You are on page 1of 106

Lecture Notes on Finite Element

Methods for Partial Differential


Equations
Endre S uli
Mathematical Institute
University of Oxford
1 December 2012
2
c _ Endre S uli, University of Oxford, 2000.
Contents
1 Introduction 5
1.1 Elements of function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.1 Spaces of continuous functions . . . . . . . . . . . . . . . . . . . . . 6
1.1.2 Spaces of integrable functions . . . . . . . . . . . . . . . . . . . . . . 8
1.1.3 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Weak solutions to elliptic problems . . . . . . . . . . . . . . . . . . . . . . . 14
2 Approximation of elliptic problems 23
2.1 Piecewise linear basis functions . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2 The self-adjoint elliptic problem . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Calculation and assembly of stiness matrix . . . . . . . . . . . . . . . . . . 35
2.4 Galerkin orthogonality; Ceas lemma . . . . . . . . . . . . . . . . . . . . . . 40
2.5 Optimal error bound in the energy norm . . . . . . . . . . . . . . . . . . . . 45
2.6 Superapproximation in mesh-dependent norms . . . . . . . . . . . . . . . . 56
3 Piecewise polynomial approximation 65
3.1 Construction of nite element spaces . . . . . . . . . . . . . . . . . . . . . . 65
3.1.1 The nite element . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.1.2 Examples of triangular nite elements . . . . . . . . . . . . . . . . . 67
3.1.3 The interpolant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.1.4 Examples of rectangular elements . . . . . . . . . . . . . . . . . . . . 73
3.2 Polynomial approximation in Sobolev spaces . . . . . . . . . . . . . . . . . . 74
3.2.1 The Bramble-Hilbert lemma . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.2 Error bounds on the interpolation error . . . . . . . . . . . . . . . . 80
3.3 Optimal error bounds in the H
1
() norm revisited . . . . . . . . . . . . . 83
3.4 Variational crimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4 A posteriori error analysis by duality 89
4.1 The one-dimensional model problem . . . . . . . . . . . . . . . . . . . . . . 89
4.2 An adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5 Evolution problems 95
5.1 The parabolic model problem . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Forward and backward Euler schemes . . . . . . . . . . . . . . . . . . . . . 98
5.3 Stability of -schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.4 Error analysis in the L
2
norm . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3
4 CONTENTS
Synopsis:
Finite element methods represent a powerful and general class of techniques for
the approximate solution of partial dierential equations; the aim of this course is
to provide an introduction to their mathematical theory, with special emphasis on
theoretical questions such as accuracy, reliability and adaptivity; practical issues
concerning the development of ecient nite element algorithms will also be dis-
cussed.
Syllabus:
Elements of function spaces. Elliptic boundary value problems: existence, unique-
ness and regularity of weak solutions.
Finite element methods: Galerkin orthogonality and Ceas lemma. Piecewise
polynomial approximation in Sobolev spaces. The Bramble-Hilbert lemma. Optimal
error bounds in the energy norm. Variational crimes.
The Aubin-Nitsche duality argument. Superapproximation properties in mesh-
dependent norms. A posteriori error analysis by duality: reliability, eciency and
adaptivity.
Finite element approximation of initial boundary value problems. Energy dissi-
pation, conservation and stability. Analysis of nite element methods for evolution
problems.
Reading List
1. S. Brenner & R. Scott, The Mathematical Theory of Finite Element Methods.
Springer-Verlag, 1994. Corr. 2nd printing 1996. [Chapters 0,1,2,3; Chapter 4:
Secs. 4.14.4, Chapter 5: Secs. 5.15.7].
2. K. Eriksson, D. Estep, P. Hansbo, & C. Johnson, Computational Dierential Equa-
tions. CUP, 1996. [Chapters 5, 6, 8, 14 17].
3. C. Johnson, Numerical Solution of Partial Dierential Equations by the Finite El-
ement Method. CUP, 1990. [Chapters 14; Chapter 8: Secs. 8.18.4.2; Chapter 9:
Secs. 9.19.5].
Chapter 1
Introduction
Partial dierential equations arise in the mathematical modelling of many phys-
ical, chemical and biological phenomena and many diverse subject areas such as
uid dynamics, electromagnetism, material science, astrophysics, economy, nancial
modelling, etc. Very frequently the equations under consideration are so compli-
cated that nding their solutions in closed form or by purely analytical means (e.g.
by Laplace and Fourier transform methods, or in the form of a power series) is either
impossible or impracticable, and one has to resort to seeking numerical approxima-
tions to the unknown analytical solution.
These notes are devoted to a particular class of numerical techniques for the
approximate solution of partial dierential equations: nite element methods. They
were proposed in a seminal work of Richard Courant
1
, in 1943; unfortunately, the
relevance of this article was not recognised at the time and the idea was forgotten.
In the early 1950s the method was rediscovered by engineers, but the mathematical
analysis of nite element approximations began much later, in the 1960s, the rst
important results being due to Milos Zl amal
2
in 1968. Since then nite element
methods have been developed into one of the most general and powerful class of
techniques for the numerical solution of partial dierential equations and are widely
used in engineering design and analysis.
In these notes we shall be concerned with the mathematical aspects of nite
element approximation, including stability, accuracy, reliability and adaptivity. We
begin by developing some of the theoretical tools: the present chapter is devoted to
summarising the elements of the theory of function spaces and reviewing some basic
results from the theory of partial dierential equations. The concepts and notational
conventions introduced here will be used systematically throughout the notes.
1
R. Courant: Variational methods for the solution of problems of equilibrium and vibrations.
Bull. Amer. Math. Soc., 49, pp. 123 (1943)
2
M. Zlamal: On the nite element method. Numerische Mathematik, 12, pp. 394402 (1968)
5
6 CHAPTER 1. INTRODUCTION
1.1 Elements of function spaces
As will become apparent in subsequent chapters, the accuracy of nite element ap-
proximations to partial dierential equations very much depends on the smoothness
of the analytical solution to the equation under consideration, and this in turn hinges
on the smoothness of the data.
Precise assumptions about the regularity of the solution and the data can be
conveniently formulated by considering classes of functions with specic dierentia-
bility and integrability properties, called function spaces. In this section we present
a brief overview of basic denitions and simple results form the theory of function
spaces. For future reference, we remark here that all functions that appear in these
notes will be assumed to be real-valued.
1.1.1 Spaces of continuous functions
In this section, we describe some simple function spaces which consist of continuously
dierentiable functions. For the sake of notational convenience, we introduce the
concept of a multi-index.
Let N denote the set of non-negative integers. An n-tuple
= (
1
, . . . ,
n
) N
n
is called a multiindex. The non-negative integer [[ :=
1
+ . . . +
n
is referred
to as the length of the multiindex = (
1
, . . . ,
n
). We denote (0, . . . , 0) by 0;
clearly [0[ = 0. Let
D

=
_

x
1
_

1
. . .
_

x
n
_

n
=

||
x

1
1
. . . x

n
n
.
Example 1 Suppose that n = 3, and = (
1
,
2
,
3
),
j
N, j = 1, 2, 3. Then
for u, a function of three variables x
1
, x
2
, x
3
,

||=3
D

u =

3
u
x
3
1
+

3
u
x
2
1
x
2
+

3
u
x
2
1
x
3
+

3
u
x
1
x
2
2
+

3
u
x
1
x
2
3
+

3
u
x
3
2
+

3
u
x
1
x
2
x
3
+

3
u
x
2
2
x
3
+

3
u
x
2
x
2
3
+

3
u
x
3
3
.
This example highlights the importance of multi-index notation: instead of labori-
ously writing out in detail the ten terms on the right-hand side of the last identity,
we can compress the information into a single entity shown on the left.
Let be an open set in R
n
and let k N. We denote by C
k
() the set of all
continuous real-valued functions dened on such that D

u is continuous on for
1.1. ELEMENTS OF FUNCTION SPACES 7
all = (
1
, . . . ,
n
) with [[ k. Assuming that is a bounded open set, C
k
(

)
will denote the set of all u in C
k
() such that D

u can be extended from to a


continuous function on

, the closure of the set , for all = (
1
, . . . ,
n
), [[ k.
C
k
(

) can be equipped with the norm


|u|
C
k
(

)
:=

||k
sup
x
[D

u(x)[.
In particular when k = 0 we shall write C(

) instead of C
0
(

) to denote the set of


all continuous functions dened on

; in this case,
|u|
C(

)
= sup
x
[u(x)[ = max
x

[u(x)[.
Similarly, if k = 1,
|u|
C
1
(

)
=

||1
sup
x
[D

u(x)[
= sup
x
[u(x)[ +
n

j=1
sup
x
[
u
x
j
(x)[.
Example 2 Consider the open interval = (0, 1) R
1
. The function u(x) = 1/x
belongs to C
k
() for each k 0. As

= [0, 1] and lim
x0
u(x) = , it is clear that
u is not continuous on

; the same is true of its derivatives. Therefore u , C
k
(

)
for any k 0.
The support of a continuous function u dened on an open set R
n
is dened
as the closure in of the set x : u(x) ,= 0. We shall write supp u for the
support of u. Thus, supp u is the smallest closed subset of such that u = 0 in
supp u.
Example 3 Let w be the function dened on R
n
by
w(x) =
_
e

1
1|x|
2
, [x[ < 1,
0, otherwise;
here [x[ = (x
2
1
+ . . . + x
2
n
)
1/2
. Clearly, the support of w is the closed unit ball
x R
n
: [x[ 1.
We denote by C
k
0
() the set of all u contained in C
k
() whose support is a bounded
subset of . Let
C

0
() =

k0
C
k
0
().
Example 4 The function w dened in the previous example belongs to the space
C

0
(R
n
).
8 CHAPTER 1. INTRODUCTION
1.1.2 Spaces of integrable functions
Next we consider a class of spaces that consist of (Lebesgue-) integrable functions.
Let p be a real number, p 1; we denote by L
p
() the set of all real-valued functions
dened on an open subset of R
n
such that
_

[u(x)[
p
dx < .
Any two functions which are equal almost everywhere (i.e. equal, except on a
set of measure zero) on are identied with each other. Thus, strictly speaking,
L
p
() consists of equivalence classes of functions; still, we shall not insist on this
technicality. L
p
() is equipped with the norm
|u|
L
p
()
:=
__

[u(x)[
p
dx
_
1/p
.
We shall also consider the space L

() consisting of functions u dened on such


that [u[ has nite essential supremum on (namely, there exists a positive constant
M such that [u(x)[ M for almost every
3
x in ; the smallest such number M is
called the essential supremum of [u[, and we write M = ess.sup
x
[u(x)[). L

()
is equipped with the norm
|u|
L

()
= ess.sup
x
[u(x)[.
A particularly important case corresponds to taking p = 2; then
|u|
L
2
()
=
__

[u(x)[
2
dx
_
1/2
.
The space L
2
() can be equipped with the inner product
(u, v) :=
_

u(x)v(x) dx.
Clearly |u|
L
2
()
= (u, u)
1/2
.
Lemma 1 (The Cauchy-Schwarz inequality) Let u and v belong to L
2
(); then
u v L
1
() and
[(u, v)[ |u|
L
2
()
|v|
L
2
()
.
3
We shall say that a property P(x) is true for almost every x in , if P(x) is true for all x
where is a subset of with zero Lebesgue measure.
1.1. ELEMENTS OF FUNCTION SPACES 9
Proof Let R; then
0 |u +v|
2
L
2
()
= (u +v, u +v)
= (u, u) + (u, v) + (v, u) + (v, v)
= |u|
2
L
2
()
+ 2(u, v) +
2
|v|
2
L
2
()
, R.
The right-hand side is a quadratic polynomial in with real coecients, and it is non-
negative for all R; therefore its discriminant is non-positive, i.e.
[2(u, v)[
2
4|u|
2
L
2
()
|v|
2
L
2
()
0,
and hence the desired inequality.
Corollary 1 (The triangle inequality) Let u and v belong to L
2
(); then u + v
L
2
(), and
|u + v|
L
2
()
|u|
L
2
()
+|v|
L
2
()
.
Proof This is a straightforward consequence of the Cauchy-Schwarz inequality:
|u +v|
2
L
2
()
= (u +v, u +v) = |u|
2
L
2
()
+ 2(u, v) +|v|
2
L
2
()

_
|u|
L
2
()
+|v|
L
2
()
_
2
.
Upon taking the square root of both sides we complete the proof.
Remark 1 The space L
p
() with p [1, ] is a Banach space
4
. In particular,
L
2
() is a Hilbert space: it has an inner product (, ) and, when equipped with the
associated norm | |
L
2
()
, dened by |u|
L
2
()
= (u, u)
1/2
, it is a Banach space.
To conclude this section, we note that a statement analogous to Corollary 1
holds, more generally, in the L
p
norm for 1 p ; namely,
|u + v|
L
p
()
|u|
L
p
()
+|v|
L
p
()
, u, v L
p
().
Furthermore, the following generalisation of the Cauchy-Schwarz inequality, known
as Holders inequality, is valid for any two functions u L
p
() and v L
p
()
with 1/p + 1/p

= 1:

u(x)v(x) dx

|u|
L
p
()
|v|
L
p
()
.
4
A normed linear space X, with norm | |
X
, is called a Banach space if, whenever u
m

m=1
is
a sequence of elements of X such that
lim
n,m
|u
n
u
m
|
X
= 0, (1.1)
there exists u X such that lim
m
|u u
m
|
X
= 0 (i.e. the sequence u
m

m=1
converges to u
in X). A sequence u
m

m=1
with the property (1.1) is called a Cauchy sequence.
10 CHAPTER 1. INTRODUCTION
1.1.3 Sobolev spaces
In this section we introduce a class of spaces, called Sobolev spaces (after the Russian
mathematician S.L. Sobolev), which play an important role in modern dierential
equation theory. Before we give the precise denition of a Sobolev space, we intro-
duce the concept of weak derivative.
Suppose that u is a smooth function, say u C
k
(), with an open subset of
R
n
, and let v C

0
(); then the following integration-by-parts formula holds:
_

u(x) v(x) dx = (1)


||
_

u(x) D

v(x) dx, [[ k,
v C

0
().
Note that all terms involving integrals over the boundary of , which arise in the
course of integrating by parts, have disappeared because v and all of its derivatives
are identically zero on the boundary of . This identity represents the starting point
for dening the concept of weak derivative.
Now suppose that u is a locally integrable function dened on (i.e. u L
1
()
for each bounded open set , with ). Suppose also that there exists a function
w

, locally integrable on and such that


_

(x) v(x) dx = (1)


||
_

u(x) D

v(x) dx v C

0
();
then we say that w

is a weak derivative of the function u of order [[ =


1
+. . . +

n
, and we write w

= D

u. In order to see that this denition is correct it has to


be shown that if a locally integrable function has a weak derivative then this must be
unique; we remark that this is a straightforward consequence of DuBois Reymonds
lemma
5
. Clearly, if u is a suciently smooth function, say u C
k
(), then its weak
derivative D

u of order [[ k coincides with the corresponding partial derivative


in the classical pointwise sense,

||
u
x

1
1
. . . x

n
n
.
In order to simplify the notation, we shall use the letter D to denote classical as
well as weak derivatives; it will always be clear from the context (by considering the
smoothness of the function dierentiated) which of the two is implied.
Example 5 Let = R
1
, and suppose that we wish to determine the weak rst
derivative of the function u(x) = (1 [x[)
+
dened on . Clearly u is not dieren-
tiable at the points 0 and 1. However, because u is locally integrable on , it may,
5
DuBois Reymonds lemma: Suppose that w is a locally integrable function dened on an
open set , R
n
. If
_

w(x)v(x) dx = 0 for all v in C

0
()
then w(x) = 0 for almost every x .
1.1. ELEMENTS OF FUNCTION SPACES 11
nevertheless, have a weak derivative. Indeed, for any v C

0
(),
_
+

u(x)v

(x) dx =
_
+

(1 [x[)
+
v

(x) dx =
_
1
1
(1 [x[)v

(x) dx
=
_
0
1
(1 + x)v

(x) dx +
_
1
0
(1 x)v

(x) dx
=
_
0
1
v(x) dx + (1 + x)v(x)[
0
1
+
_
1
0
v(x) dx + (1 x)v(x)[
1
x=0
=
_
0
1
(1)v(x) dx +
_
1
0
1 v(x) dx
_
+

w(x)v(x) dx,
where
w(x) =
_

_
0, x < 1,
1, x (1, 0),
1, x (0, 1),
0, x > 1.
Thus, the piecewise constant function w is the rst (weak) derivative of the contin-
uous piecewise linear function u, i.e. w = u

= Du.
Now we are ready to give a precise denition of a Sobolev space. Let k be a
non-negative integer and suppose that p [1, ]. We dene (with D

denoting a
weak derivative of order [[ )
W
k
p
() = u L
p
() : D

u L
p
(), [[ k.
W
k
p
() is called a Sobolev space of order k; it is equipped with the (Sobolev) norm
|u|
W
k
p
()
:=
_
_

||k
|D

u|
p
L
p
()
_
_
1/p
when 1 p <
and
|u|
W
k

()
:=

||k
|D

u|
L

()
when p = .
Letting,
[u[
W
k
p
()
:=
_
_

||=k
|D

u|
p
L
p
()
_
_
1/p
,
for p [1, ), we can write
|u|
W
k
p
()
=
_
k

j=0
[u[
p
W
j
p
()
_
1/p
.
12 CHAPTER 1. INTRODUCTION
Similarly, letting
[u[
W
k

()
:=

||=k
|D

u|
L

()
,
we have that
|u|
W
k

()
=
k

j=0
[u[
W
j

()
.
When k 1, [[
W
k
p
()
is called the Sobolev semi-norm
6
on W
k
p
().
An important special case corresponds to taking p = 2; the space W
k
2
() is then
a Hilbert space with the inner product
(u, v)
W
k
2
()
:=

||k
(D

u, D

v).
For this reason, we shall usually write H
k
() instead of W
k
2
().
Throughout these notes we shall frequently refer to the Hilbertian Sobolev spaces
H
1
() and H
2
(). Our denitions of W
k
p
() and its norm and seminorm, for p = 2,
k = 1, give:
H
1
() =
_
u L
2
() :
u
x
j
L
2
(), j = 1, . . . , n
_
,
|u|
H
1
()
=
_
|u|
2
L
2
()
+
n

j=1
|
u
x
j
|
2
L
2
()
_
1/2
,
[u[
H
1
()
=
_
n

j=1
|
u
x
j
|
2
L
2
()
_
1/2
.
Similarly, for p = 2 and k = 2,
H
2
() =
_
u L
2
() :
u
x
j
L
2
(), j = 1, . . . , n,

2
u
x
i
x
j
L
2
(), i, j = 1, . . . , n
_
,
|u|
H
2
()
=
_
|u|
2
L
2
()
+
n

j=1
|
u
x
j
|
2
L
2
()
+
n

i,j=1
|

2
u
x
i
x
j
|
2
L
2
()
_
1/2
,
6
When k 1, [ [
W
k
p
()
is only a semi-norm rather than a norm because if [u[
W
k
p
()
= 0 for
u W
k
p
() it does not necessarily follow that u(x) = 0 for almost every x in (all that is known
is that D

u(x) = 0 for almost every x , [[ = k), so [ [


W
k
p
()
does not satisfy the rst axiom
of norm.
1.1. ELEMENTS OF FUNCTION SPACES 13
[u[
H
2
()
=
_
n

i,j=1
|

2
u
x
i
x
j
|
2
L
2
()
_
1/2
.
Finally, we dene the special Sobolev space H
1
0
() as the closure of C

0
() in the
norm of | |
H
1
()
; in other words, H
1
0
() is the set of all u H
1
() such that u
is the limit in H
1
() of a sequence u
m

m=1
with u
m
C

0
(). It can be shown
(assuming that is suciently smooth) that
H
1
0
() = u H
1
() : u = 0 on ;
i.e. H
1
0
() is, in fact, the set of all functions u in H
1
() such that u = 0 on , the
boundary of the set . We shall use this space when considering a partial dierential
equation that is coupled with a homogeneous (Dirichlet) boundary condition: u = 0
on . We note here that H
1
0
() is also a Hilbert space, with the same norm and
inner product as H
1
().
We conclude the section with the following useful result.
Lemma 2 (Poincare-Friedrichs inequality) Suppose that is a bounded open set in
R
n
(with a suciently smooth boundary
7
) and let u H
1
0
(); then there exists
a constant c

(), independent of u, such that


_

[u(x)[
2
dx c

i=1
_

[
u
x
i
(x)[
2
dx. (1.2)
Proof As any function u H
1
0
() is the limit in H
1
() of a sequence u
m

m=1
C

0
(),
it is sucient to prove this inequality for u C

0
().
In fact, to simplify matters, we shall restrict ourselves to considering the special case
of a rectangular domain = (a, b) (c, d) in R
2
. The proof for general is analogous.
Evidently
u(x, y) = u(a, y) +
_
x
a
u
x
(, y) d =
_
x
a
u
x
(, y) d, c < y < d.
Thence, by the Cauchy-Schwarz inequality,
_

[u(x, y)[
2
dxdy =
_
b
a
_
d
c
[
_
x
a
u
x
(, y) d[
2
dy dx

_
b
a
_
d
c
(x a)
__
x
a
[
u
x
(, y)[
2
d
_
dy dx

_
b
a
(x a) dx
__
d
c
_
b
a
[
u
x
(, y)[
2
d dy
_
=
1
2
(b a)
2
_

[
u
x
(x, y)[
2
dxdy.
7
Say, is a polygonal domain in R
2
or a polyhedron in R
3
.
14 CHAPTER 1. INTRODUCTION
Analogously,
_

[u(x, y)[
2
dxdy
1
2
(d c)
2
_

[
u
y
(x, y)[
2
dxdy.
By adding the two inequalities, we obtain
_

[u(x, y)[
2
dxdy c

_
[
u
x
[
2
+[
u
y
[
2
_
dxdy,
where c

=
_
2
(ba)
2
+
2
(dc)
2
_
1
.
For further reference, we note that if = (0, 1)
2
R
2
then c

=
1
4
; similarly, if
= (0, 1) R then c

=
1
2
.
1.2 Weak solutions to elliptic problems
In the rst part of this lecture course we shall focus on boundary value problems for
elliptic partial dierential equations. Elliptic equations are typied by the Laplace
equation
u = 0,
and its non-homogeneous counterpart, Poissons equation
u = f,
where we used the notation
=
n

i=1

2
x
2
i
for the Laplace operator.
More generally, let be a bounded open set in R
n
, and consider the linear
second-order partial dierential equation

i,j=1

x
j
_
a
ij
(x)
u
x
i
_
+
n

i=1
b
i
(x)
u
x
i
+ c(x)u = f(x), x , (1.3)
where the coecients a
ij
, b
i
, c and f satisfy the following conditions:
a
ij
C
1
(

), i, j = 1, . . . , n;
b
i
C(

), i = 1, . . . , n;
c C(

), f C(

),
and
n

i,j=1
a
ij
(x)
i

j
c
n

i=1

2
i
, = (
1
, . . . ,
n
) R
n
, x

; (1.4)
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 15
here c is a positive constant independent of x and . The condition (1.4) is usually
referred to as uniform ellipticity and (1.3) is called an elliptic equation.
In problems that arise in applications equation (1.3) is usually supplemented by
one of the following boundary conditions, with g denoting a given function dened
on :
(a) u = g on (Dirichlet boundary condition);
(b)
u

= g on , where denotes the unit outward normal vector to (Neu-


mann boundary condition);
(c)
u

+ u = g on , where (x) 0 on (Robin boundary condition);


(d) A generalisation of the boundary conditions (b) and (c) is
n

i,j=1
a
ij
u
x
i
cos
j
+ (x)u = g on ,
where
j
is the angle between the unit outward normal vector to and
the x
j
axis (Oblique derivative boundary condition).
In many physical problems more than one type of boundary condition is imposed
on (e.g. is the union of two disjoint subsets
1
and
2
, with a Dirichlet
boundary condition on
1
and Neumann boundary condition on
2
). The study
of such mixed boundary value problems will not be pursued in these notes.
We begin by considering the homogeneous Dirichlet boundary value problem

i,j=1

x
j
_
a
ij
u
x
i
_
+
n

i=1
b
i
(x)
u
x
i
+ c(x)u = f(x), x , (1.5)
u = 0 on , (1.6)
where a
ij
, b
i
, c and f are as in (1.4).
A function u C
2
()C(

) satisfying (1.5) and (1.6) is called a classical solu-


tion of this problem. The theory of partial dierential equations tells us that (1.5),
(1.6) has a unique classical solution, provided that a
ij
, b
i
, c, f and are suciently
smooth. However, in many applications one has to consider equations where these
smoothness requirements are violated, and for such problems the classical theory is
inappropriate. Take, for example, Poissons equation with zero Dirichlet boundary
condition on = (1, 1)
n
in R
n
:
u = sgn
_
1
2
[x[
_
, x ,
u = 0, x .
_
()
This problem does not have a classical solution, u C
2
()C(

