Theory of Elasticity

Lectures on
The Finite Element Method

By
Ph. Ciarlet
Tata Institute of Fundamental Research
Bombay
1975
Lectures on
By
Ph. Ciarlet
Notes by
S. Kesavan, Akhil Ranjan
M. Vanninathan
Tata Institute of Fundamental Research
Bombay
1975
c Tata Institute of Fundamental Research, 1975
No part of this book may be reproduced in any
form by print, microlm or any other means with-
out written permission from the Tata Institute of
Fundamental Research, Colaba, Bombay 400005
Printed In India
By
Anil D. Ved At Prabhat Printers, Bombay 400004
And Published By
The Tata Institute of Fundamental Research, Bombay.
Acknowledgements
THE AUTHOR EXPRESSES his deep gratitude to Professor Ramana-
than for his very kind invitation to deliver these lectures at Bangalore,
as part of the T.I.F.R.I.I.Sc. Mathematics Programme. Also, the author
tanks the T.I.F.R. for nancial support.
Ph. Ciarlet
v
Preface
OUR BASIC AIM has been to present some of the mathematical as-
pects of the nite element method, as well as some applications of the
nite element method for solving problems in Elasticity. This is why
important topics, such as curved boundaries, mixed and hybrid meth-
ods, time-dependent problems, etc..., are not covered here. No attempt
has been made to give an exhaustive bibliography.
vii
Contents
Acknowledgements v
Preface vii
1 The Abstract Problem 1
2 Examples 9
3 The Finite Element Method in its Simplest Form 29
4 Examples of Finite Elements 35
5 General Properties of Finite Elements 53
6 Interpolation Theory in Sobolev Spaces 59
7 Applications to Second-Order Problems... 67
8 Numerical Integration 77
9 The Obstacle Problem 95
10 Conforming Finite Element Method for the Plate Problem 103
11 Non-Conforming Methods for the Plate Problem 113
ix
Chapter 1
The Abstract Problem
SEVERAL PROBLEMS IN the theory of Elasticity boil down to the 1
solution of a problem described, in an abstract manner, as follows:
Let V be a normed linear space over R. Let J : V R be a func-
tional which can be written in the form
(1.1) J(v) =
1
2
a(v, v) f (v) for all v V,
where a(, ) is a continuous, symmetric bilinear form on V and f is an
element of V
, the dual of V. Then the problem consists in nding an

element u V such that
(1.2) J(u) = Min
vV
J(v).
Usually J represents the energy of some physical system.
More often, instead of minimising J over the entire space V, we do
so over a non-empty convex subset K of V and nd a element u K
such that
(1.3) J(u) = Min
vK
J(v).
Henceforth we shall denote this abstract problem by the symbol (P).
One can ask immediately whether this problem admits of a solution and
if so, is the solution unique? We present in this section the essential
results regarding existence and uniqueness.
1
2 1. The Abstract Problem
Denition 1.1. Let V be a normed linear space. A bilinear form a(, )
on V is said to be V-elliptic if there exists a constant > 0 such that for
all v V.
(1.4) a(v, v) v
2
.
2
Theorem 1.1. Let V be a Banach space and K a closed convex subset
of V. Let a(, ) be V-elliptic. Then there exists a unique solution for the
problem (P).
Further this solution is characterised by the property:
(1.5) a(u, v u) f (v u) for all v K.
Remark 1.1. The inequalities (1.5) are known as variational inequali-
ties.
Proof. The V-ellipticity of a(, ) clearly implies that if a(v, v) = 0 then
v = 0. This together with the symmetry and bilinearity of a(, ) shows
that a(, ) denes an inner-product on V. Further the continuity and the
V-ellipticity of a(, ) shows that the norm
(1.6) v V a(v, v)
1
2
dened by the inner-product is equivalent to the existing norm on V.
Thus V acquires the structure of a Hilbert space and we apply the Riesz
representation theorem to obtain the following: for all f V
, there
exists f V such that
(1.7) f (v) = a(f , v) for all v V.
The map : V
V given by f f is linear. Now,

J(v) =
1
2
a(v, v) f (v)
=
1
2
a(v, v) a(f , v)
1. The Abstract Problem 3
=
1
2
a(v f , v f )
1
2
a(f , f ).
The symmetry of a(, ) is essential in obtaining the last equality.
For a given f , since f is xed, J is minimised if and only if a(v 3
f , v f ) is minimised. But this being the distance between v and
f , our knowledge of Hilbert space theory tells us that since K is a
closed convex subset, there exists a unique element u K such that this
minimum is obtained. This proves the existence and uniqueness of the
solution, which is merely the projection of f over K.
We know that this projection is characterised by the inequalities:
(1.8) a(f u, v u) 0 for all v K.
Geometrically, this means that the angle between the vectors (f u)
and (v u) is obtuse. See Fig. 1.1.
Figure 1.1:
Thus, a(f , v u) a(u, v u) which by virtue of (1.7) is precisely
the relation (1.5). This completes the proof.
We can state the following
Corollary 1.1. (a) If K is a non-empty closed convex cone with ver-
tex at origin 0, then the solution of (P) is characterised by:
(1.9)
_
_
a(u, v) f (v) for all v K
a(u, u) = f (u).
(b) If K is a subspace of V then the solution is characterised by 4
(1.10) a(u, v) = f (v) for all v K.
Remark 1.2. The relations (1.5), (1.9) and (1.10) are all called varia-
tional formulations of the problem (P).
Proof. (a) If K is a cone with vertex at 0, then for u, v K, u+v K.
(cf. Fig. 1.2). If u is the solution to (P), then for all v K applying
(1.5) to (u + v) we get a(u, v) f (v) for all v K. In particular
this applies to u itself. Setting v = 0 in (1.5) we get a(u, u)
f (u) which gives the reverse inequality necessary to complete
the proof of (1.9). Conversely, if (1.9) holds, we get (1.5) by just
subtracting one inequality from the other.
Figure 1.2:
(b) Applying (a) to K, since any subspace is a cone with vertex at 0,
we get (b) immediately. For if v K, then v K and applying
(1.9) both to v and v we get (1.10).
This completes the proof.
Remark 1.3. The solution u of (P) corresponding to f V
(for a xed
a(, )) denes a map V
V. Since this solution is the projection of

f on K, it follows that the above map is linear if and only if K is
a subspace. The problems associated with variational inequalities are
therefore nonlinear in general.
Exercise 1.1. Let V be as in Theorem 1.1. For f
1
, f
2
V
, let u
1
, u
2
5
be the corresponding solutions of (P). If
denotes the norm in V
,
prove that
u
1
u
2

1
f
1
f
2
.
Remark 1.4. The above exercise shows, in particular, the continuous
dependence of the solution of f , in the sense described above. This
together with the existence and uniqueness establishes that the problem
(P) is well-posed in the terminology of partial dierential equations.
Exercise 1.2. If V is a normed linear space, K a given convex subset of V
and J : V R any functional which is once dierentiable everywhere,
then (i) if u K is such that J(u) = Min
vK
J(v), u satises, J
(u)(vu) 0
for all v K. (ii) Conversely, if u K such that J
(u)(v u) 0 for
all v K, and J is everywhere twice dierentiable with J
satisfying
J
(v)(w, w) w
2
, for all v, w K and some 0, then J(u) =
Min
vK
J(v).
Exercise 1.3 (
1
). Apply the previous exercise to the functional
J(v) =
1
2
a(v, v) = f (v)
with a(, ) and f as in Theorem 1.1. If K is a subspace of V, show that
J
(u)(v) = 0 for all v K. In particular if K = V, J
(u) = 0.
It was essentially the symmetry of the bilinear form which provided
the Hilbert space structure in Theorem 1.1. We now drop the symmetry
assumption on a(, ) but we assume V to be a Hilbert space. In addition
we assume that K = V.
Theorem 1.2 (LAX-MILGRAM LEMMA). Let V be a Hilbert space. 6
1
Exercises (1.2)(i) and 1.3 together give relations (1.5)
a(, ) a continuous, bilinear, V-elliptic form, f V
. If (P) is the prob-

lem: to nd u V such that for all v V,
(1.11) a(u, v) = f (v),
then (P) has a unique solution in V
1
.
Proof. Since a(, ) is continuous and V-elliptic, there are constants M,
> 0 such that
a(u, v) Mu v,
a(v, v) v
2
,
(1.12)
for all u, v V. Fix any u V. Then the map v a(u, v) is continuous
and linear. Let us denote it by Au V
. Thus we have a map A : V V
dened by u Au.
(1.13) Au
= sup
vV
v0
Au(v)
v
= sup
vV
v0
a(u, v)
v
Mu.
Thus A is continuous and A M.
We are required to solve the equation
(1.14) Au = f .
Let be the Riesz isometry, : V
V so that
(1.15) f (v) = (( f , v)),
where ((, )) denotes the inner product in V. Then, Au = f if and only
if Au = f or equivalently.
(1.16) u = u ( Au f ),
where > 0 is a constant to be specied. We choose such that 7
g : V V is a contraction map, where g is dened by
(1.17) g(v) = v (Av f ) for v V.
1
cf. Corollary 1.1(b).
Then the solution to (P) will be the unique xed point of this con-
traction map, which exists by the contraction mapping theorem.
Let v
1
, v
2
V. Set v = v
1
v
2
. Then
g(v
1
) g(v
2
) = (v
1
v
2
) A(v
1
v
2
)
= v Av.
But,
v Av
2
= ((v Av, v Av))
= v
2
2((Av, v)) +
2
Av
2
= v
2
2Av(v) +
2
Av
2
v
2
2v
2
+
2
M
2
v
2
= (1 2 +
2
M
2
)v
2
since Av(v) = a(v, v) v
2
and A M. Choosing ]0,
2
M
2
[, we
get that
(1.18) 1 2 +
2
M
2
< 1
and hence g is a contraction, thus completing the proof.
Remark 1.5. The problem (P) of Theorem 1.2 is well-posed. The ex-
istence and uniqueness were proved in the theorem. For the continuous
dependence of u on f , we have
(1.19) u
2
a(u, u) = f (u) f
u.
8
REFERENCE. For Variational Inequalities, see Lions and Stampac-
chia [18].
Chapter 2
Examples
WE GIVE IN this section several examples of the abstract problem for- 9
mulated in Sec. 1. We interpret the solutions of these problems as so-
lutions of classical boundary value problems which often occur in the
theory of Elasticity.
Before we proceed with the examples, we summarize briey the
results (without proofs) on Sobolev spaces which will prove to be very
useful in our discussion.
Henceforth R
n
will denote an open set (more often will
be a bounded open set with a specic type of boundary which will be
described presently). A multi-index will denote an n-tuple (
1
, . . . ,
n
)
of non-negative integers, and we denote
(2.1) =
1
+ +
n
,
and call it the length of the multi-index. If v is a real-valued function on
for which all derivatives upto order m exist, for a multi-index with
m we dene
(2.2)
v
=
1
x
1
. . .
n
x
n
.
The space of test functions on is given by
(2.3) D() =
_
v C
(); supp(v) is a compact subset of

_
.
9
10 2. Examples
where
(2.4) supp(v) = x ; v(x) 0.
Denition 2.1. Let m 0 be an integer. Then the Sobolev space H
m
() 10
is given by
(2.5) H
m
() =
_
v L
2
();
v L
2
() for all m
_
,
where all derivatives are understood in the sense of distributions.
On H
m
() one can dene a norm by means of the formula
(2.6) v
m,
=
_
m
_
v
2
dx
_
_
1
2
, v H
m
().
It is easy to check that
m,
denes a norm on H
m
(), which
makes it a Hilbert space. One can also dene a semi-norm by
(2.7) v
m,
=
_
=m
_
v
2
dx
_
_
1
2
, v H
m
().
Note that since for all m 0, D() H
m
(), we may dene,
(2.8) H
m
0
() = D(),
the closure being taken with respect to the topology of H
m
(). Since
H
m
0
() is a closed subspace of H
m
(), it is also a Hilbert space under
the restriction of the norm
m,
. We also have a stronger result:
Theorem 2.1. Assume that is a bounded open set. Then over H
m
0
()
the semi-norm
m,
is a norm equivalent to the norm
m,
.
This result is a consequence of the following:
Theorem 2.2 (POINCAR
E-FRIEDRICHS INEQUALITY). If is a
bounded open set, there exists a constant C = C() such that, for all
v H
1
0
(),
(2.9) v
0,
Cv
1,
.
2. Examples 11
Henceforth, unless specied to the contrary, the following will be 11
our standing assumptions: is a bounded open subset of R
n
. If is the
boundary of , then is Lipschitz continuous in the sense of Ne cas [20].
(Essentially, can be covered by a nite number of local coordinate
systems, such that in each, the corresponding portion of is described
by a Lipschitz continuous function).
If L
2
() is dened in the usual fashion using the Lipschitz continuity
of , one has the following result:
Theorem 2.3. There exists a constant C = C() such that, for all v
C
(),
(2.10) v
L
2
()
Cv
1,
.
By virtue of Theorem 2.3 we get that if v C
(), then its restric-

tion to is an element of L
2
(). Thus we have a map from the space
C
() equipped with the norm

1,
into the space L
2
() which is
continuous. We also have:
Theorem 2.4. The space C
() is dense in H
1
(), for domains with
Lipschitz continuous boundaries.
Consequently, the above map may be extended to a continuous map
H
1
() L
2
() which we denote by tr
. It is called the trace operator.

An important result on the trace is the characterization:
(2.11) H
1
0
() =
_
v H
1
(); tr
v = 0
_
.
When no confusion is likely to occur we will merely write v instead
of tr
v. In fact if v is a smooth function then tr
v is the restriction of
v to .
Retaining our assumption on and , the unit outer normal

is 12
dened a.e. on . Let

= (
1
, . . . ,
n
). If v is smooth then we may
dene the outer normal derivative
v
by
(2.12)
v
=
n
i=1
i
v
x
i
.
12 2. Examples
We extend this denition to v H
2
(). If v H
2
(), then
v
x
i
H
1
() and hence tr
v
x
i
L
2
(). We now dene
(2.13)
v
=
n
i=1
i
tr
v
x
i
.
However when there is no confusion we write it in the form of
(2.12). Then one has the following characterization:
(2.14) H
2
0
() =
_
v H
2
(); v =
v
= 0 on
_
.
Theorem 2.5 (GREENS FORMULA IN SOBOLEV SPACES). Let u,
v H
1
(). Then we have
(2.15)
_
u
v
x
i
dx =
_
u
x
i
v dx +
_
u v
i
d,
for all 1 i n.
If we assume u H
2
(), we may replace u in (2.15) by
u
x
i
; sum-
ming over all 1 i n, we get for u H
2
(), v H
1
(),
(2.16)
_
i=1
u
x
i
v
x
i
dx =
_
u v dx +
_
v d,
where =
_
n
i=1
2
x
2
i
is the Laplacian.
If both u and v are in H
2
(), we may interchange the roles of u and
v in (2.16). Subtracting one formula from the other, we get
(2.17)
_
(uv uv)dx =
_
_
u
v
v
_
d,
for u, v H
2
(). 13
2. Examples 13
Finally replacing u by u in (2.17) we get, for u H
4
(), v
H
2
(),
(2.18)
_
uv dx =
_
2
u v dx +
_
u
v
d
_
(u)
vd.
The formulae (2.15) through (2.18) are all known as Greens formu-
lae in Sobolev spaces.
We derive two results from these formulae. These results will be
useful later.
Lemma 2.1. For all v H
2
0
(),
(2.19) v
0,
= v
2,
.
Consequently over H
2
0
(), the mapping v v
0,
is a norm equiv-
alent to the norm
2,
.
Proof. Since D() is dense in H
2
0
(), it suces to prove (2.19) for
v (). Let v D(). Then
v
2
=
n
i=1
_
2
v
x
2
i
_
_
2
+ 2
1i<jn
2
v
x
2
i
2
v
x
2
j
.
By Greens formula (2.15),
(2.20)
_
2
v
x
2
i
2
v
x
2
j
dx =
_
v
x
i
3
v
x
i
x
2
j
dx =
_
_

2
v
x
i
x
j
_
2
dx
for, the integrals over vanish for v D(). (cf. (2.11)). Now (2.19)
follows directly from (2.20). This proves the lemma.
Lemma 2.2. Let R
2
. Then for u H
3
(), v H
2
(), 14
_
2

2
u
x
1
x
2
2
v
x
1
x
2
2
u
x
2
1
2
v
x
2
2
2
u
x
2
2
2
v
x
2
1
dx
=
_
2
u
2
v
2
u
_
d,
(2.21)
where

denotes the tangential derivative.

14 2. Examples
Proof. Let

= (
1
,
2
),

= (
1
,
2
) be the unit vectors along the outer
normal and the tangent respectively. Without loss in generality we may
assume
1
=
2
,
2
=
1
. Also note that
2
1
+
2
2
= 1. The second
derivatives occurring in the right hand side are dened by
(2.22)
_
2
u
2
=
2
u
x
2
1
2
1
+ 2

2
u
x
1
x
2
2
+
2
u
x
2
2
2
2
2
u
2
u
x
2
1
1
+
2
u
x
1
x
2
(
1
2
+
2
1
) +
2
u
x
2
2
2
.
Using all these relations we get
2
u
2
v
2
u
=
_
2
u
x
1
x
2
v
x
2
2
u
x
2
2
v
x
1
_
1
+
_
2
u
x
1
x
2
v
x
1

2
u
x
2
1
v
x
2
_
2
=

X

,
(2.23)
where,

X = (X
1
, X
2
) and
(2.24)
_
_
X
1
=
2
u
x
1
x
2
v
x
2
2
u
x
2
2
v
x
1
X
2
=
2
u
x
1
x
2
v
x
1
2
u
x
2
1
v
x
2
Also note that,
(2.25) div X =
X
1
x
1
+
X
2
x
2
= 2

2
u
x
1
x
2
2
v
x
1
x
2

2
u
x
2
1
2
v
x
2
2

2
u
x
2
2
2
v
x
2
1
.
15
Now by Greens formula (2.15) applied to functions v
i
H
1
() and
to the constant function 1,
_
v
i
x
i
dx =
_
v
i
i
d.
2. Examples 15
Hence summing over all i, if

v = (v
1
, . . . , v
n
),
(2.26)
_
div

v dx =
_

v

d.
(This is known as the Gauss Divergence Theorem and also as the Os-
trogradskys formula). From (2.23), (2.25) and (2.26) the result fol-
lows.
With this background, we proceed to examples of the abstract prob-
lem of Sec. 1.
Example 2.1. Let K = V = H
1
0
(). Let a L
() such that a 0 a.e.

in . Let f L
2
(). Dene the bilinear form a(, ) and the functional
f (), by
(2.27)
_
_
a(u, v) =
_
_
_
n
i=1
u
x
i
v
x
i
+ auv
_
dx,
f (v) =
_
f v dx.
The continuity of a(, ) and f () follows from the Cauchy-Schwarz
inequality. For instance
(2.28) f (v) f
0,
v
0,
f
0,
v
1,
.
We now show that a(, ) is V-elliptic.
a(v, v) =
_
_
n
i=1
_
v
x
i
_
2
+ av
2
_
_
dx
i=1
_
v
x
i
_
2
dx (since a 0 a.e. in )
= v
2
1,
.
16
Since
1,
is equivalent to
1,
over V, this proves the V-ellipticity.
Hence by our results in Sec. 1 there exists a unique function u V such
that a(u, v) = f (v) for all v V.
16 2. Examples
Interpretation of this problem: Using the above equation satised by
u, we get
(2.29)
_
_
n
i=1
u
x
i
v
x
i
+ auv
_
_
dx =
_
f v dx, for all v H

1
0
().
From the inclusion D() H
1
0
(), we get that u satises the equa-
tion u + au = f in the sense of distributions.
If we assume that u is suciently smooth, for example u H
2
(),
then we may apply Greens formula (2.16), which gives
(2.30)
_
auv dx
_
uv dx +
_
v d =
_
f v dx.
Since v H
1
0
(), tr
v = 0. Hence the integral over vanishes.

