This action might not be possible to undo. Are you sure you want to continue?
Finite Element Methods for Partial
Differential Equations
Endre S¨ uli
Mathematical Institute
University of Oxford
2011
2
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 selfadjoint elliptic problem . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3 Calculation and assembly of stiﬀness matrix . . . . . . . . . . . . . . . . . . 35
2.4 Galerkin orthogonality; C´ea’s lemma . . . . . . . . . . . . . . . . . . . . . . 40
2.5 Optimal error bound in the energy norm . . . . . . . . . . . . . . . . . . . . 45
2.6 Superapproximation in meshdependent norms . . . . . . . . . . . . . . . . 56
3 Piecewise polynomial approximation 65
3.1 Construction of ﬁnite element spaces . . . . . . . . . . . . . . . . . . . . . . 65
3.1.1 The ﬁnite element . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.1.2 Examples of triangular ﬁnite 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 BrambleHilbert lemma . . . . . . . . . . . . . . . . . . . . . . . 75
3.2.2 Error bounds on the interpolation error . . . . . . . . . . . . . . . . 80
3.3 Optimal error bounds in the H
1
(Ω) norm – revisited . . . . . . . . . . . . . 83
3.4 Variational crimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4 A posteriori error analysis by duality 89
4.1 The onedimensional 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 θschemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 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 diﬀerential 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 eﬃcient ﬁnite 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 Cea’s lemma. Piecewise
polynomial approximation in Sobolev spaces. The BrambleHilbert lemma. Optimal
error bounds in the energy norm. Variational crimes.
The AubinNitsche duality argument. Superapproximation properties in mesh
dependent norms. A posteriori error analysis by duality: reliability, eﬃciency and
adaptivity.
Finite element approximation of initial boundary value problems. Energy dissi
pation, conservation and stability. Analysis of ﬁnite element methods for evolution
problems.
Reading List
1. S.C. Brenner & L.R. Scott, The Mathematical Theory of Finite Element Methods.
Springer, 2nd edition, 2002. [Chapters 0,1,2,3; Chapter 4: Secs. 4.1–4.4, Chapter 5:
Secs. 5.1–5.7].
2. H. Elman, D. Silvester & A. Wathen, Finite Elements and Fast Iterative Solvers.
OUP, 2005. [Mainly Chapters 1 and 5].
3. C. Johnson, Numerical Solution of Partial Diﬀerential Equations by the Finite El
ement Method. CUP, 1990. [Chapters 1–4; Chapter 8: Secs. 8.1–8.4.2; Chapter 9:
Secs. 9.1–9.5].
4. K.W. Morton, Lecture Notes on Finite Element Methods. Oxford University Com
puting Laboratory, 1991.
Chapter 1
Introduction
Partial diﬀerential equations arise in the mathematical modelling of many phys
ical, chemical and biological phenomena and many diverse subject areas such as
ﬂuid dynamics, electromagnetism, material science, astrophysics, economy, ﬁnancial
modelling, etc. Very frequently the equations under consideration are so compli
cated that ﬁnding 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 diﬀerential equations: ﬁnite 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 1950’s the method was rediscovered by engineers, but the mathematical
analysis of ﬁnite element approximations began much later, in the 1960’s, the ﬁrst
important results being due to Miloˇs Zl´amal
2
in 1968. Since then ﬁnite element
methods have been developed into one of the most general and powerful class of
techniques for the numerical solution of partial diﬀerential equations and are widely
used in engineering design and analysis.
In these notes we shall be concerned with the mathematical aspects of ﬁnite
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 diﬀerential 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. 1–23 (1943)
2
M. Zl´ amal: On the ﬁnite element method. Numerische Mathematik, 12, pp. 394–402 (1968)
5
6 CHAPTER 1. INTRODUCTION
1.1 Elements of function spaces
As will become apparent in subsequent chapters, the accuracy of ﬁnite element ap
proximations to partial diﬀerential 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 speciﬁc diﬀerentia
bility and integrability properties, called function spaces. In this section we present
a brief overview of basic deﬁnitions 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 realvalued.
1.1.1 Spaces of continuous functions
In this section, we describe some simple function spaces which consist of continuously
diﬀerentiable functions. For the sake of notational convenience, we introduce the
concept of a multiindex.
Let N denote the set of nonnegative integers. An ntuple
α = (α
1
, . . . , α
n
) ∈ N
n
is called a multi–index. The nonnegative integer [α[ := α
1
+ . . . + α
n
is referred
to as the length of the multi–index α = (α
1
, . . . , α
n
). We denote (0, . . . , 0) by 0;
clearly [0[ = 0. Let
D
α
=
_
∂
∂x
1
_
α
1
. . .
_
∂
∂x
n
_
α
n
=
∂
α
∂x
α
1
1
. . . ∂x
α
n
n
.
Example 1 Suppose that n = 3, and α = (α
1
, α
2
, α
3
), α
j
∈ N, j = 1, 2, 3. Then
for u, a function of three variables x
1
, x
2
, x
3
,
α=3
D
α
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 multiindex notation: instead of labori
ously writing out in detail the ten terms on the righthand side of the last identity,
we can compress the information into a single entity shown on the left.
Let Ω be an open set in R
n
and let k ∈ N. We denote by C
k
(Ω) the set of all
continuous realvalued functions deﬁned on Ω such that D
α
u is continuous on Ω for
1.1. ELEMENTS OF FUNCTION SPACES 7
all α = (α
1
, . . . , α
n
) with [α[ ≤ k. Assuming that Ω is a bounded open set, C
k
(
¯
Ω)
will denote the set of all u in C
k
(Ω) such that D
α
u can be extended from Ω to a
continuous function on
¯
Ω, the closure of the set Ω, for all α = (α
1
, . . . , α
n
), [α[ ≤ k.
C
k
(
¯
Ω) can be equipped with the norm
u
C
k
(
¯
Ω)
:=
α≤k
sup
x∈Ω
[D
α
u(x)[.
In particular when k = 0 we shall write C(
¯
Ω) instead of C
0
(
¯
Ω) to denote the set of
all continuous functions deﬁned on
¯
Ω; in this case,
u
C(
¯
Ω)
= sup
x∈Ω
[u(x)[ = max
x∈
¯
Ω
[u(x)[.
Similarly, if k = 1,
u
C
1
(
¯
Ω)
=
α≤1
sup
x∈Ω
[D
α
u(x)[
= sup
x∈Ω
[u(x)[ +
n
j=1
sup
x∈Ω
[
∂u
∂x
j
(x)[.
Example 2 Consider the open interval Ω = (0, 1) ⊂ R
1
. The function u(x) = 1/x
belongs to C
k
(Ω) for each k ≥ 0. As
¯
Ω = [0, 1] and lim
x→0
u(x) = ∞, it is clear that
u is not continuous on
¯
Ω; the same is true of its derivatives. Therefore u ,∈ C
k
(
¯
Ω)
for any k ≥ 0. ⋄
The support of a continuous function u deﬁned on an open set Ω ⊂ R
n
is deﬁned
as the closure in Ω of the set ¦x ∈ Ω : u(x) ,= 0¦. We shall write supp u for the
support of u. Thus, supp u is the smallest closed subset of Ω such that u = 0 in
Ω¸supp u.
Example 3 Let w be the function deﬁned 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
(Ω) the set of all u contained in C
k
(Ω) whose support is a bounded
subset of Ω. Let
C
∞
0
(Ω) =
∩
k≥0
C
k
0
(Ω).
Example 4 The function w deﬁned 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
(Ω) the set of all realvalued functions
deﬁned on an open subset Ω of R
n
such that
_
Ω
[u(x)[
p
dx < ∞.
Any two functions which are equal almost everywhere (i.e. equal, except on a
set of measure zero) on Ω are identiﬁed with each other. Thus, strictly speaking,
L
p
(Ω) consists of equivalence classes of functions; still, we shall not insist on this
technicality. L
p
(Ω) is equipped with the norm
u
L
p
(Ω)
:=
__
Ω
[u(x)[
p
dx
_
1/p
.
We shall also consider the space L
∞
(Ω) consisting of functions u deﬁned on Ω such
that [u[ has ﬁnite essential supremum on Ω (namely, there exists a positive constant
M such that [u(x)[ ≤ M for almost every
3
x in Ω; the smallest such number M is
called the essential supremum of [u[, and we write M = ess.sup
x∈Ω
[u(x)[). L
∞
(Ω)
is equipped with the norm
u
L
∞
(Ω)
= ess.sup
x∈Ω
[u(x)[.
A particularly important case corresponds to taking p = 2; then
u
L
2
(Ω)
=
__
Ω
[u(x)[
2
dx
_
1/2
.
The space L
2
(Ω) can be equipped with the inner product
(u, v) :=
_
Ω
u(x)v(x) dx.
Clearly u
L
2
(Ω)
= (u, u)
1/2
.
Lemma 1 (The CauchySchwarz inequality) Let u and v belong to L
2
(Ω); then
uv ∈ L
1
(Ω) and
[(u, v)[ ≤ u
L
2
(Ω)
v
L
2
(Ω)
.
3
We shall say that a property P(x) is true for almost every x in Ω, if P(x) is true for all x ∈ Ω¸Γ
where Γ is a subset of Ω with zero Lebesgue measure.
1.1. ELEMENTS OF FUNCTION SPACES 9
Proof Let λ ∈ R; then
0 ≤ u +λv
2
L
2
(Ω)
= (u +λv, u +λv)
= (u, u) + (u, λv) + (λv, u) + (λv, λv)
= u
2
L
2
(Ω)
+ 2λ(u, v) +λ
2
v
2
L
2
(Ω)
, λ ∈ R.
The righthand side is a quadratic polynomial in λ with real coeﬃcients, and it is non
negative for all λ ∈ R; therefore its discriminant is nonpositive, i.e.
[2(u, v)[
2
−4u
2
L
2
(Ω)
v
2
L
2
(Ω)
≤ 0,
and hence the desired inequality.
Corollary 1 (The triangle inequality) Let u and v belong to L
2
(Ω); then u + v ∈
L
2
(Ω), and
u + v
L
2
(Ω)
≤ u
L
2
(Ω)
+v
L
2
(Ω)
.
Proof This is a straightforward consequence of the CauchySchwarz inequality:
u +v
2
L
2
(Ω)
= (u +v, u +v) = u
2
L
2
(Ω)
+ 2(u, v) +v
2
L
2
(Ω)
≤
_
u
L
2
(Ω)
+v
L
2
(Ω)
_
2
.
Upon taking the square root of both sides we complete the proof.
Remark 1 The space L
p
(Ω) with p ∈ [1, ∞] is a Banach space
4
. In particular,
L
2
(Ω) is a Hilbert space: it has an inner product (, ) and, when equipped with the
associated norm  
L
2
(Ω)
, deﬁned by u
L
2
(Ω)
= (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
(Ω)
≤ u
L
p
(Ω)
+v
L
p
(Ω)
, u, v ∈ L
p
(Ω).
Furthermore, the following generalisation of the CauchySchwarz inequality, known
as H¨older’s inequality, is valid for any two functions u ∈ L
p
(Ω) and v ∈ L
p
′ (Ω)
with 1/p + 1/p
′
= 1:
¸
¸
¸
¸
_
Ω
u(x)v(x) dx
¸
¸
¸
¸
≤ u
L
p
(Ω)
v
L
p
′ (Ω)
.
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 diﬀerential
equation theory. Before we give the precise deﬁnition of a Sobolev space, we intro
duce the concept of weak derivative.
Suppose that u is a smooth function, say u ∈ C
k
(Ω), with Ω an open subset of
R
n
, and let v ∈ C
∞
0
(Ω); then the following integrationbyparts formula holds:
_
Ω
D
α
u(x) v(x) dx = (−1)
α
_
Ω
u(x) D
α
v(x) dx, [α[ ≤ k,
∀v ∈ C
∞
0
(Ω).
Note that all terms involving integrals over the boundary of Ω, 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 Ω. This identity represents the starting point
for deﬁning the concept of weak derivative.
Now suppose that u is a locally integrable function deﬁned on Ω (i.e. u ∈ L
1
(ω)
for each bounded open set ω, with ¯ ω ⊂ Ω). Suppose also that there exists a function
w
α
, locally integrable on Ω and such that
_
Ω
w
α
(x) v(x) dx = (−1)
α
_
Ω
u(x) D
α
v(x) ∀v ∈ C
∞
0
(Ω);
then we say that w
α
is a weak derivative of the function u of order [α[ = α
1
+. . . +
α
n
, and we write w
α
= D
α
u. In order to see that this deﬁnition 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 Reymond’s
lemma
5
. Clearly, if u is a suﬃciently smooth function, say u ∈ C
k
(Ω), then its weak
derivative D
α
u of order [α[ ≤ k coincides with the corresponding partial derivative
in the classical pointwise sense,
∂
α
u
∂x
α
1
1
. . . ∂x
α
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 diﬀerentiated) which of the two is implied.
Example 5 Let Ω = R
1
, and suppose that we wish to determine the weak ﬁrst
derivative of the function u(x) = (1 −[x[)
+
deﬁned on Ω. Clearly u is not diﬀeren
tiable at the points 0 and ±1. However, because u is locally integrable on Ω, it may,
5
DuBois Reymond’s lemma: Suppose that w is a locally integrable function deﬁned on an
open set Ω, Ω ⊂ R
n
. If
_
Ω
w(x)v(x) dx = 0 for all v in C
∞
0
(Ω)
then w(x) = 0 for almost every x ∈ Ω.
1.1. ELEMENTS OF FUNCTION SPACES 11
nevertheless, have a weak derivative. Indeed, for any v ∈ C
∞
0
(Ω),
_
+∞
−∞
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 ﬁrst (weak) derivative of the contin
uous piecewise linear function u, i.e. w = u
′
= Du. ⋄
Now we are ready to give a precise deﬁnition of a Sobolev space. Let k be a
nonnegative integer and suppose that p ∈ [1, ∞]. We deﬁne (with D
α
denoting a
weak derivative of order [α[ )
W
k
p
(Ω) = ¦u ∈ L
p
(Ω) : D
α
u ∈ L
p
(Ω), [α[ ≤ k¦.
W
k
p
(Ω) is called a Sobolev space of order k; it is equipped with the (Sobolev) norm
u
W
k
p
(Ω)
:=
_
_
α≤k
D
α
u
p
L
p
(Ω)
_
_
1/p
when 1 ≤ p < ∞
and
u
W
k
∞
(Ω)
:=
α≤k
D
α
u
L
∞
(Ω)
when p = ∞.
Letting,
[u[
W
k
p
(Ω)
:=
_
_
α=k
D
α
u
p
L
p
(Ω)
_
_
1/p
,
for p ∈ [1, ∞), we can write
u
W
k
p
(Ω)
=
_
k
j=0
[u[
p
W
j
p
(Ω)
_
1/p
.
12 CHAPTER 1. INTRODUCTION
Similarly, letting
[u[
W
k
∞
(Ω)
:=
α=k
D
α
u
L
∞
(Ω)
,
we have that
u
W
k
∞
(Ω)
=
k
j=0
[u[
W
j
∞
(Ω)
.
When k ≥ 1, [[
W
k
p
(Ω)
is called the Sobolev seminorm
6
on W
k
p
(Ω).
An important special case corresponds to taking p = 2; the space W
k
2
(Ω) is then
a Hilbert space with the inner product
(u, v)
W
k
2
(Ω)
:=
α≤k
(D
α
u, D
α
v).
For this reason, we shall usually write H
k
(Ω) instead of W
k
2
(Ω).
Throughout these notes we shall frequently refer to the Hilbertian Sobolev spaces
H
1
(Ω) and H
2
(Ω). Our deﬁnitions of W
k
p
(Ω) and its norm and seminorm, for p = 2,
k = 1, give:
H
1
(Ω) =
_
u ∈ L
2
(Ω) :
∂u
∂x
j
∈ L
2
(Ω), j = 1, . . . , n
_
,
u
H
1
(Ω)
=
_
u
2
L
2
(Ω)
+
n
j=1

∂u
∂x
j

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

∂u
∂x
j

2
L
2
(Ω)
_
1/2
.
Similarly, for p = 2 and k = 2,
H
2
(Ω) =
_
u ∈ L
2
(Ω) :
∂u
∂x
j
∈ L
2
(Ω), j = 1, . . . , n,
∂
2
u
∂x
i
∂x
j
∈ L
2
(Ω), i, j = 1, . . . , n
_
,
u
H
2
(Ω)
=
_
u
2
L
2
(Ω)
+
n
j=1

∂u
∂x
j

2
L
2
(Ω)
+
n
i,j=1

∂
2
u
∂x
i
∂x
j

2
L
2
(Ω)
_
1/2
,
6
When k ≥ 1, [ [
W
k
p
(Ω)
is only a seminorm rather than a norm because if [u[
W
k
p
(Ω)
= 0 for
u ∈ W
k
p
(Ω) it does not necessarily follow that u(x) = 0 for almost every x in Ω (all that is known
is that D
α
u(x) = 0 for almost every x ∈ Ω, [α[ = k), so [ [
W
k
p
(Ω)
does not satisfy the ﬁrst axiom
of norm.
1.1. ELEMENTS OF FUNCTION SPACES 13
[u[
H
2
(Ω)
=
_
n
i,j=1

∂
2
u
∂x
i
∂x
j

2
L
2
(Ω)
_
1/2
.
Finally, we deﬁne the special Sobolev space H
1
0
(Ω) as the closure of C
∞
0
(Ω) in the
norm of  
H
1
(Ω)
; in other words, H
1
0
(Ω) is the set of all u ∈ H
1
(Ω) such that u
is the limit in H
1
(Ω) of a sequence ¦u
m
¦
∞
m=1
with u
m
∈ C
∞
0
(Ω). It can be shown
(assuming that ∂Ω is suﬃciently smooth) that
H
1
0
(Ω) = ¦u ∈ H
1
(Ω) : u = 0 on ∂Ω¦;
i.e. H
1
0
(Ω) is, in fact, the set of all functions u in H
1
(Ω) such that u = 0 on ∂Ω, the
boundary of the set Ω. We shall use this space when considering a partial diﬀerential
equation that is coupled with a homogeneous (Dirichlet) boundary condition: u = 0
on ∂Ω. We note here that H
1
0
(Ω) is also a Hilbert space, with the same norm and
inner product as H
1
(Ω).
We conclude the section with the following useful result.
Lemma 2 (Poincar´eFriedrichs inequality) Suppose that Ω is a bounded open set in
R
n
(with a suﬃciently smooth boundary
7
∂Ω) and let u ∈ H
1
0
(Ω); then there exists
a constant c
⋆
(Ω), independent of u, such that
_
Ω
[u(x)[
2
dx ≤ c
⋆
n
i=1
_
Ω
[
∂u
∂x
i
(x)[
2
dx. (1.2)
Proof As any function u ∈ H
1
0
(Ω) is the limit in H
1
(Ω) of a sequence ¦u
m
¦
∞
m=1
⊂ C
∞
0
(Ω),
it is suﬃcient to prove this inequality for u ∈ C
∞
0
(Ω).
In fact, to simplify matters, we shall restrict ourselves to considering the special case
of a rectangular domain Ω = (a, b) (c, d) in R
2
. The proof for general Ω is analogous.
Evidently
u(x, y) = u(a, y) +
_
x
a
∂u
∂x
(ξ, y) dξ =
_
x
a
∂u
∂x
(ξ, y) dξ, c < y < d.
Thence, by the CauchySchwarz inequality,
_
Ω
[u(x, y)[
2
dxdy =
_
b
a
_
d
c
[
_
x
a
∂u
∂x
(ξ, y) dξ[
2
dy dx
≤
_
b
a
_
d
c
(x −a)
__
x
a
[
∂u
∂x
(ξ, y)[
2
dξ
_
dy dx
≤
_
b
a
(x −a) dx
__
d
c
_
b
a
[
∂u
∂x
(ξ, y)[
2
dξ dy
_
=
1
2
(b −a)
2
_
Ω
[
∂u
∂x
(x, y)[
2
dx dy.
7
Say, Ω is a polygonal domain in R
2
or a polyhedron in R
3
.
14 CHAPTER 1. INTRODUCTION
Analogously,
_
Ω
[u(x, y)[
2
dx dy ≤
1
2
(d −c)
2
_
Ω
[
∂u
∂y
(x, y)[
2
dxdy.
By adding the two inequalities, we obtain
_
Ω
[u(x, y)[
2
dxdy ≤ c
⋆
_
Ω
_
[
∂u
∂x
[
2
+[
∂u
∂y
[
2
_
dxdy,
where c
⋆
=
_
2
(b−a)
2
+
2
(d−c)
2
_
−1
.
For further reference, we note that if Ω = (0, 1)
2
⊂ R
2
then c
⋆
=
1
4
; similarly, if
Ω = (0, 1) ⊂ R then c
⋆
=
1
2
.
1.2 Weak solutions to elliptic problems
In the ﬁrst part of this lecture course we shall focus on boundary value problems for
elliptic partial diﬀerential equations. Elliptic equations are typiﬁed by the Laplace
equation
∆u = 0,
and its nonhomogeneous counterpart, Poisson’s equation
−∆u = f,
where we used the notation
∆ =
n
i=1
∂
2
∂x
2
i
for the Laplace operator.
More generally, let Ω be a bounded open set in R
n
, and consider the linear
secondorder partial diﬀerential equation
−
n
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 ∈ Ω, (1.3)
where the coeﬃcients a
ij
, b
i
, c and f satisfy the following conditions:
a
ij
∈ C
1
(
¯
Ω), i, j = 1, . . . , n;
b
i
∈ C(
¯
Ω), i = 1, . . . , n;
c ∈ C(
¯
Ω), f ∈ C(
¯
Ω),
and
n
i,j=1
a
ij
(x)ξ
i
ξ
j
≥ ˜ c
n
i=1
ξ
2
i
, ∀ξ = (ξ
1
, . . . , ξ
n
) ∈ R
n
, x ∈
¯
Ω; (1.4)
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 15
here ˜ c is a positive constant independent of x and ξ. 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 deﬁned
on ∂Ω:
(a) u = g on ∂Ω (Dirichlet boundary condition);
(b)
∂u
∂ν
= g on ∂Ω, where ν denotes the unit outward normal vector to ∂Ω (Neu
mann boundary condition);
(c)
∂u
∂ν
+ σu = g on ∂Ω, where σ(x) ≥ 0 on ∂Ω (Robin boundary condition);
(d) A generalisation of the boundary conditions (b) and (c) is
n
i,j=1
a
ij
∂u
∂x
i
cos α
j
+ σ(x)u = g on ∂Ω,
where α
j
is the angle between the unit outward normal vector ν to ∂Ω and
the x
j
axis (Oblique derivative boundary condition).
In many physical problems more than one type of boundary condition is imposed
on ∂Ω (e.g. ∂Ω is the union of two disjoint subsets ∂Ω
1
and ∂Ω
2
, with a Dirichlet
boundary condition on ∂Ω
1
and Neumann boundary condition on ∂Ω
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
−
n
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 ∈ Ω, (1.5)
u = 0 on ∂Ω, (1.6)
where a
ij
, b
i
, c and f are as in (1.4).
A function u ∈ C
2
(Ω)∩C(
¯
Ω) satisfying (1.5) and (1.6) is called a classical solu
tion of this problem. The theory of partial diﬀerential equations tells us that (1.5),
(1.6) has a unique classical solution, provided that a
ij
, b
i
, c, f and ∂Ω are suﬃciently
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, Poisson’s equation with zero Dirichlet boundary
condition on Ω = (−1, 1)
n
in R
n
:
−∆u = sgn
_
1
2
−[x[
_
, x ∈ Ω,
u = 0, x ∈ ∂Ω.
_
(∗)
This problem does not have a classical solution, u ∈ C
2
(Ω)∩C(
¯
Ω), for otherwise ∆u
would be a continuous function on Ω, which is not possible because sgn(1/2 − [x[)
is not continuous on Ω.
16 CHAPTER 1. INTRODUCTION
In order to overcome the limitations of the classical theory and to be able to
deal with partial diﬀerential equations with “nonsmooth” data, we generalise the
notion of solution by weakening the diﬀerentiability 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
(Ω),
−
n
i,j=1
_
Ω
∂
∂x
j
_
a
ij
∂u
∂x
i
_
v dx +
n
i=1
_
Ω
b
i
(x)
∂u
∂x
i
v dx
+
_
Ω
c(x)uv dx =
_
Ω
f(x)v(x) dx.
Upon integration by parts in the ﬁrst integral and noting that v = 0 on ∂Ω, we
obtain:
n
i,j=1
_
Ω
a
ij
(x)
∂u
∂x
i
∂v
∂x
j
dx +
n
i=1
_
Ω
b
i
(x)
∂u
∂x
i
v dx
+
_
Ω
c(x)uv dx =
_
Ω
f(x)v(x) dx ∀v ∈ C
1
0
(Ω).
In order for this equality to make sense we no longer need to assume that u ∈ C
2
(Ω):
it is suﬃcient that u ∈ L
2
(Ω) and ∂u/∂x
i
∈ L
2
(Ω), 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
(Ω), where, as in Section 1.1.3,
H
1
0
(Ω) = ¦u ∈ L
2
(Ω) :
∂u
∂x
i
∈ L
2
(Ω), i = 1, . . . , n, u = 0 on ∂Ω¦.
Therefore, we consider the following problem: ﬁnd u in H
1
0
(Ω) such that
n
i,j=1
_
Ω
a
ij
(x)
∂u
∂x
i
∂v
∂x
j
dx +
n
i=1
_
Ω
b
i
(x)
∂u
∂x
i
v dx
+
_
Ω
c(x)uv dx =
_
Ω
f(x)v(x) dx ∀v ∈ C
1
0
(Ω). (1.7)
We note that C
1
0
(Ω) ⊂ H
1
0
(Ω), and it is easily seen that when u ∈ H
1
0
(Ω) and
v ∈ H
1
0
(Ω), (instead of v ∈ C
1
0
(Ω)), the expressions on the left and righthand side
of (1.7) are still meaningful (in fact, we shall prove this below)
8
. This motivates the
following deﬁnition.
8
Note further that since the coeﬃcients a
ij
no longer appear under derivative signs in (1.7),
it is not necessary to assume that a
ij
∈ C
1
(
¯
Ω); a
ij
∈ L
∞
(Ω) will be seen to be suﬃcient. Also,
the smoothness requirements imposed on the coeﬃcients b
i
and c can be relaxed: b
i
∈ L
∞
(Ω) for
i = 1, . . . , n and c ∈ L
∞
(Ω) will suﬃce.
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 17
Deﬁnition 1 Let a
ij
∈ L
∞
(Ω), i, j = 1, . . . , n, b
i
∈ L
∞
(Ω), i = 1, . . . , n, c ∈
L
∞
(Ω), and let f ∈ L
2
(Ω). A function u ∈ H
1
0
(Ω) satisfying
n
i,j=1
_
Ω
a
ij
(x)
∂u
∂x
i
∂v
∂x
j
dx +
n
i=1
_
Ω
b
i
(x)
∂u
∂x
i
v dx
+
_
Ω
c(x)uv dx =
_
Ω
f(x)v(x) dx ∀v ∈ H
1
0
(Ω) (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
(Ω),
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
_
Ω
a
ij
(x)
∂w
∂x
i
∂v
∂x
j
dx
+
n
i=1
_
Ω
b
i
(x)
∂w
∂x
i
v dx +
_
Ω
c(x)wv dx (1.9)
and
l(v) =
_
Ω
f(x)v(x) dx. (1.10)
With this new notation, problem (1.8) can be written as follows:
ﬁnd u ∈ H
1
0
(Ω) such that a(u, v) = l(v) ∀v ∈ H
1
0
(Ω). (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, NorthHolland, 1978; K. Yosida:
Functional Analysis, Reprint of the 6th ed., SpringerVerlag, 1995.
We apply the LaxMilgram theorem with V = H
1
0
(Ω) and  
V
=  
H
1
(Ω)
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
(Ω) is a Hilbert space with the inner
product
(w, v)
H
1
(Ω)
=
_
Ω
wv dx +
n
i=1
_
Ω
∂w
∂x
i
∂v
∂x
i
dx
and the associated norm w
H
1
(Ω)
= (w, w)
1/2
H
1
(Ω)
. Next we show that a(, ) and l(),
deﬁned by (1.9) and (1.10), satisfy the hypotheses (a), (b), (c) of the LaxMilgram
theorem.
We begin with (c). The mapping v → l(v) is linear: indeed, for any α, β ∈ R,
l(αv
1
+ βv
2
) =
_
Ω
f(x)(αv
1
(x) + βv
2
(x)) dx
= α
_
Ω
f(x)v
1
(x) dx + β
_
Ω
f(x)v
2
(x) dx
= αl(v
1
) + βl(v
2
), v
1
, v
2
∈ H
1
0
(Ω);
so l() is a linear functional on H
1
0
(Ω). Also, by the CauchySchwarz inequality,
[l(v)[ = [
_
Ω
f(x)v(x) dx[ ≤
__
Ω
[f(x)[
2
dx
_
1/2
__
Ω
[v(x)[
2
dx
_
1/2
= f
L
2
(Ω)
v
L
2
(Ω)
≤ f
L
2
(Ω)
v
H
1
(Ω)
,
for all v ∈ H
1
0
(Ω), where we have used the obvious inequality v
L
2
(Ω)
≤ v
H
1
(Ω)
.
Letting c
2
= f
L
2
(Ω)
, we obtain the required bound.
Next we verify (b). For any ﬁxed w ∈ H
1
0
(Ω), the mapping v → a(v, w) is linear.
Similarly, for any ﬁxed v ∈ H
1
0
(Ω), the mapping w → a(v, w) is linear. Hence a(, )
is a bilinear functional on H
1
0
(Ω) H
1
0
(Ω). Applying the CauchySchwarz inequality,
we deduce that
[a(w, v)[ ≤
n
i,j=1
max
x∈
¯
Ω
[a
ij
(x)[ [
_
Ω
∂w
∂x
i
∂v
∂x
j
dx[
+
n
i=1
max
x∈
¯
Ω
[b
i
(x)[ [
_
Ω
∂w
∂x
i
v dx[
1.2. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 19
+max
x∈
¯
Ω
[c(x)[ [
_
Ω
w(x)v(x) dx[
≤ ˆ c
_
n
i,j=1
__
Ω
[
∂w
∂x
i
[
2
dx
_
1/2
__
Ω
[
∂v
∂x
j
[
2
dx
_
1/2
+
n
i=1
__
Ω
[
∂w
∂x
i
[
2
dx
_
1/2
__
Ω
[v[
2
dx
_
1/2
+
__
Ω
[w[
2
dx
_
1/2
__
Ω
[v[
2
dx
_
1/2
_
≤ ˆ c
_
__
Ω
[w[
2
dx
_
1/2
+
n
i=1
__
Ω
[
∂w
∂x
i
[
2
dx
_
1/2
_
_
__
Ω
[v[
2
dx
_
1/2
+
n
j=1
__
Ω
[
∂v
∂x
j
[
2
dx
_
1/2
_
(1.12)
where
ˆ c = max
_
max
1≤i,j≤n
max
x∈
¯
Ω
[a
ij
(x)[, max
1≤i≤n
max
x∈
¯
Ω
[b
i
(x)[, max
x∈
¯
Ω
[c(x)[
_
.
By further majorisation of the righthand side in (1.12) we deduce that
[a(w, v)[ ≤ 2nˆ c
_
_
Ω
[w[
2
dx +
n
i=1
_
Ω
[
∂w
∂x
i
[
2
dx
_
1/2
_
_
Ω
[v[
2
dx +
n
j=1
_
Ω
[
∂v
∂x
j
[
2
dx
_
1/2
,
so that, by letting c
1
= 2nˆ c, we obtain inequality (b):
[a(w, v)[ ≤ c
1
w
H
1
(Ω)
v
H
1
(Ω)
. (1.13)
It remains to establish (a). To do so, we shall slightly strengthen the smoothness
requirements on the coeﬃcients b
i
by demanding that b
i
∈ W
1
∞
(Ω) (see, however,
Remark 4 at the end of this chapter). Using (1.4), we deduce that
a(v, v) ≥ ˜ c
n
i=1
_
Ω
[
∂v
∂x
i
[
2
dx +
n
i=1
_
Ω
b
i
(x)
1
2
∂
∂x
i
(v
2
) dx +
_
Ω
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
_
Ω
[
∂v
∂x
i
[
2
dx +
_
Ω
_
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 ∈
¯
Ω. (1.14)
Then
a(v, v) ≥ ˜ c
n
i=1
_
Ω
[
∂v
∂x
i
[
2
dx. (1.15)
By virtue of the Poincar´eFriedrichs inequality stated in Lemma 1.2, the righthand
side can be further bounded below to obtain
a(v, v) ≥
˜ c
c
⋆
_
Ω
[v[
2
dx. (1.16)
Summing (1.15) and (1.16),
a(v, v) ≥ c
0
_
_
Ω
[v[
2
dx +
n
i=1
_
Ω
[
∂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
(Ω) 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
∞
(Ω), i, j = 1, . . . , n, b
i
∈ W
1
∞
(Ω), i = 1, . . . , n,
c ∈ L
∞
(Ω), f ∈ L
2
(Ω), 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
(Ω). In addition,
u
H
1
(Ω)
≤
1
c
0
f
L
2
(Ω)
. (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 CauchySchwarz inequality and recalling the deﬁnition of  
H
1
(Ω)
,
c
0
u
2
H
1
(Ω)
≤ a(u, u) = l(u) = (f, u)
≤ [(f, u)[ ≤ f
L
2
(Ω)
u
L
2
(Ω)
≤ f
L
2
(Ω)
u
H
1
(Ω)
.
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
Ω = (−1, 1)
n
, we see that (1.4) holds with ˜ c = 1 and (1.14) is trivially fulﬁlled.
Thus (∗) has a unique weak solution u ∈ H
1
0
(Ω) 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 DirichletNeumann mixed bound
ary value problem:
−∆u = f in Ω,
u = 0 on Γ
1
,
∂u
∂ν
= g on Γ
2
,
where Γ
1
is a nonempty, relatively open subset of ∂Ω and Γ
1
∪ Γ
2
= ∂Ω. We shall
suppose that f ∈ L
2
(Ω) and that g ∈ L
2
(Γ
2
). Following a similar reasoning as in the
case of the Dirichlet boundary value problem, we consider the special Sobolev space
H
1
0,Γ
1
(Ω) = ¦v ∈ H
1
(Ω) : v = 0 on Γ
1
¦,
and deﬁne the weak formulation of the mixed problem as follows: ﬁnd u ∈ H
1
0,Γ
1
(Ω)
such that
a(u, v) = l(v) for all v in H
1
0,Γ
1
(Ω),
where we put
a(u, v) =
_
Ω
n
i=1
∂u
∂x
i
∂v
∂x
i
dx
and
l(v) =
_
Ω
f(x)v(x) dx +
_
Γ
2
g(s)v(s) ds.
Applying the LaxMilgram theorem with V = H
1
0,Γ
1
(Ω), 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 wellposed in the sense of Hadamard; namely, for each
f ∈ L
2
(Ω) there exists a unique (weak) solution u ∈ H
1
0
(Ω), 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
(Ω) of (1.5), (1.6)
corresponding to righthand sides f
1
and f
2
in L
2
(Ω), respectively, then u
1
− u
2
is the weak solution in H
1
0
(Ω) of (1.5), (1.6) corresponding to the righthand side
f
1
−f
2
∈ L
2
(Ω). Thus, by virtue of (1.18),
u
1
−u
2

H
1
(Ω)
≤
1
c
0
f
1
−f
2

L
2
(Ω)
, (1.19)
and hence the required continuous dependence of the solution of the boundary value
problem on the righthand side. ⋄
Remark 4 The requirement b
i
∈ W
1
∞
(Ω) in Theorem 2 can be relaxed to the original
assumption b
i
∈ L
∞
(Ω), i = 1, . . . , n. To see this, note that the smoothness require
ments on b
i
are unrelated to the veriﬁcation of condition (c) in the LaxMilgram
22 CHAPTER 1. INTRODUCTION
theorem, and condition (b) can be shown with b
i
∈ L
∞
(Ω), i = 1, . . . , n, only any
way. Thus, it remains to see how condition (a) may be veriﬁed under the hypothesis
b
i
∈ L
∞
(Ω), i = 1, . . . , n. By (1.4) and the CauchySchwarz inequality,
a(v, v) ≥ ˜ c[v[
2
H
1
(Ω)
−
_
n
i=1
b
i