), for otherwise u
would be a continuous function on , which is not possible because sgn(1/2 [x[)
is not continuous on .
16 CHAPTER 1. INTRODUCTION
In order to overcome the limitations of the classical theory and to be able to
deal with partial dierential equations with non-smooth data, we generalise the
notion of solution by weakening the dierentiability requirements on u.
To begin, let us suppose that u is a classical solution of (1.5), (1.6). Then, for
any v C
1
0
(),

i,j=1
_

x
j
_
a
ij
u
x
i
_
v dx +
n

i=1
_

b
i
(x)
u
x
i
v dx
+
_

c(x)uv dx =
_

f(x)v(x) dx.
Upon integration by parts in the rst integral and noting that v = 0 on , we
obtain:
n

i,j=1
_

a
ij
(x)
u
x
i
v
x
j
dx +
n

i=1
_

b
i
(x)
u
x
i
v dx
+
_

c(x)uv dx =
_

f(x)v(x) dx v C
1
0
().
In order for this equality to make sense we no longer need to assume that u C
2
():
it is sucient that u L
2
() and u/x
i
L
2
(), i = 1, . . . , n. Thus, remembering
that u has to satisfy a zero Dirichlet boundary condition, it is natural to seek u in
the space H
1
0
(), where, as in Section 1.1.3,
H
1
0
() = u L
2
() :
u
x
i
L
2
(), i = 1, . . . , n, u = 0 on .
Therefore, we consider the following problem: nd u in H
1
0
() such that
n

i,j=1
_

a
ij
(x)
u
x
i

v
x
j
dx +
n

i=1
_

b
i
(x)
u
x
i
v dx
+
_

c(x)uv dx =
_

f(x)v(x) dx v C
1
0
(). (1.7)
We note that C
1
0
() H
1
0
(), and it is easily seen that when u H
1
0
() and
v H
1
0
(), (instead of v C
1
0
()), the expressions on the left- and right-hand side
of (1.7) are still meaningful (in fact, we shall prove this below)
8
. This motivates the
following denition.
8
Note further that since the coecients a
ij
no longer appear under derivative signs in (1.7),
it is not necessary to assume that a
ij
C
1
(

); a
ij
L

() will be seen to be sucient. Also,


the smoothness requirements imposed on the coecients b
i
and c can be relaxed: b
i
L

() for
i = 1, . . . , n and c L

() will suce.
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 17
Denition 1 Let a
ij
L

(), i, j = 1, . . . , n, b
i
L

(), i = 1, . . . , n, c
L

(), and let f L


2
(). A function u H
1
0
() satisfying
n

i,j=1
_

a
ij
(x)
u
x
i
v
x
j
dx +
n

i=1
_

b
i
(x)
u
x
i
v dx
+
_

c(x)uv dx =
_

f(x)v(x) dx v H
1
0
() (1.8)
is called a weak solution of (1.5), (1.6). All partial derivatives in (1.8) should be
understood as weak derivatives.
Clearly if u is a classical solution of (1.5), (1.6), then it is also a weak solution
of (1.5), (1.6). However, the converse is not true. If (1.5), (1.6) has a weak solution,
this may not be smooth enough to be a classical solution. Indeed, we shall prove
below that the boundary value problem () has a unique weak solution u H
1
0
(),
despite the fact that it has no classical solution. Before considering this particular
boundary value problem, we look at the wider issue of existence of a unique weak
solution to the more general problem (1.5), (1.6).
For the sake of simplicity, we adopt the following notation:
a(w, v) =
n

i,j=1
_

a
ij
(x)
w
x
i
v
x
j
dx
+
n

i=1
_

b
i
(x)
w
x
i
v dx +
_

c(x)wv dx (1.9)
and
l(v) =
_

f(x)v(x) dx. (1.10)


With this new notation, problem (1.8) can be written as follows:
nd u H
1
0
() such that a(u, v) = l(v) v H
1
0
(). (1.11)
We shall prove the existence of a unique solution to this problem by exploiting the
following abstract result from Functional Analysis.
Theorem 1 (Lax & Milgram theorem) Suppose that V is a real Hilbert space equipped
with norm | |
V
. Let a(, ) be a bilinear functional on V V such that:
(a) c
0
> 0 v V a(v, v) c
0
|v|
2
V
,
(b) c
1
> 0 v, w V [a(w, v)[ c
1
|w|
V
|v|
V
,
and let l() be a linear functional on V such that
(c) c
2
> 0 v V [l(v)[ c
2
|v|
V
.
18 CHAPTER 1. INTRODUCTION
Then, there exists a unique u V such that
a(u, v) = l(v) v V.
For a proof of this result the interested reader is referred to the books: P. Ciarlet:
The Finite Element Method for Elliptic Problems, North-Holland, 1978; K. Yosida:
Functional Analysis, Reprint of the 6th ed., Springer-Verlag, 1995.
We apply the Lax-Milgram theorem with V = H
1
0
() and | |
V
= | |
H
1
()
to show the existence of a unique weak solution to (1.5), (1.6) (or, equivalently, to
(1.11)). Let us recall from Section 1.1.3 that H
1
0
() is a Hilbert space with the inner
product
(w, v)
H
1
()
=
_

wv dx +
n

i=1
_

w
x
i

v
x
i
dx
and the associated norm |w|
H
1
()
= (w, w)
1/2
H
1
()
. Next we show that a(, ) and l(),
dened by (1.9) and (1.10), satisfy the hypotheses (a), (b), (c) of the Lax-Milgram
theorem.
We begin with (c). The mapping v l(v) is linear: indeed, for any , R,
l(v
1
+ v
2
) =
_

f(x)(v
1
(x) + v
2
(x)) dx
=
_

f(x)v
1
(x) dx +
_

f(x)v
2
(x) dx
= l(v
1
) + l(v
2
), v
1
, v
2
H
1
0
();
so l() is a linear functional on H
1
0
(). Also, by the Cauchy-Schwarz inequality,
[l(v)[ = [
_

f(x)v(x) dx[
__

[f(x)[
2
dx
_
1/2
__

[v(x)[
2
dx
_
1/2
= |f|
L
2
()
|v|
L
2
()
|f|
L
2
()
|v|
H
1
()
,
for all v H
1
0
(), where we have used the obvious inequality |v|
L
2
()
|v|
H
1
()
.
Letting c
2
= |f|
L
2
()
, we obtain the required bound.
Next we verify (b). For any xed w H
1
0
(), the mapping v a(v, w) is linear.
Similarly, for any xed v H
1
0
(), the mapping w a(v, w) is linear. Hence a(, )
is a bilinear functional on H
1
0
() H
1
0
(). Applying the Cauchy-Schwarz inequality,
we deduce that
[a(w, v)[
n

i,j=1
max
x

[a
ij
(x)[ [
_

w
x
i
v
x
j
dx[
+
n

i=1
max
x

[b
i
(x)[ [
_

w
x
i
v dx[
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 19
+max
x

[c(x)[ [
_

w(x)v(x) dx[
c
_
n

i,j=1
__

[
w
x
i
[
2
dx
_
1/2
__

[
v
x
j
[
2
dx
_
1/2
+
n

i=1
__

[
w
x
i
[
2
dx
_
1/2
__

[v[
2
dx
_
1/2
+
__

[w[
2
dx
_
1/2
__

[v[
2
dx
_
1/2
_
c
_
__

[w[
2
dx
_
1/2
+
n

i=1
__

[
w
x
i
[
2
dx
_
1/2
_

_
__

[v[
2
dx
_
1/2
+
n

j=1
__

[
v
x
j
[
2
dx
_
1/2
_
(1.12)
where
c = max
_
max
1i,jn
max
x

[a
ij
(x)[, max
1in
max
x

[b
i
(x)[, max
x

[c(x)[
_
.
By further majorisation of the right-hand side in (1.12) we deduce that
[a(w, v)[ 2n c
_
_

[w[
2
dx +
n

i=1
_

[
w
x
i
[
2
dx
_
1/2

_
_

[v[
2
dx +
n

j=1
_

[
v
x
j
[
2
dx
_
1/2
,
so that, by letting c
1
= 2n c, we obtain inequality (b):
[a(w, v)[ c
1
|w|
H
1
()
|v|
H
1
()
. (1.13)
It remains to establish (a). To do so, we shall slightly strengthen the smoothness
requirements on the coecients b
i
by demanding that b
i
W
1

() (see, however,
Remark 4 at the end of this chapter). Using (1.4), we deduce that
a(v, v) c
n

i=1
_

[
v
x
i
[
2
dx +
n

i=1
_

b
i
(x)
1
2

x
i
(v
2
) dx +
_

c(x)[v[
2
dx,
where we wrote
v
x
i
v as
1
2

x
i
(v
2
). Integrating by parts in the second term on the
right, we obtain
a(v, v) c
n

i=1
_

[
v
x
i
[
2
dx +
_

_
c(x)
1
2
n

i=1
b
i
x
i
_
[v[
2
dx.
20 CHAPTER 1. INTRODUCTION
Suppose that b
i
, i = 1, . . . , n, and c satisfy the inequality
c(x)
1
2
n

i=1
b
i
x
i
0, x

. (1.14)
Then
a(v, v) c
n

i=1
_

[
v
x
i
[
2
dx. (1.15)
By virtue of the Poincare-Friedrichs inequality stated in Lemma 1.2, the right-hand
side can be further bounded below to obtain
a(v, v)
c
c

[v[
2
dx. (1.16)
Summing (1.15) and (1.16),
a(v, v) c
0
_
_

[v[
2
dx +
n

i=1
_

[
v
x
i
[
2
dx
_
, (1.17)
where c
0
= c/(1 + c

), and hence (a). Having checked all hypotheses of the Lax-


Milgram theorem, we deduce the existence of a unique u H
1
0
() satisfying (1.11);
consequently, problem (1.5), (1.6) has a unique weak solution. We encapsulate this
result in the following theorem.
Theorem 2 Suppose that a
ij
L

(), i, j = 1, . . . , n, b
i
W
1

(), i = 1, . . . , n,
c L

(), f L
2
(), and assume that (1.4) and (1.14) hold; then the boundary
value problem (1.5), (1.6) possesses a unique weak solution u H
1
0
(). In addition,
|u|
H
1
()

1
c
0
|f|
L
2
()
. (1.18)
Proof We only have to show (1.18) as the rest of the theorem has been proved above.
By (1.17), (1.11), the Cauchy-Schwarz inequality and recalling the denition of | |
H
1
()
,
c
0
|u|
2
H
1
()
a(u, u) = l(u) = (f, u)
[(f, u)[ |f|
L
2
()
|u|
L
2
()
|f|
L
2
()
|u|
H
1
()
.
Hence the desired inequality.
Now we return to our earlier example () which has been shown to have no
classical solution. However, applying the above theorem with a
ij
(x) 1, i = j,
a
ij
(x) 0, i ,= j, 1 i, j n, b
i
(x) 0, c(x) 0, f(x) = sgn(
1
2
[x[), and
= (1, 1)
n
, we see that (1.4) holds with c = 1 and (1.14) is trivially fullled.
Thus () has a unique weak solution u H
1
0
() by Theorem 2. Similar results
are valid in the case of Neumann, Robin, and oblique derivative boundary value
problems, as well as mixed problems.
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 21
Remark 2 Consider, for example, the following Dirichlet-Neumann mixed bound-
ary value problem:
u = f in ,
u = 0 on
1
,
u

= g on
2
,
where
1
is a non-empty, relatively open subset of and
1

2
= . We shall
suppose that f L
2
() and that g L
2
(
2
). Following a similar reasoning as in the
case of the Dirichlet boundary value problem, we consider the special Sobolev space
H
1
0,
1
() = v H
1
() : v = 0 on
1
,
and dene the weak formulation of the mixed problem as follows: nd u H
1
0,
1
()
such that
a(u, v) = l(v) for all v in H
1
0,
1
(),
where we put
a(u, v) =
_

i=1
u
x
i
v
x
i
dx
and
l(v) =
_

f(x)v(x) dx +
_

2
g(s)v(s) ds.
Applying the Lax-Milgram theorem with V = H
1
0,
1
(), the existence and uniqueness
of a weak solution to this mixed problem easily follows.
Remark 3 Theorem 2 implies that the weak formulation of the elliptic boundary
value problem (1.5), (1.6) is well-posed in the sense of Hadamard; namely, for each
f L
2
() there exists a unique (weak) solution u H
1
0
(), and small changes in
f give rise to small changes in the corresponding solution u. The latter property
follows by noting that if u
1
and u
2
are weak solutions in H
1
0
() of (1.5), (1.6)
corresponding to right-hand sides f
1
and f
2
in L
2
(), respectively, then u
1
u
2
is the weak solution in H
1
0
() of (1.5), (1.6) corresponding to the right-hand side
f
1
f
2
L
2
(). Thus, by virtue of (1.18),
|u
1
u
2
|
H
1
()

1
c
0
|f
1
f
2
|
L
2
()
, (1.19)
and hence the required continuous dependence of the solution of the boundary value
problem on the right-hand side.
Remark 4 The requirement b
i
W
1

() in Theorem 2 can be relaxed to the original


assumption b
i
L

(), i = 1, . . . , n. To see this, note that the smoothness require-


ments on b
i
are unrelated to the verication of condition (c) in the Lax-Milgram
22 CHAPTER 1. INTRODUCTION
theorem, and condition (b) can be shown with b
i
L

(), i = 1, . . . , n, only any-


way. Thus, it remains to see how condition (a) may be veried under the hypothesis
b
i
L

(), i = 1, . . . , n. By (1.4) and the Cauchy-Schwarz inequality,


a(v, v) c[v[
2
H
1
()

_
n

i=1
|b
i
|
2
L

()
_
1/2
[v[
H
1
()
|v|
L
2
()
+
_

c(x)[v(x)[
2
dx

1
2
c[v[
2
H
1
()
+
_

_
c(x)
2
c
n

i=1
|b
i
|
2
L

()
_
[v(x)[
2
dx.
Assuming that
c(x)
2
c
n

i=1
|b
i
|
2
L

()
0 (1.20)
we arrive at the inequality
a(v, v)
1
2
c
n

i=1
_

[
v
x
i
[
2
dx,
which is analogous to (1.15). Thus, proceeding in the same way as in the transi-
tion from (1.15) to (1.17) we arrive at (1.17) with c
0
= c/(2 + 2c

); this veries
condition (a) in the Lax-Milgram theorem, under the assumptions that b
i
L

(),
i = 1, . . . , n, only and (1.4), (1.20) hold.
Chapter 2
Approximation of elliptic problems
In this chapter we describe the construction of nite element methods for ellip-
tic boundary value problems and outline some of their key properties. Unlike nite
dierence schemes which are constructed in a more-or-less ad hoc fashion through re-
placing the derivatives in the dierential equation by divided dierences, the deriva-
tion of nite element methods is quite systematic.
The rst step in the construction of a nite element method for an elliptic bound-
ary value problem (e.g. (1.5), (1.6)) is to convert the problem into its weak formu-
lation:
nd u V such that a(u, v) = l(v) v V , (P)
where V is the solution space (e.g. H
1
0
() for the homogeneous Dirichlet boundary
value problem), a(, ) is a bilinear functional on V V , and l() is a linear functional
on V (e.g. (1.9) and (1.10)).
The second step in the construction is to replace V in (P) by a nite-dimensional
subspace V
h
V which consists of continuous piecewise polynomial functions of
a xed degree associated with a subdivision of the computational domain; then
consider the following approximation of (P):
nd u
h
V
h
such that a(u
h
, v
h
) = l(v
h
) v
h
V
h
. (P
h
)
Suppose, for example, that
dimV
h
= N(h) and V
h
= span
1
, . . . ,
N(h)
,
where the (linearly independent) basis functions
i
, i = 1, . . . , N(h), have small
support. Expressing the approximate solution u
h
in terms of the basis functions,
i
,
we can write
u
h
(x) =
N(h)

i=1
U
i

i
(x), ()
where U
i
, i = 1, . . . , N(h), are to be determined. Thus (P
h
) can be rewritten as
follows:
nd (U
1
, . . . , U
N(h)
) R
N(h)
such that
23
24 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
N(h)

i=1
a(
i
,
j
)U
i
= l(
j
), j = 1, . . . , N(h). (P

h
)
This is a system of linear equations for U = (U
1
, . . . , U
N(h)
)
T
, with the matrix
of the system A = (a(
j
,
i
)) of size N(h) N(h). Because the
i
s have small
support, a(
j
,
i
) = 0 for most pairs of i and j, so the matrix A is sparse (in the
sense that most of its entries are equal to 0); this property is crucial from the point
of ecient solution in particular, fast iterative methods are available for sparse
linear systems. Once (P

h
) has been solved for U = (U
1
, . . . , U
N(h)
)
T
, the expansion
() provides the required approximation to u.
After this brief outline of the idea behind the nite element method, we illustrate
the construction of this numerical technique by considering some simple examples.
2.1 Piecewise linear basis functions
In this section we describe the construction of the nite element method through
two simple examples: the rst of these is a two-point boundary value problem
for a second-order ordinary dierential equation; the second model problem is the
homogeneous Dirichlet boundary value problem for Poissons equation on the unit
square in the plane. For the time being we shall assume that the nite element
space V
h
consists of continuous piecewise linear functions. Higher-degree piecewise
polynomial approximations will be discussed later on in the notes.
One-dimensional problem
Let us consider the boundary value problem
(p(x)u

+ q(x)u = f(x), x (0, 1), (2.1)


u(0) = 0, u(1) = 0, (2.2)
where p C[0, 1], q C[0, 1], f L
2
(0, 1) with p(x) c > 0 and q(x) 0 for all x
in [0, 1]. The weak formulation of this problem is:
nd u H
1
0
(0, 1) such that
_
1
0
p(x)u

(x)v

(x) dx +
_
1
0
q(x)u(x)v(x) dx =
_
1
0
f(x)v(x) dx
v H
1
0
(0, 1).
_

_
(P)
In order to construct the nite element approximation of this problem, we sub-
divide

= [0, 1] into N subintervals [x
i
, x
i+1
], i = 0, . . . , N 1, by the points
x
i
= ih, i = 0, . . . , N, where h = 1/N, N 2, as shown in Fig 2.1. We note that
in general the mesh points x
i
need not be equally spaced: here we have chosen a
uniform spacing only to simplify the exposition.
2.1. PIECEWISE LINEAR BASIS FUNCTIONS 25
0 = x
0
x
1
x
2
. . . x
i
. . . x
N
= 1
Figure 2.1: Subdivision of

= [0, 1].

Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
x
i
x
i1
1
x
i+1
Figure 2.2: The piecewise linear nite element basis function
i
(x).
The subintervals (x
i
, x
i+1
) are referred to as element domains or elements,
(hence the name nite element method). In this example, the weak solution
u H
1
0
(0, 1) of problem (P) will be approximated by a continuous piecewise linear
function on the subdivision depicted in Figure 2.1. It will be convenient to express
our approximation as a linear combination of the nite element basis functions

i
(x) =
_
1 [
x x
i
h
[
_
+
, i = 1, . . . , N 1,
shown in Figure 2.2. It is clear that
i
H
1
0
(0, 1); furthermore, supp
i
= [x
i1
, x
i+1
],
i = 1, . . . , N 1, and the functions
i
, i = 1, . . . , N 1, are linearly independent;
therefore
V
h
:= span
1
, . . . ,
N1

is an (N 1)-dimensional subspace of H
1
0
(0, 1).
The nite element approximation of (P) is:
nd u
h
V
h
such that
_
1
0
p(x)u

h
(x)v

h
(x) dx +
_
1
0
q(x)u
h
(x)v
h
(x) dx
=
_
1
0
f(x)v
h
(x) dx v
h
V
h
.
_

_
(P
h
)
Since u
h
V
h
= span
1
, . . . ,
N1
, it can be written as a linear combination
26 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
of the basis functions:
u
h
(x) =
N1

i=1
U
i

i
(x).
Substituting this expansion into (P
h
) we obtain the following problem, equivalent
to (P
h
):
nd U = (U
1
, . . . , U
N1
)
T
R
N1
such that
N1

i=1
U
i
_
1
0
[p(x)

i
(x)

j
(x) + q(x)
i
(x)
j
(x)] dx
=
_
1
0
f(x)
j
(x) dx,
for j = 1, . . . , N 1.
_

_
(P

h
)
Letting
a
ji
:=
_
1
0
[p(x)

i
(x)

j
(x) + q(x)
i
(x)
j
(x)] dx, i, j = 1, . . . , N 1;
F
j
:=
_
1
0
f(x)
j
(x) dx, j = 1, . . . , N 1,
(P

h
) can be written as a system of linear equations
AU = F,
where A = (a
ji
), F = (F
1
, . . . , F
N1
)
T
. The matrix A is symmetric (i.e. A
T
= A)
and positive denite (i.e. x
T
Ax > 0, x ,= 0). Since supp
i
supp
j
has empty
interior when [i j[ > 1, it follows that the matrix A is tri-diagonal (namely, a
ji
is
zero, unless [i j[ 1). Having solved the system of linear equations AU = F, we
substitute the values U
1
, . . . , U
N1
into the expansion
u
h
(x) =
N1

i=1
U
i

i
(x)
to obtain u
h
.
In practice the entries a
ji
of the matrix A and the entries F
j
of the vector F
are calculated approximately using numerical integration (quadrature) rules. In the
simple case when p and q are constant functions on [0, 1], the entries of A can be
calculated exactly:
a
ij
= p
_
1
0

i
(x)

j
(x) dx + q
_
1
0

i
(x)
j
(x) dx
= p
_
_
_
2/h, i = j,
1/h, [i j[ = 1,
0, [i j[ > 1,
+ q
_
_
_
4h/6, i = j,
h/6, [i j[ = 1,
0, [i j[ > 1.
=
_
_
_
2p/h + 4hq/6, i = j,
p/h + qh/6, [i j[ = 1,
0, [i j[ > 1.
2.1. PIECEWISE LINEAR BASIS FUNCTIONS 27
Figure 2.3: A subdivision (triangulation) of

.
This gives rise to the following set of linear equations:
p
U
i1
2U
i
+ U
i+1
h
2
+ q
U
i1
+ 4U
i
+ U
i+1
6
=
1
h
_
x
i+1
x
i1
f(x)
i
(x) dx,
i = 1, . . . , N 1,
with the convention that U
0
= 0 and U
N
= 0 (corresponding to the fact that
u
h
(0) = 1 and u
h
(1) = 0, respectively). This is a three-point nite dierence scheme
for the values U
i
, the values of u
h
(x) at the mesh points x
i
.
Two-dimensional problem
Let be a bounded domain in R
2
with a polygonal boundary ; thus can be
exactly covered by a nite number of triangles. It will be assumed that any pair of
triangles in a triangulation of intersect along a complete edge, at a vertex, or not
at all, as shown in Fig. 2.3. We shall denote by h
K
the diameter (longest side) of
triangle K, and we dene h = max
K
h
K
.
With each interior node (marked in the gure) we associate a basis function
which is equal to 1 at that node and equal to 0 at all the other nodes; is assumed
to be a continuous function on

and linear in each of the triangles, as shown in
Fig. 2.4.
Let us suppose that the interior nodes are labelled 1, 2, . . . , N(h); let
1
(x, y),
. . . ,
N(h)
(x, y) be the corresponding basis functions. The functions
1
, . . . ,
N(h)
are linearly independent and they span an N(h)-dimensional linear subspace V
h
of
H
1
0
().
Let us consider the elliptic boundary value problem
u = f in ,
28 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
0 0
0 0
0 0
1
Figure 2.4: A typical nite element basis function .
u = 0 on .
In order to construct the nite element approximation of the problem, we begin by
considering its weak formulation (see the discussion about weak solutions in Chapter
1 in the special case when n = 2, a
ij
(x) 1 for i = j and 0 for i ,= j, b
i
(x) 0
for all i and c(x) 0):
nd u H
1
0
() such that
_

_
u
x
v
x
+
u
y
v
y
_
dxdy =
_

fv dxdy v H
1
0
().
The nite element approximation of the problem is:
nd u
h
V
h
such that
_

