which class ofinhomogeneous terms can be permitted?
More precisely, which space of data matcJ'¡es the solution space 3C1? Roughly
speaking, as long as I(v) remains finite for all v in X1, the minimization can
proceed. This rule accepts ó-functions and their linear combinations but not
u(x)=x 0
x=x0
x=O X=7T
Fig. 1.3 Fundamental solution for a point load; f and u not in
3C 0 and 3C2.
16 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.4.
""':-
their derivatives. For example, the dipole function f = J'(x 0 ) would give "Every continuous mapping admit:

J: fv - J: v'J -v'(x 0 ),
mations." The previous sections es
1 to u, one for the differential equat
variational equivalent (11 u 11 1 u-
and this may not be finite; v' can have an unbounded peak at x 0 and still should be ripe for numerical soluti
ha ve fini 1tf energy. The dipole is too .singular to be allowed. We begin with the differential e<
The data f which are now accepted come from the space denoted by by difference quotients. The resul
:re-t; f's derivative of order -1, that is, its indefinite integral, is in X 0 • The discrete operational form. There ar
second-order operator L takes the space X 1 into X - 1 , just as it too k Xi of this difference equation:
into 3C 0 • A suitable norm in ac- 1 is
l. To compute the local trunca ti
of a Taylor series expansion.
1 Jf(x)v(x) dx 1
2. To establish that the systerr.
(12) ll/ll-t = ~a;l llvllt · Uh depends continuously on Jh as;

There is, however, a peculiar result if f happens to be a J-function concen- Together, these two steps estab
trated at the origin, namely that fv f
O for every v in the admissible space true solution u as h - O. The poin
X 1. Such a phenomenon means that the variational principie will treat this contrast the convergence theory fot
fas zero, and the solution will be u O. (This is ju.st the ramp function of used in the next section, and througl
Fig. 1.3, with x 0 0.) This means that the solution space 3C1 matches the gence in the variational problem. t
data space x- 1 only under the following proviso: There is to be no distinction steps above would be approached
between data / 1 and / 2 , if they differ by a múltiple of the J-function at the truncation error is abandoned in fa
origin. erties (or completeness) of the trial
Finally, the solution should depend continuously onf, using the solution- proof-it is automatically present, ·
space norm for u and the data-space norm for f. The proof depends on the As usual, the interval [O, x] is di,
vanishing of the first variation for every v in X 1, in particular for v = u: by the points x 1 = ih, i =O, 1, ...
-(pu')' qu + are removed in
f no p(u')2. + qu2 = fno fu.
The right si de, from the definition of the data norm ll f 11- P is bounded by
11 f 11- 1 11 u 11 1 • The left side is obviously larger than Pminll u' 115, and this is easily The result is a discrete equation -
shown to be larger than u 11 u lit for sorne positive a. Therefore, to hold at the interior mesh points
1
allulli < 11/11-tllullt or llullt -11/11-t·
a h
1
2[ -p(xi + ~ )cu:+l
(13)
This expr~sses the continuous dependence of u on f. + q(x )U: =
1 f(x;).

Since this is a second-order equa

1.4. FINITE DIFFERENCE APPROXIMATIONS
each end of the interval. At the left e

tThis section is a digression from OUI

lt is almost an article of faith in numerical analysis that whatever can be
difference methods are so closely related th:
solved abstractly can also be solved by concrete numerical computations. The reader may discern our own preferenc
 CHAP. 1
1ole function f = Q'(x 0 ) would give
~ an unbounded peak at x 0 and still
~ular to be allowed.
j come from the spa<?e denoted by
, its in definí te integral, is in 3C 0 • The
:e 3C1 into :JC- 1 , just as it took 3Ci
f(x)v(x) dx 1 2. To establish that the system is globally stable, in other wo~ds, that
llvllt . U" depends continuously on f" as the mesh size h approaches zero.
· happens to be a .:5-function cQllcen-
0 for every v in the admissible space
e variational principie will treat this
O. (This is ju'st the ramp function of
: the solution space 3C1 matches the
proviso: There is to be no distinction
y a mtiltiple of the .:5-function at the
;ontinuously onf, using the solution-
>rm for f The .proof depends on the
ry v in 3C1, in particular for v u:
J: fu.
1e data norm 11 f 11-P is bounded by
trger than Pm 1nl 1u' 11~, and this is easily
1e positive u. Therefore,
l
r llullt <-11/11-t·
q hl2 [
1ce ofu onf
~TIONS
nerical analysis that whatever can be
y concrete numerical computations.
SEC. 1.4. FINITE DIFFERENCE APPROXIMATIONS 17
"Every continuous mapping admits a convergent series of discrete approxi-
mations." The previous sections established two continuous mappings from
fto u, one for the differential equation (11 u 11 2 Cll f !lo) and the other for its
variational equivalent (ljujj 1 u- 111/11:.. 1). Therefore, both these problems
should be ripe for numerical solution.
We begin with the differential equation Lu fand replace the derivatives
by difference quotient,s. The result is a finite linear system V'U" f"-a
discrete operational form. There are two key steps in the theoretical analysis
of this difference equation:
l. To compute the local truncation error, or discretization error, by means
of a Taylor series expansion.
Together, these two steps establish the rate of convergence of U" to the
true solution u as h ~ O. The point of our discussion is really to be able tp
contrast the convergence theory for difference equations with the techniques
used in the next section, and throughout the rest ofthe book, to prove conver-
gence in the variational problem.t It is remarkable how differently the two
steps abo ve would be approached in the variational framework: the local
truncation error is abandoned in favor of verifying the approximation prop-
erties (or completeness) of the trial functions, and stability needs no special ·
proof-it is automatically present, for finite elements.
As usual, the interval [0, n] is divided ~nto equal pieces oflength h = n/N
by the points x 1 = ih, i = O, 1, ... , N. Then the deriva ti ves in the equation
-(pu')' + qu f are removed in favor of centered difference quotients:
h/2)
The result is a discrete equation -!J."(pll."U) + qU J, which we require
to hold at the interior mesh points x 1 :
-p( X¡+ ~ )(Uf+t - Uf)+ P( X;- ~)(Uf- Uf-t)J
(13)
+ q(xJU7 = f(x¡).
Since this is a second-order equafion, it requires a boundary condition at
each end of the interval. At the left end, U~ O is the obvious choi~. At the
tThis section is a digression from our main theme, but the finite element and finite
difference methods are so closely relatedthat we need to be in a position to compare them.
The reader may discern our own preference and, if he shares it, skip this section.
 SEC. 1.4. FIJ'I
18 AN INTRODUCTION TO THE THEORY CHAP. 1

other end,' there is no unique difference rep]acement for u'(n) =O, and we operator,
shall consider two alternatives: a one-sided difference

U7v- Uj._ 1
(l4a)
h
o
and a certered difference
Again, consistency with the true conditi
(14b) u~+l u~-~_ 0 term. The centered difference is of cou1
2h - .

In the first case the difference equation (13) holds for O< ¡<N, which
u(n + h) _, u(n - h)
2h
together with the two boundary conditions produces N + 1 equations in
N+ 1 unknowns. In the second case the difference equation applies also at If the difference equation and bom
xN = n, to compensate for the extra unknown Uf.+ 1 .-These boundary condi-
rather than separately, the truncation e
tions.are easy to cpmpare when p 1 and q =O; after eliminating the un-
an example the matrices Lh displayed ab
knowns at the extreme ends, the difference equations are, respectively,
have been used to eliminate the last 1J
o Taylor series, the one-sided and cente1

~r-~o
-1

-~l(LJ CJ
2
(15a) LhUh ~-~-.::--->---2_.;.h) - f(il
h2 2
o -l 2u(n) 2u(n - h)
h2
and
o
(15b)
h2
(
Lh U h = _!_ - l
2 -1
o
.
2
.
o
2
-2
-J
o .

"JCl
2
.
UN
(]
To analyze the difference equation with variable p and q, we consider the
In this form, it would appear that t
which is completely wrong.
For an estímate ofthe error Eh=
tion which it satisfies:

local truncation error 't'h(x) which arises when the true solution u is substituted
for Uh: With the centered difference at x 11
equation are

E'iHt E
E~ =0,
This is a completely formal computation, in which u(xi ± h) and p(x; ± h/2) 2h
are exp~nded in a Taylor series aroundthe central point x 1• Because u satisfies
the differential equation, the zero-order term vanishes; this cancellation Notice that this difference equation
expresses the consistency of the difference and differential equations. The problem, except that the inhomoge¡
terms which remain are truncation error. We therefore expect
from the terms which are of order h 2
't'h = - ~~ [(pu')'" + (pu'")'] + O(h 4). -(pe~)' + qez = -
The same process applies at the boundaries. For the one-sided differ.ence ez(O) =O,
 CHAP. 1 SEC. :l.4. FINITE DIFFERENCE APPROXIMATIONS 19

nce replacement for u'(n) O, and we operator,

-sided difference
u(n)- ~n - h) = u'(n) - ~ u"(n) + ...
1}:¡_1 o
= - ~ u"(n) + O(h 3 ).

Again, consistency with the true condition u'(n) O cancelled the zero-order
o. terrn. The centered difference is of course more accurate:

tion (I 3) holds for O < i < N, which
tditions produces N + 1 equations in
the difference equation applies also at
mknown U}:¡+ 1 ••These boundary con di- •
1 and q = O; after eliminating the un-
rence equations are, respectively,
o
u(n + h) lh, u(n - h} = ~ um(n) + O(h4).
If the difference equation and boundary condition are treated together,
rather than separátely, the truncation errors Iook quite different. We take as
an exarnple the matrices Lh displayed abo ve, in which the boundary conditions
have been used to elirninate the Iast unknown. Expanding the final row by

2 ][u,
-: L_,
) e) IN-1
Taylor series, the one-sided and centered conditions yield, respectively,

u(n- h)- u(n

2Ú(n) - x~(n ~
f(n

h) - f(n)
h) = -

= - ~
~ u"(n) +

u'"(n)
O(h),

+ O(hz).
o
In this forrn, it would appear that the overall errors are 0(1) and O(h),
which is completely wrong.
2
2 -rlriJ=Cl F<?r an estímate ofthe error Eh= Uh ..:__u, we Iook at the difference equa-
tion which it satisfies:

with variable p and q, we consider the .

when the true solution u is substituted
With the centered difference at x = n, the boundary conditions in this error
equation are ·

m, in which u(x; ± h) and p(x1 ± h/2) Eí =0,

the central point xi' Beca use u satisfies
der term vanishes; this cancelJation
~ence and differential equations. The
Notice that this difference equation is analogous to the original differential
problem, except that the inhomogeneous terms now come from the local
truncation error. We therefore expect the leading terrn in E\ say h 2 e2 , to arise
frorn the terms which are of order h 2 in the local error:

-(pe~)' + qe 2 -it:[(pu')'" + (pu'")'],

undaries. For the one-sided difference ez(O) =O, e~(n) = iu"'(n).
20 AN INTRODUCTION 1'0 THE THEORY CHAP. 1 SEC. 1.4.

This solution e 2 (x) is the principal error function. To compute the next term Obviously /!!.. is a finite difference
in the error, we substitute u + h 2 e 2 into the difference problem. This yields a
truncation error which statts at h 4 , and the coefficient ofthis term is theright-
hand side in the equation for e4 • /(v) = s:·p
In short, we may recursively determine an expansion
And, almost as obviously, JA is ac1
(16) the centered trapezoidal rule, whicl
replaced by sums biased to one si<
The calculation of the error terms en is entirely mechanical; it has to be This hints at a technique wt
stopped only when there is a break in the smoothness of the solution or when finite elements-the integral /(v) i~
the boundary conditions no longer permit such an expansion. One virtue of ference quotients, and then minin
(16) is that the error varies with x in the proper way; we ha ve more than just mations of a low order of accurac
an unrealistic maximum bound over the whole interval. Furthermore, this received. Its advantages tend to d
expansion justifies Richardson's extrapolation to h = O; one computes with is demanded.
two or more choices of ~, nnd chooses a combination of the results so as to Up to this point our analysis of
cancel the leading terms in the expansion. With the centered boundary ly formal, leading to the error exp
condition in our example, the linear term he 1 vanishes identically, and the to be taken: to prove that ijh conv1
combination of Uh and U 2 h which increa·3es the accuracy from second to totically valid. [The expansion itse
fourth order is would require that ijh is an analy
well posed even for complex h; al

This extrapolation technique has been disc_ussed and verified numerically = O(f7M+ 1) as h decreases to zero.:
any number of times but has not yet been widely adopted in practice. The of estima te as in the differential p1
accuracy of the boundary conditions is its most severe limitation, particular] y tinuously on the data Jh.
in more dimensions when the boundary of the region intersects the mesh in We return to the difference e<
a completely erratic way. there exist a unique solution Uh
Applying the same ideas to the one-sided boundary condition, the first Lh nonsingular? One of the most
term he 1 arises from the O(h)truncation error at the boundary: leads at the same time to a discret
Suppose LhUh =O. Let the larg
-(pe'1 )' + qe 1 =O, the sign of Uh to ·make u: > O. T
e 1 (O) = O, e'1(n) = -!u"(n).

Naturally the centered difference is to be preferred.

Since each term is nonnegative, all
This first-order accuracy appears also if the difference equation is linked
cient qn is positive, it follows imme
with a variational problem. As always, the positive-definite symmetric system
terms lead to u:+l = u:= u:_¡.
LhUh =~fh is the Euler equation for the vanishing of the first derivative, at
and the whole argument can be· re
the minimizing U\ of
Ultimately, after enough repetiti1
LhUh =O holds only if Uh is thc
Allowing inhomogeneous bounda
With the one-sided boundary condition, this functional is show that no component uf can
discrete maximum principie, from
h = :~t [Pi-ti2(V? -h V~-~)2 + q¡(V7)2- 2j¡V~J. function (Lh)- 1 is a nonnegative mt
J By the way, a similar proof l1
 CI:IAP.
Jr function. To compute the next term
o the difference prbblem. This yie]ds a
the coefficient ofthis term is the right-
nine an expansion
is entirely mechanical; it has to be
te smoothness of the solution or ~hen
·mit such an expansion. One virtue of
: proper way; we ha ve more than just
he whole interval. Furthermore, this
olation to h = O; one computes with
a combination of the results so asto
nsion. With the centered boundary
erm he 1 vanishes identically, and the
creases the accuracy from second to
=u+ O(h 4 ).
1 disc_ussed and verified numerically
Jeen widely adopted in practice. The
its most severe limitation, particularly
y of the region intersects the mesh in
e-sided boundary condition, the first
m error at the boundary:
qe 1 =O,
n) = -fu"(n).
>e preferred.
m if the difference equation is linked
he positive-definite symmetric system
e vanishing of the fir_st derivative, at
, this functional is
1
SEC. lA. FJNJTE DIFFERENCE APPROXIMATIONS 21
Obviously /t. is a finite difference analogue of the true quadratic functional
l(v) = s:·p(v') 2 + qv 2 - 2fv.
And, almost as obviously, /t. is accurate only to first order. Instead of using
the centered trapezoidal rule, ·which is of second order, the integral 1 has been
replaced by sums blased to eme side.
This hints at. a technique which ·Iies in between finite differences and
finite elements-the integral I(v) is approximated by a sum /t. involving dif-
ference quotients, and then minimized. It is an easy way to derive approxi-
mations of a low order of accuracy, and deserves more analysis than it has
received. Its advantages tend to disappear, however, when higher accuracy
is demanded.
Up to this point our analysis ofthe difference equation has been complete-
ly formal, leading to the error expansion L; hnen. There is now a second step
to be taken: to prove that Uh converges to u and that the expansion is asymp-
totically valid. [The expansion itself cannot converge for a finite h, since this
would require that Uh is an analytic function of h, and that the problem is
well posed even for complex h; all one hopes to pro ve is that U'' -
M
L; hnen
o
= O(hM+ 1) as h decreases to zero.] This second step demands the same kind
of estímate as in the differential problem: the solution Uh must depend con-
tinuously on the data fh.
We return to the difference equation and ask the first question: Does
there exist a unique solution Uh for every fh? Equivalently, is the matrix
Lh nonsingular? One of the most effective proofs of the invertibility of Lh
leads at the same time to a discrete maximum principie, as follows.
Suppose LhUh =O. Let the largest component 1 U71 be the nth, and choose
the sign of Uh to make U~ >O. Then the difference equation at xn is
Since each term is nonnegative, all three must vanish. Ifthe zero-order coeffi-
cient qn is positive, it follows immediately that U~ = O. In any case, the other
terms lead to U~+ 1 = U~ = U~- 1 • Thus these components are also maximal,
and the whole argument can be repeated with n - l or n + 1 in 'place of n.
Ultimately, after enough repetitions, it follows that U~ = Vi = O. Thus
LhUh = O holds only if ljh is the zero vector, and Lh must be invertible.
Allowing inhomogeneous boundary conditions, the same argument would
show that no compopent uf can be larger than both Vi and U'N. This is a
discrete maximum principie, from which it follows that the discrete Green's
function (Lh)- 1 is a nonnegatil'e matrix.
' By the way, a similar proof leads to Gerschgorin's theorem in matrix
22 AN INTROOUCTION TO THE THEORY CHAP. 1 SEC. 1.4.

theory: Every eigenvalue A of a matrix A líes in at least one of the circles For each choice of difference equa
It is. not an automatic conseque1
continuous inequality is necessary ,
In this sense the theory of differeJ
Choosing A to be either of the matrices h Lh displayed in· (15), all eigenvalues
2 an enormous variety of possible di
satisfy lí~ - 21 2. Thus the theorem does not rule out the possibility that or-less new proof of (18). As in th
A = Ois,·an eigenvalue, that is, that Lh is singular; it was at this point that the inequality-as true and concentrate
repetitions of the argument at i n 1, n - 2, ... , 1 were needed. in one-dimensional problems is s1
Gerschgorin's theorem becomes perfectly useless in a fourth-order (18) asserts the stability of the diJ.
problem. The leading coefficients in the simplest case are A u = 6, Ai,i± 1 depends continuously on the disc
-4, and At,t±z = 1, and the Gerschgorin circles are j A 61 < 10. Since To establish that Uh converges
A O lies inside these circl~s, the Gerschgorin argument fails to prove even of numerical analysis: Consisten
that A is semidefinite. This difficulty simply reflects the absence of a maximum theorem is proved in two steps:
principie for fourth-order problems. If we compare u" O with uav> = O,
for example, it is obvious that straight lines attain their extrema at the ends l. The error Eh satisfies the :
of the interval, whereas cubics need not. therefore by stability it depends e
The maximum principie, when it holds, can· be made to yield a simple but the local truncation error:
proof of convergence. We prefer instead to preserve a close analogy between
differential and difference equations, by· discussing the discrete inequality
which corresponds to 11 u lb < C 11 f llo· The statement of such an inequality
2. By consistency, which was
requires, first of aH, a redefinition of the norm to make it apply to grid func-
Eh(O), and A.1tE\ the right side aJ
tions. An obvious choice for the discrete el)ergy is
vergence is proved.

We call attention to one more p

For the square of the 2-norm, we introduce the energy in the function and , ti ve o'f u, the error estímate in 1
its first and second forward-difference quotients: IJEhll 2 < Ch 2 llu!l 4 • With a little t

These sums extend only over the grid points at which they are defined; the Thus the convergence is of order)
forward-difference quotient A+ft = (/;+ 1 - J¡)/h makes no sense at the Iast whenever fis in3C 0 • This is the sa
grid point. finite element method given in Sec
There is one other new feature to be introduced, inhomogeneou& bound- easier.
ary conditions. For the two-point boundary-value problem, the continuous Altogether, it appears that a
dependénce of u onf and on the boundary data is expressed by difference equations is possible b1
Iems, that is, for partial different
(17) llulb < C{IJ/IIo + ju(O)j + ju'(n)l). proof in the numerical analysis l
principie. Without this principie ti
For the finite difference equation, in which An denotes whichever boundary adequate for special problems, b1
operator is applied at the right-hand end, the corresponding inequality is oped is so delicate that it looks ex1
that a single differential equation
(18) approximations, above all at a cun
 .<::HAP.
·ix A lies in at least one of the circles
s h 2 Lh displayed in (15), all eigenvalues
1 does not rule out t~e possibility that
is singular; it was at f)lis point that the
1, n- 2, ... , 1 wete needed.
perfectly useless in a fourth-order
he simplest case are Au = 6, Ai,t±t
sorin circles are 1 A. - 61 1O. Since
schgorin argument fails to prove even
nply reflects the absence of a maximum
f we compare u" = O with u<iv> = O,
t lines attain their extrema at the ends
::>t.
holds, can· be made to yield ·a simple
d to preserve a close analogy between
by· disci.Jssing the discrete inequality
1• The statement of such an inequality
te norm to make it apply to grid func-
:te en.ergy is
oduce the energy in the function and
quotients:
points at which they are defined; the
i+t - / 1)/h makes no sense at the last
e introduced, inhomogeneous bound-
ndary-value problem, the continuous
ry data is expressed by
l u(O) 1 + !u'(n) 1).
hich A,. denotes whichever boundary
end, the corresponding inequality is
1 SEC. 1.4. FÍNITE DIFFERENCE APPROXIMATIONS 23
For each choice of difference equation, this is the basic estímate to be proved.
It is not an a u toma tic consequence of ( 17), .although it ·implies (17): The
continuous inequality is necessary but not sufficient for the discrete inequality.
In this sense the theory of difference equations is the more difficult; there is
an enormous variety ofpossible difference schemes, ánd each requires a more-
or-less new proof of (18). As in the continuous problem, we shall accept the
inequality as true and concentrate on its implications; the technique ofproof
in one-dimensional problems is summarized by Kreiss [K7]. The inequality
(18) asserts the stability of the difference equation: The discrete solution Uh
depends continuously on the discrete data f\ uniformly in h.
To establish that Uh converges to u, we need the most celebrated theorem
of numerical analysis: Consisten~y and stability imply com•ergence. This
theorem is pro ved in two steps:
l. The error Eh satisfies the same difference equation (13) as Uh, and
therefore by stability it depends continuously on its data, which ·is nothing
but the local trunéation error:
2. By consistency, which was reflected íri the Taylor expansions of t\
Eh(O), and AnEh, the right side approaches zero as h - t O. Therefore, con-
vergence is proved.
We call attention to one more point. Beca use th involves the fourth deriva-
\ tive ..of u, the error estímate in the centered difference case is effectively
11 Eh 11 2 Ch 2 ll u !1 4 • With a little extra patience this can be reduced to
11 Eh llo < C'h 2 ll U llr
Thus the convergence is of arder h 2 whenever u is in :JC2, or in other words
whenever fis in :JC 0 • This is the same rate of convergence as in the simp1est
finite element method given in Section 1.6, but the proof there is very much
easier.
Altogether, it appears that a satisfactory theory for one-dimensional
difference equations is possible but not trivial. In multidimensional prob-
1ems, that is, for partial differential equations, almost every convergence
proof in the numerical analysis Jiterature has depended on a maximum
principie. Without ~his principie there have been a few ad hoc arguments,
adequate for special problems, but the general theory now being devel-
oped is so delicate that it looks extremely difficult to apply. The problem is
that a single differential equation allows an enormous variety of difference
approximations, above all ata curved boundary. In contrast, the variational
 24 AN INTRODUCTION TO THE THEO~Y CHAP. 1 SEC. 1.5. THE F

metho~s are governed by stricter rules, and it is just these restrictions which Clearly these functions lie in 3C1; t11
permit a more complete theory. condition (vh)'(n) O, which is not n1
We shall now concentrate exclusively on this theory-the construction The weights q1 are to be determined
and convergence of finite elements. nality property of the eigenfunctions,

1.5. THE\ RITZ METHOD ANO LINEAR ELEMENTS

l(vh) t[q¡A. ..
l J

In this section we begin on the finite element method itself. The general where 11 p(j - ~) 2 + q. The minim
framework is already established, that there is a choice of approximating to the optimal values
either the individual terms in the differential equation or the underlying
variational principie. The finite element method chooses the latter. At the
same time, the discrete equations which arise in the variational approximation
are effectively difference equations.
In variationarform, the problem is to minimize the quadratic functional Therefore, the Ritz approximation is

J(v) =· J: [p(x)(v'~x)) 2 + q(x)(v(x)) 2 2f(x)v(x)] dx

over the infinite-dimensional space JCJ,. The Ritz method is to replace X1

In this case the system oflinear equati
~ in this variational problem by afinite-dimensional subspace S, or more precise/y
coordinates Q1 was trivial to sol ve; it:
by a sequen ce offinite-dimensional subspaces Sh contained in JC 1. The elements
the eigenfunctions are orthogonal. F
vh of Sh are called trialfunctions. Because they belong to JCJ,, they satisfy the
true solution
essential boundary condition vh(O) O. Over each space Sh the minimization
of 1 leads to the solution of a system of simultaneous linear equations; the u í:·1
number of equations coincides with the dimension of S". Then the Ritz
approximation is the function uh which minimizes 1 over the space Sh:
onto the jirst N eigenfunctions. '
for all v" in S 11
• In this example it is easy to comp
the Ritz sequence of approximation
The fundamental problems are (1) to determine uh, and (2) to estímate the fPN, uh is improved by the term AÑ 1(j
distance between uh and the true solution u. This section is devoted to prob- approach zero very rapidly as N·~ ex
lem 1 and the next section to problem 2. tion is atmost AÑ 1 ll/llo· In a more re.
We begin with two examples which illustrate the classical Ritz method. will not be known, but if the geomet1
The subspaces Sh will not have the special form associated with the finite and cosines as trial functions is still e
element method; instead, each subspace in the sequence will contain the have shown how the fast Fourier tra
preceding one. within reasonable bounds.) For me
Suppose in the first example that the coefficients p and q are constant, and finite elements.
let Sh be the subspace spanned by the first N= 1/h eigenfunctions of the In the second example we keep p
continuous problem. The trial functions-the members of Sh-are then the tions which are polynomials: ·
linear combinations

Again these satisfy the essential bom

 CHAP.
~s, and it is ]ust these restrictions which
ively on this theory-the construction
AR ELEMENTS
tite element method itself. The general
at there is a choice of approximating
ifferential equation or the underlying
ent method chooses the latter. At the
1 arise in the variational approximation
s to minimize the quadratic functional
r(x)(v(x)) 2 - 2f(x)v(x)] dx
C1. The Ritz method is to replace X 1
:nensional subspace S, or more precise/y
1aces Sh contained in X1. The elements
tse they belong to X1, they satisfy the
1
•Over each space Sh the minimization
of simultaneous linear equations; the
the dimension of Sh. Then the Ritz
ninimizes 1 o ver the space Sh:
for all ?/' in Sñ.
determine uh, and (2) to estímate the
on u. This section is devoted to prob-
2.
1 iUustrate the classical Ritz method.
,ecial form associated with the finite
tce in the sequence will contain the
: coefficients p and q are constant, and
: first N= I/h eigenfunctions of the
Is-the members of Sh-are then the
, qFYIn
J
. (.J -
2 sm 1)
x.
"2"
1 SEC. 1.5. THE RITZ METHOD AND LINEAR ELEMENTS 25
Clearly these functions Iie in X 1; they satisfy even the natural boundary
condition (vh)'(n) = O, which is not necessary.
The weights qi are to be determined so asto minimize l. From the orthogo-
nality property of the eigenfunctions, the integral has the special form
where li p(j ~) 2 + q. The minimizingequation a¡¡aqi Oleadsdirectly
to the optimal values
j 1, ... ,N.
Therefore, the Ritz approximation is
In this case the system of linear equations a¡¡ aq i = Odetermining the optimal
coordinates Qi was trivial to solve; its coefficient matrix is diagonal, because
the eigenfunctions are orthogonal. Furthermore, uh is the projection of the
true solution
u
onto the first N eigenfunctions.
In this example it is easy to compute how much is gained at each step in
the Ritz sequence of approximations. By including the Nth trial ·function
rpN, uh is improved by the term AÑ 1(f, rpN)'PN· Forsmoothfthesecorrections
approach zero very rapidly as N...- oo, and even for an arbitrary fthe correc-
tion is at most AÑ 1II f llo· In a more realistic problem the exact eigenfunctions
will not be known, but if the geometry remains simple, then the use of sines
and cosines as trial functions is still of great importance. (Orszag and others
have shown how the fast Fourier transform keeps the computational effort
within reasonable bounds.) For more complicated geometries we prefer
finite elements.
In the second example we keep p and q constant, and choose trial func-
tions which are polynomials:
Again these satisfy the essential houndary condition vh(O) O, but this time
26 AN INTRODUCTION TO "[HE THEORY CHAP. 1 SEC. 1.5. THE

· not the natural boundary ·condition. In this case In our example the admissible sp
ous. This rules out piecewise cons
choice for Sh is tite space of functl
[(j l)h,jh], continuous at the node.
tive of such a function is piecewise e
Differentiating 1 with• respect to the parameters qJ., we find a system of N thus Sh is a subspace of 3C1. We sh•
linear e ~.uations for the optimal parameters Q 17 . .. , QN:
1 elements.
For j l, ... , N, let rp1 be the
KQ=F.
particular node x jh, and vanish(
The unknown vector· is Q ( Q 1 , • • • , QN), the components of F are the functions constitute a basis for the st
"moments" F1 f fx 1 dx, and the coefficient matrix K is a mess: written as a combination

K,1 J: [p(ix 1- 1 )(jxF-I) + qx xi] dx

1
vh(x) =

_ pijnt+J-1 qn~+j+ 1
Notice the following irnportant fact
- i+j- 1 + i+j+ l. the val u e of v at the jth node x = jh
the nodal values of the function, tht
The matrix whose en tries are (i j + +1)- 1 is the notorious Hilbert matrix,
physical signíficance; they are the. J
with which computations are virtually impossible. The eigenvalues of this
of the string at the nodes. Notice al:
matrix are so out of scale-the matrix is so ill-conditioned-that normally
basis, because each rp' is identically
even for N ~ 6 roundoff error destroys all information in the data f With
In fact, it is obvious that ~-t is 01
the other terms of K included, the situation is even worse. The difficulty is
these functions doesn't vanish, the
that the powers xi are nearly linearly dependent; all of them ha ve essentially
not quite orthogonal, it is only a<
all their weight in the neighborhood of x n. Numerical stability will
With coefficients norrnalized to
depend on choosing a more strongly independent basis for the subspace.
The remedy is to cope with the excessive sensitivity of the problem befare
introducing the data, by orthogonalizing the original trial functions rp1
In most cases this means introducing Legendre or Chebyshev polynomials
xi. J(vh) = s: (((V")'

as a new basis for the polynomial space Sh. Such a basis is useful on an ínter- With vh :E qirp1,. this integral is
val, and remains so in more dimensions provided the geometry is very simple. q 1 , • • • , q N• and it can be computec
For a general domain, however, these orthogonal polynomials again become
unworkable.
We turn now to the construction of a finite element subspace Sh. The
domain-in this case the interval [0, n]-is divided into pieces, and on each
piece the trial functions vh are polynomials. Sorne degree of continuity is
imposed at the boundaries between pieces, but ordinarily this continuity is
no moré than is required in order that

l. The functions vh will be admissible in the variational principie.

2. The quantities of physical interest, usually displacements, stresses,
or moments, can be conveniently recovered from the approximate solution
uh.

Jt is very difficult to construct piecewise polynomials if too much con- o

tinuity is required at the boundaries between the pieces. Fig. 1.4 Piecewi
y 27
CHAP.
In this case
: parameters qi, we find a system of N
tmeters QP ... , QN~'\
'=F.
· . , QN), the components of F aré the
:oefficient matrix K is a mess:
q1Cl+j+ l
i+J+I
+- !)- 1
is the notorious Hilbert matrix
lly impossible. The eigenvalues of thi~
·ix is so ill-conditioned-that normally
)ys all information in the data f With
tuation is even worse. The difficulty is
dependent; all of them ha ve essentially
d of x 11:. Numerical stabiJity will
y independent basis for the subspace.
:essive sensitivity of the problem before
ing the original trial functions rpi = xi.
~ Legendre or Chebyshev polynomials
:e Sh. Such a basís is useful on an inter-
s provided the geometry is very simple.
orthogonal polynomials again become
of a finite element subspace Sh. The
:]-is divided into pieces, and on each
omials. Sorne degree of continuity is
ieces, but ordinarily this continuity is
ible in the variational principie.
!rest, usually displacements, stresses,
•vered from the approximate solution
:ewise polynomíals if too much con-
tween the pieces.
1 SEC. 1.5. THE RITZ METHOD AND LINEAR ELEMENTS
In our example the admissible space i·s JC1, whose members are continu-
ous. This rules out piecewise constant functions. Therefore, the simplest
choice for Sh is the space of functions which are linear over each interval
[(} I)h,jh], continuous at the nodes x jh, and zero at x =O. The deriva-
tive of such a function is piecewise constant and obviously has finite energy;
thus Sh is a subspace of JC1. We shall refer to these trial functions as linear
elements.
For j I, ... , N, Iet rp1 be the function in Sh which equals one at the
particular node x = jh, and ·vanishes at all the others (Fig. 1.4). These roof
functions constitute a basis for the subspace, since every member of Sh can be
written as a combination
Notice the following important fact about the coefficient qi: It coincides with
the value of v at thejth node x = jh. Since the coordinates qi are nothing but
the nodal val ues of the function, the optimal coordinates Qi will have direct
physical significance; they are the. Ritz approximations to the displacement
of the string at the nodes. Notice also that the rp~ form a local basis, or patch
basis, because each rp1 is identicalJy zero except in a region of diameter 2h.
In fact, it is obvious that 'P1-t is orthogonal to rp~+ 1 , since whenever one of
these functions doesn't vanish, the other does. Thus although the basis is
not quite orthogonal, it is only adjacent elements which will be coupled.
With coefficients normalized to p q = 1, the problem is to minimize
With vh :2; qirp1, this integral is a quadratic function of the coordinates
q., . .. , qN, and it can be computed over one subinterval at a time. On the
o
Fig. 1.4 Piecewise linear basis functions.
28 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.5. TI

jth subinterval, the function vh goes linearly from qi_ 1 to qi and (vh)' = single element is recorded once a
(qi- qi_ 1 )/h. (By convention q 0 = 0.) Therefore, matrix:

A slightl~. longer compu..tation gives

Now the summation over the eler

sembly of the global stiffness matriJ
are added together, after being ¡:
These terms correspond to a single piece of string, with linearly varying dis-
placement. For the whole string, the second-degree term in I(vh) is the sum global array. The matrix associa1
unknown q 0 is to be discarded beca

This is nota particularly convenient form for the result. We would prefer
to ha ve it in the matrix form qrKq, in other words (Kq, q), beca use it is the +_!_
h
matrix K which we shall eventually need. The reason is this: The expression
l(vh) is quadratic in the parameters q = (ql' . .. , qN), of the form
2 -1 o o
-1 2 -1
The mínimum of such an expression occurs (as we know from the matrix 1 o -1 -1
-¡:¡
case, Section 1.3, where we set a1¡aqm =O) at the vector Q = (Ql' ... , QN)
determined by
o -1 2
o o _!._1
KQ=F.
Again the relationship between thc
This is the system we shall have to solve, and therefore all we need to know
is the matrix K and the vector F.
The best plan is to find the contribution to K from each "element," that is,
each piece of the string. Therefore, we go back to

The integral of the undifferentiat1

and record the right side (the value of the integral) in the matrix form mass matrix K 0 (later denoted b
process:

h
The matrix k 1 is an element stiffness matrix. lt represents a computation which Ko=6
needs to be done only once, since it is independent ofthe particulardifferential
equation. Similarly, the calculation of the zero-order term f (vh)2 dx o ver a
 CHAP. 1
SEC. 1.5. THE RITZ METHOD AND LINEAR ELEMENTS 29
linearly from q j - 1 to q j and (vh)' =
lrherefore, · single ·element is recorded once and for all in terms of the element mass
matrix:

3h(qj-I
2 + qj_Iqj + qj2)-( )h(2
- qj_Iqj 6 1
1)(qj-I)
2 qj

::e of string, with linearly varying dis-
xond-degree term in I(vh) is the sum
= (q}_lq}k, (~;-')-
Now the summation over the elements j = 1, ... , N is replaced by an as-
sembly of the global stiffness matrix K. This means that the element matrices
are added together, after being placed into their proper positions in the
global array. The matrix associated with J: ((vh)') 2 dx, recalling that the
unknown q 0 is. to be discarded because· ofthe essential boundary condition, is

-1
. form for the result. We would prefer -1 1
other words (Kq, q), because it is the
d. The reason is this: The expression -1
= (q~' · .. , qN), of the form
-1
2 -1 o o
-1 2 -1 o
•ccurs (as we know from the matrix 1
= O) at the vecto. r Q = (Q 1' · · · ' QN ) -¡¡ o -1 -1 o
o -1 2 -1
F. o o -1
Again the relationship between the matrix and the integral is that
e, and therefore. all we need to know

on to Kfrom each "element," that is,

~oback to

The integral of the undifferentiated term (vh)2 is given by qTK0q, where the
he integral) in the matrix form mass matrix K 0 (later denoted by M) is formed by the same assembling
process:

4 1 o o
1 4 o
~x. It represents a computation which h
Ko=6 o o
!pendent ofthe particulardifferential
~e zero-order term f (vh)2 dx o ver a
o 4
o o 2
30 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1:5. THE

The req!Jired matrix K is now the sum K 1 + K 0 • These matrices need not be bling the mass matrix:
assembled all at one time and stored; instead the entries can be computed
when they are required in the process of solving the final system, KQ = F.
lt remains to compute the term J fvh, by which the inhomogeneous load
f enters the approximation. This integral is linear in the coordinates qi:

K0 differs from K 0 only in that the .

general f(O) =/= O.
if we define the load vectot F to ha ve components This result approximates the true
would be obtained by exact integrat
FTq; the mass matrix K 0 gives the coc
i':
In practice these numbers are computed, just like the stiffness and mass ma- · l)h)-
trices, by integrating over one element ata time. Suppose that over the jth
interval The difference between Fi and the
to estímate. If f is linear, then of cot
f 1 coincides withfitself. For a quad·
If f(x) = x 2 , then
When these integrals are summed, the coefficient of a given q k is Fk pk +
a k+ 1 • One sees he re in the most rudimentary form the bookkeeping which
must be carried out by the computer. A gíven nade kh enters as the right Fi = fjh x 2
(- j + 1 + ~) dx -!
(j-l)h
endpoint of the kth subinterval, leading toa term pkqk in the integral, and
also as the left endpoint of the next subinterval, leading to etk+ 1 qk. The as- whereas
sembly subroutine must be aware that both tttese subintervals are incident
on the kth node, and combine the results. (There seems to be no doubt
that, starting from scratch, it takes longer to program finite elements than
finite differences for a two-point boundary-value problem; again the real Therefore, the error in numerical in
benefit is felt first in more than one dimension.) arbitrary smooth f, this will be the h
For arbitrary f these integrals cannot be computed exactly, and sorne There does exist an integration
numerical quadrature will be necessary. One possibility is to approximate Fj, which gets also the quadratic (anc
f by linear interpolation at the nodes. In other words,/is replaced by its piece- Fi = h(fj-t + 10/j + /j+ 1 )/12 and is
wiseHnear interpolate / 1 = l:fktpí(x), where fk is the value off at the node [H3]. It can be derived by a quadratic
x = kh. Then the integral J/ 111' involves the same inner products J tpÍtp~ which val [(j l)h, (j + l)h], · fÓilowed t
were computed earlier in forming the mass matrices k 0 • Over the jth interval This derivation is unnatural, howevj
of computing once and for all an in
over each such interval twice, once 1
polate at the nades xj-n xi, xj+t' an
interpolate at xi, xj+P xj+l' This is
since the only difference between this and the integration of (vh)Z is that one efficient formula in a certain class is 1
pair of coefficients qj-I and qi is replaced by the two nodal values off process; given complete freedom te
Again the computer will su m these results from j = 1 to j N by assem- case, finite differences can win. But t
 CHAP.
1 K 1 + K 0 • These matrices need not be
: instead the entrí~s can be computed
; of solving the final system, KQ = F.
f'vh, by which the inhomogene6us load
ral is linear in the coordina tes q.J.·
J: frp' dx = prq,
::omponents
d, just like the stiffness and mass ma-
t ata time. Suppose that over the jth
coefficient of a given qk is Fk = pk +
1entary form the bookkeeping which
A gíven node kh enters as the right
g to a term Pkq k in the integral, and
Jbinterval, leading to rxk+ 1 qk. The as-
both these subintervals are incident
sults. (There seems to be no doubt
1ger to program finite elements than
1dary-value problem; again the real
1ension.)
1ot be computed exactly, and sorne
y. One possibility is to approximate
•ther words,f is replaced by its piece-
where fk is the_ value off at the node
; the same inner products f rpirp' which
1ss matrices k 0 • Over the jth interval
td the integration of(vh)2 is that one
d by the two nodal values off
sults from j = 1 to j = N by assem-
1 SEC. L5. THE RITZ METHOD AND LINEAR ELEMENTS 31
bling the mass matrix:
K0 differs from K 0 only in that the zeroth row has to be retained, since in
general f(O) -=F O.
This result approximates the true linear term P q, that is, the one which
would be obtained by exact integrations. We denote the approximation by
frq; the mass matrix K 0 gives the coefficients in the approximate load vector
f:
Fi = ~ [f((j- l)h) + 4f(jh) + f((j + l)h)].
The difference between Fi and the true value Fi = f frpj dx is not difficult
to estima te. lff is linear, then of course the two agree, since the interpolate
fr coincides with f itself. For a quadratic f, however, a discrepancy appears.
lf f(x) = x 2 , then
jh . f(j+ l)h ( ) h3
Fi= f x 2 (-j+ 1 + Z)dx+ . x 2 j+ 1- Z =j2h 3 +6'
(j-1 )h ¡h
whereas
Therefore, the error in numerical integration is h 3 /6 = h 3 f"(jh)/12. For an
arbitrary smooth f, this will be the leading term in Fi - Fi.
There does exist an integration formula, with the same simplicity as
Fi, which gets also the quadratic (and even cubic) terms correct. It is given by
Fi = h(fi-I + 10fi + fj+z)/12 and is known as Collatz's Mehrstellenverfahren
[H3]. It can be derived by a quadratic interpolation off over the double mter-
val [(j- l)h, (j + I)h], fÓIIowed by exact integration of Fi = f f1 rp~ dx.
This derivation is unnatural, however, for finite element programs. Instead
of computing once and for aii an integral over [jh, (j + I)h], we have to go
over each such interval twice, once to evaluate Fi using the quadratic inter-
polate at the nodes xj-P xi, xj+P and then again to evaluate Fj+t using the
interpolate at x J., x J·+z, x J'+?·
-
This is a typical instance in which the most
efficient formula in a certain class is not found by the finite element assembly
process; given complete freedom to find the best formula in each special
case, finite differences can win. But the essential point for complicated prob-
 í'

32 AN INTRODUCTlON TO THE THEORY CHAP. 1 SEC. 1.5. THE RI

lems is that the finite element formula is not far from optimal, and it is p>O,
generated in a systematic and painless way by the computer.
In practice, the replacement off by its interpola te / 1 has been superseded
by direct numerical integration. Over each subinterval, fvh is integrated by
a standard quadrature formula can be zero only if "E qlp~ is identicall~
q¡ o.
In fact, the global stiffness matrix K
matrixL\ or rather hL", ofthe previous
The most common choice, and the most effi.cient, is Gaussian quadrature, leading terms are identical, both being
for example with equal weights wt at the two symmetrica11y placed evaluation with weigh ts 2, -1. The zero-or
points et {j + 1/2 ± 1/,.J'J")h. This is again exactly correct for cubic nal entries of L". In K it appears also
knowns, and is "smoothed" with the wt
f, and therefore it will automatica11y attain the same accuracy as the Collatz
ei
formula described above. The one-point Gauss rule, with at the midpoint rule. We reemphasize that once the •
the discrete forms of different terms in
of each interval, is already suffi.cient to maintain the accuracy inherent in ·
linear trial functions: Jt yields F1 h(f1_ 1í 2 + h+ 112 )/2. [The trapezoidal Ritz method is a "package deal," and
rule, with equal intervals, gives exactly the right side F1 = hf(jh) ofthe to make independent decisions about <
simple three-point difference equation of Section 1.4.] In particular, the treatment of bom
The result of any numerical integration is to replace the true linear terms the alternatives in the difference equati<
FT q by sorne approximate expression ft q, still linear in the unknowns tion, of which order of accuracy, is
element equation. Looking at the Jast r
qp • • • 'qN•
The same ideas apply to the quadratic terms, if the coeffi.cients p(x) and x = n in the case p q = 1 is
q(x) in the differential equation actually depend on x. The integrals of
p(x)((vh)')2 and q(x)(vh)2 are again computed on each subinterval by a
numerical qtiadrature. The assembled results are stored as approximate
stiffness matrices, whose quadratic forms are el ose to the ttue integrals qrK 1 q
and qT K 0q. We shall assume for the present thcit all these integrals are com- Substituting the true value u(x1) for
puted exactly, and study in Secti~n 4.3 the effect of errors in numerical series about x n:, the truncation err
integration.
Now we try to assemble our own ideas. Writing K for the sum K 1 +K 0, h , + h '"
u, - 2u
2
6u _
3
h uiv
24
+ ~(3u
6
the computations so far have produced the formula
~¡+!!.
2 (

This is the discrete expression to be minimized in the Ritz method. Observe using the differential equation -u"+
1 that it is in exactly the standard variational form; the minimizing vector Q O. In terms of difference equations, tl
is determineO. by the linear equation happens to be third-order accurate.
The final step in compúting the fini
KQ ==F. the linear system KQ =F. We prop
methods, which are ·preferred to ite1
We refer to this as the finite element equation. lts solution is the central step majority offinite element programs. (1
in the numerical computations, and if h is small, it will be a large system. The fall of iterative methods during the 1
matrix K is guaranteed to be positive definite, and therefore in vertible; since of hard work and good mathematic
 \ CHAP.
1Ia is not far from optimal, and it is
: way by the computer.
)y its interpola te / 1 has been superseded
ea_ch subinterval, fvh is integnited by
1ost efficient, is Gaussian quadrature,
te two symmetrically placed evaluation
is is again exactly correct for cubic
ttain the same accuracy as the Collatz
1t Gauss rule, with ei at the midpoint
to maintain the accuracy inherent in
h(h-1/2 + h+ 1/2)/2. [The trapezoidal
tly the right side Fj = hf(jh) of the
of Section 1.4.]
~tio_!l is to replace the true linear terms
rt FT q, still linear in the unknowns
atic terms, if the coefficients p(x) and
tally depend on x. The integrals of
omputed on each subinterval by a
j results are stored as approximate
1s are close to the true integrals qTK1 q
esent that all these integrals are com-
·.3 the effect of errors in numerical
eas. Writing K for the sum K 1 +K
0'
the formula
nimized in the Ritz method. Observe
ional form; the minimizing vector Q
F.
tation. Its solution is the central step
is smaii, it will be a large system. The
·finite, and therefore invertible; since
1 SEC. 1.5. THE-RITZ METHOD AND LINEAR ELEMENTS 33
p >0,
can be zero only if ~ qlp~ is identically zero, and this happens only if every
qj =O.
In fact, the global stiffness matrix K is remarkably like the finite difference
matrix L\ or rather hL\ ofthe previoussection. With constant coefficients, the
leading terms are identical, both being proportional to the second difference
with weigh ts ~ 1, 2, -l. The zero-order term qu ente red only the diago-
nal entries of Lh. In Kit appears also in the coupling between adjacent un-
knowns, and is "smoothed" with the weights 1, 4, 1, reminiscent ofSimpson's
rule. We reemphasize that once the approximating subspace Sh is chosen,
the discrete fornú of different terms in the equation are al! determined. The
Ritz method is a "package deal," and neither requires nor permits the user
to make independent decisions about different parts of the problem.
In particular, the treatment of boundary conditions is fixed, and recalling
the alterna ti ves in the difference equation, it is natural to wonder which condi-
tion, of which order of accuracy, is chosen "automatically" by the finite
element equation. Looking at the .last row of K, the equation at the boundary
x = n in the case p = q = 1 is
Substituting the true value u(x) for qP and expanding u and fin Taylor
series about x = n, the truncation error is
u1 - Tu
h 11 + 2
h 111
3
h iv + h (3 u- hu1 + Tu
2
h 11)
6 u - 24 u
6
- !!:__ f + h2 JI - h3 ¡~~ + ... ~ h3 u"(n) + ...
2 6 24 24
using the differential equation -u"+ u= f and boundary condition U 1(n) =
O. In terms of difference equations, this means that the boundary condition
happens to be third-order accurate.
The final step in compúting the finite element approximation uh is to solve
the linear system KQ = F. We propose to discuss only direct elimination
methods, which are preferred to iterative techniques in an overwhelming
majority offinite elementprograms. (It is fascinating to reflect on the rise and
fall of iterative methods during the last generation. A tremendous amount
of hard work and good mathematics went into the development of over-
34 AN INTRODUCTION TO THE THEORY CHAP. 1· SEC. 1.5. THERI

relaxati9n and alternating direction methods; this was a dominant theme of a lower triangular matrix times an
in numerical analysis. Now these methods are increasingly squeezed between which is the solution Q we want to cot
elimination and devices Iike the fast Fourier transform; the latter is unques- the two triangular matrices are easy to i
tionably more efficieñt when the geometry and the equation are appropriate, geneous term as it stands after elimim
and otherwise-especially when the same linear system is to be sólved for after back substitution. (If there are r
many rigl)t-hand sides, as in design problems-elimination is convenient with different data bpt the same stiffne
and straightforward.) should be stored.)
We want to discuss very briefly the theory of Gaussian elimination. Now we consider what is special a
In applying this familiar algorithm toa general coefficient matrix K, the first is known, as in our example, to be s~
1
unknown Q is eliminated from the last N- 1 equations, then Q 2 is elimi-
nated from the last N - 2 equations, and finally QN_ 1 is eliminated from
agonal. The first point is that the proc1
carried out, and the factorization K= 1

(~ ~). since Q, could not t

the last equation. The system KQ = F is transformed in this way toan equiva-
lent system with K=
by subtracting a multiple 12 , 1 of the fl

~1N)(~1)
u11
each of the matrices in the upper left e
UQ= 2N • ~ F'.
(

UNN ºN

should ha ve. a nonzero determinan t.

The unknowns Q¡ are now determined by back substitution, solving the last
determinants are all positive, and the1
equation for QN, the next to last for QN_ 1 , and so on.
with no exchanges of rows. In fact, the
.It is important to understand in matrix terms what has happened. Suppose
u11 u22 . . . uii' so that the pivot elerr
we begin to undo the elimination process by adding back to the last equation
of U, are alf positive.
the multiple IN,N- 1 of the (N- l)st equation, which was subtracted off
Something more is required for th<
in the process of eliminating QN_ 1 • Next we add back to the last two equations
ically stable: The pivots U¡¡ must be
the multiples IN- 1 ,N_ 2 and IN,N- 2 of equation N- 2, which were subtracted
large. Otherwise, the information' in ti
in the elimination of QN:.. 2. Final~y, we recover the original system KQ = F,
stroyed as the algorithm progresses. w~
having added back to each equatio-n i the multipJes l¡¡ of the previous equa-
of these pivots, and therefore sorne ro u
tions, j = 1 through j = i - 1, which were subtracted in the elimination of
sic sensitivity of K to small perturbatio
Q 1 through Qt_ 1 • In matrix terms the system KQ = F has been recovered by
roughly the ratio of its largest eigenval
multiplying the transformed system U Q = F' by the matrix
of Chapter 5. This number will depen<
of the differential equation. In sorne
can spring not from an ill-conditionec

L= With a matrix like K= {: ~). for e:

IN-1, N-2 be exchanged, or pivoted. If we exchan!
IN1 IN, N-2 IN, N-1 the pivot element Uú, then the Gauss
stable as the condition number allow
This means that LUQ = LF' is precisely the same as KQ = F; Gaussian example is, of course, not positive-defi
elimination is nothing but the factorization of K into a product which arise in the "mixed" method (Sf
The direct stiffness method-in whi1
K=LU u, as in the greater part of this book-:
 CHAP.
1ethods; this was a dominant therne
ods are increasingly squeezed between
)Urier transform; the latter is unques-
:try and the equation are appropriate,
.me linear system is to be sólved for
problems-elimination is conveni~nt
¡·-..
:he theory of Gausslan elimination.
. general coefficient matrix K, the first
;t N 1 equations, then Q2 is elimi-
and finally QN_ 1 is eliminated from
•transformed in this way toan equiva-
~IN)(~!)
2N • = F'.
UNN ºN
by back substitution, solving the Iast
-¡, and so on.
'ix terms what has happened. Suppose
.s by adding back to the last equation
~quation, which was subtracted off
we add back to the last two equations
tation N 2, which were subtracted
recover the original system KQ = F,
1e multiples lii of the previous equa-
vere subtracted in the elimination of
vstem KQ F has been recovered by
F' by the matrix
N-2
·-2 IN,N-1
!ly the same as K Q = F; Gaussian
1 of K into a product ·
,u
1 t SEC. 1.5. THE RITZ METHOD AND LINEAR ELEMENTS 35
of a lower triangular matrix times an upper triangular matrix. Thus K- 1F,
which is the solution Q we want to compute, is identical with u- 1L -• F, and
the two triangular matrices are easy to invert. In fact,L- 1Fis F', the inhomo-
geneous termas it stands after elimination, and then Q is u-•F', the result
after back substitution. (If there are many systems KQ Fn to be solved,
with different data bpt the same stiffness matrix K, then the factors L and U
should be stored.)
Now we consider what is special about the elimination process when K
is known, as in our example, to be symmetric, positive-definite, and tridi-
agonal. The first point is that the process succeeds-the eliminations can be
carried out, and the factorization K LU exists. lt would not ha ve succeeded
with K = (~ ~ ). since Q 1 could not be ellminated from the second equatioti
by subtracting a multiple 12 , 1 of the first. The condition for success is thar
each of the matrices in the upper left corner of K,
should have a nonzero determinant. For a positive-definite matrix, these
determinants are all positive, and therefore elimination can be carried out
with no exchanges of rows. In fact, the determinant of K<i) equals the product
U 11 U 22 • • • Un, so that the pivot elements Un, lying on the main diagonal
of U, are all positive.
Something more is required for the eljmination algorithm to be numer-
ically stable: The pivots Un must be not only nonzero but also sufficiently
large. Otherwise, the information in the original coefficients Kii will be de-
stroyed as the'algorithm progresses. We ha ve only partial control overthe size
ofthese pivots, and therefore sorne roundoff error is unavoidable. The intrin-
sic sensitivity of K to small perturbations is _measured by its condition number,
roughly the ratio of its largest eigenvalue to its smallest, which is the subject
of Chapter 5. This number will depend on the mesh size h and on the order
of the differential equation. In sorne cases, however, numerical difficulties
can spring not from an ill-conditioned K but from an ill-chosen algorithm.
With a matrix like K= (: ~), for example, the two equations should first
be exchanged, or pivoted. If we exchange rows at each stage so asto maximize
the pivot element UiJ; then the Gaussian elimination algorithm becomes as
stable as the condition number allows. The troublesome matrix K in this
example is, of course, not positive-definite; it is more typical ofthe matrices
which arise in the "mixed" method (Section 2.3).
The direct stiffness method-in:which the unknowns are the displacements
u, as in the greater part of this boók-automatically yields a positive definite
36 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.5. THEI

K. In this case, elimination without row exchanges is not only possible but Correspondingly, the vector F' = L-
also numerically stable. To understand this, we first give the factorization from back substitution are
K = LU a more symmetric form, by divfding out the diagonal part of U:

F'.
:._¿
d¡

The important point is that the num

is proportional only to N.
This leaves K = LD( D- 1 U), where all three factors are uniquely deterrni{\ed: In many-dimensional problems, th
Las a lower triangular matrix with unit diagonal, n- 1 U asan upper triangu- · and symmetric positive-definite. Howc
lar matrix with unit diagonal, and D as a positive diagonal matrix. By ordering of the unknowns. In fact, ti
symmetry, n- 1 Umust be the transpose of L. Thus K LDLT, the symmetric subroutines have been developed to yie
form of the factorization. It is even possible to go one step further, by intro- of Gaussian elimination.
ducing a new lower triangular matrix L = LD 112 ; this yields the Cholesky The simplest and most popular cri
factorization K= LLT. Now we can explain why positive-definite symmetric Suppose it is known that only the firs
matrices never require pivoting: The factor L in this decomposition behaves superdiagonals, are nonzero. (For tri
like a square root of K and cannot get out of scale. In fact, the condition this definition is w = l.) Then at eacl
number of L, appropriately defined, is exactly the square root of that of K. information can be used to advantag
This is to be compared with to be eliminated below each pivot, an
tracting a multiple of one row from

(f1 o1) = ( 1 o)1 (fo -e-•

o )(o1 have only 2w + 1 nonzero entries. Tht
1
t- )
t-• 1 .' out elimination and isinherited by t
symmetric matrix, the number of open
where the factors on the right are large even though the matrix on the left is N 3 /3 required for a full matrix.
no t. The simplest illustration of good
Finally, the fact that K is tridiagonal means an enormous reduction in view of the bandwidth, is furnished in
the number of computations. "Since the first unknown Q 1 has only to be of nodes. If there are fewer nodes in tt
eliminated from the second equation-it does not appear in the others-all tion, the unknowns should be numben
the multiples / 3 ,., ••• , lN- 1, 1 ofthe first equation, normally required in elimi- than each column. The bandwidth wil
nating Q1 , are zero. At the second stage, Q 2 has only to be eliminated from for a given node we must wait that lo1
the third equation, and so on. Thus in the lower triangular matrix L, the sumably linked to it by a nonzero er
only nonzero elements líe on the main diagonal and on the first subdiagonal; in the ordering. In general, finite ele
the factors of a tri diagonal matrix are bidiagonal. structure than those from conventiom:
Since the Cholesky factorization involves square roots of the pivots and ordering is far from self-evident.
presentf some extra problems in avoiding operations with zeros, the most There is a second criterion which
popular numerical algorithm is based on the decomposition K= LDLT. The matrix; it is a little more precise and
two factorizations are equivalent, both mathematically and from the point on the profile, or skyÍine, of the matr
of view of numerical stability. (We assume no overflow.) For a tridiagona~ in the ith row. If this occurs in colum
matrix, the en tries of L and D satisfy a simple recursion formula: · ~ puter, it will not be necessary to subt
Jrom row i. The multiplying factors 11, :
d¡ =K¡,¡- d¡-1lJ,J-H d0 =0, since QP ... , Q1_ 1 do not require eli
are absent in the first place. The prc
/.+1 ' = Kl+t.J.
J • d, non-zero entries in each row, and li
 CHAP.
row exchanges is _not only possible but
nd this, we first give the factorization
y dividing out the diagonal part of U:
1three factors are uniquely determined:
1it diagonal, n- 1 U as an upper triangu-
D as a positive diagonal matrix. By
ie of L. Thus K= LDLT, the symmetric
ossible to go one step further, by intro-
x L = LD 112 ; this yields the Cholesky
explain why positive-definite symmetric
factor L in this decomposition behaves
~et out of scale. In fact, the condition
is exactly the square root of that of K.
(oe O)(1o e-1),
1 .
-f"- 1
~e even though the matrix on the left is
onal means an enormous reduction in
the first unknown Q 1 has only to be
-it does not appear in the others-all
·st equation, normally required in elimi-
tge, ·Q2 has only to be eliminated from
in the lower triangular matrix L, the
t diagonal and on the first subdiagonal;
~ bidiagonal.
nvolves square roots of the pivots and
·iding operations with zeros, the most
on the decomposition K LDLT. The
th mathematically and from the point
tssume no overflow.) For a tridiagonal
·a simple recursion formula:·
IJ,j- H
1 SEC. 1.5. THE RITZ METHOD AND LINEAR ELEMENTS 37
Correspondingly, the vector F' L - 1F and the solution Q which results
from back substitution are
F0 O,
QN+1 = 0.
The important point is that the number of arithmetic operations involved
is proportional only to N.
In many-dimensional problems, the stiffness matrix K will still be sparse
and symmetric positive-definite. However, there will not be such an obvious
ordering of the unknowns. In fact, the nodal ordering becomes crucial, and
subroutines ha ve been developed to yield an ordering that reduces the expense
of Gaussian elimination.
The simplest and most popular criterion is the bandwidth of the matrix.
Suppose it is known that only the first w subdiagonáls of K, and the first w
[ superdiagonals, are nonzero. (For tridiagonal matrices, the bandwidth by
this definition is w 1.) Then at each stage in the elimination process this
information can be used to advantage. There are only w nonzero elements
to be eliminated below each pivot, and furthermore each elimination-sub-
tracting a multiple of one row from another-operates on rows known to
have only 2w + 1 nonzero entries. The band structure is preserved through..;
out elimination and is inherited by the triangular factors L and U. For a
symmetric matrix, the number of operations is roughly Nw 2 f2 rather than the
N 3 /3 required for a full matrix.
The simplest illustration of good and bad orderings, ·from the point of
view of the bandwidth, is furnished in two dimensions by a rectangular array
of nodes. If there are fewer nodes in the horizontal than in the vertical direc-
tion, the unlmowns should be numbered consecutively along each row rather
than each column. The bandwidth will be roughly the length of a row, since
for a given node we must wait that long for the node above it (which is pre-
sumably linked to it by a nonzero entry in the stiffness matrix) to appear
in the ordering. In gerleral, finite element matrices are less systematic in
structure than those from conventional difference equations, and an optimal
ordering is far from self-evident.
There is a second criterion which takes into account the sparseness of a
matrix; it is a little more precise and subtle than the bandwidth. lt is based
on the profi/e, or sk)line, of the matrix. Consider the first nonzero element
in the ith row. If this occurs in columnj and that fact is known to the com-
puter, it will·not be necessary to subtract multiples of rows 1, 2, ... , j - 1
from row i. The multiplying factors 11, P • . • , l1, j - t would all be zero anyway,
since QP . .. , Qi_ 1 do not require elimination from the ith equation; they
are absent in the first place. The profile. is formed by locating these first
non-zero entries in each row, and like the band structure it is preserved
38 AN INTRODUCTION TO THE THEORY
.. CHAP. 1 SEC. 1.6.

thro~ghout Gaussian elimination and inherited by the factor L. In case the band scheme using Nw N 312 locati
profiie goes significantly inside the band-so that a good many rows have dimensions.
length well below 2w + 1 non-zero entries-it may be worthwhile for the Whatever the algorithm-this wil1
computer to know the profile, and even to order the unknowns according to research-the eventual result of the e
a "profile-reducing" algorithm. KQ F. With this, the finite element
We fmphasize that the number of arithmetic operations is not the only
criteriorl\in choosing an algorithm; the storage requirements may be at least
as important. For a band matrix, a standard procedure is to store successive 1.6. THE ERROR WITH LINEAR ElEf
diagonals ofthe matrix; this is probably near to optimal, and uses about Nw
How close is the Ritz approximati~
locations. For a linear or bilinear element on a 50 x 50 square mesh, N =
as possible, according to the following
2500 and w ~50; at about this point, a typicallarge computer would have
in the error u - u" is a mínimum. TIJ
to store information outside of core, and the programming and data manage-
assuming that the energy · is measure
ment become much more complex. Therefore considerable attention is being
must be one which is associated with t
paid to algorithms which take account of, ap.d· maintain as far as possible,
fact it is specified by the functional /(v
the sparseness of a matrix even within its band or profile. At the extreme,
degree term in /(v). (This definition dH
we may identify the Iocation of every nonzero entry in A and order the un-
is physically correct; it is convenient jt
knowns by a "sparse matrix algorithm" so as to minimize the number of
functional is written in the form
nonzero entries in the lower triangular L. It seems to us thát for finite elements
this is too expensive; it takes no account of the systematic structure of the (19) I(v) = a(v, v
matrices.
We prefer, if the problem is so large that the standard band or profile the energy in the function v is given by ,
algorithm goes outside of core, to follow the analysis given in a series of The ertergy therefore coincides witt
papers by Alan George. His goal, for finite elements with',~ unknown para- as (Lv, v) and been integrated by parts.
meters in the plane, is to achieve O(N 312 ) arithmetic operations ~!1 elimination symmetric expression, and this symn
and O(N log N) non-zero en tries in L. This would be essentially optimal. a{v, v). In particular, if (Lv, w) is integr
(There are sorne special direct methods, ana_logous to the fast Fourier trans- ric form
form, which need only O(N log N) operations and O(N) total storage loca~
tions; but they are restricted to _simple problems on rectangles.) His goal is a(v, w) = J: (p(x)v'(x)w
achieved by an ordering [03] which is like the mínimum degree algorithm:
at each stage, the. unknown to be eliminated should be the one currently This is the energy inner product. It is de
connected to the fewest unknowns. This is very different from Irons' frontal space :re1 and represents the inner proc
method, and it appears to demand a large program overhead-probably lem.
too large. A more recent suggestion is given in George's paper "An efficient Our goal in this section is first to
band-oriented scheme forsolving n by n grid problems." He divides the domain that the energy in the error is minimi:
into thin strips, and applies a band algorithm to the corresponding submat- apply this theorem in establishing bou
rices (thé unknowns inside the strip having been numbered to reduce the band-
width). Between two strips will come a line of unknowns, and in George's THEOREM 1.1
ordering they should appear after the strips which théy separate. It is then Suppose. that u minimizes I(v) over
the comparatively small number of unknowns on these separating lines which is any closed subspace of :Je1. Then:
contribute.to a filling out ofthe band; the bigger ~ubmatrices corresponding · (a) The mínimum of I(v") and the
to the link between one strip and another are empty. With a number of strips ranges over the subspace S", are achie1
proportional to h- 112 , the storage requirement turns out to be O(Nsf 4 )-'-not
optimal, but better by the substantial factor N 114 than a straightforward (20) a( u u", u- u")= 1
~
y 1
¡¡,CHAP.
td inherited by the factor L. In case the
band-so that a good many rows have
entries-it may be worthwhile for the
ven to order the unknowns according to
::>f arithmetic operations is not the only
:he storage requirem~nts may be at least
:tandard procedure is''to store successive
tbly near to optimal, and uses about Nw
ement on a 50 X 50 square mesh, N =
1t, a typical Iarge computer would have
and the programming and data manage-
~herefore considerable attention is being
unt of, and· maintain as far as possible,
hin its band or profile. At the extreme,
y nonzero entry in A and order the un-
hm" so as to minimize the number of
r L. lt seems to us that for finite elements
~ount of the systematic structure of the
large that the standard band or profile
:Ollow the analysis given in a series of
r finite elements with N unknown para-
312
' ) arithmetic operations in elimination
L. This would be essentially optimal.
·ds, analogous to the fast Fourier trans-
)perations and O(N) total storage loca~
'le problems on rectangles.) His goal is
is like the mínimum degree algorithm:
1iminated should be the one currently
"'his 'is very different from Irons' frontal
a large program overhead-probably
is given in. George's paper "An efficient
n grid problems." He divides the domain
.lgorithm to the corresponding submat-
.ving been numbered to reduce the band-
;, a line of unknowns, and in George's
te strips which théy separate. lt is then
tknowns on these separating lines which
1; the bigger submatrices corresponding
ther are empty. With a number ofstrips
tuirement turns out to be O(N 514 )-not
ial factor N 114 than a straightforward
SEC. 1.6. THE ERROR WITH LINEAR ELEMENTs 39
band scheme using Nw = N 312 locations. The saving is the same in three
dimensions.
Whatever the algorithm-this will obviously remain an area of active
research-the event\lal result of the elimination process is the solution of
KQ F. With this, the finite element approximation is determined.
1.6. THE ERROR WITH LINEAR ELEMENTS
How close is the Ritz approximation u" to the true solution u? As close
as possible, according to the following theorem, in the sense that the energy
in the error u u" is a mínimum. The Ritz method is therefore optimal,
assuming that the energy · is measured in the natural way. This measure
must be one which is associated with the particular problem at hand, and in
fact it is specified by the functional1(v) itself: The energy in vis the second-
degree term in f(v). (This definition differs by a factor f from the one which .
is physically cot~ct; it is convenient just to ignore this factor.) Thus if the
functional is writtén in the form
(19) I(v) a(v, v) 2(f, v),
the energy in thefunction vis given by a(v, v).
The energy therefore coincides with the term which up to now has arisen
as (Lv, v) and been integrated by parts. This integration has produced a more
symmetric expression, and this symmetry is emphasized by the notation
a(v, v). In particular, if (Lv, w) is integrated by parts, the result is the symmet-
ric form
a(v, w) J: (p(x)v'(x)w'(x) + q(x)v(x)w(x)) dx.
This is the energy inner product. It is defined for all v and w in the admissible
space X 1 and represents the inner product that is intrinsic to the given prob-
lem.
Our goal in this section is first to prove the theorem described above-
that the energy in the error is minimized by the Ritz method-and then to
apply this theorem in establishing bounds for the error with linear elements.
THEOREM 1.1
Suppose that u minimizes l(v) over the full admissible space 3C};, and S"
is any closed subspace of X};. Then:
(a) The minimum of I(v") and the minimum of a(u v", u- v"), as v"
ranges o ver the subspace Sh, are achieved by the same function u". Therefore
(20)
40 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.6. T

(b) With respect to the energy inner product, u" is the projection of u onto The right side is
S". Equivalent(y, the error u- uh is orthogonal toS":
a(uh + ev", uh + ev")- 2(f, uh + f?J'
(21)
Therefore,
(e) 'f.he minimizing function satisfies
O 2e[a(uh, v") - (/
(22) for al/ v" in S".
Since this is true for small f of either si
In particular, if S" is the whole space JC1, then This equation expresses the vanishing ~
direction of v". In particular, a(u, v) =
(23) a(u, v) (f, v) foral/ v in JC1. vanishes in every direction V. This is th
(e) is established; it is the equation ofvi
COROLLARY
The choice v = u in this equation
Jtfollowsfrom (21) that a(u u", u")= O, or'a(u, u") a(u", uh), and the at the minimum, the strain energy is the
Pythagorean theorem holds: The en¡!rgy in the .error equals the error in the
energy, (25) I(u) = a(u, u) - 2C
Similarly, /(u") = -a(u", u"). In every e¡
o ver a larger class of functions, and tll
Furthermore, since the left side is necessarily positive, the strain energy in u"
the result stated in the corollary, that
always underestimates the strain energy in u:
estimated:
(24) a(u", u")<

This whole theorem is fundamental to the Ritz theory, and its three parts The proof of the theorem is now co
ate very much interdependent. We can deduce (b) immediately from (e): the existence nor the uniqueness of uh (
If (23) holds for all v, it holds for every v" in S\ and subtracting (22), the of u itself) has been proper1y est~blish·
result is (21). omission means that the proof has ha
Also (b) follows from (a): Ín an inner product space, the function in a him by poittting to the key word in the
subspace S" Closest to a given u is always the projection of u onto S". In the space S" is required to be closed. Tbis n
opposite direction, to show that (b) implies (a), we compute its own limiting functions. If there is a s

a( u u" v", u - u" v") = a(u u", u u") - 2a(u u", vh)
+ a(v", v").
then there must exist in S 11 a limit v, suc
If (21) holds, then

Equality occurs only if a(v", v") O, in other words, only if v" =O. Thus This will always be true if S" is finite-dir.
u" is the unique minimizing function in (20), and (a) is proved. plated in the Ritz method. In general, <
The problem is now to establish (e), since that will imply (b) and there- a function uh in S" which is absolutely
fore (a). If u" minimizes 1 over S", then for any f and v", is closed. We cite the example Sh = iJC
contains functions arbitrarily close to u
the "projection" breaks down.

y CHAP.
1er product, u" is the projection of u onto
Jrthogonal toS": ·
zes
) for all v" in S". ·
:JC1, then
for all v in X1.
, u") O, or a(u, u"r a(u", u"), and the
rgy in the error equals the error in the
cessarily positive, the strain energy in u"
~y in u:
a(u, u).
al to the Ritz theory, and its three parts
can deduce (b) immediately from (e):
very v" in S", and subtracting (22), the
inner product space, the function in a
vays the projection of u onto S". In the
mplies (a), we compute
'J(u - uli, u - u") - 2a(u - u", v")
+ a(v"' v").
, in other words, only if v" = O. Thus
in (20), and (a) is proved.
(e), since that will imply (b) and there-
:n for any E and v",
' ..¡ . ,
1 SEC. 1.6. THE ERROR WITR UNEAR ELEMENTS 41
The right side is
' '----./
a(u" + EV", u" + EV") - 2(f, u" + EV") = J(u") + 2E[a(uh, ?1')- (f, v")]
+ éa(v", v").
Therefore,
Since this is true for small E of either sign, it follows that a(u", v") (f, 11').
This equation éxpresses the vanishing of the first variation of 1 at u", in the
direction of v". In'particular, a(u, v) = (f, v), so that at u the first variation
vanishes in every direction v. This is the equation (11) derived earlier. Thus
(e) is established; it is the equation of virtual work.
The choice v = u in this equation leads to an interesting result, that
at the minimum, the strain energy is· the negative of the potential energy:
(25) /(u) a(u, u) - 2(f, u) = -a(u, u).
Similarly, /(u") -a(u", u"). In every case /(u) /(u"), since u is minimizing
over a larger class of functions, and therefore a change of sign reproduces
the result stated in the corollary, that the strain energy is always under-
estimated:
a(u", u")< a(u, u).
The proof of the theorem is now complete, except for one point: Neither
the existence nor the uniqueness of u" (or, in case S" is the whole space X1,
of u itself) has been properly established. To a functional analyst, such an
omission means that the proof has hardly begun. We shall try to mollify
him by pointing to the key word in the hypothesis of the theorem: The sub-
space S" is required to be closed. This means that the subspace must contain
its own limiting functions. If there is a sequence vN in S" such that
then there must exist in S" a limit v, such that
as N~ oo.
This will always be true if S" is finite-dimensional, which is the case contem-
plated in the Ritz method. In general, one cannot guarantee that there exists
a function u" in S" whieh is absolutely the closest to u, unless the subspace
is clósed. We cite the example S"= Xi, whieh is nota closed subspace. It
contains functions arbitrarily close. to u(x) x, but there is no closest one;
the "projection" breaks down.
42 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.6.

To ,prove the existence of u, defined to be the minimizing function over Now by borrowing part ofthe right si
the whole space X1, we need to see that X1 is itself closed. This is exactly
what was achieved in the completion process, enlarging X~ to the full admis-
sible space X1. In particular, it was in compl<?ting (or closing) the admissible f p(v')' + qv' > Pmi• f( v'
space that the natural boundary condition u'(n) O was dropped. There
was one ~echnical point in that process which we hurried by: the space was This is the required inequality a(v, v) ~
completed in the natural energy norm, a(v Vm v - vN) -O as in (10), lem is el/iptic. Together with (26a), it i
and yet w~·have described the completed space in terms ofthe 3C 1 norm. This in the standard norm 11 v 11 1 or in the e
step is justified by the equivalenceof the two norms: there exist constants space is the same, 3C1.
u and K such that . Note that with a natural boundary
is entirely different. The Poincaré im
(26a) a(v, v)< Kllvl!I, v(O) = O, and will be violated by every <
and with q O, the energy a(v, v) j
(26b) a(v,v) ullvllt· that v = O. Equivalently, the differenti:
The last inequality also yields the uniqueness of u and uh, sínce it means
-(p;
that the energy is positive definite: a(v, v) =O if and only if v =O. The sur-
face l(v) is strictly convex, and can have only one stationary point, at the u'(O) =d
minlmum. has no unique solution; u is determine(
The first inequality is easy, since apure Neumann problem-physically, ~
are possible-may lead to technical d·
J[p(v') 2
+ qv 2
] dx max(p(x), q(x)) J[(v') 2
+v 2
] dx. will be indefinite.
So much for the general theory. The
Therefore, K can be chosen as max(p, q). example, with Sh composed of broken
The inequality (26b) in the other direction starts in the same way, since estímate of the error eh = u - uh. The
p is bounded below by a positive constant Pmin: (20):
a(eh, eh)< a(u vA,, u
(27) Jp(v') 2
dx > Pmin J(v')2 dx. Of course, the function u is not know1
in 3C 0 , then 'u is in X~. Therefore, the qu
The difficulty comes with the undifferentiated, or zero-order, terms, since function u in JC~ be approximated by n
q need not be bounded away from zero; in fact, we may have q O. There- to work only with the Ritz approxima1
fore, we need an inequality of Poincaré type, bounding v in terms of v'. Wíth Sh a good approximation to u, since uh v
the boundary condition v(O) = O, the natural idea is to write tion of the error eh becomes a straigh
theory: How far away are functions i
~a(v, v)?
The most convenient choice of a fu
and apply the Schwarz inequality: interpolate u1 • The two functions agree
_in between. It can be expanded in termf
N
uJ(x) =E"'
l
Integration o ver O x0 n yields a Poincaré inequality:
At any node, only the one correspondí:
We. compare u with u1 first on the
ment, leading to a pointwise estímate <
 y
CHAP. 1 SEC. 1.6. THE ERROR WITH LINEAR ELEMENTS 43

.ned to be the minimizing furiction over Now by borrowing part of the right side óf (27),
· that 3C1 is itself dosed. This is exactly
' process, enlarging JC_i to the full admis-
ín compl~ting (or closing) the admissible I p(v') 2 + qV > Pmin
2
I( v')
2
> 2~zPrnin I(
v')
2
+V
2
•

mdition u'(n) O was dropped. There

:ess which we hurried by: the space was
)fm, a(v vN, v - ~}v) O as in (10),
eted space in terms ofthe 3C 1 norm. This
Jf the two norms: there exist constants
<KJivm,
ul!vm.
e uniqueness of u and uh, since it means
{v, v) = O if and only if v O. The sur-
have only one stationary point, at the
This is the required inequality a(v, v) ull v !Ir, which asserts that the prob-
lem is elliptic. Together with (26a), it implies that whether JC_i is completed
in the standard norm 11 v 11 1 or in the energy norm ~, the completed
space is the same, 3C1.
Note that with a natural boundary condition at both ends, the situation
is entirely different. The Poincaré inequality depended on pinning down
v(O) = O, and will be violated by every constant furiction. With v = constant,
and with q O, the energy a(v, v) J p(v')Z can be zero without implying
that v = O. Equivalently, the differential equation
-(pu')' = f,
u'(O) = u'(n) O

has no unique solution; u is determined only up toa constant function. Thus

apure Neumann problem-physically, a problem in which rigid body motions
are possible-may lead to technical difficulties: the quadratic form a(v, v)
{p(x), q(x)) J[(v') + v
2 2
] dx. will be indefinite.
So much for the general theory. The goal is to apply it to the finite element
v,q). example, with Sh composed of broken-line functions, and emerge with an
direction starts in the same way, since estímate of the error eh = u - uh. The key lies in the minimizing property
lStant Prnin: (20):

~ Prnin f (v') dx.

2

Of course, the fuº"ction u is not known; we can be certain only that if f is

in 3C 0 , then'u is in JC_i. Therefore, the question is: How closely can an arbitrary
ferentiated, or zero-order, terms, since function u in :JC_i be approximated by members of Sh? We need not attempt
~ro; in fact, we may ha ve q O. There- to work only with the Ritz approximation uh; it will be sufficient to find in
ré type, bounding v in terms of v'. With
Sh a good approximation to u, since uh will always be better. Thus the estima-
: natural idea is to write
tion of the error e" becomes a straightforward problem in approximation

r:· v'(x) dx,

theory: How far away are functions in JC_i from S\ in the natural norm
,.ja(v, v)?
The most convenient choice of a function in Sh which is close to u is its
interpolate u1 • The two functions agree at every node x = jh, and u1 is linear
_in between. 1t can be expanded in terms of the roof functions as
N
u¡{x) = I; u(jh)ffJ1(x).
1
a Poincaré inequality:
At any node, only the one corresponding basis function ffJ1 is nonzero.
We compare u with u1 first on the basis of a simple Taylor-series argu-
ment, leading to a pointwise estímate of their difference.
44 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.6. T

THEOREM 1.2 A more careful proof of (29) would i

Jf u" is continuous, then and the extreme function illustrated in
to be the best possib1e.
(28) maxl u(x)- ulx) 1 < -!h 2 maxl u"(x) 1 It follows from the theorem that if z.

and

(29) maxl u'(x)- u;(x) 1 < h maxl u"(x) 1- Since the Ritz approximation uh is at le
with linear e/ements satis.fies
Proof Consider the difference .d(x) = u(x) - ulx) over a typical interval
( j - I)h < x <jh. Since .d_vanishes at both ends ofthe interval, theremust
be at least one point z where .d'(z) = O. Then for any x,
This is almost the result we want. 1
reflects the observed rate of decrease of
.d'(x) = J~ .d "(y) dy. imperfection is in the other factor, ma)
assume that u" is continuous, or even th
But .d" = u", since u1 is linear, and (29) follows immediateiy: is an estímate of the error in the enerJ
on1y that u" has finite energy in the 3C 0
1.d'(x) 1 = 1 J~ u"(y) dy 1 < h maxl u" 1- < oo. The proof of this sharper resu1t,
polation, will be based on Fourier series
The maximum ofl .d(x) 1 will occur ata point where the derivative vanishes, error estímate is given in (34) below.
= O. We look to see in which half of the interval z lies; suppose, for
L\'(z)
example, that it is closer to the right-hand endpoint, so that jh - z < h/2. THEOREM 1.3
Then expanding in a Taylor series about z, Jf u" /ies in 3C 0 , then

.d(jh) = .d(z) + (jh- z).d'(z) + J:(jh- z) 2 .d"(w), (30)

where z < w < jh. Since .d vani~hes at the endpoint jh, and .d" = u", (31)

(32)
This constant-! is the best possible, not only for the error in linear interpo-
lation, but even if we a1low an arbitrary piecewise linear approximation. Proof Consider any subinterval oJ
The most difficult function to approximate has u" alternating between + 1 O < x < h. The difference .d(x) = u(x
and -1 on successive intervals (Fig. 1.5). The best piecewise linear approxi- and we represent it as a Fourier sine se
mation in this extreme case is identically zero, and the error is h 2 /8.
L\(x) = :El
u(x) = x(h -x)/2 1

By direct computation

J: (A')' dx =-

Fig. 1.5 The extreme case in piecewise linear approxirnation.

J: (.d ") 2 dx = -
 CHAP. 1 SEC. 1.6. THE ERROR WITH LINEAR ELEMENTS 45

A more careful proof of (29) wóuld improve it to maxl A' 1 < !h maxl u" l,
and the extreme function illustrated in Fig. 1.5 also shows this constant !
to be the best possible.
1 < -kh 2 maxl u "(x) 1 It follows from the theorem that if u" is continuous, then

1 < h maxl u"(x) 1.

Since the Ritz approximation uh is at least as close as u1, the error in energy
) = u(x) - u¡(x) over a typical interval with linear elements satisfies
1t both ends of the interval, there must
O. Then for any x,
This is almost the result we want. The factor h2 is perfectly correct, and
d"(y)dy. refiects the observed rate of decrease of the error as the mesh is refined. The
imperfection is in the other factor, maxl u" 12 • It is unsatisfactory to ha ve to
assume that u" is continuous, or even that it is bounded, when the conclusion
~) follows immediately:
is an estímate of the error in the energy. It should be sufficient to assume
only that u" has finite energy in the X 0 norm, in other words that (u") 2 dx f
) dy 1 < h maxl u"l. < oo. The proof of this sharper result, in the special instance of linear inter-
polation, will be based on Fourier series rather than Taylor series. The proper
Lt a point where the deriva ti ve vanishes, error estímate is given in (34) below.
alf of the interval z lies; suppose, for
·hand endpoint, so that jh - z < h/2. THEOREM 1.3
)Ut z, 0
Jf u" lies in X , then

.1'(z) + !(jh _:_ z) 2 d"(w), (30) 11 u- U¡ < ~'h 2 ll u" llo,

llo -n
.t the endpoint jh, and .1" = u",
(31) 11 u' - u~ llo < _!__hllu"llo,
n
(32)
not only for the error in linear interpo-
:rary piecewise linear approximation. Proof Consider any subinterval of length h, for example the first one,
jmate has u" alternating between + 1 O < x < h. The difference d(x) = u(x) - u¡(x) vanishes at both endpoints
.5). The best piecewise linear approxi- and we represent it as a Fourier sine series:
lly zero, and the error is h 2 /8.
= . nnx
d(x) = ~ansm¡:z·
By direct computation

J: 2
(.1') dx = ~~ (n:ra;,
>iecewise linear approximation. J: 2
(.1") dx = ~ ~ (n:ra; .
46 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.6.

Since n I, As an alternative to Fourier serie

been derived by solving a variational
to A(O) A(h) · O, J (A")2 l. The
and the extremum falls naturally af sir
Therefore, summing on n,
The interpolate u1 is not to be con_

(33) J: (A')'
Both are piecewise linear, but u" is
simply a convenient choice close to z,
is closer still, yields the first of the fol
Here A" = u" because u1 is linear. Equality holds if and only if every coefii-
cient an after the first ís zero; A must have the form sin nxfh. CoROLLARY
The conclusion (33) holds equally well over each subinterval, say from The error e" = u - u" in the finite ~
U 1)h to jh, and we can sum over all these subintervals:
(34)
N fjh h2 N fjh
L; (A')2 < 2 L; (u")Z. The last inequality comes from (3),
1 (}-l)h 1(, 1 (J-l)h
data. The leading terms in the constar
We want to simplify this to for c2.
The final result is thus a bound of
putation shows this bound to be compl
a(u - u", u - u") i& almost exactly p
very crude meshes (h ! or ! ?). Sucl
This step, which looks completely obvious, is justified only because there is lishing an asymptotic expansion for
no trouble at the points where the subintervals meet. Notice that if A" were difference equations in (16).
still on the right side, as in (33), the equality Our convergence proofs have so far
excluding the case in which f is a ¿
f~
L;
f }h

(j-l)h
(A ")2 =
o
(A ")2 "ramp." An easy computation show~
of order h, unless a node is placed ri!
would have been completely false, the right side being in reality infinite. O(h) for a line discontinuity in two di
(A 11 is a ó-function at the nodes.) This point will recur when we consider the u lies in JC1, it is impossible to say a11
difference between "conforming" and "nonconforming" elements; when the it may be arbitrarily slow as h ~ O.
t~ial functions are not smooth enough to lie in the admissible space, l(v) pro ve.
cannot be computed element by element.
By a similar argument, THEOREM 1.4

f no Az dx
ll
h ~ 2 < h4 ¡~
2 ~ an - 1(,4
(
o u
")2
.
For any solution u in JC};-in othe.
data space-the finite element method

Therefore, a(e", e") O

Proof. Since JC1 was constructed
there is a sequence vN in JC.i convergí
fixed N, the finite element approxima
Theorem 1.3. Therefore, choosing N 1
tion v~ in S" arbitrarily close.to u. Sint
This completes the proof. the sequence uh must converge to u.
 CHAP. 1 THE ERROR WITH LINEAR ELEMENTS 47

As an alternative to Fourier series, the key inequality (33) could have

been derived by solving a variational problern: Maximize f (11')2, subject
to !1(0) !l.(h) __:_ O, f (11'') 2 l. The stationary points are !l. = sin nnx/h,
and the extremum falls naturally at sin nx/h.
The interpolare u1 is not to be confused with the Ritz approximation zi'.
Both are piecewise linear, but uh is determined variationally while u1 is
(!l. ")2 h2
11:2
fh (u ")2 .,., . simply a c~mvenient choice el ose to u. Theorem 1.1, which asserts that uh
o .
is closer still, yields the first of the following inequalities.
quality holds if and only if every coeffi-
have the form sin nxfh. CoROLLARY

y well over each subinterval, say from The error eh = u - uh in the finite element method satisfies
di these subintervals:
(34)
h 2 Nfjh
: I; (u")z. The last im!quality comes from (3), bounding the solution in terms of the
1 (/-l)h
data. The leading terms in the constants are Pmax/n 2 for C1 , and Pmax/n 2p~ 1 n
for C 2 •
The final result is thus·a bound of order h 2 for the error in energy. Com-
putaÜon shows this bound to be completely realistic, and in practice the error
a(u - u\ u - uh) is almost exactly proportional to h 2 , beginning even at
very crude meshes (h = ! or i ?). Such regularity can be explained by estab-
>vious, is justified only because there is Iishing an asymptotic expansion for the error, of the kind introduced for
bintervals meet. Notice that if !l." were difference equations in (16).
~quality
Our convergence proofs ha ve so far assumed two deriva ti ves for u, thereby
excluding the case in which f is a €5-function and u is a piecewise-linear
"ramp." An easy cQmputation shows that the resulting error in energy is
of order h, unless a no de is placed right at the discontinuity in f. [It is also
.he right side being in reality infinite. O(h) for a line discontinuity in two dimensions.] In general, given only that
s point will recur when we consider the u lies in X 1, it is irripossible to say anything about the rate of convergence;
"nonconforming" elements; when the it may be arbitrarily slow as h ~O. However, convergence itself is easy to
gh to líe in the admissible space, /(v) pro ve.
en t.
THEOREM 1.4
For any solution u in X 1-in other words for any data fin the matching
'
, a, 2
< h4
n4
f" (u ) . n 2 data space-the finite element method converges in the energy norm:
0
as h ~O.

Proof. Since X 1 was constructed in the first place by completing x¡,

there is a sequence vN in x¡ converging in the energy norm to u. For each
fixed N, the finite element approximations vR, converge to vN as h ~ O, by
Theorem 1.3. Therefore, choosing N large and then h small, there is a func-
tion v'N in Sh arbitrarily close to u. Since the projection uh will be even closer,
the sequence uh must converge to u.
48 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.6.

TQ.is argument applies unchanged to all such minimization problems and On the other hand, Theorem 1.1 asse
need not be repeated in every case. The necessary and sufficient condition for Subtracting from (36), this gives
convergence in the Ritz method is obviously this, \hat for every admissible
u, the distance to the trialspacesSh (measured by the energyj should approach (37) a(z v\ eh
zero as h ---+ O. The point of the previous theorem is that once this conver-
To the left side we apply the Schwarz .
gence h~s been verified for a "dense" subspace-one whose completion in
the ener$y norm yields all admissible functions-then convergence is auto;. la(v,w)l (a(v,~
matic for··every u. Therefore, the interesting problem is to establish the rate
of convergence in energy when u is sufficiently smooth. with v = z vh, w eh. By the coroU:
lt is equally interesting, and somewhat harder, to find this rate of conver-
gence in a different norm. The result of the corollary is that the strains-the (a( eh, é)) 112 ·

first derivatives (uh)'-are in error by O(h). We inquire next about the error
in the displacement. How quickly does eh= u- uh decrease when measured Choosing vh as the Ritz approximation
by 11 ehlfo?
The crudest answer is to apply Poincaré's inequality 1eh(x 0 ) 1< ,./1tl! eh !1 1 ,
derived above. This bounds the error at every point x 0 uniformly by O(h).
Thus the Schwarz inequa]ity applied t,
One might expect, however, to improve this estímate to O(h 2 ). Such an im-
provement appears almost obvious from (30), where the error in the interpo-
lating element u1 is of second order. This at least proves that Sh contains a
function which is within O(h 2 ) of u in displacement. The difficulty is that Finally, we can bound the solution z t
in the ;reo norm the Ritz approximation uh is not minimÚ:ing. There is no
assurance that uh is as close to u as u1 is, and we shalllater find fourth-order !lz"llo 11 zl
problems in which the error in displacement is no better than the error in
slope. Here is the key point; that to estima
In the present example, however, the displacement error is indeed O(h 2). which is variationally unnatura1, one
One possible proof is simply to forget the variational derivation of the finite solutions in the norm of 3C 2 • The Iatt
element equations KQ = F, and to compute their truncation error as differ- from the viewpoint of differential' equ'
ence equations. (At the boundary x 11: this has aiready been done.) Applying theory was exactly this, to estímate tt
the maximum principie, the result is actually a pointwise bound 1eh(x) 1 = in 3C 0 • Substituting this bound into t
O(h 2 ), which is optimal. Nevertheless, this approach by way of finite differ- the common factor 11 eh 11 0 , this argum
ences is not entirely satisfactory, because its extension to irregular finite literature as Nitsche's trick) has estab
elements in two-dimensional problems .becomes extremely ·difficult. There- displacement:
fore, it is essential to look for an argument which will establish variationally
the rate of convergence of the displacement error 11 eh llo· THEOREM 1.5
The following trick is remarkably successful. Let z be the solution to The piecewise linear finite' element
the origínal variational problem over 3C1, when eh u- uh is chosen as the method, satisfies
data. Then by the vanishing of the first variation,
(38)
(35) a(z, v) =(eh, v) for all v in 3C1.
It is interesting that this bound ~
In particular, we may choose v = e": to the fact that approximation to ord1
fact is therefore another consequenct:
(36) approximation in 3C 1 , then it can achi
 CHAP.
to all such minimization problems and
1e necessary and sufficient condition for
·bviously this, that for every admissible
1easured by the energy) should approai:h
vious theorem is that once this conver~
:" subspace-one whose completion in
e functions-then cohvergence is auto-
:resting problem is to'' establish the rate
tfficiently smooth.
what harder, to find this rate of conver-
of the corollary is that the strains-the
O(h). We inquire next about the error
s eh = u - uh dacrease when measured
ncaré's inequality 1eh(x 0 ) 1 Jnll eh IIP
r at every point x 0 uniformly by O(h).
>ve this estímate to O(h 2). Such an im-
om (30), where the error in the interpo-
This at least proves that Sh contains a
in displacement. The difficulty is that
'ion uh is not minimizing. There is no
is, and we shalllater find fourth-order
acement is no better than the error in
the displacement error is indeed O(h ). 2
: the variational derivation of the finite
>mpute their truncation error as differ-
n this has already been done.) Applying
actually a pointwise bound 1eh(x) 1
, this approach by way of finite differ-
cause its extension to irregular finite
1s beco mes extreme! y· difficult. There-
tment which will establish variationally
;ement error 11 eh llo·
r successful. Let z be the solution to
3C 1, when eh = u uh is chosen as the
st variation,
for all v in X 1.
1 THE ERROR WITH LINEAR ELEMENTS 49
On the other hand, Theorem 1.1 asserts that a(vh, eh)= O for all vh in Sh.
Subtracting from (36), this gives
(37)
To the left side we apply the Schwarz inequality in the energy norm, that is,
1 a(v, w) 1 < (a(v, v)) 112 (a(w, w))t1 2
with v z- v\ w = eh. By the corollary to Theorem 1.2,
Choosing vh as the Ritz approximation to z, the same corollary gives
Thus the Schwarz inequality applied to (37) yields
Finally, we can bound the solution z by its data eh; according to (3),
Here is the key point; that to estímate the Ritz error eh in the norm 3C 0 ,
which is variationally unnatural, one nee<;is an equally unnaturat bound for
solutions in the norm of 3C 2 • The latter bound is entirely normal, however,
from the viewpoint of differential equations; in fact, the central result of the
theory was exactly this, to estímate the solution in JC 2 in terms of the data
in 3C 0 • Substituting this bound into the previous inequality and cancelling
the common factor 11 eh 11 0 , this argument (known in the numerical analysis
Jiterature as Nitsche's trick) has established an h. 2 estímate for the error in
displacement:
THEOREM 1.5
The piecewise linear finite element approximation u\ derh•ed by the Ritz
method, satisjies
(38)
It is interesting that this bound was derived without appealing directly
to the fact that approximation to order h 2 is possible in the JC 0 norm. This
fact is therefore another consequence of the theorem: If Sh achieves O(h)
approximation' in X', then it can ·achieve O(h 2) approximation in 3C 0 •
50. AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.7. THE FINITE ELEMI

We have remarked that the rates of decrease predicted for the error-h 2 [We ha ve not been able to use the stronges1
for di'splacement and h for strain-have been repeatedly confirmed by derivative of u - ií is O(h 2 ).]
numerical experiment. Sorne experimenters have calculated only the errors The proof is completed by applying th.:
at individual mesh points, instead ofthe mean-square errors over the interval, is the projection onto Sh of u - ií, an~
again with the same rates of convergence. [To predict these pointwise errors energy,
we must either return to the maximum principie, or assume more mean-
square cÓntinuity of th.e data and refine the variational estímate. In a few a(uh - fíh, uh - üh) a(u - ií, u- ií)
important problems the Ritz solution is actually more accurate atthe nodal a((u ií)-
points than elsewhere; for -u" u(O) = u(n) =O, uh agrees identically a( u - ií, u - üj
with u at the nodes, and the accuracy is infinite.] There have been a few cases,
however, in which the expected convergence has not been confirmed, because
·. We conclude that the error due to i11
of a simple fiaw i.n the experiment: The computer has worked with the
quantity in the sense of energy) than the h2 error i
mation on linear elements.
Eh max 1eh(jh) 1·
J

This expression introduces, with every decrease in h, a new and larger set of 1.7. THE FINITE ELEMENT METHOD U
mesh points. In particular, points nearer and nearer to the boundary, where DIMENSION
the error is often greatest, will determine the computed value Eh. It is unrea-
sonable to expect that this error, occurring at a point which varíes with h, This section extends the previous om
will still display the optimal h 2 rate of decrease. inhomogeneous boundary conditions, ele
Finally, there is the error which arises when the load f is replaced by its rather than linear, and differential equati4
linear interpolate f 1 • This change, which was introduced to simplify the inte- estimates for different finite elements wil
grations J frp1 dx, alters the load vector from Fto F. It leads to approximate they Iater fall within the theory which
finite element solutions Q = K-tp and fíh I; Q,rp;, which are the exact The steps in the method itself are exactl)
jinlte element approximations for the problem with data f 1 • Therefore, we statement of the problem, the construct
have only to consider the change in the Rit~ solution due to change in the space of the admissible space, and the :
data. equations K Q = F. The pattern will t1
in one dimension.
THEOREM 1.6 We begin by retaining the same diff
Jf f is. replaced by its interpolate j¡, then the induced error uh - uh in the but with the more general boundary co1
finite eleinent approximation satisfies
u(O) = g, u'(n:

The first of these conditions is agajn ess

function v in the admissible space 3C1. 1
Prooj. The exact solution u - ií, corresponding to data f- j¡, is bounded between any two admissible functions w,
by ~ (0) O. We denote by V0 the space of
slble space when the essential condition:
(39) The boundary condition at the othe1
u' and u, and therefore the functional '
The Iast inequality gives the error in linear interpolation; it is copied from the system represents a string which is
Theorem 1.3. Now it follows that · x = n; instead it is connected to a sprin

a(u ií,u-ií) Kllu ü!!i<KIIu-iíll~

JRY CHAP. 1 SEC. 1.7. THE FINITE BLEMBNT METHOD IN ONE DIMBNSION 51

~es of decrease predicted for the error-h 2 '

[We\have not been able to use the strongest part of (39)-that even the second
.in-have been repeatedly confirmed by derivauve of u ü is O(h 2).]
rimenters have calculated only the érrors The proof is completed by applying the corollary to Theorem l.l: u11 - i?
>fthe mean-square errors overtheinterval, is the projection onto s~a of u - u, and this projection cannot increase
:rgence. [To predict these pointwis~ errors energy,
-~:.imum principie, or assume more mean-
refine the variation~,l estímate. In a few a(uh - ü\ uh - uh) a(u - Ü, u - u)
ion is actually more a:ccurate at the nodal a((u - ü) - _(uh i?), (u - ü) (uh- uh))
=J, u(O) = u(n:) O, uh agrees identically
u, u- ü) < ~ h 4 ll/nll~·
2

;:,y is infinite.] There have been a few cases, < a(u

tvergence has not been confirmed, because
nt: The computer has worked with the We conclude that the error due to interpolating, the data is smaller (h4
in the sense of energy) than the h2 error inherent in basing the Ritz approxi-
max 1eh(jh) j. mation on linear elements.
1

'ery decrease in h, a new and larger set of
1earer and nearer to the boundary, where
rmine the computed value Eh. 1t is unrea-
'ccurring at a point which varies with h,
: of decrease.
1 arises when the load f is replaced by its
vhich was introduced to simplify the inte-
:ctor from Fto f. It leads to approximate
~ and üh = L; Q¡rp~, which are the exact
'he problem with data fi· Therefore, we
n the Ritz solution due to change in the
? JI, then the induced error uh uh in the
·s
1.7. THE FINITE ELEMENT METHOD IN ONE
DIMENSION
This section extends the previous ones in three directions-it introduces
inhomogeneous boundary conditions, elements which are quadratic or cubic
rather than linear, and differential equations of order 4 as well as 2. The error
estimates for different finite elements wi11 be stated rather than proved, since
they later fall within the theory which it is our main object to describe.
The steps in the method itself are exactly the same as before: the variational
statement of the problem, the construction of a piecewise polynomial sub-
space.. of the admissible space, and the assembly and solution of the linear
equations KQ =F. The pattern will therefore be more or less complete,
in one dimension.
We begin by retaining the same differential equation -(pu')' + qu
but with the more general boundary conditions
u(O) = g, u'(n:) + tXu(n:) = b.

, corresponding toda~af- JI, is bounded
in linear inter'polation; it is copied from
· x = x; instead it is connected to a spring. The new functional is
The first of these conditions is again essential, and must be satisfied by every
function v in the admissible space 3C1. Therefore, the difference V 0 = V1 V 2
between any two admissib/e functions will satis/y the homogeneous condition
v0 (0) =O. We denote by V 0 the space ofthese differences v0 ; it was the admis-
sible space when the essential conditions were homogeneous.
, The boundary condition at the other end is óf a new kind, involving both
u' and u, and therefore the functional /(v) must be recomputed. Physically,
the system represents a string which is neither fixed nor completely free at

J(v) = J: (p(v') 2 + qv 2
) dx +tXp(n:)v (x)- 2 J:
2
fv dx 2bp(n:)d.:n;).
 1
52 AN INTRODUCTION TO THE THEORY t_ CH;AP. 1 SEC. 1.7~ THE FINITE E

Thus the new condition has introduced a boundary term both in the linear system will be identical to the one in 1
part óf the functional and in the energy the matrix, corresponding to the left ei
other row except that q 0 g. t Therefo
a(v, v) J (p(v') 2 + qv 2 ) dx + ap('n:)v 2 (n). coefficients taken to be p = q = 1) is

The last,,term represents the energy in the spring. -g + 2Q 1 - Q2 + !!_(g -t

1
We rt 9w verify that the vanishing of the. first variation in every direction
h 6
v 0 leads tó the same conditions on the minimizing u as the differential equa-
Shifting the terms involving g to the <
tion and boundary conditions. [Note that u is perturbed by functions v 0
as before. The inhomogeneous conditi
in V 0 , which assures that the essential condition (u + Ev 0 )(0) gis satisfied.
F 1 of the load vector, by the amount
The perturbations are not the functions v in 3C1.] The coefficient of 2E in
At the other end the computation
I(u + EV 0 ) is
I(v") are
J [pu'v~ + quv 0 ] + ap(n)u(n)vo(n) - J fv 0 - bp(n)v 0 (n)

J[-(pu')' + qu f]v 0 + p(n)[u~(n) + au(n)- b]v0 (n). Therefore, in the last equation alfaqN
is an extra bp(n) in the load componet
entry KNN· Again the local nature oft:
This expression vanishes for all v 0 if and only if u satisfies the differential
city of this change; the only basis fum
equation and the new boundary condition. Therefore, 'this boimdary condi-
therefore coupled to the boundary coi
tion is naturalfor the modifiedfunctional l(v).
The error estimates for this probl1
In the general Ritz method, it no longer makes sense to ask that Sh be a
sectíon, namely
subspace of JC1. X1 itselfis nota vector space; ithas been shifted away from
the origin. Therefore, we ask S" to ha ve the same form. The tria! functions a(u u\ u
1, but the difference of any tria! func-
vh need not lie in the admissible space X and
tiáns must be in the homogeneous space V 0 • These differences v~ = v~ v~
form a finite-dimensional space sg, whicli is required to be a subspace of V 0 •
For linear finite elements, the procedure is clear. There ís no constraínt The first estímate again depends on ·
at x n, where the boundary condition is natural. The trial space S" will u" than it is to u1 • In turn, this depends
therefore consist of al! piecewise linear functions which satisfy v"(O) = g. u1 in the trial space Sh. Therefore, we:
(This essentia1 condition can be imposed exactly in one dimension, since it Theorem 1.3 for the distance between
constrains vh only at a point. In two or more dimensions, a boundary condi- The second extension of the methc
tion vh(x, y) g(x, y) cannot be satisfied by a polynomial, and Sh is not are more "refined'' than the piecewise
contained in X1.) S~ is the same piecewise linear trial space introduced in approximated by quadratic or cubic
previous sections, vanishing at x = O, and every vh can be expanded as will be a corresponding improvemem
natural to construct trial spaces S" v
higher degree.
To start, let S" consist of all pi<
Thus the coefficient of qJ~ is fixed: q 0 g. continuous at the nodes x = jh and
The ~ potential energy I(v") is a quadratic in the unknowns q~' . .. , qN,
tThis corresponds to the way in which
and its minimization leads again toa linear system KQ =F. In the interior in practice. They are ignored until thematrio
of the interval-that is, for all but the first and last rows of the matrix-this (in this case q 0 ) given the value which is p1
 CHAP. 1
ced a boundary term both in the linear
rgy
n the spring.
· of the first variatioJ\in every direction
te minimizing u as the ·differential equa-
te that u is perturbed by functions v 0
1condition (u+ fv 0 )(0) =gis satisfied.
;ons v in X 1.] The coefficient of 2f in
J fvo - bp(rc)vo(rc) ..
- f]v 0 + p(rc)[u'(rc) + rtu(rc) - b]v 0 (rc).
~ and only if u satisfies the differential
.dition. Therefore, 'fhis boimdary condi-
'1al I(v).
Ionger makes sense to ask that Sh be a
:or space; it has been shifted away from
tave the same form. The tria! functions
JC 1, but the difference of any tria! func-
ce V 0 • These differences vi = v~ - v1
hich is required to be a subspace of V 0 •
'cedure is clear. There is no constraint
tion is natural. The trial space Sh will
·ear functions which satisfy vh(O) = g.
,sed exactly in one dimension, since it
r more dimensions, a boundary condi-
sfied by a polynomial, and Sh is not
cewise linear trial space introduced in
and every v" can be expanded as
N natural to construct trial spaces Sh which are composed of polynomials of
1 + I; qlp~(x).
1
= g.
Iadratic in the unknowns q 1' • • · ' qN•
linear system KQ =F. In the interior
first and last rows of the matrix-this
SEC. l. 7. THE FINITE ELEMENT METHOD IN ONE. DIMENSION 53
(
syst~m will be identical to the one in the previous sections. The first row of
the matrix, corresponding to the left end of the interval, looks just like every
other row except that q 0 = g. t Therefore the first equation in the system (with
coefficients taken to be p = q = 1) is
-g + ~·- Q, + ~ (g + 4Q, + Q,) =
2
r Jrpl dx.
Shifting the terms involving g to the other side, the first row of K is exactly
as before. The inhomogeneous condition has altered only the first component
F 1 of the load vector, by the amount gfh- gh/6.
At the other end the computation is almost as simple. The new terms in
J(vh) are
Therefore, in the last equation ai¡aqN = o, after deleting the factor 2, there
is an extra bp(rc) in the load component FN andan extra rtp(n) in the stiffness
entry KNN· Again the local nature of the basis was responsible for the simpli-
city of this change; the only basis function which is nonzero at x = n, and is
therefore coupled to the boundary condition, is the last one.
The error estimates for. this problem will be the same as in the previous
section, namely
and
The first estimate again depends on the variational theorem·: u is closer to
tih than it is to u1 • In turn, this depends on the presence ofthe linear interpolate
u1 in the trial space Sh. Therefore, we may again appeal to the approximation
Theorem 1.3 for the distance between a function and its interpolate.
The second extension of the method is the introduction of elements which
are more "refined" than the piecewise linear functions. A given u(x) is better
approximated by quadratic or cubic than by linear interpolation, and there
will be a corresponding improvement in the accuracy of uh. Thereforé, it is
higher degree.
To start, let Sh consist of all piecewise quadratic functions which are
continuous at the nodes x = jh and satisfy vh(O) = g. Our first object is to
tThis corresponds to the way in which essential boundary conditions are introduced
in practice. They are ignored until the matrices are assembled, and only then is the unknown
(in this case q 0 ) given the value which is prescribed by the boundary condition.
54 AN INTRODUCTION TO TiiE THEORY CHAP. 1 SEC. 1.7. THE FINITE EL

compu~e the dimension of Sh (the pumber of free parameters q1) and to cient of q0 is the right half of rp 1 in Fig.
determine a basis. Note that as x passes a node, continuity imposes only and to zero at x = h/2, x = h. The co,
one constraint on the parabola which begins at that node; two parameters Fig. 1.6b, and the coefficient of q 1 is th(
of the parabola remain free. Therefore, the dimension must be twice the The element matrix k 1 is computed
number of parabolas, or 2N. the result as (qoqt¡zq.)Tk.(qoqt¡zq¡)· Nc
A ba~is can be constructed by introducing the midpoints x = (j - !)h three of the parameters q appear in an
as nodesf\in addition to the endpoints x = jh. There are then 2N nodes, mined by three conditions. In the cal(
since x ~-O is excluded and Nh n; we shall denote them by zi' j = and we give only the final result:
1, ... , 2N._ To each node there corresponds a continuous piecewise quadratic
which equals one at z1 and zero at Z¡, i

(40)

These functions are of two kinds, depending on whether z1 is an endpoint Notice that k 1 is singular; applied to t
(Fig. l.6a) ot a midpoint (Fig. 1.6b). Notice that both are continuous and vector q0 1, qt 1z = 1, q 1 = 1 corresp<
4[(x/h)- (xlh) 2] a horizontal line vh = 1, so its deriva1
a useful check; k 0 would not be singuh
1- 3(x/h) + 2 (x/h) 2 These ideas extend directly to cut
functions imposes one constraint at ea
free parameters in the cubic, and the
a basis we place two nodes within eacl
the endpoints. Together with the end]
o h
The functions ·which satisfy rp1(z1) b
Fig. 1.6 The basis functions for piecewise quadratic elements.
kinds ·(Fig. l. 7). The element matrices

therefore in X1. The function rp 2 , which is n01i'zero only over one subinterval,

~
is not intrinsically determined by _the subspace; the intemal node could have
been chosen anywhere in the interval. The choice of the midpoint affected
the basis but not the space itself. -h o h
Je
The element stiffness matrix k 1 , corresponding to the integral of (v') 2
Fig. 1.7 Cubics which are
overO x < h, is computed from the quadratic which equals q0 at x O, 01

q 112 at the midpoint x = h/2, and q 1 at x =h. This quadratic is

There is anoth~r cubic element wh
1t is constructed by imposing continuit
on its first derivative. This means that
of the preceding one, with one new cor
To relate this to the basis functions, collect the coefficient of each q: nodes x = h, 2h, ... , n h. Therefo
will be 3N (N- 1) 2N + 1, a re~
' parameters to be calculated. ·The onl~
vh(x) q 0 (1-3~ +2~:)+q 112 (4~ -4~:)+q 1 (-~ +2~:)· fails to improve on the old one is in th
do not have a continuous derivative. 1
These three coefficients are exactly the three parabolas in Fig. 1.6. The coeffi- case-u is in X1 but not in Xj-for a p
 CHAP.
number of free p~rameters q1) and to
•asses a node, coritinuity imposes only
h begins at that node; two parameters
ore, the dimension must be twice the
1troducing the midp?ints x = (} - !)h
nts x = jh. There ate then 2N nodes,
n; we shall denote ·them by z j' j =
ponds a contin.uous piecewise quadratic
, i =1= j;
~pending on whether z 1 is an endpojnt
. Notice that both are c~ntinuous and
4[(x/h)- (x/h) 2]
o h
or piecewise quadratic elements.
;h is nonzero only over one subinterval,
subspace; the internal node could have
l. The choice of the midpoint affected
corresponding to the integral of (v')2
te quadratic which equals q 0 at x = O,
t x = h. This quadratic is
ollect the coefficient of each q:
: three parabolas in Fig. 1.6. The coeffi-
1 THE FINITE ELEMENT METHOD IN ONE DIMENSION 55
cient of q 0 is the right half of rp 1 in Fig. 1.6a-a parabola equal to 1 at x = O
and to zero at x = h/2, x = h. The coefficient of q 112 is the parabola qP in
I.6b, and the coefficient of q 1 is the first half of rp 1 •
The element matrix k 1 is computed by integrating (dvhjdx)2 and writing
the result as (q 0 q 112 q 1 )Tk 1 (q 0 q 112 q 1). Notice that k 1 is a 3 X 3 matrix, sincc
three of the parameters q appear in any given interval; a parabola is deter-
mined by three conditiop.s. In the calculation it is every man for himself,
and we give only the final result:
-8
16
-8
Notice that k 1 .is singular; applied to the vector (1, 1, 1) it yields zero. This
vector q 0 = 1, q 112 = 1, q 1 = 1 corresponds toa parabola vh which is actually
a horizontal line vh = l, so its derivative is zero. This singularity of k 1 is
a useful check; k 0 would not be singular.
These ideas extend directly to cubic elements. Continuity of the trial
functions imposes one constraint at each of the nodes x = jh, leaving three
free parameters in the cubic, and the dimension of Sh is 3N. To construct
a basis we place two nodes within each interval, say at a· distance h/3 from
the endpoints. Together with the endpoints this gives 3N nodes zj = jh/3.
The functions which satisfy rp /z;) = J;j constitute a basis and are of three
kinds (Fig. 1.7). The element matrices will be of order 4.
~
-h o h
Fig. l. 7 Cubics which are only continuous at the nodes.
There is another cubic element which is better in almost every respect.
It is constructed by imposing continuity not only on the function vh but also
on its first derivative. This means that the trial space is actually a subspace
of the preceding one, with one new constraint at each ofthe N- 1 internal
nodes x = h, 2h, ... , _n - h. Therefore, the dimension of this new space
will be 3N- (N- 1) = 2N + 1, a reduction of o.ne-third in the number of
parameters to be calculated. The only way in which this new cubic space
fails to improve on the old Óne is in the approximation of solutions u which
do not ha ve· a continuous derivative. We saw in Section 1.3 that'this is the
case-u is in JC1 but not in JCi-for a point loador for a discontinuous coeffi-
56 AN INTRODUCTION TO THE THEÓRY CHAP. 1 SEC. 1.7. THE FINITE ELI

cient p(x) in the differential equation. In these cases of interna! singularities

it is essential not to require excess smoothness of the trial functions; the
Mass matrix k 0 : s: (
v") 2 = qTk 1
singularity X 0 shotild be chosen as a node, and across that node the cubic Stiffness matrix k 1 : s: ((Y)
v11 2
q
should only be continuous. This will preserve "the order of convergence.
The placement of nodes for the smoother cubics beco mes more interesting:
There is a double node at each point x jh. Instead of being determined by
Bending matrix k 2 : J:((if)") 2 = ~
at
its values the four distinct points O, h/3, 2h/~, and h, the cubic is now deter- Notice that because v 11 is in 3C 2 , it catJ
mined by its values and its first derivatives at the two endpoints, that is, by at that point the bending matrix will be
V 0 , v~, v~' and v'1 • Both v 1 and v'1 will be shared by the cubic in the next There are several ways in which too
subinterval, assuring continuity of t' and v'. The basis functions are, therefore, matrix k 0 • One of the best is to form tl
of two kinds· (Fig. 1.8). These basis functions ha ve a double zero at the ends nodal parameters in the vector q to the
of the cubic polynomial v11 : A = Hq, e
l/f(x/h) w(x/h)

o/(0) =1 ao o
a. o 1
3 2
-h2 -¡¡
2 1
h3 h2
'i/=0
~~----4-----~~
-h o h
The integration of (v 11) 2 (a 0 + a x -1
1

w(x) = x(lxl-1) 2

Fig. 1.8 Hermite cubics: v and v' c9ntinuous.