2
L
∞
(Ω)
_
1/2
[v[
H
1
(Ω)
v
L
2
(Ω)
+
_
Ω
c(x)[v(x)[
2
dx
≥
1
2
˜ c[v[
2
H
1
(Ω)
+
_
Ω
_
c(x) −
2
˜ c
n
i=1
b
i

2
L
∞
(Ω)
_
[v(x)[
2
dx.
Assuming that
c(x) −
2
˜ c
n
i=1
b
i

2
L
∞
(Ω)
≥ 0 (1.20)
we arrive at the inequality
a(v, v) ≥
1
2
˜ c
n
i=1
_
Ω
[
∂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 veriﬁes
condition (a) in the LaxMilgram theorem, under the assumptions that b
i
∈ L
∞
(Ω),
i = 1, . . . , n, only and (1.4), (1.20) hold. ⋄
Chapter 2
Approximation of elliptic problems
In this chapter we describe the construction of ﬁnite element methods for ellip
tic boundary value problems and outline some of their key properties. Unlike ﬁnite
diﬀerence schemes which are constructed in a moreorless ad hoc fashion through re
placing the derivatives in the diﬀerential equation by divided diﬀerences, the deriva
tion of ﬁnite element methods is quite systematic.
The ﬁrst step in the construction of a ﬁnite 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:
ﬁnd u ∈ V such that a(u, v) = l(v) ∀v ∈ V , (P)
where V is the solution space (e.g. H
1
0
(Ω) 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 ﬁnitedimensional
subspace V
h
⊂ V which consists of continuous piecewise polynomial functions of
a ﬁxed degree associated with a subdivision of the computational domain; then
consider the following approximation of (P):
ﬁnd 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¦φ
1
, . . . , φ
N(h)
¦,
where the (linearly independent) basis functions φ
i
, i = 1, . . . , N(h), have “small”
support. Expressing the approximate solution u
h
in terms of the basis functions, φ
i
,
we can write
u
h
(x) =
N(h)
i=1
U
i
φ
i
(x), (∗∗)
where U
i
, i = 1, . . . , N(h), are to be determined. Thus (P
h
) can be rewritten as
follows:
ﬁnd (U
1
, . . . , U
N(h)
) ∈ R
N(h)
such that
23
24 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
N(h)
i=1
a(φ
i
, φ
j
)U
i
= l(φ
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(φ
j
, φ
i
)) of size N(h) N(h). Because the φ
i
’s have small
support, a(φ
j
, φ
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 eﬃcient 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 ﬁnite 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 ﬁnite element method through
two simple examples: the ﬁrst of these is a twopoint boundary value problem
for a secondorder ordinary diﬀerential equation; the second model problem is the
homogeneous Dirichlet boundary value problem for Poisson’s equation on the unit
square in the plane. For the time being we shall assume that the ﬁnite element
space V
h
consists of continuous piecewise linear functions. Higherdegree piecewise
polynomial approximations will be discussed later on in the notes.
Onedimensional 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:
ﬁnd 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 ﬁnite element approximation of this problem, we sub
divide
¯
Ω = [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
¯
Ω = [0, 1].
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
Z
x
i
x
i−1
1
x
i+1
Figure 2.2: The piecewise linear ﬁnite element basis function φ
i
(x).
The subintervals (x
i
, x
i+1
) are referred to as element domains or elements,
(hence the name ﬁnite 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 ﬁnite element basis functions
φ
i
(x) =
_
1 −[
x −x
i
h
[
_
+
, i = 1, . . . , N −1,
shown in Figure 2.2. It is clear that φ
i
∈ H
1
0
(0, 1); furthermore, supp φ
i
= [x
i−1
, x
i+1
],
i = 1, . . . , N − 1, and the functions φ
i
, i = 1, . . . , N − 1, are linearly independent;
therefore
V
h
:= span¦φ
1
, . . . , φ
N−1
¦
is an (N −1)dimensional subspace of H
1
0
(0, 1).
The ﬁnite element approximation of (P) is:
ﬁnd 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¦φ
1
, . . . , φ
N−1
¦, it can be written as a linear combination
26 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
of the basis functions:
u
h
(x) =
N−1
i=1
U
i
φ
i
(x).
Substituting this expansion into (P
h
) we obtain the following problem, equivalent
to (P
h
):
ﬁnd U = (U
1
, . . . , U
N−1
)
T
∈ R
N−1
such that
N−1
i=1
U
i
_
1
0
[p(x)φ
′
i
(x)φ
′
j
(x) + q(x)φ
i
(x)φ
j
(x)] dx
=
_
1
0
f(x)φ
j
(x) dx,
for j = 1, . . . , N −1.
_
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
_
(P
′
h
)
Letting
a
ji
:=
_
1
0
[p(x)φ
′
i
(x)φ
′
j
(x) + q(x)φ
i
(x)φ
j
(x)] dx, i, j = 1, . . . , N −1;
F
j
:=
_
1
0
f(x)φ
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
N−1
)
T
. The matrix A is symmetric (i.e. A
T
= A)
and positive deﬁnite (i.e. x
T
Ax > 0, x ,= 0). Since supp φ
i
∩ supp φ
j
has empty
interior when [i −j[ > 1, it follows that the matrix A is tridiagonal (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
N−1
into the expansion
u
h
(x) =
N−1
i=1
U
i
φ
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
φ
′
i
(x)φ
′
j
(x) dx + q
_
1
0
φ
i
(x)φ
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
¯
Ω.
This gives rise to the following set of linear equations:
−p
U
i−1
−2U
i
+ U
i+1
h
2
+ q
U
i−1
+ 4U
i
+ U
i+1
6
=
1
h
_
x
i+1
x
i−1
f(x)φ
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 threepoint ﬁnite diﬀerence scheme
for the values U
i
, the values of u
h
(x) at the mesh points x
i
.
Twodimensional problem
Let Ω be a bounded domain in R
2
with a polygonal boundary ∂Ω; thus Ω can be
exactly covered by a ﬁnite number of triangles. It will be assumed that any pair of
triangles in a triangulation of Ω 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 deﬁne h = max
K
h
K
.
With each interior node (marked ⊙ in the ﬁgure) we associate a basis function φ
which is equal to 1 at that node and equal to 0 at all the other nodes; φ is assumed
to be a continuous function on
¯
Ω 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 φ
1
(x, y),
. . . , φ
N(h)
(x, y) be the corresponding basis functions. The functions φ
1
, . . . , φ
N(h)
are linearly independent and they span an N(h)dimensional linear subspace V
h
of
H
1
0
(Ω).
Let us consider the elliptic boundary value problem
−∆u = f in Ω,
28 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
0 0
0 0
0 0
1
Figure 2.4: A typical ﬁnite element basis function φ.
u = 0 on ∂Ω.
In order to construct the ﬁnite 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):
ﬁnd u ∈ H
1
0
(Ω) such that
_
Ω
_
∂u
∂x
∂v
∂x
+
∂u
∂y
∂v
∂y
_
dxdy =
_
Ω
fv dxdy ∀v ∈ H
1
0
(Ω).
The ﬁnite element approximation of the problem is:
ﬁnd u
h
∈ V
h
such that
_
Ω
_
∂u
h
∂x
∂v
h
∂x
+
∂u
h
∂y
∂v
h
∂y
_
dxdy =
_
Ω
fv
h
dxdy ∀v
h
∈ V
h
.
Writing
u
h
(x, y) =
N(h)
i=1
U
i
φ
i
(x, y),
the ﬁnite element method can be restated as follows:
ﬁnd 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
__
Ω
_
∂φ
i
∂x
∂φ
j
∂x
+
∂φ
i
∂y
∂φ
j
∂y
_
dxdy
_
=
_
Ω
fφ
j
dxdy,
for j = 1, . . . , N(h).
Letting A = (a
ij
), F = (F
1
, . . . , F
N(h)
)
T
,
a
ij
= a
ji
=
_
Ω
_
∂φ
i
∂x
∂φ
j
∂x
+
∂φ
i
∂y
∂φ
j
∂y
_
dxdy,
F
j
=
_
Ω
fφ
j
dxdy,
the ﬁnite 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
φ
i
(x, y).
The matrix A is called the stiﬀness matrix.
To simplify matters let us suppose that Ω = (0, 1) (0, 1) and consider the
triangulation of
¯
Ω shown in Fig. 2.5. The case of a general triangulation will be
considered later. Let φ
ij
denote the basis function associated with the interior node
(x
i
, y
j
):
φ
ij
(x, y) =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
_
1 −
x−x
i
h
−
y−y
j
h
, (x, y) ∈ 1
1 −
y−y
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 −
x−x
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,
∂φ
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
¯
Ω = [0, 1] [0, 1].
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
@
si
1
2
3
4
5
6
Figure 2.6: Triangles surrounding a node.
2.2. THE SELFADJOINT ELLIPTIC PROBLEM 31
and
∂φ
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
N−1
i=1
N−1
j=1
U
ij
_
Ω
_
∂φ
ij
∂x
∂φ
kl
∂x
+
∂φ
ij
∂y
∂φ
kl
∂y
_
dxdy
= 4U
kl
−U
k−1,l
−U
k+1,l
−U
k,l−1
−U
k,l+1
, k, l = 1, ..., N −1,
the ﬁnite element approximation is equivalent to
−
U
k+1,l
−2U
k,l
+ U
k−1,l
h
2
−
U
k,l+1
−2U
k,l
+ U
k,l−1
h
2
=
1
h
2
_ _
supp φ
kl
f(x, y)φ
kl
(x, y) dxdy, k, l = 1, . . . , N −1;
U
kl
= 0 on ∂Ω.
Thus, on this special triangulation of Ω, the ﬁnite element approximation gives
rise to the familiar 5point ﬁnite diﬀerence scheme with the forcing function f aver
aged in a special way.
2.2 The selfadjoint elliptic problem
Let us consider, as in Chapter 1, the elliptic boundary value problem
−
n
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 ∈ Ω, (2.3)
u = 0 on ∂Ω, (2.4)
where Ω is a bounded open set in R
n
, a
ij
∈ L
∞
(Ω), i, j = 1, . . . , n; b
i
∈ W
1
∞
(Ω),
i = 1, . . . , n, c ∈ L
∞
(Ω), f ∈ L
2
(Ω), and assume that there exists a positive constant
˜ c such that
n
i,j=1
a
ij
(x)ξ
i
ξ
j
≥ ˜ c
n
i=1
ξ
2
i
∀ξ = (ξ
1
, . . . , ξ
n
) ∈ R
n
, ∀x ∈
¯
Ω. (2.5)
We recall from Chapter 1 that the weak formulation of (2.3), (2.4) is:
ﬁnd u ∈ H
1
0
(Ω) such that a(u, v) = l(v) ∀v ∈ H
1
0
(Ω), (2.6)
32 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
where the bilinear functional a(, ) and the linear functional l() are deﬁned by
a(u, v) =
n
i,j=1
_
Ω
a
ij
∂u
∂x
i
∂v
∂x
j
dx +
n
i=1
_
Ω
b
i
(x)
∂u
∂x
i
v dx +
_
Ω
c(x)uv dx,
and
l(v) =
_
Ω
f(x)v(x) dx.
We have shown that if
c(x) −
1
2
n
i=1
∂b
i
∂x
i
≥ 0, x ∈
¯
Ω,
then (2.6) has a unique solution u in H
1
0
(Ω), the weak solution of (2.3), (2.4). In
the special case when the boundary value problem is selfadjoint, i.e.
a
ij
(x) = a
ji
(x), i, j = 1, . . . , n, x ∈
¯
Ω,
and
b
i
(x) ≡ 0, i = 1, . . . , n, x ∈
¯
Ω,
the bilinear functional a(, ) is symmetric in the sense that
a(v, w) = a(w, v) ∀v, w ∈ H
1
0
(Ω);
in the rest of this section this will always be assumed to be the case. Thus we
consider
−
n
i,j=1
∂
∂x
j
_
a
ij
(x)
∂u
∂x
i
_
+ c(x)u = f(x), x ∈ Ω,
(2.7)
u = 0, on ∂Ω
with a
ij
(x) satisfying the ellipticity condition (2.5); a
ij
(x) = a
ji
(x), c(x) ≥ 0, x ∈
¯
Ω.
It turns out that (2.7) can be restated as a minimisation problem. To be more
precise, we deﬁne the quadratic functional J : H
1
0
(Ω) →R by
J(v) =
1
2
a(v, v) −l(v), v ∈ H
1
0
(Ω).
Lemma 3 Let u be the (unique) weak solution to (2.6) in H
1
0
(Ω) and suppose that
a(, ) is a symmetric bilinear functional on H
1
0
(Ω); then u is the unique minimiser
of J() over H
1
0
(Ω).
2.2. THE SELFADJOINT ELLIPTIC PROBLEM 33
Proof Let u be the unique weak solution to (2.6) in H
1
0
(Ω) and, for v ∈ H
1
0
(Ω), 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
(Ω)
,
where c
0
is a positive constant. Thus
J(v) −J(u) ≥
c
0
2
v −u
2
H
1
(Ω)
∀v ∈ H
1
0
(Ω), (2.8)
and therefore,
J(v) ≥ J(u) ∀v ∈ H
1
0
(Ω), (2.9)
i.e. u minimises J() over H
1
0
(Ω).
In fact, u is the unique minimiser of J() in H
1
0
(Ω). Indeed, if ˜ u also minimises J()
on H
1
0
(Ω), then
J(v) ≥ J(˜ u) ∀v ∈ H
1
0
(Ω). (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
(Ω)
= 0,
and hence u = ˜ u.
It is easily shown that J() is convex (down), i.e.
J((1 −θ)v + θw) ≤ (1 −θ)J(v) + θJ(w) ∀θ ∈ [0, 1], ∀v, w ∈ H
1
0
(Ω).
34 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
u
J(u)
J(v)
H
1
0
(Ω)
R
Figure 2.7: The quadratic functional J().
This follows from the identity
(1 −θ)J(v) + θJ(w) = J((1 −θ)v + θw) +
1
2
θ(1 −θ)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
λ→0
J(u + λv) −J(u)
λ
= 0
for all v ∈ H
1
0
(Ω). Since
J(u + λv) −J(u)
λ
= a(u, v) −l(v) +
λ
2
a(v, v),
we deduce that if u minimises J() then
lim
λ→0
[a(u, v) −l(v) +
λ
2
a(v, v)] = a(u, v) −l(v) = 0 ∀v ∈ H
1
0
(Ω),
which proves the following result.
Lemma 4 Suppose that u ∈ H
1
0
(Ω) minimises J() over H
1
0
(Ω); then u is the
(unique) solution of problem (2.6). The problem (2.6) is called the Euler–Lagrange
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
ﬁnd u ∈ H
1
0
(Ω) such that a(u, v) = l(v) ∀v ∈ H
1
0
(Ω) (W)
of the selfadjoint elliptic boundary value problem (2.7) to the associated minimisa
tion problem:
ﬁnd u ∈ H
1
0
(Ω) such that J(u) ≤ J(v) ∀v ∈ H
1
0
(Ω). (M)
We shall now use this equivalence to give a variational characterisation of the
ﬁnite element approximation u
h
to u in the selfadjoint case. Given that V
h
is a
certain ﬁnitedimensional subspace of H
1
0
(Ω) which consists of continuous piecewise
polynomials of a ﬁxed degree, the ﬁnite element approximation of (W) is:.
ﬁnd 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
(Ω) by V
h
throughout) to show the equivalence of (W
h
) to the following minimisation prob
lem:
ﬁnd 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 ﬁnite element space V
h
. This means that the ﬁnite element
solution u
h
inherits the energy minimisation property possessed by the weak solution
u ∈ H
1
0
(Ω) 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 stiﬀness matrix
Using the variational characterisation of u
h
described at the end of the previous
section we return to the construction of the ﬁnite element approximation to Poisson’s
equation −∆u = f in Ω subject to homogeneous Dirichlet boundary condition, u = 0
on ∂Ω, in the case of a general triangulation. Rather than restricting ourselves to
the special case when Ω is a square, we now suppose that Ω 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
deﬁned on such a triangulation
with the property that v
h
= 0 of ∂Ω; 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
_
Ω
[∇v
h
(x, y)[
2
dxdy −
_
Ω
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
φ
i
(x, y),
where V
i
is the value of v
h
(x, y) at the node (x
i
, y
i
), φ
i
is the continuous piecewise
linear basis function associated with this node, and N is the number of nodes internal
to Ω, we can write this minimisation problem in matrix form as follows:
ﬁnd 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) stiﬀness matrix  an N N matrix
with (i, j) entry
a(φ
i
, φ
j
) = (∇φ
i
, ∇φ
j
) =
_
Ω
∇φ
i
(x, y) ∇φ
j
(x, y) dxdy,
and F = (F
1
, . . . , F
N
)
T
is the (global) load vector, with
F
i
= (f, φ
i
) =
_
Ω
f(x, y)φ
i
(x, y) dxdy.
Consider any triangle K in the triangulation of Ω, and introduce the position
vectors r
i
= (x
i
, y
i
), i = 1, 2, 3, of its three vertices labelled in the anticlockwise
direction, say. In addition, we consider a socalled local (ξ, η) 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 −ξ −η)r
1
+ ξr
2
+ ηr
3
(2.12)
≡ r
1
ψ
1
(ξ, η) +r
2
ψ
2
(ξ, η) +r
3
ψ
3
(ξ, η).
The set ¦ψ
1
, ψ
2
, ψ
3
¦ is called the nodal basis (or local basis) for the set of
linear polynomials in terms of the local coordinates. Consider the transformation
(ξ, η) → r = (x, y) deﬁned by (2.12) from the canonical triangle to the ‘global’
(x, y) coordinate system. Let J
T
denote the transpose of the Jacobi matrix J of this
transformation; thus,
J
T
=
_
∂(x, y)
∂(ξ, η)
_
T
=
_
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
η
ξ
(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[ = [J
T
[ = 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(ξ, η)) = V
1
ψ
1
(ξ, η) + V
2
ψ
2
(ξ, η) + V
3
ψ
3
(ξ, η), (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 stiﬀness matrix, we need the
gradient of v
h
in the global coordinate system; however, from (2.12) and the form
of the matrix J
T
we have that
_
_
∂v
h
∂ξ
∂v
h
∂η
_
_
= J
T
_
_
∂v
h
∂x
∂v
h
∂y
_
_
,
_
_
∂v
h
∂x
∂v
h
∂y
_
_
= (J
T
)
−1
_
_
∂v
h
∂ξ
∂v
h
∂η
_
_
. (2.15)
Consequently,
∂v
h
∂x
=
1
[J[
_
(y
3
−y
1
)
∂v
h
∂ξ
−(y
2
−y
1
)
∂v
h
∂η
_
(2.16)
∂v
h
∂y
=
1
[J[
_
−(x
3
−x
1
)
∂v
h
∂ξ
+ (x
2
−x
1
)
∂v
h
∂η
_
.
Hence
[J[
2
[∇v
h
[
2
= [r
3
−r
1
[
2
_
∂v
h
∂ξ
_
2
+[r
2
−r
1
[
2
_
∂v
h
∂η
_
2
−2(r
3
−r
1
) (r
2
−r
1
)
∂v
h
∂ξ
∂v
h
∂η
(2.17)
38 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
and from (2.14) and (2.12) it follows that
∂v
h
∂ξ
= V
2
−V
1
,
∂v
h
∂η
= 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
_
Ω
[∇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 ﬁnd that the coeﬃcient 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 coeﬃcient 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 coeﬃcients 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 stiﬀness 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 stiﬀness 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 Ω; as
u
h
(x, y) =
N
i=1
U
i
φ
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 Ω (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
φ
i
(x, y),
with the understanding that the coeﬃcients 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 deﬁned 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
ﬁrst 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 stiﬀness matrix A
∗
is an
N
∗
N
∗
matrix deﬁned 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 Ω; 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 Ω, 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 stiﬀness 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 stiﬀness 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:
ﬁnd V
∗
= (V
1
, . . . , V
N
, 0, . . . , 0)
T
∈ R
N
∗
such that
1
2
V
∗T
A
∗
V
∗
−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 diﬃculties arise when we work with A
∗
and F
∗
.
It is worth noting that in practice it is not essential that the ﬁrst 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; C´ea’s lemma
Having described the construction of the ﬁnite element method, we now outline the
basic tools for its error analysis. Let us consider the elliptic boundary value problem
−
n
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 ∈ Ω, (2.20)
u = 0 on ∂Ω, (2.21)
where Ω is a bounded open set in R
n
, a
ij
∈ L
∞
(Ω), i, j = 1, . . . , n; b
i
∈ W
1
∞
(Ω),
i = 1, . . . , n, c ∈ L
∞
(Ω), f ∈ L
2
(Ω), and assume that there exists a positive constant
˜ c such that
n
i,j=1
a
ij
(x)ξ
i
ξ
j
≥ ˜ c
n
i=1
ξ
2
i
∀ξ = (ξ
1
, . . . , ξ
n
) ∈ R
n
, ∀x ∈
¯
Ω. (2.22)
The weak formulation of (2.20), (2.21) is:
ﬁnd u ∈ H
1
0
(Ω) such that a(u, v) = l(v) ∀v ∈ H
1
0
(Ω), (2.23)
where the bilinear functional a(, ) and the linear functional l() are deﬁned by
a(u, v) =
n
i,j=1
_
Ω
a
ij
∂u
∂x
i
∂v
∂x
j
dx +
n
i=1
_
Ω
b
i
(x)
∂u
∂x
i
v dx +
_
Ω
c(x)uv dx,
and
l(v) =
_
Ω
f(x)v(x) dx.
2.4. GALERKIN ORTHOGONALITY; C
´
EA’S LEMMA 41
We have shown that if
c(x) −
1
2
n
i=1
∂b
i
∂x
i
≥ 0, x ∈
¯
Ω,
then (2.23) has a unique solution u in H
1
0
(Ω), the weak solution of (2.20), (2.21).
Moreover,
u
H
1
(Ω)
≤
1
c
0
f
L
2
(Ω)
,
where c
0
is as in (1.17).
Now suppose that V
h
is a ﬁnitedimensional subspace of H
1
0
(Ω), without making
further precise assumptions on the nature of V
h
(although we shall implicitly assume
that V
h
consists of continuous piecewise polynomials deﬁned on a subdivision of
“ﬁneness” h of the computational domain Ω). The ﬁnite element approximation of
(2.23) is:
ﬁnd 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
(Ω) it follows from the LaxMilgram 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 ﬁnite element methods. Since by (1.17),
with v = u −u
h
∈ H
1
0
(Ω) we have that
u −u
h

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

2
H
1
(Ω)
≤
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
(Ω)
u −v
h

H
1
(Ω)
.
Combining the last two inequalities, we deduce that
u −u
h

H
1
(Ω)
≤
c
1
c
0
u −v
h

H
1
(Ω)
for all v
h
∈ V
h
.
Thus we have proved the following result.
42 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
Lemma 5 (C´ea’s lemma) The ﬁnite element approximation u
h
to u ∈ H
1
0
(Ω), the
weak solution to the problem (2.20), (2.21), is the nearbest ﬁt to u in the norm
 
H
1
(Ω)
; i.e.,
u −u
h

H
1
(Ω)
≤
c
1
c
0
min
v
h
∈V
h
u −v
h

H
1
(Ω)
.
Remark 5 We shall prove in the next chapter that, for a typical ﬁnite element space
V
h
,
min
v
h
∈V
h
u −v
h

H
1
(Ω)
≤ 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 C´ea’s lemma we shall be able to deduce that
u −u
h

H
1
(Ω)
≤ C(u)
_
c
1
c
0
_
h
s
(2.26)
which is a bound of the global error e
h
= u−u
h
in terms of the meshsize 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 reﬁning the subdivision further and further, the
sequence of ﬁnite element solutions ¦u
h
¦
h
converges to u in the H
1
(Ω) 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 diﬃcult 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 meshsize h has to be taken extremely small before any reduction in
the size of the global error is observed. Suppose that Ω is a bounded open set in R
n
.
Consider the following boundary value problem:
−ε∆u +b ∇u = f in Ω,
u = 0 on ∂Ω,
where ε > 0, b = (b
1
, . . . , b
n
)
T
, with b
i
∈ W
1
∞
(Ω) for i = 1, . . . , n. For the sake
of simplicity, we shall suppose that div b ≤ 0 almost everywhere on Ω. Such prob
lems arise in the mathematical modelling of advectiondiﬀusion phenomena. When
advection dominates diﬀusion the socalled P´eclet number
Pe =
_
n
i=1
b
i

2
L
∞
(Ω)
_
1/2
ε
2.4. GALERKIN ORTHOGONALITY; C
´
EA’S 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
=
_
ε
2
+
n
i=1
b
i

2
L
∞
(Ω)
_
1/2
and
c
0
=
ε
(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
(Ω)
≤ (1 + c
2
⋆
)
1/2
(1 + Pe
2
)
1/2
C(u)h
s
. (2.27)
Thus, when ε << 1, the constant on the righthand 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 ε through u (typically C(u) >> 1 when ε << 1).
We shall not consider the ﬁnite element approximation of advectiondominated
diﬀusion 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 reﬂection of the fact that for advection
dominated diﬀusion equations conventional ﬁnite 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 meshsize h. ⋄
In order to put this example into perspective, we now discuss the other extreme
case, when b ≡ 0 on Ω: then c
1
= c
0
= ε, so C´ea’s lemma implies that
u −u
h

H
1
(Ω)
≤ min
v
h
∈V
h
u −v
h

H
1
(Ω)
.
In fact, since the lefthand side of this inequality cannot be strictly less than the
righthand side (this can be seen by choosing v
h
= u
h
on the right), it follows that
u −u
h

H
1
(Ω)
= min
v
h
∈V
h
u −v
h

H
1
(Ω)
,
so that u
h
is the best approximation to u from V
h
in the H
1
(Ω) norm. We shall
show that a result of this kind holds in a slightly more general setting, when the
problem is selfadjoint, namely a
ij
(x) ≡ a
ji
(x) for all i, j = 1, . . . , n, b
i
(x) ≡ 0 for
i = 1, . . . , n. Let us deﬁne
(v, w)
a
:= a(v, w), v, w ∈ H
1
0
(Ω).
44 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS

6
>
u
h
0
H
1
0
(Ω)
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
(Ω) H
1
0
(Ω) and
a(v, v) ≥ c
0
v
2
H
1
(Ω)
∀v ∈ H
1
0
(Ω),
it is easily seen that (, )
a
satisﬁes all axioms of an inner product. Let  
a
denote
the associated energy norm deﬁned by:
v
a
:= [a(v, v)]
1/2
.
Since V
h
⊂ H
1
0
(Ω), 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 selfadjoint case, the error u −u
h
between the exact solution u and its
ﬁnite 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 CauchySchwarz 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 reﬁnement of C´ea’s lemma in the selfadjoint
case.
Lemma 6 The ﬁnite element approximation u
h
∈ V
h
of u ∈ H
1
0
(Ω) is the best ﬁt 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
.
C´ea’s lemma is the key to the error analysis of the ﬁnite element method for
elliptic boundary value problems. In the next section we describe how such an
analysis proceeds in the selfadjoint case, for a particularly simple ﬁnite element
space V
h
consisting of continuous piecewise linear functions on Ω. 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 C´ea’s lemma to derive an optimal error bound for
the ﬁnite element approximation (W
h
) of problem (W) in the case of piecewise linear
basis functions. We shall consider two examples: a onedimensional model problem
– a twopoint boundary value problem, and a twodimensional model problem –
Poisson’s equation subject to homogeneous Dirichlet boundary condition.
Onedimensional 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:
ﬁnd 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
deﬁned 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 ﬁnite 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
N−1
< x
N
= 1,
where the meshpoints x
i
, i = 0, . . . , N, are not necessarily equally spaced. It will be
supposed that N ≥ 2 so that we have at least one meshpoint inside (0, 1). We put
h
i
= x
i
−x
i−1
and deﬁne the mesh parameter h = max
i
h
i
. For such a subdivision,
we consider the ﬁnite element basis function
φ
i
(x) =
_
¸
¸
_
¸
¸
_
0 if x ≤ x
i−1
(x −x
i−1
)/h
i
if x
i−1
≤ 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¦φ
1
, . . . , φ
N−1
¦.
Clearly V
h
is an (N − 1)dimensional subspace of H
1
0
(0, 1). We approximate the
boundary value problem by the ﬁnite element method
ﬁnd 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 C´ea’s 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 meshpoints x
i
, i = 0, . . . , N. Thus,
J
h
u(x) =
N−1
i=1
u(x
i
)φ
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 ﬁnite 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
π
_
1 +
h
2
π
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
ﬁnite element space V
h
deﬁned above; then the following error bounds hold:
u −J
h
u
L
2
(0,1)
≤
_
h
π
_
2
u
′′

L
2
(0,1)
,
u
′
−(J
h
u)
′

L
2
(0,1)
≤
h
π
u
′′

L
2
(0,1)
.
Proof Consider a subinterval [x
i−1
, x
i
], 1 ≤ i ≤ N, and deﬁne ζ(x) = u(x) − J
h
u(x)
for x ∈ [x
i−1
, x
i
]. Then ζ ∈ H
2
(x
i−1
, x
i
) and ζ(x
i−1
) = ζ(x
i
) = 0. Therefore ζ can be
expanded into a convergent Fourier sineseries,
ζ(x) =
∞
k=1
a
k
sin
kπ(x −x
i−1
)
h
i
, x ∈ [x
i−1
, x
i
].
Hence,
_
x
i
x
i−1
[ζ(x)]
2
dx =
h
i
2
∞
k=1
[a
k
[
2
.
Diﬀerentiating the Fourier sineseries for ζ twice, we deduce that the Fourier coeﬃcients
of ζ
′
are (kπ/h
i
)a
k
, while those of ζ
′′
are −(kπ/h
i
)
2
a
k
. Thus,
_
x
i
x
i−1
[ζ
′
(x)]
2
dx =
h
i
2
∞
k=1
_
kπ
h
i
_
2
[a
k
[
2
,
_
x
i
x
i−1
[ζ
′′
(x)]
2
dx =
h
i
2
∞
k=1
_
kπ
h
i
_
4
[a
k
[
2
.
Because k
4
≥ k
2
≥ 1, it follows that
_
x
i
x
i−1
[ζ(x)]
2
dx ≤
_
h
i
π
_
4
_
x
i
x
i−1
[ζ
′′
(x)]
2
dx,
_
x
i
x
i−1
[ζ
′
(x)]
2
dx ≤
_
h
i
π
_
2
_
x
i
x
i−1
[ζ
′′
(x)]
2
dx.
48 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
However ζ
′′
(x) = u
′′
(x) − (J
h
u)
′′
(x) = u
′′
(x) for x ∈ (x
i−1
, 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
ζ
2
L
2
(0,1)
≤
_
h
π
_
4
u
′′

2
L
2
(0,1)
,
ζ
′

2
L
2
(0,1)
≤
_
h
π
_
2
u
′′

2
L
2
(0,1)
.
After taking the square root and recalling that ζ = 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
π
2
_
1 +
h
2
π
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
π
_
1 +
h
2
π
2
_
1/2
u
′′

L
2
(0,1)
. (2.34)
This shows that, provided u
′′
∈ L
2
(0, 1), the error in the ﬁnite element solution,
measured in the H
1
(0, 1) norm, converges to 0 at the rate O(h) as h → 0.
As a ﬁnal 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 diﬀerential 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)
≤ 2f
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
π
_
1 +
h
2
π
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 ﬁnite
element (a priori) error analysis. In a nutshell, it consisted of using:
a) C´ea’s 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 (multidimensional) elliptic boundary value problems proving an
elliptic regularity estimate of the form
[u[
H
2
(Ω)
≤ C
∗
f
L
2
(Ω)
(2.36)
is a highly nontrivial task (this issue will be touched on in the next section, in dis
cussion about the AubinNitsche duality argument). In fact, for multidimensional
problems (2.36) will not hold unless the boundary ∂Ω and the coeﬃcients a
ij
, b
i
and
c are suﬃciently 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 ﬁnite element approximation. Later on in the notes
we shall discuss an alternative approach, a posteriori error analysis, which resolves
this diﬃculty and provides computable bounds on the error in terms of u
h
.
Twodimensional problem
Let Ω = (0, 1) (0, 1), and consider the elliptic boundary value problem
−∆u = f in Ω, (2.37)
u = 0 on ∂Ω. (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
¯
Ω = [0, 1] [0, 1].
We recall that the weak formulation of this problem is:
ﬁnd u ∈ H
1
0
(Ω) such that
_
Ω
_
∂u
∂x
∂v
∂x
+
∂u
∂y
∂v
∂y
_
dxdy =
_
Ω
fv dxdy ∀v ∈ H
1
0
(Ω). (2.39)
In order to construct the ﬁnite element approximation, we triangulate the domain
as shown in the Fig 2.10. Let h = 1/N, and deﬁne x
i
= ih, i = 0, . . . , N, y
j
= jh,
j = 0, . . . , N. With each node, (x
i
, y
j
), contained in the interior of Ω (labelled ⊙ in
the ﬁgure), we associate a basis function φ
ij
, i, j = 1, . . . , N −1, deﬁned by
φ
ij
(x, y) =
_
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
_
1 −
x−x
i
h
−
y−y
j
h
, (x, y) ∈ 1
1 −
y−y
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 −
x−x
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
i−1
, y
j+1
)
(x
i−1
, y
j
)
(x
i
, y
j−1
)
(x
i+1
, y
j−1
)
(x
i+1
, y
j
)
Figure 2.11: Triangles surrounding the node (x
i
, y
j
).
Let V
h
= span¦φ
ij
, i = 1, . . . , N − 1; j = 1, . . . , N − 1¦. The ﬁnite element
approximation of (2.37) (and (2.39)) is:
ﬁnd u
h
∈ V
h
such that
_
Ω
_
∂u
h
∂x
∂v
h
∂x
+
∂u
h
∂y
∂v
h
∂y
_
dxdy =
_
Ω
fv
h
dxdy ∀v
h
∈ V
h
. (2.40)
Letting
l(v) =
_
Ω
f(x)v(x) dx and
(v, w)
a
= a(v, w) =
_
Ω
_
∂v
∂x
∂w
∂x
+
∂v
∂y
∂w
∂y
_
dxdy,
(2.39) and the ﬁnite element method (2.40) can be written, respectively, as follows:
ﬁnd u ∈ H
1
0
(Ω) such that a(u, v) = l(v) ∀v ∈ H
1
0
(Ω), (5.13
′
)
and
ﬁnd u
h
∈ V
h
such that a(u
h
, v
h
) = l(v
h
) ∀v
h
∈ V
h
. (5.14
′
)
According to C´ea’s 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
¯
Ω = [0, 1] [0, 1]:
(J
h
u)(x, y) =
N−1
i=1
N−1
j=1
u(x
i
, y
j
)φ
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
=
_
Ω
[
∂
∂x
(u −J
h
u)[
2
dxdy +
_
Ω
[
∂
∂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 Ω. 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 deﬁne the canonical triangle
K = ¦(s, t) : 0 ≤ s ≤ 1, 0 ≤ t ≤ 1 −s¦
and the aﬃne 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
_
1−s
0
[
∂¯ u
∂s
(s, t) −[¯ u(1, 0) − ¯ u(0, 0)][
2
ds dt
=
_
1
0
_
1−s
0
[
∂¯ u
∂s
(s, t) −
_
1
0
∂¯ u
∂s
(σ, 0) dσ[
2
ds dt
=
_
1
0
_
1−s
0
[
_
1
0
_
∂¯ u
∂s
(s, t) −
∂¯ u
∂s
(σ, t)
_
dσ +
_
1
0
_
∂¯ u
∂s
(σ, t) −
∂¯ u
∂s
(σ, 0)
_
dσ[
2
ds dt
=
_
1
0
_
1−s
0
[
_
1
0
_
s
σ
∂
2
¯ u
∂s
2
(θ, t) dθ dσ +
_
1
0
_
t
0
∂
2
¯ u
∂s ∂t
(σ, η) dη dσ[
2
ds dt
≤ 2
_
1
0
_
1−s
0
_
1
0
_
1
0
[
∂
2
¯ u
∂s
2
(θ, t)[
2
dθ dσ ds dt
+2
_
1
0
_
1−s
0
_
1
0
_
1
0
[
∂
2
¯ u
∂s ∂t
(σ, η)[
2
dη dσ ds dt
≤ 2
_
1
0
_
1
0
[
∂
2
¯ u
∂s
2
(θ, t)[
2
dθ dt +
_
1
0
_
1
0
[
∂
2
¯ u
∂s ∂t
(σ, η)[
2
dσ dη
= 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
_
Ω
_
[
∂
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
(Ω)
. (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 ﬁnite element approximation deﬁned by (2.40). Suppose
that u ∈ H
2
(Ω) ∩ H
1
0
(Ω); then
u −u
h

a
≤ 2h[u[
H
2
(Ω)
.
Corollary 2 Under the hypotheses of Theorem 4,
u −u
h

H
1
(Ω)
≤
√
5h[u[
H
2
(Ω)
.
Proof According to Theorem 4,
u −u
h

2
a
= [u −u
h
[
2
H
1
(Ω)
≤ 4h
2
[u[
2
H
2
(Ω)
.
Since u ∈ H
1
0
(Ω), u
h
∈ V
h
⊂ H
1
0
(Ω), it follows that u − u
h
∈ H
1
0
(Ω). By the Poincar´e
Friedrichs inequality,
u −u
h

2
L
2
(Ω)
≤
1
4
[u −u
h
[
2
H
1
(Ω)
; (2.47)
thus,
u −u
h

2
H
1
(Ω)
= u −u
h

2
L
2
(Ω)
+[u −u
h
[
2
H
1
(Ω)
≤
5
4
[u −u
h
[
2
H
1
(Ω)
≤ 5h
2
[u[
2
H
2
(Ω)
,
and that completes the proof.
From (2.47) and (2.46) we also see that
u −u
h

L
2
(Ω)
≤ h[u[
H
2
(Ω)
. (2.48)
The AubinNitsche duality argument. The error estimate (2.48) indicates
that the error in the L
2
norm between u and its ﬁnite 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 ﬁrst observe that if w ∈ H
2
(Ω) ∩ H
1
0
(Ω), Ω = (0, 1) (0, 1), then
∆w
2
L
2
(Ω)
=
_
Ω
_
∂
2
w
∂x
2
+
∂
2
w
∂y
2
_
2
dxdy
=
_
Ω
_
∂
2
w
∂x
2
_
2
+ 2
_
Ω
∂
2
w
∂x
2
∂
2
w
∂y
2
dxdy +
_
Ω
_
∂
2
w
∂y
2
_
2
dxdy.
Performing integration by parts and using the fact that w = 0 on ∂Ω,
_
Ω
∂
2
w
∂x
2
∂
2
w
∂y
2
dxdy =
_
Ω
∂
2
w
∂x∂y
∂
2
w
∂x∂y
dxdy
=
_
Ω
[
∂
2
w
∂x∂y
[
2
dxdy.
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 55
Thus
∆w
2
L
2
(Ω)
=
_
Ω
_
[
∂
2
w
∂x
2
[
2
+ 2[
∂
2
w
∂x∂y
[
2
+[
∂
2
w
∂y
2
[
2
_
dxdy
= [w[
2
H
2
(Ω)
.
Given g ∈ L
2
(Ω), let w
g
∈ H
1
0
(Ω) be the weak solution of the boundary value
problem
−∆w
g
= g in Ω, (2.49)
w
g
= 0 on ∂Ω; (2.50)
then w
g
∈ H
2
(Ω) ∩ H
1
0
(Ω), and
[w
g
[
H
2
(Ω)
= ∆w
g

L
2
(Ω)
= g
L
2
(Ω)
. (2.51)
After this brief preparation, we turn to the derivation of the optimal error bound in
the L
2
norm.
According to the CauchySchwarz inequality for the L
2
inner product (, ),
(u −u
h
, g) ≤ u −u
h

L
2
(Ω)
g
L
2
(Ω)
∀g ∈ L
2
(Ω).
Therefore,
u −u
h

L
2
(Ω)
= sup
g∈L
2
(Ω)
(u −u
h
, g)
g
L
2
(Ω)
. (2.52)
For g ∈ L
2
(Ω), the function w
g
∈ H
1
0
(Ω) is the weak solution of the problem (2.49),
so it satisﬁes
a(w
g
, v) = l
g
(v) ∀v ∈ H
1
0
(Ω), (2.53)
where
l
g
(v) =
_
Ω
gv dxdy = (g, v),
a(w
g
, v) =
_
Ω
_
∂w
g
∂x
∂v
∂x
+
∂w
g
∂y
∂v
∂y
_
dxdy.
Consider the ﬁnite element approximation of (2.53):
ﬁnd 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
(Ω)
,
56 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
and therefore, by (2.51),
w
g
−w
gh

a
≤ 2hg
L
2
(Ω)
. (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 CauchySchwarz 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
(Ω)
g
L
2
(Ω)
. (2.57)
Substituting (2.57) into the righthand side of (2.52), we obtain
u −u
h

L
2
(Ω)
≤ 4h
2
[u[
H
2
(Ω)
,
which is our improved error bound in the L
2
norm. The proof presented above is
called the Aubin–Nitsche duality argument.
2.6 Superapproximation in meshdependent norms
We have shown that the piecewise linear ﬁnite element approximation u
h
to the solu
tion u of the homogeneous Dirichlet boundary value problem for Poisson’s equation
obeys the following error bounds:
u −u
h

H
1
(Ω)
≤ Ch[u[
H
2
(Ω)
,
u −u
h

L
2
(Ω)
≤ Ch
2
[u[
H
2
(Ω)
,
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 MESHDEPENDENT NORMS 57
that, when sampled at certain special points, the ﬁnite 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
(Ω) 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 Ω, (2.58)
u = 0 on ∂Ω, (2.59)
where Ω = (0, 1) (0, 1). We showed in Section 2.1 that when using continuous
piecewise linear ﬁnite elements on the uniform triangulation shown in Figure 2.5,
the ﬁnite element solution u
h
(x, y) can be expressed in terms of the ﬁnite element
basis functions φ
ij
(x, y) as
u
h
(x, y) =
N−1
i=1
N−1
j=1
U
ij
φ
ij
(x, y),
where the U
ij
(= u
h
(x
i
, y
j
)) are obtained by solving the set of diﬀerence equations
−
U
i+1,j
−2U
ij
+ U
i−1,j
h
2
−
U
i,j+1
−2U
ij
+ U
i,j−1
h
2
=
1
h
2
_ _
supp φ
ij
f(x, y)φ
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) ∈ ∂Ω, 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 φ
ij
f(x, y)φ
ij
(x, y) dxdy, for i, j = 1, . . . , N −1.
Here N is an integer, N ≥ 2, and h = 1/N; the meshpoints are (x
i
, y
j
), i, j =
0, . . . , N, where x
i
= ih, y
j
= jh. These form the ﬁnite diﬀerence mesh
¯
Ω
h
= ¦(x
i
, y
j
) : i, j = 0, . . . , N¦.
We consider the set of interior mesh points
Ω
h
= ¦(x
i
, y
j
) : i, j = 1, ..., N −1¦,
and the set of boundary mesh points Γ
h
=
¯
Ω
h
¸ Ω
h
. In more compact notation, the
diﬀerence scheme can be written as follows:
−(D
+
x
D
−
x
U
ij
+ D
+
y
D
−
y
U
ij
) = F
ij
, (x
i
, y
j
) ∈ Ω
h
, (2.60)
U = 0 on Γ
h
, (2.61)
58 CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
(i,j−1)
(i,j)
(i,j+1)
(i−1,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 Ω
h
(), the boundary mesh Γ
h
(), and a typical 5point
diﬀerence stencil.
where D
+
x
and D
−
x
denote the forward and backward divided diﬀerence operators in
the x direction, respectively, deﬁned by
D
+
x
V
ij
=
V
i+1,j
−V
ij
h
, D
−
x
V
ij
=
V
ij
−V
i−1,j
h
,
and
D
+
x
D
−
x
V
ij
= D
+
x
(D
−
x
V
ij
) =
V
i+1,j
−2V
ij
+ V
i−1,j
h
2
is the second divided diﬀerence operator in the x direction. Similar deﬁnitions apply
in the y direction.
For each i and j, 1 ≤ i, j ≤ N −1, the ﬁnite diﬀerence equation (2.60) involves
ﬁve values of U: U
i,j
, U
i−1,j
, U
i+1,j
, U
i,j−1
, 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,N−1
, U
21
, U
22
, . . . , U
2,N−1
, . . . ,
. . . , U
i1
, U
i2
, . . . , U
i,N−1
, . . . , U
N−1,1
, U
N−1,2
, . . . , U
N−1,N−1
)
T
,
2.6. SUPERAPPROXIMATION IN MESHDEPENDENT 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,N−1
, F
21
, F
22
, . . . , F
2,N−1
, . . . ,
. . . , F
i1
, F
i2
, . . . , F
i,N−1
, . . . , F
N−1,1
, F
N−1,2
, . . . , F
N−1,N−1
)
T
,
and A is an (N −1)
2
(N −1)
2
sparse matrix of band structure. A typical row of
the matrix contains ﬁve nonzero entries, corresponding to the ﬁve values of U in
the ﬁnite diﬀerence 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,
deﬁned on Ω
h
, we introduce the inner product
(V, W)
h
=
N−1
i=1
N−1
j=1
h
2
V
ij
W
ij
(which resembles the L
2
inner product (v, w) =
_
Ω
v(x, y)w(x, y) dxdy.)
Lemma 7 Suppose that V is a function deﬁned on
¯
Ω
h
and that V = 0 on Γ
h
; then
(−D
+
x
D
−
x
V, V )
h
+ (−D
+
y
D
−
y
V, V )
h
=
N
i=1
N−1
j=1
h
2
[D
−
x
V
ij
[
2
+
N−1
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 ﬁnite 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 ﬁrst term on the left is equal to the ﬁrst 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
= −
N−1
i=1
N−1
j=1
(V
i+1,j
−2V
ij
+V
i−1,j
)V
ij
= −
N−1
i=1
N−1
j=1
(V
i+1,j
−V
ij
)V
ij
+
N−1
i=1
N−1
j=1
(V
ij
−V
i−1,j
)V
ij
= −
N
i=2
N−1
j=1
(V
ij
−V
i−1,j
)V
i−1,j
+
N−1
i=1
N−1
j=1
(V
ij
−V
i−1,j
)V
ij
= −
N
i=1
N−1
j=1
(V
ij
−V
i−1,j
)V
i−1,j
+
N
i=1
N−1
j=1
(V
ij
−V
i−1,j
)V
ij
=
N
i=1
N−1
j=1
(V
ij
−V
i−1,j
)
2
=
N
i=1
N−1
j=1
h
2
[D
−
x
V
ij
[
2
.
Similarly,
(−D
+
y
D
−
y
V, V )
h
=
N−1
i=1
N
j=1
h
2
[D
−
y
V
ij
[
2
,
and that completes the proof.
Returning to the analysis of the ﬁnite diﬀerence 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
N−1
j=1
h
2
[D
−
x
V
ij
[
2
+
N−1
i=1
N
j=1
h
2
[D
−
y
V
ij
[
2
, (2.64)
for any V deﬁned on
¯
Ω
h
such that V = 0 on Γ
h
. Now this implies that A is a
nonsingular matrix. Indeed if AV = 0, then (2.64) yields:
D
−
x
V
ij
=
V
ij
−V
i−1,j
h
= 0,
i = 1, . . . , N,
j = 1, . . . , N −1;
D
−
y
V
ij
=
V
ij
−V
i,j−1
h
= 0,
i = 1, . . . , N −1,
j = 1, . . . , N.
Since V = 0 on Γ
h
, these imply that V ≡ 0. Thus AV = 0 if and only if V = 0.
Hence A is nonsingular, and U = A
−1
F is the unique solution of (2.60). In summary
then, the (unique) solution of the ﬁnite diﬀerence scheme (2.60) may be found by
solving the system of linear equations (2.62).
2.6. SUPERAPPROXIMATION IN MESHDEPENDENT NORMS 61
In order to prove the stability of the ﬁnite diﬀerence scheme (2.60), we introduce
the mesh–dependent 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
N−1
j=1
h
2
[D
−
x
U
ij
[
2
_
1/2
and
D
−
y
U][
y
=
_
N−1
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
(Ω)
,
u
H
1
(Ω)
=
_
u
2
L
2
(Ω)
+
∂u
∂x

2
L
2
(Ω)
+
∂u
∂y

2
L
2
(Ω)
_
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 Poincar´eFriedrichs 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 Poincar´eFriedrichs inequality) Let V be a function deﬁned on
¯
Ω
h
and such that V = 0 on Γ
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
N−1
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 deﬁnition 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 CauchySchwarz 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 deﬁne 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) ∈ ∂Ω, 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 MESHDEPENDENT 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 φ
ij
φ
ij
(x, y)f(x, y) dxdy
=
_
1
h
2
_ _
supp φ
ij
φ
ij
(x, y)
∂
2
u
∂x
2
(x, y) dxdy −D
+
x
D
−
x
u(x
i
, y
j
)
_
+
_
1
h
2
_ _
supp φ
ij
φ
ij
(x, y)
∂
2
u
∂y
2
(x, y) dxdy −D
+
y
D
−
y
u(x
i
, y
j
)
_
≡ ϕ
ij
.
Thus,
Ae
ij
= ϕ
ij
, 1 ≤ i, j ≤ N −1,
e = 0 on Γ
h
.
By virtue of (2.68),
u −U
1,h
= e
1,h
≤
1
c
0
ϕ
h
. (2.69)
Assuming that u ∈ C
4
(
¯
Ω) and employing a Taylor series expansion of u(x, y) about
(x
i
, y
j
), we deduce that
[ϕ
ij
[ ≤ K
0
h
2
_

∂
4
u
∂x
4

C(
¯
Ω)
+
∂
4
u
∂y
4

C(
¯
Ω)
_
,
where K
0
is a positive constant independent of h. Thus,
ϕ
h
≤ K
0
h
2
_

∂
4
u
∂x
4

C(
¯
Ω)
+
∂
4
u
∂y
4

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

1,h
≤
5
4
K
0
h
2
_

∂
4
u
∂x
4

C(
¯
Ω)
+
∂
4
u
∂y
4

C(
¯
Ω)
_
. (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 ﬁnite 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
(
¯
Ω). Since this exceeds the ﬁrst order of convergence of
the global error observed in the Sobolev norm  
H
1
(Ω)
, the result encapsulated in
Theorem 6 is referred to as a superapproximation property. In fact the smoothness
requirement u ∈ H
1
0
(Ω)∩C
4
(
¯
Ω) can be relaxed to u ∈ H
1
0
(Ω)∩H
3
(Ω) while retaining
the superapproximation property u − u
h