Thus we get
(2.31)
_
(u + au f )v dx = 0 for all v H
1
0
().
Varying v over H
1
0
(), we get that u satises the equation u+au =
f in . Further since u H
1
0
(), we get the boundary condition u = 0
on . Thus we may interpret u as the solution of the classical boundary
value problem:
(2.32)
_
_
u + au = f in ,
u = 0 on .
17
This is known as the homogeneous Dirichlet problem for the opera-
tor u + au.
A particular case of this equation arises in the theory of Elasticity,
for which R
2
and a = 0. Thus u = f in and u = 0 on . This
corresponds to the membrane problem:
Consider an elastic membrane stretched over and kept xed along
. Let Fdx be the density of force acting on an element dx of . Let
u(x) be the vertical displacement of the point x R
2
, measured
2. Examples 17
in the x
3
-direction from the (x
1
, x
2
)-plane. If t is the tension of the
membrane, then u is the solution of the problem
(2.33)
_
_
u = f in
u = 0 on
where f = F/t.
Figure 2.1:
Remark 2.1. To solve the problem (2.32) by the classical approach, one
needs hard analysis involving Schauders estimates. By the above pro-
cedure viz. the variational method, we have got through with it more
easily.
The above problem is a typical example of a second-order problem. 18
Exercise 2.1. The obstacle problem. Let , be as in example (2.1).
Let X be an obstacle in this region. Let X 0 on . Let F dx be the
density of the force acting on a membrane stretched over , xed along
. The displacement u(x) at x in the vertical direction is the solution of
the following problem:
If V = H
1
0
(), K =
_
v H
1
0
(); v X a.e. in
_
,
18 2. Examples
let a(u, v) =
_
i=1
u
x
i
v
x
i
dx, f (v) =
_
f v dx, f L
2
(),
X H
2
(), X 0 on .
Show that K is a closed convex set and hence that this problem ad-
mits of a unique solution. Assuming the regularity result u H
2
()
H
1
0
(), show that this problem solves the classical problem,
_
_
u X in ,
u = f when u > X, ( f = F/t)
u = 0
(We will discuss the Obstacle Problem in Sec. 9).
Figure 2.2:
Exercise 2.2. Let V = H
1
(). Dene a(, ), f () as in example 2.1. 19
Assume further that there exists a constant a
0
such that a a
0
> 0 in .
If u
0
is a given function in H
1
(), dene
K =
_
v H
1
(); v u
0
H
1
0
()
_
=
_
v H
1
(); tr
v = tr
u
0
_
.
2. Examples 19
Check that K is a closed convex subset. Interpret the solution to be
that of the Non-homogeneous Dirichlet problem,
_
_
u + au = f in .
u = u
0
on .
1
(). Let a L
() such that a a
0
> 0,
f L
2
(). Dene a(, ) and f () as in example (2.1). The continuity
of a(, ) and f () follow as usual. For the V-ellipticity, we can no longer
prove it with the semi-norm
1,
as we did earlier. It is here we use the
additional assumption on a, since
a(v, v) =
_
_
n
i=1
_
v
x
i
_
2
+ av
2
_
_
dx
min(1, a
0
)v
2
1,
.
Thus we have a unique solution u to the abstract problem satisfying
a(u, v) = f (v). If we assume again that u is suciently smooth to
apply the Greens formula (2.16), we get
(2.34)
_
(u + au)v dx +
_
v d =
_
f v dx.
If v D(), then the integral over will vanish. Thus u satises
the equation u + au = f as in Example 2.1
1
. However we now get a 20
dierent boundary condition. In example (2.1) the boundary condition
was built in with the assumption u V = H
1
0
(). Now from (2.34), we
may write:
(2.35)
_
(u + au f )v dx =
_
v d
for all v H
1
().
But the left hand side of (2.35) is zero since u satises the dieren-
tial equation as above so that for all v H
1
(),
_
v d = 0. Thus
1
As in Example 2.1, the equation u + au = f is always satised in the sense of
distributions since D() H
1
().
20 2. Examples
u
= 0 on , and we may interpret this problem as the classical prob-

lem:
(2.36)
_
_
u + au = f in
u
= 0 on .
This is a homogeneous Neumann problem.
Exercise 2.3. With K, V, a(, ) as in example 2.2., dene
f (v) =
_
f v dx +
_
gv d,
_
_
f L
2
(),
g L
2
().
Show that the abstract problem leads to a solution of the non-homo-
geneous Neumann problem
_
_
u + au = f in
u
= g on .
Remark 2.2. In these examples one may use the more general bilinear 21
from dened by
(2.37) a(u, v) =
_
_
n
i, j=1
a
i j
u
x
j
v
x
i
+ auv
_
_
dx,
where the functions a
i j
L
() satisfy the condition that for some

> 0,
(2.38)
n
i, j=1
a
i j
j

n
i=1
2
i
for all R
n
and a.e. in . This is the classical ellipticity condition for
second order partial dierential operators. One should check (exercise!)
in this case that the abstract problem leads to a solution of the boundary
value problem
(2.39)
n
i, j=1
x
i
_
a
i j
u
x
j
_
+ au = f in
2. Examples 21
with the boundary condition
(2.40)
_
_
u = 0 on if K = V = H
1
0
(),
n
_
i, j=1
a
i j
u
x
j
i
= 0 on if K = V = H
1
().
The latter boundary operator in (2.40) is called the conormal deriva-
tive associated with the partial dierential operator,
i, j=1
x
i
_
a
i j
x
j
_
.
Notice that the term (+ au) contributes nothing.
Example 2.3. System of Elasticity. Let R
3
, with Lipschitz con-
tinuous boundary . Further assume that can be partitioned into two
portions
0
and
1
such that the d-measure of
0
is > 0. Let
K = V =
_

v = (v
1
, v
2
, v
3
); v
i
H
1
(), 1 i 3 and

v =

0 on
0
_
.
Dene 22
(2.41)
_
i j
(

v ) =
1
2
_
v
i
x
j
+
v
j
x
i
_
,
i j
(

v ) =
_
_
3
k=1

kk
(

v )
_
i j
+ 2
i j
(

v ),
for 1 i, j 3. The latter relation is usually known as Hookes law.
The constants ( 0) and (> 0) are known as Lames coecients. We
dene the bilinear form a(, ) by,
a(

u ,

v ) =
_
i, j=1
i j
(

u
i j
(

v )dx
=
_
( div

v + 2
3
i, j=1
i j
(

u )
i j
(

v )dx.
(2.42)
Let

f = ( f
1
, f
2
, f
3
), f
i
L
2
(), and

g = (g
1
, g
2
, g
3
), g
i
L
2
(), be
given.
22 2. Examples
Dene the linear functional f (), by,
(2.43) f (

v ) =
_

f

v dx +
_
1

g

v d.
The continuity of a(, ) and f () follow from the Cauchy-Schwarz
inequality. For the V-ellipticity of a(, ) one uses the inequality
a(

v ,

v ) 2
_
i, j=1
(
i j
(

v ))
2
dx
and the fact that the square root of the integral appearing in the right
hand side of the above inequality is a norm over the space V, equivalent
to the norm

v = (v
1
, v
2
, v
3
) (
_
3
i=1
v
i
2
1,
)
1
2
. This is a nontrivial fact
which uses essentially the fact that
0
has measure > 0 and an inequality
known as K orns inequality. We omit the proof here.
Again the problem a(

u ,

v ) = f (

v ) admits a unique solution. As-
suming sucient smoothness, we may apply Greens formula: 23
_
i, j=1
i j
(

u )
i j
(

v )dx =
1
2
_
i, j=1
i j
(

u )
_
v
i
x
j
+
v
j
x
i
_
dx
=
_
i, j=1
i j
(

u )
v
i
x
j
dx
(since
i j
(v) is symmetric in i and j)
=
_
i, j=1
x
j
(
i j
(

u ))v
i
dx +
_
1
3
i, j=1
i j
v
i
j
d.
Thus, the abstract problem leads to a solution of
(2.44)
_
3
_
j=1
x
j
(
i j
(

u )) = f
i
(1 i 3) in

u =

0 on
0
and
3
_
j=1
i j
(

u )
j
= g
i
on
1
(1 i 3).
2. Examples 23
Note also that
j=1
x
j
(
i j
(

u )) =
3
j=1
x
j
_
k=1
kk
(

u )
i j
+ 2
i j
(

u )
_
_
=
3
j=1
x
j
_
k=1
u
k
x
k
_
i j

3
j=1
x
j
_
2
_
u
i
x
j
+
u
j
x
i
__
= ( + )(grad div

v )
i
u
i
.
Thus the rst equation of (2.44) is equivalent to
(2.45)

u ( + ) grad div

u =

f in .
The equations (2.44) constitute the system of linear Elasticity.
Figure 2.3:
24
If we have an elastic three-dimensional body xed along
0
, acted
on by an exterior force of density f dx and force of density g d along
1
and if
i j
is the stress tensor, the displacement u satises (2.44); cf.
Fig. 2.3.
The relation a(

u ,

v ) = f (

v ), viz.,
(2.46)
_
i, j=1
i j
(

u )
i j
(

v )dx =
_

f

v dx +
_
1

g

v d
24 2. Examples
for all

v V is known as the principle of virtual work. The tensor
i j
is the strain tensor and the tensor
i j
the stress tensor. The expression
1
2
a(

v ,

v ) is the strain energy, and the functional f (

v ) is the potential
energy of exterior forces. This example is of fundamental importance in
that the nite element method has been essentially developed for solving
this particular problem or some of its special cases (membranes, plates,
shells, etc.,) and generalizations (nonlinear elasticity, etc. . . ).
2. Examples 25
Remark 2.3. The above problems are all examples of linear problems
2
: 25
The map from the right-hand side of the equation and of the boundary
conditions to the solution u is linear. The non-linearity may occur in
three ways:
(i) When K is not a subspace of V. (e.g. Exercises 2.1 and 2.4);
(ii) If in Example 2.3 we have, instead of the rst equality in (2.41):
i j
(

v ) =
1
2
_
_
_
v
i
x
j
+
v
j
x
i
_
+
3
k=1
v
k
x
i
v
k
x
j
_
_
.
This is the case for instance when one derives the so-called Von
Karmanns equations of a clamped plate;
(iii) We may replace Hookes law (the second relations in (2.41)) by
non-linear equations connecting
i j
and
i j
, which are known as
non-linear constitutive equations. (e.g. Henckys law).
Exercise 2.4. Let V = H
1
(), and a(, ) and f () be as in example 2.2,
and let
K =
_
v H
1
(); v 0 a.e. on
_
.
Show that K is a closed convex cone with vertex 0. Using the results
of Sec. 1 show that the interpretation is
u + au = f in ,
u 0,
u
0, u
u
= 0 on .
(This is called the SIGNORINI problem).
We now examine fourth-order problems.
2
0
(). Dene
(2.47)
_
_
a(u, v) =
_
uv dx,
f (v) =
_
f v dx, f L
2
(),
2
Except in Exercise 2.1.
26 2. Examples
for all u, v V. The continuity follows as usual. For the V-ellipticity 26
of a(, ) we have,
(2.48) a(v, v) =
_
(v)
2
dx = v
2
0,
= v
2
2,
.
(by Lemma 2.1. Since
2,
and
2,
are equivalent on H
2
0
(), the
V-ellipticity follows from (2.48).
Hence there exists a unique function u H
2
0
() such that
_
uv dx =
_
f v dx for all v H
2
0
().
Assuming u to be suciently smooth (say, u H
4
()), then by
Greens formula (2.18),
(2.50)
_
(
2
u f )v dx =
_
(u)
v d
_
u
v
d,
for all v H
2
0
(). Hence by varying v over H
2
0
(), we get that u satises
2
u = f in . Since u H
2
0
(), the boundary conditions are given by
(2.14). Thus we interpret this problem as the classical problem
(2.51)
_
2
u = f in ,
u =
u
= 0 on .
This is the homogeneous Dirichlet problem for the operator
2
.
When n = 2, this is an important problem in Hydrodynamics. Here
u is known as the stream function and u is the vorticity.
A slight modication of a(, ) leads to an important problem in Elas- 27
ticity. Again let n = 2. Let f L
2
() and if K = V = H
2
0
(), dene
(2.52)
_
_
f (v) =
_
f v dx,
a(u, v) =
_
_
u v + (1 )
_
2

2
u
x
1
x
2
2
v
x
1
x
2
2
u
x
2
1
2
v
x
2
2
2
u
x
2
2
2
v
x
2
1
__
dx.
The integrand occurring in the denition of a(u, v) may also be writ-
ten as
(2.53) uv + (1 )
_
2
u
x
2
1
2
v
x
2
1
+
2
u
x
2
2
2
v
x
2
2
+
2
2
u
x
1
x
2
2
v
x
1
x
2
_
_
.
2. Examples 27
Usually, from physical considerations, 0 < <
1
2
.
Note that
(2.54) a(v, v) = v
2
0,
+ (1 )v
2
2,
by (2.53) and this leads to the V-ellipticity of a(, ). By virtue of (2.21)
in Lemma 2.2, we get that the relations a(u, v) = f (v) read as
(2.55)
_
2
uv dx =
_
f v dx,
assuming sucient smoothness of u. Thus again we get the same equa-
tion as in (2.51). Notice that the additional term in the denition of a(, )
has contributed nothing towards the dierential equation.
This latter problem is known as the clamped plate problem:
Consider a plate of small thickness e lying on the x
1
x
2
-plane. Let
E be its Youngs modulus and its Poisson coecient. Let there be a
load F acting on the plate. The displacement u is the solution of (2.51),
where f is given by (cf. Fig. 2.4):
(2.56) F =
Ee
3
f
12(1
2
)
.
28
Figure 2.4:
We will return to this problem in Sections 10 and 11.
28 2. Examples
Exercise 2.5. Let K = V = v H
2
(); v = 0 on = H
2
()H
1
0
(),
and dene a(, ) and f () as in the case of the clamped plate. Assuming
the V-ellipticity of a(, ) show that the solution of the abstract problem
satises
2
u = f in and u = 0 on . What is the other boundary
condition? This is known as the problem of the simply supported plate.
Exercise 2.6. Let K = V = H
2
() H
1
0
(), and
a(u, v) =
_
uv dx;
f (v) =
_
f v dx
_
d, where f L
2
(), L
2
().
Show that we may apply the result of Sec. 1 and give an interpreta-
tion of this problem.
REFERENCES. For details on Sobolev spaces, see Ne cas [20] and
Lions and Magenes [17]. For the theory of Elasticity, one may refer to
Duvaut and Lions [10] and Landau and Lipschitz [14].
Chapter 3
in its Simplest Form
MAINTAININGOUR ASSUMPTIONS as in the Lax-Milgram Lemma 29
(Theorem 1.2 we concentrate our attention on the following problem
(P):
(P): To nd u V such that a(u, v) = f (v) for all v V.
Let V
h
be a nite-dimensional subspace of V. Then we may state
the following problem:
(P
h
): To nd u
h
V
h
such that a(u
h
, v
h
) = f (v
h
) for all v
h
V
h
.
V
h
, being a nite-dimensional subspace, is a Hilbert space for the
norm of V. Hence by Theorem 1.2, u
h
exists and is unique. We try
to approximate the solution u of (P) by means of solutions u
h
of the
problem (P
h
) for various subspaces V
h
. This is known as the internal
approximation method.
As a rst step in this direction, we prove a most fundamental result:
Theorem 3.1. There exists a constant C, which is independent of V
h
,
such that
(3.1) u u
h
C inf
v
h
V
h
u v
h
.
29
30 3. The Finite Element Method in its Simplest Form
Proof. If w
h
V
h
, then
a(u, w
h
) = f (w
h
) = a(u
h
, w
h
).
Thus for all w
h
V
h
(3.2) a(u u
h
, w
h
) = 0.
Using this and the V-ellipticity of a(, ), we get, for all v
h
V
h
,
u u
h
2
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
) by (3.2)
Mu u
h
u v
h
.
30
Hence, u u
h
M/ u v
h
for all v
h
V
h
. Setting C = M/
and taking inmum on the right-hand side the result follows.
The above result estimates the error in the solution of (P) when
instead we solve (P
h
). To get an upper bound for the error, we only need
to compute inf
v
h
V
h
u v
h
which is the distance of u from the subspace
V
h
. This is a problem in approximation theory.
Remark 3.1. If a(, ) is also symmetric then we observe the following:
(i) J(u
h
) = inf
v
h
V
h
J(v
h
) by Corollary 1.1 (b)
(ii) We saw that a(u u
h
, w
h
) = 0 for all w
h
V
h
. Since a(, ) is now
an inner product, we get that u
h
is the projection of u to the closed
subspace V
h
in the sense of this inner-product. Therefore,
u u
h

_
a(u u
h
, u u
h
)
_
a(u v
h
, u v
h
)
Mu v
h
for all v
h
V
h
. Hence the constant C in theorem 3.1 can be taken
to be here

M/ M/, since the continuity and V-ellipticity
imply jointly that M .
3. The Finite Element Method in its Simplest Form 31
We may now describe the nite element method (f.e.m.) in its sim-
plest terms. The method consists in making special choices for the sub-
spaces V
h
such that the solutions u
h
of the problems (P
h
) converge to
u.
We will outline the procedure for obtaining the spaces V
h
by consid- 31
ering, for example, a second-order problem.
Let V = H
1
0
() or H
1
(). Let us assume to be a polygonal domain
in R
n
. That is, is a polygon in R
n
. We then have the following step-
by-step procedure:
(i) We rst establish a nite triangulation k
h
of the domain such
that =
_
Kk
h
K. The sets K are called nite elements. If n = 2,
they will be, in general, triangles. They will be tetrahedral in n =
3 and n-simplices in any R
n
. These have the further property that
any side of a nite element K is either a portion of the boundary
or the side of an adjacent nite element. (See Fig. 3.1).
Figure 3.1:
(ii) The space V
h
is such that for each v
h
V
h
, its restriction v
h
K to
each K belongs to some nite-dimensional space P
k
of real valued
functions over K which are preassigned. In practice we choose P
K
to be a space of polynomials.
(iii) We then need inclusions such as V
h
H
1
0
() or H
1
(). We es-
tablish a simple criterion to realise this.
Theorem 3.2. If for every K = k
h
, P
K
H
1
(K), and V
h
C
0
(), then 32
V
h
H
1
(). If in addition v = 0 on for all v V
h
, then V
h
H
1
0
().
Proof. Let v V
h
. Since vK L
2
(K) for every K k
h
, it follows that
v L
2
(). Hence to complete the proof it only remains to show that for
1 i n, there exist v
i
L
2
() such that for each = D(), we have,
(3.3)
_
v
i
dx =
_
x
i
v dx. (1 i n).
Then it will follow that
v
x
i
= v
i
and hence v H
1
().
However, vK P
K
H
1
(K) implies that
(vK)
x
i
L
2
(K) for
1 i n. Let D(). Since the boundary K of any K of the tri-
angulation is Lipschitz continuous, we apply the Greens formula (2.15)
to get
(3.4)
_
K
(vK)
x
i
dx =
_
K
(vK)
x
i
dx +
_
K
(vK)
i,K
d
K
,
where d
K
is the measure on K and