(j 1)h. They are called Hermite-cubics, since they interpolate both function
value and derivative; the earlier examples might ha ve been associated with the
name of Lagrange.
The cubic polynomial on O x < h which takes on the four prescribed
values v 0 , v~, v 1 , and v'1 is
Denoting this matrix by N 0 , the result
(41)

Therefore, the ele~ent mass matrix m1

The element matrices, which are of order 4, are computed by the following All this is easily programmed.
integrations, where q is the column vector (v 0 , v~, v~' v'1Y: For the stiffness matrix, the link Ji
 - CHAP. 1 THE FINITE ELEMENT METHOD IN ONE DIMENSION 57

1. In these cases of interna! singularities

smoothness of tHe trial functions; the Mass matrix k 0 : J: (v") 2 qTk 0 q,
1 node, and across that node the cubic
'iii preserve the order of convergenee.
Stiffness matrix k 1 : J: ((v")') 2 qrk 1 q,
10other cubics becomes more interesting:
x = jh. Instead of being determined by
Bending matrix 1<;2 : J: ((v")") 2
=· qTk2q.

, h/3, 2hf3, and h, the;:~ubic is now deter-

'atives at the two endpoints, that is, by
ill be shared by the cubic in the next
nd v'. The basis functions are, therefore,
Jnctions have a double zero at the ends
w(x/h)
Notice that because v" is in 3e2, it can be us~d in fourth-order problems;
at that point the bending matrix will be needed. .
There are several ways in which to organize the computat10n of the mass
matrix k 0 • One of the best is to form the matrix H ~hich cortnects the four)
nodal parameters in the vector q to the four coeffictents A = (ao, al, a2, a3
of the cubic polynomial v": A = Hq, or from (41 ),

ao o o o Vo
al o 1 o o V~
J 2 3 1
-h2 h2 --¡¡ v.
2 2 1 v~.
h3 h2 -h3

The integration of ( v") 2 (a 0 + a x + a 2x 2 + a

1 3 x 3 ) 2 is completely trivial,
h2 h3
w(x) = x(!xl-1) 2 h
2
es: v and v' ct;mtinuous. h2 h3 h4

~es, since they interpolate both function J: (11')' dx (aoata2a3) h3

2
h4
4
hs
at

a2
•les might háve been associated with the 3 4
: h which takes on the four prescribed

Denoting this matrix by N 0 , the result beco mes

X _:_ h) + v 1cv (X-h-h)
+- V 1 rp ( -h- 1

2
'o - v'1h - 2v~h)~ 2
Therefore, the element mass matrix must be
1 x3
+- hvo'Jj¡J ·

·der 4, are computed by the following All this is easily programmed.

~ctor (v0 , v~, v~' v'1)T: For the stiffness matrix, the l:ink H between the nodal parameters q and
 1.7. THE FINITE ELE!\
58 AN INTRODUCTJON TO THE THEORY CHAP. 1 SEC.

the coe~cient vector a is unchanged. The only difference now is that - tion of new elements, and of combination
method and force methods, has been ovet

J: ((vh)') 2 dx = J(a 1 2a 2x + 3a 3 x 2)2 ATN¡A

stiffness matrices in their singular form,
role of all four nodal parameters v0 , v~, v
o o- o o For application to the problem -(pu
o h h2 h3 must be assembled into the global stiffi
coefficients, a typical row (or rather pairo
AT o h2 4h 3 3h 4 A. u. and u~ at the mesh point x 1 = jh) of tl
3 T }

3h 4 9h 5
o h3
·T 5 {oh( -36u1 _ 1 - 3hu~_ 1 + 12u,
The element stiffness matrix is k 1 HTN 1H. The results ofthese computa- (42a) + :;0(54u1- + I3hu~-• + 1

tions, and of a similar one for kz, are the following matrices (which are to be
completed by symmetry): F1 =· I f(x)"( ~ - j);
~[156
22h 54
4kl. 13h -!3h)
-3h 2 j0 (3u 1_ 1 hu~-• + Sh~ 3¡
ko 420 156 -22h ' qh2 .
4h 2
(42b) + 420(-13u1_ 1 - 3h~-t

F~ = Jf(x)a>( ~
3h ~36
3h) ' - i)·
e
1· 4h 2 -3h -h2
kl 30h 36 -3h
It is interesting to ·make sense of th.
4h 2 p 1, q O, and 1, so that the di:

~e
6h 12 finite element equations above become
4h 2 -6h 6h )
' 2h 2 .
1

k2 h3 6 u 1,.2;+_Lt_ ___,2u-='1 +-=--u..!..

1---!c1
12 -6h h2
i-·

-
4h 2
1
-S
Ui+t
2h
U·
J-t + S1 U, i - 3]
The matrix k 0 is positive definite, whereas k 1 has a zero eigenvalue corre-
sponding to the constant function vh 1, that is, to q = (1, O, 1, 0). The matrix
k 2 should be doubly singular, since every linear vh will have (vh)" O. Th~ After Taylor expansions, the first eqt
new null vector, corresponding to vh(x) x, is q = (0, 1, h, 1). and the second with -h 2 u"'f15 = O, wl
It is sometimes valuable to change k 1 and k 2 from the above singular the first. This is exactly the point at whi
element matrices to natural element matrices. This means that the matrices tributed a new and valuable idea to the
are made nonsingular by removing the rigid hody motions, which are the cases Instead of operating only with the unk
(vh)' O and (vh)" O. The orders of the matrices are reduced to 3 and 2, equation per mesh point, the finite elen
respective! y, and they now carry no redundant information · the fact that both displacements and slopes as unknc
' ' '
formally consistent with the original eq
(1, O, 1, O) is a null vector of k~' which is automatic because it corresponds
toa state with no strain, is taken for granted and not displayed. It appears is that high accuracy can be achieved,
that these "natural" matrices may make it simpler for a single program to approximated, without abandoning. the
accept a wide variety of elements, which is a tremendous asset; the prolifera:- equation. The more irregular the mesh
ORY
CHAP. 1 SEC. 1.7. THE FINITE ELEMENT METHOD IN ONE DIMENSION 59

!d. The on1y difference now is t~t tion of new elements, and of combinations between our present displacement
method and force methods, has been overwhelming. In the book we keep the
stiffness matrices in their singular form, because they display so clearly the
o role of all four nodal parameters v 0 , v~, Vp and v'1 •
o
For application to the problem -(pu')' + qu _.:_ J, these elementmatrices
h2 h3
must be assembled into the global stiffness matrix K. Assuming constant
4h 3 3h 4 coefficients, a typical row (or rather pair of rows, since there are two unknowns
T A.
2 u1 and u~ at the mesh point x1 jh) ofthe assembled K is
3h 4 9h 5
2 T
3~h( -36u1_ 3hu~_ 1 + 12u1 36u1+ + 3hu~+ 1 )
1 1

= nrNt H. The resu1ts of these computa-

re the fo1Jowing matrices (which are to be
(42a) + h_~(54u1 _ 1 + 13hu~_ 1 + 312u1 + 54u1 _ 13hú1-t) 1

22h 54 = F1 · J f(x)w( ~ - i)'

4h 2 13h -l3h)
-3h 2
156 -22h '
j 0 (3u 1_ 1 - hu~-t + 8hu~ 3u1+1 - hu~+t)

4h 2 (42b)
qh2 -
+ 420 ( 13u1 .,. 1 3hu~-t + 8hu~ + 13u1+ 1 -
, )
3hu1+t
3h -:-36
lh2 -3h 3h) '
-h2
36 -3h
4h 2 1t is interesting to make sense of this as a ditference equation. Suppose

6h j
12 p = 1, q =O, and f = 1, so that the d~fferential equation is -u" = 1. The
2 finit~ element equations above become
-6h 2h 2 .
12 -6h
4h 2

hereas k 1 has a zero eigenvalue corre- 2uj + u~-t) =O.

:= 1, that is, to q (l, O, 1, 0). The matrix
~very linear vh will have {vh)" o. The
•(x) x, is q = (0, 1, h, 1).
ge kt and k 2 from the above singular
natrices. This means -that the matrices
~ rigid body motions, which are the cases
tf the matrices are redu~ed to 3 and 2,
redundant information; the fact that
:h is automatic because it correspo~ds
gra?te~ and not displayed. It appears
tke. tt SJmpler for a single program to
;h ts a tremendous asset; the prolifera-
After Taylor expansions, the first equation is consistent with -u" = 1,
and the second with -h 2 u'"fl5 ~O, which can be derived by differentiating
the first. This is exactly the point at which finite element equations have con-
tributed a new and valuable idea to the established finite difference technique.
Instead of operating only with the unknown u1 , so that there is always one
equation per mesh point, the finite element difference equations may eouple
both displacements and slopes as unknowns, the equation for the slope being
formally consistent with the original equation differentiated once. The effect
is that high accuracy can be achieved, and derivatives of high arder can be
approximated, without abandoning. the strictly local nature of the difference
equation. The more irregular the mesh and the more curved the boundaries,
60 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.7. THE FINI':

the more im.portant thi§ innovation becomes. It should be taken seriously by

the developers of difference schemes, since there-without being limited by
having to derive discrete analogues from polynomial displacements and the
Ritz method-even greater efficiency might be possible.
Suppose we apply Taylor expansions to test the order of accuracy of the
Hermite ,difference equations (42). Starting with the assumptions
:•.

(43)
-2 -1
and expanding v 1± 1 and v~± 1 about the central point jh, it is possible tó use
the original differential equation and its differentiated forms to check that Fig. 1.9 The cubic E-spline:
nodes.
En= e~. and that these terms disappear for n = 1, 2, 3. In other words, the
coupled difference equation is fourth-order accurate. This matches exactly Schoenberg. In p~rticular, Schoen
the error estimatesto be derive<:f variationally, justas both the Taylor series on [0, n] can be written as a linear e'
and the variational bounds on u- u1 were O(h 2 ) in the linear case. However,
an unanticipated and slightly distressing point emerged in [M8]: The asymp-
totic expansions (43) are spoiled by a mismatch at the boundaries, and the
finite eh:~ment errors seem to be less systemátic as h ____,.O than a simple power
series in h. The errors are, however, definitely of order h 4 • These basis functions tp'
are formt
There is stÜl another important cubic space, formed from those functions scaling the independent variable fn
for which even the second derivative is continuous at the nodes. A piecewise lie at the node jh:
cubic which has continuous second derivatives is called a cubic spline. This
tJJXx) =
space is again a subspace of the preceding one, the Hermite cubics, with
one new constraint at each of the N - 1internal nodes. Therefore, the dimen-
sion of the spline subspace is 3N 2(N +
1) = N 2. This means that In the expression for vh, the cons~
there is one unknown per mesh point, including the extra points x 0 = O been included. If lf'" and lf'~* are co1
(where v0 O but the slope v~ may be regarded as the free parameter) and essential condition, the admissible
xN+t n + h (or we may prefer that the last patameter be vjy). At interna! as
mesh points the displacement vi is the unknown, and the finite element
equations will once more look exactly like conventional finite difference
equations.
1t is no longer obvious which four nodal parameters determine the shape · In this form the N+ 2 unknowr
of the cubic over a given subinterval, say U I)h x <jh. The nodes would be q*, ... , qN+t· The unk
x 1 _ 1 and x 1, which belong to the interval, account for only two conditions, boundary condition imposes no e
and the other two must come from outside. Therefore, cubic splines cannot it would be interesting to see the e:
ha ve a strtctly local basis, and the shape of vh within an element must be affect- thereby removing the last unknow
ed by displacements outside the element. In fact, the spline which vanishés small.
at all nodes except one (a cardinal spline) will be nonzero over every subinter- Two final remarks on splines:
val between nodes.
To compute with these splines, we need to construct the one which equals l. It may be that their greatest
1 at the origin and which drops to zero as rapidly as possible (Fig. 1.9). This than minimization. A given relation
function is known as a basic spline, or B-spline. lt is fundamental to the whole approximated by a spline, while the
theory of splines, and was among the many remarkable discoveries of 1. J. tion is less convenient.
 1
/

CHAP. 1 SEC. 1.7. THE FINITE ELEMENT METHOD IN ONE DIMENSIÓN 61

Jmes. lt should be taken seriously by cp(x)

.nce there-withmit being limited by
n polynomial displacements and the
ight be possible.
s to test the order of accuracy of the
:ing with the assumptions

-2 -1 o 2
central point jh, it is possible to use
:s differentiated forms to check that Fig. 1.9 The cubic B~spline: continuous second derivatives at the
for n 1, 2, 3. In other words, the nodes.
rder accurate. This matches exactly
Schoenberg. In particular, Schoenberg has proved that every cubic spline
onally, just as both the Taylor series
on [0, n] can be written as a linear combination ofB-splines:
~re O(h 2 ) in the linear case. However,
; point emerged in [M8]: The asymp-
nismatch at the boundaries, and the
!matic as h ~ O than a simple power
initely of order h 4 • These basis functions rp1 are formed from the B-spline in the figure by re-
; space, formed from those functions scaling the independent variable from x to xfh, and translating the origin to
:ontinuous at the nodes. A piecewise lie at the no de jh:
vatives is called a cubic spline. This
:ding one, the Hermite cubics, with
interna! nodes. Therefore, the dimen-
:N 1) N + 2. This means that In the expression for vh, the constraint that it vanish at the origin has not
including the extra points x 0 = O beep included. If cph and <~'!* are combinations of <p~ and cp~ which satisfy this
regarded as the free parameter) and essential condition, the admissible cubic splines can be characterized exactly
te last patameter be v~). At interna! as
~ unknown, and the finite element
· like conventional finite difference

,dal parameters determine the shape

say (j- I)h < x <jh. The nodes
tl, account for only two conditions,
side. Therefore, cubic splines cannot
f v" within an:element-must be affect-
t. In fact, the spline which vanishes
will be nonzero over every subinter-
!d to construct the one which equals
.s rapidly as possible (Fig. 1.9). This
pline. It is fundamental to the. whole
tany remarkable discoveries of l. J.
In this form the N+ 2 unknowns in the firtite element system KQ = F
would be q*,, .. , qN+l· The unknown qN+l appears because the natura~
boundary condition imposes no constraint in the variational formulation;
it would be interesting to see the effect on u" of satisfying this condition and
thereby removing the last unknown. Presumably this effect is exponentially
small.
Two final remarks on splines:
l. It may be that their greatest importance comes in approximation rather
than minimization. A given relationship between data is well and conveniently
approximated by a spline, while the search for an unknown spline by minimiza-
tion is less convenient.
62
J
AN INTRODlJCTION TO THE THEORY CHAP. 1 SEC. ].8. BOUNDARY-VJ

2. With unequally spaced nodes they are still nonzero over four intervals Equation (45) represents the bendh
(the mínimum for a cubic spline), and if every node x 2 N-t approaches x 2 N, the boundary conditions at that erid a
the E-splines degenerate into the Hermite basis functions rp and ro.
u(O) =u
It is easy to summarize the approximation properties ofthese more refined
spaces. If ~he polynomials are of degree k 1 (k = 3 for quadratics, k 4 and these are essential: Every trial fm
for cubicsi~;~then any smooth function u differs from its interpolate by origin. To see which are the natural col
in the equation a(u, v) = (J, v) for th
result is that for every v in the admiss

The errors in the derivatives lose one power of h for every differentiation: (46) f [(ru")"- (pu')' + qu- f]i
(44)
Th us if no conditions are imposed on 'l
This estímate makes sense only if u1 is known to possess s derivatives, in tions on u are those which physically
other words, only if the piecewise polynomials lie in :res. Therefore, there is end:
a limit s < q on the inequality (44): q = 1 for the eo quadratics and cubics, u"(n) =O, (p
which are in :JC 1 , q = 2 for Hermite cubics, and q = 3 for splines. For s > q
the estímate can still be proved element by element, but ó-functions appear There is an additional case of great in
at the nodes. supported: u(n) =O. This becomes :
These approximation results lead to the expected rates of convergence functions v must satisfy. Therefore, 1
for the finite element method, provided that the derivative u<k> has finite is automaÍically zero, with no conditi
energy: the slopes are in error by O(hk-t), the strain energy by O(h 2 <k-u), grated term is zero for all v only if í
and the displacements u uh by O(hk). Since these rates are confirmed must remain. Thus a simply supportc
numerically, there is good reason to compute with refined finite elements. natural boundary condition:
The final step in this section is to admit differential equations of fourth
order, u(n~ = 1

(45) Lu (ru")' 1 - (pu')' + qu =f. Other combinations of boundary c•

The operator: L is still formally self-adjoint, since u"' and u' are not allowed u'(n) = u"(n) O. But these are we
to occur by themselves, and it is positive definí te if r > rm in > O, p > O,
q o. 1.8. BOUNDARY-VALUE PROBLEM
The associated energy inner product is
DIMENSIONS

a(u, v) = f (ru"v" + pu'v' + quv) dx, This section introduces severa! pn

domain D bounded by a smooth cun
and the Euler equation for the minimization of I(v) = a(v, v) - 2(/, v) the differential forms of these probler
is Lu =f. For the Ritz method to apply, the trial functions vh must have the biharmonic operator 4 2 appear,
finite energy, which means that vh must lie in :JC 2 • The Hermite and spline This means that in the variational stat
cubics are therefore applicable, but not the polynomials which are only con- spaces in which the solution is sough
tinuous. They are "nonconforming," and to use them by ignoring the ó- boundary conditions, and as in-the t'
functions in their second derivatives at the nodes would in one dimensiort will be a distinction between Dirichl
guarantee disaster. Neumann conditions (natural con
 CHAP.
1ey are still nonzer.o over foui intervals
d if every node x 2 N-I approaches x 2 N~
mite basis functions rp and ru. ·
mation properties ofthese more refined
ree k 1 (k = 3 for quadratics, k = 4
1 u differs from its irr~erpolate by
~ Chk 11 u<k> llo·
~ power of h for every differentiation:
is known to possess s derivatives, in
ynomials 1ie in :JC.s. Therefore, there is
= 1 for the eo quadratics and cubics
Jbics, and q 3 f.or splines. For s > ~
nt by element, but J-functions appear
to the expected rates of convergence
:fed that the derivative u<k> has finite
hk- 1), the strain energy by O(h2<k-I>),
:hk). Since these rates are confirmed
compute with refined finite elements.
admit differential equations of fourth
~pu')' + qu f
joint, since u"' and u' are not allowed
sitive definite if r r min > O'P >
_ O'
::t is
+ pu'v' + quv) dx,
timization of I(v) a(v, v) - 2(/, v)
•ply, the trial functions vh must have
st lie in 3C 2 • The Hermite and spline
t the polynomials which are only con-
and to use them by ignoring the J-
tt the nodes would in one dimension
1 SEC. 1.8. BOUNDARY-VALUB PROBLEMS IN TWO DIMENSIONS 63
Equation (45) represents the bending of a beam. If it is c1amped at x = O,
the boundary conditions at that erid are
u(O) = u' (O) = O,
and these are essential: Every trial function must have a double zero at the
origin. To see which are the natural conditions at x = n, we integrate by parts
in the equation a(u, v) (f, v) for the vanishing of the first variation. The
result is that for every v in the admissible space 3C};,
(46) J [(ru")" (pu')' + qu f]v + ru"v'ln + (pu'- (ru")')??ln =O.
Thus if no conditions are imposed on v at x = n, the natural boundary condi-
tions on u are those which physically correspond to a free (or cantilevered)
end:
u"(n) =O, (pu' - ru'")(n) = O.
There is an additional case of great importance, in which the beam is simply
supported: u(n) O. This becomes an essential condition, which the trial
functions v mtist satisfy. Therefore, the last term in the first variation (46)
is automatically zero, with no condition on u at n. However, the other inte-
grated term is zero for all v only if u"(n) = O, and this natural condition
must remain. Thus a simply supported end combines one essential and one
natural boundary condition:
u(n) u"(n) O.
Other combinations of boundary conditions could be imagined, such as
u'(n) u"(n) = O. But these are weird, both physically and variationally.
1.8. BOUNDARY-VALUE PROBLEMS IN TWO
DIMENSIONS
This section introduces severa] problems in the plane, or rather in aplane
domain n bounded by a smooth curve r. Our purpose is first ofall to match
the differential forms of these problems, in which the Laplace operator 11 and
the biharmonic operator 11 2 appear, with the equivalent variational forms.
This means that in the variational statement we ha veto identify the admissible
spaces in which the solution is sought. Naturally these spaces depend on the
boundary conditions, and as inA:he two-point boundary-value problem, there
will be a distinction between Dirichlet conditions (essential conditions) and
Neumann conditions (natural conditions). The examples are extremely
 '¡
64 AN INTROJ?~TION TO T.HE THEORY CHAP. 1 SEC. 1.8. BOUNDARY-V

conventional, but they are the simplest models for plane stress and plate The sign in the differential equation is
bending~ and it seems worthwhile to illustrate once more the basic ideas:
L -11 =
l. The equivalence of differential and variational problems, the admissible
space for the latter being constructed by completion in the energy norm.
2. The vanishing of the first variation, giving the weak form a(u, v) will be associated with a positive, n
(f, v) of th~ equation, and leading to Galerkin;s method. (Lu, u).
3. The Ritz process of minimization over a subspace. As in the one.. dimensional exampl
the inhomogeneous term, in other w
Then the next section introduces the finite element method in earnest, includ- which the Dirichlet problern is to be
ing many of the most important choices of piecewise-polynomial "elements." . same as before:
The requirement that the boundary curve r be smooth introduces one
difficulty while preventing another. On the one hand, it is certain that the
11/llo (J JIJ
n
interior O cannot be subdivided into polygons, say triangles, without leaving
a· "skin" around · the boundary. The approximation theory in Chapter 3. lf this norm is finite, then f belongs
will reflect this difficulty. On the other hand, the smoothness of the boundary dimensional .Problems, this space inc
makes it reasonable to suppose that the solution is itselfsmooth. This property excludes á-functions.
follows rigorously from the· theory of elliptic boundary-value problems, As possible so1utions to the diffé1
if the coefficients in the equation and the data are also smooth. u which vanish on the boundary r at
In contrast, consider the problem uxx + uYY = 1 in a polygon, with u= O in O. The natural norm for this solut
on the boundary. For the unit square, the solution behaves radially like
r 2 log r near a corner, and the second derivatives break down. (Sorne break-
down is obvious, since at the corner uxx and uYY both vanish, while their sum
llullz [JJ (u
n
2
+u;+ u;
is one just inside.) In this problem u does ha ve second derivatives in the mean-
square sense; u lies in X 2 but not in X 3 • This can be verified directly by con- The space of all functions for whicl
structing u as a Fourier series. Therefore, the error estimates for piecewise second derivatives have finite energ
linear approximation will apply to this problem, but the accuracy associated for the Dirichlet problem is then,a sul
with higher-order elements cannot be attained without either refining the ary condition u o on r.
mesh at the corners, or introducirtg special trial functions which possess It should be clear from the defi1
the same kind of singularity as the solution u. For a nonconvex polygon, operator L -11 is a bounded oper;
for example an L-shaped region, u fails evento have second derivatives in a
l!Lullo::
mean-squáre sense. The solution must lie in X 1 , which is the admissible space
(or rather X 1 contains the admissible space, depending on the boundary_ The crucial point of the Dirichlet the
conditions), but since u behaves like r 2 13 at the interior angle of the L, it is the inverse of L-which is given by
excluded from X 2 • Such singularities, arising when r is rtot smooth, are yields a solution u which depends C4
studied in Chapter 8. that to every f there corresponds one '
We begin this section by studying the Dirichlet problem, governed inside
the domain by Poisson's equation (47)

-11u =f in n, The construction of this solutio

difference methods, most commonly
and on the boundary by the Dirichlet condition
(48)
u=O onr.
 CHAP. 1
est models for pl~ne stress ~nd plate
illustrate once more the basic ideas:
nd variational problems, the admissible
i by completion in the energy norm.
Ltion, giving the weak form a(u, v) =
Galerkin 's method.
m over a subspace.
1ite element method in earnest, includ-
~s of piecewise-polynomial "elements."
·y curve r be smooth introduces one
n the one hand, it is certain that the
olygons, say triangles, without leaving
approximation theory in Chapter 3
hand, the smoothness of the boundary
solution is itselfsmooth. This property
Jf elliptic boundary-value problems,
1e data are also smooth.
-"x + u:r:r 1 in a polygon, with u O
·e, the solution behaves radially like
lerivatives break down. (Sorne break-
' and u:r:r both vanish, while their sum
s ha ve second deriva ti ves in the mean-
:. This can be verified direct1y by con-
ue, the error estimates for piecewise
problem, but the accuracy associated
attained without either refining the
pedal trial functions which possess
lution u. For a nonconvex polygon,
even to have second derivatives in a
e in X 1 , whích is the admissible space
space, depending on the boundary
3 The crucial point of the Dirichlet theory goes in the opposite directioQ: that
' at the interior angle of the L, it is
arísing when r is riot smooth, are
e Diríchlet problem, governed inside
in n,
mdition
on r.
SEC. 1.8. BOUNDARY-VALUE PROBLEMS IN TWO DIMENSIONS 65
The sign in the differential equation is chosen so that the operator
L=-A
will be associated with a positive, rather than a negative, quadratic fórm
(Lu, u).
As in the one-dimensional example, we need to prescribe sorne norm for
the inhomogeneous term, in other words to choose sorne set of data f for
which the Dirichlet problem is to be solved. Our choice of norm will be the
same as befo re:
11 f llo (JJ1f(x, Y) 1 dxdy)
2
112
n
lf this norm is finite, then j belongs to the data space X 0 (0). ·As in one-
dimensional .Problems, this space includes any piecewise continuous f and
excludes c5-functions.
As possible solutions to the differeniial equation we admit alJ functions
u which vanish on the boundary r and have derivatives up to second order
in Q. The natural norm for this solution space is
11 ull2 = [J J(u 2+u;+ u;+ u;x + u;:r + u;:r) dxdyJ12 ·
n
The space of all functions for which this norm is finite-functions whose
second derivatives have finite energy_!_is X 2 (Q). The solution space Xi
for the Dirichlet problem is then a subspace of X 2 , determined by the bound-
ary condition u= o on r.
It should be dear from the definitions of the norms that the Laplace
operator L = -A is a bounded operator from Xi to X 0 :
IILullo < Kllull2·
the inverse of L-which is given by the Green's function of th~ problem-
yields a solution u which depends continuously on the data f. This means
that to el'ery f there corresponds one and only one u, and for some constant p,
(47) 11 U llz P 11 f llo ·
The construction of this solution could be attacked directly by finite
difference methods, most commonly by the .five-point difference equation
(48)
66 AN INTRODUcjiON TO THE THEORY CHAP. 1 SEC. 1.8. BOUNDARY-VA

Near the boundary this equation must be modified, but it is possible to pre- on r, and u satisfies Poisson's equatic
serve bÓth the second-order accuracy of the scheme and the discrete maxi- Therefore, u is minimizing.
mum principie which is obviou~ from (48): If/ =O, then U1,¡ cannot exceed Of course, this process works equal
all four ofthe values U1±t,J±t· (We do not know whether there is a theoretical Poisson equation -Au =fis put into
limit to the order of accuracy of schemes satisfying a maximum principie. v which vanishes on the boundary, inte
It is certain that as the accuracy increases, the boundary conditions on the Ieft side by Green's theorem. The we~
difference ;~quation will become extremely complicated.) 1t is to this form that Galerkin's meth
Withou't· going into details, we note that the equation implicitly requires on a subspace Sh and the solution uh is
fto be continuous, in order for J;,¡ f(i Ax,j Ay) to be well defined. When a(v, v) is self-adjoint and positive definit
f is less smooth, sorne averaging process (which will be built into the varia- vanishing of the first variation are equi
tional methods !) may be applied. The best estima te for the five-point scheme (f, v) retains its meaning, however, eve
in a square seems to be This argument depended on the th
u. Supp'ose, instead, that we begin wi
max 1U 1, ¡ - u1,¡ 1< Ch 2 1ln hl max 1/1· its J)linimum directly. Then the first tas
i,j
v are admissible in the minimization.
Here h Ax Ay. The extra factor !In h l, which was not needed ,in one The reasoning runs parallel to the on
dimension, is required by the r 2 Jog r behavior ata corner. Any smooth v which vanishes on r mus1
For us the important formulation ofDirichlet's problem is the variational happen to be _:_Av, this function v woul
one: the admissib/e functions V vanish on the boundary r, and the so/ution U On the other hand, the requirement tb
is the one which minimizes the quadratic functional is more than is necessary for I(v) to b(
admissible space: if a(v- vN, v- vN:
I(v) = JJ(v; + v; - 2fv) dxdy. then v will also be admissible. This e
n which is intuitively natural: v must va
an essential one), but it has only to poss
The first step is to check that a solution u to the differential problem does sense. In other words, the quadratic ter
minimize /. Perturbing u in any direction v, make sense,

(51) a(v, v) JJ (v; -i

We have only to show that the coefficient of E is zero. Then since the coeffi- The functions satisfying (51) lie in 3C 1 l
cient of é is positive unless vis constant, which would imply by the boundary the Dirichlet condition V = o on r is d
condition that v O, u will be the unique function which minimizes /. The would be 3C1, but in the special case
vanishing of the coefficient of f, that is, of the first variation, means that o .
3CA; the notation 3C 1 also appears in thc
sible functions for the Dirichlet problem
(49) for all admissible v. It is important to emphasize that
retained for the whole admissible spacc
To prove that the solution to the Dirichlet problem satisfies this identity, stubborn enough to remain vaJid in th
we apply Green's theorem; or in other words, we integrate by parts o ver n: condition v O is stable in the 3C 1 nO:
and converges in the natural strain en
(50) -ff (Uxx + Uyy + f)v dxdy + fr UnV ds = 0, vanishes on r.
n This stability of the boundary con<

where un is the derivative in the direction ofthe outward normal. Since v ·o (52) -Au + qu =
 CHAP)l
\t be modified, but it is possible to'pre-
' of the scheme and the discrete maxi-
(48): If/ =O, then U1,¡ cannot exceed
not know whether there is a theoretical
emes satisfying a maximum principie.
·eases, the boundary conditions on the
!mely complicated.) ~··\
te that the equation i'inplicitly requires
f(i Ax,j Ay) to be well defined. When
;ess (which will be built into the varia-
best estímate for the five-point scheme
~ Ch 2 jln h 1 max 1f j.
· jln h j, which was not needed in one
Jehavior ata corner.
of Dirichlet's problem is the variational
on the boundary r' and the solution u
ic functional
v; - 2/v) dxdy.
1tion u to the differential problem does
ion v,
.ent of E is zero. Then since the coeffi-
lt, which would imply by the boundary
ique function which minimizes l. The
is, of the first variation, means that
for all admissible v.
tirichlet problem satisfies this identity,
· words, we integrate by parts over Q:
dxdy +Ir unv ds O,
1n of the outward normal. Since v O (52)
SEC. 1.8. BOUNDARY-VALUE PROBLEMS IN TWO DIMENSIONS 67
on r, and u satisfies Poisson's equation in Q, the first variation vanishes.
Therefore, u is minimizing.
Of course, this process works equally well in the opposite direction. The
Poisson equation -Jiu = f is put into its weak form by multiplying by aily
v which vanishes on the boundary, integrating over Q, and transforming the
left side by Green's theorem. The weak form is exactly the equation (49).
It is to this form that Galerkin's method app1ies; this equation is imposed
on a subspace Sh and the solution uh is also in Sh. As we see, when the form
a(v, v) is self-adjoint and positive definite, the Ritz minimization and Galerkin's
vanishing of the first variation are equivalent. The weak equation a(u, v)
(/, v) retains its meaning, however, even without self-adjointness.
This argument depended on the theory of elliptic equations to produce
u. Suppose, instead, that we begin with the quadratic I(v) and try to find
its mínimum directly. Then the first task is to decide exactly which functions
v are admissible in the minimization.
The reasoning runs parallel to the one-dimensional example of Section 1.5;
Any smooth v which vanishes on r must be made admissible, since if f should
happen to be -Av, this function v would be the minimizing solution we want.
On the other hand, the requirement that v be smooth, say that v lie in X~,
is more than is necessary for I(v) to be defined. Therefore, we complete the
admissible space: if a(v- vN, v vN)---+ O for sorne sequence vN in X~,
then v will also be admissible. This completion process ·leads to the class
which is intuitively natural: v must vanish on r (the Dirichlet condition is
an essential one), but it has only to possessjirst derivatives in the mean-square
sense. In other words, the quadratic term a(v, v) in the functional I(v) should
mak~ sense,
(51) a(v, v) JJ(v! + v;) dxdy < oo.
The functions satisfying (51) líe in 3C 1(Q); the subspace which satisfies also
the Dirichlet condition v O on r is denoted by 3CA(Q). (Our own notation
would be 3C 1, but in the special case of the Dirichlet problem we accept
o
3CA; the notation 3C 1 also appears in the literature.) This is the space of admis-
sible functions for the Dirichlet problem.
It is important to emphasize that the boundary condition v O was
retained for the whole admissible space, not by fiat, but only because it was
stubborn enough to remain valid in ·the limit. In other words, the Dirichlet
condition v = O is stable in the 3C 1 norm: If the sequence vN vanishes on r
and converges in the natural strain energy norm to v, then the limit v also
vanishes on r.
This stability of the boundary condition is not present. for the equation
-Jiu+ qu = f in Q
68 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.8. BOUNDARY-VAI

with a natural boundary condition This differenée between the Dirid

appear also in the theory, at the poin
un o on r. ~ is equivalent to the Stf!ndard
The functional for this Neumann problem is ThlJS the crucial question for existence ,
problem is elliptíc; is there a constant (
I(p) = a(v, v)- 2(/, v) = JJ (v! + v; + qv 2
- 2fv) dxdy.
(53) a(v,v) ullvm f
The differential equation looks for a solution u in Xi, that is, possessing two For the equation -Au + qu this i
derivatives and satisfying the Neumann condition un O. However, every
function v in X 1 is the limit of a sequence taken from Xi; Xi is dense in
X 1 • After completion, therefore, the admissible space for the Neumann
variational prob/em is the who/e space X 1 (Q). As a result, no boundary condi-
tions are imposed· on the trial functions vh in the Ritz method; any subspace With q > O this ellipticity is obvious,
Sh e X 1 is acceptable. In practica} applications ofthe finite element method~ (or uh, in the Ritz method). With q =O t:
this means that the values of vh at boundary points are not constrained, yield- lem; if v 1, the left si de is zero and ·
ing a modest simplification in comparison with the Dirichlet problem. (In fact, problem, the rigid body motion v 1
the increased numerical instability due to natural boundary conditions leads of Poincaré type does hold:
us to doubt that the advantages are entirely on the side of Neumann.)
Of course the minimizing function u (not the approximate mínima uh)
must somehow automatically satisfy the Neumann condition when enough
smoothness is present. This is confirmed l?Y equation (50) for the first varia- Thus, ellipticity holds even for q = O in
tion, which now vanishes for all v in the admissible space X 1(Q). In the city simply means that q exceeds Amax' ·
first place, W = Uxx + Uyy + j must V~nish throughout Q, since V may be operator A. In the Neumann case Amax =
chosen equal to w in any small circle inside r and zero elsewhere, giving q = O fails to achieve ellipticity. In the 1
JJ w2 O. Once w O is established, it follows easily that un = O on the lem would even remain elliptic for a rar
boundary. Therefore, the Neumann conditiori holds for u, even though it It is, of course, possible to impose u
is not imposed on all admissible v. say rl, and to have un o on r2 r
If q =O in (52), then obviously no solution u is unique; u+ e will be mixed prob/em contains those v in X 1 '
a solution for any constant c. This single degree of freedom suggests, if will ha ve a singularity at the junction b
Fredholm 's theorem of the alterna ti ve is to hold, that there must be a single A further possibility is that the bom
constraint on the data f Integrating both si des of the differential equation of an oblique derivative:
- Au f o ver Q, the left si de vanishes by Green's formula (50) with v· l.
Therefore, the constraint onjis: No solution to the Neumann problem with un+ c(x, y)u$ =
q = O can exist unless JJf dx dy = O. This nonexistence can be verified
explicitly for the one-dimensional Neumann problem where us is the tangential derivative. TI
associated with the inner product
u'' 2, u'(O) = u'(l) O.

The parabola u = x 2 + Ax + B, which gives the general solution to u" = 2,

cannot be made to satisfy the bouhdary conditions. Correspondingly,
This inner product yields Poisson's equ
uxvx + uyvy. In fact, Green's identity tr
J u" u'(l) - u'(O) = O =1= J 2.
In contrast, the solution to the Dirichlet problem is unique. - JJ(Au + f)v + J
 CHAP.
on r.
blem is
f (v;_ + v; + qv 2
1,fv)
["·,
dxdy.
;olution u in x¡, that is, possessing two
ann condition u, O. flowever, every
::¡uence taken from Xi; Xi is dense in
1e admissible space for the Neumann
3C 1 (!l). As a result, no boundary condi-
ns v11 in the Ritz method; any subspace
·plications of the finite element method,
ndary points are not constrained, yield-
son with the Dirichlet problem. (fn fact,
1e to natural boundary conditions leads
mtirely on the side of Neumann.)
·n u (not the approximatf! mínima u11)
the Neumann condition when enough
ted by equation (50) for the first varía-
n the admissible space X 1(!l). In the
vanish throughout !l, since v may be
e inside r and zero elsewhere, giving
1, it follows easily that u, o on the
:=ondition holds for u, even though it
10 solution u is unique; u+ e will be
single degree of freedom suggests, if
is to hold, that there must be a single
both sides of the differential equation
s by Green's formula (50) with v· l.
olution to the Neumann problem with
). This nonexistence can be verified
tmann pro,blem
)) u'(l) =O.
1 gives the general solution to u" = 2,
ry conditions. Correspondingly,
t'(O) = O =F f 2.
et problem is unique.
1 SEC. 1.8. BOUNDARY-VALUE PROBLEMS IN TWODIMENSIONS 69
This differenée between the Dirichlet and Neumann conditions must
appear also in the theory, at the point of verifying that the energy norm
Ja(v,V) is equivalent to the st_andard norm 11 v 11 1 on the admissible space.
Thus the crucial question for existence and uniqueness is whether or not the
problem is elllptic; is there a constant a > O such that
(53) a(v,v) allvlli for all admissible v?
For the equation - Au + qu = f, this is the same as
With q > O this ellipticity is obvious, and there is a unique minimizing u
(or u\ in the Ritz method). With q ==O ellipticity failsfor the Neumann prob-
lem; if v = 1, the left si de is zero and the right si de is no t. In the Dirichlet
problem, the rigid body motion v = 1 is not admissible, and an inequality
of Poincaré type do es hoid:
JJv; + v; > a' JJvz for v in Xó.
Thus, ellipticity holds even for q O in the Dirichlet problem. In fact, eiJipti-
city simply means that q exceeds 1malo the largest eigenvalue of the Laplace
operator A. In the Neumann case Amax O, with v = 1 as eigenfunction, and
q =O fails toachieve ellipticity. In the D~richlet case Amax <O, and the prob-
]em would even rema in elliptic for a range of negative values of q.
It'is, of course, possible to impose u= O on on]y a part ofthe boundary,
say r 1 , and to ha ve u, O on r 2 r r 1 • The admissible space for this
mixed problem contains those v in X 1 which vanish on r P .and the solution
will ha ve a singularity at the junction between r 1 and r 2'
A further possibility is that the boundary condition express the vanishing
of an oblique deriJ•atil'e:
u,+ c(x, y)us o on r,
where us is the tangentia1 derivative. This is the natural boundary condition
associated with the inner product
This inner product yields Poisson's equ~tion just as surely as the integral of
uxvx + uyvy. In fact, Green's identity transforms a(u, v) = (/, v) into
-ff (Au + f)v + f (u,+ cus)v ds =o.
70 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.8. BOUNDARY-V)

This illu.strates that there is no unique way of integrating (Lv, v) by parts to We emphasize that the first new term
obt,ain an energy norm a(v, v). Different manipulations of(Lv, v) lead to differ-
ent forms for a(v, v) and to correspondingly different natural boundary
conditions. This is illustrated again below, where Poisson's ratio enters the
a( v, v) = JJ(v;
boundary condition and the energy form a(v, v) but not the operator L = A2 •
Inhom9geneous boundary conditions for the equation - Au = f are of and thus contributes to the stiffness n
two kinds/:either a prescribed displacement linear term - 2bv adds a boundary ce
admissible space is the w~hole of3C 1 , a
u = g(x, y) on r which is continuous between elements
We want to propose also three pro
ora prescribed traction, which can be included in the general case ofNewton's the biharmonic equation
botindary condition
(54)
un + d(x, y)u = b(x, y) onr.
This equation is a model for the tran
Variationally these two are completely different. 'fhe first is an inhomogene- bent by the body force f(x, y), with b
ous Dirichlet condition and must be imposed on the trial functions; the solu- As usual, the number m of boundary <
tion u minimizes JJ v; + v; - 2fv over the class offunctions v in JC 1 satisfy- · equation, so that m = 2.
ing v = g on the boundary. Notice that this admissible class JC1 is no longer A first possibility is to impose Diri
a "space" of functions; the su m of two admissible functions would equal the problem of a clamped plate:
2g on the bounda·ry and would be inadmissible. However, the difference
of any two admissible functions vanishes on the boundary ~nd líes exactly u= O and u
in the space JCJ. The easiest description of JC1 is therefore the following:
Choose any one member, say G(x, y), which agrees with g on the boundary, An alternative of great interest is the
and then every admissible v equals G + v 0 , where v 0 lies in JCJ. In short, in which one boundary condition is
natural. As in the oblique derivative 1
JC1 = G + JCJ.··. equation, the form of the natural cm
variational integral l(v). In elasticity th
In the Ritz method it is not requiréd. that the tria/ functions agree exactly with ratio v, which measures the change in
g on the boundary. lt is enough if thé trial functions are of the form vh = Iengthwise; v = 0.3 is a common choic
Gh(x_, y) + L; qitp1(x, y), where the '7'1lie in JCA and Gh nearly agrees with g. by physical reasoning are
This means that the trial class is
(55) u = O and vll.u + (J
The cases v = O and v = l, which m
and that S8 is a subspace of JCA. Section 4.4 verifies that the basic Ritz theo- course be included. Note that on a stra'
rem 1.1 applies to this case (and ultimately we remove evén the hypothesis are all zero, Au = unn' and V disapp~
that S8 e JCA, but this leads outside of the Ritz framework). remarkable and apparently paradoxic
Finally we consider the boundary condition un + du = h. This will be a for polygonal approximation of a circ:
natural boundary condition, provided the basic functional JJ v; + v; - 2fv Finally, there is the pure Neuma
is altered to take account of d and b. Both alterations appear as boundary boundary. The second of the bounda1
terms: written as
I(v) = JJv; + v; ·_ 2fv + Jr (dv 2
- 2bv)ds. -M
_
D
_nn=U
nn
+1'
Q
 CHAP.
'!way of integrating (Lv, v) by parts to
t manipulations of (Lv, v) lead to differ-
pondingly different natural boundary
dow, where Poisson's ratio enters the
~m a(11, v) but not the operator L L.\ 2 •
ons for the equatio~ - L.\u = f are of
!ment r\
·) on r which is continuous between elements can be used as a tria! function.
ncluded in the general case ofNewton's
b(x, y) on r.
' different. The first is an inhomogene-
1posed on the trial functions; the sotu-
er the class offunctions v in 3C 1 satisfy-
Lt this admissible class 3C1 is no Ionger
wo admissible functions would equal
nadmissible. However, the difference ·.
:hes on the boundary and lies exactly
:on of X 1 is therefore the following:
which agrees with g on the boundary,
+
V 0 , where V 0 lies in XA. In short,
zt the tria! functions agree exactly with
trial functions are of the form vh
ie in XA and Gh nearly agrees with g.
,y)+ Sg
n 4.4 verifies that the basic Ritz theo-
ately we remove evén the hypothesis
the Ritz framework).
ondition un + du h. This will be a
te basic functional f f v! + v; - 2f v
Both alterations appear as boundary
1 SEC. 1.8. BOUNDARY-VALUE PROBLEMS IN 1WO DIMENSIONS 71
We emphasize that the first new term enters the energy
a(v, v) = JJ(v! + v;) + Jdv 2
,
and thus contri bu tes to the stiffness matrix K in the Ritz method. The new
linear term - 2bv adds a boundary contribution to the load vector F. The
admissible space is the whole of 3C 1 , and therefore any piecewise polynomial
We want to propase also three problems of fourth order, all governed by
the biharmonic equation
(54) inn.
This equation is a model for the ·transverse displacement u of a thin plate
bent by the body force f(x, y), with bending stiffness normalized to D l.
As usual, the number m of boundary conditions will be half the order of the
equation, ,so that m = 2.
A first possibility is to impose Dirichlet conditions, leading physically to
the problem of a clamped plate:
u= o and un= o on r.
An altemative of great interest is the próblem of a simply supported plate,
in which one boundary condition is essential (or forced) and the other is
natural. As in the oblique derivative problem described above for Poisson's
equation, the form of the natural condition will depend on the form of the
variational integral J(v). In elasticity this natural condition involves Poisson's
ratio v, which measures the change in width when the material is stretched
lengthwise; v 0.3 is a common choice. The boundary conditions determined
by physicai reasoning are
(55) u = O and vi.\ u + (1 - V )unn O onr.
The cases v =O and v = 1, which may arise in other applications, will of
course be included. Note that on a straight boundarythe tangential derivatives
are all zero, L.\u unn, and v disappears from the boundary condition; the
remarkable and apparently paradoxical consequences of this disappearance,
for polygonal approximation of a circle, are discussed in Section 4.4.
Finally, there is the pure Neumann problem, corresponding to afree
boundary. The second of the boundary conditions (55) remains; it is often
written as
 SEC. 1.8. BOUNDARY~VALl
72 AN INTRODUCTION TO THE THEORY CHAP. 1

where rp is the angle between the normal and the x~axis. The condition u = O, With all edges free ellipticity must fail, si
which fixes the edge, no longer applies. To fill its place one has to compute + +
a linear function a bx cy; with the e
the coefficient of Ju, when the energy functional in (56) below is varied. This qu = f, corresponding to a plate under
coefficient is given by Landa u and Lifshitz explicitly in terms of u and rp; this case is also elliptic.
it is normally written in the contracted Kirchhoff form Qn + aMnsfas O. Elli ptici ty Qan become difficult to pr'
In practic;;tl problems it is un usual for the whole boundary to be free. elasticity-there are two or three unknm
All th¿·se conditions can be coped with somehow by directly replacing only certain combinations e.iJ !(ut,J +
derivatives by finite differences. The construction of accurate equations near inequality which establishes that this str
the boundary becomes fantastically complicated, however, anda much better ~
~ f e. ..e. .. > fJ' ~ f U¡ Jui
l} l} - .~ ,. '
¡·· 1'" '

starting point is furnished by the variational Jorm. Before proceeding to the constructic
Variationally, the problem is to minimize of these problems,we want to comment
Jn one dimension they were deceptivel
I(v) = a(v, v) 2(f, v) :res[O, lJ if it is the s~fold integral of a
(56) guarantees that v and its first s - 1 e
JJ (v!~ + v;Y + 2vv~xvyy + 2(1 - v)v;Y 2fv) dxdy sth derivative, that is, the original funct
Q
In the. plane" it is possible for v to b
di.fferentiab/e. One such function in :JC 1 i:
subject to the appropriate boundary conditions. In the Neumann problem
there are no constraints at the boundary, and the space of admissible v is
exactly 3<! 2 (0). For the clamped plate, the Dirichlet conditions v O and Jfv;+v;= Jf[v;+
vn = O are imposed; the subspace satisfying these constraints is 3<!5(0). In
the intermediate case of the simply supported plate, only the condition v = O
is essential; we denote the corresponding subspace by x;s,
noting that
= JJ(log r)
OC5 e: x;s e: OC • Of course, it is Green's theorem which yields the equiva~
2
Thus the derivatives have finite energy,
lence of these variational problems with their differential counterparts.
is not continuous at the origin. The gent
We emphasize that the fourth derivatives which appear in the term l1 2 u
ing: lf v is in :res and s > n/2, then ~ is C1
in Green's theorem are not required for the variational formulation-the
theorem of minimum potential energy-to be valid. In fact, the opposite is (57) max 1v( x 1 , ••• , ;
true. Any limitation to solutions with four continuous derivatives would be
awkward in the extreme, and the whole idea of completion is to arrive at the This is the essence of the celebrated Sol
admissible space under mínimum restrictions, requiring only that the es~ erties of continuity and finite energy o
sential boundary conditions hold and that the energy a(v, v) be finite. The s nf2, it cannot be guaranteed that
solution u then satisfies the equation in its weak form a(u, v) (f, v). By duality, Sobolev's rule will als'
F or the theory of the Ritz method, the key point is that the energy norm function lies in :re-s. Of course, this is
,.ja(v, v) should be equivalent to the standard norm llvl! 2 • This is again the after enough integrations that the J.
condition of ellipticity: For sorne fJ' > O and all admissible v, The norm on :re-s is defined as in (12):

a(V, V)> fJ' 11 V lli

or (58) llwll-s max
ti
'11'Vjl)l =
S
1

If w is a J-function, say at the origin, t

JJv!~ + v;y + 2vvxxvyy + 2(1 v)v!Y

> fJ' JJ(V + v! + v; + v!~ + v!y + v;y).

2
)RY CHAP.
·rmal and the x-axis. The co~dition u = O
?líes. To fill its place one has to comput~
gy functional in (56) below is varied. This
l Lifshitz explicitly in terms of u and rp;
.cted Kirchhoff form Qn + aMnsfas O.
or the whole boundary to be free.
ped with somehow ~y directly replacing
~ construction of accurate equations near
complicated, however, and a much better
·ariational form.
minimize
y conditions. In the Neumann problem
ndary, and the space of admissible v is
!lte, the Dirichlet conditions v = O and
atisfying these constraints is 3CMO). In
upported plate., only the condition v = O
Jonding subspace by x;n noting that
reen's theorem which yields the equiva-
v'Íth their differential counterparts.
:rivatives which appear in the term A2u
d for the variational formulation-the
v-to be valid. In fact, the opposite is
h four continuous derivatives would be
>le idea of completion is to arrive at the
!strictions, requiring only that the es-
d that the energy a(v, v) be finite. The
in its weak forro a(u, v) = (J, v).
l, the key point is that the energy norm
standard norm llv 1! 2 • This is again the
> O and aH admissible v,
1 SEC. 1.8. BOUNDARY-VALUE PROBLEMS IN 'IWO DIMENSIONS 73
With all edges free ellipticity must fail, since the solution is unique only up to
a 1inear function a+ bx + cy; with the differential problem altered to A.2u +
qu ,corresponding to a plate under steady rotation with q pro 2 > O,
this case is aJso elliptic .
Ellipticity can become difficult to prove when-as in problems of linear
dasticity-there are two or three unknowns u1, and the strain energy involves
only certain combinations s .. = zt (u . . +u,¡, 1) oftheir derivatives. It is Korn's
lj l,j
inequality which establishes that this strain energy dominates the 3C 1 norm,
"f
,¿_,¡ et.e..
J l} ,¿_,¡ fu ..u...
u"· l,j l,j , •
Befo re proceeding to the construction of finite elements for the solution
of these problems, we want to comment briefty on the function spaces acs(Q).
1 In one dim~nsion they were deceptively ..simple to describe; v belongs to
3Cs[O, IJ if it is the s-fold integral of a function f with f / 2 dx < oo, This
r guarantees that v and its first s - 1 derivatives are continuous; only the
I sth derivative, that is, the original function J, might have jumps or worse.
In the plane' it is possible for v to be discontinuous and at the same time
differentiqb/e; One such function in 3C 1 is v Iog log 1/r, on the circle r -!:
JJ v; + v; JJ[v; (v:YJrdrdO
= JJ(log r)-: d(log r) dO =
2
;: ·
1 2
Thus the derivatives have finite energy, and so does v itself, but the function
is not continuous at the origin. The general rule in n dimensions isthefollow-
ing: Jf v is in ¡res and s > n/2, then v is continuous and
(57)
This is the essence of the celebrated Sobolev inequality, relating the two prop;.
erties of continuity and finite energy of the derivatives. If v líes in acs with
s < n/2, it cannot be guaranteed that · v is continuous.
By duality, Sobolev's rule will also decide when the n-dimensional o-
function líes in ¡re-s. Of course, this is impossible unless -:S < O; it is only
after enough integrations that the o-function might have finite energy.
The norm on ¡re-s is defined as in (12):
(58) llw 11-s
If w is a o-function, say at the origin, this gives
l.¡ o¡¡ -s. = max
v
lv(O)I ·
11 V lis
74 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.9. TRIAN

Accord~ng to Sobolev's inequality (57), this is :finite and the ó-function is great importance; that is a rare and h
in x-s if and only if s > n/2. piecewise polynomials which are de1
In particular, the ó-function in the plane' is not in x- 1, and correspond- ofnodal values, and yet have the desi
ingly the fundamental solution of Laplace's equation, u = log r, is not in mation.
:JCl: There are a great many competi
writing whether it is more efficient t
00, into quadrilaterals. Triangles are ob'
boundary, but there are advantag~
Therefore a point-loaded membrane is, strictly speaking, inadmissible in the rectangles) in the interior: there an
variational prob/em. simple elements of high degree. The:
For a plate the situation is different, since the differential equation is of choice may be a mixture of the two ¡
order 4. The solution space is 3C 2 (with boundary conditions) and the data by nodes which ensure the required
space is four derivatives lower, that is, x- 2 • In the plane the ó-function does We begin by subdividing the ba
lie in ::JC- 2 , and a point-loaded plate is acceptable.
There is one additional question about these function spaces 3'C 8 which r
is fundamental for :finite élements: What is the condition for an e/ement to be
conforming? In other words, given a differential equation of order 2m in n
independent variables, which piecewise po/ynomials /ie in the admissible space
3C~? It is easy to check on the essential boundary conditiorts; the only
question is how smooth the element must be to líe in :JCm ~
The standard conforming condition is well known: The trial function and
its :first m - 1 derivatives should be continuous across the element boundaries.
This condition is clearly sufficient for admissibility, since the mth derivatives ·
have at worst a jump between elements, and their energy ·is finite. On the
other hand, the example log log 1/r may seem to put in doubt the necessity
of this condition for conformity; there exist.functions which do not have
m - 1 continuous derivatives, but which nevertheless lie in 3Cm and are
admissible. Fortunately, however; such troublesome functions cannot be Fig. 1.10 Subdivision of the 1
piecewise polynomials. If a function v is a polynomial (or a ratio of polyno-
mials) on each side of an element boundary, then v /ies in xm if and only if The union of these triangles will be
every derivative of order /ess than m is continuous across the boundary. The a curved boundary-there will be a
success ofthe finite element method lies in the construction of such elements, edge of the jth triangle will be denote
retaining both a convenient basis and a high degree of approximation. convenience that Qh is a subset of !
part way along the edge of another. Ir
has. to know the location of every
1.9. TRIÁ'NGULAR ANO RECTANGULAR should be carried out as far as possi
ELEMENTS matic subroutine can begin by cover
then make the necessary adjustmen
This section describes sorne of the most important finite elements in the very coarse in one part of n and m
plane. Their construction has been going on for 30 years, if we include Cou- has corners and other points of sin~
rant's early paper on piecewise linear elements, and it seems to have been one the triangulation by hand. In any
of the most enjoyable occupations in applied mathematics. lt involves not refinement of a coarse mesh, which
much more than high-school algebra, and at the same time the results are· of each triangle into four similar triang
 CHAP.
this is finite and the J-function is
.ane is not in X-I, and correspónd-
:tce's equation, u = log r, is not 'in
r drd8 = JJr- 1
drd~ = oo.
trictly speaking, inadmissible in the
since the differential equation is of
boundary conditions) and the data
2
- • In the plane the o-function does
ceptable.
ut these function spaces x_s which
is the condition for an element to be
erential equation of order 2m in n
lynomials lie in the admissible space
ial boundary conditioris; the only
:be to lie in xm.
well known: The trial function and
uous across the element boundaries.
1issibility, since the mth derivatives
and their energy is finite. On the
seem to put in doubt the necessity
!xist functions which do not have
:h nevertheless lie in x_m and are
troublesome functions cannot be
L polynomial (or a ratio of polyno-
uy, then v lies in x_m if and only if
mtinuous across the boundary. The
; the construction of such elements,
igh degree of approximation.
•st important finite elements in the
on for 30 years, ifwe include Cou-
ents, and it seems to have been one
plied mathematics. It involves not
at the same time the results are of
1 SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 75
great importance; that is arare and happy combination. The goal is to choose
piecewise polynomials which are determined -by a small and convenient set
of nodal values, and yet have the desired degree of continuity and of approxi-
mation.
There are a great many competing elements, and it is not clear at ·this
writing whether it is:more efficient to subdivide the region into triangles or
into quadrilaterals. Triangles are obviously better at approximating a curved
boundary, but there are advantages to quadrilaterals (and especially to
rectangles) in the interior: there are fewer of them, and they permit very
simple elements of high degree. These remarks already suggest that the best
choice may be a mixture of the two possibilities, provided they can be united
by nodes which ensure the required continuity across the junction.
We begin by subdividing the basic region Q into triangles (Fig. 1.10).
r
Fig. 1.10 Subdivision of the polygon Qh into irregular triimgles.
The union of these triangles will be a polygon 0\ and in general-if r is
a curved boundary-there will be a nonempty "skin" O- Qh. The Iongest
edge ofthe jth triangle will be denoted by hj, and h max hr We assume for
convenience that Qh is a subset of Q and that no vertex of one triangle líes
part way along the edge of another. In practice-since the computer eventually
has to know the location of every vertex zi (xj, yj)-the triangulation
should be carried out as far as possible by the computer itself. A fully auto-
matic subroutine can begin by covering O with a regular triangular grid and
then make the necessary adjustments at the boundary. If the grid is to be
very coarse in one part of Q and more refined in another, or if the domain
has corners and other points of singularity, it may be necessary to establish
the triangulation by hand. In any case, George has argued [G2] that the
reftnement of a coarse mesh, which can be achieved simply by subdividing
each triangle into four similar tria:ngles, is a mechanical procedure for which
76 AN INTRODUCTION TO THE THEORY CHAP • .1 SEC. 1.9.

the co.mputer is ideally suited. This a1lows the user to begin by describing as a linear combination
a coarse mesh to the computer, at a minimpm cost in time and effort, and
v"(x,y)
then to work eventually with a fine mesh. Furthermore, he can more easily
carry out the one check on his program which is the mostconvincingand high-
ly recommended of ali: to test whether a reduction in h leads toan acceptable In this context the coordinate q has
1

or an unficceptable change in the numerical solution. placement vh at thejth node z1 = (x1

Givert:~a triangulation, we describe now the simplest and most basic of elements which have been developt
all trial functions. It is linear inside each. triángle (vh . a 1 + a 2 x + a 3 y) each q1 can be associated with the v
and continuous across each edge. Thus the graph of vli(x, y) is a surface made specific nodal point in the domain.
up offlat triangular pieces, joined along the edges. This is an obvious generali- The optimal coordinates Q1 are
zation of broken-line functions in one dimension. The subspace Sh composed I('L: q1rp), which is quadratic in q1, .
of these piecewise linear functions was proposed by Courant [Cll] for the ing function uh = 1: Q1rp1 is indepe.
solution of variational problems; it is a subspace of 3C1 , since the first deriva- of choosing a basis is simply to pw
tives are piecewise constant Its independent development by Turner and computable (or operational) form }¡
others was the beginning of finite elements, and the trial functions are sorne- numerical calculations are strongly .
times referred to as Turner triangles. matrix K' from a different choice of
The effect of continuity is to avoid b'-functions in the first derivatives at K' SKST for sorne S, with F repl;
interelement boundaries; without this constraint the functions are not admis- is to choose a basis in which K is spa
sible, and the (infinite) energy over the domain Q cannot be found by adding the entries of K and Fas easy to com
the separate contributions from within each element. It is here tha! the finite element
The simplicity of Courant's space líes in the fact that within each triangle, values, is an effective compromise. ·
the three coefficients of vh a 1 + a 2 x + a 3 y are uniquely determined by tioned and reasonably sparse, since t1
the values of vh at the three vertices. This means that the function can be to the same element. Furthermore,
conveniently described by giving its nodal values; equivalently, Sh will have F1 (J, rp1) can be found by a very l
a convenient basis. Furthermore, along any edge, vh reduces to a linear func- is to compute, not one inner prodt
tion of one variable, and this function is obv.iously determined by its values to all the i_nner products from one
at the two endpoints of the edge. The value of vh at the third vertex has no the integrals are computed over eacl
effect on the function along this ~dge, no matter whether this third vertex ness matrices ki. Each ki involves o
belongs to the triangle on one side or the other. Therefore, the continuity of partition; its other entries are all zc
vh across the edge is assured by continuity at the J'ertices. containing all inner products, is ass
In case of an essential boundary condition, say u O on r, the simplest was carried out on an iriterval in Sect
subspace Sh 3Có is formed by requiring the functions to vanish on the poly- lar mesh in the next section.
gonal boundary rh. Extended to be zero on the skin Q Qh, such a function .Courant's triangles yield an inteJ
equation, the standard five-point d
vh is continuous o ver the whole domain Q; it lies in 3Có, and it is admissible
in the Dirichlet problem. angles are formed in a regular way, ~
The dímension of the space Sh-the number N of free parameters in the all the diagonals in the northeast di
functions vh-coincides with the number of unconstrained nodes. (A boundary scheme can be constructed in a simi
node at which vh is required to vanish, or to equal sorne other prescribed dis- this is largely a curiosity.) The appf
placement, is constrained and does not add to.the dimension ofthe subspace.) stiffness matrix means that the fast
For proof, let rpix, y) be the trial function which equals 1 at thejth node and KQ = F; this is extremély successf
zero at all other nodes. Then these pyramid functions rp 1 form a basis for the . applications are being rapidly develo
tria[ space Sh. An arbitrary vh in S 1' can be expressed in one and only one way matically, a key property of the 5-¡

ows the user to b~gin by describing
inimum cost in time and effort, and
;h. Furthermore, he can more easily
1hich is the mostconvincingandhigh-
reduction in h leads toan acceptable
·ica] solution. , placement vh at thejth node z1 = (x1, y1). This is an important feature ofthe
now the simplest antl:.most basic of
•ach triangle (vh ·+
a 1' a2 x + a 3 y)
e graph of vh(x, y) is a surface made
he edges. This is an obvious generali-
mension. The subspace Sh composed
proposed by Courant [C 11] for the
ubspace of X 1 , since the first deriva-
:nden~~evelopment by Tumer and
1ts, and the trial functions are sorne-
i-functions in the first derivatives at
nstraint the functions are not admis-
Jmain Q cannot be found by adding
ach element.
in the fact that within each triangle,
+ a 3 y are uniquely determined by
'his means that the function can be
1 J•alues; equivalently, Sh will have
ny edge, vh reduces to a linear func-
obviously determined by its values
tlue of vh at the third vertex has no
lO matter whether this third vertex
e other. Therefore, the continuity of
t the J'ertices.
lition, say u = o on r, the simplest
the functions to vanish on the poly-
Jn the skin Q - Qh, such a function
l; it lies in Xó, and it is admissible
umber N of free parameters in the
f unconstrained nodes. (A boundary
to equal sorne other prescribed dis-
d to.the dimension ofthe subspace.)
t which equals I at the jth node and
lid functions rp1 form a basis for the ·
expressed in one and only one way
SEC. 1.9. TRIANGULAR ANO RECTANGULAR ELEMENTS 77 .
' as a linear combimítion
In this context the coordinate q1 has direct physical significance, as the dis-
elements which haye been developed for engineering computations: that
each g 1 can be associated with the value of vh or one of its derivatives ata
specific nodal point in the domain.
The optimal coordinates Q1 are now determined by minimizing I(v")
I(L: q1rp1), which is quadratic in q1, . .. , qN. We emphasize that the minimiz-
ing function uh = L: Q1rp1 is independent of the particular basis; the effect
of choosing a basis is simply to put the problem of minimization into the
computable (or operational) form KQ F. On the other hand, the actual
numerical calculations are strongly dependent on this choice. The stiffness
matrix K' from a different choice of basis would be congruent to the first:
K' = SKST for sorne S, with F replaced by F' SF. Therefore, the object
is to choose a basis in which K is sparse and well conditioned, while keeping
the entries of K and Fas easy to compute as possible.
It is here thatthe finite element choice rp1 , based on interpolating nodal
values, is an effective compromise. The matrix K is reasonably we11 condi-
tioned and reasonably sparse, since two nodes are coupled only ifthey belong
to ·the same element. Furthermore, the inner products K 11 = a(rp1, rp1) and
F1 = (f, rp1) can be found by a very fast and systematic algorithm. The trick
is to compute, not one inner product. after another, but the contributions
to alLthe i_nner products from one triangle after another. This means that
the integrals are computed over each triangle, yielding a set of element stiff-
ness matrices k 1• Each k 1 involves only the nodes of the ith triangle of the
partition; its other en tries are all zero. Then ~he global stiffness matrix K,
containing all inner products, is assembled from these pieces. This process
was carried out on an interval in Section 1.5 and will be extended toa triangu-
lar mesh in the next section.
Courant's triangles yield an interesting stiffness matrix K. For Laplace's
equation, the standard :five-point difference scheme is produced if the tri-
angles are formed in a regular way, starting with a square mesh and drawing
all the diagonals in the northeast direction. (The more accurate nine-point
scheme can be constructed in a similar way from bilinear elements [F6], but
this is largely a curiosity.) The appearance of such a simple and systematic
stiffness matrix means thát the fast Fourier transform can be used to solve
KQ F; this is extremély· successful on a rectangle, and non-rectangular
applications are being rapidly developed by Dorr, Golub, and others. Mathe-
matically, a key property of the 5-point scheme is the maximum principie:
78 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.9. TRIAN<

all off-diagonal terms Ku in the stiffness Jnatrix are negative, they are domi- C0 quadrotic
nated by the diagonal en tries Km and as· a consequence the inverse matrix
K- 1 is non-negative. A simple calculation shows that this property holds for
linear elernents on any triangulation, if no angle exceeds rc/2. (The precise
condition is that the two angles subtended by any-·given edge add to no more
than :n:.) A similar result holds for linear elernents in n dimensions. Because
K- 1 and án rpix) are non-negative, the finite element approxirnation uh obeys
the same physicallaw as the true displacement u: if the load f is everywhere
positive, then so is the resulting displacemeni. Fried has asked whether this
holds also for elements of higher degree; it will not be true that all off-diago-
nal K 11 are negative, but this is not necessary for a positive K- 1 •
In the Neurnann problem there is no constraint on vh at the boundary
nodes, and the dirnension of Sh becomes the total number of interior and
boundary nodes. The basis functions rp1 again equal one at a single node and
zero at the others. In -this case, however, defining the trial functions to be
zero on the skin would not leave them all continuous. Instead, we maysimply
continue into each piece of n Qh the linear function defined in the adjacent
triangle.
There are other possibilities, of which we rnention one-to ignore the Fig. 1.11 Nodal placement f(
skin entirely and study the Neurnann problem only within the approximate
domain Qh. Of course, the approximate solution may be significantly altered, at all others. This rule specifies rp a1
producing sorne loss in accuracy in cornparison with minimizing the integral vertices and three midpoints, thereb:
over n. Such errors, introduced by modifying the boundary-value problem, (59).
are estimated in Chapter 4. The crucial point to decide is wl
We turn now toa more accurate and more refined element. This was a determined by this rule in each sepa
decisive step in the finite element techniqu<(, to generalize the basic idea edges between triangles. The proof is
behind Courant's simple trial functions. Rathér than assuming a linear func- vh is a quadratic in one variable,' and
tion within each triangle, vh will-be permitted to be a quadratic: at the rnidpoint and two at the ends).
by these three nodal values, shared b
edge. The nodal values elsewhere in
In order to lie in X 1, vh must again be continuous across the edge between along the edge, and continuity holds. 1
adjacent triangles. .rh in the Dirichlet problem, the three
To compute with this subspace, a different description is required; we quadratic) are zero.
need a basis for the space. Therefore, we must find a set of continuous piece- Any vh in Sh has a unique expam
wise quad!atic functions rp1 such that every member of S has a unique expan- q1 is the value of vh at the jth interiOI
sionas these rp1 forma basis for Sh, and the d
strained nodes.
For continuous piecewise cubics,
There is a beautiful construction of such a basis, if we add to the vertices way. A cubic in the two variables x,
a further set of node~, placed at the rnidpoints of the edges (as in the first tri- in particular by its values at the lC
angle of Fig. 1.11 ). With each node, whether it is a vertex or the midpoint of Again the 4 nodes on each edge dete
an edge, we associate the function f{Jj which equals 1 at that node and zero and continuity is assured. This is the
 CHAP.
matrix are negative, they are domi.;.
ts a consequence the inverse- marrix
n shows that this proper~y holds for
· no angle exceeds n/2. (The precise
~d by any given edge add to no more
r elements in· n dimensions. Because
nite element approxift;¡ation uh obeys
::ement u: if the load f is everywhere
ment. Fried has asked whether this
it will not be true that all off-diago-
;sary for a positive K- 1 •
o constraint on vh at the boundary
s the total number of interior and
:1gain eqpal one at a single node and
r, defining the trial functions to be
continuous. Instead, we maysimply
near function defined in the adjacent
eh we mention one-to ignore the
oblem only within the approximate
olution may be significantly altered,
,arison with minimizing the integral
ifying the boundary-value problem,
i more refined element. This was a
1ique, to generalize the basic idea
Rather than assuming a linear func-
Iitted to be a quadratic:
r4 X 2 +a xy +a y 2•
5 6
ontinuous across the edge between
ifferent description is required; we
must find a set of continuous piece-
y member of S has a unique expan-
lPix,y).
t a basis, if we add to the vertices
oints of the edges (as in the first tri-
her it is a vertex or the mid point of
ich equals 1 at that node and zero
1 SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 79
C0 quadratic C0 cubic
z3 cubic
Fig. 1.11 Nodal placement for quadratic and cubic elements.
at all others. This rule specifies rp at six points in each triangle, the three
vertices and three midpoints, thereby determining the six coefficients a1 in
(59).
The crucial point to decide is whether or not the piecewise quadratic,
determined by this rule in each separate, triangle, is continuous across the
edges.petween triangles. The proofis deceptively simple. Along such an edge
vh is a quadratic in one variable, and there are three nodes on the edge (one
at the midpoint andtwo at the ends). Therefore, the quadratic is determined
by these three nodal values, shared by the two triangles which meet at that
edge: Th'e nodal values elsewhere in the two triangles have no effect on vh
along the edge, and continuity holds. For an edge lying on the o uter boundary
rh in the Dirichlet problem, the three nodal values (and therefore the whole
quadratic) are zero.
Any vh in Sh has a unique expansion into I; qlp1, where the coordinate
q1 is the .value of vh at the jth interior node (vertex or midpoint). Therefore,
these rp1 forma basis for S\ and the dimension N equals the number of uncon-
strained nodes.
For continuous piecewise cubics, a basis can be constructed in the same
way. A cubic in the two variables x, y is determined by 10 coefficients, and
in particular by its values at the 10-node triangle of the previous figure.
Again the 4 nodes. on e~ch edge determine the cubic uniquely on that edge,
and continuity is assured. This is the triangular analogue of the one-dimen-
80 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.9. TRIAN

sional. cubic element constructed in Section 1.6 (that is, the one which was There is an additional point of
only continuous, and had two nodes inside each interval). elements: The unknowns Qc corres
The same construction extends to polynomials of any degree k - l eliminated immediately from the fir
in any number of variables x 1 , • • • , xn, provided the basic regions are always system with effectively three unknow1
simplexes-intervals for n 1, triangles for n = 2, and tetrahedra for n of unknowns is known as static cond
3. In fa1t, we can obtain discrete analogues of arbitrarily high accuracy in corresponding basis functions fPc ar
n-space./'Unfortunately, there is for practical applications a fatal flaw; the each Qc is coupled only to the other 1
dimension of Sh, which equals the total number of interior nodes, grows the cth equation can be solved for {
enormously fast with k and n. This is an essential problem with finite elements, Qc can be eJiminated from the sys
to impose further constraints on the trial functions (thereby reducing the The number of arithmetical operati o
dimension of Sh) without destroying either the approximation properties of centroids which are removed, whi
or the simplicity of a local basis. sional problems, it is impossible to el
The key is to increase proper1y the level of continuity required. In the Physically, the optimal displacemen1
cubic case, we may add the _condition that the first derivatives of vh should sively by the nine parameters whic1
be continuous at every vertex. This is clearly a subspace of the previous cubic boundary of th~, element.
case, in which only continuity of vh was required. If in triangles meet at a Mathemati~'i'ny, this, process m~
given interior vertex, the continuity of v~ and v; imp.Óses new constraints on of the basis functions rp1 with resp
the trial functions, and the dimension N is correspondingly reduced. To con- Of course, the Gram-Schmidt proce:
struct a basis we remove the midedge nodes and concentrate a triple node at the whole basis, leaving the trivial s~
each vertex. In other words, the 10 coeffi.cients in the cubic are determined insane; it is quicker just to solve KQ
by v, v.:c and V y at each vertex, together with v at the centroid; this tenth node tion ag_ainst fPc is possible since it che
has nowhere else to go. A unique cubic is determined by these 10 nodal values, within the given triangle. It appear:
and along each edge it is determined by a subset of 4-the values of v and its knowns in other than the normal •
edge derivative at each end, a combination which again assures matching of a band matrix, except in sp~ia
along the edge. The result is an extremely useful cubic trial space, which we reduction, which systematically elimi
denote by z3. Laplace difference equation. , .
Piecewise polynomial elements-,like Z 3 aré easy to describe in the interior There is an important variánt of
of the domain; at the boundary we ha ve to be more careful, and there are condensation is avoided: the centroi
various alternatives. In case of an essential condition u = O, there is first of the nine nodal parameters given by 1
all the constraint vh = O. If vh is to vanish along all of rh, then also the ingly, one degree of freedom must t
derivatives along both chords must be set to zero, leaving no parameters and a familiar device is to require eq
free at that vertex. A more satisfactory approximation is obtained by the 1

isoparametric method of Section 3.3, or-if it is preferred to stay with the vh = a1 +a x +a y +a x +a

2 3 4
2
5J

x-y variables-to impose only the condition that tl)e derivative of v vanish
in the direction tangent to r. In the latter case we must give up the Dirichlet
condition that vh vanish on rh and by extension on the true boundary r. This restriction damages ~he accura
vh
Instead, the functions will actual/y l'iolate the essential condition for admissi- rate of convergence. to the degree .e
bility, and in this version of the cubic Z 3 , the tri al space will not be a subspace reproduced in the trial space, and
of the Dirichlet space JCA. Nevertheless, it is possible to give a rigorous esti- Nevertheless, this restricted cubic i
ma te of the error induced by such inadmissible elements; this wilJ be a key and is preferred by many engineers
result ofChapter 4. We may expect vh to be nearly zero oil t, and if we go so excellent. (However, we observe in Se
far as to compute o ver the curved triangles of O rather than the real triangles of freedom are not invariant With n
of 0", the numerical results will be good. for sorne orientations they are not
 CHAP.
:tion 1.6 (that is, the one which was
1side each intervaf).
polynomials of any degree k
;->rovided the basic regions are always
~s for n = 2, and tetrahedra for n =
>gues of arbitrarily high accuracy in
a\
lctical applications fatal flaw; the
al number of interio'r, nodes, grows
!Ssential problem with finite elements,
rial functions (thereby reducing the
~ither the approximation properties
level of continuity required. In the
that the firf-L~derivatives of vh should
uly a subspace of the previous cubic
.s required. If m triangles meet at a
·~ and v~ imposes new constraints on
is correspondingly reduced. To. con-
,des and concentrate a triple node at
fficients in the cubic are determined
ith v at the centro id; this tenth node
determined by these 10 nodal values,
1 subset of 4-the values of v and its
tion which again assures matching
ly useful cubic trial space, which we
•3 are easy tó describe in the interior
e to be more careful, and there are
:ial condition u = O, there is first of
nish along aH of r\ then also the
set to zero, Ieaving no parameters
approximation is obtained by the
-if it is preferred to stay with the
tion that the derivative of v vanish
r case we must give up the Dirichlet
~xtension on the true boundary r.
te the essential condition for admissi-
:he trial space will not be a subspace
it is possible to give a rigorous esti-
tissible elements;this will be a key
:>e nearly zero on r, and if we go so
:S of a rather than the real triangles
1 SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 81
There is an additional point of practical importance about the cubic
elements: The unknowns Qe corresponding to the centro id nodes can be
elirhinated immediately from the finite element system KQ F, Jeaving a
system with effectively three unknowns per mesh point. This early elimination
of unknowns is known as static condensation. It depends on the fact that the
corresponding basis functions 'Pe are nonzero only inside a single triangle;
each Qe is coupled only to the other nine parameters ofits own triangle. Thus
the cth equation can be sol ved for Qe in terms of these nine nearby Qi, and
Qe can be,. eliminated from the system without increasing its bandwidth.
The number of arithmetical operations is directly proportional to the number
of centroids which are removed, which is exceptional; in general two-dimen-
sional problems, it is impossible to eliminate n nodes with only a,n operations.
Physically, the optimal displacement Qc at the centroid is determined exclu-
sively by the nine parameters which establish the displacement along the
boundary of th~ element.
Mathemati~~'lly, this process may be regarded as the orthogonalization .
of the basis functions 'Pi with respect to each centroid basis function 'Pe·
Of course, the Gram-Schmidt process could always be used to orthogonalize
the whole basis, leaving the trivial stiffness matrix K /, but this wou1d be
insane; it is quicker just to solve KQ Fdirectly. The special orthogonaliza-
tion ag_ainst 'Pe is possible since it changes only the nine adjacent rpi' and only
within the given triangle. It appears that in general the elimination of un-
knowns in other than the normal Gaussian order will increase the width
of a band matrix, except in special circumstances-such as the odd-even
reduction, which systematically eliminates every other node in the five-point
Laplace difference equation. ' .
There is an important variant of the cubic space Z 3 , in which this static
condensation is avoided: the centro id is not a node at all, and there are only
the nine nodal parameters given by v, vx, and vy at each vertex. Correspond-
ingly, one degree of freedom must be removed from the cubic polynomials,
and a familiar device is to require equal coefficients in the x 2 y and xy 2 terms:
vh = a1 +a x +a y +a
2 3 4 x2 +a 5 xy +a y
6
2
+ a1x + a (x y + xy + a
3
8
2 2
) 9 y3•
This restriction damages the accuracy of the element; we shall connect the
rate of convergence. to the degree of polynomial which can be completely
reproduced in the trial space, and this degree has dropped from J to 2.
Nevertheless, this restricted cubic is among the most important elements,
and is preferred by many engineers to the full Z 3 • Its nodal parameters are
excellent. (However, we observe in Section 4.2 that the remaining nine degrees
of freedom are not invaríant with respect to rotation in the x-y plane, and
for sorne orientations they are not even uniquely determined -by the nine
 82 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.9. TRIANGl

nodal parameters. Anderheggen has proposed JJ vh = O as an alternati ve

Full quintic
constráint, and Zienkiewicz [22, p. 187] analyzes a third, and very attractive,
possibility.)
None of the spaces described so far can be applied to the biharmonic
equation, since they are not subspaces of 3C 2 • (Or rather, their use would be
illegal; the restricted Z 3 is frequently used for shells. Elements which are
nonconf<>t~:ming along interior edges are discussed in 4.2.) Therefore, denoting
by ek the. . class of functions with continuous-derivatives through order k,
we want now to construct elements which belong to e 1 • The essential new
condition is that the normal slope be continuous across interelement bound-
aries. A function in e 1 is automatically in 3C 2 and therefore admissible for
fourth-order problems; the integrals over n can be computed an element at
a time, with no ó-functions at the boundaries.
,--/Severa! possibilities exist. One is to modify the cubic polynomials in
Z 3 by rational functions, chosen to cancel the edge discontinuities in the nor-
mal slopes without affecting the nodal quantities v, vx, and v>'. This makes
vh piecewise rational instead ofpiecewise polynomial, and again the accuracy
is reduced; the space does not contain añ arbitrary cubic. We are opposed
to rational functions, however, on other and more important grounds; they
Fig. 1.12 Two quintic triang
are hard to integtate, and even numerical integration has very serious diffi-
culties (Section 4.3).
lf we prefer to work exclusively with polynomials, we are led to one of An equally effective way of dispo!
the most interesting and ingenious of all finite elements-the quintics is to require that vn reduce toa cubica
(Fig. 1·.12). A polynomial of fifth degree in x and y has 21 coefficients to be leading coefficient in what would othe
determined, Of Which 18 Will COme from the Va}UeS Of V, Vx, Vy, Vxx' VxY' and to vanish. With this constraint the n<
_ v>'>' at the vertices. These second derivatives . represent "bending moments" cubic vn along the edge, and the eleme
of physical interest, which will now be continuous at the vertices and avail- reducing the displacement error 'in u
able as output from the finite element process; they are given directly by however, the dimension of Sh is sigr
the appropriate weights Qr Furthermore, with all second derivatives present see Table 1.1) a;nd this more than co
there is no difficulty in allowing an irregular triangulation. The quintic along tetms. In fact, a series of careful nun
the edge between two triangles is the same from both sides, since three condi- prize to this remarkable element. Adr
tions are shared at each'end ofthe edge-v and its edge derivatives vs and vss' with-the continuity of second deriv:
which are computable from the set of six parameters at the vertex. domain, where the true solution u is le~
lt remains to determine three more constraints in such a way that also Note that to achieve a e 1 triangul:
the normal derivative vn is continuous between triangles. One technique is
to add the value of vn at the midpoint of each edge to the list of nodal para-
tThere are important problems in whic
meters. Then since vn is a quartic in s along the edge, it will be uniquely fixed incompressible fluids, for example, it is usefu
by this parameter together with the values of vn and vns at each end. {The The French school (Crouzeix, Fortin, Glowi
normal n at the middle of an edge points into one triangle and out of the the Courant triangles are completely inac
adjacent one, adding somewhat to the bookkeeping.) This constrüction pro- the use of
duced the complete e 1 quintic, which was devised independently in at least t. standard quadratics in the plane and
four papers. lts accuracy in the displacement will be O(h 6 ), if the boundary
e
2. these 1 quintics as stream functio
velocity field of quartic polynomials constr
is successfully approximated; comparable accuracy with finite differences 3. a non-conforming element, linearon
has apparently never been achieved for fourth-order problems in the plane. of the edges (cf. Section 4.2).
 CHAP. 1 SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 83

S proposed f f Vh 0 as an alternative
87] analyzes a thi~d, and very attr"'ctive, Full quintic Reduced quintic

> far can be applied to the biharmonic

Vn is o
es of JC 2 • (Or rather, their use would be . /cubic on
ly used for she11s. Elements which are each edge
.re discussed in 4.2.) ~herefore, denoting
ontinuous derivatives.'through order k,
which belong to G1 • The essential new
continuous across inter:element bound-
Llly in JC 2 and therefore admissible for
overO can be computed an element at
undaries.
to modify the cubic polynomials in
tcel the edge discontinuities in the nor-
d quantities v, vx, and vy. This makes
ise polynomial, and again the accuracy
n an arbitrary cubic. We are opposed
ter and more important grougds; they
;ical integration has very serious diffi- Fig. 1.12 Two quintic triangles and a cubic macro-triangle.

ith polynomials, we are led to one of

An equally effective way of disposing of the three remaining constraints
of all finite elements-the quintics
is to require that v, reduce toa cubic along each edge, i? other wo~ds, that the
ee in x and y has 21 coefficients to be
Ieading coefficient in what would otherwise be a quartic polynomtal b~ made
m •the valués of v, vX' vY' vXX' vXY' and
. to v~mish. With this constraint the nodal values of v, and v,s determm~ the
•attves represent "bending moments"
cubic. v, along the edge, and the element is in e 1 • The complete quin tic is ~ost,
continuous at the vertices and avai1-
reducing the displacement error in u- uh· from h6 to hs. At the same ttme,
: process; they are given directly by
however the dimension of Sh is significantly reduced (by a factor of 9: 6;
see Tabl~ l.l) and this more than compensates for the loss ofthree quintic
re, with all second derivatives present
~ular triangulation. The quin tic along
tetms. In fact, a series of careful numerical experiments [CI4] has given first
ne frorn both sides, since three condi-
prize to this remarkable element. Admittedly, it is not always simple to.work
-v and its edge derivatives v and v
1( pararneters at the vertex. s w
with-the continuity of second derivatives must be relaxed at corners m the
domain where the true solution u is less continuous-but it is highly accurate. t
constraints in such a way that also
between triangles. One technique is
Not~ that to achieve a e 1 triangular element, it was necessary to prescribe
f each edge to the Iist.of nodal para-
mg the edge, it will be uniquely fixed tThere are important problems in which the physics introduces side co~ditions .. For
ues of v, and v,.s at each end. (The incompressible fluids, for example, it is useful to impose div vh = O on all the trtal funct10ns.
lts into one triangle and out of the The French school (Crouzeix, Fortín, Glowinski, Raviart, Témam) has found th~t althou~h
the Courant triangles are completely inadequate, convergence can be estabhshed wtth
)Okkeeping.) This construction pro- the use of
as devised independently in at Jeast 1. standard quadratics in the plane and cubics in 3-space;
ment will be O(h 6 ), if the boundary 2. these e t quintics as stream fuT).ction in two dimensions, with a divergence-free
>le accuracy with finite differences velocity field of quartic polynomials cgnstructed by differentiating the qui!ltics;
:Ourth-order problems in the plane. 3. a non-conforming element, linea'r on triangles but continuous only at the _midpoints
of the edges (cf. Section 4.2).
 \
84 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.9. TRIANG

even th~ second derivatives at the vertices. Zenisek has found a remarkable Table 1.1 TRIAI

theorem [Zl] in this direction: To achieve piecewise polynomials of class eq

Element Type Contim
on an arbitrary triangulation, the nodal parameters must include al/ derivatives
at the vertices of order less than or equal to 2q. He has constructed such ele- .Linear eo
ments in n dimensions by using polynomials of degree 2nq + 1, conjectured Quadratic eo
to be the rp.inimum possible. Cubic eo
Fortun~tely there is a way to avoid this severe limitation. and still achieve Cubic Z3 e 0 , vx, vy contin
at vertices
a conformi.ng e 1 element. It consists of building up a "macroelement" out As above, plus <
Restricted Z 3
of several standard elements. One of the best known is the Clough-Tocher coefficients of
triangle, forrned by combining different cubic polynomials in three subtri- Quintic e 1, Vxx, Vxy, Vyy
angles (Fig. 1.12). The final nodal parameters for the large triangle are v, vx, continuous at
and vy at the vertices and vn at the midpoints of the edges-12 in all. These Reduced quintic e 1, Vxx, Vxy, Vyy
continuous at
guarantee that even the normal slope will match that of the adjoining macro- e1
Cubic macrotriangle
triangle, since the slope is a quadratic and is· therefore determined along
an edge exclusively by the parameters on that edge-vn is given directly at
the midpoint, and indirectly at the two vertices as a combination of V and 2n2 triangles, each of the small squ
vy. Since each of the three cubics has 10 degrees of freedom, and thex final which has .slope +l. Only the leadi
macrotriangle has 12 nodal parameters, 18 constraints may be imposed. It be correction terms depending on th
turns out. that this is-just right to achieve e 1 continuity intermilly in the tri- but the important constant is M. T
angle. Requiring a common value of v, vX , and vy furnishes 3 constraints at element in which each vertex is a p-·
.
each externa} vertex and 6 at the interna! one, and agreement of vn at the the interior. of each triangle contaim
midpoints of the three interna! edges makes 18. angulation there are twice as many tri
Assuming that a given triangulation allows the triangles to be combined for 180° in ~ngles and each vertex for
three at a time into macrotriangles, we have just found a basis for the space triangles is 3 : 2, since each edge is sh
Sh of all piecewise cubics ofdass e 1 • This is one case of an important and 1-2-3 weighting of p, r, and q. (Seconj
.apparently very difficult problem-to determine. a basis for the space of piece- inside an éxisting triangle, it bring
wise polynomials v(xp . .. , xJ of degree k - 1 which have continuity eq of two triangles.) This 1-2-3 rule hol
between simplexes. We have no idea (for n > 1) even of the dimension of that on average there are M/2 unkno
this space: According to Table 1.1, the e 1 cubics have M= 6 parameters nodes are desirable; an increase inp
for each pair of macrotriangles, and therefore an average of one unknown for dimension and to bandwidth. Inten
each of the original triangles !t they can be removed by static co1
To summarize the polynomials described in this section, we shall tabulate worst in respect of computation tim
their essential properties. The column headed d gives the number of para- Let us remark that in any theon
meters required to determine the polynomial within each region, that is, the the one of higher degree k _;_ 1 will
null?-ber of degrees of freedom if adjacent elements imposed no constraints. cally as h ~ O. The error decreases 1
The integef k - 1 is the degree of the highest "complete polynomial" which both ~n the particular element .an
is present; this means that sorne polynomial of degree k cannot be found in fore, the only theoreticallimitation ~
the trial space, and consequently (as we shall prove) that the error in u- uh ness of u, and even this can be ave
is of order hk. Finally, N is the dimension ofthe trial space S\ assuming that The balance is altered, however
Q is a square which has been partitioned into n 2 small squares and then into asks for the element which achieves
width h will be finite, not infinitesi
whether the.calculation is to be do
tAdded in proof: We now have a conjecture about the dimension, but no inkling of a
basis, for the space of polynomials of degree k - 1 and continuity eq.
in programming is a necessity, or w
 CHAP. 1 SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 85

ces. Zenisek has found a remarkable Table 1.1 TRIANGULAR ELEMENTS

eve piecewise polyhomials of class eq
arameters must include al! derivatives Element Type Continuity d k N Mn2
l to 2q. He has constructed such ele-
mials of degree 2nq 1, conjectured Linear eo 3 2 n2
Quadratic eo 6 3 4n 2
Cubic eo lO 4 9n2
his severe limitationJt.nd stil1 achieve Cubic Z3 eo, vx, vy continous
:- building up a "macroelement" out at vertices 10 4 5n2
1e best known is the Clough-Tocher Restricted Z 3 As above, plus equal
coefficients of x2y and xy 2 9 3 3n2
cubic polynomials in three subtri-
Quintic e t, 'Vxx, 'Vxy, 'Vyy 21 6 9n 2
te~ers for the large triangle are v, vx, continuous at vertices
1omts of the edges- 12 in aH. These Reduced quintic et, 'Vxx, 'Vxy, 'Vyy
l match that ofthe adjoining macro- continuous at vertices 18 5 6n2
an~ is therefore ~etermined along Cubic macrotriangle et 12 4 M=6
m that edge-vn,lis/ given directly at
vertices as a combination of v and 2n 2 triangles, each of the small squares being cut in half by the diagonal
) degrees of freedom, and thex final which has slope + 1. On1y the leading term N Mn 2 is given; there will
18 constraints · may be imposed. It be correction terms depénding on the constraints imposed at the boundary,
e e 1 continuity internally in th/tri- but the important constant is M. The coefficient Mis p 3q + 2r for an
'x• and vY furnishes 3 constraints at element in which each vertex is a p-tuple node, each edge has q nodes, and
al one, and agreement of v at the the interior.of each triangle contains r nodes. The reason is that in any tri-
kes 18. n
angulation there are twice as many triang1es as vertices; each triangle accounts
t11ows the triangles to be combined for 180° in angles and each vertex for 360°. Furthermore, the ratio of edges· to
ave just found a basis for the space triangles 'is 3: 2, since each edge is shared by two triangles. This explains 'the
is is one case of an important and 1-2-3 weighting of p, r, and q. (Second prqof: lf one new vertex is introduced
rmine a basis for the space of piece- insid.e an existing triangle, it brings three new edges and a net increase
e k ~·1 which have continuity eª oftwo triangles.) This 1-2-3 rule holds on any set oftriangles, and it means
r n > 1) even of the dimension of that on average there are M/2 unknowhs per triangle. lt follows that multiple
e 1 cubics have M= 6 parameters nodes are desirable; an increase in p is economical in its contribution both to
ore an average of one unknown for dimension and to bandwidth. Interna! nodes are next best, especially since
they can be removed by static condensation. Nodes along edges are the
Jed in this section, we sha11 tabulate· worst in respect of computation time.
aded d gives the number of para- Let us remark that in any theoretical comparison of two finite elements,
lial within each region, that is, the the one of higher degree k - 1 will always be the more efficient, asymptoti-
elements imposed no constraints. cally as h O. The érror decreases like Chk, with a constant C which depends
hest "complete polynomial" which both Ón the particular element and on the kth derivatives of u. There-
ial of degree k cannot be found in fore, the only theoreticallimitation on the rate of convergence is the smooth-
talJ prove) that the error in u- uh ness of u, and even this can be avoided (Section 3.2) by grading the mesh.
)fthe trial space Sh, assuming that The balance is altered, however, if one fixes the accuracy required and
nto n 2 small squares and then into asks for the element which achieves that accuracy at least expense. The mesh .
width h will be finite, not infinitesimal. Furthermore, there is the question
bout the dimension, but no inkling of a whether the calculation is to be done only a few times, so that convenience
1 and continuity e11. in programming is a necessity, or whether the programming arrd preparation
86 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.9. TRIAN<

costs qm be amortized over a long program lifetime. It is our judgment seem slightly superior to linear ones,
that elc~ments of limited order, like those in Table 1.1-together with others term xy. Nevertheless, a square "m~
constructed to have s¡milar characteristics, such as quadrilateral elements, triangles and have the same element l
three-dimensional elements, and shell elements-will provide the basic range -A, the asseinbled bilinear stiffness
of selection for the practica} application of the finite element method. main diagonal and to _:_ 1 at points
We w~nt now to discuss rectangular element~, which are rapidly increasing the plane. The extension to a trililí
in popularf~ty. They are of special value in three dimensions, where each cube a 5 xy + a 6 xz + a 7 yz + a 8 xyz in 3-sp
occupies the same volume as six very elusive tetrahedra. (An irregular de- at each corner.
composition of 3-space into tetrahedra is hard even on the computer.) Just as the bilinear element on n
Furthermore, a number of important problems in the plane are defined on triangles, there are biquadratics anc
a rectangle or on a union of rectangles. The boundary of a more general quadratics and cubics on triangles.
region cannot be described sufficiently well without using triangles, but it degree k - 1 in each of the variab1
will often be possible to combine rectangular elements in the interior with freedom within each "box" element
triangular,elements near the boundary. . are easy to describe. Figure 1.13 shc
Thésimplest consiruction, in analogy with the linear element for triangles,
is based on functions which are piecewise bilinear: vh == a 1 + a 2 x + a 3 y +
a 4 xy in each rectangle. Th~se four coefficients are determined by the values
of vh at the vertices. -:fhe overall trial fuitction is continuous and may be ap-
plied to differential equations of second order. There is an obvious inter-
polating basis: rp1 equals one at the jth node and zero at the others. Its graph
•
V

looks like a pagoda, or at least like our conception of one, and rp1 will be
called a pagoda function. It is a product 'lf(X)'If(Y) of the basic piecewise linear
roof functions in one variable, so that the space Sh is a tensor product of two V
-
V
simpler spaces. This is a valuable element.
It is important to notice that for arbitrary quadrilaterals, these piecewise Fig. 1.13 Nodes for biq
-bilinear functions. would not be continuous from one element to the next. Sup-
pose that two quadrilaterals are joined by a line y = mx + h. Then along 1
coefficients of a biquadratic are dete
that edge, the bilinear function reduces toa quadratic; it is linear only if the nodes on each edge assure continuit
edge is horizontal or vertical. A quadratic cannot be determined from the ratic can be merged with the quadr
two nodal values at the endpoints ofthe edge, and in fact the other nodes do vertices and midpoints are nodes. ~
affect the value of vh. Therefore, bilinear elements may be used only on rec- A simple modification of the bi
tangles. Given a more general quadrilateral, however, it is still possible to reducing the number of nodes in e~
change coordinates in such a way that it becomes a rectangle, and then bilin- biquadratic is restricted by the rem4
ear functions in the new coordinates are admissible. In fact, this coordinate to the approximation of u was very 1
change can itself be described by a bilinear function, so that the same po/yno- element belongs to the "serendipi1
mials are úsed for the transformation of coordinares as for the shape functions particularly useful for quadrilater~
within each element. This is the simplest of the isoparametric elements, which change transforms the region into
are discussed in detail in Section 3.3. becomes, along with the bilinear (
The bilinear element on rectangles can be merged with Courant's linear element in the plane.
elements on triangles, since both are completely determined by the value of The corresponding three-dime
vh at the nodes. Other combinations are also possible: A bilinear function which again is especially valuable fe
on a triangle, with a midpoint node along one edge, is sometimes merged with or multiple nodes. Its nodes líe at
a full quadratic on a neighboring element. In general, bilinear trial functions the twelve edges of the brick~ This
 CHAP. 1
,rogram lifetime. It is our judgment
se in Table l.l-t6gether with others
stics, such as quadrilateral elements,
lements-will provide the basic range
n of the finite element method.
dement~, which are rapidly increasing
in three dimensions, tWhere each cube
elusive tetrahedra. (An irregular de-
ra is hard even on the computer.)
•roblems in the planeare defined on
:s. The boundary of a more general
well without using triangles, but it
.ngular elements in the interior with
{ with the linear element for triangles,
ise bilinear: vh = a 1 + a 2 x + a 3 y +
ficients are determined by the values
nction is continuous ancÍ may be ap-
d order. There is an obvious inter-
ode and zero at the others. Its graph
r conception of one, and rp j will be
r(x)Vf(y) ofthe basic piecewise linear
e space Sh is a tensor product of two
lt.
~itrary quadrilaterals, these piecewise
Yfrom one element to the next. Sup-
by a Iine y= mx + b. Then along
:o a quadratic; it is linear only if the
ttic cannot be determined from the
edge, and in fact the other nodes do
· elements may be used only on rec-
:eral, however, it is still possible to
becomes a rectangle, and then bilin-
admissible. In fact, this coordinate
u function, so that the same polyno-
1ordinates as for the shape functions
,f the isoparametric elements, which
~n be merged with Courant's linear
1pletely determined by the value of
also possible: A bilinear function
one edge, is sometimes merged with
. In general, bilinear trial functions
SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 87
seem slightly superior to linear ones, since they exactly reproduce the twist
term xy. Nevertheless, a square "macroelement" can be formed from two
triangles and have the same element stiffness matrix. For the Laplacian L
-A, the asseinbled bilinear stiffness matrix K is proportional to 8 along the
main diagonal and to __:: 1 at points correspondingto its eight neighbors in
the plane. The extension to a trilinear function a 1 + a 2 x + a 3 y + a4 z +
a 5 xy + a 6 xz + a 7 yz + a 8 xyz in 3-space is clear; again there is one unknown
at each corner.
Just as the bilinear element on rectangles matches the linear element on
triangles, ·thei-e are biquadratics and bicubics which correspond to the eo
quadratics and cubics on triangles. (Uitimately there are polynomials of
degree k- 1 in each of the variables x 1 , • • • , xn, possessjng kn degrees of
freedom within each "box" element.) The biquadratics are often used and
· are easy-to describe. Figure 1.13 shows the arrangement of nodes. The nine
•
V
V
-
V
Fig. 1.13 Nodes for biquadratic and Hermite bicubic.
coefflcients of a biquadratic are determined by its nodal values, and the three
nodes on each edge assure continuity between rectangles. Clearly the biquad-
ratic can be merged with the quadratic element on triangles, since the same
vertices and rhidpoints are nodes. Similar remarks apply to the bicubics.
A simple modification of the biquadratic is to rt!move the interior node,
reducing the .number of nodes in each element to eight. To compensate, the
biquadratic is restricted by the remo val of the x 2y 2 term, whose contribution
to the approximation of u was very marginal anyway. This simple and pleasant
element belongs to the "serendipity class" described by Zienkiewicz. It is
particularly useful for quadrilaterals with curved sides, after a coordinate ~
change transforms the region into a square (Section 3.3); in this context it
becomes, along with the bilinear element, the most valuable isoparametric
element in the plane.
The corresponding three-dimensional element is the 20-point brick,
which again is especially valuable for isoparametrics because it has no interna!
or multiple nodes. Its nodes Iie at the eight vertices and at the midpoints of
the twelve edges of the brick. This total matches the number of terms x•yPzr
88 SEC. 1.9. TRIAN(
AN INTRODUCTION TO THE THEORY CHÁP. 1

which 3;re of second degree in at most one ofthe variables-x 2yz is allowed, We emphasize that these nodes must ll
but not x 2y 2 z or x 3 • in one direction and y = Yz in the ot
The higher-order serendipity elements in the plane have 4p nodes equally intersections absolutely determine
spaced along the 'perimeter of the rectangle, including four at the corners. Therefore, the bicubic element is res
The shape functions include all terms xrr.yP in which neither exponent exceeds . linear transformation of the plane, t
p, and the ,smaller exponent is O or l. The first missing term is therefore x 2y 2 , On a su:fficiently rectangular regio
a~d k = 4.~ These are again especially useful in coordinate changes, where elements. Its degree of continuity fo
the boundaries of the rectangle can be transformed into arbitrary polynomial in (61); since r¡1 and ro are both in e 1 ~
curves of degree p, and there are no annoying interna! nodes. Therefore, this bicubic may be appl
Thereare·several fascinating elements (dueto Clough, Felippa, and others) functions lie in 3C 2 • Furthermore, e
in which a quadrilateral is produced as the union of two or more triangular all continuous. (This suggests a ba~
pieces, so that the polynomial varíes from piece to piece within the quadri- space Sh; it consists of all piecewisc;
lateral macroelernent. The nodal arrangement and continuity between pieces are continuous; we say that v is in <
can become rath~r subtle. However, we shall describe only one other element, extra degree of continuity does not s~
the Hermite bicubic. The trial functions are again cubic in each variable sepa- The function vxy is quadratic along
rately, vh = .L; aiixiyi for O< i,j < 3, yielding 16 degrees of freedom in rectangles which share that edge, ye
each rectangle .. The parameters are determined by the values of v ' v x' vY' points are automatically held in com
and vxyat the four vertices.t Thus the dimension of Sh ~s grea,tly reduced a quadratié!
in comparison with the ordinary bicubics described above, which are based The Hermite idea can be extend
on 16 distinct nodes. In one dimension there would be q
This element should be understood as a natural extension of the funda- and ro of the cubic case; all their de
mental ene-dimensional element, the Hermite cubic, which was discussed at the nodes, except that the pth fu
in Section l. 7. In that case v = a 1 a2 x +a3 x 2+ +
a4 x 3 was determined o ver origin. This yields the most natural b
each subinterval by the values of v and vx at the endpoints. This construction 2q - 1 having q - 1 continuous de
left vx continuous at the nodes, and therefore everywhere, so that the element products cop(x)cop.(y), which means 1
was in e 1 • The standard basis in one dimension contained two kinds of many ofthem are cross derivatives of
functions, r¡1(x) and ro(x), which interpolated function values and derivatives, too inefficient after bicubics or biquiJ
respectively: can be reduced by imposing additio
tion is achieved by the splines, whi~
1 at node x = xi, becoming a single polynomial ove
"'j = {o at other nodes,
co 1 = O at all nodes,
unknown for each node, with basis J
(60) line and by tp(x)tp(y) in the plane. TI
d• •
11
,_.,_J = 0 at a ll no d es, dco 1 {1 at node x = x 1, different elements become coupled
dx dx = O at other nodes. tions become more complicated, aJ
away from rectangles) are impossit
These functions span all piecewise cubics of class e 1 •
Because all these spaces are SJ
The Hermite bicubic space is a product of two such cubic spaces, and the
basis functions, the stiffness matric<
four parameters at a typical node z = (xi, y 1) lead to four corresponding
one-dimensional operators. Rough
basis functions:
of variables can be applied to the e
importance occurs in parabolic p1
(61)
W¡ = 'fliX)r¡lz{y), w2 = 'flix)co¡{y),
to implicit difference schemes, witl
w3 = ro/x)r¡~¡{y), · C0 4 = colx)co¡{y).
matrix, or Gram matrix, formed 1
tThis element was introduced in the engineering Iiterature by Bogner, Fox, and Schmit. inverted at each time step. Por difl
 CHAP. 1
one of the variables-x 2yz is allowed,
Hs in the plane have 4p nodes equally
angle, including four at the corners.
'yP in which neither exponent exceeds
1e first missing term is therefore xzyz,
useful in coordinat~;¡,_changes,, where
·ansformed into arbitrl;lry poly~omial
oying interna! nodes.
; (dueto Clough, Felippa, and others)
the union of two or more triangular
)ffi piece to piece within the quadri-
~ment and continuity between pieces
hall describe only one other element
ue again cubic in each vafiable sepa~
yielding 16 degrees of freedom in
:ermined by the values of v v v
, ' X' Y'
· dtmension of Sh is gre~tly reduced
:s described above, which are based
lS a natural extension of the funda-
ermite cubic, which was discussed
+ a 3 X 2 + a4 x 3 was determined o ver
at the endpoints. This construction
r-ore everywhere, so that the element
limension conütined two kinds of
ted function values and derivatives
'
w 1 = O at all nodes,
1 at node x x1
{ O at other no des.'
of class e 1 •
t of two such cubic spaces, and the
'!C1 , Yr) Iead to four corresponding
2 = Vfix)wr(Y),
wix)wr(y).
g literature by Bogner, Fox, and Schmit.
'SEC. 1.9. TRIANGULAR AND RECTANGULAR ELEMENTS 89
.we emphasize that these nodes must lie on a rectangular grid; the lines x = x 1
in one direction and y y 1 in the other can be spaced arbitrarily, but their
intersections · absolutely determine the two-dimensional array of nodes.
Therefore, the bicubic element is restricted to rectangles (or, after a simple
.linear transformation of the plane, to parallelograins).
On a sufficiently rectangular region the Hermite bicubic is one of the best
elements. Jts degree of continuity follows immediately from the basis given
in (61); since t¡1 and ro are both in e 1 , all their products inherit this property.
Therefore, this bicubic may be applied to fourth-order equations; the trial
functions lie in JC 2 • Furthermore, even the cross derivatives iJ 2v/iJx iJy are
all continu9us. (This suggests a basis-free char.acterization of the Hermite
space Sh; it consists of all piecewise bicubics v such that v, vx, vy, and vxy
are continuous; we say that vis in e 1• 1 .) The remarkable thing is that this
extra degree of continuity does not seem to follow from the usual arguments.
The function vxy is quadratic along each edge and is the same for the two
rectangles which share that edge, yet only the values of vxy at the two end-.
points are automatically held in common-and two values cannot determine
a quadratic!
The Hermite idea can be extended to elements of higher degree 2q 1.
In one dimension there would be q different functions, corresponding to t¡1
and ro of the cubic case; all their derivatives of order less than q will vanish
at the nodes, except that the pth function rop(x) has (iJ/iJx)P-trop = 1 at the
origin. This yields the most natural basis for the space of polynomials of degree
2q - 1 having q - 1 continuous derivatives. In two dimensions we need all
products rop(x)rop.(y), which means that there areq 2 unknowns at every node;
many ofthem are cross derivatives of high' order, and the construction beco mes
too inefficient after bicubics or biquintics. As always, the number ofunknowns
can be reduced by imposing additional continuity. The ultimate in this direc-
tion is achieved by the splines, which are as continuous as possible without
becoming a single polynomial over the whole domain. There is only one
unknown for each node, with basis function given by the B-spline VJ(x) on the
line and by VJ(X)VJ(Y) in the plane. The difficulty, of course, is that nodes from
different elements beco me coupled in the construction; the boundary condi-
tions' become more complicated; and isoparametric transformations (to get
away from rectangles) are impossible.
Because all these spaces are spanned by products of one-dimensional
basis functions, the stiffness matrices K may also admit a decomposition into
one-dimensional operators. Roughly speaking, this occurs when separation
of variables can be applied to the differential operator L. A case of practical
importance occurs in parabolic problems, where Galerkin's method Ieads
to implicit difference schemes, with a two-dimensional matrix M-the mass
matrix, or Gram matrix, formed from the inner products of the VJ1-to be
inverted at each time step. For difference equations it was this kind of diffi-
90 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.10. ELEMENT MAT

culty which gave rise to the alternating direction methods, in which the tic in the coordinates qi' describing
~nversion is approximated by two one-dimensional computations. For any
space formed from the products of one-dimensional elements, the same (62) a(v, v
technique can be applied, with the usual qualification that if the domain
is not precisely rectangular, then the success of the alternating direction The entries of K are the energy innc
In practice these inner products
method is observed but not proved~
computation is the evaluation of the E
In lt4ble 1.2, M = p + 2q + r for an element with p parameters at each
in other words, over the separate ~
Table 1.2 RECTANGULAR ELEMENTS Each such fraction of the overall er

Element type Continuity d k N Mn2 (63)

Bilinear eo 4 2 n2 in analogy with (62). The vector q

Biquadratic eo 9 3 4n 2 which actually contribute to the em
Restricted biqt~dratic e o 8 3 3n2
the full set of coordinates qi coult
Ordinary bicu , ic e o 16 4 9n2
Hermite bicubic et,t 16 4 4n 2 host of zeros into the element stiffm
Splines of degree k ek-z,k-z k2 k n2 those functions tp1 which vanish in
Hermite~ degree k- 1 2q 1 eq-t;q-t k2 k q2n2 the order of k e equal to the number
Serendipity, p > 2 eo 4p 4 (2p- l)n2 function within e.) For linear fun
for example, the matrix ke will be e
vertex, q along each edge, and r inside each rectangle. The weights 1, 2, 1 three components of qe will be the
come from the combinatorics of an arbitrary quadrilateral subdivision: The ratics d 6, and for the quintic e
number of vertices equals the number of quadrilaterals equals half the
number of edges. There are n subintervals in each coordinate direction.