1,h
= O(h
2
); the proof of this is more
technical and relies on the BrambleHilbert lemma (See Chapter 3).
Chapter 3
Piecewise polynomial
approximation
In the previous chapter we discussed ﬁnite 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 ﬁnite element
spaces and to formalise the concepts introduced in Chapter 2.
3.1 Construction of ﬁnite element spaces
Let us consider an elliptic boundary value problem written in its weak formulation:
ﬁnd u in V such that a(u, v) = l(v) ∀v ∈ V ,
where H
1
0
(Ω) ⊂ V ⊂ H
1
(Ω); in the case of a homogeneous Dirichlet boundary value
problem V = H
1
0
(Ω), in the case of a Neumann, Robin or oblique derivative bound
ary value problem, V = H
1
(Ω). In order to deﬁne a ﬁnite element approximation
to this problem we need to construct a ﬁnitedimensional subspace V
h
of V consist
ing of continuous piecewise polynomial functions of a certain degree deﬁned on a
subdivision of the computational domain Ω. 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 ﬁnite element
We begin by giving a formal deﬁnition of a ﬁnite element.
Deﬁnition 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 ﬁnitedimensional space of functions deﬁned 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 ﬁnite element.
In this deﬁnition T
′
denotes the algebraic dual of the linear space T.
Deﬁnition 3 Let (K, T, A) be a ﬁnite element, and let ¦ψ
1
, ψ
2
, . . . , ψ
k
¦ be a basis
for T, dual to A; namely,
N
i
(ψ
j
) = δ
ij
, 1 ≤ i, j ≤ k.
Such a basis is called a nodal basis for T.
We give a simple example to illustrate these deﬁnitions.
Example 7 (The onedimensional 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 ﬁnite element, with nodal basis ¦ψ
1
, ψ
2
¦ where
ψ
1
(x) = 1 −x and ψ
2
(x) = x. ⋄
Next we give an equivalent characterisation of condition (iii) in Deﬁnition 2.
Lemma 9 Let T be a kdimensional 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 ¦ψ
1
, . . . , ψ
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 = α
1
N
1
+. . . +α
k
N
k
.
This is equivalent to demanding that, for each i = 1, . . . , k, L(ψ
i
) can be written in a
unique fashion as a linear combination
L(ψ
i
) = α
1
N
1
(ψ
i
) +. . . +α
k
N
k
(ψ
i
).
Let us deﬁne the matrix B = (N
j
(ψ
i
))
i,j=1,...,k
and the vectors
y = (L(ψ
1
), . . . , L(ψ
k
))
T
, a = (α
1
, . . . , α
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 = β
1
ψ
1
+. . . +β
k
ψ
k
.
3.1. CONSTRUCTION OF FINITE ELEMENT SPACES 67
Now N
i
(v) = 0 for all i = 1, . . . , k if and only if
β
1
N
i
(ψ
1
) +. . . +β
k
N
i
(ψ
k
) = 0, i = 1, . . . , k. (3.1)
Thus (b) is equivalent to requiring that (3.1) implies β
1
= . . . = β
k
= 0. Let C =
(N
i
(ψ
j
))
i,j=1,...,k
; then (b) holds if and only if Cb = 0, with b = (β
1
, . . . , β
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 deﬁnition.
Deﬁnition 4 We say that A determines T if ψ ∈ T with N(ψ) = 0 for all N ∈ A
implies that ψ = 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 nondegenerate linear function. Then we
can write P in the factorised form P = LQ where Q is a polynomial of degree (d−1).
Proof Let us write ˆ x = (x
1
, . . . , x
n−1
). Suppose that we have carried out an aﬃne change
of variables such that L(ˆ x, x
n
) = x
n
; so, the hyperplane L(ˆ x, x
n
) = 0 is the hyperplane
x
n
= 0; then, by hypothesis, P(ˆ x, 0) ≡ 0. Since P is of degree d, we have that
P(ˆ x, x
n
) =
d
j=0
α≤d−j
c
αj
ˆ x
α
x
j
n
,
where α = (i
1
, . . . , i
n−1
) and ˆ x
α
= x
i
1
1
x
i
n−1
n−1
. Letting x
n
= 0 we get
0 ≡ P(ˆ x, 0) =
α≤d
c
α0
ˆ x
α
,
which implies that c
α0
= 0 for [α[ ≤ d. Hence
P(ˆ x, x
n
) =
d
j=1
α≤d−j
c
αj
ˆ x
α
x
j
n
= x
n
d
j=1
α≤d−j
c
αj
ˆ x
α
x
j−1
n
= x
n
Q = LQ
where Q is of degree (d −1).
3.1.2 Examples of triangular ﬁnite 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 ﬁgure • indicates function evaluation at the point where the dot is placed.
We verify condition (iii) of Deﬁnition 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 nontrivial linear functions
which deﬁne 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 nontrivial linear
functions which deﬁne 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 nondegenerate. 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 ﬁnally 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 ﬁnite elements, we now wish to piece them together
to construct ﬁnitedimensional subspaces of Sobolev spaces.
Deﬁnition 5 Let (K, T, A) be a ﬁnite element and let the set ¦ψ
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 deﬁned, we introduce the local interpolant J
K
v by
J
K
v =
k
i=1
N
i
(v)ψ
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 ﬁnd the local interpolant J
K
v of the function v deﬁned by v(x, y) =
(1 + x
2
+ y
2
)
−1
.
By deﬁnition,
J
K
v = N
1
(v)ψ
1
+ N
2
(v)ψ
2
+ N
3
(v)ψ
3
.
Thus we must determine ψ
i
, i = 1, 2, 3, to be able to write down the local interpolant.
This we do, using Deﬁnition 3, as follows. The line L
1
has equation y = 1 − x;
as ψ
1
vanishes at z
2
and z
3
, and thereby along the whole of L
1
, it follows that
ψ
1
= cL
1
= c(1 − x − y), where c is a constant to be determined. Also, N
1
ψ
1
= 1,
so c = ψ
1
(z
1
) = 1. Hence, ψ
1
= 1 −x −y. Similarly, ψ
2
= L
2
(x, y)/L
2
(z
2
) = x and
ψ
3
= L
3
(x, y)/L
3
(z
3
) = y.
Having found ψ
1
, ψ
2
and ψ
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)ψ
j
_
_
=
k
j=1
N
j
(v)N
i
(ψ
j
)
=
k
j=1
N
j
(v)δ
ij
= N
i
(v),
for i = 1, . . . , k, where δ
ij
= 1 when i = j and = 0 when i ,= j.
c) It follows from b) that N
i
(v−J
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.
Deﬁnition 6 A subdivision of the computational domain Ω is a ﬁnite collection
of open sets ¦K
i
¦ such that
(1) K
i
∩ K
j
= ∅ if i ,= j, and
(2)
i
¯
K
i
=
¯
Ω.
Deﬁnition 7 Suppose that Ω 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 ﬁnite el
ement. Let m be the order of the highest partial derivative involved in the nodal
variables. For v ∈ C
m
(
¯
Ω) the global interpolant J
h
v is deﬁned on
¯
Ω 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
¯
Ω. To keep the
presentation simple, we shall restrict ourselves to the case of two space dimensions,
namely when Ω ⊂ R
2
, although an analogous deﬁnition can me made in R
n
.
Deﬁnition 8 A triangulation of a polygonal domain Ω is a subdivision of Ω 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
Ω ⊂ R
2
: when Ω ⊂ R
n
and n = 2 we shall mean that the condition of this deﬁnition
is satisﬁed; when n > 2, the obvious generalisation of this condition to n dimensions
will be meant to hold.
Deﬁnition 9 We say that an interpolant has continuity of order r (or, brieﬂy,
that it is C
r
) if J
h
v ∈ C
r
(
¯
Ω) for all v ∈ C
m
(
¯
Ω). The space
¦J
h
v : v ∈ C
m
(
¯
Ω)¦
is called a C
r
ﬁnite 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 Ω, 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
(
¯
Ω) for all v in C
m
(
¯
Ω), where m = 0 for Lagrange and m = 1 for
Hermite elements.
Proof It suﬃces 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 deﬁned 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 onedimensional 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 diﬀerent ele
ments, we introduce the following deﬁnition (now for K ⊂ Ω ⊂ R
n
.)
Deﬁnition 10 Let (K, T, A) be a ﬁnite element and suppose that F(x) = Ax + b
where A is a nonsingular nn matrix and x and b are ncomponent column vectors.
The ﬁnite element (
ˆ
K,
ˆ
T,
ˆ
A) is aﬃne equivalent to (K, T, A) if:
(a) F(K) =
ˆ
K;
(b) F
∗
ˆ
T = T and
(c) F
∗
A =
ˆ
A.
Here F
∗
is the pullback of F deﬁned by F
∗
(ˆ v) = ˆ v ◦ F, and F
∗
is the pushforward
of F deﬁned by (F
∗
N)(ˆ v) = N(F
∗
(ˆ v)) = N(ˆ v ◦ F).
Example 12 Lagrange elements on triangles with appropriate choice of edge and
interior nodes are aﬃne equivalent. The same is true of Hermite elements on tri
angles. ⋄
3.1.4 Examples of rectangular elements
To conclude this section we consider ﬁnite elements deﬁned 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
signiﬁes 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 ﬁnite 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 HardyLittlewood maximal function, following
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 75
the work of Ricardo Dur´an
1
. An alternative approach which exploits the theory of
Riesz potentials is presented in the BrennerScott monograph cited in the reading
list.
3.2.1 The BrambleHilbert lemma
A key device in ﬁnite element error analysis is the following result.
Lemma 12 (BrambleHilbert lemma) Suppose that Ω is a bounded open set in R
n
and assume that Ω is starshaped with respect to every point in a set B of positive
measure contained in Ω (i.e. for all x ∈ Ω the closed convex hull of ¦x¦ ∪ B is
a subset of Ω). Let l be a bounded linear functional on the Sobolev space W
m
p
(Ω),
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
(Ω)
for all v ∈ W
m
p
(Ω).
Proof By hypothesis, there exists C
0
> 0 such that
[l(v)[ ≤ C
0
v
W
m
p
(Ω)
∀v ∈ W
m
p
(Ω).
Since l(Q) = 0 for all Q ∈ T
m−1
, we have by the linearity of l that
[l(v)[ = [l(v −Q)[ ≤ C
0
v −Q
W
m
p
(Ω)
= C
0
_
_
m
j=0
[v −Q[
p
W
j
p
(Ω)
_
_
1/p
= C
0
_
_
m−1
j=0
[v −Q[
p
W
j
p
(Ω)
+[v[
p
W
m
p
(Ω)
_
_
1/p
≤ C
0
m−1
j=0
[v −Q[
W
j
p
(Ω)
+[v[
W
m
p
(Ω)
.
In order to complete the proof it remains to prove that
∃K
j
> 0 ∀v ∈ W
j
p
(Ω) ∃Q ∈ T
m−1
such that
[v −Q[
W
j
p
(Ω)
≤ K
j
[v[
W
m
p
(Ω)
, j = 0, . . . , m−1. (3.2)
This will be done in the rest of the section. Once (3.2) has been veriﬁed, we shall have
that
[l(v)[ ≤ C
0
_
_
1 +
m−1
j=0
K
j
_
_
[v[
W
m
p
(Ω)
,
1
R. Dur´ an: On polynomial approximation in Sobolev spaces. SIAM Journal of Numerical
Analysis, 20, No. 5., (1983), pp. 985–988.
76 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
and the proof will be complete, with the constant
C
1
= C
0
_
_
1 +
m−1
j=0
K
j
_
_
.
The original proof of (3.2) given by Bramble and Hilbert was based on the use of
the HahnBanach theorem and was nonconstructive in nature in the sense that it
did not provide computable constants K
j
, j = 0, . . . , m−1; 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 ν ∈ R
n
such that [ν[ = 1, we deﬁne
g
1
(x, ν) = sup
t>0
1
t
_
t
0
[g(x + sν)[ ds
and
g
∗
(x) =
__
ν=1
g
1
(x, ν)
p
dσ
ν
_
1/p
.
Then
g
∗

L
p
(R
n
)
≤
p
p −1
ω
1/p
n
g
L
p
(R
n
)
,
where ω
n
is the measure of the unit sphere in R
n
.
Proof Since g
1
(, ν) is the HardyLittlewood maximal function of g() in the direction ν,
it follows that
2
_
R
n
g
1
(x, ν)
p
dx ≤
_
p
p −1
_
p
_
R
n
[g(x)[
p
dx,
and therefore
_
R
n
[g
∗
(x)[
p
=
_
R
n
_
_
ν=1
g
1
(x, ν)
p
dσ
ν
_
dx
=
_
ν=1
__
R
n
g
1
(x, ν)
p
dx
_
dσ
ν
≤ ω
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. Calder´on: Estimates for
singular integral operators in terms of maximal functions. Stud. Math. 44, (1972), pp.563–582.
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 77
Theorem 8 Let Ω ⊂ R
n
be a bounded open set which is starshaped with respect to
each point in a set of positive measure B ⊂ Ω. Let 1 < p < ∞, 0 ≤ j < m, and let
d be the diameter of Ω. If v ∈ W
m
p
(Ω) then
inf
Q∈P
m−1
[v −Q[
W
j
p
(Ω)
≤ C
d
m−j+(n/p)
[B[
1/p
[v[
W
m
p
(Ω)
,
where [B[ denotes the measure of B and
C = (♯¦α : [α[ = j¦)
m−j
n
1/p
p
p −1
ω
1/p
n
_
_
β=m−j
(β!)
−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 multiindex β = (β
1
, . . . , β
n
), β! = β
1
! . . . β
n
!.
Proof Because C
∞
(
¯
Ω) is dense
3
in W
m
p
(Ω), it suﬃces to prove the theorem for v ∈ C
∞
(
¯
Ω).
Given x ∈ B, we deﬁne
P
m
(v)(x, y) =
β<m
D
β
v(x)
(y −x)
β
β!
and
Q
m
(v)(y) =
1
[B[
_
B
P
m
(v)(x, y) dx.
Here we used the multiindex notation (y −x)
β
= (y
1
−x
1
)
β
1
. . . (y
n
−x
n
)
β
n
. It is easy
to prove by induction that
D
α
Q
m
(v)(y) = Q
m−α
(D
α
v)(y).
Thus,
[v −Q
m
(v)[
W
j
p
(Ω)
=
_
_
α=j
D
α
(v −Q
m
(v))
p
L
p
(Ω)
_
_
1/p
≤
α=j
D
α
(v −Q
m
(v))
L
p
(Ω)
=
α=j
D
α
v −Q
m−j
(D
α
v)
L
p
(Ω)
.
Now let us estimate D
α
v −Q
m−j
(D
α
v)
L
p
(Ω)
for each α, [α[ = j. As
(D
α
v −Q
m−j
(D
α
v))(y) =
1
[B[
_
B
[D
α
v(y) −P
m−j
(D
α
v)(x, y)] dx,
3
See R.A. Adams: Sobolev Spaces. Academic Press, 1975.
78 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
it follows, by applying Minkowski’s inequality for integrals
4
that
D
α
v −Q
m−j
(D
α
v)
L
p
(Ω)
≤
1
[B[
_
B
__
Ω
[D
α
v(y) −P
m−j
(D
α
v)(x, y)[
p
dy
_
1/p
dx. (3.3)
Now recalling the integralremainder for Taylor series, for x ∈ B, y ∈ Ω, we have that
[D
α
v(y) −P
m−j
(D
α
v)(x, y)[
=
¸
¸
¸
¸
¸
¸
(m−j)
β=m−j
(y −x)
β
β!
_
1
0
D
β
D
α
v(x +t(y −x))(1 −t)
m−j−1
dt
¸
¸
¸
¸
¸
¸
≤ (m−j)d
m−j
_
1
0
β=m−j
1
β!
[D
β
D
α
v(x +t(y −x))[ dt
= (m−j)d
m−j
1
[y −x[
_
y−x
0
β=m−j
1
β!
¸
¸
¸
¸
D
β
D
α
v
_
x +s
y −x
[y −x[
_¸
¸
¸
¸
ds.
Let g be the function that coincides with
β=m−j
1
β!
[D
β
D
α
v[ in Ω and is identically zero
outside Ω. If g
1
and g
∗
are the functions associated with g, as deﬁned in Lemma 13, we
have that
[D
α
v(y) −P
m−j
(D
α
v)(x, y)[ ≤ (m−j)d
m−j
g
1
_
x,
y −x
[y −x[
_
,
and therefore
[D
α
v(y) −P
m−j
(D
α
v)(x, y)[
p
≤ (m−j)
p
d
(m−j)p
g
p
1
_
x,
y −x
[y −x[
_
.
Noting that B ⊂ Ω, it follows for each x ∈ B that
_
Ω
[D
α
v(y) −P
m−j
(D
α
v)(x, y)[
p
dy ≤ (m−j)
p
d
(m−j)p
_
y−x≤d
g
p
1
_
x,
y −x
[y −x[
_
dy.
Thus,
__
Ω
[D
α
v(y) −P
m−j
(D
α
v)(x, y)[
p
dy
_
1/p
≤ (m−j)d
m−j
_
_
d
0
_
ν=1
g
p
1
(x, ν) dσ
ν
r
n−1
dr
_
1/p
= (m−j)d
m−j
_
d
n
n
_
1/p
_
_
ν=1
g
p
1
(x, ν) dσ
ν
_
1/p
= (m−j)d
m−j
_
d
n
n
_
1/p
g
∗
(x).
4
Minkowski’s integral inequality states that, for a function u ∈ C(
¯
B
¯
Ω),

_
B
u(x, ) dx
L
p
(Ω)
≤
_
B
u(x, )
L
p
(Ω)
dx.
3.2. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 79
Inserting this into (3.3) we get, by H¨older’s inequality (see Ch.1, Sec.1.1.2).
D
α
v −Q
m−j
(D
α
v)
L
p
(Ω)
≤
1
[B[
(m−j) d
m−j
_
d
n
n
_
1/p
_
B
g
∗
(x) dx
≤
1
[B[
(m−j) d
m−j
[B[
1−(1/p)
_
d
n
n
_
1/p
g
∗

L
p
(R
n
)
and hence, by Lemma 13,
D
α
v −Q
m−j
(D
α
v)
L
p
(Ω)
≤
m−j
n
1/p
d
m−j+(n/p)
[B[
1/p
p
p −1
ω
1/p
n
g
L
p
(R
n
)
=
m−j
n
1/p
d
m−j+(n/p)
[B[
1/p
p
p −1
ω
1/p
n
g
L
p
(Ω)
.
However,
g
L
p
(Ω)
= 
β=m−j
1
β!
[D
β
D
α
v[ 
L
p
(Ω)
≤
β=m−j
1
β!
D
β
D
α
v
L
p
(Ω)
≤
_
_
β=m−j
(β!)
−p
′
_
_
1/p
′ _
_
β=m−j
D
β
D
α
v
p
L
p
(Ω)
_
_
1/p
,
where 1/p + 1/p
′
= 1. Therefore,
D
α
v −Q
m−j
(D
α
v)
L
p
(Ω)
≤
m−j
n
1/p
d
m−j+(n/p)
[B[
1/p
p
p −1
ω
1/p
n
_
_
β=m−j
(β!)
−p
′
_
_
1/p
′
_
_
β=m−j
D
β
D
α
v
p
L
p
(Ω)
_
_
1/p
.
Recalling that [α[ = j, we deduce that
[v −Q
m
(v)[
W
j
p
(Ω)
≤ K
j
[v[
W
m
p
(Ω)
,
where
K
j
= C
d
m−j+(n/p)
[B[
1/p
.
Since Q
m
∈ T
m−1
, this completes the proof of Theorem 8, and thereby also the proof of
the BrambleHilbert lemma (Lemma 12).
Corollary 3 Let Ω ⊂ R
n
be a bounded open set of diameter d which is starshaped
with respect to every point of an open ball B ⊂ Ω of diameter µd, µ ∈ (0, 1]. Suppose
that 1 < p < ∞ and 0 ≤ j < m, m ≥ 1. If v ∈ W
m
p
(Ω) then
inf
Q∈P
m−1
[v −Q[
W
j
p
(Ω)
≤ C(m, n, p, j, µ)d
m−j
[v[
W
m
p
(Ω)
,
80 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
where
C(m, n, p, j, µ) = µ
−n/p
(♯¦α : [α[ = j¦)
p(m−j)
p −1
_
_
β=m−j
(β!)
−p
′
_
_
1/p
,
with 1/p + 1/p
′
= 1.
Proof Note that
[B[ =
(µd)
n
ω
n
n
.
In order to minimise the size of the constant C(m, n, p, j, µ), µ should be taken
as large as possible a number in the interval (0, 1] such that Ω is starshaped with
respect to each point in a ball B ⊂ Ω of radius µd.
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 ﬁnite element interpolant. We begin by estimating the norm of the local inter
polation operator.
Lemma 14 Let (K, T, A) be a ﬁnite 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 ¦ψ
1
, . . . , ψ
k
¦ ⊂ T be the basis dual to A. The
local interpolant of a function u is deﬁned by the formula
J
K
u =
k
i=1
N
i
(u)ψ
i
,
where each ψ
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)[ ψ
i

W
m
p
(K)
≤
_
k
i=1
N
i

(C
l
(
¯
K)
′ ψ
i

W
m
p
(K)
_
u
C
l
(
¯
K)
= Const. u
C
l
(
¯
K)
.
and that completes the proof.
We deﬁne
σ(K) = sup
v∈C
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 ﬁnite element satisfying the following conditions:
(i) K is starshaped with respect to some ball contained in K;
(ii) T
m−1
⊂ 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, µ, σ(
ˆ
K))h
m−j
K
[v[
W
m
p
(K)
,
where h
K
is the diameter of K,
ˆ
K = ¦x/h
K
: x ∈ K¦ and µ is the largest real
number in the interval (0, 1] such that a ball of diameter µh
K
is contained in K.
Proof It suﬃces to take K with diameter equal to 1, in which case K =
ˆ
K; the general
case follows by a simple scaling argument. Also, note that the local interpolation operator
is well deﬁned 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
m−1
⊂ 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)
+σ(K) Q
m
v −J
K
v
C
l
(
¯
K)
≤ (1 +C
m,n,p
σ(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, µ, σ(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, µ, σ(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 Ω, the constant C(m, n, p, µ, σ(
ˆ
K)) can be made
independent of σ(
ˆ
K).
5
The Sobolev embedding theorem asserts that W
m
p
(K) ⊂ C
l
(
¯
K) for m−l >
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 (
ˆ
K,
ˆ
T,
ˆ
A) is aﬃne equivalent to a single reference
element (K, T, A) through an aﬃne transformation
x → ˆ 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 deﬁnition of aﬃne equivalence yields:
ˆ
J
ˆ
K
ˆ v(ˆ x) =
N∈N
(A
∗
N)ˆ v(A
−1
)
∗
ψ
N
(ˆ x),
where
(A
∗
N)(ˆ v) = N(A
∗
ˆ v), (A
∗
ˆ v)x = ˆ v(Ax).
Thus,
[(A
∗
N)(ˆ v)[ = [N(A
∗
ˆ v)[ ≤ C
N
A
∗
ˆ v
C
l
(
¯
K)
≤ C
N,n,l
_
1 + max
1≤i,j≤n
[a
ij
[
_
l
ˆ v
C
l
(
¯
ˆ
K)
.
Also,
(A
−1
)
∗
ψ
N

W
m
p
(
ˆ
K)
≤ C
′
N,n,m
_
1 + max
1≤i,j≤n
[(a
−1
)
ij
[
_
m
[det a[
1/p
ψ
N

W
m
p
(K)
.
Since ψ
N

W
m
p
(K)
is a ﬁxed constant on the reference element K, we have that

ˆ
J
ˆ
K
ˆ v
W
m
p
(
ˆ
K)
≤ C
ref
_
1 + max
1≤i,j≤n
[a
ij
[
_
l
_
1 + max
1≤i,j≤n
[(a
−1
)
ij
[
_
m
[det a[
1/p
ˆ v
C
l
(
¯
ˆ
K)
,
where
C
ref
= [A[ max
N∈N
¦C
N,n,l
¦ max
N∈N
¦C
′
N,n,m
¦ max
N∈N
¦ψ
N

W
m
p
(K)
¦,
and [A[ denotes the number of nodal variables (i.e. the dimension of T). Thus, we
have shown that
σ(
ˆ
K) ≤ C
ref
_
1 + max
1≤i,j≤n
[a
ij
[
_
l
_
1 + max
1≤i,j≤n
[(a
−1
)
ij
[
_
m
[det a[
1/p
.
6
To avoid confusion between the ﬁnite element (K, T, A), associated with an element domain
K in the triangulation, and the aﬃne image of (
ˆ
K,
ˆ
T,
ˆ
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
aﬃne 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 ﬁnite element
on an element domain K in the triangulation.
3.3. OPTIMAL ERROR BOUNDS IN THE H
1
(Ω) NORM – REVISITED 83
Assuming that the subdivision T = ¦K¦ is regular in the sense that
∃µ > 0 ∀K ∈ T µh
K
≤ ρ
K
(≤ h
K
),
where h
K
is the diameter of K and ρ
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 σ(
ˆ
K) ≤ C
µ
, where C
µ
is a ﬁxed constant dependent on µ, but independent of
ˆ
K ∈ T . Consequently,
[v −J
K
v[
W
j
p
(K)
≤ C(m, p, n, µ)h
m−j
K
[v[
W
m
p
(K)
(3.4)
for each K ∈ T provided that T is a regular subdivision, 1 < p < ∞, m−l −(n/p) >
0 and 0 ≤ j ≤ m; from this, and recalling the deﬁnition of the global interpolant of
v ∈ W
m
p
(Ω) it follows that
_
K∈T
h
(j−m)p
K
[v −J
h
v[
p
W
j
p
(K)
_
1/p
≤ C(m, p, n, µ)[v[
W
m
p
(Ω)
. (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 < ∞, m−l −(n/p) > 0 and 0 ≤ j ≤ m, and v ∈ W
m
p
(Ω), we have that
[v −J
h
v[
W
j
p
(Ω)
≤ C(m, p, n, µ)h
m−j
[v[
W
m
p
(Ω)
, (3.6)
where h = max
K∈T
h
K
. These interpolation error estimates are of crucial importance
in ﬁnite element error analysis.
3.3 Optimal error bounds in the H
1
(Ω) norm –
revisited
In this section we return to the discussion of error estimation in the H
1
(Ω) norm.
In Chapter 2 we showed, for the ﬁnite element approximation u
h
∈ V
h
to the weak
solution u of the homogeneous Dirichlet boundary value problem for a secondorder
elliptic equation, that
u −u
h

H
1
(Ω)
≤
c
1
c
0
inf
v
h
∈V
h
u −v
h

H
1
(Ω)
. (3.7)
(c.f. C´ea’s lemma 5). Thus, restricting ourselves to the case of Poisson’s equation
and a continuous piecewise linear approximation u
h
deﬁned on a uniform triangula
tion of Ω = (0, 1)
2
, we proved that, whenever u ∈ H
2
(Ω) ∩ H
1
0
(Ω), we have
u −u
h

H
1
(Ω)
≤ Ch[u[
H
2
(Ω)
.
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 Ω ⊂ 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 Ω we shall suppose that it is regular in the sense
introduced in the previous section and we put
V
h
= J
h
(H
m
(Ω) ∩ H
1
0
(Ω)).
Then,
inf
v
h
∈V
h
u −v
h

H
1
(Ω)
≤ u −J
h
u
H
1
(Ω)
≤ C(m, n, µ)h
m−1
[u[
H
m
(Ω)
.
Substituting this into (3.7), we arrive at the following error bound:
u −u
h

H
1
(Ω)
≤ C(m, n, µ, c
1
, c
0
)h
m−1
[u[
H
m
(Ω)
, (3.8)
provided that u ∈ H
m
(Ω) ∩ H
1
0
(Ω). In particular, this will be the case if we use
continuous piecewise polynomials of degree m− 1 on a regular triangulation of Ω,
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
(Ω)
norm is of size O(h
m−1
).
3.4 Variational crimes
To conclude this chapter we brieﬂy comment on a further issue which arises in the
implementation of ﬁnite element methods. Let us consider the weak formulation of
the secondorder elliptic partial diﬀerential equation (1.5) on a bounded open set
Ω ⊂ R
n
, in the case of a homogeneous Dirichlet boundary condition (1.6):
ﬁnd u ∈ H
1
(Ω) such that a(u, v) = l(v) for all v ∈ H
1
0
(Ω),
where, as before,
a(w, v) =
n
i,j=1
_
Ω
a
ij
(x)
∂w
∂x
i
∂v
∂x
j
dx
+
n
i=1
_
Ω
b
i
(x)
∂w
∂x
i
v dx +
_
Ω
c(x)wv dx (3.9)
and
l(v) =
_
Ω
f(x)v(x) dx. (3.10)
3.4. VARIATIONAL CRIMES 85
The associated ﬁnite element method is based on choosing a ﬁnite element sub
space V
h
⊂ H
1
0
(Ω) consisting of continuous piecewise polynomials of a certain de
gree deﬁned on a subdivision of the computational domain Ω, and considering the
approximate problem
ﬁnd u
h
∈ V
h
such that a(u
h
, v
h
) = l(v
h
) for all v
h
∈ V
h
.
Unfortunately, unless the coeﬃcients a
ij
, b
i
and c and the righthand side f are
exceptionally simple functions, the integrals which appear in the deﬁnitions of a(, )
and l() will not be possible to evaluate exactly, and numerical integration rules (such
as the trapezium rule, Simpson’s rule, Gausstype rules and their multidimensional
counterparts) will have to be used to calculate a(, ) and l() approximately. Without
focusing on any particular quadrature rule, we attempt to analyse the eﬀects of this
quadratureinduced perturbation on the accuracy of the exactlyintegrated ﬁnite
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 deﬁnition of u
h
:
ﬁnd 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 ﬁnite element error analysis was the
presence of the Galerkin orthogonality property. With this new deﬁnition 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 ﬁnite 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
(Ω)
,
with c
0
a positive constant (as in Section 1.2). Thus,
c
0
u −u
h

2
H
1
(Ω)
≤ 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
(Ω)
u −v
h

H
1
(Ω)
+ sup
w
h
∈V
h
[l(w
h
) −l
h
(w
h
)[
w
h

H
1
(Ω)
v
h
−u
h

H
1
(Ω)
,
86 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION
with c
1
a positive constant (as in Section 1.2). To simplify writing, we deﬁne
l −l
h

−1,h
= sup
w
h
∈V
h
[l(w
h
) −l
h
(w
h
)[
w
h

H
1
(Ω)
.
Hence,
c
0
u −u
h

2
H
1
(Ω)
≤ c
1
u −u
h

H
1
(Ω)
u −v
h

H
1
(Ω)
+l −l
h

−1,h
_
u −u
h

H
1
(Ω)
+u −v
h

H
1
(Ω)
_
=
_
c
1
u −v
h

H
1
(Ω)
+l −l
h

−1,h
_
u −u
h

H
1
(Ω)
+l −l
h

−1,h
u −v
h

H
1
(Ω)
. (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
(Ω)
+l −l
h

−1,h
_
u −u
h

H
1
(Ω)
≤
1
2c
0
_
c
1
u −v
h

H
1
(Ω)
+l −l
h

−1,h
_
2
+
c
0
2
u −u
h

2
H
1
(Ω)
.
Substituting this into (3.11) gives
c
2
0
u −u
h

2
H
1
(Ω)
≤
_
c
1
u −v
h

H
1
(Ω)
+l −l
h

−1,h
_
2
+2c
0
l −l
h

−1,h
u −v
h

H
1
(Ω)
.
Noting that c
0
≤ c
1
, this yields
c
2
0
u −u
h

2
H
1
(Ω)
≤ 2
_
c
1
u −v
h

H
1
(Ω)
+l −l
h

−1,h
_
2
,
and therefore,
u −u
h

H
1
(Ω)
≤
c
1
√
2
c
0
u −v
h

H
1
(Ω)
+
√
2
c
0
l −l
h

−1,h
.
Equivalently,
u −u
h

H
1
(Ω)
≤
c
1
√
2
c
0
u −v
h

H
1
(Ω)
+
√
2
c
0
sup
w
h
∈V
h
[l(w
h
) −l
h
(w
h
)[
w
h

H
1
(Ω)
. (3.12)
Since v
h
∈ V
h
is arbitrary, it follows that we have proved the following perturbed
version of C´ea’s lemma:
u −u
h

H
1
(Ω)
≤
c
1
√
2
c
0
min
v
h
∈V
h
u −v
h

H
1
(Ω)
+
√
2
c
0
sup
w
h
∈V
h
[l(w
h
) −l
h
(w
h
)[
w
h

H
1
(Ω)
.
3.4. VARIATIONAL CRIMES 87
The second term on the righthand side of (3.12) quantiﬁes the extent to which the
accuracy of the exactly integrated ﬁnite element method is aﬀected 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
(Ω)
≤ C(m, n, µ, c
1
, c
0
)h
m−1
[u[
H
m
(Ω)
+
√
2
c
0
sup
w
h
∈V
h
[l(w
h
) −l
h
(w
h
)[
w
h

H
1
(Ω)
.
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
m−1
); 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 Ω by a “conveniently chosen closeby domain” Ω
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 onedimensional model problem
In order to illuminate the key ideas and avoid technical diﬃculties, we shall consider
the twopoint 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:
ﬁnd 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 ﬁnite element approximation of this problem is constructed by considering a
(possibly nonuniform) 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
N−1
< x
N
= 1 and deﬁning the ﬁnite 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 ﬁnite element approximation of the boundary
value problem is:
ﬁnd 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
i−1
, 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 deﬁnition of the dual problem and
a straightforward integration by parts yield (recall that (u−u
h
)(0) = 0, (u−u
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
N−1
< 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 righthand side no longer involves the unknown
analytical solution u. Now,
a(u
h
, z −J
h
z) =
N
i=1
_
x
i
x
i−1
u
′
h
(x) (z −J
h
z)
′
(x) dx
+
N
i=1
_
x
i
x
i−1
b(x) u
′
h
(x) (z −J
h
z)(x) dx
+
N
i=1
_
x
i
x
i−1
c(x) u
h
(x) (z −J
h
z)(x) dx.
4.1. THE ONEDIMENSIONAL MODEL PROBLEM 91
Integrating by parts in each of the (N−1) integrals in the ﬁrst sum on the righthand
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
i−1
[−u
′′
h
(x) + b(x)u
′
h
(x) + c(x)u
h
(x)] (z −J
h
z)(x) dx.
Further
(f, z −z
h
) =
N
i=1
_
x
i
x
i−1
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
i−1
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
i−1
, x
i
).
The function R(u
h
) is called the ﬁnite element residual; it measures the extent
to which u
h
fails to satisfy the diﬀerential equation −u
′′
+b(x)u
′
+c(x)u = f(x) on
the interval (0, 1). Now, applying the CauchySchwarz inequality on the righthand
side of (4.3) yields
u −u
h

2
L
2
(0,1)
≤
N
i=1
R(u
h
)
L
2
(x
i−1
,x
i
)
z −J
h
z
L
2
(x
i−1
,x
i
)
.
Recalling from the proof of Theorem 3 (with ζ = z − J
h
z and noting that ζ
′′
(x) =
z
′′
(x) for all x in (x
i−1
, x
i
), since J
h
z is a linear function on (x
i−1
, x
i
), i = 1, . . . , N)
that
z −J
h
z
L
2
(x
i−1
,x
i
)
≤
_
h
i
π
_
2
z
′′

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

2
L
2
(0,1)
≤
1
π
2
N
i=1
h
2
i
R(u
h
)
L
2
(x
i−1
,x
i
)
z
′′

L
2
(x
i−1
,x
i
)
and consequently,
u −u
h

2
L
2
(0,1)
≤
1
π
2
_
N
i=1
h
4
i
R(u
h
)
2
L
2
(x
i−1
,x
i
)
_
1/2
z
′′

L
2
(0,1)
. (4.4)
The rest of the analysis is aimed at eliminating z
′′
from the righthand 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 Poincar´eFriedrichs 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 righthand 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)
≤ Ku −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
coeﬃcients in the diﬀerential equation under consideration, and therefore it can be
computed without diﬃculty. Inserting (4.9) into (4.4), we arrive at our ﬁnal 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
i−1
,x
i
)
_
1/2
, (4.10)
where K
0
= K/π
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 reﬁnement 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 ﬁnite
element approximation u
h
to the unknown solution u (with the same deﬁnition 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 reﬁning 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
i−1
,x
i
)
_
1/2
≤ TOL
is satisﬁed. 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)
i−1
for i = 1, . . . , N
0
, and h
(0)
=
max
i
h
(0)
i
, and consider the associated ﬁnite 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, deﬁned 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)
i−1
,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)
i−1
for i = 1, . . . , N
m+1
and
h
(m+1)
= max
i
h
(m+1)
i
, and an associated ﬁnite 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)
i−1
,x
(m+1)
i
)
_
1/2
= TOL, (4.12)
and continue.
Here (4.11) is the stopping criterion and (4.12) is the mesh modiﬁcation strategy.
According to the a posteriori error bound (4.10), when the algorithm terminates the
global error u−u
h

L
2
(0,1)
is controlled to within the prescribed tolerance TOL. The
relation (4.12) deﬁnes the new meshsize 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)
i−1
,x
(m+1)
i
)
=
TOL
2
K
2
0
N
m+1
for each i = 1, . . . , N
m+1
; the implementation can be simpliﬁed by replacing N
m+1
on the righthand 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)
i−1
,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 ﬁxed, by some rootﬁnding algorithm (e.g.
successive bisection or ﬁxedpoint 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. Eﬃciency means that the computational eﬀort required to
achieve reliability is minimal. It is unclear from the present discussion whether the
adaptive algorithm described above is eﬃcient in this sense: although we have min
imised the computational eﬀort required to ensure that the righthand 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 eﬀort required to ensure
that the lefthand 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 ﬁnite element approximation of elliptic
boundary value problems. This chapter is devoted to ﬁnite element methods for
timedependent problems; in particular, we shall be concerned with the ﬁnite ele
ment approximation of parabolic equations. Hyperbolic equations will not be dis
cussed in these notes.
5.1 The parabolic model problem
Let Ω be a bounded open set in R
n
, n ≥ 1, with boundary Γ = ∂Ω, and let T > 0.
In Q = Ω(0, T], we consider the initial boundary value problem for the unknown
function u(x, t), x ∈ Ω, 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 ∈ Ω, t ∈ (0, T], (5.1)
u(x, t) = 0, x ∈ Γ, t ∈ [0, T], (5.2)
u(x, 0) = u
0
(x), x ∈
¯
Ω. (5.3)
Suppose that u
0
∈ L
2
(Ω), and that there exists a positive constant ˜ c such that
n
i,j=1
a
ij
(x, t)ξ
i
ξ
j
≥ ˜ c
n
i=1
ξ
2
i
,
∀ξ = (ξ
1
, . . . , ξ
n
) ∈ R
n
, ∀x ∈
¯
Ω, 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 diﬀerential 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 advectiondiﬀusion 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 signiﬁes 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
(Ω), the inner product (u, v) and the norm v
L
2
(Ω)
are deﬁned by
(v, w) =
_
Ω
v(x)w(x) dx,
v
L
2
(Ω)
= (v, v)
1/2
.
Taking the inner product of (5.1) with u, noting that u(x, t) = 0, x ∈ Γ, 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
(Ω)
≤ (f(, t), u(, t)).
Noting that
_
∂u
∂t
(, t), u(, t)
_
=
1
2
d
dt
u(, t)
2
L
2
(Ω)
,
and using the Poincar´eFriedrichs inequality (1.2), we obtain
1
2
d
dt
u(, t)
2
L
2
(Ω)
+
˜ c
c
⋆
u(, t)
2
L
2
(Ω)
≤ (f(, t), u(, t)).
5.1. THE PARABOLIC MODEL PROBLEM 97
Let K = ˜ c/c
⋆
; then, by the CauchySchwarz inequality,
1
2
d
dt
u(, t)
2
L
2
(Ω)
+ Ku(, t)
2
L
2
(Ω)
≤ f(, t)
L
2
(Ω)
u(, t)
L
2
(Ω)
≤
1
2K
f(, t)
2
L
2
(Ω)
+
K
2
u(, t)
2
L
2
(Ω)
.
Thence,
d
dt
u(, t)
2
L
2
(Ω)
+ Ku(, t)
2
L
2
(Ω)
≤
1
K
f(, t)
2
L
2
(Ω)
.
Multiplying both sides by e
Kt
,
d
dt
_
e
Kt
u(, t)
2
L
2
(Ω)
_
≤
e
Kt
K
f(, t)
2
L
2
(Ω)
.
Integrating from 0 to t,
e
Kt
u(, t)
2
L
2
(Ω)
−u
0

2
L
2
(Ω)
≤
1
K
_
t
0
e
Kτ
f(, τ)
2
L
2
(Ω)
dτ.
Hence
u(, t)
2
L
2
(Ω)
≤ e
−Kt
u
0

2
L
2
(Ω)
+
1
K
_
t
0
e
−K(t−τ)
f(, τ)
2
L
2
(Ω)
dτ. (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
satisﬁes (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
(Ω)
≤ e
−Kt
u
0

2
L
2
(Ω)
, t ≥ 0; (5.7)
in physical terms, the energy
1
2
u(, t)
2
L
2
(Ω)
dissipates exponentially. Since K =
˜ c/c
⋆
, we have
u(, t)
2
L
2
(Ω)
≤ e
−˜ ct/c
⋆
u
0

2
L
2
(Ω)
, t ≥ 0, (5.8)
and we deduce that the rate of dissipation depends on the lower bound, ˜ c, on the
“diﬀusion coeﬃcients” 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
(Ω)
= u
0

2
L
2
(Ω)
;
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
secondorder parabolic equation.
In the next section we consider some simple ﬁnite 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 = Ω (0, T], where Ω = (0, 1), T > 0;
ﬁnd 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 ﬁrst scheme (called the forward
Euler scheme) employs a forward divided diﬀerence in t to approximate ∂u/∂t, the
second (called the backward Euler scheme) uses a backward diﬀerence in t.
The forward Euler scheme. We begin by constructing a mesh on
¯
Q = [0, 1]
[0, T]. Let h = 1/N be the meshsize in the xdirection and let ∆t = T/M be the
meshsize in the tdirection; here N and M are two integers, N ≥ 2, M ≥ 1. We
deﬁne 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 deﬁned
on the xmesh
0 = x
0
< x
1
< . . . < x
N−1
< x
N
= 1
which vanish at the endpoints, x = 0 and x = 1.
We approximate (5.9) by the ﬁnite element method, referred to as the forward
Euler scheme:
ﬁnd 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 deﬁned 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 ﬁnd u
m+1
h
at time level t
m+1
we have to solve a system of linear
equations with symmetric positive deﬁnite matrix M of size (N −1) (N −1), with
entries (φ
i
, φ
j
) where φ
i
denotes the onedimensional piecewise linear ﬁnite element
basis function associated with the xmesh 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 diﬀerence, which gives rise to the following backward
Euler scheme:
ﬁnd 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 ﬁnd u
m+1
h
at time level t
m+1
we have to solve a system of linear
equations with symmetric positive deﬁnite matrix A of size (N −1) (N −1), with
entries (φ
i
, φ
j
) + ∆t (φ
′
i
, φ
′
j
) where φ
i
denotes the onedimensional piecewise linear
ﬁnite element basis function associated with the xmesh point x
i
; ﬁnding 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 socalled stiﬀness matrix.
100 CHAPTER 5. EVOLUTION PROBLEMS
5.3 Stability of θschemes
We shall study the stability of the schemes (5.10) and (5.11) simultaneously, by
embedding them into a oneparameter family of ﬁnite element schemes:
ﬁnd u
m
h
∈ V
h
, 0 ≤ m ≤ M, such that
_
u
m+1
h
−u
m
h
∆t
, v
h
_
+ a(u
m+θ
h
, v
h
) = (f
m+θ
, v
h
) ∀v
h
∈ V
h
(5.12)
(u
0
h
−u
0
, v
h
) = 0 ∀v
h
∈ V
h
,
where 0 ≤ θ ≤ 1, and for the sake of notational simplicity, we wrote
f
m+θ
(x) = θf(x, t
m+1
) + (1 −θ)f(x, t
m
)
and
u
m+θ
h
(x) = θu
m+1
h
(x) + (1 −θ)u
m
h
(x).
For θ = 0 this gives the forward Euler scheme, for θ = 1 the backward Euler scheme.
The method corresponding to θ =
1
2
is known as the CrankNicolson scheme.
Recall that
(w, v) =
_
1
0
w(x)v(x) dx,
v
L
2
(Ω)
= (v, v)
1/2
.
Taking the inner product of (5.12) with u
m+θ
h
we get
_
u
m+1
h
−u
m
h
∆t
, u
m+θ
h
_
+ a(u
m+θ
h
, u
m+θ
h
) = (f
m+θ
, u
m+θ
h
).
Equivalently,
_
u
m+1
h
−u
m
h
∆t
, u
m+θ
h
_
+[u
m+θ
h
[
2
H
1
(Ω)
= (f
m+θ
, u
m+θ
h
).
Since
u
m+θ
h
= ∆t
_
θ −
1
2
_
u
m+1
h
−u
m
h
∆t
+
u
m+1
h
+ u
m
h
2
,
it follows that
∆t
_
θ −
1
2
_

u
m+1
h
−u
m
h
∆t

2
L
2
(Ω)
+
u
m+1
h

2
L
2
(Ω)
−u
m
h

2
L
2
(Ω)
2∆t
+[u
m+θ
h
[
2
H
1
(Ω)
= (f
m+θ
, u
m+θ
h
). (5.13)
5.3. STABILITY OF θSCHEMES 101
Suppose that θ ∈ [1/2, 1]; then θ −1/2 ≥ 0, and therefore
u
m+1
h