K
= (
1,K
, . . . ,
n,K
) is the outer
normal on K. Summing over all the nite elements K, we get
_
v
i
dx =
Kk
h
_
K
(vK)
x
i
dx
=
_
x
i
v dx +
Kk
h
_
K
(vK)
i,K
d
K
,
(3.5)
where v
i
is the function whose restriction to each K is
(vK)
x
i
.
The summation on the right-hand side of the above equation is zero
for the following reasons:
On the boundary , since D(), the integral corresponding to
K is zero. So the problem, if any, is only on the other portions
of the boundary of each K. However, these always occur as common
boundaries of adjacent nite elements. The value of vK on the common 33
boundary of two adjacent nite elements is the same (V
h
C
0
()). But
3. The Finite Element Method in its Simplest Form 33
the outer normals are equal and opposite from orientation considera-
tions. (See Fig. 3.2).
Figure 3.2:
Hence the contributions from each K along the common boundaries
cancel one another. Thus the summation yields only zero. Hence v
i
satises (3.3) for 1 i n, and clearly v
i
L
2
(). The last part of the
theorem follows the characterization (2.11).
Exercise 3.1. If for all K k
h
, P
K
H
2
(K) and V
h
C
1
(), then
show that V
h
H
2
(). Also if v =
v
= 0 on , for all v V
h
, then
V
h
H
2
0
().
We nally describe the system of linear equations associated with
the space V
h
. Suppose w
j
; 1 j M is a basis for V
h
. Let u
h
be the
solution of (P
h
). If u
h
is given by
(3.6) u
h
=
M
j=1
u
j
w
j
.
then we have, since a(u
h
, w
i
) = f (w
i
) for 1 i M,
(3.7)
M
j=1
a(w
j
, w
i
)u
j
= f (w
i
), 1 i M.
To nd u
h
, the above system of linear equations must be solved. The
matrix for this system has for its (i, j)-coecient the value a(w
j
, w
i
). 34
Note that the symmetry of a(, ) implies the symmetry of the matrix
and the V-ellipticity says that the matrix is positive denite. In practical
computations these observations are important.
Since we have to handle the matrix of the system, it would be ideal
of course to have a diagonal matrix. We could in principle achieve this
through a Gram-Schmidt orthogonalisation procedure applied to the ba-
sis functions. However such a process is not feasible since it is highly
numerically unstable. So the best we may hope for is a matrix with a
lot of zeros in it - what is known as a sparse matrix.
For example in the problem given by
_
_
u + au = f in
u = 0 on
the (i, j)-coecient of the matrix is
(3.8) a(w
j
, w
i
) =
_
_
n
k=1
w
j
x
k
w
i
x
k
+ aw
j
w
i
_
_
dx.
The matrix will be sparse if the supports of the basis functions are
as small as possible so that their inner-products will be most often
zero. We will study subsequently methods to achieve this. This trivial
criterion extends, of course, to all types of problems.
Chapter 4
Examples of Finite Elements
WE SUMMARIZE BELOW our requirements regarding the nite el- 35
ement subspaces V
h
(i) Let R
n
be a polygonal domain. Let k
h
be a triangulation of
as in Sec. 3. Then V
h
is a nite-dimensional vector space such
that for all v V
h
, vK P
K
for every nite element K, where
P
K
is a vector space of nite dimension. Usually, P
K
is a space
of polynomials. This is of practical importance in computing the
matrix of the system. We shall see later that it is of theoretical
importance as well. Observe for the moment that if P
K
consists
of polynomials, then we automatically have that P
K
H
1
(K) or
P
K
H
2
(K).
(ii) By Theorem 3.2, V
h
C
0
() implies that V
h
H
1
() and by
Exercise 3.1 V
h
C
1
() implies that V
h
H
2
(). Thus we must
choose a proper basis for the local spaces P
K
such that these
global inclusions hold.
(iii) There must exist at least one basis w
j
of V
h
which consists of
functions with small support.
We bear these points in mind when constructing examples of nite
elements. Before we proceed we need a few denitions.
35
36 4. Examples of Finite Elements
Denition 4.1. An n-simplex is the convex hull in R
n
of (n + 1) points
a
j
n+1
j=1
such that if a
j
= (a
k j
)
n
k=1
and A is the matrix
(4.1) A =
_
_
a
11
a
12
. . . a
1,n+1
.
.
.
.
.
.
.
.
.
a
n1
a
n2
. . . a
n,n+1
1 1 . . . 1
_
_
then det A 0. 36
The above denition generalises the notion of a triangle to n dimen-
sions. Geometrically the condition det A 0 simply means that the
points a
j
n+1
j=1
do not lie in the same hyperplane. For det A is equal to,
by elementary column operations, the determinant of the matrix
_
_
(a
11
a
1,n+1
) . . . (a
1,n
a
1,n+1
)
.
.
.
.
.
.
(a
n1
a
n,n+1
) . . . (a
n,n
a
n,n+1
)
_
_
and that this is non-zero means that (a
1
a
n+1
), . . . , (a
n
a
n+1
) are
linearly independent vectors in R
n
, which is the same as saying that
a
1
, . . . , a
n+1
do not lie in the same hyperplane.
Denition 4.2. Let a
j
n+1
j=1
be (n + 1)-points in R
n
satisfying the condi-
tions of denition 4.1. The barycentric coordinates of any x R
n
with
respect to these points are numbers
j
n+1
j=1
such that
(4.2)
_
_
x =
n+1
_
j=1
j
a
j
,
1 =
n+1
_
j=1
j
.
The barycentric coordinates exist because they are merely the com-
ponents of the unique solution vector

of the system of (n + 1) linear
equations in (n + 1) unknowns given by
A

=
_

x
1
_
where

x =
_
_
x
1
.
.
.
x
n
_
_
4. Examples of Finite Elements 37
The functions
j
=
j
(x) are all ane functions of x. Also
j
(a
i
) =
i j
, where is the Kronecker symbol. 37
Remark 4.1. Given points a
j
n+1
j=1
as in the denition (4.1), the corre-
sponding n-simplex is given by
K =
_
_
x =
n+1
j=1
j
a
j
; 0
j
1;
n+1
j=1
j
= 1
_
_
.
Denition 4.3. Let k 0 be an integer. Then, P
k
is the space of all
polynomials of degree k in x
1
, . . . , x
n
.
We now proceed with examples of nite elements.
Example 4.1. The n-simplex of type (1).
Let K be an n-simplex. Let P
K
= P
1
. We dene a set
_
K
=
p(a
i
); 1 i n + 1 of degrees of freedom for p P
K
, where a
i
n+1
i=1
are the vertices of K: The set
_
K
determines every polynomial p
P
K
uniquely. For, note that dimP
K
= dim P
1
= n + 1. Consider
1
, . . . ,
n+1
P
1
, the barycentric coordinate functions. These are lin-
early independent since
_
k
= 0 implies that its value at each vertex
is zero. Since
k
(a
j
) =
k j
we get that
j
= 0 for all j. Thus these
functions form a basis for P
1
. Let us write
p =
n+1
i=1
i
.
Then
p(a
j
) =
n+1
i=1
i
(a
j
) =
n+1
i=1
i j
=
j
.
Thus,
(4.4) p =
n+1
i=1
p(a
i
)
i
.
Let K be an n-simplex with vertices a
i
n+1
i=1
. Let a
i j
(i < j) be the 38
mid-points of the line joining a
i
and a
j
, i.e. a
i j
=
1
2
(a
i
+ a
j
).
Figure 4.1:
Let P
K
= P
2
. We dene for p P
2
, the set
_
K
= p(a
i
), 1 i
n + 1; p(a
i j
), 1 i < j n + 1 of degrees of freedom. Again
_
K
determines p P
2
completely. To see this note that dim P
K
=
_
n+2
2
_
and
there are as many functions in the set
i
(2
i
1), 1 i n+1;
i
j
, 1
i j n + 1. There are all functions in P
2
. Further since
i
(a
j
) =
i j
,
i
(a
k j
) =
_
_
1
2
if i = k or j,
0 otherwise,
we see again that these are linearly independent in P
2
. Let us write
p =
n+1
i=1
i
(2
i
1) +
1i<jn+1
i j
j
.
Then
p(a
k
) =
n+1
i=1
ik
(2
ik
1) =
k
.
Further,
p(a
kl
) =
n
i=1
i
(2
2
i
(a
kl
)
i
(a
kl
))
+
1i<jn+1
i j
i
(a
kl
)
j
(a
kl
).
But since
i
(a
kl
) = 0 or
1
2
, the rst sum
n
_
i=1
is zero. Further, 39
i
(a
kl
)
j
(a
kl
) =
_
_
1
4
if (i, j) = (k, l) or (l, k),
0 otherwise.
Hence
kl
= 4p(a
kl
). Thus we have
(4.5) p =
n+1
i=1
i
(2
i
1)p(a
i
) +
1i<jn+1
4
i
j
p(a
i j
).
Let K be an n-simplex with vertices a
i
n+1
i=1
. Let a
ii j
=
2a
i
+ a
j
3
,
i j. Let a
i jk
=
a
i
+ a
j
+ a
k
3
for i < j < k.
Figure 4.2:
Set P
K
= P
3
. Dene the set of degrees of freedom
K
=
_
p(a
i
), 1 i n + 1; p(a
ii j
), 1 i j n + 1;
p(a
i jk
), 1 i < j < k a + 1
_
.
Note that
i
(a
ii j
) =
2
3
;
j
(a
ii j
) =
1
3
;
1
(a
ii j
) = 0 if 1 i, 1 j;
1
(a
i jk
) =
1
3
if 1 = i, j or k and 0 otherwise, etc. Using these, one
checks the linear independence of the functions
_
i
(3
i
1)(3
i
2), 1 i n + 1;
i
j
(3
i
1), 1 i j n + 1;
k
1 i < j < k n + 1
_
.
These then form a basis for P
3
, for there are as many functions in the
above collection as dim P
K
. Using the values of
i
at the special points
described above, we get
p =
n+1
i=1
i
(3
i
1)(3
i
2)
2
p(a
i
)
+
1ijn+1
9
2
j
(3
i
1)p(a
ii j
)
1i<j<kn+1
27
i
k
p(a
i jk
).
(4.6)
40
Thus
_
K
completely determines p P
3
.
The points of K at which the polynomials are evaluated to get
_
K
are known as the nodes of the nite element. The set
_
K
is the set of
degrees of freedom of the nite element.
Exercise 4.1. Generalize these ideas and describe the n-simplex of type
(k) for any integer k 1.
We now show how these nite elements may be used to dene the
space V
h
.
First of all we show the inclusion V
h
C
0
(). Consider for instance
a triangulation by n-simplices of type (1). Number all the nodes of the
triangulation by b
j
. Let us dene
_
h
= p(b
j
); b
j
is a node.: This is
the set of degrees of freedom of the space V
h
: A function v in the space
V
h
is, by denition, determined over each K k
h
by the values v(b
j
) for
those nodes b
j
which belong to K.
Let us examine the two-dimensional case, for simplicity. If K
1
and
K
2
are two adjacent triangles with common side K
(cf. Fig. 4.3), we

need to show that vK
1
= vK
2
along K
for any v V
h
. Let t be an
abscissa along K
.
Figure 4.3:
41
Now vK
1
along K
is a polynomial of degree 1 in t. So is vK
2
along
K
. But these two agree at the nodes b

1
and b
3
. Therefore, they must be
identical and hence the continuity of v follows.
This argument can be extended to any simplex of type (k). These
simplices, by Theorem 3.2 yield the inclusion V
h
H
1
() and hence
we may use them for second order problems.
Exercise 4.2. The triangle of type (3
).
Let K be a triangle in R
2
. Dene
_
K
to be the values of p at the
points a
i
, 1 i 3, and the points a
ii j
, 1 i j 3. If we dene
P
3
= p P
3
; 12p(a
123
) + 2
_
3
i=1
p(a
i
) 3
_
ij
p(a
ii j
) = 0, then show
that
_
K
uniquely determines p P
3
= P
K
. Further show that P
2
P
3
.
We now relax our terminological rules about triangulations and
admit rectangles (and in higher dimensions, hyper rectangles or hyper-
cubes) in triangulations. We describe below some nite elements which
are rectangles.
We need another space of polynomials.
Denition 4.4. Let k 1 be an integer. Then
Q
k
=
_
_
p; p(x) =
0i
j
k
1jn
a
i
1
,...i
n
x
i
1
1
. . . x
i
n
n
_
_
.
42
We have the inclusions P
k
Q
k
P
nk
.
Example 4.4. The Rectangle of type (1).
Let K be the unit square in R
2
, i.e., K = [0, 1]
n
. Let P
K
= Q
1
.
The set of degrees of freedom is given by
_
K
= p(a
i
), 1 i 4; cf.
Fig. 4.4 in the case n = 2.
Figure 4.4:
To show that
_
K
indeed determines p Q
1
uniquely we adopt a
dierent method now. (There are essentially two methods to show that
_
K
completely determines P
K
; the rst was used in the previous ex-
amples where we exhibited a basis for P
K
such that the corresponding
coecients in the expansion of p in terms of this basis came from
_
K
;
the second is illustrated now).
Observe rst that dimP
K
= card
_
K
= 2
n
. To determine a polyno-
mial completely in terms of the elements of
_
K
we must solve 2
n
linear
equations in as many unknowns. That every polynomial is determined
this way is deduced from the existence of a solution to this system. But
for such a system the existence and uniqueness of the solution are equiv- 43
alent, and one establishes the latter. Thus we show that if p P
K
such
that all its degrees of freedom are zero, the p 0.
Returning to our example, consider a polynomial p Q
1
such that
p(a
i
) = 0 for all 1 i 4. On each side p is a polynomial of degree
1 either in x
1
alone or in x
2
alone. Since it vanishes at two points, the
polynomial p vanishes on the sides of the square. Now consider various
lines parallel to one of the axes. Here too p is a polynomial of degree 1
in one variable only. Since it vanishes at the points where the line meets
the side, it also vanishes on this line. Varying the line we get p 0
1
.
Example 4.5. The Rectangle of type (2).
Again consider the unit square (or hypercube in R
n
) to be the nite
element K. Set P
K
= Q
2
, and
_
K
= p(a
i
), 1 i 9 where the a
i
are
as in the gure below.
Figure 4.5:
Here again one can prove the unisolvency as above. Now let p Q
2
be given such that p(a
i
) = 0 for all 1 i 9; then p = 0 on the four
1
It is not necessary to restrict ourselves to a square. Any rectangle with sides parallel
to the coordinate axes would do.
sides and on the two central (dotted, in Fig. 4.5) lines. Now take lines
parallel to one of the axis and p vanishes on each of these. Thus p 0 44
on K and we get that
_
K
uniquely determines p Q
2
.
Exercise 4.3. Describe the rectangle of type (3) and generalize to hyper-
rectangles of type (k).
Exercise 4.4. Prove that in all the preceding examples, we get V
h

C
0
().
Exercise 4.5. The Rectangle of type (2
).
Let K be as in example 4.5. However omit the node a
9
(the centroid
of K). Let
_
K
= p(a
i
), 1 i 8 and show that this determines
uniquely a function in the space
P
K
=
_
_
p Q
2
; 4p(a
9
) +
4
i=1
p(a
i
) 2
8
i=5
p(a
i
) = 0
_
_
.
and that P
2
P
K
.
We now turn to dierent types of nite elements. They dier from
the preceding ones in the choice of degrees of freedom as will be seen
presently.
Example 4.6. The Hermite Triangle of Type (3).
Let K R
2
be a triangle with vertices a
1
, a
2
, a
3
. Let
i
, 1 i 3,
be the barycentric coordinate functions. Then one can check that any
polynomial p P
3
= P
K
can be expanded as
p =
3
i=1
(2
3
i
+ 3
2
i
7
1
3
)p(a
i
) + 27
1
3
p(a
123
)
+
3
i=1
j=1
ji
j
(2
i
+
j
1) Dp(a
i
)(a
j
a
i
).
Thus,
_
K
= p(a
i
), 1 i 3; Dp(a
i
)(a
j
a
i
), 1 i j 3; p(a
123
)
is the corresponding set of degrees of freedom.
Note that Dp(a
i
) is the Frechet derivative of p evaluated at a
i
: If 45
(e
1
, . . . , e
n
) is the standard basis for R
n
, then for v : R
n
R, we have,
(Dv)(x)(e
i
) =
v
x
i
(x), the usual partial derivative.
Notice that we may replace
_
K
by the set
K
=
_
p(a
i
), 1 i 3; p(a
123
);
p
x
1
(a
i
),
p
x
2
(a
i
), 1 i 3
_
.
Remark 4.2. The term Hermite means that we assume knowledge of
derivatives at some of the nodes. If only the values of p at the nodes ap-
pear in the set of degrees of freedom, as was the case upto Example 4.5,
we refer to the nite elements as of Lagrange type. These ideas will
be made precise in Sec. 5. We usually indicate degrees of freedom in-
volving derivatives by circling the nodes - one circle for rst derivatives,
two for rst and second derivatives and so on. Thus the nite element
of example 4.6 may be pictured as in Fig. 4.6.
Figure 4.6:
Exercise 4.6. The Hermite Triangle of Type (3
).
This is also known as the Zienkiewicz triangle in Engineering lit-
erature. Set
_
K
=
_
p(a
i
),
p
x
1
(a
i
),
p
x
2
(a
i
), 1 i 3
_
. Show that
_
K
uniquely determines a function in the space
P
K
=
_
_
p P
3
; 6p(a
123
) 2
3
i=1
p(a
i
) +
3
i=1
Dp(a
i
)(a
i
a
123
) = 0
_
_
.
All examples cited upto now yield the inclusion V
h
C
0
() and 46
consequently are useful to solve second-order problems. In order to
solve problems of fourth order, we need the inclusion V
h
C
1
(). Our
subsequent examples will be in this direction.
Remark 4.3. Consider a 1-simplex K R
1
. A triangulation is merely
a subdivision of into subintervals. In any subinterval K not only vK
but also
d(vK)
dx
must be continuous at both end points. Thus we get
4 conditions on vK. Consequently P
K
must contain all polynomials
of degree 3 in it. The analogous result (which is non-trivial) is due
to A.