T,
Here, for example, v; is the x-deri
it equals the weight qi attached t•
1.10. ELEMENT MATRICES IN
at this vertex equals one and wl
TWO-DIMENSIONAL PROBLEMS
In the following we shall use th1
In this section we shall outline the sequence of operations performed by basic example, since they illustrat•
the computer in assembling the stiffness matrix K, that is, in setting up the There are now two problems,
discrete finite element system KQ F. The calculation of the consistent and second to assemble them intc
load vector F will also be briefly described. Our iíitention is not to give the a(v, v) =
intiníate details which a programmer would require, but rather to clarify the
crucial point in the success of the finite element method: polynomial elements
like those ofthe previous section are astonishingly convenient in the practica! The latter is a questioq of efficien
application of the Ritz method. Perhaps un'iquely so. relative size of the computer and ·
We recall first the source ofthe matrix K. The functional /(v) to be mini- handled in core, one possible m<
mized has as leading term a quadratic expression a(v, v), representing in in which the elements (~ubregior
many cases the strain energy (or strictly speaking, twice the strain energy). The matrices for each element in
The Ritz method restricts v toa subspace S by introducing as trial functions element · containing sorne particu
a finite-dimensional family v = ~ q1rpr (The superscript h will be dropped tThe nine entries of the element sti
in this section, while the Ritz method is applied toa fixed choice of elements.) from the dot products of the si des of th
The substitution ofthe trial functions v into the strain energy yields a quadra- Every off-diagonal k¡¡ :S O unless the t1
 CHAP.
zting direction methods, in which .the
;e-dimensional computations. Por any
one-dimensional elements, the same
sual qualification that if the1 domain
~ success of the alternating direction
an element with p p~rameters at each
'lGULAR ELEMENTS
d k N Mn2
4 2 n2
9 in analogy with (62). The vector q, now contains only those parameters q1
3 4n2
8 3 3n2
16 4 9n2
16 4 4n2
~-2
k2 k nZ
r-l k2
k q2n2
4p 4
e. each rectangle. The weights 1, 2, 1
11trary quadrilateral subdivision: The
:r of quadrilaterals equals half the
·als in each coordinate direction.
equen~e of operations performed by
; matnx K, that is, in setting up the
: The calculation of the consistent
•ed. Our intention is not to give the
uld require, but rather to clarify the
ement method: polynomial elements
nishingly convenient in the practical
uniquely so. _ relative size of the computer and the problem. ·For problems too large to be
x K. The functional/(v) to be mini-
expression a(v, v), representing in
speaking, twice the strain energy).
:S by introducing as trial functions
(T~e superscript h will be dropped
>phed toa fixed choice of elements.)
to the strain energy yields a quadra-
l SEC. 1.10. ELEMENT MATRICES IN TWO-DIMENSIONAL PROBLEMS 91
tic in the coordinates qi' describing the energy on the subspa~ S:
(62) a(v, v) qTKq.
The entries of K are the energy inner products KN, a(rpk, rp1).
In practice these inner products are calculated only indirectly. The basic
computati6n is the evaluation ofthe energy integral a(v, v) over each element e,
in other words, over the separate subdomains into which O is partitioned.
Each such fraction of the overall energy takes the form
(63)
which actually contribute to the energy in the particular region e. (Of course,
the full ser of coordinates q1 could be included, at the cost of inserting a
host of zeros into the element stiffness matrix k,. It is much neater to suppress
those functions rp1 which vanish in the particular region e, and to maintain
the order of k, equal to·the number of degrees offreedom d ofthe polynomial
function _within e.) For linear functions v = a 1 + a 2 x + a 3 y on triangles,
for example, the matrix k e will be of order d 3. t With the usual basis, the
three components of q, will be the values of v at the vertices of e. For· quad-
ratics d 6, and for the quintic element with 18 degrees of freedom,
Here, for example, v; is the x-derivative at the second vertex of the triangle;
it eq~lls the weight q1 attached to the 'basis function rp1 whose x-derivative
at tnis vertex equals one and whose other nodal parameters equal zero.
In the following we shall use these quintic functions on triangles as the
basic example, since they illustrate most of the difficulties that can arise.
There are now two problems, first to compute the element matrices k,,
and second to assemble them into the overall strain energy
a(v, v) = qTKq "_E q~keqe.
e
The latter is a question of efficient bookkeeping and depends in part on the
handled in core, one possible mode of organization is the frontal method,
in which the elements (~ubregions) are ordered rather than the unknowns.
The matrices for each element in turn are assembled, and as soon as every
element · containing sorne particular unknown Qn has been accounted for,
tThe nine entries of the element stiffness matrix for Laplace's equation come directly
from the dot products ofthe sides ofthe triangle: kiJ = s1 •s¡/2A, A= area ofthe element.
Every off-diagonal k¡¡ ~ O unless the' triangle is obtuse.
92 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.10. ELEMENT MATRI

eliminati9n is ..carried 6ut in the corresponding row of K and the results are For the 18-degree element, the las1
stored. In this way, the only unknowns iQ core ata given moment are those geneous constraints on the coefficient
belonging to several elements, sorne of which have already been assembled slope vn is a cubic along each edge. T
and sorne of which ha ve not. \
In this section we shall concentrate on the computation of the element
matrices, e~tending to two dimensions the technique used in Section 1.7 for
the Hermitftcubic calculations. At the end we comment on the use of numer-
ical integratfon.
The essential point is that with a polynomial basis rpi and polygonal
(not curved) elements, the energy is a weighted sum of integrals of the form
Inverting, the matrix H which conne1
columns:
\'- ' prs = JI xrys dxdy. A

These integrals will depend on the location of the element e and on constants For any element, the jirst step is toco
which can be tabulated. Therefore, the problem is to find a convenient coordi- the nodal parameters q and the polym
nate system in which to describe. the geometry of e, and to relate. the nodal Now we calculate the energy inb
parameters qe to the coefficients in the polynomial v. There is general agree- the plate-bending example in Sectior
ment that the "global" x-y coordinates are inappropriate, but there seems to
be much less agreement on which local system is the best. We shall therefore
describe two or three of the possibilities. ·
aeCv, v) JJ(vix + Vfy +
The first is to translate the coordinates so that the origin lies at the cen- e
troid (x 0 , y 0 ) of the triangle e. Bell [8] refers to this as the local-global system.
Substituting the polynomial v, this i
Since the transformation is linear, a quintic polynomial in x, y remains
a quintic in the new variables X= x X 0 , Y =~y- y 0 :
(66)
v = a1 +a 2X +a 3Y + a4X +L.+ a
2 5
21 Y •
where the matrix N requires the ev:
8
It is easy to find the nodal parameters of q~ in terms of these coefficients are computed, the calculation of thf
a1 : If the new coordinates of the vertices are (X1, Y 1),. then with A= Hqe,

V
1
= a 1 + a 2 X 1 + a 3 Y 1 + a 4 Xi + · · · + a 21 Yi,
(64) v! = v_l. a2 + 2a 4 X 1 + ... , so that finally

(67)
and so on. Eor the 21-degree element the three midedge slopes ha ve al so to be
Thus the element stiffness matrix i
computed; the nodal parameters for this element areq; 1 (q; 8 , v~, v;, v;).
connection matrix H between the
Their relationship with the coefficient vector A = (a 1 ~ ••• , a 21 ) can be written
coefficients, and a matrix N giving t
in matrix formas A different set of local coordina1
(65) Lindberg, and Olson [C14]. In tb
b, e, and ()in Fig. 1.14 become p.aJ
whe,rethe first 18 rows of G are given by (64); the last three rows involve nót the coordinates (x¡, y 1) of the vertlc
only the coordinates X 1, Yi of the vertices but also the orientation of the· polynomials in the rotated coor~in
triangle. lnverting (65), we have A G- 1 1 Hq; 1 • q; of a new matrix N, giving the m1
 CHAP~ 1 1
ponding row of K and the results are
, in core at a giveri moment are those
_. which have already been ass~mbled
~ on the computation of the element
the technique used in Section 1.7 for
nd we comment on th1~ use of numer-
polynomial basis rpj and polygonal
·eighted su m of integrals of the form
·ys dxdy.
on of the element e and on constants
roblem is to find a convenient coordi-
ometry of e, and to relate the nodal
tolynomial v. There is general agree-
tre inappropriate, but there seems to r.,
ystem is the best. We shall therefore ,
:es so that the origin lies at the cen-
ers to this as the local-global system.
¡uintic polynomial in x, y remains
Xo, Y= Y- Yo:
)f q} 8 in terms of these coefficients
: are (X¡, Y;), then
.. '
hree midedge slopes ha ve also to be
elementareq 2 t =(qta vt vz v3)
)r A = (a~' .. ~, a 21 ) ;a~ b~ wnritt~~
4,
64); the last three rows involve not
:s but also the orientation of the
q¡t = Hq;t.
SEC. 1.10. ELEMENT MATRICES IN TWO-DIMENSIONAL PROBLEMS 93
For the 18-degree element, the last three rows of G are changed to homo-
geneous constraints on the coefficients a¡, in order to ensure that the normal
slope vn is a cubic along each edge. This alters the relationship to
qJ8)
(~
= G0 A.
Inverting, the matrix H which connects qe toA this time has 21 rows and 18
columns,:
For any element, the jirst step is to compute this connection matrix H between
the nodal parameters q and the polynomial coefficients a.
Now we calculate the energy integral in terms of the coefficients a;. For
the plate-bending example in Section 1.8, this integral is
ae(v, v) = JJ (vix + vJ;y + 2vvxxVyy + 2(1 - v)viy) dXdY.
e
Substituting the polynomial v, this is easily put in the form
(66)
where the matrix N requires the evaluation of the integrals P,s· Once these
are computed, the calculation of the stiffness matrix is complete. From (66),
with A= Hqe,
so that finally
(67)
Thus the element stiffness matrix involves two separate computations, the
connection matrix H between the nodal parameters and the polynomial
coefficients, and a matrix N giving the integrals of polynomials.
A different set of local coordinates was recommended by Cowper, Kosko,
Lindberg, and Olson [C14]. In their system the geometrical quantities a,
b, e, and O in Fig. 1.14 become paramount; they are easily computed from
the coordina tes (xi' y¡) of the vertices. The trial functions will still be qui~tic
e
polynomials in the rotated coordinates and '1· Therefore, the computation
of a new matrix N, giving the integrals of polynomials over the triangle,
 1 •, ELEMENT MAl
94 AN INTRODUCTION TO THE THEORY CHAP. 1 SEC. 1.10.

y _ length of PQ
t1 - length of AQ =
~ + C + ~ = orea of (BPC +
1 2 3 · orea of (
y

X
Fig. 1.14 Another choice of local coordinates.

depends on the existence of a convenient formula:

Fig. 1.1~ Area c•

II .er,¡saea11 = cs+t(ar+l (-by+ t)

(r + s''+ 2)!
r .s. . Geometrically, C1 is the fraction A
fraction ztfh 1 of the distance to th
e cording to (68), the integrals Pr, be
We also need a new connection matrix H, and this is computed in two steps.
First, entirely in the e-11 variables, H' connects the nodal parameters to the
polynomial coefficients. Then there is a rotation matrix R to link the local
and the global coordina tes, and H = H' R. The·details are clearly explained
and we need only the formulas for
in [C14l and leave the impressiori that a cubic variation of the normal slope
and BeU [8], p. 84):
is simplest to impose in this system.
The computation of Prs = JJ xr ys dX dY can be carried out without
any rotations by introducing area coordinates. These are the most natural
parameters and are also known to engineers as triangular coordinates and to
II CiC1C~ = (ñ

mathematicians as barycentric coordinates. In this system each point has

This takes info ·account the Jacob
actually three coordinates, whose sum is always ( 1 + ( 2 + (3 = l. The ver-
n p =O, the coefficient 2! appe
tices of e become the points (1, O, 0), (0, 1, 0), and (0, O, 1). If the same ver-
nant ofthe matrix B in (68) is reco!
tices havé rectangular coordinates (Xp Y1), (X2 , Y 2 ), and (X3 , Y 3 ), then by
element.
]inearity the coordinates (X, Y) and ((p ( 2 , ( 3) of an arbitrary point Pare
Holand and Bell [8] have give
related by
keeping the origin of the X- Y sy!
xl remarkable that for r + s < 6, t1
(68) yt
( 1
 CHAP. 1 SEC. 1.10. ELEMENT MATRICES IN TWO-DIMENSIONAL PROBLEMS . 95

~ = length of PQ = orea of BPC

1 length of AQ orea of BAC
~ + ~ + ~ = orea of (BPC + CPA + APB) =
1 2 3 orea of (BAC)
y

X 8=(0,1,0)
;e of local coordinates.

nt formula: X
Fig. 1.1~ Area coordinates for a triangle.
(-by+t) r!s! . Geometrically, ' 1 is the fraction A 1 /A of the area in Fig. 1.15 and also the
(r + s + 2)! fraction z 1/h 1 of the distance to the opposite side. In area coordinates, ac-
cording to (68), the integrals Prs become ·
H, and this is computed in two steps.
connects the nodal parameters to the
t rotation matrix R to link the local
rt R. The ·details are clearly explained

. cubic variation of the normal slope
r dX dY can be carried out without
dinates. These are the· most natural
teers as triangular coordinates and to
rtes. In this system each point has
; always ' 1 + ' 2 + ' 3 = l. The ver-
, l, 0), and (0, O, l ). If the same ver-
Y1), (X2, Y2 ), and (~, Y3), then by
, '2' '3) of an arbitrary point P are
and we need only the formulas for integration in area coordinates (Holand
and Bell [8], p. 84):
m!n!p!
(m+ n + p + 2) 1detB.
This takes into account the Jacobian of the coordinate change; with m =
n = p O, the coefficient 2! appears in the denominator, and the determi-
nant ofthe matrix B in (68) is recognized as twice the area A ofthe triangular
element.
Holand and Bell [8] ha ve given explicit formulas for the integrals Prs'
keeping the origin of the X- Y system at the centroid of the triangle. lt is
remarkable that for r +· s < 6, these integrals are all of the simple form
 1.10. ELEMENT MATRI<
96 AN INTRODU:C'I10N TO THE THEORY CHAP.l SEC.

The c<;>nstants are c 1 O, c 2 = 1/12, c 3 c4 = 1/30, c 5 2/105. From over the elements; if there are variab
these formulas Bell d~rives the stiffness matrix k" for the complete quintic taken into account, then numerical in
applied to the plate-bending pr()blem. Finally, there is the calculation oft
We mention that another local coordinate system could be based entirely geneous.~erm in the finite element eqt
on the natural coordinates C1• In this case the parameters q", which involve element ~t a time:
nodal der~vatives like vl.)l, must be converted into derivatives with respect
to the Ct· I~is our impression that algebraic manipulations in this system take (70)
a consideráble amount of time (and practice!), although Argyris and his
coworkers did succeed with this system in forming H = G- 1 analytically.
This natural coordin_ate system really comes into its own with numerical Again the calculation starts with A=
quadrature. of the element to the coefficients a1
Next we turn to the computation of the mass matrix M. This is the matrix coefficients,
which arises from an undifferentiated term JJ v 2 dx dy in the energy. In
other words, writing V. I; qlPi' M' is the Gram matrixt formed from th~ (71)
inner pro4ucts of the basis functions rp i:
where the components of the d-dimer
Mik = (rpk, rpi) JJ ffJk'Pi dxdy.
a

This matrix plays a central part in eigenvalue calculations ..

Just as with K, the mass matrix is assembled from the. separate elements:
Matching (71) with (70), the require<
(69) JJv2 = qTMq I; q~meqe = ~ JJv2.
a

Two special cases in the calcula

The element mas~ 111atrix, by ~our previous computations, can be written as
The first is that of a uniform l9ad J
integrals appear again:

Here H is the same matrix connecting the parameters qe to the coefficients of

the polynomial ·
d
V L a¡Xm;yn;. The second is that of a point load
i=l
(X0 , Y0). If this point is outside e, t
Z comes directly from the inner products of these monomials:
u,=

Again these calculations can be do1

For a more general loading f,
by a quadrature formula over each
These numbers are tabulated, and any ofthe local coordinate systems should
element which líes in the subspace ~
be acceptable. Thus Z plays the role for mass matrices which N played for
parameters fr'-th~ values off and i
stiffness matrices. In both cases we have tacitly assumed constant coefficients
the interpolating element
tWe have tried to think of a pun on the coincidence of mass matrix and Gram mafrix,
but the editor says it will have to wait. (72)
y
CHAP. 1
/12, c3 C 4 = 1/30, c 5 = 2/105. From
'ness matrix k e for the complete ,quintic
\1.
)Ordinate system could be based entirely
s case the parameters qe, which involve
converted into derivatives with respect
ebraic manipulationsc;jn this svstem take
:l practice !), although, Argyrls and his
~tem in forming H G- 1 analytica11y.
Iy comes into its own with\ numerical
)f the mass matrix M. This is the 1inatrix
~~ t<?rm JJ V 2 dx dy in the energy. In
1s the Gram matrixt formed from the
rp j:
= II 'Pk'Pj dxdy.
o
senvalue calculations.
1ssembled from the separa te elements:
wus computations, can be written as
~he parameters qe to the coefficients of
:ts of these monomials:
:- the local coordina te systems should
r mass matrices which N played for
tacitly assumed constant coefficients
tcidence of mass matrix and Gram matrix,
SEC. 1.10. ELEMENT MATRICES IN TWO·DIMENSIONAL.PROBLEMS 97
o ver the elements; if there are variable material properties which must be
taken into account, then numerical integration is recommended.
Finally, there is the calculation ofthe load vector F, whichis the inhomo-
geneous term in the finite element equations K Q = F. This, too, is done an
element at a time:
(70) ff f
o
V qT F ~ q~ Fe ~ ff f V.
Again the calculation starts with A = Hqe, connecting the nodal parameters
of the element to the coefficients a1 in the polynomial. In terms of these
coefficients,
(71)
where the components of the d-dimensional vector a are ju_st
Matching (71) with (70), the required element load vector is
Two special cases in the calculation of the a 1 are pointed out in [C14].
The first is that of a uniform load f(x, y) / 0 , in which case the tabulated
integrals appear again:
The second is that of a point load f(x, y) = f0 ~, concentrated at the point
(X0 , Y 0 ). If this point is outside e, then of course Fe = O; if it is inside, then
Again these calculations can be done in any local system.
For a more general Ioading J, the integrals a1 may be computed either
by a quadrature formula over each element, or else by interpolating f by an
element which Iies inthe subspace S 11 • In the latter case we compute the nodal
parametersfr-~th~ values offand its derivatives at the nodes-andconstruct
the interpolating element
(72)
98 AN IN'{RODUCTION TO THE THEORY CHAP. 1 SEC. 1.10. ELEMENT MATRICE

The change from f to / 1 gives a perturbed load vector F with components In other words, uh may exceed u in som
smaller derivatives in the mean.:.square
mates of both displacement and slope al
can be changed either by a fundamente:
stress, mixed, and hybrid methods are
Mis the mass matrix described above and f' is the vector formed from the
intentionally committing numerical erre
nodal para¡:p.eters
'·.
f J.. This is a comparatively easy computation. The latter effect is the one produce(
So far we have considered only analytical computation of the element
ically, this effect appears as a decrease
matrices, based on exact integration formulas for polynomials over polygons.
approximate K satisfies
In more and more cases, the element matrix computations are being carried out
approximately, by someform of Gaussian quadrature. For curved elements, (74)
arising from shell problems or from curved boundaries in planar problems,
this numerical quadrature is virtually a necessity. In this case there will be It may seem paradoxical that this leads
a perturbation not only from F to F but also from K to K. This means that the Ritz solution, but the paradox is e
the corresponding Q is the finite element solution for a perturbed problem. · Q = K.- 1F. Its strain energy·is QrK{
In a later chapter we shall estimate the error Q - Q; it must depend on the energywas QTKQ FTK- 1 F. Sincethe
accuracy of the quadrature formula. is equivalent (after sorne argument) 1
We regret to report that these inexact numerical integrations have even inverses,
been shown in sorne cases to improve the quality of the solution. This is one
instance (nonconforming elements are another) in which computational
experiments yield results which are frustrating to the mathematical analyst and it follows that the strain energy is 1
but nevertheless numerically valid and important. The improvement for In case of a point load P at the jth
finite h is due partly to the following effect: The strict Ritz procedure always case to a similar conclusion about the é
corresponds to an approximation which is too stiff, and the quadrature error ponent of the load vector F is Fi = p,
reduces this excess stiffness. is Q. (K- 1F)i ft(K- 1)ir If K is ~
J -
Stiffness is an intrinsic property of the Ritz method. By limiting the dis- 1
increased to ft(K- )ii.
placements v to a finite number of m o des rp 1 , •• ·• , rpN' instead of allowing all It remains to see why Gaussian qu:
admissible functions, the numerical structure is more constrained than the effect of reducing K. Obviously K will
real one. In eigenvalue problems, the result of this constraint is that ít? always k is reduced. A rigorous proof of the
lies above the true Ar In static problems, the potential energy l(uh) exceeds Razzaque [16] in the one-dimensional ca
l(u), since uh is obtained by minimization over the subspace spanned by in terms of Legendre polynomials. The
(/Jp .•• , 'PN· This overestimate of 1 corresponds to an underestimate of the
strain energy a, as proved in the corollary to the fundamental theorem 1.1 : qTkeq = f ((vh)')z f (~o +
(73)
2(~5 + ~} + ... +
In the special case of a point load ó(x0 ), where the displacement u(x 0 )
is proportional to the strain energy, this displacement is al so underestimated The effect of n-point Gaussian quadratu
by the Ritz method; the overstiff numerical structure has less "give" under ~;_ 1 , since they come from polynomials
a point load than the true' structure. For\ distributed loads, the tendency is late the term in~;, since Pn vanishes a1
the same: The finite element displacement uh(x) is generally below the true are defined.) Therefore in this special
displacement u(x). This is not a rigorous theorem, because the Ritz method and the same tendency reappears in me
minimizes en-ergy, and its link to displacement is not strictly monotone~ The least sti.f!ness matrix criterion i:
 CHAP. 1 SEC. 1.10. ELEMENT MATRICES IN 1WO-DIMENSIONAL PROBLEMS 99

:urbed load vector F with components
~ and f' is the vector formed from the
atively easy computation.
malytical computatidn of the element
•rmulas for polynomütls over polygons.
ltrix computations are being carried out
sian quadrature. For curved ~lements
:urved boundaries in planar prbqlems:
a necessity. In this case there wlll
be
mt also from K to K. This means that
ent solution. for a perturbed problem.
~error Q Q; it must depend on the
mct numerical integrations have even
he quality of the solution. This is one
·e another) in which computational
ustrating to the mathematical analyst
td important. The improvement for
fect: The strict Ritz procedure always
h is too stijf, and the quadrature error
the Ritz method. By Iimiting the dis-
~es 'PP ... , rpN, instead of allowing all
ucture is more constrained than the
mlt ofthis constraint is that A.1 always
as, the potential energy l(uh) exceeds
Ltion over the subspace spanned by
Tesponds to an underestimate of the
ary to the fundamental theorem 1.1 :
In other words, uh may exceed u in sorne part of the structure and _still ha~e
smaller derivatives in the mean-square sense. Nevertheless, one-stded esti-
mates of both displacement and slope are common with finite elements. This
can be ·changed either by a fundamental alteration of the Ritz process (the
stress mixed, and hybrid methods are described in the next chapter) or by
inten{ionally committing numerical errors in the right direction.
The latter effect is the one produced by Gaussian quadrature. Algebra-
ically, this effect appears as a decrease in the positive definiteness of K; the
approximate K satisfies
(74) for all q.
It may seem paradoxical that this Ieads to an increase in the strain ene~gy ~s
the Ritz solution, but the paradox is easily explained. The new solut10n Is
Q K.-tF. Its strain energy is QTKQ PK- 1F, whereas t~e old stra~n
energy was QTKQ = prK- 1F. Sin ce the inequality (74! on. the stt~ness matn.x
is equivalent (after sorne argument) to the opposlte mequahty on thetr
in verses,
for all F,
and it follows that the strain energy is increased.
In case of a point load p at the jth node, this leads as in the continuous
caseto a similar conclusion about the displacement. The only nonzero com-
ponent of the load vector F is Fi p, and the displa~eme~:lt ~t the jth no~e
is Q. = (K-tF)i p(K- 1 )ir If K is decreased to K, th1s dtsplacement IS
J - '
increased to P(K- 1) ir . .
It:remains to see why Gaussian quadrature tends to have thts desuable
effect of reducing K. Obviously K will be reduced if every element matrix
k is reduced. A rigorous proof of the latter effect was found by Irons and
Razzaque [16] in the one-dimensional case, by imagining that (vh)' is. expanded
in terms of Legendre polynomials. The strain energy over [-1, 1] ts
qTkeq J ((vh)') 2 J (rt 0 + rt P (x) + · · · + rtnPn(x))
1 1
2

a(u, u).

= J(x 0 ), where the displacement u(x0 )

= 2(a;~ + ~t + . . . + 2na¡ 1) .
displacement is also underestimated
The effect of n-point Gaussian quadrature is to preserve all the terms thro~g?
rical structure has less "give" under
a; 2 since they come from polynomials of degree less than 2n, and to anmht-
>r'-distributed loads, the tendency is
~nt uh(x) is generally below the true
l;t~1~he termina;, since Pn vanishes at the Gauss points. (This is how they
are defined.) Therefore in this special case the integral is clearly reduced,
s theorem, because the Ritz method
and the same tendency reappears in more general problems.
lacement is not strictly monotone.
The least stiffness matrix criterion is a valuable basis for the comparison
100 AN INTRODUCTION TO THE THEORY CHAP. 1

of different elements. Mathematically, it must reflect the approximation

qualiÜes of these elements, and in particular the numerical constants which
enter approximation of polynomials one degree higher than those repro-
duced exactly by the element. Computationally, it is a clearly visible effect,
and a number of theoretically useful elements have been discarded as too
stiff. The criterion raises the possibility of optimizing the stiffness matrix
under t]¡}:~ constraint of reproducing polynomials of degree k - l.

2 A SUMMARY OF 1

2.1. B.ASIS FUNCTIONS FOR THE FINI

\

In this chapter we collect in one pla~

ing the finite element method. The gm
for the analysis, in which each of the
Then succeeding chapters will take up
detail.
The first step is to decide which sub~
at the examples already given in Chaptt
which are mathematically essential.
One general description which fits t
we call the nodal finite element method. l
analysis. Every trial function vh is detet
are the unknowns qi of the discrete pro
is the value, ata given node z i' of either
tives. Thus the unknowns are

where the differential operator Di is

parameter is just the function value, an
az¡ax ay, and so on.
To each of these nodal parameteJ
determined by the following property •
all other nodal parameters of rp i are
node z J. to be shared by severa! paran
11
JRY CHAP. 1

:ally, it must reflect the approximation

particular the nunierical constants which
ls one degree higher than those repro-
putationally, it is a clearly visible· effect,
fui elements .have been discarded as· too
>ility of optimizing the stiffness inatrix
; polynomials of degtye k 1.

2 A SUMMARY OF THE THEORY

2.1. BASIS FUNCTIONS FOR THE FINITE ELEMENT SUBSPACES Sh

In this chapter we collect in one place sorne of the main results of analyz-
ing the finite element method. The goal is to describe a general framework
for the analysis, in which each of the various error estimates has its place.
Then succeeding chapters will take up these estimates one by one, in more
detail.
The first step is to decide which subspa~es Sh to study. Therefore, we look
at the ..examples aiready given in Chapter 1 and try to draw out the properties
which ·are mathematically essential.
One general description which fits these examples is the following, which
we call the nodalfinite element method. It will be the foundation for our whole
analysis. Every trial function vh is determined by its nodal parameters, which
are the unknowns q1 of the discrete problem. Each of these nodal parameters
is the value, at a given no de z 1, of either the function i tself or one of its deriva-
tives. Thus the unknowns are

where the differential operator D 1 is of order zero (D1vh = vh) in case the
pararneter is just the function value, and otherwise it rnay be a¡ax, a¡ay, a¡an,
a2 jax ay, and so on.
To each of these nodal pararneters q1 we associate a trial function rp1,
deterrnined by the following property: At z1 the value of D1rp1 is 1, whereas
all other nodal pararneters of rp1 are zero. Note that it is possible for the
node z 1 to be shared by severa! parameters, in other words to be a rnultiple

101
102 A SUMMARY OF THE THEORY CHAP. 2 SEC. 2.1. BASIS FUNCTIONS FOR TI

node;it is not z 1 but the pair (z1, D 1) which uniquely defines the parameter to use in the analysis. This property e
ql' Thus the key property is that each rp1 satisfies Dlp/z1) 1 but gives zero formed in a geometrically regular pattet
when matched with a different pair (zn D,): of as covered by a regular mesh of wié
positions in each mesh square. The pro¡
(1) lation invariance: If a basis function rp is
node z is translated to Iie at the new pe
These fu,JJ,.ctions rp1 form an interpolating basis for the trial space, since every
tion rp* associated with (z*, D) is just th
trial funbiion can be expanded as
(2)
vh = ~qJ'Pr
In n dimensions, l = (/ 1, ••• , In} is a vec
By applying D1 to bJth sides and evaluating at z1, the parameters q1 are ex-
actly what they shoul~ be:
Clearly Drp~(z +lh) = Drp(z) = 1; the i
to give a function rp* interpolating at
lation will haye to break down at the
conditions happen to be periodic; in th
The whole object of the Ritz method is to find the optimum values Q1 for has a completely periodic pattem.
these parameters, by minimizing I(vh). Then the particular tria! function Let us take as an example the spa(
which possesses these optimal parameters is the finite element approximation functions which was deséribed in Sect
triangular mesh, as indicated in Fig. 2.1

Let us fit two or three examples into this general framework. For the basic
piecewise linear functions on triangles (Turner triangles, or Courant trian-
gles) the nodes z 1 are the vertices in the triangulation, and all derivatives D 1
are of order zero, D 1vh vh. The unknowns are q1 = vh(z¡), and the basis is
formed from the pyramid functions determined by rpiz1) ~w The same is
true for bilinear functions on quadrilatera1s and for quadratics on triangles,
this time with midedge nodes included in the Zr For the Hermite cubics in one
dimension, the derivatives do enter: Every node z1 appears in both pairs Fig. 2.1 Nodes for CO q
(z 1, I) and (zp dfdx). We distinguished the two kinds of basis functions
rp1 by t¡J1 and ror For Hermite bicubics there are four parameters per node, to unity. Note that we can associate .VI
corresponding to v, vx, V y, and vxY' that is, to D = [, a¡ax, a¡ay, and iJ 2/ilx ay. four nodes; the four associated with tl
For the cubic space Z 3 on triangles, the vertex nodes are triple and the mesh point (1, O) are crossed. Note a
centroids are only simple. gives the number of basis functions fP.
To complete the description of the nodal finite element method, we say square, differs from the number d of d
something about the geometry. The domain Q is partitioned into a union of angle. M is the coefficient in the last
closed subdomains, ot elements, overlapping only at the interelement bound- giving the dimension of the space Sh. O
aries. Each element e contains no more than d of the nodal points z1 , and al! dimension is M/h 2 •
the basis functions rp1 , except those which correspond to these d nodes, are zero We shall denote by W1 , W2 , W3 , and
throughout e. Thus rp1 is also a local basis. This framework seems to be suffi- late at the nodes z 1 to z 4 , respective!~
ciently general to include most of the finite element spaces in current use. vanish at every no de except their own;
We note, however, that cubic splines are not included, since their basis func- basis functio:hs at the four nodes assoi
tions (the B-splines) are nonzero over severa! elements. found just by translation as describe
We call attention to one additional property of the rp1 which can be put W1(x- lp y- 12 ), i = 1, 2, 3, 4. lf the
 CHAP~
•) which uniquely defines the parameter
h rpi satisfi.es Di'P¡(zi) = 1 but gives zero
(z¡, D 1):
!¡) Oo·
~ting basis for the tri~~ spibe, since every
aluating at z 1, the parameters q1 are ex-
d is to find the optimum values Qi for
v"). Then the particular trial function
~ters is the finite element approximation
to this general framework. For the basic
es (Turner triangles, or Courant trian-
he triangulation, and all deriva ti ves D.J
(nowns are qi = v"(z ), and the basis is
:etermined by rp¡(z1) Ow The same is
aterals and for quadratics on triangles,
. in the Zr For the Hermite cubics in one
Every node zi appears in both pairs
.ed the two kinds of basis functions
:;s there are four parameters per node,
.t is, to D /, a¡ax, a¡ay, and ai¡ax ay.
, the vertex nades are triple and the
e nodal finite element method, we say
omain n is partitioned into a union of
1pping only at the interelement bound-
e than d of the nodal points zi, and al!
-:h correspond to these d nodes, are zero
asis. This framework seems to be suffi-
: finite element spaces in current use.
.re not included, since their basis func-
several elements.
Ll property of the rpi which can be put
2 SEC. 2.1. BASIS FUNCTIONS FOR THE FINITE ELEMENT SUBSPACES Sh 103
to use in the analysis. This property emerges only when the elements are
formed in a geometrically regular pattern, where the domain can be thought
of as covered by a regular mesh of width h, and the nodes fall in the same
positions in each mesh square. The property of the basis is then one of trans-
lation invariance: If a basis function rp is associated with a pair (z, D), and this
node z is translated to lie at the new point z* = z + lh, then the basis func-
tion rp* associated with (z*, D) is just the translate of rp:
(2) rp*(x) = rp(x - lh).
In n dimensions, l (l P . . • , !J is a vector ofintegers, that is, a multiinteger.
Clearly Drp~(z + lh) = Drp(z) = 1; the interpolating function rp is just shifted
to give a function rp* interpolating at z*. Óf course, this pattern of trans-
lation will have to break down at the boundary of n unless the boundary
conditions happen to be periodic; in that case we can imagine that the basis
has a complet~ly periodic pattern.
Let us take as an example the space of continuous piecewise quadratic
functions which was described in Section 1.7. The nades fall on a regular
triangular mesh, as indicated in Fig. 2.1, where the mesh width is normalized
Fig. 2.1 Nodes for C0 quadratics on a square mesh .
to unity. Note that we can associate with each mesh point a unique set of
four no des; the four associated with the origin are circled and those for the
mesh point (1, O) are crossed. Note also that this number M 4, which
gives the number of basis functions 'Pi and unknown weights qi per mesh-
square, differs from the number d of degrees of freedom; d 6 in each tri-
angle. M is the coefficient in the last column of the table in Section 1.9,
giving the dimension of the space Sh. On a unit square, with periodicity, this
dimension is Mjh 2 •
We shall denote by <ll 1, <ll 2 , <ll 3 , and (1) 4 the basis functions whichinterpo-
late at the nades z 1 to z 4 , respectively. They are piecewise quadratic, and
vanish at every node except their own; this determines them completely. The
basis functions at the four nodes associated with the mesh point (1 1, 12 ) are
found just by translation as described abo ve; these shifted functions are
<Plx- lp y - / 2 ), i 1, 2, 3, 4. if the meshwidth is reduced to h, this just
104 A SUMMARY OF THE THEORY CHAP.2 SEC. 2.2.

rescales the independent variables; the four basis functions associated with over each subinterval by v and v' at t
the mesh point (/ 1h, / 2 h) are CfJ;(xjh 11 , yjh- / 2). parameters for each meshpoint. Their ger
This pattern of tra.nslation is so useful and important that we shall make t¡l and m of Fig. 1.8.
it the foundation for a second general description of the finite element Experiments have been carried out wi
method. This-.description applies on a regular mesh, and we call it the the reduction in M is still greater, but tl
abstract finite e1ement method. In n dimensions it begins with the choice of regular if there is to be any compariso
M functi:~ms w·1(x), ... , CfJM(x). These will eventually lead to M unknowns method. The bandwidth of the stiffness 11
per mesh cube, and the finite element equations KQ F will take the form decreased by a reduction in M; for cubic
of M coupled finite difference equations. seven nonzero entries in each row of K, :
To form the basis functions associated with one particular grid point, the four intervals. This is virtually the same
one at the origin, we simply rescale the variable x = (x¡, ... , xn), obtain- where the usual ordering of unknowns gi
ing W1(x/h), ... , CfJM(xjh). To form the basis functions associated with a
different grid point lh (/~' ... , !Jh, we translate the functions just con-

(~ ~) (~ ~)
structed. Thus to denote all the basis functions constructed in this way, by
scaling and translation, we need two subscripts i and /:
K=
l, ... ,M.

We must add the requirement that the original functions CfJ;(x) shall vanish
outside sorne sphere 1 x 1 <R. Then rptz will vanish outside the local sphere The order of the matrix K is of course
1 x - lh 1 < Rh, and we shall again have a local (but possibly not a strictly and splines do appear to be more efficien
local) basis. domain O is conveniently rectangular. 1
Note that in the abstract finite element method, we do not require the (f)1 of Chapter 8.
to interpolate at sorne node zi' and we do not require them to be piecewise We frankly believe that the nodal n
polynomials. (We prove, however, thatthe latter are the most efficient.) The geometry that it will continue to domina
interpolating property will be possessed by every basis which falls also within polating property is not essential t<;> par
the scope of the nodal finite element method, but our theory in the abstract convenient when on a regular mesh to be
case will not use this property. Asan example we consider the case of cubic tion in terms of functions CfJ 1 , ••• , CfJM·
splines in one dimension. In this case M 1, and a suitable choice for CfJ 1 is of Sh depends on these functions, they m1
the B-spline of Fig. 1.9. The space Sh of all combinations ¡: q1rp?,z is then ex- to all our questions about approximat
actly the spline space of all twice differentiable piecewise cubic functions with fore, this abstract approach turns a si
joints at x O, ±h, ±2h, .... It is essential to see that the B-spline CfJ 1 is analysis into an agreeable problem in ft
not interpolating; it is nonzero over four intervals instead of the two allowed
in the nodal method, and in particular it is nonzero at three nodes instead
ofone. 2.2. RATES OF CONVERGENCE
The spline subspaces Sh included in the abstract finite element method
have been used with sorne success in one-dimensional applications. The Suppose that u is the solution to (
assembly of the equation KQ F and the treatment of the boundary problem of order m. This means that ;
require modifications in the technique which is standard in the nodal case, class JC}l, restricted by homogeneous 01
but the ground rules are still those common to any form of the Ritz conditions, and (by ellipticity) that tp
method. The chief advantage in splines is that the extra continuity reduces a(v, v) crllvll~· Suppose also that uh
the dimension of the trial space, without destroying the degree of approxi- trial class Sh and that üh is the solutio1
mation. In the nodal case, the Hermite cubic polynomials are determihed numerical integration; the coordinate
 CHAP.2
the four basis functions associated with
- 1¡, yjh - /2).
useful and important that we \)shall make
~neral description of the finf~e element
on a regular mesh, and we ""call it the
· dimensions it begin~ with the choice of
.ese will eventually le:ad to M unknowns
nt equations KQ = Fwill take the form
1ns.
ciated with one particular grid point, the
~ the Variable X = (x 1 , ••• , Xn), obtain-
1 the basis functions associated with a
)h, we translate the functions just con-
;is functions constructed in this way, by
' subscripts i and /:
/), i = 1, ... , M .
.t the original functions «l>¡(x) shall vanish
Ptrwill vanish outside the local sphere
have a local (but possibly not a strictly
!ment method, we do not require the «l>i
.ve do not require them to be piecewise
at the latter are the most efficient.) The
ed by every basis which falls also within
method, but our theory in the abstract
t example we consider the case of cubic
M= 1, anda suitable choice for Cl> 1 is
of all combinations .L; q1rpf, 1 is then ex-
rentiable piecewise cubic functions with
essential to see that the B-spline el> 1 is
'our intervals instead of the two allowed
ar it is nonzero at three nodes instead
in the abstract finité element method
in one-dimensional applications. The
and the treatment of the boundary
e which is standard in the nodal case,
! common to any form of the Ritz
tes is that the extra continuity reduces
wut destroying the degree of approxi-
tite cubic polynomials are determined
SEC. 2.2. RATES OF CONVERGENCE 105
over each subinterval by v and v' at the endpoints; this means M = 2
parameters for each mesh point. Their generating functions el> 1 a11d tl> 2 are the
IJf and ro of Fig. 1.8.
Experiments have been carried out with splines in two dimensions, where
the reduction in Mis still greater, but the boundary is almost forced to be
regular if there is to be any comparison with the simplicity of the nodal
method. The bandwidth of the stiffness matrix K is of course not necessarily
decreased by a reduction in M; for cubic splines in one variable there will be
seven nonzero en tries in each row of K, since the B-spline el> 1 stretches over
four intervals. This is virtually the same bandwidth as in the Hermite case,
where the usual ordering of unknowns gives a matrix of the form
"(~ ~) (~ ~) (~ ~)
K=
(~-~) (~ ~) (~ ~)
The order of the matrix K is of course directly affected by the number M,
and splines do appear to be more efficient (at least in solution time) when the
domain n is conveniently rectangular. This is confirmed in the calculations
of Chapter 8.
We frankly believe that the nodal method allows such flexibility in the
geometry that it will continue to domínate the use of splines. Since the inter-
polatihg property is not essential to part of the analysis, however, it will be
convenient when on a regular mesh to be able to use the more general descrip-
tion in terms of functions el> 1 , • • • , el> M· Since everything in the construction
of Sh depends on these functions, they must contain in themselves the answers
to all our questions about approximation and numerical stability. There-
fore, this abstract approach turns a significant part of the finite element
analysis into an agreeable problem in function theory.
2.2. RATES OF CONVERGENCE
Suppose that u is the solution to an n-dimensional elliptic variational
problem of order m. This means that u minimizes l(v) over an admissible
class JC~, restricted by homogeneous or inhomogeneous essential boundary
conditions, and (by ellipticity) that the strain energy is positive definite:
a(v, v) >a 11 v 11!: Suppose also that uh is the minimizing function over the
trial class Sh and that uh is the solution when the problem is perturbed by
numerical integration; the coordinates of uh satisfy KQ = F. Suppose,
106 A SUMMARY OF THE THEORY CHAP. 2 SEC. 2.2.
'~

finally, that üh is tHe solution which is actually computed, differing from üh Iated. (Most notably in the biharmonic e
because of roundoff error in the numerical solution. Obviously the three twist term xy was left out of Sh.) We sha:
approximations u\ ü\ üh include progressively more sources of error. We we know, the first proof.:_that this con<
want to summarize, for problems with smooth solutions and for typical the case of a regular mesh. Such a the
finite elements, what are the orders of magnitude of these errors. finite element theory, which admits th1
The s~arting point is always the same: lf Sh is a subspaceof 3C~, thenby regular subdivision.
the fund~mental Theorem 1.1 We have stated the constant strain <
for irregular meshes, but it is not. At leé
(3) a(u u\ u- uh) = min a(u- v\ u vh). be mapped into another domain n' b~
v" in S"
mation T which does not preserve pol
The theorem still holds with inhomogeneous essential conditions, where the satisfy the constant strain condition on l
difference of any two functions in JC~ líes in V0 and the difference of any two However, convergence on n implies cor
functions in Sh lies in Sg, as 1\)ng as Sg is a subspace of V 0 • (For proof see le m on n', since we can go freely back ¡
Section 4.4.) Therefore, the strain energy in u-uh is a question in pure.approxi- on errors for isoparametric elements is
mation theory, to estímate the distance between u and Sh. The basic hypo- Convergence in strain energy is esse
theses will be, first, that Sh is of degree k - l-it contains in each element atives of uh to those of u. This deriva
the complete polynomial of this degree, restricted only at the boundary by Ritz method is minimizing in energy. F
essential conditions-and second, that its basis is uniform as h O. The smaller or larger than m, convergence 1
1atter is effectively a geometrical condition on the element regions: lf the the order of the best approximation to
diameter of ei is h1, then e1 should contain a sphere of radius at least -ch1, This rate of convergence will normally
where T is bounded away from zero. This forbids arbitrarily small angles in · Using the "Nitsche triek," which gave tl
triangular elements. Angles near n must also be forbidden in quadrilaterals. dimensional linear element~ in Section
Under these conditions, the distance between u and Sh is
(4)
a(u - u\ u - uh) C 2h 2 <k-m)¡ u 1¡.
In almost all cases the first exponent i
Therefore, h 2 <k-m) is the rate of convergence in strain energy. governs the convergence rate. [For s
The form of this error bound is typical of the results of numerical analysis. error in displacement over an eleinent
There are three factors. The power of h is the simplest to find, since it depends order better (hk+l) than the displaceme
only on the degree ofthe polynomials, and it indicates the rate of convergence which the h 2 <k-m) term would enter: lf'
as the mesh is refined; this effect should be clearly observable numerically. sixth-ordet: problem, or more realist
The constant C depends on the construction of the element and its nodal equation) element complete only throu:
parameters. For regular geometries a good asymptotic value of C can be limited to 2(k m)= 2 even in disph;
computed as the error in approximating polynomials of the next higher The convergence rate at individual]
degree k (Section 3.2). The third factor, 1 u !k, reflects the properties of the provided u has k derivatives at every p
problem itself-the degree to which the solution is smooth, and therefore easy differentiability and therefore the rate
to approximate accurately. This norm is the mean-square value of the kth the mean-square and pointwise senses.
derivatives of u and therefore-according to partial-differential-equation optimum pointwise error bounds.) At
theory-directly related to the derivatives of order k 2m of the data f converge faster than it does in the mear,
Notice that convergence occurs if and only if k > m; this is 'the constant for example, the nodes themselves wer
strain condition, that the elements should reproduce exactly any solution tion is exact at the nodes. This is true):
which is a polynomial of degree m. This requirement for convergence appeared to the homogeneous differential equat
very gradually in the engineering literature, partly developing from intuition Thomée noticed a special convergence
and partly fron the numerical failures which were observed when it was vio- and Dupont are extending this princi¡
 CHAP.
1 is actually computed, differing&rom uh
wmerical solution. Obviously the three
rogressively more sources of error. We
with smooth solutions and for typical
of magnitude of these errors.
same: If Sh is a subspace of JC~, thenby
min a(u - v\ u - vh).
¡¡hin Sh
)geneous essential conditions, where the
~ lies in V0 and the difference of any two
: S~ is a subspace of V 0 • (For proof see
• h. • • .
rgy m u-u 1s a questwn m pure approxi-
mce between u and Sh. The basic hypo-
~ree k - l-it contains in each element
gree, restricted only at the boundary by
hat its basis is uniform as h ~ O. The
mdition on the element regions: If the
:ontain a sphere of radius at Ieast rh¡,
This forbids arbitrarily smaJI angles in
mst also be forbidden in quadrilaterals.
between u and Sh is
!rgence in strain energy.
Ji cal of the results of numerical analysis.
h is the simplest to find, since it depends
, and it indica tes the rate of convergence
)Uld be clearly observable numerically.
struction of the element and its nodal
a good asymptotic value of C can be
ating polynomials of the next higher
tor, 1u lk, reflects the properties of the
e solution is smooth, and therefore easy
n is the mean-square value of the kth
ording to partial-differential-equation
ltives of order k - 2m of the data f
r- and only if k > m; this is 'the constant
r;hould reproduce exactly any solution
requirement for convergence appeared
ature, partly developing from intuition
which were observed when it was vio-
2 SEC. 2.2. RATES OF CONVERGE~CE 107
Iated. (Most notably in the biharmonic case m = 2 of plate bending, when the
twist term xy was left out of Sh.) We shaJI describe a rigorous proof_:_as far as
we know, the first proof-that this condition is necessary for convergence in
the case of a regular mesh. Such a theorem fits naturaJly into the abstract
finite element theory, which admits the most general trial functions on a
regular subdivision.
We have stated the constant strain condition as if it were necessary also
for irregular meshes, but it is not. At least it is not quite necessary. n could
be mapped into another domain Q' by a fixed smooth invertible transfor-
mation T which does not preserve polynomials, and then elements which
satisfy the constant strain condition on n will not do so in the new variables.
However, convergence on n implies convergence of the corresponding prob-
lem on n', since we can go freely back and forth. The effect of the Jacobian
on errors for isoparametric elements is considered in Section 3.3.
Convergence in strain energy is essentiaJly convergence of the mth deriv-
atives of uh to those of u. This derivative is therefore special, because the
Ritz method is minimizing in energy. For the sth derivative, where s may be
smaller or larger than m, convergence cannot be [aster than O(hk-s); this is
the order of the best approximation to u from a space Sh of degree k - l.
This rate of convergence will normaJly be attained by the Ritz solution uh.
Using the "Nitsche triék," which gave the O(h 2 ) rate in displacement for one-
dimensional linear elements in Section 1.6, we shaJI show that
(4) 11 u - uh lis = O(hk-s + hz<k-ml).
In almost all cases the first exponent is smaller, and approximation theory
governs thé convergen ce rate. [For s = -1, the leff si deis effectively the mean
error in displacement over an element, and we see that it may weJI be one
order better (hk+l) than the displacement itself.] There are, however, cases i~
which the h 2 <k-mJ term would enter: If we imagined cubic splines applied toa
sixth-order problem, or more realistically a plate-bending (fourth-order
equation) element complete only through quadratics, then the rate would be
limited to 2(k- m) = 2 even in displacement.
The convergence rate at individual points can be expected to be the same,
provided u has k derivatives at every point. (With singularities, the order of
differentiability and therefore the rate of convergence are quite different in
the mean-square and pointwise senses. We shall not give a detailed proof of
optimum pointwise error bounds.) At special points the error may actually
converge faster than it does in the mean. With linear elements for -u" =f,
for example, the nodes themselves were special: uh = u1, and the Ritz solu-
tion is exact at the nodes. This is true in general if the elements are solutions
to the homogeneous differential equation [H6, T4]. For the heat equation,
Thomée noticed a special convergence rate at the nodes of splines; Douglas
and Dupont are extending this principie to their collocation methods.
 ) )

)¿ !

1os A SUMMARY OF THE THEORY CHAP. 2 SEC. 2.2.

The rate of convergence is maintained with inhomogeneous boundary Iow-is normally of arder h 3 • Roughly
conditions (Section 4.4), provided the boundary data are interpolated (or triangles and interpolate the essential
approximated) by polynomials of at least the same degree k- l. these cubics will still violate the cor
For a domain Q ofirregular shape,a new type of approximation error may in Section 4.4, we deduce from this tht
enter. 1t will generally be necessary to approximate the boundary r by a There is still another alternative, wh
piecewis~ polynomial boundary rh. In the simplest case rh is piecewise linear; Curved elements can be straightened m
Q is repÚtced by a polygon Qh. Such a polygon can be carved into triangles, transformation may even be necessary t
and the finite element method may proceed by ignoring the skin Q - Qh Iaterals whose sides are already straight
r
between the true boundary and the polygon. Therefore,it is as ifthe original These coordinate changes are a central t~
differential problem were moved to Qh. In Section 4.4 we investiga te the effect In theory we can straighten almost :
of this change of domain. Briefly, the mth derivative is in error by O(h) at the practice that is absurd. Piecewise polyr
boundary, but this error decays rapidly in the interior. There is a boundary- ríes for the same reasons that they are th
layer effect, and the mean error is O(h 312 ). Since the strain energy depends on ments: they can be handled efficientl)
the square of this derivative, the error in energy due to computing on Qh describe the coordinate changes by th~
(with natural or essential boundary conditions) is O(h 3 ). This estima te applies used as trial functions; this is the metl
to Q as well as Qh ifthe finite element solution is extended in the natural way, It is a brilliant idea. Coordinate chang
by extending each polynomial out to r. Otherwise all the energy in the skin trial functions: The mapping must be e
will be lost, and this amounts to O(h 2 )-proportional to the volume of the so that elements which are adjacent i
skin. adjacent in the f,-r¡ plane. If the transfo
Note that the h 3 error due to change of domain will domínate when the ted from nodal parameters in the sta
finite elements used in the interior are of degree higher than m. Ifthe polyno- e
continuity in and r¡ (as we are for tl
mials are of the mínimum degree m which is re'quired for convergence, then elements), then the isoparametric mai
the change of domain effect will be submerged (at least in the interior of Qh) whose boundaries are polynomials of d
by the h 2 ck-m> = h 2 error which arises from ordinary approximation theory. raises new questions of approximation
If the boundary r
is approximated by piecewise polynomials of higher are no longer polynomials in x and y.
degree 1, there is a corresponding reduction in the change of domain error. mations need not decrease the orde1
The error in the mth derivative (the strain) at the boundary is O(N), and in the sth derivative is achieved, pro'vide<
the overall strain energy in Qh it is O(h 21 + 1). This assumes, in case of an es- smooth (Section 3.3). In this sense the
sential condition u = O, that the condition is exactly satisfied by the polyno- for second-order equations and curve<
mial trial functions on the approximate boundary. Mi.tchell has found a condition u = g can be handled natur;
neat construction of cubic elements which vanish on a boundary made up fundamental order of accuracy.
of piecewise hyperbolas, with h 5 error in strain energy and h 3 in displacement. The errors enumerated so far hav
For an essential condition on a curved boundary, say u = O, an alterna- assumed the Ritz approximation to b<
tive is to use any of the standard elements and to require that they interpo- ever there will also be errors in nun:
late the conditions at boundary nades. In this case the trial functions will not of u~) and in the solution of the finall
satisfy tlre essential condition along the whole boundary; each trial function It is essential to know the scale of the i
may vanish along sorne curve clase to r, but this zero curve will vary from rule can be chosen which is neither .~
one function to the next. As a result, the Ritz theory will not apply; the trial curacy. We mentioned in the·exampl
functions are not in JC~, either on the exact Q or on an approximate Qh. lt bilities: (1) the inhomogeneous data a1
no longer follows that u\ which is still eh osen to minimize /( v), is the closest with x can be replaced by interpola.ti
trial function to u. Nevertheless, it is possible to estímate the error, taking the resulting problem computed exactl
into account that each vh is nonvanishing but small on the boundary. The can be carried out from the beginnir
best error bound on the strain energy-which surprised us by being rather Gaussian quadrature. For a simple pr
 ' CHAP~
ntained with inhomogeneous boundary
the boundary datá are interpolated (or
: least the same degree k - 1.
:::,a new type of approximation'error may
V to approximate the boundary r by a
rh
n the simplest case is piecewise linear;
a polygon can be ca,t.yed into triangles,
proceed by ignoring· the skin n - Qh
polygon. Therefore, it is as if the original
~h. In Section 4.4 we investiga te the effect
mth derivative is in error by O(h) at the
Hy in the interior. There is a boundary-
:73/2). Since the strain energy depends on
~or in energy due to computing on Qh
mditions) is O(h 3 ). This estimate applies
solution is extended in the natural way,
r. Otherwise all the energy in the skin
12 )-proportional to the volume of the
nge of domain will dominate when the
of degree higher than m. If the polyno-
vhich is required for convergence, then
1bmerged (at least in the interior of Qh)
s from ordinary approximation theory.
:::d by piecewise polynomials of higher
uction in the change of domain error.
.train) at the boundary is O(h1), and in
'(h 21 + 1 ). This assumes, in case of an es-
ition is exactly satisfied by the polyno-
nate boundary. Mitchell has found a
.vhich vanish on a boundary made up
in strain energy and h 3 in displacement.
rved boundary, say u = O, an alterna-
lents and to require that they interpo-
In this case the trial functions will not
e whole boundary; each tri al function
r, but this zero curve will vary from
he Ritz theory will not apply; the trial
: exact n or on an approximate Qh. It
chosen to minimize I(v), is the closest
possible to estimate the error, taking
hing but small on the boundary. The
-which surprised us by being rather
2 RATES OF CONVERGENCE 109
low-is normally of order h 3 • Roughly speaking, if we work with cubics on
triangles and interpolate the essential boundary conditions, the worst of
these cubics will still violate the conditions by O(h 312 ) between nodes;
in Section 4.4, we deduce from this the O(h 3 ) error. in energy.
There is still another alternative, which is undoubtedly the most popular.
Curved elements can be straightened out by a change of coordinates. Such a
transformation may even be necessary to achieve continuity between quadri-
laterals whose sides are already straight, unless they happen to be rectangles.
These coordina te changes are a central technique in the finite element method.
In theory we can straighten almost any boundary curve, but of course in
practice that is absurd. Piecewise polynomials are the best element bounda-
ries for the same reasons that they are the best approximations to the displace-
ments: they can be handled efficiently by the computer. In fact, we may
describe. the coordinate changes by the same class of polynomials that are
used as trial functions; this is the method of isoparametric transformations.
It is a brilliant idea. Coordinate changes lead to the same difficulties as for
trial functions: The mapping must be continuous across element boundaries,
so that elements which are adjacent in the original x-y plane will remain
adjacent in the e-t¡ plane. If the transformations x(e, t¡), y(e, t¡) are construc-
ted from nodal parameters in the standard way, and if we are assured of
e
continuity in and t¡ (as we are for the standard rectangular or triangular
elements), then the isoparametric mappings will succeed even for elements
whose boundaries are polynomials of degree k - 1 in x and y. This technique
raises new questions of approximatiort theory, since polynomials in and t¡e
are no longer polynomials in x and y. N~vertheless, isoparametric transfor-
mations need not decrease the order of accuracy; the full order hk-s in
the sth derivative is achieved, provided these transformations are uniformly
smooth (Section 3.3). In this sense the isoparametric technique is the best one
for second-order equations and curved boundaries. An essential boundary
condition u= g can be handled naturally and efficiently, with no loss in the
fundamental order of accuracy.
The errors enumerated so far have all contributed to u- uh; we have
assumed the Ritz approximation to be calculated exactly. In practice, how-
ever, there will also be errors in numerical integration (yielding U.h instead
of uh) and in the solution of the final linear system (yielding üh instead of U.h).
1t is essential to know the scale of the integration errors, so that a quadrature
rule can be chosen which is neither wasteful of time nor destructive of ac-
curacy. We mentioned in the·example of Chapter 1 the two leading possi-
bilities: (1) the inhomogeneous data and any material coefficients which vary
with x can be replaced by interpolating polynomials and the integrals for
the resulting problem computed exactly; or (2) integrations over all elements
can be carried out from the beginning by a standard quadrature rule, say
Gaussian quadrature. For a simple problem like -(pu')' + qu = J, the user
110 A SuMMARY OF THE THEORY CHAP. 2 SEC. 2.2.

has sorne freedom of choice, and we mention one result ofthe analysis: Both Galerkin's method applies directly to
interpolatíon by polynomials of degree k - 1, and Gauss quadrature. with uh = ~-Qitpi in the subspace Sh such th(
k - 1 points, yield errors in the strains of order hk. For more complicated
problems,'method (2) is strongly indicated, and in fact numerical integration (M(uh), rpk) =O
has become one of the major components of the finite element method. It The number of (nonlinear) equations eqt
produces a solution üh whose convergence to u requires a certain accuracy
cients; that is, it equals the dimension
in the nuliherical quadrature: The mth derivatives of al! tria! functions must be classes described abo ve, it is possible to p.
integrated'exactly. For each additional degree of exactness, the error ül¡- uh
uh and its convergence to u-provided
dueto numerical quadrature improves by a power of h. The proof depends on imposed on the operator. In fact, om
an identity given in Section 4.3, and we also determine the extra accuracy solution u follows these lines; the existe
required for isoparametric elements. diinensions, as well as an a priori boum
Finally we come to the roundoff error. This has a character completely sorne cornpact set. Then the uh must ha"
different from the others; it is proportional to a negative power of h. As h u. Ciarlet, Schultz, and Varga [C4] have ~
decreases there is a region of crossover, prior to which roundoff is negl.igible very little between linear and rnonotone
and uninteresting, and after which it is all-imp,ortant. The roundoff does not The finite element method is in cot
depend too strongly either on the degree of the polynomials u sed or on the for example in·elastic-plastic materials o
number of dimensions; the key factor h- 2 m is set by the mesh width and the tion to the text by Oden [15], there is a raJ
order of the equation itself. Por second-order equations the crossover usual- we rnention among many others the early
ly occurs below the mesh widths which are typical in practica! problems, but Marcal [MI]. 1t seems fruitless for us ton
for equations of order 4 it is not so delayed; accurate calculations may well formulations, distinguishing between th
require double precision. We discuss in Chapter 5 both a priori estimates of earities dueto large deflections and thost
the condition number and a posteriori estimates of the roundoff actually doing something substantial about the
committed in a given problem.
that the convergence problems are we1
These are the principal errors to be analyzed for linear static problems ripe for solution. We hope that ultimate
Lu =f. In every case, the analysis is based on the fundamental variational
come sufficiently complete to justify a h
equation, the vanishing of a(u, v) - (f, v). In comparison with finite diff~r­ We do want to include one warning e:
ences, there can be no dispute that this produces a more coherent and satts- to problems such as nonlinear elasticity,
factory rnathematical theory. In part, the same techniques can be extended to
minimization of a functional like
nonlinear equations; this rnust surely be at present the outstanding theoret-
ical problem, to isolate classes of nonlinear problems which are both phys-
ically important and mathematically amenable, and to study the dependence
of the approxirnation on the trial space, domain, and coefficients. Lions has
rnade trernendous progress in this direction (see [9]). Since J(v) is not quadratic in v, the vanis
Two classes have already been isolated, as yielding the simplest general- a linear equation for u. Therefore son
izations of linear elliptic equations. We cannot say how comprehensive are simplest being the method of successJve
their applications in engineering and physics, but they appear to be very coe:fficient at the nth approximation un, e:
natural. One is the class of strictly monotone operators, satisfying a linear problem. .
Our warning is this: if such an Ite
J [M(u)
n M(v)](u v) dx >a 11 u v 11!.
problem, so that u,.+ 1 minimizes

The other (closely related) class contains the potential. operators, those for (d)
which Mis the gradient of sorne nonquadratic convex functional l(v). We
refer to the book of Vainberg [18] for an introduction to the theory. then the iteration will converge to the l1
 CHAP.
e mention one result ofthe analysis: Both
~g~ee k l, and Gauss quadrature with
r~ms of ord~r hk. For more complicated
i1cated, and m fact numerical integration
ponents of the finite element method. It
ergence to u requires a certain accuracy
th derivatives of all ttfipl functions must be
nal degree of exactne·ss, the error ü" uh
es by a power of h. The proofdepends on
d we also determine the extra accuracy
f er:or. This has a character completely
~ortiOnai to a negatil'e power of h. As h
ter, prior to which roundoff is negligible
· is all-important. The roundoff does not
!gree of the polynomials used or on the
)r h- zm is set by the mesh width and the
•nd-order equations the crossover usual-
eh are typical in practica! problems, but
~elayed; accurate calculations may weii
m Chapter 5 both a priori estimates of
iori estimates of the roundoff actually
be analyzed for linear static problems
i based on the fundamental variational
(/, v). In comparison with finite differ-
tis produces a more coherent and satis-
the same techniques can be extended to
~e at present the outstanding theoret-
1hnear problems which are both phys-
tmenable, and to study the dependence
::e, domain, and coefficients. Lions has
·ection (see [9]).
lated, as yieiding the sirnplest general-
le cannot say how comprehensive are
physics, but they appear to be very
monotone operators, satisfying
1ins the potential operators, those for
tquadratic convex functiona1 I(v). We
an introduction to the theory.
2 SEC. 2.2. RATES OF CONVERGENCE 11f
Galerkin's method applies directly to these operators; we seek a function
uh Í:~iqJi in the subspace Sh such that \
for all rpk.
The number of (nonlinear) equations equals the number of unknown coeffi-
cients; that is, it equals the dimension of the trial space Sh. For the two
classes described abo ve, it is possible to pro ve the existence of such a solution
uh and its convergence to u-provided a suitable continuity condition is
imposed on the operator. In fact, one possible existence proof for the
solution u follows these lines; the existence of uh is demonstrated in finite
dimensions, as well as an a priori bound which establishes that all uh lie in
sorne compact set. Then the uh must ha ve a limit point as h......-¡. O, and this is
u. Ciarlet, Schultz, and Varga [C4] have shown that the error estimates differ
very little between linear and monotone nonlinear problems.
The finite element method is in constant use for nonlinear problems,
for example in elastic-plastic materials or in thermo-viscoelasticity. In addi-
tion to the text by Oden [15], there is. a rapidly growing engineeringliterature;
we mention among many others the early survey articles by Martín [M2] and
Marcal [Ml]. It seems fruitless for us to repeat here the possible finite element
formulations, distinguishing between the techniques for geometric nonlin-
earities due to large defiections and those for material nonlinearities, without
doing something substantial about the mathematics. It is absolutely clear
that the convergence problems are well posed, extremely interesting, and
ripe for solution. We hope that ultimately this mathematical.theory will be-
come sufficiently complete to justify a book (by someone else}
We do want to include one warning about nonlinear equations. It applies
to problems such as nonlinear elasticity, and we take as a simple model the
minimization of a functional like
I(v) J [p(v, vx)v! - 2fv] dx.
Since I(v) is not quadratic in v, the vanishing of the first variation will not be
a linear equation for u. Therefore sorne iteration will be considered, the
simplest being the method of successive substitution: evaluate the nonlinear
coefficient at the nth approximation un, and determine un+ 1 as the solution of
a linear pro blem. ·
Our warning is this: if such an iteration is applied to the variational
problem, so that un+ 1 minimizes
(d)
then the iteration will converge to the wrong answer. The reader can easily
112 A SUMMARY OF THE THEORY CHAP.2 SEC. 2.2.

verify that the limit u* ?r such an iteration satisfies

J
At a given t, the coefficient ph(x) \
modulus p. Instead it will depend o
* *) du*)
d ( p(u '
dx dx -f,
Ux - '
u~(-r), 1' t. To prove convergence,
strains ha ve remained elose to the
given instant is close to the true p:
which is not the equation ofvirtual work for I(v). (Take p vx.) The mistake
was to p1~.rform successive substitutions before taking the first variation; if
we first establish the nonlinear equation for the minimizing function u, and max l p(x) - ph(x) 1 <
X

then solve that equation by iteration, the limit will be the right one.
There is one particular nonlinear problem, describing the deformation To prove convergence, we must
of an elastic-plastic material, which illustrates both the possibilities and the mizes /(vh), and its x-derivative the
difficulties in nonlinear convergence proofs. lt might be useful to present coefficient p in the variational pri
sorne of the details. For simplicity of notation we consider a one-dimensional time t into two parts,
model, with strain dufdx; the same arguments apply to the system ofstrains
e1¡ occurring in two- or three-dimensional elasticity. The deformation is a
nonlinear function of the externa! forces, and cannot be determined from a wh is the function which minimizes
knowledge only of the final loading. Instead it depends on the history of the the true elastic coefficient p at time ¡
problem, that is on the chronological order in which the forces are applied approximation, and the second to a
over the domain., This introduces an artificial "time" parameter, and at a first we can appeal to the standa1
given instant t the rate of change ü of the deformation is the function which problem at each instant of time, ar
minimizes space of degree k 1 is
I(v) f [p(vJ 2
- 2jv] dx.

The crucial quantity is the elastic modulus p(x). If this coefficient is in- For the second part we must turn to
dependent of the stresses within the material, then there is no need for the asserts that the effect of a change i1
artificial time; the final deformation u(T) can be determined by a single mini- 11 w!(t)- u!(t) 11
mization, as in the rest of this book, using the final load f(T). In the case of a
nonlinear stress-strain law, the coefficientp at time t depends on the stresses, Substituting these last two bounds,
and not only on their values at the given instant: it depends on the whole
stress history. This "path-dependence" enters physically in severa! ways. 11 e(t) 11 < e hk-1
For example, once the elastic Iimit has been exceeded, a loading followed by
an equal unloading leaves a net change in the state of stress in the material. This is exactly the situation in whic
This phenomenon actually creates a nonlinear problem for v at each instant Gronwall's lemma: dividing throug
of time, since the elastic modulus then depends not only on the past history multiplying by CC", and integratin!
but also on the current rate of change; p is influenced by the sign of v xv x'
and has one value for loading and another for unloading. For simplicity we
shall avoid this extra difficulty, and assume there is no unloading. We do
log ( C' hk- 1 + CC" s: 11 e(í

not· require, however, that the stress be a· single-valued function of the Taking the exponential, and substit
strain-only that it can be computed from a knowledge of the strains ux(-r)
at all times -r < t. lle(t)ll e
In the Ritz approximation, Ü' is the function in the trial suhspace Sh
Finally, an integrati~n with respect ·
which minimizes
 CHAP. 2 SEC. 2.2. RATES OF CONVERGENCE ·113

on satisfies At a given t, the coefficient ph(x) will not have the same value as the true
modulus p.· Instead it will depend on the history of the Ritz approximations
du*)
-
dx
=/, u~(r), 't' t. To prove convergence we :rp_ust assume that ifthese approximate
strains have remained close to the true ones, then the coefficient ph at the
given instant is el ose to the true p:
for I(v). (Take p = vx.) The mist~ke
befo re taking the fir~t variation; if
for the minimizing fdnction u, and
limit will be the right une.
To prove convergence, we must estímate how the function úh which mini-
oblem, describing the deformation
:rates both the possibilities and the mizes l(iJh), and its x-derivative the rate of change of strain, depend on the
ofs. It might be useful to present coefficient p in the variational principie. Our plan is to split the error at
.tion we consider a ene-dimensional time t into two parts,
1ents apply to the system of strains úx(t) - ü!(t) (ú/t) w!(t)) + (w!(t) - ü!(t));
al elasticity. The deformation is a
, and cannot be determined from a wh is the function which minimizes I(iJh) over all trial functions in S\ using
!ad it depends on the history of the the true elastic coefficient p at time t. In other words, the first error is dueto
ier in which the forces are applied apptoximation, and the second to a change in the elastic coefficient. For the
:ificial "time" parameter, and at a first we can appeal to the standard approximation theory; it is a linear
· deformation is the function which próblem at ·each instant of time, and the error in the first derivati ves for a
space of degree k 1 is
2/v] dx.
For the second part we must turn to Section 4.3. The corollary in that section
dulus p(x). If this coefficient is in-
asserts that the effect of a change in coefficient is bounded by
erial, then there is no need for the
::an be determined by a single mini- 11 w!(t) ü!(t) 11 < C" max lp(x) ph(x) 1.
the final load f(T). In the case of a
p at time t depends on the stresses, Substituting these last two bounds, and writing e úx - ü!,
n instant: it depends on the whole
~nters physically in severa! ways.
m exceeded, a loading followed by
the state of stress in the material. This is exactly the situation in which to apply the argument which leads to
near problem for iJ at each instant Gronwall's lemma: dividing through by the right side of the inequality,
pends not only on the past history multiplying by CC", and integrating with respect to t,
· is influenced by the sign of v xiJ x'
r for unloading. For simplicity we
me there is no unloading. We do
log ( C'hk-t + CC" s: 11 e('t') 11 d't') log C'hk-t < CC"t.

a single-valued function of the
a a knowledge of the strains uit')
function in the trial subspace Sh
Taking the exponential, and substituting into the previous inequality,
11 e(t) 11 C' hk-t exp( CC" t).
Finally, an integration with respect to t yields

lle(t)ll < f He(t)ll dt < C"'hk-t.

 ( )
' (, l 2.2.
114 A SUMMARY OF THE THEORY ·cHAP.2 SEC.

Therefore. the rate of convergence in h is the same for this nonlinear plastic Galerkin's rule: The residual u~+ Luh
problem as it is for linear elasticity. the true solution u happens to lie in tt
Notice that we have assumed a Ritz ápproximation computed continu- Sh should ·vanish. Thus the original
ously in tin:e: the only discretization so far has been the change from the full In practice, time must also be made
admissible space to its subspace Sh. This is in keeping with the presentation a finite difference method. The Crar
ofinitial ~alue problems in Chapter 7, where the Ritz error is separated from centered at the time tn+ 1¡ 2 when uh(t,+ 1)
the error due to a finite difference method (or other procedure) in the time accurate to order tlt 2 • Thus the final,
direction. For the nonlinear problem there has been considerable discussion error as well as the Galerkin error dut
of the best "incremental method," but we believe all of the leading possibil- one which we analyze in detail, in ord
ities to be convergent. They should simply contribute a new error-propor- ofthe optimal order ftk-s in the sth der
tional to a power of tlt, in case of a difference equation. of parabolic type, for example the he~
There is one technical difficulty in the proof, however, which an over- liptic operator that occurs in static prc
zealous conscience will not allow us to ignore. It is a question of the choice tions, which include no dissipative t
of norm: if the double bars representa mean-square non~, then the pointwise method is somewhat diminished; the
bound given for p - ph is not strictly valid. On the other hand, to use maxi- explicit difference methods may be to•
mum norms throughout requires a fresh examination of the hk-t estimate for the comparatively automatic treatmen1
üx - w!. This bound followed from mean-square approximatio~ theory, and tage, and we include in Chapter 7 a s
perhaps the simplest remedy is to establish instead a pointwise estlmate of theory.
the Ritz error in static linear problems. Another possibility is to use an idea For want of space, the analysis of
proposed by the first author (for subsequent applications see [W3]) which tion, and roundoff will be limited to
permits switching backand forth between these two norms when the solutión eigenvalue and initial-value problems
is smooth; such a technique is frequently required in nonlinear problems the case of quadrature errors these e]
when the error estimates are global but instabilities can arise locally. Or a ried out by Fix (Baltimore Finite Eler
third possibility is to improve the corollary in Section 4.3, to depend instead In the final chapter, we present th
on the mean-square norm of the perturbation p - ph. We are confident that numerical experiments. One of the me
the basic proof is correct, and that a combination of experiment and analysis strong singularities, produced by a e
will soon lead to a much fuller understanding of nonlinear errors. is a classical problem in fracture mee
This summary of the theory must include also eigem•a!ue problems and factor at the head of the crack; aroun
initial-1•alue problems. The finite element method applies directly and suc.cess- Therefore, a number of questions aris
fully to both. For self-adjoint eigenvalue problems, the computation of
upper bounds by minimizing the Rayleigh quotient over a subspace is a l. Do our error estimates yield ti
classical technique; it leads to a discrete eigenvalue problem KQ J..MQ, pointwise and in mean-square?
where K and M are precisely the same stiffness and mass matrices already 2. The singularity reduces the sm(
encountered. We devote Chapter 6 to deriving this discrete form and to convergence; does a reduced rate ap¡:
estimating the errors in eigenvalue and eigenfunction which depend on so that the singularity pollutes the wh
approximation theory~they are due to replacing the true admissible space 3. Is it possible, by gradin'g the me
JC~ by Sh. The results are simple to descibe: J..- ).h is of order fl2Ck-m>, tions at the singularity, to recover tl
and for k 2m the eigenfunction errors in the sth derivative are of the method?
maximum order hk-s permitted by approximation theory.
+
For initial-value problems u, Lu = f, the position is equally favorable. The answer in each case is "yes,"'
The finite element solution has the form uh(t, x) 2: Q/t)rp/x); the time vincing. Both for static and for eigenv
variable is left continuous, while the dependence on x is discretized in terms of singularities which are introduced by
the standard piecewise polynomial basis functions fPr The coefficients Q:(t) duction of trial functions which mirr
are determined by a system of N ordinary differential equations, expressing Straight interfaces between mater
 r__,.Q;IAP. 2
is the same for this nonlinear plastic
- (
tz approximation CO}nputed continu-
) far has been the change from the full
:is is in keeping with the presentation
.vhere the Ritz error is separated from
.hod (or other procei:l,ure) in the· time
1ere has been considerable discussion
we believe all of the leading possibil-
Iply contribute a new error-propor-
fference equation.
the proof, however, which an over-
ignore. It is a question of the choice
nean-square norm, then the pointwise
alid. On the other hand, to use maxi-
1 examination ofthe hk-_1 estímate for
an-square approximation theory, and
blish instead a pointwise estímate of
Another possibility is to use an idea
equent applications see [W3]) which
:n these two norms when the solutión
ntly required in nonlinear problems
t instabilities can arise locally. Or a
lary in Section 4.3, to depend. instead
bation p - ph. We are confident that
nbination of experiment and analysis
:rstanding of nonlinear errors.
nclude also eigenl'alue problems and
method applies directly and success-
:tlue problems, the computation of
!eigh quotient over a subspace is a
:e eigenvalue problem KQ ).MQ,
stiffness and mass matrices already
deriving this discrete -·form and to
nd eigenfunction which depend on
• replacing the true admissible space
jescibe: ). - ).h is of order h 2 (k-m),
)rs in the sth derivative are of the
·oximation theory.
= f, the position is equally favorable.
·m uh(t, x) 2: Q/t)rp/x); the time
:ndence on x is discretized in terms of
, functions rp r The coefficients Q/t)
ary differential equations, expressing
SEC. 2.2. RATES OF CONVERGENCE 115
Galerkin 's rule: The residual u~ + Luh - f will not be. identically zero unless
the true solution u happens to líe in the trial space Sh, but its component in
Sh should vanish. Thus the original equation holds "on the subspace."
In practice, time m ust al so be made discrete. This is assumed to be done by -
a finite difference method. The Crank-Nicolson scheme, for example, is
centered at the time tn+ 112 when uh(t n+ 1) is computedJrom uh(tJand is therefore
accurate to order /lt 2 • Thus the final computed approximation includes this
error as well as the Galerkin error due to discretization in x. The latter is the
one which we analyze in detail, in order to show that for k > 2m, it, too, is
ofthe optimal order hk-s in the sth derivative. This result applies to equations
of parabolic type, for example the heat equation; L is the same kind of el-
liptic operator that occurs in static problems. In the case of hyperbolic equa-
tions, which include no dissipative term, the power of the finite element
method is somewhat diminished; the price to be paid in comparisqn with
explicit difference methods may be· too high. Nevertheless, even in this case
the comparatively automatic treatment of boundaries is a tremendous advan-
tage, and we include in Chapter 7 a sketch of the hyperbolic finite element
theory.
For want of space, the analysis of change of domain, numerical integra-
tion, and roundoff will be limited to the static case Lu = f The results for
eigenvalue and initial-value problems are certain to be very similar, and in
the case of quadrature errors these extensions of the theory have been car-
ried out by Fix (Baltimore Finite Element Symposium).
In the final chapter, we present the results of a rather extensive series of
numerical experiments. One of the most interesting concerns a problem with
strong singularities, produced by a crack which runs into the material. It
is a ciassical problem in fracture mechanics to compute the stress intensity
factor at the head of the crack; around that point the stress varíes like r- 112 •
Therefore, a number of questions arise:
l. Do our error estimates yield the observed rate of convergence, both
pointwise and in mean-square?
2. The singularity reduces the smoothness of u and therefore the rate of
convergence; does a reduced rate apply even where the solution is smooth,
so that the singularity pollutes the whole computation?
3. Is it possible, by grading the mesh or by introducing special trial func-
tions at the singularity, to recover the normal rate of convergence of the
method?
The answer in each case is "yes," and the numerical results are very con-
vincing. Both for static and for eigenvalue problems, an excellent remedy for
singularities which are introduced by sharp angles in the domain is the intro-
duction of trial functions which mirror the singularity correctly.
Straight interfaces between materials create a slightly different problem.
 (-
('
116 A SUMMARY OF THE THEORY CHAP.2 SEC. 2.3. GALERKIN'S METHOD, COLI

There is a jump in the derivatives of u across the interface, and we strongly at every point. The discrete form of
recommend the following simple solution: relax any continuity imposed on method, discussed below. At the othe1
the derivatives of the trial functions, so that uh is free to copy the singularity smooth functions which vanish in a b
in u. We do not believe that in normal circumstances the jump condition- parts · shifts all the deriva ti ves off u a
or any other natural boundary condition-should be imposed.
Finally, we have chosen to give the theoretical background for one com- (u, L *v) (f, v), L
paratively\new computational technique-the Peters-Wilkinson algorithm
In this weakest form, u satisfies the «:
for the ma:trix eigenvalue problem Kx lAfx. The solution of the linear sys-
tions." No doubt a discrete form hal
tem KQ = F is of course a still more fundamental problem, and it is subject
Between these extremes there lie a
to considerable refinement in the ordering of the unknowns or in the choice
the equation is said to hold in the stn
of a gradient procedure; but it is comparatively well understood. The eigen-
More generally the choice V = xs(
value problem is more subtle, and without an efficient algorithm the num-
shifted from u to v; if L is of order 2
ber of unknowns will be artificially limited-below the number required
be sought in JC 2 m-s. The cases m
to represent the physics of the problem. Therefore, we have described the
ready been given the notation
Peters-Wilkinson idea (as well as sorne more established algorithms) in
Chapter 6, and applied it to the numerical experiments of Chapter 8. a(u, v

In all these manipulations, the bom

2.3. GALERKIN'S METHOD, COLLOCATION, . With the strong form (s = 0), the full
ANO THE MIXED METHOD on u; the solution must lie in x;m. A
derivatives to qualify for the solutio
As we ha ve described it, the Ritz technique applies only to problems of the which make sense are those of order
classical variational type, in which a convex functional is to be minimized. the essential boundary conditions, t a
The corresponding Euler differential equation is self-adjoint and elliptic. number of conditions which are imiJ
It is well known, however, that equations of quite general type can also be emed by the deriva ti ves of order less 1
written in a weak form, and that this form suggests a generalization from the for s m, v too lies in JC']i. Thus the
Ritz to the Galerkin method. Applied to initial-value problems, this is the space and the test space.
subject ofChapter 7. Here we discuss two types ofstatic problems, first those Galerkin 's method is the obviou
in which derivatives of odd order spoil the self-adjointness of an elliptic general it involves two families of ft
equation, and then those in which the associated functional is not positive space and a subspace Vh of the test
definite-the problem is to find a stationary point rather than a minimum of is the element of Sh which satisfies
I(v). This arises in the Hellinger-Reissner principie in elasticity, and in the
corresponding mixed method for finite elements. It leads to sorne difficult (5)
mathematicat questions in the proof of convergence. t In matheroatical terros, an essential e<
First sorne comments on the weak forms of a differential equation Lu f linear operator B on thé space 3Cm of all fu
There are several, but they all share the following basic principie: The equa- which satisfy Bv O forro a closed subspat
tion is mu1tiplied by test functions v(x) and integrated o ver the domain n, to and cannot be, iroposed on every v; but 1
yield are satisfied by the roiniroizing u. The ust
involve only derivatives of order less thar
(Lu, v) = (/, v). sufficient. In a two-diroensional problero, ft
v O ata given single point P. The valut
This is to holp for each function v in sorne test space V, and everything hinges arbitrarily large, as with log log r, while th
tions which satisfy v(P) = O do not forro ~
on the choice of V. If V includes all o-functions, then the equation Lu = f close in strain energy to trial functions wl
will have to hold in the most classical (sorne would say old-fashioned) sense, thero would give the saroe resultas to roin
 CHAP.
i. across the interface, and?we strongly
twn: relax any continuity, 'imp<fSed on
o that uh is free to copithe sirlgularity
l circumstances the jump condition-
on-should be imposed.
: theoretical background for one com-
¡ue-the Peters-Wil~inson algorithm
A.Mx. The solution ·Úf the linear sys-
llndamental problem, and it is subject
ring of the unknowns or in the choice
>aratively well understood. The eigen-
thout an efficient algorithm the num-
:imited-below the number required
m. Therefore, we have described the
me more estabiished algorithms) in
!rical experiments of Chapter 8.
B.TION,
mique applies only to problems of the
mvex functional is to be minimized.
!quation is self-adjoint and eliiptic.
ms of quite general type can also be
~m suggests a generalization from the
to initial-value problems, this is the
'O types of static problems, first those
il the self-adjointness of an elliptic
associated functional is not positive
wry point rather than a minimum of
1er principie in elasticity, and in the
elements. It Ieads to sorne difficult
f convergence.
~ms of a differenti,al equaÚon Lu =f..
following basic principie: The equa-
md integrated over the domain n to .·
' are satisfied by the minimizing u. The usual test for essential conditions is that B must
(/, v).
e test space V, and everything hinges
"unctions, then the eqmttion Lu
)me would say old-fashioned) sense,
2 117
SEC. 2.3. GALERKIN'S METHOD, COLLOCA TION, AND THE MIXED METHOD
at every point. The discrete form of this test space leads to the collocation
method, discussed below. At the other extreme, V may contain only infinitely
smooth functions which vanish in a boundary strip. A formal integration by
parts shifts all the derivatives off u and onto v, leading to the equation
(u, L*v) (/, v), L* = formal adjoint of L.
In this weakest form, u satisfies the equation only "in the sense of distribu-
tions." No doubt a discrete form has been studied.
Between these extremes there lie a great many possibilities. If V= 3C 0 (0),
the equation is said to hold in the strong sense, or sometimes in the L 2 sense.
More generally the choice V xs(Q) permits S of the derivatives to be
shifted frorri u to v; if L is of order 2m, then presumably the solution u will
be sought in xzm-s. The cases= m is of prime importance, and it has al-
ready beert given the notation
a(u, v) = (/, v).
In all these manipulations, the boundary conditions must play their part.
With the strong form (s = 0), the full set of boundary conditions is imposed
on u; the solution must lie in Xim. As s increases and u needs only 2m s
derivatives to qualify for the solution space, the only boundary conditions
which make sense are those of order less than 2m - s; for s = m, these are
the essential boundary conditions,t and u lies in X'::. At the sarne time, the
number of conditions which are imposed on v is increasing. These are gov-
erned by the deriva ti ves of order less than s which appear in Green's formula;
for s = m, v too lies in 3C'I:. Thus the Ritz case is symmetric between the trial
space and the test space.
Galerkin's method is the obvious discretization of the weak form. In
general it involves two families of functions-a subspace Sh of the solution
space anda subspace Vh of the test space V. Then the Galerkin solution uh
is the element of Sh which satisfies
(5)
tin mathematical terms, an essential condition Bv =gis one specified by a bounded
linear operator B on the space Jem of all functions with finite strain energy. The functions
which satisfy Bv = O form a closed subspace. Natural conditions are those which are not,
and cannot be, imposed on every v; but because of the special form given to I(v), they
involve only derivatives of order less than m, but this condition is neither necessary nor
sufficient. In a two-dimensional problem, for example with m 1, we cannot impose that
v = O at a given single point P. The value at P is ·not a bounded functional (it may be
arbitrarily Iarge, as with Jog log r, while the function has unít strain energy) and the func-
tions which satisfy v(P) O do not form .. a closed subspace. In fact, they come arbitrarily
close in strain energy to trial functions which do not vanish at P, and to minimize over
them would give the same result as to' minimize over all of JC t.
 (
118 A SUMMA_RY ÓF THE THEORY CHAP.2 SEC. 2.3. GALERKIN'S METHOD, COLI

The left side will need s integrations by parts, as in the continuous problern. Problems which are not positive definí¡
When Sh and Vli are of equal dirnension N, this Galerkin equation goes We propose to analyze in more d1
into operational forrn in the usual way: lfqJ 1, • : . , qJNis a basis for Sh and Sh coincides with the test space Vh-tl
f// 1 , • • • , t¡JN is a basis for V\ the solution uh í: QiqJi satisfies weak form is essentially the a(u, v)
problern departs from the classiéal
(Í: QiLqJ i' f/1 k) = (/, f/1 k), k l, ... ,N. functional. The first departure arises
i~ GQ = as in the constant-coefficient example
1
In matrix form this F, with
Lu = -pu"

The adjoint operator L * will ha ve thi

The Ritz case is of course the one in which Sh Vh e 3Ce, qJi fJ1 i' and
(Lrpi, qJk) = a(rpi, rpk); G becomes the stiffness matrix K. tive of odd order:
Suppose that Sh is a finite elernentspace of degree k - 1, and Vh a finite
element space of degree 1 - l. Then the· expected rate of convergen ce of the J(-pu" + 1:u' + qu)v =

sth derivatives in the Galerkin method rnight be

This suggests an energy inner produc
(6)

As in (4), the first exponent of h reflects the best order of approxirnation

which is possible from Sh, and the second exponent is inftuenced also by
approxirnation in the test subspace and the order 2m ofthe differential equa- which itself is not syrnrnetric: a( u, v)
tion. In theory it appears possible to rnake sorne econornies by choosing adjoint; it corresponds to the imagin
1 < k, say k 4 (cubic splines) and 1 = 2 (linear test functions) in a second- part - pu" + qu is as positive as eve1
order problem ([B21]). The convergence rate is as good as with 1 = 4 and the product is extended to complex func1
bandwidth is reduced; however, Gis no longer syrnrnetric, even in self-adjoint
problems, and we are dubious. a(u, v) Spu'
A related possibility is to "lurnp" sorne terrns in the discretization, thereby
departing frorn the Ritz equations K Q = F and K Q A-M Q, by using ·
The real part of a(u, u) is f p l u'\ 2 + •
subspaces of lower degree on the lower-order terrns in the equation. In the
nary.
past this was useful in eigenvalue problerns, in order to replace Jhe con-
We want to show that the rate of
sistent mass matrix M by a diagonal lumped mass matrix. In the newest established as hk-t, justas before,
algorithms for the eigenvalue problern (Chapter 6) ít is no longer so impor-
method may be very poor. The argut
tant that the mass rnatrix be diagonal, and we shall refer to [T8] for an error
proof when the real part (self-adjoint
analysis of the lurnping process.
as a special case, but the Galerkin rr
In the collocation method, the test subspace Vh has a basis of J-functions:
odd-derivative term (the irnaginary
fJfix) J(x x). The Galerkin equation (5) then dernands that the differ-
size.
ential equation hold at each node: Lu(x) = f(x). In the error bound (6), the
method will norrnally actas if 1 =0 and converge at the rate !zk-zm. There are
THEOREM 2.1
special collocation points, however, which increase this order of convergence
and rnake the method very interesting [07, B24]. G has a srnaller bandwidth Suppose that 1a(u, v) \ < Kll u llmll'
than K, and there are no inner products or elernent matrices to compute; for is elliptic: Re a( u, u) O' !1 u ll~for u¡
complicated nonlinear problerns, these advantages may weii cornpensate for which satisfies the Ritz-Galerkin equ
the heavier approxirnation and srnoothness demands on Sh. Then the order of convergence in ener~
 CHAP.
>y parts, as i.n th,e con!inuouS'probl~m.
:nswn N, th1s Galerkm equation goes
1y: If 'l't• ~ .. , 'PN is a basis for Sh and
tion uh = L: Q1rp1 satisfies
rk), k= 1, ... , N.
which
.
Sh = Vh e xmE• 'f'm.=
J
tiF
'Y}'
and
;tiffness matrix K.
space of degree k - 1, and Vh a finite
1e expected rate of convergence of the
might be
cts the best order of approximation
:cond exponent is influenced also by
the order 2m of the differential equa-
mak.e sorne economies by choosing
= 2 (hnear test functions) in a second-
: rate is as good as with l = 4 and the
longer symmetric, even in se1f-adjoint
ne terms in the discretization, thereby
Q = F and KQ = A.MQ, by using ·
·-order terms in the equation. In the
•b1ems, in order to replace Jhe con-
lumped mass matrix. In the newest
:chapter 6) it is no Jonger so impor-
nd we shall refer to [T8] for an error
bspace Vh has a basis of J-functions :·
on (5) then demands that the differ-
) = f(x¡). In the error bound (6), the
converge at the rate !zk-zm. There are
h increase this order of convergence
>7, B24]. G has a smaller bandwidth
or element matrices to compute; for
dvantages may well compensate for
!ss demands on Sh.
2 SEC. 2.3. GALERKIN'S METHOD, COL LOCATION, AND THE MIXED METHOD 119
Problems which are not positive de.finite symmetric.
We propose to analyze in more detail two cases in which the trial space
S 11 coincides with the test space Vh-there are m integrations by parts, so the
weak form is essentially the a(u, v) = (/, v) of the Ritz method-but the
problem departs from the classical one of minimizing a positive-definite
functional. The first departure arises when the equation is not self-adjoint,
as in the constant-coefficient example
Lu = --pu" + 1:u' + qu = f
The adjoint operator L * will ha ve the opposite sign multiplying the deriva-
tive of odd order:
J(-pu" + 1:u' + qu)v = Ju( -pv"- tv' + qv).
This suggests an energy inner product
a(Zf, v) == Jpu'v' + 1:u'v + quv,
which itself is not symmetric: a(u, v) -=1=- a(v, u). In fact, the new term is skew-
adjoint; it corresponds to the imaginary part of the operator L, whose real
part -pu" + qu is as positive as ever. This appears most clearly if the inner
product is extended to complex functions,
a(u, v) = Jpu'v' + 1:u'v + quv.
The real part of a(u, u) is J p 1u'1 2 + q 1u 12 , and the new term is purely imagi-
nary.
We want to show that the rate of convergence in slope can be rigorously
established as hk-t, just as before, but that for large 1: the finite element
method may be very poor. The argument is quite general. The convergence
proof when the real part (self-adjoint part) is elliptic includes the Ritz method
as a special case, but the Galerkin method may be unsatisfactory when the
odd;.deri vative term (the imaginary or skew-adjoint part) is of significant
size.
THEOREM 2.1
Suppose that 1a( u, v) 1 < K11 u 11m11 v 11m and that the real part ofthe problem
is elliptic: Re a( u, u)> u 11 u 11~for u in X~. Let uh be thefunction in Sh e X~
which satisfies the Ritz-Galerkin equation a(uh, vh) = (f, vh) for al/ vh in Sh.
Then the order of convergence in energy (equivalently, the order of convergence
120 A SUMMARY OF THE THEORY CHAP.2 SEC. 2.3. GALERKIN'S METHOD, COl

of the mth derivatfVr_s) equals the best possible order of approximation which from x 2 + y 2 to xy, which has a sa•
can be achieved kY 8-h: · origin.
A good éxample is given by the e
(7) 11 U - uh ll m K min 11 u - vh 11m. a beam under loading. Suppose that
(JI SA
new unknown. Then the original eq1
For afini{e element space of degree k l, this order is hk-m. system of two second-order equatior

Proof Since a(u, v) (J, v) for all v in :rer;z, subtraction yields a(u - u\
(8)
vh) = O for all vh. Therefore,

(J 11 u uhll! Re a(u - u\ u uh)

This reduction of order brings sever
Rea(u u\ u) Re a(u u'¡, u
variational form need only be cont
fourth-order equation was associat
Dividing by the common factor 11 u uhllm' the proof is complete. Nitsche's
method could now be used to establish the usual rate of convergence (4) for
f (v")2 .- 2fv, and this is _finite only
thermore, the condition number of tl
the displacement.
by a reduction in the ordet ofthe diffe
In spite of this convergence, the Galerkin method may in practice be a are replaced by second differences,
bad choice. Suppose that Sh is the standard piecewise linear subspace. Then O(h- 4 ) to O(h- 2 ). This may seem mü
in our example the Galerkin equations for Q1 vh(x1) will be just the be achieved by a purely f9rmal intrc
difference equations unknown, but the improvement is q1
We examine this roundoff error in
beam, with u and M vanishing at b•
and w" M can be solved separatel
ther finite differences or finite elemen
tion of M" = f includes a roundoff e
h- 2 2-t if the computer word leQ.gth
Notice that the first derivative is replaced by a centered difference, regard- w" = -fwilllnclude first ofall its o"
less of the sign of -r. The true solution, however, may depend very and also an inherited error E 3 • This en
strongly on this sign; as 1-r 1--¡. <XJ, the advection term -ru' dominates the exactly the discretization of this equa
second derivative, and the solution u is of boundary-layer type. Over most of E 3 is also of order h- 2 • The roundoff,
the interval u is essentially the solution toan initial-value problem, for which As a second check, we shall compt
the centered difference is completely inappropriate; at the far end there is a system. In the finite difference case, t
rapid variation in order to satisfy the other boundary condition, and an
extremely fine mésh is required. The need for one-sided (upstream) differences
(9)
is well known in chemic'al engineering. Mathematically the dominance of -r is
reflected fn a larg~ value for K/u in the error estímate (7).
A second departure from the classical Ritz formulation arises when the and the eigenvalues p, of this block
potential-energy functional l(v) is not convex-a(u, u) is not positive defi- A. of 0 2 by
nite-and the task is to find a stationary point rather than a mínimum. This
occurs naturally in the mixed method, when both the displacement and its
derivatives &re taken as independent unknowns. The potential energy involves We know that the eigenvalues l of tl
products which are just as likely to be negative as positive; it is like going from 0(1) to O(h- 2). Solving the quad
 r
CHAP.2
possible order of approximation which
- 1, this order is hk-m.
l v in X'E, subtractioh.yields a(~
uh!lm, the proof is complete. Nitsche's
t the usual rafe of convergence (4) for
alerkin method may in practice be a
1dard piecewise linear subspace. Then
~ns for Qi = vk(x) will be just the
ced by a centered difference, regard-
!ution, however, may depend very
te advection term ru' dominates the
of boundary-layer type. Over most of
toan initial-value problem, for which
1ppropriate; at the far end there is a
: other boundary condition, and an
d for one-sided (upstream) differences
~athematically the dominance oiT is
error estímate (7).
;;al Ritz formulation arises when the
convex-a(u, u) is not positive defi-
y point rather than a mínimum. This
when both the displacement and its
nowns. The potential energy involves
negative as positive; it is like going
SEC. 2.3. GALERKIN'S METHOD, COLLOCATION, AND THE MIXED METHOD 121
from x 2 + y 2 to xy, which has a saddle point instead of a mínimum at the
origin.
A good example is given by the equation wuv> f(x), for the bending of
a beam under loading. Suppose that the moment M = w" is introduced as a
new unknown. Then the original equation becomes M" J, and we have a
system of two second-order equations:
(8)
This reduction of order brings several advantages. The trial functions in the
variational form need only be continuous between elements, whereas the
· fourth-order equation was associatttd with the potential energy l(v) =
f( v")2 - 2fv, and this is finite only if the slope v' is also continuous. Fur-
thermore, the condition number of the stiffness matrix is completely altered
by a reduction in the order ofthe differential equations; the fourth differences
are replaced by second differences, and the condition number goes from
O(h- 4 ) to O(h- 2 ). This may seem miniculous, that such an improvement can
be achieved by a purely formal introduction of the derivative M as a new
unknown, but the improvement is quite genuine.
We examine this roundoff error in two ways. Considera simply supported
beam, with u and M vanishing at both ends. Then the equations M" f
and w" = M can be solved separately, first for M and then for w, using ei-
ther finite differences or finite elements. Suppose that the approximate solu-
tion of M"= f includes a roundoff error E 1 , which will normally be of order
h- 2 2-:t if the computer word length is t. Then the approximate solution of
w" - f will in elude first of all its own roundoff error E2 , of this same order,
and al so an inherited error E 3 • This error satisfies E~' = E 1 , or rather it satisfies
exactly the discretization of this equation which is being used, and therefore
E 3 is also of order h- 2 • The roundoff error is not compounded to h- 4 •
As a second check, we shall compute the condition number of the discrete
system. In the finite difference case, the matrix form is
(9)
and the eigenvalues p of this block matrix are related to the eigenvalues
l of Ó2 by
We know that the eigenvalues l of the second difference operator Ó2 range
from 0(1) to O(h- 2 ). Solving the quadratic, the eigenvalues p fall in the same
122 A SUMMARY OF THE THEORY CHAP. 2 SEC. 2.3. GALERKIN'S METHOD, COLL

range, and the condition number llmax..l ftmin of the coupled system is indeed functional I(v) has no definite sign; th
of order h- 2 • and nota mínimum.
We wish to offer an explanation for this miracle. lt rests on the observa- Let us note the computational cons
tion that the usual measute of condition number for these matrices is un- piecewise linear elements and Nh = n
natural. We are regard_ing them for numerical purposes as transforming
Euclidean space (discrefe 3C 0 ) into itself, and therefore we use the same norm
for both the residual in the equation and the resulting error in the solution.
This is completely contrary to what is done in the differential problem, or
(Because M is unconstrained, it requ
for that matter in estimating discretization error; in such a case f is measured
are now 2N + 2 unknowns, but the s.
in 3C 0 , M and its error in 3C 2 , and w and its error in .:fe 4 • (In the variational
first for Mh and then for wh. With N
problem, these may be x-2, 3C 0 , and 3C 2 , respectively.) In fact, the operator
L = d 2 /dx 2 , with any of the usual boundary conditions, is perfectly condi-
tioned as a map from 3C 2 to 3C 0 • This was the essential point of Section 1.2,
that both L and L -t are .bounded. We can show that the same is true ofthe
difference operator Ó2 , and of any reasonable finite .element analogue, pro-
(
-G
D*
D)
O'
where G !!_
6
(~o '

vided these natural norms are retained. Therefore, there ought to exist an o
algorithm for solving KQ = F which reflects this property, and in this case
the miracle would disappear; the errÓrs in M and w would be appropriate to Gis the mass matrix (or Gram matrix)
their station. As long as a standard elimination is used, however, there will difference matrix. For large N the m
be a dífference in roundoff error between fourth-order and second-order z P q 2 , ••• to reduce the band width.
prob.lems. with G replaced by the identity, the 1
Before going further, the coupled differential equations of the mixed roots of p 2 + p = A, where A is an ei
method should be put into variational form. Multiplying the first equation matrix D* D. The errors in Mh and wh ·
in (8) by M and the second by w, and integrating by parts, will be accurate to order h. Of course tl
M h.
For two-dimensional problems, l
J (w"M + M"w M 2fw) dx 2
simple elements which are appropriah
catalogue these elements, however, v
- J (2w'M' + M + 2fw) dx + w'M + M'wl~·
2
mathematically the main point: the
when the functional is indefinite. The 1
In the simply supported case w and M vanish at each end-the admissible an infinite symmetric system Lu b · =
space satisfies full Dirichlet conditions-and the integrated term disappears. pose that Sh is a finite-dimensional sul
For the clamped beam a remarkable thing occurs: The condítion w = O is operator onto Sh: Phv is the compone
imposed at each end, but tbe vanishing of w' yields a natural boundary condi- (5) is identical with the problem of fin
tion for M. The stationary point of which satisfies

I(v) - f (2w'M' + M 2
+ 2fw),
(10)

The operator PhLPh ofthe discrete pr(

if the admissible space V contains all pairs (M, w) with M in the Neumann K, will be positive definite if L is; thü
space X 1 and w in the Dirichlet space Xó, solves exactly the problem of the even more positive than L itself; the n
clamped beam. Roughly speaking, if Mis unconstrained at the endpoints and problem is constrained toa subspace. ~
the first variation of 1 is to vanish, then the factor w' which multiplies M S\ then (PhLPhv, v) = (LPhv, Phv) =
must be zero. We emphasize that as in the Reissner principie of elasticity, fhe on the whole space: (Lv, v) > a(v, v) f

u!Jlmin of the coupled system is indeed
"or this miracle. It rests on the observa-
tion number for these matrices is un-
numerical purposes as transforming
~lf, and therefore we use the same norm
and the resulting ernb.r in the solution.
is done in the differential problem or
.tion error; in such a case f is meas~red
1nd its error in 3C 4 • (In the variational
3C 2 , respectively.) In fact, the operator
nmdary conditions, is perfectly condi-
¡ was the essential point of Section 1.2,
~can show that the same is true ofthe
Lsonable finite .element analogue,. pro-
!d. Therefore, there ought to exist an
reflects this property, and in this case
·s in M and w would be appropriate to
im.ination is used, however, there will
tween fourth-order and second-order
l differential equations of the mixed
l form. Multiplying the first equation
integrating by parts,
2fw) dx
2/l·v) dx + w'M + M'w l'ó·
f vanish at each end-the admissible
~and the integrated term disappears.
hmg occurs: The condition w = O is
of w' yields a natural boundary condi-
· + M + 2fw),
2 (10)
'airs (M, w) with M in the Neumann
fCó, sol ves exactly the problem of the
s unconstrained at the endpoints and
m the factor w' which multiplies M
1e Reissner principie of elasticity, the
SEC. 2.3. GALERKIN'S METHOD, COLLOCATION, AND THE MIXED METHOD 123
functional /(v) has no definite sign; the problem is one of a stationary point,
and not a mínimum.
Let us note the computational consequences ofthis variational form. With
piecewise linear elements and Nh n, Mh and wh are expressed by
N+l
Mh = L.; Z/fl/x),
o
(Because M is unconstrained, it requires two extra basis functions.) There
are now 2N + 2 .unknowns, but the system can no longer be solved in series,
first for Mh and then for wh. With N= 2 the coefficient matrix is
-G D) 1 2
-:-1 o)1
where G
( D* O' h -1 2.
(
o -1
Gis the mass matrix (or Gram matrix) ofthe rpi, aríd D is a rectangular second
difference matrix. Por large N the unknowns should be ordered by z 0 , q u
z 1, q2 , ••• to reduce the band width. The condition number is again O(h- 2 );
with G replaced by the identity, the eigenvalues are p, = 1, twice, and the
roots of p, 2 p, = A., where A. is an eigenvalue of the usual fourth difference
matrix D* D. The errors in Mh and wh will be of order h 2 , and their derivatives
will be accurate to order h. Of course the second derivative of wh will not equal
Mh. .
Por two-dimensional problems, HeÜan and Herrmann have developed
simple elements which are appropriate for the mixed method. Ratherthan to
catalogue these elements, however, we prefer to return instead to what is
mathematically the main point: the difficulty of establishing convergence
when the functional is in definí te. The problem is easiest to explain in terms of
an infinite symmetric system Lu == b of simultaneous linear equations. Sup-
pose that Sh is a finite-dimensional subspace and Ph the symmetric projection
operator onto Sh: Phv is the component of v in S h. Then the Galerkin method
(5) is identical with the problem of jinding the approximate solution uh in Sh
which satisfies
The operator PhLPh ofthe discrete problem, which is just our stiffness matrix
K, will be positive definite if Lis; this is the Ritz case. In fact, PhLPh will be
even more positive than L itself; the mínimum eigenvalue increases when the
problem is constrained toa subspace. This is easy to see: If vis in the subspace
S\ then (PhLPhv, v) = (LPhv, Phv) = (Lv, v). Suppose L is pósitive definitt~
on the whole space: (Lv, v) > a(v, v) for all v. Then this positive definiteness
124 A SUMMARY OF THE THEORY CHAP. 2
SEC. 2.3. GALERKIN'S METHOD, COI

is inher.ited by PhLPh on the subspace S\ and with no decrease in u.