2
L
2
(Ω)
−u
m
h

2
L
2
(Ω)
2∆t
+[u
m+θ
h
[
2
H
1
(Ω)
≤ (f
m+θ
, u
m+θ
h
)
≤ f
m+θ

L
2
(Ω)
u
m+θ
h

L
2
(Ω)
.
According to the Poincar´eFriedrichs inequality,
u
m+θ
h

2
L
2
(Ω)
≤
1
2
[u
m+θ
h
[
2
H
1
(Ω)
.
Thus
u
m+1
h

2
L
2
(Ω)
−u
m
h

2
L
2
(Ω)
2∆t
+ 2u
m+θ
h

2
L
2
(Ω)
≤
1
2
f
m+θ

2
L
2
(Ω)
+
1
2
u
m+θ
h

2
L
2
(Ω)
,
so that
u
m+1
h

2
L
2
(Ω)
≤ u
m
h

2
L
2
(Ω)
+ ∆tf
m+θ

2
L
2
(Ω)
.
Summing through m, m = 0, . . . , k, we get that
u
k
h

2
L
2
(Ω)
≤ u
0
h

2
L
2
(Ω)
+
k−1
m=0
∆tf
m+θ

2
L
2
(Ω)
, (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
1≤k≤M
u
k
h

2
L
2
(Ω)
≤ u
0
h

2
L
2
(Ω)
+
M−1
m=0
∆tf
m+θ

2
L
2
(Ω)
,
i.e.
max
1≤k≤M
u
k
h

L
2
(Ω)
≤
_
u
0
h

2
L
2
(Ω)
+
M−1
m=0
∆tf
m+θ

2
L
2
(Ω)
_
1/2
, (5.15)
which expresses the continuous dependence of the solution to the ﬁnite element
scheme (5.12) on the initial data and the righthand side. This property is called
stability.
Thus we have proved that for θ ∈ [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 θ ∈ [1/2, 1].
Now let us consider the case θ ∈ [0, 1/2). According to (5.13),
u
m+1
h

2
L
2
(Ω)
−u
m
h

2
L
2
(Ω)
2∆t
+[u
m+θ
h
[
2
H
1
(Ω)
= ∆t(
1
2
−θ)
u
m+1
h
−u
m
h
∆t

2
L
2
(Ω)
+ (f
m+θ
, u
m+θ
). (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
(Ω)
=
_
f
m+θ
,
u
m+1
h
−u
m
h
∆t
_
−a
_
u
m+θ
h
,
u
m+1
h
−u
m
h
∆t
_
.
Therefore,

u
m+1
h
−u
m
h
∆t

2
L
2
(Ω)
≤ f
m+θ

L
2
(Ω)

u
m+1
h
−u
m
h
∆t

L
2
(Ω)
+[u
m+θ
h
[
H
1
(Ω)
[
u
m+1
h
−u
m
h
∆t
[
H
1
(Ω)
. (5.17)
Next we shall prove that, for each w
h
∈ V
h
,
[w
h
[
H
1
(Ω)
≤
√
12
h
w
h

L
2
(Ω)
. (5.18)
We shall then use this inequality to estimate the terms appearing on the righthand
side of (5.17). Let W
i
denote the value of the piecewise linear function w
h
∈ V
h
at
the meshpoint x
i
, i = 0, . . . , N, and note that W
0
= W
N
= 0. A simple calculation
reveals that
[w
h
[
2
H
1
(Ω)
=
N
i=1
h[
W
i
−W
i−1
h
[
2
≤
4
h
2
N−1
i=1
h[W
i
[
2
. (5.19)
On the other hand,
w
h

2
L
2
(Ω)
=
h
6
N−1
i=1
(W
i
W
i−1
+ 4W
2
i
+ W
i
W
i+1
)
≥
h
6
N−1
i=1
_
−
1
2
W
2
i
−
1
2
W
2
i−1
+ 4W
2
i
−
1
2
W
2
i
−
1
2
W
2
i+1
_
≥
1
3
N−1
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
(Ω)
≤ f
m+θ

L
2
(Ω)

u
m+1
h
−u
m
h
∆t

L
2
(Ω)
+
√
12
h
[u
m+θ
h
[
H
1
(Ω)

u
m+1
h
−u
m
h
∆t

L
2
(Ω)
and hence

u
m+1
h
−u
m
h
∆t

L
2
(Ω)
≤ f
m+θ

L
2
(Ω)
+
√
12
h
[u
m+θ
h
[
H
1
(Ω)
(5.21)
5.3. STABILITY OF θSCHEMES 103
By (5.21), for any ǫ ∈ (0, 1),

u
m+1
h
−u
m
h
∆t

2
L
2
(Ω)
≤
_
√
12
h
[u
m+θ
h
[
H
1
(Ω)
+f
m+θ

L
2
(Ω)
_
2
≤ (1 + ǫ)
12
h
2
[u
m+θ
h
[
2
H
1
(Ω)
+ (1 + ǫ
−1
)f
m+θ

2
L
2
(Ω)
,
where the inequality (a + b)
2
≤ (1 + ǫ)a
2
+ (1 +
1
ǫ
)b
2
, a, b ≥ 0, ǫ > 0, has been
applied. Substituting into (5.16),
u
m+1
h

2
L
2
(Ω)
−u
m
h

2
L
2
(Ω)
2∆t
+
_
1 −∆t(
1
2
−θ)
12(1 + ǫ)
h
2
_
[u
m+θ
h
[
2
H
1
(Ω)
≤ f
m+θ

L
2
(Ω)
u
m+θ
h

L
2
(Ω)
+ ∆t(
1
2
−θ)(1 + ǫ
−1
)f
m+θ

2
L
2
(Ω)
.
(5.22)
According to the Poincar´eFriedrichs inequality,
u
m+θ
h

2
L
2
(Ω)
≤
1
2
[u
m+θ
h
[
2
H
1
(Ω)
,
and therefore,
f
m+θ

L
2
(Ω)
u
m+θ
h

L
2
(Ω)
≤
1
8ǫ
2
f
m+θ

2
L
2
(Ω)
+ 2ǫ
2
u
m+θ
h

2
L
2
(Ω)
≤
1
8ǫ
2
f
m+θ

2
L
2
(Ω)
+ ǫ
2
[u
m+θ
h
[
2
H
1
(Ω)
. (5.23)
Substituting (5.23) into (5.22),
u
m+1
h

2
L
2
(Ω)
−u
m
h

2
L
2
(Ω)
2∆t
+
_
1 −∆t
6(1 −2θ)(1 + ǫ)
h
2
−ǫ
2
_
[u
m+θ
h
[
2
H
1
(Ω)
≤
1
8ǫ
2
f
m+θ

2
L
2
(Ω)
+ ∆t(
1
2
−θ)(1 + ǫ
−1
)f
m+θ

2
L
2
(Ω)
.
Let us suppose that
∆t ≤
h
2
6(1 −2θ)
(1 −ǫ), θ ∈ [0, 1/2),
where ǫ is a ﬁxed real number, ǫ ∈ (0, 1). Then
1 −∆t
6(1 −2θ)(1 + ǫ)
h
2
−ǫ
2
≥ 0,
so that
u
m+1
h

2
L
2
(Ω)
≤ u
m
h

2
L
2
(Ω)
+
∆t
4ǫ
2
f
m+θ

2
L
2
(Ω)
+ ∆t
2
(1 −2θ)(1 + ǫ
−1
)f
m+θ

2
L
2
(Ω)
.
104 CHAPTER 5. EVOLUTION PROBLEMS
Letting c
ǫ
= 1/(4ǫ
2
) + ∆t(1 − 2θ)(1 + ǫ
−1
), upon summation through all m this
implies that
max
1≤k≤M
u
k
h

2
L
2
(Ω)
≤ u
0
h

2
L
2
(Ω)
+ c
ǫ
M−1
m=0
∆tf
m+θ

2
L
2
(Ω)
.
Taking the square root of both sides, we deduce that for θ ∈ [0, 1/2) the scheme
(5.12) is conditionally stable in the sense that
max
1≤k≤M
u
k
h

h
≤
_
u
0
h

2
L
2
(Ω)
+ c
ǫ
M−1
m=0
∆tf
m+θ

2
L
2
(Ω)
_
1/2
, (5.24)
provided that
∆t ≤
h
2
6(1 −2θ)
(1 −ǫ), 0 < ǫ < 1. (5.25)
To summarise: when θ ∈ [1/2, 1], the method (5.12) is unconditionally stable. In
particular, the backward Euler scheme, corresponding to θ = 1, and the Crank
Nicolson scheme, corresponding to θ = 1/2, are unconditionally stable, and (5.15)
holds. When θ ∈ [0, 1/2), the scheme (5.12) is conditionally stable, subject to the
time step limitation (5.25). The forward Euler scheme, corresponding to θ = 0, is
only conditionally stable.
5.4 Error analysis in the L
2
norm
In this section we investigate the accuracy of the ﬁnite 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 (θ = 1); for other values of
θ ∈ [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
= η
m
+ ξ
m
,
where
η
m
= u(, t
m
) −Pu(, t
m
), ξ
m
= Pu(, t
m
) −u
m
h
,
and for t ∈ [0, T], Pu(, t) ∈ V
h
denotes the Dirichlet projection of u(, t) deﬁned
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 LaxMilgram theorem.
Hence,
a(η
m
, v
h
) = 0 ∀v
h
∈ V
h
,
5.4. ERROR ANALYSIS IN THE L
2
NORM 105
and therefore, by C´ea’s lemma,
[η
m
[
H
1
(Ω)
≤ [u(, t
m
) −J
h
u(, t
m
)[
H
1
(Ω)
≤
h
π
[u(, t
m
)[
H
2
(Ω)
,
where J
h
u(, t
m
) ∈ V
h
denotes the continuous piecewise linear interpolant of u(, t
m
)
from V
h
. By the AubinNitsche duality argument,
η
m

L
2
(Ω)
≤
h
2
π
2
[u(, t
m
)[
H
2
(Ω)
. (5.26)
Since also,
a
_
η
m+1
−η
m
∆t
, v
h
_
= 0 ∀v
h
∈ V
h
,
by an identical argument we deduce that

η
m+1
−η
m
∆t

L
2
(Ω)
≤
h
2
π
2
¸
¸
¸
¸
u(, t
m+1
) −u(, t
m
)
∆t
¸
¸
¸
¸
H
2
(Ω)
. (5.27)
For m = 0,
(ξ
0
, v
h
) = (e
0
h
, v
h
) −(η
0
, v
h
) = −(η
0
, v
h
)
and therefore, choosing v
h
= ξ
0
and applying the CauchySchwarz inequality on the
right,
ξ
0

L
2
(Ω)
≤ η
0

L
2
(Ω)
≤
h
2
π
2
[u
0
[
H
2
(Ω)
. (5.28)
It is easily seen that ξ
m
∈ V
h
satisﬁes the following identity:
_
ξ
m+1
−ξ
m
∆t
, v
h
_
+ a(ξ
m+1
, v
h
)
=
_
u(, t
m+1
) −u(, t
m
)
∆t
−
∂u
∂t
(, t
m+1
) −
η
m+1
−η
m
∆t
, v
h
_
According to the stability result proved earlier on,
max
1≤m≤M
ξ
m

L
2
(Ω)
≤
_
ξ
0

2
L
2
(Ω)
+
M−1
m=0
∆tϕ
m+1

2
L
2
(Ω)
_
1/2
, (5.29)
where
ϕ
m+1
=
u(, t
m+1
) −u(, t
m
)
∆t
−
∂u
∂t
(, t
m+1
) −
η
m+1
−η
m
∆t
.
By (5.28),
ξ
0

L
2
(Ω)
≤
h
2
π
2
[u
0
[
H
2
(Ω)
. (5.30)
106 CHAPTER 5. EVOLUTION PROBLEMS
It remains to estimate ϕ
m+1

L
2
(Ω)
. Now
ϕ
m+1

L
2
(Ω)
≤ 
u(, t
m+1
) −u(, t
m
)
∆t
−
∂u
∂t
(, t
m+1
)
L
2
(Ω)
+
η
m+1
−η
m
∆t

L
2
(Ω)
≡ I + II. (5.31)
For term I, Taylor’s 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
(Ω)
dt
_
1/2
.
Further, by (5.27),
II ≤
h
2
π
2
¸
¸
¸
¸
u(, t
m+1
) −u(, t
m
)
∆t
¸
¸
¸
¸
H
2
(Ω)
=
h
2
π
2
¸
¸
¸
¸
¸
1
∆t
_
t
m+1
t
m
∂u
∂t
(, t) dt
¸
¸
¸
¸
¸
H
2
(Ω)
≤
h
2
π
2
√
∆t
_
_
t
m+1
t
m
¸
¸
¸
¸
∂u
∂t
(, t)
¸
¸
¸
¸
2
H
2
(Ω)
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
1≤m≤M
ξ
m

L
2
(Ω)
≤ C
1
(h
2
+ ∆t), θ ∈ (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
1≤m≤M
u(, t
m
) −u
m
h

L
2
(Ω)
≤ max
1≤m≤M
ξ
m

L
2
(Ω)
+ max
1≤m≤M
η
m

L
2
(Ω)
.
Thus, by (5.32) and (5.26), we deduce that
max
1≤m≤M
u(, t
m
) −u
m
h

L
2
(Ω)
≤ C
2
(h
2
+ ∆t),
where C
2
is a positive constant independent of h and ∆t.
The CrankNicolson scheme (θ = 1/2) can be shown to converge in the norm
 