Zeni sek [24] that is case of R
2
, and K a triangle of R
2
, at least
polynomials of degree 5 must be contained in P
K
.
Example 4.7. The Argyris triangle.
This is also known as the 21-degree-of-freedom-triangle. We set
P
K
= P
5
and
K
=
_
_
p(a
i
),
p
x
1
(a
i
), . . . ,
2
p
x
2
2
(a
i
), 1 i 3;
p
(a
i j
), 1 i j 3
_
.
Figure 4.7:
The knowledge of the normal derivative
p
is indicated by a line 47
perpendicular to the side at the appropriate point; cf. Fig. 4.7.
We now show that any p P
K
is uniquely determined by
_
K
. Let
p P
K
= P
5
be given such that all its degrees of freedom are zero. If K

is any side of K and t is an abscissa along K
then pK
is a polynomial
p
1
(t) of degree 5. The vanishing of p,
dp
dt
,
d
2
p
dt
2
at the end points, say b,
b
, of K
imply that all the 6 coecients of p

1
are 0 and hence p
1
0.
Thus p = 0 =
dp
dt
on K
. The polynomial r(t) =

p
(t) is of degree 4 on
K
and we also have r(b) = r(b
) =
dr
dt
(b) =
dr
dt
(b
) = r
_
b + b
2
_
= 0
which imply that r 0 on K
. Hence p,
p
x
1
,
p
x
2
all vanish on the
sides of the triangle K. The sides of K are dened by the equations
i
(x
1
, x
2
) = 0, (i = 1, 2, 3) where
i
are the barycentric coordinate func-
tions. We claim that
2
i
divides p for i = 1, 2, 3. To see this it is enough
to prove that if p is a polynomial such that p,
p
x
1
,
p
x
2
vanish on any
straight line L = (x
1
, x
2
); (x
1
, x
2
) = 0 then
2
divides p. In the special
case, when (x
1
, x
2
) = x
1
writing p(x
1
, x
2
) =
_
5
j=0
a
j
(x
2
)x
j
1
(with deg.
a
j
5 j) it follows that a
0
(x
2
) = a
1
(x
2
) = 0 since p =
p
x
1
= 0 on
L. Thus x
2
1
divides p. The general case reduces to this case by an ane
transformation. In fact, by translating the origin to a point P, xed ar-
bitrarily on L and by rotation of the coordinate axes we can assume that
L = (X
1
, X
2
); X
1
= 0 in the new coordinates. If p
is the image of p
under this transformation then p
is also a polynomial (of degree 5) and

p
,
p
X
1
,
p
X
2
vanish on L by chain rule for dierentiation. Hence X
2
1
di-
vides p
. This is the same thing as saying

2
divides p which proves the
claim. Since
i
are mutually coprime we may now write p = q
2
1
2
2
2
3
.
Then we necessarily have q(x
1
, x
2
) 0 for, otherwise deg. p 6 which
is impossible since p P
5
. Hence p 0 on K which proves that
_
K
determines p P
5
.
To dene the corresponding space V
h
, we number all the vertices of
the triangles by b
j
and all midpoints of the sides by c
k
. The the set
of degrees of freedom of the space V
h
is 48
h
=
_
_
v(b
j
),
v
x
1
(b
j
),
v
x
2
(b
j
),
2
v
x
2
1
(b
j
),
2
v
x
1
x
2
(b
j
),
2
v
x
2
2
(b
j
),
v
k
(c
k
)
_
_
where

k
is one of the two possible normal derivatives at the mid-point
c
k
.
We nowshow that V
h
C
1
(). Consider two adjacent Argyris trian-
gles K
1
and K
2
with common boundary K
along which t is an abscissa

(Fig. 4.8). Let v V
h
. Now vK
1
and vK
2
are polynomials of degree 5 in
t along K
and they agree together with their rst and second derivatives
at the end points. Thus vK
1
= vK
2
on K
, proving continuity.
Now
(vK
1
)
and
(vK
2
)
along K
are polynomials of degree 4

agreeing in their values with rst derivatives at end points and agree at
the mid-point in their values. Thus
(vK
1
)
=
(vK
2
)
on K
. Similarly,
(vK
1
)
t
=
(vK
2
)
t
on K
and hence v C
1
(). Thus V
h
C
1
().
Figure 4.8:
Exercise 4.7. The 18-Degree-of-Freedom-Triangle
Let K be a triangle in R
2
. Let P
K
consist of those polynomials of
degree 5 for which, along each side of K, the normal derivative is
a polynomial of degree 3, in one variable of course. Show that a
polynomial in P
K
is uniquely determined by the following set of degrees
of freedom:
K
=
_
_
p(a
i
),
p
x
1
(a
i
), . . . ,

2
p
x
2
2
(a
i
), 1 i 3
_
_
.
49
Note that P
4
P
K
and dimP
K
= 18.
Exercise 4.8. The HCT-Triangle.
This element is due to Hsieh, Clough and Tocher. Let a be any inte-
rior point of the triangle K with vertices a
1
, a
2
, a
3
. With a as common
vertex subdivide the triangle into triangles K
1
, K
2
, K
3
; cf. Fig. 4.9. De-
ne
P
K
=
_
p C
1
(K); PK
i
P
3
, 1 i 3
_
.
Obviously, P
3
P
K
. The degrees of freedom are given by
K
=
_
p(a
i
),
p
x
1
(a
i
),
p
x
2
(a
i
), 1 i 3;
p
(a
i j
), 1 i < j 3
_
.
Figure 4.9:
Show that
_
K
uniquely determines p P
K
.
Note: Since we have to determine 3 polynomials p
i
= pK
i
each of
degree 3, we need to determine 30 coecients on the whole. For this
we have the following conditions:
(i) The values at the vertices together with rst derivatives and also
the normal derivative at the mid points give 7 conditions for each
p
i
= pK
i
. Thus we have 21 conditions from these.
(ii) p
1
(a) = p
2
(a) = p
3
(a) gives 2 conditions.
(iii)
p
1
x
i
(a) =
p
2
x
i
(a) =
p
3
x
1
(a) for i = 1, 2, gives 4 more conditions. 50
(iv)
p
1
=
p
2
along a
1
a and two more similar conditions give 3
conditions.
Thus we have 30 conditions to determine the 30 coecients. But,
of course this is no proof, which is left as an exercise!
Exercise 4.9. The Bogner-Lox-Schmidt Rectangle; cf. Fig. 4.10.
Figure 4.10:
Let P
K
= Q
3
, the degrees of freedom being given by
K
=
_
p(a
i
),
p
x
1
(a
i
),
p
x
2
(a
i
),
2
p
x
1
x
2
(a
i
), 1 i 4
_
.
Show that
_
K
determines uniquely a polynomial p Q
3
(a double
dotted arrow indicates that the mixed second derivative is a degree of
freedom). Show also that in this case V
h
C
1
().
So far, we have veried requirements (i) and (ii) mentioned at the
beginning of this Section. Let us now examine requirement (iii), which
will be fullled by a canonical choice for the basis functions. Let
_
h
be the set of degrees of freedom of the space V
h
derived in an obvious
way from the sets
_
K
, K k
h
; Examples of such sets
_
h
have been
given for n-simplices of type (k) and for Argyris triangles. Then if 51
h
=
_
jh
, 1 j M
_
,
we let the basis functions w
j
, 1 j M, be those functions in the space
V
h
which satises
i
(w
j
) =
i j
, 1 i M.
Then it is easily seen that this choice will result in functions with
small support: in Fig. 4.11, we have represented there types of sup-
ports encountered in this fashion, depending upon the position of the
node associated
support of basis
function associated
with node
support of basis
function associated
with node
support of basis
function associated
with node
Figure 4.11:
with the degree of freedom.
Chapter 5
General Properties of Finite
Elements
IT WOULD HAVE been observed that upto now we have not dened - 52
nite elements in a precise manner. Various polygons like triangles, rect-
angles, etc. were loosely called nite elements. We rectify this omission
and make precise the ideas expressed in the previous sections.
Denition 5.1. A nite element is a triple (K, , P) such that
(i) K R
n
with a Lipschitz continuous boundary K and Int K .
(ii) is a nite set of linear forms over C
(K). The set is said to be

the set of degrees of freedom of the nite element.
(iii) P is a nite dimensional space of real-valued functions over K
such that is P-unisolvent: i.e. if =
i
N
i=1
and
i
, 1 i N
are any scalars, then there exists a unique function p P such that
(5.1)
i
(p) =
i
, 1 i N.
Condition (iii) of denition (5.1) is equivalent to the conditions that
dimP = N = card and that there exists a set of functions p
j
N
j=1
with
i
(p
j
) =
i j
(1 i, j N), which forms a basis of P over R. Given any
53
54 5. General Properties of Finite Elements
p P we may write
(5.2) p =
N
i=1
i
(p)p
i
.
Instead of (K,
_
, P) one writes at times (K,
_
K
, P
K
) for the nite
element.
In the various examples we cited in Sec. 4 our set of degrees of free-
dom for a nite element K (which was an n-simplex or hyper-rectangle)
had elements of the following type: 53
Type 1:
0
i
given by p p(a
0
i
). The points a
0
i
were the vertices, the
mid-points of sides, etc...
Type 2:
1
i,k
given by p Dp(a
1
i
)(
1
i,k
). For instance, in the Hermite
triangle of type (3) (cf. Example 4.6), we had a
1
i
= a
i
,
1
i,k
= a
i
a
k
,
where a
1
, a
2
, a
3
were the vertices.
Type 3:
2
i,kl
given by p D
2
p(a
2
i
)(
2
i,k
,
2
i,l
). For example, in the 18-
degree-of-freedom triangle, a
2
i
= a
i
,
2
i,k
= e
1
=
2
i,1
, the unit vec-
tor in the x
1
-direction so that we have D
2
p(a
i
)(e
1
, e
1
) =
2
p
x
2
1
(a
i
)
as a degree of freedom. (cf. Exercise 4.7).
In all these cases the points a
s
i
for s = 0, 1 and 2, are points of K
and are called the nodes of the nite element.
Denition 5.2. A nite element is called a Lagrange nite element if its
degrees of freedom are only of Type 1. Otherwise it is called a Hermite
nite element. (cf. Remark 4.2)
Let (K, , P) be a nite element and v : K R be a smooth
function on K. Then by virtue of the P-unisolvency of , there exists a
unique element, say, v P such that
i
(v) =
i
(v) for all 1 i N,
where =
i
N
i=1
. The function v is called the P-interpolate function
of v and the operator : C
(K) P is called the P-interpolation

5. General Properties of Finite Elements 55
operator. If p
j
N
j=1
is a basis for P satisfying
i
(p
j
) =
i j
for 1 i,
j N then we have the explicit expression
(5.3) () =
N
i=1
i
()p
i
.
Example 5.1. In the triangle of type (1) (see Example 4.1), P = P
1
,
=
i
;
i
(p) = p(a
i
), 1 i 3 and p
i
=
i
, the barycentric coordinate 54
functions. Thus we also have
(5.4) v =
3
i=1
v(a
i
)
i
.
Exercise 5.1. Let K be a triangle with vertices a
1
, a
2
and a
3
. Let a
i j
(i <
j) be the mid-point of the side joining a
i
and a
j
. Dene
K
= p
p(a
i j
), 1 i j 3. Show that is P
1
-unisolvent and that in general
V
h
C
0
() for a triangulation made up of such nite elements.
Exercise 5.2. Let K be a rectangle in R
2
with vertices a
1
, a
2
, a
3
, a
4
. Let
a
5
, a
6
, a
7
, a
8
be the midpoints of the sides as in Fig. 4.5. If = p
p(a
i
), 5 i 8, show that is not Q
1
-unisolvent.
Let us now consider a family of nite elements of a given type. To
be more specic, we will consider for instance a family of triangles of
type (2) (see Example 4.2), but our subsequent descriptions extend to all
types of nite elements in all dimensions.
Pick, in particular, a triangle

K with vertices a
1
, a
2
, a
3
from this
family. Let the mid-points of the sides be a
12
, a
23
, a
13
. Set

P = P

K
=
P
2
and dene accordingly the associated set of degrees of freedom for
K as
K
= p p( a
i
), 1 i 3; p p( a
i j
), 1 i < j 3.
In as much as we consider the nite element (

K,

,

P) as xed in the
sequel, it will be called the reference nite element of the family.
Given any nite element K with vertices a
1
, a
2
, a
3
in this family,
there exists a unique invertible ane transformation of R
2
i.e. of the
form F
K
( x) = B
K
x + b
K
, where B
K
is an invertible 2 2 matrix and
b
K
R
2
, such that F
K
(

K) = K and F
K
( a
i
) = a
i
, 1 i 3. It is 55
easily veried that F
K
( a
i j
) = a
i j
for 1 i < j 3. Also, the space
p : K R; p = p F
1
K
, p

P is precisely the space P
K
= P
2
. Hence
the family (K, , P) is equivalently dened by means of the following
data:
(i) A reference nite element (

K,

,

P),
(ii) A family of ane mappings F
K
such that F
K
(

K) = K, a
i
=
F
K
( a
i
),
1 i 3, a
i j
= F
K
( a
i j
), 1 i < j 3, and
K
= p p(F
K
( a
i
)); p p(F
K
( a
i j
)),
P
K
= p : K R; p = p F
1
K
, p

P.
This special case leads to the following general denition.
Denition 5.3. Two nite elements (

K,

,

P) and (K, , P) are ane
equivalent if there exists an ane transformation F on R
n
such that
(i) F( x) = B x + b, b R
n
, B an invertible n n matrix,
(ii) K = F(

K),
(iii) a
s
i
= F( a
s
i
), s = 0, 1, 2,
(iv)
1
i,k
= B
1
i,k
,
2
i,k
= B
2
i,k
,
2
i,l
= B
2
i,l
and
(v) Pp : K R; p = p F
1
, p

P.
This leads to the next denition.
Denition 5.4. A family (K,
K
, P
K
) of nite elements is called an
ane family if all the nite elements (K,
K
, P
K
) are equivalent to a
single reference nite element (

K,

,

P).
5. General Properties of Finite Elements 57
Let us see why the relations given by (iv) must be precisely of that 56
form. We have by (v), p(x) = p( x). This must be valid when we use the
basis functions as well. We have:
p( x) =
i
p( a
0
i
) p
0
i
( x) +
i,k
D p( a
1
i
)(
1
i,k
) p
1
i,k
( x)
+
i,k,l
D
2
p( a
2
i
)(
2
i,k
,

2
i,l
) p
2
i,kl
( x).
Now D p( a
1
i
)(
1
i,k
) = Dp(a
1
i
)B
1
i,k
by a simple application of the chain
rule and therefore
D p( a
1
i
)(
1
i,k
) = Dp(a
1
i
)
1
i,k
, by (iv).
By a similar treatment of the second derivative term, we get
p( x) =
i
p(a
0
i
)p
0
i
(x) +
i,k
Dp(a
1
i
)(
1
i
)(
1
i,k
)p
i,k
(x)
+
i,k,l
D
2
p(a
2
i
)(
2
ik
,
2
il
)p
ikl
(x) = p(x).
Thus the relations (iv) and (v) are compatible.
Theorem 5.1. Let (K, , P) and (

K,

,

P) be ane equivalent with F
K
as the ane transformation. If v : K R induces v :

K R by
v( x) = v(x) for x

K, (x = F
K
( x)), then v = v.
Proof. Let

=
i
N
i=1
, =
i
N
i=1
. By denition,

i
( v) =
i
(v) =
i
(v), 1 i N.
i
( v) =
i
( v) =
i
(v), 1 i N.
Thus,
i
( v) =
i
( v) for 1 i N. Hence v = v by uniqueness
of the function v. 57
Let us consider a polygonal domain with a triangulation t
h
. Sup-
pose to each K t
h
is associated a nite element, (K,
K
, P
K
),
K
being
the set of degrees of freedom, and P
K
the nite dimensional space such
that
K
is P
K
-unisolvent. Then we have dened the interpolation opera-
tor
K
. All these make sense locally i.e. at a particular nite element K.
We now dene the global counterparts of these terms. The comparison
is given in the following table.
Table 5.1.
Local denition Global denition
1. Finite element K. 1. The set =
_
K=t
n
K
2. The boundary of K, K. 2. The boundary of , = .
3. The space P
K
of functions
K R, which is nite-
dimensional.
3. The space V
h
of functions
R, which is also nite-
dimensional.
4. The set
_
K
=
i,K
N
i=1
of
degrees of freedom of K.
4. The set of degrees of free-
dom
_
h
=
i
N
i=1
, where
i
(pK) =
i,K
(pK).
5. Basis functions of P
K
are
p
i,K
N
i=1
.
5. Basis function of V
h
are w
j
.
6. The nodes of K are
a
0
i
, a
1
i
, a
2
i
, . . ..
6. The nodes of t
h
are by def-
inition,
_
Kt
h
Nodes of K =
b
j
say.
7.
K
is the P
K
-interpolation
operator, dened by
i,K
(
K
v) =
i,K
(v), for all
i,K

_
K
.
7. The V
h
interpolation opera-
tor
h
is dened by
h
v
V
h
such that,
i,K
(
h
vK) =
i,K
(vK) for all
i,K

_
K
.
58
Notice that, by denition,
(5.5) (
h
v)K =
K
(vK) for all K t
h
.
It is this property and the conclusion of theorem 5.1 that will be
essential in our future error analysis.
Denition 5.5. We say that a nite element of a given type is of class
C
0
, resp. of class C
1
, if, whenever it is the generic nite element of a
triangulation, the associated space V
h
satises the inclusion V
h
C
0
(),
resp. V
h
C
1
(). By extension, a triangulation is of class C
0
, resp. of
class C
1
if it is made up of nite elements of class C
0
, resp. of class C
1
.
Reference: A forthcoming book of Ciarlet and Raviart [5].
Chapter 6
Interpolation Theory in
Sobolev Spaces
WE OUTLINED THE internal approximation method in Sec. 3. We are 59
naturally interested in the convergence of the solutions u
h
V
h
to the
global solution u V. As a key step in this analysis we obtained the
error estimate (cf. Theorem 3.1):
(6.1) u u
h
C inf
v
h
V
h
u v
h
.
To be more specic let us consider an example. Given R
2
a
polygon, consider the solution of the following problem, which is there-
fore posed in the space V = H
1
0
():
(6.2)
_
_
u + au = f in ,
u = 0 on .
Let t
h
be a triangulation of by triangles of type (1), (2) or (3).
Then u
h
V
h
H
1
0
() and (6.1) reads as
(6.3) u u
h
1,
C inf
v
h
V
h
u v
h
1,
.
We know a priori that u H
1
0
(). Let us assume for the moment
that u C
0
(). (Such assumptions are made possible by the various
59
60 6. Interpolation Theory in Sobolev Spaces
regularity theorems. For instance, u H
2
() C
0
() if f L
2
()
and is a convex polygon). If u C
0
(), then we may dene the
V
h
-interpolate of u, i.e.,
h
u by
h
u(b
j
) = u(b
j
) for the nodes b
j
of the
triangulation. Note also that
h
u
K
=
K
u (cf. (5.5)). Now from (6.3)
we get,
u u
h
1,
Cu
h
u
1,
= C
_
Kt
h
u
h
u
2
1,K
_
_
1
2
= C
_
Kt
h
u
K
u
2
1,K
_
_
1
2
60
Thus the problem of estimating uu
h
1,
is reduced to the problem
of estimating u
K
u
1,K
. This is one central problem in the nite ele-
ment method and motivates the study of interpolation theory in Sobolev
spaces.
We consider more general types of Sobolev spaces for they are no
more complicated for this purpose than those dened in Sec. 2.
Denition 6.1. Let m 0 be an integer, and 1 p +. Then the
Sobolev space W
m, p
() for R
n
, open, is dened by
W
m, p
() = v L
p
();
v L
p
() for all m.
Remark 6.1. H
m
() = W
m,2
().
On the space W
m, p
() we dene a norm
m, p,
by
(6.4) v
m, p,
=
_
_
_
v
p
dx
_
_
1/p
and the semi-norm
m, p,
by
(6.5) v
m, p,
=
_
_
_
=m
v
p
dx
_
_
1/p
6. Interpolation Theory in Sobolev Spaces 61
If k 1 is an integer, consider the space W
k+1, p
()/P
k
. If v stands
for the equivalence class of v W
k+1, p
() we may dene the analogues
of (6.4) and (6.5) respectively by
(6.6) v
k+1, p,
= inf
pP
k
v + p
k+1, p,
and 61
(6.7) v
k+1, p,
= v
k+1, p,
.
These are obviously well-dened and
k+1, p,
denes the quotient
norm on the quotient space above. We then have the following key
result, whose proof may be found in Ne cas [20] for instance.
Theorem 6.1. In W
k+1, p
()/P
k
, the semi-norm v
k+1, p,
is a norm
equivalent to the quotient norm v
k+1, p,
, i.e., there exists a constant
C = C() such that for all v W
k+1, p
()/P
k
(6.8) v
k+1, p,
v
k+1, p,
C v
k+1, p,
.
Equivalently, we may state
Theorem 6.2. There exists a constant C = C() such that for each
v W
k+1, p
()
(6.9) inf.
pP
k
v + p
k+1, p,
Cv
k+1, p,
.
(Note: This result holds if has a continuous boundary and if it is
bounded so that P
k
W
k+1, p
().)
We now prove the following
Theorem 6.3. Let W
k+1, p
() and W
m,q
() be such that W
k+1, p
()
W
m,q
() (continuous injection). Let L(W
k+1
, (), W
m,q
()), i.e.
a continuous linear map, such that for each p P
k
, p = p. Then there
exists C = C() such that for each v W
k+1, p
()
v v
m,q,
CI
L(W
k+1
,P(),W
m,q
())
v
k+1, p,
62
Proof. For each v W
k+1, p
() and for each p P
k
, we can write
v v = (I )(v + p).
Thus,
v v
m,q,
v v
m,q,
= I
L(W
k+1,p
(),W
m,q
())
v + p
k+1, p,
,
for all p P
K
. Hence,
v v
m,q,
I
L(W
k+1,p
(),W
m,q
())
inf
pP
k
v + p
k+1, p,
CI
L(W
k+1,p
(),W
m,q
())
v
k+1, p,
By theorem 6.2, this completes the proof.
Denition 6.2. Two open subsets ,

of R
n
are said to be ane equiv-
alent if there exists an invertible ane map F mapping x to B x +b, B an
invertible (n n) matrix and b R
n
, such that F(