_
u
h
x
v
h
x
+
u
h
y
v
h
y
_
dxdy =
_

fv
h
dxdy v
h
V
h
.
Writing
u
h
(x, y) =
N(h)

i=1
U
i

i
(x, y),
the nite element method can be restated as follows:
nd U = (U
1
, . . . , U
N(h)
)
T
R
N(h)
such that
2.1. PIECEWISE LINEAR BASIS FUNCTIONS 29
N(h)

i=1
U
i
__

i
x

j
x
+

i
y

j
y
_
dxdy
_
=
_

f
j
dxdy,
for j = 1, . . . , N(h).
Letting A = (a
ij
), F = (F
1
, . . . , F
N(h)
)
T
,
a
ij
= a
ji
=
_

i
x

j
x
+

i
y

j
y
_
dxdy,
F
j
=
_

f
j
dxdy,
the nite element approximation can be written as a system of linear equations
AU = F.
Solving this, we obtain U = (U
1
, . . . , U
N(h)
)
T
, and hence the approximate solution
u
h
(x, y) =
N(h)

i=1
U
i

i
(x, y).
The matrix A is called the stiness matrix.
To simplify matters let us suppose that = (0, 1) (0, 1) and consider the
triangulation of

shown in Fig. 2.5. The case of a general triangulation will be
considered later. Let
ij
denote the basis function associated with the interior node
(x
i
, y
j
):

ij
(x, y) =
_

_
1
xx
i
h

yy
j
h
, (x, y) 1
1
yy
j
h
, (x, y) 2
1
x
i
x
h
, (x, y) 3
1
x
i
x
h

y
j
y
h
, (x, y) 4
1
y
j
y
h
, (x, y) 5
1
xx
i
h
, (x, y) 6
0 otherwise,
where 1, 2, . . . , 6 denote the triangles surrounding the node (x
i
, y
j
) (see Fig. 2.6.)
Thus,

ij
x
=
_

_
1/h, (x, y) 1
0, (x, y) 2
1/h, (x, y) 3
1/h, (x, y) 4
0, (x, y) 5
1/h, (x, y) 6
0, otherwise,
30 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
6
- @
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
si si si
si si si
si si si
y
N
= 1
x
N
= 1
y
x
x
1
. . . x
0
= 0
y
1
.
.
.
Figure 2.5: Triangulation of

= [0, 1] [0, 1].
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
si
1
2
3
4
5
6
Figure 2.6: Triangles surrounding a node.
2.2. THE SELF-ADJOINT ELLIPTIC PROBLEM 31
and

ij
y
=
_

_
1/h, (x, y) 1
1/h, (x, y) 2
0, (x, y) 3
1/h, (x, y) 4
1/h, (x, y) 5
0, (x, y) 6
0, otherwise.
Since
N1

i=1
N1

j=1
U
ij
_

ij
x

kl
x
+

ij
y

kl
y
_
dxdy
= 4U
kl
U
k1,l
U
k+1,l
U
k,l1
U
k,l+1
, k, l = 1, ..., N 1,
the nite element approximation is equivalent to

U
k+1,l
2U
k,l
+ U
k1,l
h
2

U
k,l+1
2U
k,l
+ U
k,l1
h
2
=
1
h
2
_ _
supp
kl
f(x, y)
kl
(x, y) dxdy, k, l = 1, . . . , N 1;
U
kl
= 0 on .
Thus, on this special triangulation of , the nite element approximation gives
rise to the familiar 5-point nite dierence scheme with the forcing function f aver-
aged in a special way.
2.2 The self-adjoint elliptic problem
Let us consider, as in Chapter 1, the elliptic boundary value problem

i,j=1

x
j
_
a
ij
(x)
u
x
i
_
+
n

i=1
b
i
(x)
u
x
i
+ c(x)u = f(x), x , (2.3)
u = 0 on , (2.4)
where is a bounded open set in R
n
, a
ij
L

(), i, j = 1, . . . , n; b
i
W
1

(),
i = 1, . . . , n, c L

(), f L
2
(), and assume that there exists a positive constant
c such that
n

i,j=1
a
ij
(x)
i

j
c
n

i=1

2
i
= (
1
, . . . ,
n
) R
n
, x

. (2.5)
We recall from Chapter 1 that the weak formulation of (2.3), (2.4) is:
nd u H
1
0
() such that a(u, v) = l(v) v H
1
0
(), (2.6)
32 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
where the bilinear functional a(, ) and the linear functional l() are dened by
a(u, v) =
n

i,j=1
_

a
ij
u
x
i
v
x
j
dx +
n

i=1
_

b
i
(x)
u
x
i
v dx +
_

c(x)uv dx,
and
l(v) =
_

f(x)v(x) dx.
We have shown that if
c(x)
1
2
n

i=1
b
i
x
i
0, x

,
then (2.6) has a unique solution u in H
1
0
(), the weak solution of (2.3), (2.4). In
the special case when the boundary value problem is self-adjoint, i.e.
a
ij
(x) = a
ji
(x), i, j = 1, . . . , n, x

,
and
b
i
(x) 0, i = 1, . . . , n, x

,
the bilinear functional a(, ) is symmetric in the sense that
a(v, w) = a(w, v) v, w H
1
0
();
in the rest of this section this will always be assumed to be the case. Thus we
consider

i,j=1

x
j
_
a
ij
(x)
u
x
i
_
+ c(x)u = f(x), x ,
(2.7)
u = 0, on
with a
ij
(x) satisfying the ellipticity condition (2.5); a
ij
(x) = a
ji
(x), c(x) 0, x

.
It turns out that (2.7) can be restated as a minimisation problem. To be more
precise, we dene the quadratic functional J : H
1
0
() R by
J(v) =
1
2
a(v, v) l(v), v H
1
0
().
Lemma 3 Let u be the (unique) weak solution to (2.6) in H
1
0
() and suppose that
a(, ) is a symmetric bilinear functional on H
1
0
(); then u is the unique minimiser
of J() over H
1
0
().
2.2. THE SELF-ADJOINT ELLIPTIC PROBLEM 33
Proof Let u be the unique weak solution to (2.6) in H
1
0
() and, for v H
1
0
(), consider
J(v) J(u):
J(v) J(u) =
1
2
a(v, v) l(v)
1
2
a(u, u) +l(u)
=
1
2
a(v, v)
1
2
a(u, u) l(v u)
=
1
2
a(v, v)
1
2
a(u, u) a(u, v u)
=
1
2
[a(v, v) 2a(u, v) +a(u, u)]
=
1
2
[a(v, v) a(u, v) a(v, u) +a(u, u)]
=
1
2
a(v u, v u).
Thence
J(v) J(u) =
1
2
a(v u, v u).
Because of (1.17),
a(v u, v u) c
0
|v u|
2
H
1
()
,
where c
0
is a positive constant. Thus
J(v) J(u)
c
0
2
|v u|
2
H
1
()
v H
1
0
(), (2.8)
and therefore,
J(v) J(u) v H
1
0
(), (2.9)
i.e. u minimises J() over H
1
0
().
In fact, u is the unique minimiser of J() in H
1
0
(). Indeed, if u also minimises J()
on H
1
0
(), then
J(v) J( u) v H
1
0
(). (2.10)
Taking v = u in (2.9) and v = u in (2.10), we deduce that
J(u) = J( u);
but then, by virtue of (2.8),
| u u|
H
1
()
= 0,
and hence u = u.
It is easily shown that J() is convex (down), i.e.
J((1 )v + w) (1 )J(v) + J(w) [0, 1], v, w H
1
0
().
34 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
u
J(u)
J(v)
H
1
0
()
R
Figure 2.7: The quadratic functional J().
This follows from the identity
(1 )J(v) + J(w) = J((1 )v + w) +
1
2
(1 )a(v w, v w)
and the fact that a(v w, v w) 0. Moreover, if u minimises J() then J() has
a stationary point at u; namely,
J

(u)v := lim
0
J(u + v) J(u)

= 0
for all v H
1
0
(). Since
J(u + v) J(u)

= a(u, v) l(v) +

2
a(v, v),
we deduce that if u minimises J() then
lim
0
[a(u, v) l(v) +

2
a(v, v)] = a(u, v) l(v) = 0 v H
1
0
(),
which proves the following result.
Lemma 4 Suppose that u H
1
0
() minimises J() over H
1
0
(); then u is the
(unique) solution of problem (2.6). The problem (2.6) is called the EulerLagrange
equation for this minimisation problem.
This lemma is precisely the converse of the previous lemma, and the two results
together express the equivalence of the weak formulation:
2.3. CALCULATION AND ASSEMBLY OF STIFFNESS MATRIX 35
nd u H
1
0
() such that a(u, v) = l(v) v H
1
0
() (W)
of the self-adjoint elliptic boundary value problem (2.7) to the associated minimisa-
tion problem:
nd u H
1
0
() such that J(u) J(v) v H
1
0
(). (M)
We shall now use this equivalence to give a variational characterisation of the
nite element approximation u
h
to u in the self-adjoint case. Given that V
h
is a
certain nite-dimensional subspace of H
1
0
() which consists of continuous piecewise
polynomials of a xed degree, the nite element approximation of (W) is:.
nd u
h
V
h
such that a(u
h
, v
h
) = l(v
h
) v
h
V
h
. (W
h
)
We can repeat the argument presented above (or simply replacing H
1
0
() by V
h
throughout) to show the equivalence of (W
h
) to the following minimisation prob-
lem:
nd u
h
V
h
such that J(u
h
) J(v
h
) v
h
V
h
. (M
h
)
Thus, u
h
can be characterised as the unique minimiser of the functional
J(v
h
) =
1
2
a(v
h
, v
h
) l(v
h
)
as v
h
ranges over the nite element space V
h
. This means that the nite element
solution u
h
inherits the energy minimisation property possessed by the weak solution
u H
1
0
() in the sense that:
J(u
h
) = min
v
h
V
h
J(v
h
).
Of course, in general J(u) < J(u
h
).
2.3 Calculation and assembly of stiness matrix
Using the variational characterisation of u
h
described at the end of the previous
section we return to the construction of the nite element approximation to Poissons
equation u = f in subject to homogeneous Dirichlet boundary condition, u = 0
on , in the case of a general triangulation. Rather than restricting ourselves to
the special case when is a square, we now suppose that is a bounded polygonal
domain in the plane, subdivided into M triangles K, so that any pair of (closed)
triangles intersect only along a complete edge, at a vertex or not at all. We consider
36 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
the set of all continuous piecewise linear functions v
h
dened on such a triangulation
with the property that v
h
= 0 of ; the linear space consisting of all such functions
v
h
is denoted V
h
. Thus, u
h
is characterised as the unique minimiser of the functional
J(v
h
) =
1
2
_

[v
h
(x, y)[
2
dxdy
_

f(x, y)v
h
(x, y) dxdy
as v
h
ranges over V
h
. Equivalently, writing
v
h
(x, y) =
N

i=1
V
i

i
(x, y),
where V
i
is the value of v
h
(x, y) at the node (x
i
, y
i
),
i
is the continuous piecewise
linear basis function associated with this node, and N is the number of nodes internal
to , we can write this minimisation problem in matrix form as follows:
nd V R
N
such that
1
2
V
T
AV V
T
F is minimum, (2.11)
where V = (V
1
, . . . , V
N
)
T
, A is the (global) stiness matrix - an N N matrix
with (i, j) entry
a(
i
,
j
) = (
i
,
j
) =
_

i
(x, y)
j
(x, y) dxdy,
and F = (F
1
, . . . , F
N
)
T
is the (global) load vector, with
F
i
= (f,
i
) =
_

f(x, y)
i
(x, y) dxdy.
Consider any triangle K in the triangulation of , and introduce the position
vectors r
i
= (x
i
, y
i
), i = 1, 2, 3, of its three vertices labelled in the anti-clockwise
direction, say. In addition, we consider a so-called local (, ) coordinate system
and the canonical triangle depicted in Figure 2.8. The coordinate r = (x, y) of any
point in the triangle K can be written as a convex combination of the coordinates
of the three vertices:
r = (1 )r
1
+ r
2
+ r
3
(2.12)
r
1

1
(, ) +r
2

2
(, ) +r
3

3
(, ).
The set
1
,
2
,
3
is called the nodal basis (or local basis) for the set of
linear polynomials in terms of the local coordinates. Consider the transformation
(, ) r = (x, y) dened by (2.12) from the canonical triangle to the global (x, y)
coordinate system. The Jacobi matrix J of this transformation is given by
J =
(x, y)
(, )
=
_
x
2
x
1
y
2
y
1
x
3
x
1
y
3
y
1
_
2.3. CALCULATION AND ASSEMBLY OF STIFFNESS MATRIX 37
-
@
@
@
@
@
@
@
6
s
s
s

(0, 1)
(1, 0) (0, 0)
1 2
3
Figure 2.8: Canonical triangle and local coordinates.
from which it follows that the Jacobian is
[J[ = det
_
x
2
x
1
y
2
y
1
x
3
x
1
y
3
y
1
_
= det
_
_
x
1
y
1
1
x
2
y
2
1
x
3
y
3
1
_
_
, (2.13)
namely,
[J[ = 2A
123
where A
123
is the area of the triangle K = (r
1
, r
2
, r
3
). Similarly, for any function
v
h
V
h
,
v
h
(x, y) = v
h
(r(, )) = V
1

1
(, ) + V
2

2
(, ) + V
3

3
(, ), (2.14)
where V
i
is the value of v
h
at the node of the triangle K with position vector r
i
,
i = 1, 2, 3. In order to determine the entries of the stiness matrix, we need the
gradient of v
h
in the global coordinate system; however, from (2.12) and the form
of the Jacobi matrix J we have that
_
_
v
h

v
h

_
_
= J
_
_
v
h
x
v
h
y
_
_
,
_
_
v
h
x
v
h
y
_
_
= J
1
_
_
v
h

v
h

_
_
. (2.15)
Consequently,
v
h
x
=
1
[J[
_
(y
3
y
1
)
v
h

(y
2
y
1
)
v
h

_
(2.16)
v
h
y
=
1
[J[
_
(x
3
x
1
)
v
h

+ (x
2
x
1
)
v
h

_
.
Hence
[J[
2
[v
h
[
2
= [r
3
r
1
[
2
_
v
h

_
2
+[r
2
r
1
[
2
_
v
h

_
2
2(r
3
r
1
) (r
2
r
1
)
v
h

v
h

(2.17)
38 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
and from (2.14) and (2.12) it follows that
v
h

= V
2
V
1
,
v
h

= V
3
V
1
. (2.18)
As v
h
(x, y) is linear on each triangle K in the triangulation, v
h
is constant on K
so the contribution to
_

[v
h
(x, y)[
2
dxdy =

K
_
K
[v
h
(x, y)[
2
dxdy
from triangle K is
_
K
[v
h
(x, y)[
2
dxdy = A
123
[v
h
[
2
=
1
2
[J[[v
h
[
2
=
1
4A
123
[J[
2
[v
h
[
2
.
Substitution of (2.17) and (2.18) into this formula yields a quadratic form in the
nodal values V
1
, V
2
, V
3
; after a little algebra, we nd that the coecient of V
2
1
is
[r
3
r
1
[
2
+[r
2
r
1
[
2
2(r
3
r
1
) (r
2
r
1
) = [r
3
r
2
[
2
and the coecient of V
1
V
2
is
2[r
3
r
1
[
2
+ 2(r
3
r
1
) (r
2
r
1
) = 2(r
2
r
3
) (r
3
r
1
)
with similar expressions for the coecients of V
2
2
, V
2
3
and V
2
V
3
, V
3
V
1
, obtained by
cyclic permutations of the indices in these expressions, respectively. Thus we deduce
that
_
K
[v
h
(x, y)[
2
dxdy = [V
1
, V
2
, V
3
]A
k
_
_
V
1
V
2
V
3
_
_
,
where k 1, . . . , M is the number of the triangle K in the global numbering and
A
k
is the symmetric 3 3 element stiness matrix:
A
k
=
1
4A
123
_
_
[r
2
r
3
[
2
(r
2
r
3
) (r
3
r
1
) (r
2
r
3
) (r
1
r
2
)
[r
3
r
1
[
2
(r
3
r
1
) (r
1
r
2
)
symm. [r
1
r
2
[
2
_
_
.
Assembly of the global stiness matrix entails relating the local numbering of
the nodes to the global numbering system. Let us denote by N the number of nodes
internal to ; as
u
h
(x, y) =
N

i=1
U
i

i
(x, y),
N is precisely the number of unknowns: U
1
, . . . , U
N
. Let us label by N + 1, N +
2, . . . N

the nodes that lie on the boundary of (thus N

is the total number of


nodes of which N are internal and N

N are on the boundary). As u


h
= 0 on the
2.3. CALCULATION AND ASSEMBLY OF STIFFNESS MATRIX 39
boundary, we can adopt the notational convention that U
N+1
= U
N+2
= . . . U
N
= 0,
and write
u
h
(x, y) =
N

i=1
U
i

i
(x, y),
with the understanding that the coecients U
j
, j = N + 1, . . . , N

are, in fact,
known (to be zero) from the boundary condition.
For the kth triangle K, we consider the Boolean matrix
1
L
k
of size N

3 whose
entries are dened as follows: if in calculating the matrix A
k
the node with position
vector r
1
is the ith node in the global numbering, i 1, . . . , N, . . . , N

, then the
rst column of L
k
has unit entry in the ith row; similarly, the second and third
column depend on the global numbering of the nodes with position vectors r
2
and
r
3
appearing in the matrix A
k
. Then, the so called full stiness matrix A

is an
N

matrix dened as a sum over the elements K in the triangulation of the


domain:
A

=
M

k=1
L
k
A
k
(L
k
)
T
,
where (L
k
)
T
is the transpose of the matrix L
k
.
When programming this, instead of working with M Boolean arrays L
k
, k =
1, . . . , M, it is more economical to store the information contained in the arrays L
k
in a single connectivity array LNODS which has dimension M3, where M is the
number of triangles in the triangulation of ; LNODS(k, j) 1, . . . , N

is equal
to the global number of the node r
j
in the kth triangle. By letting k = 1, . . . , M, we
loop through all the triangles in the triangulation of , and calculate A
k
ij
for i, j =
1, 2, 3 from the formula for A
k
given above; once the value A
k
ij
has been calculated it
is added into the full stiness matrix A

at position (LNODS(k, i), LNODS(k, j)).


The full load vector F

= (F
1
, . . . , F
N
, . . . F
N
)
T
is built up in the same way.
Once A

and F

have been found, we erase the last N

N rows and columns


of A

to obtain the global stiness matrix A, and the last N

N entries of F

to obtain to global load vector F, and then solve the linear system
AU = F
to determine the vector of unknowns U = (U
1
, . . . , U
N
)
T
.
In order to justify more clearly the compression of A

to A and F

to F, let us
note that the minimisation problem (2.11) can be restated in the following equivalent
form:
nd V

= (V
1
, . . . , V
N
, 0, . . . , 0)
T
R
N

such that
1
2
V
T
A

V
T
F

is minimum.
(2.19)
Since the last N

N entries of V

are equal to 0, the last N

N rows and columns


of A

and the last N

N entries of F

can be discarder since they are all multiplied


1
i.e. a matrix whose entries are 0s and 1s
40 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
by entries of V

that are equal to zero. Even though it may seem that we are doing
unnecessary work when computing entries of A

and F

which are then thrown away


when A

is compressed to A and F

is compressed to F, the assembly of A

and F

,
followed by compression, is typically a faster process than the direct assembly of A
and F, since in the latter case special care has to be taken for nodes which belong
to triangles with at least one boundary point, leading to a slower assembly process.
No such diculties arise when we work with A

and F

.
It is worth noting that in practice it is not essential that the rst N indices in
the set 1, . . . , N

correspond to the interior nodes and the last N

N to the
boundary nodes: indeed, the nodes may be numbered in any order; the only thing
that matters is that rows and columns of A

and entries of F

corresponding to
boundary nodes are discarded when A and F are formed. Here we have chosen the
last N

N nodes of a total of N

to be those on the boundary simply for ease of


presentation.
2.4 Galerkin orthogonality; Ceas lemma
Having described the construction of the nite element method, we now outline the
basic tools for its error analysis. Let us consider the elliptic boundary value problem

i,j=1

x
j
_
a
ij
(x)
u
x
i
_
+
n

i=1
b
i
(x)
u
x
i
+ c(x)u = f(x), x , (2.20)
u = 0 on , (2.21)
where is a bounded open set in R
n
, a
ij
L

(), i, j = 1, . . . , n; b
i
W
1

(),
i = 1, . . . , n, c L

(), f L
2
(), and assume that there exists a positive constant
c such that
n

i,j=1
a
ij
(x)
i

j
c
n

i=1

2
i
= (
1
, . . . ,
n
) R
n
, x

. (2.22)
The weak formulation of (2.20), (2.21) is:
nd u H
1
0
() such that a(u, v) = l(v) v H
1
0
(), (2.23)
where the bilinear functional a(, ) and the linear functional l() are dened by
a(u, v) =
n

i,j=1
_

a
ij
u
x
i
v
x
j
dx +
n

i=1
_

b
i
(x)
u
x
i
v dx +
_

c(x)uv dx,
and
l(v) =
_

f(x)v(x) dx.
2.4. GALERKIN ORTHOGONALITY; C

EAS LEMMA 41
We have shown that if
c(x)
1
2
n

i=1
b
i
x
i
0, x

,
then (2.23) has a unique solution u in H
1
0
(), the weak solution of (2.20), (2.21).
Moreover,
|u|
H
1
()

1
c
0
|f|
L
2
()
,
where c
0
is as in (1.17).
Now suppose that V
h
is a nite-dimensional subspace of H
1
0
(), without making
further precise assumptions on the nature of V
h
(although we shall implicitly assume
that V
h
consists of continuous piecewise polynomials dened on a subdivision of
neness h of the computational domain ). The nite element approximation of
(2.23) is:
nd u
h
in V
h
such that a(u
h
, v
h
) = l(v
h
) for all v
h
V
h
. (2.24)
As, by hypothesis, V
h
is contained in H
1
0
() it follows from the Lax-Milgram theorem
that (2.24) has a unique solution u
h
in V
h
. Moreover, (2.23) holds for any v = v
h

V
h
; namely,
a(u, v
h
) = l(v
h
) for all v
h
V
h
.
Subtracting (2.24) from this identity we deduce that
a(u u
h
, v
h
) = 0 for all v
h
V
h
. (2.25)
The property (2.25) is referred to as Galerkin orthogonality and will be seen to
play a crucial role in the error analysis of nite element methods. Since by (1.17),
with v = u u
h
H
1
0
() we have that
|u u
h
|
2
H
1
()

1
c
0
a(u u
h
, u u
h
),
it follows from (2.25) that
|u u
h
|
2
H
1
()

1
c
0
a(u u
h
, u v
h
);
further, by (1.13),
a(u u
h
, u v
h
) c
1
|u u
h
|
H
1
()
|u v
h
|
H
1
()
.
Combining the last two inequalities, we deduce that
|u u
h
|
H
1
()

c
1
c
0
|u v
h
|
H
1
()
for all v
h
V
h
.
Thus we have proved the following result.
42 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
Lemma 5 (Ceas lemma) The nite element approximation u
h
to u H
1
0
(), the
weak solution to the problem (2.20), (2.21), is the near-best t to u in the norm
| |
H
1
()
; i.e.,
|u u
h
|
H
1
()

c
1
c
0
min
v
h
V
h
|u v
h
|
H
1
()
.
Remark 5 We shall prove in the next chapter that, for a typical nite element space
V
h
,
min
v
h
V
h
|u v
h
|
H
1
()
C(u)h
s
where C(u) is a positive constant, dependent on the smoothness of u, h is the mesh-
size parameter (the maximum diameter of elements in the subdivision of the com-
putational domain) and s is a positive real number, dependent on the smoothness of
u and the degree of piecewise polynomials comprising the space V
h
. Hence, with the
aid of Ceas lemma we shall be able to deduce that
|u u
h
|
H
1
()
C(u)
_
c
1
c
0
_
h
s
(2.26)
which is a bound of the global error e
h
= uu
h
in terms of the mesh-size parameter
h. Such a bound on the global error is called an a priori error bound (the terminol-
ogy stems from the fact that (2.26) can be stated prior to computing u
h
). It shows,
in particular, that as h 0 when rening the subdivision further and further, the
sequence of nite element solutions u
h

h
converges to u in the H
1
() norm. While
this result is reassuring from the theoretical point of view, it is of little practical rel-
evance as the constant C(u) involved in (2.26) is dicult to quantify (given that it
depends on the unknown analytical solution u). Later on we shall discuss a poste-
riori error bounds which make explicit use of the computed solution u
h
and provide
computable bounds on the global error.
Example 6 In this example we highlight a further point concerning the a priori
error bound (2.26): for certain elliptic problems the ratio c
1
/c
0
can be very large,
and then the mesh-size h has to be taken extremely small before any reduction in
the size of the global error is observed. Suppose that is a bounded open set in R
n
.
Consider the following boundary value problem:
u +b u = f in ,
u = 0 on ,
where > 0, b = (b
1
, . . . , b
n
)
T
, with b
i
W
1