L
2
(Ω)
unconditionally, with error O(h
2
+(∆t)
2
). For θ ∈ (1/2, 1] the scheme con
verges unconditionally with error O(h
2
+∆t). For θ ∈ [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 onedimensional
case, is beyond the scope of these notes.
2
Contents
1 Introduction 1.1 Elements of function spaces . . . . . . 1.1.1 Spaces of continuous functions 1.1.2 Spaces of integrable functions . 1.1.3 Sobolev spaces . . . . . . . . . 1.2 Weak solutions to elliptic problems . . 5 6 6 8 10 14 23 24 31 35 40 45 56 65 65 65 67 70 73 74 75 80 83 84
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . .
2 Approximation of elliptic problems 2.1 Piecewise linear basis functions . . . . . . . . . 2.2 The selfadjoint elliptic problem . . . . . . . . . 2.3 Calculation and assembly of stiﬀness matrix . . 2.4 Galerkin orthogonality; C´a’s lemma . . . . . . e 2.5 Optimal error bound in the energy norm . . . . 2.6 Superapproximation in meshdependent norms
3 Piecewise polynomial approximation 3.1 Construction of ﬁnite element spaces . . . . . . . . . 3.1.1 The ﬁnite element . . . . . . . . . . . . . . . 3.1.2 Examples of triangular ﬁnite elements . . . . 3.1.3 The interpolant . . . . . . . . . . . . . . . . . 3.1.4 Examples of rectangular elements . . . . . . . 3.2 Polynomial approximation in Sobolev spaces . . . . . 3.2.1 The BrambleHilbert lemma . . . . . . . . . . 3.2.2 Error bounds on the interpolation error . . . 3.3 Optimal error bounds in the H 1 (Ω) norm – revisited 3.4 Variational crimes . . . . . . . . . . . . . . . . . . .
4 A posteriori error analysis by duality 89 4.1 The onedimensional model problem . . . . . . . . . . . . . . . . . . . . . . 89 4.2 An adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 5 Evolution problems 5.1 The parabolic model problem . . . . . 5.2 Forward and backward Euler schemes 5.3 Stability of θschemes . . . . . . . . . 5.4 Error analysis in the L2 norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 95 98 100 104
3
4. with special emphasis on theoretical questions such as accuracy. A posteriori error analysis by duality: reliability. the aim of this course is to provide an introduction to their mathematical theory. Chapter 8: Secs. [Chapters 1–4. The AubinNitsche duality argument. S.4. Elman. Analysis of ﬁnite element methods for evolution problems.3. The BrambleHilbert lemma.R. C. reliability and adaptivity. 2nd edition. 2. Syllabus: Elements of function spaces. 2002.7]. 9. Variational crimes. Silvester & A. 8.2. 4. practical issues concerning the development of eﬃcient ﬁnite element algorithms will also be discussed. Chapter 5: Secs. [Chapters 0.5]. H. Energy dissipation. Johnson. Reading List 1. CUP. Finite element approximation of initial boundary value problems.1–9. OUP.1–4. Piecewise polynomial approximation in Sobolev spaces. Oxford University Computing Laboratory. Elliptic boundary value problems: existence.W.4. Numerical Solution of Partial Diﬀerential Equations by the Finite Element Method.1–8. K. 1990.C. Chapter 9: Secs. . Finite element methods: Galerkin orthogonality and Cea’s lemma.2. 2005. Scott. D. The Mathematical Theory of Finite Element Methods. Superapproximation properties in meshdependent norms. 5. Morton. Brenner & L. conservation and stability. 3.1. Chapter 4: Secs. uniqueness and regularity of weak solutions. 1991. Lecture Notes on Finite Element Methods.1–5. Optimal error bounds in the energy norm. Wathen. Springer. eﬃciency and adaptivity. [Mainly Chapters 1 and 5]. Finite Elements and Fast Iterative Solvers.4 CONTENTS Synopsis: Finite element methods represent a powerful and general class of techniques for the approximate solution of partial diﬀerential equations.
accuracy. Zl´mal: On the ﬁnite element method. etc. Soc. 49. Very frequently the equations under consideration are so complicated that ﬁnding their solutions in closed form or by purely analytical means (e. These notes are devoted to a particular class of numerical techniques for the approximate solution of partial diﬀerential equations: ﬁnite element methods. reliability and adaptivity. Math. Courant: Variational methods for the solution of problems of equilibrium and vibrations. They were proposed in a seminal work of Richard Courant1 . including stability. In these notes we shall be concerned with the mathematical aspects of ﬁnite element approximation.. Since then ﬁnite element s a methods have been developed into one of the most general and powerful class of techniques for the numerical solution of partial diﬀerential equations and are widely used in engineering design and analysis. but the mathematical analysis of ﬁnite element approximations began much later.Chapter 1 Introduction Partial diﬀerential equations arise in the mathematical modelling of many physical. R. economy. Bull. pp. the ﬁrst important results being due to Miloˇ Zl´mal2 in 1968. and one has to resort to seeking numerical approximations to the unknown analytical solution. 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 diﬀerential equations. by Laplace and Fourier transform methods. unfortunately. in 1943. the relevance of this article was not recognised at the time and the idea was forgotten. Amer. ﬁnancial modelling. chemical and biological phenomena and many diverse subject areas such as ﬂuid dynamics. Numerische Mathematik. or in the form of a power series) is either impossible or impracticable. 1–23 (1943) 2 M. material science. 394–402 (1968) a 1 5 . electromagnetism.g. In the early 1950’s the method was rediscovered by engineers. pp. in the 1960’s. 12. astrophysics. The concepts and notational conventions introduced here will be used systematically throughout the notes.
we remark here that all functions that appear in these notes will be assumed to be realvalued. 1. + αn is referred to as the length of the multi–index α = (α1 . α3 ). + 2 + ∂x1 ∂x2 ∂x3 ∂x2 ∂x3 ∂x2 ∂x2 ∂x3 3 + ⋄ This example highlights the importance of multiindex notation: instead of laboriously writing out in detail the ten terms on the righthand side of the last identity. . Let Ω be an open set in Rn and let k ∈ N.1. 3.. Let D = α ∂ ∂x1 α1 . ∂ ∂xn αn = ∂ α . In this section we present a brief overview of basic deﬁnitions and simple results form the theory of function spaces.. 0) by 0. INTRODUCTION 1. The nonnegative integer α := α1 + . . ∂xα1 . x2 . Dα u = α=3 ∂ 3u ∂ 3u ∂ 3u + 2 + 2 ∂x3 ∂x1 ∂x2 ∂x1 ∂x3 1 ∂ 3u ∂ 3u ∂ 3u + + 3 ∂x1 ∂x2 ∂x1 ∂x2 ∂x2 2 3 ∂ 3u ∂ 3u ∂ 3u ∂ 3u + + 3. . . ∂xαn 1 n Example 1 Suppose that n = 3. the accuracy of ﬁnite element approximations to partial diﬀerential equations very much depends on the smoothness of the analytical solution to the equation under consideration. For future reference. . 2. An ntuple α = (α1 . j = 1. We denote by C k (Ω) the set of all continuous realvalued functions deﬁned on Ω such that Dα u is continuous on Ω for . . . For the sake of notational convenience. . αn ). αn ) ∈ Nn is called a multi–index. x3 . clearly 0 = 0. α2 . . . . and this in turn hinges on the smoothness of the data. . called function spaces.1 Spaces of continuous functions In this section. and α = (α1 . . Precise assumptions about the regularity of the solution and the data can be conveniently formulated by considering classes of functions with speciﬁc diﬀerentiability and integrability properties. . we introduce the concept of a multiindex.1 Elements of function spaces As will become apparent in subsequent chapters. αj ∈ N. Then for u. we describe some simple function spaces which consist of continuously diﬀerentiable functions. . a function of three variables x1 .6 CHAPTER 1. Let N denote the set of nonnegative integers. We denote (0. we can compress the information into a single entity shown on the left. .
The function u(x) = 1/x ¯ belongs to C k (Ω) for each k ≥ 0. Example 3 Let w be the function deﬁned on Rn by w(x) = e 0. . ⋄ . Therefore u ∈ C k (Ω) ¯ u is not continuous on Ω. Assuming that Ω is a bounded open set. here x = (x2 + . . . for any k ≥ 0. x∈Ω ¯ ¯ In particular when k = 0 we shall write C(Ω) instead of C 0 (Ω) to denote the set of ¯ all continuous functions deﬁned on Ω. ELEMENTS OF FUNCTION SPACES 7 ¯ all α = (α1 . αn ). in this case. + x2 )1/2 . . . it is clear that ¯ the same is true of its derivatives. u ¯ C 1 (Ω) ¯ C(Ω) = sup u(x) = max u(x). 0 Example 4 The function w deﬁned in the previous example belongs to the space ∞ C0 (Rn ). Let ∞ C0 (Ω) = k≥0 ∩ C k (Ω).1. . the closure of the set Ω.1. . otherwise. − 1 1−x2 . α ≤ k. 1) ⊂ R1 . . 1] and limx→0 u(x) = ∞. Clearly. We shall write supp u for the support of u. ⋄ The support of a continuous function u deﬁned on an open set Ω ⊂ Rn is deﬁned as the closure in Ω of the set {x ∈ Ω : u(x) = 0}. As Ω = [0. αn ) with α ≤ k. x < 1. k ¯ C (Ω) can be equipped with the norm u ¯ C k (Ω) := α≤k sup Dα u(x). u Similarly. . ⋄ k We denote by C0 (Ω) the set of all u contained in C k (Ω) whose support is a bounded subset of Ω. C k (Ω) k α will denote the set of all u in C (Ω) such that D u can be extended from Ω to a ¯ continuous function on Ω. if k = 1. for all α = (α1 . . x∈Ω ¯ x∈Ω = α≤1 sup Dα u(x) x∈Ω n = sup u(x) + x∈Ω j=1 sup  x∈Ω ∂u (x). the support of w is the closed unit ball 1 n {x ∈ Rn : x ≤ 1}. supp u is the smallest closed subset of Ω such that u = 0 in Ω\supp u. ∂xj Example 2 Consider the open interval Ω = (0. Thus.
Lp (Ω) consists of equivalence classes of functions. there exists a positive constant M such that u(x) ≤ M for almost every3 x in Ω. Ω Any two functions which are equal almost everywhere (i. v) ≤ u 3 L2 (Ω) v L2 (Ω) . we denote by Lp (Ω) the set of all realvalued functions deﬁned on an open subset Ω of Rn such that u(x)p dx < ∞. if P (x) is true for all x ∈ Ω\Γ where Γ is a subset of Ω with zero Lebesgue measure. except on a set of measure zero) on Ω are identiﬁed with each other. and we write M = ess. we shall not insist on this technicality. still. then u v ∈ L1 (Ω) and (u.1. . Lemma 1 (The CauchySchwarz inequality) Let u and v belong to L2 (Ω).supx∈Ω u(x)). equal. INTRODUCTION 1. Let p be a real number. L∞ (Ω) is equipped with the norm u L∞ (Ω) = ess. strictly speaking. the smallest such number M is called the essential supremum of u. u)1/2 . v) := Ω u(x)v(x) dx. A particularly important case corresponds to taking p = 2. p ≥ 1. We shall say that a property P (x) is true for almost every x in Ω. then 1/2 u L2 (Ω) = Ω u(x)2 dx .8 CHAPTER 1.2 Spaces of integrable functions Next we consider a class of spaces that consist of (Lebesgue) integrable functions. We shall also consider the space L∞ (Ω) consisting of functions u deﬁned on Ω such that u has ﬁnite essential supremum on Ω (namely.e. Clearly u L2 (Ω) = (u.supx∈Ω u(x). The space L2 (Ω) can be equipped with the inner product (u. Thus. Lp (Ω) is equipped with the norm 1/p u Lp (Ω) := Ω u(x) dx p .
L2 (Ω) is a Hilbert space: it has an inner product (·. v)2 − 4 u and hence the desired inequality. the sequence {um }∞ converges to u m=1 in X). whenever {um }m=1 is lim un − um X = 0. namely. λ ∈ R.1) is called a Cauchy sequence. u+v Lp (Ω) ≤ u Lp (Ω) + v Lp (Ω) .e. u) + (λv. m=1 . In particular. we note that a statement analogous to Corollary 1 holds. 2(u. ·) and. 2 L2 (Ω) v 2 L2 (Ω) ≤ 0. v ∈ Lp (Ω).1. ∞] is a Banach space4 . Lp (Ω) v Ω 4 Lp′ (Ω) . i.1.1) there exists u ∈ X such that limm→∞ u − um X = 0 (i. u + λv) = (u. and u+v L2 (Ω) ≤ u L2 (Ω) + v L2 (Ω) . when equipped with the associated norm · L2 (Ω) . Remark 1 The space Lp (Ω) with p ∈ [1. known as H¨lder’s inequality. Upon taking the square root of both sides we complete the proof. Furthermore. u)1/2 .e. then 0 ≤ u + λv 2 L2 (Ω) 9 = (u + λv. deﬁned by u L2 (Ω) = (u. λv) = u 2 L2 (Ω) + 2λ(u. A sequence {um }∞ with the property (1. u. u) + (u. and it is nonnegative for all λ ∈ R. ELEMENTS OF FUNCTION SPACES Proof Let λ ∈ R. is valid for any two functions u ∈ Lp (Ω) and v ∈ Lp′ (Ω) o with 1/p + 1/p′ = 1: u(x)v(x) dx ≤ u X. then u + v ∈ L2 (Ω). v) + λ2 v 2 L2 (Ω) . A normed linear space X. (1. Proof This is a straightforward consequence of the CauchySchwarz inequality: u+v 2 L2 (Ω) = (u + v.m→∞ ∞ is called a Banach space if. v) + v + v 2 L2 (Ω) 2 L2 (Ω) L2 (Ω) . u + v) = ≤ u u 2 L2 (Ω) + 2(u. ⋄ To conclude this section. therefore its discriminant is nonpositive. The righthand side is a quadratic polynomial in λ with real coeﬃcients. λv) + (λv. more generally. with norm · a sequence of elements of X such that n. in the Lp norm for 1 ≤ p ≤ ∞. the following generalisation of the CauchySchwarz inequality. Corollary 1 (The triangle inequality) Let u and v belong to L2 (Ω). it is a Banach space.
Clearly u is not diﬀerentiable at the points 0 and ±1. . If w(x)v(x) dx = 0 Ω ∞ for all v in C0 (Ω) 5 then w(x) = 0 for almost every x ∈ Ω. . and suppose that we wish to determine the weak ﬁrst derivative of the function u(x) = (1 − x)+ deﬁned on Ω. with ω ⊂ Ω). In order to see that this deﬁnition is correct it has to be shown that if a locally integrable function has a weak derivative then this must be unique. if u is a suﬃciently smooth function. This identity represents the starting point for deﬁning the concept of weak derivative. + αn . ∂xαn 1 n In order to simplify the notation.L. with Ω an open subset of ∞ Rn . ∂ α u . and we write wα = Dα u. which play an important role in modern diﬀerential equation theory.e. we remark that this is a straightforward consequence of DuBois Reymond’s lemma5 .1. then the following integrationbyparts formula holds: Ω Dα u(x) · v(x) dx = (−1)α Ω u(x) · Dα v(x) dx. u ∈ L1 (ω) for each bounded open set ω. it will always be clear from the context (by considering the smoothness of the function diﬀerentiated) which of the two is implied. locally integrable on Ω and such that wα (x) · v(x) dx = (−1)α u(x) · Dα v(x) ∞ ∀v ∈ C0 (Ω). called Sobolev spaces (after the Russian mathematician S. Sobolev). have disappeared because v and all of its derivatives are identically zero on the boundary of Ω. then its weak derivative Dα u of order α ≤ k coincides with the corresponding partial derivative in the classical pointwise sense. we introduce the concept of weak derivative. and let v ∈ C0 (Ω). which arise in the course of integrating by parts. Ω ⊂ Rn . DuBois Reymond’s lemma: Suppose that w is a locally integrable function deﬁned on an open set Ω. Suppose that u is a smooth function. Ω Ω then we say that wα is a weak derivative of the function u of order α = α1 + . ∞ ∀v ∈ C0 (Ω). However. INTRODUCTION 1. . Suppose also that there exists a function ¯ wα . α ≤ k. Note that all terms involving integrals over the boundary of Ω. say u ∈ C k (Ω). we shall use the letter D to denote classical as well as weak derivatives.3 Sobolev spaces In this section we introduce a class of spaces. . Example 5 Let Ω = R1 . because u is locally integrable on Ω. it may. ∂xα1 . Before we give the precise deﬁnition of a Sobolev space. . say u ∈ C k (Ω).10 CHAPTER 1. Clearly. Now suppose that u is a locally integrable function deﬁned on Ω (i.
We deﬁne (with Dα denoting a weak derivative of order α ) k Wp (Ω) = {u ∈ Lp (Ω) : Dα u ∈ Lp (Ω). for any v ∈ C0 (Ω). w(x)v(x) dx. −∞ where 0. have a weak derivative. ∞].1. := Dα u α≤k p Lp (Ω) when 1 ≤ p < ∞ k W∞ (Ω) := α≤k Dα u L∞ (Ω) when p = ∞. w(x) = −1. +∞ +∞ 1 11 u(x)v ′ (x) dx = −∞ 0 −∞ (1 − x)+ v ′ (x) dx = 1 0 −1 (1 − x)v ′ (x) dx = −1 (1 + x)v ′ (x) dx + 0 (1 − x)v ′ (x) dx 1 0 =− 0 −1 v(x) dx + (1 + x)v(x)0 + −1 1 v(x) dx + (1 − x)v(x)1 x=0 +∞ = −1 (−1)v(x) dx + 0 1 · v(x) dx ≡ − x < −1. for p ∈ [1. x ∈ (−1. Indeed. k Wp (Ω) is called a Sobolev space of order k. ⋄ Now we are ready to give a precise deﬁnition of a Sobolev space. we can write u uWp (Ω) := k Dα u α=k p Lp (Ω) 1/p . 1). ∞). 1. 0). it is equipped with the (Sobolev) norm 1/p u k Wp (Ω) and u Letting. ELEMENTS OF FUNCTION SPACES ∞ nevertheless. Let k be a nonnegative integer and suppose that p ∈ [1. Thus. 0. k k Wp (Ω) 1/p = j=0 up j (Ω) Wp .e. . x > 1. i. x ∈ (0. α ≤ k}. the piecewise constant function w is the ﬁrst (weak) derivative of the continuous piecewise linear function u.1. w = u′ = Du.
. u k W∞ (Ω) = j=0 j uW∞ (Ω) . j = 1.12 Similarly. INTRODUCTION Dα u α=k L∞ (Ω) . for p = 2. . letting uW∞ (Ω) := k we have that k CHAPTER 1. . n . ∂xj ∂ 2u ∈ L2 (Ω). . give: H 1 (Ω) = u = u ∈ L2 (Ω) : u n 2 L2 (Ω) ∂u ∈ L2 (Ω). . j = 1. Our deﬁnitions of Wp (Ω) and its norm and seminorm. ∂xi ∂xj n u H 2 (Ω) = u 2 L2 (Ω) + j=1 n ∂u ∂xj 2 L2 (Ω) 1/2 2 L2 (Ω) + i. . When k ≥ 1. . n . the space W2 (Ω) is then a Hilbert space with the inner product (u. Dα v). k When k ≥ 1. Throughout these notes we shall frequently refer to the Hilbertian Sobolev spaces 1 k H (Ω) and H 2 (Ω). uH 1 (Ω) = j=1 ∂u ∂xj 2 L2 (Ω) .j=1 6 ∂ 2u ∂xi ∂xj . n. k k An important special case corresponds to taking p = 2. ∂xj n H 1 (Ω) + j=1 ∂u ∂xj 1/2 1/2 2 L2 (Ω) . j = 1. . . α = k). . .  · Wp (Ω) is only a seminorm rather than a norm because if uWp (Ω) = 0 for k k k u ∈ Wp (Ω) it does not necessarily follow that u(x) = 0 for almost every x in Ω (all that is known is that Dα u(x) = 0 for almost every x ∈ Ω. . v)W2 (Ω) := k α≤k (Dα u. so  · Wp (Ω) does not satisfy the ﬁrst axiom k of norm. for p = 2 and k = 2. k For this reason. H 2 (Ω) = u ∈ L2 (Ω) : ∂u ∈ L2 (Ω). i. Similarly. k = 1. . ·Wp (Ω) is called the Sobolev seminorm6 on Wp (Ω). we shall usually write H k (Ω) instead of W2 (Ω).
we shall restrict ourselves to considering the special case of a rectangular domain Ω = (a. The proof for general Ω is analogous.e. y) = u(a. Ω ∂x Say. y)2 dξ (x − a) ∂x a c  d b a d x dy dx (x − a) dx c  ∂u (ξ. We note here that H0 (Ω) is also a Hilbert space. In fact. 1 i. y)2 dξ dy ∂x 1 (b − a)2 2 ∂u  (x. y)2 dx dy. H0 (Ω) is the set of all u ∈ H 1 (Ω) such that u 1 ∞ is the limit in H (Ω) of a sequence {um }∞ with um ∈ C0 (Ω). in other words. We conclude the section with the following useful result. It can be shown m=1 (assuming that ∂Ω is suﬃciently smooth) that 1 H0 (Ω) = {u ∈ H 1 (Ω) : u = 0 on ∂Ω}. 1 ∞ Finally. then there exists a constant c⋆ (Ω). Thence. by the CauchySchwarz inequality. with the same norm and inner product as H 1 (Ω). ∂x c < y < d. the set of all functions u in H 1 (Ω) such that u = 0 on ∂Ω. we deﬁne the special Sobolev space H0 (Ω) as the closure of C0 (Ω) in the 1 norm of · H 1 (Ω) . y) dξ2 dy dx c a ∂x x d ∂u  (ξ. y) dξ. ELEMENTS OF FUNCTION SPACES n 13 1/2 2 L2 (Ω) uH 2 (Ω) = i. y) + a ∂u (ξ. y) dξ = ∂x x a ∂u (ξ. m=1 ∞ it is suﬃcient to prove this inequality for u ∈ C0 (Ω). the boundary of the set Ω. Evidently x u(x. . to simplify matters. ∂xi (1. u(x. such that n Ω u(x) dx ≤ c⋆ 2 i=1 Ω  ∂u (x)2 dx.1. H0 (Ω) is. y)2 dx dy = ≤ ≤ = 7 b a b a b a Ω ∂u (ξ. Ω is a polygonal domain in R2 or a polyhedron in R3 . d) in R2 .2) ∞ 1 Proof As any function u ∈ H0 (Ω) is the limit in H 1 (Ω) of a sequence {um }∞ ⊂ C0 (Ω). independent of u. Lemma 2 (Poincar´Friedrichs inequality) Suppose that Ω is a bounded open set in e n 1 R (with a suﬃciently smooth boundary7 ∂Ω) and let u ∈ H0 (Ω).j=1 ∂ 2u ∂xi ∂xj .1. in fact. b) × (c. We shall use this space when considering a partial diﬀerential equation that is coupled with a homogeneous (Dirichlet) boundary condition: u = 0 1 on ∂Ω.
¯ f ∈ C(Ω). n.14 Analogously. bi . and its nonhomogeneous counterpart.4) i. ∂xi x ∈ Ω. and consider the linear secondorder partial diﬀerential equation ∂ − ∂xj i.2 Weak solutions to elliptic problems In the ﬁrst part of this lecture course we shall focus on boundary value problems for elliptic partial diﬀerential equations. ¯ x ∈ Ω. . j = 1. . ¯ bi ∈ C(Ω). . (1. 1. Poisson’s equation −∆u = f. . . . 1 u(x. and n n i. if 1 Ω = (0.j=1 aij (x)ξi ξj ≥ c ˜ i=1 ξi2 . . y)2 dx dy. c and f satisfy the following conditions: ¯ aij ∈ C 1 (Ω). ∀ξ = (ξ1 . . (1. INTRODUCTION Ω  ∂u (x. Elliptic equations are typiﬁed by the Laplace equation ∆u = 0. y)2 dx dy ≤ (d − c)2 2 Ω By adding the two inequalities. i = 1. where we used the notation ∆= i=1 n ∂2 ∂x2 i for the Laplace operator. similarly. 1) ⊂ R then c⋆ = 2 . ξn ) ∈ Rn . y)2 dx dy ≤ c⋆ 2 (d−c)2 −1 CHAPTER 1. where c⋆ = 2 (b−a)2 + . n. we obtain u(x. we note that if Ω = (0.3) where the coeﬃcients aij . ∂y Ω Ω  ∂u ∂u 2  +  2 ∂x ∂y dx dy.j=1 n ∂u aij (x) ∂xi n + i=1 bi (x) ∂u + c(x)u = f (x). More generally. . 1 For further reference. . 1)2 ⊂ R2 then c⋆ = 4 . let Ω be a bounded open set in Rn . ¯ c ∈ C(Ω). . . .
1. c. provided that aij . Poisson’s equation with zero Dirichlet boundary condition on Ω = (−1.3) is called an elliptic equation. (∗) ¯ This problem does not have a classical solution. for otherwise ∆u would be a continuous function on Ω. ¯ A function u ∈ C 2 (Ω)∩C(Ω) satisfying (1.j=1 ∂u cos αj + σ(x)u = g ∂xi on ∂Ω. which is not possible because sgn(1/2 − x) is not continuous on Ω. n (d) A generalisation of the boundary conditions (b) and (c) is aij i. We begin by considering the homogeneous Dirichlet boundary value problem ∂ − ∂xj i. with a Dirichlet boundary condition on ∂Ω1 and Neumann boundary condition on ∂Ω2 ). . x ∈ Ω.4).4) is usually ˜ referred to as uniform ellipticity and (1. for example. In problems that arise in applications equation (1.5) (1.6) has a unique classical solution. u ∈ C 2 (Ω) ∩ C(Ω).6) is called a classical solution of this problem. with g denoting a given function deﬁned on ∂Ω: (a) u = g on ∂Ω (Dirichlet boundary condition). ∂Ω is the union of two disjoint subsets ∂Ω1 and ∂Ω2 . The theory of partial diﬀerential equations tells us that (1. where σ(x) ≥ 0 on ∂Ω (Robin boundary condition). in many applications one has to consider equations where these smoothness requirements are violated. bi .3) is usually supplemented by one of the following boundary conditions.5) and (1. ∂xi u=0 on ∂Ω. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 15 here c is a positive constant independent of x and ξ. where ν denotes the unit outward normal vector to ∂Ω (Neumann boundary condition). In many physical problems more than one type of boundary condition is imposed on ∂Ω (e. (1. c and f are as in (1. The study of such mixed boundary value problems will not be pursued in these notes. x ∈ Ω.2. (1. bi .5).g.6) where aij . where αj is the angle between the unit outward normal vector ν to ∂Ω and the xj axis (Oblique derivative boundary condition).j=1 n ∂u aij ∂xi n + i=1 bi (x) ∂u + c(x)u = f (x). x ∈ ∂Ω. However. (b) (c) = g on ∂Ω. f and ∂Ω are suﬃciently smooth. ∂u ∂ν ∂u ∂ν + σu = g on ∂Ω. Take. The condition (1. 1)n in Rn : −∆u = sgn u = 0. and for such problems the classical theory is inappropriate. 1 2 − x .
and righthand side of (1.7) are still meaningful (in fact. we obtain: n i. c(x)uv dx = Ω f (x)v(x) dx (1.5).7). . where. we generalise the notion of solution by weakening the diﬀerentiability requirements on u. . Upon integration by parts in the ﬁrst integral and noting that v = 0 on ∂Ω. 8 . Also. u = 0 on ∂Ω}.j=1 ∂u ∂v dx + aij (x) ∂xi ∂xj Ω + Ω n bi (x) i=1 Ω ∂u v dx ∂xi 1 ∀v ∈ C0 (Ω). INTRODUCTION In order to overcome the limitations of the classical theory and to be able to deal with partial diﬀerential equations with “nonsmooth” data. . ¯ it is not necessary to assume that aij ∈ C 1 (Ω).16 CHAPTER 1. n. and it is easily seen that when u ∈ H0 (Ω) and 1 1 v ∈ H0 (Ω). Note further that since the coeﬃcients aij no longer appear under derivative signs in (1. the smoothness requirements imposed on the coeﬃcients bi and c can be relaxed: bi ∈ L∞ (Ω) for i = 1. (1. . . remembering that u has to satisfy a zero Dirichlet boundary condition. c(x)uv dx = Ω f (x)v(x) dx In order for this equality to make sense we no longer need to assume that u ∈ C 2 (Ω): it is suﬃcient that u ∈ L2 (Ω) and ∂u/∂xi ∈ L2 (Ω). the expressions on the left. Thus. . (instead of v ∈ C0 (Ω)). we consider the following problem: ﬁnd u in H0 (Ω) such that n i.7) 1 1 1 We note that C0 (Ω) ⊂ H0 (Ω). let us suppose that u is a classical solution of (1. n. n − i. . aij ∈ L∞ (Ω) will be seen to be suﬃcient. To begin. . as in Section 1. . for 1 any v ∈ C0 (Ω). Then.j=1 Ω ∂ ∂xj ∂u aij ∂xi + Ω n · v dx + bi (x) i=1 Ω ∂u · v dx ∂xi c(x)uv dx = Ω f (x)v(x) dx. . n and c ∈ L∞ (Ω) will suﬃce. it is natural to seek u in 1 the space H0 (Ω). 1 H0 (Ω) = {u ∈ L2 (Ω) : ∂u ∈ L2 (Ω).6).j=1 ∂u ∂v aij (x) · dx + ∂xi ∂xj Ω + Ω n bi (x) i=1 Ω ∂u v dx ∂xi 1 ∀v ∈ C0 (Ω). i = 1.3. ∂xi 1 Therefore. .1. This motivates the following deﬁnition. we shall prove this below)8 . . i = 1.
despite the fact that it has no classical solution. problem (1.8) is called a weak solution of (1.8) can be written as follows: 1 ﬁnd u ∈ H0 (Ω) such that a(u. . .2. . . (1. n. Indeed. this may not be smooth enough to be a classical solution. A function u ∈ H0 (Ω) satisfying n i. w ∈ V (c) ∃c2 > 0 ∀v ∈ V and let l(·) be a linear functional on V such that l(v) ≤ c2 v V a(w. Theorem 1 (Lax & Milgram theorem) Suppose that V is a real Hilbert space equipped with norm · V .5). . v) ≤ c1 w . WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 17 Deﬁnition 1 Let aij ∈ L∞ (Ω). we look at the wider issue of existence of a unique weak solution to the more general problem (1.5).8) should be understood as weak derivatives. and let f ∈ L2 (Ω). i. we shall prove 1 below that the boundary value problem (∗) has a unique weak solution u ∈ H0 (Ω).5). (1. If (1. i = 1. Let a(·. j = 1.9) and l(v) = Ω f (x)v(x) dx. bi ∈ L∞ (Ω). . Before considering this particular boundary value problem.10) With this new notation. v V.5). (1.11) We shall prove the existence of a unique solution to this problem by exploiting the following abstract result from Functional Analysis. c ∈ 1 L∞ (Ω).6). Clearly if u is a classical solution of (1. All partial derivatives in (1. v) ≥ c0 v 2 V .1. ·) be a bilinear functional on V × V such that: (a) ∃c0 > 0 ∀v ∈ V a(v. . For the sake of simplicity.6).6) has a weak solution.j=1 Ω n aij (x) + i=1 Ω ∂w ∂v dx ∂xi ∂xj ∂w v dx + ∂xi c(x)wv dx Ω bi (x) (1. . However. (1. (1.6). then it is also a weak solution of (1. n.6).5). (1. v) = l(v) 1 ∀v ∈ H0 (Ω). . (1.j=1 ∂u ∂v dx + aij (x) ∂xi ∂xj Ω + Ω n bi (x) i=1 Ω ∂u v dx ∂xi 1 ∀v ∈ H0 (Ω) c(x)uv dx = Ω f (x)v(x) dx (1. the converse is not true. we adopt the following notation: n a(w. V (b) ∃c1 > 0 ∀v. v) = i.
deﬁned by (1. NorthHolland.6) (or. = α f (x)v2 (x) dx = αl(v1 ) + βl(v2 ). Also. (b). 1/2 1/2 Ω l(v) =  = f Ω f (x)v(x) dx ≤ L2 (Ω) Ω f (x) dx v H 1 (Ω) . Letting c2 = f L2 (Ω) . v)H 1 (Ω) = Ω wv dx + i=1 1/2 Ω ∂w ∂v · dx ∂xi ∂xi and the associated norm w H 1 (Ω) = (w. l(αv1 + βv2 ) = Ω f (x)(αv1 (x) + βv2 (x)) dx f (x)v1 (x) dx + β Ω Ω 1 v1 .9) and (1. w) is linear. INTRODUCTION ∀v ∈ V. ·) 1 1 is a bilinear functional on H0 (Ω) × H0 (Ω). Ciarlet: The Finite Element Method for Elliptic Problems. where we have used the obvious inequality v L2 (Ω) ≤ v H 1 (Ω) . for any ﬁxed v ∈ H0 (Ω). Applying the CauchySchwarz inequality. 1 Next we verify (b). v2 ∈ H0 (Ω). ·) and l(·). 1 so l(·) is a linear functional on H0 (Ω). 2 v(x) dx 2 v L2 (Ω) ≤ f L2 (Ω) 1 for all v ∈ H0 (Ω). The mapping v → l(v) is linear: indeed. we deduce that n a(w. Hence a(·.10). v) = l(v) CHAPTER 1.11)). to 1 (1.3 that H0 (Ω) is a Hilbert space with the inner product n (w. 1 We apply the LaxMilgram theorem with V = H0 (Ω) and · V = · H 1 (Ω) to show the existence of a unique weak solution to (1. by the CauchySchwarz inequality. for any α. (c) of the LaxMilgram theorem. the mapping w → a(v. Next we show that a(·. For any ﬁxed w ∈ H0 (Ω). K. 1995.. SpringerVerlag. Yosida: Functional Analysis. β ∈ R.1. the mapping v → a(v. we obtain the required bound.j=1 n max aij (x)  ¯ x∈Ω Ω ∂w ∂v dx ∂xi ∂xj ∂w v dx ∂xi + i=1 max bi (x)  ¯ x∈Ω Ω . v) ≤ i. Let us recall from Section 1. Reprint of the 6th ed. w)H 1 (Ω) . w) is linear. 1978. satisfy the hypotheses (a). (1. there exists a unique u ∈ V such that a(u. equivalently. We begin with (c).18 Then. For a proof of this result the interested reader is referred to the books: P.5). 1 Similarly.
To do so. max c(x) .4). . 2 ∂xi Integrating by parts in the second term on the 1 c(x) − 2 n a(v.1. v) ≤ c1 w H 1 (Ω) v H 1 (Ω) . we obtain inequality (b): c a(w. j=1 so that.j=1 n Ω  ∂w 2  dx ∂xi  ∂v 2  dx ∂xj 1/2 2 1/2 i=1 ∂w 2  dx  Ω ∂xi 1/2 2 v dx 1/2 + Ω w dx 1/2 Ω v dx n ≤ c ˆ × where c = max ˆ ¯ 1≤i. by letting c1 = 2nˆ.2.13) It remains to establish (a).12) we deduce that n a(w. max max bi (x). v) ≥ c ˜ i=1 ∂v 2   dx + Ω ∂xi Ω i=1 ∂bi ∂xi v2 dx. however. we shall slightly strengthen the smoothness 1 requirements on the coeﬃcients bi by demanding that bi ∈ W∞ (Ω) (see. (1. v) ≤ 2nˆ c × Ω w dx + 2 2 i=1 n ∂w 2  dx  Ω ∂xi ∂v 2   dx Ω ∂xj 1/2 1/2 Ω v dx + . WEAK SOLUTIONS TO ELLIPTIC PROBLEMS + max c(x)  ¯ x∈Ω n 19 w(x)v(x) dx Ω 1/2 Ω 1/2 Ω 2 ≤ c ˆ + i. Using (1. v) ≥ c ˜ i=1 ∂v ∂xi ∂v 2  dx +  Ω ∂xi · v as n n bi (x) i=1 Ω 1 ∂ 2 (v ) dx + 2 ∂xi c(x)v2 dx.j≤n x∈Ω Ω w dx 1/2 2 + i=1 n ∂w 2   dx Ω ∂xi ∂v 2   dx Ω ∂xj 1/2 1/2 Ω v dx 2 + j=1 (1. we obtain 1 ∂ (v 2 ). Ω where we wrote right. ¯ 1≤i≤n x∈Ω ¯ x∈Ω By further majorisation of the righthand side in (1.12) max max aij (x). Remark 4 at the end of this chapter). we deduce that n a(v.
applying the above theorem with aij (x) ≡ 1. However. i = 1. n.16). c ∈ L∞ (Ω). i = 1. f (x) = sgn( 1 − x).15) By virtue of the Poincar´Friedrichs inequality stated in Lemma 1. We encapsulate this result in the following theorem. (1. ∂xi (1. . . c0 u 2 H 1 (Ω) ≤ (f. (1. v) ≥ c ˜ i=1 Ω  ∂v 2  dx.14) a(v.11). problem (1. and hence (a). ∂xi ¯ x ∈ Ω. . .2.17). j = 1. consequently. as well as mixed problems.14) hold. Robin. we see that (1. By (1. bi ∈ W∞ (Ω). . we deduce the existence of a unique u ∈ H0 (Ω) satisfying (1.5). i.17) where c0 = c/(1 + c⋆ ). and oblique derivative boundary value problems. .11). and c satisfy the inequality 1 c(x) − 2 Then n n i=1 ∂bi ≥ 0. . .18) as the rest of the theorem has been proved above.6) possesses a unique weak solution u ∈ H0 (Ω). aij (x) ≡ 0. the righthand e side can be further bounded below to obtain a(v.18) Proof We only have to show (1. . (1. . i = j. u) ≤ f ≤ Hence the desired inequality. INTRODUCTION Suppose that bi . f ∈ L2 (Ω). then the boundary 1 value problem (1. j ≤ n. Having checked all hypotheses of the Lax˜ 1 Milgram theorem. u) L2 (Ω) H 1 (Ω) u L2 (Ω) u . 1)n .14) is trivially fulﬁlled.4) and (1. (1. bi (x) ≡ 0. ∂xi (1. i = j. . (1. n.4) holds with c = 1 and (1. In addition. f L2 (Ω) ≤ a(u.15) and (1. n c ˜ c⋆ Ω v2 dx.16) a(v. v) ≥ c0 Ω v dx + 2 i=1 Ω  ∂v 2  dx . v) ≥ Summing (1.20 CHAPTER 1.5). Similar results are valid in the case of Neumann. the CauchySchwarz inequality and recalling the deﬁnition of · H 1 (Ω) . ˜ 1 Thus (∗) has a unique weak solution u ∈ H0 (Ω) by Theorem 2.6) has a unique weak solution. u H 1 (Ω) ≤ 1 f c0 L2 (Ω) . Now we return to our earlier example (∗) which has been shown to have no classical solution. n. 1 Theorem 2 Suppose that aij ∈ L∞ (Ω). 1 ≤ i. (1. . and 2 Ω = (−1. u) = l(u) = (f. . c(x) ≡ 0. and assume that (1.
v) = l(v) for all v in H0. n.Γ1 (Ω).18). To see this. we consider the special Sobolev space 1 H0. note that the smoothness requirements on bi are unrelated to the veriﬁcation of condition (c) in the LaxMilgram .2. ⋄ 1 Remark 4 The requirement bi ∈ W∞ (Ω) in Theorem 2 can be relaxed to the original assumption bi ∈ L∞ (Ω).19) and hence the required continuous dependence of the solution of the boundary value problem on the righthand side. (1. WEAK SOLUTIONS TO ELLIPTIC PROBLEMS 21 Remark 2 Consider. on Γ1 . relatively open subset of ∂Ω and Γ1 ∪ Γ2 = ∂Ω. u1 − u2 H 1 (Ω) ≤ 1 f1 − f2 c0 L2 (Ω) . respectively. Following a similar reasoning as in the case of the Dirichlet boundary value problem. on Γ2 . then u1 − u2 1 is the weak solution in H0 (Ω) of (1.5). (1. .1. (1.Γ1 (Ω) = {v ∈ H 1 (Ω) : v = 0 on Γ1 }. .5). for example.Γ1 (Ω) such that 1 a(u. . ⋄ Remark 3 Theorem 2 implies that the weak formulation of the elliptic boundary value problem (1. where Γ1 is a nonempty.6) is wellposed in the sense of Hadamard. the following DirichletNeumann mixed boundary value problem: −∆u = f u = 0 ∂u = g ∂ν in Ω.5). We shall suppose that f ∈ L2 (Ω) and that g ∈ L2 (Γ2 ).Γ1 (Ω). Thus. namely. and “small” changes in f give rise to “small” changes in the corresponding solution u.6) corresponding to the righthand side f1 − f2 ∈ L2 (Ω). The latter property 1 follows by noting that if u1 and u2 are weak solutions in H0 (Ω) of (1. where we put a(u. by virtue of (1. (1. the existence and uniqueness of a weak solution to this mixed problem easily follows. i = 1. 1 and deﬁne the weak formulation of the mixed problem as follows: ﬁnd u ∈ H0. . Γ2 and l(v) = Ω f (x)v(x) dx + 1 Applying the LaxMilgram theorem with V = H0. v) = n Ω i=1 ∂u ∂v dx ∂xi ∂xi g(s)v(s) ds. for each 1 f ∈ L2 (Ω) there exists a unique (weak) solution u ∈ H0 (Ω).6) corresponding to righthand sides f1 and f2 in L2 (Ω).
22 CHAPTER 1. ⋄ . n. this veriﬁes condition (a) in the LaxMilgram theorem.4) and the CauchySchwarz inequality. .4). . . only and (1. Thus. n.17) we arrive at (1. Thus. i = 1.15) to (1. . and condition (b) can be shown with bi ∈ L∞ (Ω). it remains to see how condition (a) may be veriﬁed under the hypothesis bi ∈ L∞ (Ω).20) hold. . n 1/2 a(v. c(x) − we arrive at the inequality 2 c ˜ bi i=1 2 L∞ (Ω) ≥0 (1. . INTRODUCTION theorem. v) ≥ c ˜ 2 i=1 n Ω  ∂v 2  dx. i = 1. . . proceeding in the same way as in the transi˜ tion from (1. i = 1. ∂xi which is analogous to (1. .15). under the assumptions that bi ∈ L∞ (Ω). n. .17) with c0 = c/(2 + 2c∗ ). By (1. . v) ≥ cv2 1 (Ω) ˜ H − bi 2 ∞ (Ω) L i=1 vH 1 (Ω) v n L2 (Ω) + Ω c(x)v(x)2 dx 1 cv2 1 (Ω) + ˜ H ≥ 2 Assuming that Ω 2 c(x) − c ˜ n bi i=1 2 L∞ (Ω) v(x)2 dx. only anyway.20) 1 a(v. . (1.
10)). (∗∗) where Ui . have “small” support. i = 1.g. . . The ﬁrst step in the construction of a ﬁnite element method for an elliptic boundary value problem (e. Suppose. (P ) 1 where V is the solution space (e. v) = l(v) ∀v ∈ V . for example.Chapter 2 Approximation of elliptic problems In this chapter we describe the construction of ﬁnite element methods for elliptic boundary value problems and outline some of their key properties.9) and (1. Expressing the approximate solution uh in terms of the basis functions. the derivation of ﬁnite element methods is quite systematic. . . .6)) is to convert the problem into its weak formulation: ﬁnd u ∈ V such that a(u. that dim Vh = N (h) and Vh = span{φ1 . . . vh ) = l(vh ) ∀vh ∈ Vh . ·) is a bilinear functional on V ×V . we can write N (h) uh (x) = i=1 Ui φi (x). H0 (Ω) for the homogeneous Dirichlet boundary value problem). φN (h) }. . .g. . i = 1. N (h). (1. (1. then consider the following approximation of (P ): ﬁnd uh ∈ Vh such that a(uh .5). (Ph ) where the (linearly independent) basis functions φi . are to be determined. Unlike ﬁnite diﬀerence schemes which are constructed in a moreorless ad hoc fashion through replacing the derivatives in the diﬀerential equation by divided diﬀerences.g. . The second step in the construction is to replace V in (P ) by a ﬁnitedimensional subspace Vh ⊂ V which consists of continuous piecewise polynomial functions of a ﬁxed degree associated with a subdivision of the computational domain. . . Thus (Ph ) can be rewritten as follows: ﬁnd (U1 . . φi . and l(·) is a linear functional on V (e. a(·. . (1. UN (h) ) ∈ RN (h) such that 23 . . N (h).
u(0) = 0. j = 1. N ≥ 2.1 Piecewise linear basis functions In this section we describe the construction of the ﬁnite element method through two simple examples: the ﬁrst of these is a twopoint boundary value problem for a secondorder ordinary diﬀerential equation. . . . . so the matrix A is sparse (in the sense that most of its entries are equal to 0). . φj )Ui = l(φj ). Higherdegree piecewise polynomial approximations will be discussed later on in the notes. 1). N . . 1).1) (2. 1]. 1] into N subintervals [xi . 1]. f ∈ L2 (0.2) In order to construct the ﬁnite element approximation of this problem. . 2. φi )) of size N (h) × N (h). Once (Ph ) has been solved for U = (U1 . . the expansion (∗∗) provides the required approximation to u. N (h). 1) such that 1 1 1 ′ ′ f (x)v(x) dx q(x)u(x)v(x) dx = p(x)u (x)v (x) dx + (P ) 0 0 0 1 ∀v ∈ H0 (0. 1) with p(x) ≥ c > 0 and q(x) ≥ 0 for all x ˜ in [0. fast iterative methods are available for sparse ′ linear systems. φi ) = 0 for most pairs of i and j. u(1) = 0. . q ∈ C[0. Because the φi ’s have small support. i = 0. a(φj . Onedimensional problem Let us consider the boundary value problem −(p(x)u′ )′ + q(x)u = f (x). . by the points xi = ih. where p ∈ C[0.1. . We note that in general the mesh points xi need not be equally spaced: here we have chosen a uniform spacing only to simplify the exposition. we illustrate the construction of this numerical technique by considering some simple examples. . we sub¯ divide Ω = [0. APPROXIMATION OF ELLIPTIC PROBLEMS N (h) a(φi . where h = 1/N . . UN (h) )T .24 CHAPTER 2. x ∈ (0. i = 0. 1]. . . . . N − 1. xi+1 ]. with the matrix of the system A = (a(φj . For the time being we shall assume that the ﬁnite element space Vh consists of continuous piecewise linear functions. . . After this brief outline of the idea behind the ﬁnite element method. UN (h) )T . . this property is crucial from the point of eﬃcient solution – in particular. (2. the second model problem is the homogeneous Dirichlet boundary value problem for Poisson’s equation on the unit square in the plane. The weak formulation of this problem is: 1 ﬁnd u ∈ H0 (0. i=1 ′ (Ph ) This is a system of linear equations for U = (U1 . . as shown in Fig 2.
2. . N − 1. furthermore. . . (hence the name ﬁnite element method). . xi+1 ]. xi+1 ) are referred to as element domains or elements. are linearly independent. and the functions φi . . the weak solution 1 u ∈ H0 (0. . . xN = 1 ¯ Figure 2. . The subintervals (xi . 1 shown in Figure 2.1: Subdivision of Ω = [0. PIECEWISE LINEAR BASIS FUNCTIONS 25 0 = x0 x1 x2 .. It will be convenient to express our approximation as a linear combination of the ﬁnite element basis functions φi (x) = 1− x − xi  h . i = 1. i = 1. . 1). .. 1].. + i = 1. xi . it can be written as a linear combination ∀vh ∈ Vh . . . . N − 1.1. .2: The piecewise linear ﬁnite element basis function φi (x). Z Z Z Z Z Z 1 Z Z Z Z Z Z xi−1 xi xi+1 Figure 2. N − 1. It is clear that φi ∈ H0 (0. . . . The ﬁnite element approximation of (P ) is: ﬁnd uh ∈ Vh such that 1 0 ′ p(x)u′h (x)vh (x) dx + 1 q(x)uh (x)vh (x) dx 0 1 = 0 f (x)vh (x) dx Since uh ∈ Vh = span{φ1 ..2. . therefore Vh := span{φ1 . In this example. (Ph ) . . . supp φi = [xi−1 .1. 1) of problem (P ) will be approximated by a continuous piecewise linear function on the subdivision depicted in Figure 2. 1). φN −1 }. φN −1 } 1 is an (N − 1)dimensional subspace of H0 (0.