) = .
If ,

are ane equivalent, then we have a bijection between their
points given by x x = F( x). Also we have bijections between smooth
functions on and

dened by (v : R) ( v :

R) where
v(x) = v( x).
The following theorem gives estimates of v
m, p,
and v
m, p,
each
in terms of the other.
Theorem 6.4. Let ,

R
n
be ane equivalent. Then there exist 63
constants C,

C such that for all v W
m, p
()
(6.11) v
m, p,
CB
m
det B
1/p
v
m, p,
and for all v W
m, p
(

)
(6.12) v
m, p,
CB
1
m
det B
1/p
v
m, p,

.
Note:
(i) It suces to prove either (6.11) or (6.12). We get the other by
merely interchanging the roles of and

. We will prove the
former.
(ii) B is the usual norm of the linear transformation dened by B,
viz. B = sup
xR
n
x0
Bx
x
. (Recall that F(

) = , F( x) = B x + b).
Proof. Let e
1
, . . . , e
n
be the standard basis for R
n
. Let be a multi-
index with = m. By choosing a suitable collection e
1
, . . . , e
m
with appropriate number of repetitions from the basis, we may write,

(
v)( x) = (D
m
v)( x)(e
1
, . . . , e
m
),
where D
m
v is the m
th
order Fr echet derivative of v and D
m
v( x) is conse-
quently an m-linear form on R
n
. Thus,
v( x) D
m
v( x) = sup
i
=1
1im
D
m
v( x)(
1
, . . . ,
m
).
Since this is true for all = m, we get
(6.13) v
m, p,
C
1
__
D
m
v( x)
p
d x
_
1/p
C
2
v
m, p,

.
The rst inequality is a consequence of our preceding argument.
The second follows by a straightforward argument. By composition of
functions in dierentiation:
(6.14) D
m
v( x)(
1
, . . . ,
m
) = D
m
v(x)(B
1
, . . . , B
m
);
This gives 64
(6.15) D
m
v( x) D
m
v(x) B
m
.
Hence the rst inequality in (6.13) may be rewritten as
v
p
m, p,
C
p
1
B
mp
_
D
m
v(F( x))
p
d x
= C
p
1
B
mp
det B
1
_
D
m
v(x)
p
dx
C
p
B
mp
det B
1
v
m, p,
.
by an inequality similar to the second inequality of (6.13). Raising to
power 1/p on either side we get (6.11). This completes the proof.
We now estimate the norms B and B
1
in terms of the sizes of
and

. More precisely, if h, (resp.

h) the supremum of the diameters
of all balls that can be inscribed in , (resp.

), we have the following:
Theorem 6.5. B h/ , and B
1

h/.
Proof. Again it suces to establish one of these. Now,
B = sup
=
_
1

B
_
.
Let R
n
with = . Choose y, z

such that = y z. Then
B = B y B z = y z, where F( y) = y, F( z) = z. But y, z and hence
y z h. Thus B h. Hence B h/ , which completes the
proof.
We conclude this section with an important, often used, result.
Theorem 6.6. Let (

K,

,

P) be a nite element. Let s(= 0, 1 or 2) be
the maximal order of derivatives occurring in

. Assume that:
(i) W
k+1, p
(

K) C
s
(

K) 65
(ii) W
k+1, p
(

K) W
m,q
(

K)
(iii) P
k

P W
m,q
(

K)
Then there exists a constant C = C(

K,

,

P) such that for all ane
equivalent nite elements (K, , P) we have
(6.16) v
K
v
m,q,K
C(meas K)
1
q
1
p
h
k+1
K
m
K
v
k+1, p,K
for all v W
k+1, p
(K), where h
K
is the diameter of K and
K
is the
supremum of diameters of all balls inscribed in K.
Proof. Since P
k

P, for any polynomial p P
k
we have p = p. We
may write
v =
i
v( a
0
i
) p
0
i
+
i,k
(D v( a
1
i
)(
1
i,k
)) p
1
i,k
= +
i,k,l
(D
2
v( a
2
i
)(
2
i,k
,

2
i,l
)) p
2
i,k,l
,
. (6.17)
all these sums being nite (the second and third may or may not be
present). We claim that L(W
k+1
, p(

K), W
m,q
(

K)). Since

P
W
m,q
(

K) all the basis functions in (6.17) are in W
m,q
(

K). Thus,
v
m,q,

K

i
v( a
0
i
) p
0
i
m,q,

K
+
i,k
D v( a
1
i
)(
1
i,k
) p
1
ik
m,q,

K
+
i,k,l
D
2
v( a
2
i
)(
2
i,k
,

2
i,l
) p
2
ikl
m,q,

K
(6.18)
Since W
k+l, p
(

K) C
s
(

K) and all the numbers v( a
0
i
), etc. . . , are
bounded by their essential supremum over

K,
v
m,q,

K
C v
k+1, p,

K
.
66
Hence the claim is valid. Now by virtue of (ii) and also our obser-
vation on preservation of polynomials, we may apply theorem 6.3 to .
Hence there exists C = C(

K,

,

P) such that
v v
m,q,

K
C v
k+1, p,

K
for v W
k+1, p
(

K).
Notice that v =
K
v by Theorem 5.1. Thus v v = v
K
v. Thus
if F
K
(

K) = K where F
K
( x) = B
K
x + b
K
, we get
(6.19) v
K
v
m,q,K
C
1
B
1
K

m
det B
K
1/q
v v
m,q,

K
,
by Theorem 6.4. Also by the same theorem
(6.20) v
k+1, p,

K
C
2
B
K
k+1
det B
K
1/p
v
k+l, p,K
.
Further det B
K
being the Jacobian of the transformation, we have
det B
K
=
meas K
meas

K
and B
K

h
K

, B
1
K

h/
K
by theorem 6.5.
Since

h, , meas

K are constants, combining (6.19), (6.20) and the pre-
ceding observation we complete the proof of the theorem.
References: See Bramble and Hilbert [28], Bramble and Zl qmal [2],
Ciarlet and Raviart [7], Ciarlet and Wagschal [8], Strang [21],

Zeni sek
[23, 24] and Zl amal [25, 32].
Chapter 7
Applications to
Second-Order Problems
Over Polygonal Domains
WE APPLYTHE results of the preceding section in studying the conver- 67
gence of the nite element method, i.e. the convergence of the solutions
u
h
of (P
h
) to the solution u of a problem (P) which corresponds to the
choice V = H
1
() or H
1
0
(), which we saw in Sec. 2 led to second-order
problems.
Let be a polygonal domain throughout.
Denition 7.1. A family (t
h
) of triangulations of is regular is
(i) for all t
h
and for each K t
h
, the nite elements (K, , P) are
all ane equivalent to a single nite element, (

K,

,

P) called the
reference nite element of the family;
(ii) there exists a constant such that for all t
h
and for each K t
h
we have
(7.1)
h
K
K

where h
K
,
K
are as in Theorem 6.6;
67
68 7. Applications to Second-Order Problems...
(iii) for a given triangulation t
h
if
(7.2) h = max
Kt
h
h
K
,
then h 0.
Remark 7.1. The condition (ii) in denition 7.1 assures us that as h 0
the triangles do not become at; cf. Exercise 7.1.
Exercise 7.1. If n = 2 and the sets K are triangles, show that condition
(ii) of denition 7.1 is valid if and only if there exists
0
> 0 such that
for all t
h
and K t
h
,
K

0
> 0,
K
being the smallest angle in K. 68
Exercise 7.2. Consider the space V
h
associate with t
h
. Since V
h
is nite
dimensional all norms are equivalent and hence
v
h
0,,
C
h
v
h
0,
for all v
h
V
h
,
for some constant C
h
, a priori dependent upon h, which we may evaluate
as follows: If (t
h
) is a regular family of triangulations, show that there
exists a constant C, independent of h, such that
(7.3) v
h
0,,

C
h
n/2
v
h
0,
for v
h
V
h
.
Also show that there exists a constant C such that
(7.4) v
h
1,

C
h
v
h
0,
for all v
h
V
h
.
We now obtain an estimate for the error u u
h
1,
when the family
of triangulations is regular, which also gives convergence.
Theorem 7.1. Let (t
h
) be a regular family of triangulations on of
class C
0
(i.e. V
h
C
0
()) with reference nite element (

K,

,

P). We
assume that there exists an integer k 1 such that
(i) P
K

P H
1
(

K)
(ii) H
k+1
(

K) C
s
(

K) where s(= 0, 1, or 2) is the maximal order of
derivatives in

_
.
7. Applications to Second-Order Problems... 69
(iii) u H
k+1
() (Regularity assumption).
Then there exists a constant C (independent of V
h
) such that
(7.5) u u
h
1,
Ch
k
u
k+1,
.
Proof. Since V
h
C
0
(), P
K
H
1
(K), we have V
h
V. By (ii) and 69
(iii) of the hypothesis we have that the V
h
-interpolate of u, viz.
h
u is
well-dened. Since
h
u V
h
, by our fundamental result (see Theo-
rem 3.1 or relation (6.1)), it suces to estimate u
h
u
1,
. Now,
H
k+1
(

K) C
s
(

K),
H
k+1
(

K) H
1
(

K) (k 1),
P
k
P H
1
(

K),
and we may apply Theorem 6.6 with p = q = 2, m = 1 to get
u
K
u
1,K
Cu
k+1,K
h
k+1
K
K
Cu
k+1,K
h
k
K
(Since
h
K
K
).
(7.6)
Similarly with m = 0 we get
(7.7) u
K
u
0,K
Cu
k+1,K
h
k+1
K
.
These together give
(7.8) u
K
u
1,K
Ch
k
K
u
k+1,K
.
Now since h
K
h,
u
h
u
1,
=
_
Kt
h
u
K
u
2
1,K
_
_
1
2
C h
k
_
Kt
h
u
2
k+1,K
_
_
1
2
= C h
K
u
k+1,
.
Example 7.1. Consider triangulations by triangles of type (1). Then 70
k = 1,

P = P
1
and if n = 2 or 3, H
2
(

K) C
0
(

K). If u H
2
(),
Theorem 7.1 says that
u u
h
1,
C hu
2,
.
We conclude the analysis of convergence in the norm
1,
with
the following result.
Theorem 7.2. Let (t
h
) be a regular family of triangulations of , of
class C
0
. Let s = 0 or 1 and let P
1

P H
1
(

K). Then (with the
assumption that u V = H
1
() or H
1
0
()) we have
(7.9) lim
h0
u u
h
1,
= 0
Proof. Let V = V W
2,
(). Since s 1, W
2,
() C
s
() and
W
2,
() H
1
(). The second inclusion follows a fortiori from the
rst with s = 1. Also, P
1

P H
1
(

K). Thus we may apply Theorem
6.6 with k = 1, p = , m = 1, q = 2. Then for all v V ,
v
K
v
1,K
C(meas K)
1
2
hv
2,,K
C(meas K)
1
2
hv
2,,
.
Summing over K, we get
v
h
v
1,
C hv
2,,
_
Kt
h
meas K
_
_
1
2
= C hv
2,,
,
since
_
Kt
h
meas K = meas , a constant. Thus, for all v V ,
(7.10) lim
h0
v
h
v
1,
= 0
71
Notice that V = V. Hence choose v
0
V such that uv
0
1,
/2
where > 0 is any preassigned quantity. Then once v
0
is chosen, by
(7.10) choose h
0
such that for all h h
0
, v
0

h
v
0
1,
/2. Now, by
(6.1)
u u
h
1,
Cu
h
v
0
1,
C
_
u v
0
1,
+ v
0

h
v
0
1,
_
C, for h h
0
.
This gives (7.9) and completes the proof.
We now have, by Theorem 7.1, u u
h
0,
u u
h
1,
= 0(h
k
).
We now show, by another argument that u u
h
0,
= 0(h
k+1
), (at least
in some cases) there by giving a more rapid convergence than expected.
This is done by the Aubin-Nitsche argument (also known as the duality
argument). We describe this in an abstract setting.
Let V be a normed space with norm denoted by . Let H be a
Hilbert space with norm and inner product (, ) such that
(7.11)
_
_
(i) V H, and
(ii) V = H.
For second-order problems: V = H
1
() or H
1
0
() and H = L
2
().
Since H is a Hilbert space, we may identify it with its dual. Further
since V is dense in H, we have that H may be identied with a subspace
of V
, the dual of V. For, if g H, dene g V
by g(v) = (g, v) g V
since g(v) Cg v. If g(v) = 0 for all v V, then (g, v) = 0 for all

v H as well since V = H. Thus g = 0. This proves the identication.
In the sequel we will set g = g. 72
Recall that u and u
h
are the solutions of the problems:
a(u, v) = f (v) for all v V, (P)
a(u
h
, v
h
) = f (v
h
) for all v
h
V
h
V, (P
h
)
and that the assumptions on (P) are as in the Lax-Milgram lemma. Then
we have the following theorem.
Theorem 7.3. Let the spaces H and V satisfy (7.11). Then with our
above mentioned notations,
(7.12) u u
h
Mu u
h
_
sup
gH
_
1
g
inf
h
V
h

h
__
_
,
where for each g H, V is the corresponding unique solution of the
problem
(P
) a(v, ) = (g, v) for all v V,

and M the constant occurring in the inequality giving continuity of
a(, ).
Remark 7.2. Note that unlike in (P), we solve for the second argument
of a(, ) in (P
). This is called the adjoint problem of (P). The existence

and uniqueness of the solution of (P
) are proved in an identical manner.

Note that if a(, ) is symmetric, then (P) is self-adjoint in the sense that
(P) = (P
).
Proof. From the elementary theory of Hilbert spaces, we have
(7.13) u u
h
= sup
gH
g0
(g, u u
h
)
g
.
For a given g H,
(7.14) (g, u u
h
) = a(u u
h
, )
Also if
h
V
h
we have, 73
(7.15) a(u u
h
,
h
) = 0.
Thus (7.14) and (7.15) give
(7.16) (g, u u
h
) = a(u u
h
,
h
),
which gives us
(g, u u
h
) Mu u
h

h
,
and hence
u u
h
Mu u
h
sup
gH
g0
_

h
g
_
.
by (7.13). Since this is true for any
h
V
h
we may take inmum over
V
h
to get (7.12), which completes the proof.
For dimensions 3 and Lagrange nite elements we now show that
u u
h
0,
= 0(h
k+1
). For this we need one more denition.
Denition 7.2. Let V = H
1
() or H
1
0
(), H = L
2
(). The adjoint
problem is said to be regular if the following hold:
(i) for all g L
2
(), the solution of the adjoint problem for g
belongs to H
2
() V;
(ii) there exists a constant C such that for all g L
2
()
(7.17)
2,
Cg
0,
,
where is the solution of the adjoint problem for g.
Theorem 7.4. Let (t
h
) be a regular family of triangulations on with
reference nite element (

K,

,

P). Let s = 0 and n 3. Suppose there
exists an integer k 1 such that u H
k+1
(), P
k

P H
1
(

K). Assume 74
further that the adjoint problem is regular in the sense of Denition 7.2.
Then there exists a constant C independent of h such that
(7.18) u u
h
0,
C h
k+1
u
k+1,
.
Proof. Since n 3, H
2
() C
0
(). Also, H
2
() H
1
() and P
1

P
H
1
(

H). Thus for H
2
(), by Theorem 7.1,

h
1,
C h
2,
.
Hence
(7.19) inf
h
V
h

h
1,
C h
2,
.
By (7.12) and (7.19).
u u
h
0,
Mu u
h
1,
sup
gL
2
()
_
1
g
0,
Ch
2,
_
.
By the regularity of (P
),
(7.20)

2,
g
0,
2,
g
0,
constant.
Thus, u u
h
0,
C hu u
h
1,
C h(h
k
u
k+1,
) (by theorem 7.1).
This gives (7.18) and completes the proof.
We nally give an estimate for the error in the L
-norm.
Theorem 7.5. Let (t
h
) be a regular family of triangulations on R
n
,
where n 3. Assume further that for all t
h
and K t
h
.
(7.21) 0 <
h
K
h
, f rm[o], being a constant.
Let u H
2
() and P
1

P H
1
(

K) L
(

K). If (P
) is regular, 75
then there exists a constant C independent of h such that
(7.22)
_
_
u u
h
0,,
C hu
2,
; if n = 2
u u
h
0,,
C
hu
2,
if n = 3.
Proof. Assume n = 2. Now
(7.23) u u
h
0,,
u
h
u
0,,
+
h
u u
h
0,,
.
Note that since (u
h

h
u) V
h
, we may apply Exercise 7.1 to get
(7.24) u
h

h
u
0,,

C
h
u
h

h
u
0,
.
Thus,
u
h

h
u
0,,

C
h
_
u
h
u
0,
+ u
h
u
0,
_
C
h
_
C
1
h
2
u
2,
+ C
2
h
2
u
2,
_
C hu
2,
(by Theorem 7.4 and Theorem 6.6).
Also H
2
() C
0
(); H
2
() L
() and P
1

P L
(

K). Thus,
by Theorem 6.6 with k = 1, p = 2, m = 0, q = ,
u
K
u
0,,K
C(meas K)
1
2
h
2
u
2,K
.
Since n = 2,
meas K C
2
K