If Lis symmetric but indefinite, then the same is to be expected of the
Galerkin operator PhLPh. (This assumes that the test space Vh and trial space
Sh are the same. Otherwise, if Qh is the projection onto V\ the Galerkin
L=
equation is QhLPhuh = Q"f, and QhLPh is not even symmetric.) It is natural
to hope tl).at with increasing S\ uh will approach u. This convergence is not
automatic,':··however, and in searching for the right hypotheses, we are led
back to the fundamental theorem of numerical analysis: Consistency and
stability imply convergence. is a good example; for every odd N, 1

last row is zero, and for f = (1, 1/2,

THEOREM 2.2 on the subspace Sh will not have a
Suppose that Galerkin' s method is McCarthy for his help with these e
(a)
consisten!: for every v, llv- Phv 11 ~O indefinite quadratic 2xy breaks dow
and Note that a reordering of the
(b) stable: the discrete operators are uniformly invertible, 11 (PhLPh)- 1 11 <
very el ose to the coupled system of 1
c.
convergence in that example, but tl
Then the method converges: 11 u - u" 11 ~ O.
choice of the subspáces Sh. Appare
Proof Denote (PhLPh)- 1 by R. Since Lu = J, we have yield such a choice. C. Johnson hasjt
two of the most important mixed ele
PhLPhu + PhL(u- Phu) = Phj,
tively constant and linear in each t
or property, after the displacement ur
moment unknowns are seen to be e
Subtracting uh = RPhJ,
for a positive definite expression (the
exactlythe familiar Ritz condition, t
are contained in the space of admiss
the prescribed load) in the full contil
Note that the rate of convergence depends both on the stabi!ity constant method, convergence depends on ap
C and on the approximation properties of S\ exactly as in Theorem 2.1. Convergence for mixed and hy
(That was effectively a special case of the present theorem, with C = 1/a and clearer, thanks to the work of Brezzi :
11 Lll = K. A more general result is given by Babuska [B4].) We could extend is that of stability.
the theory to show that consistency and stability are also necessary for con- To achieve numerical stability in
vergence, and that the existen ce of the solution u need not be assumed; Brow- matrix is indefinite, pivoting (row ar
der and Petryshyn ha ve shown how to deduce the invertibility of the original during the Gauss elimination proces
operator L. the predicted decrease in conditio11
We pólnt out again the special role of positive definiteness: It makes Connor and Will have reported son:
stability automatic. That is why the Ritz method is so safe. In the indefinite high degree: the mixed method has t
case, suppose that Sh is the subspace formed from the first N coordinate reduction of''order is. such a valuab
directions; Phv is given by the first N components of v. Then P"LPhis the Nth idea must continue.
principal minor of L-the ~ubmatrix in the upper left corner of L-and stabil- t We believe that for o.ther mixed (am
ity means that these N X N submatrices are uniformly invertible. It sounds verification of consistency (or approxim:
as if this might follow from the invertibility of the whole matrix L, but it theorem [6, B4] yields convergence. Brezzi
does not. The following invertible matrix and his technique needs to be generalized;
Miyoshi, Kikuchi, and Ando in Japan.
 )
.<mAP. 2
Sh, and with,no decrease in u.
en the same ikto ·be expected of the
s that the test space Vh and trial space
he projection onto V\ the Galerkin
is not even symmetric.) It is natural
1approach u. This convergence is not
for the right hypoth;€ses, we are led
numerical analysis: 'consistency and
uniformly invertible, 11 (PhLPh)-1
11 o.
e Lu we have
CIJLJJ!Iu- Pul! O.
!pends both on the stability constant
s of S\ exactly as in Theorem 2.1.
e present theorem, with C Ifu and
t by Babuska [B4].) We could extend
t stability are also necessary for con-
lution u need not be assumed; Brow-
educe the invertibility of the original
e of positive definiteness: It makes
z method is so safe. In the indefinite
:Ormed from the first N coordinate
nponents of v. Then PLP' is the Nth
1e upper left corner of L-and stabil-
; are uniformly invertible. It sounds
bility of the whole matrix L, but it
J{
SEC. 2.3. GALERKIN'S METHOD, COLLOCATION, AND THE MIXED METHOD 125
L=
is a good example;for every odd N, the leading principal minor is singular.lts
last row is zero, and for f = (1, 1/2, 1/4, ... ), the functional (Lv, v)- 2(/, v)
on the subspace Sh will not have a saddle (stationary) point. (We thank C.
McCarthy for his help with these questions.) Even in the 2 by 2 case, the
indefinite quadratic 2xy breaks down completely .on the subspace x = O.
Note that a reordering of the unknowns yields L (~ ~), which is
11 <
very el ose to the coupled system of the mixed method example. We claimed
convergence in that example, but this is actually justified only for a proper
choice of the subspaces Sh. Apparently the jinite element construction does
yield 1uch a choice. C. Johnson has just succeeded in proving convergence for
two of the most important mixed elements, in which the moments are respec-
tively constant and linear in each triangle. His proof establishes a special
property, after the displacement unknowns have been eliminated and the
moment unknowns are seen to be determined as the minimizing functions
for a positive definite expression (the complementary energy). This property is
exactlythe familiar Ritz condition, that the trial moments in the discrete case
are contained in the space of admissible moments (those in equilibrium with
the prescribed load) in the full continuous problem. Therefore, as in the Ritz
method, convergence depends on approximation theory and can be proved. t
Convergence for mixed and hybrid elements. is becoming very much
clearer, thanks to the work of Brezzi and others; the essential extra:hypothesis
is that of stabílity.
To achíeve numerical stability in the mixed method, where the coefficient
matrix is indefinite, pivoting (row and column exchanges) must be permitted
during the Gauss elimination process. Computational results ha ve borne out
the predicted decrease in condition number and roundoff error, although
Connor and Will have reported sorne unsatisfactory results with elements of
high degree: the mixed method has enjoyed only mixed success. Nevertheless
reduction of order is such a valuable property that the development of the
idea must continue.
t We bdieve that for o.ther mixed (and hybrid) elements, the natural proof ís a direct
verífication of consistency (or approximability) and stability. Then Babuska's general
theorem [6, B4] yields convergence. Brezzi has established stability for one hybrid element,
and hís technique needs to be generalized; these methods have recently been discussed by
Miyoshi, K.ikuchi, andAndo in Japan.
126 A SUMMARY OF T~ THEORY CHAP.2 SEC. 2.4. SYSTEMS OF EQUATIONS; SHE
(
There are also a number of curve
2.4. S.:YSTEMS OF EOUATIONS; ·sHELL variables, to which our analysis can 1
PROBLEMS; VARIATIONS ON THE involve derivatives of different orders.
FINITE ELEMENT METHOD basis of a fully compatible shell theor:
parametrically by a set of three equa1
It may, be objected that our finite element theory deals always with a dent variables in tpe shell equations ar1
single unkriown, while most applications involv~ a system of r equations for standard plane domain Q. The curvat
a vector uriknown u (u~' ... , ur). Fortunately, the distinction is often not the deriva ti ves of the X¡, which en ter a:
essentia1. The variational principie for a system again minimizes a quadratic equations. Numerical integration will
I(v), and the error estimates depend exactly as before on the approximation evaluation of the element stiffness r
properties of Sh. essential respect like any other proble
A typical example is furnished by two- or three-dimensional elasticity. The have constructed a conforming and
unknowns at each point are the displacements in the coordinate directions, CURSHL [C12], by combining the ~
and the fit:tite element solution íi!' is again the trial function closest to the true. cubics of Section l. 9 for the normal at
solution in the,sense of strain energy-which is now a quadratic function Since the difference k -m equals 3 f<
involving all the unknowns. The approximation theory in the next chapter gen ce in strain energy will be h 6 •
will,show that, independent of the number of dimensio_ns and unknowns, the Such a construction, whose accura~
rate of convergence depends,on the degree of -the finite elements and on the periments, ought in theory to end the
approximation to the domain. has not. For most applications and rr
Sorne new and much more difficult problems arise for shells. The theory quite complicated; it has not been wi
is normally constructed as a limiting case of three-dimensional elasticity, in are many problems with special symn
which the domain n becomes very thin in one direction (normal to the shell elements can be used [04]. On tl
surface of the shell). The result of the limiting process is to introduce second mate a general shell by an assembla
deriva ti ves of the transverse displacement w into the potential-energy func- becomes effectively a simple plate ele
tional; the differential equations are offourth order in w, and ofsecond order deformations are coupled only by the
in the in-plane displacements. This imbalance is the price to be paid for the variant of this approach-which ,inevi
reduction from a three-dimensional toa two-dimensional problem. from the viewpoint of pure shell theo
Let us note immediately the increasing popularity of three-dimensional plest and most practica! way to deal 1
elements, in which this reduction· is not made. It is by no means automatic, We would very much like to ar
when a limiting process simplifies the search for exact solutions to problems gence of these combinations of ftat ¡:
with special symmetries-as it certainly does in shell theory-that the same that under reasonable conditions the
process will simplify the numerical solution of more general problems. (The as a geodesic dome) approaches that
same question arises with the Airy stress functions of plate bending; is it unable to verify that conjecture. The
numerically sensible to reduce the nufuber of unknowns and increase the extremely interesting.
order of the equations? We do,ubt it.) Obviously a thin shell will never pre- In the one-dimensional case, wht
sent a typlcal three-dimensional problem; there will always be difficulties gonal frame, we can be more explicít.
with a nearly degenerate domain. Not only the isoparametric technique but of bending and extensional energies,
also special choices of nodal unknowns and of reduced integration formulas
in the normal direction are being tried experimentally. From a theoretical
viewpoint, it will beco me necessary to estímate the effect of a small thickness
a(v, v) J
=e (~~ + ~2
parameter t (Fried has done so, in respect of numerical stability and the con-
dition number), but in general the approximation-theoretic approach can The radius of curvature is r, the thickr
2
proceed in the usual way. are that v 1 and V 2 lie in :JC 1 and :JC '
 CHAP.2 SEC. 2.4. SYSTEMS OF EQUATIONS; SHELLS; FORCE A.ND HYBRID METHODS 127

There are also a number of curved shell elements, in two independent

Ell
variables, to which our analysis can be directly applied-even though they
"HE
involve derivatives of different orders. These elements are constructed on the
basis of a fully compatible shell theory: The surface of the shell is ~escribed
', parametrically by a set of three equations X¡= X¡((}¡, _0 2), _and the mdepen-
~ element theory deal~ ~lways \vith a
dent variables in the shell equations are(} 1 and 0 2-whtch _stmply vary o ver a
ns involvf? a system clf r equations for
standard plane domain Q. The curvature of the shell appe~rs onl~ throu?h
'ortunately, the distinction is often not
the derivatives ofthe x¡, which enteras variable coefficients m the dtfferenttal
a system again minimizes a quadratic
equations. Numerical integration will almost certainly be requi~ed_ for the
mctly as before on the approximation
evaluation of the element stiffness matrices, but the problem Is m every
essential respect like any other problem in the plane. Cowper and Lindberg
'0- or three-dimensional elasticity. The
ha ve cÓnstructed a conforming and highly accurate element .known as
1cements in the coordinate directions,
CURSHL [C12], by combining the e 1 reduced quintics and eo com_pleté
Lin the trial function closest to the true
cubics of Section 1.9 for the normal and in-plane displacements, respecttvely.
-which is now a quadratic function
Since the difference k -m ~quals 3 for all components, the rate of conver-
"Oximation theory in the next chapter
gence in strain energy will be h 6 • •
.ber of dimensio_ns and unknowns, the
Such a construction, whose accuracy has been confirmed by numencal ex-
~ree of the finlte elements and on the
periments, ought in theory to end the search for good shell elements-but it
has not. For most applications and most programs, CURSHL appears to be
problems arise for shells. The theory
quite complicated; it has not been widely accepted. On the one hand, there
ase of three-dimensional elasticity, in
are many problems with special symmetries, in which cylindrical or shallow
hin in one direction (normal to the
shell elements can be used [04]. On the other hand, it is possible to approxi-
imiting process is to introduce second
mate a general shell by an assemblage of flat pieces. E~ch of these pie~es
tent w into the potential-energy func-
becomes effectively a simple plate element, and the bendmg and stretchmg
fourth order in w, and of second order
deformations are coupled only by the fitting together of these plates. Sorne
mlance is the price to be paid for the
variant of this approach-which inevitably involves nonconforming elements
a two-dimensional problem.
from the viewpoint of pure shell theory.-:.appears to be surviving as the sim-
sing popularity of three-dimensional
plest and most practica! way to deal with complicated shells.
t made. It is by no means automatic,
We would very much Jike to analyze the convergence or nonconver-
:!arch for exact solutions to problems
gence ~f these combinations of flat plate elements. It has been conjectured
r does in shell theory-that the same
that under reasonable conditions the deformation of a polyhedral shell (such
1tion of more general problems. (The
as a geodesic dome) approaches that of a genuinely curved shell, but we are
·ess functions of plate bending; is it
unable to verify that conjecture. The mathematical problems are novel and
mber of unknowns and increase the
extremely interesting. ;
)bviously a thin shell will never pre-
In the one-dimensional case, when a curved arch is replaced by a poly-
em; there will always be difficulties
gonal frame, we can be more explicit. The strain energy f?r the arch is a s~m
Jnly the isoparametric technique but
of bending and ext~nsional energies, and can be normahzed to
and of reduced integration formulas
l experimentally. From a theoretical
;tima te the effect of a small thickness a(v, v) C J(dv 1
ds
+ Vr
2)
2
+ f_(_l dV¡
12 r ds
2
_ d Vz)
ds 2
2

;::t of numerical stability and the con-

proximation-theoretic approach can The radius of curvature is r, the thickness is t, and the admissibility conditions
are that vt and 112 líe in 3<: 1 and 3<: 2 , respectively; the first derivative avzfas
128 A SUMMARY OF THE TIÍEORY CHAP. 2 SEC. 2.4. SYSTEMS OF EQUATIONS; SHE

is effectivel:y,_ the rotation of the vector normal to the arch, and this is not tions on the shell) are taken to be ir
perJllitted to have a jump discontinuity. The pair v 1 = u and v2 w, which relationship to the normal displacerm
minimlze the potential energy [I(v) = a(v, v) +linear terms], are the tangen- the nodes, not over the entire surface.
tial and normal displacements ofthe arch. Note that they are determined by tinue to be practica} and valuable; ther,
differential equations, the Euler equations of l(v), which are of different and too many alternative elements, to
orders. . Even conforming elements such as Cl
For ah' approximating frame, made up of straight lines, the radius of the ground that rigid body motions are
curvature becomes r oo. This uncouples the two components of in the v is that a tall structure leaning in the wi
strain energy, and therefore Ieads to simple and independent differential as a consequence such large finite eler
equations for u and w. The coupling reappea¡s, however, at the joints of the bending, which is of prime importan<
frame, in order to prevent the structure from coming apart in its deformed curacy: and to be concerned rather '
v
state. In other words, each trial function = (v 0 v 2 ), as s s 0 is approached elements.
from the left and from the right, must move the joint to the same place and Obviously shells have provided fin
thereby leave the frame continuous; this represents two conditions relating alllinear problems perhaps the most
both components of v_ and v+ to the joint angle B. (The conditions are essen- matical analysis. But these are probl
tial, not natural.) Furthermore, there is an essential continuity condition by earlier techniques, and we believt
on the rotation av 2 jas, preserving the angle at each joint. (Note that v 2 itself under8tood.
need not be continuous; it will include t5-functions at the joints, whose In addition to systems of equatio1
derivatives from left and right are both zero.) the finite element analysis described S<
With finite elements these conditions constrain the polynomials meeting displacement method, in which an ex]
at the nodes. Typically v 2 is a modified Hermite cubic: the continuity of the placement, and the optimal coefficien1
rotation is left unchanged, but discontinuities in function value are allowed, kin principie. The numerical effective
and related to those in v 1 • Obviously such trial functions are inadmissible in and this is largely a period of perfe<
the arch problem, but since the strain energy is also modified by the removal input and output-rapid developme1
of r, the convergence question remains open. Direct evidence against con- quired to make the output intelligibl€
vergence was provided by Walz, Fulton, and Cyrus (Wright-Patterson Con- fections which can be removed only 1
ference Il) for the case of a circular arch and a regular polygon. The finite itself. In the remainder of this sectiOJ
element equations are simply difference equations, and they were found to posed changes, still remaining in tht
be consistent with the wrong differential equation. The leading terms were cor- residuals. This means that the apprm
rect-the radius of curvature reappears through the agency of the angle B tion of trial functions, but the rule 1
in the frame continuity conditions-but there were, for the particular ele- cients q.1 may be different, or the un
ment chosen, also unwanted terms of zero order in h. t This suggests the pos- The most Ímportant variant is the
sibility of a convergence condition resembling the patch test of Section 4.2. derivatives of u, rather than u itself-
For a shell the problems are inevitably more complex. Even with a poly- lems these ~re the quantities of gn
hedron bujlt from flat plates~ the continuity requirements are difficult to im- approximate them directly. The resul
pose on polynomial elements; continuity of the deriva ti ves is always difficult od, in which both displacements and s
to achieve with anything less than quintics. One useful but mathematically is indefinite. Here tlíere is a compll
cloudy approximation has been the discrete Kirchhoff hypothesis: The rota- mized, and apart from a change in
tions (the rate of change of normal displacement in the two coordina te direc- functions ha ve new interelement and
is the same. The order of approxirr
tThese unw~ted terms are, however, comparable to others which appear in relating still decisive.
the arch equations to a full two-dimensional elasticity theory. Thus the leading terms in As an example we take Laplace':
all theories, including the frame approximation, may coincide-and the same may apply
to shells. Au =O in !l,

CHAP.2
) SEC.
(../
tor normal to the arch, and this is not
ty. The pair vt = u and v = w which
a(v, v) +linear terms], ar~ the ~angen-
ar~h. Note that they are determined by
tatiOns of I(v), which are o?-mtrerent
de up of straight litÚ~s, the radius of
mp~es the two components of v
in the
J stmple and independent differential
·eappears, however, at the joints of the
lre from coming apart in its deformed
v
on = (v¡, v2), as s S 0 is approached
t r:'ove the joint to the same place and
:~ts represents two conditions relating
n~t angle B. (The conditions are essen-
~ IS an essential continuity condition
:~ngle at eac~ joi,nt. (Note that v2 ítself
ude ó-functiOns at the joints, whose
1 zero.)
•ns cons.train t?e polynomials meeting
~ H~~mt~e cubtc: the continuity of the
mmties m function value are allowed
uch trial functions are inadmissible ¡~
mergy is also modified by the removal
lS open. Direct evidence against con-
l, and Cyrus (Wright-Patterson Con-
·ch and a regular polygon. The finite
:e equations, and they were found to
equation. The leading terms were cor-
·s through the agency of the angle (}
ut there were, for the particular ele-
ero order in h. t This suggests the pos-
lbling the patch test of Section 4.2.
Jl~ more ~omplex. Even with a poly-
tmty reqUirements are difficult to im-
is
t~ of the derivatives always difficult
ttcs. O~e useful but mathematically
crete Ktrchhoff hypothesis: The rota-
Iacement in the two coordina te direc-
par~b.Ie to others which appear in relating
:lastJcity t?e~ry. Thus the leading terms in
1, may comc1de-and the same may apply
1
2.4. SYSTEMS OF EQUATIONS; SHELLS; FORCE AND HYBRID METHODS 129
.
tions on the shell) are taken to be independent unknowns, and their true
relationship to the normal displacement is imposed only as a constraint at
the nodes, not over the entire surface. Approximations of this kind will con-
tinue to be practica! and valuable; there are too many kinds of shell problems,
and too many alternative elements, to decide that a single approach is best.
Even conforming elements such as CURSHL have had their detractors, on
the ground that rigidbody motions are not reproduced exactly, Theargument
is that a tall st.ructure leaning in the wind might have such large motions, and
as a consequence such large finite element errors, as to obscure the interna]
bending, which is of prime importance. We are inclined to trust in the ac-
curacy, and to be concern,ed rather about the convenience, of high-degree
elements.
Obviously shells have provided finite elements with a very severe test-of
all linear problems perhaps the most severe. The same is true of the mathe-
matical analysis. But these are problems· which were virtually inaccessible
by earlier techniques, and we believe they are now on their way to being
understood.
In addition to systems of equations, there is another major omission in
the finite element analysis described so far; we have concentrated only on the
displacement method, in which an expression I; qj(/Jj is assumed for the dis-
p~acen;ten~, and the optimal coefficients Qj are determined by the Ritz-Galer-
km prmctple. The numerical effectiveness of this method is well established
and this is largely a period of perfecting the elements and automating th~
input and output-rapid developments in computer graphics have been re-
quired to make the output intelligible. there remain, however, sorne imper-
fections which can be removed only by modifying the mathematical method
itself. In the remainder of this section we want to describe sorne of the pro-
posed changes, still remaining in the framework of the method of weighted
residuals. This means that the approximate solution will still be a combina-
tion of trial functions, but the rule for choosing among the possible coeffi-
cients qj may .be different, or the unknown itself may be altered.
The most important variant is theforce method, in which the strains-the
derivatives of u, rather than u itself-are taken as unknowns. In many prob-
lems these are the quantities of greatest importance, and it is natural to
approximate them directly. The result is quite different from the mixed meth-
od, in which both displacements and strains are unknowns, and the functional
is indefinite. Here there is a complementary energy functional to be inini-
mized, and apart from a change in the admissibility conditions-the trial
functions ha ve new interelement and boundary constraints-the mathematics
is the same. The order of approximation achieved by the trial subspace is
still decisive.
As an example we take Laplace's equation
Jiu o in n, u= g on r,
 .,
130 A SUMMARY OF THE THEORY CHAP.2 SEC. 2.4. SYSTEMS OF EQUATIONS; Sl

in which u is-,~termined by minimizing /(v) = JJ v; + v;. In the comple- the tangential component of the g1
mentary~-energy principie these derivatives v x = f 1 and vy = f 2 are taken as component of (f" f 2 ), must be contiJ
the primary dependent variables. The pair (f ~' f 2 ) is no longer constrained the greatest dijficulty in approximatinJ
to be the gradient of sorne v; in other words the cross-derivative identity rate functions f 1 and f 2 need not be
(f 1) 7 (f 2)x, or compatibility condition, is no longer imposed. (It follows, of ofthe vector (f 1, f 2 ) must be.
course, that when approximations f~ and f; are determined, there is no It is important to establish this
unique .Jt~y ofintegrating to find a corresponding uh. In the exact application condition (11), without arguing by w
of the complementary principie the optimal f 1 and f 2 will be the deriva ti ves of general problem the idea of a strea
the true displacement u, but in the discrete approximation this link between have to attach sorne meaning to eq
gradient and displacement is lost.) and f2 lie only in creo. They may ha'
In the Laplace example, the trial functions for f 1 and f 2 need only to líe dition must be understood in a weak
in creo for the quadratic functional to be finite. There is a constraint ofequilib- z which vanishes on r' and integrate
rium, however, provided by the differential equation itself: uxx + uyy O given by this weak form of equatior
leads to

(11)

This actually means that the pair (f 2 , - f 1) will be the gradient of sorne func- For a tria! space of piecewise polyn
tion w: wx = f 2 and W7 - f 1 • We may interpret the complementary process to this form of the constraint, takii
in this special example as follows: Instead of Iooking for the harmonic func- each element the original condition j
tion u, we are computing the associated stream function w. The function w is boundary, there must be a canceilaü
conjugate to u, meaning that u + iw is analytic. arising from the two adjacent regio1
This leaves the two principies very nearly parallel in the interior of O. On requires the continuity of the norm~
the boundary, however, there is a marked distinction; the Dirichlet condition The literature on the complerr.
on.u is exchanged for a Neumann condition on w. To see this, recall that along Trefftz [T9}, who established that it
any curve, u and its stream function are related by w, = -us. (This follows gral. Friedrichs [F19} then dis~ove1
from the Cauchy-Riemann equation wx = -u7 when the curve is vertical, underlying idea was exactly the Ritz
from W7 = ux when it is horizontal, and from a linear combination in the and that boundary conditions whi<
general case.) Therefore, on r the essential condition u g .is replaced by natural in its conjugate. The ortho¡
w, = -gs- As in any inhomogeneous Neumann problem, this means that the two problems was developed by Syr
admissible space is not constrained at the boundary, but the functional lis a fundamental paper with Prager [P~
changed to gate principies leads to upper as we
and also for the displacement u; ar:
Burchard [AJO}, and the idea is ap
1' = JJ(w; + w;) dx dy + Jgsw ds. berger [W2}. It has been programme
The finite element method was f
Integrating the last term by parts to obtain gws ds, w can be eliminated complementary principie, by de Veu
in favor of its derivatives wx = f 2 and W7 - f 1 • The complementary prin-
literature. t
cipie now asserts that the pair (f ~' € 2 ), which minimizes /', is the gradient of Always the interelement conditi
the displacement u which minimizes /. has led to the construction of mult
In finite element approximation the trial functions for w must lie in X 1 ,
in order for 1' to be finite. What does this imply about the variables f 1 ~nd f 2 ? tThe analysis of the force method pr<
First, since the tria! functions for w are continuous across inter-elem.ent same approximation theorems, tbat are a
boundaries, so are their derivatives in the direction along the edge. Therefore, There is not space to develop tbis whole
 CHAf.
~
~i~g I(v) = JJ v; + v;. In the co~ple-
ttives vx = E 1 and vy E2 are taken as
~ pair (E 1, E2) is no longer constrained
1er words the cross-derivative idenilty
n, is no longer imposed. (It fóllows, of
~ and E1 are determ,ined, thére is no
respcnding uh. In the'.exact application
timal E 1 and E2 will be the deriva ti ves of
:crete approximation this link between
~nctions for E 1 and E2 need only to líe
e finite. There is a constraint of equilib-
·entiai equation itself·• u XX + uYY o given _by this weak form of equation (11):
-E 1) will be the gradient of sorne func-
t interpret the complementary process
!ad of looking for the harmonic func-
~ stream function w. The function w is
analytic.
Iearly parallel in the interior of n. On
~d distinction; the Dirichlet condition
ion on w. To see this, recall that along
: related by wn = -u•. (This follows
'x -uy when the curve is vertical
1d from a linear ~ombination in th~
ntial condition u= gis replaced by
!umann problem, this means that the
:he boundary, but the functional lis
x dy + Jg.w ds.
ain - J gw. ds,
w can be eliminated
'y The complementary prin-
-E 1·
hich minimizes /', is the gradient of
:rial functions for w must líe in X 1
imply about the Variables E 1 and E2 ? tThe analysis of the force method proceeds from the same principies, and in fact the .
re continuous across inter-element
direction along the edge. Therefore,
2 SEC. 2.4. SYS'fEMS OF EQUATIONS; SHELLS; FORCE AND HYBRID METHODS 131
the tangential component of the gradient (E 2 , -E 1), which is the normal
component of(E 1 , E2 ), must be continuous. lt is this constraint which imposes
the greatest dijficulty in approximating the complementary principie. The·sepa-
rate functions E 1 and E 2 need not be continuous, but the normal component
of the vector (E 1 , E2 ) must be.
It is important to establish this constraint directly from the equilibrium
condition (11), without arguing by way of a stream function, since in a more
general problem the idea of a stream function is not so relevant. First we
have to attach sorne meaning to equation (11), (E 1)x + (E 2 )y =O, when E1
and E2 líe only in ;reo. They may ha ve no derivatives, and therefore the con:-
dition must be understood in a weak sense; we multiply by a smooth function·
z which vanishes on r, and integrate by parts. The proper constraint is then
ff (E 1 ~~ + E2 ~~) dx dy 0 for all such z.
For a trial space of piecewise polynomials, Green's theorem can be applied
to this form of the constraint, taking it one element at a time. Then inside
each element the original condition (11) must hold, and on eagh interelement
boundary, there must be a cancellation ofthe boundary integrals or tractions,
arising from the two adjacent regions. It is precisely this cance1Iation which
requires the continuity of the normal compdnent of (E 1 , E2 ).
The' literature on the complementary method apparently begins with
Trefftz [T9], who established that it yields a lower bound to the energy inte-
gral.. Friedrichs [F19] then discovered that as in the Laplace example, the
underlying idea was exactly the Ritz method applied to a conjugate problem·,
and that boundary conditions which are essential in one problem become
natural in its conjugate. The orthogonality of the admissible spaces for the
two problems was developed by Synge [17] into the hypercircle method, after
a fundamental paper with Prager [P9]. A combination ofthe direct and conju-
gate principies leads to upper as well as lower bounds for the strain energy,
and also for the displacement u; an abstract account is given by Aubin and
Burchard ·[AIO], and the idea is applied to sorne model problems by Weitf-
berger [W2]. It has been programmed by Mote and Y oung.
The finite element method was first applied to the stresses, that is, to the
complementary principie, by de Veubeke [F11]. By now there is an enorrrious
literature. t '
Always the interelement conditions present practica! dif_licuties, and this
has led to the construction of multiplier and hybrid methods. For example,
same approximation theorems, that are applied in this book to the displacement method.
There is not space to develop this whole theory in parallel.
 132 A SUMMARY OF THE"THEORY CHAP. 2 SEC. 2.4. SYSTEMS OF EQUATIONS; SHEI

Andetheggen [A3]rfílapplying to a fourth-order plate problem the standard The Ritz method now minimizes ÜJ
Ritz method qf minimizing the potential energy functional I(v), proposes term, over a space Sh with no bounda
to use cubic polynomials for which the normal slope will ordinarily be dis- piecewise polynomial space, containin.
continuous between elements. Imposing the constraint of slope continuity k - l. Then there will be a balance l
associates with each element edge a Lagrange multiplier and alters the Ritz unsatisfied boundary condition and
method;i,p.to a constrained minimization. The required programming changes involved in minimizing only over a sul
are very· simple. The stiffness matrix becomes indefinite, however, and (be- to determine an optimal dependence o
cause of the edge unknowns) the computation time for cubics seems to be Babuska has also given rigorous err
comparable to the usual stiffness method for the reduced quintics proposed Lagrange multipliers, in which again th
in Section 1.9. boundary. For Poisson's equation, th
A second interesting modification ís the hybrid method pioneered by Pian point (u(x, y), 1(s)) of the indefinite fu
and Tong [P4, P5]. This copes with the interelement difficulties in an ingenious
way, by constructing a family of approximations both to the stress field F(v, A)~ So (v; + 2fv
within each element and to the displacement on the element boundaries. The
stress fields satisfy the differential equation within each element (so that the
The Lagrange multiplier runs over al11
homogeneous case/= Ois by far the simplest) and the displacements-which
at the true stationary point it is relate
are given by an independent set of piecewise polynomials-are continuous.
error in the stationary point (uh, A,h) o
For each displacement pattern the complementary energy is first minimized
estimated [B6].
within the elements separately. This leads to a family of displacements v\
The essential boundary conditions
defined now not only on the element boundaries but throughout the region
quite different approach, one with a lo1
n, to which the Ritz method can be applied; the final hybrid approximation Instead of minimizing I(v) = (Lv, v)
is foi.md by minimizing over these vh. The energy generally Iies between the
residualll Lw f W. lgnoring boundar
lower bound provided by the pure Ritz method and the upper bound of its
lying functional is
conjugate, and in many cases will yield a substantially better approximation
than either. llLw- f UZ (Lw, Lw:
We return once more to the basic Ritz method and reconsider the follow-
= (L*Lw,"
ing question: Can the variational principie be altered in such a way that
essential boundary conditions need not be imposed? The complementary The last term is independent of w, a1
principie is one possible ánswer, but there are others. In fact, there is now a actually minimizes 1' = (L*Lw, w)- 2(
standard device for dealing with unsatisfied constraints: to insert a penalty tional for the problem L *Lu L *f. Tl
.function into the expression to be minimized. (This was the main theme of adjoint, but we note that the order of t
Courant's remarkable lecture [C 11]; the finite element method was an after- With boundary conditions, the lea
thought!) For -llu f, with u g on r, l(v) is altered to lyzed for the first time in the recent w<
Taking as an example the Dirichlet p
Jh(v) SScv; + v; - 2fv) + eh fr cv - g)2 ds. th~y introduce the functional
o
The exact mínimum over the_unconstrained admissible space X 1
is achieved (12) I"(w) SJ(llw + /) 2
dx
by the fu'nction U\ which satisfies
tWe believe this method to be a prorr:
-C¡; 1 V! + Uh =o on r. tnethod, because it attacks directly the prob
normal derivative) at the boundary. Often it
· Therefore, if eh~ oo, these solutions Uh converge to the solution u, which and to find it from an approximate solutiot
vanishes on r. not entirely satisfactory.
 CHAP.2
fourth-order plate problem the startdard
•tential energy furictional l(v), proposes
the normal slope will ordinarily be dis-
osing the constraint of slope continuíty
Lagrange multiplier and alters the Ritz
:ion. The required programming,clianges
x b~comes indefinit~}.however, and (be-
>mputation time for cubics seems to be
thod for the reduced quintics proposed
t is the hybrid method pioneered by Pian
e interelement difficulties in an ingenious
.pproximations both to the stress field
tcement on the element boundaries. The
¡uation within each element (so that the
simplest) and the displacements-which
>Íecewise polynomials-are continuous.
Jmplementary energy is first minimized
Jeads. to a family of displacements vh,
: boundaries but throughout the region
tpplied; the final hybrid approximation
. The energy generaHy Jies between the
tz method and the upper bound of its
:Id a substantiaUy better approximation
Ritz method and reconsider the follow-
·inciple be altered in such a way that
not be imposed? The complementary
there are others. In fact, there is now a
atisfied constraints: to insert a penalty
nimized. (This was the main theme of
he finite element method was an after-
on r, l(v) is altered to
~fv) + eh fr (v g) 2 ds.
rained admissible space X 1 is achieved
r~ + uh =o on r.
; Uh converge to the solution u, which
SEC. 2.4. SYSTEMS OF EQUATIONS; SHELLS; FORCE AND HYBRID METHODS 133
The Ritz method now minimizes the functionalJ\ including the penalty
term, over a space Sh with no boundary constraints. We suppose it to be a
piecewise polynomial space, containing the complete polynomial of degree
k - l. Then there will be a balance between the error u U\ due to the
unsatisfied boundary condition and the penalty,· and the error Uh uh
involved in minimizing only over a subspace. This balance led Babuska [B5]
to determine an optimal dependence of eh on h:eh = ch 1 -k.
Babuska has also given rigorous error estimates for the related method of
Lagrange multipliers, in which again the trial functions are unrestricted at the
boundary. For Poisson's equation, this method searches for a stationary
point (u(x, y), l(s)) ofthe indefinite functional
F(v, A)= In (v; + v; 2fv)dx dy 2 Ir A(v- g)ds. ·
The Lagrange multiplier runs o ver all admissible functions defined on r, and
at the true stationary point it is related to the solution by 1 = au¡an.t The
error in the stationary point (uh, lh) over a finite element subspace can be
estimated [B6].
The essential boundary conditions can also be made to fall away by a
quite different approach, one with a long history: the method of least squares.
Instead of minimizing J(v) (Lv, v) 2(f, v), this method minimizes the
residuali!Lw- f 11 2 • Ignoring bouridary conditions for a moment, the under-
lying functional is
ll Lw f 11 2 (Lw, Lw) ~ 2(f, Lw) + (f,f)
= (L*Lw, w)- 2(L* f, w) + (f,f).
The last term is independent of w, and therefore the least-squares method
actual/y minimizes 1' = (L*Lw, w) 2(L*J, w), which is exactly the Ritzfunc-
tiona/ for the problem L *Lu L *f This new problem is automatically self-
adjoint, but we note that the order of the equation has doubled.
With boundary conditions, the least-squares method was properly ana-
Iyzed for the first time in the recent work of Bramble and Schatz [B30, B31].
Taking asan example the Dirichlet problem -Au =fin O, u g on r,
they introduce the functional
(12) I"(w) = JJ(Aw + f) 2
dx dy + ch- 3
J (wr g) 2 ds.
tWe believe this method to be a promising variant of the standard displacement
inethod, because it attacks directly the problem of computing the solution (or rather its
normal derivátive) at the boundary. Often it is exactly this information which is desired,
and to find it from an approximate solution in the interior is numerically unstable and
not entirely satisfactory.
134 A SUMMARY OF THE THEORY CHAP.2 SEC. 2.4. SYSTEMS OF EQUATIONS; SHE

The factor h- 3 is not related to the degree k of the subspace S", as it was in again the serious difficulties are at the l
the penalty method. Instead, it achieves a natural balance between the dimen- [11-13] for a clearand deta}led discussi'
sions of the interior and:~ boundary terms. The minimization of 1" beco mes A second way to use exact inform2
now a problem of -simultaneous approxlmation in n and on the lower- is to introduce its Green's function ar
dimensional manifold r. gral equation over the boundary r ..
It is t~asonable to expect piecewise polynomial spaces to give the same solution on r has been taken as a pi!
degree of approximation on ras in Q. Ifr were straight, to take the simplest determined by collocation. And a th
case, then acomplete poJynomial in X p . • • , Xn reduces on r to a complete the old-fashioned global trial functio:
polynomial in the boundary variables s 1 , • • • , sn-t· With a curved boundary ments at the boundary. This makes ~
we believe the approximation théory to be effectively the same; u differs from the interior permits good approximat
its interpolate on r by O(hk-s) in the sth derivative. However, given only that tions, and the hard work is done in
approximation of degree k can be achieved in n, simultaneous approximation exist almost no theoretical analysis o1
with coefficient h- 3 is a much more subtle problem; its solution by Bramble
and Schatz showed that the balance of powers in (12) is exactly right. Pro-
vided that k > 4m, their error estímate for the least-squares solution u'Ls is
the optimal

(13)

The practica} difficulties in the least-squares method.are exactly those as-

sociated with an increase in the order ofthe differential equation frorn 2m to
4m;The trial functions rnust lie in 3C 2 m in order for the new functional to be
finite. This means that the admissibility condition is continuity of all deriva-
tives through order 2m 1, which is ditp.cult to achieve. Furtherrnore, the
bandwidth of K is increased and its condition number is essentially squared,
going from O(h~-zm) to O(h- 4 m).t Therefore,_ the nurnerical solution must
inevitably proceed more slowly. The rate ofconvergence in xs will normally
be hP, where p is the srnaller of k s and 2k 4m.
Finally, we rnention three of the techniques recently invented for problems
which have homogeneous, differential equatións, say Au = O, and inhorno-
geneous boundary conditions. There are a nurnber of important applications
in which u and dufdn are wanted only on the boundary, and it seerns rather
inefficient to compute the solution everywhere in n.
One possibility is to find a farnily of exact solutions rp¡, .•. , rpN to the
differential equation, and choose the combination .!; Qirpi which satisfies
the boundary conditions as closely as possible. This means that sorne expres-
sions for the boundary error is to be rninimized. With least squares this
leads to a linear equation for the Q¡; Fox, Henrici, and Moler [FlO] had re-
rnarkable success by using instead a minirnax principie, minirnizing the error
o ver a discrete set of boundary points. If the rp i are eigenfunctions of the
differential operator, there may be- enorrnous simplifications at sorne points;

tThis objection has been met by a modificat1on due to Bramble and Nitsche.
 (
l
CHAP.2 SEC. 2.4. SYSTEMS OF EQUATIONS; SHELLS; FORCE AND HYBRID METHODS 135

egree k of the subspace S", as it was in again the serious difficulties are at the boundary. We refer t~ Mi~hlin's ~ooks
es a natural balance between the dimen- [11-13] for a clear and deta}led discussion ofthis more classtcal ctrcle of td~as.
~rms. The minimization of 1" bedomes A second way to use exact information about the homogeneous equat10n
proximation in n and on the lower- is to introduce its Green's function and transform the problem into an inte-

se polynomial space~ to give the same

r
. If were straight, tS:'take the simplest
( t' . . . ' xn reduces on r to a complete
so ... , sn- r: Wíth a curved boundary
:> be effectively the same; u differs from
th derivative. However, given only that
:ved in n, simultaneous approximation
1btle problem; its solution by Bramble
,f powers in (12) is ·exactly right. Pro-
te for the least-squares solution u'Ls is
gral equation over the boundary r. In sorne initial trials the approxi~ate
solution on r has been taken as a piecewise polynomial and its coeffictents
determined by collocation. And a third idea, suggested by Mote, is to use
the old-fashioned global trial functions in the interior of n, with finite ele-
ments at the boundary. This makes sense, because the smoothness of u in
the interior permits good approximation by comparatively few global func-
tions, and the hard work is done in a "boundary layer". There appears to
exist almost no theoretical analysis of this idea, but its time will come.

t-squares method are exactly those as-

fthe differential equation from 2m to
' in order for the new functional to be
V condition is continuity of all deriva-
dilp.cult to achieve. Furthermore, the
tdition number is essentially squared,
erefore, the numerical solution must
te of convergence in Jes will normally
.nd 2k- 4m .
.niques recently invented for problems
!quatións, say Au = O, and inhomo-
e a number of important applications
on the boundary, and it seems rather
ywhere in n.
:>f exact solutions rp 1 , • • • , 'PN to the
combination 2:; Q1rp1 which satisfies
•ssible. This means that sorne expres-
minimized. With least squares this
ox, Henrici, and Moler [FlO] had re-
timax principie, minimizing the error
If the rp1 are eigenfunctions of the
mous simplifications at ~o me points;

.tion due to Brarnble and Nitsche.

 )
ÍJ SEC. 3.1.

The degree of S is usually trivial t

proximation the degree is one (in otl
cubics k = 4, for the reduced quintic 1<
From these hypotheses we want t
achieved by S. Recalling that the de
ep e 2 , ••• , we shall use the distances

h¡ = diameter of i

3
in measuring the accuracy of approx
of nodal subspaces S\ parameterized
APPROXIMATION approximation error decreases like a p<
assumption on the functions QJ~, whid
functions· QJ~ are uniform to order q prc
for al! h, i, and j,

(2) max 1Da.QJ~(:x

x in e¡
la.l=s
3.1. POINTWISE APPROXIMATION
This condition is imposed on al! derivat
This section begins our discussion of the question which líes mathemati- q; that is, it holds for al/ ex sueh that
cally at the center ofthe whole finite element analysis-approximation by the 1 Di 1 is the order of the deriva ti ve whi
spaces Sh. We start with approximation in a pointwise sense, where the special QJ~ at its node z~; 1D¡ 1 = Oif QJ~ is asso
role of polynomials will be easy to recogrtize. Then in the remainder of the 1 Di 1 1 if it is associated with 'V x or
chapter, the same pattern is extended to the spaces Jes(Q), in other words, A good example in one dimensi
to the approximations in energy norms on which the Ritz-Galerkin-finite polynomials (k = 4), determined ove
element method is based. derivatives at the endpoints. On a 1
Suppose to start with that we are given a smooth function u = u(x), and ro are displayed in Fig. 1.8. Over
defined at every pOÍnt X (X 1 , • • • , Xn) Ín the n-dimensional domain Q. 'l'h 'l'(xfh) and roh = hro(x/h), res]
Suppose also that S is a nodal finite element space, spanned by q1 1 (x), ... , appearing in roh, in order to keep its
QJN(x). As in Section 2.1, this means that to each QJ i there is associated a node the reason for the term hlD;l in (2),
z¡ anda derivative D¡ such that correct. These basis functions are un
tives a J-function enters at the node
(1) typical of finite elements; the qth dt
we accept in (2), but higher derivative
Suppose, finally, that the space has degree k 1: Every polynomial in X¡, .•. ' xn
associated with the smoothness of th
of total degree less than k' is a combination of the basis functions rp i' and
eq-t . of functions with q - 1 contim
therefore lies in S. (The total degree of x 1 x 2 , for example, is 2; its presence
it is contained in the space 3Cª of fu
is required in a space of degree k- 1 2, but not in a space of first degree
square sense. The conforming condi
even though it is linear in x 1 and x 2 separately.) If P(x) is such a polynomial,
to a differential equation of order 2r.
and the linear combination ')'hich produces it is P L; p ¡QJP then by the
The uniformity condition becom
interpolation property (1) the weights are just the nodal values of the poly-
dimensions. Normally it can be tra1
nomial: ·
straint on the elemental regions ei.
P¡ D¡P(z¡). say the right triangle Twitli vertices
The nodal derivatives Di should be of order to T the basis functions and theiJ

136

of the question which Iies mathemati-
!e?Ient a?aly~is-approximation by the
:I m a pomtwise sense, where the special
cogrtize. Then in the remainder of the
to the spaces xs(Q), in other words,
ns on which the Ritz-Galerkin-finite
e gi:en a smooth function u u(x),
x,) m the n-dimensional domaín n.
~lement space, spanned by rp 1 (x), ... ,
tt to each rp i there is associated a no de
:e k. l : E very palynamial in x 1 , ••• , x,
~atwn af the basis fynctians rpi' and
for example, is 2; its presence
· X¡X 2 ,
2, but not in a space of first degree
arately.) If P(x) is such a polynomial
·duces
•
it is P I; p J.rpJ., then by th~
ue JUSt the nodal values of the poly-
P(zi).
Jrder <k.
SEC. 3.1. POIN1WISE APPROXIMATION 137
The degree of S is usually trivial to compute. For linear or bilinear ap-
proximation the degree is one (in other words, k = 2), for cubics and bi-
cubics k = 4, for the reduced quintic k is only 5, and so on.
From these hypotheses we want t~ deduce the order of approximation
achieved by S. Recalling that the domain Q is partitioned into elements
e 0 e2 , ••• , we shall use the distances
in measúring the accuracy of approximation. We have in mind a sequence
of nodal subspaces S\ parameterized by h, and we hope to establish that the
approximation error decreases Iike a power of h. This will require a uniformity
assumption on the functions rp?, which can be expressed as follows: The basis
functians· rpj are unifarm ta arder qpravided there exist canstants es such that
far al! h, i, and j,
(2) max 1Da.rp?(x) 1 < cshJDJI-s.
x in e¡
lo:l=s
This canditian is impased anal! derivatives Da. = aro:ljax~ 1 • •• ax~· up ta arder
q; that is, it halds far al! a such that 1a 1 a 1 + ... + a, < q. Recall that
1 Di 1 is the order of the derivative which is interpolated by the basis function
rpj at its node zj; 1 Di 1 = Oif rp? is associated with a function value ata vertex,
1 Di 1 1 if it is associated with vx or vy or v,, and so on.
A good example in one dimension is furnished by the Hermite cubic
polynomia]s (k = 4), determined over each interval by their values and their
derivatives at the endpoints. On a unit interval the two basis functions r¡¡
and w are displayed in Fig. 1.8. Over the reduced interval [0, h] these become
r¡¡h r¡¡(xfh) and wh = hw(x/h), respectively; one sees the extra factor h
appearing in ro\ in order to keep its slope equal to 1 at the origin. This is
the reason for the term h1DJI in (2), to make the inequality dimensionally
correct. These basis functions are uniform up to q 2, but for third deriva-
tives a <5-function enters at the nodes and uniformity breaks down. This is
typical of finite elements; the qth deriva ti ves involve step functions, which
we accept in (2), bu t higher derivatives are disallowed. Th us q is the parameter
associated with the smoothness of the subspace; Sh is contained in the space
eq-t of functions with q - 1 continuous derivatíves, and more importantly,
it is contained in the space Jeq of functions with q derivatives in the mean-
square sense. The conforming condition, in order to apply the Ritz method
toa differential equation of order 2m, is simply q m.
The uniformity condition becomes especially significant in two or more
dimensions. Normally it can be translated directly into a géometrical con-
straint on the elemental regions e¡. First one considers a standard element,
say the right triangle T with vertices at (0, 0), (1, 0), and (O, 1). With respect
to T the basis functions and their derivatives up to order q shou]d be
 ¿
138 APPROXIMATION CHAP. 3 SEC. 3.1.
\

bounded. Then ~'"a change of coordinates, taking T into the given triangle e, derivative: c0 l. Consider, however,
and involving the Jacobian, leads to the following resul~: The basis is uniform rp 1 equals 1 at z1 and zero at P, its slopt
provided that as h O, all angles in the triangulation exceed sorne lower
---¡.

e
bound o. In this case it is not difficult to find constants C¡; bounded by

constant
(sin 80 )s Comparing this with the uniformity cot
Thus, if the triangles are allowed to deg
We emphasize that the effect ofthe element·geometries on approximation will not be uniform.
is. entirely contained in this estímate. For quadrilaterals the angles must also On such a triangle the linear ínter]
be bounded away from 180°, to avoid degeneration into triangles. Consider, for example, the function u(x
We should note a situation in which this bound on es is not valid. It will at z1 and equals h2/4 at the other two
certainly fail if the polynomial is not uniquely determined within the element quently the derivative along the x-axi:
by the nodal parameters, that is, ifthe matrix Hin Section 1.10 which links is independent of x, may be seriously i
the polynomial coefficients a, to the nodal parameters q1 beco mes unbounded
for sorne element configuration. Such singular cases, once recognized, have
been avoided in the literature. The danger is greatest when e; is originally
curved, and is mapped into a polygon in order to use one of the standard This derivative error is O(h) only if th1
finíte element constructions. If the Jacobian of this mapping is allowed to that u u1 = O(h 2 ) regardless of 8; thc
vanish (see Section 3.3, for example), the mathematical properties ofthe con- less factor "hmaxfhmin" introduced by th
struction are lost. There is also another reason to av4
The geometrical condition is easy to understand for linear functions on They may injure the numerical stabilit
triangles. The pyramid rpj equals 1 at the jth vertex and zero at the others. In the condition number of K. Therefon:
between one always has 1 rpj(x) 1 < 1, so that uniformity holds for the zeroth structed to triangulate an arbitrary dor
y
than 80 ~ n/8. For a more general n-
req uiremen t can be ex pressed in ten
angles: e, should contain a sphere of
We are now ready for pointwise
The given function u will be approxirr
in other words,. by the function ui wh
as u itself:
uix) 1:; 1

We suppose that u has k derivatives

Zj = (h tan 8/2, 0} X estima te the sth derivatives of u- ui.

THEOREM 3.1
Suppose that S is of degree k - l
condiiion (2). Then for s < q,

(0, -h/2)
(3) max!Dccu(x) Dccu¡{x)i
x in e,
Je~;["'S

Fig. 3.1 Linear shape function in a thin triangle. Proof We choose an arbitrary po
 CHAP.
tates, taking T int9 the given triangle e;
the following result: The basis is unifortA
in the triangulation exeeed soine lower
llt to find constants es bounded by
constant
(sin 8 0 ) 9 •
le element geometries on approximation
. For quadrilª'terals the angles must also
id degeneration into triangles.
tich this bound on es is not valid. It will
uniquely determined within the element
l1e matrix Hin Section 1.10 which 1inks
tod~I parameters q¡ becomes unbounded
h smgu~ar cases, once recognized, have
da~ger Is greatest when e, is origínally
)n m order to use one of the standard
racobian of this mapping is allowed to
the mathematical properties of the con-
' to understand for linear functions on
the jth vertex and zero at the others. In
so that uniformity holds for the zeroth
' - (h tan 812, o)· X
tnction in a thin triangle.
3c.. SEC. 3.1. POINlWISE APPROXIMATION 139
derivative: e 0 = L Consider, however, the x-derivative in Fig. 3.1. Since
rp¡ equals 1 at z¡ and zero at P, its slope between these points is
Comparing this with the uniformity condition (2), e 1 will be at Ieast 2/tan 8.
Thus, if the triangles are allowed to degenerate into thin "needles," the basis
will not be uniform .
On such a triangle the linear interpolate can lead to significant errors.
Consider, for example, the function u(x, y)= y 2 • The interpolate u1 vani~hes
at z¡ and equals h 2 j4 at the other two nodes (and therefore at P). Conse-
quently the derivative along the x-axis, which should vanish since u y 2
is independent of x, may be seriously in error for the interpolate:
a h 2 /4 h
ax(u- U¡)= (h/2)tan8 2tan8·
This derivative error is O(h) only if the angle 8 is bounded below. We note
that u - u1 = O(h 2 ) regardless of 8; the difficulty comes from the dimension-
less factor "hmax/hmin" introduced by the derivative.
There is also another reason to avoid degenera te triangles of any kind:
They may injure the -numerical stability of the Ritz method, as reflected in
the condition number of K. Therefore, severa! algoríthms have been con-
structed to triangulate an arbitrary domain n, keeping all the angles greater
than 80 ~ n/8. For a more general n-dimensional element, the geometrical
requirement can be expre~sed ·in term¿ of inscribed spheres rather than
angles: el should contain a sphere of radius at least vh,, v a fixed constan t.
We are now ready for pointwise approximation in the nodal method.
The given function u will be approximated by its interpolate in the space S,
in other words, by the function u1 which shares the same nodal parameters
as u itself:
We suppose that u has k derivatives in the ordinary pointwise sense, and
estímate the sth derivatives of u- u1 •
THEOREM 3.1
Suppose that S is of degree k - 1 and its basis satisfies the uniformity
eondiiion (2). Then for s q,
(3) max 1 Drt.u(x) D"'u¡{x)l Cshf-s max 1 DPu(x) [.
x in e 1
¡,.¡=s
x in e,
IPl=k
Proof We choose an arbitraty point x 0 in e¡, and expand u in a Taylor
140 APPROXIMA~tON CHAP. 3 SEC. 3.1.
\ ~s,
finite element.solution uh, sinceit is ob
series: \_

I(v), the potential energy over the wh'

· u(x) P(x)+ R(x),.
u· is indeed própagated by uh-somet
where P is a polynomial of degree k 1. and R is the remainder term. We
throughout the entire domain. This is
shall need the standard estimatefor the Taylor remainder R and its deriva- There is a similar approximation t
tives, method. We shall take time to prove
two overlap for the usual nodal methc
(4) max !Da-R(x)! Ch7-s max 1 DPu(x) ¡. covered by the previous theorem, be
x in e; x in e,
lrx-l=s IPl""k basis, yet their approximation prope1
In fact, the special regularity of the 1
This can be proved by expressing R as a line integral from x 0 to x. elegant result. Approximation on a re~
The interpola te of u decomposes, by linearity, into the sum of two inter- tria! function 'fl with the following re~
polates: P of degree less than k,
u¡{x) P¡{x) + R¡{x).
The crucial point is that P1 is identical with P; any polynomial of degree
(5) P(x) =L;ll

k 1 is exactly reproduced by its interpolate. This is where polynomials play Such a superfunction 'fl can be found
a special role. In other words, P --'- P1 coincides in e1 with a trial function for S is of degree k - l. 'fl will be nonzen
which all the determining nodal parameters are zero; therefore, P P1 in e1• In the simplest one-dimensional e
This means that u - u1 = R R1 , and the only. terms left to estímate piecewise linear functions. There is on
are the derivati ves of R1 : and the basis is generated by the roof
functions are rp?(x) <1> 1(x/h - /), a1
<1> 1 itself will serve as the function r¡.
(f.x+P,
At most d nonzero terms can appear in this sum, since the other basis func-
tions vanish in the element. Therefore; if (4) is combined with the unifor- P(x) = L;P
mity condition (2),
The two sides agree at every node and
!Da-R¡(x) 1 f/Ch~-IDJI max 1 DPu !· cshJD1 1-s identical. The same will be true in tW(
functions on triangles, where the gra
C'h;-s max 1DPu ¡.
bilinear functions on rectangles, whe
Thus R1 is bounded exactly as R was, and it follows immediately that cases the approximation uQ in the folle
polate ur
The construction of 'fl is not so 1
functions for Hermite cubics, as well
This is the estima te (3), and the proof is complete.t Section 1.7. None of these functions
which is required of 'fl· However, ther
Note that this is a completely local estímate; u1 imitates the properties it is nonzero over six intervals (Fig.
of u within each element. This is a property which we cannot expect_ of the E-spline and its two neighbors, and a
a Hermite cubic) it will serve as the
tThe uniformity of the basis is not a necessary condítion for the approximation theo- too. We must emphasize here that it
rems to hold. Por bilinear or bicubic approximation on rectangles, the choice of a very fine what matters is only that somewhere
mesh in one coórdinate direction will spoil the uniformity but not the order of approxima-
tion. Even for triangles, Synge's condition in [17] is that the largest angle be bounded below is responsible, by itself, for the apprm
n, i.e., that the second smallest angle in each triangle exceed sorne 80 • Given the existence of rp(x), vani
 CHAP.J
d + R(x),
1 and R is the remainder term. We
1e Taylor remainder R and its deriva-
1z7-s max l DPu(x) J.
x in e 1
IPI=k
a line integral from x 0 to x.
•Y Iinearity, into the sum of two in ter-
e)+ Rlx).
:tl with P; any polynomial of degree
;>olate. This is where polynomials play
oincides in e¡ with a trial function for
eters are zero; therefore, P = PI in e;.
, and the only terms left to estímate
t this sum, since the other basis func-
if (4) is combined with the unifor-
max 1DPu 1· cshJD;I-s
ax 1 DPul.
td it follows immediately that
(C + C')h1-s max 1DPu J.
complete.t
estímate; ui imitates the properties
perty which we cannot expect of the
;ary condition for the approximation theo-
ltion on rectangles, the choice of a very fine
uniformity but not the order.of approxima-
7] is that the largest angle be bounded below
iangle exceed sorne 8 0 •
SEC. 3.1. POINTWISE APPROXIMATION 141
finite element, solution uh, since it is obtained by minimizing a global function
I(v), the potential energy over the whole domain O. In fact, a singularity in
u is indeed propagated by uñ-sometimes with frustratingly slow decay-
throughout the entire domain. This is analyzed in Chapter 8.
There is a similar approximation theorem for the abstract finite element
method. We shall take time to prove this second theorem, even though the
two overlap for the usual nodal method on a regular mesh. Splines were not
covered by the previous theorem, because they lack a local interpolating
basis, yet their approximation properties are of considerable importance.
In fact, the special regularity of the mathematical structure allows a more
elegant result. Approximation on a regular mesh hinges on the presence of a
tria/ function t¡1 with the following remarkable property: For any polynomial
P of degree less than k,
(5) P(x) 2; P(l)t¡!(x - /).
l
Such a superfunction t¡1 can be found in the tr_ial subspace S if and only if
S is of degree k l. t¡1 will be nonzero o ver a small patch of elements.
In the simplest one-dimensional example, Sh is made up of continuous
piecewise linear functions. There is one unknown per mesh interval (M= 1),
and the basis is generated by the roof function <I> 1 • In other words, the basis
functions are qJf(x) <1> 1 (xfh l), and Sh is of degree l. In this example
<1> 1 itself will serve as the function t¡1: Por any linear polynomial P(x) =
rxx + p,
The two sides agree at every node and are linear in between, so they must be
identical. The same. will be true in two dimensions both in the case of linear
functions on triangles, where the graph of <I> 1 = t¡1 is a pyramid, and · of
bilinear functions on rectangles, where the graph is a "pagoda." In these
cases the approximation uQ in the following theorem coincides with the inter-
polate ui.
The construction of t¡1 is not so trivial for cubics. The two generating
functions for Hermite cubics, as well as the B-spline <1> 17 were sketched in
Section 1.7. None of these functions satisfies the identity (5) up to k = 4,
which is required of t¡J. However, there exists a cubic spline which does, and
it is nonzero over six intervals (Fig. 3.2). This t¡1 is a combination of the
E-spline and its two neighbors, and also (since every spline is automatically
a Hermite cubic) it will serve as the superfunction for the Hermite space,
too. We must emphasize here that it is never essential to know what t¡1 is;
what matters is only that somewhere in the trial space lies a function which
is responsible, by itself, for the approximation properties of the space.
Given the existence of t¡J(x), · vanishing for 1x 1 > p and satisfying the
142 APPROXIMATION CHAP. 3 SEC. 3.2.

This is the identity (5) with x scaled by

the polynomial p(x) = P(xh), and then
The Taylor remainder R was éstir
this leaves only R 0 :

For x in the mesh cube O X¡< h,

where C' bounds the derivatives of Vf, 1

entiation D«, and the rest is an upper
it is essential that the sum is finite; s
'l'(x) = -(1/6)cp(x+1) + (4/3)cp(x)- (116kp(x -1) only finitely many values of l actually
Fig. 3.2 The superfu~ction for cubics.
estímate, with lrt 1 s:
1D«(u uQ) 1 l D«(R
superfunction identity, we approximate a given function u by the following
member of Sh: The same result follows in any other rr
of the vertices.

We refer to uQ as a quasi-interpolate based on r¡¡. lt depends as u1 did on the 3.2. MEAN-SOUARE APPROXIMATIO
local properties of u, but does not quite interpolate u at the nodes. lts ad-
vantage is that it can be written down so easily and explicitly, whereas We want to prove that also in the
interpolation by splines requires the solution of simultaneous equations of approximation is possible und~r th'
(several B-splines are nonzero at each node). Since approximation only degree k - 1 and a uniform basis. It is
demands that there exists sorne fuhction in Sh which is clase to u, we are free it natural and inevitable to work wit
to work with a convenient choice. itself is exactly such an expression, ar
of u. Since uh is the element closest to
THEOREM 3.2 estímate in the whole subject is this or
Suppose the abstract finite element space Sh is of degree k 1 and r¡¡ has In general, the Jem norm involves
bounded derivatives up to order q. Thenfor any derivative DrA. of order lrt 1 < q,

(6) max !Drku(x)- DrkuQ(x)l

- o:::;;xt<h
(rA.( "'S

We shall ha ve use also for the seminorr.

Proof. The argument is nearly the same as in Theorem 3.1. A Taylor of order exactly lrt 1 m are included i1
expansion around the origin gives u(x) = P(x) R(x), where Pis of degree it shares the properties l cv 1 1e 11 v 1a
k 1 and R is the remainder. Splitting uQ into PQ + RQ, we note first that it fails to be positive definite: For a tr
PQ coincides :with P: v O, but for these seminorms, IPlm
degree less than m. Obviously 11 v 11! =
P(x) = PQ(x) ~ P(lh)r¡¡( ~ t) · In one sense it is completely trivial1
 /
\·
CHAP.3
4/3)cp( x)- {1/6)cp(x 1)
fu(\ction for cubics.
e a given function u by the following
.sed on 'JI· It depends as u1 did on the
ite interpolate u at the nodes. Its ad-
;vn so easily and explicitly, whereas
solution of simultaneous equations
h node). Since approximation only
1 in Sh which is close to u, we are free
· estímate in the whole subject is this one.
pace Sh is of degree k 1 and VI has
or any derivative D« of order IIX 1 q,
same as in Theorem 3.1. A Taylor
P(x) + R(x), where Pis of degree
~ uº into PQ + Rº, we note first that
SEC. 3.2. MEAN-SQUARE APPROXIMATION 143
This is the identity (5) with x scaled by the factor h; that is, (5) is applied to
the polynomial p(x) = P(xh), and then x-+ xjh. ,
The Taylor remainder R was estimated in the previous theorem, and
this leaves pnly Rº:
For x in the mesh cube O X¡ < h,
where C' bounds the deriva ti ves of 'JI, the factor h-1«1 comes from the differ-
entiation D«, and the rest is an upper bound for R(lh). As in Theorem 3.1,
it is essential that the su m is finite; since 'JI is nonzero only for 1x 1 < p,
only finitely many values of 1 actually appear. We now have the required
estimate, with 1a 1 s:
1D«(u uQ) 1 1D"(R - RQ) 1< cshk-s max 1DPu 1.
The same result follows in any other mesh cube by expanding u around one
of the vertices .
3.2. MEAN-SQUARE APPROXIMATION
We want to prove that also in the mean-square norms, the same degree
of approximation is possible under the same hypotheses on Sh-that it has
degree k 1 and a uniform basis. It is the variational principie which makes
it natural and inevitable to work with these Jes norms; the strain energy
itself is exactly such an expression, an integral of the square of derivatives
of u. Since uh is the element closest to u in this energy norm, the key error
In general, the Jem norm involves all deriva ti ves D" of order 1a 1 m:
We shall ha ve use also for the seminorm 1v lm,n, for which only the derivatives
of order exactly 1a 1 = m are included in the su m. This is a seminorm, because
it shares the properties 1cv 1 1e 11 v 1and 1v + w 1 1v 1+ 1w 1of a norm, but
it fails to be positive definite: For a true norm, 11 v 11 O can happen only if
v =O, but for these seminorms, IP!m =O whenever P is a polynomial of
degree less than m. Obviously 11 v 11; 1v 1~ + . . . 1v ¡;.
In one sense it is completely trivial to extend the pointwise results to these
144 APPROXIMATION CHAP. 3 SEC. 3.2.

new nqrms. Using the interpolate u1 of the previous section, we may simply that since u(x¡, ... , x") has k mean-~
square the pointwise errÓr and integrate: If 1 et 1 = s; is well defined at any point zi:

fn 1Drr.u - Drr.u1 12 dx <(vol Q) max 1Drr.u - Drr.u1 12

(9)

Fortunately, this assumption 1 Di 1 <

< c;(vol Q)h 2<k-s> max 1DPu(x) 12 •
xinn elements, and the theorem leads direc
IPI=k
finite element method.
Thus if the kth derivatives of u are bounded, the error is of the same order To find a more awkward case, wt
hk-s as befare. This settles the rate of convergence to smooth solutions. constants (k = 1) in the plane (n = ~
Such an estímate, however, can never be entirely satisfactory. It involves in Section 1.8, Sobolev does not guar
two quite different norms, and we have made a pointwise assumption on the sen se; if a node happened to fall at
kth derivatives in order to obtain a mean-square error bound on the sth Even in this case, however, Sh does con
derivatives. One has a right to expect a more symmetric theorem, in which order h of least-square approximation
only norms of the same kind appear. Furthermore, such an improvement is is interpolated. The same constructior
made necessary by singularities. For a plate with a fracture, the solution u S h has degree k - 1 and a uniform baJ
increases o ni y like the square root of the distan ce from the end of the crack.
The function r 117.: is not even differentiable; the pointwise error will behave
like h 112 or perhaps h11 2 log h. However, the s_olution does lie in JC 1 ; it must.
Thus each finite element space Sh e
In fact, u nearly has "one and a half derivatives" in the mean-square sense;
necessary Ü1 , which approximates u to
the sth derivative behaves like r< 112 >-s, and for any s < !,
is much simpler to prove and is suffi
would be, if it were extended also to fr
J1r<1!2)-s 12 r dr d() < oo. example of a cracked plate. This exten
of the theorem, given the theory of "
In· short, we would like to pro ve that the mean-square error is roughly of as being too technical.
order h 3 l 2, and such a result can only follow from a theorem in which Remark 2. The proof even of the
hypotheses are made on the mean-square differentiability of u. technical lemma. The Taylor expansi<
was based will not work here; u may t
THEOREM 3.3
late to be defined, but not for a Taylo
Suppose that Sh is of degree k - 1, and its basis is ·uniform to order q. fore, we must rely on functional analy~
Suppose also that the derivatives Di' associated with the nodal parameters el ose to u in the mean-square sense a
are al/ of order 1 Di 1 <k'-- n/2. Then for any function u(xp ... , x") which in the pointwise sense. The key result,
has k derivatives in the mean-square sense, and any derivative nrr. of order In e very element there exists a polyn01
s<q, u - Pk-t satisfies
c;h¡<k-s) 1u l~.e,·
(7)
f e¡
1nrr.u(x) - nrr.u¡(x) 12 dx < (lOa) 1 R ls,e, < ch~-:s 1 ¡
and at every node zi,
Since the integral o ver n is the sum of the integrals o ver ei,
(lOb)
(8)
On a region of diameter hi = 1, (1 Oa) :
Remark l. The assumption 1 Di 1 <k- n/2 is necessary in order that the from (lOa) and Sobolev's ineq~ality e
interpolate u1 be defined. Sobolev's lemma (mentioned in Section 1.8) assures when the independent variables are re
 CHAP.
of the previous section, we may simply
ate:Ifi~XI=s, · is well defined at any point z 1 :
>1 !l) max 1Dau - Dau1 l2
(vol !l)h 2 <k-s) max 1 DPu(x) j2 •
x in .O
lPI=k
ounded, the error is of the same order
of converg¡;nce to smooth solutions.
:ver be entirely satisfactory. It involves
'e made a pointwise assumption on the
mean-square error bound on the sth
:a more symmetric theorem, in which
Furthermore, such an improvement is
a plate with a fracture, the solution u
the distance from the end of the crack.
tiable; the pointwise error will behave
!r, the solution does lie in 3C 1 ; it must.
:lerivatives'' in the mean-square sense;
and for any s < !,
· dr dO< <X>.
t the mean-square error is roughly of
tly follow from a theorem in which
:are differentiability of u.
', and its basis is. uniform to order q.
associated with the nodal parameters
: for any function u(xl' ... ' xn) which
~ense, and any derivative Da of order
dx
he integrals over e¡,
k - n/2 is necessary in order that the
ma (mentioned in Section 1.8) assures
3 SEC. 3.2. MEAN~SQUARE APPROXIMATION 145
that since u(x 1' .•• ' xn) has k mean-square derivatives, the derivative D ¡U
(9)
Fortunately, this assumption 1 D 1 1 < k n/2 is fulfilled for all practica!
elements, and the theorem leads directly to the rate of convergence of the
finite element method.
To find a more awkward case, we turn to approximation by piecewise
constants .(k 1) in the plane (n 2). Then, as for the function log log r
in Section 1.8, Sobolev does not guarantee that the interpolate of u makes
sense; if a node happened to fall at the origin, what would be log log O?
Even in this case, however, Sh does contain a function which gives the correct
order h of least-square approximation; u can be suitably smoothed befo re it
is interpolated. The ·same construction succeeds in the general case [S6]: lf
Sh has degree k 1 anda uniform basis, thenfor any dimension n,
ü smoothed u.
Thus each finite element space Sh contains a function, usually u1 but if
necessary Ü1 , which approximates u to the expected order. The bound for u1
is much simpler to prove and is sufficient for all practica! purposes-or it
would be, if it were extended also to fractional derivatives, as required by the
example of a cracked plate. This extension is actually an immediate corollary
of the theorem, given the theory of "interpolation spaces"-whith we omit
as being too technical.
Remark 2. The proof even of the simpler theorem does still require one
technical lemma. The Taylor expansion on which pointwise approximation
was based will not work here; u may have enough derivatives for its interpo-
late to be defined, but not for a Taylor expansion with remainder hk. There-
fore, we must rely on functional analysis to produce a polynomial which is as
· close to u in the mean-square sense as the leading Taylor-series terms were
in the pointwise sense. The key result, simplified from [14] and [827], is this:
In every element there exists a polynomial Pk_ 1 such that the remainder R =
u Pk-t satisfies
(lOa) S <k,
and at every node z 1,
(lOb) 1DjR(z¡) 1 <e' hf-IDJI-n/ZI u lk,e¡•
On a region ofdiametet hi 1, (lOa) is a standard lemma [14]; (lOb) follows
from (lOa) and Sobolev's ineq~ality (9). Then the given powers of h1 appear
when the independent variables are rescaled so asto shrink the region down
 ,)
146 APPROXIMATwN CHAP.3 SEC. 3.2.

to e1• t The constants e and e' will depend on the smallest angle of ei; it is terms of the next higher degree k: A
simplest to assume this angle to be bounded below, as is normally required on the right si de of (7)? Bilinear elerr
for the basis to be uniform. \- · term xy, and therefore it seems supe
Proof of the theorem. The argument will be exactly the same as in the ux:v in the error bound ch 2 -s 1u b· This
pointwise case; that is the happy effect of the lemma. In each element e¡, bért [B28]: It is indeed sufficient to incl
we writ~ u Pk_ 1 + R; R satisfies the inequalities (10). The interpolate This will have valuable consequences
u1 is, by"-linearity, the sum of the two interpolates: of rectangular isoparametric elements
It may well happen that (7) holds
U¡= (Pk_l)l + RI = pk-1 + R¡, all orders s < k. The limitation to s ::
·since the polynomial Pk-t is exactly reproduced by interpolation. Therefore derivatives across element boundarie~
u- u1 = R R1 , and we consider It would be useful to know sometl
R¡{x) = 2: (D 1RXz1)rp1(x). rem. They are a direct indication of ti
if they are larger for one element tha
Only a fi.nite number d of these terms can be ·nonzero in e;. Applying the the first element is comparatively ina
uniformity ofthe basis, as well as (lOb), any derivative of order la 1~ s gives approximation on the line, with equa
Du.R1(x)l < dc'hf::-IDJI-nl 2 lu! k,et ·esh!DJI-s. constants can be computed. There
! - t t
th~ first trigonometric polynomial J
Now ifwe square, integrate over e1, and take the square root, meshpoint, and the first algebraic pe

(J e, 1 Da: R¡(x) 12 dx r 12
e" hr-s 1u !k, e¡·
its linear interpolate. In the case ofj
approximation are identically zero.
easy computation gives
\.

This is exactly the estimate given for R by (1 Oa), and exactly the same as the
right side of (7). Therefore, the proof is complete. The technique, depending
on (1 O) and on the special role of polynomials, is known as the Bramble-
Hilbert lemma. It follows from the proof of Theoret
The same theorem and proofhold for abstract finite elements on a regular are optimal, since for any u the interJ
mesh, in particular for splines; we again use uQ instead of u1 [S5J.
The theory of partial differential inequalities leads to questions of one-
sided approximation. We have established for the standard linear elements
that, given u> O, the estimates of Theorems 3.1-3.3 continue to hold if the Therefore, C0 l/n 2 and C1 1/n
approximate vh is required to satisfy O< vh u. (The interpolate vh = u1 What role is left for the quadn
is of course no use, since it need not líe below u.) We note that Duvaut-Lions function u, the constant which is 1

were able to formulate as variational inequalities several important physical previous case we fixed h and sough1
problems, including elastic-plastic phenomena-which in differential. form case we fix u and ask for the limits
lead to ~xtremely awkward elliptic-hyperbolic systems separated by an un-
known free boundary. Mosco and Strang, and independently Falk, have . min l u - v" lo
(11) Co = lIm h2j u 12 .
confirmed the usual h 2 error in energy for linear approximation of the St. h-0

Venant torsion problem, which is typical of the class of variational inequalities

The mínimum is taken o ver all v" in
known as obstacle problems.
tions. For every h these ratios are
An additional question arises if the elements contain a few polynomial
and therefore the limits cannot exc<
thf arises from rescaling the kth derivatives, and h~12 from the square root of the to be smaller, since a fixed u cann
volume of e1• sin nx/h at once. These new constan
 CHAP.
epend on the smallest angle of e.· it is
Jounded below, as is nonnally re~uired
1ent will be exactly the same as in the
fect of the lemma. In each element e¡,
; the inequalities (10). The interpolate
) interpola tes:
R¡ = pk-t + Rn
r~produced by interpolation. Therefore
:D1RXz1)rp1(x).
1s can be nonzero in e¡. Applying the
J), any deriva ti ve of order l (X 1 ~ s gives
td take the square root,
1/2
1 <
- e" hlf-s
.z l u 1k,e1.•
R by (lOa), and exactly the same as the
is complete. The technique, depending
)]ynomials, is known as the Bramble-
·or abstract :finite elements on a regular
.in use uQ instead of u1 [S5].
~nequalities leads to questions of one-
;;hed for the standard linear elements
!orems 3.1-3.3 continue to hold if the
O < vh u. (The interpolate vh u1
below u.) We note that Duvaut-Lions
nequalities severa! important physical
!nomena-which in differential. form
perbolic systems separated by an un-
trang, and independently Falk, ha ve
r for linear approximation of the St.
l of the class of variational inequalities
~ elements contain a few polynomial
ves, and h?12 from the square root of the
3 SEC. 3.2. MEAN-SQUARE APPROXIMATION 147
· terms of the next higher degree k: Are all deriva ti ves of this order needed
on the right si de of (7)? Bilinear elements, for example, reproduce the twist
term xy, and ·therefore it seems superfluous to include the cross derivati ve
uxy in the error bound eh -s l u !2 • This point was settled by Bramble and Hil-
2
bert [B28]: It is indeed sufficient to include only u;u arid Uyy in the error bound.
This will have valuable consequences in the following section for the theory
of rectangular isoparametric elements.
It may well happen that (7) holds inside each element for derivatives of
all orders s k. The limitation tos q in (8) arises when u1 has on]y q 1
derivatives across element boundaries.
1t would be useful to know sometliing about the constants es in the theo-
rem. They are a direct indication of the properties of the particular element;
if they are larger for one element than for another of the same degree, then
the :first element is comparatively inaccurate or "stiff." For piecewise l_inear
approximation on the line, with equally spaced nodes x 1 jh, these optimal
constants can be computed. There are two functions of special interest:
the :first trigonometric polynomial f(x) = sin nxfh that vanishes at every
meshpoint, and the first algebraic polynomial g(x) x 2 not reproduced by
its linear interpolate. In the case off, both its interpola te and its best linear
approximation are identically zero. Therefore, the error is f itself, and an
easy computation giv~s
1
-h!flz·
íC
It follows from the proof of Theorem l.l in Section 1.6 that these constants
are optimal, since for any u the interpolate u1 was shown to satisfy
Therefore, e 0 == Iín 2 and e 1 1/n for linear approximation.
What role is left for the quadratic g(x) x 2 ? It yields, for each fixed
function u, the constant which is asymptotieally eorreet as h O. In the
previous case we fixed h and sought out the worst function sin nxfh. In this
case we fix u and ask for the limits
lim min !zu - vh lo and . min u
(11) e1 = l1m --+-:---.----'-"-
h-o h 1 u lz h-+o
The mínimum is taken over all vh in S\ that is, over all piecewise linear func-
tions. For e~ery h these ratios are bounded by 1/n 2 and 1/n, respectively,
and therefore the limits cannot exceed these constants; we may expect them
to be smaller, since a fixed u cannot resemble all the oscillatory functions
sin nxfh at once. These new constants e 0 and e 1 (ifthey exist) seem to be more
148 APPROXIMATION CHAP. 3 SEC. 3.2.

natural in asses~ng the improvement to be' expected from mesh refinement

in a pi-actical prOblem, since it is the solutionwhich is fixed á.nd h which
changes.
Consider first the particular function u = g = x 2 • On the interval [-1, 1],
the hest linear approximations to displacement and slope minimize

~espectively. These expressions equal

2 4a 1 + 2a + 2ai
2
and 8 + 2a 2•
-1/3
3 1
3 2

In both cases a2 should be zero; the best approximatión to an even function

on a symmetric interval [~ 1, 1] is even. The optimal value of a 1 is j-, and the
ratios which enter e0 and e1 (with an interval oflength h = 2) become
Fig. 3.3 Error in linea
1
12,JT' The remarkable thing is that the
of the function u.
The latter happens not to be very different from 1/n; it is smaller, as it should
be. THEOREM 3.4
Notice that x 2 - j- is the second Legendre polynomial; it is the error For an arbitrary funetion u(x) in ~
after least-square approximation by linear functions ofthe quadraticg(x) h ~ O are the sqme as for the specia
x 2 •. Notice also that this error has the same value %at the two ends of the 1/2,..;-T.
interval. This makes it easy to connect it with the best approximation to x 2
on the next interval 1 x < 3. On this interval the error function after This theorem means that in the li
optimal approximation is just (x - 2)2 - j-, the same Legendre polynomial like a piecewise quadratic; the erro:
translated two units along the axis. This pattern continues: On every interval locally resembles Fig. 3.3. In other '
[2n 1, 2n + 1], the optimal error function x 2 vh is given by (x- 2n)2- j-, funetion, the more it looks like a polync
and the error function is periodic w.th period h 2 (Fig. 3.3). The con- for Taylor-series expansions. The c<
stants 1/12~ and 1/2¿'3- are the same on any union of these intervals. correct for least-square approximatic
Furthermore, these eonstants are dimensionally eorreet, and they will not u but rather to all choices. Similarly, e
ehange if the independent variable is resealed to give an arbitrary initial ínter val element error u - uh iri the second-01
[-h/2, h/2], and then translated to give an arbitrary posltion of the origin. e 1 will be dec_isive e ven in displaeemen
The x-mf;ls in the figure is scaled by a factor of h/2: 1 and the error function minimize the strain energy of the ern
on the y-axis is scaled by (h/2)2: l. It is easy to check that the ratios are in va~ er
indicates how rather than e0 enters
riant; the reason x 2 is so special is that translation to (x x 0 )2 alters'it by Theorem 3.4 will be proved first f1
a linear expression 2xx 0 - xfi, which can be taken up by the trial functions. of length h, with midpoint X 0 ,
Note that .the optimal approximation is by no means the linear interpolate;
the interpolate yields the right exponent of h but too large a constant. The
Ritz procedure achieves the best constant, beca use it minimizes o ver Sh.
 CHAP. 3 SEC. 3.2. MEAN-SQUARE APPROXIMATION 149

to be expected from mesh refinement

! is
solution which fixed and h which

nu= g x 2 • On the interval [- 1, 1],

lacement and slope minimize
x2 -1/3 (x- 2) 2 1/3

:ai 8
3 and
3 + 2ai. -1/3

:st approximatión to an even function

. The optimal value of a1 is -j-, and the
lterval of length h = 2) beco me
1
2,/j.
~nt from 1/n; it is smaller, as it should
Legendre polynomial; it is the error
tear functions ofthe quadraticg(x)
;ame value %at the two ends of the
it with the best approximation to x 2
:his interval the error function after
· - j-, the same Legendre polynomial
s pattern continues: On every interval
tion x 2 - vh is given by (x- 2n) 2 f,
:h period h = 2 (Fig. 3.3). The con-
tme on any union of these intervals.
·nsionally correct, and they will not
2/ed to give an arbitrary initial interval
'! an arbitrary position of the origin.
tctor of h/2: 1 and t~ error function
easy to check that the ratios are inva-
. translation to (x - x 0 )2 alters it by
m be taken up by the trial functions.
~ by no means the linear interpolate;
tt of h but too large a constant. The
.nt, beca use it minimizes o ver Sh.
. Fig. 3.3 Error in linear approximation to x 2 •
The remarkable ·thing is that the asymptotic constants are independent
of the function u.
THEOREM 3.4
For an ·arbitrary function u(x) in JC 2 , the limiting values of the ratios as
h O are the same as for the special function x 2 : c 0 1/12¿) and C1 =
1/2~.-
This theorem means that in the Iimit as h O, every function behaves
Iike a piecewise quadratic; the error function after linear approximation
locally resembles Fig. 3.3. In other words, the more closely you look at a
function, the more it looks like a polynomial. This is the underlying foundation
. for Taylor-series expansions. The constant c 0 is therefore asymptotically
correct for least-square approximation, not just to sorne extreme choice of
u but rather to all choices. Similarly, e1 is asymptotically correct for the finite
element error u - uh in the second-order example of Chapter l. In general,
c 1 will be dec.isive even in displacement error for u- u\ since uh is chosen to
minimize the strain energy of the error; Nitsche's argument in Theorem 1.5
indicates how cf rather than c0 enters the displacement error.
Theorem 3.4 will be proved first for smooth functions u. In each interval
of length h, with midpoint X 0 ,
150 APPROXIMATION CHAP. 3 SEC. 3.2.
1
The optimal (least-squares) linear approximation to the first three terms is boring. to a mathematician, property :
pro ved.
2
l(x) u(x 0 ) + (x- Xo)u'(xo) T·)
+ 31 ( 2h ) u"(x We point out now (and shall do so
x2 ~ j- are special points. Of course,
polynomial, and therefore appear .in
Using these linear functions l in each subinterval, we are effectively back to
[jh, (j + l)h) they move to (j + 1/2 ±
the quadr~tic case. The ratios in each interval, taking account of the O(h 3 )
however, they are special in another Wé
error from the remainder in the Taylor expansion, still satisfy
ratic vanishes at these points, and uh i~
(In collocation this is known to be tl
lu -ll 1 i~(l + O(h)). them as displacement points. There are a
which are of still greater importance. TI
Squaring, and summing over all the intervals, the piecewise linear function tives of the error function vanish (x =
L formed from all these pieces l satisfies in Section 3.4 that the stress errors a.
power o[ h.
(13) 1 + O(h) The previous theorem extends to a
12~.· n dimensions and more generally toan~
method. In n variables, there are sev'
There remains one difficulty: L is not continuous at the nades. The linear possibly associated with different erro
approximations l depend on Taylor expansions in their own interval and if xP happened to be present in S\ the
cannot be expected to connect with their neighbors. The discrepancy, how- Locally, we imagine the function u as
ever, is only of the order terms of degree k; those of degree k
space, and the approximation depends
_!_(!!_)
3 2
2
(u"(x 0 )
2
_ u"(x 0
2
+ h)) _
-
O(h 3 )
'
DPu of order k. This generalization OJ
following way, with matrices Ks in ph

and therefore an O(h 3 ) alteration in each linear piece l will produce a continu- THEOREM 3.5
ous L. (In the 1-norm the alteration is O(h 2 ).) After such an alteration, both Jf Sh is of degree k 1 on a reguü
(13) and its analogue in the 1-norm continue to hold for L. Therefore, as
matrices Ks such that for every u in X
h ---7 O, the ratios approach the constants 1/12~ and 1/2~. It is clear
that no other choice of vh in (11) could yield a smaller constant, beca use these (14) h 2 (s-k> min 1u if!;---+
constants ·were aiready correct when L was formed from the optimal l on sn la.

each subinterva1. The theorem is therefóre proved when u is smooth enough

The diagonal entries K~P can be
to permit the Taylor expansion (12).
The extension to all u in 3C 2 is a standard technical problem. We define
monomial u xP x1• ... x~n; for ~
derivatives D«u of order k are zero.
the linear operator P~, from 3C 2 to ocs, by
1t was convenient to square the e'
hs- 2 P~u = component of u perpendicular to Sh.
3.4; thus K 0 d = -rlo and K1 =
Two properties of this operator are airead y pro ved, that in two dimensions, the matrices Ks w
three derivatives a2 jax 2 , u2 /ux ay, ar
1. 1P!u ls < Cs! u 12 (Theorem 3.3, with k 2). spaces of arbitrary degree the const:
2. 1P!u ls ~es 1u 12 for smooth u (see previous paragraph). terms ofthe Bernoulli numbers; we bt:
spaces of corresponding degree. Also
By a standard density argument, which is uninteresting to an engineer and h and varying u, are known for splin
 CfiAP. 3 SEC. 3.2. MEAN-SQUARE APPROXIMATION 151

proximation to the first three terms is
r)
subinterval, we are\effectively back to
interval, taking acc9unt of the O(h 3 )
r expansion, still sadsfy
!u 11 .=
1 h u (1 + O(h)).
ntervals, the piecewise linear function
es
_ 1 + O(h)
. 12,J"S
wt continuous at the nodes. The linear
~xpansions in their own interval and
teir neighbors. The discrepancy, how-
h linear piece t'will produce a continu-
O(h2).) After such an alteration, both
·ontinue to hold for Therefore, as
mts l/12,J"S and 1/2~. It is clear
yield a smaller constant, because these
[ was formed from the optimal 1 on
:Ore proved when u is smooth enough
andard technical problem. We define
'by
boring to a mathematician, property 2 holds for all u and Theorem 3.4 is
pro ved.
We point out now (and shall do so again) that the zeros of the function
x 2 - t are special points. Of course, they are the zeros of the Legendre
polynomial, and therefore appear in Gauss quadrature-on an interval
[jh, (j + l)h] they move to (j + 1/2 ± 1/,.Jj")h. For finite element purposes,
however, they are special in another way: The best approximation toa quad-
ratic vanishes at these points, and uh is likely to be of exceptional accuracy.
(In collocation this is known to be the case; see Section 2.3.) We refer to
them as displacement points. There are also stress points, discovered by Barlow,
which are of still greater importance. These are the points at which the deriva-
tives of the error function vanish (x O in our simple example) and we show
in Section 3.4 that the stress errors at these points are smaller by an extra
power o[ h.
The previous theorem extends to any finite element on a regular mesh in
n dimens.ions and more generally to any example ofthe a'bstract finite element
method. In n variables, there are severa! derivatives DP of order l P1 = k,
possibly associated with different error constants in approximation. In fact,
if xP happened to be present in S\ the corresponding constant would vanish.
Locally, .we imagine the function u as expanded in a Taylor series through
terms of degree k; those of degree k - 1 are reproduced exactly by the trial
space, and the approximation depends asymptotically only on the derivatives
DPu of order k. This generalization of Theorem 3.4 can be expressed in the
following way, with matrices Ks in place of the scalar constants cs.
THEQREM 3.5
/f Sh is of degree k l on a regular mesh, there exist nonnegatil•e definite
matrices Ks such that for e very u in :JCk,
(14) h 2 (s-k) min 1 u - vh ¡; ~
Sh
L;
!cr.! =IPI=k
K~P J(Dcr.u)(DPu) dx.
The diagonal entries K~P can be determined by approximation of the
monomial u xP = x1' ... x~n, fór which DPu is a constant and the other
derivatives vcr.u of order k are zero.

:u perpendicular toS h. It was convenient to square the expressions which occurred in Theorem
3.4; thus K 0 e~= -,-k and K 1 ct = -(z-. For linear approximation in
ready proved, that in two dimensions, the matrices Ks will be of order 3, corresponding to the
three derivatives a2 jax 2 , a2 jax ay, and a2 jay 2 of order k 2. For spline
'ith k 2). spaces of arbitrary degree the constants K:P have been computed [S5] in
!e previous paragraph). terms ofthe Bernoulli numbers; we believe they are the same for the Hermite
spaces of corresponding degree. Also the minimal constants Cn for a fixed
h is uninteresting to an engineer and h and varying u, are known for ·splines (Babuska). In this case the extreme
152 APPROXIMAT~ON CHAP. 3 SEC. 3.2.

functions are again sines of wavelength 2h, their best approximations are turns out to be
identically zero, and the constants are es res-k.
It is possible to use Theorem 3.5 for a quantitative comparison of two (15) lu lflt
different1 polynomial elements, or of the same element with two different
geometr(cal configurations. Consider two regular triangulations of the plane, On this basis, the two configuration!
one involving diagonals in both directions, the other in only one direc- case (A) is twice as great; in other wo
tion (Fig,>·, 3.4). Combinatorially the two are quite different. In (A) sorne side h/.../2 instead of h. With this el
and configura/ion (B) is better by ara¡
error, with the number of free param
(B) by a factor of .JT. By symmetr:
For the twist term xy, the two configu
and there is nothing to choose.
These calculations are confirmed
in the engineering literature, which fa·
of higher order a computer could de1
(a) (b) finite element problem, for which tht
same time be computing the leading
Fig. 3.4 Two possible tríangulations. finite difference scheme which arise~
and 3.4).
nodes are connected to four neighbors, and others to eight; in (B) every node Theorem 3.5 has also an importa
has six neighbors. Consider Courant's space Sh of continuous piecewise linear
functions on these triangles. Since (A) has twice as many nodes as (B), the COROLLARY
dimension of S~ will be twice that of S'B. Furthermore, the space S~ contains To achieve approximation of order
S'B and is therefore at least as good in approximation. The question is Sh on a regular mesh must be of degr
whether it is twice as good, to compensate for having twice the number of element method converges, in case' of,
parameters. if k > m; this is the constan! strain e
In two dimensions, the three quadratic monomials are x 2 , xy, and y 2 • m must be present in Sh.
For each of them we propose to find the element uh of Sh which minimizes
1u - uh j 1 • This uh will be exactly the finite element solution for Poisson's This corollary is a converse, on a r
equation, when the true solution u is a quadratic. it, suppose that Sh were only of deg
We begih with u= x 2 , on a square of side h = 2 centered at the origin. and not in Sh. Then by Theorem 3.4,
From considerations of symmetry, the optimal uh takes the same value at the
four vertices, say rt in (A) and p in (B); at the center, it must equal P in case
(B) but may ha ve a different value y in (A). This means that in configuration
(B), uh is a constant on this particular square~the same constant P ! Therefore, the order of approximatio
as in the -ene-dimensional case computed abo ve. The error over the mesh corollary is proved. Clearly the theor'
square is the former applies to all u, while for t
approximation of the lowest-degree
1u - uh Ir = ff (x 2
- P)~ + (x 2 - p); = fJ(2x)2 dx dy = l.
1
space.
The conclusion is rather remarkat
This equals h2 1 u IV12, exactly as in one dimension: K~ 1 = fi. In case (A) of polynomials. It implies that piece1
there is a minimization with respect to y; the constant must be smaller, and tions, not only for their convenience bu,
 , CHAP.
:gth
es 2h,ns-k.
their best approximations are
' ,....,
for a quantitative comparison of two
the same element with two different
wo regular triangulations of the plane,
:ctions, the other in only one direc-
two are quite differe:nt. In (A) sorne
. .r..,
!
( and configuration (B) is better by a ratio of 12: 9. The coefficient of x 2 in the
~--------,/t&-
(b}
le triangulations.
and others to eight; in (B) every node
;pace Sh of continuous piecewise linear
1 has twice as many nodes as (B), the
·~. Furthermore, the space S~ contains
l in approximation. The question is
1sate for having twice the number of
iratic monomials are x 2 , xy, and y 2 •
:he element uh of Sh which minimizes
finite element solution for Poisson's
quadratic.
: of side h = 2 centered at the origin.
optimal uh takes the same value at the
: at the center, it must equal p in case
(A). This means that in configuration
r square-the same constant p t
1ted above. The erro-r over the mesh
p); = JJ(2x) 2
dx dy i3l •
e dimension: K~ 1 fz-.. In case (A)
y; the constant must be smaller, and
3 SEC •.3.2. MEAN-SQUARE APPROXIMATION 153
turns out to be
(15)
On this basis, the two configurations can be compared. The dimension in
case (A) is twice as great; in other words, it effectively works with squares of
side h/VT instead of h. With this change in (15), the constant becomes .¿-,
error, with the number of free parameters equalized, will be smaller in case
(B) by a factor of By symmetry this is true also of the y 2 coefficient.
For the twist term xy, the two configurations turn out to be equally effective,
and there is nothing to choose.
These calculations are confirmed by numerical experiments, reported
in the engineering literature, which favor the configuration (B). For elements
of higher order a computer could determine the constants K«P by solving a
finite element problem, for which the true solution is u= xP. lt will at the
. same time be computing the leading terms in the truncation error for the
finite difference scheme which arises on a regular mesh (see Sections 1.3
and 3.4).
Theorem 3.5 has also an important theoretical consequence.
COROLLARY
To achieve approximation of order hk-s to the sth derivative, a tria! space
Sh on a regular mesh must be of degree at least k - l. Therefore, the finite
element Ínethod converges, in case of a differential equation of order 2m, only
if k > m; this is the constant strain condition, that al! polynomials of degree
m must be present in Sh.
This corollary is a converse, on a regular mesh, to Theorem 3.2. To prove
it, suppose that Sh were only of degree l 1, l < k. Let x« be of degree l
and not in Sh. Then by Theorem 3.4,
hs-t min 1xa: - vh ls -+ constant :;t:O.
Therefore, the order of approximation to xa: is only l s, not k s, and the
corollary is proved. Clearly the theorem is much stronger than its corollary;
the former applies to all u, while for the latter it is only necessary to consider
approximation of the lowest-degree polynomial which is not in the trial
space.
The conclusion is rather remarkable: everything depends on the presence
of polynomials. It implies that piecewise polynomials are the best tria! func-
tions, not only for their convenience but also for their approximation properties.
154 APPROXIMATlON
K.rv-., ,.
· CHAP. 3 SEC. 3.2.

From the beginning~ the finite element method has worked with subspaces We conclude this section with soJJ
of the ·optimal kind. regular mesh, in other words, on appt
The result of the corollary can be proved directly [810], and because of element method. It is natural to appro
its importance we shall do so in the simplest case. We show that for approxi- transform. The mean-square norms f
mation of order h, the constant function 1 must lie in the trial space. Suppose, seval's formula to JJ1e~~u. 12 de,and thl
for examp!e, that the roof function were replaced by the cosine (Fig. 3.5). produces zeros in its Fourier transf<
generates alllinear polynomials; its tr•
Roof function Cos function has zeros of order k = 2 at all the p
spline of arbitrary degree k 1, the
the result of convolving the box funct
between polynomials of degree k 1
was discovered by Schoenberg in his
covered three times in the finite elemt
In the papers [SS, SIO] and in Au
the abstract finite element method in .
The corollary stated earlier, that on
Fig. 3.5 A non-polynomial shape function. k - 1 to achieve approximation of
methods-together with the existenc
The trial space contains all combinations v11 2:: qi Cos [n(x- jh)/2h], in Section 3.1. We confess to a prof
the notation Cos implying that the, cosine is truncated outside its central arch cannot be included in full detail, but
1(J 1 n/2. It might appear that, in the limitas h ~O, there would be little results which are unique to the case e
difference in comparison with piecewise linear approximation. important to the general theory-the
Suppose, however, that we attempt to approximate u= 1 on a unit ínter- lary, the finite difference aspect of K~
val. In the linear case, this function lies in the space. In the new case it does discussion of condition number in Cl
not, and the error function Eh will be periodic of period h. We ignore boun- Permit us to return to the main n:
dary conditions, or rather assume them to be periodic, since in any case they e1 the differerice between u and it~ in te
have only a secondary effect on approximation in the interior of the interval.
Thus the error is (16)

min 111 -
S"
vh 115 = J (E"(x)) dx.
1

O
2
What happens if there is a singulari
from belonging to the space JCk? Wi
This integral covers 1/h periods, and over each period its value is K 0 h, where gence will definitely be reduced; if u 1
K 0 is a constant independent of h. If h is changed, the error function (justas sq uare sense, the error in energy will
in Fig. 3.3) is simply scaled to its new period. Therefore, the approximation And the pointwise error in the strains
error is a constant K 0 (as predicted in Theorem 3.5) and does not decrease any improvement is possible by "g1
with h. The Cosine space has no approximation properties whatever. varying the mesh width h to produce
Of course, this adverse conclusion does not apply to the ordinary cosines, · A useful rule of thumb, when it is
which are among the most valuable trial functions in the Ritz method. In a lar trial functions, is suggested by fo
sense they are of infinite accuracy, k = oo, since they take advantage of each so as to keep h7-m 1u lk,e, roughly the St
additiona1 degree of smoothness of the solution u to be approximated. Since dimension, with an x~~ singularity at
each cosine i~ nonzero over the whole interval, this case does not fall in the should be approximately constant; th
finite element framework and the restriction that the space must contain to be. It appears that this rule haE
polynomials to be effective in approximation is no longer in force. order of accuracy can be achieved Jo
 CHAP.
nt method has worked with subspaces
~ preved directly [SIO], and because of
mplest case. We show that for approxi-
)n 1 must lie in the trial space. Suppose,
Nere replaced by the cosine (Fig. 3.5).
Cos function
Iomíal shape function.
1ations v" 2::: q¡ Cos [n(x- jh)/2h],
ine is truncated outside its central arch
e limit as h O, there would be little
se linear approximation.
to approximate u =
1 on a unit inter-
s in the space. In the new case it does
periodic of period h. We ignore boun-
1 to be periodic, since in any case they
jmation in the interior of the interval.
' J~ (E"(x)) 2
dx.
ver each period its value is K 0 h, where
s changed, the error function (just as
period. Therefore, the approximation
Theorem 3.5) and does not decrease
dmation properties whatever.
oes not apply to the ordinary cosines,
al functions in the Ritz method. In a
oo, since they take advantage of each
solution u to be approximated. Since
interval, this case does n.ot fall in the
riction that the space must contain
1ation is no longer in force.
3 SEC. 3.2. MEAN-SQUARE APPROXIMATION 155
We conclude this section with sorne historical remarks on the case of a
regular mesh, in other words, on approximation theory in the abstract finite
element method. It is natural to approach this problem by use ofthe Fourier
transform. The mean-square norms JJ l D<~-u !2 dx can be converted by Par-
seval's formula to JJ lea-a !2 de, and the condition that rp generate polynomials
produces zeros in its Fourier transform. The roof function, for example,
generates alllinear polynomials; its tra~lsform is tp((j (sin e/2/e/2) 2 , which
has zeros of order k = 2 at all the points e ±4n, .... For the B-
spline of arbitrary degree k 1, the exponent in tp simply becomes k; it is
the result of convolving the box function with itself k times. The connection
between polynomials Of degree k - 1 Ín X and zeros Of order k at e 2nn
was discovered by Schoenberg in his first paper on splines [S2] and redis-
covered three times in the finite element context [G5, B2, F6].
In the papers [SS, S lO] and in Aubin's book [4], the Fourier analysis of
the abstract finite element method in n dimensions is pursued very seriously.
The corollary stated earlier, that on a/ regular mesh Sh must be of degree
k - 1 to achieve approximation of order hk, was first proved by Fouri~r
methods-together with the existence of the superfunction t¡1 referred to
in Section 3.1. We confess to a profound regret that this Fourier analysis
cannot be included in full detail, but it is only realistic for us to select those
results which are unique to the case of a regular mesh and at the same time
important to the general theory-the asymptotic theorem 3.5 and its corol-
lary, the finite difference aspect of KQ F described in Section 3.4, and the ·
discussion of condition number in Chapter 5.
Permit us to return tq the main resuJt of this section, that in the element
e¡ th.e difference between u and its interpola te satisfies
(16)
What happens if there is a singularity in the solution u, which prevents it
from belonging to the space X k? With a regular mesh, the order of conver-
gence will definitely be reduced; if u possesses only r derivatives in the mean-
square sense, the error in energy will decrease like h 2 <r-m) instead of h 2 <k-m).
And the pointwise error in the strains is visibly worse. The question is whether
any improvement is possible by "grading" the mesh, in other words, by
varying the mesh width h to produce a finer mesh near the singularity.
A useful rule of thumb, when it is inconvenient to introduce special singu-
lar trial functions, is suggested by formula (16): The grading should be done
so as to keep hf-m 1 u lk, e, roughly the same from one element to the next. In one
dimension, with an xrx singularity at the origin, this means that hk-m+ 112 xa-~k
should be approximately constant; the Iarger x is, the larger h= L\x is a~lowed
to be. It appears that this rule has a remarkable consequence: the same
order of accuracy can be achieved for a singular as for a regular solution u,
156 APPROXIMATION , CHAP.3 SEC. 3.3. CURVED ELEMENTS AND 1~

by properly grading the mesh. In other words, suppose that for an irregular to be satisfied? The isoparametric tec
n-dimensional mesh with N elements we compute an average mesh size ii polynomials to define the coordinate tran.
from the formula Niin vol n. Then the correct grading can achieve speaking, isoparametric means that the
1u u1 lm,n O(iik-m), even though u is singular and 1u lk,n = oo. for coordinate change as for the tria]
means that a subset of lower-degree ~
require continuity between elements at
3.3. CURV~,D ELEMENTS AND ISOPARAMETRIC The fundamental isoparametric ex
TRANS.FORMATIONS
from a square to a quadrilateral (Fig.
The basic idea is simple. Suppose it is intended to use a standard polyno- y
mial element, for example one of those defined on triangles or rectangles in
Section 1.8. Suppose also that the regions into which n is subdivided are
not of the proper shape; they may have one or more curved sides, or they (0,1)
may be nonrectangular quadrilaterals. By changing toa new /;-11 coordinate
system, the elements can be given the correct shape. The element stiffness
matrices are then evaluated by integrations in the new variables, over triangles (0,0}
or rectangles, and minimization leads to a finite element solution uh(l;, 11)
which can be transformed back to x and y.t
There are several points to watch. First, since a typical two-dimensional
element integral is transformed by

Fig. 3.6 Isoparametric rr

(17)
--j.- ff p(x(l;, 11), e, 11))(vel:x + v,11x)2J(I;, 11) di; d11,
E¡
y(
S and Q is given by

(18)
the coordinate change and its derivatives must be easily computable. Further-
more, the change must not distort_the element excessively, or the Jacobian
determinant J XeY'l - x,ye may vanish within the region of integration; Each edge of the square S goes linea
it is surprisingly easy for this to happen. Excessive distortion will also destroy For example, if 11 O and 1; varies fr·
the accuracy built into the polynomial element. Polynomials in the new vari- linearly from one corner (xp y 1 ) to the
ables do not correspond to polynomials in the old, and the requirement in the location ofthe other vertices (x 3 , J
order to preserve the approximation theory is that the coordinate changes This guarantees that any conforming
should be uniformly smooth. Finally, in order that conforming elements in bicubic Hermite element, for which 1
e-11 shall be conforming in x-y, there is a global continuity condition on the parametric, respectively-will still be
coordinate cllange:Ifthe energy involves mth derivatives, then the coordinate We must check that the mapping
change must be of class em-l between elements. For the present we discuss each point (x, y) in Q corresponds too
only the case m 1, arising from second-order differential equations, where the equations (18) for 1; and 11 in term
the mapping must be continuous between elements: A point common to e; square roots and lead nowhere. Inste
and ei must not split into two separate points when ei E¡, ej--+ boundaries of S and Q correspond, v
How are these properties, especially the requirement of computability, is nonvanishing inside S:

tThe isoparametric technique is equally irnportant in three dirnensions. It is sirnpler

x2
J = XeY, x,Ye =
to discuss exarnples in the plane, but there is no difference in the theory. l Y2-
 CHAP. 3 SEC. 3.3. CURVED ELEMENTS AND ISOPARAMETRIC TRANSFORMATIONS 157

words, suppose that for a;n irregular to be satisfied? The isoparametric technique consists in choosing piecewise
we compute an average ~esh size fí polynomials to define the coordina{e transformations x(e, t¡) and y(e, t¡). Strictly
~n the correct grading can achieve
speaking, isoparametric means that the same polynomial elements are chosen
s singular and 1 u lk,n = oo. for coordinate change as for the trial functions themselves; subparametric
means that a subset of lower-degree polynomials is used. In either case we
~RAMETRIC
require continuity between elements and nonvanishing of the Jacobian.
The fundamental isoparametric example is the bilinear transformation
from a square to a quadrilateral (Fig. 3.6). The coordinate change between
is intended to use a standard polyno-
defined on triangles or reot~ngles in y 7] y
ions into which Q is subdivl'ded are
1
1e one or more curved sides, or they (0,1) ( 1' 1)
By changing toa new e-11 coordinate
correct shape. The eleinent stiffness
ms in the new variables, o ver triangles
1
(0,0)
S
,, X

to a finite element solution uh(e, t¡) s'

dy.t
;irst, since a typical two-dimensional

Fig. _3.6 Isoparametric mappings to quadrilaterals.

S and Q is given by

: must be easily computable. Further-
element excessively, or the Jacobian
sh within the region of integration;
Excessive distortion will also destroy
lement. Polynomials in the new vari-
s in the old, and the requirement in
eory is that the coordinate changes
1 order that conforming elements in
a global continuity condition on the
: mth derivatives, theq_the coordina te
:lements. For the present we discuss
j-order differential equations, where
en elements: A point common to e;
>oints when e; --)oE¡, ei Er
--)o
' the requirement of computability,
(18},
x(e, t¡) = x 1 + (x x1)e + (x3 -
2 - x1)11 + (x4- x3 - Xz + x1)e11,
y(e, t¡) = Y1 + (Yz- Y1)e + (y3 - Y1)t¡ + (y4- Y3 - Yz + Y1)et¡.
Each edge of the square S goes linearly into the corresponding edge of Q.
e
For example, if t¡ = O and varies from O to 1, then the point (x, y) m oves
linearly from one corner (xP y 1) to the next corner (x 2, y 2). On this boundary,
the location of the other vertices (x 3 , y 3 ) and (x 4 , y 4 ) has absolutely no effect.
This guarantees that any conforming element in and t¡-say a bilinear ore
bicubic Hermite element, for which the mapping is isoparametric or sub-
parametric, respectively-will still be conforming in x and y.
We must check that the mapping (18) is invertible, in other words, that
each point (x, y) in Q corresponds to one and only one pair (e, t¡) in S. Solving
e
the equations (18) for and t¡ in terms of X and y Will introduce COmplicated
square roots and lead nowhere. Instead, since it is already verified that the
boundaries of S and Q correspond, we have only to show that the Jacobian
is nonvanishing inside S:

portant in three dimensions. It is simpler

to difference in the theory.
_ _ -j
J - Xr:Y, x,Yr: - .
x2 - x1 + At¡ x3 - x1 + AeJ: j,
Yz - Y1 + B17 Y3 - Y1 + B~.,
158 APPROXIMATION CHAP. 3 SEC. 3.3. CURVED ELEMENTS AND

where A= x 4 x 3 - x 2 + x 1 and B y 4 - y 3 - y 2 + y 1 • The Jacobian a 6 r¡ 2 is determined by its values at tl

is actualiy linear, and not bilinear, since the coefficient of er¡ in this 2 X 2 vertices and three at the midpoints.
determinant is AB AB O. Therefóre, if J has the same sign at al! four We imagine two mappings of ti
.corners of S, it cannot vanish inside. At the typical corner e = O, r¡ = O, the plane. First a simple linear map non
Jacobian is two straight edges onto the coordina
is a constant, and this step is simpl
is to connect x'-y' to e-r¡, associa·
point (X', Y') with the midpoint (~
By the cross-product formula this equals it~_~in (}, where the lengths 1 and 1' (X', Y') will often be chosen halfway
and the angle (} a~e shown in the figure. Therefore, J is positive at this corner The mapping is bilinear:
provided the interiQr angle (} is less than n. The same is true at every other
x'=e+!
corner. Consequently J is nonvanishing if and only if the quadrilateral Q is (19)
convex: AH its angles must be less than n. Otherwise, as for Q" in the figure, y' r¡ + 1

J will change sign somewhere inside S. In this case the coordinate change
is illegal. It is easy to verify that the straight e
Notice that even though polynomials in x and y do not generally trans- adjacent elements is assured. The Ja1
form into polynomials in e and r¡, the linear po~ynomia/s ], X, and y are specia/.
The coordinate transformation itself expresses x and y as bilinear functions
1 + (4X'-:
J
e
of and r¡, and of course constant functions remain constant. Since these 1 (4Y' 2)t¡
three polynomials líe in the trial space, the convergence condition is guar- = 1 + (4X'- 2
anteed to hold-all solutions u =a + Px + yy of constant strain are
reproduced exactly in sn. This is always the case for isoparametric mappings, Again the Jacobian is linear. It equal
and convergence is assured. The subparametric case is even better: If the in the triangle if and only if it is posit
e
trial space contains all biquadratics or bicubics in and r¡, then it contains tion, first shown to us by Mitchell, i
all quadratics or cubics in x and y, and Sh has degree 2 or 3, respectively.
Therefore, assuming that the angles of Q arebounded away from O and n, At (0, 1): 4X'
approxtfnation is possible to the full degree and the error in the strain is At (1, 0): 4Y'
O(hk- 1), as it should be.
The most important example for triangles is one with a curved edge, such Therefore, (X', Y') may lie anywhe1
as might arise at the boundary of r (Fig. 3.7). The simplest curve possible lines, and correspondingly (X, Y) sh
is one of second degree, and a natural choice· of element is the quadratic. On Notice that even if the original trian
the e-r¡ triangle, the trial function vh al + a2e + a3r¡ + a4e 2 + aserJ + lie in the middle half of its edge or s
a vanishing Jacobian. (Of course ir.
y y' shift it; on a straight triangle we e
variables, even with arbitrarily plac
is really intended for the case when
In this example the curved si de w
ric case, either with triangles or qw
are given by the same type of poly11
~ents, and all edges may be polynt
will be the same as those for the ele
sorne derivatives at a node, this me~
Fig. 3.7 Mappings of a curved triangle. the boundary curves must be contim

, -\~ .
'= Y4- YJ- Yz + Y 1 • The Jacobian
tce the coefficient of ~r¡ in this 2 x 2
ore, if J has the same sign at al! four
t the typical corner ~ = O, r¡ = O, the
:tls 11' sin B, where the lengths 1 and /'
. Therefore, J is positive at this corner
an re. The same is true at every other
~ if and only if the quadrilaterh_l Q is
1 re. Otherwise, as for Q" in the figure,
~. In this case the coordinate change
1ls in x and y do not generally trans-
rear polynomials 1, x, and y are special.
(presses x and y as bilinear functions
nctions remain constant. Since these
!, the convergence condition is guar-
+ Px + f'Y of constant strain are
s the case for isoparametric mappings,
arametric case is even better: If the
bicubics in ~ and r¡, then it.contains
.d Sh has degree 2 or 3, respectively.
· Q are bounded away from O and re,
jegree and the error in the strain is
mgles is one with a curved edge, such
;'ig. 3.7). The simplest curve possible
hoice of element is the quadratic. On
= al + a 2 ~ + a 3 r¡ + a 4 ~ 2 + a 5 ~r¡ +
r¡
---x'
· a curved triangle.
SEC. 3.3. CURVED ELEMENTS ANO ISOPARAMETRIC TRANSFORMATIONS 159
a6 r¡ 2 is determined by its values at the six nodes of the figure, three at the
vertices and three at the midpoints.
We imagine two mappings of the original curved triangle in the x-y
plan e. First ·a simple linear map normalizes the curved triangle, putting the
two straight edges onto the coordinate axes in the x'-y' plane. The Jacobian
is a constant, and this step is simply for convenience. The important step
is to connect x'-y' to ~-r¡, associating an arbitrarily specified boundary
point (X', Y') with the midpoint (!, !). In practice, (X, Y) and therefore
(X', Y') will often be chosen halfway along the curve, but this is not required .
The mapping is bilinear:
x' = ~ + (4X'- 2)~r¡,
(19)
y'= r¡ + (4Y'- 2)~r¡.
It is easy to verify that the straight edges are preserved, and continuity with
adjaceut elements is assured. The Jacobian is simply
J = 11 + (4X' - 2)r¡ (4X' - 2)~ -~
(4Y'- 2)r¡ 1 + (4Y'- 2)~
= 1 + (4X'- 2)r¡ + (4Y'- 2)~.
Again the Jacobian is linear. lt equals 1 at ~ = O, r¡ = O, and will be nonzero
in the triangle if and only if it is positive at the other two vertices. This condi-
tion, first shown to us by Mitchell, is simply
At(O,l): 4X'-2>-l, or X'>t,
At (1, 0): 4Y'- 2 > -,---1, or Y'>!-
Therefore, (X', Y') may lie anywhere in the quadrant formed by the dashed
lines, and correspondingly (X, Y) should lie in the sector shown in the figure.
Notice that even if the original triangle were straight, the point (X, Y) must
lie in the middle half of its edge or shifting it to the midpoint would produce
a vanishing Jacobian. (Of course in this case there would be no reason to
shift it; on a straight triangle we could use quadratic elements in the x-y
variables, even with arbitrarily placed midedge nodes. The mapping to ~-r¡
is really intended for the case when a curved edge needs to be straightened.)
In this example the curved si de was a para bola. In the general isoparamet-
ric case, either with triangles or quadrilaterals, the mappings x(~, r¡), y(~, r¡)
are given by the same type of polynomial elements as are used for displace-
ments, and all edges may be polynomials of degree k - l. The constraints
will be the same as those for the element itself-when the urtknowns include
sorne derivatives at a node, this means that the corresponding derivatives of
the boundary curves must be continuous there. The Lagrange case is therefore
160 APPROXIMATION CHAP. 3 SEC. 3.3. CURVED ELEMENTS AN

much the simplest for isopararnetric transformations, since the only un- of terms, for example,
knowris are function values and the only constraint is the continuity between
elements, which is required in any case. Iri fact, it is especially simple if, as
in the serendipity rectangular e1ement in Fig. 3.8, there are no interna! nodes. ·
y y
~=1
~=1/2 \
~=O \
\
In general a derivative of order l Pl
' ' ..... ,,7] = 1
/'~ = 1/2
' ',,7] = 1/2
where the constant ll Fllk is comp1
//~= 1 y~, x,l' .... within the element up to
argument we are following, write ot
of the Jacobian J also enters the tra
~--------------------x ~-----------------------x

Fig. 3.8 Common isoparametric elements.