() for i = 1, . . . , n. For the sake


of simplicity, we shall suppose that div b 0 almost everywhere on . Such prob-
lems arise in the mathematical modelling of advection-diusion phenomena. When
advection dominates diusion the so-called Peclet number
Pe =
_

n
i=1
|b
i
|
2
L

()
_
1/2

2.4. GALERKIN ORTHOGONALITY; C

EAS LEMMA 43
is very large (say, of the order 10
6
to 10
8
).
A simple calculation shows that for the present problem
c
1
=
_

2
+
n

i=1
|b
i
|
2
L

()
_
1/2
and
c
0
=

(1 + c
2

)
1/2
.
Therefore
c
1
c
0
= (1 + c
2

)
1/2
(1 + Pe
2
)
1/2
,
and (2.26) gives
|u u
h
|
H
1
()
(1 + c
2

)
1/2
(1 + Pe
2
)
1/2
C(u)h
s
. (2.27)
Thus, when << 1, the constant on the right-hand side in this error bound is made
very large through the presence of the Peclet number; in fact, things are even worse:
the constant C(u) also depends on through u (typically C(u) >> 1 when << 1).
We shall not consider the nite element approximation of advection-dominated
diusion problems any further. The point that we wish to make is merely that care
should be taken when attempting to draw practically relevant conclusions from theo-
retical results of the kind (2.26). As it happens, the poor quality of the a priori error
bound (2.27) when Pe >> 1 is merely a reection of the fact that for advection-
dominated diusion equations conventional nite element methods are genuinely
badly behaved: on coarse meshes the numerical solution exhibits large unphysical
oscillations which can only be eliminated by severely reducing the mesh-size h.
In order to put this example into perspective, we now discuss the other extreme
case, when b 0 on : then c
1
= c
0
= , so Ceas lemma implies that
|u u
h
|
H
1
()
min
v
h
V
h
|u v
h
|
H
1
()
.
In fact, since the left-hand side of this inequality cannot be strictly less than the
right-hand side (this can be seen by choosing v
h
= u
h
on the right), it follows that
|u u
h
|
H
1
()
= min
v
h
V
h
|u v
h
|
H
1
()
,
so that u
h
is the best approximation to u from V
h
in the H
1
() norm. We shall
show that a result of this kind holds in a slightly more general setting, when the
problem is self-adjoint, namely a
ij
(x) a
ji
(x) for all i, j = 1, . . . , n, b
i
(x) 0 for
i = 1, . . . , n. Let us dene
(v, w)
a
:= a(v, w), v, w H
1
0
().
44 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
-
6

>
u
h
0
H
1
0
()
u u
h
u
V
h
s
s
Figure 2.9: The error u u
h
is orthogonal to V
h
.
Because a(, ) is a symmetric bilinear functional on H
1
0
() H
1
0
() and
a(v, v) c
0
|v|
2
H
1
()
v H
1
0
(),
it is easily seen that (, )
a
satises all axioms of an inner product. Let | |
a
denote
the associated energy norm dened by:
|v|
a
:= [a(v, v)]
1/2
.
Since V
h
H
1
0
(), taking v = v
h
V
h
in the statement of (W), we deduce that
a(u, v
h
) = l(v
h
), v
h
V
h
; (2.28)
also by, (W
h
),
a(u
h
, v
h
) = l(v
h
), v
h
V
h
. (2.29)
Subtracting (2.29) from (2.28) and using the fact that a(, ) is a bilinear functional,
we deduce the Galerkin orthogonality property
a(u u
h
, v
h
) = 0 v
h
V
h
,
i.e.
(u u
h
, v
h
)
a
= 0 v
h
V
h
. (2.30)
Thus, in the self-adjoint case, the error u u
h
between the exact solution u and its
nite element approximation u
h
is orthogonal to V
h
in the inner product (, )
a
(see
Figure 2.9). By virtue of the orthogonality property (2.30),
|u u
h
|
2
a
= (u u
h
, u u
h
)
a
= (u u
h
, u)
a
(u u
h
, u
h
)
a
= (u u
h
, u)
a
= (u u
h
, u)
a
(u u
h
, v
h
)
a
= (u u
h
, u v
h
)
a
v
h
V
h
.
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 45
Thence, by the Cauchy-Schwarz inequality,
|u u
h
|
2
a
= (u u
h
, u v
h
)
a
|u u
h
|
a
|u v
h
|
a
v
h
V
h
;
therefore
|u u
h
|
a
|u v
h
|
a
v
h
V
h
.
Consequently,
|u u
h
|
a
= min
v
h
V
h
|u v
h
|
a
.
Thus we have proved the following renement of Ceas lemma in the self-adjoint
case.
Lemma 6 The nite element approximation u
h
V
h
of u H
1
0
() is the best t to
u from V
h
in the energy norm | |
a
, i.e.
|u u
h
|
a
= min
v
h
V
h
|u v
h
|
a
.
Ceas lemma is the key to the error analysis of the nite element method for
elliptic boundary value problems. In the next section we describe how such an
analysis proceeds in the self-adjoint case, for a particularly simple nite element
space V
h
consisting of continuous piecewise linear functions on . The general case
is very similar and will be considered later on in the notes.
2.5 Optimal error bound in the energy norm
In this section, we shall employ Ceas lemma to derive an optimal error bound for
the nite element approximation (W
h
) of problem (W) in the case of piecewise linear
basis functions. We shall consider two examples: a one-dimensional model problem
a two-point boundary value problem, and a two-dimensional model problem
Poissons equation subject to homogeneous Dirichlet boundary condition.
One-dimensional problem
Consider, for f L
2
(0, 1), the boundary value problem
u

+ u = f(x), 0 < x < 1,


u(0) = 0, u(1) = 0.
Its weak formulation is:
nd u in H
1
0
(0, 1) such that a(u, v) = l(v) v H
1
0
(0, 1),
46 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
where
a(u, v) =
_
1
0
(u

(x)v

(x) + u(x)v(x)) dx
and
l(v) =
_
1
0
f(x)v(x) dx.
The symmetric bilinear functional a(, ) induces the energy norm | |
a
dened by
|w|
a
= (a(w, w))
1/2
=
__
1
0
_
[w

(x)[
2
+[w(x)[
2
_
dx
_
1/2
= |w|
H
1
(0,1)
.
The nite element approximation of this problem, using piecewise linear basis
functions, has been described in Section 2.1 (take p(x) 1 and q(x) 1 there
to obtain the present problem). Here, instead of restricting ourselves to uniform
subdivisions of [0, 1], we consider a general nonuniform subdivision:
0 = x
0
< x
1
< . . . < x
N1
< x
N
= 1,
where the mesh-points x
i
, i = 0, . . . , N, are not necessarily equally spaced. It will be
supposed that N 2 so that we have at least one mesh-point inside (0, 1). We put
h
i
= x
i
x
i1
and dene the mesh parameter h = max
i
h
i
. For such a subdivision,
we consider the nite element basis function

i
(x) =
_

_
0 if x x
i1
(x x
i1
)/h
i
if x
i1
x x
i
(x
i+1
x)/h
i+1
if x
i
x x
i+1
0 if x
i+1
x,
for i = 1, . . . , N 1. We put
V
h
= span
1
, . . . ,
N1
.
Clearly V
h
is an (N 1)-dimensional subspace of H
1
0
(0, 1). We approximate the
boundary value problem by the nite element method
nd u
h
in V
h
such that a(u
h
, v
h
) = l(v
h
) v
h
V
h
.
Now since the bilinear functional a(, ) is symmetric it follows from Ceas lemma
that
|u u
h
|
H
1
(0,1)
= |u u
h
|
a
= min
v
h
V
h
|u v
h
|
a
= min
v
h
V
h
|u v
h
|
H
1
(0,1)
. (2.31)
Let J
h
u V
h
denote the continuous piecewise linear function on the subdivision
x
0
, x
1
, . . . , x
N
which coincides with u at the mesh-points x
i
, i = 0, . . . , N. Thus,
J
h
u(x) =
N1

i=1
u(x
i
)
i
(x).
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 47
The function J
h
u is called the interpolant of u from the nite element space V
h
.
Choosing v
h
= J
h
u in (2.31), we see that
|u u
h
|
H
1
(0,1)
|u J
h
u|
H
1
(0,1)
. (2.32)
Thus, to derive a bound on the global error u u
h
in the H
1
(0, 1) norm, we shall
now seek a bound on the interpolation error u J
h
u in the same norm. The rest of
this subsection is devoted to the proof of the following estimate:
|u J
h
u|
H
1
(0,1)

h

_
1 +
h
2

2
_
1/2
|u

|
L
2
(0,1)
. (2.33)
Theorem 3 Suppose that u H
2
(0, 1) and let J
h
u be the interpolant of u from the
nite element space V
h
dened above; then the following error bounds hold:
|u J
h
u|
L
2
(0,1)

_
h

_
2
|u

|
L
2
(0,1)
,
|u

(J
h
u)

|
L
2
(0,1)

h

|u

|
L
2
(0,1)
.
Proof Consider a subinterval [x
i1
, x
i
], 1 i N, and dene (x) = u(x) J
h
u(x)
for x [x
i1
, x
i
]. Then H
2
(x
i1
, x
i
) and (x
i1
) = (x
i
) = 0. Therefore can be
expanded into a convergent Fourier sine-series,
(x) =

k=1
a
k
sin
k(x x
i1
)
h
i
, x [x
i1
, x
i
].
Hence,
_
x
i
x
i1
[(x)]
2
dx =
h
i
2

k=1
[a
k
[
2
.
Dierentiating the Fourier sine-series for twice, we deduce that the Fourier coecients
of

are (k/h
i
)a
k
, while those of

are (k/h
i
)
2
a
k
. Thus,
_
x
i
x
i1
[

(x)]
2
dx =
h
i
2

k=1
_
k
h
i
_
2
[a
k
[
2
,
_
x
i
x
i1
[

(x)]
2
dx =
h
i
2

k=1
_
k
h
i
_
4
[a
k
[
2
.
Because k
4
k
2
1, it follows that
_
x
i
x
i1
[(x)]
2
dx
_
h
i

_
4
_
x
i
x
i1
[

(x)]
2
dx,
_
x
i
x
i1
[

(x)]
2
dx
_
h
i

_
2
_
x
i
x
i1
[

(x)]
2
dx.
48 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
However

(x) = u

(x) (J
h
u)

(x) = u

(x) for x (x
i1
, x
i
) because J
h
u is a linear
function on this interval. Therefore, upon summation over i = 1, . . . , N and letting h =
max
i
h
i
, we obtain
||
2
L
2
(0,1)

_
h

_
4
|u

|
2
L
2
(0,1)
,
|

|
2
L
2
(0,1)

_
h

_
2
|u

|
2
L
2
(0,1)
.
After taking the square root and recalling that = u (J
h
u) these yield the desired
bounds on the interpolation error.
Now (2.33) follows directly from this theorem by noting that
|u J
h
u|
2
H
1
(0,1)
= |u J
h
u|
2
L
2
(0,1)
+|(u J
h
u)

|
2
L
2
(0,1)

h
2

2
_
1 +
h
2

2
_
|u

|
2
L
2
(0,1)
.
Having established the bound (2.33) on the interpolation error, we arrive at the
following a priori error bound by inserting (2.33) into the inequality (2.32):
|u u
h
|
H
1
(0,1)

h

_
1 +
h
2

2
_
1/2
|u

|
L
2
(0,1)
. (2.34)
This shows that, provided u

L
2
(0, 1), the error in the nite element solution,
measured in the H
1
(0, 1) norm, converges to 0 at the rate O(h) as h 0.
As a nal note concerning this example, we remark that our hypothesis on f
(namely that f L
2
(0, 1)) implies that u

L
2
(0, 1). Indeed, choosing v = u in the
weak formulation of the boundary value problem gives
_
1
0
[u

(x)[
2
dx +
_
1
0
[u(x)[
2
dx =
_
1
0
f(x)u(x) dx

__
1
0
[f(x)[
2
dx
_
1/2
__
1
0
[u(x)[
2
dx
_
1/2
. (2.35)
Hence,
__
1
0
[u(x)[
2
dx
_
1/2

__
1
0
[f(x)[
2
dx
_
1/2
,
i.e.
|u|
L
2
(0,1)
|f|
L
2
(0,1)
.
Thereby, from (2.35),
|u

|
L
2
(0,1)
|f|
L
2
(0,1)
.
Finally, as u

= u f from the dierential equation, we have that


|u

|
L
2
(0,1)
= |u f|
L
2
(0,1)
|u|
L
2
(0,1)
+|f|
L
2
(0,1)
2|f|
L
2
(0,1)
.
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 49
Thus we have proved that u

L
2
(0, 1) (in fact, as we also know that u and u

belong to L
2
(0, 1), we have proved more: u H
2
(0, 1)). Substituting this bound on
|u

|
L
2
(0,1)
into (2.34) gives
|u u
h
|
H
1
(0,1)

2h

_
1 +
h
2

2
_
1/2
|f|
L
2
(0,1)
.
This now provides a computable upper bound on the global error u u
h
in the
H
1
(0, 1) norm, since f is a given function and h = max
i
h
i
can be easily calculated
for any given subdivision of [0, 1].
The argument presented in this example is representative of a general nite
element (a priori) error analysis. In a nutshell, it consisted of using:
a) Ceas lemma, together with
b) an interpolation error bound.
These two ingredients then lead us to the error bound (2.34). Finally, if we are
fortunate enough to have a bound of the type |u

|
L
2
(0,1)
C

|f|
L
2
(0,1)
(or, in
other words, [u[
H
2
(0,1)
C

|f|
L
2
(0,1)
), which is called
c) an elliptic regularity estimate,
then, at least in principle, we obtain a computable bound on the global error. Un-
fortunately, for (multi-dimensional) elliptic boundary value problems proving an
elliptic regularity estimate of the form
[u[
H
2
()
C

|f|
L
2
()
(2.36)
is a highly non-trivial task (this issue will be touched on in the next section, in dis-
cussion about the Aubin-Nitsche duality argument). In fact, for multi-dimensional
problems (2.36) will not hold unless the boundary and the coecients a
ij
, b
i
and
c are suciently smooth. Worse still, even when (2.36) holds precise estimates of the
size of the constant C

are only available in rare circumstances. The upshot is that


an a priori error bound will usually not provide a computable estimate of the global
error. This is a serious drawback from the point of view of practical computations
where one would like to have precise information about the size of the error between
the analytical solution and its nite element approximation. Later on in the notes
we shall discuss an alternative approach, a posteriori error analysis, which resolves
this diculty and provides computable bounds on the error in terms of u
h
.
Two-dimensional problem
Let = (0, 1) (0, 1), and consider the elliptic boundary value problem
u = f in , (2.37)
u = 0 on . (2.38)
50 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
6
- @
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
si si si
si si si
si si si
y
N
= 1
x
N
= 1
y
x
x
1
. . .
y
0
= 0
x
0
= 0
y
1
.
.
.
Figure 2.10: Triangulation of

= [0, 1] [0, 1].
We recall that the weak formulation of this problem is:
nd u H
1
0
() such that
_

_
u
x
v
x
+
u
y
v
y
_
dxdy =
_

fv dxdy v H
1
0
(). (2.39)
In order to construct the nite element approximation, we triangulate the domain
as shown in the Fig 2.10. Let h = 1/N, and dene x
i
= ih, i = 0, . . . , N, y
j
= jh,
j = 0, . . . , N. With each node, (x
i
, y
j
), contained in the interior of (labelled in
the gure), we associate a basis function
ij
, i, j = 1, . . . , N 1, dened by

ij
(x, y) =
_

_
1
xx
i
h

yy
j
h
, (x, y) 1
1
yy
j
h
, (x, y) 2
1
x
i
x
h
, (x, y) 3
1
x
i
x
h

y
j
y
h
, (x, y) 4
1
y
j
y
h
, (x, y) 5
1
xx
i
h
, (x, y) 6
0 otherwise.
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 51
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
si
1
2
3
4
5
6
(x
i
, y
j+1
)
(x
i1
, y
j+1
)
(x
i1
, y
j
)
(x
i
, y
j1
)
(x
i+1
, y
j1
)
(x
i+1
, y
j
)
Figure 2.11: Triangles surrounding the node (x
i
, y
j
).
Let V
h
= span
ij
, i = 1, . . . , N 1; j = 1, . . . , N 1. The nite element
approximation of (2.37) (and (2.39)) is:
nd u
h
V
h
such that
_

_
u
h
x
v
h
x
+
u
h
y
v
h
y
_
dxdy =
_

fv
h
dxdy v
h
V
h
. (2.40)
Letting
l(v) =
_

f(x)v(x) dx and
(v, w)
a
= a(v, w) =
_

_
v
x
w
x
+
v
y
w
y
_
dxdy,
(2.39) and the nite element method (2.40) can be written, respectively, as follows:
nd u H
1
0
() such that a(u, v) = l(v) v H
1
0
(), (5.13

)
and
nd u
h
V
h
such that a(u
h
, v
h
) = l(v
h
) v
h
V
h
. (5.14

)
According to Ceas lemma,
|u u
h
|
a
= min
v
h
V
h
|u v
h
|
a
|u J
h
u|
a
, (2.41)
where J
h
u denotes the continuous piecewise linear interpolant of the function u on
52 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
the set

= [0, 1] [0, 1]:
(J
h
u)(x, y) =
N1

i=1
N1

j=1
u(x
i
, y
j
)
ij
(x, y).
Clearly (J
h
u)(x
k
, y
l
) = u(x
k
, y
l
). Let us estimate |u J
h
u|
a
:
|u J
h
u|
2
a
=
_

[

x
(u J
h
u)[
2
dxdy +
_

[

y
(u J
h
u)[
2
dxdy
=

__

[

x
(u J
h
u)[
2
dxdy +
_

[

y
(u J
h
u)[
2
dxdy
_
(2.42)
where is a triangle in the subdivision of . Suppose, for example, that
= (x, y) : x
i
x x
i+1
; y
j
y y
j+1
+ x
i
x.
In order to estimate
_

[

x
(u J
h
u)[
2
dxdy +
_

[

y
(u J
h
u)[
2
dxdy,
we dene the canonical triangle
K = (s, t) : 0 s 1, 0 t 1 s
and the ane mapping (x, y) (s, t) from to K, by
x = x
i
+ sh, 0 s 1,
y = y
j
+ th, 0 t 1.
Let u(s, t) := u(x, y). Then,
u
x
=
u
s

s
x
+
u
t

t
x
=
1
h

u
s
,
u
y
=
u
s

s
y
+
u
t

t
y
=
1
h

u
t
.
The Jacobian of the mapping (s, t) (x, y) is
[J[ =

(x, y)
(s, t)

x
s
x
t
y
s
y
t

= h
2
.
Thus
_

[

x
(u J
h
u)[
2
dxdy = (P.T.O.)
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 53
=
_
K
[

s
( u(s, t) [(1 s t) u(0, 0) + s u(1, 0) + t u(0, 1)]) [
2
ds dt
=
_
1
0
_
1s
0
[
u
s
(s, t) [ u(1, 0) u(0, 0)][
2
ds dt
=
_
1
0
_
1s
0
[
u
s
(s, t)
_
1
0
u
s
(, 0) d[
2
ds dt
=
_
1
0
_
1s
0
[
_
1
0
_
u
s
(s, t)
u
s
(, t)
_
d +
_
1
0
_
u
s
(, t)
u
s
(, 0)
_
d[
2
ds dt
=
_
1
0
_
1s
0
[
_
1
0
_
s

2
u
s
2
(, t) d d +
_
1
0
_
t
0

2
u
s t
(, ) d d[
2
ds dt
2
_
1
0
_
1s
0
_
1
0
_
1
0
[

2
u
s
2
(, t)[
2
d d ds dt
+2
_
1
0
_
1s
0
_
1
0
_
1
0
[

2
u
s t
(, )[
2
d d ds dt
2
_
1
0
_
1
0
[

2
u
s
2
(, t)[
2
d dt +
_
1
0
_
1
0
[

2
u
s t
(, )[
2
d d
= 2
_
x
i+1
x
i
_
y
j+1
y
j
[

2
u
x
2
(x, y)[
2
[h
2
[
2
h
2
dxdy
+
_
x
i+1
x
i
_
y
j+1
y
j
[

2
u
x y
(x, y)[
2
[h
2
[
2
h
2
dxdy.
Therefore,
_

[

x
(u J
h
u)[
2
dxdy 2h
2
_
x
i+1
x
i
_
y
j+1
y
j
_
[

2
u
x
2
[
2
+
1
2
[

2
u
x y
[
2
_
dxdy.
(2.43)
Similarly,
_

[

y
(u J
h
u)[
2
dxdy 2h
2
_
x
i+1
x
i
_
y
j+1
y
j
_
[

2
u
y
2
[
2
+
1
2
[

2
u
x y
[
2
_
dxdy.
(2.44)
Substituting (2.43) and (2.44) into (2.42),
|u J
h
u|
2
a
4h
2
_

_
[

2
u
x
2
[
2
+[

2
u
x y
[
2
+[

2
u
y
2
[
2
_
dxdy. (2.45)
Finally by (2.41) and (2.45),
|u u
h
|
a
2h[u[
H
2
()
. (2.46)
Thus we have proved the following result.
54 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
Theorem 4 Let u be the weak solution of the boundary value problem (2.37), and
let u
h
be its piecewise linear nite element approximation dened by (2.40). Suppose
that u H
2
() H
1
0
(); then
|u u
h
|
a
2h[u[
H
2
()
.
Corollary 2 Under the hypotheses of Theorem 4,
|u u
h
|
H
1
()

5h[u[
H
2
()
.
Proof According to Theorem 4,
|u u
h
|
2
a
= [u u
h
[
2
H
1
()
4h
2
[u[
2
H
2
()
.
Since u H
1
0
(), u
h
V
h
H
1
0
(), it follows that u u
h
H
1
0
(). By the Poincare-
Friedrichs inequality,
|u u
h
|
2
L
2
()

1
4
[u u
h
[
2
H
1
()
; (2.47)
thus,
|u u
h
|
2
H
1
()
= |u u
h
|
2
L
2
()
+[u u
h
[
2
H
1
()

5
4
[u u
h
[
2
H
1
()
5h
2
[u[
2
H
2
()
,
and that completes the proof.
From (2.47) and (2.46) we also see that
|u u
h
|
L
2
()
h[u[
H
2
()
. (2.48)
The Aubin-Nitsche duality argument. The error estimate (2.48) indicates
that the error in the L
2
-norm between u and its nite element approximation u
h
is
of the size O(h). It turns out, however, that this bound is quite pessimistic and can
be improved to O(h
2
); the proof of this is presented below.
Let us rst observe that if w H
2
() H
1
0
(), = (0, 1) (0, 1), then
|w|
2
L
2
()
=
_

2
w
x
2
+

2
w
y
2
_
2
dxdy
=
_

2
w
x
2
_
2
+ 2
_

2
w
x
2


2
w
y
2
dxdy +
_

2
w
y
2
_
2
dxdy.
Performing integration by parts and using the fact that w = 0 on ,
_

2
w
x
2


2
w
y
2
dxdy =
_

2
w
x y


2
w
x y
dxdy
=
_

[

2
w
x y
[
2
dxdy.
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 55
Thus
|w|
2
L
2
()
=
_