C
2
h
2
K

C
2
2
h
2
by (7.1), so that (meas K)
1
2
C h
1
and therefore,
u
K
u
0,,K
C hu
2,K
.
Hence we obtain (7.22) for n = 2 since
u
h
u
0,,
= max
Kt
h
u
K
u
0,,K
C hu
2,
.
76
For n = 3, the only variation in the proof occurs in the fact that
u
h

h
u
0,,

C
h
3/2
u
h

h
u
0,
as in Exercise 7.1 and that now
meas K C
3
K

C
3
h
3
K

C
3
3
h
3
.
References: One may refer to Ciarlet and Raviart [6] for 0(h) conver-
gence in the norm
0,,
for any n. See also Bramble and Thom ee
[1].
Chapter 8
Numerical Integration
LET US START with a specic problem. Let be a polygonal domain 77
in R
n
. Consider the problem
(8.1)
_
n
_
i, j=1
x
i
_
a
i j
u
x
j
_
= f in ,
u = 0 on = .
where the (a
i j
) and f are functions over which are smooth enough.
Let us further assume that there exists > 0 such that, for all R
n
,
(8.2)
n
i, j=1
a
i j
j

n
i=1
2
i
.
It has been seen earlier (cf. Remark 2.2) that the above problem (8.1)
is obtained from a problem (P) with a(, ) and f () being dened by
(8.3)
_
_
a(u, v) =
_
_
n
i, j=1
a
i j
u
x
j
v
x
i
dx
f (v) =
_
f v dx
for u, v V = H
1
0
().
Approximating the solution by the nite element method, i.e. by
constructing a regular family of triangulations (t
h
) with reference nite
77
78 8. Numerical Integration
element (

K,

,

P), we get the problems (P
h
), i.e., to nd u
h
V
h
such
that
(8.4) a(u
h
, v
h
) = f (v
h
), for all v
h
V
h
.
If we choose a basis w
k
M
k=1
for V
h
, then we may write
(8.5) u
h
=
M
k=1
u
k
w
k
.
Thus to solve (P
h
) we have to solve the linear system
(8.6)
M
k=1
a(w
k
, w
m
)u
k
= f (w
m
), 1 m M.
78
Notice that
a(w
k
, w
m
) =
Kt
h
_
K
n
i, j=1
a
i j
w
k
x
j
w
m
x
i
dx
f (w
m
) =
Kt
h
_
K
f w
m
dx.
(8.7)
Thus we have ended up with the computations of integrals over K
t
h
. These are, in general, dicult or impossible to evaluate exactly and
one thus has to resort to numerical methods. We now study briey how
this may be done.
Let us assume F
K
(

K) = K, where, F
K
( x) = B
K
x+b
K
, with det B
K
>
0. There is no loss in generality in the last assumption. Then if is a
function over K, we have
(8.8)
_
K
(x)dx = (det B
K
)
_

K
( x)d x
the functions and being in the usual correspondence. We then re-
place the expression in the right-hand side by the following:
(8.9)
_

K
( x)
dx
L
1=1

1
(
b
1
).
8. Numerical Integration 79
In this section will denote the right-hand side replacing the ex-
pression in the left hand side in similar relations). In (8.9) the quan-
tities
1
are called the weights and the points

b
1
are called the nodes
of the quadrature scheme. While in general we may assume
1
R
and

b
1
R
n
, we will restrict ourselves to the most common case where

1
> 0 and

b
1

K, 1 l L.
We may now dene the error functional

E by
(8.10)

E ( ) =
_

K
( x)d x
L
1=1

1
(
b
1
).
79
We will be interested in nding spaces of polynomials for which
E ( ) = 0, i.e., again we need polynomial invariance, an idea already

found in interpolation theory. The above quadrature scheme for

K in-
duces one on K as well since if we set
(8.11)
_
1,K
= (det B
K
)
1
,
b
1,k
= F
K
(
b
1
),
we then deduce the numerical quadrature scheme
(8.12)
_
K
(x)dx
L
1=1
1,K
(b
1,K
).
We shall therefore dene the error functional
(8.13) E
K
() =
_
K
(x)dx
L
l=1
1,K
(b
1,K
).
Notice that the following relation holds:
(8.14) E
K
() = (det B
K
)

E ( ).
Example 8.1. Consider

K to be an n-simplex in R
n
. Let a be its centroid.
Let
_

K
( x)d x (meas

K) ( a).
Exercise 8.1. Show that E ( ) = 0 for P
1
in Example 8.1.
Example 8.2. Let

K be a triangle in R
2
. With the usual notations, set
_

K
( x)d x
1
3
(meas

K)
1i<j3
( a
i j
).
Exercise 8.2. Show that

E ( ) = 0 for P
2
in Example 8.2. 80
Example 8.3. Let

K be as in Example 8.2. Let (cf. Fig. 8.1):
_

K
( x)d x
1
60
(meas

K)
_
_
3
3
i=1
( a
i
) + 8
1i<j3
( a
i j
) + 27 ( a)
_
_
.
Figure 8.1:
Exercise 8.3. Show that

E ( ) = 0 for P
3
, in Example 8.3.
Let us now review the whole situation. We had the original ap-
proximation problem (P
h
): To nd u
h
V
h
such that a(u
h
, v
h
) = f (v
h
)
for all v
h
V
h
.
This led to the solution of the linear system (8.6). By virtue of the
quadrature scheme we arrive at a solution of a modied approximation
problem (P
h
): To solve the linear system
(8.15)
M
k=1
a
h
(w
k
, w
m
)u
k
= f
h
(w
m
), 1 m M,
where
(8.16)
_
_
a
h
(w
k
, w
m
) =
_
Kt
h
_
L
l=1

1,K
_
n
_
i, j=1
a
i j
w
k
x
j
w
m
x
i
(b
1,K
)
_
f
h
(w
m
) =
_
Kt
h
_
L
_
l=1
1,K
( f w
m
)(b
1,K
)
_
.
While u
h
was given by (8.5) we now obtain
(8.17) u
h
=
M
k=1
u
k
w
k
.
Thus the problem (P
h
) (not to be confused with any adjoint prob-
lem!) consists in nding u
h
V
h
such that, for all w
h
V
h
, 81
(8.18) a
h
(u
h
, w
h
) = f
h
(w
h
)
where
(8.19)
_
_
a
h
(v
h
, w
h
) =
_
Kt
h
L
_
l=1
1,K
_
n
_
i, j=1
a
i j
v
h
x
j
w
h
x
i
(b
1,K
)
_
f
h
(v
h
) =
_
Kt
h
L
_
l=1
1,K
( f v
h
)(b
1,K
),
for v
h
, w
h
V
h
.
Remark 8.1. The bilinear form a
h
(, ) : V
h
V
h
R and the linear
form f
h
: V
h
R are not dened over V in general. For instance if
V = H
1
0
()(n = 2) in one of the examples, then as they require point
values of the nodes, we see that they are not in general dened over V.
Having obtained the approximate solution u
h
by numerical integra-
tion, we are naturally interested in its ecacy. Thus we require to know
the error uu
h
. We now carry out the error analysis, rst in an abstract
setting.
Let us maintain our assumptions as in the Lax-Milgram lemma and
consider the problem (P). Then we have problems (P
h
) to nd u
h
V
h

V such that for all v
h
V
h
, a
h
(u
h
, v
h
) = f
h
(v
h
) where f
h
V
h
and a
h
(, )
is a bilinear form on V
h
. Then we would like to answer the following
questions:
(i) What are sucient conditions such that (P
h
) have unique solu-
tions?
(ii) Can we nd an abstract error estimate for u u
h
?
(iii) If uu
h
= 0(h
k
), i.e., without numerical quadrature, under what
conditions is this order of convergence preserved, i.e. when can
we say u u
h
= 0(h
k
)?
The assumption of V
h
-ellipticity of the bilinear forms a
h
(, ) answers 82
the rst question (by the Lax-Milgram lemma) and we will see in The-
orem 8.2 under which assumptions it is valid. The following theorem
answers the second question.
Theorem 8.1. Let the bilinear forms a
h
(, ) be V
h
-elliptic uniformly with
respect to h, i.e., there exists a constant > 0, independent of h, such
that for all h and for all v
h
V
h
,
(8.20) a
h
(v
h
, v
h
) v
h
2
.
Then the approximate problems (P
h
) all have unique solutions u
h
,
and further we have the estimate:
u u
h

C
_
inf
v
h
V
h
_
u v
h
+ sup
w
h
V
h
a(v
h
, w
h
) a
h
(v
h
, w
h
)
w
h
_
+ sup
w
h
V
h
f (w
h
) f
h
(w
h
)
w
h
_
.
(8.21)
Remark 8.2. If a = a
h
, f = f
h
then we get our original estimate (3.1).
Thus (8.21) generalizes our previous result.
Remark 8.3. The terms involving a, a
h
and f , f
h
merely mean that if u
h
is to converge to u, then a
h
and f
h
must be close to a and f respec-
tively. Their convergence to 0 with h may be viewed as consistency
conditions which are so often found in Numerical Analysis.
Proof. The existence and uniqueness of the u
h
are obvious by the Lax-
Milgram lemma applied to the V
h
. Since, for all v
h
V
h
, we have
a
h
(u
h
, u
h
v
h
) = f
h
(u
h
v
h
),
a(u, u
h
v
h
) = f (u
h
v
h
),
we have the identity 83
a
h
(u
h
v
h
, u
h
v
h
) = a(u v
h
, u
h
v
h
) + a(v
h
, u
h
v
h
)
a
h
(v
h
, u
h
v
h
) + f
h
(u
h
v
h
) f (u
h
v
h
). (8.22)
Hence by (8.20) we get
u
h
v
h
2
Mu v
h
u
h
v
h
+ a(v
h
, u
h
v
h
) a
h
(v
h
, u
h
v
h
)
+ f
h
(u
h
v
h
) f (u
h
v
h
).
Thus,
u
h
v
h
Mu v
h
+
a(v
h
, u
h
v
h
) a
h
(v
h
, u
h
v
h
)
u
h
v
h
+
f
h
(u
h
v
h
) f (u
h
v
h
)
u
h
v
h
Mu v
h
+ sup
w
h
V
h
a(v
h
, w
h
) a
h
(v
h
, w
h
)
w
h
+ sup
w
h
V
h
f (w
h
) f
h
(w
h
)
w
h
since (u
h
v
h
) V
h
. Hence,
u u
h
u v
h
+ u
h
v
h
_
1 +
M
_
u v
h
+
1
sup
w
h
V
h
a(v
h
, w
h
) a
h
(v
h
, w
h
)
w
h
+
1
sup
w
h
V
h
f (w
h
) f
h
(w
h
)
w
h
.
Varying v
h
over V
h
and taking the inmum, and replacing (1+M/), 84
(1/) by a larger constant C, we get (8.21), which completes the proof.
The following theorem tells us when the uniform V
h
-ellipticity as-
sumption of Theorem 8.1 is satised in the example we started with.
Theorem 8.2. Let a
h
(, ) and f
h
() be as in (8.19). Assume further that
(i)
1
> 0, 1 l L, (ii)

P P
k
, and (iii)
L
_
l=1
b
l
contains a P
k
1
unisolvent subset. Then the a
h
(, ) are V
h
-elliptic uniformly with respect
to h.
Proof. We must produce an > 0, free of h, such that a
h
(v
h
, v
h
)
v
h
2
1,
for all v
h
V
h
. We have
a
h
(v
h
, v
h
) =
Kt
h
L
l=1
1,K
_
_
n
i, j=1
a
i j
v
h
x
j
v
h
x
i
_
_
(b
l,K
)

Kt
h
L
l=1
l,K
_
_
n
i=1
_
p
K
x
i
(b
l.K
)
_
2
_
_
(8.23)
where p
K
= v
h
K
. The inequality (8.23) is a result of the ellipticity
condition (8.2) on the matrix (a
i j
) and the fact that
l,K
> 0, since
1
>
0 and we assumed without loss in generality that det B
K
> 0. Now let
p
K
(x) = p
K
(x), where x = B
K
x + b
K
. Let B
K
= (b
i j
), so that
x
j
=
n
l=1
b
jl
x
l
+ b
K, j
.
Then
p
K
x
i
=
n
j=1
p
K
(x)
x
j
x
j
x
i
=
n
j=1
p
K
(x)
x
j
b
ji
.
Thus is
D =
_
p
K
( x)
x
1
, . . . ,
p
K
( x)
x
n
_
and D =
_
p
K
(x)
x
1
, . . . ,
p
K
(x)
x
n
_
,
we have

D = DB
K
. Hence

D
2
D
2
B
K
2
. Thus,
(8.24)
n
i=1
_
p
K
x
i
(b
l,K
)
_
2
B
K
2
n
i=1
_
p
K
x
i
(
b
1
)
_
2
.
85
Now suppose p

P is such that
(8.25)
L
l=1

1
n
i=1
_
p
x
i
(
b
1
)
_
2
= 0.
Then since
1
> 0, we have
p
x
i
(
b
l
) = 0 for all 1 l L and 1 i
n. Since

P P
k
, we have
p
x
i
P
k
1
and hence
p
x
i
= 0 by the P
k
1
-
unisolvency. Thus p P
0
, and on the nite dimensional space

P/P
0
(in practice we always have P
0

P) we have a norm dened by the
square-root of the left hand side of (8.25). By the nite dimensionality
this is equivalent to the norm dened by the
1,

K
norm on P. Hence we
have a constant

> 0 such that
(8.26)
L
l=1

l
n
i=1
_
p
x
i
(
b
l
)
_
2

p
2
l,

K
.
We will apply this to p
K
. We also have
(8.27) p
K
2
l,K
CB
1
K

2
(det B
K
)
1
p
K
2
l,K
,
by Theorem 6.4. Combining the inequalities (8.23), (8.24), (8.26) and
(8.27), we get
a
h
(v
h
, v
h
)
Kt
h
(det B
K
)B
K
CB
1
K

2
(det B
K
)
1
p
K
2
l,K
=
Kt
h
(B
K
B
1
K
)
2
p
K
2
l,K

Kt
h
p
K
2
l,K
=
Cv
h
2
1,
= v
h
2
l,
since (B
K
B
1
K
)
2
by Theorem 6.5. This proves the theorem.
Let us now review our Examples 8.1 through 8.3 to see if the condi- 86
tions of Theorem 8.2 are satised.
Let n = 2 and consider Example 8.1. Clearly = meas

K > 0. Also
P = P
1
for triangles of type (1). Since
_
K
= p( a) is P
0
-unisolvent, we
have that for triangles of type (1) and the quadrature scheme of Example
8.1 the corresponding a
h
(, ) are V
h
-elliptic uniformly with respect to h.
For triangles of type (2),

P = P
2
. The weights
1
are all > 0 in
Example 8.2. Further we saw in Exercise 5.1 that a
i j
1i<j3
is P
1
-
unisolvent. Hence Theorem 8.2 is valid for this quadrature scheme as
well.
For triangles of type (3), consider the quadrature scheme of Exam-
ple 8.3. We have
1
> 0 and

P = P
3
. It was seen in Example 4.2 that
the set a
i
; 1 i 3 a
i j
; 1 i < j 3 is P
2
-unisolvent. Hence
the corresponding bilinear forms a
h
(, ) are V
h
-elliptic uniformly with
respect to h.
Exercise 8.4. Let (H, ) be a Hilbert space and V a subspace with norm
such that V H and V = H cf. Sec. 7. Then with the usual
notations show that
u u
h
sup
gH
_
1
g
inf
h
V
h
(Mu u
h

h
+ a(u
h
,
h
) a
h
(u
h
,
h
)
+ f (
h
) f
h
(
h
))
_
where is the solution of adjoint problem for g.
We now turn our attention to the evaluation of the bound for uu
given by (8.21). For second-order problems, for which the norm is
1,
, we will take as usual for v
h
V
h
the element
h
u V
h
so that we
now get the bound
u u
1,
C
_
u
h
u
1,
+ sup
w
h
V
h
a(
h
u, w
h
) a
h
(
h
u, w
h
)
w
h
1,
+ sup
w
h
V
h
f (w
h
) f
h
(w
h
)
w
h
1,
_
.
(8.28)
Let us assume that we may apply Theorem 7.1, so that 87
(8.29) u
h
u
1,
C h
k
u
k+1,
.
In order to keep the same accuracy, we will therefore try to obtain
estimates of the following form:
(8.30)
_
_
sup
w
h
V
h
a(
h
u, w
h
) a
h
(
h
u, w
h
)
w
h
1,
C(u)h
k
,
sup
w
h
V
h
f (w
h
) f
h
(w
h
)
w
h
1,
C( f )h
k
,
and these will in turn be obtained from local estimates. (cf. Theo-
rem 8.4 and Exercise 8.5).
As a preliminary step, we need two results which we prove now.
The rst of these is a historically important result in the interpolation
theory in Sobolev spaces.
Theorem 8.3 (BRAMBLE-HILBERT LEMMA; cf. Bramble and
Hilbert [27]). Let R
n
be open with Lipschitz continuous bound-
ary . Let f W
k+1, p
()) which vanishes over P
k
. Then there exists a
constant C = C() such that, for all v W
k+1, p
(),
(8.31) f (v) C f
k+1, p,
v
k+1, p,
.
Proof. For v W
k+1, p
() and all p P
k
, we have
f (v) = f (v + p),
so that
f (v) = f (v + p) f
k+1, p,
v + p
k+1, p,
,
and thus, 88
f (v) f
k+1, p,
inf
pP
k
v + p
k+1, p,
C f
k+1, p,
v
k+1, p,
,
by Theorem 6.2, which completes the proof.
Lemma 8.1. Let W
m,q
(), w W
m,
(). Then w W
m,q
(), and
there exists a numerical constant C, independent of and w such that
(8.32) w
m,q,
C
M
j=0
mj,q,
w
j,,
.
Proof. The result is an immediate consequence of the Leibniz formula:
For any = m,
(w) =
m
j=0
=j
C
,
()
(w)
which yields (8.32).
We may now apply Lemma 8.1 and Theorem 8.3 to get the estimates
(8.30). We do this in two stages (Theorem 8.4 and Exercise 8.5) in
which, for the sake of simplicity, we present our results for the special
case P
K
= P
2
.
Theorem 8.4. Let P
K
= P
2
and consider a quadrature scheme such that
for all P
2
,

( ) = 0. Then there exists a constant C, independent of
K, such that for all a
i j
W
2,
(K) and for all p, p
P
K
we have
(8.33) E
K
_
(a
i j
)
p
x
j
p
x
i
_
C h
2
K
a
i j
2,,K
p
x
j
1,K
x
i
0,K
.
Proof. Since we have
p
x
j
,
p
x
i
P
1
, it suces to nd a suitable esti-
mate for E
K
(avw), for a W
2,
(K), v, w P
1
. Further, since
(8.34) E
K
(avw) = (det B
K
)

E ( a v w),
we will rst nd an estimate for

E ( a v w), with a W
2,
(

K) and v, 89
w P
1
. Let
0
w be the orthogonal projection of w onto the subspace P
0
in the sense of L
2
(

K). Then we may write
(8.35)

E ( a v w) =

E ( a v
0
w) +

E ( a v( w
0
w)).
(i) Estimate for

E ( a v
0
w).
Consider the functional

E : W
2,
(

K) R dened by

E (
) =
_

K
( x)d x =
L
l=1

1

(
b
1
).