i − j > 1. . . FN −1 )T . i − j > 1. . 1 Fj := 0 f (x)φj (x) dx. i − j = 1. + q 0. . .e. UN −1 into the expansion N −1 uh (x) = i=1 Ui φi (x) to obtain uh . . . . ′ (Ph ) can be written as a system of linear equations AU = F. AT = A) and positive deﬁnite (i. i = j. . . 0. Substituting this expansion into (Ph ) we obtain the following problem. In practice the entries aji of the matrix A and the entries Fj of the vector F are calculated approximately using numerical integration (quadrature) rules. i − j > 1. i − j = 1. equivalent to (Ph ): ﬁnd U = (U1 . . we substitute the values U1 . N − 1. N − 1. . 1]. APPROXIMATION OF ELLIPTIC PROBLEMS of the basis functions: N −1 uh (x) = i=1 Ui φi (x). . h/6.e. In the simple case when p and q are constant functions on [0. . = p −1/h. 0 0 φ′i (x)φ′j (x) dx + q . = 0. j = 1. = 0 for j = 1. −p/h + qh/6. the entries of A can be calculated exactly: 1 1 aij = p φi (x)φj (x) dx 2/h. Since supp φi ∩ supp φj has empty interior when i − j > 1. Having solved the system of linear equations AU = F . The matrix A is symmetric (i. Letting 1 aji := 0 [p(x)φ′i (x)φ′j (x) + q(x)φi (x)φj (x)] dx. j = 1. i − j = 1. N − 1.26 CHAPTER 2. where A = (aji ). i = j. . aji is zero. . UN −1 )T ∈ RN −1 such that N −1 1 ′ ′ [p(x)φi (x)φj (x) + q(x)φi (x)φj (x)] dx Ui 0 i=1 (P ′ ) 1 h f (x)φj (x) dx. it follows that the matrix A is tridiagonal (namely. i = j. xT Ax > 0. . . 4h/6. unless i − j ≤ 1). . x = 0). 2p/h + 4hq/6. . . . F = (F1 . i. . .
. φN (h) are linearly independent and they span an N (h)dimensional linear subspace Vh of 1 H0 (Ω). N − 1. . as shown in Fig. 2. . and we deﬁne h = maxK hK . let φ1 (x. thus Ω can be exactly covered by a ﬁnite number of triangles. φN (h) (x. or not at all. Let us suppose that the interior nodes are labelled 1. . . . y) be the corresponding basis functions. .2. the values of uh (x) at the mesh points xi .3. at a vertex. respectively). Twodimensional problem Let Ω be a bounded domain in R2 with a polygonal boundary ∂Ω.1. N (h). This gives rise to the following set of linear equations: Ui−1 + 4Ui + Ui+1 1 Ui−1 − 2Ui + Ui+1 +q = −p 2 h 6 h xi+1 f (x)φi (x) dx. φ is assumed ¯ to be a continuous function on Ω and linear in each of the triangles. with the convention that U0 = 0 and UN = 0 (corresponding to the fact that uh (0) = 1 and uh (1) = 0. . as shown in Fig. . .3: A subdivision (triangulation) of Ω.4. y). This is a threepoint ﬁnite diﬀerence scheme for the values Ui . We shall denote by hK the diameter (longest side) of triangle K. . xi−1 i = 1. . With each interior node (marked ⊙ in the ﬁgure) we associate a basis function φ which is equal to 1 at that node and equal to 0 at all the other nodes. Let us consider the elliptic boundary value problem −∆u = f in Ω. 2. PIECEWISE LINEAR BASIS FUNCTIONS 27 ¯ Figure 2. . . . It will be assumed that any pair of triangles in a triangulation of Ω intersect along a complete edge. The functions φ1 . . 2. .
uh (x. . APPROXIMATION OF ELLIPTIC PROBLEMS 1 0 0 0 0 0 0 Figure 2. UN (h) )T ∈ RN (h) such that . .4: A typical ﬁnite element basis function φ.28 CHAPTER 2. y) = i=1 Ui φi (x. . In order to construct the ﬁnite element approximation of the problem. the ﬁnite element method can be restated as follows: ﬁnd U = (U1 . we begin by considering its weak formulation (see the discussion about weak solutions in Chapter 1 in the special case when n = 2. aij (x) ≡ 1 for i = j and ≡ 0 for i = j. Ω The ﬁnite element approximation of the problem is: ﬁnd uh ∈ Vh such that ∂uh ∂vh ∂uh ∂vh + ∂x ∂x ∂y ∂y Ω Writing N (h) dx dy = Ω f vh dx dy ∀vh ∈ Vh . bi (x) ≡ 0 for all i and c(x) ≡ 0): 1 ﬁnd u ∈ H0 (Ω) such that ∂u ∂v ∂u ∂v + dx dy = ∂x ∂x ∂y ∂y Ω 1 f v dx dy ∀v ∈ H0 (Ω). . y). u = 0 on ∂Ω.
y) = i=1 Ui φi (x. . . and hence the approximate solution N (h) uh (x. Solving this. y) ∈ 1 1 − y−yj . 2.6. . (x. dx dy. y) ∈ 6 h 0 otherwise. . To simplify matters let us suppose that Ω = (0. y) ∈ 1 0. aij = aji = Ω ∂φi ∂φj ∂φi ∂φj + ∂x ∂x ∂y ∂y f φj dx dy. ∂φij ∂x .5. y) ∈ 3 h yj −y xi −x φij (x. (x. .) Thus. we obtain U = (U1 . y) ∈ 6 0. PIECEWISE LINEAR BASIS FUNCTIONS N (h) 29 Ui i=1 Ω ∂φi ∂φj ∂φi ∂φj + ∂x ∂x ∂y ∂y dx dy = Ω f φj dx dy. where 1. . The case of a general triangulation will be considered later. yj ) (see Fig. for j = 1. y) ∈ 3 1/h. 6 denote the triangles surrounding the node (xi . yj ): y−yj x−x 1 − h i − h . 1) and consider the ¯ triangulation of Ω shown in Fig. .1. 1) × (0. . (x. 2. . y) ∈ 4 = 0. 2. . . UN (h) )T . Letting A = (aij ). y) ∈ 5 h 1 − x−xi . (x. (x. FN (h) )T . (x. (x.2. (x. . . (x. F = (F1 . −1/h. y) ∈ 4 1 − yj −y . (x. y) ∈ 5 −1/h. (x. N (h). (x. Let φij denote the basis function associated with the interior node (xi . y) ∈ 2 h 1 − xi −x . y) = 1 − h − h . y) ∈ 2 1/h. y). . . Fj = Ω the ﬁnite element approximation can be written as a system of linear equations AU = F. otherwise. . The matrix A is called the stiﬀness matrix.
@ @ @ @ @ @ 2 @ @ 3 @ 1 @@ @ i s @ @ @ @ @ 4 @ 6 @ @ @ @ 5 @ @ @ @ @ Figure 2. y1  x x0 = 0 x1 . .5: Triangulation of Ω = [0.6: Triangles surrounding a node. .. xN = 1 ¯ Figure 2.30 CHAPTER 2. 1] × [0.. APPROXIMATION OF ELLIPTIC PROBLEMS y 6 yN = 1 @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @s @s i i i @s @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ i i i @ @s @s @s @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ i i i @ @s @s @s @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ . . 1].
. N − 1. (x. y) ∈ 2 0.2.. bi ∈ W∞ (Ω). l = 1. y)φkl (x. . (2. x ∈ Ω.l−1 − h2 h2 1 f (x. l = 1. 31 ∂φij ∂y Since N −1 N −1 Uij i=1 j=1 Ω ∂φij ∂φkl ∂φij ∂φkl + ∂x ∂x ∂y ∂y = 4Ukl − Uk−1. the ﬁnite element approximation is equivalent to − Uk+1. . (x. y) ∈ 3 1/h. .4) is: 1 1 ﬁnd u ∈ H0 (Ω) such that a(u.j=1 aij (x)ξi ξj ≥ c ˜ i=1 ¯ ξi2 ∀ξ = (ξ1 .l+1 . (x.l + Uk−1. the elliptic boundary value problem − aij (x) ∂u ∂xi n + i=1 bi (x) ∂u + c(x)u = f (x). (x. the ﬁnite element approximation gives rise to the familiar 5point ﬁnite diﬀerence scheme with the forcing function f averaged in a special way. . f ∈ L2 (Ω). .2 The selfadjoint elliptic problem ∂ ∂xj i. ξn ) ∈ Rn . (2.j=1 n Let us consider. dx dy k. Thus. on this special triangulation of Ω. y) dx dy. n. i = 1. = 2 h supp φkl Ukl = 0 on ∂Ω. otherwise.l − 2Uk. . y) ∈ 6 0. as in Chapter 1.l + Uk. ∂xi u = 0 on ∂Ω.l − Uk+1. y) ∈ 4 = 1/h. k. aij ∈ L∞ (Ω).. . j = 1. 2. .l Uk. . (x. (2. . .5) We recall from Chapter 1 that the weak formulation of (2. and assume that there exists a positive constant c such that ˜ n n i. c ∈ L∞ (Ω). i. (x. ∀x ∈ Ω. .l+1 − 2Uk. y) ∈ 1 −1/h. ..3) (2. (2. . THE SELFADJOINT ELLIPTIC PROBLEM and −1/h. y) ∈ 5 0. n. . v) = l(v) ∀v ∈ H0 (Ω).4) 1 where Ω is a bounded open set in Rn .3). . N − 1.6) .l−1 − Uk.l − Uk.2.
6) in H0 (Ω) and suppose that 1 a(·. . v) − l(v). In the special case when the boundary value problem is selfadjoint.4). APPROXIMATION OF ELLIPTIC PROBLEMS where the bilinear functional a(·. x ∈ Ω. . To be more 1 precise.6) has a unique solution u in H0 (Ω). in the rest of this section this will always be assumed to be the case. . we deﬁne the quadratic functional J : H0 (Ω) → R by 1 J(v) = a(v. Ω and l(v) = Ω f (x)v(x) dx.32 CHAPTER 2. the weak solution of (2. ¯ aij (x) = aji (x). ·) is symmetric in the sense that 1 a(v. v) ∀v.e. x ∈ Ω. aij (x) = aji (x). n. .7) can be restated as a minimisation problem. i. i = 1. i. . j = 1.j=1 n aij (x) ∂u ∂xi + c(x)u = f (x). v) = i. w ∈ H0 (Ω).3). x ∈ Ω. x ∈ Ω. w) = a(w.7) u = 0. x ∈ Ω. ∂xi 1 then (2. ·) is a symmetric bilinear functional on H0 (Ω).j=1 Ω aij ∂u ∂v dx + ∂xi ∂xj n bi (x) i=1 Ω ∂u v dx + ∂xi c(x)uv dx. on ∂Ω ¯ with aij (x) satisfying the ellipticity condition (2. Thus we consider − ∂ ∂xj i. . (2. and ¯ bi (x) ≡ 0.5). then u is the unique minimiser 1 of J(·) over H0 (Ω). We have shown that if c(x) − 1 2 n i=1 ∂bi ¯ ≥ 0. c(x) ≥ 0. 2 1 v ∈ H0 (Ω). It turns out that (2. 1 Lemma 3 Let u be the (unique) weak solution to (2. (2. . . . ·) and the linear functional l(·) are deﬁned by n a(u. the bilinear functional a(·. n.
u) + l(u) 2 2 1 1 a(v. c0 v−u 2 2 H 1 (Ω) 1 ∀v ∈ H0 (Ω).6) in H0 (Ω) and. u minimises J(·) over H0 (Ω). v − u). u) − a(u.9) and v = u in (2. v) − a(u. i. u) + a(u. for v ∈ H0 (Ω).e. (2. v − u). Thus J(v) − J(u) ≥ and therefore. Indeed.10) Taking v = u in (2. then on H0 1 J(v) ≥ J(˜) ∀v ∈ H0 (Ω). v) − a(u.10). v − u) 2 2 1 [a(v. if u also minimises J(·) ˜ 1 (Ω). u) − l(v − u) 2 2 1 1 a(v. v) − a(v. 2 1 J(v) − J(u) = a(v − u. v) − a(u.e. u)] 2 1 [a(v. v) + a(u. 2 Because of (1. consider J(v) − J(u): J(v) − J(u) = = = = = = Thence 1 1 a(v. a(v − u. J((1 − θ)v + θw) ≤ (1 − θ)J(v) + θJ(w) ∀θ ∈ [0.2. 1 In fact. u−u ˜ and hence u = u. u is the unique minimiser of J(·) in H0 (Ω). v) − l(v) − a(u.9) 1 i. 2 H 1 (Ω) .2. .8). w ∈ H0 (Ω). THE SELFADJOINT ELLIPTIC PROBLEM 33 1 1 Proof Let u be the unique weak solution to (2. v − u) ≥ c0 v − u where c0 is a positive constant. u (2. u but then. v) − 2a(u. 1 J(v) ≥ J(u) ∀v ∈ H0 (Ω). 1 ∀v. we deduce that ˜ J(u) = J(˜). by virtue of (2. It is easily shown that J(·) is convex (down).8) (2. ˜ H 1 (Ω) = 0. u)] 2 1 a(v − u. 1].17).
v − w) 2 and the fact that a(v − w. v). v) − l(v) = 0 ∀v ∈ H0 (Ω). v − w) ≥ 0. if u minimises J(·) then J(·) has a stationary point at u.7: The quadratic functional J(·). v)] = a(u.34 CHAPTER 2. This follows from the identity 1 (1 − θ)J(v) + θJ(w) = J((1 − θ)v + θw) + θ(1 − θ)a(v − w.6) is called the Euler–Lagrange equation for this minimisation problem. This lemma is precisely the converse of the previous lemma. J ′ (u)v := lim 1 for all v ∈ H0 (Ω). then u is the (unique) solution of problem (2. Since J(u + λv) − J(u) =0 λ→0 λ λ J(u + λv) − J(u) = a(u.6). λ→0 2 which proves the following result. namely. and the two results together express the equivalence of the weak formulation: . The problem (2. v) − l(v) + a(v. 1 1 Lemma 4 Suppose that u ∈ H0 (Ω) minimises J(·) over H0 (Ω). APPROXIMATION OF ELLIPTIC PROBLEMS R J(v) J(u) 1 H0 (Ω) u Figure 2. λ 2 we deduce that if u minimises J(·) then λ 1 lim [a(u. v) − l(v) + a(v. Moreover.
7) to the associated minimisation problem: 1 1 ﬁnd u ∈ H0 (Ω) such that J(u) ≤ J(v) ∀v ∈ H0 (Ω). at a vertex or not at all. we now suppose that Ω is a bounded polygonal domain in the plane. 2.3 Calculation and assembly of stiﬀness matrix Using the variational characterisation of uh described at the end of the previous section we return to the construction of the ﬁnite element approximation to Poisson’s equation −∆u = f in Ω subject to homogeneous Dirichlet boundary condition. vh ) = l(vh ) ∀vh ∈ Vh . in general J(u) < J(uh ). Thus. We consider . Given that Vh is a 1 certain ﬁnitedimensional subspace of H0 (Ω) which consists of continuous piecewise polynomials of a ﬁxed degree. CALCULATION AND ASSEMBLY OF STIFFNESS MATRIX 35 1 1 ﬁnd u ∈ H0 (Ω) such that a(u. so that any pair of (closed) triangles intersect only along a complete edge. v) = l(v) ∀v ∈ H0 (Ω) (W ) of the selfadjoint elliptic boundary value problem (2. uh can be characterised as the unique minimiser of the functional 1 J(vh ) = a(vh . (Wh ) 1 We can repeat the argument presented above (or simply replacing H0 (Ω) by Vh throughout) to show the equivalence of (Wh ) to the following minimisation problem: ﬁnd uh ∈ Vh such that J(uh ) ≤ J(vh ) ∀vh ∈ Vh .3.2. u = 0 on ∂Ω. (M ) We shall now use this equivalence to give a variational characterisation of the ﬁnite element approximation uh to u in the selfadjoint case. vh ) − l(vh ) 2 (Mh ) as vh ranges over the ﬁnite element space Vh . This means that the ﬁnite element solution uh inherits the energy minimisation property possessed by the weak solution 1 u ∈ H0 (Ω) in the sense that: J(uh ) = min J(vh ). in the case of a general triangulation. subdivided into M triangles K. vh ∈Vh Of course. the ﬁnite element approximation of (W ) is:. Rather than restricting ourselves to the special case when Ω is a square. ﬁnd uh ∈ Vh such that a(uh .
Thus. Equivalently. y)φi (x. i = 1. y) ∂(ξ. y) coordinate system. η). φj ) = (∇φi . and N is the number of nodes internal to Ω. Consider any triangle K in the triangulation of Ω. y)2 dx dy − f (x. y) = i=1 Vi φi (x. . η) coordinate system and the canonical triangle depicted in Figure 2. ψ3 } is called the nodal basis (or local basis) for the set of linear polynomials in terms of the local coordinates.12) The set {ψ1 . y) · ∇φj (x. j) entry a(φi .an N × N matrix with (i. writing N vh (x. y).8. . thus. 2. VN )T . we consider a socalled local (ξ. . . φi ) = Ω f (x. the linear space consisting of all such functions vh is denoted Vh . φi is the continuous piecewise linear basis function associated with this node. (2. where Vi is the value of vh (x. say. 3. y) dx dy Ω Ω as vh ranges over Vh . In addition. Consider the transformation (ξ. 2 (2. η) → r = (x. . y) deﬁned by (2. with Fi = (f. FN )T is the (global) load vector. we can write this minimisation problem in matrix form as follows: ﬁnd V ∈ RN such that 1 V T AV − V T F is minimum. ∇φj ) = ∇φi (x. APPROXIMATION OF ELLIPTIC PROBLEMS the set of all continuous piecewise linear functions vh deﬁned on such a triangulation with the property that vh = 0 of ∂Ω. η) + r3 ψ3 (ξ. Ω and F = (F1 . . y) dx dy. J = T ∂(x. yi ). η) + r2 ψ2 (ξ. .12) from the canonical triangle to the ‘global’ (x. y) at the node (xi .11) where V = (V1 . A is the (global) stiﬀness matrix . uh is characterised as the unique minimiser of the functional J(vh ) = 1 2 ∇vh (x. . η) T = x2 − x1 y2 − y1 x3 − x1 y3 − y1 . y)vh (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 − ξ − η)r1 + ξr2 + ηr3 ≡ r1 ψ1 (ξ. Let J T denote the transpose of the Jacobi matrix J of this transformation.36 CHAPTER 2. yi ). The coordinate r = (x. of its three vertices labelled in the anticlockwise direction. y) dx dy. and introduce the position vectors ri = (xi . ψ2 .
for any function vh ∈ Vh . J = 2A123 where A123 is the area of the triangle K = ∆(r1 . η).8: Canonical triangle and local coordinates. i = 1. from (2. 3. ∂vh ∂y = (J T )−1 ∂ξ ∂vh ∂η . 2. 0) ξ Figure 2.16) + r2 − r1 2 ∂vh ∂vh ∂ξ ∂η ∂vh ∂η 2 −2(r3 − r1 ) · (r2 − r1 ) (2. however. x3 y3 1 (2.12) and the form of the matrix J T we have that ∂vh ∂vh ∂vh ∂vh ∂ξ ∂vh ∂η = JT ∂x ∂x ∂vh ∂y . 1 ∂vh ∂vh ∂vh = − (y2 − y1 ) (y3 − y1 ) ∂x J ∂ξ ∂η ∂vh 1 ∂vh ∂vh = + (x2 − x1 ) −(x3 − x1 ) .3. CALCULATION AND ASSEMBLY OF STIFFNESS MATRIX η6 s (0. from which it follows that the Jacobian is J = J T  = det namely. y) = vh (r(ξ. r2 . 1) @ 3@ @ @ @ @ 2 @s  37 (0. Similarly.2. 0) s1 (1. vh (x. (2. η)) = V1 ψ1 (ξ. we need the gradient of vh in the global coordinate system.15) Consequently. r3 ). η) + V3 ψ3 (ξ. (2.13) where Vi is the value of vh at the node of the triangle K with position vector ri .17) .14) x2 − x1 y2 − y1 x3 − x1 y3 − y1 x1 y1 1 = det x2 y2 1 . η) + V2 ψ2 (ξ. ∂y J ∂ξ ∂η Hence J2 ∇vh 2 = r3 − r1 2 ∂vh ∂ξ 2 (2. In order to determine the entries of the stiﬀness matrix.
. N is precisely the number of unknowns: U1 .38 CHAPTER 2. after a little algebra. K V3 where k ∈ {1. N ∗ the nodes that lie on the boundary of Ω (thus N ∗ is the total number of nodes of which N are internal and N ∗ − N are on the boundary). obtained by cyclic permutations of the indices in these expressions. As uh = 0 on the . respectively. . . . V3 . y) = i=1 Ui φi (x. y)2 dx dy = [V1 .12) it follows that ∂vh = V2 − V1 . ∇vh (x. N + 2. M } is the number of the triangle K in the global numbering and Ak is the symmetric 3 × 3 element stiﬀness matrix: r2 − r3 2 (r2 − r3 ) · (r3 − r1 ) (r2 − r3 ) · (r1 − r2 ) 1 r3 − r1 2 (r3 − r1 ) · (r1 − r2 ) . Ak = 4A123 symm. r1 − r2 2 Assembly of the global stiﬀness matrix entails relating the local numbering of the nodes to the global numbering system. . APPROXIMATION OF ELLIPTIC PROBLEMS and from (2. . y)2 dx dy Ω K K from triangle K is 1 1 J2 ∇vh 2 . y)2 dx dy = A123 ∇vh 2 = J∇vh 2 = 2 4A123 K Substitution of (2. Let us denote by N the number of nodes internal to Ω. .18) into this formula yields a quadratic form in the nodal values V1 . . y)2 dx dy = ∇vh (x. y). V3 V1 . ∇vh is constant on K so the contribution to ∇vh (x. V32 and V2 V3 . V2 . we ﬁnd that the coeﬃcient of V12 is r3 − r1 2 + r2 − r1 2 − 2(r3 − r1 ) · (r2 − r1 ) = r3 − r2 2 and the coeﬃcient of V1 V2 is −2r3 − r1 2 + 2(r3 − r1 ) · (r2 − r1 ) = 2(r2 − r3 ) · (r3 − r1 ) with similar expressions for the coeﬃcients of V22 . .18) As vh (x. ∂ξ ∂vh = V3 − V1 . . . V3 ]Ak V2 .17) and (2. as N uh (x. Thus we deduce that V1 ∇vh (x. Let us label by N + 1. ∂η (2. y) is linear on each triangle K in the triangulation. V2 . UN .14) and (2.
. When programming this. M . j)). . CALCULATION AND ASSEMBLY OF STIFFNESS MATRIX 39 boundary. we loop through all the triangles in the triangulation of Ω. it is more economical to store the information contained in the arrays Lk in a single connectivity array LN ODS which has dimension M ×3. . . Once A∗ and F ∗ have been found. . . LN ODS(k. . 0. in fact. . known (to be zero) from the boundary condition. . N ∗ }. For the kth triangle K. . we can adopt the notational convention that UN +1 = UN +2 = . . . 0)T ∈ RN such that 1 V ∗ T A∗ V ∗ − V ∗ T F ∗ is minimum. j = ij k 1. M . we consider the Boolean matrix1 Lk of size N ∗ × 3 whose entries are deﬁned as follows: if in calculating the matrix Ak the node with position vector r1 is the ith node in the global numbering.19) ∗ ∗ ∗ Since the last N − N entries of V are equal to 0. . The full load vector F ∗ = (F1 .11) can be restated in the following equivalent form: ﬁnd V ∗ = (V1 . and the last N ∗ − N entries of F ∗ to obtain to global load vector F . . similarly. UN )T . . FN ∗ )T is built up in the same way. . LN ODS(k. . . where M is the number of triangles in the triangulation of Ω. we erase the last N ∗ − N rows and columns of A∗ to obtain the global stiﬀness matrix A. 2. In order to justify more clearly the compression of A∗ to A and F ∗ to F . . . and write ∗ N uh (x. By letting k = 1. UN ∗ = 0. once the value Aij has been calculated it is added into the full stiﬀness matrix A∗ at position (LN ODS(k. . . the second and third column depend on the global numbering of the nodes with position vectors r2 and r3 appearing in the matrix Ak .e. . . N ∗ are. . . then the ﬁrst column of Lk has unit entry in the ith row.3. 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. the so called full stiﬀness matrix A∗ is an N ∗ × N ∗ matrix deﬁned as a sum over the elements K in the triangulation of the domain: M A = k=1 k T ∗ Lk Ak (Lk )T . . . 3 from the formula for Ak given above. . 2 (2. j = N + 1. instead of working with M Boolean arrays Lk . and calculate Ak for i. FN . y) = i=1 Ui φi (x. . i ∈ {1. . . . N. . and then solve the linear system AU = F to determine the vector of unknowns U = (U1 . . VN . . .2. i). y). with the understanding that the coeﬃcients Uj . . j) ∈ {1. . Then. . . . N ∗ } is equal to the global number of the node rj in the kth triangle. . where (L ) is the transpose of the matrix Lk . let us note that the minimisation problem (2. a matrix whose entries are 0s and 1s . . k = 1. .
22) The weak formulation of (2. . Let us consider the elliptic boundary value problem − ∂ ∂xj i.j=1 Ω aij ∂u ∂v dx + ∂xi ∂xj n bi (x) i=1 Ω ∂u v dx + ∂xi c(x)uv dx. Ω and l(v) = Ω f (x)v(x) dx.21) 1 where Ω is a bounded open set in Rn . N ∗ } correspond to the interior nodes and the last N ∗ − N to the boundary nodes: indeed. . . aij ∈ L∞ (Ω). we now outline the basic tools for its error analysis. 2. . . leading to a slower assembly process.20) (2.21) is: 1 1 ﬁnd u ∈ H0 (Ω) such that a(u. ξn ) ∈ Rn . ·) and the linear functional l(·) are deﬁned by n a(u. i = 1. It is worth noting that in practice it is not essential that the ﬁrst N indices in the set {1. No such diﬃculties arise when we work with A∗ and F ∗ . j = 1. followed by compression. v) = i. n. (2.20). 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 ∗ . . since in the latter case special care has to be taken for nodes which belong to triangles with at least one boundary point. APPROXIMATION OF ELLIPTIC PROBLEMS by entries of V ∗ that are equal to zero. 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. . ∂xi u = 0 on ∂Ω.j=1 aij (x)ξi ξj ≥ c ˜ ξi2 i=1 ¯ ∀ξ = (ξ1 . f ∈ L2 (Ω). and assume that there exists a positive constant c such that ˜ n n i. (2. . C´a’s lemma e Having described the construction of the ﬁnite element method.40 CHAPTER 2.4 Galerkin orthogonality. x ∈ Ω.23) where the bilinear functional a(·. 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. i. is typically a faster process than the direct assembly of A and F . . . n. the nodes may be numbered in any order. bi ∈ W∞ (Ω). . . ∀x ∈ Ω. . c ∈ L∞ (Ω). (2.j=1 n aij (x) ∂u ∂xi n + i=1 bi (x) ∂u + c(x)u = f (x). . . . (2. . v) = l(v) ∀v ∈ H0 (Ω).
by hypothesis. u − uh ). the weak solution of (2. 1 with v = u − uh ∈ H0 (Ω) we have that u − uh it follows from (2. Moreover. The ﬁnite element approximation of (2. we deduce that u − uh H 1 (Ω) ≤ H 1 (Ω) for all vh ∈ Vh . (2.24) has a unique solution uh in Vh . c0 ≤ 1 a(u − uh . u − vh ) ≤ c1 u − uh c1 u − vh c0 H 1 (Ω) 2 H 1 (Ω) 2 H 1 (Ω) ≤ 1 a(u − uh . GALERKIN ORTHOGONALITY. x ∈ Ω. (2. without making further precise assumptions on the nature of Vh (although we shall implicitly assume that Vh consists of continuous piecewise polynomials deﬁned on a subdivision of “ﬁneness” h of the computational domain Ω).13). Moreover.17).20). (2. vh ) = l(vh ) for all vh ∈ Vh . Vh is contained in H0 (Ω) it follows from the LaxMilgram theorem that (2. Subtracting (2. Combining the last two inequalities.24) ﬁnd uh in Vh such that a(uh . 1 f L2 (Ω) . a(u − uh .23) is: (2. Since by (1.´ 2. 1 As. u H 1 (Ω) ≤ c0 where c0 is as in (1.23) has a unique solution u in H0 (Ω). u − vh ). CEA’S LEMMA We have shown that if 1 c(x) − 2 n 41 i=1 ∂bi ¯ ≥ 0.17). a(u. 1 Now suppose that Vh is a ﬁnitedimensional subspace of H0 (Ω). vh ) = l(vh ) for all vh ∈ Vh .25) is referred to as Galerkin orthogonality and will be seen to play a crucial role in the error analysis of ﬁnite element methods. namely.25) that u − uh further.24) from this identity we deduce that a(u − uh . vh ) = 0 for all vh ∈ Vh . ∂xi 1 then (2. Thus we have proved the following result. .23) holds for any v = vh ∈ Vh .4. c0 u − vh H 1 (Ω) .21). by (1.25) The property (2.
on ∂Ω. bn )T . with bi ∈ W∞ (Ω) for i = 1. min u − vh H 1 (Ω) ≤ C(u)hs vh ∈Vh where C(u) is a positive constant.20). (2. . . 1 where ε > 0.26) is diﬃcult to quantify (given that it depends on the unknown analytical solution u). When advection dominates diﬀusion the socalled P´clet number e n i=1 Pe = bi ε 2 L∞ (Ω) 1/2 . Later on we shall discuss a posteriori error bounds which make explicit use of the computed solution uh and provide computable bounds on the global error. Such problems arise in the mathematical modelling of advectiondiﬀusion phenomena. Consider the following boundary value problem: −ε∆u + b · ∇u = f u = 0 in Ω. h is the meshsize parameter (the maximum diameter of elements in the subdivision of the computational domain) and s is a positive real number.26) can be stated prior to computing uh ). the e weak solution to the problem (2. It shows. . .26) which is a bound of the global error eh = u − uh in terms of the meshsize parameter h. the sequence of ﬁnite element solutions {uh }h converges to u in the H 1 (Ω) norm.21). Such a bound on the global error is called an a priori error bound (the terminology stems from the fact that (2. u − uh H 1 (Ω) ≤ c0 vh ∈Vh Remark 5 We shall prove in the next chapter that. we shall suppose that div b ≤ 0 almost everywhere on Ω. for a typical ﬁnite element space Vh . dependent on the smoothness of u and the degree of piecewise polynomials comprising the space Vh . Suppose that Ω is a bounded open set in Rn . n. that as h → 0 when reﬁning the subdivision further and further. . APPROXIMATION OF ELLIPTIC PROBLEMS 1 Lemma 5 (C´a’s lemma) The ﬁnite element approximation uh to u ∈ H0 (Ω). it is of little practical relevance as the constant C(u) involved in (2. . i. While this result is reassuring from the theoretical point of view. c1 min u − vh H 1 (Ω) .42 CHAPTER 2. For the sake of simplicity. ..26): for certain elliptic problems the ratio c1 /c0 can be very large. Hence. and then the meshsize h has to be taken extremely small before any reduction in the size of the global error is observed. ⋄ Example 6 In this example we highlight a further point concerning the a priori error bound (2. dependent on the smoothness of u. is the nearbest ﬁt to u in the norm · H 1 (Ω) .e. with the aid of C´a’s lemma we shall be able to deduce that e u − uh H 1 (Ω) ≤ C(u) c1 c0 hs (2. b = (b1 . in particular. .
´ 2.4. GALERKIN ORTHOGONALITY; CEA’S LEMMA is very large (say, of the order 106 to 108 ). A simple calculation shows that for the present problem
n 1/2
43
c1 = and
ε2 +
i=1
bi
2 L∞ (Ω)
c0 = Therefore
ε . (1 + c2 )1/2 ⋆
c1 = (1 + c2 )1/2 (1 + P e2 )1/2 , ⋆ c0
and (2.26) gives u − uh
H 1 (Ω) 2 ≤ (1 + c⋆ )1/2 (1 + P e2 )1/2 C(u)hs .
(2.27)
Thus, when ε << 1, the constant on the righthand 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 ε through u (typically C(u) >> 1 when ε << 1). We shall not consider the ﬁnite element approximation of advectiondominated diﬀusion 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 theoretical results of the kind (2.26). As it happens, the poor quality of the a priori error bound (2.27) when P e >> 1 is merely a reﬂection of the fact that for advectiondominated diﬀusion equations conventional ﬁnite 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 meshsize h. ⋄ In order to put this example into perspective, we now discuss the other extreme case, when b ≡ 0 on Ω: then c1 = c0 = ε, so C´a’s lemma implies that e u − uh
H 1 (Ω)
≤ min u − vh
vh ∈Vh
H 1 (Ω) .
In fact, since the lefthand side of this inequality cannot be strictly less than the righthand side (this can be seen by choosing vh = uh on the right), it follows that u − uh
H 1 (Ω)
= min u − vh
vh ∈Vh
H 1 (Ω) ,
so that uh is the best approximation to u from Vh in the H 1 (Ω) norm. We shall show that a result of this kind holds in a slightly more general setting, when the problem is selfadjoint, namely aij (x) ≡ aji (x) for all i, j = 1, . . . , n, bi (x) ≡ 0 for i = 1, . . . , n. Let us deﬁne (v, w)a := a(v, w),
1 v, w ∈ H0 (Ω).
44
CHAPTER 2. APPROXIMATION OF ELLIPTIC PROBLEMS
1 H0 (Ω)
0
u s > 6 u − uh s uh
Vh
Figure 2.9: The error u − uh is orthogonal to Vh .
1 1 Because a(·, ·) is a symmetric bilinear functional on H0 (Ω) × H0 (Ω) and
a(v, v) ≥ c0 v
2 H 1 (Ω)
1 ∀v ∈ H0 (Ω),
it is easily seen that (·, ·)a satisﬁes all axioms of an inner product. Let the associated energy norm deﬁned by: v
a
·
a
denote
:= [a(v, v)]1/2 .
1 Since Vh ⊂ H0 (Ω), taking v = vh ∈ Vh in the statement of (W ), we deduce that
a(u, vh ) = l(vh ), vh ∈ Vh ; also by, (Wh ), a(uh , vh ) = l(vh ), vh ∈ Vh .
(2.28)
(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 − uh , vh ) = 0 ∀vh ∈ Vh , i.e. (u − uh , vh )a = 0 ∀vh ∈ Vh . (2.30)
Thus, in the selfadjoint case, the error u − uh between the exact solution u and its ﬁnite element approximation uh is orthogonal to Vh in the inner product (·, ·)a (see Figure 2.9). By virtue of the orthogonality property (2.30), u − uh
2 a
= = = = =
(u − uh , u − uh )a (u − uh , u)a − (u − uh , uh )a (u − uh , u)a (u − uh , u)a − (u − uh , vh )a (u − uh , u − vh )a ∀vh ∈ Vh .
2.5. OPTIMAL ERROR BOUND IN THE ENERGY NORM Thence, by the CauchySchwarz inequality, u − uh therefore u − uh Consequently, u − uh
a a 2 a
45
= (u − uh , u − vh )a ≤ u − uh a u − vh
a
∀vh ∈ Vh ;
≤ u − vh
a
∀vh ∈ Vh .
= min u − vh a .
vh ∈Vh
Thus we have proved the following reﬁnement of C´a’s lemma in the selfadjoint e case.
1 Lemma 6 The ﬁnite element approximation uh ∈ Vh of u ∈ H0 (Ω) is the best ﬁt to u from Vh in the energy norm · a , i.e.
u − uh
a
= min u − vh a .
vh ∈Vh
C´a’s lemma is the key to the error analysis of the ﬁnite element method for e elliptic boundary value problems. In the next section we describe how such an analysis proceeds in the selfadjoint case, for a particularly simple ﬁnite element space Vh consisting of continuous piecewise linear functions on Ω. 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 C´a’s lemma to derive an optimal error bound for e the ﬁnite element approximation (Wh ) of problem (W ) in the case of piecewise linear basis functions. We shall consider two examples: a onedimensional model problem – a twopoint boundary value problem, and a twodimensional model problem – Poisson’s equation subject to homogeneous Dirichlet boundary condition. Onedimensional problem Consider, for f ∈ L2 (0, 1), the boundary value problem 0 < x < 1, −u′′ + u = f (x), u(0) = 0, u(1) = 0. Its weak formulation is:
1 1 ﬁnd u in H0 (0, 1) such that a(u, v) = l(v) ∀v ∈ H0 (0, 1),
Thus. (2. 1]. N . . . has been described in Section 2. v) = 0 (u′ (x)v ′ (x) + u(x)v(x)) dx 1 and l(v) = 0 f (x)v(x) dx. i=1 . . . xN } which coincides with u at the meshpoints xi . . . We put Vh = span{φ1 . Now since the bilinear functional a(·. we consider a general nonuniform subdivision: 0 = x0 < x1 < . . . instead of restricting ourselves to uniform subdivisions of [0. . N . N −1 Ih u(x) = u(xi )φi (x). 1). 1). w)) 1/2 = 0 w (x) + w(x) ′ 2 2 dx = w H 1 (0. It will be supposed that N ≥ 2 so that we have at least one meshpoint inside (0. i = 0.31) Let Ih u ∈ Vh denote the continuous piecewise linear function on the subdivision {x0 . . . for i = 1.1) . APPROXIMATION OF ELLIPTIC PROBLEMS 1 a(u. Here. < xN −1 < xN = 1. . vh ) = l(vh ) ∀vh ∈ Vh . are not necessarily equally spaced. i = 0. N − 1. . ·) is symmetric it follows from C´a’s lemma e that u − uh H 1 (0. . .1) . x1 . · a The symmetric bilinear functional a(·. . where the meshpoints xi . . φN −1 }. using piecewise linear basis functions. We put hi = xi − xi−1 and deﬁne the mesh parameter h = maxi hi . .1) = u − uh a = min u − vh vh ∈Vh a = min u − vh vh ∈Vh H 1 (0. The ﬁnite element approximation of this problem. 1 Clearly Vh is an (N − 1)dimensional subspace of H0 (0.46 where CHAPTER 2. . We approximate the boundary value problem by the ﬁnite element method ﬁnd uh in Vh such that a(uh .1 (take p(x) ≡ 1 and q(x) ≡ 1 there to obtain the present problem). . . we consider the ﬁnite element basis function 0 if x ≤ xi−1 (x − xi−1 )/hi if xi−1 ≤ x ≤ xi φi (x) = (xi+1 − x)/hi+1 if xi ≤ x ≤ xi+1 0 if xi+1 ≤ x. ·) induces the energy norm 1 1/2 deﬁned by w a = (a(w. . For such a subdivision.
we deduce that the Fourier coeﬃcients of ζ ′ are (kπ/hi )ak . The rest of this subsection is devoted to the proof of the following estimate: u − Ih u H 1 (0.31).1) ≤ ≤ h π h ′′ u π 2 u′′ L2 (0. Choosing vh = Ih u in (2. and deﬁne ζ(x) = u(x) − Ih u(x) for x ∈ [xi−1 . xi ) and ζ(xi−1 ) = ζ(xi ) = 0.32) Thus. OPTIMAL ERROR BOUND IN THE ENERGY NORM 47 The function Ih u is called the interpolant of u from the ﬁnite element space Vh . 4 ak 2 .1) L2 (0. hi [ζ(x)] dx = 2 xi−1 2 xi ak 2 . hi ∞ k=1 x ∈ [xi−1 .1) .1) h ≤ π h2 1+ 2 π 1/2 u′′ L2 (0. (2. we see that u − uh H 1 (0.5. Thus. Then ζ ∈ H 2 (xi−1 .1) . to derive a bound on the global error u − uh in the H 1 (0.1) .2. Because k 4 ≥ k 2 ≥ 1. ∞ ζ(x) = k=1 ak sin kπ(x − xi−1 ) . [ζ ′′ (x)]2 dx. (2.33) Theorem 3 Suppose that u ∈ H 2 (0. 1) norm. then the following error bounds hold: u − Ih u u′ − (Ih u)′ L2 (0. Therefore ζ can be expanded into a convergent Fourier sineseries. 1) and let Ih u be the interpolant of u from the ﬁnite element space Vh deﬁned above. Diﬀerentiating the Fourier sineseries for ζ twice. xi ]. xi ]. it follows that xi xi−1 xi xi−1 [ζ(x)]2 dx ≤ hi π hi π 4 xi xi−1 [ζ ′′ (x)]2 dx. xi ]. while those of ζ ′′ are −(kπ/hi )2 ak .1) ≤ u − Ih u H 1 (0. [ζ ′ (x)]2 dx ≤ 2 xi xi−1 . Proof Consider a subinterval [xi−1 .1) . 1 ≤ i ≤ N . we shall now seek a bound on the interpolation error u − Ih u in the same norm. L2 (0. xi xi−1 xi xi−1 [ζ ′ (x)]2 dx = [ζ ′′ (x)]2 dx = hi 2 hi 2 ∞ k=1 ∞ k=1 kπ hi kπ hi 2 ak 2 . Hence.
As a ﬁnal note concerning this example. 1 0 0 f (x) dx 1/2 1 u(x) dx 1/2 2 . Now (2.33) into the inequality (2. APPROXIMATION OF ELLIPTIC PROBLEMS However ζ ′′ (x) = u′′ (x) − (Ih u)′′ (x) = u′′ (x) for x ∈ (xi−1 .1) 2 ≤ ≤ f ≤ f 0 f (x) dx L2 (0.e. (2. we arrive at the following a priori error bound by inserting (2. from (2. 2 L2 (0. as u′′ = u − f from the diﬀerential equation.1) . converges to 0 at the rate O(h) as h → 0. 1) norm.1) . u′ L2 (0. provided u′′ ∈ L2 (0. 2 L2 (0. . 2 After taking the square root and recalling that ζ = u − (Ih u) these yield the desired bounds on the interpolation error.1) ≤2 f L2 (0. Therefore.1) .35).1) = ≤ h2 π2 u − Ih u 1+ h π2 2 L2 (0. .1) .1) = u−f L2 (0. 1)) implies that u′′ ∈ L2 (0. measured in the H 1 (0.1) + f L2 (0. .1) L2 (0. we obtain ζ ζ′ 2 L2 (0. xi ) because Ih u is a linear function on this interval.32): u − uh H 1 (0. Thereby. i.1) .1) . 1).1) ≤ u L2 (0. choosing v = u in the weak formulation of the boundary value problem gives 1 0 u′ (x)2 dx + 1 1 0 1 u(x)2 dx = 1/2 2 f (x)u(x) dx 0 1 0 1/2 ≤ Hence. the error in the ﬁnite element solution.34) This shows that.1) 2 + (u − Ih u)′ 2 L2 (0. 1).1) 2 L2 (0. . (2.1) . we have that u′′ L2 (0. 2 .33) follows directly from this theorem by noting that u − Ih u 2 H 1 (0. upon summation over i = 1. Indeed.48 CHAPTER 2. . we remark that our hypothesis on f (namely that f ∈ L2 (0. Finally. N and letting h = maxi hi .1) ≤ ≤ h π h π 4 u′′ u′′ 2 L2 (0.33) on the interpolation error.1) ≤ h π 1+ h2 π2 1/2 u′′ L2 (0.35) u(x) dx u L2 (0.1) u′′ Having established the bound (2.
The argument presented in this example is representative of a general ﬁnite element (a priori) error analysis. Finally. 1).1) (or.5. In a nutshell. Unfortunately. since f is a given function and h = maxi hi can be easily calculated for any given subdivision of [0. 1].34).37) (2. Worse still. and consider the elliptic boundary value problem −∆u = f u = 0 in Ω.2. In fact. 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 ﬁnite element approximation. for multidimensional problems (2.34) gives u − uh H 1 (0. (2.36) is a highly nontrivial task (this issue will be touched on in the next section. we obtain a computable bound on the global error.1) into (2. together with e b) an interpolation error bound. These two ingredients then lead us to the error bound (2. 1).1) ). uH 2 (0. 1)). for (multidimensional) elliptic boundary value problems proving an elliptic regularity estimate of the form uH 2 (Ω) ≤ C∗ f L2 (Ω) (2. 1) × (0. OPTIMAL ERROR BOUND IN THE ENERGY NORM 49 Thus we have proved that u′′ ∈ L2 (0.36) holds precise estimates of the size of the constant C∗ are only available in rare circumstances.1) . a posteriori error analysis. 1) norm. 1) (in fact. it consisted of using: a) C´a’s lemma. as we also know that u and u′ belong to L2 (0.36) will not hold unless the boundary ∂Ω and the coeﬃcients aij . This now provides a computable upper bound on the global error u − uh in the H 1 (0. in discussion about the AubinNitsche duality argument). which resolves this diﬃculty and provides computable bounds on the error in terms of uh . bi and c are suﬃciently smooth. then. The upshot is that an a priori error bound will usually not provide a computable estimate of the global error. Later on in the notes we shall discuss an alternative approach. we have proved more: u ∈ H 2 (0.1) ≤ C∗ f L2 (0.1) ≤ C∗ f L2 (0. in other words.1) 2h ≤ π h2 1+ 2 π 1/2 f L2 (0.38) . which is called c) an elliptic regularity estimate. at least in principle. Twodimensional problem Let Ω = (0. if we are fortunate enough to have a bound of the type u′′ L2 (0. even when (2. Substituting this bound on u′′ L2 (0. on ∂Ω.
1]. . N . . y) ∈ 6 otherwise. y1 y0 = 0 x0 = 0  x x1 .10: Triangulation of Ω = [0. Let h = 1/N . h xi −x . . i. h y −y xi −x − jh .10. contained in the interior of Ω (labelled ⊙ in the ﬁgure). . . . we triangulate the domain as shown in the Fig 2.50 CHAPTER 2. y) ∈ 3 (x. h yj −y . h x−xi . j = 1.. . yj ). . yj = jh. . . xN = 1 ¯ Figure 2. We recall that the weak formulation of this problem is: 1 ﬁnd u ∈ H0 (Ω) such that ∂u ∂v ∂u ∂v + dx dy = ∂x ∂x ∂y ∂y Ω Ω 1 f v dx dy ∀v ∈ H0 (Ω). y) ∈ 1 (x. (xi . 1] × [0. . deﬁned by 1− 1− 1− φij (x. . h (x.. . y) ∈ 5 (x. j = 0. With each node. y) ∈ 4 (x. . y) ∈ 2 (x. N . we associate a basis function φij . . and deﬁne xi = ih. y) = 1− 1− 1− 0 y−y x−xi − h j.39) In order to construct the ﬁnite element approximation. i = 0. APPROXIMATION OF ELLIPTIC PROBLEMS y 6 yN = 1 @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @s @s @s i i i @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @s @s @s @ i i i @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @s @s @ @s i i i @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ @ . N − 1. (2. h y−yj .
. . y (xi−1 . (v. . yj+1 ) @ @ @ @ @ @ 2 @ @ 3 @ 1 @@ @ i s @ (x .40) can be written.39) and the ﬁnite element method (2. i = 1. . w) = Ω (2. N − 1}.14′ ) = min u − vh vh ∈Vh a ≤ u − Ih u a . respectively. (2.40) f (x)v(x) dx and ∂v ∂w ∂v ∂w + ∂x ∂x ∂y ∂y dx dy. . as follows: 1 1 ﬁnd u ∈ H0 (Ω) such that a(u. yj−1 ) i+1 j−1 ) Figure 2. j = 1. .2. Let Vh = span{φij . yj+1 ) (xi . . e u − uh a (5. OPTIMAL ERROR BOUND IN THE ENERGY NORM 51 (xi−1 . (5.13′ ) and ﬁnd uh ∈ Vh such that a(uh . yj ).5. (2.39)) is: ﬁnd uh ∈ Vh such that ∂uh ∂vh ∂uh ∂vh + dx dy = ∂x ∂x ∂y ∂y Ω Letting l(v) = Ω Ω f vh dx dy ∀vh ∈ Vh . y ) @ i+1 j @ @ @ 4 @ @ 6 @ @ @ 5 @ @ @ @ @ (x .37) (and (2. N − 1. The ﬁnite element approximation of (2. v) = l(v) ∀v ∈ H0 (Ω). vh ) = l(vh ) ∀vh ∈ Vh . According to C´a’s lemma. yj ) (xi . . w)a = a(v.41) where Ih u denotes the continuous piecewise linear interpolant of the function u on .11: Triangles surrounding the node (xi .
0 ≤ t ≤ 1. 0 ≤ t ≤ 1 − s} and the aﬃne mapping (x. yj )φij (x. Let us estimate u − Ih u a : u − Ih u = △ ∂ ∂ (u − Ih u)2 dx dy +  (u − Ih u)2 dx dy Ω ∂x Ω ∂y ∂ ∂  (u − Ih u)2 dx dy +  (u − Ih u)2 dx dy (2. for example. t) xs xt ys yt = h2 . yl ) = u(xk . by x = xi + sh. that △ = {(x. y) = i=1 j=1 u(xi . t) → (x. ¯ ∂ u ∂s ∂ u ∂t ¯ ¯ 1 ∂u ¯ ∂u = · + · = · . APPROXIMATION OF ELLIPTIC PROBLEMS ¯ the set Ω = [0. △ . t) : 0 ≤ s ≤ 1. yl ). y = yj + th. t) := u(x.52 CHAPTER 2. ∂x ∂s ∂x ∂t ∂x h ∂s ∂u ∂ u ∂s ∂ u ∂t ¯ ¯ 1 ∂u ¯ = · + · = · . t) from △ to K.T. Then. Clearly (Ih u)(xk . Suppose. y). 0 ≤ s ≤ 1. y) → (s.) ∂(x. ∂y ∂s ∂y ∂t ∂y h ∂t The Jacobian of the mapping (s. Let u(s.O. In order to estimate  ∂ (u − Ih u)2 dx dy + ∂x  ∂ (u − Ih u)2 dx dy. y) : xi ≤ x ≤ xi+1 .42) △ ∂x △ ∂y 2 a =  where △ is a triangle in the subdivision of Ω. ∂y △ △ we deﬁne the canonical triangle K = {(s. yj ≤ y ≤ yj+1 + xi − x}. 1] × [0. y) is J = Thus  ∂ (u − Ih u)2 dx dy = ∂x (P. y) = ∂(s. y). 1]: N −1 N −1 (Ih u)(x.
t) − [(1 − s − t)¯(0. 0) + s¯(1. t)2 dθ dσ ds dt 2 ∂s 0 0 0 0 1 1 1−s 1 ∂ 2u ¯  (σ. u − Ih u 2 a ≤ 4h2 Ω  ∂ 2u 2 ∂ 2u ∂ 2u 2  +  2 2  + ∂x2 ∂x ∂y ∂y dx dy.41) and (2. (2.45).43) and (2. t) − (σ. 0) − u(0. ∂  (u − Ih u)2 dx dy ≤ 2h2 △ ∂x Similarly. y)2 · h2 2 · h−2 dx dy ∂x yj xi xi+1 yj+1 yj + xi  ∂ 2u (x.45) Finally by (2.44) Substituting (2. (2. u − uh a ≤ 2huH 2 (Ω) . η) dη dσ2 ds dt 2 σ ∂s 0 0 0 ∂s ∂t 0 0 1 1 1−s 1 ∂ 2u ¯  2 (θ. 0) + t¯(0. t) − [¯(1. y)2 · h2 2 · h−2 dx dy.42). t) dθ dt + ∂s ∂s ∂t 0 0 0 0 yj+1 xi+1 2 ∂ u 2  2 (x.46) Thus we have proved the following result. η)2 dη dσ ds dt +2 ∂s ∂t 0 0 0 0 1 1 1 1 ∂ 2u ¯ ∂ 2u ¯ 2 (σ. (2. . (2. 0) dσ2 ds dt ∂s ∂s ∂s ∂s 0 0 0 0 s 2 1 1−s t 1 1 ∂ u ¯ ∂ 2u ¯  (θ.  ∂ (u − Ih u)2 dx dy ≤ 2h2 ∂y xi+1 xi yj+1 yj xi+1 xi yj+1 yj ∂ 2u 2 1 ∂ 2u 2   2 +  ∂x 2 ∂x ∂y dx dy. η)2 dσ dη 2   2 (θ. 0)]2 ds dt ∂s 0 0 1−s 1 1 ∂u ¯ ∂u ¯  (s. t) dσ + (σ.5. t) − (σ. 1)]) 2 ds dt u u u u K ∂s 1−s 1 ∂u ¯ u ¯  (s. ∂x ∂y Therefore.43) △  ∂ 2u 2 1 ∂ 2u 2   +  ∂y 2 2 ∂x ∂y dx dy. t) − (σ.44) into (2.2. 0) dσ2 ds dt ∂s ∂s 0 0 0 1 1−s 1 1 ∂u ¯ ∂u ¯ ∂u ¯ ∂u ¯  (s. OPTIMAL ERROR BOUND IN THE ENERGY NORM = = = = = ≤  53 ≤ = ∂ (¯(s. t) dθ dσ + (σ.