The mapping between the boundaries then determines the entire coordinate This holds for alll P1 k, and the1
change, which is otherwise highly sensitive to the movement of interna! nodes.
We must emphasize that the whole isoparametríc technique depends on
the use of numerical integration (in the e-r¡ variables) to compute the entries
of K and F. From the change of variables in the element integral (1 7), it is
obvious even for an isotropic material (p constant) that the mathematical This yields an upper bound for the
equivalent of variable material properties is introduced by the functions For the left sirle we want, a lo
ex, 1'/x, and J(e, r¡). In general the first two are rational functions and the last which led to (21) is reversed:
is a polynomial, with smoothnessdepending on the distortión of the element.
We shall establish in Section 4.3 how the numerical integration error affects (22)
the final result and what order of accuracy is required.
Here we deal with the question of approximation: How closely can iso- This time the factor 11 p-t¡¡s depen
parametric elements match the true solution u(x, y)? The answer must depend s of the in verse mapping p-t, takin!
on the size of the derivatives in the coordinate change F: J is nonvanishing, as we assume,
in terms of derivatives ofF. For th~
FCe, r¡) = cxce, r¡), y( e, r¡)). identity for the Jacobian matrices

Let ú(e, 11) denote the solution u(x, y) transformed to the new coordinates.
Then if Sh is of degree k e
1 in and r¡, the interpolate ú1 satisfies

(20) For higher derivatives this identity i

again becomes the boundedness of
Changing to the x-y variable's, each derivative of ú is transformed into a sum Now we assemble the results.
 CHAP. 3 SEC. 3.3. CURVED ELEMENTS AND ISOPARAMETRIC TRANSFORMATIONS 161

transformations, since. the only un- of terms, for example,

y constraint is tb;e continuity between
:. In fact, it is especia1Iy simple if, as
1 Fig. 3.8, there are no interna! nodes.

y
!=112 !=~
!=O \
\
\
\
1
1
In general a derivative of order l p1 k will be bounded by

k
l DPu(e, 11) 1 < 11 Fllk L: l D"u(x, y) l,
1)11=1

where the constant 11 Fllk is computed from powers of the derivatives x,,
y,, x,, ... whhin the element up to order k. Ciarlet and Raviart [C7], whose
argument we are following, write out this expression in detail. The reciproca!
of the Jacobian J also enters the transformation of the integral:
~~---------------------x

>arametric elements. f E¡
IDPul2ded11<
-
.IIFII~
~~n IJ(e, 'f/) 1
f e¡
i:ID)Iul2dxdy.
1

hen determines the entire coordinate

ve to the movement of interna! nodes. This holds for alll P1 k, and therefore
isoparametric technique depends on
;-11 variables) to compute the entries (21) 11 u 112k, E,
A 11 Fll~
min IJ l 11 u 112k, e,·
>les in the element in.tegral (17), it is
p constant) that the mathematical This yields an upper bound for the right side of (20).
ties is introduced by the functions For the left side we want a lower bound, and therefore the argument
vo are rational functions and the last which led to (21) is reversed:
Iing on the distortión of the element.
e numerical integration error affects (22)
:tcy is required.
pproximation: How dosel y can iso- This time the factor 11 p-t lis depends on powers of derivatives up to order
ion u(x, y)? The answer must depend S ofthe inverse mapping p-t, taking X, y toe, 11· Ifthe Jacobian determinant
·dinate change F: J is nonvanishing, as we assume, then derivatives of p-t can be bounded
11), y( e, r¡)). in terms of derivatives ofF. For the first derivatives, this is expressed by the
identity for the Jacobian matrices
ransformed to the new coordinates.
(Yexe Y,x,)-
1
the interpolate íi1 satisfies •

For higher derivatives this identity is differentiated, and the essential question
again beco mes the boundedness of the deriva ti ves ofF.
ative of íi is transformed into a su m
1
Now we assemble the results. Substituting (21) and (22), the original
162 APPROXIMA TI ON CHAP. 3 SEC. 3.3. CURVED ELEMENTS AND 1:

late u1 in Sh satisfies
inequality 11 u u1 lls < Chf-s 11 u]]k on the element E¡ is con verted toa similar
inequality"on e¡: ,
(Jn!u U¡]2dx;
(23)

where the constant is now

(J n 1grad(u U1 )1 2 dx

c(m~x]JI)
112
C'
mmiJI IIFII k IIF- 1 11·s The constan! C' is given in (23). The re
solution uh, for a second-order differenti
in strain energy.
This is the fundamental result. The order of approximation is the same
with isoparametrics as with the standard polynomials, provided C' remains The inequalities in the theorem ar
bounded. Note that T1¡ is still the diameter of the element in the 1;-r¡ plane, derivatives, that is, to the cases s = O
even though the approximation inequality (23) is now in x-y. We may, how- order hk-s holds also for higher derivat
ever, suppose the diameters to be the same in the two coordinate systems. elements are joined the interpolate u1 is
(A change of scale in 1;,-r¡ lea ves the inequality unchanged, as it must; if h¡ only to ;JC 1 over the whole domain O.
becomes rxh¡, then the norms ofF and F- 1 are multiplied by rx-k and rxs, The uniformity required by the the
respectively.) This will be a convenient normalization, since with equal amounts essentially to this:
diameters the isoparametric approximation problem reduces exactly to the
boundedness ofF and its derivatives, and the nonvanishing of the Jacobian l. The Jacobians should be bound
J, uniformly as h O. to be bounded away from O and n, an
There is an important modification in the quadrilateral case. 1t is needed 2. The edges of the elements in the
even for the map defined by uniformly bounded derivatives, i.e., thc
r¡ - l;,r¡. along the edges must remain bounde
x(l;,, r¡) = ?;, y(l;,, r¡)
2h displaced only by O(h 2 ) from straight
ha ve 11 Fllk < constant.
This transformation takes the three points (0, 0), (h, O) and (0, h) into them-
selves and moves the fourth corner (h, h) ofthe square toa new point (h, h/2). Computations by Fried [F14] have
In other words, this is a completely typical map, transforming the square tion 2. He increased the curvature <
in the 1;,-r¡ plane into a quadrilateral of comparable size and shape. Neverthe- from the original square was ofthe sar
less, the cross derivative y;; 11 , which enters the norm 11 F\lk in the approximation On a unit square this still implies a b
inequality, is of order 1/h. This would destroy the order of approximation. . size h the curvature (being a second de
Ciarlet and Raviart were able, however, to make use of the presence of the results were correspondingly poor. P
twist term l;,r¡ in the trial space; bilinear interpolation reproduces ~t exactly, between these overly distorted element
in addition to the linear terms a 1 + a 2 /; + a 3 r¡ of the triangular case. Their by the uniformity hypothesis in theth
result is that-for quadrilaterals, the cross derivatives Xe 11 and Ye 11 do not appear On the happy side, suppose that a <
in the factor 11 Fl]k, and the expected order of approximation is possible. A of Q; its one curved si de follows the
similar conclusion holds for biquadratics and bicubics. a polynomial approximation to r,
The consequences of the inequality (23) are summarized in the following specified nodes.) Then if r is smooth
wi/1 automatically be, satisfied. The in
THEOREM 3.6 only by O(h2 ) from a straight line, a
Suppose the ¿;-r¡ tria! space is of degree k - 1, and the element transforma- terms ofthe curvature ofr. Therefore,
tions Finto x-y are uniformas h ----> O. Then if u belongs to ;JCk(Q), its interpo.:. order equations, allows essential boundc
 CHAP.3
1e eJement E, is converted to a ~-iJar
/2
11 Fllk 11 p-lJJs•
order of approximation is the same
·d polynomials, provided C' remains
ter of the eJement in the e-1(}:>-lane,
ty (23) is now in x-y. We may, how-
;ame in the two coordinate systerhs.
~quality unchanged, as it must; if h;
F- 1 are multiplied by rx-k and r:t.S,
nt normalization, since with equal
tion problem reduces exactly to the
1d the nonvanishing of the Jacobian
n the quadrilateral case. It is needed
;, 11) e11.
2h
1ts (0, 0), (h, O) and (0, h) into them-
of the square to a new point (h, h/2).
pical map, transforming the square
>mparable size and shape. Neverthe-
the norm 11 Fllk in the approximation
lestroy the order of approximation.
to make use of the presence of the
interpolati~n reproduces it exactly,
+ a 3r¡ of the triangular case. Their
derivatives x~, and Y~11do not appear
jer of approximation is possible. A
~ and bicubics.
~3) are summarized in the following
~k 1, and the element transforma-
en if u belongs to Xk(Q), its interpo-
SEC. 3.3. CURVED ELEMENTS AND ISOPARAMETRIC TRANSFORMATIONS 163
late u1 in Sh satisjies
The constan! C' is given in (23). The rate of convergence of the finite element
solution uh, for a second-order dif.ferential equation, is therefore the full Jzz<k-n
in strain energy.
The inequalities in the theorem apply onJy to the function and its first
derivatives, that is, to the cases s = O and s 1 of (23). Approximation of
order Jzk-s holds also for higher derivatives within each element, but when the
elements are joined the interpola te u1 is no more than continuous, and belongs
only to 3C 1 over the whole domain n.
The uniformity required by the theorem is a rather severe condition, and
amounts essentially to this:
l. The Jacobians should be bounded away from zero, forcing all angles
to be bounded away from O and n, and
2. The edges ofthe elements in the x-y plane should be polynomials with
uniformly bounded derivatives, i.e., the curvatures and the higher derivatives
along the edges must remain bounded. In particular, the edges should ~e
displaced only by O(h 2 ) from straight Iines. Then a well-chosen map F wtll
have-,¡¡ Fllk < constant.
Computations by Fried [Fl4] have demonstrated the necessity of condi-
tion 2. He increased the curvature of the edges until their displacement
from the original square was ofthe same order as the size ofthe square itself.
On a unit square this still implies a bounded curvature, but scaled down to
. size h the curvature (being a second derivative) is of order 1/h. The numerical
results were correspondingly poor. Practical problems will fall somewhere
between these overly distorted eJements and the nearly straight ones required
by the uniformity hypothesis in the theorem.
On the happy side, suppose that a curved element occurs at the boundary
of Q; its one curved side follows the true boundary r. (It will normally be
a polynomial approximation to r, interpolating the true boundary at
specified nodes.) Then if r is smooth, the uniformity cond!tion on th~s edge
will automatically be satisfied. The interpolating polynomtal edge wlll vary
only by O(h 2 ) fro~ a straight line, and the curvatures will be bounded in
terms ofthe curvature ofr. Therefore, the isoparametric technique,for second-
order equations, allows essential bbundary conditions to be handled with no loss
164 APPROXIMATION CHAP. 3 SEC. 3.4.

in accuracy or in simp/icity over natural boundary conditions. The same remarks coordinates. The mapping in the figure (
apply to an interna! boundary. mite elements, for example, allowing thc.:
The improvement which this technique allows, in comparison with ap- tangents, and in the Hermite case also ti
proxim~ting the boundary by a polygon, is enormous. In his experiments be continuous at the vertices. This is a
with cubic triangular elements, for example, Zlamal [ZlO] found a difference theoretical limitation on the isoparame1
of an ordyr of magnitude in displacements, and even more in strains. There Iems.
can be nordoubt ofthe technique's success in second-order problems.t The message seems to be this, that
For fo~rth-order equations-plate and shell problems, for example- order equations rather than second-ord
the position is much less happy. The theory insists that the coordina te change tion of unknowns may be of overwhelm
be e 1 : Its first derivatives should be continuous between elements, or the The construction of e 1 polynomials o
trial functions are nonconforming. (Convergence is still possible for non- difficult, and curved elements-especiall
conforming elements, as we prove in the next chapter. However, it appears cal integration-are infinitely worse.
. ---
(
that for plate elements even this hope is dashed, because they cannot pass
the required patch test.) A e 1 coordinate change is theoretically possible,
but the extra continuity requirement is extremely restrictive (Fig. 3.9). Once 3.4. ERROR ESTIMATES

In this section we put the previous

b order to achieve the main goal of our '
element error u- uh. The function u
elliptic boundary-value problem of ord
S tion computed in a finite element spa<
element equations KQ F become a S)
shall find at the same time the order o
e p a tions.
The basic question is how well th€
admissible space :re~. In the energy non
as possible to u, and the error in energy
x-y plane { r¡ ptane (24)

Fig. 3.9 The constraints on a C 1 mapping. This is the simplest and still the most j
expression a(v, v) for strain energy is p1
two directions are known at a point-at P, the tangents to PA and P B are the problem is elliptic: a(v, v) ¿ ullvl
established when S is mapped to Q-all other directions are completely
determined. [For any e 1 function f(x, y), all directional derivatives can be (24')
computed from the gradient, that is, from the derivatives fx and !Y in two
directionsl In the figure, this means that the tangent to PC is determined. In displacement, or more generally in t
In fact, the curve CPA must have a continuous tangent at P, since epa is whether u - uh is again of the optím(]
straight. It follows that a general quadrilateral or triangular mesh, e~·en one everything is determined by the degree
with straight edges, cannot be carried into a regular mesh by a e 1 change of is not quite true. If we fix Sh and increa
eventually the trial functions will no lonJ
tGordon an_d Hall recently proposed to map n all at once, rather than element by will collapse. Therefore, the order of a•
element, into a square. For this they use blending functions, a variant of the standard finite
as well as on k and s.
elements. If n is itself not too unlike a square (the mapping of a circular region would
create artificial singularities at the corners) this should save time as compared to the The correct order can be determine
isoparametric method on each element. due to Aubin and Nitsche. The argm
 CHAP. 3
al boundary conditions. The same remarks
:~nique allows, in_~9mparison with ap-
lygon, is enormous. In his experiments
xample, Zlamal [ZIO] found a difference
~ments, and even more in strains. There
1ccess in second-ordet, problems. t
lte and shell probleins;;Jor example-
theory insists that the coordina te change
1e continuous between elenients, or the
(Convergence is still possible for non-
1 the next chapter. However, it appears
pe is dashed, because they éarmot pass
dinate change is theoretically possible,
: is extremely restrictive (Fig. 3.9). Once
b
S tion computed in a finite element space Sh. On a regular mesh, the finite
e p
~- "7 plane
aints on a C 1 mapping.
-at P, the tangents to PA and P B are
?-all other directions are completely
~, y), all directional derivatives can be
from the derivatives fx and fy in two
that the tangent to PC is determined.
continuous tangent at P, since epa is
zdrilateral or triangular mesh, e~'en one
into a regular mesh by a e 1 change of
> map Q all at once, rather than element by
tding functions, a variant of the standard finite
tare (the mapping of a circular region would
) this should save time as compared to the
SEC. 3.4. ERROR ESTlMATES 165
coordinates. The mapping in the figure could be carried out by bicubic Her-
rnite elements, for example, allowingthe curved edges to be cubics; but their
tangents, and in the Hermite case also the cross-derivatives xe, and Ye,, musí-
be contimwus at the vertices. This is a severe, and almost an unacceptable,
theoretical limitation on the isoparametric technique for fourth-order prob-
Iems.
The rnessage seems to be this, that it is a mistake to work with fourth-
order equations rather than second-order systems. Analytically the elimina-
tion of unknowns may be of overwhelming importance, but not numerically.
The construction of e 1 polynomials on straight-sided elements is already
difficult, and curved elements-especially when we cometo consider numeri-
cal integration-are infinitely worse.
3.4. ERROR ESTIMATES
In this section we put the previous approximation theorems to use, in
order to achieve the main goal of our whole analysis-to estímate the finite
element error u uh. The function u is the solution to an n-dimensional
elliptic boundary-value problem of order 2m, and uh is its Ritz approxirna-
element equations K Q F become a system of difference equations, and we
shall find at the same time the order of accuracy of these difference equa-
a tions.
The basic question is how well the subspaces Sh approximate the full
admissible space 3C~. In the energy norm~ nothing else matters: uh is as close
as possible to u, and the error in energy must be of the optimal order h2 <k-m):
(24)
This is the simples! and still the most fundamental error estímate. Since the
expression a(v, v) for strain energy is positive definite (in other words, since
the problem is elliptic: a(v, v) > a 11 v 11!), (24) is equiyalent to
(24')
In displacement, or more generally in the sth derivative, the first question is
whether u- uh is again of the optimal order hk-s. This would imply that
everything is determined by the degree k - 1 of the finite elements, but it
is not quite true. If we fix Sh and increase the order 2m of the problem, then
eventually the tria! functions will no longer be admissible and the Ritz method
will collapse. Therefore, the order of accuracy must somehow depend on m
as well as on k and s.
The correct order can be determined by an elegant variational argument
due to Aubin and Nitsche. The argument has come to be known as the
166 APPROXIMATION CHAP. 3 SEC. 3.4.

Nitsche trick; it was exten~ed by Schultz, and is now the standard approach Substituting (24') and (29) into (28)
to errors in displacement. Ín Section 1.6, we proved an h2 error with linear
elements; the general case is going to be more technical, but the results (25-
j(g, u if)!
26) are very simple. The upper limit 2(k- m) on the rate of convergence was
found by the first author [SIO].

THEORENI, 3. 7
Suppose that the finite element space Sh is of degree k - 1 and that the By duality in the definition (equation 1
strain energy has smooth coefficients and satisfies the ellipticity condition
v)
u 11 V 11! a(v, < Kll V 11!. Then the finite element approximation uh differs
from the true solution u by

(25) jjcí('CL Uh lis< Chk-s 11 Ullk if s >2m- k, The proof is complete. Tt is most st
(26) !!u- uhlls Ch (k-m)l!ul!k
2
if s 2m k. s = O, because in this case the auxiliar:
was made in Section 1.6.
These exponents are optimal, so that the order of accuracy never exceeds
2(k- m) in ~my norm; in almost al! realistic cases the order is k s. A similar result holds with inhomo~
[S6]. The error estimates have also b
Proof Nitsche's trick is to introduce an auxiliary problem Lw = g, whose vibration, in which the basic ellipticity
variational form (equation of virtual work) is · by the addition of a new zero-order ter:
equation Lu f is altered to Lu- fZU
(27) a(w, v) (g, v) for all v in Xe. values of L, then L fZ is no longer p·
has a solution. Schultz has proved [S'
From the theory of partial differential equations, there exists a uniquesolution changed.
w which is 2m derivatives smoother than the data g: 11 w llzm-s e 11 gll-s·
We choose v u - uh in (27): Remarks. The case s m, corres
yields the correct power !Jk-m. The erre
!(g, u- uh)! = !a(w, u uh)j !a(w- v\ u uh)! included in the theorem as it stands, si
(28)
<Kllw vhllmllu-uhllm· sense. Nevertheless the exponent k
provided the subspaces. Sh satisfy an iJ
This holds for any vh in S\ since a(v\ u - uh) = O by the fundamental Ritz of a tria! function vh increases its mm
theorem 1.1. Now suppose that vh is the closest approximation to w in the hypothesis is almost the same as the
:rem norm. [Or, what is effectively the same, let vh be the finite element solution there it was a case of c/h;; therefore t
to the auxiliary problem (27) and therefore the best approximation to w in elements are of comparable size. Othj
the energy. Note that only approximation in energy is used in the proof, in the error estímate for derivatives of
nevera direct approximation in the :res norm.] According to the approxima- The theorem and proof are valid v
tion theorem 3.3, ingly, the rafe of convergence in negativt
if k s,
2m The reason is this: the error in the I-
(29) o ver the domain:
if k< 2m- s.

In the first case k was reduced to 2m - s before the approximation theorem 11 u ifli-1
was applied; if the subspace is complete through degree k - l, then it is
certainly complete through any lower degree. if we choose v l. Therefore, in the
 CHAP. 3 SEC. 3.4. ERROR ESTIMATES 167

hultz, and is now the standard approach Substituting (24') and (29) into (28),
n l. 6, we pro ved a[LS error with linear
o be more technical, but the results (25-
2(k m) on the rate of convergence was

pace Sh is of degree k - 1 and that the By duality in the definition (equation 1.58) of negative norms,
1ts and satisfies the ellipticfty condition
1e finite element approximation uh differs max l(g, u- uh)i
g 11 gil-s

if s >2m- k, The proof is complete. It is most straightforward for the displacernent,

if s <2m- k. s = O, because i·n this case the auxiliary data gis exactly u uh; this choice
was made in Section I .6.
'1t the order of accuracy never exceeds
·ealistic cases the order is k s. A similar result holds wi_th inhomogeneous essential boundary conditions
[S6]. The error estimates have also been extended to problems of forced
luce an auxiliary problem Lw = g, whose l'ibration, in which the basic ellipticity condition a(v, v) > a 11 v ll! is violated
1 work) is by the addition of a new zero-order term. Effectively, the original differential
equation Lu = f is altered to Lu - rxu = f: if a fáHs in between two eigen-
for all v in X~. values of L, then L a is no longer positive definite-but the equation still
has a solution. Schultz has proved [S4] that the rate of convergence is un-
1equations, there exists a unique solution changed.
than the data g: 11 w ll 2 m-s e 11 gll-s·
.Remarks. The case s = m, corresponding to error in energy, always
yield~ the correct power hk-m. The error in higher derivatives, s >m, is not
uh) 1 = 1a(w vh, u uh)l
included in the theorem as it stands, since w líes outside xm and (29) is non-
uhllm· sense. Nevertheless the exponent k s will be correct within each element,
provided the subspaces Sh satisfy an inverse hypothesis: each differentiation
·h, u - uh) O by the fundamental Ritz of a trial function vh increases its maximum by at most a factor cfh. This
the closest approximation to w in the hypothesis is almost the same as the uniformity condítion (2), except that
;ame, let vh be the finite element solution there it was a case of c/h¡; therefore the in verse hypothesis is fulfilled if all
erefore the best approximation to w in elements are of comparable size. Otherwise a factor (hmaxlhminyr-m appears
nation in energy is used in the proof, in the error estímate for derivatives of order s > m.
es norm.] AccordingJo the approxima- The theorem and proof are valid without change in cases< O. Surpris-
ingly, the rafe of convergence in negative norms is not only of academic interest.
if k> 2m- s, The reason is this: the error in the 1-norm is a bound for the average error
if k 2m s. o ver the domain:

- s before the approximation theorem 11 u - uh 11 = max 1

J (u- uh)v dx
0
j
..!...---'==---....,.,----,-,------'-
j J(u ~ if) dx 1'

)Iete through degree k - 1, then it is -t llvllt (vol Q) 112

· degree.
if we choose v __ l. Therefore, in the usual case k > 2m, Theorem 3.7 has
168 APPROXlMATION CHAP.3 SEC. 3.4.

the following consequence for s = 1: The .average error is much smaller mesh this problem can be solved exac
than the typical displacement error ata point. More precisely,. by the property that the true stresses (
approximations (derivatives of a lower-
J
but j (u- uh) dx 1
2
,....., hl<k+l). elements in one dimension, equality oc<
that is where the slope of a quadratic
polate. (Equivalently, that is where t
This mu~t mean that the error alternates rapidly in sign. In fact, it does so horizontal tangent.) Symmetry suggest
within eaeh element, and the practica} problem is to discover even approxi- tional for elements ofhigher degree. (Tl
mately the "special points" where these sign changes occur. Near such points lie elsewhere, and in the simplest cast
the displacements uh will be of exceptional accuracy. Legendre polynomial. They seem more
(We had imagined that -u" == 1, with linear elements, would provide than the stress points.) In two qimens
the perfect example: uh coincides with theinterpolate u1 , so there is exceptional
choice of Pk, and exceptional points fe
accuracy at the nod.es.t Looking more closely, however, the average error exceptional for the others. The midpoin
is only of the same order h2 as the error at a typical point. In fact, the linear tional for derivatives along the edge bt
interpolate never goes above the true (quadratic) solution, and the error is the normal. This is still an area for resc
completely one-sided. The explanation is that the condition k > 2m for the will be more completely understood.
hk+ 1 phenomenon does not hold: k 2m = 2. The no des are exceptional,
but it is because the trial functions solve the homogeneous differential equa-
Three additional problems arise on
tion -u" O [H6, T4]; this is not an example of rapid alternations in sign.)
l. To interpret KQ = Fas a syster
The stresses are in a similar position. In second-order problems they are corresponding local truncation errors.
in error by hk- 1 at typical points, and hk on average; therefore, these errors
2. To prove that the exponents in '
also must alternate in sign, and there must exist exceptional stress points.
3. To show that for smooth soluti
Their presence was noticed in actual computations by Barlow, and it appears apply not only in the mean, but at ever:
for quadratics on triangles that the midpoints of the edges are exceptional.
The accuracy at these midpoints is better than at the vertices, where even
W e shall not attempt a technical d
after averaging the results from adjacent elements, the stress approximations
Roughly speaking, once the behavior o;
are not good. · Since the midpoints are also nodes for quadratic elements,
the central question is the stability ofth~
the situation is extremely favorable; it is only spoiled by errors dueto change
is difficult to establish in the maximum 1
in domain, which do not necessarily alternate in sign.
mentioned above it is almost certain to l
We believe that the stress points can be located in the following way. The
coefficient model problems, and the er
leading term in the error is govemed by the problem of approximating poly-
then ofthe correct arder. More precise r
nomials pk of degree k, in energy, by the trial functions in Sh. On a regular
Bramble, and Ciarlet and Raviart, but 1
tThe phenomenon of "super-convergence" at the nodes was recently made clearer. by For problem 1 suppose first that ther
an elegant observation of Dupont and Douglas. Let Go(x) be the fundamental solut10n unknown) for every mesh square; this v
corresponding to the point x 0 ; G0 is the response toa point load /o = ó(x xo). Then right triangles, bilinear elements on s
1u(x 0) uh(x 0) 1:::;;;; C 11 u - uh llm 11 Go - vh llm for any vh in S h. will Iook exactly like a conventional d
Proof: u(x 0) uh(x 0) (u- uh,/0) a(u- uh, Go) a(u uh, Gd- vh). The most
to innumerable discussions. about the
interesting case occurs if x 0 is a node, since the approximation of Go is likely to be finite differences. It is clear that not all ~
especially good; in the, one-dimensional case -u" /, Go is linea~ wit~ a change i~ slo~e by an appropriate choice of element; t
at x 0, and can be reproduced identically by vh. This confirms the mfimte accuracy m this positive definite, and even under these
special case. N:ormally, the term 11 Go - vh 11 will add sorne finite power of h to the hk~m may not exist~ On the other hand, a suffi
coming from 11 u - uh llm· The whole question of pointwise convergence, and of these.m-
to regard all finite element equations-e
creases in the power of h (super-convergence) at special points, is now being studied very
intensively ~ nodal unknowns-as finite difference e
 CHAP. 3 SEC. 3.4. ERROR ESTIMATES 169

-1 : The average error is much smaller mesh this problem can be solved exactly. Then stress points are identified
a point. More precisely, , by the property thatthe true stresses (derivatives of Pk) coincide with their
approximations (derivatives of a lower-degree polynomial.) With first-degree
lt J
1 (u- uh) dx r 1'1 h2(k+1).
elem:nts in one dimension, equality occurs at the midpoints of the intervals;
that IS where the slope of a quadratic matches the slope of its linear inter-
polate. (Equivalently, that is where the error function in Fig. 3.3 has a
rzates rapidly in sign .. Jn fact, it does so
horizontal tangent.) Symmetry suggests that the midpoints are also excep-
al problem is to discover even approxi-
tional for elements of higher degree. (The exceptional points for displacement
:se sign changes occur. Near such points
líe elsewhere, and in the simplest case they were the zeros of the second
>tional accuracy.
Legendre polynomial. They seem more sensitive to the boundary conditions
1, with linear elements, wouÍd',provide
than the stress points.) In two qimensions, the results may depend on the
the interpolate u¡, so there is exceptional
choice of Pk, and exceptional points for one stress component need not be
)fe closely, however, the averag~ error
exceptional for the others. The midpoints of an edge seem likely to be excep-
ror at a typical point. In fact the linear
tional for derivatives along the edge but not for stresses in the direction of
e (quadratic) solution, and the error is
the normal. This is still an area for research; eventually these special points
m is that the condition k > 2m for the
will be more completely understood.
2m 2. The nodes are exceptional,
)}ve the homogeneous differential equa-
Three additional problems arise on a regular mesh:
~ example of rapid alternations in sign.)
l. To interpret KQ = Fas a system of finite difference equations, with
IOn. In second-order problems they are
corresponding local truncation errors.
i fzk on average; therefore, these errors
2. To pro ve that the exponents in Theorem 3. 7 are optimal.
re must exist exceptional stress points.
3. To _show that for smooth solutions, the same rates of convergence
;omputations by Barlow, and it appears
apply not only in the mean, but at every individual point.
midpoints of the edges are exceptional.
better than at the vertices, where even
We shall not attempt a technical discussion, particularly of problem 3.
ent elements, the stress approximations
Roughly speaking, once the behavior of the truncation errors is established,
Lre also nodes for quadratic elements
~he central question is the stability ofthe difference operator K. This property
t is only spoiled by errors dueto chang~
IS difficult to establish in the maximum norm, but with the inverse hypothesis
alternate in sign.
mentioned above it is almost certain to hold. We have verified it for constant-
an be located in the following way. The
coefficient model problems, and the error at every point of the domain is
by the problem of approximating poly-
then ofthe correct order. More precise results ha ve been achieved by Nitsche,
the trial functions in Sh. On a regular
Bramble, and Ciarlet and Raviart, but the general problem is unsolved.
;e" at the nodes was recently made clearer by For problem 1 suppose first that there is only one node (and one associated
Jglas. Let Go(x) be the fundamental solution unknown) for every mesh square; this will be the case for linear elements on
sponse to a point load / 0 = ó(x - x 0 ). Then right triangles, bilinear elements on squares, and splines. Then KQ = F
m11 Go vh llm for any ~hin S h. will look exactly like a conventional difference equation. This fact has led
rz(u - uh, Go) a(u- ub, G~- vh). The most to innumerable discussions about the relation between finite elements and
ce the approximation of G0 is likely to be finite differences. It is clear that not all difference equations can be produced
: -u''=/, Go is linear with a change in slope by an appropriate choice of element; the matrix K must be symmetric and
vh. This confirms the infinite accuracy in this positive definite, and even under these restrictions a corresponding element
will add sorne finite power of h to the hk-m may not exist. On the other hand, a sufficiently tolerant reader inay be willing
•n of pointwise convergence; and of these in-
:::) at special points, is now being studied very to regard all finite element equations-even on an irregular mesh, with many
nodal unknowns..:_._as finite diffetence equations. We agree with this view.
170 APPROXIMATION CHAP. 3 SEC. 3.4.

The system KQ F is in general a novel kind of coupled difference equa- We may regard Sh as generated by th
tion, which in principie could' have been devised without the variational Iess useless functions. Forming the co
principie as an intermediary. Historically, of course, that almost never ence equations KQ F, we can rewr
happened; the finite element method leads systematically toa special class of difference equations having a very sp
equations-the intersection between al! possible difference equations and al/ is an accurate analogue of the origit
possible ~itz-Galerkin equations-which are astonishingly successful in com- M 1 (coming from the functions in
putations) are completely inconsistent. (This C(
lf the original differerítial equation Lu f has constant coefficients the large M, and Fourier analysis is the (
order of the localtruncation error may be tested as follows: Let f(x) be a is that the Ritz method attaches almo
pure exponential é'x, so that u can be found explicitly and substituted into tion which corresponds to 'JI, and tha
the difference equation. The truncation error wi.ll have the forro eu:xE(h, .;). agrees with the exponent min (k, 2(
With -u" f, fo~ example, the solution is u= eiexgz, and the truncation order of accuracy (in displacement) (
error of h- 2 (-uj;t "=r 2uj- uj_¡) is

In general the coefficient E can be computed in terms of the Fourier trans-

forms of the trial functions. (Here we are right at the heart of the abstract
finite element method; this technique would be impossible on an irregular
mesh.) Expanding E as a pbwer series in h, the leading exporient is exactly
the order of accuracy of the difference equatiQn. We have computed this
exponent and verified that it is the smaller of k and 2(k m); the rate of
convergence given in Theorem 3.7 is correct.
Suppose, finally, that there are M unknowns associated with each mesh
square, so that the finite element equation KQ F becomes a coupled sys-
tem of M difference equations. The unknowns may be the values of uh at
distinct nodes, or function values and derivatives at a multiple node. This
introduced no complications when the error was estimated by a variational
argument, in Theorem 3.7; that result depends only on the order of approxi-
mation achieved by Sh, and any additional facts about the subspace were
irrelevant. The difference equation aspect, however, becomes much more
subtle when M > l.
Briefly, the problem is that the truncation errors in the difference equations
are not al/ of the expected order. Therefore, it will not do simply to estímate
these errÓrs and then, applying stability to invert the matrix K, to convert
them into estimates of the error u - uh. The point is that uh is given by a
special combination of the trial functions, and if other combinations make
little or no contribution to the problem of approximation, then also their
contribution to uh turns out to be small. We recall that in the abstract method,
(])~' •.• , (])M- genera te approximation of order k if and only if it is possible
to construct from them a single function 'JI with the property (5) required in
Theorem 3.2, that is, a function which by itself is adequate for approximation.
 CHAP. 3 SEC. 3.4. ERROR ESTIMATES 171

tovel kind of coupled d\fference equa-
been devised without tHe variational
·ically, of course, that · almost never
eads systematically to a special class of
ll possible difference equatfons and all
:h are astonishingly ~uccessful in éom-
~ ' :-.
n Lu = f has constarit coefficients the
:ty be tested as follows: Let f(x) be a
! found explicitly and substituted into
n error wiH ha ve the form e'd: E(h,, . ~).
tion is u = e't:x /e 2 , and the truncation
~ is
We may regard S" as generated by this superfunction t¡1 and M 1 more or
less useless functions. Forming the corresponding combinations ofthe differ-
ence equations KQ = F, we can rewrite our finite element system as a set of
difference equations having a very special form: one equation of the system
is an accurate analogue of the original differential equation, and the other
M- 1 (coming from the functions in S" which are useless for approximation)
are completely inconsistent. (This construction becomes very technical for
large M, and F ourier analysis is the only possible tool.) The final conclusion
is that the Ritz method attaches almost all weight to the one difference equa-
tion which corresponds to tp, and that the order of accuracy of this equation
agrees with the exponent min (k, 2(k m)) in Theorem 3.7. This is the
order of accuracy (in displacement) of the finite element method.

mputed in terms of the Fourier trans-

. are right at the heart of the abstract
would be impossible on an irregular
; in h, the leading exponent is exactly
ce equation. We have computed this
naller of k and 2(k- m); the rate of
orrect.
unknowns associated with each mesh
ction KQ F becomes a coupled sys-
mknowns may be the values of uh at
l derivatives at a multiple node. This
: error was estimated by a variational
depends only on the order of approxi-
tional facts about the subspace were
;pect, however, becomes much more

cation errors in the difference equations

fore, it will not do simply to estímate
ity to invert the matfix K, to convert
uh. The point is that uh is given by a
ons, and if other combinations make
·m of approximation, then also their
We recall that in the abstract method,
)[ order k if and only if it is possible
)n t¡1 with the property (5) required in
y itself is adequate for approximation.
 SEC. 4.1. VIOLAT

following way: The essential boundaJ

boundary nodes. Between the nodes, a
of following a general boundary and
interpolated. We show in Section 4.4 t
Condition 3 is the most violated of
I(v) exactly and very easy to compute
enter in two ways: The integrals over 1
cal integration, and the domain of intt
union of simple element shapes. In bo1

4 VARIATIONAL CRIMES
tional /*(v): Therefore, the mathemati<
ence of the minimizing function uh 01
paraboloid /(v) is deformed, how far
Each of these three possible violat
are to justify the finite element methoc
it may be useful to call attention speci
and o ver throughout this chapter; th
tLree problems. The first idea is to cor
4.1. VIOLATIONS OF THE RAYLEIGH-RITZ CODE variation-in physical terms, the equ<
rules are violated, the minimizing fun'
The one cardinal rule in the Ritz theory is that the trial functions should
be admissible in the variational principie. In our notation, each vh should
belong to the space :JC'E, and uh should minimize l(vh). This rule is simple ~o
state, but it is broken every day, and for good reason. In fact, there are three In the Ritz case this is compared with
conditions to the rule, and all three present computational difficulties-per-
haps not insuperable difficulties, but seríous ones: a(u, v) ([, v)

l. The trial functions should possess m derivatives in the mean-square If Sh e 3C'E, and if 1 = l * so that the a:
sense, and therefore be of class e m- 1 across element boundaries. tion can be subtracted from the other;
2. Essential boundary conditions should be respected.
3. The functional I(vh) = qr Kq 2qr F should be computed exactly.

Our goal is to analyze the consequences of violating these conditions. In our case this expression is not i:ero
The first is violated by nonconforming elements, and we show in the next that the quaqtity which does appear is
section that in this case convergence mayor may not occur. It is by no means The second idea applies specifically
automatic (or even probable) that the discrete problem is consistent with the the special properties of polynomials.
continuous one. Instead, there is a patch test to be applied to the trial func- test applies to polynonÍial solutions. ·
tions, whiéh determines whether or not they consistently reproduce states of integration, where accuracy depends o
constant strain. If so, convergence takes place within each element. are integrated exactly. And we note
Essential boundary conditions are fulfilled, at least along an approxi- analyzing changes of domain: Polynm
mate boundary, by the isoparametric or subparametric elements. However, region between the given domain n al
there are many circumstances in which these elements are not available, We emphasize that the analysis doe
either because they are too complicated for.the programat hand or because bation in the stiffness matrix, from K
the problem itself is too complicated__.:a full fourth-order system of shéll large-the equations may seem compl
equatíons, for example. In such cases the rule may be partly satis:fied, in the polynomials of degree m are properl~

172

IMES
-RITZ CODE
1eory is that the trial functions should
:iple. In our notation, each ?J't should
l minimize l(vh). This rule is simple to
::>r good reason. In fact, there are three
~esent computational difficulties-per-
~rious ones:
ess m derivatives in the mean-square
cross element boundaries.
hould be respected.
2qr F should be computed exactly.
uences of violating these condítions.
ing elements, and we show in the next
:zy or may not occur. It is by no means
di serete pro blem is consistent with the
eh test to be applied to the trial func-
;t they consistently reproduce states of
es place within each element.
: fulfilled, at Ieast along an approxi-
or subparametric elements. However,
eh these elements are not available,
d for.the program at hand or because
-a full fourth-order system of shell
.he rule may be partly satisfied, in the
2
SEC. 4.1. VIOLATIONS OF THE RAYLEIGH-RITZ CODE 173
following way: The essential boundary conditions may be imposed at the
boundary nodes. Between the nodes, a polynomial trial function is incapable
of following a general boundary and thus the essential conditions are only
interpolated. We show in Section 4.4 that éonvergence still occurs.
Condition 3 is the most violated of all, since it is inconvenient to compute
J(v) exactly and very easy to compute it approximately. The approximations
en ter in two ways: The integrals o ver each element are computed by numeri-
cal integration, and the domain of integration Q is itself altered to become a
union of simple element shapes. In both cases I(v) is replaced by a new func-
tional.J*(v). Therefore, the mathematical problem is to determine the depend-
ence of the minimizing function uh on the functional itself: If the convex
paraboloid I(v) is deformed, how far does the mínimum move?
Each of these three possible violations requires a careful analysis, if we
are to justify the finite element method as it is actually used. Before we start;
it may be useful to call attention specifically to two ideas which appear over
and o ver throughout this chapter; they are common to the analysis of all
tLree problems. The first idea is to consider always the vanishing of the first
variation-in physical terms, the equation of virtual work. Even when the
rules are violated, the minimizing function u~ still satisfies
for all ?J't in Sh.
In the Ritz case this is compared with
a(u, v) = (/, v) for all v in 3C'E.
If Sh e 3C'E, and if 1 = 1* so that the asterisk can be removed, then one equa-
tion can be subtracted from the other; as in Theorem 1.1,
for all ?J't in Sh.
In our case this expression is not zero, and the whole problem is to show
that the quaQtity which does appear is small.
The second idea applies specifically to finite elements; it consists in using
the special properties of polynomials. We have noted already how the patclf
test applies to polynomial solutions. The situation is similar in numerical
integration, where accuracy depends on the degree of the polynomials which
are integrated exactly. And we note one property which will be useful in
analyzing changes of domain: Polynomials cannot vary tremendously in the
region between the given domain Q and its approximation Qh.
We emphasize that the analysis does not depend on estimating the pertur-
bation in the stitfness matrix, from K to K or K*. This change may be very
Iarge-the equations may seem completely altered-while still the piecewise
polynorttials of degree m are properly dealt with. In abstract terms, these
 \
J
174 VARIATIONAL CRIMES CHAP.4 SEC. 4.2.

polyn<;>mials become dense in the admissible space as h -? O, and their finite element so/ution ui is sti/1 identi
behavior is therefore decisive. that the problem has constant coeffic
have only an O(h) effect on ui.
There is a celebrated example in [
4.2. NONCONFORMI.NG ELEMENTS ANO ments of shape (B), but is violated b:
THE PATCH TEST (This is in the large-scale strain enei
25 %.) The element in question is a e
A n~~ber of frequently used elements are nonconforming:_their deri-
at the vertices. The tenth unknown
vatives of .order m 1 ha ve discontinuities at the element boundaries-and
coefficients; the authors disapproved
nevertheless they work quite well. Or r-ather, they sometimes work well and
xy 2 , beca use this constraint can bre~
sometí mes not. It is a risk that is taken most often in fourth-order problems,
nodal parameters of xy(I -x-y):
where the elements should lie in e 1 ; matching the slopes between elements
the constraint is satisfied, but the poi
can be difficult for the normal displacement w of a plate under bending, and
fore the authors chose a condition wt
extremely difficult for shells. Yherefore, the technique has been to compute
and never singular.
the energies within each separate element in the usual way and then simply to
add together the results. This has the effect of replacing the true functional
I(v) by a sum of element integrals

I*(v) = I; [ae(v, v) 2(f, v)el a*(v, v)- 2(f, v)*.

e

The difference between 1 and /* is that the singularities on the element bounda-
ries are ignored in/*, whereas I(v) = oo for nonconforming elements.
The Ritz approximation u~ is the (probably nonconforming) function in
the trial space which minimizes 1*(v11 ). This property of u~ is expressed as
(a)
usual by the vanishing of the first variation of 1*:
Fig. 4.1 Success and .
(1)
With irregular triangles, the odds a
This is the approximate equation of virtual work. It is identical to the usual
test. Nevertheless, an error of a few p
equation, except that again the integrals are computed an element at a time
when it is committed in the directio
and then summed, ignoring discontinuities between elements.
explained in Section 1.10, the latter a
Our goal is to find conditions under which this finite element approxi-
a(u11 , uh) < a(u, u); the nonconformin
mation ut in spite of its illegitimate construction, converges to u. This ques-
displacements as well as the strains a1
tion was almost completely obscure until Irons [B9] had a simple but
One approach to the theory is te
brilliant idea, now known as the patch test. Suppose that an arbitrary patch
KQ =Fas a finite difference equati<
of elements is in a state of constant strain: u(x, y)= P m(x, y), a polynomial of
rived. If this equation is consistent a
degree ni~ Then since this polynomial is present in Sh-ev~n on nonconform-
For a regular mesh, this is exactly th
ing elements the constant strain condition is imposed, that the degree k 1
for consistency. t Normally the noncon:
of the subspace must be at least m-a true Ritz solution uh would coincide
derivative, so that v 11 lies in em-z, anc
identically with P m· (At the boundary of the patch, the conditions imposed
are chosen to be consistent with constant strain; e.g. uh = P m is imposed on tRemember that for difference equatior
the displacements at the patch boundary.) Then the test is to see whether, first few terms of a Taylor series-in other
in spite of shifting from 1 to 1* by ignoring the inter-element boundaries, the certain degree. The patch test does exactly t
 CHAP.4 SEC. 4.2. NONCONFORMING ELEMENTS 175

tdmissible space as h -----4 O, and their finite element solution ui is still identic_al with. P m· W~ o:ay and shall assume
that the problem has constant coeffic1ents, smce vanatwns over an element
have only an O(h) effect on ui.
There is a celebrated example in [B9] in which this test is p&ssed for ele-
ANO ments of shape (B), but is violated by about 1.5% for those of shape (A).
, (This is in the large-scale strain energy; the pointwise stress errors reached
-~ .
25 %.) The element in question is a cubic, with v, vx, and vy as ~arameters
rnents are nonconfor.ining-their deri-
at the vertices. The tenth unknown is elirninated by a constramt on the
mities at the elernent boundaries-and
coefficients · the authors disapproved of equating the coefficients of X 2Y and
xy2 becau~e this constraint can break down on a right triangle-all nine
rather, they sometirnes work well and
n rnost often in fourth-order pro_plerns,
nodal parameters of xy(I -x-y) are zero on the standard triangle, and
rnatching the slopes between elements
the constraint is satisfied, but the polynomial is not identically zero. There-
:ernent w of a plate under bending, and
fore the authors chose a condition which would be invariant under rotation,
re, the technique has been to compute
and never singular.
ent in the usual way and then simply to
: effect of replacing the true functional

the singularities on the element bounda-

oo for nonconforming elernents.
(probably nonconforrning) function in
'). This property of ui is expressed as
:iation of 1*: {o) {b)

Fig. 4.1 Success and fail~re in the patch test.

With irregular triangles, the odds are strongly against success in the patch
irtual work. lt is identical to the usual
test. Nevertheless, an error of a few per cent may be acceptable, particularly
als are computed an elernent at a time
when it is committed in the direction opposite to the true Ritz error. As
mities between elements.
explained in Section 1.10, the latter always comes from too stiff a structure,
der which this finite element approxi-
a(u\ uh) a(u, u); the nonconforming solution ui is more relaxed, and the
onstruction, converges to u. This ques-
displacements as well as the strains are frequently overestimated.
until Irons [B9] had a simple but
One approach to the theory is to look directly at the discrete system
1 test. Suppose that an arbitrary patch
KQ F as a finite difference equation, forgetting that it is improperly de-
tÍn: u(x, y) p m(X, y), a polynomial of
rived. If this equation is consistent and stable, then ui must converge to u.
is present in Sh-eveá on nonconforrn-
For a regular mesh, this is exactly the content of the patch test: It is a test
:tion is imposed, that the degree k 1
for consistency.t Normally the nonconforming elements will be short only one
1 true Ritz solution uh would coincide
derivative, so that if lies in em- 2 , and consistency is ensured for the lower-
of the patch, the conditions irnposed
:ant strain; e.g. uh Pm is imposed on tRemember that for difference equations, consistency is checked by looking at the
lary.) Then the test is to see whether, first few terms of a Taylor series-in other words, by considering polynomials up to a
)ring the inter-element boundaries, the certain degree. The patch test does exactly the same for finite elements.
176 VARIATIONAL CRIMES CHAP.4 SEC. 4.2.

order ~e:rms in the differential equation. Then the patch test in case (B) abo ve are fJJ = 1 - x 2 and t¡1 1 - y 2 , anl
confirined consistency for the highest-order difference quotient, while in case therefore, they are not continuous af t.
(A) the equation KQ F proved to be the analogue of a wrong differential ·the element, the solution ofthe final Ii1
equation, with a new Ieading term. This link between consistency and the patch and, the effect of qJ and r¡/ should be
test can be established by Fourier methods; we omit the details, beca use it is within each element. Because of the
effectively limited to a regular mesh and requires a good deal of preparation. is required.
The ¿econd approach to the nonconforming theory is variational and Suppose for simplicity that the e
therefore 'more general. We assert that the variational meaning of success in energy inner product is a(u, v) JJ
the patch test is this: For each polynomial P m and each (nonconforming) is an arbitrary polynomial of degree ,
basis function qJr ,

(2)
aiR, rp) = Jt Jt
-¡ -t

This equality holds ifthe patch test is passed, and vice versa. Notice·that the On the other hand, Green's formula~
right si de is well defined; the main terms are ·mth derivati ves of P m' which
are constants, multiplied by mth derivatives of the nonconforming functions
(/Jp which are o-functions when the derivative is normal to the element bound-
a(P, rp) JJ
aries. The boundary contributions are therefore :finite; they are computable
Choosing a domain large enough to e
from the jumps in the normal derivative of order m- 1, as we shall show by
qJ will· vanish on the boundary and tl
example. These contributions need not be zero; the patch test may fail, and
· - Pyy = O for any linear polyr
in this case convergence will not occur.
O = a*( P, rp) and the patch test is pas
To prove the equivalence of (2) with the patch test, consíder a patch
The patch test is evidently a very s
outside of which qJ1 vanishes identically. We assume that the true solution is
is zero, as we ha ve just seen, and there
u = P m' and note that the corresponding load/ (if any) will certainly ha ve no
computed in the nonconforming way, b
o-functions on the element boundaries:
required to be zero. This can be checke'
without a computer. The size of the~
decides the degree of inconsistendy of
Introducing the two equations ofvirtual work, this is the same as There is another way to achieve ti
o ver a single square:

lf the patch test is passed, so that the approximáte solution is exactly u~ = P m,

then obviously (2) holds. Conversely, suppose that (2) holds for every fJJl' Since rp vanishes on the vertical s
Then a*(u~, rp1) a*(Pm, rp), and necessarily u~ = Pm; the patch test is integral along the bottom and one ba'
passed. one case, and +e in the other, so tha
We rrfust mention one technical difficulty in this argument: Ordinarily the
first variation might not vanish in the direction of qJ1 , which lies outside Jem.
At the smooth solution u P m' however, we are able to show (based on an
integration by parts [S7]) that it does vanish: a(u, qJ1) = (f, qJ1), and the
argument given above is valid. In short? (2) is what the patch test tests. Thus the element again passes the test
Perhaps the simplest instance in which the patch test can be verified occurs rilateral of arbitrary shape, where it i
with Wilson's rectangular elements [W5]. Beginning with the standard e
1 - 2 , '11 = 1 r¡ 2 , and x, y are bil
bilinear functions-on the square 1 x, y 1 there are the four basis standard square into the given quadril
functions (1 ± x)(l y)/4-he adds two new ones. Within the square these strongly violated, even for quadrilatera
 CHAP.4
1. Then the patch test in case (B) above
·order difference quotient, while in case
be the analogue of a wrong dif.ferential
link between consistency and the patch
hods; we omit the details, because it is
td requires a good de,al of preparation.
conforming theory is variational . and
t the variational mean:ing of success in
vnomial P m and eaeh (noneon[orming)
passed, and vice versa. Notice that the
!rms are mth derivatives of p m' which
atives of the nonconforming functions
ivative is normal to the element bound-
~ therefore finite; they are computable
ve of order m 1, as we shall show by
1t be zero; the patch test may fail, and
tr.
with the patch test, consider a patch
Iy. We assume that the true solution is
ng Ioadf(if any) will certainly ha ve no
s:
= (/, rp,).
ual work, this is the same as
pproximate solution is exactly u~ p m'
suppose that (2) holds for every rp r
ecessarily u~ P m ; the patch test is
ficulty in this argument: Ordinarily the
direction of rp1' which Iies outside xm.
ver, we are able to show (based on an
es vanish: a(u, rp 1) (f, rp 1), and the
~t, (2) is what the patch test tests.
ich the patch test can be verified occurs
[W5]. Beginning with the standard
l x, y 1 there are the four basis
wo new ones. Within the square these
SEC. 4.2. NONCONFORMING ELEMENTS 177
are rp x 2 and t¡t = 1 y 2 , and outside they are defined to be zero;
therefore, they are not eontinuous at the boundary. Since they vanish outside
the element, the solution of the final linear system admits static condensation,
and, the effect of rp and t¡t. should be to permit an improved representation
within each element. Because of the discontinuity, however, the patch test
ís required.
Suppose for simplicity that the differential equation is -Au = f The
energy inner product is a(u, v) = JJ uxvx + uyvy, and if P =a+ bx + ey
is an arbitrary polynomial of degree m= 1, then
On the other hand, Green's formula yields
a(P, tp) = Jf ( -AP)rp f rp~~ ds.
Choosing a domain large enough to contain the given square in its interior,
rp will·vanish on the boundary and the Iine integral is zero. Since -AP =
- Pxx - PYY O for any linear polynomial P, we conclude that a(P, rp)
O = a*(P, rp) and the patch test is passed.
The patch test is evidently a very simple rule: The true value of a(Pm, rp)
is zero, as we ha ve just seen, and therefore the integral of every strain Dmrp1-
eomputed in the noneonforming way, by ignoring element boundaries-is also
required to be zero. This can be checked analytically; the patch test can be run
withQut a computer. The size of these integrals, in case they don't vanish,
decides the degree of inconsistency of the nonconforming equatíons.
There is another way to achieve the same result, using Green's theorem
o ver a single square:
Since rp vanishes on the vertical si des x ± 1, there remains only an
integral along the bottom and one back along the top. aPjan equals -e in
one case, and +e in the other, so that
f rp aP/an ds f -¡
l
( l - x 2 X-e)dx + s-1 (1
1
x 2 )e(-dx) = O.
Thus the element again passes the test. lt does not pass the test for a quad-
rilateral of arbitrary shape, where it is constructed isoparametrically: rp
1 - <! 2 , t¡1 1 r¡ 2 , and X, y are bilinear functions of e,
r¡ mapping the
standard square into the given quadrilateral. In fact the patch test can be so
strongly violated, even for quadrilaterals of reasonable shape, that the ele-
 \

178 VARIATIO NAL CRIMES CHAP.4 SEC. 4.2.

ment is useless. To pass the test with isoparametrics, Taylor altered the non- This inequality reflects, as it must, bot
conforming rp and '11 and then also modified the numerical quadrature; two term) and the effect of the nonconform
wrongs do make a right, in California. ing element. To prove (3) we corripare
There is a second non-conforming element which is equally simple and ing that f is smooth,
useful. It is composed of piecewise linear functions on triangles. Rather than
placing nodes at the vertices, however, which produces continuity across
each int~telement edge, the nodes are placed instead at the midpoints of the
edges. Interelement continuity is therefore lost (except at these midpoints) It follows that for all vh in S\
and we have a Iarge.r trial space-roughly three times the dimension of the
standard Courant space, because this is ihe ratio of edges to vertices. This (4)
larger space has made it possible to impose the side condition div vh = O
and still to retain enough degrees of freedom for approximation. Here we get a valuable result right a"
For this space to succeed, it must pass the patch test. We impose the test equality, is bounded by 1u- u~ 1*1 vh 1*.
exactly as above, by computing J rp aP¡an ds along every edge. ln fact, it. is over S\ we h..ave (with thanks to R. Se
the jump in rp which we integrate, to compute the effect of going along an error:
edge in one direction and then back along the other side. The jump in rp is
(5)
a linear function, since rp is linear in each of the two triangles which meet at
the edge. The jump is zero at the midpoint, where rp is continuous. Finally, This shows that the estímate (3), once 1
aPjan is a constant because Pis linear. Since the integral of a linear function it implies that on a regular mesh any c<
is zero, given that the function vanishes at the midpoint of the range of test. Only if the test is passed will .d-
integration, we conclude that each trial function passesthe patch test. To complete the proof of the upp
No ti ce that the mesh was not even required to be regular! Témam has element in Sh to u. By the triangle inej
established a Poincaré inequality 11 vh 11~ < Ca*(v\ vh) for this element.
We intend to prove that the finite element solutions based on Wilson's
nonconforming element converge to u. The rate of convergence will be only
1u - u~ 1* < 1u - w 1* + 1w - u~
minimal-O(h 2 ) in energy-although this may not give a fair description of
its accuracy for large h. (An essential feature offinite elements is their success Therefore (3) is proved once the last
on a coarse mesh; even sorne elements which fail the patch test and are term is the maximum over all vh o[ th
nonconvergent give very satisfactory results for realistic h.) If the element because w is the closest element in Sh
also passes the test for polynomials of higher degree Pn, this rate of conver- it satisfies a*(w, vh) = a*(u, vh) for all
gence in energy would be increased to h 2 <n-m+ 1 J; but it doesn't. numerator of R, followed by the ident
Our plan is to begin with a general error estímate, applicable to any with .d. This establishes the upper bot
nonconforming element, and thereby to isolate the quantity which is decisive For Wilson's rectangular element,
in determining the error. This is the quantity .d, defined by show that .d = O(h); this gives the coJ
To achieve this estímate, we write the

Then, for the particular element studied above, we shall esdmate .d and The functions (/J¡ and lj/¡ are the new 1
deduce the rate of convergence. to lie on the ith element. Thus (/J¡ =
The err~r bound on which everything is based is the following: coordinate of the center of the ith S<
standard basis for the conforming b
(3) 1u - u~ 1* < .d + min 1u - wh 1*.
wh in sn "pagoda functions." Since they are co
 CHAP.4
isoparametrics, Taylor altered the non-
10dified the numerical quadrature; two
l.
tg element which is equiQy simple and
tear functions on triangles. J:{ather than
ver, which produces contihuity across
e placed instead at tñ~ midpoints of the
refore lost (except at these midpoints)
ughly three times the dimension of the
s is the ratio of edges to vertices. This
• impose the side condition div vh O
freedom for approximation.
pass the patch test. We impose the test
iPjan ds along every .edge. In fact, it is
compute. the effect of going a long an
along the other side. The jump in rp is
!ach of the two triangles which meet at
jpoint, where rp is continuous. Finally,
~. Since the integral of a linear function
ishes at the midpoint of the range of
ial function passes ·the patch test.
!n required to be regular! Témam has
11~ < Ca*(vh, vh) for this element.
e element solutions based onWilson's
1. The rate of convergence will be only
this may not give a fair description of
eature offinite elements is their success
nts which fail the patch test and are
results for realis~ic h.) If the element
,f higher degree Pn, this rate of conver-
o f¡ZCn-m+ u; but it doesn't.
eral error estímate, applicable to any
to isolate the quantity which is decisive
¡uantity 1\, defined by
tdied above, we shall esdmate L\ and
1ing is based is the following:
- min
w' in s•
!u- wh¡*.
SEC. 4.2. NONCONFORMING ELEMENTS 179
This inequality reflects, as it must, both the order of approximation (the last
term) and the effect ofthe nonconformity; a= a*and L\ =O for a conform-
ing element. To pro ve (3) we compare the equations of virtual work: assum-
ing that f is smooth,
It follows that for all vh in S\
(4)
Here we get a valuable result right away: The left si de, by the Schwarz in-
equality, is bounded by l u- ui 1*1 vh 1*. Dividing by¡ vh 1* and then maximizing
over S\ we have (with thanks to R. Scott for his help) a lower boundfor the
error:
(5)
This shows that the estímate (3), once proved, wiii be extremely realistic, and
it implies that on a regular mesh any convergent element must pass the patch
test. Only if the test is passed will L\ ---+ O.
To complete the proof of the upper bound (3}, we let w be the closest
element in Sh to u. By the triangle inequality
Therefore (3) is proved once the last term. is identified as equal to 1\. This
term is the maximum o ver all vh of the ratio R 1a*( w - ui, vh) 1/1 vh 1*. But
because w is the closest element in Sh to u (i.e., the projection of u onto Sh),
it satisfies a*(w, vh) = a*(u, vh) for all vh in Sh. With this substitution in the
numerator of R, followed by the identity (4), the maximum of R is identical
with 1\. This establishes the upper bound (3).
For Wilson's rectangular element, which passed the patch test, we shall
show that L\ O(h); this gives the correct rate of convergence of l u ui 1*.
To achieve this estímate, we write the trial functions as
The functions rp¡ and t¡/¡ are the new nonconforming basisfunctions, shifted
to líe on the ith element. Thus fPt = 1 - ((x x¡)/2h1) 2 , where Xc is the
coordinate of the center of the ith square. The functions Wr make up the
standard basis for the conforming bilinear element; we ha ve called them
"pagoda functions." Since they are conforming, they make no contribution
 \
180 VARIAN..._ONAL CRIMES CHAP.4
SEC. 4.3.

to ~- In fact, over a typical square, the numerator of ~ amounts to for the error in eigenvalues, which cor
static problems. This agrees with our pr1
(a* - a)(u, ?A)= (a* - a)(u, a 1rp1 + b1tp¡) =(a* - a)(u -- Pp a1rp1 + b1tp 1).
~ . formity which is responsible; the elem
theory would have permitted O(h 2<k-m>)
Here we use the patch test, or rather the equivalent identity (2); the inclusion
It appears that element 3 passes the t
of the linear polynomial P 1 has no effect. With a little patience, we can esti-
is rather remarkable. In fact, the vanish
ma te this 1~;~ontribetion to ~ from the ith square by
test requires would seem in general to
regular mesh, and one must expect tha
convergent elements will continue to a
Here we have chosen P~' as approximation theory allows, to·be within O(h) future.
of u in the first deriva ti ve (recall that m = 1 in our example). Now we sum the The list also contained a number of
contributions from all the squares, and use the Schwarz ihequality: we prefer not to regard numerical inte,
elements. The effect ofthis integration is
la*(u, vh)- a(u, vh)l < C'h 2: llull2,etlla¡rp¡ + b;t¡/;llt I(v) to a new one, bu't the difference b
boundary integrals. Therefore this effec1
< C"h(2: llull~.eY 12 (2: llvhlli,eY 12 = C"hllull21vhl*·
finite element approximation u\ will be
Dividing by 1vh 1*' we reach the right inequality:
4.3. NUMERICAL INTEGRATION

It now follows from the basic estimate (3) that the Wilson nonconforming Numerical integration has become a
elements give an error in the strain energy norm ofl u ---:.ui 1* = O(h). We are finite element technique. In the early sta
confident that this is the correct rate of convergence, although we must finite elements appeared to lie in preci
emphasize two points: (1) It is obvious experimentally that the constant integration of polynomials over triangl1
which multiplies h is very much smaller than it would be without the extra exact formulas. At the present stage, it r
nonconforming tria} functions, and (2) the strain energy in ui is no longer of polynomials is no longer essential, an
necessarily smaller thari that in u. In fact, convergerice from above is the general shape functions, are equally ce
rule rather than the exceptiori, both for strain energy and for displacements. further from the truth: The key to the s1
The convergence proof after a successful patch test is discussed further in finite element method is the presence oJ
the Index of Notations. Irons and Razzaque ha ve described in the Baltimore The fundamental question is this : W
Symposium volume [6] a number of other elements which pass the test, in--- tion formula is required for convergem
cluding nomial which appears be integrated ex
a(v, v) involves the squares of polynomi
l. the 12-degree-of-freedom rectangular element due to Ari, Adini, this degree may simply cost too much.
Clough, and Melosh. It has v, vx, vy at the corners as nodal parameters, and perly the fraction of computer time whi1
the complete cubic plus x 3y and xy 3 as shape function. Mathematically, we are again faced
2. a constant curvature bending element proposed by Morley, which can this is the effect of numerical quadratm
be viewed as conforming for the complementary energy principie.
3. a new element of Allman-Pian type. . I(v) = a(v, v)- 2b(f, v) = JJ [¡
Extensiv~ computations with element (1) on eigenvalue problems for
plates (the element is continuous, but nonconforming for fourth order equa- [Notice the new notation b(f, v) for the
tions) were carried out in [L3]. They reporta definite convergence rate .of h 2
I*(v) = a*(v, v)- 2b*(f, v) = 2: wJ
 CHAP.4
the ·numerator of A amounts ~
. - L~
the equivalent identity (2); the inclusion
ffect. With a little patience, we can esti-
~ ith square by
nation theory allows, to be within O(h)
n = 1 in our example). Now we sum the
nd use the Schwarz iileq uality:
,e! + b¡fjl¡ llt
11 a¡tp¡
LJ CL: 11 v" lli.eY 12
112
C"hll U llzl vh 1*.
inequality:
ate (3) that the Wilson nonconforming
lergy norm ofl u ---:<u~ 1* = O(h). We are
Lte of convergence, although we must
1ious experimentally that the constant
ller than it would be without the extra
~2) the strain energy in u~ is no longer
n fact, convergerice from above is the
:Or strain energy and for displacements.
:ssful patch test is discussed further in
tzzaque have described in the Baltimore
other elements which pass the test, in-
tangular element due to Ari, Adini,
lt the corners as nodal parameters, and
as shape function.
~Iement proposed by Morley, which can
tplementary energy principie.
type.
:ment (1) on eigenvalue problems for
_nonconforming for fourth order equa-
report a definite convergence rate of h 2
SEC. 4.3. NUMERICAL INTEGRAtiON 181
for the err.or in eigenvalues, which corresponds to the error in energy for
static problems. This agrees with our prediction. Note that it is the noncon-
formity which is responsible; the element has k = 4, and approximation
theory would have permitted O(h 2ck-m>) O(h 4 ).
It appears that element 3 passes the test even on an irregular mesh, which
is rather remarkable. In fact, the vanishing of boundary integrals which the
test requires would seem in general to be a rather lucky chance even on a
regular mesh, and one must expect that not only nonconforming but non-
convergent elements will continue to achieve engineering accuracy in the
future.
The list also contained a number of numerically integrated elements, but
we prefer not to regard numerical integration as producing nonconforming
elements. The effect ofthis integration is indeed to change the true functional
I(v) to a new one, bu't the difference between the two does not consist of
boundary integrals. Therefore this effect, and the error it introduces into the
finite element approximation uh, will be analyzed separately.
4.3. NUMERICAl INTEGRATION
Numerical integration has beco me an increasingly important part of the
finite element technique. In the early stages, one of the crucial advantages of
finite elements appeared to líe in precisely the opposite direction, that the
integration of polynomials over triangles and rectangles could be based on
exact formulas. At the present stage, it m~ght seem that the unique simplicity
of polynomials is no longer essential, and that rational functions, or still more
general shape functions, are equally convenient. In fact, nothing could ·be
further from the truth: The key to the success of numerical integration in the
finite element method is the presence of polynomials.
The ftmdamental question is this : What degree of accuracy in the integra-
tion formula is required for convergence? It is not required that every poly-
nomial which appears be integrated exactly. The integrand in the energy
a(v, v) involves the squares of polynomials, and a formula which is exact to
this degree may simply cost too much. It is very important to control pro-
perly the fraction of computer time which is spent on numerical integration.
Mathematically, we are again faced with a change in the function"J:1(v);
this is the effect of numerical quadrature. Suppose the true functional is
I(v) = a(v, v)- 2b(f, v) JJ[p(x, y)(v; + v;)- 2fv] dx dy.
[Notice the new notation b(f, v) for the linear term.] Then typically
182 VARIATIO~AL CRIMES CHAP.4 SEC. 4.3.

Each element region contri bu tes a certain number of evaluation points. .terms of degree higher than k - 1 in
e¡ (X¡, y¡), WÍth W~Íghts W¡ which depend on the size and shape of the for rectangular elements, this will inj1
element and on the rule adopted for numerical int'egration. The rule is called the correct test is expressed by a*(Pn'
exact to degree q, if the integral of every polynomial Pq is given correctly by mials of degree n m, multiplied b)
Í: W¡Pq{et). integrated correctly.t
Suppose that we minimize 1* over all trial functions vh. Then the minimiz- No attempt will be made to devi
ing funct~lon üh 2: Q/p1 is determined by an approximate finite element sys- astonishing that even on triangles an~
tem 'f(Q ~ P, in which the stiffness matrix and load vector are computed by never been completely solved. Irons [1
numerical rather than exact integrations. This is the system which (apart from this direction, by achieving a given ac,
roundoff error) the computer actuaÍly so/ves. Our goal is to estima te the differ- standard product-Gauss rules. The R
ence uh üh, and we repeat the main point: lt is not necessary that the ener- and others has made a profound stud
gies 1 and 1* be clase in order for uh- üh to be small. regions, and has discovered sorne ren
We shall summarize the main theorem before giving examples and proof. interested in a 14-point formula on a e
The essential condition on the quadrature formula is identical with the patch degree q 5:
test for nonconforming elements: üh converges to uh in strain energy Cll uh üh 11m
---7 O) if and only if for al/ polynomials of degree m and al/ tria/ functions, 2-J5.(121
361 8
±!('
1 •

(6)
e¡
where the are the vertices of a simi
the r¡¡ are the vertices of a regular
There is also a supplementary condition of positive definiteness, reqmrmg sphere of radius ,Jf§760! We repeat, 1
that the approximate strain energy a* should be elliptic on the subspaces finite elements must take account of
Sh: a*(vh, vh) Bll vh 11!. The term O(h) enters the condition only because, if occur in trial functions on rectangl€
the material properties vary within an element, this will not be treated exactly nothing to approximation theory. On ·
by the quadrature. This is a second-order effect. plete polynomial of degree k 1, or ~
The test (6) actually applies only to the leading terms in the strain energy, tion of correctly integrating complet
involving mth derivatives. Since such derivatives of P m are constants, the possible. We owe Table 4.1 on the fol
inner product a(Pm' vh) involves only integrals of mth deriva ti ves of v\ added several new quadrature rules
and the convergence condition reduces to this: the mth derivatives of every Zienkiewicz [22]. The formulas are syr
tria/ function should be integrated exactly. If the trial functions are polyno- ei
íf a sampltng point = ((¡. Cz, C3) o
mials of degree k 1, this means that the quadrature formula should (¡ are distinct, there are six samplin¡
be correct at least through degree k m - l. Rectangular elements are three; the point (j, -j-, j) at the centro
much more demanding; there are terms in the trial functions which add once.
nothing to the degree of approximation, but whose derivatives must never- The isoparametric method could r
theless be integrated correctly. With bilinear elements in second-order prob- since the integrand is a rational funct
lems, for example, the twist term xy has linear derivatives; therefore the first it looks impossible that even 1
quadratute formula must be correct for these terms, and not only for the
constant strains which come from the linear terms a + bx + cy. tit is interesting to compute the quadra
In practice the test (6) will often hold for all polynomials Pn ofsome hígher 'will be of the same order hk-m in the strains <
degree n > m. In this case the accuracy is more than just minimal: the error exponent is to agree with n - m + l, then n
in the strains is of order hn-m+ 1 • Every additional degree of accuracy in the als of degree k - 1, multiplied by the mth (
integrated exactly. If the vh themselves compr
quadrature scheme contributes an additional power of h to the error estima te. true on triangles, and the material coefficien1
In other words, if the tri al functions are of degree k - 1, and the quadrature terms in this energy-the squares of mth derb
is exact to degree q, the error is of order q k + m + 2. If there are any the full accuracy!
 CHAP.4
~~\.. .....) ~
certain number of evaluation points
iepend on the and shape of the.
.umerical integration. The rule i.s called
~ry polynomial Pq is given correctly by
111 trial functions v". Then the minimiz-
i by an approximate finite element sys-
Ltrix and load vector 'are computed by
1s. This is the system whích (apart from
oh•es. Our goal is to estima te the differ-
Joint: It is not necessary that the ener-
i? to be small.
rem before giving examples and proof.
1re formula is identical with the patch
~·erges to u" in strain energy (!1 u"- u" 11m
of degree m and al! trialfunctions,
ion of positive definiteness, reqmrmg
!* should be elliptic on the subspaces
) enters the condition only because, if
!lement, this will not be treated exactly
·der effect.
the leading terms in the strain energy,
derivatives of P m are constants, the
y integrals of mth derivatives of v\
:s to this: the mth derivatives of every
-:tiy. If the trial functions are polyno-
that the quadrature formula should
- m l. Rectangular elements are
~ms in the trial functions which add
m, but whose derivatives must never-
tlinear elements in second-order prob-
has linear deriva ti ves; therefore the
~or these terms, and 1iot only for the
linear terms a + +
bx cy.
:i for all polynomials Pn of sorne higher
'f is more than just minimal: the error
additional degree of accuracy in the
tional power of h to the error estima te.
~ of degree k 1, and the quadrature
der q - k + + m 2. If there are any
SEC. 4.3. NUMERICAL INTEGRATION 183
terms of degree higher than k 1 in the trial functions, as there always are
for rectangular elements, this will injure the order of the error. In all cases
the correct test is expressed by a*(Pn, v") = a(Pn, vh); it is complete polyno-
mials of degree n - m, multiplied by trial strains Dmvh, which have to be
in tegrated correctly. t
No attempt will be made to devise new quadrature formulas, but it is
astonishing that even on triangles and rectangles this classical problem has
never been completely solved. Irons [I5] has shown what can still be done in
this direction, by achieving a given accuracy q with far fewer points than the
standard product-Gauss rules. The Russian school of Sobolev, Liusternik,
and others has made a profound study of "cubature formulas" over regular
regions, and has discovered sorne remarkable formulas: the reader may be
interested in a 14-point formula on a cube of edge which is accurate to
degree q 5:
2-JS (121 8 6
+ 4o ~f(11J
361 8 ~JceJ
)
,
el
where the are the vertices of a similarly placed cube of edge ,J.38l33 and
the 11; are the vertices of a regular octahedron touching a circumscribed
sphere of radius ,Jf9765! We repeat, however, that numerical integration for
finite elements must take account of the higher-degree terms which often
occur in trial functions on rectangles, even though they may contribute
nothing to approximation theory. On triangles the element is normally a com.:.
plete polynomial of degree k 1, or el ose to it, and then it is purely a ques-
tion of correctly integrating complete polynomials of as high a degree as
possible. We owe Table 4.1 on the following page to Cowper [CI3], who has
added several new quadrature rules over triangles to those reproduced in
Zienkiewicz [22]. The formulas are symmetric in the area coórdinates, so that
if a sampling point ( (Cp C2 ,.¡; 3 ) occurs, so do al! its permutations. lf the
C; are distinct, there are six sanipling points; if two C; coincide, there are
three; the point (j, j, j) at the centroid, if used in the formula, is taken only
once.
The isoparametric method could not exist without numerical integration,
since the integrand is a rational function of the new coordina tes e; and 11· At
first it looks impossible that even numerical integration should succeed,
flt is interesting to compute the quadrature accuracy required ín order that u" - iih
'will be of the same order hk-m in the strains as the basic approximation error u u11 • lf this
exponent is to agree wi th n m + 1, then n = k 1 ; the mth derivatives of all polynomi-
als of degree k - 1, multiplied by the mth derivatives of all trial polynomials vh, must be
integrated exactly. If the vh themselves comprise al 1polynomials of degree k - 1, as is often
true on triangles, and the material coefficients in the strain energy are constant, the leading
terms in this energy-the squares o! mth derivatives-must be computed exactly to maintain
the full accuracy!
 Table 4.1

W¡
0.33333 33333 33333
0.33333 33333 33333
-0.56250 00000 00000
0.52083 33333 33333
0.16666 66666 66667
0.10995 17436 55322
0.22338 15896 78011
0.37500 00000 00000
......
00 0.10416 66666 66667
~
0.22503300003300000
0.12593 91805 44827
0.13239 41527 88506
0.20595 05047 60887
0.06369 14142 86223
0.05084 49063 70207
0.11678 62757 26379
0.08285 10756 18374
-0.14957 00444 67670
0.17561 52574 33204
0.05334 72356 08839
0.07711 37608 90257
'1
3-point formula
0.66666 66666 66667
3-point formula
0.50000 00000 00000
4-point formula
0.33333 33333 33333
o. 60000 00000 00000
6-point formula
0.65902 76223 74092
6-point formula
0.81684 75729 80459
0.10810 30181 68070
7-point formula
0.33333 33333 33333
0.73671 24989 68435
7-point formula
0.33333 33333 33333
0.79742 69853 53087
0.47014 20641 05115
9-point formula
0.12494 95032 33232
0.79711 26518 60071
12-point formula
0.87382 19710 16996
0.50142 65096 58179
0.63650 24991 2139~
13-point formula
0.33333 33333 33333
0.47930 80678 41923
0.86973 97941 95568
0.63844 41885 69809
Multiplicity
'2 '3
degree of precision 2
0.16666 66666 66667 0.16666 66666 66667 3
degree of precision 2 r
0.50000 00000 00000 0.00000 00000 00000 3
degree of precision 3
0.33333 33j33 33333 0.33333 33333 33333 1
0.20000 00000 00000 0.20000 00000 00000 3
degree of precisio'n 3
0.23193 33685 53031 0.10903 90090 72877 6
degree of precision 4
0.09157 62135 09771 0.09157 62135 09771 3
0.44594 84909 15965 0.44594 84909 15965 3
degree of precision 4
0.33333 33333 33333 0.33333 33333 33333 1
0.23793 23664 72434
0.02535 51345 51932 6
degree of precision 5
0.33333 33333 33333 0.33333 33333 ,33333 1
0.10128 65073 23456 0.10128' 65073 23456 3
0.47014 20641 05115 0.05971 58717 89770 3
degree of precision 5
0.43752 52483 83384 0.43752 52483 83384 3
0.16540 99273 89841 0.03747 74207 50088 6
degree of precision 6
0.06308 90144 91502 0.06308 90144 91502 3
0.24928 67451 70910 0.24928 67451 70911 3
0.31035 24510 33785 0.05314 50498 44816 6
degree of precision 7
0.33333 33333 43333 0.33333 33333 33333 1
0.26034 59660. 79038 0.26034 59660 79038 3
0.06513 01029 02216 0.06513 01029 02216 3
0.31286 54960 04875 0.4S69 03154 253160 6

-..
-..:¡ a"I:T'J .-+- m §"· .§ ~ g" ~
~ o ~ ~ ~ 3 ..0 ~ ~ ~ ~
o t:;· e;;· ~ S. §'· s¡· ~ ...._, o < s·
.-+- ...., 1 w ¡::: . (') .,. -· o ..... e;;·
- ~ ==' ·"' (fe¡ (') (') r
..... ~~...., o~~~(J'C¡.-+-~ e~a" ~~~ (')
o ~ -· ,J::.,
::ro.,._,S~(óo~~~ .::: ~~"'o 0 m~""t (')
....,
o>¡-<goo
o'"1-a(ó.-+-2'";;J~o~~ ;;,"g~O~I:l~ 3 o ¡::: g. .-+- ~ .-+- o 3 ~· ~
o ~
s· m ::r' o' .-+- fl - ...., . ;. ~ ..., o Q o :::: ~- m (') m O ::r' ~ "' -·
::Sa § ~ E; m ~ ..0 • o t.., g.~ ~ ;:::¡ t.., § -<
- .-+- ~ -· .-+- o .-+- (D m
3m "'a.~. "'§om -· o .-+- 3 ¡::: ~ "' -· (";) ~ ~
.~
~~
§ ~ a" (t e;;· (t ~
~
o.,.,g~~~.,._,~om~e;~~~- s~e-
a.-+- ..... -o:e. -<"<
,_,...,o-..oso ~ ~om~ - ..... ~~~ o "::~ . (";. ) q ;. .... -g ~~
~ O ::r' ~ ~ m :=.: ..., (') ;;J - o ~ 0\ O - · '"ti S: o -..
.-+- ;:.< ... 1 -·· ::::;,· - m ~ -· e o w m ;:::¡ _,..,-..... 1 ¡:g¡:g-eEr~eio
o §' ~ .::t ~ -· a" - ~ ::::. ~ o ""t . (') ~ ~ (";) '-.:::! cr
m (";)
!< l:3 ~ ..... o -·o ~
"' • ~ ...._, .-+- §' ~ o ~ (') (') g ....., ~ :;: - O" "O ~ ~- ""e::; .::t ~~
~ o
..., '-.:::!
...._, ~~~o~(ó~
~. a e: ¡:g ~ ~ 3 o g_"' . '"'t
. c:r ""t o :::1 ~ ~ ~:l.. e t'l~
<
o ~~ ~-
~ ~¡~ .9. , [ 2.
(') .-+- ~
@ ~ ~ ~ g ~ Er,% ~ E ~ ~ So~ ~ ~ ~ (";)• ~
..::::t :['tt ?<
:!~ ~-e -·
~'.:::!o¡:;~oo.-+-~~~~ "::~ z-.
g·. §"· [ ..... ¡:;m" ""t~ ~ "'_.. ::r' ,--..,
::::n ;.....3 ~ -.. O o -· ~ e;;· ::r' o ~ e
g t'lt.-, o ~,~
::::. ~ e. ~ o o ..,
~t.-, ~
::r 0 m ~ ~ ~ -
~::r-~~~oSg~oa~,_,""t~~as.~~ ..::::t -.. ?< <:::l.... o¡:;.,.,oaS
~
..... o G ..::::t ¡:g ....., ~ e o "¡ -· "::~ _,..,-..... O> ~ ~ - o ::¡ .-+- ::::.
6a(J'C¡...._,~o~...,-...,e.~c§- ... "'"!)~·
....., ~ :::1 ~ e;;· '-.:::! ~ ..., ~ ::r :r ::r ::r o
~ 11 ~ ..::::t ~ ~o-g,.o~
...,~a::r'mom~i-m~~r...,::r' ..... ~ 0 ~~ ~ a ';::-"'
..::::t
...._, 1
"' o m m o o E.
g. g o ~ S ~ o S. ::= ñ'' ~ E. ~ iS' o ~ ..... _0 o ~ ~ m ~
'-.:::! -· m
...., o o ~ ~ o ....., - ~ e.. . . . ~ s· ;:::¡ ~ ::r ~ ...., ~ -.. '-.:::! '...: ?<
....., -..
~....., m,_,~::rm~ ..... oo~~~ ~.a ..... Y> q ?
....~ "' ~
~ o '"2 x~ ~ ~ ¡; a . . , ~ ~ Q ~· [ ~ ~ S: § '-.:::! s "" "" ~ ~
?ti o [ ~ r.p·~
-- '
 SEC. 4.3. NUMERICAL INTEGRATION 185

since it is never exact for rational functions. The entries a*(rpi' rpk) of K will
be completely different from the entries Kik a(rpi, rpk), and no perturbation
argument is possible. Nevertheless, we compute a(Pm' vh) - a*(Pm' vh) and
apply the test; the crucial point is that the test involves only one trial func-
tion_:_not both rpi and rpk at the same time-and this will save us.
A typical transformation was given in Section 3.3, with m = 1; aPmfox
is a constant e, and

..
'1

~
M\00
M 11') t'--
MVt'--
MM O\
voo
00 00
MO
MO
N....,. \0
o--
11') 0\ 00
-ov
M00\00
MM....,. \O
MON....,. ff p(x, y) aPm avh
ax ax dx dy
r¡ M N 00 00 11') MO\<'IM
0\t--v Mf'--011')
r¡ M M t'-- M t'-- M00\('1
:t M t-....,. v-oo
000 VIl') O\ M \O <'IV
')
..., MOt-- VN -vv M\0011')
M 11') 00 <'IV or--o M 0\....,.,.....
.... M \0 11') 11') r-- 0\\011') Mil') OM
? M 00....,. N t'-- oooov MVMO
M N t'-- V') V 0<'1....,. MM-O\
M....,. 0\ r--r-- M 0\ M
~ M011') MOII')\0
MM \OVIl') M\0\0oo
M,.....O vo ONO It is crucial to observe the form ofthe transformation matrix:
::) 000 oo ooo MNOv
oooo
-x")·
X.;

Evidently ('1: = YrtfJ and f/x -Y.;/1. With this substitution, the integral
beco mes

(7) e JJp(!;, r¡)(v~y"- v~Y.;) dt; dr¡.

Et

The rational functions have disappeared, and convergence will occur if this
integral is computed correctly. (For fourth order equations the rational func-
tions -do not disappear, and we cannot justify numerical integration unless
the coordinate changes satisfy the smoothness conditions 11 Fllk <.e des-
cribed on p. 163. In that case, the elements are only slightly distorted, and J
is apparently no worse than the variable coefficient p; the integration error
is of the same arder as for constant coefficients without isoparametrics.)
In the case of bilinear functions on quadrilaterals, all the derivatives
Ot--\0
ONO
r-- 1'1')
OON
t-0\V
o t- t-
OVO\t-- v~, y", ... are linear, and it would appear that the quadrature must be cor-
oo N t'--0MII')
80011') <'IMM \0('100('1
voo
ovoo
0\0
\0 00
0\0oo
t'--('1-
f'--MOOO
\O M O O\
rect for quadratic polynomials. It just happens, however, that the second-
M
M 11') t- t'--N M t'-- \0
\011')11')
vv\Ooo
Vt'--11')0
degree terms cancel in the particular combination K= v~y" v:y.;, so that
§ 000
< '11')
1 """"""
o-
V') V
or--r-- VIl') M \O
exactness to first degree is actually sufficient for convergence. In practice the
a:~ o-
O\ N O
":t:.\0 .....
ONNt--
011')f'--M
M
M M 0\
00\1'1')
11') 0\
0\ \0
VOOII')
00 t- 00
r- ..... v ..... quadrature will be more accurate than that (it will probably have to be, just
11') 11')('1 V') M
11') \0 M -
<'IN M
<'~,.....,.....
o \0
NO
O \O N
11') ..... 00
0\ll')l")t-
Vt-ll')f'-- to achieve positive definiteness of a*) and the benefits come in establishing
o-o --oo
000 00 ooo oooo more than a mini mal rate of convergence. Suppose the conditions of Section
1
3.3 are satisfied-the Jacobian stays away from zero, and the coordinate trans-
formations x(t;, r¡) and y(!;, r¡) have bounded coefficients. Then justas in the
x-y plane, each additional degree ofpolynomials in t;, r¡ contributes one more
order of approximation. Therefore we expect the integration error to improve
at the same rate; since first-orde'r quadrature is sufficient for convergence
14
186 VARJATIONAL CRJMES CHAP.4 SEC. 4.3.

in the bilinear case, the error in strains should be O(hq) if the quadrature is Then the error in strains due to approx.
exact to degree q.
One final isoparametric remark: since the coordinate transformations (10)
x(c;, r¡) and y(<!', r¡) have the same form as the shape function v"(c;, r¡), it fol-
lows that the combination K has the same form as the Jacobian J. Therefore, · Proof. The left side of the identity
our rule ~hat K must be correctly integrated coincides with the rule [22], and the right side is bqunded above bj
based on/iirons' intuition, that the volume of each element (the integral of J) common factor, the result is (10).
must be correctly computed by the quadrature. In three dimensions this
requires a higher order of exactness: K and J will involve products of three The lemma and theorem are not
rather than two derivatives. For subparametric elements the two rules di- They apply also toa change in the co~
verge; the stiffness matrix depends on K, the mass matrix and load vector on tion; in other words, they describe th
J. To repeat, these rules apply to highly distorted isoparametrics; for small continuous as well as the discrete pr<
distortions J is smooth and can be discounted in the convergence test. on the inhomogeneous term. Consider
We begin now on the theory, which is entirely based on the following qu = f, and suppose thatp, q, andjar1
simple identity. #: tity, applied on the whole space JC1 i

LEMMA 4.1 f p(u' u') 2 + ij(u - u) 2

Suppose that uh and u" minimize thefunctionals I(v") and I*(v"), respective/y, (11)
so that the equations of virtual work (the Euler equations ~1 = ~!* =O) become f (¡;- p)u'(u' u')+ (

The left si de, which is a*, will be posi

Then righ t si de, each term is less than 11 u
perturbation. This yields a simple b01
(8) a*(uh u\ uh- uh) =(a* - a)(uh, uh u") (b* - b)(f, uh u"). on the resulting perturbation in the s~

Proof The left si de of the identity is COROLLARY

a*(u", uh uh) a*(ü", uh- u") Suppose that the coefficients and
(a* - a)(u", uh - u") + a(u", uh- u") a*(u", u" u"). less than f:
max(lp- pj, Jq --
With v" = uh - u" in the equations of virtual work, the last two terms yield
x

(b b*)(f, u,. u"), and the proof is complete. There is a similar identity, Then the solution is also perturbed by (
with a, b, and u,. replaced by a*, b*, and u\ but it is not so useful. Notice
also that the terms in a and b actually cancel in (8), but it is "important to
keep them and to work with the differences a* - a and b* - b.
(12) !tu u11
Our principal theorem follows immediately from the identity. In terms of finite elements this ha
that p, q, andfare replaced by their i
THEOREM 4.1 This is a perturbation of order hk. lf
Suppose that the approximate strain energy is positive definite, a*(v", v") solved exactly (there will be produc1
811 v,. 11~,
and that integrals), then by the corollary 11 u"
possible alternative to numerical in
(9) 1(a* a)(u\ vh) 1 + l (b* - b)(f, vh) 1 ChP 11 v" llm· share of attention in the numerical ar
 CHAP.4
ns should be O(htJ) if the quadrature is
since the coordinate transformations
m as the shape function vh({, 17), it fol-
mme form as the Jacobian J. :Therefore
ntegrated coincides ~ith the rule [22]:
ume of each element¡{the integral of J)
quadrature. In three· dimensions this
K and J will in vol ve products of three
)parametric 'elements the twi
rules di-
K, the mass matrix and load vector on
hly dístorted isoparametrics; for small
iscounted in the convergence test.
lich is entirely based on the following
(unctionals 1(vh) and l*(vh), respective/y,
ze Euler equations ~1 = ~1* =O) become
r is
ih)
z(u\ uh uh)- a*(ii\ uh- uh).
·virtual work, the last two terms yield
complete. There is a similar identity,
:md ii\ but it is not so useful. Notice
y cancel in (8), but it is important to
~ences a*- a and b* b.
1ediately from the idéntity.
energy is positive definite, a*(vh, vh)
SEC. 4.3. NUMERICAL INTEGRATION 187
Then the error in strains due to approximate integration is of order
(lO)
Proof. The left side ofthe identity (8) is bounded below by Bii uh- uh 11!,
and the right side is bqunded above by (9), with if = uh uh. Cancelling the
common factor, the result is (lO).
The lemma and theorem are not limited only to numerical quadrature.
They apply also to a ohange in the coefficients of the original differential equa-
tion·; in other words, they describe the manner in which the solution, to the
continuous as well as the discrete problem, depends on the coefficients and
on the inhomogeneous term. Considera one-dimensional problem -(pu')' +
qu = f, and suppose that p, q, and f are changed to p, ij, and /. Then the--iden-
tity, applied on the whole space X1 instead of the subspace S\ becomes
f ft(u' u')2 + ij(u - u)2
(11)
= f (ft p)u'(u' u')+ (ij q)u(u u)-(/ f)(u u).
The left si de, which is a*, will be positive definite if ft > O and ij >O. On the
right side, each term is less than 11 u- u11 1 times a constant multiple of the
perturbation. This yields a simple bound, not the most precise one possible,
on the resulting perturbation in the solution.
COROLLARY
Suppose that the coefficients and inhoinogeneous term are perturbed by
less than E:
max (J p - ft J, 1q - ij J, 1f ID< E.
X
Then the solution is also perturbed by O(E):
(12)
In terms of finite elements this has the following interpretation. Suppose
that p, q, and f are replaced by their interpolates ·in the finite element space.
This is a perturbation of order hk. If the resulting finite element problem is
solved exactly (there will be products of three polynomials in the element
integrals), then bythecorollary !luh- i?ll 1 O(hk). Thus interpolation is a
possible alternative to numerical integration, and it has had the lion's
share of attention in the numerical analysis literature; Douglas and Dupont
 )
1.88 VARIATI<fNAI: CRIMES CHAP.4
SEC. 4.3.

[D8] have successfully explored even nonlinear parabolic problems. In engi~ This is the 5-point scheme rotated th
neering calculations, however, direct numerical integration has consistently unstable; in fact, our oscillating twist ~
been preferred; with isoparametric or shell elements, there is effectively no equation KQ = O.t
choice. A rough 'operation count suggests that also in other problems direct A similar one-point rule at the cen
quadrature is the more efficient, and we therefore intend to concentrate on
this techni,que.
functions, will not be indefinite. The str:
at the centroid without being identical
The simplest example is, as usual, the most illuminating. Therefore, we rise to the normal 5-point scheme.
begin with the e¡¡uation -u" = f, and apply numerical integration to l(v) = The basic test for definiteness is to
f (v') 2 - 2fv. Consider first the requirement of positive definiteness: trial functions which, under numerical
energy. In practice this is decided fror
(13) matrices; if the only zero eigenvalues
quadrature is all right. If there are add
The interval is divided into elements, in other words into subintervals, and a tu re might still be acceptable: it has to t
standard quadrature rule is applied o ver each element. If vh is a polynomial of ing polynomials in separate elements e~
degree k l, and· the quadrature weights wi are positive, then the definite- of the twist described above-into a t
ness requirement amounts to this: There must be at least k 1 integration energy over the whole domain. For ex
points- c!1 in the subintervals. Otherwise, there will exist in each element a non- (2 x 2) does not satisfy our stability e
zero polynomial of degree k - 2 which vanishes at every c!1• Joining these q uadratic. With Gauss points ( ±~
polynomials and integrating once, we have constructed a trial function vh (x2 _ c;2)(yz - c! 2) has zero strain ener
which contradicts (13): to give trouble over the whole domain
singular, if this· pattern does not fit the
the problem. In this case one eould live ·
· integration, even though K is mueh n
demands.)
Ifthere are negative weights, then even more integration points will be needed The question of definiteness beeorr
for definiteness. We do not expect such formulas to be popular; they are also eight-parameter element, obtained fron
vulnerable to roundoff errors. x 2y 2 term and the no de at the eenter
In two dimensions, with bilinear trial functions on rectangles, the definite- functions no Ionger include (x 2 c! 2)(
ness condition would fail if we were to choose the midpoint rule for quad- apparently enough for a stable finite
rature: equation. Taylor has pointed out, ho 1
h/2 fh/2 two dependent variables, the situatio1
J -h/2 -h/2
g(x, y) dx dy "' h 2 g(O, 0).
x(yz _ c;2), v - y(xz c! 2) líes in the
ishes for the 2 x 2 rule. But Taylor h~
This is the one-point Gauss rule on a square, and it is exact for the polyno- cannot be continued into a neighboring
mials g 1, x, y, xy. However, it is indefinite; for the trial function vh = xy matrix is too low, but after assembly 1
numerical Íntegration of (v~) 2 + (v;) 2 gives zero. If we set vh +1 or -1 in nonsingular and the numerical integra·
a checkerboard pattern O'Ver the whole set of nodes in O, the result is a high So mueh for definiteness, whieh is a
frequency oscillation (pure twist, with the smallest wavelength 2h which the points. We turn now to aeeuracy, whi
mesh can accept) whose numerical energy is exactly zero. This is reflected in whieh the quadrature is exact. There ar
the discrete approximation to the Laplacian which would arise from this is to compute directly the exponent p
midpoint rule; a typical row of the stiffness matrix K gives
tA recent paper by V. Girault analyzes the
rature rule. It turns out to be surprisingly usf
our hypothesis of positive definiteness.
 CHAP.4
~ nonlinear parabo_lic problems. In engi-
t numerical integration ~as1:on~istently ·
)r shell elements, there 1s (ffectively no
:gests that also in other problems direct
l we therefore intend to concentrate on
1, the most illuminadng. Therefore, we
d apply numerical integration to I(v) = rise to the normal 5-point scheme. '
irement of positive definiteness:
in other words into subintervals, anda
'er each element. If vh is a polynomial of
:ights W¡ are positive, then the definite-
'zere must be at least k - 1 integration
, there will exist in each element a non-
ich vanishes at every ej. Joining these
e have constructed a tria! function vh
but v~(eJ = O.
t more integration points will be needed
h. formulas to be popular; they are al so
ial functions on rectangles, the definite-
to choose the midpoint rule for quad-
dx dy "' h 2 g(O, 0).
square, and it is exact for the polyno-
definite; for the tria! function vh = xy
gives z~ro. If we set Vh = + 1 or - 1 in
;! set of nodes in Q, the result is a high
the smallest wavelength 2h which the
:rgy is exactly zero. This is reftected in
Lplacian which would arise from this
ffness matrix K gives
SEC. 4.3. NUMERICAL INTEGRATION 189
This is the 5-point scheme rotated through 45°, and seems dangerously
unstable; in fact, our oscillating twist gives a solution to the homogeneous
equation KQ = O.t
A similar one-point rule at the centroids of triangles, with linear trial
functions, will not be indefinite. The strains are constant, and cannot vanish
at the centroid without being identically zero. In fact, the Laplacian gives
The basic test for definiteness is to determine whether or not there are
trial functions which, under numerical quadrature, give up all their strain
energy. In practice this is decided from the rank of the element stiffness
matrices; if the only zero eigenvalues come from rigid body motions, the
quadrature is all right. If there are additional zero eigenvalues, the quadra-
ture might still be acceptable: it has to be checked whether or not the offend-
ing polynomials in separate elemehts can be fitted together-as in the case
of the twist described above-into a trial function vh which has too little
energy over the whole domain. For example, four-point Gauss integration
(2 X 2) does not ·satisfy our stability condition for the nine-parameter bi-
quadratic. With Gauss points ( ±e, ±e) on the square centered atthe origin,
(x 2 - e 2
e
)(y 2 - 2
) has zero strain energy, and this pattern can be translated
to give trouble over the whole domain. (The matrix K may not be actually
singular, if this pattern does not fit the boundary conditions (say vh = O) of
the problem. In this case one could live dangerously and try such a four-point
integration, even though K is much nearer .to singularity than the theory
demands.)
The question of definiteness becomes rather delicate for the important
eighÍ:·parameter element, obtained from the biquadratic by eliminating-the
x 2 y 2 term and the node at the center of each mesh square. Sin ce the trial
functions no longer include (x 2 - <!' 2 )(y 2 - <!' 2 ), the four Gauss points are
apparently enough for a stable finite element approximation to Laplace's
equation. Taylor has pointed out, however, that for plane elasticity with
two dependent variables, the situation is different: the combination u=
x(y 2 - e 2
), V = - y(x 2 - <!' 2 ) lies in the trial space, and its strain energy van-
ishes for the 2 X 2 rule. ~ut Taylor has also demonstrated that this pattern
cannot be continued into a neighboring element; the rank of a single element
matrix is too low, but after assembly the global stiffness matrix is perfectly
nonsingular and the numerical integration is all right.
So much for definiteness, which is a question of having enough integration
points. We turn now to accuracy, which is decided by the polynomials for
which the quadrature is exact. There are two ways to develop the theory. One
is to compute directly the exponent p which appears in equation (9) .of the
t A recent paper by V. Girault analyzes the discrete system which results from this quad-
rature rule. It turns out to be surprisiRgly useful in sorne contexts, even though it violates
our hypothesis of positive definiteness.
190 VARIATIONAL CRIMES CHAP.4 SEC. 4.3.
"\ "

theorem; w~)ake)this·approach in the next two paragraphs, writing clown In two or three dimensions the I
explicit boutJ.ds (14-15) for the errors in numerical integration. Then in the nonials of degree q + 1, say x11.yP, th
final paragraphs of this section we describe a simpler' and neater argument, rise to corresponding terms in the en
which leads immediately to the connection between the exactness of the
quadrature rule and the order of the resulting error uh- üh.
First 1!/proach. Por a quadrature exact to degree q, the error in comput-
ing f g(iJdx numerically will be bounded by C hq+t f 1g<q+o (x) 1 dx. This is
the exact counterpart of the Approximation Theorem 3.3, and is proved in The strain Dmuh is smooth, but ever:
the same way. Applied to the expressions (9) which appear in Theorem 4.1, factor h- 1 • The convergence conditit
this becomes pear. Therefore the full q + 1 dil
hilate nmvh for any trial function if.
(14) l(a*- a)(u', u'- ii')l Ch•+l ~ f..iCfxt\p(xM(u~ íí~) any of the terms x~~.yp for which the
the convergence criterion (6) given e
+ q(x)u"(u"- U')]' dx. trial function should be integrated e:
Second approach. Suppose that tl

(15) 1(b* b)(f, u'- íí') 1 < Ch•+ ~f.. 1Cfxtucu• íí')]l dx.
1 n - m, multiplied by any mth deriv;
puted exactly by the quadrature. Wt
exponent p in Theorem 4.1 is p n
Suppose the datafis smooth, as well as the variable coefficient p(x) we have dueto numerical integration will be
now allowed. Then u is also smooth, and so is its finite element approxima- the two quantities a - a* and b 1
tion uh. Therefore the only uncontrolled terms on the right sides are uh - fíh A typical term in (a- a*) (u\ v"
and its derivative. Every differentiation ofthese trial functions can introduce a a coefficient c(x, y), looks like
factor h- 1.t It would appear that q + 1 differentiations could completely
cancel the factor hq+ 1 , and destroy the proof of convergence. Here is the
point at which it is essential for the tria! functions to be polynomials. Since
u~ u~ is of degree k - 2, this is the maximum number of factors h- 1 which
catt appear; further differentiations would annihilate the polynomiaL There- The essential point, coming from m
fore the expression (14) is of order hcq+tJ-<k- 2 ) lluh- ühll 1 • The same is true exactness of the quadrature, is that
ofthe other expression (15); the first differentiation gives u~- ü~, and we are
back to the same argument. (We shall return to this point, and show in more
detail why b* bis of the same order as a* - a if the same quadrature rule
is applied.) We conclude that p = q- k+ 3 in Theorem 4.1, and therefore Since numerical integration is carric
that the effect of numerical quadrature is bounded by centrate on a specific element and el
between the two, according to our 1
be of order hn-,m+t_provided cDmuh
This coincides with the expgnent q - k + m 2 given earlier. Por N-point case if the original problem is smoot
Gauss quadrature, the degree of accuracy is q = 2N 1, and the resulting elements, the error (16a) is ofthe rig
error is O(h 2 N-k+ 2 ). Therefore the use of k 1 Gauss points should be com- A variation of this argument yi~
pletely successful in one dimension: it is enough to satisfy the requirement of below order m in the energy, and als
definiteness, ~nd it leads to an error of order hk. This is even of lower order the steps involved in estimating thi
than the approximation error in the strains. ' simplicity that the mth derivatives
polynomials ofsome degree t. Then e
tMore precisely, lvhlS+t < Ch- 1 lvhls for aJI s. amounts to correct integration of
 \)
' CHAP.4
1~ next two paragraphs, writing down
• m numerical integration. Then in the
:scribe a simpler and neater argument
nection between the exactness oJ) th~
resulting error uh- fíh. ·· rY
~xact to degree q, the,. error in comput-
tded by e hq+ 1 f 1 gCq+.'o (xJ1 dx. This is
mation Theorem 3.3, and is proved in
ions (9) which appear in Theorem 4.1,
., ~ J../C~T'rpCxM(u~
ü!J
•(x)u'(u"- ü')]l dx.
'~J../ (fxf'ucu" - ü'll/ dx.
:ts the variable coefficient p(x) we ha ve
1nd so is its finite element approxima-
~d terms on the right sides are uh - ilh
of these trial functions can introduce a
+- 1 differentiations could completely
1e proof of convergence. Here is the
·ial functions to be polynomials. Since
naximum num ber of factors h- 1 which
lUid annihilat~the polynomial. There-
l-1)-(k-z> 11 uh fíhllt· The same is true
fferentiation gives u! - u~, and we are
~eturn to this point, and show in more
as a* a if the same quadrature rule
k + 3 in Theorem 4.1, and therefore
! is bounded by
Chq-k+3.
( + m + 2 given earlier. For N-point
acy is q = 2N_ 1, and the resulting
.f k - 1 Gauss points should be com-
; enough to satisfy the requirement of
order hk. This is even of lower order
·ains. · ' simplícity that the mth derivatives of the trial functions vh consist of all
)) S.
SEC. 4.3. NUMERICAL INTEGRATION 191
In two or three dimensions the principie is the same. For certain poly-
nonials of degree q + 1, say x(l;yP, the quadrature will be inexact. This gives
rise to corresponding terms in the error a* - a, of the form
The strain Dmuh is smooth, but every differentiation of Dmvh can introduce a
factor h- 1• The convergence condition is that at most q of these factors ap-
pear. Therefore the full q + 1 differentiations (ajax)(l;(a¡ay)P must anni-
hilate Dmvh for any trial function vh. In other words, Drrivh must not include
any of the terms xa.yP for which the quadrature is inexact. This is precisely
the convergence criterion (6) given earlier, that the mth derivatives of every
trial function should be integrated exactly.
Second approach. Suppose that the integral of any polynomial of degree
n m, multiplied by any mth derivative Dmvh of any trial function, is com-
puted exactly by the quadrature. We want to show, following [S7], that the
exponent p in Theorem 4.1 is p n - m + 1 ; then the error in the strains
dueto numerical integration will be of this order hP. To do so we consider
the two quantities a - a* and b b* which appear in the inequality (9).
A typical term in (a a*) (uh, vh), with material properties expressed by
a coefficient c(x, y), looks like
The essential point, coming from our condition a(P vh) a*( P.,, vh) on the
11
,
exactness of the quadrature, is that this term can be rewritten as
Since numerical integration is carried out an element at a time, we can con-
centrate on a specific element and choose Pn-m close to cDmuh. The difference
between the two, according to our mean-square approximation theory, will
be of order hn-m+ 1-provided cDmuh is sufficiently smooth, which will be the
case if the original problem is smooth. Summing the results over the separate
elements, the error (16a) is of the right order hn-m+ 1 •
A variation of this argument yields the same bound for any derivatives
below order m in the energy, and also for the term b - b*. We shall describe
the steps involved in estimating this ·error in the latter term, assuming for
polynomials of sorne degree t. Then our exactness condition on the quadrature
amounts to correct integration of all polynomials of degree n - m + t.
192 VARIATIONAL CRIMES CHAP.4 SEC. 4.4. APPROXIMATION OF D(

Therefore This section discusses four proble1

1. Change of domain with a ho1
(b - b*)(f, vh) = JJfvh dx dy - 2: w¡(fvh)(e) onr.
= JJ (fvh Pn-m+t) dxdy 2: w¡(fvh - Pn-m-t)(e¡)· on
2. Change of domain with a horn
r.
3. Approximation of an inhomog
With the ·right choice of P in each element, these quantities are of order on r.
hn-m+t+ 1 , multiplied by an integral of the absolute value of derivatives up to 4. Approximation of an arbitrar:
this order of fvh. But vh can be differentiated only t + m times, after which un + ct(x, y)u b(x, y) on r.
(being a polynomial) it disappears. Therefore, assumingf is smooth enough
to allow n m + t + 1 differentiations, the error b - b* is of order In each case we work with a sec
say Poisson's equation -Au-= fIn
simplicity of exposition: very little is E
three space dimensions. For apure B
At the last st'ep, the removal of t differentiations from vh was paid for, as in higher order, saya plate with its edge
the footnote on p. 190, by a factor h-e. error in
strain energy is the same as
Thus both a- a* and b b* are of the correct order hn-m+ 1 , and by A new and quite different situatio
Theorem 4.1 the strains are in error to this order. This is the main result of boundary condition are combined. r

the section: /f a(Pn, vh) a*(Pn, vh), then 11 uh üh 11m O(hn-m+ 1). Ciarlet the biharmonic equation A2 U f acc
and Raviart have been able to show, even allowing for the use of iso- Poisson's ratio v enters the natural t:
parametrics [6], that the error in displacement shows the usual improvement
over the error in strains. (Their proof is a subtle variation of Nitsche's (17) u O and vAu +
trick.) Therefore the displacement error is of order hn+ 1 , and the theory
of numerical quadrature is in a satisfactory state: n = m is necessary for Jt is no longer possible to guarantee,
convergence, and n k - 1 is sufficient to reduce the errors due to quad- matically well-posed problem, that !
rature to the same level as the errors due to.approximation by polynomial solution on the polygon is close to the
tria! functions. exact solution Uh on Qh and to the f
required than simple convergence .of
The di:fficulty is easy to see when
4.4. APPROXIMATION OF DOMAIN ANO dition Uh =O immediately forces 1
BOUNDARY CONDITIONS Therefore, the second boundary con
edge to AUh O, and the depende1
At the same time that the admissible functions in X'; are being approxi- propases introducing Vh AUh as
mated by piecewise polynomials, sorne other and quite different approximations order equations:
are being_made in the finite element method. In the first place, the domain
itself may be changed: !l is replaced by a nearby polygon !l\ or in the iso-
parametric method by a domain whose boundary is a piecewise polynomial. It
For such a second-order system '
would bé extremely di:fficult to handle an arbitrary domain in any other way.
approach the solutions of
Second, the boundary conditions themselves are subject to approximation;
if the problem specifies that U g(x, y) On r, Or that Un+ (tU b(x, y), AU = V and AV= f
then these fuñctions g and b will almost inevitably be interpolated at the nodes
on the boundary r (or on its approximation). We want to estimate the errors This is precise! y the problem of the
involved. . the limiting function is independent (
 CHAP.
dy - 2: wlfvh)(e¡)
dy 2: w¡(fvh - Pn-m-r)(e1).
element, these quantities are of order
th~ absolute value of derivatives up to
mtlated only t + m times, after which
1erefore, assumingf is smooth enough
ns, the error b b* is of order
~entiations from vh was paid for, as in
t higher order, saya plate with its edges either clamped or free, the order ofthe
: o f. the correct order hn-m+ 1 ~ and by
> thts order. This is the main result of
then 11 uh íJh 11m = O(hn-m+ 1). Ciarlet
v, even allowing for the use of iso-
Lcement shows the usual improvement
of is a subtle variation of Nitsche's
ror is of order hn+ 1 , and the theory
factory state: n m is necessary for
mt to reduce the errors due to quad-
due to- approximation by polynomial
~NO
le functions in Jejl are being approxi-
:her and quite different approximations
ethod. In the firs t p~ce, the domain
r a nearby polygon í!\ or in the iso-
>oundary is a piecewise polynomial. It
n arbitrary doínain in any other way.
selves are subject to approximation;
y) On r, Or that Un + rtU = b(x, y),
nevitably be interpolated at the nodes
.tion). We want to estímate the errors
4 SEC. 4.4. APPROXIMATION OF DOMAIN AND BOUNDARY CONDITIONS 193
This section discusses four problems in detail:
1. Change of domain with a homogeneous Dirichlet condition, u O
on r.
2. Change of domain with a homogeneous Neumann condition, un = O
onr.
3. Approximation ofan inhomogeneous Dirichlet condition, u g(x, y)
on r.
4. Approximation of an arbitrary inhomogeneous Neumann condition,
Un+ rt(X, y)u = b(x, y) on r.
In each case we work with a second-order equation in two dimensions,
say Poisson's equation -!:,.u-= f In large part this is for convenience and
simplicity of exposition: very little is altered ifthere are severa] unknowns and
three space dimensions. For a pure Dirichlet ora pure Neumann problem of
er~or in strain energy is the same as the one demonstrated below.
A new and quite different situation arises when an essential anda natural
boundary condition are combined. This occurs in a slmply supported plate:
the biharmonic equation ll. 2 u faccounts for the loading in the interior, and
Poisson's ratio v enters the natural boundary condition:
(17) u = O and vl\u + (1 - v)unn O on r.
It is no longer possible to guarantee, in this physically important and mathe-
matically well-posed problem, that if thé polygon Qh is clase to n, then the
soluiion on the polygon is close to the solution u on Q. This applies both to the
exact solution Uh on Qh and to the finite element approximation uh. More is
required than simple convergence of the boundaries.
The difficulty is easy to see when Qh is a polygon. On each edge, the con-
dition Uh =O immediately forces the tangential derivative U~t to vanish.
Therefore, the second boundary condition in (17) is equivalent on a straight
edge to fl.Uh O, and the dependence on v has disappeared. Babuska [B1]
proposes introducing Vh = !:,.Uh as a new unknown, yielding two second-
order equations:
For such a second-order system convergence is guaranteed, and Uh, Vh
approach the solutions of
ll.U =V and :ll.V =f in n, U V=O onr.
This is precisely the problem ofthe clamped plate with v 1, and therefore
the limiting function is independent of the Poisson's ratio actually specified in
 ·.~

194 VA~IA?ONAL CRIMES CHAP.4 SEC. 4.4. APPROXIMATION OF DOM

the boundary conditions. Convergence occurs, but nearly always to the wrong Q
· answer. Corresponding difficulties with finite élement calculations are report-
ed in [Rl] and discussed in [Bll]. We would anticípate success in the isopara-
metric method, on the other hand, if the approximation of r is at least
piecewise quadratic; the curvature of the boundary converges in this case.
AlternativY.ly, suppose the essential condition u O were replaced at bound-
auhjat = o, using the cubic trial space Z 3 (see Section 1.9)
1
ary nodes by uh
and taking the tangent to the true boundary r. Then convergen ce is expected
even on a polygon. At this writing, however, the required theory does not
exist.
Fig. 4.2 Polygonal appro:x
We turn to the four problems listed above. In each case the analysis is
rather technical, but the conclusions are straightforward.
boundary r' the mean-value theorem :
l. Suppose that Q is replaced by an inscribed polygon Qh, and the trial
functions are made to vanish on the straight edges of rh. Then imagining (20) !u(M)I Ch 2 :
that they are defined to vanish everywhere outside rh, they are admissible in
the variational problem; they vanish on tl:ie true boundary r, and. the trial It is no longer permissible to choose v
space Sh is a bona fide subspace of 3Có(Q). Therefore, the fundamental Ritz vh will be the piecewise quadratic uj ~
theorem 1.1 assures that uh minimizes the error in strain energy: only and vanishes on rh. The error in
using the triangle inequality, is
(18) JJ(u uh); + (u - uh); n;~n JJ(u- vh)~ + (u vh);.
a a JJ(u uj); + (u - uj);
Qh