E (
)

C
0,,

K

C
2,,

K
. Thus

E is a continuous linear func-
tional on W
2,
(

K). Hence by Theorem 8.3, since

E vanishes on P
1
(
P
2
)
1
, we have a constant

C such that
(8.36)

E (
)

C
2,,

K
.
Thus

E ( a v
0
w)

C a v
0
w
2,,

K

C a v
2,,

K

0
w
0,,

K
since
0
w P
0
is a constant function. By Lemma 8.1 (recall that v
P
1
),

E ( a v
0
w)

C
0
w
0,,

K
_
a
1,,

K
v
1,,

K
+ a
2,,

K
v
0,,

K
_
.
By the equivalence of the L
2
and L
norms on P
0
, and since the
projection has norm less than that of the vector itself in any Hilbert
space we have the chain of inequalities

0
w
0,,

K

C
0
w
0,

K

C w
0,

K
.
90
Similarly we may replace v
1,,

K
by v
1,

K
and v
0,,

K
by v
0,

K
since
the L
2
and L
norms are equivalent on P

1
. Thus we get
(8.37)

E ( a v
0
w)

C( a
1,,

K
v
1,

K
+ a
2,,

K
v
0,

K
) w
0,

K
.
1
In this estimate, we do not use the full polynomial invariance of the quadrature
scheme.
(ii) Estimate for

E ( a v( w
0
w)).
Let w P
1
be xed and let W
2,
(

K). Then

E ( ( w
0
w))

C ( w
0
w)
0,,

K

C
0,,

K
w
0
w
0,,

K

C w
0
w
0,,

K

2,,

K
.
Thus the functional on W
2,
(

K) dened by

E ( ( w
0
w)) is
continuous, linear with norm

C w
0
w
0,,

K
. Since for P
1
, ( w

0
w) P
2
, we have that the functional vanishes on P
1
. By Theorem 8.3,

E ( ( w
0
w))

C w
0
w
0,,

K

2,,

K
.
Set = a v. Now,
a v
2,,

K

C( a
2,,

K
v
0,,

K
+ a
1,,

K
v
1,,

K
).
Again we may use the equivalence between the L
-norms of v and
w
0
w and the L
2
-norms of the same functions as in (i) since they be-
long to the nite dimensional space P
1
. Also, by the triangle inequality,
w
0
w
0,

K

C w
0,

K
.
Thus we get
(8.38)

E ( a v( w
0
w))

C( a
2,,

K
v
0,

K
+ a
1,,

K
v
1,

K
) w
0,

K
.
91
(iii) We can now complete the proof. Recall that E
K
(avw) = (det B
K
)
E ( a v w). Also,
a
m,,

K
C h
m
K
a
m,,K
v
2m,

K
C h
2m
K
(det B
K
)
1
2
v
2m,K
.
w
0,

K
C(det B
K
)
1
2
w
0,K
,
(8.39)
by Theorems 6.4 and 6.5. Combining (8.37), (8.38) and (8.39),
we get
E
K
(avw) C h
2
K
(a
1,,K
v
1,K
+ a
2,,K
v
0,K
)w
0,K
C h
2
K
a
2,,K
v
1,K
w
0,K
.
Setting a = A
i j
, v =
p
x
j
, w =
p
x
i
we obtain (8.33), thus completing
the proof.
We leave the second stage as an exercise:
Exercise 8.5. Let P
K
= P
2
and let the quadrature scheme be such that
E ( ) = 0 for all P
2
. Then show that for q such that W
2,q
(K)
C
0
(K), there exists C independent of K such that for all f W
2,q
(K)
and all p P
K
,
E
K
( f p) C h
2
K
(det B
K
)
1
2
1
q
f
2,q,K
p
1,K
.
[Hint: If
1
is the orthogonal projection to P
1
in the L
2
-sense then write
E (

f p) = E (

f
1
p) +

E (

f ( p
1
p))].
Remark 8.4. The inclusion W
2,q
(K) C
0
(K) is true if, for instance,
2
n
q
> 0, by the Sobolev imbedding theorem.
We now come to the nal stage in the estimation of u u
h
.
Theorem 8.5. Let (t
h
) be a regular family of triangulations on by 92
n-simplices of type (2). Let us assume that the V
h
-ellipticity is uniform
with respect to h. Let

E ( ) = 0 for all P
2
. Then if u H
3
()
C
0
()(n 5), a
i j
W
2,
() and f W
2,q
() for some q 2, we have
the estimate
(8.40) u u
1,
C h
2
[u
3,
+ f
2,q,
].
Proof. We estimate the various quantities in (8.28). We have:
a(
h
u, w
h
) a
h
(
h
u, w
h
)
Kt
h
n
i, j=1
E
K
(a
i j
(
h
uK)
x
j
(w
h
K)
x
i
Kt
h
n
i, j=1
C h
2
K
a
i j
2,,K
(
h
u
K
)
x
j
1,K
(w
h
K)
x
i
0,K
h
2
n
i, j=1
a
i j
2,,
_
Kt
h
(
h
uK)
x
j
2
1,K
_
_
1
2
_
Kt
h
(w
h
K)
x
i
2
0,K
_
_
1
2
(since h
K
h, and we may apply the Cauchy-Schwarz inequality)
C h
2
h
u
2,
w
h
1,
Now,
h
u
2,
u
2,
+ u
h
u
2,
Cu
2,
,
using Theorem 6.3 with P
1
P
K
= P
2
. Therefore, for all w
h
V
h
, we
have
(8.41)
a(
h
u, w
h
) a
h
(
h
u, w
h
)
w
h
1,
Ch
2
u
2,
.
Similarly, we have
f (w
h
) f
h
(w
h
)
Kt
h
E
K
( f w
h
K
)
Kt
h
C h
2
K
(meas K)
1
2
1
q
f
2,q,K
w
h
1,K
.
93
Since q 2,
1
2

1
q
0 and by the general H olders inequality,
Kt
h
(meas K)
1
2
1
q
f
2,q,K
w
h
1,K
Kt
h
meas K
_
_
1
2
1
q
_
Kt
h
f
q
2,q,K
_
_
1
q
_
Kt
h
w
h
2
1,K
_
_
1
2
= C f
2,q,
w
h
1,
.
Hence we get, fot all w
h
V
h
,
(8.42)
f (w
h
) f
h
(w
h
)
w
h
1,
Ch
2
f
2,q,
.
Combining (8.28), (8.29), (8.41) and (8.42) we get (8.40), thus com-
pleting the proof.
Remark 8.5. The condition n 5 (needed for the continuous inclusion
H
3
() C
0
()) was already necessary for the denition of
h
u.
References: For a survey on numerical quadrature in general one may
refer to Habers survey article [13]. For application to the nite element
method, the Sec. 4.2 of the book by Strang and Fix [22] or Chapter 2 of
the forthcoming book of Ciarlet and Raviart [5].
Chapter 9
The Obstacle Problem
In Sec. 2 we cited the Obstacle Problem as an example of a non-linear 94
abstract problem of Sec. 1. Let us recall a few facts about this to start
with.
Consider an elastic membrane (cf. Fig. 9.1) stretched over an open
set R
2
and xed along the boundary which is assumed to be
Lipschitz continuous. Let a force of density Fdx act on the membrane.
Let us assume the existence of an obstacle given by (x), for x .
Then vertical displacement given by u is the solution of the abstract
problem where
(9.1)
_
_
a(u, v) =
_
2
_
i=1
u
x
i
v
x
i
dx
f (v) =
_
f v dx, f L
2
()
for u, v V = H
1
0
(), where f = F/t, t being the tension. The subset K
is given by
K = v H
1
0
(); v a.e. in.
If v
1
, v
2
are in K and v
i
< in A
i
with meas A
i
= 0 for i = 1, 2, then
v
1
+ (1 )v
2
on (A
1
A
2
)
c
i.e. the complement of A
1
A
2
and
meas (A
1
A
2
) = 0. Thus K is convex. If v K, let v
n
K such that
v
n
v in H
1
0
(). Let v
n
in A
c
n
, meas A
n
= 0. Then all the v
n
are
on (
n
A
n
)
c
and meas (
n
A
n
) = 0. Hence v a.e. as well. Thus
95
96 9. The Obstacle Problem
v K and K is closed as well. We have the regularity assumption that
H
2
(). Of course it is
Figure 9.1:
95
The solution u satises
(9.2) J(u) = min
vK
J(v),
where J(v) =
1
2
a(v, v) f (v) and is also characterized by the variational
inequalities (cf. Theorem 1.1):
(9.3) a(u, v u) f (v u), for all v K.
We proposed as a problem to show that this problem is interpreted
as the following classical problem (assuming u H
1
0
() H
2
()).
(9.4)
_
_
u in ,
u = f where u > ,
u = 0 on .
We have a few regularity results which are listed below:
(i) If is convex and is a C
2
-boundary then u H
1
0
() H
2
().
(ii) If f = 0 and a convex polygon then also u H
1
0
() H
2
().
(iii) The norm u
2,
is bounded above by a function of f
0,
and
2,
in cases (i) and (ii). 96
9. The Obstacle Problem 97
Our aim in this section is to use the nite element method to approx-
imate this problem and obtain error estimates. We list our assumptions
now:
Let be a convex polygon, f L
2
(), H
2
() and let u
H
2
() H
1
0
().
Remark 9.1. We cannot assume any more smoothness on u other than
H
2
(). For instance in the 1-dimensional case if f = 0, and even if
the function is very smooth the points of contact of u with will have
discontinuous second derivatives in general; cf. Fig. 9.2.
Figure 9.2:
With the above assumptions we proceed to the approximate prob-
lems, rst in the abstract setting, as usual.
We have the problems (P
h
) associated with the subspaces V
h
V =
H
1
0
(). We now choose closed convex subsets K
h
V
h
. One has to bear
in mind that, in general, K
h
K (we will see that this is the case in our
approach, sub-sequently).
We nd u
h
K
h
such that for all v
h
K
h
,
(9.5) a(u
h
, v
h
u
h
) f (v
h
u
h
).
The existence and uniqueness of the u
h
follow from Theorem 1.1.
Let H be a Hilbert space with norm and inner-product (, ). Let 97
(V, ) be a subspace such that V H, V = H. Then, as usual, if we
identify H
and H, then H will be identied with a subspace of V
. (We
will take V = H
1
0
() and H = L
2
()). Also as in Sec. 1 (cf. proof of
Theorem 1.2), for all u, v V we have
(9.6) a(u, v) = (Au)(v),
where A : V V
is a linear map. We now pass on to an abstract error

bound.
Theorem 9.1 (FALK). Assume that f H, Au H. Then there exists a
constant C, independent of V
h
and K
h
, such that
(9.7) u u
h
C
_
inf
v
h
K
h
(u v
h
2
+ u v
h
) + inf
vK
u
h
v
_ 1
2
.
(Note: The condition Au H = L
2
() is satised if u H
2
() since
Au = u L
2
().)
Proof. Let stand for the V
h
-ellipticity constant. Then
u u
h
2
a(u u
h
, u u
h
)
= a(u, u) + a(u
h
, u
h
) a(u, u
h
) a(u
h
, u). (9.8)
For any v K and any v
h
K
h
, by (9.3) and (9.5), we have
(9.9)
_
_
a(u, u) a(u, v) + f (u v),
a(u
h
, u
h
) a(u
h
, v
h
) + f (u
h
v
h
).
Substituting in (9.8) we get
u u
h
2
a(u, v) + f (u v) + a(u
h
, v
h
)
+ f (u
h
v
n
) a(u, u
h
) a(u
h
, u)
= a(u, v u
h
) f (v u
h
) + a(u, v
h
u) f (v
h
u)
+ a(u
h
u, v
h
u)
= ( f Au, u v
h
) + ( f Au, u
h
v) + a(u
h
u, v
h
u)
f Au u v
h
+ f Au u
h
v + Mu
h
u v
h
u.
98
Notice that since (
_

M
u u
h

_
M
u v
h
)
2
0, we have
u u
h
u v
h

1
2
_

M
u u
h
2
+
M
u v
h
2
_
,
and hence
u u
h
2
C[u v
h
+ u
h
v] +
2
u u
h
2
+
M
2
2
u v
h
2
Or,
u u
h
2
C[(u v
h
+ u v
h
2
) + u
h
v].
Varying v
h
K
h
and v K and extracting the square root after
taking the inma we get (9.7). This completes the proof.
Remark 9.2. If we have a linear problem then f = Au gives the solution
and we get the original bound (3.1).
Remark 9.3. From (9.8) we see that this estimate holds even if a(, ) is
not symmetric.
We now apply this to the specic membrane problem. Maintaining
our assumptions on , let t
h
be a triangulation by triangles of type (1),
and let V
h
be the corresponding subspace of V = H
1
0
().
Remark 9.4. It is of no practical use if we go to more sophisticated nite
elements, unlike the linear problem. Since u H
2
() is the maximum
smoothness, we may atmost use our abstract estimate theorems only on
the spaces P
1
.
One may be tempted to try for K
h
those v
h
which are a.e. in .
However this is not of value from numerical and computational points
of view for we do not easily know where exactly our piecewise linear 99
solution functions would touch . We set instead
(9.10) K
h
= v
h
V
h
; At all nodes b of t
h
, v
h
(b) (b).
Remark 9.5.
Figure 9.3:
As seen in Fig. 9.3, though for nodes b, v
h
(b) (b), it does not
guarantee that v
h
a.e. Thus we see that K
h
K. Now the relation
(9.10) is very easy to implement using the computer.
We now have our main result on the error bound.
Theorem 9.2 (FALK). There exists a constant C depending on f
0,
and
2,
such that for a regular family of triangulations (t
h
) as above
we have
(9.11) u u
h
1,
C h.
Remark 9.6. The order of convergence is therefore the same as that for
the linear problems when we use piecewise linear approximations.
Proof. By Theorem 9.1,
u u
h
1,
C
_
inf
v
h
K
h
_
u v
h
2
1,
+ u v
h
0,
_
+ inf
vK
u
h
v
0,
_ 1
2
.
(i) We rst estimate the inmum over K
h
. Note that if v
h
=
h
u, then 100
v
h
V
h
. Also for all nodes b, v
h
(b) =
h
u(b) = u(b) (b). Thus
v
h
K
h
as well. Thus,
inf
v
h
K
h
_
u v
h
2
1,
+ u v
h
0,
_
u
h
u
2
1,
+ u
h
u
0,
C h
2
_
u
2
2,
+ u
2,
_
.
(ii) For the inmum over K, consider v
1
= max(u
h
, ). Clearly v
1

and v belongs to H
1
() because u
h
and also belong to H
1
()
(this is a non-trivial result which we assume here). Hence v
1
K,
and thus,
inf
vK
u
h
v
0,
u
h
v
1
0,
.
We have
u
h
v
1
2
0,
=
_
h
u
h

2
dx, where
h
= x; (x) u
h
(x).
If
h
is the V
h
-interpolate of , then for all nodes b, u
h
(b) (b) =
h
(b). Since both u
h
and
h
are piecewise linear, we may now assert
that u
h

h
everywhere. Thus u
h
h
0 on . Thus for all x
h
,
we have
0 < ( u
h
)(x) = ( u
h
)(x) (
h
)(x)
(
h
)(x)
and for x
h
, ( u
h
)(x) = 0, so that
u
h
v
1
0,

_
_
_
2
dx
_
_
1
2
=
h
=0,
C h
2
2,
.
Hence
u u
h
1,
C(h
2
)
1
2
= Ch,
where C depends on
2,
and u
2,
. However the regularity result (iii) 101
helps us to bound u
2,
above by a constant C depending on f
0,
and
2,
which completes the proof of the theorem.
References: Two important references are Falk [11] and [12]. For
regularity results refer Brezis and Stampacchia [3] and Lewy and Stam-
pacchia [16].
Another references is Mosco and Strang [19].
Chapter 10
Conforming Finite Element
Method for the Plate
Problem
In Sec. 2 (cf. Example 2.4), as an example of fourth-order problem, we 102
described the plate problem. In abstract terms it is to nd the solution of
a(u, v) = f (v), for all v V,
where
(10.2)
_
_
K = V = H
2
0
(), R
2
,
a(u, v) =
_
_
u v + (1 )
_
2

2
u
x
1
x
2
2
v
x
1
x
2

2
u
x
2
1
2
v
x
2
2

2
u
x
2
2
2
v
x
2
2
__
dx,
f (v) =
_
f v dx, f L
2
().
The problem was interpreted as the classical boundary value prob-
lem
(10.3)
_
2
u = f in ,
u =
u
= 0 on = .
103
104 10. Conforming Finite Element Method for the Plate Problem
Remark 10.1. It was commented in Sec. 2 that the second term of the
integrand in the denition of a(, ) does not contribute to the dierential
equation. Our method here will be equally applicable to both the cases.
viz. with the second term present (the plate problem) or with that term
absent (as it happens in Hydro-dynamics). In our next section, on non-
conforming methods, we will see that the second term is essential in
order that we may apply that method.
We assume that is a polygonal domain in R
2
. We saw in Sec. 3
that for fourth-order problems we need the inclusion V
h
C
1
(). (cf.
Exercise 3.1). Thus we need to use nite elements of class C
1
, such as
the Argyris triangle, the Bogner-Fox-Schmidt rectangle and so on (cf.
Sec. 4).
When such nite elements can be imbedded in an ane family, then 103
we have the approximation theory, for regular families of triangula-
tions, available to us. We show that this is the case for the Bogner-
Fog-Schmidt rectangle. However for the Argyris triangle or for the 18-
degree-of-freedom triangle such an imbedding is not possible and we
have to modify the usual argument to obtain error estimates. The min-
imal assumptions for 0(h) convergence in the
2,
norm are that
P
2
P and that u H
3
() H
2
0
(). We have that if is a convex
polygon and if f L
2
(), then u H
3
() H
2
0
(). This result is due
to Knodrat ev.
We will go through the various examples of triangulations of class
C
1
and study convergence in these cases.
Example 10.1. The Bogner-Fog-Schmidt rectangle (cf. Exercise 4.9).
Let P
K
= Q
3
(dim P
K
= 16). We then have (cf. Fig. 10.1):
(10.4)
K
=
_
p(a
i
),
p
x
1
(a
i
),
p
x
2
(a
i
),
2
p
x
1
x
2
(a
i
); 1 i 4
_
.
10. Conforming Finite Element Method for the Plate Problem 105
Figure 10.1:
Equivalently, one may also use the P
K
-unisolvent set
K
=
_
p(a
i
), Dp(a
i
)(a
i+1
a
i
), Dp(a
i
)(a
i1
a
i
),
D
2
p(a
i
)(a
i+1
a
i
, a
i1
a
i
); 1 i 4
_
(10.5)
(all indices being read modulo 4). 104
Recall that for an ane family of nite elements, the degrees of
freedom p(a
0
i
), Dp(a
1
i
)(
1
ik
), (D
2
p(a
2
i
)(
2
ik
,
2
il
) are such that (cf. Sec. 5):
(10.6)
_
_
a
0
i
= F( a
0
i
), . . . , a
2
i
= F( a
2
i
),
1
i,k
= B
K
1
i,k
, . . . ,
2
i,1
= B
K
2
i,1
for then
K
v = v which is essentially what we need for the abstract
error analysis.
In
_
K
note that
(10.7) a
i+1
a
i
= F( a
i+1
) F( a
i
) = B
K
( a
i+1
a
i
),
and so on. Thus it is clear that this rectangle can be imbedded in an
ane family of nite elements. Now P
K

P Q
3
for k = 3. By our
abstract error analysis, we therefore have
(10.8) u u
h
2,
Ch
2
u
4,
.
assuming sucient smoothness on u as usual.
A word about the boundary conditions. As in Exercise 3.1, we get
that V
h
H
2
0
() if v =
v
= 0 on , for v V
h
. Thus in choosing our
basis functions we must assure ourselves that this condition is satised.
This in turn depends on the values at the boundary nodes. Let b and c be
two nodes on such that the line joining them is parallel to (say) the x
1
-
axis. Since we need v = 0 on this line, and since v will be a polynomial
in x
1
of degree 3 on this line we must have v(b) = v(c) = 0,
v
x
1
(b) =
v
x
1
(c) = 0. Also since we need
v
= 0 on this line and since

v
=
v
x
2
is a polynomial in x
1
of degree 3, we need to set
v
x
2
(b) =
v
x
2
(c) = 0
and