_
[

2
w
x
2
[
2
+ 2[

2
w
x y
[
2
+[

2
w
y
2
[
2
_
dxdy
= [w[
2
H
2
()
.
Given g L
2
(), let w
g
H
1
0
() be the weak solution of the boundary value
problem
w
g
= g in , (2.49)
w
g
= 0 on ; (2.50)
then w
g
H
2
() H
1
0
(), and
[w
g
[
H
2
()
= |w
g
|
L
2
()
= |g|
L
2
()
. (2.51)
After this brief preparation, we turn to the derivation of the optimal error bound in
the L
2
-norm.
According to the Cauchy-Schwarz inequality for the L
2
-inner product (, ),
(u u
h
, g) |u u
h
|
L
2
()
|g|
L
2
()
g L
2
().
Therefore,
|u u
h
|
L
2
()
= sup
gL
2
()
(u u
h
, g)
|g|
L
2
()
. (2.52)
For g L
2
(), the function w
g
H
1
0
() is the weak solution of the problem (2.49),
so it satises
a(w
g
, v) = l
g
(v) v H
1
0
(), (2.53)
where
l
g
(v) =
_

gv dxdy = (g, v),


a(w
g
, v) =
_

_
w
g
x
v
x
+
w
g
y
v
y
_
dxdy.
Consider the nite element approximation of (2.53):
nd w
gh
V
h
such that a(w
gh
, v
h
) = l
g
(v
h
) v
h
V
h
. (2.54)
From (2.53), (2.54) and the error bound (2.46), we deduce that
|w
g
w
gh
|
a
2h[w
g
[
H
2
()
,
56 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
and therefore, by (2.51),
|w
g
w
gh
|
a
2h|g|
L
2
()
. (2.55)
Now
(u u
h
, g) = (g, u u
h
) = l
g
(u u
h
)
= a(w
g
, u u
h
) = a(u u
h
, w
g
). (2.56)
Because w
gh
V
h
, (2.30) implies that
a(u u
h
, w
gh
) = 0,
and therefore, by (2.56),
(u u
h
, g) = a(u u
h
, w
g
) a(u u
h
, w
gh
)
= a(u u
h
, w
g
w
gh
)
= (u u
h
, w
g
w
gh
)
a
.
Applying the Cauchy-Schwarz inequality on the right,
(u u
h
, g) |u u
h
|
a
|w
g
w
gh
|
a
,
and thence by (2.46) and (2.55)
(u u
h
, g) 4h
2
[u[
H
2
()
|g|
L
2
()
. (2.57)
Substituting (2.57) into the right-hand side of (2.52), we obtain
|u u
h
|
L
2
()
4h
2
[u[
H
2
()
,
which is our improved error bound in the L
2
-norm. The proof presented above is
called the AubinNitsche duality argument.
2.6 Superapproximation in mesh-dependent norms
We have shown that the piecewise linear nite element approximation u
h
to the solu-
tion u of the homogeneous Dirichlet boundary value problem for Poissons equation
obeys the following error bounds:
|u u
h
|
H
1
()
Ch[u[
H
2
()
,
|u u
h
|
L
2
()
Ch
2
[u[
H
2
()
,
where C denotes a generic positive constant and h is the maximum element size
in the subdivision. It is possible to show that these error bounds are sharp in the
sense they cannot in general be improved. However, it was observed by engineers
2.6. SUPERAPPROXIMATION IN MESH-DEPENDENT NORMS 57
that, when sampled at certain special points, the nite element approximation u
h
is
more accurate than these error bounds might indicate. Indeed, we shall prove here
that when measured in a discrete counterpart of the Sobolev H
1
() norm, based
on sampling at the mesh points, the error u u
h
is O(h
2
). A result of this kind is
usually referred to as a superapproximation property.
We consider the model problem
u = f in , (2.58)
u = 0 on , (2.59)
where = (0, 1) (0, 1). We showed in Section 2.1 that when using continuous
piecewise linear nite elements on the uniform triangulation shown in Figure 2.5,
the nite element solution u
h
(x, y) can be expressed in terms of the nite element
basis functions
ij
(x, y) as
u
h
(x, y) =
N1

i=1
N1

j=1
U
ij

ij
(x, y),
where the U
ij
(= u
h
(x
i
, y
j
)) are obtained by solving the set of dierence equations

U
i+1,j
2U
ij
+ U
i1,j
h
2

U
i,j+1
2U
ij
+ U
i,j1
h
2
=
1
h
2
_ _
supp
ij
f(x, y)
ij
(x, y) dxdy, i, j = 1, . . . , N 1;
U
ij
= 0 when i = 0 or i = N or j = 0 or j = N.
Since u
h
(x, y) = 0 when (x, y) , we have adopted the convention that U
ij
= 0
when i = 0 or i = N or j = 0 or j = N. For simplicity, we shall write
F
ij
=
1
h
2
_ _
supp
ij
f(x, y)
ij
(x, y) dxdy, for i, j = 1, . . . , N 1.
Here N is an integer, N 2, and h = 1/N; the mesh-points are (x
i
, y
j
), i, j =
0, . . . , N, where x
i
= ih, y
j
= jh. These form the nite dierence mesh

h
= (x
i
, y
j
) : i, j = 0, . . . , N.
We consider the set of interior mesh points

h
= (x
i
, y
j
) : i, j = 1, ..., N 1,
and the set of boundary mesh points
h
=

h

h
. In more compact notation, the
dierence scheme can be written as follows:
(D
+
x
D

x
U
ij
+ D
+
y
D

y
U
ij
) = F
ij
, (x
i
, y
j
)
h
, (2.60)
U = 0 on
h
, (2.61)
58 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS



(i,j1)
(i,j)
(i,j+1)
(i1,j) (i+1,j)
s s s s s s s
s s s s s s s
s s s s s s s
s s s s s s s
s s s s s s s
s s s s s s s
s s s s s s s
u u u
u
u
Figure 2.12: The mesh
h
(), the boundary mesh
h
(), and a typical 5-point
dierence stencil.
where D
+
x
and D

x
denote the forward and backward divided dierence operators in
the x direction, respectively, dened by
D
+
x
V
ij
=
V
i+1,j
V
ij
h
, D

x
V
ij
=
V
ij
V
i1,j
h
,
and
D
+
x
D

x
V
ij
= D
+
x
(D

x
V
ij
) =
V
i+1,j
2V
ij
+ V
i1,j
h
2
is the second divided dierence operator in the x direction. Similar denitions apply
in the y direction.
For each i and j, 1 i, j N 1, the nite dierence equation (2.60) involves
ve values of U: U
i,j
, U
i1,j
, U
i+1,j
, U
i,j1
, U
i,j+1
. It is possible to write (2.60) as a
system of linear equations
AU = F, (2.62)
where
U = (U
11
, U
12
, . . . , U
1,N1
, U
21
, U
22
, . . . , U
2,N1
, . . . ,
. . . , U
i1
, U
i2
, . . . , U
i,N1
, . . . , U
N1,1
, U
N1,2
, . . . , U
N1,N1
)
T
,
2.6. SUPERAPPROXIMATION IN MESH-DEPENDENT NORMS 59
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
A =

j
j
Figure 2.13: The sparsity structure of the band matrix A.
F = (F
11
, F
12
, . . . , F
1,N1
, F
21
, F
22
, . . . , F
2,N1
, . . . ,
. . . , F
i1
, F
i2
, . . . , F
i,N1
, . . . , F
N1,1
, F
N1,2
, . . . , F
N1,N1
)
T
,
and A is an (N 1)
2
(N 1)
2
sparse matrix of band structure. A typical row of
the matrix contains ve non-zero entries, corresponding to the ve values of U in
the nite dierence stencil shown in Fig. 2.12, while the sparsity structure of A is
depicted in Fig. 2.13.
Next we show that (2.62) has a unique solution
2
. For two functions, V and W,
dened on
h
, we introduce the inner product
(V, W)
h
=
N1

i=1
N1

j=1
h
2
V
ij
W
ij
(which resembles the L
2
-inner product (v, w) =
_

v(x, y)w(x, y) dxdy.)


Lemma 7 Suppose that V is a function dened on

h
and that V = 0 on
h
; then
(D
+
x
D

x
V, V )
h
+ (D
+
y
D

y
V, V )
h
=
N

i=1
N1

j=1
h
2
[D

x
V
ij
[
2
+
N1

i=1
N

j=1
h
2
[D

y
V
ij
[
2
. (2.63)
2
The uniqueness of the solution to the linear system (2.62) is a trivial consequence of the
uniqueness of solution u
h
to the nite element method. The argument that follows is an alternative
way of verifying uniqueness; we present it here since some of its ingredients will be exploited in
the course of the proof of the superapproximation property.
60 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
Proof We shall prove that the rst term on the left is equal to the rst term on the right,
and the second term on the left to the second term on the right.
(D
+
x
D

x
V, V )
h
=
N1

i=1
N1

j=1
(V
i+1,j
2V
ij
+V
i1,j
)V
ij
=
N1

i=1
N1

j=1
(V
i+1,j
V
ij
)V
ij
+
N1

i=1
N1

j=1
(V
ij
V
i1,j
)V
ij
=
N

i=2
N1

j=1
(V
ij
V
i1,j
)V
i1,j
+
N1

i=1
N1

j=1
(V
ij
V
i1,j
)V
ij
=
N

i=1
N1

j=1
(V
ij
V
i1,j
)V
i1,j
+
N

i=1
N1

j=1
(V
ij
V
i1,j
)V
ij
=
N

i=1
N1

j=1
(V
ij
V
i1,j
)
2
=
N

i=1
N1

j=1
h
2
[D

x
V
ij
[
2
.
Similarly,
(D
+
y
D

y
V, V )
h
=
N1

i=1
N

j=1
h
2
[D

y
V
ij
[
2
,
and that completes the proof.
Returning to the analysis of the nite dierence scheme (2.60), we note that by
(2.63) we have
(AV, V )
h
= (D
+
x
D

x
V D
+
y
D

y
V, V )
h
= (D
+
x
D

x
V, V )
h
+ (D
+
y
D

y
V, V )
h
=
N

i=1
N1

j=1
h
2
[D

x
V
ij
[
2
+
N1

i=1
N

j=1
h
2
[D

y
V
ij
[
2
, (2.64)
for any V dened on

h
such that V = 0 on
h
. Now this implies that A is a
non-singular matrix. Indeed if AV = 0, then (2.64) yields:
D

x
V
ij
=
V
ij
V
i1,j
h
= 0,
i = 1, . . . , N,
j = 1, . . . , N 1;
D

y
V
ij
=
V
ij
V
i,j1
h
= 0,
i = 1, . . . , N 1,
j = 1, . . . , N.
Since V = 0 on
h
, these imply that V 0. Thus AV = 0 if and only if V = 0.
Hence A is non-singular, and U = A
1
F is the unique solution of (2.60). In summary
then, the (unique) solution of the nite dierence scheme (2.60) may be found by
solving the system of linear equations (2.62).
2.6. SUPERAPPROXIMATION IN MESH-DEPENDENT NORMS 61
In order to prove the stability of the nite dierence scheme (2.60), we introduce
the meshdependent norms
|U|
h
= (U, U)
1/2
h
,
and
|U|
1,h
= (|U|
2
h
+|D

x
U][
2
x
+|D

y
U][
2
y
)
1/2
,
where
|D

x
U][
x
=
_
N

i=1
N1

j=1
h
2
[D

x
U
ij
[
2
_
1/2
and
|D

y
U][
y
=
_
N1

i=1
N

j=1
h
2
[D

y
U
ij
[
2
_
1/2
.
The norm | |
1,h
is the discrete version of the Sobolev norm | |
H
1
()
,
|u|
H
1
()
=
_
|u|
2
L
2
()
+|
u
x
|
2
L
2
()
+|
u
y
|
2
L
2
()
_
1/2
.
With this new notation, the inequality (2.64) takes the following form:
(AV, V )
h
|D

x
V ][
2
x
+|D

y
V ][
2
y
. (2.65)
Using the discrete Poincare-Friedrichs inequality stated in the next lemma, we shall
be able to deduce that
(AV, V )
h
c
0
|V |
2
1,h
,
where c
0
is a positive constant.
Lemma 8 (Discrete Poincare-Friedrichs inequality) Let V be a function dened on

h
and such that V = 0 on
h
; then there exists a constant c

, independent of V
and h, such that
|V |
2
h
c

_
|D

x
V ][
2
x
+|D

y
V ][
2
y
_
(2.66)
for all such V .
Proof Writing
[V
ij
[
2
=
_
i

k=1
hD

x
V
kj
_
2
,
62 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
we deduce that
[V
ij
[
2

_
i

k=1
h
_ _
i

k=1
h[D

x
V
kj
[
2
_
i
N

k=1
h
2
[D

x
V
kj
[
2
.
Multiplying both sides by h
2
and summing through i = 1, . . . , N 1 and j = 1, . . . , N 1,
on noting that
h
2
N1

i=1
i = h
2
(N 1)N
2

1
2
,
we deduce that
|V |
2
h

1
2
|D

x
V ][
2
x
.
Analogously,
|V |
2
h

1
2
|D

y
V ][
2
y
.
Adding these two inequalities we complete the proof of (2.66) with c

=
1
4
.
Now (2.65) and (2.66) imply that
(AV, V )
h

1
c

|V |
2
h
.
Finally, combining this with (2.65) and recalling the denition of the norm | |
1,h
,
we obtain
(AV, V )
h
c
0
|V |
2
1,h
, (2.67)
where c
0
= (1 + c

)
1
.
Theorem 5 The scheme (2.60) is stable in the sense that
|U|
1,h

1
c
0
|F|
h
. (2.68)
Proof The proof is simple: it follows from (2.67) and the Cauchy-Schwarz inequality that
c
0
|V |
2
1,h
(AV, V )
h
= (F, V )
h
|F|
h
|V |
h
|F|
h
|V |
1,h
,
and hence the result.
Having established stability, we turn to the question of accuracy. We dene the
global error e
h
by e
h
(x, y) = u(x, y) u
h
(x, y) and note that u
h
(x
i
, y
j
) = U
ij
for
i, j = 1, . . . , N 1. Since u
h
(x, y) = 0 when (x, y) , we have adopted the
convention that U
ij
= 0 when i = 0 or i = N or j = 0 or j = N. Thus, writing
e
ij
= e
h
(x
i
, y
j
), we have that
e
ij
= u(x
i
, y
j
) U
ij
, 0 i, j N,
with e
ij
= 0 when i = 0 or i = N or j = 0 or j = N.
2.6. SUPERAPPROXIMATION IN MESH-DEPENDENT NORMS 63
Now,
Ae
ij
= Au(x
i
, y
j
) AU
ij
= Au(x
i
, y
j
) F
ij
= Au(x
i
, y
j
)
1
h
2
_ _
supp
ij

ij
(x, y)f(x, y) dxdy
=
_
1
h
2
_ _
supp
ij

ij
(x, y)

2
u
x
2
(x, y) dxdy D
+
x
D

x
u(x
i
, y
j
)
_
+
_
1
h
2
_ _
supp
ij

ij
(x, y)

2
u
y
2
(x, y) dxdy D
+
y
D

y
u(x
i
, y
j
)
_

ij
.
Thus,
Ae
ij
=
ij
, 1 i, j N 1,
e = 0 on
h
.
By virtue of (2.68),
|u U|
1,h
= |e|
1,h

1
c
0
||
h
. (2.69)
Assuming that u C
4
(

) and employing a Taylor series expansion of u(x, y) about


(x
i
, y
j
), we deduce that
[
ij
[ K
0
h
2
_
|

4
u
x
4
|
C(

)
+|

4
u
y
4
|
C(

)
_
,
where K
0
is a positive constant independent of h. Thus,
||
h
K
0
h
2
_
|

4
u
x
4
|
C(

)
+|

4
u
y
4
|
C(

)
_
. (2.70)
Finally (2.69) and (2.70) yield the following result.
Theorem 6 Let f L
2
() and suppose that the corresponding weak solution u
H
1
0
() belongs to C
4
(

); then
|u u
h
|
1,h

5
4
K
0
h
2
_
|

4
u
x
4
|
C(

)
+|

4
u
y
4
|
C(

)
_
. (2.71)
Proof Recall that c
0
= (1 + c

)
1
, c

=
1
4
, so that 1/c
0
=
5
4
, and combine (2.69) and
(2.70).
According to this result, the piecewise linear nite element approximation of the
homogeneous Dirichlet boundary value problem on uniform triangular subdivision
of size h is O(h
2
) convergent to the weak solution in the discrete Sobolev H
1
norm,
64 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
| |
1,h
, provided that u C
4
(

). Since this exceeds the rst order of convergence of


the global error observed in the Sobolev norm | |
H
1
()
, the result encapsulated in
Theorem 6 is referred to as a superapproximation property. In fact the smoothness
requirement u H
1
0
()C
4
(

) can be relaxed to u H
1
0
()H
3
() while retaining
the superapproximation property |u u
h
|
1,h
= O(h
2
); the proof of this is more
technical and relies on the Bramble-Hilbert lemma (See Chapter 3).
Chapter 3
Piecewise polynomial
approximation
In the previous chapter we discussed nite element approximations to elliptic bound-
ary value problems using piecewise polynomials of degree 1. The purpose of this
chapter is to develop, in a more general setting, the construction of nite element
spaces and to formalise the concepts introduced in Chapter 2.
3.1 Construction of nite element spaces
Let us consider an elliptic boundary value problem written in its weak formulation:
nd u in V such that a(u, v) = l(v) v V ,
where H
1
0
() V H
1
(); in the case of a homogeneous Dirichlet boundary value
problem V = H
1
0
(), in the case of a Neumann, Robin or oblique derivative bound-
ary value problem, V = H
1
(). In order to dene a nite element approximation
to this problem we need to construct a nite-dimensional subspace V
h
of V consist-
ing of continuous piecewise polynomial functions of a certain degree dened on a
subdivision of the computational domain . We have already discussed the special
case when V
h
consists of continuous piecewise linear functions; here we shall put this
construction into a general context.
3.1.1 The nite element
We begin by giving a formal denition of a nite element.
Denition 2 Let us suppose that
(i) K R
n
is a simply connected bounded open set with piecewise smooth bound-
ary (the element domain);
65
66 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
(ii) T is a nite-dimensional space of functions dened on K (the space of shape
functions);
(iii) A = N
1
, . . . , N
k
is a basis for T

(the set of nodal variables).


Then (K, T, A) is called a nite element.
In this denition T

denotes the algebraic dual of the linear space T.


Denition 3 Let (K, T, A) be a nite element, and let
1
,
2
, . . . ,
k
be a basis
for T, dual to A; namely,
N
i
(
j
) =
ij
, 1 i, j k.
Such a basis is called a nodal basis for T.
We give a simple example to illustrate these denitions.
Example 7 (The one-dimensional Lagrange element) Let K = (0, 1), T the set
of linear polynomials, and A = N
1
, N
2
, where N
1
(v) = v(0) and N
2
(v) = v(1)
for all v T. Then (K, T, A) is a nite element, with nodal basis
1
,
2
where

1
(x) = 1 x and
2
(x) = x.
Next we give an equivalent characterisation of condition (iii) in Denition 2.
Lemma 9 Let T be a k-dimensional linear space of functions on R
n
, and suppose
that N
1
, N
2
, . . . , N
k
is a subset of the dual space T

. Then the following two


statements are equivalent:
(a) N
1
, N
2
, . . . , N
k
is a basis for T

;
(b) Given that v T and N
i
(v) = 0 for i = 1, . . . , k, then v 0.
Proof Let
1
, . . . ,
k
be a basis for T. Now N
1
, . . . , N
k
is a basis for T

if and only
if any L T

can written in a unique fashion as a linear combination of the N


i
s:
L =
1
N
1
+. . . +
k
N
k
.
This is equivalent to demanding that, for each i = 1, . . . , k, L(
i
) can be written in a
unique fashion as a linear combination
L(
i
) =
1
N
1
(
i
) +. . . +
k
N
k
(
i
).
Let us dene the matrix B = (N
j
(
i
))
i,j=1,...,k
and the vectors
y = (L(
1
), . . . , L(
k
))
T
, a = (
1
, . . . ,
2
)
T
.
Then the last condition is equivalent to requiring that the system of linear equations
Ba = y has a unique solution, which, in turn, is equivalent to demanding that the matrix
B be invertible. Given any v T, we can write
v =
1

1
+. . . +
k

k
.
3.1. CONSTRUCTION OF FINITE ELEMENT SPACES 67
Now N
i
(v) = 0 for all i = 1, . . . , k if and only if

1
N
i
(
1
) +. . . +
k
N
i
(
k
) = 0, i = 1, . . . , k. (3.1)
Thus (b) is equivalent to requiring that (3.1) implies
1
= . . . =
k
= 0. Let C =
(N
i
(
j
))
i,j=1,...,k
; then (b) holds if and only if Cb = 0, with b = (
1
, . . . ,
k
)
t
, implies
that b = 0, which is equivalent to demanding that C be invertible. However C
t
= B and
therefore (a) and (b) are equivalent.
Motivated by this result, we introduce the following denition.
Denition 4 We say that A determines T if T with N() = 0 for all N A
implies that = 0.
We shall need the following Lemma.
Lemma 10 Suppose that P is a polynomial of degree d 1 that vanishes on the
hyperplane x : L(x) = 0 where L is a non-degenerate linear function. Then we
can write P in the factorised form P = LQ where Q is a polynomial of degree (d1).
Proof Let us write x = (x
1
, . . . , x
n1
). Suppose that we have carried out an ane change
of variables such that L( x, x
n
) = x
n
; so, the hyperplane L( x, x
n
) = 0 is the hyperplane
x
n
= 0; then, by hypothesis, P( x, 0) 0. Since P is of degree d, we have that
P( x, x
n
) =
d

j=0

||dj
c
j
x

x
j
n
,
where = (i
1
, . . . , i
n1
) and x

= x
i
1
1
x
i
n1
n1
. Letting x
n
= 0 we get
0 P( x, 0) =

||d
c
0
x

,
which implies that c
0
= 0 for [[ d. Hence
P( x, x
n
) =
d

j=1

||dj
c
j
x

x
j
n
= x
n
d

j=1

||dj
c
j
x

x
j1
n
= x
n
Q = LQ
where Q is of degree (d 1).
3.1.2 Examples of triangular nite elements
Let K be a triangle and let T
k
denote the set of all polynomials of degree k in
two variables. The dimension of the linear space T
k
is displayed in Table 3.1.2.
68 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
u u z
1
L
3
z
2

u
z
3
L
2
C
C
C
C
C
C
C
C
C
C
C
C
L
1
Figure 3.1: Linear Lagrange triangle with edges L
1
, L
2
, L
3
and vertices z
1
, z
2
, z
3
.
k dim T
k
1 3
2 6
3 10
... ...
k
1
2
(k + 1)(k + 2)
Table 3.1.2: The dimension of the linear space T
k
.
Lagrange elements
Example 8 Let k = 1 and take T = T
1
, A = A
1
= N
1
, N
2
, N
3
(and therefore
the dimension of T
1
is 3), where N
i
(v) = v(z
i
) and z
1
, z
2
, z
3
are the vertices of the
triangle K, as shown in Figure 3.1.
In the gure indicates function evaluation at the point where the dot is placed.
We verify condition (iii) of Denition 2 using part (b) of Lemma 9; namely, we prove
that A
1
determines T
1
. Indeed, let L
1
, L
2
and L
3
be non-trivial linear functions
which dene the lines that contain the three edges of the triangle. Suppose that a
polynomial P T
1
vanishes at z
1
, z
2
and z
3
. Since P[
L
1
is a linear function of
one variable that vanishes at two points, it follows that P 0 on L
1
. By virtue of
Lemma 10, we can write P = cL
1
where c is a constant (i.e. a polynomial of degree
1 1 = 0). However, because L
1
(z
1
) ,= 0,
0 = P(z
1
) = cL
1
(z
1
)
implies that c = 0; thus P 0. Hence, according to Lemma 9, A
1
determines T
1
.
Example 9 Now take k = 2, let T = T
2
and A = A
2
= N
1
, N
2
, . . . , N
6
(so we
have that dim T
2
= 6), where
N
i
(v) =
_
v(at the ith vertex of the triangle), i = 1, 2, 3
v(at the midpoint of the (i 3)rd edge), i = 4, 5, 6.
3.1. CONSTRUCTION OF FINITE ELEMENT SPACES 69
u u u z
1
L
3
z
6
z
2

u
z
3
z
5
L
2
C
C
C
C
C
C
C
C
C
C
C
C
z
4
L
1
u u
Figure 3.2: Quadratic Lagrange triangle with edges L
1
, L
2
, L
3
, vertices z
1
, z
2
, z
3
,
and z
4
, z
5
and z
6
denoting the midpoints of L
1
, L
2
and L
3
, respectively.
We have to show that A
2
determines T
2
. Let L
1
, L
2
and L
3
be non-trivial linear
functions which dene the lines containing the edges of the triangle. Let P T
2
be such that P(z
i
) = 0 for i = 1, . . . , 6. As P[
L
1
is a quadratic function of one
variable that vanishes at three points, it follows that P 0 on L
1
. By Lemma 10,
P = L
1
Q
1
, where the degree of Q
1
is one less than the degree of P, so Q
1
is of
degree 1. However, by an analogous argument P also vanishes along L
2
. Therefore,
L
1
Q
1
[
L
2
0. Thus, on L
2
, either L
1
0 or Q
1
0. But L
1
can be equal to zero
only at one point of L
2
because the triangle is non-degenerate. Thus Q
1
0 on
L
2
, except possibly at one point. However, by continuity of Q
1
, we then have that
Q
1
0 along the whole of L
2
.
Now applying again Lemma 10, we deduce that Q
1
= L
2
Q
2
, where the degree of
Q
2
is one less than the degree of Q
1
, so Q
2
is of degree 0. Thus, Q
2
c, where c
is a constant. Hence, P = cL
1
L
2
. However P(z
6
) = 0 and z
6
does not lie on either
L
1
or L
2
. Consequently,
0 = P(z
6
) = cL
1
(z
6
)L
2
(z
6
).
Therefore, c = 0 since L
1
(z
6
) ,= 0 and L
2
(z
6
) ,= 0. This nally implies that P 0,
so A
2
determines T
2
.
Hermite elements
Example 10 Let us suppose that k = 3, and let T = T
3
. Let denote evaluation
at a point and let _ signify evaluation of the gradient at the centre point of the
circle. We shall prove that A = A
3
= N
1
, N
2
, . . . , N
10
, as depicted in Figure 3.3,
determines T
3
(whose dimension is precisely 10). Let, as before, L
1
, L
2
and L
3
be
the lines corresponding to the three sides of the triangle and suppose that P T
3
and
N
i
(P) = 0 for the i = 1, 2 . . . , 10. The restriction of P to L
1
is a cubic polynomial
of one variable with double roots at z
2
and z
3
. Hence P 0 along L
1
. Similarly,
P 0 along the edges L
2
and L
3
. By Lemma 10, we can write P = cL
1
L
2
L
3
, where
c is a constant. However,
0 = P(z
4
) = cL
1
(z
4
)L
2
(z
4
)L
3
(z
4
),
70 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
u
_
u
_
z
1
L
3
z
2

u
_
z
3
L
2
C
C
C
C
C
C
C
C
C
C
C
C
L
1
u
z
4
Figure 3.3: Cubic Hermite triangle with edges L
1
, L
2
, L
3
and vertices z
1
, z
2
and z
3
,
and centroid z
3
.
and so c = 0, since L
i
(z
4
) ,= 0 for i = 1, 2, 3. Thus P 0 and we deduce that A
uniquely determines T
3
.
3.1.3 The interpolant
Having described a number of nite elements, we now wish to piece them together
to construct nite-dimensional subspaces of Sobolev spaces.
Denition 5 Let (K, T, A) be a nite element and let the set
i
: i = 1, . . . , k,
be a basis for T dual to A. Given that v is a function for which all N
i
A,
i = 1, . . . , k, are dened, we introduce the local interpolant J
K
v by
J
K
v =
k

i=1
N
i
(v)
i
.
In order to illustrate the idea of local interpolant, we give a simple example.
Example 11 Consider the triangle K shown in Figure 3.4, let T = T
1
, A =
N
1
, N
2
, N
3
as in the case of the linear Lagrange element (k = 1), and suppose
that we wish to nd the local interpolant J
K
v of the function v dened by v(x, y) =
(1 + x
2
+ y
2
)
1
.
By denition,
J
K
v = N
1
(v)
1
+ N
2
(v)
2
+ N
3
(v)
3
.
Thus we must determine
i
, i = 1, 2, 3, to be able to write down the local interpolant.
This we do, using Denition 3, as follows. The line L
1
has equation y = 1 x;
as
1
vanishes at z
2
and z
3
, and thereby along the whole of L
1
, it follows that