Since every vh vanishes outside. Qh, the integral over the skin Q Qh is
fixed; it is just the integral of u; + u;. Therefore, uh is minimizing over Qh as
well as Q:
The term 1u u1 II is completely famili:
(19) JJ(u uh)~ + (u uh); ~}n JJ(u vh); + (u - v");.
ter m u1 - uj is a piecewise quadratic
~ ~
gles: the only nodes at which u1 ::f.= u~
these points uj =o and ul u= e
The question is simply to estímate (18) and (19). U p to now we ha ve chosen triangles is 0(1/h), and
vh to be the interpola te u1 • If Sh is Courant's piecewise linear space, with nodes
at the vertices of the triangles, this is still a good choice. Since it interpola tes
u O at boundary nodes, it vanishes along the boundary of the polygon and
1u1 uj 11 L JJ 1u(M¡) P

lies in Sh. _The standard approximation theorem 3.3 yields an h 2 error in

energy. No doubt the approximation is very poor near the boundary.
Suppose a more refined element is used, for example a piecewise quadratic
(Fig. 4.2). As usual, the nodes are placed at the vertices and midpoints of the Here rp is the quadratic with value 1 at
edges. Each trial function will vanish along rh provided it is zero at aii bound- five nodes ofthe triangle. Its first derh
ary nodes, but this excludes u1 from the tria! space Sh. The true solution u the triangle is of order h2 , so that fina
vanishes at the vertices P and Q but not at the midpoint M, so the interpola te
is not zero along the boundary. Since Mis a distance O(h 2 ) from the true !u1 ujj
 CHAP.4 SEC. 4.4. APPROXIMATION OF DOMAIN AND BOUNDARY CONDITIONS 195

occurs, but near1y always to the wrong

th finite element calculations a~~ report-
would anticípate success in tfie ?i~opara­
if the approximation of r is )¡t least
f the boundary converges in this case.
Jndition u o were replaced at bound-
he cubic trial space Z~ (see Section 1.9)
mdary r. Then convergence is expected
however, the required theory does not

ted above. In each case the analysis is Fig. 4.2 Polygonal approximation of the boundary.
are straightforward.
boundary r, the mean-value theorem yields
an inscribed polygon !l\ and the trial
straight edges of rh. Then imagining (20) [u(M)[ Ch 2 max 1 grad u[.
here outside rh, they are admissible in o
on the true boundary r, and·the trial
It, is no Ionger permissible to choose vh u1 in (19). Instead, a convenient -
:J(Q). Therefore, the fundamental Ritz
vh will be the piecewise quadratic u1 which interpolates u at interior nodes
the error in strain energy:
only and vanishes ·on rh. The error in strain energy over Qh for this choice,
using thetriangle inequality, is
~!n JJ(u .vh); + (u vh);.
o
un;+ (u u1); =[u- U¡+ U¡ uJ'Ii
he integral over the skin n - Qh is
'¡
<([u u1 l1 + 1u1 uJ'I1 )2
Therefore, uh is minimizing over Qh as
<liu u1 11 + 2[ u1 - uJ'[i ·

The term 1u- u1 121 is completely familiar; it is O(h 4 ) by Theorem 3.3.• The new
~}n JJ(u
•
vh); +(u- vh);. term u - u* is a piecewise quadratic which vanishes over all intenor tnan-
o~<
gles: the onÍy no des at which u1 *
u1 are the boundary midpoints M;; at
these points ur o and U¡ u O(h 2 ). The number of such boundary
8) and (19). Up to now we ha ve chosen triangles is 0(1/h), and
mt's pfecewise linear space, with nodes
ill a good choice. Since it interpolates
.ong the boundary of the polygon and
1U¡ u11i = JJ u(M¡)I (cp; + cp;) dx dy
2: 1
2

e¡
m theorem 3.3 yield§ an h 2 error in
s very poor near the boundary. <o( k) c h max grad u [ JJ(cp; + cp;) dx dy.
2 4
1
2

5ed, for example a piecewise quadratic

d at the vertices ahd midpoints of the
Here cp is the quadratic with value 1 at the midpoint M; and zero at the other
ong rh provided it is zero at aU bound-
five nodes ofthe triangle. Its first derívatives are of order 1/h, and the area of
e tria! space Sh. The true solution u
the triangle is of order h 2 , so that finally
at the midpoint M, so the interpolate
M is a distance O(h 2) from the true
196 VARIA TIONAL CRIMES CHAP.4
SEC. 4.4. APPROXIMATION OF D

THEOREM4.2 St. Venant principie, but one which a 1

more customary, to the boundary da
The error in energy over Q\ produced by polygonal approximation of the boundary, it is the average which ma
domain, satisfies oscillations.
The situation is not unlike that (
(21) 1u - uh li < 1u - .u11i = O(h 3 ).
with a wavy boundary. Suppose the 1
Then in the conforma! map betwee
The error 'In function values is u - uh ·= O(h 2 ).
into (r + h2 r 11h cos 0/h,O). All the
The latter estímate (which we shall not prove) is suggested by the contin-
r > 1 - Ch; for smaller values of r,
tion term is exponentially small,
uous problem on the approximate domain: -dUh = f on Q\ Uh =O on
rh. Since d(u - Uh) = - f + f = O, the maximum principie applies, and
u - Uh attains its maximum on rh. But Uh = O and u= O(h 2 ) on this
boundary, so that u- U 11 is of order h 2 everywhere. The theorem asserts the
same result for u - uh [B20]. · The derivatives of the mapping are ~
they are virtually zero, because of r
1t follows that if we loo k only at the exponent of h, and approximate Q by
same area; if one were inscribed in t
a polygon, then there is no point in going beyond quadratic shape functions
term rh 2 , whose derivati ves do not de~
in computing stresses. Furthermore, even linear polynomials will be correct
the domain. In St. Venant terms, tl
to the best possible order h 2 in the displacements. Probably this is an instance
nonzero and is propagated. Furthern
in which the exponents of h are not sufficient to indicate the true accuracy;
of wavy, with 1cos 0/h 1 instead of c1
the actual computed error with linear elements may be excessive.
possess weak singularities at the cus]
A similar estímate applies to three-dimensional and to higher-order
however, these are smeared over and
Dirichlet problems. 1t is most unusual that the exponent should be odd;
h at the boundary and h2 inside.
th~ average error in the deriva ti ves of u - uh is of the fractional order h 3 12 • The
There are a number of ways to n
estímate itself is nevertheless correct, and in fact the exact solution Uh to the
quadratics. Coordinate change by i
equation on the polygon [which is closer to u than uh is, because ids minimiz-
1 x-y element which has piecewise hy
ing over the space Xb(Qh), which contains S,;] is also in error to order h31 2 . !
been rrientioned. Another possibility
[BlO]. For a rectangular approximation ofthe boundary, which is much more
polygonal approximation · uh [826],
crude, the order becomes h 112 ; the computational results would be com-
l(v) [N3]. In all these techniques it is 1
pletely unsatisfactory. That is the reason for triangular elements.
the error dueto change of domain is
. The explanation for the fractional exponent ! appears to be this: There
atically in and out of r. This is St
1s a boundary !ayer, a couple of elements thick, within which the derivatives
term depends on the area of n - Q
e
are in error by O(h). It is easy to verify that the angle in the figure is of this
further possibility: to average the a
order, so that the true deriva ti ve of u along the chord is O(h) rather than zero.
circumscribed polygons. All these pro
This layer alone accounts for the O(h 3 ) error in energy. Beyond the boundary
are largely untested in practica! prot
layer, the error is quite different: it is of the optimal order h 2 in the displace-
ment, that is, u - uh = O(h 2 ), and it is so smooth that the first derivatives
2. Consider next the equatiort -
are also of order h 2 • The error is much better behaved in the interior than in
condition un= O. In this case there
the boundary, illustrating the smoothing property common to all elliptic
on the wrong boundary; the trial fm
problems. In the finite element approximation, u - uh oscillates from zero to
However, a change of domain does
O(h 2 ) and ba~k along each c-hordPMQ; this singular behavior is damped out
integrations over the true domain ~
rapidly in the direction normal to the boundary. The damping is very visible
instead over an approximate domain
in computations and has been explicitly verified. It represents a kind· of
 CHAP.4
ced by polygonal approximation of the
not prove) is suggested by the contin-
)main: -JlUh f on 0\ Uh O on
the maximum principie applies, and
But Uh O and u = O(h 2 ) on this
~ everyw~ere. The theorem asserts the
: exponentof h, and approximate n by
·ing beyond quadratic shape functions
ven linear polynomials will be correct
lacements. Probably this is an instance
tfficient to indica te the true accuracy;
elements may be excessive.
~ee-dimensional and to higher-order
al that the exponent should be odd;
uh is ofthefractional order h 31 2 • The
nd in fact the exact solution Uh to the
:r to u than uh is, beca use ifis minimiz-
ains Sh] is also in error to order h 3 12
of the boundary, which is much more
omputational results would be com-
tn for triangular elements.
!Xponent fappears to be this: There
,ts thick, within which the derivatives
that the angle (} in the figure is of this
mg the chord is O(h) rather than zero.
error in energy. Beyond the boundary
f the optimal order ft2 in the displace-
s so- smooth that the first derivatives
better behaved in the interior than in
ing property common to all elliptic
mation, u uh oscillates from zero to
; this singular behavior is damped out
)oundary. The damping is very v.isible
tly verified. It represents a kind of
SEC. 4.4. APPROXIMATION OF DOMAIN ANO BOUNDARY CONDITIONS 197
St. Venant principie, but one which applies to the geometry rather than, as is
more customary, to the boundary data. The rule is the same: Away from the
boundary, it is the average which matters and not the local short-wavelength
oscillations.
The situation is not unlike that of mapping the unit circle onto a circle
with a wavy boundary. Suppose the latter boundary is R((}) = 1 + h2 cos (Jjh.
Then in the conforma} map between these "circles," the point (r, (}) goes
into (r + h2 r 11h cos Ofh,O). A11 the activity lies inside a boundary layer
r > 1 Ch; for smaller values of r, the amplitude of the "wavy" perturba-
tion term is exponential1y small,
The derivatives of the mapping are O(h) near the boundary. In the interior
they are virtually zero, because of r 11h. Note that the two circles have the
same area; if one were inscribed in the other, there ~ould be an additional
term rh 2 , whose deriva ti ves do not decay but are of lower order h2 throughout
the doinain. In St. Venant terms, the average over the local oscillations is
nonzero and is propagated. Furthermore, if the circle were scalloped instead
of wavy, with 1 cos 0/h 1 instead of cos 6/h, then the conforma} map would
possess weak singularities at the cusps. In the finite element approximation,
however, these are smeared over and the mean error in derivatives is of order
h at the boundary and h2 inside.
There are a number of ways to recover the optimal convergence rate for
quadratics. Coordinate change by isoparametric elements, and Mitchell's
x-yelement which has piecewise hyperbolas (!) as boundary, have already
been nientioned. Another possibility is to compute a correction term to the
polygonal approximation uh [B26], or to modify the original functional
l(v) [N3J. In all these techniques it is not essential that Qh lie inside Q; in fact,
the error dueto change of domain is partly self-cancelling, ifrh goes system-
atically in and out of r. This is St. Venant again, that the Ieading error
term depends on the area of Q - Qh (with algebraic sign), and suggests a
further possibility: to average the approximate solutions for inscribed and
circumscribed polygons. All these proposals, except for the isoparametric one,
are largely untested in practica! problems.
2. Consider next the equation -A u + qu = f, with natural boundary
condition un O. In this case there is no question of imposing conditions
on the wrong boundary; the trial functions are unrestricted at the boundary.
However, a change of domain does enter, if it is inconvenient to carry out
integrations over the true domain Q. If the potential energy is computed
instead over an approximate domain Qh, the effect is to introduce a new defi-
198 VARIATIONAL CRJMES CHAP. 4 SEC. 4.4. APPROXIMATION OF D(

nition of the potential energy: LEMMA 2.2

Suppose that p = area(T Th)fa
J"(v) ah(v, v)- 2(/, v)h depending only on the degree of the p(
= JJcv; + v;+ qv 2 2fv)dxdy.
Qll
-
ff p;
T-T 4
P;+pz
Assuming.that uh minimizes the true functional I over Sh and that uh mini-
mizes 1\ the problem is to estímate eh uh- uh. Summing over all boundary tria1
Mathematically, this is essentially the same question which arises for include also the interior triangles, the
Gaussian quadrature: the integrals are computed inexactly. Therefore, we
apply a variant ofthe identity ofLemma 4.1; it follows directly from the van-
ishing of the first variations at u, u\ and uh that
JJ (e~)2 + (e~Y
n-íl"

Substituting back into the estímate fo

The first term on the right, by the Schwarz inequality, is not larger than

In other words, the error in strain ener

n satisfies
The other terms on the right amount to an integral o ver the skin n- Qh:

B J J(-uxe~ u:ve~ queh + feh) dx dy. The first term reflects the strain enei
n-n4 new. The second term is the interes1
Suppose, for example, that Qh is a poi
Assuming that ux, u:r, q, andf are bounded, the Schwarz inequality this time ratio p to the area of the neighboring
involves the area A of the skin: error in changing to a polygon is of tht.
tia! boundary conditions. We conjectut
For isoparametrics instead of subpar
and their product h2 k-I is dominated
strain energy of order h2 k- 2 • Therefo
The remaining problem is to estímate this last integral in terms of in with curved regions, whether the boun
other words, to relate the size of eh in the skin to its size in the interior. For
arbitrary functions this would be impossible. The function uh u\ however, 3. The next problem is that of a
is by no means arbitrary: -¡n each triangle it agrees with sorne polynomial, condition, u= g(x, y) on r. This con
and polynomials cannot suddenly explode-a bound in Qh implies a bound the true admissible space 3C1. Any tw
in n. by a function v0 in Xó; the boundary
We denote by T a typical curv~d triangle at the boundary, and by Th Suppose that the situation is the s;
the inscribed straight ·triangle; their difference T -:- Th is one of the pieces trial functions assume the same valm
that go to make up n Qh. The trial functions are polynomials on each vh = g, that is too much to expect fron
triangle T, and we art: dealing with the error eh = uh ijh which is caused by Then any two trial functions differ b:
integrating only over Th. This error is itself a polynomial P(x, y) on each T, these functions vi form a space Si, w
and obeys the following lemma given by Berger. neous space Xá.
 CHAP.
!(/, v)h
: + qv 2 - 2/v) dx dy.
: functional 1 over S,; and that fth mini-
?h = uh _ fth.
ly the same question which arises for
ue computed inexactly. Therefore, we
ma 4.1; it follows directly from the van-
and i? that
hwarz inequality, is not larger than
t to an integral o ver the skin Q - Qh:
mded, the Schwarz inequality this time
~)' +(e;)' + (e')'T'·.
te this last integral in terms of E, in
the skin to its size in the interior. For
>ssible. The function uh - u\ however,
.ngle it agrees with ~ome polynomial,
1lode-a_ bound in Qh implies a bound
triangle at the boundary, and by Th
iifference T - Th is one of the pieces
al functions are polynomials on each
· error eh = uh - fth which is caused by
.itself a polynomial P(x, y) on each T,
by Berger.
4 SEC. 4.4. APPROXIMATION OF DOMAIN AND BOUNDARY CONDITIONS 199
LEMMA 2.2
Suppose that p = area(T- Th)jarea(Th). Then there is a constant e,
depending only on the degree of the polynomial P(x, y), su~h that
JJ P2; + P; + pz < eP JJP2; + P; + pz.
T-Th Th
Summing over all boundary triangles and increasing the right side to
include also the interior triangles, the lemma gives
JJ (e~)z + (e;)z + (eh)2 < cpEz.
Q-Qh
Substituting back into the estimate for 1 B ¡, the identity now reads
In other words, the error in strain energy dueto integration over Qh instead of
Q satisfies
The first term reflects the strain energy in u - uh and contributes nothing
new. The second term is the interesting one, and it is purely geometrical.
Suppose, for example, that Qh is a polyg'on. Then the area A is O(h 2), and the
ratio p to the area of the neighboring triangles is O(h). Thus the strain energy
error in changing to a polygon is of the same arder h 3 for natural as for essen-
tial boundary conditions. We conjecture that there is again a boundary layer.
For isoparametrics instead of subparametrics, A is O(hk) and pis O(hk- 1 ),
and their product h2 k-J is dominated by the ordinary approximation error in
strain energy of order h 2 k-z. Therefore isoparametrics should be successful
with curved regions, whether the boundary conditions are essential or natural.
3. The next problem is that of an inhomogeneous essential,boundary
condition, u= g(x, y) on r. This condition is satisfied by every member v of
the true admissible space JC1. Any two admissible functions therefore differ
by a function v 0 in JC¿; the boundary values of v 0 are zero.
Suppose that the situation is the same for the trial functions vh in Sh: All
trial functions assume the same values on the boundary r (not necessarily
vh = g, that is too much to expect from polynomials, but say vh = gh instead).
Then any two trial functions differ by a function V~, which vanishes on r;
these functions v~ forma space S~, which is a subspace of the true homoge-
neous space JC¿.
200 VARIA TIONAL CRIMES CHAP.4 SEC. 4.4. APPROXiMA TION OF DO

THEOREM 4.3 b. All tria] functions vh must assur

that their differences belong to 3CJ. ]
Suppose that u minimizes J(v) over 3C'; and uh minimizes over Sh. Then the
should be determined on r by their va
vanishing of the first variation is expressed by composed of flat edges, and we assum
(22) a(u, V0 ) = (f, V0) for all v 0 in 3CJ, this will be the case.

(23) a(u\ vi)=(/, vi) for aíl vi in Sí. We hesitated to give the usual estima
As in Theorem 1. 1, uh has the additional mJnimizing property requires that the solution be smooth (
domains, a singularity in the derivative
(24) a(u- u\ u- uh) = min a(u- v\ u- vh). order equations, u normally just fails
vh in Sh ·
interior angle. The first derivatives v.
The proof is copied from Theorem 1.t. Since u and uh are minimizing, any h"ia.; the strain energy error will be h2 '
perturbations EV 0 and fVÍ must increase the functional /: usual order f1Hk- 1 > can be preserved, 1
(better) by introducing. special trial fu
/(u)< /(u + fVo) and I(uh) < l(uh + fVÍ). Suppose that a nonpolygonal Q i
directly in the x-y plane, or else in the
Expanding, the coefficients off must vanish, and these are the virtual work formation. Jf the computations ate <l
equations (22) and (23). nodes completely determine all the be
To prove that uh minimizes the error in strain energy, write be applied: The final error is dominat
proximate domain, plus the error stud¡
a(u - v\ u - vh) = a(u - uh + uh - v\ u - uh + uh - vh)
There remains a further possibility
= a(u - u\ u - uh) + 2a(u - u\ uh - vh) given x-y coordinates and to interpoh
+ a(uh - v\ uh - vh). ably the integrations would be done r
elements at the boundary, although 1
The middle term must vanish automatically, subtracting (23) from (22) and out by exact algebraic operations. In
choosing vi = uh - vh. The last term will be positive unless vh happens to from one tria] function to anothex, ex<
equal uh; therefore, the minimum occurs at this point, and the theorem is the differences between the trial funct
pro ved. Ritz rules are therefore violated, and
are:
COROLLARY

Suppose that n is a polygon, and the essential condition u = g(x, y) on r a. How large can these differences
is interpolated in the finite element method: at al! boundary nodes, the tria! b. Js the error u - uh affected by
functions satisfy vh(z) = g(zj) [or more generally Djvh(z) = Djg(z)]. does uh still yield an optimal approxi
Then
These questions are relevant even w
(25)
u= g =O on r. In this case the tri~
Therefore, the error estimate in the finite element method is reduced to the nodes; question a asks how large the)
standard approximation estimate of u - u1 • the answer is independent oftheir deg1
same whether vh is linear, vanishing
Proof The conditions that must be checked are triangles, or quadratic or cubic, vanisl
a. u1 must lie in S\ so that the choice vh = u1 is possible in (24). Since u1 each piece of r. It is even the same fe
satisfies the boundary conditions imposed in the corollary, that is, u¡(z) = true tangential derivatives set to zero
u(z j) = g(z) at boundary no des, this requirement is me t. There exists a particular tria! function
 CHAP.4 SEC. 4.4. APPROXIMATION ·OF DOMAIN ANO BOUNDARY CONDITIONS 201

b. All trial functions vh must assume the same boundary values on r, so

that their differences be1ong to X~. In other words, the trial functions vh
:JC1t and uh minimizes over Sh. Then the
s·sed by
r
should be determined on by their values at the boundary nodes. Since is r
composed of flat edges, and we assume that a conforming eleinent is used,
for all v0 in Xó, this will be the case.
for all vi in Si. We hesitated to give the usual estima te h2 (k-t >for (25), since such an es ti mate
'11 minimizing property requires that the solution be smooth (u in :JCk). At the corners of polygonal
domairis, a singularity in the deriva ti ves of u is almost a u toma tic. For second-
min a(u vh, u - vh). order equations, u normally just fails to lie in :JCl+nlo:, where rx is the largest
'in S"
interior angle. The first derivatives will actualJy be approximated to order
1.1. Since u and uh are minimizing, any hnlo:; the strain energy error will be h 2 n1o:. In Chapter 8 we consider how the
:ase the fu'nctional 1: usual order h 2 (k-n can be preserved, by refining the mesh at the corners or
(better) by introducing special trial functions with the right singularities.
Suppose that a nonpolygonal Q is ·replaced by a polygonal Qh-either
vanish, and these are _the virtual work
dir~ctly in the x-y plane, or else in the e-r¡ plane after an isoparametric trans-
formation. Jf the computations are done on 0\ then again the boundary
nodes completely determine all the boundary values, and the corollary can
·or in strain energy, write
be applied: The final error is dominated by the error in u- u1 over the ap-
+ uh v\ u- uh + uh vh) proximate domain, plus the error studied earlier due to change in domain.
There remains a further possibility: to work on a curved domain Q in the
u uh) + 2a(u u\ uh- vh) given x-y coordinates and to interpolate vh = g at boundary nodes. Presum-
v\ uh- vh). ably the integrations would be done numerically, especially o ver the curved
elements at the boundary, although the experiments in [Cl4] were carried
tically, subtracting (23) from (22) and
out by exact algebraic operations. In this method the boundary values vary
will be positive unless vh happens to
from one trial function to another, except of course at the nodes themselves;
;urs at this point, and the theorem is
the differences between the trial functions are small but not zero on r. The
Ritz rules are therefore violated, and the theoretical questions which arise
are:

'w essential condition u = g(x, y) on r a. How large can these differences on the boundary beco me?
?thod: at al! boundary nodes, the tria! b. Is the error u - uh affected by this worst possible behavior on r, or
!Ore generally Divh(z¡) Dig(z¡)]. does uh still yield an optima) approximation to u from Sh?

- a(u - u1 , u u1 ). These questions are relevante ven with homogeneous boundary conditions,
u =g =O on r. In this case the trial functions are zero at the boundary
nite element method is reduced to the nodes; question a asks how large they can be on the rest of r. Surprisingly,
U¡. the answer is independent of their degree. The maximum size of vh on r is the
same whether vh is linear, vanishing only at the vertices of the boundary
e checked are triangles, or quadratic or cubic, vanishing at one or two additional nodes on
1ice vh u1 is possible in (24). Since u1 each piece of r. It is even the same for the cubic space Z 3 , with values and
osed in the corollary, that is, u¡{z¡) = true tangential derivatives set to zero at boundary vertices. The answer is:
requirement is met. There exists a particular tria! function Vh of unit energy whose mean value on
202 VARIATIONAL CRIMES CHAP.4 SEC. 4.4. APPROXIMATION OF DC

r is of.order h 312 : paper, a second rather surprising fact

Iimited by the existence of an undesira
f f (V~)z + (V;)z
o
l'
the true displacement error u u" is
words, the boundary values of the Ri1
lt must be the normal derivatives at t
To an~wer question b, that is, to find a bound on u - uh, it will be neces- We consider now the possibility of
sary to eitend the classical Ritz theory. It is true that the first variations the right side of (26). If the boúndary
still vanish ·at the minimizing functions u- and uh: zero, then even though their average
the integral itself will tend to be an o
a(u, v) (/, v) for v in Je8, lation does occur for sorne elements, bz
for example, when the condition u O
However, even with g = O, Sh is nota subspace of JeA; in the inhomogeneous vertices, and at the points on r which
case g -::f::. O, Sg is not contained in Jeó. Nevertheless, we may still appeal to in vh is a cubic s(s h/2)(s - h) in th
Green's formula: Sin~e - Au f, at boundary nodes, and this cubic ha
vh oscillates around zero. Correspond
a(u, vh) = JJ uxv~ + u~v; JJfvh + Jr unvh ds. in energy dueto interpolating u= g i
occur for cubics [810], unless the nod
o o
give vh an average val ue of zero.) Sinct
Subtracting a(u\ vh) = (f, vh), the boundary term remains: quadratics isO(h 2 ) in energy, the boUJ
ceptable. We don't know whether t
(26) boundary triangles; namely integratio
x-y plane, has a useful future.
lt is this term which controls the error due to violating the essential boundary 4. The final boundary problem on
conditions. natural condition
We propose to show that question b usually has the worst possible answer:
The order of error will be decided by the trial function Vh which is largest un + tX(x, y)u =' b(x,
on r. In fact, with vh Vh in (26), that is · easy to see. Since Vh has unit
energy, the Ieft side is bounded by the square root of a(u- uh, u uh); For Poisson's equation, this Ieads to ti
this is the Schwarz inequality. On the right side, Vh is of average value· h3lz, the potential energy:
and unless there is sorne special cancellation (see below !) we must expect the
integral of unVh to be of the same order. Therefore, /( v) JJ v; + v; + Jr
o

Strictly speaking, there are three

This woul~ mean that interpolating. the boundary conditions and integrat- these boundary integrals: polynomial
ing over Q is no better, even for an element of high degree, than replacing gration, and change of domain. The la
Q by a polygon Qh. (At least, the exponent of h is no better: we ha ve no estí- it wiJI be nearly impossible to compute
mate of the constants involved.) One explanation is this: Each polynomial numerical procedure will be adopted, a
which interpolates u Oat boundary nodes will vanish along a curve which on Qh rather than n, the integrals will
runs close to the true boundary, but the two curves may differ by as much as a detailed analysis of these errors on
O(h 2 )-th~ same distance as for a polygon Qh. are no larger than those already studie1
It is shown in [810] that h312 is also an upper bound for the error due to very desirable.
interpolating the boundary conditions. From the manipulations in that The novel question is that of polyn
 CHAP.4
tnd a bound on u tJ\ it will be neces-
)ry. 1t is true that the first variations
ts u and uh: · zero, then even though their average modulus 1vh 1 may be as large as h 312 ,
subspace of x¿; in the inhomogeneous
~. Nevertheless, we may still appeal to
); = ff fvh +Ir unvh ds.
o
.ndary term remains:
ds for all vh in Sh.
jue to violating the essential boundary
' usually has the worst possible answer:
the trial function Vh which is Jargest
1at is easy to see. Since Vh has unit
he square root of a(u uh, u uh);
right side, Vh is of average value h 31z,
lation (see below!) we must expect the
~r. Therefore,
.e boundary conditions and integrat-
ement of high degree, than replacing
1ent of h is no better: we ha ve no esti-
explanation is this: Each polynomial
1odes will vanish along a curve which
! two curves may differ by as much as
gon .Qh.
an upper bound for the error due to
s. From the manipulations in that
SEC. 4.4. APPROXIMATION OF DOMAIN ANO BOUNDARY CONDITIONS 203
paper, a second rather surprising fact emerges: Even though the accuracy is
limited by the existence of an undesirable Vh of order h 312 on r, nevertheless
the true displacement error u- uh is actually of order h3 on r. In other
words, the boundary values of the Ritz solution will look deceptively good.
It must be the normal derivatives at the boundary which are most in error.
We consider now the possibility of cancellation in the integral J unvh ds on
the right side of (26). If the boundary values vh happen to oscillate around
the integral itself will tend to be an order of magnitude smaller. This oscil-
lation does occur for sorne elements, but not for al!. It occurs for quadratics,
for example, when the condition u O(or u g) is interpolated at boundary
vertices, and at the points on r which lie halfway between. The leading term
in vh is a cubic s(s - hf2)(s h) in the are length variables on r, vanishing
at boundary nodes, and this cubic, has average value zero. In other words,
vh oscillates around zero. Correspondingly, Scott has shown that the error
in energy due tointerpolating u = g is improved to O(h 5 12 ). (This does not
occur for cubics [BIOJ, unless the nodal point~ are specially placed on r to
give vh an average value of zero.) Since the ordinary approximation error for
quadratics isO(h 2 ) in energy, the boundary error now appears perfectly ac-
ceptable. We don't know whether this alternative to isoparametrics for
boundary triangles, namely integration over curved elements directly in the
x-y plane, has a useful future.
4. The final boundary problem on our list arises from an inhomogeneous
natural condition
un+ a(x, y)u = b(x, y) on r, a O.
For Poisson's ·equation, this Ieads to the appearance of boundary integrals in
the potential energy:
l(v) JJ v; + v; + Jr av 2
- 2 JJfv - 2 Jr bv.
n n
Strictly speaking, there are three questions to consider in relation to
these boundary integrals: polynomial approximation on r, numerical inte-
gration, and change of domain. The last two arise beca use, when r is curved,
it will be nearly impossible to compute the boundary integrals exactly. Sorne
numerical procedure will be adopted, and in case the trial functions are defined
on ,Qh rather than n, the integrals will be shifted to rh. we pro pose to omit
a detailed analysis of these errors on rh, because we are confident that they
are no larger than those already studied on the interior .Qh. A proof would be
very desirable.
The novel question is that of polynomial approximation on r. It arises in
204 VARIATIONAL CRIMES CHAP.4

the ust,1al way: The method minimizes the error in energy,

a( u - u\ u uh) = min JJ(u - vh); + (u - vh); + Jr rx(u - vh) 2 ds.

n

For a finite element space of degree k - 1, and the choice vh u1 , the inte-
gral overr',a is of order h2ck-u. Fortunately, the integral o ver r is even of
higher order; the rate of convergence is not reduced by the presence of boundary
integrals. This is obvious if r is composed of straight lines; the restriction of
the trial functions to r yields a complete polynomial of degree k - 1 in the
boundary variables, and the integral o ver r is of order h 2 k. Nitsche [6] has
obtained the same result for a curved boundary.
5 STABILITY

With isoparametric elements the boundary in the t;-r¡ plane is straight,

and all the boundary integrals can be computed directly by numerical inte-
gration. In fact, this seems to be the main conclusion of the boundary theory
for second-order problems: The isoparametric technique establishes a local
change of coordinates into normal and tangential directions, which is more
accurate and com•enient than was el'er achiel'ed with finite differences. 5.1. INDEPENDEÑCE OF THE BASI:

In one sense, there should be no 1

elliptic variational problem depends
f and all pres~ribed boundary displa,
strain energy in u is small. In. other wo
more, regardless of the choice of the
Ritz approximation úh is automaticall:
projects u onto S\ and this can only r
1.1). Therefore, the approximate pro
should always be possible to construc
calculation of uh.
The difficulty is that to achieve tot
ing as the mesh size is reduced to zer
be too fancy. For the standard fiv
systematically use all the constraints
from consistency with the Laplace o
the matrix, as well as the first momc
elimination process. For finite elemen
responding constraints will be extrem
analyze the standard e1imihation algm
error as h O but intending that tht
proof as possible; unnecessary nume
It is known that the key to stabilit
of the basis elements tp r Even thoug
choice of basis, the roundoff which en

2
 CHAP.4

s the error in energy,

: - 1, and the choice vh = u1, the inte-

mately, the integraltbver r is even of
not reduced by the pre~ence of boundary
o sed of straight lines; the restriction of
lete polynomial of degree k - 1 in the
over r is of order h 2 k. Nitsche [6] has
boundary.
5 STABILITY

Joundary in the e-t¡ plane is straight,

! computed directly by numerical inte-
min conclusion of the boundary theory
zrametric technique establishes a local
i tangential directions, which is more
chieved withfinite differences. 5.1. INDEPENDENCE OF THE BASIS

In one sense, there should be no problem of stability. The solution of an

elliptic variational problem depends continuously on the data; if the load
f and all pres~ribed boundary displacements and forces are small, then the
strain energy in u is small. In.other words, the problem is well posed. Further-
more, regardless of the choice of the subspace Sh, the strain energy in the
Ritz approximation üh is automatically bounded by that in u; the Ritz method
projects u onto S\ and this can only reduce the energy (corollary to Theorem
1.1) .. Therefore, the approximate problems are uniformly well posed, and it
should always be possible to construct a "stable" procedure for the numerical
calculation of uh.
The difficulty is that to achieve total numerical stability-conceding noth-
ing as the mesh size is reduced to zero-the required algorithm may simply
be too fancy. For the standard five-point difference scheme, one could
systematically use all the constraints on the coefficient matrix which come
from consistency with the Laplace operator: The sums along each row of
the matrix, as well as the first moments, vanish at all stages of the Gauss
elimination process. Por finite elements of irregular shape, however, the cor-
'
responding constraints will be extremely difficult to use. Therefore, we shall
analyze the standard elimination algorithm, accepting an increase in roundoff ·
error as h - t O but intending that the numerical stability should be as fool-
proof as possible; unnecessary numerical instabilities will not be accepted.
It is known that the key to stability lies in the uniform linear independence
of the basis elements rp r E ven though uh is complete] y independent of the
choice of basis, the roundoff which enters its computed value uh does depend

205
206 STABILITY CHAP. 5 SEC. 5.1.

on this choice. (Mikhlin refers to the "strong minimality'' of the basis, and roundoff error would not mean a lat
Soviet authors are generally careful to consider Ílumerical methods from is small, but at the particular node z1
this point of view, as well as for stability with respect to coefficient changes tively poor .
in the differential equation.) To quantify the linear independence of the basis, . The standard rule in scaling aspa
the standard procedure is to consider the Gram matrix (in physical terms, the all the diagonal en tries equa/: in the f
mass ma~rix) whose entries are the inner products of the basis elements: governed by the stiffness matrix, thc
1\ in the basis elements are equal. This
D which is nearly optimal [V3]. The qt
lar mesh for finite elements of Hermi
and deriva ti ves appear among the unk1
In many cases we shall prefer, because the Ritz method operates always with is to keep the unknowns dimensional
the energy a(v, v) which is intrinsic to the problem, to work with the stiffness hv~ rather than v~ itself.
matrix: K1k = a(rp1 , rpJ. Both matrices are Hermitian and positive definite. Fried [F17] has observed that sorne
As a first measure of the independence of the basis, we propase the while others cannot. Take, for example
condition number · say -u" f with piecewise linear ei
K(M) = Amax(M). length h, except that the first is of len1
Amin(M) a natural condition u'(O) = O or an es,
tiónal respectively to
If the basis were orthonormal, M would be the identity matrix and K l.
This is not the case for finite elements, but the important fact on a regular -e
mesh is that thefinite element basisfunctions are uniformly linear/y independent: 1 +e
K(M) constant. In other words, the eigenvalues of M are all of the same
order. As Schultz observed, this makes piecewise polynomials infinitely more
stable for least-squares approximation (which is just the Ritz method applied
to the differential equation u= f of order zero) than the sequence 1, x, y,
-1 2
J
x 2 , ••• of ordinary polynomials. The condition number for this sequence, The largest eigenvalue grows with e,
whose mass matrix is the Hilbert matrix (1.5), would increase exponentially. order N- 2 , N order of the matrix.
There are applications in whieh a more realistic meas u re of independence as e oo.

is provided by the optimal condition number Scaling produces diagonal entries

the matrices unchanged:
c(M) = min K(DMD).
D 2
Here D may be any positive diagonal matrix, corresponding to a rescaling -2.../c/1 +e
of the basis elements; e = 1 if the original basis is only orthogonal rather (
than orthonormal. With irregular elements, sorne trial functions may be
much smaller than others, and this ·rescaling could make a significant differ-
ence to th.-e condition number. We view rescaling in the following way:
When the condition number of Mor of K is improved by rescaling, that suggests
DK,.,D =( -,J2~1 +e -A/
that the numerical difficulties which are thereby cured were never likely to
propagate throughout the whole problem. Rescaling cures local difficulties;
if a particular rp1 is badly out of scale, say the diagonal entry Kn is too small, The Iargest eigenvalues are now bou
then roundoff will destroy confidence in the computed value of the weight the size of Amia· In this case there is
Q1 • The effect which is measureddepends on the choice of norm; such a e has very little effect on Amia• which i
 CHAP. 5 SEC. 5.1. INDEPENDENCE OF THE BASIS 207

: "strong minimality" of the basis, and
to consider numerical methods from
,ility with respect to coefficient changes
tify the linear independence of the basís,
the Gram matrix (in physical terms, the
mer products of the basis elements:
e the Ritz method operates always with
the problem, to work with the stiffness
es are Hermitian and positive definite.
~ndence of the basis, we propose the
>Uld be the identity matrix and K l.
ts, but the important fact on a regular
-::tions are uniformly linear/y independent:
e eigenvalues of M are all of the same
:s piecewise polynomials infinitely more
1 ( which is just the Ritz method applied
order zero) than the sequence 1, x, y,
! condition number for this sequence,
rix (1.5), would increase exponentially.
no re realistic measure of independence
umber
~n K(DMD).
roundoff error would not mean a large mistake in the energy, because rp1
is small, but at the particular node z1 the approximation may be compara-
tively poor.
The standard rule in scaling a sparse positive definite matrix is to keep
al! the diagonal entries equal: in the finite element case, which is primarily
governed by the stiffness matrix, that means that the strain energies Kn
in the basis elements are equal. This rule yields a diagonal scaling matrix
D which is nearly optimal [V3]. The question of scaling arises even on a regu-
lar mesh for finite elements of Hermite type, in which both function values
and derivatives appear among the unknowns Qr Perhaps a natural procedure
is to keep the unknowns dimensionally correct, by using the rotation (}
hv~ rather than v~ itself.
Fried [F17] has observed that sorne problems can be successfully rescaled,
while othe,rs cannot. Take, for exarnple, a two-point boundary-value problem,
say -u" f with piecewise linear elements. If the subintervals are all of
length h, except that the first is of length h/c, then the stiffness matrices with
a natural condition u' (O) = O 9r an essential condition u(O) == O are propor-
tional respectively to
-e
l +e 1 +e -1 )
-1 2 : or K,.,= ( 1 ~ :
)
The Iargest eigenvalue grows with e, whereas the smallest is of the usual
order N- 2 , N order of the matrix. Therefore, the condition deteriorates
as e~ oo.
Scaling produces diagonal entries all equal to 2, and leaves the tail of
the matrices unchanged:

.l matrix, corresponding to a rescaling 2

-,.}2/1 +e )'
~iginal basis is only orthogonal rather 2 -1
ements, sorne trial functions may be -1
scaling could make ásignificant differ-
view rescaling in the;. following way:
K is improved by rescaling, that suggests DK,,D ( -,.)2~1 +e 2
we thereby cured were never likely to
'em. Rescaling cures local difficulties;
say the diagonal entry Kn is too small, The largest eigenvalues are now bounded; the diffiéult question is always
: in the computed value of the weight the size of Amín· In this case there is no problem with the second matrix;
pends on the choice of norm; such a e has very little effect on Amin' which is again of order N- 2 • The first matrix,
 208 STABILITY. CHAP. 5 SEC. 5.2.

' howev~r, begins with a block of order 2 which is nearly singular for large e, it may require double precision to pro
and the smallest eigenvalue for the whole matrix is necessarily below the It has been objected that the cond
eigenvalues for this block. Thus Amín O as e - We conclude that
(X),
with the load vector in a particular p
rescaling tends to be successful with an essential condition but not with a it must be, at least partly pessimistic.
naturai condition. tive numbers, which the computer e
This is the numerical anaÍogue of a physical situation which is well under- which automatically take into accoun
stood: A,r\stiff system connected to earth by a soft spring is extremely un- problem. We accept that for on-the·
stable, whereas a firm connection (essential condition) is stable. It should be calculation or changing to double pt
emphasized that while ill-conditioning from numerical sources is to be kind are the best. However, our goa
watched and degenerate elements are to be avoided, there will arise cases of of sensitivity, and for this purpose th(
physical ill-conditioning which cannot be altered-except perhaps by a change The rule to which it leads, that the ro u
from stiffness method to force method, with stresses as unknowns. This situa- with h-zm, is definitely obeyed in noJ
tion occurs when there is a sharp change in the stiffness of the medium, or the dependence is not on the total nut
when Poisson's ratio approaches the limit v = f of incompressibility [Fl8]. the number of elements per side. In oth
For shells there will be difficulties with large stiffness in the thickness direc- ence on the number of spatial. dimen:
tion, or with extremely thin shells. Roughly speaking, extensional modes can
involve the ratio (rfth) 2 [20], whereas bending modes bring out the fourth-
5.2. THE CONDITION NUMBER
order (plate-like) aspect of the problem and the roundoff is proportional to
h-4.
Our goal is to estímate the ratio 1<
To conclude this introduction, we must clarify the connection between and mínimum eigenvalues of the stifi
the condition number of a matrix and its sensitivity to perturbations. The
extreme case occurs when the given linear system KQ = F has a load vector THEOREM 5.1.
F which coincides with a unit eigenvector of K, in particular with the eigen-
For each variational problem and
vector Vmax corresponding to lmax· The solution is then Q = Vmax/lmax·
constan! e such that
Suppose this load vector is slightly perturbed by the eigenvector at the
other extreme, so that F = Vmax + EVmin· Then the solution becomes Q (1) ,K
Vmax/Amax + EVmin/Amin· The relative change in the solution is therefore
The constant depends inversely on the
tinuous problem, and it increases if
degenerate.

Thus a perturbation in F of arder E is amplified to a perturbation in Q of We give two proofs, both correc1
arder KE. only to the special case of a regular
lt is easy to see that this case is extreme, and that always Toeplitz) analysis permits a quite prec
condition number. The second applü
IO.QI K!O.FI. To begin, suppose that the mesh i
TQT IFI given problem Lu = f are constants.
Proof jFj jKQj < AmaxlQI, and jO.Fj = jKO.QI > Amin!O.Qj. also the mass matrices m1, will all be iC
K and M are essentially Toeplitz ma;
The consequences of this simple inequality, and a parallel one with are constant along each diagonal: K
1O. K lf! K 1 on the right si de [20], are very far-reaching. If the condition number between the column and row índice!
is K "'"' 1o-s, as many as s significan! digits may be lost during the solution of case to an integral operator f K(s t
KQ F. If this is clase to the number of digits carried by the computer, in which the kernel depends only on
 eHAP. 5
r 2 which is nearly singular for large e,
whole matrix is necessarily below the
in ~O as e~ co. We conclude that
an essential condition but not with a
1 physical situation which is well under-
arth by a soft spring,is extremely un-
ential condition) is sút:ble. It should be
ing from numerical sources is to be
to be avoided, there will arise cases of
be altered-except perhaps by a change
, with stresses as unknowns. This situa-
mge in the stiffness of the medium or
limit v = f of incompressibility [F,l8].
:h large stiffness in the thickness direc-
ughly speaking, extensional modes can
bending modes bring out the fourth-
m and the roundoff is proportional to
! must clarify the connection between
td its sensitivity to perturbations. The
near system K Q F has a load vector
ctor of K, in particular with the eigen-
The solution is then Q--:- Vmax/lmax-
perturbed by .the eigenvector at the
Vmin· Then the solution becomes Q =
hange iD the solution is therefore
Ké.
s amplified to a perturbation in Q of
reme, and that always
K ióFJ.
IFI
!óFI IK óQI > ArninlóQI.
inequality, and a parallel one with
far-reaching. Ifthe condition number
;gits may be lost during the solution of
er of digits carried by the computer,
SEC. 5.2. THE CONDITION NUMBER 209
it may require double precision to protect the accuracy ofthe computedresult.
It has been objected that the condition number might have nothing to do
with the load vector in a particular problem. The constant K may be, in fact
it must be, at least partly pessimistic. Irons [12] has proposed several alterna-
tive numbers, which the computer can form as elimination proceeds, and
which automatically take into account the scaling and the data/ ofthe given
problem. We accept that for on-the-spot decisions-terminating a specific
calculation or changing to double precision-computable quantities of this
kind are the best. However, our goal here is to find sorne a priori measure
of sensitivity, and for this purpose the condition number is very satisfactory.
The rule to which it leads, that the roundoff error will increase proportionately
with h- 2 m, is definitely obeyed in normal computations. We emphasize that
the dependence is not on the total number of elements in the domain; it is on
the number of elements per side. In other words, there is no significant depend-
ence on the number of spatial. dirriensions.
5.2. THE CONDITION NUMBER
Our goal is to estímate the ratio K lN(K)/l 1 (K) between the maximum
and mínimum eigenvalues of the stiffness matrix.
THEOREM 5.1.
For each variational prob!em and each choice of finite element there is a
constan! e such that
(1)
The constan! depends in verse/y on the smallest eigenvalue l 1 · of the given con-
tinuous problem, and it increases if the geometry of the elements becomes
degenerate.
We give two proofs, both correct but rather informal. The first applies
only to the special case of a regular mesh, and illustrates how Fourier (or
Toeplitz) analysis permits a quite precise computation of the eigenvalues and
condition number. The second applies to finite elements of arbitrary shape.
To begin, suppose that the mesh is regular and that the coefficients in the
givert problem Lu f are constants. The element stiffness matrices k 1, and
also the mass matrices m 1, will all be identical. This means that after assembly,
K and M are essentially Toeplitz matrices. The entries of a Toeplitz matrix
are constant along each diagonal: K¡¡ depends. only on the difference j i
between the column and row índices. (This corresponds in the continuous
caseto an integral operator J K(s- t)f(t) dt, in other words toa convolution,
in which the kernel depends only on s- t.) Although this property may be·
 210 STABILITY CHAP. 5 SEC. 5.2.

lost at boundaries, particularly with natural boundary· conditions, we want The same technique applies to the
to show how calculations with. Toeplitz matrices are still useful and easy. h/6 along each row. The eigenvalues i1
Suppose we compare two stiffhess matrices, both arising from -u" f
with piecewise linear elements. The first is of order N, and it is constrained J.l(O) = ~ e-:te + ~h +
by u(O) u'(n) O; it has been the basic example throughout the book.
The second is formed with no boundaries whatsoever, extending the interval Since cos (} ranges from 1 to 1, the (
[0, n] all the way to oo, oo) by adding on more and more element matrices: h. Notice the way in which the mass m
2 -1 the correct upper bound on the large
bound on Amin(M). The condition nur
-1
1 1 2 -1
K x:(M) x:(M«
h 2 -1
2 -1
Since this is independent of h, the piect
independent uniformly as h - ¡ . O. Tht:
Since K is formed by constraining K=-eliminating all elements outside
from the rigid-body motions still allow
[0, n] and imposing the essential condition Q0 0-the extreme eigenvalues
-u"= O on the whole 1ine, corres¡:
o.f K are enclosed by those of Koo:
(... 111 ...) which is eliminated from t
boundary cÓndition. Therefore, we shc
the mass matrices to achieve a rigorou
The same will be true for the mass matrices, and therefore linear independence First, to give a better understandin:
of the basis can be tested first on an infinite interval.
It is simple to work with the Toeplitz matrix K""', which can also be de- l. Bilinear elements on a square n
scribed as a discrete convolution matrix.Its eigenvectors are pure exponentials, known at every node, and a typical ro
just like the eigenfunctions of any constant-coefficient differential equation. on the main diagonal and eight l's o:
We try the vector whose components are v1 efi8 , and apply the matrix KQ Fis
do it:

. K oovj l ( -ie +' 2

=h-e where (j', k') represent the eight near
square grid. Again the eigenvectors v
Therefore,, the eigenvalue is there are two frequencies: the comp¡
the eigenvalues,

3hKoovjk = (8 é8 e-te-
er<e-'~'> e-i(B-!p
Since this_number ranges between O and 4/h, we conclude that
The extreme eigenvalues come from m
o sion in parentheses. Since it is linear
occur where these equal ± 1, and we f
In this case the result coincides with the conclusion of Gerschgorin's theorem
(1.4): Every eigenvalue lies in the circle with center given by the main diagonal
Ku 2/h and radius given by the absolute row sums Ei~i 1Kii 1 = 2/h. In
general, Gerschgorin will not be nearlyso good: Even for the bilinearelements
below, with off-diagonal entries all ofthe same sign, it is not precise. Gerschgorin's argument would give An
 CHAP.
1 natural boundary conditions, we want
:plitz matrices are· still useful and easy.
ss matrices, both arising from -u" = f
first is of order N, and it is constrained
1e basic example throughout the book.
iaries whatsoever, extending the interval
ing on more and m o~~ element matrices:
-1 2 -1
-1 2 -1
Koo-eliminating all elements outside
iition Q 0 = 0-the extreme eigenvalues
:rices, and therefore linear independence
infinite interval.
plitz matrix Koo, which can also be de-
x. Its eigenvectors are pure exponentials,
mstant-coe:fficient differential equation.
cts are vj = éj8 , and apply the matrix
_ eio) = 2(1 - cos {}).
h
nd 4/h, we conclude that
e conclusion of Gerschgorin's theorem
with center given by the main diagonal
1solute row sums I:j*i 1 Kij 1 = 2/h. In
so good: E ven for the bilinear elements
he same sign, it is not precise.
5 SEC. 5.2. THE CONDITION NUMBER 211
The same technique applies to the mass matrices, which have h/6, 4h/6,
h/6 alóng each row. The eigenvalues In the infinite case are
Since cos (} ranges from -1 to 1, the eigenvalues of M lie between h/3 and 00
h. Notice the way in which the mass matrices are better: We obtain not only
the correct upper bound on the largest eigenvalue, but also a good lower
bound on Amin(M). The condition number of M is given very accurately by
h
K(M) < K(Moo) = h/ 3 = 3.
Since this is independent of h, the piecewise linear roof functions are linearly
independent uniformly as h ------+ O. The zero lower bound for Amin(K) arose
from the rigid-body motions still allowed in Koo-a constant function satisfies
-u" = O on the whole line, corresponding to a discrete eigenvector of
(... 111 ... ) which is eliminated from the finite matrix K only by the essential
boundary condition. Therefore, we shall eventually have to argue by way of
the mass matrices to achieve a, rigorous estima te for K( K).
First, to give a better understanding of Koo, we study two more examples:
l. Bilinear elements on a square mesh, for -ll.u = f There is one un-
known at every node, anda typical row of K (multiplied by 3h) shows an 8
on the main diagonal and eight -1 'son t~e adjacent diagonals. The equation
KQ =:=Pis
where (j', k') represent the eight nearest neighbors of the point (j, k) on a
square grid. Again the eigenvectors v of Koo are pure exponentials, but now
there are two frequencies: the components are vjk = eiUB+krpl. Calculating
the eigenvalues,
3hKoovjk = (8- e¡o- e-io- é"'- e-irp- ei<B+rpl - e-i(B+rpl
- ei<o-"'> - e-i(B-"'>)vjk"
The extreme eigenvalues come from maximizing and minimizing the expres-
sion in parentheses. Since it is linear in cos (} and cos rp, the extrema must
occur where these equal ± 1, and we find
(at (} = re, rp = 0).
Gerschgorin's argument would give Amax < I6j3h.
 212 STABILITY CHAP.5 SEC. 5.2.

2. Cubic elements in pne dimension for u<iv) f There are now two bound is useless. At this point we ha·
unknowns at every node, and KQ F Ís a coupled system of two difference return to inequalities. According to J
equations. Correspondingly, K and Koo are block Toeplitz matrices: there is value of K is
a 2 x 2 block appearing o ver and o ver on the main diagonal, flanked on one
side by another such block and on the other side by its transpose. From (2) A. (K)= min xrKx
1 XTX
Section 1.7,
With the appearance of M :- 1 K, it bec<
The eigenproblem KQ A1lf Q is exc:
continuous problem Lu = lu, and we
eigenvalue A1 (M- 1K) in the discrete
fundamental eigenvalue A1(L) in the co¡
which is independent of h.
In the example Lu -u", with ·

A 1(24
h3 o
o)
8h 2 '
B 1(-12
h 3 -6h
6h)
2h 2 •
the lowest eigenvalue is A1(L) l Si
> A1(M"")> h/3, it follows· that 11
smallest eigenvalue of K is sin 2 (h/4)/(
In analogy with. the previous examples, the eigenvalues now come from the mate (2) misses by a fact~r of 3, es
blockl(O) Bre-te + +
A Be10 • This is itself a 2 x 2 matrix; its eigenvalues, because the true mínimum of xrKx;
maximized over the whole range -n O n, yield AmaxCK=). Again the whose components are all
of one si
extreme case must occur at O = O or O n, since the Rayleigh quotient Amax(M) than to the Amin(M) which a
xTA.(O)xfxTx is linear in cosO. We compute The essential point, however, is t
number is correctly determined:

l(O)
1
h3
(oo o)
12h 2 '
l(n) 1 (48
h3 o
o)
4h 2 •
K( K)
Therefore, Amax(K"") is the larger of 12/h and 48/h 3 • The result can be influ-
, enced by altering the relative scaling of the two types of basis functions, For an equation of order 2m, in' a re~
corresponding to displacement and slope. approach [combined with (2)} yields:
constant involves A. 1(L), as it should;
The extreme eigenvalues of any block Toeplitz matrix K"" can be found tioned, that must be reflected in its fi
in the same way. If the blocks are of order M (the number of unknowns
per mesh square), then 1(0) will be a matrix of this order; Orepresents O1 , • • • , Second proof ofTheorem 5.1.
On in an n-dimensional problem. Since each square is connected only to its The remaining problem is to exte
nearest neighbors in the nodal finite element method, the matrix A. will be number to the case of irregular elemc
linear in cos 0 0 . •• , cos On. Therefore, Amax(K,) can be computed by trying and the whole argument must be ba~
all possible ± 1 combinations for these cosines and evaluating the largest For this second proof we follow Fri
eigenvalues of the 2n resulting matrices l. The condition number for apure AmaxCK), we recall that the global ~
Toeplitz matrix can thus be computed exactly by working only with matrices element matrices k1 by
of order M (which is less than the arder of the element matrices and is given (3)
for common elements in the tables of Section 1.8). For the mass matrices
(also denoted by M!) we obtain K(M) K(M""), and this is a constant depend- The vectorx and matrix K are of ord
ing only on the element. Finite elements are uniformly independent. order d¡, ·the number of degrees of fJ
Por the stiffness.matrix there is a good upper bound on Amax' but the lower x 1 is formed from x just by striking o
 CHAP.
ion for u(iv) = f There are now two
!;' is a coupled systém of two difference
~" are block Toeplitz matrices: there is
~ on the main diagonal, flanked on one
he other side by its transpose. From
~, the eigenvalues now come from the
is itself a 2 x 2 matrix; its eigenvalues,
: < O < n, yield Amax(Koo). Again the
r (} = n, since the Rayleigh quotient
tpute
A(n) = 1 (48
h3 O
O)
4h 2
•
/ h and 48/h 3 • The result can be influ-
of the ,two types of basis functions,
,pe.
JCk Toeplitz matrix Koo can be found
order M (the number of unknowns
.trix of this order; (} represents (} 1 , • • • ,
: each square is connected only to its
~lement method, the matrix A will be
, AmaxCKoo) can be computed . by trying
se cosines and evaluating the largest
;¡ A. The condition number for a pure
~xactly by working only with matrices
:r of the element matrices and is given
· Section 1.8). For the mass matrices
: K(Moo), and this is a constant depend-
's are uniformly independent.
od upper bound on Amax' but the lower
5 SEC. 5.2. THE CONDITION NUMBER 213
bound is useless. At this point we have to abandon exact computations and
return to inequalities. According to Rayleigh's principie, the smallest eigen-
value of K is
TK TK TM
(2) A1(K)= min~ > min~min~ =A 1(M- 1K)A 1(M).
xT x - xT Mx xT x
With the appearance of M- 1K, it becomes easy to use variational arguments.
The eigenproblem KQ = AMQ is exactly the finite element analogue ofthe
continuous problem Lu = Au, and we show in Chapter 6 that thefundamental
eigehvalue A. 1 (M- 1 K) in the discrete problem always equals or exceeds the
fundamental eigenvalue A1 (L) in the continuous problem. This is a lower bound
which is independent of h.
In the example Lu = -u", with boundary conditions u(O) = u'(n) = O,
the lowest eigenvalue is A1 (L) =t. Since we have already proved that A1 (M)
> A1 (Moo) > h/3, it follows- that A1 (K) > h/12. The exact value for the
smallest eigenvalue of K is sin 2 (h/4)/(h/4), and therefore about h/4. The estí-
mate (2) misses by a factor of 3, essentially the condition number of M, ..
because the true mínimum of xT KxfxTx occurs at a fundamental mode x
whose components are all of one sign. This corresponds more closely to
Amax(M) than tothe Amin(M) which appears in (2).
The essential point, however, is that the exponent of h in the condition
number is correctly determined:
= AmaxCK) < 4/h = 4~.
K(K)
Amin(K) , h/12 h
For an equation of order 2m, in a regular domain, the same Toeplitz matrix
approach [combined with (2)] yields a condition number of order h- 2 m. The
constant involves A1(L), as it should; ifthe physical problem is poorly condi-
tioned, that must be reflected in its finite element analogue.
Second proof of Theorem 5.1 .
The remaining problem is to extend this h- 2 m estímate of the condition
number to the case of irregular elements. The Toeplitz structure falls apart,
and the whole argument must be based on the individual element matrices.
For this second proof we follow Fried [F14]. To find an upper bound for
AmaxCK), we recall that the global stiffness matrix is assembled from the
element matrices k¡ by
(3) xTKx = L; xfk¡X¡.
i
The vector x and matrix K are of order N; the vector X¡ and matrix ki are of
order d¡, the number of degrees of freedom within the ith element. In fact,
X¡ is formed from x just by striking out all except the appropriate d¡ compo-
214 STABILITY CHAP. 5 SEC. 5.2.

nents; we may think of inultiplying x by an incidence matrix, made up of Therefore, (K- 1 )jj is less than the tr
O's and 1's, to obtain x 1• pendently of the distribution of nodes.
If A is the maximum 't:;igenvalue of all the element stiffness matrices, so unknowns, and a second-order prol
that xfk¡x1 <A xTxr by Rayleigh's principie, we conclude from (3) that positive definite and thus (K- 1)f1 ~
oecur on the main diagonal. Therefor
· absolute values is bounded by eN; N
N- 1 to be the average mesh width ii.
Now suppose that q is the maximum number of elements which meet at any sum is bounded by cfhmin' since hi (
node. Then no component of x will appear in more than q ofthe abbreviated element matrix. Therefore, the condi
vectors x 1, and :E xTx1 qxTx. Therefore, of absolute row sums-correspondin
bounds-is less than Cfh hmin· With a
The important conclusion is thi
To find a lower bound on Amin(M), we proceed in exactly the same way: strongly on the degree of the polynm
h and on the order and fundamental
Therefore, the ·way to achieve num<
is to increase the degree of the tria
where e is the smallest eigenvalue of any of the element mass matrices and r cubics is only slightly worse than fo
is the mínimum number of elements which meet at a node. (Often r 1 errors for a given h are cómparable
or r 2 beca use of elements at the boundary; certainly r = 1 if there are an order oJ: magnitude smaller for c1
nodes interna! to the elements.) Rayleigh's principie again gives where roundoff prohibits any furthe1
in h, the cubic element is much more
computation of stresses, where diffe
ments introduces an extra factor h
second-order problems the roundof
Therefore, according to (2), the condition number of the stiffness matrix
in the degree of the trial functions is
is less than
(4)

If the geometry of the elements does not degenerate, so that the basis is
uniform in the sense of Chapter 3, a direct calculation yields the expected
K(K) = O(h- 2 m). If there is a degeneracy, so that triangles become very thin
or rectangles approach triangles, it is reflectedin the parameters .A ande.
Since these are eigenvalues of small matrices, the effects of this degeneracy
can be rigorously estimated. Fried [F14] has computed thís dependence on
the geometry for several examples-the estímate in Theorem 5.1 is sometimes
pessimistic-and gíves also a 1ower bound on the condition number.
He has also given a níce estímate for irreg~lar meshes in one dimension,
using the fact that a(uh, uh) a(u, u); the discrete structure is always stiffer
than the continuous one. Suppose we apply a point load at the jth node,
f ó(x- zj). Then it was observed in Section 1.1Ó that the load vector F
has only one nonzero component and that
 CHAP. 5 SEC. 5.2. THE CONDITION NUMBER 215

: x by an incidence mairix, made up of
of all the element stiffness mátrices so
'rinciple, we conclude from (J) that '
· absolute values is bounded by eN; N is the order ofthe mat.rix. We may take
number of elements which meet at any
.ppear in more than q of the abbreviated
refore,
and .A.mu.(K) Aq.
'), we proceed in exactly the same way:
any of the element mass matrices and r
s which meet at a node. (Óften r 1 errors for a given h are comparable. The discretization error, however, is
boundary; certainly r = 1 if there are
!Ígh's principie again gives
tdition number of the stiffness matrix
K) Aq
(M) A.JL)Or.
Therefore, (K- 1)n is less than the true energy a(u, u), and is bounded inde-
pendently of the distribution of nodes. [We assume function values to be the
unknowns, and a second-order problem -(pu')' + qu =f.] Since K- 1 is
positive definite and thus (K- 1)fj (K- 1)u(K- 1 )n, the largest entry must
occur on the main diagonal. Therefore, along each row of K- 1 , the sum ofthe
N- 1 to be the average mesh width ii. Along each row of K, the corresponding
sum is bounded by cjhmin• since h¡ appears in the denominator of the ith
element matrix. Therefore, the condition number, even in the stronger sense
of absolute row sums-corresponding to pointwise rather than mean-square
bounds-is less than Cfh hmin· With a regular mesh we recover C/h 2 •
The important conclusion is this: The roundo.ff error does not depend
strongly on the degree of the polynomial element. It depends principally on
h and on the order and fundamental eigenvalue of the continuous problem.
Therefore, the way to achieve numerical accuracy in the face of roundoff
is to increase the degree of the trial functions. The condition number for
cubics is only slightly worse than for linear elements, so that the roundoff
an order of magnitude smaller for cubics. Therefore, at the crossover point
where roundoff prohibits any further improvement cohling from a decrease
in h, the cubic element is much more accurate. This applies especially to the
computation of stresses, where differentiation (or differencing) of displace-
ments introduces an extra factor h- 1 into the numerical error. Even in
second-order problems the roundoff becomes significant, and an increase
in the degree of the trial functions is th~ best way out.

·s not degenerate, so that the basis is

. direct calculation yields the expected
acy, so that triangles become very thin
; reflected in the parameters A and (}.
natrices, the effects of this degeneracy
'14] has computed this dependence on
e estimate in Theorem 5.1 is sometimes
mnd on the conditioñ number.
br irreg'flar meshes in one dimension,
the discrete structure is always stiffer
~ apply a point load at the jth node,
.n Section 1.1 ó that the load vector F
that
 SEC. 6.1. VARIATIONAL FORMUI

element subspace Sh. Therefore, the

can be put directly to use. Mathem~
is to deduce, starting from these appn
for the error in the eigenvalues and t
our main goal.
We shall precede the general theo:
in order to illustrate the behavior o
we link the approximation theory
find bounds on A1 - A.~ and u1 - u?, tl

6 EIGENVALUE PROBLEMS
function, which are sharp enough to
as on h.
We begin with the eigenvalue prc

(1) Lu _!!_(p(x)du) +,
d~ dx

This is the same operator L which wa

city lies in the fact that it is only on(
6.1. VARIATIONAL FORMULATION ANO
THE MINMAX PRINCIPLE
the. distinction between natural and <
ducing one of each:
Eigenvalue problems-which we shall write as Lu A.u, or more gener-
(2} u(O) =O,
ally as Lu = A.Bu-arise in a tremendous variety of applications. We mention
the buckling of columns and shells, the vibration of elastic bodies, and multi-
It ís known that such a Sturm-Liot
group diffusion in nuclear reactors. Fortunately the Rayleigh-Ritz idea is
as useful for these problems as for steady-state equations Lu = f In fact, of real eigenvalues
the idea had its beginning in Rayleigh's description of the fundamental
frequency as the mínimum value of the Rayleigh quotient. Therefore, the
-step which has been taken in the last 15 years was completely natural and
and an associated complete set of on
inevitable: to apply the new finite element ideas to this long-established

f"'
variational form of the eigenvalue problem.
From a practica} point of view, this means that piecewise polynomial (3) o
U.'
J
functions can be substituted directly as trial functions into the Rayleigh
quotient. The evaluation ofthis quotient becomes exactly the problem which It follows immediately that these eig
has already been discussed and which large-scale computer systems have energy inner product
been developed to carry out: the assembly of the stiffness and mass matrices
K and M: The next step, however, leads to a different and more difficult (4)
computational problem in linear algebra: instead of a linear system KQ = F,
there arises a ·discrete eigenvalue problem, KQ A.MQ. Fortunately, it is or, with the usual integration by par
now known how the properties of the two matrices-symmetry, sparseness,
positive definiteness of M-can be used to speed up the numerical algorithm.
We shall discuss several effective numerical processes in Section 6.4.
From a theoretical point of view, the basic steps in establishing error
bounds depend once again on the approximation properties of the finite In the case of constant p and q, wh

216
 SEC. 6.1. VARIÁTIONAL FORMULATION AND THE MINMAX PRINCIPLE 217

element subspace Sh. Therefore, the approximation theorems of Chapter 3

can be put directly to use. Mathematically, the new step which is required
is to deduce, starting from these approximation theorems, satisfactory bounds
for the error in the eigenvalues and the eigenfunctions. This step is therefore
o ur main goal.
We shall precede the general theory by a study of sorne specific examples,
in order to illustrate the behavior of finite element approximations. Then
we link the approximation theory to the eigenvalue problem, aiming to
find bounds on A1 A? and u1 - uf, the errors in the lth eigenvalue and eigen-
function, which are sharp enough to give the right dependence on las well
as on h.
We begin with the eigenvalue problém

(1) ~(p(x)~~) + q(x)u == Au, O< x <n.

This is the same operator L which was studied in the first chapter; its simpli-
~NO city líes in the fact that it is only one-dimensional. We- shall again illustrate
the. distinction between natural and essential boundary conditions by intro-
ducing one of each:
1all write as Lu .A.u, or more gener-
us variety of applications. We mention (2) u(O) O, u'(n) =O.
vibration of elastic bodies, and multi-
'ortunately the Rayleigh-Ritz idea is It is known that such a Sturm-Liouville problem has an infinite sequence
teady-state equations Lu f. In fact, of real eigenvalues
gh's description of the fundamental
the Rayleigh quotient. Therefore the A<
2-
... ... -+oo,

I 5 years was c;ompletely natural' and

lement ideas to this ·long-established and an associated complete set of orthonormal eigenfunctions:
blem.
1is means that piecewise polynomial (3)
as trial functions into the Rayleigh
1t becomes exactly the problem which It follows immediately that these eigenfunctions are orthogonal also in the
1 Jarge-scale computer systems have energy inner product
bly of the stiffness and mass matrices
:1ds to a different ami more difficult (4)
:t: instead of a linear system KQ F,
1
lem, KQ .A.MQ. Fortunately, it is or, with the usual integration by parts,
two matrices-symmetry, sparseness,
to speed up the numerical algorithm. j=k,
rical processes in Section 6.4. (5)
j=Fk.
the basic steps in establishing error
proximation properties of the finite In the case of constant p and q; which will serve as a model problem, the
 218 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.1. VARIATIONAL FORMU

eigenfunctions are sinusoidal: cross section of the egg cup which .

(v, v) = l. The lowest point on thi
ulx) = !{- sin(j !)x, · ;ti p(j- !)2 + q. fundamental frequency A- 1 •
Alternatively, we can fix the nur
The first step in the Rayleigh-Ritz method is to rewrite Lu = lu as a the egg cup is cut by a horizontal pl
variatio11al problem. There are two possibilities, corresponding to the Ritz cross section is the ellipse a(v, v) = :
minimizátion and the Galerkin weak form of steady-state equations Lu = f 1 in infinitely many dimensions.
Both lead ·to the same result. The first is to introduce the Rayleigh quotient, first eigenfunction u1 , since it is in ti
which is defined by from the axis. In other words, with tl
leigh quotient is minimized when thc
a(v, v) sider the cross section of the ellipst
- (6) R(v)
(v, v) fix fX 1 = O and look in a space of 01
of this ellipse has length ,.;rrr;_.
[AIJ
ellipsoid will have. a major axis who:
We claim that the stationary (or critical) points of this functional R(v), to the minmax principie described
which are the points where the gradient of R vanishes, are exactly the eigen- 1'1 under any one given constraint :
functions of the problem. To understand this, Jet the trial function v be It appears that in four dimensioJ
replaced by its eigenfunction expansion 2:; fXiui, fXi = (v, ui): shell of a whole egg. ·
We have not yet specified whicl
(7) R(v) = a(2; fXiui, 2:; akuk) = 2:; AifXJ,
(2:; fXiui, 2:; fXkuk) 2:; fXJ variational, or Rayleigh quotient, el
functions. This is exactly the same ~
using the orthogonality conditions (3) and (5). Ata stationary point of R(v), problem in Section 1.3. In that case
the derivative with respect to each fXk must be. zero: tions in 3<! 1 satisfying the essential
This space arose naturally through a
o. functions that satisfied alt bounda1
which was a limit of such vN in the
we shall do exactly the same thing,
If v is one of the eigenfunctions ui' this is certainly satisfied: Every ock = O
changed; all stationary points rema
witñ the exception of the term with k = j, and this remainíng survivor is
p(x) were discontinuous, this proces
cancelled by the vanishing of lj lk. There are no stationary points other
R(v) actually fills in the eigenfuncti
than the eigenfunctions. (For a repeated eigenvalue ;tj = lj+P all combina-
were initially excluded.J
tions v +
fXiui ocj+Juj+J are eigenfunctions, and there is a whole plane of
The resulting admissible space is
stationary points.) The value of R(v) at a stationary point v = u 1 is easy
that all the same finite element subsi
to determine: it is exactly the eigenvalue Ar
mation, the difference being that wt
rather than the mínimum of I(v).
(8) 1t was mentioned earlier that tl
eigenvalue problem. This is to put t
lt is useful to try to visualize the graph of R(v) .. The nume~ator a(v, v) or Galerkin form, which is produc
corresponds toa convex surface, an '"egg cup," exactly as for I(v) in the first Lu lu by a function v and integr~
chapter; the only difference is that the linear term -2(/, v) is now absent,
so that the bottom of the egg cup lies at the origin. There are two ways to
consíder the effect of the denomínator (v, v). Since it makes the quotient
(9) J: (pu'v' + qu
homogeneous, R(fXv) R(v), one possibility is to consider only unit vectors · The eigenvalue problem is then to
v. In other words, we can fix the denominator to equal 1, and loo k at the such that (9) holds for all v in 3<!1. '
 CHAP.6
itz method is to rewrite Lu Au as a
possibilities, corresp<;mding to the Ritz
form of steady-statel~quations Lu f.
;t is to introduce the Rayleigh quotient,
itical) ·points of this functional R(v),
nt of R vanishes, are exactly the eigen-
stand this, Jet the trial function v be
mI; rx¡uj, rxi (v; u¡):
and (5). At a stationary point of R(v),
must be zero:
o.
1is is certainly satisfied: Every ak O we shall do exactly the same thing, since it still leaves the surface R(v) un-
k j, and this remaining survivor is
. There are no stationary points other
ed eigenvalue A.i A¡+t> all combina-
lctions, and there is a whole plane of
at a stationary point v u¡ is easy
ue A.i.
'¡,u)= A. .•
'u¡) ' It was mentioned earlier that there is a second way to reformulate the
~raph of R(v). The numetator a(v, v)
gg cup," exactly as for l(v) in the first
~ linear term - 2(/, v) is now absent,
at the origin. There are two ways to
•r (v, v). Since it makes the quotient
ibility is to consider only unit vectors
)minator to equal 1, and loo k at the
SEC. 6.1. VARIATIONAL FORMULATION ANO THE MINMAX PRINCIPLE 219
cross section of the egg cup which is cut out by the right circular cylinder
(v, v) l. The lowest point on thi& cross section will correspond to the
fundamental frequency A1 •
Alternatively, we can fix the numerator at the value l. This means that
the egg cup is cut by a horizontal plane, lying one unit above the base. The
cross section is the ellipse a(v, v) = 1, or more precisely the ellipsoid L; A¡IXJ
= 1 in infinitely many dimensions. Its major axis is in the direction of the
first eigenfunction u 1 , sin ce it is in this direction that the ellipsoid is farthest
from the axis. In other words, with the numerator of R(v) fixed at 1, the Ray-
leigh quotient is minimized when the denominator is largest. If we then con-
sider the cross section of the ellipse normal to this major axis, that is, we
fix tX 1 O and look in a space of one lower dimension, then the major axis
of this ellipse has length ~. [And any other cross section of the original
ellipsoid will ha ve a major axis whose length falls in between, corresponding
to the minmax principie described in (13) below: The smallest eigenvalue
A.'1 under any one given constraint satisfies A1 < .A.'1 A2 .]
It appears that in four dimensions, the cross section of an egg cup is the
shell of a whole egg.
We have not yet specified which functions V are to be admitted in the
variational, or Rayleigh quotient, characterization of eigenvalues and eigen-
functions. This is exactly the same question which arose for the steady-state
problem in Section 1.3. In that case I(v) was finally minimized over all func-
tions in X 1 satisfying the essential boundary conditions, that is, o ver X 1.
This space aro se naturally through a completion process, startingwith smooth
functions that satisfied al! boundary conditions and then including any v
which was a limit of such vN in the sense that a(v vN, v- vN)---+ O. Here
changed; all stationary points remain so in the completion process. [And if
p(x) were discontinuous, this process of filling in the holes on the surface of
R(v) actually fills in the eigenfunctions; because they are not smooth, they
were initially excluded.]
The resulting admissible space is then X1, precisely as before. This means
that all the same finite element subspaces may be used in the discrete approxi-
mation, the difference being that we now look for stationary points of R(v)
rather than the mínimum of /(v).
eigenvalue problem. This is to put the equation Lu Au into its weak form,
or Galerkin form, which is produced in the following way: We multiply
Lu AU by a function v and integrate by parts to obtain
(9) J: (pu'v' + quv) dx A J: uv dx.
The eigenvalue problem is then' to find a scalar A and a function u in X 1
such that (9) holds for all v in X1. The boundary conditions on u come out
220 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6J. VARTATIONAL FORMUL

correctly, since equation (9) is actually equivalent to There is an a1ternative formula f<
ing u~' ... , u1_ 1 • It will be of fundan
J Lu(x )v(x) dx pu'vln: A Juvdx. was discovered by Poincaré, Courant
"' o Minmax Principie: /f R(v) is ma
At x = O, the integrated term pu'v automatically vanishes. Equality of the S 1, then the minimum possible value fm
remaining t~rms, for all v in 3C1, -forces both the natural condition u' O
to hold at x.=
n and the differential equation Lu Au to hold over the (13) A1 mir
s,
interval.
The equation (9) is a specific case ofthe standard weak form for the eigen- The mínimum is taken over alll-dimel
value problem: Find a scalar A, anda function u in the admissible space V, subspace S1 = Eh the maximum of R
such that To prove the minmax formula (1
choice of S 1,
(10) a(u, v) = A(u, v) for all v in V.
(14) maxR
This resembles the ·condition a(u, v) (J, v) for the vanishing of the first v in St

variation in the steady-state problem ofminimizíng /(v). Indeed, the equation

represents exactly the vanishing of the first variation of R at the stationary The argument depends on selecting
pomtu: - that is, so as to satisfy the l - 1 f
exists such a v*, since we are imposi:
R(u -+· Ev) = a(u + EV, u+ Ev) = a(u, u)+ 2Ea(u, v) + ...
on an l-parameter space. Now it fo
(u + EV, u + Ev) (u, u) + 2E(u, v) + · · ·
A¡< R(v*). In other words, (14) ht
R(u) + lE a(u, vXu, u) a( u, u)(u, v} + ... pro ved.
(u, u) 2 One useful consequence of this fo
R(u) + lE a( u, v) A( u, v) + O(Ez). value A1, based on a comparison v
(u, u) clear that for any v,
Thus the weak formulation and the stationary point formulation are equiva-
lent.
In the case of A1 and u 1-thai is, for the fundamental frequency and
its associated normal mode-the stationary point is actually a mínimum: Dividing by f v 2 , this beco mes a c<
the variable-coefficient case sand"'
(11) 1 1 = "!in, R(v). eigenvalues are known explicitly. By
v m·:JC 8
A1 of the middle problem must lie bt
This is obvious if vis expanded as 2: rxiui' since then R(v) = 2: Airx7fí: a; >
A¡.
It is valuable to describe also the higher eigenfunctions in terms of a
minimization, since convergence to a mínimum is so much simpler to analyze In particular, A1 is of order J2 as l
than convergence to a stationary point. One possibility in this direction to Finally, we cometo the maino~
constrain v to be orthogonal to the first 1- 1 eigenfunctions: a 1 = (v, u 1 ) the Rayleigh-Ritz principie for appro
O, ... , rx 1_ 1 (v, u1_ 1 ) O. Under these constraints the mínimum of the with the weak forma(u, v) A(u, v)'
Rayleigh quotieht beco mes A1 : as critical (stationary) points of R(v~
is to work only within a finite-dimen
(12) A1 = min R(v).
v_I_Et-1 space 3C1. In this subspace, we look
Here E 1_ 1 is the space spanned by the eigenfunctions up ... , u 1_ 1 • (16) a(tl', if) = A"(tf, 1
 CHAP.
equivalent to
~' v /: = J
A uv dx.
tomatically vanishes. Equality ·of the
~s both the natural q,pndition u' = O
equation Lu = Au tb' hold over the
the standard weak form for the eigen-
function u in the admissible space V,
for all v in V.
(f, v) for the vanishing of the first
minimizing I(v). Indeed, the equation
first variation of R at the stationary
_ a(u, u)+ 2ea(u, v) + · ..
- (u, u) + 2e(u, v) + ...
1, u) - a(u, u)(u, v) + pro ved.
(u, u) 2
. u;(u, v) + O(é).
ionary point formulation are equiva-'
for the fundamental frequency and
mry point is actually a mínimum:
~. R(v).
"E
tigher eigenfunctions in terms of a
1imum is so much simpler to analyze
One possibility in this direction to
1 eigenfunctions: a 1 = (v, u 1) =
;e constraints the mínimum of the
R(v).
igenfunctions u 1 , ••• , u1_ 1 •
6
SEC. 6.1. VARIATIONAL FORMULATION AND THE MINMAX PRINCIPLE 221
There is an alternative formula for A1 which does not depend on know..
ing u., ... , u1_ 1 • lt will be of fundamental importance in what follows, and
was discovered by Poincaré, Courant, and Fischer; we refer to itas the
Minmax Principie: Jf R(v) is maximized over an !-dimensional subspace
S1, then the minimum possible value for this maximum is A1 :
(13) A1 = min maxR(v).
St v in S¡
The mínimum is taken o ver all/-dimensional'subspaces of 3C1. For the special
subspace S 1 = E 1, the maximum of R(v) is exactly A1•
To prove the minmax formula (13), we have to show that for any other
choice of S1,
(14) max R(v) > A1•
v inSt
The argument depends on selecting v* in S1 so as to be orthogonal to E1_ 1 ,
that is, so as to satisfy the 1- 1 equations (v, uJ =O, 1 i <l. There
exists such a v*, since we are imposing only 1 1 homogeneous constraints
on an /-parameter space. Now it follows from (12), since v* j_ E 1_p that
A1 R(v*). In other words, (14) holds, and the minmax formula (13) is
One useful consequence of this formula is a rough estímate for the eigen-
value A1, based on a comparison with the constant-coefficient case. It is
clear that for any v,
Dividing by J v 2 , this beco mes a comparison of Rayleigh quotients, with
the variable-coefficient case sandwiched between two problems whose
eigenvalues are known explicitly. By the minmax principie, each eigenvalue
A1 of the middle problem must lie between the two known eigenvalues,
In particular, A1 is of order f2 as 1 oo.
Finally, we come to the main object of this section, which is to establish
the Rayleigh-Ritz principie for approximating the eigenvalues. It begins either
with the weak forma( u, v) = A(u, v) or with the description ofthe eigenvalues
as critica! (stationary) points of R(v) = a(v, v)f(v, v). In either case, the idea
is to work only within a finite-dimensional subspace Sh of tlÍe full admissible
space 3C1. In this subspace, welook for a pair Ah and uh such that
(16)
222 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.2.

In other words, If the basis functions rpi are orthc

identií:y and the discrete problem is ·
this orthogonality condition on the rp
property of finite elements, that rp¡ s·
Alternatively, the approximate eigenvectors are the critica/ points of R(vh) taining the corresponding node z¡. 1
m•er the spape Sh. 1 or do violence to Rayleigh's idea l:
• To see t'hat these methods Iead to the same approximations, choose a former, since numerical algorithms
basis qJp ... ~ rpN for Sh. Then any vh in S" can be expanded as are now appearing which are compar
We emphasize that in all cases the
(17) definite; it is the Gram matrix for t
(/JN•
where q¡ are the generalized coordinates (the nodal parameters of vh, if Sh The fundamental frequency A1 is o
is a finite element space). Substituting into the Rayleigh quotient, and we especially hope that A1 prov.
.:tt always lies above A1 , l1 A¡, sine
the subspace Sh and l 1 is the minimt
(18)
It is natural to expect that if the tru
mated by the ~ubspace Sh, then A1 is
key result of the theory.
The integrals in the numerator and denominator are by now familiar: They The minmax principie applies eq
are the entries in the stiffness matrix Kh and the mass matrix Mh. Thus the the same proof, so that the approx
Rayleigh quotient, in terms of the vector q = (q 1 , • • • , qN), is exactly by

(19) (22) Af = mir

St

The critica) points of this discrete quotient are the solutions to the matrix Here S 1 ranges over al/1-dimensional
eigenvalue problein the definition makes sense only for
N, there exist only N approximatc:
(20)
principies (22) and (13), it follows i
Therefore, this is the eigenproblem which has to be assembled and solved. proximated from above:
The eigenvalues Af are expected to approximate the continuous eigenvalues
A~> at least for small values of 1, and the eigenvectors Qf lead toa correspond-
(23)
ing approximate eigenfunction
Every space S1 which is allowed in t
(13), and therefore the mínimum A1 i
(21)
The minmax principie extends ve
B is positive definite; the Rayleigh
Thus the components of the discrete eigenvectors, in the matrix problem The mass matrix in the discrete prob
KQ = AMQ, yield the nodal values of the finite element eigenfunctions.
The weak form' of the efgenproblem Ieads directly to the same result.
Suppose that in,equation (16) we take vh rpk: 6.2. SOME ELEMENTARY EXAMPLI

In this section we shall treat son

This is simply the kth row ofthe matrix equation KhQh AhMhQh. the patterns which occur in the gene1
 CHAP. 6
·ctors are the critica! points of R(vh)
the same approximi~:tions, choose a
1 S'' can be ex panded as
es (the nodal· parameters of vh, if Sh
nto the Rayleigh quotient,
J(p(-x)rp~rp~ + q(x)rpjrpk) dx
:¿; qjq k Jrpjrpk dx
wminator are by now familiar: They
h and the mass matrix Mh. Thus the
1r q = (q~' ... , qN), is exactly
fKhq
rTMhq.
tient are the solutions to ~he matrix
ich has to be assembled and solved.
roximate the continuous eigenvalues
:!igenvectors Q? lead to a correspond-
!igenvectors, in the matrix problem
f the finite element eigenfunctions.
n leads directly to the same result.
h = rpk:
SEC. 6.2. SOME ELEMENTARY EXAMPLES 223
If the b-asis functions rp j are orthonormal, then the mass matrix Mh is the
identity and the discrete problem is to find the eigenvalues of K''. However,
this orthogonality condition on the rp j is incompatible with a more important
property of flnite elements, that rpj should vanish over all elements not con-
taining the corresponding node Zr Therefore, we must either accept Mh *
1 or do. violence to Rayleigh's idea by "lumping" the masses. We prefer the
former, since numerical algorithms for the general problem KQ = A.MQ
are now appearing which are comparable in efficiency to those for KQ = A.Q.
We emphasize that in all cases the mass matrix is symmetric and positive
definite; it is the Gram matrix for the linearly independent vectors rpp •.• ,
fPN•
The fundamental frequency A. 1 is often the q1,1antity of greatest significance,
and we especially hope that A.~ provides a good approximation. Notice that
A.~ always lies above A. 1 , A.~ >- A. 1 , since A.~ is the mínimum value of R(v) over
the subspace Sh and A. 1 is the mínimum over the whole admissible space 3C1.
It is natural to expect that if the true eigenfunction u 1 can be well approxi-
mated by the subspace S\ then A.~ is automatically close to A. 1 ; this will be a
key result of the theory.
The minmax principie applies equally well to the discrete problem, with
the same proof, so that the approximate eigenvalues can be characterized
by
(22) A.? = min max R(vh),
s, vh in St
Here S1 ranges over alll-dimensional subspace of the subspace Sh. Of course,
thé· definition makes sense only for l < N, since _if the dimension of Sh is
N, there exist only N approximate eigenvalues. ·Comparing the. minmax
principies (22) and (13), it follows immediately that every eigenvalue is ap-
proximated from abo ve:
(23) for all/.
Every space S1 which is allowed in the minimization (22) is also allowed in
(13), and therefore the mínimum A.1 in (13) is at least as small as A,7.
The minmax principie extends verbatim to the case Lu = A.Bu as long as
B is positive definite; the Rayleigh quotient is then R(v) = (Lv, v)f(Bv, v).
The mass matrix in the discrete problem becomes Mjk = (Brpj, rpk).
6.2. SOME ELEMENTARY EXAMPLES
In this section we shall treat sorne specific examples in order to isolate
the patterns which occur in thé general theory of eigenvalue approximation.
224 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.2.

We shap concentrate on the constant-coefficient problem Then

(25)
0 <X< 1l,
In this special problem the eigenfu11
with the boundary conditions u(O) O, u'(n) = O.
The fi~t trial space to consider is made.up ofpiecewise linear functions,
with equally spaced nodes xi = jh. Using th~ standard roof functions rpi (26) ~uf(jh) u/.jh)
as a basis, the key matrices are
It follows that uf is equal to the i
4 2 -1 theorems give
1 4 -1 2 -1
(27)
Mh K~=!
4 1 -1 2 -1 Although the exactness of the Ritz
the estimate (27) is typical of the !
1 2 -1
functions u1 - uf are of the same ord
The stiffness matrix can be written as Kh pK~ + qMh. The optimal weights to steady-súzte problems Lu -=f.
for the Rayleigh-Ritz eigenfunction Turning now to the approxima1

kh(l) (l !)2
h2 (1
are the values of uh at the nodes, and, in fact, the system KhQh = J..hMhQh 6
is precisely the difference equation
Recalling that 1 1 p(l - !) + q,
2

~ (2Q~
(24)

Observe that the eigenvalues art

For this to be valid at the boundary point j = N we put QJ¡.+ 1 Q'N~ 1 • functions. This is true in general:
Recall that the Dirichlet condition u(O) O gives Q0 O.
The eigensystem (24) has a Toeplitz structure away from the boundaries,
in the sense that the (i, j) en tries of the stiffness and mass matrices depend for 2mth-order problems. The rea:
only on the differeríce i - j. 1t is therefore reasonable to expect trigonometric near a critica! point. Moderately ac
eigenvector..s, and indeed the components of the /th eigenvector are eigenvalues.
The eigénvalues are actually ir
(Qf)j jf sin((/ f)jh). addition of the term qu to the op
a 'constant amount: l 1 = p(l -f
To write out an expression for the associated eigenvalue A.?, put in general. The crucial point is that i
approximation J..?. It will be conven
kil) 2h~ 2 (1 - cos(/- !)h),
lf - A¡, by adding a sufficiently la
this chapter that 1 1 > O.
 CHAP. 6 SEC. 6.2. SOME ELEMENTARY EXAMPLES 225

-coefficient problem Then

.u, O< x < n, (25)

J, u'(n) =O. In this special problem the eigenfunctions are infinitely accurate at the nodes,
nade. up of piecewis~;;.,linear functions,
Jsing th~ standard rbof functions rpj (26) u?(jh) = u¡(jh) = Jf sin((/- f)jh).

It follows that u? is equal to the interpolate of u1, and the approximation

2 -1
theorems give
-1 "2 -1
(27)

-1 2 -1 Although thé exactness of the Ritz eigenfunction u? at the nodes is special,

-1 the estimate (27) is typical of the general case. The errors in the Ritz eigen-
functions u1 - u? are of the same order as the errors in the approximate solution
~h = pK1 + qMh. The optimal weights to steady-state problems Lu = f
Turning now to the approximate eigenvalue M given by (25), we expand

;1 Q1.rp.(x)
1
kh(l) = ( / - f); - ~; (1- f) 4 + 0(/6 h4),
mh(l) = 1 - ~2 (1 - !)2 + 0(14h4).
Recalling that A1 = p(l - !F + q, the .eigenvalue error is
(28)

point j = N we put QhN +1 = QhN-!·
= O gives Q0 = O.
structure away from the boundaries
~ stiffness and mass matrices depend
re reasonable to expect trigonometric
ts of the lth eigenvector are
n((/ - f)jh).
iated eigenvalue A?, put
mh(l) =2 + cos(l- f)h.
3 this chapter that A1 > O.
Observe that the eigenvalues are as accurate as the energies in the eigen-
functions. This is true in general:
A?- A¡< Cllu?- u11l~
for 2mth-order problems. The reason is that the Rayleigh quotient is flat
near a critica! point. Moderately accurate tria] functions yield very accurate
eigenvalues.
The eigenvalues are actually independent of the con.stant q, since the
ad,dition of the term qu to the operator simply translates the spectruin by
a constant amount: A1 = p(l - !F when q = O and A1 = p(l - f) 2 + q
in general. The crucial point is that it has the same effect on the Ritz-Galerkin
approximation A?. It will be convenient to exploit this invariance ofthe error
A? - A1, by adding a sufficiently large constant term to ensure throughout
226 EIOENVAÜJE PROBLEMS CHAP. 6 SEC. 6.2.

In a lumping process the mass matrix Mh is replaced by a diagonal matrix, valid at the boundaries if we put
in this. simple case by the identity l. In our example it happe_ns that the
eigenfunctions remain. unchanged. The eigenvalues, on the other hand, are U0 = U_ 1 +U 1 U'1 u'_ 1 ÚN::::::

altered to
[Observe that (30) reduces to O = O ,
atj N.]
As in the piecewise linear case, tb
simple problem happen to agree at the 1
a
Thus 1; is lower bound to 1 1 and has the same O(h 2 ) accuracy as l?.
Lumping has a very seductive physical interpretation in terms of the s tiff-
ness ofthe system KhQh = lhMhQh. From this point ofview, the replacement (32) ui = ,jf sin((! -1)jh),
of the mass matrix Mh with the identity matrix 1 has the effect of making the
structure "softer" and hence reducing the magnitude of the approximate Indeed, with v1 = 1 - !, the substitut
eigenvalues. Since the Rayleigh-Ritz. approximations are necessarily upper 2 X 2 eigenvalue problem
bounds, l? l, the hope is that this reduction in magnitude will increase ·
the accuracy of the approximation. Our concern, on the other hand, is that . COS V 1h)
sin v ,
--5-1
this violation of the Ritz rules may make the structure too soft and thereby
adversely affect the accuracy. In the present case the damage has not been sin v1h h
great, but a more typical example at the end of this section shows that a far -5- T5 (4 cos
(33)
more serious loss of accuracy is possible.
We turn toan example of a higher-degree piecewise polynomial approxi- ;o (52.+ 18 cos v,J
mation, namely the cubic Hermite space Sh on a uniform mesh. There are
basis functions VI i corresponding to function values, and roi to slopes, so
l?
[ 13 h2 sm
210
. vh
1

thatt
N N-1 Observe that (33) gives two eigenvah
vh(x) = 2:1 ui'fllx) + 2:
j= j=O
u~rolx). l <N. Since the system (30)-(31) is o
all the finite element eigenvalues. By a
The matrix eigenvalue problem for the optimal weights, with p = 1, q O,
beco mes (34)

The corresponding eigenfunctions are

a ti ve.
(30)
The second eigenvalues lt 1 are of
to any eigenvalue of the differential equ
a serious disadvantage of using cubic .fl
tions, that at least half of the eigem•alu
(31) as approximations. However, this phen
ment approximations. A closer look at
tions shows that for each h there is a1
Strictly speaking (30) and (31) hold only for 1 <j <N- 1, but they remain eigenvalues are reasonable; moreover,)
implications for the choice of a metho(
tFor simplicíty we shall require that all functions vh in Sh satisfy the natural boundary of K Q lM Q; only the leading eigenv¡
condition at n as well as the essential boundary condition at O. We clase this section by noting that
 CHAP. 6 SEC. 6.2. SOME ELEMENTARY EXAMPLES 227

·ix Mh is replaced by a diagonal matrix, valid at the boundaries if we put

In our example ·it happe.ns that the '
1e eigenvalues, on the other hand, are o.
(Observe that (30) reduces to O= O atj O, and (31) reduces to O= O
atj =N.]
As in the piecewise linear case, the approximate eigenfunctions for this
simple problem happen to agree at the nodes with trigonometric polynomials:
the same O(h 2 ) accuracy as A.?.
:ical interpretation in terms of the stiff-
~om this point ofview, the replacement (32) ui /{sin((/ !)jh), U~=/{(/- !)a¡ GOS((/- i)jh).
ty matrix 1 has the effect of making the
.g the magnitude of the approximate Indeed, with v1 l - !, the substitution of (32) into (30)-(31) produces a
approximations are necessarily upper 2 X 2 eigenvalue problem
; reduction in magnitude will increase

[
ur concern, on the other hand, is that
ake the structure too soft and thereby
~~ (1 COS V 1h)

Jresent case the damage has not been sin v1h

he end of this section shows that a far --5-
(33)
ble.

[ 7~ (52+ 18
-degree piecewise polynomial approxi- COS v1h)
ace Sh on a uniform mesh. There are A.?
13 h2 sm
. vh
function values, and roj to slopes, so 210 1

N-1 Observe that (33) gives two eigenvalues, A.?, o and A.t 1 , for each integer !,
e)+ L; u~rolx). l <N. Since the system (30)-(31) is of order 2N- 2, these are necessarily
j=O
all the'finite element eigenvalues. By a direct calculation,
! optimal weights, with p = 1, q O,
(34)

The corresponding eigenfunctions are in error by l 4 h 4 -s in the sth deriv-

ative.
The second eigenvalues 1?, 1 are of order O(h- 2 ), and they are not close
to any eigenPalue of the differential equation. At first notice this seems to be
a serious disadvantage of using cubic Hermite elements in eigenvalue calcula-
tions, that at least half of the eigenvalues are spurious and complete/y useless
as approximations. However, this phenomenon is quite typical of finite ele-
ment approximations. A closer look at even the piecewise linear approxima-
tions shows that for each h there is an integer lh such that only the first lh
y for 1 j <N- 1, but they remain eigenvalues are reasonable; moreover, hlh ~O as h O. This has iinportant
implications for the choice of a method to use in computing the eigenvalues
nctions vh in Sh satisfy the natural boundary of KQ = lMQ; only the leading eigenvalues are wanted.
uy condition at O. We close this section by noting that "lumping" can lead to a serious Ioss

l
228 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.3. EIGENVAI
"'
of accuracy [T8]. For example, a typical approach would be to replace Now we begin on the eigenvalue es
projection, defined as follows: If u is i1
1( f 1 ) subspace Sh (with tespect to the energy
10 Uj+! - Ui-1

(38) a(u- Pu, vh) O

in (30) with ui and O, respectively, and to replace
This means that in the energy norm a(
to the given u. In other words, if u were 1
Lu J, Pu would be exactly its Ritz app
in (31) with u~/210 andO. This leads toa problem KQ = XQ, which appears ous work on approximation (Theorem
simpler because M is replaced by the ·.identity. However, a direct calculation
shows that the O(h 6 ) eigenvalue error is increased to O(h 2 ). (39)

Our strategy for estimating lt A., j

spanned by the true eigenfunctions u 1 ,
6.3. EIGENVALUE ANO EIGENFUNCTION ERRORS
subspace of Sh in which to use the min
In this section we give a general theory of Rayleigh-Ritz approximations the trial functions Pu 1 , • • • , Pu1• These a
to elliptic eigenvalue problems Lu = A.u. The setting is by now familiar: eigenfunctions ut, ... , ut, but the essen
Integration by parts converts (Lv, v) into a symmetric form a(v, v), which is el ose.,
defihed for all v in the admissible space 3C';. The eigenfunctions are the points
u1 at which the Rayleigh quotient R(v) a(v, v)f(v, v) is stationary, and the
LEMMA 6.1
corresponding eigenvalues are l 1 R(u 1). These eigenfunctions are ortho- Let e1 be the set of unit vectors in E1 ~
gonal and, since L is symmetric, the eigenvalues are real.
On the subspace S\ the Rayleigh quotient becomes (40) at max j2(u, u
u in er.
Pu)

Then provided that at < 1, the app~oxin

and the stationary points Q yield the approximations ut and A.t. These points (41)
are determined by the matrix eigenvalue problem KQ = l_hMQ, and any
two eigenvectors Qi and Q1 satisfy the stan<;lard orthogonality relations
Proof To apply the minmax princi¡
(35) is !-dimensional. Certainly E1 itself is l-•
tion is whether possibly Pu* = Ofor so
Translated in terms ofthe eigenfunctions u~= l:(Q;)jlpj and ut = l:(Q1)i'Pb to be a unit vector, and therefore in e1, ..4

this means that

af j2(u*, u* - Pu*)- (u* P;
(36)
Since this contradicts a7 < 1, S 1 must
Thus the approximations mirror the basic properties of the true solutions. Now from the minmax principie (3í
They share also the minmax principie:

(37) A; = min max R(vh), A.t max.R(vk) =

v" in St
S¡cS" v" in s,

where S 1 denotes any subspace of dimension l. The numerator is bounded above by a(-
 CHAP. 6 SEC. 6.3. EIGENV ALUE AND EIGENFUNCTION ERRORS 229

pical approach would be to replace Now we begin on the eigenvalue estiinates. Let P be the Rayleigh-Ritz
projection, defined as follows: If u is in X 'E, then Pu is its component in the
'1¡-t ) , l ( u¡+t
' - ' 1)
u¡_ subspace Sh_ (with respect to the energy inner product):
10
(38) a(u Pu, vh) =O
.nd to replace
This means that in the energy norm a(v, v), Pu is the closest function in Sh
to the given u. In other words, if u were the solution to a steady-state problem
Lu J, Pu would be exactly its Ritz approximation uh. This link to our previ-
toa problem KQ 1Q, which appears ous work on approximation (Theorem 3.7) guarantees that
e ·identity. However, a direct calculation
•r is increased to O(h 2 ). (39)

CTION ERRORS
theory of Rayleigh-Ritz approximations
lu. The setting is by now familiar:
1
into a symmetric form a(v, v), which is
ce X'E. The eigenfunctions are the points
~v) = a(v, v)f(v, v) is stationary, and the
R(u1). These eigenfunctions are ortho-
eigenvalues are real.
1 quotient becomes
Our strategy for estimating l?- l 1 is this: We let E1 denoté the subspace
spanned by the true eigenfunctions up . .. , Uz, and choose S 1 = PE1 as the
subspace of Sh in which to use the minmax principie. Thus S 1 is spanned by
the trial functions Pu 1 , • • • , Pu1• These are not identical with the approximate
eigenfunctioris u1, ... , uf, but the essential point of the proof is that they are
el ose.
LEMMA 6.1
Let e1 be the set of unit vectors in E 1, and let
(40) (}'7 = max 12(u, u- Pu)- (u- Pu, u- Pu)l.
u in e¡

Then prpvided that (}'7 < 1, the approximate eigenvalues are bounded abo ve by
approximationsu? and l?. These points (41)
tvalue problem KQ _lhMQ, and any
.1e standard orthogonality relations
Proof To apply the minmax principie we must make sure that S 1 PE1
QTKQz = Móu. is /-dimensional. Certainly E1 itself is /-dimensional, and therefore the ques-
tion is whether possibly Pu* O for sorne nonzero u* in E1• Normalizing u*
ions u?= í:(Qi)jQ'j and uj = í:(Q1)¡Q'¡, to be a unit vector, and therefore in e1, Pu* = O would imply that

(}'; > 12(u*, u* - Pu*) - (u* Pu*, u Pu*) 1 1(u*, u*) 1 l.

Since this contradicts (}'7 < 1, S1 must be /-dimensional.

basic properties of the true solutions.
Now from the minmax principie (37), .
e:

max R(vh), max.R(vh) = max a(Pu, Pu) ·

'vh in St vhinS1 uine 1 (Pu, Pu)

nension l. The numerator is bounded above by a(Pu, Pu) <a(u, u), since Pis a projec-
 )

230 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.3. EIGENV

tion Ít:J. the energy norm.t The denominator is bounded below by Proof We want to estímate a~. It~

(Pu, Pu) (u, u) 2(u, u Pu) +(u Pu, u- Pu) > 1 a~.
2l(u, u- Pu)l = 2¡ ~ c1A.¡

Therefore, the lemma is pro ved: 2K[!(/

max a(u, u) _ _A._

1 _. since !a(v, w)l < Kllvllmllwllm· Appl~
u in e¡ 1 af - 1 a?
The problem is now to estímate cr;; and for this we need the following 2l(u, u- Pu)l < 2KCZhZ(k-m>
identity.
C'h2(k-m) JI~
(44)
LEMMA 6.2
C' h2(k-m) A_jk,

This holds for all functions u = ~ e

(42)
Proof Since u1 is a true eigenfunction,
(u 1, u - Pu)
by (10). Furthermore, a(Pu¡, u- Pu) O, if we take 11'
tion (38) of the projection P. Subtracting,
case occurs at c1 = 1, because it is
eigenvalue A.1• The next-to-last step w
quálity 11 v llk <e 11 Lk!Zmv llo· For k
11 v 11~ c 2 a(v, v), and the inequality:
= A.¡ 1 a(u1, u- Pu),
differential operator L are sinooth.
= Pu1 in the defini-
The other term in af is of highe
theorem with s = O yields

(45) (u Pu, u Pu) < (

Therefore, if we increase the constan1
Multiplying by c1 and summing on i, the result is (42).
expression will exceed a? for all 'smal
Now we have the basic error estímate for eigenvalues. The theorern is now proved; if h
then
THEOREM 6.1
/f Sh is a finite element space of degree k - 1, then there is a constant o
su eh that the approximate eigenvalues are bounded for small h by Such error bounds for the Rayle
particularly in the Russian literature.
(43) theory, including eigenvalue and eig1
self-adjoint case. Combined with tl
This is in agreement with the explicit bounds (28) and (34) for linear and cubic elements, his analysis establishes th
elements in one dimension. pro ved. for self-adjoint problems by
principie must be replaced by a con1
tThis is in the corollary to Theorem 1.1 and is easy to see directly: followed by an application of the (
a(v, v) a(Pv, Pv) 2a(v - Pv, Pv) + a(v Pv, v - Pv). given in Section 2.3. The integral of (
The last term is nonnegative, and the next-to-last term is zero by the definition (38) of produces exactly 2niuju1 ; Bramble an
the projection P. have computed the error in using fin:
 CHAP.6 SEC. 6.3. EIGENVALUE ANO EIGENFUNCTION ERRORS 231

inator is bounded below by Proof We want to estímate u?. Its first term is
) + (u - Pu, u Pu) 1 ~ u?. 21 (u, u Pu) 1 21 t e1).;: 1a(u1 - Pu1, u - Pu) 1

< 2Kjl(l- P) ~ e¡Aj 1Utllm!l(l P)ullm

since 1a(v, w) l K 11 v 11m 11 w 11m· Applying the approximation theorem,

rr, and for this we need the following 2j (u, u Pu) j 2KC 2h 2<k-m> 11 te,).¡ 1U¡ llk ll u llk
(44)
C' h2<k-m> 11 t e1).}k/2m>- 1 u1 llo 11 t e1).f12mu1 llo
C' h2(k-m) .A,}k/m}-1.

This holds for all functions u ~ e¡u1 in e1, since ~e[ 1; the extreme
case occurs at e1 1, because it is multiplied by a power of the largest
eigenvaltie A-1• The next-to-last step was more delicate and required the ine-
ion,
quálity ll v llk <e ll Lktzmv !lo· For k m this is exactly the ellipticity condition
11 v !1~ < e2a(v, v), and the inequality holds for all k if the coefficients of the
differential operator L are stnooth.
= O, if we take vh Pu/in the defini- The other term in af is of higher order in h, since the approximation
ing, theorem with s = O yields

l(u1 - Pu1, u- Pu). (45)

the result is (42). Therefore, if we in crease the constant C' in (44) to a larger constant 8, that
expression will exceed uf for all small h.
.ate for eigenvalues. The theorem is now pro ved; if h is small enough to ensure that uf !,
then

·gree k 1, then there is a eonstant 8

zre bounded for small h by Such error bounds for the Rayleigh-Ritz method have a long history,
particularly in the Russian literature. Vainikko [7] has given a very complete
theory, including eigenvalue and eigenfunction estimates even in the non-
self-adjoint case. Combined with the approximation theorems for finite
Junds (28) and (34) ¡ó'r linear and eubie elements, his analysis establishes the same error bounds which we have
pro ved. for self-adjoint problems by means of the minmax principie. This
principie must be replaced by a contour integration in the complex plane,
and is easy to see directly: followed by an application of the Galerkin estimates for static problems
Pv, Pv) + a(v - Pv, v Pv). given in Section 2.3. The integral of (L- zl)- 1 around a true eigenvalue ).1
o-last term is zero by the definition (38) of produces exactly 2niufu1 ; Brambleand Osborn [B29], and Babuska and Fix,
· have computed the error in using finite elements instead.
232 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.3. ElGEN'

The minmax argument, which we ha ve preferred for self-adjoint problems quotient is R(v) = a(v, v)fb(v, v), and
beca use it is mo~e elementary, has been applied to finite elements by a number where the mass matrix Mb is formed
of authors; the basic reference is [816]. We have made the argument more (Mb)jk = b(rpi, rpk). One change occurs
precise, in order to determine not only the power h 2 ck-mJ but also the correct (45) in the 0-norm i~ replaced by
dependence on l:
b(u--'- Pu, u- Pu) <C
However, the first term in u? remains •
The last factor means that higher eigenvalues are progressively more difficult
final estímate in the theorem is unch<:
to compute. This prediction has been tested experimentally, ancl the appear-
The effect on the computed eigen
ance of Af1m is confirmed by the calculations described in [Cl4J with a quintic
finite element technique-change of d
element, those in Section 8.4 for the eigenvalues of an L-shaped membrane,
rature, nonconformity of elements-
and the following numerical results which are reproduced from [L3]. The
energy in static problems. We cautio
bicubic Hermite element was applied to a square plate which is simply sup-
polygonal domain Q\ the error in en
ported either on all sides (SSSS) or on only two sides, with the other two left
fore, this error is not the O(h 3 ) estab:
free (SFSF); the number of elements is fixed at 10 per side (Fig. 6.1). Since
if all trial functions vanish over n -
k = 4 and m = 2, the prediction is a relative error of
completely lost, with_an effect which.
We turn now to the eigenfunctio1
A.?-
A
Az
""'e
h4,
Az. errors are of the same magnitude as
l
It must be expected, of course, that tl
accuracy will deteriorate for higher
V Ah- A
-2 log
1 1 involve two simple identities, stated
A¡ technical computation.

LEMMA 6.3

-3
With the normalizations (u1, u1) =

o
(46) a(u1 - u?, u1 - u?)=)
o
o-SFSF Proof Observe that
-4

D w
Wx
Wy
Wxy
A-SSSS
a(uz - u7, Uz - uf)= a(u1:
=Az-
= 11[2

- 5 1~--~--~3--~--~~~--L---~--~--~9----L---,L,--~--_jl3 But the quantity in brackets is exactl:

l
2 - 2(u1, u?) = (u 1, u1) - 2(u
Fig. 6.1 Errors in the higher eigenvalues of a square plate.

The generalized eigenvalue problem Lu = ABu yields to exactly the same With this identity, and the estim
analysis. Suppose that B is a symmetric operator of order 2m' < 2m, corre- is necessary only to bound the error i
sponding to the quadratic form (Bv, v) = b(v, v). Then the true Rayleigh in energy will be an immediate conse1
 CHAP. 6 SEC. 6.3. EIGENV ALUE AND EIGENFUNCTION ERRORS 233.

ave preferred for self-adjoint,problems quotient is R(v) a(v, v)fb(v, v), and the discrete one becomes qr Kqfqr Mbq,
applied to finite elements by a nwmber ,where the mass matrix Mt is formed with respect to the new inner product:
)]. We have made the argument more (Mb)jk = b(rpj, rpk). One change occurs in the last term of u?, where the bound
· the power h2 (k-m> but also the correct (45) in the 0-norm is replaced by

However, the first term in uf rernains larger, 2b(u, u- Pu) ,...., h2 (k-m>, and the
nvalues are progressively more difficult
final estímate in the theorern is unchanged. .
:ested experimentally, and the appear-
The effect on the computed eigenvalues of other approximations in the
ttions described in [Cl4] with a quintic
finite elernent technique-change of dornain or coefficients, numerical quad-
igenvalues of an L-shaped membrane,
rature, nonconforrnity of elernents-is exactly comparable to the effect on
vhich are reproduced from [L3]. The
energy in static problerns. We caution only that, for a change frorn Q to a
toa square plate which is simply sup-
po1ygonal dornain Qh, the error in energy must be measured over Q. There-
only two sides, with the other two ieft
~ore, th~s error i.s not th~ O(h 3) established ih Section 4.4 over the polygon;
is fixed at 10 per side (Fig. 6,1). Since
tf all tnal functwns vamsh over Q Qh, then the energy in this region is
elative error of
cornpletely lost, with an effect which is proportional to its area, of order hz.
We turn now to the eigenfunctions. We should like to prove that their
errors are of the sarne rnagnitude as those in steady-state problems Lu f
It must be expected, of course, that there will be sorne dependence on /; the
accuracy will deteriorate for higher eigenfunctions. Again the arguments
involve two simple identities, stated in the following lernrnas, and a rather
technical computation.

LEMMA 6.3
/:¡.
With the normalizations (u1, u¡) (uf, ut) 1, we have
/:¡. o o

~ o
(46)

o-SFSF Proof Observe tha t

J
w t:..-SSSS
Wx a(u¡ - uf, Uz - uf) a(u1, u1) - 2a(uz, uf) + a(uf, uf)
Wy A¡- 2A.¡(u¡, uf)+ A.?
Wxy = 11[2 2(u1, uf)]+ A.?- A.1•

11 13 But the quantity in brackets is exactly

l

eigenvalues of a square plate.