APPROXIMATION OF ELLIPTIC PROBLEMS Theorem 4 Let u be the weak solution of the boundary value problem (2. (2. √ u − uh H 1 (Ω) ≤ 5huH 2 (Ω) . 1 Let us ﬁrst observe that if w ∈ H 2 (Ω) ∩ H0 (Ω). uh ∈ Vh ⊂ H0 (Ω).48) indicates that the error in the L2 norm between u and its ﬁnite element approximation uh is of the size O(h).48) The AubinNitsche duality argument. the proof of this is presented below.47) = ≤ 5 u − uh 2 1 (Ω) ≤ 5h2 u2 2 (Ω) . then u − uh a ≤ 2huH 2 (Ω) . 1).46) we also see that u − uh L2 (Ω) ≤ h uH 2 (Ω) . H 4 (2. Ω = (0. u − uh 2 a = u − uh 2 1 (Ω) ≤ 4h2 u2 2 (Ω) . H H 1 1 1 e Since u ∈ H0 (Ω).47) and (2. By the Poincar´Friedrichs inequality. ∂ 2w ∂ 2w · dx dy = ∂x2 ∂y 2 = ∂ 2w ∂ 2w · dx dy Ω ∂x ∂y ∂x ∂y ∂ 2w 2  dx dy. Proof According to Theorem 4.40). that this bound is quite pessimistic and can be improved to O(h2 ). and let uh be its piecewise linear ﬁnite element approximation deﬁned by (2. however. u − uh 2 H 1 (Ω) 2 L2 (Ω) 1 ≤ u − uh 2 1 (Ω) . It turns out.  Ω ∂x ∂y Ω . then ∆w 2 L2 (Ω) = Ω ∂ 2w ∂ 2w + ∂x2 ∂y 2 ∂ 2w ∂x2 2 2 dx dy ∂ 2w ∂ 2w · dx dy + ∂x2 ∂y 2 ∂ 2w ∂y 2 2 = Ω +2 Ω dx dy. 1) × (0. From (2. Ω Performing integration by parts and using the fact that w = 0 on ∂Ω.54 CHAPTER 2. u − uh thus. The error estimate (2. H H 4 u − uh 2 L2 (Ω) 2 + u − uh H 1 (Ω) and that completes the proof. Suppose 1 that u ∈ H 2 (Ω) ∩ H0 (Ω). it follows that u − uh ∈ H0 (Ω).37). Corollary 2 Under the hypotheses of Theorem 4.
54) and the error bound (2.5.51) After this brief preparation. g L2 (Ω) g∈L2 (Ω) (2.52) 1 For g ∈ L2 (Ω). g) ≤ u − uh Therefore. = sup (u − uh . u − uh L2 (Ω) L2 (Ω) g L2 (Ω) ∀g ∈ L2 (Ω). so it satisﬁes a(wg .2.46). . According to the CauchySchwarz inequality for the L2 inner product (·. ·). (2. v) = Ω Consider the ﬁnite element approximation of (2.50) wg H 2 (Ω) = ∆wg L2 (Ω) = g L2 (Ω) . (2. (2. the function wg ∈ H0 (Ω) is the weak solution of the problem (2. v) = lg (v) where lg (v) = Ω 1 ∀v ∈ H0 (Ω). let wg ∈ H0 (Ω) be the weak solution of the boundary value problem −∆wg = g wg = 0 1 then wg ∈ H 2 (Ω) ∩ H0 (Ω). g) . and in Ω. =  dx dy 1 Given g ∈ L2 (Ω).53) gv dx dy = (g.49) (2. vh ) = lg (vh ) ∀vh ∈ Vh .49). (2. OPTIMAL ERROR BOUND IN THE ENERGY NORM Thus ∆w 2 L2 (Ω) 55 ∂ 2w 2 ∂ 2w ∂ 2w 2  + 2  +  2 2 ∂x2 ∂x ∂y ∂y Ω 2 = wH 2 (Ω) .53): ﬁnd wgh ∈ Vh such that a(wgh . we turn to the derivation of the optimal error bound in the L2 norm.54) From (2.53). we deduce that wg − wgh a ≤ 2hwg H 2 (Ω) . (u − uh . a(wg . on ∂Ω. (2. v). ∂wg ∂v ∂wg ∂v + ∂x ∂x ∂y ∂y dx dy.
wg − wgh Now (u − uh . g) ≤ 4h2 uH 2 (Ω) · g L2 (Ω) . by (2. wgh ) = 0. wg − wgh )a . wg − wgh ) = (u − uh . (u − uh .56) wg − wgh a . Applying the CauchySchwarz inequality on the right. (u − uh .51). a a ≤ 2h g L2 (Ω) .30) implies that a(u − uh . we obtain u − uh L2 (Ω) ≤ 4h2 uH 2 (Ω) . u − uh ) = lg (u − uh ) = a(wg . 2. g) ≤ u − uh and thence by (2.55) (u − uh . by (2. g) = (g. it was observed by engineers . (2. u − uh ) = a(u − uh . where C denotes a generic positive constant and h is the maximum element size in the subdivision. wg ).57) Substituting (2.52). However. It is possible to show that these error bounds are sharp in the sense they cannot in general be improved.6 Superapproximation in meshdependent norms We have shown that the piecewise linear ﬁnite element approximation uh to the solution u of the homogeneous Dirichlet boundary value problem for Poisson’s equation obeys the following error bounds: u − uh u − uh H 1 (Ω) L2 (Ω) ≤ ChuH 2 (Ω) . (2. g) = a(u − uh .56). APPROXIMATION OF ELLIPTIC PROBLEMS and therefore.56 CHAPTER 2.55) (2. wgh ) = a(u − uh . (2. Because wgh ∈ Vh .57) into the righthand side of (2. which is our improved error bound in the L2 norm. ≤ Ch2 uH 2 (Ω) .46) and (2. and therefore. The proof presented above is called the Aubin–Nitsche duality argument. wg ) − a(u − uh .
y) ∈ ∂Ω. N − 1. and h = 1/N . N − 1}. y) dx dy. where the Uij (= uh (xi . j = 1. Indeed. yj ) : i. the diﬀerence scheme can be written as follows: + − + − −(Dx Dx Uij + Dy Dy Uij ) = Fij .6. yj ) ∈ Ωh . We showed in Section 2. . we shall write Fij = 1 h2 supp φij f (x. . . . y) as N −1 N −1 uh (x. j = 0. yj = jh. SUPERAPPROXIMATION IN MESHDEPENDENT NORMS 57 that.1 that when using continuous piecewise linear ﬁnite elements on the uniform triangulation shown in Figure 2. y)φij (x. Since uh (x.59) where Ω = (0. . the meshpoints are (xi . 1) × (0. A result of this kind is usually referred to as a superapproximation property. N }...2. yj ). . We consider the set of interior mesh points Ωh = {(xi . U = 0 (xi . the error u − uh is O(h2 ). These form the ﬁnite diﬀerence mesh ¯ Ωh = {(xi . the ﬁnite element solution uh (x. . (2. N. i. we have adopted the convention that Uij = 0 when i = 0 or i = N or j = 0 or j = N . We consider the model problem −∆u = f u = 0 in Ω. yj ) : i. . .j − 2Uij + Ui−1.60) (2.j+1 − 2Uij + Ui. ¯ and the set of boundary mesh points Γh = Ωh \ Ωh . y) dx dy. y) = 0 when (x. on Γh .5. Here N is an integer. In more compact notation. j = 0. when sampled at certain special points. = 2 h supp φij Uij = 0 when i = 0 or i = N or j = 0 or j = N . . j = 1. y)φij (x. .61) . 1). .j−1 − h2 h2 1 f (x. N − 1. y). . where xi = ih. (2. . for i. . yj )) are obtained by solving the set of diﬀerence equations − Ui+1. y) = i=1 j=1 Uij φij (x. j = 1. the ﬁnite element approximation uh is more accurate than these error bounds might indicate. based on sampling at the mesh points.58) (2.j Ui. y) can be expressed in terms of the ﬁnite element basis functions φij (x. . N ≥ 2. . For simplicity. we shall prove here that when measured in a discrete counterpart of the Sobolev H 1 (Ω) norm.. on ∂Ω. i.
j ≤ N − 1.N −1 . . . Ui1 . U1. .N −1 . Ui. UN −1. . .j . . U2. UN −1. + − where Dx and Dx denote the forward and backward divided diﬀerence operators in the x direction. . and a typical 5point diﬀerence stencil. the ﬁnite diﬀerence equation (2. Similar deﬁnitions apply in the y direction.j h2 is the second divided diﬀerence operator in the x direction. U21 .j . UN −1. . . . .58 CHAPTER 2. Ui.j . . respectively. .j) (i. . . APPROXIMATION OF ELLIPTIC PROBLEMS × × × × × × × × × × s s s s s s s × s s s s u s s s × s s s u × s s s s u s s s × s s s s s s s × s s s s s s s × s s s s s s s × × × × × × × × × (i. .j) (i.60) involves ﬁve values of U : Ui.j−1 . deﬁned by + Dx Vij = Vi+1.j) s u s u s s (i+1. the boundary mesh Γh (×). .j+1 .2 .1 . . Ui2 . . where U = (U11 . h Vi+1. . For each i and j. . It is possible to write (2. .N −1 )T . 1 ≤ i. Ui−1.j − Vij . . . U22 . .j .62) . (2.j − 2Vij + Vi−1. . h − Dx Vij = Vij − Vi−1.j−1) × × × × × × × Figure 2. Ui. . Ui+1.12: The mesh Ωh (·). .j+1) (i−1.60) as a system of linear equations + − + − Dx Dx Vij = Dx (Dx Vij ) = and AU = F.N −1 . U12 .
. The argument that follows is an alternative way of verifying uniqueness. . F12 . 2 . . . . 2. FN −1. . SUPERAPPROXIMATION IN MESHDEPENDENT NORMS 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 j @ s s @ @ @ @ s@@ s @s s s @ @ @ @ s s @ s @s s 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 59 A= Figure 2. F1.N −1 )T . Fi1 .62) is a trivial consequence of the uniqueness of solution uh to the ﬁnite element method. . . we present it here since some of its ingredients will be exploited in the course of the proof of the superapproximation property. . (2. and A is an (N − 1)2 × (N − 1)2 sparse matrix of band structure.) (which resembles the L2 inner product (v.1 . while the sparsity structure of A is depicted in Fig.12.2 . . F22 . FN −1. corresponding to the ﬁve values of U in the ﬁnite diﬀerence stencil shown in Fig. V )h N N −1 N −1 N = i=1 j=1 h 2 − Dx Vij 2 + i=1 j=1 − h2 Dy Vij 2 . .6. V )h + (−Dy Dy V. . For two functions. .2. . then + − + − (−Dx Dx V. . F = (F11 . W )h = i=1 j=1 h2 Vij Wij v(x.N −1 . . we introduce the inner product N −1 N −1 (V. A typical row of the matrix contains ﬁve nonzero entries.13: The sparsity structure of the band matrix A. . F21 . . . V and W .13. . . .N −1 . . . deﬁned on Ωh . Next we show that (2. F2.N −1 . FN −1.63) The uniqueness of the solution to the linear system (2. .62) has a unique solution2 . Fi2 . 2. Fi. y)w(x. w) = Ω ¯ Lemma 7 Suppose that V is a function deﬁned on Ωh and that V = 0 on Γh . . y) dx dy. . .
we note that by (2. − Dy Vij = Since V = 0 on Γh .64) yields: − Dx Vij = Vij − Vi−1. (Vij − Vi−1. . .j + i=1 j=1 N N −1 i=1 j=1 (Vij − Vi−1. In summary then.60) may be found by solving the system of linear equations (2. Now this implies that A is a nonsingular matrix. V )h = (−Dx Dx V − Dy Dy V.j )Vi−1. N − 1. V )h = i=1 j=1 − h2 Dy Vij 2 . .60 CHAPTER 2.j )Vij N −1 N −1 =− =− =− = i=1 j=1 N N −1 i=2 j=1 N N −1 i=1 j=1 (Vi+1. N. APPROXIMATION OF ELLIPTIC PROBLEMS Proof We shall prove that the ﬁrst term on the left is equal to the ﬁrst term on the right. then (2. .j )Vij N −1 N −1 (Vij − Vi−1.j )Vi−1. . . .j − Vij )Vij + i=1 j=1 (Vij − Vi−1. N. .60).j )2 = Similarly.j + (Vij − Vi−1.63) we have + − + − (AV.j−1 = 0. and U = A−1 F is the unique solution of (2. h Vij − Vi. N −1 N −1 + − (−Dx Dx V. the (unique) solution of the ﬁnite diﬀerence scheme (2.j = 0. Hence A is nonsingular. . i = 1. these imply that V ≡ 0. Thus AV = 0 if and only if V = 0. j = 1. . V )h = − N −1 N −1 i=1 j=1 (Vi+1. j = 1. Returning to the analysis of the ﬁnite diﬀerence scheme (2. .60). h i = 1. . . N −1 N + − (−Dy Dy V. . and that completes the proof. (2. N − 1.64) ¯ for any V deﬁned on Ωh such that V = 0 on Γh . . V )h + − + − = (−Dx Dx V. V )h + (−Dy Dy V. Indeed if AV = 0.j − 2Vij + Vi−1.j )Vij N N −1 i=1 j=1 N N −1 i=1 j=1 − h2 Dx Vij 2 . .j )Vij (Vij − Vi−1. V )h N N −1 N −1 N = i=1 j=1 h 2 − Dx Vij 2 + i=1 j=1 − h2 Dy Vij 2 .62). . and the second term on the left to the second term on the right.
. independent of V and h. · 1/2 2 L2 (Ω) H 1 (Ω) . V )h ≥ Dx V ]2 + Dy V ]2 . Lemma 8 (Discrete Poincar´Friedrichs inequality) Let V be a function deﬁned on e ¯ Ωh and such that V = 0 on Γh . ≤ c∗ − − Dx V ]2 + Dy V ]2 x y (2. x y (2. The norm · 1.60).h is the discrete version of the Sobolev norm H 1 (Ω) u = u 2 L2 (Ω) ∂u + ∂x 2 L2 (Ω) ∂u + ∂y . V )h ≥ c0 V where c0 is a positive constant. With this new notation. U )h .2. we introduce the mesh–dependent norms U and U where N N −1 − Dx U ]x 1/2 1. the inequality (2. = i=1 j=1 h 2 − Dx Uij 2 and N −1 N − Dy U ]y = i=1 j=1 − h2 Dy Uij 2 1/2 .h h = (U. Proof Writing i 2 − hDx Vkj k=1 2 h 2 1.6. SUPERAPPROXIMATION IN MESHDEPENDENT NORMS 61 In order to prove the stability of the ﬁnite diﬀerence scheme (2. such that V for all such V .h .64) takes the following form: − − (AV. then there exists a constant c∗ .65) Using the discrete Poincar´Friedrichs inequality stated in the next lemma.66) Vij  = 2 . we shall e be able to deduce that (AV. 1/2 =( U 2 h − 2 − 2 + Dx U ]x + Dy U ]y )1/2 .
we turn to the question of accuracy. Having established stability.62 we deduce that CHAPTER 2. Finally. V )h ≤ F h V h ≤ F h V 1. 2 x 1 − 2 D V ]y . N − 1. . N − 1.67) ≤ 1 F c0 h. V )h ≥ 1 V c∗ 2 h. . . on noting that N −1 1 (N − 1)N ≤ . V 2 h 2 h ≤ ≤ 1 − 2 D V ]x . writing eij = eh (xi . . we have adopted the convention that Uij = 0 when i = 0 or i = N or j = 0 or j = N .68) Proof The proof is simple: it follows from (2. . (2.65) and recalling the deﬁnition of the norm we obtain (AV. .67) and the CauchySchwarz inequality that c0 V 2 1. yj ). Since uh (x.66) with c∗ = 1 . y) and note that uh (xi . j = 1. (2. y) ∈ ∂Ω. . . with eij = 0 when i = 0 or i = N or j = 0 or j = N . j ≤ N. . N − 1 and j = 1. . 4 Now (2. y) = u(x.60) is stable in the sense that U 1. Thus.h . . and hence the result. combining this with (2. h2 i = h2 2 2 i=1 we deduce that V Analogously. 0 ≤ i. yj ) − Uij . V )h ≥ c0 V where c0 = (1 + c∗ )−1 . we have that eij = u(xi . Multiplying both sides by h2 and summing through i = 1. y) = 0 when (x.h ≤ (AV. . We deﬁne the global error eh by eh (x. . y) − uh (x.h 2 1.66) imply that (AV. yj ) = Uij for i. 2 y Adding these two inequalities we complete the proof of (2. V )h = (F. APPROXIMATION OF ELLIPTIC PROBLEMS i i N − hDx Vkj 2 Vij 2 ≤ h k=1 k=1 ≤i k=1 − h2 Dx Vkj 2 .h .65) and (2.h . Theorem 5 The scheme (2. · 1.
y) about (xi . yj ) − Fij 1 φij (x. j ≤ N − 1. we deduce that ϕij  ≤ K0 h2 ∂ 4u ∂x4 ¯ C(Ω) + ∂ 4u ∂y 4 ¯ C(Ω) .71) Proof Recall that c0 = (1 + c∗ )−1 . Thus. ¯ C(Ω) + ∂ 4u ∂y 4 ¯ C(Ω) 5 4.70) Finally (2. c∗ = (2.69) and According to this result. SUPERAPPROXIMATION IN MESHDEPENDENT NORMS Now.68). y) dx dy − Dx Dx u(xi . (2. the piecewise linear ﬁnite element approximation of the homogeneous Dirichlet boundary value problem on uniform triangular subdivision of size h is O(h2 ) convergent to the weak solution in the discrete Sobolev H 1 norm. y) dx dy = Au(xi .h 5 ≤ K0 h2 4 ∂ 4u ∂x4 1 4. then u − uh 1. Aeij = Au(xi . .69) ¯ Assuming that u ∈ C 4 (Ω) and employing a Taylor series expansion of u(x. Thus. Theorem 6 Let f ∈ L2 (Ω) and suppose that the corresponding weak solution u ∈ 1 ¯ H0 (Ω) belongs to C 4 (Ω). yj ) 2 ∂y 1 ≤ i. yj ) − 2 h supp φij = 1 h2 + ≡ ϕij . yj ). on Γh . (2.h ≤ 1 ϕ h.70). y)f (x.70) yield the following result.2. u−U 1. e = 0 By virtue of (2. so that 1/c0 = and combine (2. y) 2 (x. ϕ h ≤ K0 h2 ∂ 4u ∂x4 ¯ C(Ω) + ∂ 4u ∂y 4 ¯ C(Ω) . y) dx dy − Dy Dy u(xi . y) ∂ 2u + − (x. . yj ) − AUij = Au(xi . = e 1.69) and (2. yj ) ∂x supp φij 1 h2 supp φij φij (x. c0 (2.h 63 ∂ 2u + − φij (x. Aeij = ϕij .6. where K0 is a positive constant independent of h.
Since this exceeds the ﬁrst order of convergence of the global error observed in the Sobolev norm · H 1 (Ω) .h . In fact the smoothness 1 1 ¯ requirement u ∈ H0 (Ω)∩C 4 (Ω) can be relaxed to u ∈ H0 (Ω)∩H 3 (Ω) while retaining the superapproximation property u − uh 1.h = O(h2 ). provided that u ∈ C 4 (Ω). APPROXIMATION OF ELLIPTIC PROBLEMS ¯ · 1. . the result encapsulated in Theorem 6 is referred to as a superapproximation property. the proof of this is more technical and relies on the BrambleHilbert lemma (See Chapter 3).64 CHAPTER 2.
here we shall put this construction into a general context. 3.1 The ﬁnite element We begin by giving a formal deﬁnition of a ﬁnite element.1. V = H 1 (Ω).1 Construction of ﬁnite element spaces ﬁnd u in V such that a(u. The purpose of this chapter is to develop. Robin or oblique derivative boundary value problem. in the case of a Neumann. 3. Let us consider an elliptic boundary value problem written in its weak formulation: 1 where H0 (Ω) ⊂ V ⊂ H 1 (Ω). We have already discussed the special case when Vh consists of continuous piecewise linear functions. v) = l(v) ∀v ∈ V . in a more general setting. the construction of ﬁnite element spaces and to formalise the concepts introduced in Chapter 2. In order to deﬁne a ﬁnite element approximation to this problem we need to construct a ﬁnitedimensional subspace Vh of V consisting of continuous piecewise polynomial functions of a certain degree deﬁned on a subdivision of the computational domain Ω.Chapter 3 Piecewise polynomial approximation In the previous chapter we discussed ﬁnite element approximations to elliptic boundary value problems using piecewise polynomials of degree 1. in the case of a homogeneous Dirichlet boundary value 1 problem V = H0 (Ω). Deﬁnition 2 Let us suppose that (i) K ⊂ Rn is a simply connected bounded open set with piecewise smooth boundary (the element domain). 65 .
dual to N . N2 . . Nk } is a subset of the dual space P ′ . for each i = 1. 1 ≤ i.. . . . Such a basis is called a nodal basis for P. . . we can write v = β1 ψ1 + . .j=1. α2 )T . . ψk } be a basis for P. P. Nk } is a basis for P ′ if and only if any L ∈ P ′ can written in a unique fashion as a linear combination of the Ni ’s: L = α1 N1 + . Then (K.. . . . Nk } is a basis for P ′ . (iii) N = {N1 . Then the last condition is equivalent to requiring that the system of linear equations Ba = y has a unique solution. + αk Nk (ψi ). This is equivalent to demanding that. . . N2 . . N ) is a ﬁnite element. . . Then (K. . . N2 }. L(ψi ) can be written in a unique fashion as a linear combination L(ψi ) = α1 N1 (ψi ) + . + αk Nk . Given any v ∈ P. Let us deﬁne the matrix B = (Nj (ψi ))i. . ψk } be a basis for P. . P the set of linear polynomials. P. . . N ) is called a ﬁnite element. and N = {N1 . . is equivalent to demanding that the matrix B be invertible. then v ≡ 0. Deﬁnition 3 Let (K. . . ⋄ Next we give an equivalent characterisation of condition (iii) in Deﬁnition 2. . with nodal basis {ψ1 . L(ψk ))T . Ni (ψj ) = δij . . ψ2 } where ψ1 (x) = 1 − x and ψ2 (x) = x. . j ≤ k. which.66 CHAPTER 3. . k. . where N1 (v) = v(0) and N2 (v) = v(1) for all v ∈ P. P. . Example 7 (The onedimensional Lagrange element) Let K = (0. N ) be a ﬁnite element. . Lemma 9 Let P be a kdimensional linear space of functions on Rn .. . . In this deﬁnition P ′ denotes the algebraic dual of the linear space P. ψ2 . k. . 1). We give a simple example to illustrate these deﬁnitions. . in turn. . . (b) Given that v ∈ P and Ni (v) = 0 for i = 1. Nk } is a basis for P ′ (the set of nodal variables). . a = (α1 . Proof Let {ψ1 . and let {ψ1 . . Then the following two statements are equivalent: (a) {N1 . . PIECEWISE POLYNOMIAL APPROXIMATION (ii) P is a ﬁnitedimensional space of functions deﬁned on K (the space of shape functions).. . namely. . Now {N1 . . . + βk ψk . . .k and the vectors y = (L(ψ1 ). and suppose that {N1 .
then. . Deﬁnition 4 We say that N determines P if ψ ∈ P with N (ψ) = 0 for all N ∈ N implies that ψ = 0. ...1) Thus (b) is equivalent to requiring that (3. . 3. .j=1. 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 nondegenerate linear function. Since P is of degree d. then (b) holds if and only if Cb = 0. . Motivated by this result. + βk Ni (ψk ) = 0.1) implies β1 = . . The dimension of the linear space Pk is displayed in Table 3. in−1 ) and xα = xi1 · · · xn−1 . CONSTRUCTION OF FINITE ELEMENT SPACES Now Ni (v) = 0 for all i = 1. = βk = 0. . Then we can write P in the factorised form P = LQ where Q is a polynomial of degree (d−1). by hypothesis. . Suppose that we have carried out an aﬃne change ˆ of variables such that L(ˆ. . . xn−1 ).1. However C t = B and therefore (a) and (b) are equivalent. Letting xn = 0 we get ˆ 1 0 ≡ P (ˆ. the hyperplane L(ˆ. . We shall need the following Lemma. βk )t . xn ) = x j=0 α≤d−j i cαj xα xn . Let C = (Ni (ψj ))i. . xn ) = 0 is the hyperplane x x xn = 0. . .2 Examples of triangular ﬁnite elements Let K be a triangle and let Pk denote the set of all polynomials of degree ≤ k in two variables.1. ˆ j n−1 where α = (i1 .. Hence d cα0 xα . 0) = x which implies that cα0 = 0 for α ≤ d. .3.1. . . so. k if and only if β1 Ni (ψ1 ) + . with b = (β1 . .2. . . i = 1. implies that b = 0. . 67 (3. xn ) = xn .. . we introduce the following deﬁnition. ˆ α≤d d P (ˆ. k. . xn ) = x j=1 α≤d−j cαj xα xn = xn ˆ j j=1 α≤d−j cαj xα xn = xn Q = LQ ˆ j−1 where Q is of degree (d − 1). 0) ≡ 0. Proof Let us write x = (x1 . which is equivalent to demanding that C be invertible. . we have that x d P (ˆ.k . . P (ˆ.
. L3 and vertices z1 . k 1 2 3 . 3 v(at the midpoint of the (i − 3)rd edge). 1 (k + 1)(k + 2) 2 Table 3. Lagrange elements Example 8 Let k = 1 and take P = P1 .1. L2 . .68 CHAPTER 3. it follows that P ≡ 0 on L1 .. Indeed. z2 . . PIECEWISE POLYNOMIAL APPROXIMATION u3 C C C C C L1 L2 C C C C C L3 CCuz u 2 z z1 Figure 3. 0 = P (z1 ) = cL1 (z1 ) implies that c = 0. i = 4. k dim Pk 3 6 10 . according to Lemma 9. We verify condition (iii) of Deﬁnition 2 using part (b) of Lemma 9. ⋄ Example 9 Now take k = 2. as shown in Figure 3. i = 1. where Ni (v) = v(zi ) and z1 . we can write P = cL1 where c is a constant (i. namely.. However.. z2 . let L1 . 6.2: The dimension of the linear space Pk . N1 determines P1 . z2 and z3 . 5. N2 . z3 are the vertices of the triangle K. Hence. . z3 . In the ﬁgure • indicates function evaluation at the point where the dot is placed. where Ni (v) = v(at the ith vertex of the triangle). Since P L1 is a linear function of one variable that vanishes at two points. . 2. a polynomial of degree 1 − 1 = 0).1. N3 } (and therefore the dimension of P1 is 3).e. because L1 (z1 ) = 0. we prove that N1 determines P1 . N6 } (so we have that dim P2 = 6). thus P ≡ 0. . By virtue of Lemma 10. N2 .1: Linear Lagrange triangle with edges L1 . N = N1 = {N1 . let P = P2 and N = N2 = {N1 . L2 and L3 be nontrivial linear functions which deﬁne the lines that contain the three edges of the triangle. Suppose that a polynomial P ∈ P1 vanishes at z1 .
and z4 . . and let P = P3 . z3 . The restriction of P to L1 is a cubic polynomial of one variable with double roots at z2 and z3 . on L2 . where the degree of Q2 is one less than the degree of Q1 . However. . . Let • denote evaluation at a point and let signify evaluation of the gradient at the centre point of the circle.3. Let L1 . where c is a constant. Therefore. L1 . 6. ⋄ Hermite elements Example 10 Let us suppose that k = 3. where c is a constant. However. determines P3 (whose dimension is precisely 10). P ≡ 0 along the edges L2 and L3 . Let P ∈ P2 be such that P (zi ) = 0 for i = 1. P = cL1 L2 . L2 . By Lemma 10. Consequently. Thus. However. L2 and L3 . Hence. 10. so Q2 is of degree 0. vertices z1 . 2 . except possibly at one point. We shall prove that N = N3 = {N1 . so Q1 is of degree 1. This ﬁnally implies that P ≡ 0. . Similarly. either L1 ≡ 0 or Q1 ≡ 0. c = 0 since L1 (z6 ) = 0 and L2 (z6 ) = 0.3. by an analogous argument P also vanishes along L2 . Thus Q1 ≡ 0 on L2 . L2 and L3 be nontrivial linear functions which deﬁne the lines containing the edges of the triangle. But L1 can be equal to zero only at one point of L2 because the triangle is nondegenerate. Q2 ≡ c. Thus. . We have to show that N2 determines P2 . as before. we deduce that Q1 = L2 Q2 . . . Therefore. where the degree of Q1 is one less than the degree of P . As P L1 is a quadratic function of one variable that vanishes at three points. by continuity of Q1 . z5 and z6 denoting the midpoints of L1 . Let. as depicted in Figure 3. N2 . we can write P = cL1 L2 L3 . Hence P ≡ 0 along L1 . . However P (z6 ) = 0 and z6 does not lie on either L1 or L2 . CONSTRUCTION OF FINITE ELEMENT SPACES u3 C C C L2 C C z4 z5 u Cu C L1 C C C L3 CCuz u u 2 69 z z6 Figure 3. P = L1 Q1 . 0 = P (z6 ) = cL1 (z6 )L2 (z6 ). z2 . so N2 determines P2 .2: Quadratic Lagrange triangle with edges L1 . By Lemma 10. Now applying again Lemma 10. L2 and L3 be the lines corresponding to the three sides of the triangle and suppose that P ∈ P3 and Ni (P ) = 0 for the i = 1. . L3 . we then have that Q1 ≡ 0 along the whole of L2 . L1 Q1 L2 ≡ 0. . . respectively. it follows that P ≡ 0 on L1 .1. 0 = P (z4 ) = cL1 (z4 )L2 (z4 )L3 (z4 ). N10 }. z1 .
4. y)/L3 (z3 ) = y. PIECEWISE POLYNOMIAL APPROXIMATION z3 u C C C C C L1 L2 C z4 C u C C C L3 CCu z u 2 z1 Figure 3. k}. . 3. and centroid z3 . This we do. Also. we now wish to piece them together to construct ﬁnitedimensional subspaces of Sobolev spaces. let P = P1 .70 CHAPTER 3. it follows that ψ1 = cL1 = c(1 − x − y). and so c = 0. Similarly. By deﬁnition. Thus P ≡ 0 and we deduce that N uniquely determines P3 . as ψ1 vanishes at z2 and z3 . z2 and z3 .1. N3 } as in the case of the linear Lagrange element (k = 1). The line L1 has equation y = 1 − x.3 The interpolant Having described a number of ﬁnite elements. y)/L2 (z2 ) = x and ψ3 = L3 (x. . we have that IK (v) = N1 (v)(1 − x − y) + N2 (v)x + N3 (v)y. N ) be a ﬁnite element and let the set {ψi : i = 1. . k. i = 1. we introduce the local interpolant IK v by k IK v = Ni (v)ψi . y) = (1 + x2 + y 2 )−1 . are deﬁned. L3 and vertices z1 . so c = ψ1 (z1 ) = 1. . Having found ψ1 . N2 . ψ1 = 1 − x − y. using Deﬁnition 3. to be able to write down the local interpolant. Hence. 3. be a basis for P dual to N . ψ2 = L2 (x. i = 1. Thus we must determine ψi . P. where c is a constant to be determined. and thereby along the whole of L1 . . ψ2 and ψ3 . N = {N1 .3: Cubic Hermite triangle with edges L1 . we give a simple example. 2. and suppose that we wish to ﬁnd the local interpolant IK v of the function v deﬁned by v(x. 2. ⋄ 3. Given that v is a function for which all Ni ∈ N . Deﬁnition 5 Let (K. Example 11 Consider the triangle K shown in Figure 3. i=1 In order to illustrate the idea of local interpolant. . as follows. . . N1 ψ1 = 1. IK v = N1 (v)ψ1 + N2 (v)ψ2 + N3 (v)ψ3 . since Li (z4 ) = 0 for i = 1. L2 . .
i = 1. z2 and z3 where the local interpolant is evaluated. . which implies that IK (v) = v 2 for all v ∈ P. where δij = 1 when i = j and = 0 when i = j. k. indeed. 2 c) IK (v) = v for v ∈ P. .3. b) Ni (IK v)) = Ni (v). IK v = IK (IK v) = IK v since IK v ∈ P. . k. . . c) It follows from b) that Ni (v −IK (v)) = 0. . . . i = 1. L2 . b) Clearly k j=1 Ni (IK (v)) = Ni k Nj (v)ψj = k Nj (v)Ni (ψj ) j=1 = j=1 Nj (v)δij = Ni (v). . consequently IK is idempotent on P. is a linear functional. L3 . . k. The second assertion follows from this. and the vertices z1 . . in our case N1 (v) = v(z1 ) = 1. . ⋄ 2 1 2 1 and N3 (v) = v(z3 ) = 2 . CONSTRUCTION OF FINITE ELEMENT SPACES z3 = (0. 1) u @ @ @ @ @ @L1 @ K @ @ @ L3 @ uz @ 2 71 L2 z1 = (0. 0) Figure 3. . . IK = IK . . Lemma 11 The local interpolant has the following properties: a) The mapping v → IK v is linear.4: Linear Lagrange triangle with edges L1 . . v → IK v has the same property. i = 1. In fact. k. that is.1. 0) u = (1. for i = 1. . N2 (v) = v(z2 ) = and so 1 IK (v) = 1 − (x + y). The next lemma summarises the key properties of the local interpolant. Proof a) Since each Ni : v → Ni (v).
Deﬁnition 8 A triangulation of a polygonal domain Ω is a subdivision of Ω consisting of triangles which have the property that (3) No vertex of any triangle lies in the interior of an edge of another triangle. brieﬂy. and merge the local interpolants to obtain a global interpolant. when n > 2. such that (K. We can now glue together the element domains to obtain a subdivision of the computational domain. the next result is stated and proved in the case of n = 2. and (2) i ¯ ¯ Ki = Ω. To keep the presentation simple. again. namely when Ω ⊂ R2 . Assume that each element domain K in the subdivision is equipped with some type of shape functions P and nodal variables N . PIECEWISE POLYNOMIAL APPROXIMATION That completes the proof of the lemma. For simplicity. ¯ ¯ that it is C r ) if Ih v ∈ C r (Ω) for all v ∈ C m (Ω). . Deﬁnition 9 We say that an interpolant has continuity of order r (or. Next we shall formulate a simple condition ¯ which ensures that the global interpolant is a continuous function on Ω. N ) forms a ﬁnite element. although an analogous deﬁnition can me made in Rn .72 CHAPTER 3. The space ¯ {Ih v : v ∈ C m (Ω)} is called a C r ﬁnite element space. we shall restrict ourselves to the case of two space dimensions. P. In the absence of further conditions on the subdivision it is not possible to assert the continuity of the global interpolant. Deﬁnition 6 A subdivision of the computational domain Ω is a ﬁnite collection of open sets {Ki } such that (1) Ki ∩ Kj = ∅ if i = j. the obvious generalisation of this condition to n dimensions will be meant to hold. an analogous result holds for n > 2. we shall use the word triangulation without necessarily implying that Ω ⊂ R2 : when Ω ⊂ Rn and n = 2 we shall mean that the condition of this deﬁnition is satisﬁed. Let m be the order of the highest partial derivative involved in the nodal ¯ ¯ variables. Deﬁnition 7 Suppose that Ω is a bounded open set in Rn with subdivision T . For v ∈ C m (Ω) the global interpolant Ih v is deﬁned on Ω by Ih vKi = IKi v ∀Ki ∈ T . From now on.
Let Qk = j cj pj (x)qj (y) : pj and qj are polynomials of degree ≤ k . Now w is a polynomial of degree k and its restriction to the edge e vanishes at the onedimensional Lagrange (or Hermite) nodes. such that the global interpolant Ih v ¯ ¯ belongs to C 0 (Ω) for all v in C m (Ω). N ) be a ﬁnite element and suppose that F (x) = Ax + b where A is a nonsingular n×n matrix and x and b are ncomponent column vectors. where we interpret IK1 v and IK2 v to be deﬁned everywhere by extension outside K1 and K2 . where Pk denotes 1 the set of all polynomials of a single variable of degree ≤ k and dim Pk signiﬁes its dimension. 1 1 It can be shown that Qk is a linear space of dimension (dim Pk )2 . it follows that the edge nodes on e for the elements on both K1 and K2 are at the same location in space. K ∈ T . v v v Example 12 Lagrange elements on triangles with appropriate choice of edge and interior nodes are aﬃne equivalent. P.) Deﬁnition 10 Let (K. Here F ∗ is the pullback of F deﬁned by F ∗ (ˆ) = v ◦ F .3. Proof It suﬃces to show that continuity holds across each edge. More precisely. be two triangles in the triangulation T with common edge e. Let w = IK1 v − IK2 v. P. the global interpolant is continuous across each edge. given a triangulation T of Ω. CONSTRUCTION OF FINITE ELEMENT SPACES 73 Theorem 7 The Lagrange and Hermite elements on triangles are all C 0 elements. The same is true of Hermite elements on triangles. Let Ki .1. Assuming that we choose nodes interior to e in a symmetric way. i.4 Examples of rectangular elements To conclude this section we consider ﬁnite elements deﬁned on rectangles. ˆ ˆ ˆ The ﬁnite element (K. respectively. it is possible to choose edge nodes for the corresponding elements (K. we introduce the following deﬁnition (now for K ⊂ Ω ⊂ Rn . N ) if: ˆ (a) F (K) = K. In order to be able to compare global interpolation operators on diﬀerent elements. as polynomials. without going into details. 2. i = 1. N ) is aﬃne equivalent to (K. .1. P. ˆ (b) F ∗ P = P and ˆ (c) F∗ N = N . We give two examples. where m = 0 for Lagrange and m = 1 for Hermite elements. P.e. Therefore we ≡ 0. N ). ⋄ 3. and F∗ is the pushforward v ˆ ∗ of F deﬁned by (F∗ N )(ˆ) = N (F (ˆ)) = N (ˆ ◦ F ). Hence IK1 e v = IK2 ve .
z2 . . We shall adopt a constructive approach which will enable us to calculate the constants in the error estimates explicitly.2 Polynomial approximation in Sobolev spaces In this section we shall develop the approximation theory for the ﬁnite element spaces described in the previous section. z9 . using Lemmas 9 and 10 that N determines P = Q1 (the dimension of Q1 is equal to 4). L2 u z7 u uz9 z8 L3 z4 u L1 u L4 z5 u uz6 uz3 z2 Figure 3. ⋄ Example 14 (Biquadratic Lagrange rectangle) Let k = 2 and suppose that K is a rectangle. and centroid z5 . z8 . . z3 and z4 . . It is left as an exercise to show that N determines P = Q2 (the dimension of Q2 is equal to 9).6: Biquadratic Lagrange rectangle with edges L1 . . . z6 . . .74 CHAPTER 3. PIECEWISE POLYNOMIAL APPROXIMATION z3 u L2 uz4 L3 L4 z1 u L1 uz2 Figure 3. . L4 and the vertices z1 . i = 1. midpoints of edges z2 . z1 u Example 13 (Bilinear Lagrange rectangle) Let k = 1 and suppose that K is a rectangle. . i = 1. L3 . L2 . The technique is based on the use of the HardyLittlewood maximal function.5.5: Bilinear Lagrange rectangle with edges L1 . as in Figure 3. . ⋄ 3. Further. z3 . . We let P = Q2 and put N = {N1 . let P = Q1 and let N = {N1 . . . . following . L2 . We leave it as an exercise to the reader to show. as in Figure 3. 4. L4 and the vertices z1 .6. N9 } with Ni (v) = v(zi ) with zi . 9. L3 . N4 } with Ni (v) = v(zi ) with zi . z7 . . z4 . .
for all x ∈ Ω the closed convex hull of {x} ∪ B is m a subset of Ω). . An alternative approach which exploits the theory of a Riesz potentials is presented in the BrennerScott monograph cited in the reading list.2. j In order to complete the proof it remains to prove that j ∃Kj > 0 ∀v ∈ Wp (Ω) ∃Q ∈ Pm−1 such that j = 0. 20.e. Proof By hypothesis. Once (3.1 The BrambleHilbert lemma A key device in ﬁnite element error analysis is the following result. (3.3. 1 < p < ∞. we shall have that m−1 j=0 R. . m − 1.. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 75 the work of Ricardo Dur´n1 .2. . Dur´n: On polynomial approximation in Sobolev spaces. Since l(Q) = 0 for all Q ∈ Pm−1 . 985–988. (1983). j This will be done in the rest of the section. we have by the linearity of l that l(v) = l(v − Q) ≤ C0 v − Q m m Wp (Ω) = C0 j=0 v − Qp j Wp (Ω) 1/p + vp m (Ω) W p = C0 ≤ C0 m−1 j=0 v − Qp j Wp (Ω) 1/p m−1 j=0 m v − QWp (Ω) + vWp (Ω) . there exists C0 > 0 such that l(v) ≤ C0 v m Wp (Ω) m ∀v ∈ Wp (Ω). Lemma 12 (BrambleHilbert lemma) Suppose that Ω is a bounded open set in Rn and assume that Ω is starshaped with respect to every point in a set B of positive measure contained in Ω (i.2) m v − QWp (Ω) ≤ Kj vWp (Ω) .2) has been veriﬁed. SIAM Journal of Numerical a Analysis. 5. Then there exists a constant C1 > 0 such that m l(v) ≤ C1 vWp (Ω) m for all v ∈ Wp (Ω). 3. Let l be a bounded linear functional on the Sobolev space Wp (Ω). pp. . . No. 1 l(v) ≤ C0 1 + m Kj vWp (Ω) . such that l(Q) = 0 for any polynomial Q of degree ≤ m − 1. m ≥ 1.
2). Our main tool is the following lemma. . ν)p dσν ν=1 dx dσν ν=1 Rn g1 (x. . PIECEWISE POLYNOMIAL APPROXIMATION and the proof will be complete. g1 (x. ν)p dx p Rn ≤ ωn p p−1 g(x)p dx. or A.2) given by Bramble and Hilbert was based on the use of the HahnBanach theorem and was nonconstructive in nature in the sense that it did not provide computable constants Kj . Lemma 13 Let g ∈ Lp (Rn ). Now we are ready to prove (3. Stud. 1993. we deﬁne g1 (x.P. The remainder of this section is devoted to the (constructive) proof of (3. ν) is the HardyLittlewood maximal function of g(·) in the direction ν.2). ν)p dx ≤ p−1 n n R R and therefore Rn g ∗ (x)p = = Rn g1 (x. .S. it follows that2 p p g(x)p dx. ν) dσν p ω 1/p g p−1 n p .563–582. pp. The original proof of (3. Stein and T. Given ν ∈ Rn such that ν = 1. j = 0. where ωn is the measure of the unit sphere in Rn . Orthogonality and Oscillatory Integrals. Then g∗ Lp (Rn ) ≤ Lp (Rn ) .M. Math. Murphy: Harmonic Analysis: Real Variable Methods. 1 < p < ∞. 2 . only the existence of such constants was proved.76 CHAPTER 3. 44. . Calder´n: Estimates for o singular integral operators in terms of maximal functions. Upon taking the pth root of the two sides in this inequality we arrive at the desired result. (1972). with the constant C1 = C0 1 + m−1 j=0 Kj . Proof Since g1 (·. See E. ν) = sup t>0 1 t t 0 g(x + sν) ds 1/p and g (x) = ν=1 ∗ g1 (x. m−1. Princeton University Press.
· βn !. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 77 Theorem 8 Let Ω ⊂ Rn be a bounded open set which is starshaped with respect to each point in a set of positive measure B ⊂ Ω. . y) = β<m D β v(x) (y − x)β β! and Qm (v)(y) = Here we used the multiindex notation (y − x)β = (y1 − x1 )β1 · . ♯A denotes the number of elements in A and. β=m−j with 1/p + 1/p′ = 1. . As 1 B B [D α v(y) − Pm−j (D α v)(x. y) dx. · (yn − xn )βn . β! = β1 ! · . . . It is easy to prove by induction that D α Qm (v)(y) = Qm−α (D α v)(y). Academic Press. y)] dx. D α (v − Qm (v)) p Lp (Ω) 1 B Pm (v)(x. for a multiindex β = (β1 . we deﬁne Pm (v)(x. Now let us estimate D α v − Qm−j (D α v) (D α v − Qm−j (D α v))(y) = 3 Lp (Ω) for each α. B v − Qm (v)Wp (Ω) = j ≤ = α=j 1/p α=j D α (v − Qm (v)) Lp (Ω) α=j D α v − Qm−j (D α v) Lp (Ω) . βn ). . m ¯ ¯ Proof Because C ∞ (Ω) is dense3 in Wp (Ω). Adams: Sobolev Spaces. it suﬃces to prove the theorem for v ∈ C ∞ (Ω). Thus. 0 ≤ j < m. 1975. Here. . α = j. Let 1 < p < ∞. . and let m d be the diameter of Ω. . See R. B1/p 1/p′ where B denotes the measure of B and C = (♯{α : α = j}) m−j p ω 1/p n1/p p − 1 n (β!)−p ′ .A. If v ∈ Wp (Ω) then Q∈Pm−1 inf j v − QWp (Ω) ≤ C dm−j+(n/p) m vWp (Ω) . for a set A. .3.2. Given x ∈ B.
Noting that B ⊂ Ω. y)p ≤ (m − j)p d(m−j)p g1 x.3) Now recalling the integralremainder for Taylor series.78 CHAPTER 3. as deﬁned in Lemma 13. If g1 and g ∗ are the functions associated with g. ν) dσν rn−1 dr 1/p p g1 (x. 1/p Ω D α v(y) − Pm−j (D α v)(x. y)p dy dx. ·) Lp (Ω) dx. . y ∈ Ω. B u(x. it follows for each x ∈ B that Ω y−x y − x . y−x y − x dy. we have that y−x D α v(y) − Pm−j (D α v)(x. y) ≤ (m − j)dm−j g1 x. for a function u ∈ C(B × Ω). y) = (m − j) β=m−j 1 0 β=m−j (y − x)β β! 1 0 D β D α v(x + t(y − x))(1 − t)m−j−1 dt ≤ (m − j)dm−j = (m − j)dm−j 1 β α D D v(x + t(y − x)) dt β! 1 y−x Dβ Dα v x + s β! y − x ds. (3. Thus. 1 y − x y−x 0 β=m−j 1 Let g be the function that coincides with β=m−j β! D β D α v in Ω and is identically zero outside Ω. ¯ ¯ Minkowski’s integral inequality states that. y)p dy ≤ (m − j)d = (m − j)d m−j 0 m−j d ν=1 1/p ν=1 1/p 1/p p g1 (x. by applying Minkowski’s inequality for integrals4 that D α v − Qm−j (D α v) ≤ 1 B B Ω Lp (Ω) 1/p D α v(y) − Pm−j (D α v)(x. y)p dy ≤ (m − j)p d(m−j)p y−x≤d p g1 x. ν) dσν dn n dn n = (m − j)dm−j 4 g ∗ (x). D α v(y) − Pm−j (D α v)(x. y − x and therefore p D α v(y) − Pm−j (D α v)(x. for x ∈ B. we have that D α v(y) − Pm−j (D α v)(x. ·) dx Lp (Ω) ≤ B u(x. . PIECEWISE POLYNOMIAL APPROXIMATION it follows.
µ)dm−j vWp (Ω) . j. n. .3. Therefore. g Lp (Ω) = β=m−j 1 β α D D v β! Lp (Ω) ≤ β=m−j 1 Dβ Dα v β! (β!)−p m−j n1/p ′ Lp (Ω) where 1/p + 1/p′ = 1. However. If v ∈ Wp (Ω) then Q∈Pm−1 inf m j v − QWp (Ω) ≤ C(m. D α v − Qm−j (D α v) ≤ ≤ β=m−j 1/p′ Dβ Dα v β=m−j p Lp (Ω) 1/p . B1/p Since Qm ∈ Pm−1 .1. o D α v − Qm−j (D α v) Lp (Ω) 79 ≤ ≤ 1 (m − j) dm−j B dn n 1/p B g ∗ (x) dx dn n 1/p 1 (m − j) dm−j B1−(1/p) B g∗ Lp (Rn ) and hence. ′ 1/p′ v − Qm (v)Wp (Ω) ≤ Kj vWp (Ω) . µ ∈ (0. by Lemma 13. 1]. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES Inserting this into (3. m ≥ 1. D α v − Qm−j (D α v) Lp (Ω) ≤ = m − j dm−j+(n/p) p ω 1/p g p−1 n n1/p B1/p m − j dm−j+(n/p) p ω 1/p g p−1 n n1/p B1/p Lp (Rn ) Lp (Ω) . Kj = C Corollary 3 Let Ω ⊂ Rn be a bounded open set of diameter d which is starshaped with respect to every point of an open ball B ⊂ Ω of diameter µd.2). by H¨lder’s inequality (see Ch. and thereby also the proof of the BrambleHilbert lemma (Lemma 12). dm−j+(n/p) B1/p Lp (Ω) p ω 1/p p−1 n Dβ Dα v β=m−j Recalling that α = j. we deduce that where × β=m−j p Lp (Ω) 1/p (β!)−p . p. m j dm−j+(n/p) .2.3) we get. Suppose m that 1 < p < ∞ and 0 ≤ j < m.1. this completes the proof of Theorem 8.1. Sec.