1
= cL
1
= c(1 x y), where c is a constant to be determined. Also, N
1

1
= 1,
so c =
1
(z
1
) = 1. Hence,
1
= 1 x y. Similarly,
2
= L
2
(x, y)/L
2
(z
2
) = x and

3
= L
3
(x, y)/L
3
(z
3
) = y.
Having found
1
,
2
and
3
, we have that
J
K
(v) = N
1
(v)(1 x y) + N
2
(v)x + N
3
(v)y.
3.1. CONSTRUCTION OF FINITE ELEMENT SPACES 71
u u z
1
= (0, 0)
L
3
z
2
= (1, 0)
u
z
3
= (0, 1)
L
2
@
@
@
@
@
@
@
@
@
@
@
@
L
1
K
Figure 3.4: Linear Lagrange triangle with edges L
1
, L
2
, L
3
, and the vertices z
1
, z
2
and z
3
where the local interpolant is evaluated.
In fact, in our case N
1
(v) = v(z
1
) = 1, N
2
(v) = v(z
2
) =
1
2
and N
3
(v) = v(z
3
) =
1
2
,
and so
J
K
(v) = 1
1
2
(x + y).
The next lemma summarises the key properties of the local interpolant.
Lemma 11 The local interpolant has the following properties:
a) The mapping v J
K
v is linear.
b) N
i
(J
K
v)) = N
i
(v), i = 1, . . . , k.
c) J
K
(v) = v for v T; consequently J
K
is idempotent on T, that is, J
2
K
= J
K
.
Proof
a) Since each N
i
: v N
i
(v), i = 1, . . . , k, is a linear functional, v J
K
v has the
same property.
b) Clearly
N
i
(J
K
(v)) = N
i
_
_
k

j=1
N
j
(v)
j
_
_
=
k

j=1
N
j
(v)N
i
(
j
)
=
k

j=1
N
j
(v)
ij
= N
i
(v),
for i = 1, . . . , k, where
ij
= 1 when i = j and = 0 when i ,= j.
c) It follows from b) that N
i
(vJ
K
(v)) = 0, i = 1, . . . , k, which implies that J
K
(v) = v
for all v T. The second assertion follows from this; indeed, J
2
K
v = J
K
(J
K
v) = J
K
v
since J
K
v T.
72 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
That completes the proof of the lemma.
We can now glue together the element domains to obtain a subdivision of the
computational domain, and merge the local interpolants to obtain a global inter-
polant.
Denition 6 A subdivision of the computational domain is a nite collection
of open sets K
i
such that
(1) K
i
K
j
= if i ,= j, and
(2)

K
i
=

.
Denition 7 Suppose that is a bounded open set in R
n
with subdivision T . As-
sume that each element domain K in the subdivision is equipped with some type of
shape functions T and nodal variables A, such that (K, T, A) forms a nite el-
ement. Let m be the order of the highest partial derivative involved in the nodal
variables. For v C
m
(

) the global interpolant J


h
v is dened on

by
J
h
v[
K
i
= J
K
i
v K
i
T .
In the absence of further conditions on the subdivision it is not possible to assert
the continuity of the global interpolant. Next we shall formulate a simple condition
which ensures that the global interpolant is a continuous function on

. To keep the
presentation simple, we shall restrict ourselves to the case of two space dimensions,
namely when R
2
, although an analogous denition can me made in R
n
.
Denition 8 A triangulation of a polygonal domain is a subdivision of con-
sisting of triangles which have the property that
(3) No vertex of any triangle lies in the interior of an edge of another triangle.
From now on, we shall use the word triangulation without necessarily implying that
R
2
: when R
n
and n = 2 we shall mean that the condition of this denition
is satised; when n > 2, the obvious generalisation of this condition to n dimensions
will be meant to hold.
Denition 9 We say that an interpolant has continuity of order r (or, briey,
that it is C
r
) if J
h
v C
r
(

) for all v C
m
(

). The space
J
h
v : v C
m
(

)
is called a C
r
nite element space.
For simplicity, again, the next result is stated and proved in the case of n = 2;
an analogous result holds for n > 2.
3.1. CONSTRUCTION OF FINITE ELEMENT SPACES 73
Theorem 7 The Lagrange and Hermite elements on triangles are all C
0
elements.
More precisely, given a triangulation T of , it is possible to choose edge nodes for
the corresponding elements (K, T, A), K T , such that the global interpolant J
h
v
belongs to C
0
(

) for all v in C
m
(

), where m = 0 for Lagrange and m = 1 for


Hermite elements.
Proof It suces to show that continuity holds across each edge. Let K
i
, i = 1, 2, be
two triangles in the triangulation T with common edge e. Assuming that we choose nodes
interior to e in a symmetric way, it follows that the edge nodes on e for the elements on
both K
1
and K
2
are at the same location in space.
Let w = J
K
1
v J
K
2
v, where we interpret J
K
1
v and J
K
2
v to be dened everywhere
by extension outside K
1
and K
2
, respectively, as polynomials. Now w is a polynomial of
degree k and its restriction to the edge e vanishes at the one-dimensional Lagrange (or
Hermite) nodes. Therefore w[
e
0. Hence J
K
1
[
e
v = J
K
2
v[
e
, i.e. the global interpolant
is continuous across each edge.
In order to be able to compare global interpolation operators on dierent ele-
ments, we introduce the following denition (now for K R
n
.)
Denition 10 Let (K, T, A) be a nite element and suppose that F(x) = Ax + b
where A is a non-singular nn matrix and x and b are n-component column vectors.
The nite element (

K,

T,

A) is ane equivalent to (K, T, A) if:
(a) F(K) =

K;
(b) F


T = T and
(c) F

A =

A.
Here F

is the pull-back of F dened by F

( v) = v F, and F

is the push-forward
of F dened by (F

N)( v) = N(F

( v)) = N( v F).
Example 12 Lagrange elements on triangles with appropriate choice of edge and
interior nodes are ane equivalent. The same is true of Hermite elements on tri-
angles.
3.1.4 Examples of rectangular elements
To conclude this section we consider nite elements dened on rectangles. Let
Q
k
=
_

j
c
j
p
j
(x)q
j
(y) : p
j
and q
j
are polynomials of degree k
_
.
It can be shown that Q
k
is a linear space of dimension (dim T
1
k
)
2
, where T
1
k
denotes
the set of all polynomials of a single variable of degree k and dim T
1
k
signies its
dimension.
We give two examples, without going into details.
74 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
u u z
1
L
1
z
2
L
2
u z
3
L
3
L
4
uz
4
Figure 3.5: Bilinear Lagrange rectangle with edges L
1
, L
2
, L
3
, L
4
and the vertices
z
1
, z
2
, z
3
and z
4
.
u u z
1
L
1
u z
3
L
2
u
z
2
z
8
u z
7
L
3
L
4
u z
4
uz
6
u
z
5
uz
9
Figure 3.6: Biquadratic Lagrange rectangle with edges L
1
, L
2
, L
3
, L
4
and the vertices
z
1
, z
3
, z
7
, z
9
, midpoints of edges z
2
, z
4
, z
6
, z
8
, and centroid z
5
.
Example 13 (Bilinear Lagrange rectangle) Let k = 1 and suppose that K is a
rectangle. Further, let T = Q
1
and let A = N
1
, . . . , N
4
with N
i
(v) = v(z
i
) with
z
i
, i = 1, . . . , 4, as in Figure 3.5. We leave it as an exercise to the reader to show,
using Lemmas 9 and 10 that A determines T = Q
1
(the dimension of Q
1
is equal
to 4).
Example 14 (Biquadratic Lagrange rectangle) Let k = 2 and suppose that K is a
rectangle. We let T = Q
2
and put A = N
1
, . . . , N
9
with N
i
(v) = v(z
i
) with z
i
,
i = 1, . . . , 9, as in Figure 3.6. It is left as an exercise to show that A determines
T = Q
2
(the dimension of Q
2
is equal to 9).
3.2 Polynomial approximation in Sobolev spaces
In this section we shall develop the approximation theory for the nite element
spaces described in the previous section. We shall adopt a constructive approach
which will enable us to calculate the constants in the error estimates explicitly. The
technique is based on the use of the Hardy-Littlewood maximal function, following
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 75
the work of Ricardo Duran
1
. An alternative approach which exploits the theory of
Riesz potentials is presented in the Brenner-Scott monograph cited in the reading
list.
3.2.1 The Bramble-Hilbert lemma
A key device in nite element error analysis is the following result.
Lemma 12 (Bramble-Hilbert lemma) Suppose that is a bounded open set in R
n
and assume that is star-shaped with respect to every point in a set B of positive
measure contained in (i.e. for all x the closed convex hull of x B is
a subset of ). Let l be a bounded linear functional on the Sobolev space W
m
p
(),
m 1, 1 < p < , such that l(Q) = 0 for any polynomial Q of degree m 1.
Then there exists a constant C
1
> 0 such that
[l(v)[ C
1
[v[
W
m
p
()
for all v W
m
p
().
Proof By hypothesis, there exists C
0
> 0 such that
[l(v)[ C
0
|v|
W
m
p
()
v W
m
p
().
Since l(Q) = 0 for all Q T
m1
, we have by the linearity of l that
[l(v)[ = [l(v Q)[ C
0
|v Q|
W
m
p
()
= C
0
_
_
m

j=0
[v Q[
p
W
j
p
()
_
_
1/p
= C
0
_
_
m1

j=0
[v Q[
p
W
j
p
()
+[v[
p
W
m
p
()
_
_
1/p
C
0
m1

j=0
[v Q[
W
j
p
()
+[v[
W
m
p
()
.
In order to complete the proof it remains to prove that
K
j
> 0 v W
j
p
() Q T
m1
such that
[v Q[
W
j
p
()
K
j
[v[
W
m
p
()
, j = 0, . . . , m1. (3.2)
This will be done in the rest of the section. Once (3.2) has been veried, we shall have
that
[l(v)[ C
0
_
_
1 +
m1

j=0
K
j
_
_
[v[
W
m
p
()
,
1
R. Duran: On polynomial approximation in Sobolev spaces. SIAM Journal of Numerical
Analysis, 20, No. 5., (1983), pp. 985988.
76 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
and the proof will be complete, with the constant
C
1
= C
0
_
_
1 +
m1

j=0
K
j
_
_
.

The original proof of (3.2) given by Bramble and Hilbert was based on the use of
the Hahn-Banach theorem and was non-constructive in nature in the sense that it
did not provide computable constants K
j
, j = 0, . . . , m1; only the existence of such
constants was proved. The remainder of this section is devoted to the (constructive)
proof of (3.2). Our main tool is the following lemma.
Lemma 13 Let g L
p
(R
n
), 1 < p < . Given R
n
such that [[ = 1, we dene
g
1
(x, ) = sup
t>0
1
t
_
t
0
[g(x + s)[ ds
and
g

(x) =
__
||=1
g
1
(x, )
p
d

_
1/p
.
Then
|g

|
L
p
(R
n
)

p
p 1

1/p
n
|g|
L
p
(R
n
)
,
where
n
is the measure of the unit sphere in R
n
.
Proof Since g
1
(, ) is the Hardy-Littlewood maximal function of g() in the direction ,
it follows that
2
_
R
n
g
1
(x, )
p
dx
_
p
p 1
_
p
_
R
n
[g(x)[
p
dx,
and therefore
_
R
n
[g

(x)[
p
=
_
R
n
_
_
||=1
g
1
(x, )
p
d

_
dx
=
_
||=1
__
R
n
g
1
(x, )
p
dx
_
d


n
_
p
p 1
_
p
_
R
n
[g(x)[
p
dx.
Upon taking the pth root of the two sides in this inequality we arrive at the desired result.

Now we are ready to prove (3.2).


2
See E.M. Stein and T.S. Murphy: Harmonic Analysis: Real Variable Methods, Orthogonality
and Oscillatory Integrals. Princeton University Press, 1993; or A.P. Calderon: Estimates for
singular integral operators in terms of maximal functions. Stud. Math. 44, (1972), pp.563582.
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 77
Theorem 8 Let R
n
be a bounded open set which is star-shaped with respect to
each point in a set of positive measure B . Let 1 < p < , 0 j < m, and let
d be the diameter of . If v W
m
p
() then
inf
QP
m1
[v Q[
W
j
p
()
C
d
mj+(n/p)
[B[
1/p
[v[
W
m
p
()
,
where [B[ denotes the measure of B and
C = ( : [[ = j)
mj
n
1/p
p
p 1

1/p
n
_
_

||=mj
(!)
p

_
_
1/p

,
with 1/p + 1/p

= 1. Here, for a set A, A denotes the number of elements in A


and, for a multi-index = (
1
, . . . ,
n
), ! =
1
! . . .
n
!.
Proof Because C

) is dense
3
in W
m
p
(), it suces to prove the theorem for v C

).
Given x B, we dene
P
m
(v)(x, y) =

||<m
D

v(x)
(y x)

!
and
Q
m
(v)(y) =
1
[B[
_
B
P
m
(v)(x, y) dx.
Here we used the multi-index notation (y x)

= (y
1
x
1
)

1
. . . (y
n
x
n
)

n
. It is easy
to prove by induction that
D

Q
m
(v)(y) = Q
m||
(D

v)(y).
Thus,
[v Q
m
(v)[
W
j
p
()
=
_
_

||=j
|D

(v Q
m
(v))|
p
L
p
()
_
_
1/p

||=j
|D

(v Q
m
(v))|
L
p
()
=

||=j
|D

v Q
mj
(D

v)|
L
p
()
.
Now let us estimate |D

v Q
mj
(D

v)|
L
p
()
for each , [[ = j. As
(D

v Q
mj
(D

v))(y) =
1
[B[
_
B
[D

v(y) P
mj
(D

v)(x, y)] dx,


3
See R.A. Adams: Sobolev Spaces. Academic Press, 1975.
78 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
it follows, by applying Minkowskis inequality for integrals
4
that
|D

v Q
mj
(D

v)|
L
p
()

1
[B[
_
B
__

[D

v(y) P
mj
(D

v)(x, y)[
p
dy
_
1/p
dx. (3.3)
Now recalling the integral-remainder for Taylor series, for x B, y , we have that
[D

v(y) P
mj
(D

v)(x, y)[
=

(mj)

||=mj
(y x)

!
_
1
0
D

v(x +t(y x))(1 t)


mj1
dt

(mj)d
mj
_
1
0

||=mj
1
!
[D

v(x +t(y x))[ dt


= (mj)d
mj
1
[y x[
_
|yx|
0

||=mj
1
!

v
_
x +s
y x
[y x[
_

ds.
Let g be the function that coincides with

||=mj
1
!
[D

v[ in and is identically zero


outside . If g
1
and g

are the functions associated with g, as dened in Lemma 13, we


have that
[D

v(y) P
mj
(D

v)(x, y)[ (mj)d


mj
g
1
_
x,
y x
[y x[
_
,
and therefore
[D

v(y) P
mj
(D

v)(x, y)[
p
(mj)
p
d
(mj)p
g
p
1
_
x,
y x
[y x[
_
.
Noting that B , it follows for each x B that
_

[D

v(y) P
mj
(D

v)(x, y)[
p
dy (mj)
p
d
(mj)p
_
|yx|d
g
p
1
_
x,
y x
[y x[
_
dy.
Thus,
__

[D

v(y) P
mj
(D

v)(x, y)[
p
dy
_
1/p
(mj)d
mj
_
_
d
0
_
||=1
g
p
1
(x, ) d

r
n1
dr
_
1/p
= (mj)d
mj
_
d
n
n
_
1/p
_
_
||=1
g
p
1
(x, ) d

_
1/p
= (mj)d
mj
_
d
n
n
_
1/p
g

(x).
4
Minkowskis integral inequality states that, for a function u C(

B

),
|
_
B
u(x, ) dx|
L
p
()

_
B
|u(x, )|
L
p
()
dx.
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 79
Inserting this into (3.3) we get, by Holders inequality (see Ch.1, Sec.1.1.2).
|D

v Q
mj
(D

v)|
L
p
()

1
[B[
(mj) d
mj
_
d
n
n
_
1/p
_
B
g

(x) dx

1
[B[
(mj) d
mj
[B[
1(1/p)
_
d
n
n
_
1/p
|g

|
L
p
(R
n
)
and hence, by Lemma 13,
|D

v Q
mj
(D

v)|
L
p
()

mj
n
1/p
d
mj+(n/p)
[B[
1/p
p
p 1

1/p
n
|g|
L
p
(R
n
)
=
mj
n
1/p
d
mj+(n/p)
[B[
1/p
p
p 1

1/p
n
|g|
L
p
()
.
However,
|g|
L
p
()
= |

||=mj
1
!
[D

v[ |
L
p
()

||=mj
1
!
|D

v|
L
p
()

_
_

||=mj
(!)
p

_
_
1/p
_
_

||=mj
|D

v|
p
L
p
()
_
_
1/p
,
where 1/p + 1/p

= 1. Therefore,
|D

v Q
mj
(D

v)|
L
p
()

mj
n
1/p
d
mj+(n/p)
[B[
1/p
p
p 1

1/p
n
_
_

||=mj
(!)
p

_
_
1/p

_
_

||=mj
|D

v|
p
L
p
()
_
_
1/p
.
Recalling that [[ = j, we deduce that
[v Q
m
(v)[
W
j
p
()
K
j
[v[
W
m
p
()
,
where
K
j
= C
d
mj+(n/p)
[B[
1/p
.
Since Q
m
T
m1
, this completes the proof of Theorem 8, and thereby also the proof of
the Bramble-Hilbert lemma (Lemma 12).
Corollary 3 Let R
n
be a bounded open set of diameter d which is star-shaped
with respect to every point of an open ball B of diameter d, (0, 1]. Suppose
that 1 < p < and 0 j < m, m 1. If v W
m
p
() then
inf
QP
m1
[v Q[
W
j
p
()
C(m, n, p, j, )d
mj
[v[
W
m
p
()
,
80 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
where
C(m, n, p, j, ) =
n/p
( : [[ = j)
p(mj)
p 1
_
_

||=mj
(!)
p

_
_
1/p
,
with 1/p + 1/p

= 1.
Proof Note that
[B[ =
(d)
n

n
n
.

In order to minimise the size of the constant C(m, n, p, j, ), should be taken


as large as possible a number in the interval (0, 1] such that is star-shaped with
respect to each point in a ball B of radius d.
3.2.2 Error bounds on the interpolation error
We shall apply Corollary 3 to derive a bound on the error between a function and
its nite element interpolant. We begin by estimating the norm of the local inter-
polation operator.
Lemma 14 Let (K, T, A) be a nite element such that the diameter of K is equal to
1, T W
m

(K) and A (C
l
(

K))

(i.e. the nodal variables in A involve derivatives


up to order l, and each element of the set A is a bounded linear functional on
C
l
(

K)). Then the local interpolation operator is bounded from C
l
(

K) into W
m
p
(K)
for 1 < p < .
Proof Let A = N
1
, . . . , N
k
, and let
1
, . . . ,
k
T be the basis dual to A. The
local interpolant of a function u is dened by the formula
J
K
u =
k

i=1
N
i
(u)
i
,
where each
i
W
m

(K) W
m
p
(K), 1 < p < , by hypothesis. Thus
|J
K
u|
W
m
p
(K)

k

i=1
[N
i
(u)[ |
i
|
W
m
p
(K)

_
k

i=1
|N
i
|
(C
l
(

K)
|
i
|
W
m
p
(K)
_
|u|
C
l
(

K)
= Const. |u|
C
l
(

K)
.
and that completes the proof.
We dene
(K) = sup
vC
l
(

K)
|J
K
v|
W
m
p
(K)
|v|
C
l
(

K)
,
the norm of the local interpolation operator J
K
: C
l
(

K) W
m
p
(K).
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 81
Theorem 9 Let (K, T, A) be a nite element satisfying the following conditions:
(i) K is star-shaped with respect to some ball contained in K;
(ii) T
m1
T W
m

(K);
(iii) A (C
l
(

K))

.
Suppose that 1 < p < and m l (n/p) > 0. Then, for 0 j m and
v W
m
p
(K) we have that
[v J
K
v[
W
j
p
(K)
C(m, n, p, , (

K))h
mj
K
[v[
W
m
p
(K)
,
where h
K
is the diameter of K,

K = x/h
K
: x K and is the largest real
number in the interval (0, 1] such that a ball of diameter h
K
is contained in K.
Proof It suces to take K with diameter equal to 1, in which case K =

K; the general
case follows by a simple scaling argument. Also, note that the local interpolation operator
is well dened on W
m
p
(K) by the Sobolev embedding theorem
5
and there exists a constant
C = C
m,n,p
such that, for all v W
m
p
(K),
|v|
C
l
(

K)
C
m,n,p
|v|
W
m
P
(K)
.
Let Q
m
v be as in the proof of Theorem 8. Since J
K
f = f for any f T, we have that
J
K
Q
m
v = Q
m
v,
because Q
m
v T
m1
T. Thus,
|v J
K
v|
W
m
p
(K)
|v Q
m
v|
W
m
p
(K)
+|Q
m
v J
K
v|
W
m
p
(K)
= |v Q
m
v|
W
m
p
(K)
+|J
K
(Q
m
v J
K
v)|
W
m
p
(K)
|v Q
m
v|
W
m
p
(K)
+(K) |Q
m
v J
K
v|
C
l
(

K)
(1 +C
m,n,p
(K)) |v Q
m
v|
W
m
p
(K)
,
by the Sobolev embedding theorem. Finally, by (3.2) we deduce that
|v J
K
v|
W
m
p
(K)
C(m, n, p, , (K)) [v[
W
m
p
(K)
,
and hence, for 0 j m, we have that
[v J
K
v[
W
m
p
(K)
C(m, n, p, , (K)) [v[
W
m
p
(K)
.
That completes the proof.
Next we show that, under a certain regularity condition on the subdivision T =
K of the computational domain , the constant C(m, n, p, , (

K)) can be made
independent of (

K).
5
The Sobolev embedding theorem asserts that W
m
p
(K) C
l
(

K) for ml >
n
p
, 1 p < and
the identity operator Id : v W
m
p
(K) v C
l
(

K) is a bounded linear operator.
82 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
Let us suppose that each (