2 2(u¡, uf) = (uz, u¡) 2(u1, uf) + (uf, uf) 11 U¡ - uf 11~-

n Lu = A.Bu yields to exactly the same With this identity, and_the estímate already established for ;.; - J., it
ic operator of order 2m' <2m, corre- . is necessary only to bourtd the error in the norrn 11 u1 u? llo; then the error
, v) = b(v, v). Then the true Rayleigh in energy will be an imrnediate consequence.
Q34 EIGENV ALUE PROBLEMS CHAP. 6 SEC. 6.3. EIGENVAJ

LEMMA 6.4 The eigenfunction error is now bounde1

error u1 Pu1 ; it follows from (39) tha-
For al/ j and l,

(47) (52)
11 U¡ - Pu7 ilo C'(l +,
C"[hk -1
Proo,f Since the term - A1(Pu1, uj) appears on both sides, we ha ve only
to show·.that This is the essential part of our theorer

THEOREM 6.2
Because uj and u1 are eigenfunctions, these two sides can be rewritten as Jf Sh is a.finite element space ofdegré
a(Pu1, u') and a(u1, uj), respectiv~ly. Now equality comes from the definition thenfor sma/1 h,
(38) of the projection P.
The expression (47) resembles a truncation error in the eigenvalue equa- (53) 1lu1 u?llo <'
tion; it coincides with a(Puz, uj) A1(Pu1, uj). (54) a(u1 - u;, u1 ufr) <,
The set u~, ... , u';¡ forms an orthogonal basis for S\ and in particular
Jf A1 is a repeated eigem•ahie, then the
(48) chosen so that these estimates still lwh
and are consisten! with the special case (
J
The ,identities (47) and (48) can be interpreted in the following way. First,
Proof The bound (53) is virtually
from (47) it is seen that the coefficient (Pu1, uj) is small if Aj is not close to
proved. It remains only to show that
.A1• Then from (48) it follows that Pu1 is close to u;. This will be our strategy
comes from fiddling about with the tri
for estimating uf - Pu1 (and hence uf - u1), but to make the process rigorous
it is convenient to consider separately the cases of distinct and repeated
eigenvalues.
If A1 is distinct from the other eigenvalues, then according to our eigen-. Recallíng that u1 and u; are unit vect
value bounds (43) there is a separation constant p such that for small h, p >O, this is the same as IP 1] ll
(49) p for all j. !Ju¡ u?llo llu¡- Pu?llo + 11
Now the right side is given by (52), a1
Now come the computations. Writing p for the key coefficient (Puz, uf) in
is pro ved: e= 2C". The error in energy
(48), the size of the other terms is given by
6.3. This is the simpJest argument we l
The case of a repeated eigenval
awkward, but not essentially different.

(50)
#1 (;.? ~ ;.)\u 1 Pu¡, uj)
2 p between these eigenvalues and the a
(49). The constant p becomes a matrix
'< p 2
(u1. - Pu1, uj) 2

The long computatio~ (50) is now carri1

since the square of the norm is the sum of the squares of the components. +
1 R on the.left side rather than the
Thus we have the crucial estímate
 CHAP. 6 SEC. 6.3. EIGENVALUE AND EJGENFUNCTION ERRORS 235

The eigenfunction error is now bounded in terms of the Ritz approximation

error u1 Pu1; it follows from (39) that

11 U¡- Puf !lo C'(1 p)[hk + h 2Ck-m>] 11 Utllk

(52)
< C"[hk + h2(k-m>]Af/2m~
~?) appears on both s~des, we have only
This is the essential part of our theorem.

THEOREM 6.2
s, these two sides can be rewritten as Jf Sh is afinite element space of degree k 1 and A1 is a distinct eigem•alue,
~ ow equality comes from the definition thenfor small h,

:runcation error in the eigenvalue equa- (53) 11 U¡ uf llo c[hk + h2Ck-m>]Af12m,

1(Pu1, uj).
hogonal basis for S\ and in particular
:nterpreted in the following way. First,
nt (Pu1, uj) is small if Aji's not close to
is close to uf. This will be our strategy
u1), but to make the process rigorous
ely the cases of distinct and repeated
(54) a(U¡ Uf, U¡- u?)< c'h 2 Ck-m> Af/m.
Jf A1 is a repeated eigenl'alue, then the orthonormal eigenfunctions uj can be
eh osen so that these estimates still hold. The -estimates are the best possible
and are consistent with the special case (27) of linear elements.
Proof The bound (53) is virtually the same as (52), which is already
pro ved. It remains only to show that the factor p is el ose to 1, and this
comes from fiddlíng about with the triangle inequality:

;envalues, then according to our eigen-

Recalpng that u1 and uf are unit vector~, and choosing their sign so that
tion constant p such that for small h,
P >O, this is the same as 1P 11 < 11 u1 Pu7llo· Therefore,
p for all j. jju,- uf!!o < llu, Pu?ilo + li(P- l)ufllo < 2llu¡ Pu?llo·

1g P for the key coefficient (Pu1, uf) in Now the right side is given by (52), and the first statement of the theorem
1en by is pro ved: e= 2C". The error in energy (54) follows immediatelyfrom Lemma
6.3. This is the simplest argument we have seen for eigenfunctions.
The case of a repeated eigenvalue A1 = Al+ 1 = · · · AI+R is more
awkward. but not essentially different. There is stiii a separation constant
A¡ .
Aj - A¡
)2( U¡ - '
P
U¡,
b.)2
uj
p between these eigenvalues and the approximations Aj to the others, as in
(49). The constant p becomes a matrix of order R 1;-
: ( u1 Pu¡, uj) 2
;¡
O< i,r <R.
U¡- Pu¡ll~,
'
The long computation (50) is now carried out with all ofthe termsj 1, ... ,
.um of the squares of the components. 1 + R on the left side rather thanthe right and Jeads to

~
~
priUl+ih
Ut+r
'11. C[hk + h2(k-m>]A_k/2m.
1
11 0
236 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.4.

Inverting p, this means that new eigenfunctions U1+r-linear combinations 1970 is [21]. Before describing a ne
of the ·old ones-can be chosen so that widely u sed· algorithm in this class
method. In its simplest form, for tl
iteration proceeds by solving a li1
Then the approximation to ). is ).n+ 1
Since the uf+r are known to be orthonormal, it is even possible to keep the to X ÍS the normalized vector Xn+ 1
U1.;., als~:,prthonormal, without any harm to the estímate. were an eigenvector. If we imagine
in terms of the true eigenvectors v.J'
The same theorem and proof-with the inner product (u, v) replaced iterations is to amplify each com¡:
throughout by b(u, v)-apply to the generalized eigenvalue problem Lu = 2: c/l)-nvr If A1 is distinctly smal
).Bu.
first component will become domim
vector v 1 • The convergence is like tha
6.4. COMPUTATIONAL TECHNIOUES
fraction ).1 1Az; the error xn X is (
becomes more effective when this ra
The Rayleigh-Ritz principie has led to the matrix eigenvalue problem One tec'hnique for reducing this
KQ = lMQ, which now remains to be solved. This is nota trivial problem, at the nth step by A lnl. This shi
and it is scarcely discussed in a typical text on linear algebra. An efficient amount ln. If ).,n is close to a true ei1
algorithm should take a"dvantage of the fact that K and M are symmetric the corresponding corríponent of th
positive definite and also of the fact that they are sparse. The last property quantity (). - ).n)- 1 • There is no nU1
would be lost, (or example, if we were to factor M into LLT by a Cholesky though A - li is nearly singular.
elimination, and compute the eigenvalues of L- 1K(L- 1 )r by a standard algo- approximation to the eigenvalue th'
rithm (we would choose QR, ora Givens method which begins by reduction Ieigh quotient ln (Ayn, Yn)f(yn, YJ
to a triangular matrix, rather than the older Jacobi method); This Ioss of tient is very ftat in the neighborhoc
sparseness will not be so serious for a srrÍall problem, which can be handled stationary point), and the algorithr
within the core of the computer; but for a large system it is inefficient. cubic convergence: ).n+l l "' ().n-
We propose to find the eigenvalues-or more precisely the first few eigen- A - lJ means that Gauss eliminat
values, since it would be useless to compute higher eigenvalues which have the triangular factors of A cam1ot b
no physical significan~e-directly from the equation KQ = ).MQ. We shall done with the simple iteration Ayn+ 1
reject Iumping, since a diagonal Mis not enormously better than a banded In the generalized eigenvalue prot
M. Kyn+ 1 = Mxn at each stage and the
There is also a technique of economization, which reduces the order of this form, however, the iteration mat
the system by working only with a small number of master ~'ariables., The fore, if the Cholesky factorization
dependence of the other slave variables is assumed a priori, and these degrees s- 1 M(BT)- 1 as the (symmetric) iterai
of freedom are thereby eliminated [22]. Fried has described this idea in the nonsparse matrices we wished to a
following terms: Placing a point load ata node zj corresponding toa master not begin by multiplying these ma1
variable, -let cJ>j be the finite element solution in the static problem. Then is determined from Brvn+ 1 u, Bw
these functions cJ>j are a basis (not local, as the original rpj were) for the trial Convergence of the normalizing fact
space in the economized problem. These functions can be expected to repre- eigenvector x~' again depends on the
sent low-frequency modes fairly well, and those are the ones which are cal- A very useful variant of inverse
culated. Nevertheless, out impression is tñat as efficient algorithms are introduced by Bauer and improved l
constructed for the original problem, this economization will become less kinson.t The idea is to find several ei!
necessary and less popular.
An authoritative reference for eigenvalue algorithms prior to about tit is known in the engineering literatur
 CHAP.
;'unctions ul~r-iine¡u combinations
rmal, it is even possible to keep the
rm to the estímate.
h the inner product (u, v) replaced
neralized eigenvalue problem Lu
j to the matrix eigenvalue problem
solved. This is not a trivial problem,
text on linear algebra. An efficient
e fact that K and M are symmetric
tt they are sparse. The last property
to factor M into LLT by a Cholesky
es of L _, K(L -~y by a standard algo-
lS method which begins by reduction
older Jacobi method). This loss of
a1aii problem, which can be handled
· a Iarge system it is inefficient.
-or more precisely the first few eigen-
tpute higher eigenvalues which have
the equation KQ = }.,MQ. We shall
ot enormously better than a banded
1ization, which reduces the order of
all number of master variables., The
; assumed a priori, and these degrees
Fried has described this idea in the
a node z1 corresponding toa master
)]ution in the static l)roblem. Then
as the original rp1 were) for the trial
: functions can be expected to repre-
td those are the ones which are cal-
is that as efficient algorithms are
his economization will become less
mvalue algorithms prior to about
6 SEC. 6.4. COMPUTATIONAL TECHNIQUES 237
1970 is [2 1]. Befo re describing a newer method, we want to review a very
widely u sed algorithm in this class: the in verse iteration or in verse power
method. In its simplest form, for the eigenvalue problem Ax = 1x, inverse
iteration proceeds by solving a linear system at each step: Ayn+t xn.
Then the approximation toA. is }.,n+t = 1/IIYn+t 11, and the new approximation
to X is the nonnalized vector xn+l = }.,n+tYn+J• These would be exact if xn
were an eigenvector. If we imagine that the starting vector x 0 is expanded
in terms of the true eigenvectors v1, x 0 .2: c1v1, then the effect of n inverse
iterations is to amplify each component by (}.,)-n; xn is proportional to
.2: c1(1 1)-nvr lf A1 is distinctly smaller than the other eigenvalues, then the
first component will become dominant, and xn wiii approach the unit eigen-
vector v 1 • The convergence is Iike that of a geometric series, depending on the
fraction A1 /A 2 ; the error xn x is of order (1 1 / A2 )n. Obviously the method
becomes more effective when this ratio is small.
One technique for reducing this ratio is to shift the origin, replacing A
at the nth step by A }.,ni. This shifts all the eigenvalues of A by the same
amount An. If }.,n is close toa true eigenvalue A., so that }., - }.,n is small, then
the corresponding component of the vector Yn+l is amplified by the Iarge
quantity (}., }.,)- 1 • There is no numerical difficulty with this process, even
though A }.,J is nearly singular. In fact, it is useful to have a better
approximation to the eigenvalue than }.,n 1/IIYn 11. For example, the Ray-
Ieigh quotient An (Ayn, Yn)f(yn, y) will be much more accurate. This quo-
tient is very ftat in the neighborhood of a true eigenvalue (where it has a
stationary point), and the algorithm with these improved shifts possesses
cubic convergence: }.,n+ 1 }.,
('V (}.,n 1) 3 ; On the other hand, shifting A to
A 4nl means that Gauss elimination has to be redone at each iteration;
the triangular factors of A cannot be stored and used over and over, as is
done with the simple iteration Ayn+ 1 xn.
In the generalized eigenvalue problem, the simplest idea would be to solve
Kyn+J = Mxn at each stage and then obtain xn+t by normalizing Yn+t· In
this form, however, the iteration matrix K- 1 M will be nonsymmetric. There-
fore, if the Cholesky factorization of K is BBr, it is preferable to use
B- 1M(B'r' as the (symmetric) iterating matrix. This appears to involve the
nonsparse matrices we wished to avoid, but obviously the iteration does
not begin by multiplying these matrices. Instead the approximation un+t
is determined from BTvn+! =u, Bwn+t Mvn+l' un+l normalized wn+t•
Convergence of the normalizing factors to Ap and of the vn (not u)) to the
eigenvector x 11 again depends on the ratio 1 2 /1 1 •
A very useful variant of inverse iteration is the block power method,
introduced by Bauer and improved by Rutishauser and by Peters and Wil-
kinson.t The idea is to find severa] eigenvalues at once, by carrying l approxi-
tlt is known in the engineering literature as subspace iteration.
238 EIGENVALUE PROBLEMS CHAP. 6 SEC. 6.4.

mate ftigenvectors in the iterations. (Obviously they must be coupled as the Parlett recommend an accelerated se4
alg~rithm proceeds, or the result would only be l different approximations
to the same fundamental mode.) Convergence toA¡ occurs at the rate lJli+P
and repeated eigenvalues can be handled without difficulty.
Block iteration for Ax = lx proceeds as follows. Suppose the l starting
vectors, ~ssumed to be orthonormal, are the columns of an N x l matrix This is a variant of the usual Newtc
P 0 • The ,first step is to solve the l equations AZ 1 P 0 • Then, rather than replaced by a difference quotient; th
normalizíng separately each column of Zl' we orthonormalize all/ columns. is el ose toa true A, the computer may
For this we form the l x 1 matrix Z{Z 1 and find its eigenvalues J.l¡- 2 (which turn to ordinary inverse iteration. We
are the first approximation to X¡- 2 , • • • , l¡ 2 ) and corresponding eigenvectors algorithm was very successful in the n
W¡. The new approximation P 1 is the product of Z 1 and the matrix W with ter 8.
columns W 1 J.lp ... , w1p 1• The columns of P 1 are orthonormal-they are the We want to give the theoretical
approximate eigenvectors of A-since P{P 1 = wrz;z 1 W l. At the next can be based on a classical theorem
step AZ2 . P 1 , and so on. two real symmetric matrices A and D •
This algorithm would reproduce exactly the eigenvectors vl' ... , v1 of tion A BDBr, where Bis any nons
A. More precisely, let each column of P0 be a linear combination of vl' ... , number ofnegative eigenvalues, ofposJ
v 1-say P 0 VQ, where Q is an 1 x 1 orthogonal matrix and Vis the N x l The proof is especially neat when
matrix whose columris are Vp •.• 'V¡. We note that vrv = 1 and AV VA, eigenvalues. Let Be be a family of nc
if A is the diagonal matrix whose entries are l 1 , • • • , A1• ually from the identity matrix (at (}
Thus the first block power step gives Z 1 A- 1P0 = A- 1 VQ VA- 1 Q, cannot be sure that the particular [(¡
and Z{Z 1 becomes QrA -zQ. Since Q is orthogonal, it follows that the singular, but the construction of a su
eigenvalues p,; 2 of Z{Z 1 are equal to the en tries 1; 2 of the diagonal matrix B9 DB~ are always symmetric and al~
A -z. In other words, if P 0 = VQ then the first step of the method would values A.((}) are real, change graduall
produce P 1 V, and the correct eigenvalues . quently, the number of eigenvalues o
. The changes involved in working with A Q = lM Q rather than Ax = Ax () = 1 as at () O. In other words, A ;
are described in the final paragraph of this chapter. The block power algo- eigenvalues and of positive eigenvalu
rithm as restated there is completely straightforward to program. the same argument for the mat'rices
These inverse power methods are very simple and practica!, especially limit as E ----"' O.
when only modest accuracy is required. However, there is a new technique, Now we apply this Iaw of inertia
based on a more subtle matrix theorem, which has also become a most is factored into LDLT (as in Section
highly recommended algorithm for the band eigenvalue problem KQ lM Q. 1's on the diagonal and D is the di:
It is due to Peters and Wilkinson [PI, P2], and depends on the following of the pivots determine the signs of
lovely matrix theorem: The number of eigenvalues less than a given l 0 can exchanged during the elimination pt
be determinedjust by counting the num/Jer of negative pivots when Gauss early pivots happens to be zero, thc
elimination is applied to K- l 0 M. exchanged to preserve the symmetr)
This suggests an algorithm based on bisection. Suppose it is determined of both rows and columns is again
that there are n 0 eigenvalues below the first guess l 0 • Then the Gaussian about by a permutation matrix B. TI
pivots for K (l 0 /2)M reveal the number n 1 which is below l 0 /2; and the and after such an exchange the Gaus
remaining n0 - n1 eigenvalues must líe between l 0 /2 and l 0 • Repeated bi- of the eigenvalues correctly.)
section will isolate any eigenvalue with tremendous accuracy, but the process The application to the generalize
requires a Gauss elimination at each step and is rather expensive. It must be lows. Let A= K 1 0 M, and count ·
speeded up by using the values of the. pivots (or their product, the determi- elimination. We claim that this equa
nant d(l) = det (K lM)) rather than just their signs. Clough, Bathe, and 1 0 in the given problem K Q = 1MQ
 CHAP.
wiously they must be coupled as the
d only be 1 different approximations
rgence to A¡ occu~s al the rate Aj Al+ 1,
:d without difficulty.
ds as follows. Suppose the 1 starting
tre the columns of an N X 1 matrix
ations AZ 1 P 0 • TH:en, rather than
Z 1 , we orthonormaliie all/ columns.
2
1 and find its eigenvalues J-li (which
A1- ) and corresponding eigenvectors
2
roduct of Z 1 and the matrix W with
of P 1 are orthonormal-they are the
P~P 1 WTZ~Z 1 W l. At the next
xactly the eigenvectors v~' . .. , v1 of
0 be a linear combination of v 1 , ••• ,
•rthogonal matrix and Vis the N X l
Ve note that VTV 1 and AV= VA,
tries are A1, ••• , A1•
$ Z1 A- 1P0 A- 1 VQ = VA- 1 Q,
? is orthogonal, it follows thatthe
1e entries Ai 2 of the diagonal matrix
t the first step of the method would
;alues.
th AQ = AMQ rather than Ax Ax
this chapter. The block power algo-
traightforward to program.
very simple and practica!, especially
. However, there is a new technique,
em, which has also become a most
,and eigenvalue problem KQ AMQ.
, P2], and depends on the following
· eigenvalues less than a given A0 can
mber of negative pivots when Gauss
n bisection. Suppose--it is determined
.e first guess A0 • Then the Qaussian
1ber n 1 which is below A0 /2; and the
: between A0 /2 and A0 • Repeated bi-
tremendous accuracy, but the process
:p and is rather expensive. It must be
Jivots (or their product, the determi-
1 just their signs. Clough, Bathe, and
6 SEC. 6.4. COMPUTATIONAL TECHNIQUES 239
Parlett recommend an accelerated secant iteration
This is a variant of the usual Newton's method, in which the derivative is
replaced by a difference quotient; the factor 2 is for acceleration. Once Ak
is close to a true A, the computer may cease further matrix factorizations, and
turn to ordinary inverse iteration. We can confirm that the Peters-Wilkinson
algorithm was very successful in the numerical experiments reported in Chap-
ter 8.
We want to give the theoretical background for this algorithm, which
can be based on a classical theorem known as Sylvester's law of inertia: Jf
two real symmetric matrices A and D are related by a congruence transforma-
tion A= BDBT, where Bis any nonsingular matrix, then they have the same
number of negative eigenvalues, of positive eigenvalues, and of zero eigenvalues.
The proof is especially neat whe11 D is nonsingular; that is, it has .no zero
eigenvalues. Let B9 be a family of nonsingular matrices which changes grad-
ually from the identity matrix (at fJ O) to the matrix B (at fJ = 1). [We
cannot be sure that the particular family Be fJ B + (1 - fJ)l will be non-
singular, but the construction of a suitable Be is never difficult.] The matrices
BeDBI are always symmetric and always nonsingular. Therefore, their eigen-
values A(fJ) are real, change gradually with fJ, and never cross zero. Conse-
quently, the number of eigenvalues on each side of zero must be the same at
fJ = 1 as at fJ = O. In other words, A and !{ have the same number ofnegative
eigenv.alues and of positive eigenvalues. If D happens to be singular, we use
Jhe same argument for the matrices D ± El and prove the theorem in the
limit as E ·-----" O.
Now we apply this law of inertia to Gauss elimination. If the matrix A
is factored into LDLT (as in Section 1.5), where L is lower triangular with
1's on the diagonal and D is the diagonal matrix of pivots, then the signs
of the pivots determine the signs of the eigenvalues. (In case two rows are
exchanged during the elimination process, as will be required if one of the
early pivots happens to be zero, the corresponding two columns must be
exchanged to preserve the symmetry of the matrix. Bu_t such an exchange
of both rows and columns is again a congruence transformation, brought
about by a permutation matrix B. Therefore, the law of inertia still applies,
and after such an exchange the Gauss elimination pivots still give the signs
of the eigenvalues correctly.)
The application to the generalized eigenproblem KQ AMQ is as fol-
lows. Let A K- A0 M, and count the num~er of negative pivots in Gauss
elimination. We claim that this equals the number n 0 of eigenvalues below
A0 in the given problem KQ AMQ.
240 EIGENVALUE PROBLEMS CHAP.6

Prpof. Jlewrite this given problem as M- 112KM- 112(M112Q) l(MII2Q),

so th'\-t ng is the number of eigenvalues of M-11 2KM- 112 which lie below l 0 •
This is the number of negative eigenvalues of M- 112KM- 112 l 0 /. But by
the law of inertia (again! Choose this last matrix as D, and B = M 1' 2 ) this
coincides with the number of negative eigenvalues of A K- l 0 M. There-
fore, n0 !s easily determined by applying Gauss elimination to A. The eigen-
vectors fóllow in one or at most two steps of inverse iteration with the matrix
C K , lcomputedM. A good initial guess for the eigenvector is described in

7
[P2].
Experiments are now going on (see the supplementary bibliography)
to compare the performance .of eigenvalue algorithms. Clough and Bathe INITIAL-VALUE P
ha ve found the determinant search techniques, of which the Peters-Wilkinson
method is one variant, to be most competitive when the bandwidth is not
large; about five triangular factorizations were required for each eigenvalue,
and the cost of every such iteration is proportional to the square of the band-
width. The most generally successful technique was the inverse block power
method, alias subspace iteration. They express the algorithm in the following
equivalent form: starting from an eigenvector guess xn-t' which is a matrix 7.1. THE GALERKIN-CRANK-NICOI
with l orthonormal columns; sol ve KYn = M Xn_ 1 • Then sol ve the /-dimen- FOR THE HEAT EOUATION
sional eigenproblem
So far we have discussed only st
(55) and eigenvalue problems. Galerkin'
also to initial-value problems, and ir
V¡ are the approximate eigenvalues, and the new matrix Xn of approximate element approximations. They still
eigenvectors is the product of Yn with a square matrix of order l, formed from differences, for general geometries at
the eigenvectors Q of (55). Two computational problems have been consid- tively slowly in time (low Reynolds
ered in detail by Bathe and Parlett-the solution of the small eigenproblem high-speed wave problems, we shall
(55), for which they use a Jacobi-type sweep since the matrices become near The main theorem for finite dim
to diagonal, and the selection of the starting X0 • There is also a compromise sistent method is convergent if and <
to be made on the choice of l; with a large l there are few iterations, but each to Galerkin approximations. In fact,
one is expensive. They chose l = min (2p, p + 8) in computing the first consistency are often easier to verify
p eigenvalues, and found that eight iteration steps gave excellent results. ences. We show below how the Gal
This technique is effective even on problems too large to be handled in core, the differential equation, and we rej:
and it may well become the accepted algorithm for finite element eigenvalue by consistency: The Galerkin method
problems. dense in the admissible space. This m
mated more and more closely, as h
mation theorems of Chapter 3 establ
even the degree of approximation-v..
sistency, or the order of accuracy, o
A natural setting in which to illus
equation

(1) au au =
at - ax2
2
f(x, t),
 CHAP. 6
ts M-ti2KM-ti2(A{ti2Q) =,l(Mti2Q),
of M- 112KM- 112 which lie below 10 •
tlues of M- 112KM- 112 - lo!· But by
ast matrix as D, and 1$ M 112 ) this
igenvalues of A= K 10 M. There-
~ Gauss elimination to A. The eigen-
)S of inverse iterationwith the matrix
:ss for the eigenvector is described in
ee the supplementary bibliography)
1 alue algorithms. Clough and Bathe
tiques, of which the Peters-Wilkinson
npetitive when the bandwidth is not
ns were required for each eigenvalue,
·oportional to the square of the band-
chnique was the inverse block power
:xpress the algorithm in the following
lVector guess Xn-t, which ÍS a matrÍX
. M Xn_ 1 • Then sol ve the /-dimen-
d the new matrix xn of approximate
square matrix of order /, formed from
1tational problems have been consid-
e solution of the small eigenproblem
;weep since the matrices become near
.rting X 0 • There is also a compromise
rge 1 there are few iterations, but each
n (2p, p + 8) in computing the first
teration steps gave excellent results.
)lems too large to be handled in core,
lgorithm for finite element eigenvalue
7 INITIAL-VALUE PROBLEMS
7.1. THE GAlERKIN-CRANK-NICOlSON METHOD
FOR THE HEAT EQUATION
So far we have discussed only steady-state problems-elliptic equations
and eigenvalue problems. Galerkin's principie is flexible enough to apply
also to initial-value problems, and in this chapter we want to consider finite
element approximations. They still have important advantages over finite
differences, for general geometries and for problems which evolve compara-
tively slowly in time (low Reynolds numbers, in the case of fluid flow). For
high-speed wave problems, we shall poi1;1t out sorne disadvantages.
The main theorem for finite differences is well known: A formally con-
sistent method is convergent if and only if it is stable. This applies verbatim
to Galerkin approximations. In fact, the only difference is that stability and
consistency are often easier to verify for finite elements than for finite differ-
ences. We show below how the Galerkin principie imitates the stability of
the differential equation, and we repeatnow the disguise which is assumed
by consistency: The Galerkin method is consistent if the subspaces Sh become
dense in the admissible space. This means that every admissible vis approxi-
mated more and more closely, as h ---:--* O, by trial functions vh. The approxi-
mation theorems of Chapter 3 establish exactly this property, aq.d they give
even the degree of approximation-which translates into the degree of con-
sistency, or the order of accuracy, of the Galerkin equations.
A natural setting in which to illustrate these ideas is provided by the heat
equation
(1) O< x < n, t > O.
241
 242 INITIAL·V ALUÉ ·PROBLEMS CHAP. 7 SEC. 7.1. THE GALE

This parabolic differential equation represents heat conduction in a rod; Notiee that the time variable is still e
u u(i, t) is the temperature at the point x and time t > O, and f(x, t) correctly, Faedo-Galerkin) formulation
is a heat-source term. As in our previous examples, we impose a Dirichlet yields a system of ordinary differential1
boundary condition at x O and a natural condition at x n: which have to be solved numerically. T·
choose a basis rp¡, .•• , rpN for the trial
(2) u(O, t) = ~~(re, t) = O. solution as
N
if(x, t) I;
The first condition means physically that the temperature at the left end of l

the rod is held at O. The Neumann condition, on the other hand, means that
the right end of the rod is insulated; there is no temperature gradient across The optimal weights Qi are determine<
x = n. To complete the statement of the problem we specify the initial
temperature

(3) u(x, O) = u 0 (x), O x n.

Since every vh is a combination of the:
This classical formulation of heat conduction is fraught with difficulties. apply the principie only to the basis ft
For example, (2) and (3) are contradictory if u 0 fails to vanish at x =O, or ordinary differential equations for the
au
if 0 f8x does not vanish at x n. In addition, f could be a point source the boundary conditions are already i
x
which is singular at sorne 0 , and so (1) is not literally true at that point. In initial condition u = u0 is still to be acc
all these cases the underlying physical pro blem still makes sense, to find possibilities. Mathematically, a natural
the temperature distributü;m associated with a given initial temperature u 0 tion ui is the best least-squares approx
and heat so urce f We therefore seek an alternative integral formulation as (ui, vh) (u 0 , vh) for all vh, or in other
in the steady-state case.
Since there is no natural mínimum energy principle-the equation is
not self-adjoint-we turn to the weak formulation: At each time t > O,

(4) J: (ut Uxx- f)v dx =O.

In practice, this means that the i11¡tegn
and the mass matrix on the left has to l
efficient to use the interpolate of u0 as
In the steady state, with ut O and/ f(x), this coincides with the earlier system: ui (u 0 ) 1 • The order of accur
Galerkin formulation. Now we put the Galerkin equatior
To achieve greater symmetry between trial function and test function, (Q P •.• , QN). The striking point is tha1
we integrate -uxxv by parts. The integrated term uxv vanishes naturally at and load vector which appear in the st
x n, and we are led back to the essential condition v(O) = O, in other Q' is the mass matrix M, and the co
words to the space X 1:
MQ'+K
(5) J: (utv + uxvx fv) dx O · for all v in X1.
(8)

The components of the right hand sic

This is the starting point for the finite element approximation. Given were inhomogeneous boundary cond
an N-dimensional subs{mce Sh of X1, the Galerkin principie is to jind a fun'c- effect would also appear in F.
tion uh(x, t) with thefollowing property: At each t > O, uh lies in Sh and satisjies It is natural to ask why finite eleme
tion. This has certainly been attempte
matically, it is perfectly reasonable to
(6)
first to analyze the finite element erro
 CHAP.
represents heat conduction in a rod;
~ point x and time' t >
O, and f(x, t)
vious examples, we impose a Dirichlet
1atural condition at x = n:
~u
r;;(n,t ) =0.
that the temperature át the left end of
ndition, on the other hand, means that
:here is no temperature gradient across
Jf the problem we specify the initial
O< x <n.
conduction is fraught with difficulties.
ctory if u 0 fails to vanish at x = O, or
:n addition, f could be a point source
1) is not literally true at that point. In
al problem still makes sense, to find
!d with a given initial temperature u 0
an alternative integral formulation as
1m energy principle-the equation is
tk formulation: At each time t > O,
- f)v dx =O.
= f(x), this coincides with the earlier
Neen trial function and test function,
!grated term uxv vanishes naturally at
:!ssential condition v(O) = O, in other
·=O for all v in'X1.
finite element approximation. Given
the Galerkin principie is to find a fun'c-
. At each t > O, uh lies in Sh and satisjies
7 SEC. 7.1. THE GALERKIN-CRANK-NICOLSON METHOD 243
No ti ce that the time variable is still continuous: the Galerkin (or, more
correctly, Faedo-Galerkin) formulation is discrete in the space variables and
yields a system of ordinary differential equations in time. lt is these equations
which have to be solved numerically. To make this formulation operational,
choose a basis rpp ... , rpN for the trial space Sh and expand the unknown
s·olution as
The optimal weights Qj are determined by the Galerkin principie (6):
(7) k= 1, ... ,N.
Since ev~ry vh is a combination of these basis functions rpk, it is enough to
apply the principie only to the basis functions. The result is a system of N
ordinary differential equations for the N unknowns Q 1 (t), ... , QN(t), and
the boundary condítions are already incorporated in these equations. The
initial condition u = u 0 is still to be accounted for, and here there are severa}
possibilities. Mathematically, a natural choice of approximate initial condi-
tion u~ is the best least-squares approximation to u 0 ; u~ is in Sh and satisfies
(u~, vh) = (u 0, vh) for all v\ or in other words,
k= 1, ... ,N.
In practice, this means that the integrals on the right have to be computed,
and the mass matrix on the left has to be inverted. Therefore, it is often more
efficient to use the interpolate of u 0 as initial condition in the finite element
system: u~ = (u 0 ) 1 • The order of accuracy is not affected.
Now we put the Galerkin equations (7) into vector notation, with Q =
(Q¡, ... , QN). The striking point is that they involve exactly the same matrices
and load vector which appear in the steady-state problem; the coefficient of
Q' is the mass matrix M, and the coefficient of Q is the stiffness matrix:
(8) MQ' + KQ = F(t).
The components of the right hand side are Fk = f f(x, t)rpk(x) dx. If there
were inhomogeneous boundary conditions, time-dependent or not, their
e:ffect would also appear in F.
It is natural to ask why finite elements are not used also in the time direc-
tion. This has certainly been attempted, but not with great success. Mathe-
matically, it is perfectly reasonable to study the discretization in two steps-
first to analyze the finite element· error u(x, t) - uh(x, t), and then the error
244 INITIAL-VALUE PROBLEMS CHAP. 7 SEC. 7.2. STABILITY AND COl'

incurred In solving the ordiriary differentia.l equations for uh. In the time vari- elimination into LLT, where Lis Chol
able there are no geometrical difficulties to overcome by finite elements, and Q11 + 1 would be computed at each ste1
in fact a straightforward application of the Galerkin principie may couple all
the time levels, and destroy the crucial property of propagation forward in LQn+l/2 M-Kilt Qn + (/
_ 11
"

time. We see no reason to forego the extra flexibility of finite differences. 2

First we must mention the technique of mode superposition, a competitor
of conveb;tional finite differences. The id~a is simply to analyze the initial u0 If the coefficients in the problem are
and the forcing term f into the natural modes of the problem-the eigen- the strict Galerkin theory the matrice
functions uf of Chapter 6. All the computati.onal work is transferred to the time step. It seems very likely that a
eigenproblem. Only the Iower eigenvalues are followed forward in time; necessary, leading to sorne hybrid of ·
Nickell suggests that fewer than 30 modes out of 1000 can produce very good order to produce a stiffness matrix,
results, unless/is unusually rich in the higher harmonics. recomputing every integral. In large
A conventional difference scheme in time, which couples all the modes, to solve exactly for Qn+ 1 ; an iteraH
has to contend with the extreme stiffness of the equation (8); the condition guess) may be more efficient. Dougla
number of M- 1K can easily exceed 1000, so that the modes are decayíng at several iteration techniques for non]
radically different rates. A "trapezoidal rule" scheme (Crank-Nicolson, the solution of a large nonlinear sy
Neumark fi) automatically filters out the useless high modes, if ilt is properly completely justifies these modificatio
chosen. Of course these schemes are implicit, but so is Galerkin's differential The two remaining sections are o
equation; M cannot be inverted in (8) without destroying the band structure. the expected rates of convergence .f<
(The alternative of lumping is discussed at the end of the chapter.) In one For the simple equation ut ux in
respectan implicit equation is not such a serious drawback for parabolic equa- finally occm;s.
tions, such as the heat equation, since the stability requirements on explicit
difference schemes are in any case severe: the time step must be limited to
7.2. STABILITY ANO CONVERGEN<
ilt < Ch 2 or there will be exponential instabilities in the difference equation.
PARABOLIC PROBLEMS
By contrast, an implicit scheme can be unconditionally stable; the size of ilt
is restricted only by the demands of accuracy, and not by stability. This dis-
There are two approaches tothe tt
tinction between implicit and explicit methods is natural for the heat equa-
the other may be more general. In t
tion, where there is an infinite speed of propagation: the temperature Q"+ 1 at
followed forward in time in all th
any point x 0 depends on the previous temperature Q" at all points of the
equation for u, the ordinary different
medium, however small the time step may be. This dependence is reflected in
ence equation for Q11 • lf the coeffic
the fact that M- 1 is not sparse in the Galerkin equation, and that ilt .- h 2
equations are independent of time 1

is required for stability in an explicit difference equation.

approach is extreme!y successful; th
We shall analyze the Crank-Nicolson scheme, which is centered at (n +
on the eigenfunction errors, and t]
i)At and therefore achieves second-order accuracy in time:
mentary. t In the nonstationary cast
Qn+l _ Q11 Q11+l + Q11 ¡11+1 + ¡11 but parabolic equations are so stron~
(9) M ilt +K 2 = 2 . even nonlinear) effects can ultimatel
proach, which is based on energy im
Rewritten, the approximation Q11 + 1 is determined by becomes comparatively simple.
We consider a parabolic equatio
M +KAtQn+l ilt kind of elliptic operator studied in t
2 (m 1 is by far the most common)
In an actual computation, the matrix on the left can be factored by Gauss tThis is also the approach to use in ana
 CHAP.
ntial equations for uh. In the time vari-
::s to overcome by finite e'lements, and
~ the Galerkin principie may couple all
il property of propagation forward in
;!Xtra flexibility of finite differences.
1e of mode superposition, a competitor
idea is simply to an~lyze the initial u0
al modes of the problem~the eigen~
1putational work is transferred to the
·alues are followed forward in time;
des out of 1000 can produce very good
higher harmonics.
in time, which couples all the modes,
ess of the equation (8); the condition
1
00, so that the modes are decaying at
idal rule" scheme (Crank-Nicolson,
1e useless high modes, if ilt is properly
zplicit, but so is Galerkin's differential
~vithout destroying the band structure.
ed at the end of the chapter.) In one
a serious drawback for parabolic equa~
the stability requirements on explicit
·ere: the time step must be Iimited to
·nstabilities in the difference equation.
unconditionally stable; the size of ilt
curacy, and not by stability. This dis-
methods is natural for the heat equa-
·propagation: the temperature Qn+l at
i temperature Qn at all points of the
1ay be. This dependence is reflected in
Galerkin equation, and that ilt h2 <"V
fference equation.
·m scheme, which is centered at (n +
der accuracy in time: ·
¡n+l + ¡n
2 ·-
determined by
on the left can be factored by Gauss
7 SEC. 7.2. STABILITY AND CONVERGii.NCE IN PARABOLIC PROBLEMS 245
elimination into LLT, where Lis Choleskts lower triangular matrix, and then
Qn+ 1 would be computed at each step by two back substitutions:
If the coefficients in the problem are time-dependent (or nonlinear), then in
the strict Galerkin theory the matrices M and K must be recoinputed at each
time step. It seems very likely that another variational crime will be found
necessary, leading to sorne hybrid of finite elements and finite differences-in
order to produce a stiffness matrix which is approximately correct without
recomputing every integral. In large problems it may also be too expensive
to solve exactly for Qn+ 1 ; an iteration (possibly starting with Qn as initial
guess) may be more efficient. Douglas and Dupont [D6, D9] ha ve introduced
severa! iteration techniques for nonlinear problems in order to circumvent
the solution of a large nonlinear system at every time step. Their analysis
completely justifies these modifications of the pure Galerkin method.
The two remaining sections are occupied with verifying the stability and
the expected rates of convergence for parabolic and hyperbolic equations.
For the simple equation ur = ux in the last section, something unexpected
finally occm;s.
7.2. STABILITY ANO CONVERGENCE IN
PARABOLIC PROBLEMS
There are two approaches to the theory. One is more explicit and revealing;
the other may be more general. In the first approach each eigenfunction is
followed forward in time in all three equations-the partial differential
equation for u, the ordinary differential equation for uh, and the finite differ-
ence equation for Qn. If the coefficients and boundary conditions in the
equations are independent of time (the stationary case), then this simple
approach is extremely successful; the previous chapter gave precise bounds
on the eigenfunction errors, and the arguments become completely ele-
mentary.t In the nonstationary case the analysis is much more technical,
but parabolic equations are so strongly dissipative that time-dependent (and
even nonlinear) effects can ultimately be accounted for. In the second ap-
proach, which is based on energy inequalities at each time, this accounting
becomes comparatively simple.
We consider a parabolic equation ut + Lu where L is exactly the
kind of elliptic operator studied in the previous chapters. It is of order 2m
(m = 1 is by far the most common) and its coefficients may depend on the
tThis is also the approach to use in analyzing mode superpositi'on.
246 INITIAL-VALUE PROBLEMS CHAP. 7 SEC. 7.2. STABILITY AND CO

position vectof x. Suppose first that f =

O, and that the initial function u0 or
is expanded into the orthonormal eigenfunctions:
00

L;
l
c.u.(x),
) ) The solution after n steps is Qn
factor J.t1 is
Each eig~nfunction decays at its own rate in time, and the solution evolves
accordirl~ to (12) h 1
j.lj=r

(10)
Since every 11 is nonnegative, it fol
For t > O, this solution lies in the admissible space :JC1. E ven if the initial the Crank-Nicolson scheme is autor
function u 0 is discontinuous, it is easy tb see that as time goes on u becomes pal eigenfunction u1 is governed by
increasingly smooth; the deriva ti ves at any positive time satisfy be compared with the true amplifica·
of the Galerkin equation o ver a sin~

J.l1 ,..., 1 ;.1át +

e-;." 14r ,..., 1 A1 Át +
The exponential term makes this sum finite, and it does the same for the spa-
tial derivatives. These norms are monotonically decreasing in t, in particular
Since J.l1 is the smaller, this compon
in the finite difference equation. Th
the second-order accuracy of the Cr
The fundamental frequency A1 > O gives the rate of decay of the solution. In the finite difference case, ho'
The Galerkin equation MQ' + KQ O is actually a little more stable not decay at fas ter and fas ter rates.
than the equation it approximates. Suppose the initial Q0 is expanded in converges to 1, and the weights ¡
terms of the discrete eigenvectors Qi of M- 1K, or, equivalently, the initial sign at each time step. This does n
ui is expanded in terms of the approximate eigenfunctions u1: a fully implicit difference scheme. TI
and óbviously J.lj O as A1 ~ oo.
J.l1 will become negative and start t<
A1 2/.At. The highest frequency t:
normally exceeds 2/á t. Therefore
Then the solution at a later time t is present only in small amounts) are
moderate frequencies. If this should
bolic problems it is possible to add
Therefore, the rate of decay is A1, which is slightly larger than A1 : The rate of convergence is eas
expansions

(13)
Finally, the Crank-Nicolson scheme is also stable-with no restriction
on the size of .At. The difference operator at each step has the same eigen- There are two sources of error: tl
vectors as M- 1 K, since it is given by Whether ui is computed as a best
eh osen to be its interpolate, the initi
M+ K .AtQn+1 _ M- K .AtQn like any other initial condition it will
2 - 2 error arises when ui is used as initia

= O, and that the initial. function
' uo
:functions:
ate in time, and the ,solutíon evolves
1issible space X1. Even if the initial
o see that as time goes on u becomes
any positive time satisfy
lÍte, and it does the same for the spa-
onically decreasing in t, in particular
es the rate of decay of the solution.
O is actually a little more stable
ppose the initial Q0 is expanded in
.f M- 1K, or, equivalently, the initial
na te eigenfunctions u1:
t is slightly larger than A1 :
~~t) 11 u~ !lo exp ( -1 1 t).
e is also stable-with no restriction
~or at each step has the same eigen-
_M- Kf:..tQn
- 2
SEC. 7.2. STABILITY AND CONVERGENCE IN PARABOLIC PROBLEMS 247
or
The solution after n steps is Qn 2; d/f.l1)nQi, where the amplification
factor fl-1 is
(12)
Since every A1 is nonnegative, it follows that every 1 f.l11 < 1, and therefore
the Crank-Nicolson scheme is automatically stable. The decay of the princi-
pal eigenfunction u1 is governed by its amplification factor Ji~, which is to
be compared with the true amplification (or contraction) factor exp (-A~ f:..t)
of the Galerkin equation o ver a single time step:
f.l1 "' 1 - A1 t:..t +!<A~ t:..t)2 - -HA1 t:..t) 3 • •• ,
e-).htt:.t "' 1 - A1 dt +!(A~ dt) 2 -k{A1 dt) 3 • • • •
Since f.l1 is the smaller, this component of the solution decays slightly fas ter
in the finite difference equation. The discrepancy is of order dt 3 , reflecting
the second-order accuracy of the Crank-Nicolson scheme.
In the finite difference case, however, the high-frequency components do
not decay at faster andfaster rates. As A1 ~ oo, the amplification factor J.t1
converges to 1, and the weights attached to the high frequencies change
sign at each time step. This does not oq::ur in the Galerkin equation or in
a fully implicit difference scheme. The latter would have f.l1 (1 + A1 at)- 1 ,
and obviously f.l1 O as A1 ~ oo. In the Crank-Nicolson case, however,
fl-1 will become negative and start to increase in magnitude at the frequency
A1 2/f:..t. The highest frequency the mesh can hold is A~ "' eh- 2 m, which
normally exceeds 2/f:..t. Therefore the very high frequencies (presumably
present only in small amounts) are actually damped less strongly than the
moderate frequencies. If this should present any difficulty, then as in hyper-
bolic problems it is possible to add a simple dissipation term.
The rate of convergence is easy to determine from the eigenfunction
expansions
(13) u
There are two sources of error: the initial error and the evolution error.
Whether u3 is computed as a best approximation to the true u0 or simply
chosen to be its interpolate, the initial error u0 u3 will be of order hk-and
like any other initial condition it will decay with time like e--<~r. The remaining
error arises when u3 is used as initial condition in both equations. In the true
248 INITIAL-VALUE PROBLEMS CHAP. 7 SEC. 7.2. STABILITY AND CO

equati~m it is expanded into the eigenfunctions up and in Galerkin's equation The error u uh will still be of ord<
it is expanded into the uj-with a different evolution in time. We know from then the decay factors e-uwill no Ion
the previous chapter that at times -r so near to t that their deca
W e try now a different idea-the
mentioned at the beginning of the
[In case lf < 2m, which is not usual in practice, hk should be replaced by rate of change ofthe error u- uh. R
h 2(k-m>.] These estimates show that the difference in the weights is only function expansions back to their
ci- dj = Ju~(ui- u1) dx "' hk. Therefore, comparing u and uh in {13), (or t + lit, in case of the difference
the evolution error is also of order hk. The error in the derivatives will be of at time t. This technique has been pi o
the usual order hk-s, again decaying at the rate e-;. 11 • following early papers by Swartz a1
We emphasize the simplicíty of this technique for deriving error bounds. [P11]; Wheeler, Dendy, and others
It would lead equally easily to the O(!l.t 2 ) errors in the Crank-Nicolson Suppose that the differential equ
process. The simplicity would seem to require that the spatial part of the written in variational form at each 1
problem be self-adjoint, in order to apply the eigenvalue and eigenfunction
estimates of the previous chapter, but this hypothesis is actually inessential.
(15) (u 0 v) + a(u, v) = (.
In fact, there is a simple formula which bypasses the eigenvalue theory entirely (16) (u7, vh) + a(uñ, vh)
and relates the evolution error directly to the basic estimates for steady-state
problems. With the same initial function ui in both equations, the solutions In the stationary case, the energy
differ at time t by time; it arises from (Lv, w). We reta
of the admissible space X1 onto i
(14) a(u Pu, vh) = O for all if.
Splitting the Galerkin error in
Here z is a complex number, uz solves the (non-self-adjoint) steady-state size of u Pu is known from appr
problem (L + z)uz ui, and u~ is its Galerkin approximation. [In effect, inequalities required for the Pu - u
uz and y~ are the Laplace transforms of u(t) and uh(t), respectively, and the identity.
integral (14) inverts the Laplace transform; the contour C runs along two
rays arg z = ±(n/2 + e) in the left half-plane, so that the exponential ezt LEMMA 7.1
produces a converg~nt integral.} With this formula-which reduces to the Jf e(t) = Pu(t) uh(t), then
eigenfunction expansions in the setf-adjoint, discrete spectrum case-the
evolution error at time t follows directly from the steady-state errors of (17) (e0 e) + a(e,
Theorem 2.1. The result is the expected arder hk, even for non-self-adjoint
equations. Proof Since e lies in S\ we ma
To complete the discussion of this first approach, we note that Duhamel's subtract:
principie places the inhomogeneous case O into the same framework. (ut- u~, e)+
According to this principie, the source term f entering at any time 1: acts
like an iriitial condition over the remaining time t - 7:; If the solution u is The second term is just a(e, e), afte
originally connected to u 0 by sorne rule .u(t) = E(t)u 0 , then in the nonhomo- Rearranging, we have
geneous case
u(t) = E(t)u 0 + J: E(t -r)f(-r) dT. (Put- u7, e) + t
It remains to identity the first term a
In the eigenfunction expansions, this becomes
to know that Put is the same as (Pu
the differentiation a¡at. This is true
inner product were to depend on t, ti
 CHAP. 7 SEC. 7.2. STABILITY ANO CONVERGENCE IN PARABOLIC PROBLEMS 249

ctions ui' and in Galerkin's equation The error u - uh will still be of order hk, but if the so urce f acts at all times,
nt evolution in time. We know from then the decay factors e-u will no longer apply; there will be errors committed
at times r so near to t that their decay has not begun.
We try now a different idea-the second approach to parabolic problems
11 u?- Ui llo "' hkljf 2 m.
mentioned at the beginning of the section-to estimate at each time t the
practice, hk should :pe replaced by rate of change of the error u - uh. Rather than connecting u and uh by eigen-
~ difference in the weights is only function expansions back to their data at t = O, the error at time t + dt
'ore, comparing u and uh in (13), (or t + 6.t, in case of the difference equation) is determined from the error
he error in the derivatives will be of at time t. This technique has been pioneered by Douglas and Dupont [06, D9],
he rate e-;.~c. following early papers by Swartz and Wendroff [S11] and Price and Varga
echnique for deriving error bounds. [Pll]; Wheeler, Dendy, and others have made more recent contributions.
6.t 2) errors in the Crank-Nicolson Suppose that the differential equation and its Galerkin approximation are
require that the spatial part of the written in variational form at each time· t,
ly the eigenvalue and eigenfunction
is hypothesis is actually inessential. (15) (uc, v) + a(u, v) = (f, v) for all v in JC1,
'passes the eigenvalue theory entirely (16) (u~, vh) + a(u\ vh) = (f, vh) for all vh in Sh.
> the basic estimates for steady-state
u~ in both equations, the solutions In the stationary case, the energy inner product a(v, w) is independent of
time; it arises from (Lv, w). We retain the notation P for the Ritz projection
of the admissible space JC1 onto its subspace S\ defined as in (6.38) by
r ezt(uz -
e
u~) dz. a(u - Pu, vh) = O for all vh.
Splitting the Galerkin error into u - uh = (u - Pu) + (Pu - uh), the
: the (n<;m-self-adjoint) steady-state size of u - Pu is known from approximation theory (6.39), and the energy
:ialerkin approximation. [In effect, inequalities required for the Pu- uh term are consequences of the following
u(t) and uh(t), respectively, and the identity.
•rm; the con tour C runs along two
!f-plane, so that the exponential ezt LEMMA 7.1
his formula-which reduces to the Jf e(t) = Pu(t) - uh(t), then
ljoint, discrete spectrum case-the •
tly from the steady-state errors of (17)
order hk, e ven for non-self-adjoint
Proof Since e lies in S\ we may put v = e in (15), vh = e in (16), and
.t approach, we note that Duhamel's subtract:
.e f-=!= O into the same framewórk.
(uc- u~, e)+ a(u- u\ e)= O.
term f entering at any time r acts
ting time t - !'; If tlú~ solution u is The second term is just a( e, e), after applying the identity a(u - Pu, e) = O.
J(t) = E(t)u 0 , then in the nonhomo- Rearranging, we have

~(t - r)f(r) dr. (Puc - u~, e) + a(e, e) = (Puc- un e).

;o mes lt remains to identity the first termas the inner product (e0 e), in other words
to know that Puc is the same as (Pu)r: The Ritz projection P commutes with
the differentiation a¡at. This is true only in the stationary case. If the energy
··d dr], .· ~(r) = Jn f(x, r)ui dx. inner product were to depend on t, the term (Puc - (Pu)c, e) would also appear
250. INITIAL-V ALUE PROBLEMS CHAP. 7
SEC. 7.3.

in the identity and would ha ve to be estimated. This difficulty is only technical

According to the theorem just prove
(we om.it the details) if the operator L depends smoothly on time! In the sta-
. suggests that in the energy norm e is ·
tionary case it is obvious that P is independent of time: Differentiating the
and that the Galerkin error u - u11
identity (u Pu~ vh) O, we have (ut (Pu)n vh) = O for all v\ so that
(Pu)r must coincide with Put. This completes the proof ofthe lemma. (21) a(u u\ u- uh) rv a(u Pí

From i!J,1is identity the· rate of change of the error is easy to fin d. The first
This completes the technical ern
term in (17) can be rewritten as
there are no surprises in the results. (
problems, finite elements are particulaJ
(e0 e) :t (ele)= 11 ello :t 11 e !lo· with a large value of h. Iri this situatio
represented by Galerkin's principie, <
The terma(e, e) is at least as large as A1 11 e 115, since A1 is the mínimum ofthe based, than by supposing difference q
Rayleigh quotient. Finally, the right side of the identity is bounded by Because of the integrals to be evaluatt
11 ur - Pur llo 11 ello. Cancelling the common factor 11 e 11 0 , the identity leads to in computing time. Perhaps there will
tion of finite elements and finite diffe1
(18)

7.3. HYPERBOLIC EQUATIONS

Multiplying by e-' 1\ and integrating with respect to -r from Oto t,
It is natural to experiment with fi1
lems. In fact, the Galerkin principie <
the equation M(u) Ois replaced by
The initial error e(O) Pu 0 - u~ is of order Chkll u 0 llk, whether u3 is com- expect finite elements to be tested in
puted by interpolation or least-squares approximation of u0 • The main Sorne preliminary tests are already 1
theorem, giving the correct order of convergence although not necessarily yet in (and may never be, given the.
the most precise estimates, follows immediately from (19). experiments). Maybe all that ca~ be 1

guidelines.
THEOREM 7.1 Mathematically, one property th<
Suppose that Sh is a jinite element space of degree k - l. Then the error is conserved in the true problem, the
in the Galerkin approximation satisjies and that if it is decreasing with time i
in the Galerkin approximation. For
11 u(t) - uh(t) llo 11 u(t)- Pu(t) llo + 11 e(t) llo easy to see. The rate ofchange ofthe e
(20)
Chk[i! u(t) 11 k + e-).¡r 11 Uo llk + J: ét<~-c) 11 ur(-r) llkd-r J. by multiplying the equation by u and

· (up u)+ (Lu, u) ..,e

Thus the error is of the same order hk as in steady-state problems~ and it decays ()

as fast as the fundamental mode if there is no source term.

The equation is conservative-(u, u:
This theorem agrees with the hk estímate derived earlier by means of is identically zero, and it is dissipat
eigenfunction expansions. An averaged error bound in energy follows easily (Lu, u) O for all possible states u. l
by direct integration of the identity (17): tive, since (Lu, u) is in that case a

2 J: a(e(-r), e(-r)) d-r 11 e(O) 115 - 11 e(t) 115 +2 J: 1(Pur Up e) 12 d-r.

.derivatives of u. Hyperbolic equaü

tProvided the test space Vh coincides '

 CHAP. 7
SEC. 7.3. HYPERBOLIC EQUATIONS 251

mated. This difficulty is o,nly technical

:lepends smoothly on time. In the sta-
cpendent of time: Differentiating the
e - (Pu)t, vh) = O fDr aii v\ so that
,]etes the proof of the lemma.
e of the error is easy/t,o fin d. The first
According to the theorem just proved, the right side is of order h2 k. This
suggests that in the energy norm e is negligible in comparison with u- Pu,
and that the Galerkin error u- uh = u- Pu + e satisfies
(21) a(u- u\ u- uh)"' a(u- Pu, u- Pu) < C 2h 2<k--:m> 11 ullf.
This completes the technical error estimates for parabolic problems;
there are no surprises in the results. Our impression is that just as in static

=
a
llello at llello·
problems, finite elements are particularly effective incoarse mesh calculations,
with a large value of h. In this situation the physics is often more adequately
represented by Galerkin's principie, on which the finite element method is
,llell~, since ..t, is theminimumofthe based, than by supposing difference quotients to be close to the derivatives.
side of the identity is bounded by Beca use of the integrals to be evaluated, however, there is a price to be paid
!lon factor 11 e 11 0 , the identity leads to in computing time. Perhaps there will ultimately be a satisfactory combina-
tion of finite elements and finite differences.

h respect to r from O to t, 7.3. HYPERBOLIC EQUATIONS

It is natural to experiment with finite elements also for hyperbolic prob-

e-<~.-¡¡ uc(r) - Puc(r) llo dr.
lems. In fact, the Galerkin principie can be formulated in such generality-
the equation M(u) = Ois replaced by (M(uh), vh) = O for all vh-that we may
· order Chkli U 0 iik, whether u~ is com- expect finite elements to be tested in an increasing variety of applications.
es approximation of u0 • The main Sorne preliminary tests are already going on, but the conclusions are not
onvergence although not necessarily yet in (and may never be, given the a:m,biguous nature of most numerical
1ediately from ( 19).
exper.jments). Maybe all that can be expected is agreement on sorne general
gui'delines.
Mathematically, one property that can be guaranteedt is that if energy
pace of degree k - l. Then the error is conserved in the true problem, then it is conserved in Galerkin's method,
and that if it is decreasing with time in the true problem, then it is decreasing
in the Galerkin approximation. For a first-order system ut + Lu = O this is
1- 11 e( t) 11 o
easy to see. The rate of change ofthe energy (u, u) = f u 2 dx can be computed
-.<¡r 11 Uo llk + J: e-<~<.--t> 11 ukc) llkdr J. by multiplying the. equation by u and integrating:

in steady-state problems, and it decays (ut, u) + (Lu, u) = :t (u2u) + (Lu, u) = O.

is no source term.
The equation is conservative-(u, u) is constant in time-in case (Lu, u)
stimate derived earlier by means of
is identically zero, and it is dissipative-the energy (u, u) is decreasing-if
error bound in energy foiiows easily
(Lu, u) > O for all possible states u. Parabolic equations are strongly dissipa-
):
tive, since (Lu, u) is in that case a positive-definite expression in the mth

e(t) 11~ +2 J: 1(Put - Un e) 12 dr.

derivatives of u. ·Hyperbolic equations are either conservative or at best

tProvided the test space Vh coincides with the trial space Sh.
252 INITIAL-VALUE PROBLEMS CHAP. 7 SEC. 7.3.

weakly dissipative; energy may leak out at the boundaries, but. only very order:
slowly. This makes their analysis much more delicate. In the Galerkm method,
Jet Q denote the projection of 3C 0 onto the súbspace Sh-Qu is the best least- (24) MQ"
squares approximation in Sh to u, where earlier Pu was t.he best ~ppr~xi­
The starting values are approxim
mation in the strain energy norm a(v, v). Then the Galerkm approxtmatwn
displacement uo(x) and initial ve14
uh is determined by projecting the differential equations onto the subspace:
is completely at variance with the
of Q". In that case, with F _ O, th
vector is associated with an expone
Since uh is required to lie in Sh, it is automatic that uh Qu\ and the Galerkin disappear. In the hyperbolic cas<
equation can be written more symmetrically as oscillates rather than decays. The
data, and discontinuities are prop;
u~+ QLQuh O, For the one-dimensional wave
approximation becomes
In other words, the true generator Lis replaced by QLQ. But then a conserva-
tive equation remains conservative, (QLQu, u) (LQu, Qu) = O, and a Q~'+ 1 + 4Q~' + Q~'-
dissipative equation remains dissipative: (QLQu, u) (LQu, q_u) >O. 6
The corresponding nonlinear operators are caUed monotone (Sectwn 2.4),
We remar k again on the coupled fo
and the same result holds.
leads to an implicit difference equ
It is interesting that the conservative property is not always desirable,
Clough and others, and the theo
particularJy in nonlinear hyperbolic equations. The simplest example is the
U.S.-Japan Seminar, that there'"
conservation Iaw ur = (u 2 )x. The solutions of these problems may develop
(through a suitable lumping proe<
spontaneous disco_ntinuities (shocks) and the conservation of energy is lost,
be the case for elements of all de:
even though sorne other conservation laws of mass and moment~m are re-
low degree (generally" piecewise C4
tained. In the Galerkin equation these shocks apparently never qmte appear,
not differentiated with respect to .
and the approximate equation remains conservative-from which it follows
other terms in the equation ar,e tre
that convergence to the correct solution is iinpossible. The st~n~ard .rem~dy
Fujii has also given a valuable
for finite differences is to dissipate energy by means of arttficml vtscosity,
tions (in the time direction) of thc
and apparently that wi11 also be necessary for finite elements. .
example that the terins Q" are
There are two forms in which hyperbolic equations may appear-either
at-2(Qn+1 2Qn + Qn-1). Then,
as a first-order system in time, say wt + Lw = f with a vector unknown. w,
step size ..1t must be restricted <
or as a second-order equation uu + Lu = f We begin with the latter case,
explode exponentially with n. Fo
in which Lis elliptic; a typical example is the wave equation utt - c 2 un = O.
stability conditions are cA.t h/'\1
The weak form of such an equation is just
c..1t h for the lumped mass e<
stability with lumping.) Fujii ha
(22) (um v) + a(u, v) = (f, v) for v in 3C 1, t > O.
and more general hyperbolic inith
In the Galerkin approximation u and vare replaced by uh and vh; this means equations of elasticity.
that uh = I; Q/t)rp/x) is determined by The Galerkin approximation h
of energy (if f = O) and convergenc1
for k= 1,.:., N, t >O. hyperbolic problem we add the ki

This is again an 'ordinary differential equation in the time variable. In this

E(t) = t[
case the same mass and stiffness matrices appear, but the equation is of second In the wave equation this energ~
 CHAP.
ut at the boundar.ies, but only very
wre delicate. In the Galerkin method,
:he súbspace Sh-Qu is the best least-
re earlier Pu was the best approxi-
·). Then the Galerkin approximation
·ential equations ont<? the subspace:
1atic that uh = Qu\ and the Galerkin
cally as
placed by QLQ. But then a conserva-
QLQu, u) (LQu, Qu) O, and a
tive: (QLQu, u)= (LQu, Qu) >O.
s are called monotone (Section 2.4),
ve property is not always desirable,
rations. The simplest example is the
ons of these problems may develop
d the conservation of energy is lost,
1ws of mass and momentum are re-
hocks apparently never quite appear,
conservative-from which it follows
is impossible. The standard remedy
:rgy by means of artificial viscosity,
uy for finite elements.
rbolic equations may appear-either
Lw f with a vector unknown w,
=f. We begin with the latter case,
s the wave equation uu c2 uxx O.
ust
for v in JC1, t >-0.
are replaced by uh and if; this means
for k 1, ... , N, t > O.
quation in the time variable. In this
. appear, but the equation is of second
7 SEC. 7.3. HYPERBOLIC EQUATIONS 253
order:
(24) MQ" + KQ = F(t).
The starting values are approximations from within Sh to the true initial
displacement u 0 (x) ~nd init~al velocity u~(~). The behavior, of the so.lutions
is completely at vanance wrth the parabohc case, where Q appears mstead
of Q". In that case, with F- O, the solution decays very rapidly; each eigen-
vector is associated with an exponential e-).jc, and discontinuities immediately
disappear. In the hyperbolic case the exponent changes to ±iA-1 ~,, a.n~
oscillates rather than decays. The solution is no smoother than Its mrtral
·º
data and discontinuities are propagated indefinitely in time.
For the one-dimensional wave equation and linear elements, Galerkin's
approximation becomes
Q~'+t + 4Q~ + Q~' -¡ ºi-1.
6
We remar k again on the coupled form ofthese equations, which automatically
Ieads to an implicit difference equation. It appears from the experiments of
Clough and others, and the theoretical discussion by Fuji~ at t~e Second
U.S.-Japan Seminar, that there will be no loss in accuracy tf M 1s replaced
(through a suitable lumping process) by a diagonal matrix. This would not
be the case for elements of all degrees; lumping implicitly uses elements of
Iow degree (generally piecewise constants) in dealing with terms which are
not differentiated with respect to x, and .this loses overall accuracy once the
other. terms in the equation are treated with high accuracy. .
Fujii has also given a valuable stability analysis for difference approxima-
tions (in the time direction) of the finite erement equation (24). Suppose for
example that the terms Q are replaced by centered second differences
11
6.r2(Qn+t - 2Qn Qn-t). Then, as is well known for fini~e di~erences, t~e
step size ó.t must be restricted or the computed approxrmatrons Qn wtll
explode exponentially with n. For the one-dimensional wave equation, his
stability conditions are có.t h/,.JT for the consistent mass matrix M, and
có.t< h for the lumped mass case. (Tong [TS] also observed the added
stability with lumping.) Fujii has studied other finite difference ·schemes
and more general hyperbolic initial-boundary-value problems,. including the
equations of elasticity. . .
The Galerkin approximation has two important properttes: ~onservatwn
of energy (if f O) and convergence. To measure the energy in a second-order
hyperbo1ic problem we add the kinetic and potential energies:
E(t) = j'[(un ut) + a(u, u)].
In the wave equation this energy is -!J (u'f c2 u;) dx. The quantity E is
254 INITIAL-VALUE PROBLEMS CHAP. 7 SEC. 7.3.

independent of time, since with v = U1 in the weak form (22), Evidently the equation is again im1
problems. The truncation error is 1
(25) vergence for linear elements.
Dupont [DIO] happened to b
For the wave equation, this becomes simply convergence for cubic trial function
not appear. lnstead, the error u - u
J (U 1Uu + C
2
UxUx 1) dx = JU1(Uu - c 2uxx) dx = O.
of magnitude larger than the best
culations, which were greeted with
were carried out for the Hermite c1
Conservation of energy in the Galerkin equation can be verified in the same
node. For cubic splines his comput
way:
of convergence does not depend e
(u'¡¡, u~) + a(u\ u~) O. polynomials, and in fact, the large
approximation than its subspace (
order. ·
Thus the approximation, like the true equation, is only neutrally stable.
We propose to compute the tn
We shall sketch the proof of convergence, which follows from an identity
the true solution of U 1 + Lu = O in
analogous to Lemma 7.1 : With e Pu - u\
O. In our case L = -a¡ax, and Q
(26) The truncation error is

The left side is the derivative of the energy E(t, e) in the quantity e. This Lu- QLQu = (I
expression is not quite conserved, but the right side is less than
This first term on the right is the
Lu = -ux. If Sh is of degree k - 1
usual order hk. It is the other te1
Thus E' Chk ,Jlf. Integrating from O to t, L(l Q)u is the derivative of the
:JC 1 nÓrm and cannot be better, tha11
term L(I Q)u is annihilated by ti
does happen for linear elements e
The initial error E 0 will be of order h 2 <k-l) and so will the energy in u - Pu. cubics. Since such a cancellation m1
rate of convergence in a first-order
u - Pu + e is of the
i
Therefore, the energy in the Galerkin error u - uh
optimal order h 2 <k-u. Provided the initial data are smooth, this continues to
hk. This rate of convergence has be
of hyperbolic systems.
hold even for large times, t "' Ifh.
To understand how this cancel
We turn now to the trivial but interesting example U 1 ux. The equation
to the polynomial of lowest degree ,
itself is certainly not very exciting; it describes a wave traveling to the left
space-x 2 for linear elements and J
with unit.velocity, u(x, t) = Uo(X + t). There is no distortion in the wave, and
J:=u 2
dx is obviousfy conserved; this is the energy in a first-order problem.
is the error function illustrated in
multiple of the sawtooth function; 1
The Galerkin approximation at each time is (u~, vh) =(u~, vh), and it, too, over each subinterval. The best ca
conserves energy. With linear elements uh(t, x) = ~ u/t)rp/x), where rpi is of such a function is identically z¡
the roof function centered at the node jh, this approximation becomes has occurred. In the cubic case, L(J
and the ñnal projection Q leaves it
(27) ' u~+l + 4u~ + u~_ 1 _ uj+ 1 ui- 1 the Galerkin error is of order 11 u
6 - 2h .
In one sense the exponent k
 CHAP. 7' SEC. 7.3. HYPERBOLIC EQUATIONS 255

n the weak form (22), Evidently the equation is again implícit-a serious drawback for hyperbolic
problems. The truncation error is found to be O(h 2 ), the usual rate of con-
vergence for linear elements.
Dupont [D 1O] happened to be computing the corresponding rate of
nply convergence for cubic trial functions, and the expected power h4 simply would
not appear. Instead, the error u- uh turned out to be O(h 3 ), which is an order
f urCuu - c 2 uxJ dx ~ O. of magnitude larger than the best approximation to u by a cubic. His cal-
culations, which were greeted with surprise and perhaps even sorne disbelief,
were carried out for the Hermite cubics with u and ux as unknowns at each
equation can be verified in the same
node. For cubic splines his computations did give O(h 4 ). Therefore, the rate
of convergence does not depend only on the ciegree of the finite element
fEh
dt -_ (l{m
.. U h)
1
+ a(Uh, U h) _
. 1 -
O• polynomials, and in fact, the larger space of Hermite cubics gives a worse
approximation than its subspace of cubic splines. Sorne explanation is in
order.
equation, is only neutrally stable.
We propase to compute the truncation error in general, by substituting
ence, which follows from an identity
- uh, the true solution of u1 + Lu =O into the Galerkin equation u;+ QLQuh =
O. In our case L = -ajax, and Q is the projection onto the subspace Sh.
The truncation error is

tergy E(t, e) in the quantity e. This
.e right side is less than
to t,
- > and so will the energy in u - Pu.
1 cubics. Since such a cancellation must be regarded as exceptional, the normal
-ror u - uh = u - Pu + e is of the
.1 data are smooth, this continues to
sting example U¡ = u x· The equation
escribes a wave traveling to the left
here is no distortion in the wave, and
the energy in a first-order problem.
me is (u;, vh) = (u~, vh), and it, too,
uh(t, x) = :E u/t)rp/x), where rp1 is
h, this approximation becomes
_U¡+ 1 -u1 _ 1
- 2h .
Lu- QLQu = (1- Q)Lu + QL(l- Q)u.
This first term on the right is the error in least-squares approximation of
Lu = -ux. If Sh is of degree k - 1 and u is smooth, then this error is of the
usual order hk. It is the other term QL(I - Q)u which is decisive. Since
L(I- Q)u is the derivative of the least-squares error, it is an error in the
JC 1 norm and cannot be better than hk- 1 • The question is whether or not this
term L(I - Q)u is annihilated by the final projection Q; we believe that this
does happen for linear elem~nts on a regular mesh but not for the usual
rate of convergence in a first-order hyperbolic system will be hk- 1 rather than
hk. This rate of convergence has been established by Lesaint for a wide class
of hyperbolic systems.
To understand how this cancellation could occur, we apply QL(I- Q)
to the polynomial of lowest degree which is not identically present in the sub-
space..:._x2 for linear elements and x 4 for cubics. In the linear case (J- Q)x 2
is the error function illustrated in Section 3.2 (Fig. 3.3). Its derivative is a
multiple of the sawtooth function; L(I - Q)x 2 goes linearly from + 1 to -1
over each subinterval. The best continuous piecewise linear approximation
of such a function is identically zero: QL(l - Q)x 2 = O, and cancellation
has occurred. In the cubic case, L(l- Q)x 4 happens to be a Hermite cubic,
and the final projection Q leaves it unchanged; there is no cancellation, and
the Galerkin error is of order 11 u- uh llo ,..., h3 •
In one sense the exponent k...:.__ 1 might have been anticipated. Ifthe wave
256 INITIAL-VALUE PROBLEMS CHAP. 7

equation u,, c2 uxx is reduced to a first-order system, the vector unknown

is made up of first derivatives u, and cux:

Thereforé, the ordinary energy llu, 115 + 11 cux 115 in the vector unknown is
precisely twice the energy E(t); it comes from the same sum of the kinetic
and potential energies. Since the error in this energy was of order h 2 <k-I)
for the single equation, the exponent k
for a system.
1 is exactly what we should expect

From a practica! point ofview, these error bounds as h ---lo O are subordi-
8 SINGULARITIES

nate to the problem of obtaining reasonable accuracy af reasonable expense.

With hyperbolic equations, we are not sure that this is achieved most effec-
tively by finite elements. The finite speed of propagation in the true solution
means that explicit finite difference equations are possible, with time steps
ilt of the same order as h, and it is known how artificial viscosity can be
introduced to promote stability. For finite elements the difference equations 8.1. CORNERS ANO INTERFACE
will be implicit and almost too conservative. (They can be explicit only if
we lump the mass matrix, or if, as Raviart has proposed, we choose the nodes Perhaps the most characteri:
as the evaluation points e!¡ for numerical integration. By conserving mass in problems like
the lumping process-or by using lower degree trial polynomials as in [T8]
for the element mass matrices-we achieve a consistent difference equation, (1) -V ·(p Vu) + qu
with a typical Courant condition for numerical stability. Stability is not un-
is that the solution u is smooth as
conditional, as it was for tlie implicit pure Galerkin processes described ear-
lier in the chapter.)
f are smooth. In fact, Weyl's femn
The one important advantage of finite elements in hyperbolic problems-
nl ofQprovidedp,q, andfareai
"up to the boundary" provided th
which must somehow be imitated-by finite difference schemes in the future-
Singularities can therefore occ
is the systematic achievement of high accuracy, even at curved boundaries.
of the data: is not smooth. Unfortt
in fracture mechanics problems,
completely unsatisfactory to proct
As with difference approximation
we ha ve discussed in earlier chapte
for dealing with singularities. HO\
ture of singularities which arise in
the variational method invites one
kin approximation. This chapter
establishing the analytical form oJ
Starting first with the case of n<
equation

(2) -A.u fir

 CHAP.7

t-order system, the vector' unknown

'::) ( Ur)
) CUX X

- 11 cux 115 in the vectdr unknown is

s from the same sum ·of the kinetic
in this energy was of order h2<k-u
- 1 is exactly what we should expect

error bounds as h _____, O are subordi-

:tble accuracy af reasonable expense.
;ure that this is achieved most effec-
l of propagation in the true solution
ations are possible, with time steps
.own how artificial viscosity can be
.te elements the difference equations
ative. (They can be explicit only if
t has proposed, we choose the nodes
integration. By conserving mass in
degree trial polynomials as in [T8]
:ve a consistent difference equation,
nerical stability. Stability is not un-
re Galerkin processes described ear-
! elements in hyperbolic problems-
te difference schemes in the future-
:curacy, even at curved boundaries.
8 SINGULARITIES
8.1. CORNERS ANO INTERFACES
Perhaps the most characteristic property of elliptic boundary-value
problems like
(1) -V·(pVu)+qu finQ, u o on r,
is that the solution u is smooth as long a~ the boundary r and data p, q, and
f are.smooth. In fact, Weyl's lemma states that u is analytic in any subregion
Q 1 of Q provided p, q, andf are analytic in Q 1 • Similar conclusions are valid
"up to the boundary" provided the boundary of Q itself is analytic.
Singularities can therefore occur only when the boundary or sorne part
of the data is not smooth. Unfortunately, these cases often arise, for example
in fracture mechanics problems, and in the presence of singularities it is
completely unsatisfactory to proceed with finite elements on a regular mesh.
As with difference approximations, local mesh refinement in the sense that
we ha ve discussed in earlier chapters has been a popular and effective method
for dealing with singularities. However, a great deal is known about the na-
'ture of singularities which arise in elliptic problems, and the special form of
the variational method invites one to use this information in the Ritz-Galer-
kin approximation. This chapter is devoted to this task, and we begin by
establishing the analytical form of the singularities which can arise.
Starting first with the case of nonsmooth boundaries, consider the La place
equation

(2) -Au fin Q, u= o on r

257
 258 SINO ULARITIES CHAP. 8 SEC. 8.1.

defined in a region n which has a corner (Fig. 8.1). To fix ideas 1et us assome use the orthogonality properties
that f is analytic in the el o sed region ñ and that r
is analytic except at P.
Then Weyl's lemma states that u is analytic in n except at P, and we seek a

Then (5) becomes, after the first te

(8)
f ro
0
dr[!!_(rdui)- .
dr dr

fir) J:n frpiO).

Since this holds for a11 t¡J, the expres
tutes rhe basic differential equatio
its general solution is
Fig. 8.1 A domain with comer angle rtn.

description of u near that point. In particular, we shall study the behavior of

u in the sector where we agree to replace

(3) (10)

where (r, 0) denotes polar coordinates at P. The weak form of (2) is whenever vj is of the form (/ 2) 2 +
set pi = O, since otherwise that ter
(4) for all v in X 1(0). rti is chosen so that at r r0 , (9) is
Altogether, the solution neaL the 1

If vis chosen to vanish near the corner and also outside the sector 0 0, this
2: rtirvjrp¡(O) + 2
00 "'

(11) u(r, 0)
reduces to j= 1 j=

(5) O= J·
Clo
[Vu·Vv -
f V}- f'"o r dr f"'lt [au av . -2au av
. o ar ar + r a(} a(} -
¡;
JV
Jd(} . Suppose for the moment that 11
(11) that the leading term in the si

Now, since u is analytic away from the corner and u(r, O) = u(r, rtn) = O,
u may be expanded for each fixed r > O into
00
Observe tbat this singularity becc
(6) u(r, O) 2: uir)rpiO),
j=t
increases, and if tbe corner at P is
derivatives of u are unbounded. Th
where ning into it; tbis is one of the prc
the last section. Around tbe point.
(7)
is a full interior angle of 2n, and ti
We can easily determine tbe d
The Fourier coefficients uir) are determined by (5): Let v = t¡~(r)rp/0), and singularity. By direct calculation,
 CHAP. 8 SEC. 8.1. CORNERS AND INTERFACES 259

r (Fig. 8.1).'To fix ideas Jet us assome use the orthogona1ity properties
l and that r is analytic except at P.
lytic in n except at P, and we seek a ,

Then (5) becomes, after the first term is integrated by parts in r,

(8)

Since this holds for all t¡J, the expression in brackets must vanish. This consti-
tutes the basic differentia1 equation for u/r). Expanding f/r) into ~hi,
its general solution is
th corner angle rx,Tl.

(9)
icular, we shall study the behavior of
where we agree to replace

),O<e<an} n, (10)

Lt P. The weak form of (2) is whenever v;

is of the form (l + 2) 2 • In addition, we reject the term r-v¡ and
set pj = O, since otherwise that term has infinite energy. The other constant
ex j is chosen so that at r r 0 , (9) is the correct Fourier coefficient of u(r 0 , e).
for all v in X1(Q).
Altogether, the solution near the corner takes the form (see Lehman [L2])

and also outside the sector no, this

(11)
00

u(r, e) = ~ ajr• 1 rp/e)

= 00

+ ~ ~ fj 1[(1 + 2)2 - vy]- 1 qJ/e)r+ 2 •

j=l j=ll=O

Suppose for the moment that 1/a is not an integer. Then it follows from
'r Jocn [au av + r_ 2 au av fv] de.
(1 1) that the leading term in the singularity of u is
o arar ae ae

e corner and u(r, O) = u(r, an) o, , •

r ¡ ,ocsm-·
e
) into a

Observe that this singularity becomes more pronounced as the angle an

u/r)rp/e), increases, and if the corner at P is not convex, that is, a > 1, even the first
deriva ti ves of u are unbounded. The worst case is a region with a crack run-
ning into it; this is one of the problems which we examine numerically in
the last section. Arotmd the point P at the head of the crack (Fig. 8.3) there
is a full interior angle of 2n, and the solution behaves like r 11z sin e¡2.
We can easily determine the degree of smoothness of u near any such
nined by (5): Let v = t¡J(r)rp/e), and singularity. By direct calculation, the function rv sin ve falls just short of
260 SlNGULARITIES CHAP. 8 SEC. 8.1.

possessing 1 + v derivatives in the mean-square sense (and only v in the piecewise constant:
pointwise sense-here we absolutely need the mean-square approximation
theory to predict any reasonable convergence). Thus for any [3 < v 1
the solution has 1 + [3 fractional derivatives. At a reentrant corner, where
1/oc, (13) p={
n is not convex and oc > I, it follows that u líes in 3C 1 but not in 3C 2 • Around The classical formulation of the pr
a crack u l~ not quite in 3C 3 12 •
equation (12) hold separately in nl a
When '1¡rx is an integer the first sum iQ (11) is analytic. However, except
ous across the interface r (v is the r.
for the case a I in which r .is a straight line near P, we cannot conclude
that u is analytic in 0 0 ; logarithms gene rally appear in the second sum. For
example, if rx = f, so that r makes a right angle at P, then v] (1 + 2)2 = 4 (14) au¡
Pt uv r-
in case j = 1 and 1 = O. This is an instance in which the replacement (1 O)
is required, and a term (f1 0 r 2 In r sin 2B)/6 appears in the sol ution u. Observe As we noted in Section I .3, the weal
that this term is in 3C 2 but it is not in :JC,l for any l > 2.
Simiiar. arguments give the behavior of the solution to other problems
near corners. In the variable coefficient problem (1) a more complicated
calculation shows that the singularity is still ofthe form (l 1). More generally,
the singularity in a '2mth-order problem is determined by the principal part and (14) is a natural boundary conditJ
of the operator. The leading term in u is of the form r;.rp;.(B), where 'PA is a tri al functions in the Ritz method ..Ir
smooth function of () and A. is an eigenvalue of an auxiliary problem-:-both at a corner of r.
of which depend on the boundary conditions. We refer the reader to Kel- Except when the interface is a
logg's work [K2], which also includes the three-dimensional case, and to the the solution u will be singular at the ¡
Russian work [K3], as well as to standard engineering references [18, P6, tion analogous to (11) for the beh
W4, H5]. Birkhoff [Bl3] and Kellogg [Kl], we
We now permit the second type of sitigularity, when the boundary is system
smooth yet one or more ofthe data is not smooth-. Such a singularity typically
, (15) d ( drp) lprp, p
arises in interface problems, anda simple example is provided by d() p d()
i

(12) -V·(p Vu) fin n, u= o on r, The eigenfunctions rp(B) are require(

to satisfy the interface conditions
where Q is the region shown in Fig. 8.2. We take the coefficient p to be
(16) ~¡~ [Pt ~~(B)]
(17) ~i~ [P ~~(rxn- B)]
1

There is an infinite sequence of posi

ciated eigenfunctions rpiB) are ortho
2n:

J
o p(B)rp~(B)rp~(B) dO = VJ

Fig. 8.2 An angular interface. For each fixed r > O, the solutior

l
 CHAP. 8 SEC. 8.1. CORNERS AND INTERFACES 261

an-square sense (and only v in the piecewise constant:

:ed the mean-square approximation
gence). Thus for any fJ < v 1 = 1/tL, (13)
1tives. At a reentrant corner, where
at u lies in JC 1 but not in JC 2 • Around
The classical formulation of the problem is to require that the differential
equation (12) hold separately in 0 1 and 0 2 , with u and p aujav being continu-
in (1 1) is analytic. However, except
ous across the interface r (vis the normal to r). Thus, referring to Fig. 8.2,
ght line near P, we cannot conclude
?rally appear in the second sum. Por
ht angle at P, then v; (! + 2)2 = 4 (14)
tan ce in which the replacement (1 O)
•/6 appears in the solution u. Observe As we noted in Section 1.3, the weak form of this equation is still
1
for any l > 2.
r of the solution to other problems
for all v in X1,
rt problem (l) a more complicated
;till ofthe form (11). More generally,
is determined by the principal part and (14) is a natural boundary condition. It will not have to be satisfied by the
is of the form r;.rp;.(fJ), where rp;. is a trial functions in the Ritz method. Indeed, it would be very -difficult to satisfy
ralue of an auxiliary problem-:-both at a corner of r.
litions. We refer the reader to Kel- Except when the interface is a straight line (ex = 1) or when p 1 p 2 ,
e three-dimensional case, and to the the solution u will be singular at the point P in Pig. 8.2, and we seek a descrip-
iard engineering references [18, P6, tion analogous to (11) for the behavior of u near this point. Following
Birkhoff [B13] and Kellogg [Kl], we introduce the periodic Sturm-Liouville
' singularity, when the boundary is system
: smooth-. Such a singularity typically P1 if O< fJ < cxn,
e example is provided by (15) ', d ( drp) lpqJ, p p(fJ)
dfJ p dfJ {p i f cxn < fJ < 2n.
2

u= o on r, The eigenfunctions rp(fJ) are required to be periodic, rp(fJ) rp(fJ + 2n), and
to satisfy the interface conditions
.2. We take the coefficient 1? to be
(16) Wt¿ [Pt ~;({})] ~¡~ [Pz~(-fJ)}
(17) ~~ [Pt ~;(cxn- fJ)J = ~i~ [Pz ~;(cxn + fJ)J.
There is an infinite sequence of positive eigenvalues }.,i v;, and the asso-
ciated eigenfunctions rp ifJ) are orthogonal:

2n J2lt
S p(fJ)rpj(fJ)rp~(fJ) dfJ = VJ p(fJ)rp/fJ)rp¡{fJ) dfJ
0 0

lar interface. Por each fixed r> O, the solution u(r, fJ) satisfies the jump conditions
262 SINGULARITIES CHAP. 8 SEC. 8.2.

( 16)-(17), and hence fied:

2>t
(18) u(r, 8) 2: uir)rpi8),
}= 1
uir) = f 0
p(8)u(r, 8)rpi8) d8.
The eigenvalue v is determined by
We now proceed in exactly the same way as in the case of boundary singular- shown that either tan vn/4 O, t
it~s. Subi~·ituting (18) into the differential equation (12) and using orthogo- v 1 is the smallest positive root of
nality, we bbtain a differential equation for the Fourier coefficients uir).
These equations can be solved exactly and we obtain an expression Iike (11) (20)
for u, except that in this case the exponents {vj} are the square roots of the
eigenvalues of (15) and the {rp1} are the associated eigenfunctions.
Similar formulas are valid for the
The dominant singularity in th
Ps Q?Y
r-------------~,--------~----,
P4=(t, t) v 1 líes between O and 2. Note th•
Otherwise u lies only in the fractio
1
1 it is not in 3C 2 when v 1 < l.
le= 37r
: 2 There is no difficulty in extend
1 number of interfaces meet at the pe
1
1 several discontinuities, with a jum:
Crack i Kellogg has observed that ata cross
P1t---------\--l-____.. P____ --_:::._e. _1T.;.__ .____ o2 = ( t ,o J
11
chosen coefficients p 1 in each of th1
TR2 =(0- .l) can be arbitrarily smalL Therefore,
1
1 24
1 sorne extra attention to the trial
1
1 the finite element rnethod will giv€
1
le 1T
1 2
1
1 8.2. SINGULAR FUNCTIONS '.
1
1

l The expansions given in Sectio

ment spaces which will improve 1
Suppose that we can construct ind
Fig. 8.3 A domain with a crack-an interna! angle of 2n. u 2: c1r¡J1 is smooth, say in JeA
e~' ... , cs. Then why, not add r¡J 1 ,
For this simple problem one can obtain exact formulas for the eigen-
The idea is obviously to let the sin¡
functions and eigenvalues-in a more complicated problem numerical
near the singularity, with the conve
techniques will be required-and for illustration we consider the specific
elsewhere. As a result, it is necess
case rt t. The eigenfunctions of (15) fall into two symmetry groups: those locally near each singularity. We tl
which are symmetric about 8 O and those which are antisymmetric. In
interfaces,
the former case the eigenfunctions have the form
r"1 sin

(19)
cos v8 for 181 < n/4, (21) tflir, 8)
{
pir) s
rp.(8) { rt. cos v(n - 8) for 181 > n/4. o
The constant a. is chosen so that the interface conditions (16)-(17) are satis- The transition points r 0 and r1 afi
 CHAP.
= J:n p(fl) u(r, fl) rp~(fl) d().
as in the case of bouQdary singular-
11 equation (12) and ~·sing orthogo-
l for the Fourier coefficients u/r).
d we obtain an expression like (11)
:nts {v 1} are the square roots of the
;sociated eigenfunctions.
8=1r
--------
;-an internal angle of 2n.
'tain exact formulas· for the eigen~
~ complicated problem numerical
lustration we consider the specific
Lll into two symmetry-groups: those
those which are antisymmetric. In
the form
for ()! < n/4,
1
for 1 fll > n/4.
~rface conditions (16)-(17) are satis-
8 SEC. 8.2. SINGULAR FUNCTIONS 263
fied:
av=-p¡Slll4
[ . 3vn]
. vn]/[ P2Slll4·
The eigenvalue vis determined by substituting (19) into (15), and it can be
shown that either tan vn/4 O, that is, v 4n, or v = ± 4n, where
v 1 is the smallest positive root of
(20) tan2 v:J/[ 1 3 tan 2 vn] =
4
Similar formulas are valid for the odd symmetry class.
The dominant singularity in the solution is therefore r"rpv(O), where v
v_1 lies between O and 2. Note that v 1 = 1 in (20) if and only if p 1 =-p 2 •
Otherwise u líes only in the fractional spaces :Jel+P, p < Vp and in particular
it is not in 3C 2 when v1 < l.
There is no difficulty in extending this analysis to the case in which any
number of interfaces meet at the point P. The coefficient p(O) in (15) will ha ve
several discontinuities, with a jump condition of the form (16) at each one.
Kellogg has observed that ata crossing oftwo straight interfaces, with suitably
chosen coefficients pi in each of the four quadrants, the leading exponent v1
can be arbitrarily small. Therefore, the singularity can be very severe; witho.ut
sorne extra attention to the trial functions at all such exceptional points,
the finite element method will give disappointing results.
8.2. SINGULAR FUNCTIONS
The expansions given in Section 8.1 suggest a modification of finite ele-
ment spaces which will improve the approximation of singular solutions.
Suppose that we can construct independent functions t¡/p ... , tfls such that_
u - L; C¡t¡/¡ is smooth, say in X\ for suitable (but unknown) constants
e~' ... , cs. Then why not add t¡/p . . . , tfls to the finite element space Sh7
The idea is obviously to let the singular functions t¡lp ... , tfls approximate u
near the singularity, with the conventional finite elements carrying the burden
elsewhere. As a result, it is necessary to define the singular functions only
locally near each singularity. We therefore choose, both for corners and for
interfaces,
forO r < r0,
(21) for r 0 < r < r P
for r 1 < r. ·
The transition points r 0 and r 1 are fixed (independent of h), and the poly-
264 SINGULARITIES CHAP. 8 SEC. 8.2.

nomials p1 are chosen so as to mergethe coefficient rv1 smoothly into zero. by high-order numerical iiltegratior,
Jf we' want the trial functions 'flt to lie in ek-l (and therefore in Jek), then the is totally inappropriate.
polynomials are of degree 2k 1 and are determined by the Hermite con- The inversion of the stiffness
ditions addition of singular functions destt
may lead to extra operations in elin
(22) d (r)
rVj] 1r = r o_ - drzP¡ =0, l O,l, ... ,k-l.t requirements. In addition, the cor
much in question. The singular ft
by the other finite elements, and he1
For example, suppose that we want to use cubic elements (k 4) to
dependent."
solve Laplace's equation in a region with a crack. The exponents are v1 =
In practice, both difficulties c<;
j !, and according to (ll) there exist constants IXp ••• , 1X 4 such that unknowns. Let rpp ... , rpN be aba

u-
1~ IX}/f¡. = IX 4 r
712
sin J+ · ·· +
7
analytic terms.
functions. Then·we order the unkno
appear last; that is, the vector of m
The stiffness matrix for S~ is thereJ
The left side belongs to 3C 4 , and therefore it can be approximated to the opti-
mal order h4 by cubic elements. The best possible degree of approximation
has been recovered by the inclusion of three singular functions '!Ir
More generally, suppose that we have constructed t¡/p ... , 'lis such that
u - 2: C;'f/; is in Jek for appropriate constants e 1 , ••• , cs. In addition, let where K 11 is the standard stiffness 1
Sh be a standard finite element space of degree k :_ l. Then if S~ is spanned the singular functions. The entri(
by Sh together with r¡Jp • •• , 'f!s, there is an approximation Uh in S~ to the a(r¡¡¡, r¡¡ J
singular solutiorru with error Faddeev and Faddeeva [6] the
matrices, which we write in block:
(23)

This is clear: Since u- 2: C;'f/; is in 3C\ we can approximate it by its interpo-:-

late vh in the finite element space Sh. Then Uh = vh + 2: C;'t/f;·
Therefore,
This extra accuracy in approximation, obtained by adding singular func-
tions to the trial space, will be reflected in extra accuracy ofthe Ritz-Galerkin
approximation uh. We shall postpone the analysis until the next section and
concentrate here ón the computational problems that accompany singular Obviously L 11 and U11 are the fa(
functions. sociated with Sh. As a consequence
Two difficulties arise in applying these ideas to an actual computation, K 11 'and the much smaller matrices
e ven after the forros of the singularities are known. The first is the evaluation in the symmetric case.) The addi
of inner products involving singular functions, and the second is the inversion which is orders of magnitude less
of the stiffness matrix. For the former there are a variety of tricks [F7, F9] addition, the factorization and th
which exploit the special forro of the singular functions. General.ly speaking, by back substitution represent only
the most singular part-the radial dependence in the energy integrals a(r¡¡1, r¡¡1 ) width of K 11 • This is the same as
-must be done analytically. The integration in(}, and if absolutely necessary the bulk of the work is in the facto
also the integrals a(rp¡, r¡¡1) involving only one singular function, can be done ities are isolated in the smaller m~
the rounding errors. Typically the ~
tA possible alternative is to multiply the singular function rv sin v9 direct1y by a
polynomial q(r) which merges smoothly into zero. This eliminates the condition at r 0
tion ofX22 L 21 U 12 , and it is oft1
in favor of qm(o) = () 01 • in high precision.
 CHAP.
coefficient rvj smoothly jnto zero.
~k-t (and therefore in :Jek), then the
·e determined by the Hermite con-
o, o, 1'. . 'k l.t
to use cubic elements (k = 4) to
1 a crack. The exponents are v1 =
mstants IX¡, .•. , tX 4 such that ·
. · . + analyti~ terms.
it can be approximated to the opti-
- possible degree of approximation
.ree singular functions 'flr
: constructed IJ/ 1 , • • • , 'fls such that
1stants e~' ... ,cs. In addition, let
legree k · l. Then if S~ is spanned
an approximation Uh in S~ to the
·e can approximate it by its interpo-
:n Uh = vh + 1: C¡'fl¡·
, o btained by adding singular func-
~xtra accuracy-ofthe Ritz-Galerkin
analysis until the next section and
Jroblems that accompany singular
;e ideas to an actual computation,
e known. The first is the evaluation
ons, and the second is the inversion
tere are a variety of fricks [F7, F9]
ular functions. General.ly speaking,
~nce in the energy integrals a('f/¡, 'f/¡)
:on in 9, and if absolutely necessary
one singular function, can be done
;íngular function rv sin v(J directly by a
ero. Thís eliminates the condition at r 0
8 SEC. 8.2. SINGULAR FUNCTIONS 265
by high-order numerical integration. A fixed quadrature -rule of low accuracy
is totally inappropriate ..
The inversion of the stiffness matrix is a more serious problem. The
addition of singular functions destroys the band structure of the matrix and
may lead to extra operations in elimination ("fill in") as well as extra storage
requirements. In addition, the conditioning of the stiffness matrix is very
much in question. The singular functions 1{1¡, ••• , 1{1 8 can be approximated
by the other finite elements, and hence the basis for S~ will be "nearly linearly
dependent."
In practice, both difficulties can be avoided by correctly ordering the
unknowns. Let (/Jp .•. , rpN be a basis for Sh and 'fl P ••• , 'fls be the singular
functions. Thenwe order the unknowns so that the components of 'f/ 1 , ••• , 'fl s.
appear last; that is, the vector of unknowns is (Q 1 • • • QNP1 ••• Ps) = (QP).
The stiffness matrix for S~ is therefore
where K 11 is the standard stiffness matrix for Sh and the other blocks involve
the singular functions. The entries of K 22 are the energy inner products
a('f/¡, 'fl).
Faddeev and Faddeeva [6] then factor Ks into a product of triangular
matrices, which we write in block forro as
L u
Therefore,
Obviously L 11 and U 1 1 are the factors of the usual stiffness matrix K 11 as-
sociated with Sh. As a consequence, it is necessary only to store the bands ·of
K 11 and the much smaller matrices K 12 , K 2 p and K 22 • (Note that K 12 = KI1
in the symmetric case.) The additional storage required is only sN + s 2 ,
which is orders of magnitude less than the storage requirement for K 11 • In
addition, the factorization and the calculation of the unknowns Q and P
by back substitution Fepresent only O(w 2 N) operations, where w is the band-
width of K 11 • This is the same as is required to sol ve K 11 Q = F 1 • In fact,
the bulk of the work is in the factorization.. The effects of numerical instabil-
ities are isolated in the smaller matrices, and it is relatively easy to control
the rounding errors. Typically the only real so urce of trouble is in the forma-
tion of K 22 L 21 U 12 , and it is often desirable to do the latter multiplication
in high precision.
 . :.·
.~l
·':'.

SEC. 8.3. ERRO

266 SINGULARITIES CHAP. 8

.)
Suppose that we attempt to ca
These ideas can also be used for eigenvalue problems. The bisection and ! and 3.4, to estimate the error in d
(block) inverse power method require factorizations of the mass and stiff-
As before, we take u - ti' to be t1
, ness matrix into LU, and if the Faddeev-Faddeeva bordering is u sed, no
solution is denoted by z:Lz =u
extra problems of storage and numerical stability are created by the addition
· unchanged, except at one crucial 1
of singular functions.
may no longer be true. Indeed, in 1
nonzero components of the singu1
8.3. ERRÓRS IN THE PRESENCE OF be content to work with theweakt
SINGULARITIES
llzlla ~
Let L be a 2mth-order self-adjoint elliptic operator with homogeneous
boundary data, and let a(v, w) be the associated inner product on the energy which follows from (25) (replacin
space JC~. If the problem Lu = fhas interface or boundary singularities such quality is that a factor h 2m-a has.
as those described in Section 8.1, the error estimat~s derived in 3.4 are no
longer valid. In this section we shall modify the earlier analysis to obtain the
correct rates of convergence in the presence of singularities.
Using expansions analogous to those derived in Section 8.1, we can write For example, in the torsion expe
the exact solution as a sum of singular functions plus a smoo.th function: section, u= -5;. For any choice of
to be O(h 112 ) and the error in disp
(24) U= :t
i=l
C¡t¡/¡ + W. This is the error over the whc
it is reasonable tó hope for sometl
ha ve a strong smoothing effect in tl
Each singular function t¡l; is in Jea for sorne u > m, and is ·independent of of ordinary least-squares approxi1
the data f; it depends only on the geometry of Q in the case of corners in apparently no pollution from the
the bound.ary, and on the coefficients of L for interface problems. We may domain Q', then even without spec
use either r¡¡i = rv rp;(O), as in (11), or ti]. e function constructed in (21); all
1
tion over Qis correct to order hk i
that matters is to keep the correct behavior near P. The smooth function w is no longer tr:ue, and sorne polluti
and the coefficients e 1 , • • • , es' op the other hand, do depend onf Accotding is better within Q' than near the ~
to the fundamental work of Kondrat'ev [K3], it is not only possible to ensure solution behaves like ra. near a e
that w is in JCk (by including sufficiently many r¡¡¡) but .even to estimate its Nitsche and Schatz, the error in th
size: to the singularity is of order h 2 a..
error in JC 1 of order h away fro
(25)
over the whole domain n.
Now we turn to the importa11
Observe that if Q and the coefficients of L are smooth, then- the associated singular functions are introduced
singular functions r¡¡ 1 , • • • , r¡¡ s are -zero and hence (25) reduces to the usual we note that by construction the s
bound for the solution in terms of the data. tion U 11 satisfying (according to (:
As a first case let us derive the rate of convergence for the finite element
method when no special tricks-mesh refinement or singular functions- (26)
are used. The error in strain energy presents no difficulties. As always, d•
is the closest trial function to u, and if u lies in JCk, then this error is of order
Therefore, the same bound holds f
h 2 <k-m>. In general, however, u will lie only in sorne less smooth space Jea,
Proceeding to the JC 0 estimat,
and the rate of convergen ce in energy is reduced to h2 <a-m>. This erro,r is
Lz =u -ti'. The crucial ~oint of
likely to be unacceptably large.

CHAP.
value problems. The bisection and
:torizations of the mass and stiff-
i-Faddeeva bordering is used, no
.tability are created by the addition
liptic operator with homogeneous
ciated inner product on the energy
face or boundar,y singularities such
or estimates derived in 3.4 are no
fy the earlier analysis to obtain the
ce of singularities.
ierived in Section 8.1, we can write
functions plus a smooth function:
+w.
tme u > m, and is ·independent of
!try of n in the case .of corners in
L for interface problems. We may
~ function constructed in (21); all
or near P. The smooth function w
~r hand, do depend onf According
C3], it is not orily possible to ensure
many r¡t¡) .but ,even to estímate its
L are smooth, then the associated
nd hence (25) reduce.s. to the usual
.ta.
~ convergence for the finite element
efinement or singular functions-
~ents no difficulties. As always, uh
es in X\ then this error is of order
1ly in sorne less smooth space :JCO',
~ reduced to h 2 <0'-m). This error is
SEC. 8.3. ERRORS IN THE PRESENCE OF SINGULARITIES 267
8
Suppose that we atternpt to carry out Nitsche's trick, as in Sections 1.6
and 3.4, to estímate the error in displacements without singular functions.
As before, we take u - uh to be the data g in an auxiliary problem, whose
solution is denoted by z: Lz u uh. Then the argument can proceed
unchanged, except at one crucial point: The estímate llzll 2 m Cllu- ifllo
may no longer be true. Indeed, in light of (24), it may happen that z contains
nonzero components of the singular terms r¡t 1 , ••• , r¡t s' and hence we must
be content to work with the weaker inequality
which follows frorn (25) (replacing u with z). The effect of this weaker ine-
quality is that a factor h 2 m-O' has to be sacrificed, and the optimal bound is
r min(k,a).
For example, in the torsion experiment for a crack described in the next
section, u = l For any choice of element this forces the error in the slopés
to be O(h 112 ) and the error in displacement to be O(h).
This is the error over the whole domain Q. Away from the singularity
it is reasonable to hope for something better, since elliptic equations always
ha ve a strong smoothing effect in the interior ofQ. In fact, ifit were a question
of ordinary least-squares approximation by piecewise polynomials, there is
apparently no pollution from the singularity; if u has k derivatives in a sub-
dornajo Q', then even.without special tricks the best least-squares approxima-
tion over Q is correct to order !zk in Q' [N6]. For second-order equations this
is no longer true, and sorne pollution does occur. However, the exponen{ still
is better within Q' than near the singularity. Suppose, for example, that the
solution behaves like rrx near a corner in the domain. Then, according to
Nitsche and Schatz, the error in the energy norm over Q' which is attributable
to the singularity is of order h2 rx. For the region with a crack this means an
error in :JC 1 of order h away from the singularity, as compared with h 112
over the whole domain n.
Now we turn to the important question-the rate of convergence ·when
singular functions are introduced into the trial space. For the energy norm
we note that by construction the singular space S~ contains at least one func-
tion Uh satisfying (according to (23))
(26) a( u - U\ u Uh) Ch 2 (k-m) 11 u - :t m.
i=l
C¡'t/f¡
Therefore, the same bound holds for u - uh.
Proceeding to the :JC 0 estímate we again consider the auxiliary problem
Lz u uh. The crucial :point of Kohdrat'ev's theory is that z can be written
268 ~ 1 SINGULARITIES CHAP. 8 SEC. 8.4.

as ~f=;' 1 d¡'f/¡ + v, where v lies in Jek and In the torsion problem we give
refinement as against the use of sin~
(27) their rates of convergence are the s
number of unknowns required. In
Since v is smoóth, we can approximate it with a trial function vh in Sh to
we shall be less concerned with si1
order hk-m in the energy norm. (We assume k> 2m, as is normal for finite
efficient methods for treating inte1
elementsi~ On the other hand, the function ~ d¡'f/¡ is in our singular space
is included beca use of its long histo
S~, and so. is zh = L d¡'f/¡ + vh. Therefore,
a singularity. In fact, special methc
d uced extremely accurate approxi
We shall compare these results wi
Repeating the (rather complicated) argument of Section 4.4, method.
The differential equation govern
(29) in normalized form as

(30) -fl.u = 1 in
This means that by including enough singular functions, it is possible to
obtain the same rate of convergence as for smooth problems. A similar conclu- n is the region shown in Fig. 8.3 a
sion applies to mesh refinement, if the mesh size is taken to be an average sions about P in Section 8.1 reduce
ii = N- 112 , based on the dimension N of the trial space. In this case the
singular functions are not in the space, but according fo the final comments u(r, 8) = L cirv; 5
in Section 3.2, a suitable mesh refinement permits their approximation to j= 1

order hk. With this estimate the Nitsche argument goes as befo re.
.It is obvious that all these theoretical predictions must be thoroughly plus analytic terms. Thus the domiJ
tested. In. complicated physical problems, it may be extremely difficult to is r 1i 2 sin (8/2). lts coefficient
identify the singularities and to incorporate them into special trial functions.
(31) c 1 = lim ,- 1
Therefore, such a construction will be carried out only if the benefits are r--+0 ··

correspondingly great. Even mesh refinement introduces sorne complications,

_although it is usually much simpler than the construction of singular func- has great engineering significance;
)tiOns. All we can do here is to try- each of these possibilities on a number of the torsion which the beam can en
siiJ?-plified physical problems and report the results; By glancing at the graphs the stress intensity factor [18]. We n
in the following section, the reader can anticipate our conclusion: High just short off derivatives in the m
accuracy has to be paid for, either by computer time on a simple method or In the problem as stated there
by programm~ng time with a more subtle -technique. The prices vary with the at each of the corners P2 , P3 , P4 , a
problem. But almost certainly, from his own experience, the reader already paratively insignificant, we shall r
knew that. conditions to

8.4. EXPERIMENTAL RESULTS

where V denotes the normal to r. ~
We conclude this chapter with three examples drawn from physics: (1)
value problem is analytic away fron
the computation of the rigidity and deformation of a cracked squan! e las tic
Section 8.I.t
beam under torsion; (2) the criticality computation in an idealized square
We shall compute approximati
nuclear reactor consisting of a homogeneous square core surrounded by a
square reflector, in the one-group diffusion approximation; and (3) the com- tBy the Saint-Venant principie this el
a·
putation of the fundamental frequency of vibrating L-shaped membrane. singularity at P, a fact which is also evid
 CHAP.
it with a trial function 1i in Sh to
Jme k > 2m, as is normal for finite
ion ~ d1t¡11 is in our;~:~ingular space
~e,
nent of Section 4.4,
singular functions, it. is possible to
smooth problems. A similar conclu-
nesh size is taken to be an average ·
of the tria! space. In this case the
•ut according to the final" comments
:nt permits their approximation to
ugument goes as before.
al predictions must be thoroughly
LS, it may be extremely difficult to
ite them into special trial fu_nctions.
:;arried out only if the benefits are
1ent introduces sorne complications,
1 the construction of singular func-
.f these possibilities on a number of
he results. By glancing at the graphs
n anticipate our conclusion: High
mputer time on a simple method or
' technique. The prices vary with the
own experienc.e, the reader already
examples drawn from physics: (1)
rmation of a cracked square elastic
:omputation in an idealized square
1eous square core surrounded by a
)n approximation; and (3) the com-
)f a vibrating L:·shaped membrane.
8 SEC. 8.4. EXPERIMENTAL RESULTS 269
In the torsion problem we give a numerical comparison of local mesh
refinement as against the use of singular functions. For a given finite element
their rates of convergence are the same, and efficiency hinges largely on the
number of unknowns required. In· the reactor problem, on the other hand,
we shall be Iess concerned with singu1arities and concentrate attention on
efficient methods for treating interfaces. Finally, the L-shaped membrane
is included beca use of íts long history as a model of an ellfptic problem with
a singularity. In fact, specia1 inethods developed for this problem have pro-
duced extremely accurate approximations to the vibrational frequencies.
We shall compare these results with those obtained by the finite element
method.
The differential equation governing the torsion problem (1) can be written
in normalized form as
(30) -ll.u = 1 in n, u= o on r.
n is the region shown in Fig. 8.3 and r includes the crack P 1 P. Our expan-
sions about P in Section 8.1 reduce to
2j-
vi= __ 1
2_,
plus analytic terms. Thus the dominant term in the singularity at the point P
is r 1l 2 sin (8/2). Its coefficient
(31) c1 Iim ,-u 2 [u(r, n)- u(O, n)]
r-~o
has great engineering significance; it is the commonly accepted measure of
the torsion which the beam can endure before fracture occurs and is caBed
the stress intensity factor [I8]. We note that because of the factor r 112 , u falls
just short of! derivatives in the mean-square sense.
In the problem as stated there are also singularities of the form p 2 ln p
a,t each of the corners P 2 , P 3 , P4 , and P 5 • Since these singularities are com-
paratively insignificant, we shall remove them by changing the boundary
conditions to
(32)
where V denotes the normal to r. That the solution u of the new boundary-
value problem is analytic away from P can be verified using the techniques of
Section 8.l.t
We shaJI compute approximations from four different spaces; the first
tBy the Saint-Venant principie this'change of boundary conditions will not affect the
singularity at P, a fact which is also evident from our analysis in Section 8.1.
270 SINGULARIT!ES CHAP.8 SEC. 8.4.

three deal with a uniform. mesh of length h and use singular functions con- is O(h 4 )'for the cubic spaces, and C
sttu~ted from rv1 sin v/l, vi= (2j 1)/2, as in Section 8.2. These are the fol- this error directly, however, since th
lowing: Therefore, we also used a quintic
More precise! y, we computed with t
l. Si denotes the space of continuous functions on a uniform mesh which across the lines PQ 1 , PQ 2 , and PQ
reduce to a bilinear polynomial a + bx + cy dxy on each subdivision. mented by six singular functions. )
This spacé~also includes the singular function fJ/ 1 constructed from r 112 sin(B/2) is obtained with this space; the ratee
so that we''have the approximation property (26) with k 2. · solutions for h !, i, and Tu- differ
figures were taken to. be a correct '
2. Si: denotes the bicubic Hermite space (Section 1.8) together with the
Figures 8.4 and 8.5 show the err<
singular functions fJfp fJ/ 2 , and fJ1 3 , so that (26) holds with k 4. an average mesh length ii N-llz
These examples fall into the category of nodal.finite element spaces in the "efficient" space, meeting a given e
sensé that we have discussed earlier, ~nd so it is perhaps appropriate to unknown parameters. In this regare
include also an example of an abstract.finite element space. The space of bicu- space S~ 6 of triangular elements w
bic splines which are of class ez in n is the most likely candidate; however, we used h !, ~ -(o and h = t•
it is a difficult space to work with for this problem. The troubles are related . ·many extra elements are needed to
to the essential boundary condition u = O on the.crack PP 1 • A bicubic spline only one extra unknown-the coc
which vanishes along this line will be .so constrained at the point P that it required. The spaces S'¡¡, S}L of cu
cannot possihly match the true solution u, whose derivatives are all singular. the space of mínimum dimension.
To avoid this, we impose only simple continuity (e 0 ) across the lines PQ 2 , appears to be the most efficient of
PQ 3 , and PQ 4 and obtain what is usually called a spline-Lagrange space: The graphs also indicate the
without singular functions. There a
3. S}L consists of piecewise bicubic polynomials which are ez everywhere these figures. The first, naturally, h
exc~pt at the lines PQ 2 , PQ 3 , and PQ4 • Across these lines the normal deriva- functions-at R1 and h = !, for
tives are permitted to be discontinuous. As with Si:, we include the singular guiar functioris is about 40 per ce
functions fJ/ 1 , fJ/ 2 , and fJ/ 3 so that (26) holds with k 4. singular functions ·are added. The :
functions, the pointwise errois ·an:
The advantage of the spline-Lagrange space is that it has only one un- that these errors are of order O(h 1
known per node, except along the lines PQ 1 , PQ 2 , PQ 3 where there are three. third is that the standard Hermite
Thus the dimension of S}L is nearly four times less than that of the Hermite linear elements. The cubics are toe
space Si:, and virtually the same as the dimension of the piecewise bilinear FinalJy, in Fig. 8.6, we give ap]
space Si. Incidentally, a basis for S}L is readily obtained from the standard (31). For the spaces Si, S'H, and S
spline formulas, by considering the lines PQ 1 , PQ 2 , and PQ 3 as a triple conflu- function fJ/ 1 , since
ence of nodallines.
Our final space consists of triangular elements and uses a graded mesh of é1 = lim r- 11
r-+0
maximum length h and mínimum length ~.
For the space S~ 6 this quantity is
4. S~ 6 denotes the space of continuous functions which reduce to linear
polynomials on each triangle. We assume that ó = O(h 2 ), so that (26) holds tWhile the error estimates cited above
with k= 2.t the rates of convergence appear to be th
the graphs.
The estimates in Section 8.3 imply that the square root of f 1u uf' 12 :j:An analysis similar fo that in Secti
case is O(h), hence the area of the regi<
tin the calculations, the transition from meshsize h to á = h2 was achieved by progres- must be quite small and converge to ze1
sively halvin~ the triangles. with linear functions; R 2 is within the 1
 CHAP.
th h and use singular functions con-
' as in Section 8.i These are the fol-
s functions on a uniform mesh. which
~~: + cy + dxy on each subdivision.
tion t¡t 1 constructed f*9m r tlz sin(O/2)
~rty (26) with k 2. · .
pace (Section 1.8) together with the
tt (26) holds with k 4.
of nodal finite element spaces in the
nd so it is perhaps appropriate to
!ite element space. The spacé ofbicu-
the most likely can di date; however,
is problem. The troubles are related
O on the crack PP1 • A bicubic spline
o constrained at the point P that it
u, whose derivatives are all singular.
ontinuity (e 0 ) across the lines PQZ,
ally called a spline-Lagrange space:
·olynomials which are ez everywhere
\cross these lines the normal deriva-
As with S'¡¡, we in elude the ~singular
!ds with k 4.
~e space is that it has only one un-
'Q1, PQ 2 , PQ 3 where there are three.
r times less than that of the Hermite
dimension of the piecewise bilinear
; readily obtained from the standard
0
Qp PQ 2 , and PQ 3 as a triple conflu-
elements and uses a graded mesh of
J.
ms functions which reduce to linear
;: that J O(h 2 ), so that (26) holds
that the square root of J1 u - u!' ¡z
.hsize h to o = hZ was achieved by progres-
8 SEC. 8.4. EXPERIMENTAL J.'ESULTS 271
is O(h 4 }for the cubic spaces, and O(h 2 ) for Si and S~ 6 • We cannot measure
this error directly, however, since the solution u is not known in closed form.
Therefore, we also used a quintic element with several singular functions.
More precisely, we computed with the quintic splines, of class e4 in n except
across the lines PQ~' PQ 2 , and PQ 3 -a quintic spllne-Lagrange space, aug-
mented by six singular functions. A very rapidly convergent approximation
is obtained with this space; the rate of convergence is h6 , and the approximate
solutions f9r h .¡,, !, and -(o- differed only in the seventh place. The first six
figures were taken to. be a correct value for u.
Figures 8.4 and 8.5 show the errors at the points R 1 and R 2 plotted against
an average mesh 1ength ii = N- 112 .t This permits the selection of the most
"efficient" space, meeting a given error tolerance with the least number of
unkpown parameters. In this regard Si appears to be more efficient than the
space S~ 8 of triangular elements using local mesh. refinement. For the latter
we used h = !, J = -fo and h !, J -h according to the rule in 4. Thus
many extra elements are needed to obtain the h312 convergence, yet with Si
only one extra unknown-the coefficient of the singular fqnction t¡t 1-is
required. The spaces S'¡¡, Sh of cubic elements were far superior to Si, and
the space of mínimum dimension, namely the spline-Lagrange space S~L,
appears to be the most efficient of all.
The graphs also indicate the (larger) errors for finite element spaces
without singular functions. There are three interesting thirtgs to note about
these figures. The first, naturally, is the improvemetit obtained with singular
functions-at R 1 and h !, for example, the relative error without sin-
gular functions is about 40 per cent, aQd this drops to 0.1 per cent when
sing~lar functions are added. The second is that in the absence of singular
functions, the pointwise errors are largest near P. In particular, it appears
that these errors are of order O(h 112 ) near P and O(h) everywhere else.t The
third is that the standard Hermite cubics come out worse than the simplest
linear elements. The cubics are too smooth to cope with the singularity.
Finally, in Fig. 8.6, we give approximations to the stress intensity factor
(31). For the spaces Si, S'¡¡, and S~L we use the coefficient.c 1 ofthe singular
function t¡t 1 , since
lim r- 112 [zf(r, n)- ·u11(0, n)].
r->0
For the space S~ 6 this quantity is zero, and it is necessary to choose sorne
tWhile the error estimates cíted above give no information concerning pointwise errors,
the rates of convergence appear to be the same for this problem, as can be verified from
the graphs.
tAn analysis similar to that in Section 8.3: shows that the mean-square error in ttiis
case is O(h), hence the area of the region about P wh~re the pointwise error is O(hliZ)
must be quite small and converge to zero as h ---¡. O. This explains the kink in the ·error
with linear functions; R 2 is within the boundary layer, when h is small.
 U(R1) = 0.071023 V
10,000
8000 1 1
20,000
'6000
'- i± 1-f--
¡-r-
"~
'~
Bicubic hermite
4000 ~ No singular functions 10,000
Slope = 1 t===" Bilinear

,,'"'
3000 1
8000 t - - - - No singular funct
~~ 7000
6000
¡-
11...
2QO.O 5000

~
t•';l

\~~
o 4000

3000
1000 '\.
1'
- ~
800

600
\'\.. "
\J.t "'"' e-Si linear
2000

Sf•8 A "Y No singular functions

400
Slope:::; 2 \,'\ Slope: 1 1\
1000
300 1 \ 1"' 800 1

shL
\' 700 1
-t-
1-

U) 200
.,...
""1
U)
+
600
500
1 ·

Slope =2 t-
+
Q
X
1-
Í\
~
\ º )(

lo.
400
o1-
1-
100
g
Q)
300
CIJ
~ 80 1\. .!!?
::::1 ::::1 200
o111 1\.. o
..e 60 (/)
.o
sh
e::[ 1 '\--
L c:::r

40 DÍ \ Slope:::; 2 1
100
30
, o 1\ \ 30

20
\1\ \ 20
~
s:L_:..--
~
~
Slope:: 4
\ s~L -
10 10 Slope::4=
1\ 8
8
6
' \\
ttt 7
6
\ \ 5
4 1 4
~~~\
3
~~h
S~ope =4
2 2
\

3 4 5 6 7 8 10 20 30 40 60 80 100 3 4 5
1/h
Fig. 8.4

272
0.071023 V(R2)= 0.027425

.
20,000

=~Bicubic hermite
-r-- No singular functions
Slope = 1
10,000
t===Sil inear "'
~~~
8000 1----- No singular _functions
'
Bicubic hermite
No singular functions
Slope =t

,,
~"~:"'Ct
7000
-t--- ~y-
r--- Slope =1
6000
'!\..
' ~""'
5000 \.