n. 1 < p < ∞. and each element of the set N is a bounded linear functional on m ¯ ¯ C l (K)).2. µ).80 where CHAPTER 3. j. j. (µd)n ωn . u and that completes the proof. . We deﬁne σ(K) = sup ¯ v∈C l (K) m IK v Wp (K) . and let {ψ1 . by hypothesis. ψi m Wp (K) u ¯ C l (K) i=1 = Const.e. . N ) be a ﬁnite element such that the diameter of K is equal to m ¯ 1. Then the local interpolation operator is bounded from C l (K) into Wp (K) for 1 < p < ∞. . v C l (K) ¯ m ¯ the norm of the local interpolation operator IK : C l (K) → Wp (K). P. i=1 m m where each ψi ∈ W∞ (K) ⊂ Wp (K). Proof Let N = {N1 . µ should be taken as large as possible a number in the interval (0. n. . n In order to minimise the size of the constant C(m. Lemma 14 Let (K. the nodal variables in N involve derivatives up to order l.2 Error bounds on the interpolation error We shall apply Corollary 3 to derive a bound on the error between a function and its ﬁnite element interpolant. Nk }. ψk } ⊂ P be the basis dual to N . p. The local interpolant of a function u is deﬁned by the formula k IK u = Ni (u)ψi . . Proof Note that B = p(m − j) p−1 β=m−j (β!)−p ′ 1/p . P ⊂ W∞ (K) and N ⊂ (C l (K))′ (i. . Thus k IK u m Wp (K) ≤ ≤ i=1 k Ni (u) ψi Ni m Wp (K) ¯ (C l (K)′ ¯ C l (K) . . We begin by estimating the norm of the local interpolation operator. µ) = µ−n/p (♯{α : α = j}) with 1/p + 1/p′ = 1. p. 1] such that Ω is starshaped with respect to each point in a ball B ⊂ Ω of radius µd. . 3. . PIECEWISE POLYNOMIAL APPROXIMATION C(m.
p.2) we deduce that v − IK v m Wp (K) ≤ C(m. m ¯ The Sobolev embedding theorem asserts that Wp (K) ⊂ C l (K) for m − l > n . Finally. K ˆ where hK is the diameter of K. P. Let Qm v be as in the proof of Theorem 8. Then.3. m and hence. Also.p such that.n. Thus.n. because Qm v ∈ Pm−1 ⊂ P. for 0 ≤ j ≤ m. the general case follows by a simple scaling argument. Since IK f = f for any f ∈ P. N ) be a ﬁnite element satisfying the following conditions: (i) K is starshaped with respect to some ball contained in K. under a certain regularity condition on the subdivision T = ˆ {K} of the computational domain Ω.p v m WP (K) . in which case K = K. note that the local interpolation operator m is well deﬁned on Wp (K) by the Sobolev embedding theorem5 and there exists a constant m C = Cm.2. for 0 ≤ j ≤ m and m v ∈ Wp (K) we have that ˆ m j v − IK vWp (K) ≤ C(m. That completes the proof. Suppose that 1 < p < ∞ and m − l − (n/p) > 0.p σ(K)) v − Qm v v − Qm v + IK (Qm v − IK v) m Wp (K) . n. σ(K)) can be made ˆ independent of σ(K). + σ(K) Qm v − IK v ¯ C l (K) by the Sobolev embedding theorem. by (3. m (ii) Pm−1 ⊂ P ⊂ W∞ (K). µ. n. n. p.n. 1 ≤ p < ∞ and p m l ¯ the identity operator Id : v ∈ Wp (K) → v ∈ C (K) is a bounded linear operator. σ(K))hm−j vWp (K) . v − IK v m Wp (K) ≤ = ≤ v − Qm v m Wp (K) m Wp (K) m Wp (K) v − Qm v + Q m v − IK v m Wp (K) m Wp (K) ≤ (1 + Cm. σ(K)) vWp (K) . µ. σ(K)) vWp (K) . ¯ (iii) N ⊂ (C l (K))′ . µ. v ¯ C l (K) ≤ Cm. n. we have that IK Qm v = Qm v. 5 . we have that m m v − IK vWp (K) ≤ C(m. ˆ Proof It suﬃces to take K with diameter equal to 1. Next we show that. p. the constant C(m. K = {x/hK : x ∈ K} and µ is the largest real number in the interval (0. p. for all v ∈ Wp (K). µ. 1] such that a ball of diameter µhK is contained in K. POLYNOMIAL APPROXIMATION IN SOBOLEV SPACES 81 Theorem 9 Let (K.
N ). ˆ C l (K) (A ) ψN Since ψN −1 ∗ m ˆ Wp (K) ≤ ′ CN. it would have been ˜ ˜ ˜ better to use a new symbol (K. instead of (K. and N  denotes the number of nodal variables (i. N ) is aﬃne equivalent to a single reference element (K.l } max{CN. ˆ ˆ ≤ CN. ˆ C l (K) ′ Cref = N  max{CN. signify a ﬁnite element on an element domain K in the triangulation.82 CHAPTER 3. Thus. m v ˆ ¯ . N ) through an aﬃne transformation ˆ x → x = Ax ≡ ax + b. we have that l ˆˆˆ IK v m ˆ Wp (K) ≤ Cref 1≤i. the dimension of P). . P. again.j≤n −1 det a1/p ψN m Wp (K) .n. N ) and x instead of x. to denote the ˜ aﬃne image. Here and in the next 17 lines we shall. associated with an element domain ˆ ˆ ˆ K in the triangulation. The deﬁnition of aﬃne equivalence yields: ˆˆˆ x IK v (ˆ) = where (A∗ N )(ˆ) = N (A∗ v ). Thereafter. P. say. v x N ∈N (A∗ v )x = v (Ax). we have shown that l m ˆ σ(K) ≤ Cref 6 1 + max aij  1≤i.j≤n × 1 + max (a )ij  1≤i. P. m Wp (K) is a ﬁxed constant on the reference element K.l 1 + max aij  Also.j≤n m × 1 + max (a−1 )ij  where det a1/p v ˆ ¯ . (A∗ N )(ˆ) = N (A∗ v ) ≤ CN A∗ v v ˆ ˆ 1≤i.n. P. P.e.j≤n ¯ C l (K) l (A∗ N )ˆ(A−1 )∗ ψN (ˆ). N ).j≤n −1 det a1/p . where a = (aij ) is an invertible N × N matrix and b is a column vector of size N . PIECEWISE POLYNOMIAL APPROXIMATION ˆ ˆ ˆ Let us suppose that each (K.m 1 + max (a )ij  1≤i. P. N ) will. of the same length as the column vector x.6 We shall denote the entries of the matrix a−1 by (a−1 )ij . N ) considered here. To avoid confusion between the ﬁnite element (K. but this would have complicated the notation.n. and the aﬃne image of (K. temporarily. P. v ˆ Thus. (K. adopt this sloppy notation.m } max{ ψN N ∈N N ∈N N ∈N m Wp (K) }.j≤n 1 + max aij  1≤i.n.
7) (c. (3. for the ﬁnite element approximation uh ∈ Vh to the weak solution u of the homogeneous Dirichlet boundary value problem for a secondorder elliptic equation. µ)hm−j vWp (Ω) . (3. . still supposing that the triangulation T is regular. that u − uh H 1 (Ω) ≤ c1 inf u − vh c0 vh ∈Vh H 1 (Ω) . we proved that. 3. m j v − IK vWp (K) ≤ C(m. n. it is a straightforward exercise in geometry to show ˆ that σ(K) ≤ Cµ .6) where h = maxK∈T hK . p.3. whenever u ∈ H 2 (Ω) ∩ H0 (Ω). we have that m j v − Ih vWp (Ω) ≤ C(m. n. 1 < p < ∞. (3. m − l − (n/p) > 0 and 0 ≤ j ≤ m. n. Consequently. These interpolation error estimates are of crucial importance in ﬁnite element error analysis.4) for each K ∈ T provided that T is a regular subdivision.3. C´a’s lemma 5). p. restricting ourselves to the case of Poisson’s equation e and a continuous piecewise linear approximation uh deﬁned on a uniform triangula1 tion of Ω = (0. m 1 < p < ∞. 83 where hK is the diameter of K and ρK is the radius of the largest sphere (largest circle for n = 2) contained in K.4). and recalling the deﬁnition of the global interpolant of m v ∈ Wp (Ω) it follows that 1/p (j−m)p hK v K∈T − Ih vp j (K) Wp m ≤ C(m. where Cµ is a ﬁxed constant dependent on µ.3 Optimal error bounds in the H 1(Ω) norm – revisited In this section we return to the discussion of error estimation in the H 1 (Ω) norm. In Chapter 2 we showed. µ)hm−j vWp (K) K (3. from this. OPTIMAL ERROR BOUNDS IN THE H 1 (Ω) NORM – REVISITED Assuming that the subdivision T = {K} is regular in the sense that ∃µ > 0 ∀K ∈ T µhK ≤ ρK (≤ hK ). Thus. p. m−l−(n/p) > 0 and 0 ≤ j ≤ m.5) We shall also need the following somewhat cruder statement which is a straightforward consequence of (3. but independent of ˆ K ∈ T . µ)vWp (Ω) . 1)2 . and v ∈ Wp (Ω).f. we have u − uh H 1 (Ω) ≤ ChuH 2 (Ω) .
PIECEWISE POLYNOMIAL APPROXIMATION Now. we can generalise to higher degree piecewise polynomial approximations. in general. also. Let us consider the weak formulation of the secondorder elliptic partial diﬀerential equation (1.84 CHAPTER 3. Suppose that Ω ⊂ Rn and that it can be represented as a union of element domains K such that conditions (i). we arrive at the following error bound: u − uh H 1 (Ω) ≤ C(m. In particular.8) 1 provided that u ∈ H m (Ω) ∩ H0 (Ω). The inequality (3. c0 )hm−1 uH m (Ω) . Substituting this into (3. (3. µ.j=1 Ω n aij (x) + i=1 Ω ∂w ∂v dx ∂xi ∂xj ∂w v dx + ∂xi c(x)wv dx Ω bi (x) (3.8) is usually referred to as an optimal error bound. v) = l(v) for all v ∈ H0 (Ω). equipped with the interpolation error estimate (3.10) . n a(w. n. vh ∈Vh inf u − vh H 1 (Ω) ≤ u − Ih u H 1 (Ω) ≤ C(m.6) we can derive an analogous error bound in a more general setting. where. µ)hm−1 uH m (Ω) .4 Variational crimes To conclude this chapter we brieﬂy comment on a further issue which arises in the implementation of ﬁnite element methods.6): 1 ﬁnd u ∈ H 1 (Ω) such that a(u.5) on a bounded open set Ω ⊂ Rn . Then. n. Given such a triangulation T of Ω we shall suppose that it is regular in the sense introduced in the previous section and we put 1 Vh = Ih (H m (Ω) ∩ H0 (Ω)). with m = 2 corresponding to our earlier result with piecewise linear basis functions. as before. c1 . (ii) and (iii) of Theorem 9 hold with p = 2. in the case of a homogeneous Dirichlet boundary condition (1. since for a given m the smallest possible error that can be. v) = i. achieved in the H 1 (Ω) norm is of size O(hm−1 ). 3. (3. this will be the case if we use continuous piecewise polynomials of degree m − 1 on a regular triangulation of Ω.9) and l(v) = Ω f (x)v(x) dx.7).
and Galerkin orthogonality no longer holds. We say that we have committed a variational crime by replacing l(·) by lh (·).3. vh ) = l(vh ) − lh (vh ) = 0. ·) and l(·) will not be possible to evaluate exactly. ·) is still calculated exactly. and considering the approximate problem ﬁnd uh ∈ Vh such that a(uh . We wish to study the extent to which the accuracy of the basic ﬁnite element approximation is disturbed by this variational crime. vh ) − a(uh . ·) and l(·) approximately. c(x) − 2 i=1 ∂xi we have that a(v. ≤ = = ≤ a(u − uh . thereby leading to the following deﬁnition of uh : ﬁnd uh ∈ Vh such that a(uh . unless the coeﬃcients aij . u − vh ) + l(vh − uh ) − lh (vh − uh ) c1 u − uh H 1 (Ω) u − vh H 1 (Ω) l(wh ) − lh (wh ) vh − uh H 1 (Ω) . Assuming that n ∂bi 1 ≥ 0. VARIATIONAL CRIMES 85 The associated ﬁnite element method is based on choosing a ﬁnite element sub1 space Vh ⊂ H0 (Ω) consisting of continuous piecewise polynomials of a certain degree deﬁned on a subdivision of the computational domain Ω. we attempt to analyse the eﬀects of this quadratureinduced perturbation on the accuracy of the exactlyintegrated ﬁnite element method. let us suppose that the bilinear form a(·. however. u − vh ) + a(u − uh . To keep the discussion simple. Simpson’s rule. vh ) = lh (vh ) for all vh ∈ Vh . we have that a(u − uh . but that l(·) has been replaced by an approximation lh (·). Thus. vh ∈ Vh . Unfortunately. Without focusing on any particular quadrature rule. vh − uh ) a(u − uh .4. Gausstype rules and their multidimensional counterparts) will have to be used to calculate a(·. bi and c and the righthand side f are exceptionally simple functions. the integrals which appear in the deﬁnitions of a(·. vh ) = a(u. With this new deﬁnition of uh . with c0 a positive constant (as in Section 1.2). v) ≥ c0 v c 0 u − uh 2 H 1 (Ω) 2 H 1 (Ω) . We recall that a key step in developing the ﬁnite element error analysis was the presence of the Galerkin orthogonality property. vh ) = l(vh ) for all vh ∈ Vh . u − uh ) a(u − uh . and numerical integration rules (such as the trapezium rule. + sup wh H 1 (Ω) wh ∈Vh .
√ c1 2 u − vh ≤ c0 H 1 (Ω) + √ 2 c0 sup wh ∈Vh l(wh ) − lh (wh ) . it follows that we have proved the following perturbed version of C´a’s lemma: e √ c1 2 u − uh H 1 (Ω) ≤ min u − vh H 1 (Ω) c0 vh ∈Vh √ l(wh ) − lh (wh ) 2 sup . 2c0 2 ≤ c 1 u − uh H 1 (Ω) u − vh H 1 (Ω) (3. 1 2 c0 2 a + b. To simplify writing.12) Since vh ∈ Vh is arbitrary. √ c1 2 ≤ u − vh c0 H 1 (Ω) + c0 l − lh −1. u − uh Equivalently. this yields c 2 u − uh 0 and therefore. + l − lh √ 2 2 −1.h = sup wh ∈Vh l(wh ) − lh (wh ) .11) gives c 2 u − uh 0 2 H 1 (Ω) a.h u − uh H 1 (Ω) 1 c0 2 ≤ c1 u − vh H 1 (Ω) + l − lh −1.11) Now applying the elementary inequality ab ≤ we have that c1 u − vh H 1 (Ω) + l − lh −1. u − uh H 1 (Ω) H 1 (Ω) 2 H 1 (Ω) ≤ 2 c1 u − vh H 1 (Ω) . + c0 wh ∈Vh wh H 1 (Ω) . PIECEWISE POLYNOMIAL APPROXIMATION with c1 a positive constant (as in Section 1.h u − uh H 1 (Ω) + l − lh −1. we deﬁne l − lh Hence.h +2c0 l − lh −1.h u − vh H 1 (Ω) .h u − uh H 1 (Ω) + u − vh H 1 (Ω) = c1 u − vh H 1 (Ω) + l − lh −1. c 0 u − uh 2 H 1 (Ω) −1.2).h u − vh H 1 (Ω) . wh H 1 (Ω) + l − lh −1. ≤ c1 u − vh H 1 (Ω) + l − lh −1. 2 H 1 (Ω) .h + u − uh 2c0 2 Substituting this into (3.86 CHAPTER 3. wh H 1 (Ω) (3.h 2 Noting that c0 ≤ c1 . b ≥ 0.h .
µ. c1 .4.12) quantiﬁes the extent to which the accuracy of the exactly integrated ﬁnite element method is aﬀected by the failure of Galerkin orthogonality. sup c0 wh ∈Vh wh H 1 (Ω) Thus. We shall not discuss variational crimes which arise from replacing the computational domain Ω by a “conveniently chosen closeby domain” Ωh .3 will lead to the error bound u − uh H 1 (Ω) ≤ C(m. in order to retain the accuracy of the exactly integrated method. ·) is analysed in a similar manner. arguing in the same manner as in Section 3. for the details of the analysis the reader is referred to the books on the reading list and references therein. n. Indeed. c0 )hm−1 uH m (Ω) √ 2 l(wh ) − lh (wh ) + . ·) is perturbed to ah (·. . The situation when a(·.3. in the case of continuous piecewise linear basis functions (m = 2) this means that the additional error should be at most O(h). VARIATIONAL CRIMES 87 The second term on the righthand side of (3. the numerical quadrature rule has to be selected so that the second term on the right is also of size O(hm−1 ).
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. v) = 0 [w′ (x)v ′ (x) + b(x)w′ (x)v(x) + c(x)w(x)v(x)] dx 1 and l(v) = 0 f (x)v(x) dx. u(0) = 0. 1] by the points 0 = x0 < x1 < 89 Assuming that . 1). 1 where b ∈ W∞ (0. u ∈ H0 (0. 4. 1). 0 < x < 1. Letting 1 a(w. 1 (4.1 The onedimensional model problem In order to illuminate the key ideas and avoid technical diﬃculties. the weak formulation of this problem can be stated as follows: 1 1 ﬁnd u ∈ H0 (0.1) c(x) − b′ (x) ≥ 0. for x ∈ (0. u(1) = 0. 1). 1) such that a(u. c ∈ L∞ (0. 1). 2 1 there exists a unique weak solution. The ﬁnite element approximation of this problem is constructed by considering a (possibly nonuniform) subdivision of the interval [0. 1). v) = l(v) for all v ∈ H0 (0. we shall consider the twopoint boundary value problem −u′′ + b(x)u′ + c(x)u = f (x). 1) and f ∈ L2 (0.
z − Ih z) = i=1 u′h (x) (z − Ih z)′ (x) dx b(x) u′h (x) (z − Ih z)(x) dx c(x) uh (x) (z − Ih z)(x) dx. that is. Thus. and put h = maxi hi . z − Ih z) − a(uh . i = 1. choosing zh = Ih z ∈ Vh . . u − uh ) = (u − uh . xi xi−1 xi xi−1 + i=1 N + i=1 . we have that a(u − uh . Now. as in an a priori error analysis). 0 < x < 1. N . To keep matters simple. . We let hi = xi − xi−1 .90 CHAPTER 4. associated with the subdivision 0 = x0 < x1 < . z − Ih z). zh ) = 0 ∀zh ∈ Vh . 1) consisting of continuous piecewise polynomials of a certain degree on this subdivision. The ﬁnite element approximation of the boundary value problem is: ﬁnd uh ∈ Vh such that a(uh . .2) We observe that by this stage the righthand side no longer involves the unknown analytical solution u. as described in Chapter 2. . A POSTERIORI ERROR ANALYSIS BY DUALITY 1 . z − Ih z) = (f. vh ) = l(vh ) for all vh ∈ Vh . (4. z − Ih z) − a(uh . called the dual or adjoint problem. N xi xi−1 N a(uh . −z ′′ − (bz)′ + cz) = a(u − uh . < xN −1 < xN = 1 and deﬁning the ﬁnite element space Vh ⊂ H0 (0. let us suppose that Vh consists of continuous piecewise linear functions. z(1) = 0. a(u − uh . (u−uh )(1) = 0): u − uh 2 L2 (0. . the continuous piecewise linear interpolant of the function z. < xN −1 < xN = 1. We begin our error analysis by noting that the deﬁnition of the dual problem and a straightforward integration by parts yield (recall that (u−uh )(0) = 0.1) = a(u − uh . We wish to derive an a posteriori error bound. Ih z) = 0. . By virtue of the Galerkin orthogonality property. we aim to quantify the size of the global error u − uh in terms of the mesh parameter h and the computed solution uh (rather then the analytical solution u. u − uh 2 L2 (0.1) = (u − uh . z). . To do so. . we consider the following auxiliary boundary value problem −z ′′ − (b(x)z)′ + c(x)z = (u − uh )(x). z − Ih z) = a(u. In particular. z(0) = 0.
Now. .xi ) z ′′ L2 (0. 1). 2 L2 (0. for i = 1. .3) R(uh )(x) = f (x) + u′′ (x) − b(x)u′h (x) − c(x)uh (x). N ) that 2 hi z ′′ L2 (xi−1 .xi ) z ′′ L2 (xi−1 . .2). .1) .1) L 1 ≤ 2 π N h2 R(uh ) i i=1 L2 (xi−1 . .4. THE ONEDIMENSIONAL MODEL PROBLEM 91 Integrating by parts in each of the (N −1) integrals in the ﬁrst sum on the righthand side. . z − Ih z L2 (xi−1 . Recalling from the proof of Theorem 3 (with ζ = z − Ih z and noting that ζ ′′ (x) = z ′′ (x) for all x in (xi−1 . xi ). . z − Ih z) = Further i=1 [−u′′ (x) + b(x)u′h (x) + c(x)uh (x)] (z − Ih z)(x) dx. .xi ) . h N xi xi−1 (f. we deduce that N u − uh where. noting that (z − Ih z)(xi ) = 0.xi ) z − Ih z L2 (xi−1 . .1) xi xi−1 = i=1 R(uh )(x) (z − Ih z)(x) dx. N . we deduce that N xi xi−1 a(uh .1) L ≤ R(uh ) i=1 L2 (xi−1 .xi ) ≤ π we deduce that u− and consequently.xi ) .1. since Ih z is a linear function on (xi−1 . N. .3) yields N u− uh 2 2 (0. z − zh ) = i=1 f (x) (z − Ih z)(x) dx.xi ) 1 ≤ 2 π N 1/2 h4 R(uh ) i i=1 2 L2 (xi−1 . . . We recall that z ′′ = uh − u − (b z)′ + c z = uh − u − b z ′ + (c − b′ ) z. . . u − uh 2 L2 (0. (4.4) The rest of the analysis is aimed at eliminating z ′′ from the righthand side of (4. The function R(uh ) is called the ﬁnite element residual. applying the CauchySchwarz inequality on the righthand side of (4. i = 1. xi ). i = 0. Substituting these two identities into (4. N .1) uh 2 2 (0. . (4. i = 1. . . xi ). h x ∈ (xi−1 . it measures the extent to which uh fails to satisfy the diﬀerential equation −u′′ + b(x)u′ + c(x)u = f (x) on the interval (0.4).
1) and z L2 (0. z ′ ) + (cz.6) By the Poincar´Friedrichs inequality. z′ 2 ≤ (u − uh . Integrating by parts and noting that z(0) = 0 and z(1) = 0 yields (−z ′′ − (bz)′ + cz. b(x)[z 2 (x)]′ dx + = z ′ 2 2 (0. again. 2 Inserting this into the righthand side of (4.1) 1 ≤ √ u − uh 2 L2 (0. in the second term on the right gives (−z − (bz) + cz.9) 1 K =1+ √ b 2 + 1 c − b′ 2 L∞ (0.7) and (4.1) .5). z ′ ) + (bz. e z Thus.1) z ′ + c − b′ L∞ (0.1) L2 (0.1) . noting (4.1) ≤ (4.1) .1) .7) z′ L2 (0.92 CHAPTER 4.1) can be bounded in terms of u − uh L2 (0.1) . A POSTERIORI ERROR ANALYSIS BY DUALITY and therefore. by virtue of (4. Let us observe that (−z ′′ − (bz)′ + cz.1) + L 2 0 0 Integrating by parts. z) = Hence. 1 u − uh L2 (0. L2 (0. we shall deduce that the same is true of z ′′ L2 (0.1) + b L∞ (0. .6) yields z L2 (0. z) 1 1 1 c(x)[z(x)]2 dx. (4.5) We shall show that both z ′ L2 (0. (4. z) = (z ′ . z ′ 2 2 (0.1) ≤ 2 L2 (0.1) (4.1) . z ′′ L2 (0.8) into (4. 2 and thereby.1) z L2 (0.1). z) = (u − uh .1) L 1 1 − 2 1 0 b (x)[z (x)] dx + 0 ′ 1 2 c(x)[z(x)]2 dx. z).1) z L2 (0. z).5) to deduce that z ′′ Where L2 (0. z) ≤ u − uh 1 ′ z 2 L2 (0.8) Now we substitute (4.6) gives 2 L2 (0. 1 c(x) − b′ (x) [z(x)]2 dx = (u − uh .1) and then.1) . (4.1) ≤ K u − uh L∞ (0.1) .1) ≤ u − uh L2 (0. (4.1) L + 0 ′′ ′ z ′ 2 2 (0.
10). Inserting (4. AN ADAPTIVE ALGORITHM 93 It is important to note here that K0 involves only known quantities. and h(0) = (0) maxi hi . namely the coeﬃcients in the diﬀerential equation under consideration. Compute the corresponding solution uh(0) ∈ Vh(0) . with hi = xi − xi−1 for i = 1.2 An adaptive algorithm Suppose that T OL is a prescribed tolerance and that our aim is to compute a ﬁnite element approximation uh to the unknown solution u (with the same deﬁnition of u and uh as in the previous section) so that u − uh L2 (0. The name a posteriori stems from the fact that (4. 3. stop if Nm K0 i=1 (m) hi 4 1/2 R(uh(m) ) 2 (m) (m) L2 (xi−1 . (4. We shall use the a posteriori error bound (4. deﬁned on a subdivision Tm .4.1) ≤ T OL. 1].xi ) ≤ T OL is satisﬁed. Given a computed solution uh(m) ∈ Vh(m) for some m ≥ 0.2. until the stopping criterion N 1/2 K0 i=1 h4 R(uh ) i 2 L2 (xi−1 . .10) where K0 = K/π 2 .xi ) ≤ T OL. < xN0 −1 < xN0 = 1 (0) (0) (0) (0) (0) (0) (0) of the interval [0.9) into (4. N0 . (4.11) . . 2. .1) ≤ K0 h4 R(uh ) i i=1 2 L2 (xi−1 . and consider the associated ﬁnite element space Vh(0) (of dimension N0 − 1). 4. we arrive at our ﬁnal result. and computing a succession of numerical solutions uh on these subdivisions.xi ) . In the next section we shall describe the construction of an adaptive mesh reﬁnement algorithm based on the bound (4. Choose an initial subdivision T0 : 0 = x0 < x1 < . and therefore it can be computed without diﬃculty.4). The algorithm proceeds as follows: 1. . N 1/2 u − uh L2 (0. .10) can only be employed to quantify the size of the approximation error that has been committed in the course of the computation after uh has been computed. the computable a posteriori error bound. .10) to achieve this goal by successively reﬁning the subdivision.
94
CHAPTER 4. A POSTERIORI ERROR ANALYSIS BY DUALITY 4. If not, then determine a new subdivision Tm+1 : 0 = x0
(m+1)
< x1
(m+1)
< . . . < xNm+1 −1 < xNm+1 = 1
(m+1)
(m+1)
(m+1)
of the interval [0, 1], with hi = xi − xi−1 for i = 1, . . . , Nm+1 and (m+1) (m+1) h = maxi hi , and an associated ﬁnite element space Vh(m+1) (of dimension Nm+1 − 1), with h(m+1) as large as possible (and consequently Nm+1 as small as possible), such that
Nm+1 1/2
(m+1)
(m+1)
K0
i=1
hi
(m+1)
4
R(uh(m) )
2 (m+1) (m+1) L2 (xi−1 ,xi )
= T OL,
(4.12)
and continue. Here (4.11) is the stopping criterion and (4.12) is the mesh modiﬁcation strategy. According to the a posteriori error bound (4.10), when the algorithm terminates the global error u−uh L2 (0,1) is controlled to within the prescribed tolerance T OL. The relation (4.12) deﬁnes the new meshsize 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: hi
(m+1) 4
R(uh(m) )
2 (m+1) (m+1) L2 (xi−1 ,xi )
=
T OL2 2 K0 Nm+1
from which the hi can be found by treating this as an equation in hi , and solving it numerically, for m and i ﬁxed, by some rootﬁnding algorithm (e.g. successive bisection or ﬁxedpoint 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 computational algorithm. Eﬃciency means that the computational eﬀort required to achieve reliability is minimal. It is unclear from the present discussion whether the adaptive algorithm described above is eﬃcient in this sense: although we have minimised the computational eﬀort required to ensure that the righthand 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 eﬀort required to ensure that the lefthand 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 .
for each i = 1, . . . , Nm+1 ; the implementation can be simpliﬁed by replacing Nm+1 (m+1) on the righthand side by Nm . Then, we have a simple formula for hi : 1/4 2 T OL (m+1) , i = 1, . . . , Nm+1 , hi = 2 K0 Nm R(uh(m) ) 2 (m+1) (m+1)
L2 (xi−1 ,xi ) (m+1) (m+1)
Chapter 5 Evolution problems
In previous chapters we considered the ﬁnite element approximation of elliptic boundary value problems. This chapter is devoted to ﬁnite element methods for timedependent problems; in particular, we shall be concerned with the ﬁnite element approximation of parabolic equations. Hyperbolic equations will not be discussed in these notes.
5.1
The parabolic model problem
Let Ω be a bounded open set in Rn , n ≥ 1, with boundary Γ = ∂Ω, and let T > 0. In Q = Ω × (0, T ], we consider the initial boundary value problem for the unknown function u(x, t), x ∈ Ω, t ∈ (0, T ] : ∂u ∂ ∂u − (aij (x, t) )+ ∂t i,j=1 ∂xj ∂xi u(x, t) = 0, u(x, 0) = u0 (x),
n n
bi (x, t)
i=1
∂u + c(x, t)u = f (x, t), ∂xi (5.1) (5.2) (5.3)
x ∈ Ω, t ∈ (0, T ], x ∈ Γ, t ∈ [0, T ], ¯ x ∈ Ω.
Suppose that u0 ∈ L2 (Ω), and that there exists a positive constant c such that ˜
n n
i,j=1
aij (x, t)ξi ξj ≥ c ˜
ξi2 ,
i=1
¯ ∀ξ = (ξ1 , . . . , ξn ) ∈ Rn , ∀x ∈ Ω, t ∈ [0, T ]. We shall also assume that aij ∈ L∞ (Q), c ∈ L∞ (Q),
1 bi ∈ W∞ (Q), i, j = 1, . . . , n, f ∈ L2 (Q),
(5.4)
95
96 and that 1 c(x, t) − 2
n
CHAPTER 5. EVOLUTION PROBLEMS
i=1
∂bi (x, t) ≥ 0, ∂xi
¯ (x, t) ∈ Q,
(5.5)
as in the elliptic case. A partial diﬀerential equation of the form (5.1) is called a parabolic equation (of second order). Simple examples of parabolic equations are the heat equation ∂u = ∆u ∂t and the unsteady advectiondiﬀusion equation ∂u − ∆u + ∂t
n
bi
i=1
∂u = 0. ∂xi
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 signiﬁes time), and discuss the question of continuous dependence of the solution on the initial datum u0 and the forcing function f . We recall that, for v, w ∈ L2 (Ω), the inner product (u, v) and the norm v L2 (Ω) are deﬁned by (v, w) =
Ω
v(x)w(x) dx,
v
L2 (Ω)
= (v, v)1/2 .
Taking the inner product of (5.1) with u, noting that u(x, t) = 0, x ∈ Γ, integrating by parts, and applying (5.4) and (5.5), we get ∂u ∂u (·, t), u(·, t) + c ˜ (·, t) ∂t ∂xi i=1 Noting that ∂u (·, t), u(·, t) ∂t = 1 d u(·, t) 2 dt
2 L2 (Ω) , n 2 L2 (Ω)
≤ (f (·, t), u(·, t)).
and using the Poincar´Friedrichs inequality (1.2), we obtain e 1 d u(·, t) 2 dt
2 L2 (Ω)
+
c ˜ u(·, t) c⋆
2 L2 (Ω)
≤ (f (·, t), u(·, t)).
6).6) implies that the solution is unique. THE PARABOLIC MODEL PROBLEM Let K = c/c⋆ . by (5. we restrict ourselves to the heat equation in one space dimension. In the next section we consider some simple ﬁnite element methods for the numerical solution of parabolic initial boundary value problems.1)–(5. (5. however since c > 0 by hypothesis. on the ˜ “diﬀusion coeﬃcients” aij (i. ˜ Conservation of energy would correspond to u(·. (5. u ≡ 0. (5. (5.1. In this case (5.3) has a solution.e. t) K t 2 L2 (Ω) .6) yields u(·. then.1)–(5. d Kt e u(·. t) K 2 L2 (Ω) . t) 2 2 (Ω) + K u(·. Since K = 2 L2 (Ω) . eKt u(·.5. this will only occur by formally setting c = 0. the smaller c. we have ˜ u(·. t) dissipates exponentially. t) 2 L2 (Ω) = u0 2 L2 (Ω) . the energy c/c⋆ . but the analysis that we shall present also applies in the general setting. ≤ e−Kt u0 2 L2 (Ω) + 1 K e−K(t−τ ) f (·. In order to simplify the presentation. This corresponds to considering the evolution of the solution from the initial datum u0 in the absence of external forces. t) L L 2 dt K 1 f (·. 2 L2 (Ω) c ≤ e−˜t/c⋆ u0 t ≥ 0. the slower the decay of the energy). if u1 and u2 are solutions to (5. τ ) 2 L2 (Ω) dτ. t) 2 2 (Ω) . t) dt Integrating from 0 to t. t) Hence u(·. t) L2 (Ω) ≤ 1 f (·. t) 2 L2 (Ω) 1 2 ≤ e−Kt u0 u(·. c. ˜ ˜ conservation of energy will not be observed for a physical process modelled by a secondorder parabolic equation.6) Assuming that (5. t) 2 2 (Ω) ≤ f (·.7) in physical terms.1). τ ) 0 t 0 2 L2 (Ω) dτ. t) 2 L2 (Ω) 2 L2 (Ω) 2 L2 (Ω) 2 L2 (Ω) L2 (Ω) 97 u(·. t ≥ 0. ≤ eKt f (·. t) L dt Multiplying both sides by eKt . i. by the CauchySchwarz inequality. ≤ L L 2K 2 Thence.e. − u0 2 L2 (Ω) ≤ 1 K eKτ f (·. then u = u1 −u2 satisﬁes (5. u1 ≡ u2 . therefore. t) 2 L2 (Ω) 2 L2 (Ω) . d u(·. Indeed. t) 2 2 (Ω) + K u(·.8) and we deduce that the rate of dissipation depends on the lower bound.3) with f ≡ 0 and u0 ≡ 0. . ˜ 1 d u(·.1)–(5.3). Let us also look at the special case when f ≡ 0 in (5. t) 2 2 (Ω) + u(·.
the second (called the backward Euler scheme) uses a backward diﬀerence in t. We approximate (5.9) We describe two schemes for the numerical solution of (5. x = 0 and x = 1. tm ). h where um represents the approximation of u(·. Let Q = Ω × (0. t) such that ∂ 2u ∂u = + f (x. referred to as the forward Euler scheme: ﬁnd um ∈ Vh . vh ) = (um . t ∈ (0. 0) = u0 (x). ¯ The forward Euler scheme. tm ). and a(·. M ≥ 1. 1]. EVOLUTION PROBLEMS 5. We begin by constructing a mesh on Q = [0. Clearly. < xN −1 < xN = 1 which vanish at the endpoints. x ∈ [0. ·) is deﬁned by h 1 a(w. . 0 ≤ m ≤ M − 1. v) = 0 w′ (x)v ′ (x) dx. t) = 0. We ¯ ¯ deﬁne the uniform mesh Q∆t on Q by h ¯ Q∆t = {(xj . T > 0. 1).9) by the ﬁnite element method. here N and M are two integers. where Ω = (0. vh ) = (f (·. such that h um+1 − um h h . vh ) h h ∀vh ∈ Vh . T ]. vh ) ∀vh ∈ Vh . ﬁnd u(x. vh ) − ∆t a(uh . t ∈ [0.10) (u0 − u0 . 1] × [0. 1) denote the set of all continuous piecewise linear functions deﬁned on the xmesh 0 = x0 < x1 < . 1). They both use the same discretisation in the x variable but while the ﬁrst scheme (called the forward Euler scheme) employs a forward divided diﬀerence in t to approximate ∂u/∂t.98 CHAPTER 5.9). x ∈ (0. T ]. tm = m · ∆t. Let h = 1/N be the meshsize in the xdirection and let ∆t = T /M be the meshsize in the tdirection. vh + a(um . N ≥ 2. vh ) + ∆t (f (·. (5. t). . t) = 0. tm ). tm ) : xj = jh. 0 ≤ m ≤ M. vh ) = 0 ∀vh ∈ Vh . h ∆t (5.10) can be rewritten as follows: m (um+1 . u(x. ∂t ∂x2 u(0. (5. 0 ≤ m ≤ M }. T ]. h 1 Let Vh ⊂ H0 (0. .2 Forward and backward Euler schemes We consider the following simple model problem for the heat equation in one space dimension. 0 ≤ j ≤ N . T ]. u(1.
0 1 4 1 0 .. ﬁnding u0 still h only involves inverting the mass matrix M . Alternatively. φ′j ) where φi denotes the onedimensional piecewise linear ﬁnite element basis function associated with the xmesh point xi . 0 h M = 0 1 4 1 . where 2 −1 0 0 .. to ﬁnd um+1 at time level tm+1 we have to solve a system of linear h h equations with symmetric positive deﬁnite matrix A of size (N − 1) × (N − 1). with Thus... such that h m+1 uh − u m h . vh ) h h h ∀vh ∈ Vh . The backward Euler scheme... vh ) = (u0 . h is the socalled stiﬀness matrix. with entries (φi . to ﬁnd um+1 at time level tm+1 we have to solve a system of linear h h equations with symmetric positive deﬁnite matrix M of size (N − 1) × (N − 1). FORWARD AND BACKWARD EULER SCHEMES with 99 Thus. . .. one can approximate the time derivative by a backward diﬀerence. .. 0 1 0 −1 2 −1 ... . 0 0 . vh ) = 0 ∀vh ∈ Vh .5. It is a simple matter to show that this matrix is tridiagonal and h has the form 4 1 0 0 .. φj ) where φi denotes the onedimensional piecewise linear ﬁnite element basis function associated with the xmesh point xi . . tm )... vh ) = (um .. tm+1 ).. φj ) + ∆t (φ′i . with entries (φi . vh ) + ∆t a(um+1 .. 0 −1 2 (u0 ... 6 . (5.... vh + a(um+1 . which gives rise to the following backward Euler scheme: ﬁnd um ∈ Vh . vh ) ∀vh ∈ Vh .. . vh ) ∀vh ∈ Vh . . ... 0 K= h . It is clear that A = M + ∆t K. h ∆t (5. vh ) ∀vh ∈ Vh .. vh ) + ∆t (f (·... 0 1 4 The matrix M is usually referred to as the mass matrix. 0 0 .11) can be h written (um+1 .11) (u0 h where um represents the approximation of u(·. h − u0 . 0 −1 2 −1 0 . . . 0 ≤ m ≤ M. 0 . vh ) = (u0 . given um . tm+1 ). given um . 0 ≤ m ≤ M − 1... (u0 .2.... ..... the same matrix arises when determining u0 . . ... Equivalently. vh ) = (f (·.
11) simultaneously.3 Stability of θschemes We shall study the stability of the schemes (5. H h h 1 = ∆t θ − 2 um+1 − um um+1 + um h h h + h . v)1/2 . um+θ ) = (f m+θ . 2 L2 (Ω) m+θ + a(um+θ . vh + a(um+θ . EVOLUTION PROBLEMS 5. v) = 0 w(x)v(x) dx.100 CHAPTER 5. v L2 (Ω) = (v.10) and (5. such that h m+1 uh − u m h . uh ). 0 ≤ m ≤ M. and for the sake of notational simplicity. vh ) = (f m+θ . uh ). uh ∆t Equivalently. for θ = 1 the backward Euler scheme. vh ) = 0 ∀vh ∈ Vh . h h h For θ = 0 this gives the forward Euler scheme. Taking the inner product of (5. ∆t 2 (5.12) (w. um+θ ). uh ∆t Since um+θ h it follows that 1 ∆t θ − 2 um+1 2 2 (Ω) − um um+1 − um 2 h h L h h L2 (Ω) + ∆t 2∆t m+θ 2 m+θ m+θ +uh H 1 (Ω) = (f . vh ) ∀vh ∈ Vh h ∆t (u0 − u0 . tm ) and um+θ (x) = θum+1 (x) + (1 − θ)um (x). we wrote f m+θ (x) = θf (x. 1 The method corresponding to θ = 2 is known as the CrankNicolson scheme. um+1 − um m+θ h h . by embedding them into a oneparameter family of ﬁnite element schemes: ﬁnd um ∈ Vh . h where 0 ≤ θ ≤ 1. Recall that 1 (5. h h + um+θ 2 1 (Ω) = (f m+θ . tm+1 ) + (1 − θ)f (x.13) .12) with um+θ we get h um+1 − um m+θ h h .
12) on the initial data and the righthand side. then θ − 1/2 ≥ 0.3. k. 1]. and therefore um+1 h 2 L2 (Ω) 101 2∆t − um h 2 L2 (Ω) + um+θ 2 1 (Ω) ≤ (f m+θ . Now let us consider the case θ ∈ [0. 1].14) can be thought of as the discrete version of (5. If follows from (5. STABILITY OF θSCHEMES Suppose that θ ∈ [1/2. (5. e um+θ h Thus um+1 h so that um+1 h 2 L2 (Ω) 2 L2 (Ω) 2 L2 (Ω) 1 ≤ um+θ 2 1 (Ω) .12) is stable. 1/2). um+θ ). 1]. we get that k−1 uk 2 2 (Ω) h L ≤ u0 2 2 (Ω) h L + m=0 ∆t f m+θ 2 L2 (Ω) .14) that M −1 1≤k≤M max uk 2 2 (Ω) h L ≤ u0 2 2 (Ω) h L + m=0 ∆t f m+θ 2 L2 (Ω) . (5. .6). According to (5.13). . According to the Poincar´Friedrichs inequality. without any limitations on the time step in terms of h. 1 ≤ k ≤ M . = ∆t( − θ) h L2 (Ω) + (f 2 ∆t 2 L2 (Ω) − um h 2 L2 (Ω) (5.5. M −1 1≤k≤M 1/2 max uk h L2 (Ω) ≤ u0 h 2 L2 (Ω) + m=0 ∆t f m+θ 2 L2 (Ω) . m = 0. the scheme (5. . Thus we have proved that for θ ∈ [1/2. In other words. Summing through m.e. . This property is called stability.14) for all k. 2 h H 1 m+θ f 2 1 m+θ u 2 h 2∆t − um h 2 L2 (Ω) + 2 um+θ h 2 L2 (Ω) ≤ 2 L2 (Ω) + 2 L2 (Ω) . i.15) which expresses the continuous dependence of the solution to the ﬁnite element scheme (5. the scheme (5. um+1 h + um+θ 2 1 (Ω) H h 2∆t m+1 u − um 2 1 h m+θ .16) . um+θ ) H h h ≤ f m+θ L2 (Ω) um+θ h L2 (Ω) . ≤ um h 2 L2 (Ω) + ∆t f m+θ 2 L2 (Ω) .12) is unconditionally stable for θ ∈ [1/2. The inequality (5.
. we have that h h um+1 − um h h ∆t Therefore.18) with wh = (um+1 − um )/∆t.20) and (5. Now.17) Next we shall prove that. EVOLUTION PROBLEMS Recalling (5.19) we deduce (5.18). (5.19)  ≤ 2 wh H 1 (Ω) = h h h i=1 i=1 On the other hand. and note that W0 = WN = 0. .102 CHAPTER 5.12) with vh = (um+1 − um )/∆t. √ 12 wh H 1 (Ω) ≤ wh h L2 (Ω) . N . we continue the stability analysis. i = 0. (5.20) i=1 From (5. we deduce that h h um+1 − um h h ∆t 2 L2 (Ω) m+1 uh − um h ≤ f L2 (Ω) L2 (Ω) ∆t √ 12 m+θ um+1 − um h + uh H 1 (Ω) h h ∆t m+θ L2 (Ω) and hence um+1 − um h h ∆t L2 (Ω) ≤ f m+θ L2 (Ω) + √ 12 m+θ uh H 1 (Ω) h (5. equipped with the inequality (5. A simple calculation reveals that N −1 N 4 Wi − Wi−1 2 2 hWi 2 . Applying (5. . h m um+1 − uh h ∆t . . um+1 − um h h ∆t 2 L2 (Ω) 2 L2 (Ω) = f m+θ . for each wh ∈ Vh . (5. ≤ m um+1 − uh h L2 (Ω) L2 (Ω) ∆t um+1 − um h +um+θ H 1 (Ω)  h H 1 (Ω) .17). Let Wi denote the value of the piecewise linear function wh ∈ Vh at the meshpoint xi . h ∆t f m+θ (5.18).21) . m+1 m uh − uh ∆t − a um+θ .18) We shall then use this inequality to estimate the terms appearing on the righthand side of (5. wh 2 L2 (Ω) = ≥ ≥ h 6 h 6 1 3 N −1 (Wi Wi−1 + 4Wi2 + Wi Wi+1 ) i=1 N −1 i=1 N −1 1 1 2 1 1 2 − Wi2 − Wi−1 + 4Wi2 − Wi2 − Wi+1 2 2 2 2 hWi 2 .
1). um+1 h 2 L2 (Ω) 2∆t ≤ f m+θ − um h 2 L2 (Ω) L2 (Ω) um+θ h 1 12(1 + ǫ) + 1 − ∆t( − θ) · um+θ 2 1 (Ω) H h 2 h2 1 −1 m+θ 2 L2 (Ω) + ∆t( − θ)(1 + ǫ ) f L2 (Ω) .22).22) According to the Poincar´Friedrichs inequality. for any ǫ ∈ (0. H h (5. Substituting into (5. ≤ (1 + ǫ) 12 m+θ 2 u H 1 (Ω) + (1 + ǫ−1 ) f m+θ 2 h h where the inequality (a + b)2 ≤ (1 + ǫ)a2 + (1 + 1 )b2 . b ≥ 0. a. ǫ ∈ (0.21). where ǫ is a ﬁxed real number. has been ǫ applied.23) Substituting (5. 2 (5. 6(1 − 2θ) θ ∈ [0.3. 2 h H 1 f m+θ 8ǫ2 and therefore. 2 h ≤ um h 2 L2 (Ω) + ∆t m+θ f 4ǫ2 2 L2 (Ω) + ∆t2 (1 − 2θ)(1 + ǫ−1 ) f m+θ 2 L2 (Ω) .5. . 1). um+1 − um h h ∆t 2 L2 (Ω) 103 ≤ √ 12 m+θ uh H 1 (Ω) + f m+θ h 2 L2 (Ω) 2 L2 (Ω) . f m+θ ≤ L2 (Ω) um+θ h L2 (Ω) 1 f m+θ 8ǫ2 ≤ 2 L2 (Ω) m+θ + 2ǫ2 uh 2 L2 (Ω) 2 L2 (Ω) + ǫ2 um+θ 2 1 (Ω) . um+1 h ≤ 2 L2 (Ω) 2∆t − um h 2 L2 (Ω) + 1 − ∆t 6(1 − 2θ)(1 + ǫ) − ǫ2 um+θ 2 1 (Ω) H h h2 2 L2 (Ω) . Then 1 − ∆t so that m+1 uh 2 L2 (Ω) 6(1 − 2θ)(1 + ǫ) − ǫ2 ≥ 0. STABILITY OF θSCHEMES By (5.23) into (5. 1 f m+θ 8ǫ2 2 L2 (Ω) 1 + ∆t( − θ)(1 + ǫ−1 ) f m+θ 2 Let us suppose that ∆t ≤ h2 (1 − ǫ). ǫ > 0.16). e m+θ uh 2 L2 (Ω) 1 ≤ um+θ 2 1 (Ω) . 1/2).
Taking the square root of both sides. vh ) = 0 ∀vh ∈ Vh . 1/2). Hence.25) To summarise: when θ ∈ [1/2.12) is unconditionally stable. 5. upon summation through all m this implies that M −1 1≤k≤M max uk h 2 L2 (Ω) 0 ≤ uh 2 L2 (Ω) + cǫ m=0 ∆t f m+θ 2 L2 (Ω) . h h where η m = u(·. t) ∈ Vh follows by the LaxMilgram theorem. EVOLUTION PROBLEMS Letting cǫ = 1/(4ǫ2 ) + ∆t(1 − 2θ)(1 + ǫ−1 ). tm ) − um . the backward Euler scheme. For simplicity. We decompose the global error eh as follows: em = u(·. and the CrankNicolson scheme. the method (5. ξ m = P u(·. The forward Euler scheme. corresponding to θ = 1.104 CHAPTER 5. In particular. t). 6(1 − 2θ) 0 < ǫ < 1. When θ ∈ [0. t) ∈ Vh denotes the Dirichlet projection of u(·. P u(·. 1] the analysis is completely analogous. a(η m . t). subject to the time step limitation (5. corresponding to θ = 0. tm ). the scheme (5. t) deﬁned by a(P u(·.12) is conditionally stable in the sense that M −1 1≤k≤M 1/2 max uk h h ≤ u0 2 2 (Ω) h L + cǫ m=0 ∆t f m+θ 2 2 (Ω) L . and (5.12) for the numerical solution of the initial boundary value problem (5.25). h and for t ∈ [0.15) holds.24) provided that ∆t ≤ h2 (1 − ǫ). vh ) = a(u(·.4 Error analysis in the L2 norm In this section we investigate the accuracy of the ﬁnite element method (5. we deduce that for θ ∈ [0.12) is conditionally stable. 1]. 1/2) the scheme (5. tm ) − P u(·. corresponding to θ = 1/2. T ]. we shall restrict ourselves to the backward Euler scheme (θ = 1). (5. (5. is only conditionally stable. vh ) ∀vh ∈ Vh . The existence and uniqueness of P u(·.9). for other values of θ ∈ [0. are unconditionally stable. tm ) − um = η m + ξ m . .
29) where ϕm+1 = By (5. tm ) ∈ Vh denotes the continuous piecewise linear interpolant of u(·. vh + a(ξ m+1 . π 105 where Ih u(·. (ξ 0 . choosing vh = ξ 0 and applying the CauchySchwarz inequality on the right. h2 (5. tm ) π2 ∆t .5. tm+1 ) − u(·. M −1 1≤m≤M 1/2 L2 (Ω) ≤ h2 u(·. tm ) ∂u − (·. tm ) − Ih u(·. tm+1 ) − u(·. ηm Since also. vh ∆t = 0 ∀vh ∈ Vh . vh ∆t ∂t ∆t According to the stability result proved earlier on. tm+1 ) − . vh ) h and therefore. tm+1 ) − .28). (5. vh ) = −(η 0 .28) ξ 0 L2 (Ω) ≤ η 0 L2 (Ω) ≤ 2 u0 H 2 (Ω) . L2 (Ω) ≤ h2 u(·.4. vh ) ∆t u(·. η m+1 − η m u(·.27) max ξ m L2 (Ω) ≤ ξ 0 2 2 (Ω) L + m=0 ∆t ϕm+1 2 2 (Ω) L . tm )H 1 (Ω) ≤ h u(·. by C´a’s lemma.26) by an identical argument we deduce that η m+1 − η m ∆t For m = 0. π2 (5. tm ) from Vh . tm+1 ) − u(·. By the AubinNitsche duality argument. π2 (5. tm ) ∂u η m+1 − η m = − (·. tm )H 2 (Ω) . ERROR ANALYSIS IN THE L2 NORM and therefore. a η m+1 − η m . ∆t ∂t ∆t ξ0 ≤ h2 u0 H 2 (Ω) . e η m H 1 (Ω) ≤ u(·. tm )H 2 (Ω) .30) L2 (Ω) . vh ) = (e0 . vh ) − (η 0 . π It is easily seen that ξ m ∈ Vh satisﬁes the following identity: ξ m+1 − ξ m . H 2 (Ω) (5.
H 2 (Ω) 2 h2 1 = 2 π ∆t 1/2 tm+1 tm ∂u (·. t) ∂t2 2 L2 (Ω) dt . For θ ∈ (1/2. independent of h and ∆t. is beyond the scope of these notes. ∂t2 √ tm+1 ∆t tm ∂ 2u (·. tm ) II ≤ π2 ∆t h2 ≤ √ π 2 ∆t tm+1 tm (tm+1 − t) 1/2 ∂ 2u (x. while very similar to the onedimensional case. t) ∂t dt H 2 (Ω) . EVOLUTION PROBLEMS Now L2 (Ω) u(·. tm+1 ) − u(·.31) For term I. 1≤m≤M max u(·. t) dt ∂t H 2 (Ω) ∂u (·. tm ) ∂u − (·. Thus.30) into (5. 1]. but the exposition of that theory. by (5. but only conditionally. and depending only on norms of the analytical solution u. tm+1 ) ≤ ∆t ∂t η m+1 − η m + L2 (Ω) ≡ I + II. t) dt. tm+1 ) − u(x. Substituting the bounds on terms I and II onto (5.26). (5. tm ) − um h L2 (Ω) ≤ max 1≤m≤M ξm L2 (Ω) + max 1≤m≤M ηm L2 (Ω) . tm ) − um h L2 (Ω) ≤ C2 (h2 + ∆t). The stability and convergence results presented here can be extended to parabolic equations in more than one space dimension. For θ ∈ [0.27). tm+1 ) − u(·. .29). The CrankNicolson scheme (θ = 1/2) can be shown to converge in the norm · L2 (Ω) unconditionally. tm ) ∂u 1 − (x. Taylor’s formula with integral remainder yields that u(x.32) where C1 is a positive constant. with error O(h2 + (∆t)2 ). θ ∈ (1/2. tm+1 ) = − ∆t ∂t ∆t and therefore I≤ Further. h2 u(·.32) and (5.106 It remains to estimate ϕm+1 ϕ m+1 L2 (Ω) L2 (Ω) . 1/2) the scheme converges with error O(h2 + ∆t). by (5. where C2 is a positive constant independent of h and ∆t. ∆t tm+1 tm (5. 1] the scheme converges unconditionally with error O(h2 + ∆t). CHAPTER 5. we deduce that 1≤m≤M max u(·. we obtain the following error bound: 1≤m≤M max ξm L2 (Ω) ≤ C1 (h2 + ∆t).31) and inserting the resulting inequality and (5. But.