K,

T,

A) is ane equivalent to a single reference
element (K, T, A) through an ane transformation
x x = Ax ax + b,
where a = (a
ij
) is an invertible N N matrix and b is a column vector of size N, of
the same length as the column vector x.
6
We shall denote the entries of the matrix
a
1
by (a
1
)
ij
. The denition of ane equivalence yields:

K
v( x) =

NN
(A

N) v(A
1
)

N
( x),
where
(A

N)( v) = N(A

v), (A

v)x = v(Ax).
Thus,
[(A

N)( v)[ = [N(A

v)[ C
N
|A

v|
C
l
(

K)
C
N,n,l
_
1 + max
1i,jn
[a
ij
[
_
l
| v|
C
l
(

K)
.
Also,
|(A
1
)

N
|
W
m
p
(

K)
C

N,n,m
_
1 + max
1i,jn
[(a
1
)
ij
[
_
m
[det a[
1/p
|
N
|
W
m
p
(K)
.
Since |
N
|
W
m
p
(K)
is a xed constant on the reference element K, we have that
|

K
v|
W
m
p
(

K)
C
ref
_
1 + max
1i,jn
[a
ij
[
_
l

_
1 + max
1i,jn
[(a
1
)
ij
[
_
m
[det a[
1/p
| v|
C
l
(

K)
,
where
C
ref
= [A[ max
NN
C
N,n,l
max
NN
C

N,n,m
max
NN
|
N
|
W
m
p
(K)
,
and [A[ denotes the number of nodal variables (i.e. the dimension of T). Thus, we
have shown that
(

K) C
ref
_
1 + max
1i,jn
[a
ij
[
_
l

_
1 + max
1i,jn
[(a
1
)
ij
[
_
m
[det a[
1/p
.
6
To avoid confusion between the nite element (K, T, A), associated with an element domain
K in the triangulation, and the ane image of (

K,

T,

A) considered here, it would have been
better to use a new symbol (

K,

T,

A), say, instead of (K, T, A) and x instead of x, to denote the
ane image; but this would have complicated the notation. Here and in the next 17 lines we shall,
temporarily, adopt this sloppy notation. Thereafter, (K, T, A) will, again, signify a nite element
on an element domain K in the triangulation.
3.3. OPTIMAL ERROR BOUNDS IN THE H
1
() NORM REVISITED 83
Assuming that the subdivision T = K is regular in the sense that
> 0 K T h
K

K
( h
K
),
where h
K
is the diameter of K and
K
is the radius of the largest sphere (largest
circle for n = 2) contained in K, it is a straightforward exercise in geometry to show
that (

K) C

, where C

is a xed constant dependent on , but independent of

K T . Consequently,
[v J
K
v[
W
j
p
(K)
C(m, p, n, )h
mj
K
[v[
W
m
p
(K)
(3.4)
for each K T provided that T is a regular subdivision, 1 < p < , ml (n/p) >
0 and 0 j m; from this, and recalling the denition of the global interpolant of
v W
m
p
() it follows that
_

KT
h
(jm)p
K
[v J
h
v[
p
W
j
p
(K)
_
1/p
C(m, p, n, )[v[
W
m
p
()
. (3.5)
We shall also need the following somewhat cruder statement which is a straight-
forward consequence of (3.4); still supposing that the triangulation T is regular,
1 < p < , ml (n/p) > 0 and 0 j m, and v W
m
p
(), we have that
[v J
h
v[
W
j
p
()
C(m, p, n, )h
mj
[v[
W
m
p
()
, (3.6)
where h = max
KT
h
K
. These interpolation error estimates are of crucial importance
in nite element error analysis.
3.3 Optimal error bounds in the H
1
() norm
revisited
In this section we return to the discussion of error estimation in the H
1
() norm.
In Chapter 2 we showed, for the nite element approximation u
h
V
h
to the weak
solution u of the homogeneous Dirichlet boundary value problem for a second-order
elliptic equation, that
|u u
h
|
H
1
()

c
1
c
0
inf
v
h
V
h
|u v
h
|
H
1
()
. (3.7)
(c.f. Ceas lemma 5). Thus, restricting ourselves to the case of Poissons equation
and a continuous piecewise linear approximation u
h
dened on a uniform triangula-
tion of = (0, 1)
2
, we proved that, whenever u H
2
() H
1
0
(), we have
|u u
h
|
H
1
()
Ch[u[
H
2
()
.
84 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
Now, equipped with the interpolation error estimate (3.6) we can derive an anal-
ogous error bound in a more general setting; also, we can generalise to higher degree
piecewise polynomial approximations.
Suppose that R
n
and that it can be represented as a union of element
domains K such that conditions (i), (ii) and (iii) of Theorem 9 hold with p = 2.
Given such a triangulation T of we shall suppose that it is regular in the sense
introduced in the previous section and we put
V
h
= J
h
(H
m
() H
1
0
()).
Then,
inf
v
h
V
h
|u v
h
|
H
1
()
|u J
h
u|
H
1
()
C(m, n, )h
m1
[u[
H
m
()
.
Substituting this into (3.7), we arrive at the following error bound:
|u u
h
|
H
1
()
C(m, n, , c
1
, c
0
)h
m1
[u[
H
m
()
, (3.8)
provided that u H
m
() H
1
0
(). In particular, this will be the case if we use
continuous piecewise polynomials of degree m 1 on a regular triangulation of ,
with m = 2 corresponding to our earlier result with piecewise linear basis functions.
The inequality (3.8) is usually referred to as an optimal error bound, since for a
given m the smallest possible error that can be, in general, achieved in the H
1
()
norm is of size O(h
m1
).
3.4 Variational crimes
To conclude this chapter we briey comment on a further issue which arises in the
implementation of nite element methods. Let us consider the weak formulation of
the second-order elliptic partial dierential equation (1.5) on a bounded open set
R
n
, in the case of a homogeneous Dirichlet boundary condition (1.6):
nd u H
1
() such that a(u, v) = l(v) for all v H
1
0
(),
where, as before,
a(w, v) =
n

i,j=1
_

a
ij
(x)
w
x
i
v
x
j
dx
+
n

i=1
_

b
i
(x)
w
x
i
v dx +
_

c(x)wv dx (3.9)
and
l(v) =
_

f(x)v(x) dx. (3.10)


3.4. VARIATIONAL CRIMES 85
The associated nite element method is based on choosing a nite element sub-
space V
h
H
1
0
() consisting of continuous piecewise polynomials of a certain de-
gree dened on a subdivision of the computational domain , and considering the
approximate problem
nd u
h
V
h
such that a(u
h
, v
h
) = l(v
h
) for all v
h
V
h
.
Unfortunately, unless the coecients a
ij
, b
i
and c and the right-hand side f are
exceptionally simple functions, the integrals which appear in the denitions of a(, )
and l() will not be possible to evaluate exactly, and numerical integration rules (such
as the trapezium rule, Simpsons rule, Gauss-type rules and their multi-dimensional
counterparts) will have to be used to calculate a(, ) and l() approximately. Without
focusing on any particular quadrature rule, we attempt to analyse the eects of this
quadrature-induced perturbation on the accuracy of the exactly-integrated nite
element method. To keep the discussion simple, let us suppose that the bilinear form
a(, ) is still calculated exactly, but that l() has been replaced by an approximation
l
h
(), thereby leading to the following denition of u
h
:
nd u
h
V
h
such that a(u
h
, v
h
) = l
h
(v
h
) for all v
h
V
h
.
We recall that a key step in developing the nite element error analysis was the
presence of the Galerkin orthogonality property. With this new denition of u
h
,
however, we have that
a(u u
h
, v
h
) = a(u, v
h
) a(u
h
, v
h
) = l(v
h
) l
h
(v
h
) ,= 0, v
h
V
h
,
and Galerkin orthogonality no longer holds. We say that we have committed a
variational crime by replacing l() by l
h
(). We wish to study the extent to which
the accuracy of the basic nite element approximation is disturbed by this variational
crime.
Assuming that
c(x)
1
2
n

i=1
b
i
x
i
0,
we have that
a(v, v) c
0
|v|
2
H
1
()
,
with c
0
a positive constant (as in Section 1.2). Thus,
c
0
|u u
h
|
2
H
1
()
a(u u
h
, u u
h
)
= a(u u
h
, u v
h
) + a(u u
h
, v
h
u
h
)
= a(u u
h
, u v
h
) + l(v
h
u
h
) l
h
(v
h
u
h
)
c
1
|u u
h
|
H
1
()
|u v
h
|
H
1
()
+ sup
w
h
V
h
[l(w
h
) l
h
(w
h
)[
|w
h
|
H
1
()
|v
h
u
h
|
H
1
()
,
86 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
with c
1
a positive constant (as in Section 1.2). To simplify writing, we dene
|l l
h
|
1,h
= sup
w
h
V
h
[l(w
h
) l
h
(w
h
)[
|w
h
|
H
1
()
.
Hence,
c
0
|u u
h
|
2
H
1
()
c
1
|u u
h
|
H
1
()
|u v
h
|
H
1
()
+|l l
h
|
1,h
_
|u u
h
|
H
1
()
+|u v
h
|
H
1
()
_
=
_
c
1
|u v
h
|
H
1
()
+|l l
h
|
1,h
_
|u u
h
|
H
1
()
+|l l
h
|
1,h
|u v
h
|
H
1
()
. (3.11)
Now applying the elementary inequality
ab
1
2c
0
a
2
+
c
0
2
b
2
, a, b 0,
we have that
_
c
1
|u v
h
|
H
1
()
+|l l
h
|
1,h
_
|u u
h
|
H
1
()

1
2c
0
_
c
1
|u v
h
|
H
1
()
+|l l
h
|
1,h
_
2
+
c
0
2
|u u
h
|
2
H
1
()
.
Substituting this into (3.11) gives
c
2
0
|u u
h
|
2
H
1
()

_
c
1
|u v
h
|
H
1
()
+|l l
h
|
1,h
_
2
+2c
0
|l l
h
|
1,h
|u v
h
|
H
1
()
.
Noting that c
0
c
1
, this yields
c
2
0
|u u
h
|
2
H
1
()
2
_
c
1
|u v
h
|
H
1
()
+|l l
h
|
1,h
_
2
,
and therefore,
|u u
h
|
H
1
()

c
1

2
c
0
|u v
h
|
H
1
()
+

2
c
0
|l l
h
|
1,h
.
Equivalently,
|u u
h
|
H
1
()

c
1

2
c
0
|u v
h
|
H
1
()
+

2
c
0
sup
w
h
V
h
[l(w
h
) l
h
(w
h
)[
|w
h
|
H
1
()
. (3.12)
Since v
h
V
h
is arbitrary, it follows that we have proved the following perturbed
version of Ceas lemma:
|u u
h
|
H
1
()

c
1

2
c
0
min
v
h
V
h
|u v
h
|
H
1
()
+

2
c
0
sup
w
h
V
h
[l(w
h
) l
h
(w
h
)[
|w
h
|
H
1
()
.
3.4. VARIATIONAL CRIMES 87
The second term on the right-hand side of (3.12) quanties the extent to which the
accuracy of the exactly integrated nite element method is aected by the failure of
Galerkin orthogonality. Indeed, arguing in the same manner as in Section 3.3 will
lead to the error bound
|u u
h
|
H
1
()
C(m, n, , c
1
, c
0
)h
m1
[u[
H
m
()
+

2
c
0
sup
w
h
V
h
[l(w
h
) l
h
(w
h
)[
|w
h
|
H
1
()
.
Thus, in order to retain the accuracy of the exactly integrated method, the
numerical quadrature rule has to be selected so that the second term on the right
is also of size O(h
m1
); in the case of continuous piecewise linear basis functions
(m = 2) this means that the additional error should be at most O(h).
The situation when a(, ) is perturbed to a
h
(, ) is analysed in a similar manner.
We shall not discuss variational crimes which arise from replacing the computational
domain by a conveniently chosen close-by domain
h
; for the details of the
analysis the reader is referred to the books on the reading list and references therein.
88 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
Chapter 4
A posteriori error analysis by
duality
In this chapter we shall derive a computable bound on the global error and indicate
the implementation of this result into an adaptive algorithm with reliable error
control.
4.1 The one-dimensional model problem
In order to illuminate the key ideas and avoid technical diculties, we shall consider
the two-point boundary value problem
u

+ b(x)u

+ c(x)u = f(x), 0 < x < 1,


u(0) = 0, u(1) = 0,
where b W
1

(0, 1), c L

(0, 1) and f L
2
(0, 1). Letting
a(w, v) =
_
1
0
[w

(x)v

(x) + b(x)w

(x)v(x) + c(x)w(x)v(x)] dx
and
l(v) =
_
1
0
f(x)v(x) dx,
the weak formulation of this problem can be stated as follows:
nd u H
1
0
(0, 1) such that a(u, v) = l(v) for all v H
1
0
(0, 1).
Assuming that
c(x)
1
2
b

(x) 0, for x (0, 1), (4.1)


there exists a unique weak solution, u H
1
0
(0, 1).
The nite element approximation of this problem is constructed by considering a
(possibly non-uniform) subdivision of the interval [0, 1] by the points 0 = x
0
< x
1
<
89
90 CHAPTER 4. A POSTERIORI ERROR ANALYSIS BY DUALITY
. . . < x
N1
< x
N
= 1 and dening the nite element space V
h
H
1
0
(0, 1) consisting
of continuous piecewise polynomials of a certain degree on this subdivision. To keep
matters simple, let us suppose that V
h
consists of continuous piecewise linear func-
tions, as described in Chapter 2. The nite element approximation of the boundary
value problem is:
nd u
h
V
h
such that a(u
h
, v
h
) = l(v
h
) for all v
h
V
h
.
We let h
i
= x
i
x
i1
, i = 1, . . . , N, and put h = max
i
h
i
.
We wish to derive an a posteriori error bound; that is, we aim to quantify the
size of the global error u u
h
in terms of the mesh parameter h and the computed
solution u
h
(rather then the analytical solution u, as in an a priori error analysis).
To do so, we consider the following auxiliary boundary value problem
z

(b(x)z)

+ c(x)z = (u u
h
)(x), 0 < x < 1,
z(0) = 0, z(1) = 0,
called the dual or adjoint problem.
We begin our error analysis by noting that the denition of the dual problem and
a straightforward integration by parts yield (recall that (uu
h
)(0) = 0, (uu
h
)(1) =
0):
|u u
h
|
2
L
2
(0,1)
= (u u
h
, u u
h
) = (u u
h
, z

(bz)

+ cz)
= a(u u
h
, z).
By virtue of the Galerkin orthogonality property,
a(u u
h
, z
h
) = 0 z
h
V
h
.
In particular, choosing z
h
= J
h
z V
h
, the continuous piecewise linear interpolant of
the function z, associated with the subdivision 0 = x
0
< x
1
< . . . < x
N1
< x
N
= 1,
we have that
a(u u
h
, J
h
z) = 0.
Thus,
|u u
h
|
2
L
2
(0,1)
= a(u u
h
, z J
h
z) = a(u, z J
h
z) a(u
h
, z J
h
z)
= (f, z J
h
z) a(u
h
, z J
h
z). (4.2)
We observe that by this stage the right-hand side no longer involves the unknown
analytical solution u. Now,
a(u
h
, z J
h
z) =
N

i=1
_
x
i
x
i1
u

h
(x) (z J
h
z)

(x) dx
+
N

i=1
_
x
i
x
i1
b(x) u

h
(x) (z J
h
z)(x) dx
+
N

i=1
_
x
i
x
i1
c(x) u
h
(x) (z J
h
z)(x) dx.
4.1. THE ONE-DIMENSIONAL MODEL PROBLEM 91
Integrating by parts in each of the (N1) integrals in the rst sum on the right-hand
side, noting that (z J
h
z)(x
i
) = 0, i = 0, . . . , N, we deduce that
a(u
h
, z J
h
z) =
N

i=1
_
x
i
x
i1
[u

h
(x) + b(x)u

h
(x) + c(x)u
h
(x)] (z J
h
z)(x) dx.
Further
(f, z J
h
z) =
N

i=1
_
x
i
x
i1
f(x) (z J
h
z)(x) dx.
Substituting these two identities into (4.2), we deduce that
|u u
h
|
2
L
2
(0,1)
=
N

i=1
_
x
i
x
i1
R(u
h
)(x) (z J
h
z)(x) dx, (4.3)
where, for i = 1, . . . , N,
R(u
h
)(x) = f(x) + u

h
(x) b(x)u

h
(x) c(x)u
h
(x), x (x
i1
, x
i
).
The function R(u
h
) is called the nite element residual; it measures the extent
to which u
h
fails to satisfy the dierential equation u

+b(x)u

+c(x)u = f(x) on
the interval (0, 1). Now, applying the Cauchy-Schwarz inequality on the right-hand
side of (4.3) yields
|u u
h
|
2
L
2
(0,1)

N

i=1
|R(u
h
)|
L
2
(x
i1
,x
i
)
|z J
h
z|
L
2
(x
i1
,x
i
)
.
Recalling from the proof of Theorem 3 (with = z J
h
z and noting that

(x) =
z

(x) for all x in (x


i1
, x
i
), since J
h
z is a linear function on (x
i1
, x
i
), i = 1, . . . , N)
that
|z J
h
z|
L
2
(x
i1
,x
i
)

_
h
i

_
2
|z

|
L
2
(x
i1
,x
i
)
, i = 1, . . . , N,
we deduce that
|u u
h
|
2
L
2
(0,1)

1

2
N

i=1
h
2
i
|R(u
h
)|
L
2
(x
i1
,x
i
)
|z

|
L
2
(x
i1
,x
i
)
and consequently,
|u u
h
|
2
L
2
(0,1)

1

2
_
N

i=1
h
4
i
|R(u
h
)|
2
L
2
(x
i1
,x
i
)
_
1/2
|z

|
L
2
(0,1)
. (4.4)
The rest of the analysis is aimed at eliminating z

from the right-hand side of (4.4).


We recall that
z

= u
h
u (b z)

+ c z = u
h
u b z

+ (c b

) z,
92 CHAPTER 4. A POSTERIORI ERROR ANALYSIS BY DUALITY
and therefore,
|z

|
L
2
(0,1)
|u u
h
|
L
2
(0,1)
+|b|
L

(0,1)
|z

|
L
2
(0,1)
+|c b

|
L

(0,1)
|z|
L
2
(0,1)
. (4.5)
We shall show that both |z

|
L
2
(0,1)
and |z|
L
2
(0,1)
can be bounded in terms of |u
u
h
|
L
2
(0,1)
and then, by virtue of (4.5), we shall deduce that the same is true of
|z

|
L
2
(0,1)
. Let us observe that
(z

(bz)

+ cz, z) = (u u
h
, z).
Integrating by parts and noting that z(0) = 0 and z(1) = 0 yields
(z

(bz)

+ cz, z) = (z

, z

) + (bz, z

) + (cz, z)
= |z

|
2
L
2
(0,1)
+
1
2
_
1
0
b(x)[z
2
(x)]

dx +
_
1
0
c(x)[z(x)]
2
dx.
Integrating by parts, again, in the second term on the right gives
(z

(bz)

+ cz, z) = |z

|
2
L
2
(0,1)

1
2
_
1
0
b

(x)[z
2
(x)] dx +
_
1
0
c(x)[z(x)]
2
dx.
Hence,
|z

|
2
L
2
(0,1)
+
_
1
0
_
c(x)
1
2
b

(x)
_
[z(x)]
2
dx = (u u
h
, z),
and thereby, noting (4.1),
|z

|
2
(u u
h
, z) |u u
h
|
L
2
(0,1)
|z|
L
2
(0,1)
. (4.6)
By the Poincare-Friedrichs inequality,
|z|
2
L
2
(0,1)

1
2
|z

|
2
L
2
(0,1)
.
Thus, (4.6) gives
|z|
L
2
(0,1)

1
2
|u u
h
|
L
2
(0,1)
. (4.7)
Inserting this into the right-hand side of (4.6) yields
|z

|
L
2
(0,1)

1

2
|u u
h
|
L
2
(0,1)
. (4.8)
Now we substitute (4.7) and (4.8) into (4.5) to deduce that
|z

|
L
2
(0,1)
K|u u
h
|
L
2
(0,1)
, (4.9)
Where
K = 1 +
1

2
|b|
L

(0,1)
+
1
2
|c b

|
L

(0,1)
.
4.2. AN ADAPTIVE ALGORITHM 93
It is important to note here that K
0
involves only known quantities, namely the
coecients in the dierential equation under consideration, and therefore it can be
computed without diculty. Inserting (4.9) into (4.4), we arrive at our nal result,
the computable a posteriori error bound,
|u u
h
|
L
2
(0,1)
K
0
_
N

i=1
h
4
i
|R(u
h
)|
2
L
2
(x
i1
,x
i
)
_
1/2
, (4.10)
where K
0
= K/
2
.
The name a posteriori stems from the fact that (4.10) can only be employed
to quantify the size of the approximation error that has been committed in the
course of the computation after u
h
has been computed. In the next section we shall
describe the construction of an adaptive mesh renement algorithm based on the
bound (4.10).
4.2 An adaptive algorithm
Suppose that TOL is a prescribed tolerance and that our aim is to compute a nite
element approximation u
h
to the unknown solution u (with the same denition of u
and u
h
as in the previous section) so that
|u u
h
|
L
2
(0,1)
TOL.
We shall use the a posteriori error bound (4.10) to achieve this goal by succes-
sively rening the subdivision, and computing a succession of numerical solutions
u
h
on these subdivisions, until the stopping criterion
K
0
_
N

i=1
h
4
i
|R(u
h
)|
2
L
2
(x
i1
,x
i
)
_
1/2
TOL
is satised. The algorithm proceeds as follows:
1. Choose an initial subdivision
T
0
: 0 = x
(0)
0
< x
(0)
1
< . . . < x
(0)
N
0
1
< x
(0)
N
0
= 1
of the interval [0, 1], with h
(0)
i
= x
(0)
i
x
(0)
i1
for i = 1, . . . , N
0
, and h
(0)
=
max
i
h
(0)
i
, and consider the associated nite element space V
h
(0) (of dimension
N
0
1).
2. Compute the corresponding solution u
h
(0) V
h
(0) .
3. Given a computed solution u
h
(m) V
h
(m) for some m 0, dened on a subdi-
vision T
m
, stop if
K
0
_
N
m

i=1
_
h
(m)
i
_
4
|R(u
h
(m) )|
2
L
2
(x
(m)
i1
,x
(m)
i
)
_
1/2
TOL. (4.11)
94 CHAPTER 4. A POSTERIORI ERROR ANALYSIS BY DUALITY
4. If not, then determine a new subdivision
T
m+1
: 0 = x
(m+1)
0
< x
(m+1)
1
< . . . < x
(m+1)
N
m+1
1
< x
(m+1)
N
m+1
= 1
of the interval [0, 1], with h
(m+1)
i
= x
(m+1)
i
x
(m+1)
i1
for i = 1, . . . , N
m+1
and
h
(m+1)
= max
i
h
(m+1)
i
, and an associated nite element space V
h
(m+1) (of di-
mension N
m+1
1), with h
(m+1)
as large as possible (and consequently N
m+1
as small as possible), such that
K
0
_
N
m+1

i=1
_
h
(m+1)
i
_
4
|R(u
h
(m) )|
2
L
2
(x
(m+1)
i1
,x
(m+1)
i
)
_
1/2
= TOL, (4.12)
and continue.
Here (4.11) is the stopping criterion and (4.12) is the mesh modication strategy.
According to the a posteriori error bound (4.10), when the algorithm terminates the
global error |uu
h
|
L
2
(0,1)
is controlled to within the prescribed tolerance TOL. The
relation (4.12) denes the new mesh-size by maximality. The natural condition for
maximality is equidistribution; this means that the residual contributions from
individual elements in the subdivision are required to be equal:
_
h
(m+1)
i
_
4
|R(u
h
(m) )|
2
L
2
(x
(m+1)
i1
,x
(m+1)
i
)
=
TOL
2
K
2
0
N
m+1
for each i = 1, . . . , N
m+1
; the implementation can be simplied by replacing N
m+1
on the right-hand side by N
m
. Then, we have a simple formula for h
(m+1)
i
:
h
(m+1)
i
=
_
_
TOL
2
K
2
0
N
m
|R(u
h
(m) )|
2
L
2
(x
(m+1)
i1
,x
(m+1)
i
)
_
_
1/4
, i = 1, . . . , N
m+1
,
from which the h
(m+1)
i
can be found by treating this as an equation in h
(m+1)
i
,
and solving it numerically, for m and i xed, by some root-nding algorithm (e.g.
successive bisection or xed-point iteration), starting from i = 1.
Reliability means that the computational error is controlled in a given norm
on a given tolerance level. Thus what we have described above is a reliable com-
putational algorithm. Eciency means that the computational eort required to
achieve reliability is minimal. It is unclear from the present discussion whether the
adaptive algorithm described above is ecient in this sense: although we have min-
imised the computational eort required to ensure that the right-hand side in the
error bound (4.10) is below the given tolerance, the extent to which this implies
that we have also minimised the amount of computational eort required to ensure
that the left-hand side in (4.10) is less than TOL depends on the sharpness of the
inequality (4.10), and this will vary from case to case, depending very much on the
choice of the functions b, c and f.
Chapter 5
Evolution problems
In previous chapters we considered the nite element approximation of elliptic
boundary value problems. This chapter is devoted to nite element methods for
time-dependent problems; in particular, we shall be concerned with the nite ele-
ment approximation of parabolic equations. Hyperbolic equations will not be dis-
cussed in these notes.
5.1 The parabolic model problem
Let be a bounded open set in R
n
, n 1, with boundary = , and let T > 0.
In Q = (0, T], we consider the initial boundary value problem for the unknown
function u(x, t), x , t (0, T] :
u
t

n

i,j=1

x
j
(a
ij
(x, t)
u
x
i
) +
n

i=1
b
i
(x, t)
u
x
i
+ c(x, t)u = f(x, t),
x , t (0, T], (5.1)
u(x, t) = 0, x , t [0, T], (5.2)
u(x, 0) = u
0
(x), x