~"
¡':
¡ ¡

4000

: 3000 ' ~ '~

'\
~
\.'- ' 2000 1\ ~
\~' '
~ -Si linear
----\ ~ No singular functions
\)\ Slope~: 1 \ s~a
1000 -¡-¡-

\ ~
)..
..._Siope = 2 -f-+-
'
.

800 -¡-¡-

\ ' 700
600
st -""-
\
'
U) Slo pe= 2
+ 500 \ 1

\ º)(
.....
400
\1
\ o
g 300

\.

o
Q)

Q)
:; 200 \
~
'\-- -
\.
\
;
sh·
L
Slope =2
(/)
..0
<{

100
30
:>'
\ ' "7

\ 20 \

\ 10
s~L
Slope = 4
\ \ \
S~
·slope = 4
\
i\ \
e
1\
\ 1
7
_\
,~,
..,,.\ 1

6
1

\ \ 5 \
1 4 1\ \
~ .. \ 3 \
\Lsh
\
sfope =
4
2
1\
\
\
\>
\
10 20 30 40 60 eo 100 2 3 4 5 678910 20 30- 40 5060 eo 100
1/h 1/h
1 Fig. 8.5

273
 Stress intensity factor =0.1917
SEC. 8.4.
1000
800
700
IE difference quotient
600
'500
,.
400

300
=Hii Since the errors in the finite elemem
tions grow quite rapidly as the poin
,,,200
1'\ Bicubic hermite choice { 11 O(h) is as good as any.
~!"a
No singular functíons The conclusions are the same as in th(

roo
~Bilinear
'~ :--.,
l"'tl ¡......
V
A'r"o
V ~
I"''..L. ...........
Oifference quotíent
approximation
Slope =t f-
f-
refinement is even less competitive 1
difference quotient.
80 -No singular functions .... ~
f- Our second example is the one-J
70
- Oifference quotient ............ "' ~
60
- approximation '~ (33)
50 -
=t "" .................
......
-V·(p Vu)
40 _Siope

' This differential equation holds in th

30

20
1

sh,s
1 1

Oi¡ference quotient
~
. " ~ i
8.7), and we require the energy .flu:

(34) u=O

"''
appro:dmatíon
V
... Slope 1 = _
º ><
... i
·e ('¡)
ro
8
!'...
""
1
1
1
7 1
Q) 1
'5
o
6
5 S~ -
' ''" 1
r

~
.o
<(
(/)

4 Slope =1 _..._" P 1
1
1

Os - - - - - - -4- - - -
.

2 """1~ '

01 .... -·----·--·----t---
1
!).,. \
P,l
0.8 1
0.7
0.6 s~L
Slope =3
__ ¡...,
'\ \
1
1
1
1
0.5 1

' \ 1
1
0.4 1

-0.3
~ \--S~ 1

\ Slope"' 3
\
0.2 Fig. 8.7 Square core sun
\\ The coefficients p, q, and pare regio1
0.1 rating n 1 and 02 is an interface, and ,
1 2 3 4 5 6 7 8 910 20 30 40 50 60 80 lOO
ous across r:
1/ h
. (35) iJu\ =p-
p- iJu\
Fig. 8.6 iJv r+ iJv r
274
1sity factor =0.1917
SEC. 8.4. EXPERIMENTAL RESULTS 275
1 1
1 l

¡ difference. q uotient
:
1

1
Since the errors in the finite element approximation without singular func-
! tions grow quite rapidly as the point P is approached, it turns out that the
!

Bicubic herr¡fu,ite choice eh = O(h) is as good as any. These are the values given in Fig. 8.6.
No singular f.unctions The conclusions are the same as in the previous experiments, except that mesh

·'~
Difference quotient
approximation
refinement is even less competitive because of the extra error made in the
t-

l~ ~
...........
Slope ~ t difference quotient.
Our second example is the one-group, two-region reactor govemed by
1 "o. .....
r... ............ ............
(33) -V ·(p Vu) + qu = A.pu.
'"" "
............ ¡..,
1-'
t This differential equation holds in the core 0 1 and the reflector 0 2 (see Fig.

~nt ~ ~ "" 8. 7), and we require the energy flux u to vanish on the o u ter boundary:

(34) u=O

i
!
(
" ~
1
1
1
1
1
1
1
1
.n2
1 1
1 1
1 t
~', ' 1 !
1 1

1 - "
1
P4 1 r+ IP3

" :""
Q6
! r_

i .... r+ r_ .n, r_ r+
!
1

r_ --
Q1
P1 1
1
r+ tP2
1
1 1
1 1
j 1 1
1 1
! 1 1
1
i 1\ 1
1
1
1
f

;\
\--S~
\ Slope= 3
,. , 1
1 l

Fig. 8.7 Square core surrounded by a square reflector.

¡\ \ Thé coefficients p, q, and p are regionwise constants. Thus the curve r sepa. .
rating 0 1 and 0 2 is an interface, and we require that u and p au¡av be continu-
~lO 20 30 40 5060 80 100
ous across r:
1/ h
. (35) au¡
p- au¡ uln ulr-·
.6 av r+ =p-
av r-' =
 ' '
276 ' SINGULA'RITIES CHAP. 8 10,00o
800o
700o
The _most important quantity to calculate is the lowest eigenvalue l, which 600o-
meastires the criticality of the reactor. 500 o
Independ(mt of any corner, the existence of interfaces fundamentally alters 400o
the best choice of an appropriate finite element space Sh. To use piecewise 300o
polynomials which are e 1 across r would clearly lead to poor approxima-
tions, ~ince u has discontinuous derivatives. Moreover, the use of trial func- 200 o "
tions si~isfying the jump condition (35) leads to extra difficulties at the corners
Pi,j = l, 2, 3, 4. If we force the trial functions to satisfy the jump conditions
\ 1\
along P 1 P 2 , then they will still be influential on a portion of Q 1 P 1 where the 100o
solution is smooth. 80o
Because of the convenient geometry, a spline-Lagrange space like that 70 o
60o
described in the torsion problem appears to be the most appropriate. The
50o
idea is to subdivide the region n into squares, with the interfaces lying on
40 o
mesh lines. The trial spac~ Sh consists of piecewise bicubic polynomials which
are C 2 everywhere except on the lines Q1 Q4 , Q2 Q7 , Q 3 Q 6 , and Q 5 Q 8 , across 30 o
which the elements are only continuous. We ignore the jump condition (35),
20o
which is a natural boundary condition, and allow the trial functions to ha ve ID
+
arbitrary jump discontinuities in their normal derivatives across the inter-
faces. Since the Galerkin method gives us in sorne sense a best approximation,
ºX

it will presumably work out the jump discontinuities across the interfaces in g
Q)
10o

a satisfactory way! Q) 8o
"S 7o
Approximations to the first eigenvalue using this space are given in Fig. o
1/)
6o '
.o 5o
8.8 for the case
~
e::(

4ó
Pt = 5, Pz = 1, q =O, p=l. 3o
t-
1 \

The approximations A_h are reasonably accurate, yet the rate of their conver- 20
gence to l is very slow. In particular,
~
(36) V= 0.78.
10
8
The reason is that the eigenfunction u has unbounded first derivatives at the 7
points Pp P 2 , P 3 , and P 4 • In fact, the analysis in Section 8.1 shows that the 6
5
dominant term in the singu_larity is of the form rvrpv(O), where 'Pv is a periodic
4
function of (} and v = 0.78. Therefore, the eigenvalue error cannot be bettyr
than h~x~ By including rvrpv in the trial spac·e, the convergence rate can be 3
Í!lcreased to approximately h 6 - 2 v. This is also confirmed by the numerical
data. 2
Because of the singularity, it may seem surprising that l\ on a uniform
mesh without singular functions, is reasonably accurate. This is in striking
contrast to the situation for the torsion problem. The reason is that the coef- 1
2 3 4 5
ficient of rvrpv is quite small; computer plots ofthe eigenfunction uh show that
it is virtually a constant in the core 0 1 • We have checked this property (which
is a consequence of the physics) by computing also the critica! eigenvalue for
 CHAP. 8 10,00o
800 o
700o
e is the lowest eig~nva1ue ,t, which

ce of interfaces fundamentaHy alters

600 O'
500 o
400o
t
element space Sh. To use piecewise 300o
Id clearly lead to poor approxima-
es. Moreover, the use of trial func- 200o
lds to extra difficulti~\at the corners
"
:tions to satisfy the juntp conditions 1\ 1\. Cubic space without
r-r--singular functions
tial on a portion of Q1 P1 where the \~
100o '\. Slope ~2v
80 o
. a spline-Lagrange space like that
·s to be the most appropriate. The
70o
60o ''
1uares, with the ·interfaces lyirig on
liecewise bicubic polynomials which
50o
40o '1\
Q4, Q2Q,, Q 3Q6, and Q 5Q 8, across 30o \
We ignore the jump condition (35),
20o
'
ld allow the trial functions to have u;¡
+
tormal derivatives across the ínter-
in sorne sense a best approximation, º
;continuities across the interfaces in 10o
8o 1

7o i
!

te using this space are given in Fig. 6o ! l

.

5o
4o ~
\ Cubic space with one
q =0, p=1. 3o
{ singular function

;curate, yet the rate of their conver- 2o

\ Slope~6-2ll

V= 0.78:
10
8
~
s unbounded first derivatives at the 7 !\
\
alysis in Section 8.1 shows that the 6
· form r rpv{8), where 'Pv is a periodic
11

1e eigenvalue error cannot be bett;r

5
4
'
ipac·e, the convergenee rate can be 3 \
is also confitmed by the numerical
2
:m surprising that lh, on a uniform
onably accurate. ·This is in striking
)roblem. The reason is that the coef- 1
2 3 4 5 6 7 8 910 20 30 40 5060 80 100
ts of the eigenfunction uh show that
e ha ve checked this property (which 1/h
uting also the critica! eigenvalue for Fig. 8.8

277
278 SING ULARITIES CHAP.8 SEC. 8.4.

the caset like

p 1 = 500, Pz = 1, q= O, p=l.
(37) u(r, 8) = E E ci
j= 1 1=0
1J

This eigenvalue is shifted only to 5.582 and is therefore virtually independent

of p 1 ; the contribution to the Rayleigh quotient from the inner square íl 1 (plus analytic terms) near the reent
is almo~t zero. The most accurate ,approxima
Frofu a physical point of view this "weak" coupling to a "strong" singu- Fox-Henrici-Moler [FIO]. The ide~
larity is quite satisfactory. In fact, unlike the torsion problem, the singularity tions to Au + lu = O, in this case
at P has no physical meaning-equation (33) should be replaced with a v = vi' and to determine the coeffi1
transport equation in this region. It would appear, therefore, that singular tion on the boundary r. This me
functions are not absolutely necessary for such problems, and a great deal method: It works with exact solutio
of mesh refinement may not be worth the effort. This conclusion does not apply tions. The essential point, however.
to all interface problems. It was mentioned in the first section that a crossing · It is known that the eigenfunction
of interfaces could produce a dominant term in the singularity of any order by linear combinations of the K,,
r€rp€(8).t lt seems reasonable that a small E will be reflected in poor approxi- produce approximations with simil
mations, as in the tQrsion problem, and that singular functions or local mesh were used. Strictly speaking, the fu
refinement may be necessary to obtain acceptable results. violate the essential boundary cm
Our final example is the perennial L-shaped membrane-an overworked closely related functions (1 - x 2 )
but nevertheless effective model (Fig. 8.9). We seek the eigenvalues of -Au = which are comparable with those e
lu in Q, u= Oon r, and note from Section 8.1 that the eigenfunctions behave These calculations reflect the re1
by knowing exactly the righ t trial f1
polynomials! (Fourteen singular f1
( -1 '1) ( 1' 1)
r-----------~-.--------------. bicubics, to improve on the resu
however, is to obtain good accurac.
finite element program. Thereforel
eigenvalues, using a cubic spl~ne s¡
(r,8)
Figure 8.1 O illus~rates how the slo
case is greatly increased by the intr
There is no question that the cons
the whole program more efficient.
P=(O,O) (1,0)
The conclusion which we draV~
Even for coarse meshes and for sii
which the theory predicts are clea
engineering literature contains a
also leading to the same conclusic
the steps of the finite element me1
achieved. We hope that this analy:
(-1,-1) (0,-1) fu tu re development of the method
nomial elements was already clear,
Fig. 8.9 An L-shaped membrane. is mathematically confirmed.
tv in this case in approximately j.
tFor the one-group, two-region reactor, the eigenvalues are the roots of (20), and
E2j. ,
 CHAP.8
q=O, p l.
(37)
td is therefore virtually independent
quotient from the inner 'square nl
veak" coupling toa 'fstrong" singu-
the torsion problem, the singularity
m (33) should be replaced with a
Jld appear, therefore, that singular
Jr such problems, and a great deal
·ffort. This conclusion does not apply
~d in the first section that a crossing
:erm in the singularity of any order
f will be refiected in poor approxi-
hat singular functions or local mesh
;ceptable results.
shaped membrane-an overworked
, We seek the eigenvalues of -Au
n 8.1 that the eigenfunctions behave
( 1' 1)
----------,
' the whole program more efficient.
::>:(0,0} (1,0}
),-1)
ted membrane.
1e eigenvalues are the roots of (20), and
SEC. 8.4. EXPERIMENTAL RESULTS 279
like
2'
u(r, e) V = _l_,
j 3
(plus analytic terms) near the reentrant corner P.
The most accurate approximations to date have been computed by
Fox-Henrici-Moler [FIO]. The idea is to take a linear combination of solu-
tions to Au + lu =O, in this case combinations of Kv li,JI:r) sin ve,
v vi, and to determine the coefficients by minimizing the linear combina-
tion on the boundary r. This method is in a sense dual to the Galerkin
method: It works with exact solutions and approximates the boundary condi-
tions. The essential point, however, is th.e class of functions which are used.
It is known that the eigenfunctions u can be very accurately approximated
by linear combinations of the Kv, and therefore a Galerkin method would
produce approximations with similar accuracy if the same class of functions
were used. Strictly speaking, the functions Kv are not admissible, since they
violate the essential boundary. conditions. However, calculations with the
closely related functions (l - x 2 )(l y2 )rv+ 21 sin ve are reported in [F9]
which are comparable with those obtained by Fox-Henrici-Moler.
These calculations reflect the remarkable accuracy which can be achieved
by knowing exactly the right trial functions-which are not always piecewise
polynomials! (Fourteen singular functions would be needed, together with
bicubics, to improve on the results described above.) Our chief interest,
however, is to obtain good accuracy with simple modifications of a standard
finite element program. Therefore, we have computed the first and fourth
eigenvalues, using a cubic spline space with and without singular functions.
Figure 8.1 O illustrates how the slow rate of convergence (h 413 ) in the latter
case is greatly· in~reased by the introduction of three singular trial functions.
There is no question that the construction of these special functions makes
The conclusion which we draw from all the numerical evidence is this:
Even for coarse meshes and for singular problems, the rates of convergence
which the theory predicts are clearly reproduced by the computatioris. The
engineering literature contains a large number of numerical experiments,
also leading to the same conclusion. This means that our goal, to analyze
the steps of the finite element method and to explain its success, is largely
achieved. We hope that this analysis will provide a theoretical basis for the
future development of the method. The simplicity and convenience of poly-
nomial elements was already clear, and now the accuracy which they achieve
is niathematically confirmed.
 A., = 9.639723844
A.4 = 41.474516

2,000,000

"
•
1,000,000
8oo,qoo
600,000 ' '
~

400,000 1111
l 11
300~000
8A.4
"
"
~·:.

2oo;ooo __ Cubics-no singular functions

ax 1 ~'\ !\. Slope =4 BIBLIOGRAPHY
100,000
80,000
=: Cubics- no singular
-:- functions 1
GQOOO -
r- Slope =4
40,000
1 1
30,000
20,000
(
11 t1t111
\111 8A.4 The number of publications on t
10,000 1 / V Cubics-3 singular functions both the engineering and the numeric
8000 Slope"" 6 -impossible any attempt at completen
6000 1\ prior to 1971 is contained in the exce
4000 ZIENKIEWlCZ, 0. C. (1970), The finite
ID
1\
º...
)(
3000
2000
1 (
Appl. Mech. Rev. 23, 249-256.
The following books have been refe1
g
for the three which appeared too rece
Q)
Q)
:; 1000 \
1. AGMON, S. (1965), Lectures on
oI/)
.a "> trand Reinhold, New York .
<1:. 400
300 2. ÁRGYRIS, J. H. (1954-1955), Ene
worth, London, 1960.
200
3. ARGYRIS, J. H. (1964), Recent
100 \ Analysis, Pergamon Press, Elms
80 4. AuBIN, J. P. (1972), Approxim
60 8X1
11" Cubics- 3 singular functions Wiley, New York.
40 -,\+ Slope ~ 6 5. DESAI, C., and J. ABEL (1972),
30
' Van Nostrand Reinhold, New)

w
20 ' 6. FADDEEV, D. K., and V. N. F.

• -
Linear Algebra, Dover, New Yo
10
8 7. KRASNOSEL'SKII, M. A., G. M. \
6 of Operator Equations, Nauka, l'
4 8. HoLAND, 1., and K. BELL, eds. (1
3 sis, Tapir, Trondheim, Norway.
2 LIONS, J. L. (1969), Quelques mé
1 2 34 6810 20 3040 60 80 100 200 400 1000 9.
non linéaires, Dunod, Paris.
1/h

Fig. 8.10
l.639723844
.474516

m~
TT

8>..4 i\
!--
t- Cubics- no singular functions
t-
BIBLIOGRAPHY
Slope =%

8>..4 The number of publications on the finite element method is growing, within
Cubics- 3 singular functions ¡...... both the engineering and the numerical analysis literatures, at arate which renders
Slope 1::6 t-
impossible any attempt at completeness. A bibliography which includes 170 titles

lffi prior to 1971 is contained in the excellent survey

ZIENKIEWICZ, O. C. (1970), The finite element method: from intuition to generality,
Appl. Mech. Rev. 23, 249-256.
The following books have been referred to in preparing the present text, except
for the three which appeared too recently (1972) for us to consult:
l. AoMON, S. (1965), Lectures on Elliptic Boundary Value Problems, Van Nos-
trand Reinhold, New York.
'
2. ARGYRIS, J. H. (1954-1955), Energy Theorems and Structural Analysis, Butter-
worth, London, 1960.
3. ARGYRIS, J. H. (1964), Recent Advances in Matrix Methods of Structural
Analysis, Pergamon Press, Elmsford, N.Y.
4. AUBIN, J. P. (1972), Approximation of E/liptic Boundary-Value Problems,
singular functions
Wiley, New York.
5. DESAI, C., and J. ABEL (1972), Introduction to the Finite Element Method,
\
Van Nostrand Reinhold, New York.
6. FADDEEV, D. K., and V. N. FADDEEVA (1963), Computational Methods of
- Linear Algebra, Dover, New York.

-
..._
1-- f-
1= 1- 7. KRASNOSEL'SKII, M. A., G. M. VAINIKKO, et al. (1969), Approximate Solution

1
!--1-
1--1- of Operator Equations, Nauka, Moscow (in Russian).
1
8. HOLAND, 1., and K. BELL, eds. (1969), Finite Element Methods in Stress Analy-
sis, Tapir, Trondheim, Norway.
11
040 60 80 100 200 400 1000 9. LIONS, J. L. (1969), Quelques méthodes de résolution iles problemes aux limites
non linéaires, Dunod, Paris.
1/ h
281
o
282 BlBLIOGRAPHY

10. LIONS, J: L., and E. MAGENES (1968), Problemes aux limites non homogenes
et applications, Dunod, París. 5. Numerical Solution of Partial Dfj
11. MIKHLIN, S. G. (1964), Variational Methods in Mathematical Physics, Perga- of Maryland), ed. B. Hubbard, 1
mon Press, Elmsford, N.Y. 6. The Mathematical Foundations ~
12. MIKHLIN, S. G. (1965), The Problem of the Mínimum of a Quadratic Func- Maryland at Baltimore), Acaden
tional, Holden-Day, San Francisco. 7. Numerical Solution of Field Prol
13. MI~tiLIN, S. G. (1971), The Numerical Performance of Variational Methods, Duke University, SIAM-AMS P
translated from the 1966 Russian edition, Wolters-Noordhoff, Groningen. 8. SM D Symposium on Computer·
14. NECAS, J. (1967), Les méthodes directes en théorie des équations el/iptiques, University of Waterloo, May 19'
Academia, Prague. 9. Finite Element Techniques in St,
15. ODEN, J. T. (1972), Finite Elements of Nonlinear Continua, McGraw-Hill, C. Brebbia, Southampton Unive
New York. 10. Conference on Variational Metht
16. PRZEMIENIECKI, J. S. (1968), Theory of Matrix Structural Analysis, McGraw- England, 1972.
Hill, New York. 11'. Conference on the Mathematics
17. SYNGE, J. L. (1957), The Hypercircle in Mathematical Physics, Cambridge University, England (to be publi
University Press, New York. 12. Proceedings of the Internationa
18. VAINBERG, M. M. (1964), Variational Methods for the Study of Nonlinear Methods in Engineering, Univers
Operators, Holden-Day, San Francisco. 13. Proceedings of the American Nu(
19. VISSER, M, (1968), The Finite Element Method in Deformation and Heat Con- 14. Proceedings of the First Interna¡
duction Problems, Delft, Holland. tures, Berlín, 1971.
20. WILKINSON, J. H. (1963), Rounding Errors in Algebraic Processes, Prentice- 15. Proceedings o/ the First Sympl
Hall, Englewood Cliffs, N.J. J. H. Goodier and H. J. Hoff, P
21. WILKINSON, J. H., and C. REINSCH (1971), Linear Algebra, Springer-Verlag, 16. Proceedings of the Symposium o
Berlin. 1969.
22. ZIENKIEWICZ, O. C. (1971), The Finite Elem.ent Method in Engineering Science, 17. Recent Advances in Matrix Met
2nd ed. McGraw-Hill, New York. J. T. Oden, R. H. Gallagher, an
23. ZIENKIEWICZ, O. C. and G. S.. HOLISTER, eds. (1965), Stress Analysis, Wiley, 1971. (First Japan-U.S. seminar
New York. 18. App/ication of Finite Element A1.
It will also be useful to list sorne of the recent conferences and symposia which Engineering, ISPRA, Italy, 1971.
have concentrated on the theory and applications of finite elements. Their published 19. Con/erence on Computer Oriente,
proceedíngs contain a large number of valuable papers. Those to which we have Alto Research Laboratories, Pal
referred specifically in the text will reappeai:' in the list of individual papers below, Journal of Computers and Struct~
with an indication such as "Wright-Patterson Conference 11."
20. Symposium on the App/ication oj
l. Proceedings ofthe First Conference on Matrix Methods in Structural Mechanics, Vanderbilt University, 1969.
WrÍiht-Patterson AFB, Ohio, 1965.
21. Symposium on Application of th
2. Proceedings of the Second Conference on Matrix Methods in Structural Swiss Society of Architects and
Mechanics, Wright-Patter~on AFB, Ohio, 1968.
22. National Symposium on Computt
3. Proceedings of the Third Conference on Matrix Methods in Structural Me-
Washington University, 1972.
chanics, Wright-Patterson AFB, Ohio, 1971 (to appear).
23. On General Purpose Finite Eler
4. Proceedings IUTAM Symposium, High Speed Computing of Elastic Struc-
Amer. Society of Mechanical Er
tures, Liege, Bel.gium, 1970.
24. Proceedings of the N ATO Advan

Problemes aux limites non 'homogenes
5. · Numerical Solution of Partial Differential Equations (SYNSPADE, University
·thods in Mathematical Physics, Perga-
6.
if the Minimum of a Quadratic Pune-
7.
·Performance of Varü/;t,ional Methods,
on, Wolters-Noordhoff,. Groningen. 8.
es en théorie des équations elliptiques,
9.
if Nonlinear Continua, McGraw-Hill,
10.
Matrix Structural Analysis, McGraw-
1L · Conference on the Mathematics of Finite Elements and Applications, Brunel
in ftt.!athematical Physics, Cambridge
12.
Methods for the Study of Nonlinear
13.
Method in Deformation and Heat Con-
14.
·rors in Algebraic Processes, Prentice-
15.
)71), Linear Algebra, Springer-Verlag,
16.
~lement Method in Engineering Science,
17.
:R, eds. (1965), Stress Analysis, Wiley,
18.
ecent conferences and symposia which
:ions of finite elements. Their published
19.
table papers. Those to which we have
in the list of individual papers below,
Jn Conference 11."
20.
fatrix Methods in Struc!_ural Mechanics,
21.
~e on Matrix Methods in Structural
io, 1968.
m Matrix Methods in Structural Me-
22.
1971 (to appear).
BIBLIOGRAPHY 283
of Maryland), ed. B. Hubbard, Academic Press, New York, 1971.
The Mathematical Foundations of the Finite Element Method. (University of
Maryland at Baltimore), Academic Press, New York, 1973.
Numerical Solution of Field Problems in Continuum Physics, ed. G. Birkhoff,
Duke University, SIAM-AMS Proceedings, Vol. 2, 1970.
SMD Symposium on Computer-Aided Engineering, ed. G. L. M. Gladwell,
University of Waterloo, May 1971.
Finite Element Techniques in Structural Mechanics, eds. H. Tottenham and
C. Brebbia, Southampton University Press, 1970.
Conference on Variational Methods in Engineering, Southampton University,
England, 1972.
University, England (to be published byAcademic Press), 1972.
Proceedings of the International Symposium on Numerical and Computer
Methods M Engineering, University of Illinois, 1971.
Proceedings of the American Nuclear Society Meeting, Boston, 1971.
Proceedings of the First International Conference on Nuclear Reactor Struc-
tures, Berlín, 1971.
Proceedings of the First Symposium on Naval Structural Méchanics, eds.
J. H. Goodier and H. J. Hoff, Pergamon Press, 1960.
Proceedings of the Symposium on Finite Element Techniques, ISD, Stuttgart,
1969.
Recent Advances in Matrix Methods of Structural Analysis and Design, eds.
J. T. Oden, R. H. Gallagher, and Y. Yainada, University of Alabama Press,
1971. (First Japan-U.S. seminar; the second will be held at Berkeley in 1972.)
App!ication of Finite Element A1ethods to Stress Analysis Problems in Nuclear
Engineering, ISPRA, Italy, 1971.
Conference on Computer Oriented Analysis of Shell Structures, Lockheed Palo
Alto Research Laboratories, Palo Alto, Calif., 1970 (papers to appear in the
Journal of Computers and Structures).
Symposium on the Application of Finite Element Methods in Civil Engineering,
Vanderbilt University, 1969.
Symposium on Application of the Finite Element Method in Stress Analysis,
Swiss Society of Architects and Engineers, Zurich, 1970.
National Symposium on Computerized Structural Analysis and Design, George
' Washington University, 1972.

L Speed Computing of Elastic Struc- 23. On General Purpose Finite Element Computer Programs, ed. P. V. Marcal,
Amer. Society of Mechanical Engineers.
24. Proceedings of the NATO Advanced Study Institute, Lisbon, 1971.
284 BIBLIOdRAPHY

BS. BABUSKA, J. (1971 ), Finite

2S . .Compuiational Approaches in Applied Mechanics, ASME Joint Computer Univ. Maryland.
Conference, Chicago, 1969.
B6. BABUSKA, J. (1973), The finit
The following bibliography contains those papers which are comment~d on in Numer. Math. 20, 179-192.
the text, together with many others which we have found to be valuable references.
B7. BACKLUND, J. (1971), Mixea
At this point especially, it must be repeated that a search of the literature would
plastic Piafes in Bending, (
yield a 01uch Iarger number .of important titles. This is particularly true of engi-
Sweden.
neering ~'~apers. The list below does indicate those journals in which the :finite
element method is strongly represented, and together with. the conference proceed- B8. BAUER, F. L. (1963), Optima
ings it should be a reasonable guide to the significant analytical and theoretical B9. BAZELEY, G.P., Y. K. CHI
work on the method. (l96S), Triangular elements
Al. AHMAD~ S., B. M. IRONS, andO. C. ZIENKIEWICZ (1968), Curved thick shell forming solutions, Wright-Pt
and membrane elements with particular reference to axisymmetric problems, BtO. BERGER, A., R. Scorr, and
Wright-Patterson ll. conditions in the finite eleme
A2. ALLMAN, D .. J. (1971), Finite element analysis of plate buckling using a Istituto Nazionale di Alta
mixed variational principie, Wright-Patterson III. Mathemat_ica, Academic Pr~
Bll. BIRKHOFF, G. (1969), Piecev
A3. ANDERHEGGEN, E. (1 970), A conforming triangular finite element plate
in polygons, Approximation ·
bending solution, Int. J. for Num. Meth. in Eng. 2, 2S9-264.
J. Schoenberg, Academic Pre
A4. ARGYRIS, J. H., and S. KELSEY (1963), Modern Fuselage Analysis and the
Bl2. BIRKHOFF, G. (1971), Num
Elastic Ai~cra/t, Butterworth, London.
Regional Conference Series, '
AS. ARGYRIS, J~ H., I. FRIED, and D. W. ScHARPF (1968), The Hermes eight Bl3. BIRKHOFF, G. (1972), Angula
element for the matrix displacement method, J. Royal Aero. Soc., 613-617. Th. 6, 21S-230.
A6. ARGYRIS, J. H., and I. FRIED (1968), The LUMINA element for the matrix Bl4. BIRKHOFF, G., and C. de Boo1
displacement method, J. Royql Aero Soc., Sl4-S17. J. o/ Math. and Mech. 13, 82
A7. ARGYRIS, J. H., O. E. BRONLUND, I. GRIEGER, and M. SORENSEN (1971), BIS. BiRKHOFF, G., and C. DE Bo1
A survey of the application of finite element methods to stress analysis and approximation, Approx
problems with particular emphasis on their application to nuclear engi- Elsevier, Amsterdam.
neering problems, /SPRA -Con/erence. ·
Bl6. BIRKHOFF, G., c. DE BooR, R
A8. AUBIN, J~ P. (1967), Approximation des espaces de distributions et des Ritz approximation by piece
opérateurs différentiels,, Bull. Soc. Math. France, Mémoire 12. 13, 188-203.
A9. AuBIN, J. P. (1968), Eva1uation des erreurs de troncature des approxima- 817. BIRKHOFF, .G., and G. FIX (l
tions des espaces de Sobolev, J. Math. Anal. Appl. 21, 3S6-368. nometric polynomials, Indiar.
A10. AUBIN, J. P., and H. BuRCHARD (1971), Sorne aspects of the method of the 818. BIRKHOFF, G., and G. Ftx (
hypercircle applied to elliptic variational problems, SYNSPADE, 1-67. elliptic problems, Duke Unü
Bl. BABUSKA, l. (1961), Stability of the domain of definition ... (in Russian), Bl9. BIRKHOFF, G., M. H. SCHUL~
Czech. Math. J. 11(86), 76-10S, )6S-203. · interpolation in one and two
ential equations, Numer. Mat
B2. BABUSKA, l. (1970), Approximation by hill functions, Tech. Note 648, Univ.
Maryland. B20. BLAIR, J. J. (1971), Bounds fe
elliptic PDEs when the boun
B3. BABUSKA, J. (1970), Finite element method for domains with corners,
Computing 6, 264-273. B21. DE BooR, C. (1968), The mett
solution of two-point bounda
B4. BABUSKA, I. (1971), Error bounds for the finite element method, Númer. University of Michigan.
Math. 16, 322.:._333.

Mechanics, ASME· Joint Computer
~ papers which are commentc;:d on in
have found to be valuable references.
that a search of the literature would
les. This is particularly true of engi-
~ those journals in w&ich the finite
ogether with the conference proceed-
significant analytical and theoretical
ENKIEWICZ (1968), Curved thick shell
r reference to axisymmetric problems,
t analysis of plate buckling using a
1tterson 111.
ning trilmgular finite element plate
?th. in Eng. 2, 259-264.
), Modern Fuselage Analysis and the
ScHARPF (1968), The Hermes eight
1ethod, J. Royal Aero. Soc., 613-617.
'he LUMINA element for the matrix
~oc.,514-517.
GRIEGER, and M. SORENSEN (1971),
element methods to stress analysis
n théir application to nuclear engi-
des espaces de distributions et des
th. France, Mémoire 12.
rreurs de troncature des approxima-
·. Anal. Appl. 21, 356-368.
), Sorne aspects of the method of the
tnal problems, SYNSPADE, 1-67.
omain of definition ..-:(in Russian),
:o3.·
' hill functions, Tech. Note 648, Univ.
method for domains with corners,
r the finite element method, Nume,r.
BIBLIOGRAPHY 285
B5. BABUSKA, J. (1971), Finite element method with penalty, Rept. BN-710,
Univ. Maryland.
B6. BABUSKA, l. (1 973), The finite element method with Lagrangian multipliers,
Numer. Math. 20, 179-192.
B7. BACKLUND, J. (1971), Mixed Finite Element Analysis of Elastic and Elasto-
plastíc Plates in Bending, Chahners lnstitute of Technology, Goteborg,
Sweden.
B8. BAUER, F. L. (1963), Optimally scaled matrices, Numer. Math. 5, 73-87.
B9. BAZELEY, G. P., Y. K. CHEUNG, B. M. IRoNs, and O. C. ZIENKIEWICZ
(1965), Triangular elements in plate bending-conforming and noncon-
forming solutions, Wright-Patterson l.
Bto. BERGER, A., R. Scorr, and G. STRANG (1972), Approximate boundary
conditions in the finite element method, Symposium on Numerical Analysis,
Istituto Nazionale di Alta Matematica, Rome; to appear in Symposia
Mathemat.ica, Academic Press, New York.
B11. BIRKHOFF, G. (1969), Piecewise bicubic interpolation and approximation
in polygons, Approximation with Special Emphasis on Spline Functions, ed.
l. Schoenberg, Academic Press, New York, 185-221.
B12. BIRKHOFF, G. (1971), Numerical solutions of elliptic equations, SJAM
Regional Conference Series, Vol. l.
B13. BIRKHOFF, G. (1972), Angular singularities of elliptic problems, J. Approx.'
Th. 6, 215-230.
B14. BIRKHOFF, G., and C. deBooR (1964), Error bounds forspline interpolation,
J. of Math. and Mech. 13, 827-836.
B15. BtRKHOFF, G., and C. DE BooR (1965), Piecewise polynomial interpolation
and approximation, Approximation of Functions, ed. H. L. Garabedian,
Elsevier, Amsterdam. ·
Bl6. BIRKHOFF, G., C. DE BooR, B. SwARTZ, and B. WENDROFF (1966), Rayleigh-
Ritz approximation by piecewise cubic polynomials, SIAM J. Num. Anal.
13, 188-203.
.817. BIRKHOFF, G., and G. F1x (1967), Rayleigh-Ritz approximation by trigo-
nometric polynomials, Jndian J. of Math. 9, 269-277.
.BIS. BIRKHOFF, G., and G. FIX (1970), Accurate eigenvalue computations for
elliptic problems, Duke University SIAM-AMS Symposium.
B19. BIRKHOFF, G., M. H. ScHULTZ, and R. VARGA (1968), Piecewise Hermite
interpolation in one and two variables with applications to partial differ-
ential equations, Numer. Math. 11, 232-256.
B20. BLAIR, J. J. (1971), Bounds for the change in the solutions of second order
elliptic PDEs when the boundary is perturbed, to appear.
B21. DE BooR, C. ( 1968), The method of projections as -applied to the numerical
solution of two-point boundary value problems using. cubic splines, Thesis,
University of Michigan.
286 BIBÜO,GRAPHY

B22. DE BooR, C. (1968), On loca1 spline approximation by moments, J. of Math. C6. CIARLET, P. G., and P. A. R.a
and Mech. 11, 729-736. interpolation in Rn with app
B23. DE BooR, C., and G. F1x (1972), Spline approximation by quasi-interpolants, Rat. Mech. Anal. 46, 177-199
to appear in J. of Approx. Theory. C7. CIARLET, P. G., and P. A. RA'
Bi4. DE BooR, C., and B. SWARTZ (1972), Collocation at -Gaussian points, elements, with applications t
fos Alamos Rept. 72-65. in Appl. Mech. and Eng. 1, 2
B25. ~RAMBLE, J. H. (1971), Variational methods for the numerical solution of cs. CLouoH, R. W., and J.- L. Te
el1iptic problems, Lecture notes, Chalmers Institute of Technology, Gote- for analysis of plates in bend
borg, Sweden. C9. CLOUGH, R. W., and C. A. FJ
B26. BRAMBLE, J. H., T. DuPóNT, and V. J'HOMÉE (1972), Projection methods for analysis of plate bending,
for Dirichlet's problem in approximating polygonal domains with boundary ClO. CLOUGH, R. W. (1969), Comp
value corrections, MRC Tech. Rept. 1213, Univ. Wisconsin. bilt Symposium.
B27. BRAMBLE, J. H., and S. R. HILBERT (1970), Estimation of linear functionals Cll. CoURANT, R. (1943), Variati<
on Sobolev spaces with application to Fourier transforms and spline irtter- equilibrium and vibrations, 1
polation, SIAM J. Num. Anal. 1, 113-124. CoWPER, G. R. (1972), CUI
C12.
B28. BRAMBLE, J. H., and S. R. HILBERT (1971), Bounds for a class of linear Shells of Arbitrary Shape, N:
functionals with applications to Hermite interpolation, Numer. Math. 16, CowPER, G. R. (1972), Gaus~
Cl3.
362-369. script.
B29. BRAMBLE, J. H., and J. OsaORN (1972), Rate of convergence estimates for CoWPER, G. R., E. KosKo,
C14.
nonselfadjoint eigenvalue approximations, MRC Tech. Rept. 1232, Univ. Static and dynamic applicati<
Wisconsin. ing element, AIAA J. 1, 1957-
B30. BRAMBLE, J. H., andA. H. ScHATZ (1970), Rayleigh-Ritz-Galerkin methods
Dl. DEMJANOVIC, J. K. (1964), Tl
for Dirichlet) problem using subspaces without boundary conditions, matical physics, Dokl. Akad.
Comm. Pure Appl. Math. 23, 653-675.
D2. DEMJANOVIC, J. K. (1966), j
B3l. BRAMBLE, J. H., and A. H. ScHATZ (1971), On the numerical solution of method in elliptic problems,
elliptic boundary value problems by least square approximation of the Dokl. 1.
data, SYNSPADE, 107-133.
D3. DENDY, J. (1971), Thesis, Ric
B32. BRAMBLE, J. H., and M. ZLAMAL (1970), Triangular elements in the finite
element method, Math. of Comp. 24, 809-821. D4. DESCLOUX, J. (1972), On fini
260-265.
Cl. CARLSON, R. E., and C. A. HALL (1971), Ritz approximations to two-
D5. DESCLOUX, J. ( 1970), On the
dimensional bounda,ry value problems, Numer. Math. 18, 171-181.
1 Numer. Math. 15, 371-381.
C2. CÉA, J. (1964), Approximation variationelle des problemes aux limites, DouoLAS, J., and T. Duror
D6.
Ann. lnst. Fourier, 14, 345-444. problems, SIAM J. Numer. /:
C3. €HERNUKA, M. w., G. R. COWPER, G. M. LINDBERG, and M. n~ ÜLSON D7. DouGLAS, J., and T. DuPON1
(1972), Finite elemeni analysis.,of plates with curved edges, In t. J. for Num. for quasilinear parabolic equ:
Methods in Eng. 4, 49-65.
D8. DOUGLAS, J., and T. DUPONT
C4. CIARLET, P. G., M. H. ScHUL~z. and R. S. VARGA (1967), Numerical cients, unpublished.
methods of higher order accuracy for nonlinear bo.undary value problems,
D9. DouGLAS, J., atÍd T. DuPONT
Numer. Math. 9, 394-430; !:"umer. Math. 13, 51-77.
ti~ns with nonlinear boundar:
C5. CIARLET, P. G., and C. WAGSCHAL {1971), Multipoint Taylor.formulas and
DIO. DuP~NT, T. (1973), Galerk
applications to the finite element method, Numer. Math. 11, 84-100.
example, SIAM J. Numer. A~

pproximation by m?ments, J. of Math.
1e approximatio:o by quasi-interpolants,
72), Collocation at Gaussian points,
\..
. í\
nethods for the numer:ical solution of
.almers Institute of Technology, Gote-
'. THOMÉE (1972), Projection methods
ting polygonal domains with boundary
1213, Univ. Wisconsin.
1970), Estimation of linear functionals
:o Fourier transforms and spline inter-
3-124.
· (1971), Bounds for a class of linear
rmite interpolation, Numer. Math. 16,
72), Rate of convergence estimat~s for
ations, MRC Tech. Rept. 1232, Univ.
970), Rayleigh-Ritz-Galerkin methods
)paces without boundary conditions,
5.
(1971), On the numerical solution of
,Y least square approximation of the
970), Triangular elements in the finite
'809-821.
(1971), Ritz approximations to two-
ns, Numer. Math. 18, 171-181.
' Numer. Math. 15, 371-381.
·iationelle des problemes aux ·limites,
G. M. LINDBERG, and M. n~ ÜLSON
ttes with curved edges, In t. J. for Num.
and R. S. VARGA (1967), Numerical
•r nonlinear boundary value próblems,
1ath. 13, 51-7i.
1971), Multipoint Taylor formulas and
thod, Numer. Math. 17, 84-100.
BIBLIOGRAPHY 287
C6. CIARLET, P. G., and P. A. RAVIART (1972), General Lagrange and Hermite
interpolation in Rn with applications to the finite element method, Arch.
Rat. Mech. Anal. 46, 177-199.
C7. CIARLET, P. G., and P. A. RAVIART (1972), Interpolation theory over curved
elements, with applications to finite element methods, Computer Methods ·
in Appl. Mech. and Eng. 1, 217-249.
CS. CLOUGH, R. W., and J; L. ToCHER (1965), Finite element stiffness matrices
for analysis of plates in bending, Wright-Patterson l.
C9. CLOUGH, R. W., and C. A. FELIPPA (1968), A refined quadrilateral element
for analysis of plate bending, Wright-Patterson JI.
ClO. CLOUGH, R. W. (1969), Comparison of three-ditnensional elements, Vander-
bilt Symposium.
C11. CoURANT, R. (1943), Variational methods for the solution of problems of
equilibrium and vibrations, Bull. Amer. Math. Soc. 49, 1-23.
Cl2. COWPER, G. R. (1972), CURSHL: A High-Precision Finite Element for
Shells of Arbitrary Shape, National Research Council of Canada Report.
C13. CowPER, G. R. (1972), Gaussian quadrature formulas for triangles, manu-
script.
Ct4. CowPER, G, R., E. KosKo, G. M. LINDBERG, and M. D. ÜLSON (1969),
Static and dynamic applications of a high-precision triangular plate bend-
ing element, AIAA J. 1, 1957-1965.
Dl. DEMJANOVIC, J. K. (1964), The net method for sorne problems in mathe-
matical physics, Dokl. Akad. Nauk SSSR 159, Soviet Math. Dokl. 5.
02. DEMJANOVIC, J. K. (1966), Approxirnation and convergence of the net
··method in elliptic problems, Dokl. Akad. Nauk SSSR 170, Soviet Math.
Dokl. 7.
D3. DENDY, J. (1971), Thesis, Rice University.
04. DESCLOUX:, J. (1972), On finite element matrices, S/AM J. Num. Anal. 9,
260-265.
D5. DESCLOUX, J. (1970), On the numerical integration of the heat equation,
06. DouoLAS, J., and T. DuPONT (1970), Galerkin methods for parabolic
problems, SIAM J. Numer. Anal. 4, 575-626.
D7. DouGLAS, J., and T. DuPONT (1972), A finite element collocation method
for quasilinear parabolic equations, manuscript.
08. DouGLAS, J., and T. DuPONT (1972), manuscript on interpolation of coeffi-
cients, unpublished.
D9: DouGLAS, J., and T. DuPONT (1973), Galerkin methods for parabolic equá-
tions with nonlinear boundary conditions, Numer. Math., 20, 213-237.
\
DIO. DuPONT, T. (1973), Galer~in methods for first-order hyperbolics: an
example, SJAM J. Numer. Anal.; to appear.
288 BIDLIOQRAPHY
F16. FRIED, l. (1971 ), Basic comput
D1 C DUPUIS, G., and J. J. GoEL-(1969), A curved element for thin elastic shel1s, of shells, lnt. J. Solids Struct.
Tech. Rept., Brown Univ.
F17. FRIED, l. (1971), Discretizati
D12. DUPUIS, 'G., and J. J. GoEL (1969), Eléments finis raffinés en élasticité finite elements, AIAA J. 9; 20
bidimensionelle, ZAMP 20, 858-881.
F18. FRIED, l. (1972), The /1 and loo
D13. DUPUIS, G., and J. J. GoEL (1970), Finite element with high degree of University.
ll~gularity, lnt. J. for Num. Meth. in Eng. 2, 563-577.
F19. FRIEDRICHS, K. (1928), Die
El. E~GATOUDIS, 1., B. I~ONS,andO. C. ZIENKIEWICZ (1968), Curved isopara-
Theorie ·der elastischen Platte
m~tric quadrilateral elements for finite element analysis, Int. J. Solids
Struct. 4, 31-42. F20. FRIEDRICHS, K. 0., and H. B.
generalized Neumann problen
Fl. FELIPPA, C. A. (1966), Refined finite element analysis of linear and non-
tia! Equations, ed. J. Bramble,
linear two-dimensional structures, Rept., Univ. California at Berkeley.
F2. FELIPPA, C. A. (1969), Analysis of plate bending problems by the finite Gl. GALLAGHER, R. H., and A. K
element method, SESM Rept. (Dept. Civil Eng.), Univ. California at ment elastoplastic analysis, Bt
Berkeley. in Vanderbilt and Jap~n-U.S

F3. FELIPPA, C. A., and R. W. CLOUGH (1970), The finite element method in G2. GEORGE, A. (1971), Computer
solid mechanics, Duke University SIAM-AMS Symposium, 210-252. Thesis, Stanford University.

F4. Fix, G. (1968), Orders of convergence of the Rayleigh-Rítz and Weinstein- G3. GEORGE, A. (1971), Block elim
Bazley methods, Proc. Nat. Acad. Sci. 61, 1219-1223. manuscript.

F5. FIX, G~ Ü969), Higher-order Rayleigh-Ritz approximations, J. Math. G4. GoEL, J. J. (1968), Constructi<
Mech. 18, 645-658. of Ritz's method, Numer. M1

F6. F1x, G., and G. STRANG (1969), Fourier analysis of the finite element G5. DIGUGLIELMO, F. (1969), C(
method in Ritz-Galerkin theory, Studies in Appl. Math. 48, 265-273. Sobolev sur des réseaux en si

F7. Fix, G., and S. GuLATI (1971), Computational problems arising from the Hl. HANNA, M. S., and K. T. SN
use of singular functions, Rept., Harvard Univ. problem in piecewise smoot
575-593.
F8. Fix, G., and N. NASSIF (1972), On finite element approximations to time-
dependen! problems, Numer. Math. 19, 127-135. H2. HARRICK, l. l. (1955), On the
boundary of a region by fun'
F9. FIX, G., S. GULATI, and G.-1. WAKOFF (1972), On the use of singular func- 37(79), 353-384.
tions with the finite element method, J. Comp. Physics, to appear.
H3. HERBOLo, R. J., M. H. Scm
FlO. Fax, L., P .. HENRICI, and C. MoLER (1967), Approximation and bounds for schemes for the numerical sol
eigenvalues of elliptic óperators, SIAM J. Numer. Anal. 4, 89-102. tional techniques, Aequ. Math
Fil. FRAEIJS DE ,VEUBEKE, B. (1965), Displacement and equilibrium modeis in H4. HERRMANN, L. R. (1967), Fini1
the finite element method, Chap. 9 of Stress Analysis, eds. O. C. Zienkiewicz Mech. Div. ASCE 94, 13-25.
and G. S_. Holister, Wiley, New York.
H5. HILTON, P. D., and J. HUTc
F12. FRAEI.JS DE VEUBEKE, B. (1968), A conforming finite element for plate cracked· plates, Eng. Fract: M
bending, Int. J~ Solids Structures 4, 96-108.
H6. HULME, B. L. (1968), lnterpol:
F13. FREDERICKSON, P. O. (1971), Generalized triangular splines, Math. Rept. 7, 18, 337-342.
Lakehead Univ., Canada.
11. IRONS, B. M. (1966), Engineel
F14. FRIED, l. (1971), Condition of fi.nite element matrices generated from stiffness methods, AIAA J. 4,
nonuniform meshes, AIAA J. 10, 219-221.
12. IRONS, B. M. (1968), ~oundotl
F15. FRIED, l. (1971), Accuracy of finite element eigenproblems, J. of Soun,d and
Vibration 18, '289-295. ~308-12.

curved element for. thin elastic shells,
Eléments finis." raffinés en élasticité
. Finite element with high degree of
'::ng. 2, 563-577.
1\
ZIENKIEWICZ (1968), G4rved isopara-
nite element analysis, Int. J. Solids
' element analysis of linear and non-
~pt., Univ. California at Berkeley.
Jlate bending problems by the finite
pt. Civil Eng.), Univ. California at
(1970), The finite element method in
tM-AMS' Symposium, 210-254.
~ of the Rayleigh-Ritz and Weinstein-
i. 61, 1219-1223.
eigh-Ritz approximations, J. Math.
ourier analysis of the finite element
dies in Appl. Math. 48, 265-273.
putational problems arising from the
•ard Univ.
nite element approximations to time-
19, 127-135.
:p (1972), On the use of singular func-
J. Comp. Physics, to appear.
1967), Approximation and bounds for
1\1 J. Numer. Anal. 4, 89-102. '
llacement and equilibrium modc~s in
Stress Analysis, eds. O. C. Zienkiewicz
k.
conforming finite elément for plate
6-108.
1
ized triangular splines, Math. Rept. 7,
te element matrices generated from
f-221.
ement eigenproblems, J. of Sound and
BIBLIOGRAPHY 289
F16. FRIED, l. (1971), Basic computational problems in the finite element analysis
of shells, lnt. J. Solids Struct. 7, 1705-1715.
F17. FRIED, l. (1971), Discretization and computational errors in high-order
finite elements, AIAA J. 9, 2071-2073.
Fl8. FRIED, l. (1 972), The 12 and /"" condition numbers ... , Conference at Brunel
University.
Fl9. FRIEDRICHS, K. (1928), Die Randwert- und I;:igenwertprobleme aus der
Theorie der elastischen Platten, Math. Ann; 98, 205-247.
F20. FRIEDRICHS, K. 0., and H. B. KELLER (1966), A finite difference scheme for
generalized Neumann problems, in Numerical Solutions of Partial Di/feren-
tial Equations, ed. J. Bramble, Academic Press,· New York.
01. 0ALLAGHER, R. H., andA. K. DHALLA (1971), Direct flexibility finite ele-
ment elastoplastic analysis, Berlin Symposium (see also papers by Oallagher
in Vanderbilt and Jap:;¡.n-U.S. Symposia).
02. OEORGE, A. (1971), Computer implementation of the finite element method,
Thesis, Stanford University.
O 3. 0EORGE, A. ( 1971 ), Block elimination of finite element systems of equations, ·
manuscript.
G4. O o EL, J. J. (1968), Construction of basic functions for numerical utilization
of Ritz's méthod, Numer. Math. 12, 435-447.
05. DIOUGLIELMO, F. (1969), Construction d'approximations des espaces de
Sobolev sur des réseaux en simplexes, Calco/o 6, 279-331.
Hl. HANNA, M. S., and K. T. SMITH (1967), Sorne remarks on the Dirichlet
problem in piecewise smooth domai,ns, Comm. Pure Appl. Math. 20,
?75-593.
H2. HARRICK, l. l. (1955), On the approximation of functions vanishing on the
boundary of a region by functions of a special form, Mat. Sbornik, N.S.
37(79), 353-384.
H3. HERBOLO, R. J., M. H. ScHULTZ, and R. S. VARGA (1969), Quadrature
schemes for the numerical solution of boundary value problems by varia-
tional techniques, Aequ. Math. 3, 96-119.
H4. HERRMANN, L. R. (1967), Finite-element bending analysis for plates, J. Eng.
Mech. Div. ASCE 94, 13-25. ,
H5. HILTON, P. D., and J. HUTCHINSON (1971), Plastic intensity factors for
cracked plates, Eng. Fract: Mech. 3, 435--451.
H6. HuLME, B. L. (1968), Interpolation by Ritz approximation, J. Math. Mech.
18, 337-342.
1 Il. IRONS, B. M. (1966), Engineering applications of numerical integration in
l
J stiffness methods, AIAA J. 4, 2035-2037. ·
l 12. lRONS, B. M. (1968), Rpundoff:criteria in direct stiffness solutions, AIAA J.
6, 1308-12. ~
1
290 BIBLIOGRAPHY

13. 1RoNs, B. M. (1969), Economical computer techniques for numerically M4. McLAY, R. W. (1968), Complet
integrated finite elements, lnt. J. for Num. Meth. in Eng. 1, 201-203. element displacement functions-
5th Aerospace Science Meeting.
, 14. 1RONS, B. M. (1970), A frontal solution program for finite element analyses,
Jnt. J. /ór Num. Meth. in Eng. 2, 5-32. M5. McLAY, R. W. (1971), On cert
method, Trans. ·o! the ASME 58-
15. 1RoNs, B. M. (1971), Quadrature rules for brick based finite elements,
AIAA J. 9, 293-294. M6. MELOSH, R. J. (1966), Basis for
ness method, AIAA J. 34, 153-l
16. 1~'Q,NS, B. M., andA. RAZZAQUE (1971 ), A new formulation for plate bending
eletnents, manuscript. M7. MIKHLIN, S. G. (1960), The sh
Dokl. 1, 1230-1233.
17. 1RONS, B. M., E. A. DE ÜLIVEIRA, and O. C. ZIENKIEWICZ (1970), Comments
on the paper: Theoretical foundations of the finite element method, lnt. M8. MILLER, C. (1971), Thesis, M.I.'
J. Solids Struct. 6, 695-697. M9. MITCHELL, A. R., G. PHILLIPS, a1
18. IRWIN, G. R. (1960), Fracture mechanics, Symposium on Naval Structural in the finite element method, J.
Mechanics. M lO. MORLEY, L. S. D. (1969), A ffiQ(
K l. KELLOGG, B. (1970), On the Poisson equation with intersecting interfaces, stress concentration problems i
Tech. Note BN-643, Univ. Maryland. Solids 17, 73-82.
K2. KELLOGG, B. (1971), Singularities in interface problems, SYNSPADE, N l. NITSCHE, J. (1968}, Ein Kriteriu:
351-400. . Verfahrens, Numer. Math. 11, 3•
N2. NITSCHE, J. (1968), Bemerkunge
K3. KONDRAT'Ev, V. A. (1968), Boundary problems fcr elliptic equations with
Verfahren, Math. Zeit. 106, 327,
conical or an~ular points, Trans. Moscow Math. Soc. 11.
N3. NITSCHE, J. (1970), Über ein V
K4. KouKAL, S. (1970}, Piecewise polynomial interpolations and their applica- Problemen bei Verwendung von
tions to partial differential equations, Czech. Sbornik VAJ.(Z, Brno, 29-38. unterworfen sind, Abh. Math. Se
K5. KRASNOSELSKII, M. A. (1950), The convergence of the Galerkin method for N4. NITSCHE, J. (1970), Lineare S¡
nonlinear equations, Dokl. Akad. Nauk SSSR 73, 1121-1124. Ritz für elliptische Randwertpro
K6. KRATOCHVIL, J., A. ZENISEK, and M. ZLAMAL (1971), A simple algorithm N5. NITSCHE, J. (1971), A projecti'
for the stiffness matrix of triangular plate bending finite elements, Int. J. su bspaces with almost zero bou
for Num. Meth. in Eng. 3, 553-563.
N6. NITSCHE, J., andA. ScHATZ (19
K7. KREISS, H. O. (1971), Difference Approximations for Ordinary Differential tions on spline-subspaces, Appli
Equations, Computer Science Department, Uppsala U~iversity.
01. OGANESJAN, L. A. (1966), Com
Ll. LAASONEN, P. (1967), On the discretization error of the Dirichlet problems under improved approximation
in a plane r~gion with corners, Ann. Acad. Scient. Fennicae 408, 3-15. 1146-1150~
L2. LEHMAN, R. S. (1959), Developments near an analytic corner of solutions 02. ÜGANESJAN, L. A., and L. A. R1
of elliptic partial differential equations, J. Math. Mech. 8, 727-760. vergence of variational differen•
L3. LINDBERG, G. M., and M. D. ÜLSON (1970), Convergencestudies of eigen- tions in two-dimensional regio
value solutlons using two finite plate bending elements, lnt. J. for Num. Matem. 9, 1102-1120.
Mcth. in Eng. 2, 99-116. 03. ÜLIVEIRA, E. A. DE (1968), The
Ml. MARCAL, P. V. (1971), Finite element analysis with nonlinearities-theory method, /nt. J. Solids Struct. 4,
and practice, Fir'st Japan-U.S. Seminar. ~N, M. D., and G. M. LIND
M2. MARTIN, H. C. (1971), Finite elements anc;i the analysis of geometrically shells with a doubly-curved triar
nonlinear problems, First Japan-U.S. Seminar. 299""7318.
M3. McCARTHY, C., and G. STRANG (1973), Optimal conditioning of matrices, Pl. PETERS, G., and J. H. WILKINS•
S/AM J. Num. 1-nal., to appear. eigenproblem, SIAM J. Num. A;

1 computer technÍques for numerically
r-or Num. Meth. i~ Eng. 1, 201-203.
1tion program for finite element analyses,
-32.
rules for brick based finite elements,
n l ), A new formulatio;h, for plate bending
nd 0. C. ZIENKIEWICZ (1970), Comments
tions of the finite element method, Int.
:hanics, Symposium on -Naval Structural
::m equation with intersecting interfaces,
Id.
~ in interface problems, _S YNSPA DE,
uy problems fcr ellipti~ equations with
toscow Math. Soc. 17.
1omial interpolations and their applica-
ns, Czech. Sbornik VAA;Z, Brno, 29-38.
:onvergence of the Galerkin method for
lauk SSSR 73, 1121-1124.
\11. ZLAMAL (1971), A simple algorithm
1r plate bending finite elem~nts, Int. J.
pproximations for Ordinary Differential
·tment, Uppsala U~iversity.
tization error of the Dirichlet problems
'l. Acad. Scient. Fennicae 408, 3-15.
ts near an analyti¿ corner of solutions
ons, J. Math. Mech. 8, 727-76fi.
-: (1970), Convergence studies ofeigen-
tte bending elements, ]nt. J. for Num.
nt analysis with nonlinearities-theory
rtar.
mts and the ana1ysis of geometrically
:; Seminar.
73), Optimal conditioning of matrices,
l
BffiLIOGRAPHY 291
M4. MeLAY, R. W. (1968), Completeness and convergence properties of finite
element displacement functions-a general treatment, AIAA Paper 67-143,
5th Aerospace Science Meeting.
M5. McLAY, R. W. (1971), On certain approximations in the finite-element
method, Trans.·ofthe ASME 58-61.
M6. MELOSH, R. J. (1966), Basis for derivation of matrices for the direct stiff-
ness method, AIAA J. 34, 153-170.
M7. MIKHLIN, S. G. (1960), The stability of the Ritz method, Soviet Math.
Dokl. l, 1230-1233.
M8. MILLER, C. (1971), Thesis, M.I.T.
M9. MITCHELL, A. R., G. PHILLIPS, and R. WACHPRESS (1971), Forbidden shapes
in the finite element method, J. /nst. Maths. Appl. 8.
MIO. MoRLEY, L. S. D. (1969), A modification of the Rayleigh-Ritz method for
stress concentration problems in elastostatics, J. of Mech. and Phys. o!
Solids 17, 73-82.
Nl. NITSCHE, J. (1968), Ein Kriterium für die Quasi-OptimaliHit des Ritzschen
Verfahrens, Numer. Math. 11, 346-348.
N2. NITSCHE, J. (1968), Bemerkungen zur Approximationsgüte bei projektiven
Verfahren, Math. Zeit. 106, 327-331.
N3. NITSCHE, J. (1970), Über ein Variationsprinzip zur Losung von Dirichlet
Problemen bei Verwendung von Teildiumen, die keinen Randbedingungen
unterworfen sind, Abh. Math. Sem. Univ. Hamburg 36.
N4. NITSCHE, J. (1970), Lineare Spline-Funktionen und die Methoden von
Ritz für elliptische Randwertprobleme, Arch. Rat. Mech. Anal. 36, 348-355.
N5. NITSCHE, J. (1971), A projection method for Dirichlet problems using
subspaces with almost zero boundary conditions, manuscript.
N6. NITSCHE, J., andA. ScHATZ (1972), On local approximation of L2-projec-
tions on spline-subspaces, Applicable Analysis 2, 161-168.
01. ÜGANESJAN, L. A. (1966), Convergence of variational difference schemes
under improved approximation to the boundary, Soviet Math. Dokl. 7,
1146-1150~
02. ÜGANESJAN, L. A., and L. A. RUKHOVETS (1969), Study of the rate of con-
vergence of variational difference schemes for second-order elliptic equa-
tions in two-dimensional regions with smooth boundaries, Zh. Vychisl.
Matem. 9, 1102-1120.
03. ÜLIVEIRA, E. A. DE (1968), Theoretical foundations of the finite element
methód, Int. J. Solids Struct. 4, 929-952.
ÜLSON, M. D., a~ M. LINDBERG (1971), Dynamic analysis of shallow
• 04.
shells with·a doubl;-~~~ed triangular finite element, J. Sound Vibration ·19,
299-318.
Pl. PETERS, G., and J. H. WILKiNSON (1970), Ax J..Bx and the generalized
eigenproblem, SIAM J. Num. Anal. 7, 479-492.
292 BIBLIOGRAPHY

, P2.·,' 'PETERS, G., and J. H. WILKINSON (1971), Eigenvalues of Ax = J.Bx with

band symmetric A and B, Comput. J. 14. S6. STRANG, G. (1972), Approx
Math. 19, 81-98.
P3. 'PHI~LIPS, Z., ~nd D. V. PHILLIPS (1971), A.n automatic generation scheme
for plane and curved surfaces by isoparametric coordinates, Int. J.for Num. S7. STRANG, G. (1972), Variatio
Meth. in Eng. 3, 519-528. land Symposium.

P4. PIA~, T. H. H.,. and P. TONO (1969), Basis of finite element methods for SS. STRANG, G. (1973), Piecewi
solí~,, continua, Int. J. for Num. Meth. in Eng. 1, 3-28.
Bull. Amer. Math. Soc. 19, 1

PS. PIAN,' T. H. H. (1970), Finite element stiffness methods by different varia- S9. STRANG, G., and A. E. BE
tional principies in elasticity, Duke University SIAM-AMS Symposium, change in domain, Proc. A~
253-271. ' Berkeley.
P6. PIAN, T. H. H., P. ToNO, and C. H. LuK (1971), Elastic crack analysis by StO. STRANG, G., and G. FIX (1
a finite element hybrid method, Wright-Patterson Ill. method, Proc. GIME Sumn

P7. PtERCE, J. G., and R. S. VARGA (1972), Higher order convergence results Sil. SWARTZ, B., and B. WENDRC
for the Rayleigh-Ritz method applied to eigenvalue problems I, SIAM J. Math. of Comp. 23, 37-50.
Num. Anal. 9, 137-151. Tl. TAYLOR, R. L. ( 1972), Once
P8. PóLYA, G. (1952), Sur une interprétation de la méthode des différences analysis, Int. J. for Num. M
finies, qui peut fournir des bornes supérieures o u inférieures, Comptes T2. THOMÉE, V. (1964), Elliptic
Rendus 235, 995-997. Diff. Eqns. 3, 301-324.
P9. PRAGER, W., and J. L. SYNGE (1947), Approximations in elasticity based T3. THOMÉE, V. (1971), Polygonl
on the concept of function space, Quart. Appl. Math. 5, 241-269. MRC Tech. Rept. 1188, Un:
1

PI O. PRAGER, W. (1968), Variational principies for elastic plates with relaxed T4. TONO, P. (1969), Exact Sl
continuity requirements, Int. J. Solids Struct. 4, 837-844. method, AIAA J. 1, 178-18
PI J. PRICE, H. S., and R. S. VARGA (1970), Error bounds for semidiscrete Galer- TS. TONG, P. (1971), On the nu
kin approximations of parabolic problems, Duke University SIAM-AMS · Waterloo Conference.
Symposium, 14-94. T6. ToNO, P., and T. H. H. PI
Rl. RAI, A. K., and K. RAJAIAH (196J), Polygon-circle paradox of simply sup- in solving linear elastic pr<
ported thin plates under uniform pressure, AIAA J. 6, 155-156. T7. · ToNO, P., and T. H. H. PI
R2. REID, J. K. O972), On the construction and convergence of a finite-element by the assumed stress metl
solution of Laplace's equation,J. Inst. Maths. Appl. 9, 1-13. T8. ToNo, P., T. H. H. PIAN, a
SI. SANDER, G. (1,970), Application of the dual analysis principie, IUTAM frequencies by the finite •
Symposium. masses, J. Comp. Struct. l,

S2. ScnoENBERa, l. J. (1946), Contributions to the problerofpproxit;nation T9. TREFFTZ, E. (1926), Ein ~
of equidistant data by analytic functions, Quart. Appl Math. 4, 45-99, Congress Applied Mechanic
. 112..:Í41. T10. TURNER, M. J., R. w. CL
S3. ScHULTZ, M. H. (1969), R~yleigh-Ritz methods for multidimensional Stiffness and deflection am
problems, SIAM J. Num. Anal. 6, 523-538. V l. VAINIKKO, G. M. (1964), A
\

S4. ScHuLTZ, M. H. (1971), L2

error bounds for the Rayleigh-Ritz-Galerkin in the eigenvalue problem,
method, SIAM J. Num. Anal. 8, 737-748. V2. ·VAINIKKO, G. M. (1967),
SS. STRANG, G. (1971), The finite element method and approximation theory, methods in the eigenvalue
SYNSPADE, 547-584. 7, 18-32.
V3. VAN DER SLUIS, A. (1970),
algebraic systems, Numer.

), Eigenvalues of Ax J.Bx with
. An automaÜc generation scheme
metric coordinates, Ittt~ J.for Num.
asís of finite element methods for
'lEng. 1, 3-28. }.,
itfness methods by ditferent varia-
'liversity SIAM-AMS Symposium,
K (1971 ), Elastic crack analysis by
f}¡atterson III.
Higher order convergence results
[) eigenvalue problems 1, SIAM J.
on de la méthode des ditférences
·érieures ou inférieures, Comptes
~pproximations in elasticity based
. Appl. Math. 5, 241-269.
les for elastic plates with relaxed
truct. 4, 837-844.
ror bounds for semidiscrete Galer-
ms, Duke Univérsity SIAM-AMS .
rgon-circle paradox of simply sup-
tre, AIAA J. 6, 155-156.
.nd convergence of a finite-element
.faths. Appl. 9, 1-13. '
dual analysis principie, IUT1M
to the problem of approxiQlation
ns, Quart. Appl. Mat'!: 4, 45-~9,
z methods for multidiinensional
38.
.
ls for the Rayleigh-Ritz-Galerkin
1,
1ethod and approximation theory,
BIBLIOGRAPHY 293
S6. STRANO, G. (1972), Approximation in the finite element method, Numer.
Math. 19, 81-98 .
S7. STRANO, G. (1972), Variational crimes in the finite element method, Mary-
land Symposium.
SS. STRANO, G. (1973), Piecewise polynomials and the finite element method,.
Bull. Amer. Math. Soc. 19, 1128-1137.
S9. STRANO, G., and A. E. BEROER (1971), The change in solution due to
change in domain, Proc. AMS Symposium on Partial Differential Equations,
Berkeley.
S lO. STRANO, G., and G. F1x (1971), A Fourier· analysis of the finite element
method, Proc. GIME Summer School, Ita/y, to appear.
S 11. SwARTZ, B., and B. WENDROFF (1969), Generalized finite ditference schemes, ·
Math. of Comp. 23, 37-50.
Tl. TAYLOR, R. L. (1972), On completeness ofshape functions for finite element
analysis, Int. J. for Num. Meth. in Eng. 4, 17-22.
T2. THoMÉE, V. (1964), Elliptic ditference operators and Dirichlet's problem,
Diff. Eqns. 3, 301-324.
T3. THOMÉE, V. (1971), Polygonal domain approximation in Dirichlet's problem,
MRC Tech. Rept. 1188, Univ. Wisconsin.
T4. ToNo, P. (1969), Exact solution of certain problems by finite-element
method, AIAA J. 1, 178-180.
T5. ToNo, P. (1971), On the numerical problems of the finite element methods,
j Waterloo Conference.
T6~- ToNo, P., and T. H. H. PIAN, The convergence of finite element method
in solving linear elastic problems, Int. J. Solids Struct. 3, 865-879.
j T7. ·ToNo, P., and T. H. H. PIAN (1970), Bounds to the infl.uence coefficients
by the assumed stress method, Int. J. Solids Struct. 6, 1429-1432.
(
T8. ToNo, P., T. H. H. PIAN, and L. L. BucciARELLI (1971), Mode shapes and
frequencies by the finite element method using consistent and lumped
masses, J. Comp. Struct. 1, 623-638.
1 T9. TREFFTZ, E. (1926), Ein Gegenstück zum Ritzschen Verfahren, Second
Congress Applied Mechanics, Zurich.
TlO. TuRNER, M. J., R. W. CLOUOH, H. C. MARTIN, and L. J. ToPP (1956),
1
,! Stitfness and defl.ection analysis of complex structures, J. Aero. Sciences 23.
i
V l. VAINIKKO, G. M. (1964), Asymptotic error estimates for projective methods
~
1
V2.
in the eigenvalue problem, Zh. Vychisl. Mat. 4, 404-425.
VAINIKKO, G. M. (1967), On the speed of convergence of approximate
methods in the eigenvalue .problem, USSR Comp. Math. and Math. Phys.
! .
7, 18-32.
V3. VAN DER SLUIS, A. (1970), Condition, equilibration, and pivoting in linear
algebraic systems, Numer. Math. 15, 74-86.
294 · .~~BLIOGRAPHY

V4. VARGA, R. S. (1965), Hermite interpolation and Ritz-type methods for Z10. ZLAMAL, M. (1972), The
two-point boundary value problems, in Numerical Solutions of Partía[ boundaries, Int. J. for Nur.
'Dijferential Equations, ed. J. H. Bramble, Academic Press, New York.
V5. VARGA, R. S. (1970), Functional analysis and approximation theory in SUPPLEMENTARY BIBLIOGRAI
numerical analysis, SIAM Regional Conference Series, Vol. 3.
V6. VIss~:~'W. (1969), A refined mixed-type plate bending element, AIAA J. SI. FINLAYSON, T. (1972), The
7, 1801-1803. Principies, Academic Press, 1
Wl. WAIT, R., andA. R. MITCHELL (1971), Corner singularities in elliptic prob- S2. Conference on Numerical Ar
lems by finite element methods, J. Comp. Physics 8, 45-52. S3. Applications of the finite ele
W2. WEINBERGER, H. F. (1961), Variational Methods in Boundary Value Prob- Army Engineers Symposium
lems, University of Minnesota. S4. BATHE, K.-J. (1971), Solutior
W3. WIDLUND, O. B. (1971), Sorne recent applications of asymptotic error lems in structural engineerin~
expansions to finite-difference schemes, Proc. Roy. Soc. Lond. A323, S5. BATHE, K.-J., and E. L. WI
167-177. direct integration methods, té
, W4. WILLIAMS, M. L. (1952), Stress singularities resulting from various boundary S6. BIRMAN, M. S., and M. Z. Sor
conditiohs in angular corners of plates in extension, J. Appl. M ech. 526-527. tion of functions of the class·
W5. VviLSON, E. L., R. L. TAYLOR, W. P. DoHERTY, and J. GHABoussi (1971), S7. CLOUGH, R. W., and K.-J.B
Incompatible displacement models, University of Ilinois Symposium. response, Second U.S.-Japar
Yl. YAMAMOTO, Y., and N. ToKUDA (1971), A note on convergence of finite S8. FRIED, I., and S. K. YANG
element solutions, Int. _J. for Num. Meth. in Eng. 3, 485-493. a singularity, AIAA J. 10, 12
Zl. ZENISEK, A. (1970), Polynomial approximations on tetrahedrons in the S9. FRIED, l., and S. K. YANG(l
finite element method, manuscript. · bending element of quadratic
Z2. ZENISEK, A. (1970), Higher degree tetrahedral finite elements, manuscript. S10. Fum, H. (19?2), Finite elem
U.S.-Japan Seminar. ,
Z3. ZENISEK, A. (1970), Interpolation polynomials on the triangle, Numer.
Math. 15, 283-296. Sil. GIRAULT, V. (1972), A finite
J. Numer. Anal., to appear.
Z4. ZIENKIEWICZ? O. C. (1971), Isoparametric and allied numerically integrated
elements-a review, University of Illinois Symposium. S12. ÜOLUB, G. H., R. UNDERW(
algorithm for the symmetric
Z5. ZIENKIEWICZ, O. C., B. M. lRONS, et al. (1969), Isoparametric and associated ·
element families for two and three-dimensional analysis, Holand and Sl3. JEROME, J. W. (1973), Topic~
Bell [8]. sium on Approximation at A

Z6. ZIENKIEWICZ, Q. C., R. L. TAYLOR, and J. M. Too (1971), Reduced integra- S14. JoHNSON, C. (1972), On the e·
tion technique in general analysis of plates and shells, In t. J.for Num. Meth. plate bending problems, N un;
in Efíg. 3, 275-290. S15. JoHNSON, C. (1972), Converg
Z7. ZLAMAL, M. (1968), On the finite element method, Numer. Math. 12, plate bending problems, unp1
394-409. S16. KoRNEEV, V.G. (1970), The e
Z8. ZLAMAL, M. (1970), A finite elerpent procedure of the second order of a high order of accuracy, Ve
accuracy, Numer. Math. 14, 394-402. S17. McLEOD, R., and A. R. MI
Z9. ZLAMAL, M. (1972), Curved elements in the finite element method I, SIAM tions for curved elements in t
J. Num. Anal., to appear. 1o, 382-393.

1
1

~-

·'-,
ation and Ritz-type methods for
n Numerical Solutions of Partial
• Academic Press, New York.
sis and approximation theory in
rerence Series, Vol. 3.
~ plate bending eleme,t:\,t, AIAA J.
:orner singularities in elliptic prob-
'· Physics 8, 45-52.
f\,fethods in Boundary Value Prob-
applications of asymptotic error
;, Proc. Roy. Soc. Lond. A323,
es resulting from various boundary
extension, J. Appl. Meoh. 526-527.
JHERTY, and J. GHABOUSSI (1971),
ersity of llinois Symposium. 1
, A note on convergence of finite
. in Eng. 3, 485-493.
<imations on tetrahedrons in the
1edral finite elements, manuscript.
·nomials on the triangle, Numer.
i.
e and allied numerically integrated
s Symposium.
.969), Isoparametric and associated ·
imensional analysis, Holand and
J. M. Too (1971), Reduced integra-
~s and shells, Int. J.for Num. Meth.
ment method, Numer. Math. 12,
Jrocedure of the second order of
the finite element method I, SIAM
BIBUOGRAPHY 295
Z10. ZLAMAL, M. (1972), The finite element method in domains with curved
boundaries, Int. J. for Num. Meth. in Eng., to appear.
SUPPLEMENTARY BIBLIOGRAPHY
Sl. FINLAYSON, T. (1972), The Method of Weighted Residuals and Variational
Principies, Academic Press, New York.
S2. Conference on Numerical Analysis, Royal Irish Academy, Dublin, 1972.
S3. Applications of the finite element method in geotechnical engineering, U. S.
Army Engineers Symposium at Vicksburg, Mississippi, 1972.
S4. BATHE, K.-J. (1971), Solution methods for large generalized eigenvalue prob-
lems in structural engineering, Thesis, Univ. of Calif. at Berkeley.
S5. BATHE, K.-J., and E. L. WILSON (1972), Stability and accuracy analysis of
direct integration methods, to appear.
S6. BIRMAN, M. S., and M. Z. SoLOMZAK (1967), Piecewise-polynomial approxima-
tion of functions of the classes W~, Math. USSR-Sbornik 2, 295-317.
S7. CLOUGH, R. W., and K.-J. BATHE (1972), Finite element analysis of dynamic
response, Second U.S.-Japan Seminar.
SS. FRIED, I., and S. K. YANG (1972), Best finite elements distribution around
a singularity, AIAA J. 10, 1244-1246.
S9. FRIED, 1., and S. K. YANG (1972), Triangular, 9 degrees of freedom, eo plate
bending element of quadratic accuracy, Q. Appl. Math., to appear.
SlO. Fum, H. (1972), Finite element schemes: stability and convergence, Second
U.S.-Japan Seminar.
S11. GIRAULT, V. (1972). A finite difference method on irregular networks, SIAM
J. Numer. Anal., to appear.
S12. GoLUB, G. H., R,. UNDERWOOD, and J. H. WILKINSON (1972), The Lanczos
algorithm for the symmetric Ax = lBx problem, to appear .
S13. JEROME, J. W. (1973), Topics in multivariate approximation theory, Sympo-
sium on Approximation at Austin, Texas.
S14. JOHNSON, C. (1972), On the convergence of a mixed finite-element method for
plate bending problems, Numer. Math., to appear.
S15. JoHNSON, C. (1972), Convergence of another mixed finite-element method for
plate bending problems, unpublished.
S16. KORNEEV, V.G. (1970), The construction of variational difference schemes of
a high order of accuracy, Vestnik Leningrad Univ. 25, 28-40.
S17. McLEOD, R., and A. R. MITCHELL (1972), The construction of basis func-
tions for curved elements in the finite element method, J. Inst. Maths. App/ics.
10, 382-393.
296 BIBLIOGRAPHY

S18. MoTE, C. D. (1971), Global-local finite element, Int. J. for Numer. Meth. in
'Eng. 3, 565-574.
S19. PIAN, T. H. H., and P. ToNo (1972), Finite element methods in continuum
mechanics, Adv. in Appl. Mech. 12, 1-58.
S20. VANDERGRAFT, J. S. (1971), Generalized Rayleigh methods with applications
to ~ding eigenvalues of large matrices, Lin. Alg. and Applics. 4, 353-368.
S21. WEA\}ER, W., JR. (1971), The eigenvalue problem for banded matrices, Com-
puters·and Structures 1, 651-664.

INDEX OF NC

This will be a little more than an

way the ideas which are essential ir
book:

l. Norms, function spaces, ar

11. Energy inner products, elli
111. Finite element spaces and ;

The definitions in 1 and 11 are 1

the subject of finite e~ements,.

l. NORMS. FUNCTION SPACE

BOUNDARY CONDITIONS

A norm is a measure of the siz€

two functions (11 u - v 11). It satisfi€
inequality llu +vil< llull + llvll.
zero only if u is the zero function. 11
are bounded in terms of their inhc
are unique: u= O if /=O, or (by
the samef.
Sorne familiar norms are

i) the maximum norm = "sup

ii) the L 2 norm = 3C 0 norm =

In the discrete case, for vector

integrals are .replaced by correspo
the exponents 2 and 1/2 are replact:
: element, lnt. J. for Numer. Meth. in

'inite element methods in continuum

" .
. Rayleigh methods with applications
Lin. Alg. and Applics. 1, 353-368.
problem for banded Matrices, Com-

INDEX OF NOTATIONS

This will be a little more than an index. We shall try to summarize in a convenient
way the ideas which are essential in understanding three of the main themes of this
book:

l. Norms, function spaces, and boundary conditions

II. Energy inner products, ellipticity, and the Ritz projection
III. Finite element spaces and the patch test.

The definitions in I and II are more or less standard; those in III are special to
the subject of finite elements..

l. NORMS. FUNCTION SPACES, ANO

BOUNDARY CONDITIONS

A norm is a measure of the size of a function (11 u 11) or of t\le distance between
two functions (11 u - v 11). It satisfies the condition 11 cu 11 1e 111 u 11 and the triangle
inequality 11 u + v 11 < 11 u 11 + 11 v 11. Furthermore, unlike a seminorm, the norm is
zero only if u is the zero function. It follows that if all solutions u to a linear problem
are bounded in terms of their inhomogeneous terms <11 u 11 < e 11/11) then solutions
are unique: u O if f O, or (by $Uperposition) u 1 - u2 O if u 1 and u2 share
the sa111e f.
Sorne familiar norms are

i) the maximum norm "sup norm" Leo norm = sup 1u(x) j;

· xinn
ii) the L 2 norm =:reo norm (J 1u(x) ¡2 dx 1 ••• dx,.)ttz.
n

In the discrete case, for vectors u = (uh u 2 , ••• ) instead of functions u(x), the
integra1s are .replaced by corresponding sums. One generalizaÜon is to Lp norms;
the exponents 2 and 1/2 are replaced by p and 1/p. The triangle inequality is satisfied

297
298 INDEX OF NOTATIONS
xs contains eq if and only if s - q > 1:
(and we have a norm) for p > l. These spaces become interesting and valuable in of trace theorems: suppose u is in 3Cs
nonlinear problems; we have not found them essential to the linear theory. sidered as a function on r? Rough
' A se'cond· generalization, ·absolutely basic, is to include derivatives as well as xm-( 1/2) is a suitable "data space"
function values of u in computing a norm (pages 5, 143). The 3Cs norm combines u = g; it matches the solution space
the 3C 0 (or L 2 ).norms of all partial derivatives The central problem of partial dif
space of data to a space ·or solution
example -Au =fin the L"" norm, i1
and the pointwise theory suffers from
problems of any order 2m: 11 u lis < e
we take/in 3C 0 and look for u in 3Cj
the variational problem s = m: u is J
The seminorm 1u ls includes only those terms of order exactly 1rx 1= s; it is zero if only by the essential boundary condi1
u is a polynomial of degree s - l. The squares are introduced in equation (1) in
order to have an inner product structure, in other words, to make 3Cs a Hilbert
space (see II below). Fractional derivatives (non-integer s) also have important 11. ENERGY INNER PRODUCTS.
applications (pages 144-5, 260), but their definition is rather technical [1, 10, 14]- ANO THE RITZ PROJECTION
.except Wh<:fnn is the whole of n-space, when we use Fourier transforms,
Linear variational problems are 1
The classical case is to minimize a pl
we separate the terms of second and
tions v. The second-degree term is ti
Negative norms (the index s is negative, not the norm itself!) are defined by duality
energy inner product
(pages 16, 73, 167):

max lf uvl. (2) a(v, w) = ! (a(v + wl

llull-s v in :re• llvlls In terms of this inner product, the ce
·The functions in these "Sobolev spaces';--as distinct from their norms-are of the first variation, alias the equatü
usually defined by starting with a set of comparatively simple functions and then a(u, v) (f, v)
(3)
completing the space (page 11). The result is a Banach space; it contains the limit
point of any sequence for which ¡-uN -uM 1--4 O as N, M --4 co. Intuitively, the Integration by parts alters this wea1\
boles are filled in .. the Euler differential equation for u-
The completed space will depend on the original set of simple functions. If we with jump conditions appearing as in
start with the set e o of all continuous functions, then in the maximum norm, this NOTE: An inner product is bili
set contains all its limit functions and there are no boles to fill in. lf the original
set includes also1 the piecewise continuous functions, the final space Loo is much a(u + v, w + z) = a(u,
larger (and more difficult to describe). This point reappears in connection with
boundary conditions: if none are applied to the original set of all infinitely differen- and it is only when the strain energy
tiable functioris, then its completion is the whole space Xs(O); this isthe admissible · If this energy depended on the poin1
space forthe Neumann probleÍn, with free edges. If each member of the original set it would fail. Among Lp norms only
is required to vanish in a strip near the boundary r -the strip may be smaller for such an inner product; it is therefo
one function than another-then the completed space in the same norm is 3Ct;, with property extends to the spaces 3Cs, '
derivatives of order less than S vanishing on r (page 67). This is the adtnissible linear elasticity and other applicatio
space for the Dirich/et problem. ~ ' Problems which are not self-adjoiJ
We mention two groups of important but technical theorems about these spaces. work, not with a minimization. The
One group is typified by the Sobolev inequality (pages 73, 142), and answers the a(u, v) =F a(v, u), and complex~value
question: ifthe derivatives of order s 1 (integer or not) are in Lp 1 , are those of qrder if the real part of a(v, v) is elliptic
s 2 in Lp 2 ? In other words, which function spaces contain which others? (Sobolev:

beco me interestjng :and valuable in
es sen tia{ to the linear theory.
is to include derivatives as well as
:es 5, 143). The :Jes norm combines
1iX 1 = iX 1 + · · · + iX~ }< S:
dx1· · · dxn.
f order exactly !tX l = s; it is zero if
s are introduced in equation (1) in
•ther words, to make :Jes a Hilbert
.on-integer s) also have important
.tion is rather technical [1, 10, 14]-
we use Four~er transforms,
norm itself!) are defined by duality
uvj
vrr;·
-as distinct from their norms-are
uatively simple functions and then
Banach space; it contains the limit
~O as N, M---¡. oo. Intuitively, the
iginal set of simple functions. If we
.s, then in the maximum norm, this
·e no boles to fill in. If the original
::tions, the final space L"" is much
•oint reappears in connection' with
original set of all infinitely differen-
~ space X"(Q}; this is.the admissible
:. If each member of tlíeoriginal set
lry r -the strip may be smaller for
¡ space in the same norm is x~. with
r (page 67). This is the adtnissible
:hnical theorems about these spaces.
y (pages 73, 142), and answers the
lr not} are in Lp 1 , are those of order
:es contain which others? (Sobolev:
INDEX OF NOTATIONS 299 ·
, :Jes contains eq if and only if s q > n/2.) The other technical results are the group
of 'trace theorems: suppose u is in :Jes, how smooth are its ·boundary values, con-
sidered as a function on r? Rough answer: they are in :Jes-012>(r). Therefore
:Jem-(112> is a suitable "data space" for the inhomogeneous boundary condition
u g; it matches the solution space Xe for u.
·The central problem of partial differential equations (pages 4-6) is to match a
space of data to a space 'of solutions. Such a match is not automatic. For the
example -Au /in the Loo norm, it is not true that lluxx!! + lluyyll ~ Cll/11-
and the pointwise theory suffers from it. The :Jea norms are well matched for elliptic
problems of any order 2m: lluils ~ Cllflls- 2 m. For the Euler equation s 2m:
we take/in X 0 and Iook for u in :Jejm, satisfying all m boundary conditions. For
the variational problem s = m: u is found in the admissible space Xe, restricted
only by the essential boundary conditions.
11. ENERGY INNER PRODUCTS, ElLIPTICITY,
ANO THE RITZ PROJECTION
Linear variational problems are posed in terms of quadratic functionals l(v).
The classical case is to minimize a potential energy l(v) a(v, v) 2{!, v)- here
we separate the terms of second and first degree--over a space of admissible solu-
tions v. The second-degree term is the strain energy, and it is associated with an
energy inner product
(2) a(v, w) = }(a(v + w, v + w)- a(v- w, v w))A
In terms of this inner product, the condition that u minimize I{v) is the vanishing
· of the first variation, alias the equation of virtual work:
{3) a(u, v) (f, v) for all admissible v.
Integration by parts alters this weak form (or Gaferkin form) of the problem into
the Euler differential equation for u-of order 2m, without the presence of v, and
· with jump conditions appearing as integrated terms at any discontinuities.
NOTE: An inner product is bilinear,
a(u + v, w + z) a( u, w) + a(v, w) +a( u, z) +a( v, z),
and it is only when the strain energy has a favorable form that th,is property holds.
If this energy depended on the point of maximum strain, a(v, v) = max 1grad v 12 ,
it would fail. Among LP norms only the case p 2 yields, through equation (2),
such an inner product; it is therefore the only Hilbert space. The inner product
property extends to the spaces :;es, and also (thank God) to the strain energies in
linear elasticity and other applications.
Problems which are not self-adjoint begin directly with the equation {3) ofvirtual
work, not with a minimization. The bilinear form a(u, v) is no longer symmetric,
a(u, v) :f= a(v, u), and complex-valued functions must be admitted. Nevertheless,
if the real part of a(v, v) is elliptic (see below), then the results of the Galerkin
 300 INDEX OF NOTATIONS

,theory (page 119) completely paralleJ thc;>se ofthe Ritz theory. They coincide, when
a(u, 1J)-'is symmetric. . u. The interpolate assumes the san
The solvability of the fundamental variational equation (3) is guaranteed if the it is one ofthe trial polynomials: 1
form is elliptic·: Re a(v, v) > al! v 1!~. (So is the solvability of the -corresponding approximation theorems of Chapt4
parabolic equation.) In the case of systems of equations, with a vector of unknowns described earlier. The error u
u (u¡(x), ... , ur(x))-as is typical in applications-there appear several varieties degree k - 1 to which the shape i
of elliptic!ty. One possibility [1, 10, 14] asks that the eigenvalues of certain ~atrices The dimension of the trial spa<
of órder r\have positive real parts; this is too weak to guarantee success when the free parameters q1• Obviously N
Galerkin Ínethod is applied on a subspace. Strong ellipticity is a condition, not on quantity is the number M of pa1
the eigenvalues, but on the matrices themselves-and a still stronger condition is assembled matrix K to be comp~
the familiar one on Re a(v, v), which applies as successfully to systems asto single number of degrees of freedom (si
equations; v becomes an admissible vector of functions. Also for boundary condi- M is smaller, depending on the C4
tions there is a catalogue of possibilities, clearly described by Kellogg in the (Our conjecture for the trial spacc:
Baltimore Symposium volume [6]; for applications the central step is still to dis- k 1 and of continuity class eq
tinguish the essential conditions, and thereby the admissible space-and to require We come now to the patch tes¡
the definiteness of Re a(v, v) over that space. . recently it was hardly known (at·
The I,titz method is to minímize the functional/(v) over a sequence of subspaces created in the appendix to [B9], i
Sh. The fundamental theorem (page J9) establishes that the minimizing uh is the was convergent in one configurad'
projection of u opto Sh, in other words, uh is the closest function to u in the strain faced, under its official name in a
energy norm a(v, v). Therefore if each subspace Sh is contained in the next-as was story was told more recently by Ir
supposed in the classical Ritz method, and usually occurs for finite elements when still assailed by doubts that the te
new elements are formed by subdividing the old ones-the convergence in strain We are convinced that, under.
energy is monotonic as h --+ O. So is the convergence of eigenvalues. This may be on page 174, and is very simple te
useful, but it is not critical to the Ritz theory; monotonicity of the Sh is an extra 177; if the strain energy involves e
hypothesis, and monotonicity of convergence is an extra conclusion. JJ Dmrp1 should be calculated corre
are ignored in the nonconforminJ
higher-order patch test asks that
degree n - m; this generalizatio1
111~ FINITE ELEMENT SPACES ANO
convergence in strain energy of ori
THE PATCH TEST
A word about the proof. On
The usual description of a finite element specifies the form of the shape function estimating the inner product erro
(trial polynomial), and the location as well as· the parameter (function value v, the linear terms involvingfalso a
or sorne derivative D 1v) assigned to each node. Section> 1.9 contains a number of - Au for which the energy ir
examples. This is enough information to compute the element matrices, and to in the patch test means that for a
assemble them into the global stiffness and mass matrices K and M.
(4)
In Section 2.1, for mathematical reasons, we took one more step in describing
the trial space Sh: we defined a set of basis functions rp 17 ••• , 'PN for the space. Choosing P close to u over the se
The function rp 1 \Vas directly associated with the node z¡, and with the particular the basis by a(rp¡, rp1) = 1, (4) is lt
nodal pa_rameter D 1v. Jf the shape functions are uniquely defined by the values of vh as 2; q1rp¡,
the nódal parameters (whrch they must be!), then there is a unique trial function rp.
for which D1rpj(z1) 1, and all its other nodal parameters Dtrplzt) are zero, Thes~ (5)
· functions form a basis, because any trial function can be expanded in terms of
its nodal parameters as

vh = 2; q1rpb whe~;e the weight q1 D1vh(z1).

If it were true that
This leads immediately to the definition of the interpolare u1 of any given function
(6)
 INDEX OF NOTATIONS 301

1e Ritz theory. They coincide, when

u. The interpolate assumes the same nodal parameters as u but inside each element
i t is on~ of ~he trial polynomials: u1 = ~ q,rp¡, where the ~eight qJ = n,u(zj). The
tal equation (3) is guaranteed if the
approxtmatwn theorems of Chapter 3 establish that u1 is close to u in the norms
described earlier. The error u- u1 depends on the element sizes ~. and on the
e solvability of the .corresponding
uations, with a vector of unknowns
degree k - 1 to which the shape functions are complete (page 136).
ions-there appear several varieties
The dimension of the trial space Sh is N, the number of basis functions rp and
: the eigenvalues of cer,tain matrices
free parameters q1• Obviously N depends on the number of elements. A crltical
~<'eak to guarantee sucd~ss when the
quantity is the number M of parameters per vertex; this permits the size of the
mg ellipticity is a condltion, not on
assembled matrix K to be compared for two competing elements. Let d be the
s-and a still stronger condition is
number of degrees of freedom (shape ·runction coefficients) within each element·
successfully to systems as to single
M is smaller, depending on the continuity constraints imposed between elements:
unctions. Also for boundary condi-
(Our conjecture for the trial space described on page 84, complete through degree
!arly described by Kellogg in the
k - 1 and of continuity class eq on triangles, is M = (k - 1 - q)(k 1 2q).)
ions the central step is still to dis-
We come now to the pateh test. This test has had very little publicity and until
te adm:issible space-and to require
recently it was hardly known (at least by that name) even to the expe;ts. It was
created in the appendix to [B9], in order to explain why the Zienkiewicz triangle
al/(v) over a sequence of subspaces
was convergent in one configuration but not in another (page 175). The test resur-
ishes that the minimizing uh is the
faced, under its official name, in a brief comment [19] on an earlier paper. The full
1e closest function to u in the strain
story was told more recently by Irons at the Baltimore Symposium but there he is
Sh is contained in the next-as was
still assailed by doubts that the test is sufficient for convergence. '
ally occurs for fi.nite elements when
We are convinced that, under reasonable hypotheses, it is. The te~t is described
•ld ones-the convergence in strain
on page 174, and is very simple to conduct. We recall the equivalent form on page
rgence of eigenvalues. This may be
177: if the strain energy involves derivatives Dmv of order m, then all the integrals
monotonicity of the Sh is an extra
is an extra conclusion.
f f Dmrp1 should be calculated correctly-even though interelement boundclry terms
are ignored in the nonconforming case, or numerical quadrature is applied. The
higher-order patch test asks that f f Pn-mDmrp1 be correct for a11· polynomials of
degree n-m; this generalization was made by the first author, and produces
convt?rgence in strain energy of order h2<n-m..;.o.
A word about the proof. On pages 179 and 186, the problem is reduced to
estimating the inner product error a*(u, vh) a(u, vh). (Por numerical integration
cifi.es the form of the shape function
the linear terms involving falso appear.) Suppose we stay with the model problem
s the parameter (function value v,
-:-Au f. for which the energy inner product is a(u, v) = f f Uxvx + u;yvy. Success
:. Section 1.9 contains a nuJilber of
m the patch test means that for any linear polynomial P,
1pute the element matrices, and to
ss matrices.K and M. (4)
te took one more step in descr,ibing
unctions rph .. . , rpN for the space. Choosing P close to u over the set E1 on which rp1 is non-zero, and renormalizing
he node Z¡, and with the particular t~ebasis by a(rp¡, rp) = 1, (4) is less than eh 11 u 1! 2 , E,· Therefore if we expand any
~e uniquely defi.ned by the values of
v as ~ q1tp¡,
en there is a unique triái function rp1 -
parameters D¡rpiz¡) are zero. These (5) [a*(u, vh) - a(u, vh)[ <eh~ [[ uii2,E, [q1 1
;;tion can be expanded in terms of
<eh(~ llulli.EY 12 (~qJ) 1 ' 2
~ c'h 11 u[b(~ qJ)ttz.

If it were true that

interpolate u1 of any given function
(6)
302 INDEX OF NOTATIONS

th:eh the convergence proof would be complete. The expression A is' less than
c"'ChiiÚ!Iz, and according to the estimateat the top of page 179, the strains are in
'error by O(l!).
The result is correct, but regrettably the inequality (6) is false. It amounts to
asking that the condition number of the stiffness matrix K be bounded-or that the
rp1 be uniformly independent in the energy norm. This is true of the mass matrix
(page 212?.,, but not of K. There is, however, a way out. Any conforming rp1 can be
ignored in.:'our calculations; the difference in (4) is identically zero. Therefore if the
trial space 'can be regarded as a conforming space to which sorne uniformly inde-
pendent nonconforming elements are added, the proof succeeds. This was obvious INDEX
in Wilson's case, since he began with the standard bilinear elements, and super-
imposedtwo nonconforming quadratics within each square. (Such interna! degrees
of freedom are called nodeless variables.) Because the squares never overlap, uniform
independence was automatic.
Crouzeix and Raviart have recently given a beautiful treatment of nonconform-
ing elements for diverge.nce-free fluids. Their technique applies to all elements whi.ch
pass the test in the following way: over each interelement edge or face, the integral
of the nonconforming jump is zero. Their technique for deducing convergence has
been formalized, and applied to plate elements, by both Ciarlet and Lascaux. A
The increasing application of finite elements - to the Navier-Stokes equations,
control problems, earthquake prediction, nonlinear elasticity and plasticity in soils Abstract method, 104, 270
as well as metals, and the design of tankers and reactors - promises a happy future Admissible space, 12,66
for both the analyst and the engineer. Anderheggen, 82, 132
Arch, 127
Area coordinates, 94, 96
Argyris, 2
Assembly, 29
Aubin, 155, 166

Babuska,124,133,151,193
Bandwidth, 105, 118, 134
Barlow, 168
Bathe, 238, 240
Bauer,237
Beam, 63, 121
Bell,96
Berger,198
Birkhoff, 261
Block power method, 237, 240
Boundary layer, 108, 196
Bramble,133, 134, 146; 147,169, 23
1 '
e
Cholesky, 36, 236
Ciarlet, 111, 161,169, 192
:te. The expression A is Iess than
top of page !79, trie strains are in

equality (6) is false. It amounts to

s matrix K be bounded-or that the
m. This is true of the mass matrix
.vay out. Any conform(l}g (/Ji can be
'is identically zero. THe'refore if the
Jace to which sorne uni/ormly inde-
proof succeeds. This was obvious IN DEX
.dard bilinear elements, and super-
each square. (Such interna! degrees
~ the squares never overlap, uniform

beautiful treatment of nonconform-

hnique applies to allelements which
terelement edge or face, the integral
1ique for deducing convergence has
, by both Ciarlet and Lascaux.
; - to the Navier-Stokes equations,
1ear elasticity and plasticity in soils
reactors- promises a happy future
A Clough, 2, 84, 238, 240
Collocation, 117, 118
Abstract method, 104, 270 Complementary energy, 129
Admissible space, 12, 66 Completion, 13
Anderheggen, 82, 132 Condition number, 35, 121, 206, 208, 244
Arch, 127 Conforming, 74, 137
Area coordinates, 94, 96 Conservative, 251 , 25 3
Argyris, 2 Consistency, 18, 23, 175, 241
Assembly, 29 Courant,74, 76,132,221,256
Aubin, 155, 166 Cowper, 127, 183
Crack, 115,144,260,269
B

Babuska,124,133,151,193 D
Bandwidth, 105, 118, 134
Barlow, 168 Degree, 136
Bathe, 238, 240 de Veubeke, 131
Bauer, 237 Dirichlet, 4, 64, 67
Beam, 63, 121 Douglas, 107, 168, 187,245,249
Be11,96 Dupont,107, 168, 1~7, 245,249,255
Berger, 198
Birkhoff, 261
E
Block power method, 237, 240
Boundary layer, 108, 196
Elastic-plastic, 111, 146
Bramble, 133, 134, 146; 147, 169, 231
Elements:
bicubic, 88, 102, 165, 270
e bilinear, 86, 157, 211
brick, 87
Cholesky, 36, 236 cubic, 55, 79, 226, 255
Ciarlet, 111, 161, 169, 192 linear, 27, 224

303
304 INDEX

linearon triangles, 75, 102, 152 lnterpolate, 4 3, 97

N
~quadratic, 5.3, 78, 103, 194 Interpolating basis, 102, 136
quintic, 82 Inverse hypothesis, 167
serendipity, 87, 160 lnverse iteration, 237 Natural condition, 117
spline,60,104,270 hons~99,174,183,186,209 Negative norms, 16, 73, 167
trilinear, 87 lsopara~etric, 109, 157, 199 Neumann, 4, 68
Element stiffness matrix, 28, 58, 81, 213 Nickell, 244
Elliptic, 6~. Nitsche, 169; 204, 267
Energy inner product, 39 Nitsché trick, 49, 107,166,268
Error: · J Nodal method, 101, 270
in displacement, 107, 166 Nonconforming, 178
in strains, 106, 165 Non-linear, 11 O, 111, 25 2
Johnson, 125
linear eleinents, 45 Norms, 5, 65
Jump condition, 14, 116, 261, 275
one-dimensional, 62 Num.erical integration, 32, 98, 109, 181
Essential condition, 4, 8, 12,117, 193

K
F o
Force method, 129 Kellogg,261,263
Oden, 111
Forced vibration, 167 Kondrat'ev, 266, 267
One-sided approximations, 146
Fried, 126, 163,207,213 Korn, 73
Osbom, 231
Fujii, 253'
Fundamental theorem, 39, 106

L p
G
Lagrange multiplier, 132, 133 Parlett, 240
Galerkin, 9, 115, 117, 119, 123, 125,219, Least squares, 133 Patch test, 164, 172, 174, 182
231 Lesaint, 25 5 Penalty, 132
Gauss elimination, 34, 238, 265 Lindberg, 127 Peters, 116, 238
Gauss quadrature, 99, 183, 188, 190 Load vector, 30 Plate, 71
George, 38,15 L-shaped membrane, 27 8 Poincaré, 42, 221
Gerschgorin, 21, 210, 211 Lumped matrix, 118,223, 226, 228,.244, Point load, 14, 73
Graded mesh, 115, 155 256

H a
M
Hermite, 56, 88 Quasi-interpolate, 14 2
Hilbert, 14_6, 147 Marcal, 111
Hybríd, 131, 132 Martín, 2
Hyperbolic, 251 Mass matrix, 29, 51, 96
McCarthy, 125 R
Minmax principie, 219, 221, 229
Mitchell, 108, 159, 197
Raviart, 161, 169, 192, 256
Mixed method, 120, 125
Rayleigh quotient, 216
Mode superposition, 244, 245
Reactor, 275
lnhomogeneous condition, 51, 70, 193, 199 Monotone, 11 O
Ritz method, 24
Interface, 260 , Mote, 131, 135
Roundoff, 110, 121, 215
 INDEX 305

rpolate, 43, 97 N
rpolating basis, 102, 136
S
rse hypothesis, 167
rse iteration, 237 ~ Natural condition, 117 Saint~Venant, 197, 269
s;99,174,183, 186,209 Negative norms, 16, 73, 167 Schatz,133,134,267
arametric, 109, 157, 199 Neumann, 4, 68 Schultz, 111, 166,167, 206
Nickell, 244 Scott, 179, 203
Nitsche, 169, 204, 267 Shell,127,128
Nitsché trick, 49, 107, 166, 268 Singular functions, 263, 270
J Nodal method, 101, 270 Singularity, 13
Nonconforming, 17 8 Sobolev, 73, 144, 145, 183 /
Non~linear, 110, 111,252 Spline,60,89,104,141,155
lSOn, 125 Norms, 5, 65 Stability, 23, 20~, 244, 247,256
.p condition, 14, 116, 261, 275 Num.erical integration, 32, 98, 109, 181 Static condensation, 81
Stationary point, 116
Stiff, 98, 175, 244
Stiffness matrix, 29
K o Strain energy, 39
Stress intensity, 269
.ogg, 261, 263 Stress points, 151, 168
Oden, 111
tdrat'ev, 266, 267 Subspace' iteration, 237, 240
One~sided approximations, 146
n,73 Superconvergence, 168
Osborn, 231
Superfunction, 141, 15 5, 170
¡
l
Swartz, 249

L p
1 T
range multiplier, 132, 133 t¡
Parlett, 240
st squares, 13 3 ~ Taylor, 189
. ~ Patch test, 164, 172, 174, 182
aint, 255 Penalty, 132 Temam, 178
dberg, 127 ~ Peter~~ 116, 238 Thomée, 107
td vector, 30
1aped membrane, 27 8
nped matrix, 118, 223, 226, 228, 244,
i
'
~
Plate, 71
Poincaré, 42, 221
Point load, 14, 73
Toéplitz, 209,212, 224
Tong,253
Torsion, 269
256 Triangular coordinates, 94, 96
Truncation error, 18, 170
Turner, 76
Q
M
Quasi~interpolate, 142 .U
real, 111
rtin, 2
Uniform basis, 106, 137
ss matrix, 29, 57, 9.6
Carthy, 125 R
1max principie, 219, 221, 229
tchell, 108, 159, 197 Raviart,161, 169,192,256 V
l{ed method, 120, 125 Rayleigh quotient, 216
'de superposition, 244, 245 Reactor, 27 5 Vainikko, 231
tnotone, 11 O Ritz method, 24 Varga, 111, 249
•te, 131, 135 Roundoff,110,121,215 Virtual work, 41, 17 3
306 INDEX

w
Weak form, 9; 117, 219
Wendioff, 249
Wilkinson, 116, 238
Wilson, 176,179

z
Zenisek, 84
Zienkiewicz, 183
Zlamal, 164