. (5.3)
Suppose that u
0
L
2
(), and that there exists a positive constant c such that
n

i,j=1
a
ij
(x, t)
i

j
c
n

i=1

2
i
,
= (
1
, . . . ,
n
) R
n
, x

, t [0, T]. (5.4)
We shall also assume that
a
ij
L

(Q), b
i
W
1

(Q), i, j = 1, . . . , n,
c L

(Q), f L
2
(Q),
95
96 CHAPTER 5. EVOLUTION PROBLEMS
and that
c(x, t)
1
2
n

i=1
b
i
x
i
(x, t) 0, (x, t)

Q, (5.5)
as in the elliptic case.
A partial dierential equation of the form (5.1) is called a parabolic equation (of
second order). Simple examples of parabolic equations are the heat equation
u
t
= u
and the unsteady advection-diusion equation
u
t
u +
n

i=1
b
i
u
x
i
= 0.
The proof of the existence of a unique solution to a parabolic initial boundary
value problem is more technical than for an elliptic boundary value problem and
it is omitted here. Instead, we shall simply assume that (5.1)(5.3) has a unique
solution and investigate its decay in t (t typically signies time), and discuss the
question of continuous dependence of the solution on the initial datum u
0
and the
forcing function f.
We recall that, for v, w L
2
(), the inner product (u, v) and the norm |v|
L
2
()
are dened by
(v, w) =
_

v(x)w(x) dx,
|v|
L
2
()
= (v, v)
1/2
.
Taking the inner product of (5.1) with u, noting that u(x, t) = 0, x , integrating
by parts, and applying (5.4) and (5.5), we get
_
u
t
(, t), u(, t)
_
+ c
n

i=1
|
u
x
i
(, t)|
2
L
2
()
(f(, t), u(, t)).
Noting that
_
u
t
(, t), u(, t)
_
=
1
2
d
dt
|u(, t)|
2
L
2
()
,
and using the Poincare-Friedrichs inequality (1.2), we obtain
1
2
d
dt
|u(, t)|
2
L
2
()
+
c
c

|u(, t)|
2
L
2
()
(f(, t), u(, t)).
5.1. THE PARABOLIC MODEL PROBLEM 97
Let K = c/c

; then, by the Cauchy-Schwarz inequality,


1
2
d
dt
|u(, t)|
2
L
2
()
+ K|u(, t)|
2
L
2
()
|f(, t)|
L
2
()
|u(, t)|
L
2
()

1
2K
|f(, t)|
2
L
2
()
+
K
2
|u(, t)|
2
L
2
()
.
Thence,
d
dt
|u(, t)|
2
L
2
()
+ K|u(, t)|
2
L
2
()

1
K
|f(, t)|
2
L
2
()
.
Multiplying both sides by e
Kt
,
d
dt
_
e
Kt
|u(, t)|
2
L
2
()
_

e
Kt
K
|f(, t)|
2
L
2
()
.
Integrating from 0 to t,
e
Kt
|u(, t)|
2
L
2
()
|u
0
|
2
L
2
()

1
K
_
t
0
e
K
|f(, )|
2
L
2
()
d.
Hence
|u(, t)|
2
L
2
()
e
Kt
|u
0
|
2
L
2
()
+
1
K
_
t
0
e
K(t)
|f(, )|
2
L
2
()
d. (5.6)
Assuming that (5.1)(5.3) has a solution, (5.6) implies that the solution is unique.
Indeed, if u
1
and u
2
are solutions to (5.1)(5.3), then u = u
1
u
2
satises (5.1)(5.3)
with f 0 and u
0
0; therefore, by (5.6), u 0, i.e. u
1
u
2
.
Let us also look at the special case when f 0 in (5.1). This corresponds to
considering the evolution of the solution from the initial datum u
0
in the absence of
external forces. In this case (5.6) yields
|u(, t)|
2
L
2
()
e
Kt
|u
0
|
2
L
2
()
, t 0; (5.7)
in physical terms, the energy
1
2
|u(, t)|
2
L
2
()
dissipates exponentially. Since K =
c/c

, we have
|u(, t)|
2
L
2
()
e
ct/c

|u
0
|
2
L
2
()
, t 0, (5.8)
and we deduce that the rate of dissipation depends on the lower bound, c, on the
diusion coecients a
ij
(i.e. the smaller c, the slower the decay of the energy).
Conservation of energy would correspond to
|u(, t)|
2
L
2
()
= |u
0
|
2
L
2
()
;
this will only occur by formally setting c = 0, however since c > 0 by hypothesis,
conservation of energy will not be observed for a physical process modelled by a
second-order parabolic equation.
In the next section we consider some simple nite element methods for the nu-
merical solution of parabolic initial boundary value problems. In order to simplify
the presentation, we restrict ourselves to the heat equation in one space dimension,
but the analysis that we shall present also applies in the general setting.
98 CHAPTER 5. EVOLUTION PROBLEMS
5.2 Forward and backward Euler schemes
We consider the following simple model problem for the heat equation in one space
dimension. Let Q = (0, T], where = (0, 1), T > 0;
nd u(x, t) such that
u
t
=

2
u
x
2
+ f(x, t), x (0, 1), t (0, T],
u(0, t) = 0, u(1, t) = 0, t [0, T], (5.9)
u(x, 0) = u
0
(x), x [0, 1].
We describe two schemes for the numerical solution of (5.9). They both use the
same discretisation in the x variable but while the rst scheme (called the forward
Euler scheme) employs a forward divided dierence in t to approximate u/t, the
second (called the backward Euler scheme) uses a backward dierence in t.
The forward Euler scheme. We begin by constructing a mesh on

Q = [0, 1]
[0, T]. Let h = 1/N be the mesh-size in the x-direction and let t = T/M be the
mesh-size in the t-direction; here N and M are two integers, N 2, M 1. We
dene the uniform mesh

Q
t
h
on

Q by

Q
t
h
= (x
j
, t
m
) : x
j
= jh, 0 j N; t
m
= m t, 0 m M.
Let V
h
H
1
0
(0, 1) denote the set of all continuous piecewise linear functions dened
on the x-mesh
0 = x
0
< x
1
< . . . < x
N1
< x
N
= 1
which vanish at the end-points, x = 0 and x = 1.
We approximate (5.9) by the nite element method, referred to as the forward
Euler scheme:
nd u
m
h
V
h
, 0 m M, such that
_
u
m+1
h
u
m
h
t
, v
h
_
+ a(u
m
h
, v
h
) = (f(, t
m
), v
h
) v
h
V
h
,
(5.10)
(u
0
h
u
0
, v
h
) = 0 v
h
V
h
,
where u
m
h
represents the approximation of u(, t
m
), and a(, ) is dened by
a(w, v) =
_
1
0
w

(x)v

(x) dx.
Clearly, (5.10) can be rewritten as follows:
(u
m+1
h
, v
h
) = (u
m
h
, v
h
) t a(u
m
h
, v
h
) + t (f(, t
m
), v
h
)
v
h
V
h
, 0 m M 1,
5.2. FORWARD AND BACKWARD EULER SCHEMES 99
with
(u
0
h
, v
h
) = (u
0
, v
h
) v
h
V
h
.
Thus, given u
m
h
, to nd u
m+1
h
at time level t
m+1
we have to solve a system of linear
equations with symmetric positive denite matrix M of size (N 1) (N 1), with
entries (
i
,
j
) where
i
denotes the one-dimensional piecewise linear nite element
basis function associated with the x-mesh point x
i
; the same matrix arises when
determining u
0
h
. It is a simple matter to show that this matrix is tridiagonal and
has the form
M =
h
6
_
_
_
_
_
_
4 1 0 0 . . . 0
1 4 1 0 . . . 0
0 1 4 1 . . . 0
. . . . . . . . . . . . . . . . . .
0 0 . . . 0 1 4
_
_
_
_
_
_
.
The matrix M is usually referred to as the mass matrix.
The backward Euler scheme. Alternatively, one can approximate the time
derivative by a backward dierence, which gives rise to the following backward
Euler scheme:
nd u
m
h
V
h
, 0 m M, such that
_
u
m+1
h
u
m
h
t
, v
h
_
+ a(u
m+1
h
, v
h
) = (f(, t
m+1
), v
h
) v
h
V
h
,
(5.11)
(u
0
h
u
0
, v
h
) = 0 v
h
V
h
,
where u
m
h
represents the approximation of u(, t
m
). Equivalently, (5.11) can be
written
(u
m+1
h
, v
h
) + t a(u
m+1
h
, v
h
) = (u
m
h
, v
h
) + t (f(, t
m+1
), v
h
)
v
h
V
h
, 0 m M 1,
with
(u
0
h
, v
h
) = (u
0
, v
h
) v
h
V
h
.
Thus, given u
m
h
, to nd u
m+1
h
at time level t
m+1
we have to solve a system of linear
equations with symmetric positive denite matrix A of size (N 1) (N 1), with
entries (
i
,
j
) + t (

i
,

j
) where
i
denotes the one-dimensional piecewise linear
nite element basis function associated with the x-mesh point x
i
; nding u
0
h
still
only involves inverting the mass matrix M. It is clear that A = M + t K, where
K =
1
h
_
_
_
_
_
_
2 1 0 0 . . . 0
1 2 1 0 . . . 0
0 1 2 1 . . . 0
. . . . . . . . . . . . . . . . . .
0 0 . . . 0 1 2
_
_
_
_
_
_
is the so-called stiness matrix.
100 CHAPTER 5. EVOLUTION PROBLEMS
5.3 Stability of -schemes
We shall study the stability of the schemes (5.10) and (5.11) simultaneously, by
embedding them into a one-parameter family of nite element schemes:
nd u
m
h
V
h
, 0 m M, such that
_
u
m+1
h
u
m
h
t
, v
h
_
+ a(u
m+
h
, v
h
) = (f
m+
, v
h
) v
h
V
h
(5.12)
(u
0
h
u
0
, v
h
) = 0 v
h
V
h
,
where 0 1, and for the sake of notational simplicity, we wrote
f
m+
(x) = f(x, t
m+1
) + (1 )f(x, t
m
)
and
u
m+
h
(x) = u
m+1
h
(x) + (1 )u
m
h
(x).
For = 0 this gives the forward Euler scheme, for = 1 the backward Euler scheme.
The method corresponding to =
1
2
is known as the Crank-Nicolson scheme.
Recall that
(w, v) =
_
1
0
w(x)v(x) dx,
|v|
L
2
()
= (v, v)
1/2
.
Taking the inner product of (5.12) with u
m+
h
we get
_
u
m+1
h
u
m
h
t
, u
m+
h
_
+ a(u
m+
h
, u
m+
h
) = (f
m+
, u
m+
h
).
Equivalently,
_
u
m+1
h
u
m
h
t
, u
m+
h
_
+[u
m+
h
[
2
H
1
()
= (f
m+
, u
m+
h
).
Since
u
m+
h
= t
_

1
2
_
u
m+1
h
u
m
h
t
+
u
m+1
h
+ u
m
h
2
,
it follows that
t
_

1
2
_
|
u
m+1
h
u
m
h
t
|
2
L
2
()
+
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
2t
+[u
m+
h
[
2
H
1
()
= (f
m+
, u
m+
h
). (5.13)
5.3. STABILITY OF -SCHEMES 101
Suppose that [1/2, 1]; then 1/2 0, and therefore
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
2t
+[u
m+
h
[
2
H
1
()
(f
m+
, u
m+
h
)
|f
m+
|
L
2
()
|u
m+
h
|
L
2
()
.
According to the Poincare-Friedrichs inequality,
|u
m+
h
|
2
L
2
()

1
2
[u
m+
h
[
2
H
1
()
.
Thus
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
2t
+ 2|u
m+
h
|
2
L
2
()

1
2
|f
m+
|
2
L
2
()
+
1
2
|u
m+
h
|
2
L
2
()
,
so that
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
+ t|f
m+
|
2
L
2
()
.
Summing through m, m = 0, . . . , k, we get that
|u
k
h
|
2
L
2
()
|u
0
h
|
2
L
2
()
+
k1

m=0
t|f
m+
|
2
L
2
()
, (5.14)
for all k, 1 k M.
The inequality (5.14) can be thought of as the discrete version of (5.6). If follows
from (5.14) that
max
1kM
|u
k
h
|
2
L
2
()
|u
0
h
|
2
L
2
()
+
M1

m=0
t|f
m+
|
2
L
2
()
,
i.e.
max
1kM
|u
k
h
|
L
2
()

_
|u
0
h
|
2
L
2
()
+
M1

m=0
t|f
m+
|
2
L
2
()
_
1/2
, (5.15)
which expresses the continuous dependence of the solution to the nite element
scheme (5.12) on the initial data and the right-hand side. This property is called
stability.
Thus we have proved that for [1/2, 1], the scheme (5.12) is stable, without
any limitations on the time step in terms of h. In other words, the scheme (5.12) is
unconditionally stable for [1/2, 1].
Now let us consider the case [0, 1/2). According to (5.13),
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
2t
+[u
m+
h
[
2
H
1
()
= t(
1
2
)|
u
m+1
h
u
m
h
t
|
2
L
2
()
+ (f
m+
, u
m+
). (5.16)
102 CHAPTER 5. EVOLUTION PROBLEMS
Recalling (5.12) with v
h
= (u
m+1
h
u
m
h
)/t, we have that
|
u
m+1
h
u
m
h
t
|
2
L
2
()
=
_
f
m+
,
u
m+1
h
u
m
h
t
_
a
_
u
m+
h
,
u
m+1
h
u
m
h
t
_
.
Therefore,
|
u
m+1
h
u
m
h
t
|
2
L
2
()
|f
m+
|
L
2
()
|
u
m+1
h
u
m
h
t
|
L
2
()
+[u
m+
h
[
H
1
()
[
u
m+1
h
u
m
h
t
[
H
1
()
. (5.17)
Next we shall prove that, for each w
h
V
h
,
[w
h
[
H
1
()

12
h
|w
h
|
L
2
()
. (5.18)
We shall then use this inequality to estimate the terms appearing on the right-hand
side of (5.17). Let W
i
denote the value of the piecewise linear function w
h
V
h
at
the mesh-point x
i
, i = 0, . . . , N, and note that W
0
= W
N
= 0. A simple calculation
reveals that
[w
h
[
2
H
1
()
=
N

i=1
h[
W
i
W
i1
h
[
2

4
h
2
N1

i=1
h[W
i
[
2
. (5.19)
On the other hand,
|w
h
|
2
L
2
()
=
h
6
N1

i=1
(W
i
W
i1
+ 4W
2
i
+ W
i
W
i+1
)

h
6
N1

i=1
_

1
2
W
2
i

1
2
W
2
i1
+ 4W
2
i

1
2
W
2
i

1
2
W
2
i+1
_

1
3
N1

i=1
h[W
i
[
2
. (5.20)
From (5.20) and (5.19) we deduce (5.18).
Now, equipped with the inequality (5.18), we continue the stability analysis.
Applying (5.18) with w
h
= (u
m+1
h
u
m
h
)/t, we deduce that
|
u
m+1
h
u
m
h
t
|
2
L
2
()
|f
m+
|
L
2
()
|
u
m+1
h
u
m
h
t
|
L
2
()
+

12
h
[u
m+
h
[
H
1
()
|
u
m+1
h
u
m
h
t
|
L
2
()
and hence
|
u
m+1
h
u
m
h
t
|
L
2
()
|f
m+
|
L
2
()
+

12
h
[u
m+
h
[
H
1
()
(5.21)
5.3. STABILITY OF -SCHEMES 103
By (5.21), for any (0, 1),
|
u
m+1
h
u
m
h
t
|
2
L
2
()

_

12
h
[u
m+
h
[
H
1
()
+|f
m+
|
L
2
()
_
2
(1 + )
12
h
2
[u
m+
h
[
2
H
1
()
+ (1 +
1
)|f
m+
|
2
L
2
()
,
where the inequality (a + b)
2
(1 + )a
2
+ (1 +
1

)b
2
, a, b 0, > 0, has been
applied. Substituting into (5.16),
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
2t
+
_
1 t(
1
2
)
12(1 + )
h
2
_
[u
m+
h
[
2
H
1
()
|f
m+
|
L
2
()
|u
m+
h
|
L
2
()
+ t(
1
2
)(1 +
1
)|f
m+
|
2
L
2
()
.
(5.22)
According to the Poincare-Friedrichs inequality,
|u
m+
h
|
2
L
2
()

1
2
[u
m+
h
[
2
H
1
()
,
and therefore,
|f
m+
|
L
2
()
|u
m+
h
|
L
2
()

1
8
2
|f
m+
|
2
L
2
()
+ 2
2
|u
m+
h
|
2
L
2
()

1
8
2
|f
m+
|
2
L
2
()
+
2
[u
m+
h
[
2
H
1
()
. (5.23)
Substituting (5.23) into (5.22),
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
2t
+
_
1 t
6(1 2)(1 + )
h
2

2
_
[u
m+
h
[
2
H
1
()

1
8
2
|f
m+
|
2
L
2
()
+ t(
1
2
)(1 +
1
)|f
m+
|
2
L
2
()
.
Let us suppose that
t
h
2
6(1 2)
(1 ), [0, 1/2),
where is a xed real number, (0, 1). Then
1 t
6(1 2)(1 + )
h
2

2
0,
so that
|u
m+1
h
|
2
L
2
()
|u
m
h
|
2
L
2
()
+
t
4
2
|f
m+
|
2
L
2
()
+ t
2
(1 2)(1 +
1
)|f
m+
|
2
L
2
()
.
104 CHAPTER 5. EVOLUTION PROBLEMS
Letting c

= 1/(4
2
) + t(1 2)(1 +
1
), upon summation through all m this
implies that
max
1kM
|u
k
h
|
2
L
2
()
|u
0
h
|
2
L
2
()
+ c

M1

m=0
t|f
m+
|
2
L
2
()
.
Taking the square root of both sides, we deduce that for [0, 1/2) the scheme
(5.12) is conditionally stable in the sense that
max
1kM
|u
k
h
|
h

_
|u
0
h
|
2
L
2
()
+ c

M1

m=0
t|f
m+
|
2
L
2
()
_
1/2
, (5.24)
provided that
t
h
2
6(1 2)
(1 ), 0 < < 1. (5.25)
To summarise: when [1/2, 1], the method (5.12) is unconditionally stable. In
particular, the backward Euler scheme, corresponding to = 1, and the Crank-
Nicolson scheme, corresponding to = 1/2, are unconditionally stable, and (5.15)
holds. When [0, 1/2), the scheme (5.12) is conditionally stable, subject to the
time step limitation (5.25). The forward Euler scheme, corresponding to = 0, is
only conditionally stable.
5.4 Error analysis in the L
2
norm
In this section we investigate the accuracy of the nite element method (5.12) for
the numerical solution of the initial boundary value problem (5.9). For simplicity,
we shall restrict ourselves to the backward Euler scheme ( = 1); for other values of
[0, 1] the analysis is completely analogous.
We decompose the global error e
h
as follows:
e
m
h
= u(, t
m
) u
m
h
=
m
+
m
,
where

m
= u(, t
m
) Pu(, t
m
),
m
= Pu(, t
m
) u
m
h
,
and for t [0, T], Pu(, t) V
h
denotes the Dirichlet projection of u(, t) dened
by
a(Pu(, t), v
h
) = a(u(, t), v
h
) v
h
V
h
.
The existence and uniqueness of Pu(, t) V
h
follows by the Lax-Milgram theorem.
Hence,
a(
m
, v
h
) = 0 v
h
V
h
,
5.4. ERROR ANALYSIS IN THE L
2
NORM 105
and therefore, by Ceas lemma,
[
m
[
H
1
()
[u(, t
m
) J
h
u(, t
m
)[
H
1
()

h

[u(, t
m
)[
H
2
()
,
where J
h
u(, t
m
) V
h
denotes the continuous piecewise linear interpolant of u(, t
m
)
from V
h
. By the Aubin-Nitsche duality argument,
|
m
|
L
2
()

h
2

2
[u(, t
m
)[
H
2
()
. (5.26)
Since also,
a
_

m+1

m
t
, v
h
_
= 0 v
h
V
h
,
by an identical argument we deduce that
|

m+1

m
t
|
L
2
()

h
2

u(, t
m+1
) u(, t
m
)
t

H
2
()
. (5.27)
For m = 0,
(
0
, v
h
) = (e
0
h
, v
h
) (
0
, v
h
) = (
0
, v
h
)
and therefore, choosing v
h
=
0
and applying the Cauchy-Schwarz inequality on the
right,
|
0
|
L
2
()
|
0
|
L
2
()

h
2

2
[u
0
[
H
2
()
. (5.28)
It is easily seen that
m
V
h
satises the following identity:
_

m+1

m
t
, v
h
_
+ a(
m+1
, v
h
)
=
_
u(, t
m+1
) u(, t
m
)
t

u
t
(, t
m+1
)

m+1

m
t
, v
h
_
According to the stability result proved earlier on,
max
1mM
|
m
|
L
2
()

_
|
0
|
2
L
2
()
+
M1

m=0
t|
m+1
|
2
L
2
()
_
1/2
, (5.29)
where

m+1
=
u(, t
m+1
) u(, t
m
)
t

u
t
(, t
m+1
)

m+1

m
t
.
By (5.28),
|
0
|
L
2
()

h
2

2
[u
0
[
H
2
()
. (5.30)
106 CHAPTER 5. EVOLUTION PROBLEMS
It remains to estimate |
m+1
|
L
2
()
. Now
|
m+1
|
L
2
()
|
u(, t
m+1
) u(, t
m
)
t

u
t
(, t
m+1
)|
L
2
()
+|

m+1

m
t
|
L
2
()
I + II. (5.31)
For term I, Taylors formula with integral remainder yields that
u(x, t
m+1
) u(x, t
m
)
t

u
t
(x, t
m+1
) =
1
t
_
t
m+1
t
m
(t
m+1
t)

2
u
t
2
(x, t) dt,
and therefore
I

t
_
_
t
m+1
t
m
|

2
u
t
2
(, t)|
2
L
2
()
dt
_
1/2
.
Further, by (5.27),
II
h
2

u(, t
m+1
) u(, t
m
)
t

H
2
()
=
h
2

1
t
_
t
m+1
t
m
u
t
(, t) dt

H
2
()

h
2

t
_
_
t
m+1
t
m

u
t
(, t)

2
H
2
()
dt
_
1/2
.
Substituting the bounds on terms I and II onto (5.31) and inserting the resulting
inequality and (5.30) into (5.29), we obtain the following error bound:
max
1mM
|
m
|
L
2
()
C
1
(h
2
+ t), (1/2, 1], (5.32)
where C
1
is a positive constant, independent of h and t, and depending only on
norms of the analytical solution u. But,
max
1mM
|u(, t
m
) u
m
h
|
L
2
()
max
1mM
|
m
|
L
2
()
+ max
1mM
|
m
|
L
2
()
.
Thus, by (5.32) and (5.26), we deduce that
max
1mM
|u(, t
m
) u
m
h
|
L
2
()
C
2
(h
2
+ t),
where C
2
is a positive constant independent of h and t.
The Crank-Nicolson scheme ( = 1/2) can be shown to converge in the norm
| |
L
2
()
unconditionally, with error O(h
2
+(t)
2
). For (1/2, 1] the scheme con-
verges unconditionally with error O(h
2
+t). For [0, 1/2) the scheme converges
with error O(h
2
+t), but only conditionally. The stability and convergence results
presented here can be extended to parabolic equations in more than one space di-
mension, but the exposition of that theory, while very similar to the one-dimensional
case, is beyond the scope of these notes.

You might also like