2
v
x
1
x
2
(b) =
2
v
x, x
2
(c) = 0. Thus the degrees of freedom on all 105
boundary nodes must be zero. The only free or unknown parameters
are the degrees of freedom at the interior nodes. This takes care of the
boundary conditions.
Let us now turn to the Argyris triangle (cf. Example 4.7).
Figure 10.2:
We recall that P
K
= P
5
, dimP
K
= 21, and
_
K
is given by (cf.
Fig. 10.2)
(10.9)
K
=
_
_
p(a
i
),
p
x
1
(a
i
), . . . ,
3
p
x
2
2
(a
i
), 1 i 3;
p
(a
i j
), 1 i < j 3
_
_
.
We may replace the rst and second derivative values at the vertices
by Dp(a
i
)(a
i+1
a
i
), Dp(a
i
)(a
i1
a
i
), D
2
p(a
i
)(a
i+1
a
i
, a
i+1
a
i
),
D
2
p(a
i
)(a
i+1
a
i
, a
i1
a
i
), D
2
p(a
i
)(a
i1
a
i
, a
i1
a
i
) in order to get
degrees of freedom for which the relations of the type (10.6) may be
satised. However, one cannot replace the normal derivatives
p
(a
i j
),
1 i < j 3, by such quantities since ane transformations do not
preserve orthogonality.
In order to estimate the errors we describe an intermediary nite
element:
Example 10.2. The Hermite Triangle of Type (5).
Let P
K
= P
5
(dim P
K
= 21). Dene
K
=
_
p(a
i
), Dp(a
i
)(a
i+1
a
i
), Dp(a
i
)(a
i1
a
i
),
D
2
p(a
i
)(a
i+1
a
i
, a
i+1
a
i
), D
2
p(a
i
)(a
i+1
a
i
, a
i1
a
i
),
D
2
p(a
i
)(a
i1
a
i
, a
i1
a
i
), 1 i 3;
Dp(a
i j
)(a
k
a
i j
), 1 i < j 3, k 1, j
_
.
106
That is to say, the only change compared to the Argyris triangle is
that we have replaced the normal derivatives at a
i j
by the derivatives
along the line joining a
i j
to a
k
, the opposite vertex.
Symbolically we can represent such a triangle as in Fig. 10.3.
Figure 10.3:
This element can be put in an ane family as is readily seen. If
K
is the associated interpolation operator, our error analysis yields
(10.10) v
K
v
m,K
C
h
6
K
m
K
v
6,K
, 0 m 6,
for v H
6
(K).
Remark 10.2. Though the Hermite triangle of type (5) yields an ane
family, one cannot use it since V
h
C
0
() but, in general, V
h
C
1
()
as is necessary for fourth-order problems. This is so because the adja-
cent triangles will not patch up, in general, in their derivatives along the
medians; cf. 10.4.
Figure 10.4:
107
Again we show how to take care of the boundary conditions in the
Argyris triangle. We need again v =
v
= 0 on . Let us have two nodes

b, b
, the vertices of a triangle lying on with mid-point c. On this line

v will be a polynomial of degree 5 in , an abscissa along this line.
v
will be a polynomial in of degree 4 on this line. Hence for v = 0

on we need to set, v(b) = v(b
) = 0,
v
(b) =
v
(b
) = 0,

2
v
2
(b) =
2
v
2
(b
) = 0. For
v
= 0 on we set,
v
(b) =
v
(b
) =
v
(c) = 0,
2
v
(b) =
2
v
(b
) = 0. Thus the only free or unknown parameters

are

2
v
2
at vertices on and the degrees of freedom at all interior nodes.
We now get an error estimate when we have triangulations of Ar-
gyris triangles. We use our usual terminology more loosely here
1
. By a
regular family of triangulations made up of Argyris triangles we mean
that all t
h
consist only of Argyris triangles and that for all K,
h
K
K
, a
constant. We also assume that if h = max
Kt
h
h
K
, then h 0.
Theorem 10.1. For a regular family (t
h
) of triangulations made up of
Argyris triangles
(10.11) v
h
v
m,
C h
6m
v
6,
, 0 m 6.
108
Proof. Let us denote the opposite vertex of a
i j
(i < j) by a
k
. Let

K
be the unit outernormal at a
i j
and

K
be the unit vector along the line
[a
i
, a
j
], at a
i j
(cf. Fig. 10.5).
Figure 10.5:
Let
K
be the interpolation operator for the Argyris triangle K and
let
K
be that for the corresponding Hermite triangle of type (5).
1
Because we have to drop the assumption that all the nite elements are ane equiv-
alent to a reference nite element.
Set =
K
v
K
v. Then P
5
. Now
K
(a
i j
) =
K
(
K
v
K
v)(a
i j
) =
K
(v
K
v)(a
i j
).
Also since
K
v =
K
v along any side of K (since the values of
these polynomials of degree 5 as well as those of their rst and second
derivatives agree at the end-points), we have

K
= 0.
Since
D(a
i j
)(a
k
a
i j
) =
K
(a
i j
)(a
k
a
i j
,

K
) +
K
(a
i j
)(a
k
a
i j
,

K
),
where (, ) is the Euclidean inner-product, substituting for

K
and

K
at a
i j
, we get
(10.12) D(a
i j
)(a
k
a
i j
) =
(v
K
v)(a
i j
)(a
k
a
i j
,

K
)
Since P
5
, using the unisolvency in the Hermite triangle we may 109
express in terms of its basis functions. Since all degrees of freedom
except those of the type D(a
i j
)(a
k
a
i j
) are zero for , we have
(10.13) =
1i<j3
ki,kj
K
(v
K
v)(a
i j
)(a
k
a
i j
,

K
)p
i jk
.
Now
(10.14) (a
k
a
i j
,

K
) a
k
a
i j

K
h
K
,
and
K
(v
K
v)(a
i j
) v
K
v
1,,K
C(meas K)
1
2
h
6
K
K
v
6,K
.
(Theorem 6.6 with m = 1, k = 5, p = 2, q = ). Also, meas K C
2
K
,
and we have
(10.15)
K
(v
K
v)(a
i j
) C
h
6
K
2
K
v
6,K
.
Finally, by Theorem 6.4 and 6.5
(10.16) p
i jk
m,K
C
h
K
m
K
p
i jk
m,

K
.
Combining (10.14), (10.15) and (10.16), we get
m,K

1i<j3
ki,kj
K
(v
K
v)(a
i j
) (a
k
a
i j
,

K
) p
i jk
m,K
C
h
8
K
m+2
K
v
6,K
, (10.17)
and hence,
v
K
v
m,K
v
K
v
m,K
+
m,K
C
h
6
K
m
K
_
_
1 +
h
2
K
2
K
_
_
v
6,K
(Using (10.10) and (10.17))
C h
6m
v
6,K
(since h
K
h,
h
K
K
).
110
This on summing over K gives (10.11), thus completing the proof.
Exercise 10.1. Perform the same analysis for the 18-degree-of freedom
triangle. (cf. Exercise 4.7).
For the interpolation theory of the HCT-triangle (cf. Exercise 4.8),
the normal derivatives are handled as in the present case. However the
arbitrariness of the interior point is an obstacle to be overcome. For a
discussion of this, see Ciarlet [4].
Another nite element, similar in its principle to the HCT-triangle
used in the conforming nite element method for the plate problem is
the Fraeijs de Veubeke and Sander Quadrilateral. See Ciavaldini and
N ed elec [9].
These are all essentially the nite elements used in the conforming
methods to approximate the plate problem (we will dene such methods
at the beginning of Sec. 11).
Chapter 11
Non-Conforming Methods
for the Plate Problem
WE START WITH a brief classication of nite element methods. The 111
rst class of methods are called conforming methods, which we have
described upto now, except when we considered numerical integration.
The second class consists of methods other than conforming. In the
latter class we have the Non-conforming methods included:
Given the abstract problem, the conforming methods deal with the
nding of subspaces V
h
V and solving the problems
(P
h
) a
h
(u
h
, v
h
) = f
h
(v
h
), for all v
h
V
h
,
where a
h
= a and f
h
= f for all h and u
h
V
h
is the required solution.
When we employ methods other than conforming we commit, in
the terminology of G. Strang, Variational Crimes. (See Strang and
Fix [22]). These may occur in the following ways:
(i) When performing numerical integration, we may have a
h
and f
h
dierent from a and f respectively. However, V
h
is a subspace of
V;
(ii) The boundary of may be curved. In this case triangles lying
in the interior will be triangles of straight edges while those meet-
ing the boundary will have curved edges like parabolas. These are
113
114 11. Non-Conforming Methods for the Plate Problem
the so-called isoparametric nite elements. Hence if
h
is the
union of the nite elements of the triangulation t
h
, then, in gen-
eral,
h
and consequently V
h
V (where V
h
is a space of
functions dened over
h
), a
h
a, f
h
f ; for a discussion of
these, see Ciarlet and Raviart [30], [31].
(iii) When employing non-conforming methods (which will be dealt
with subsequently) though
h
= , f
h
= f , we will have V
h
V
and a
h
a.
(iv) One may employ any combination of the above three. 112
Let us return to the plate problem. For a conforming method we
need the inclusion V
h
H
2
0
() which essentially results from the in-
clusion V
h
C
1
(). Because of this necessity, when compared with
second-order problems, we either have the dimension of P
K
large (as
in the case of the Argyris triangle) or that the structure of P
K
is compli-
cated (as in the HCT-triangle). Also one would like to have just P
K
= P
2
since u is only in H
3
() in most cases, but this is impossible by the
Zeni sek result (cf. Remark 4.3) which stresses that at least polynomials
of degree 5 must be present in P
K
.
Hence the desire to surmount these diculties led to the devising of
non-conforming methods, essentially developed by the Engineers.
Since the root of all trouble is the inclusion V
h
H
2
0
(), we drop
this condition. Thus we start with V
h
C
0
() and it is much easier
from the computer programme view point. This of course, works only
for a few nite elements, and we describe one of them.
Example 11.1. The Adinis rectangle; cf. Fig. 11.1.
11. Non-Conforming Methods for the Plate Problem 115
Figure 11.1:
The element K consists of a rectangle with vertices a
i
, 1 i 4;
the space P
K
is given by P
K
= P
3
x
1
x
3
2
x
3
1
x
2
, by which we mean
polynomials of degree 4 whose only fourth-degree terms are those
involving x
1
x
3
2
and x
3
1
x
2
. Thus P
3
P
K
. We have the set of degrees of
freedom:
K
=
_
p(a
i
),
p
x
1
(a
i
),
p
x
2
(a
i
), 1 i 4
_
.
113
Of course this element can be used only for plates with sides parallel
to the coordinate axes, such as rectangular plates.
Exercise 11.1. Show that in Example 11.1,
_
K
is P
K
-unisolvent and
that Adinis rectangle is a nite element of class C
0
, and, in general, not
of class C
1
.
Thus we get a priori that V
h
H
1
(). For the boundary condition,
we set all degrees of freedom on the boundary nodes as zero. This gives
us that V
h
H
1
0
(). Thus the only unknown or free parameters are
the degrees of freedom at the interior nodes. Note that
v
h
is zero only
at the boundary nodes, in general.
In the abstract problem, we have a(, ) and f () given by
(11.1)
_
_
a(u, v) =
_
_
uv + (1 )
_
_
2

2
u
x
1
x
2
2
v
x
1
x
2
2
u
x
2
1
2
v
x
2
2
2
u
x
2
2
2
v
x
2
1
_
_
_
_
dx,
f (v) =
_
f vdx, f L
2
().
The second integral is dened over V
h
as well. Thus for the discrete
problem (P
h
) we may set f
h
= f . However while for u, v V
h
the rst
integral is dened over each K t
h
, we cannot dene it over , since
we get Dirac measure-like terms along the boundary. To get over this,
we now dene
a
h
(u
h
, v
h
) =
Kt
h
_
K
_
u
h
v
h
+ (1 )
_
2

2
u
h
x
1
x
2
2
v
h
x
1
x
2
2
u
h
x
2
1
2
u
h
x
2
2
2
v
h
x
2
1
__
dx
=
Kt
h
_
K
_
u
h
v
h
+ (1 )
_
2
u
h
x
2
1
2
v
h
x
2
1
+
2
u
h
x
2
2
2
v
h
x
2
2
+ 2

2
u
h
x
1
x
2
2
v
h
x
1
x
2
__
dx,
(11.2)
and we have the discrete problem (P
h
): To nd u
h
V
h
such that for all
v
h
V
h
(11.3) a
h
(u
h
, v
h
) = f (v
h
).
114
We now prove the existence and uniqueness of the solution u
h
for
(P
h
). We dene on V
h
the seminorm
(11.4) v
h
h
=
_
Kt
h
v
h
2
2,K
2
2,K
_
_
1
2
.
Notice that this may be dened over V = H
2
0
() as well and for
v V, v
h
= v
2,
. In the same way for u, v V, a
h
(u, v) = a(u, v).
We now show that the seminorm
h
is indeed a norm on V
h
. Let
v
h
V
h
with v
h
h
= 0. This gives that
v
h
x
1
= constant over any K. But
given adjacent nite elements the value of
v
h
x
1
at the common vertices
coincide and hence
v
h
x
1
is constant over . But this is zero on the
boundary nodes. Hence
v
h
x
1
= 0 on . Similarly
v
h
x
2
= 0 on . Since
V
h
C
0
() and v
h
= 0 on , the above conditions give that v
h
= 0 over
. Thus (11.4) denes a norm on V
h
.
To show the existence and uniqueness of the solution of (P
h
), we
show that a
h
(, ) is V
h
-elliptic. In fact we do more than this. We show
that the a
h
(, ) are V
h
-elliptic uniformly with respect to h.
Recall that from physical considerations, 0 < <
1
2
(see Sec. 2).
Now
a(v
h
, v
h
) =
Kt
h
_
K
(v
h
)
2
dx + (1 )v
h
2
h
(1 )v
h
2
h
.
(11.5)
Remark 11.1. It was mentioned in passing in Sec. 10 that in order to
apply non-conforming methods one needed the second term involving
in the integral dening a(, ). The uniform V
h
-ellipticity could not be
got in the Hydrodynamical case where this term is absent.
We now proceed with the abstract error analysis. 115
Theorem 11.1 (STRANG). Let a
h
(, ) be V
h
-elliptic uniformly with re-
spect to h with > 0 so that for all v
h
V
h
(11.6) a
h
(v
h
, v
h
) v
h
h
.
Let in addition, there exist

M such that for all u
h
, v
h
V
h
(11.7) a(u
h
, v
h
)

Mu
h
h
v
h
h
.
Assume that a
h
= a and
h
= on V. (These are needed to
extend the denition of a
h
and
h
to V). Then there exists a constant
C, independent of h, such that
(11.8) u u
h
h
C
_
inf
v
h
V
h
u v
h
h
+ sup
w
h
V
h
f (w
h
) a
h
(u, w
h
)
w
h
h
_
u
h
being the solution of (P
h
).
Proof. For all v
h
V
h
we have
(11.9) u u
h
h
u v
h
h
+ u
h
v
h
h
.
Now for any v
h
V
h
, f (u
h
v
h
) = a
h
(u
h
, u
h
v
h
), so that we may
write
u
h
v
h
2
h
a
h
(u
h
v
h
, u
h
v
h
)
= a
h
(u v
h
, u
h
v
h
) + f (u
h
v
h
) a
h
(u, u
h
v
h
)

Mu v
h
h
u
h
v
h
h
+ f (u
h
v
h
) a
h
(u, u
h
v
h
),
and thus,
u
h
v
h
u v
h
h
+
1
f (u
h
v
h
) a
h
(u, u
h
v
h
)
u
h
v
h
u v
h
h
+
1
sup
w
h
V
h
f (w
h
) a
h
(u, w
h
)
w
h
h
.
116
Substituting in (11.9) and varying v
h
V
h
and taking the inmum
we get (11.8). This completes the proof.
Remark 11.2. In case the method is conforming, then a
h
(u, w
h
) = a
(u, w
h
) = f (w
h
) and the second term disappears in (11.8), leaving us
with the original bound (3.1).
We note that for the a
h
(, ) dened for the plate problem by (11.2),
the conditions of Theorem 11.1 are satised. The condition (11.6) is
embodied in (11.5). The condition (11.7) follows from the similar (con-
tinuity) condition on a(, ) and an application of the Cauchy-Schwarz
inequality.
Exercise 11.2. Let (H) be a Hilbert space with innerproduct (, ) and
norm . Let (V, ) be a subspace such that V H and V = H. Let
V
h
H. Dene
E
h
(u, v) = ( f , v) a
h
(u, v), for all u, v V
h
V.
Then show that
u u
h

Mu u
h
h
_
_
sup
gH
1
g
inf
h
V
h

h
h
_
_
+ sup
gH
_
1
g
inf
h
V
h
(E
h
(u,
h
) + E
h
(, u u
h
))
_
where for all v V, a(v, ) = g(v).
We now go on with the error analysis and study the order of conver-
gence. We assume that u H
3
() H
2
0
() which is quite realistic from
the regularity results.
Now, since
h
u V
h
, we have that
inf
v
h
V
h
u v
h
h
u
h
u
h
.
117
Again applying error bounds for each K and then summing over all
K we get
u
h
u
h
Chu
3,
.
Thus
(11.10) inf
v
h
V
h
u v
h
h
Chu
3,
.
Our aim is to get a similar estimate for the second term in (11.8). In
fact we show that
(11.11) sup
w
h
V
h
f (w
h
) a
h
(u, w
h
)
w
h
h
Chu
3,
.
This entails more work. We dene
(11.12) E
h
(u, w
h
) = f (w
h
) a
h
(u, w
h
)
for u H
3
() H
2
0
(), w
h
V
h
. Since w
h
H
1
0
(), there exists a
sequence w
n
h
in D() converging to w
h
in H
1
0
(). Hence,
_
f w
n
h
dx =
_
_
uw
n
h
+ (1 )
_
2

2
u
x
1
x
2
2
w
n
h
x
1
x
2

2
u
x
2
1
2
w
n
h
x
2
2

2
u
x
2
2
2
w
n
h
x
2
1
__
dx
=
_
(grad u)(grad w
n
h
)dx,
by Greens formula. The term involving (1 ), by Lemma 2.2, can
be converted to an integral over . All integrals over vanish since
w
n
h
D(). Since both sides of the above relation are continuous linear
functionals on H
1
0
(), we can pass to the limit to obtain
(11.13) f (w
h
) =
_
f w
h
dx =
_
(grad u)(grad w
h
)dx
for all w
h
V
h
. Now,
a
h
(u, w
h
) =
Kt
h
_
K
_
uw
h
+ (1 )
_
2

2
u
x
1
x
2
2
w
h
x
1
x
2

2
u
x
2
1
2
w
h
x
2
2

2
u
x
2
2
2
w
h
x
2
1
__
dx
=
Kt
h
_
_
K
grad u grad w
h
dx
+
_
K
u
w
K
h
K
d + (1 )
_
K
_
2
u
2
K
w
h
K
+
2
u
K
w
h
K
_
d
_
_
, (11.14)
by Greens formula (2.15) and Lemma 2.2 again. Notice however that 118
by standard orientation arguments and the continuity of w
h
over ,
(11.15)
Kt
h
_
K
2
u
K
w
h
K
d = 0.
Using (11.13), (11.14) and (11.15), we substitute in (11.12) to get
(11.16) E
h
(u, w
h
) =
Kt
h
_
K
_
_
u + (1 )
2
u
2
K
_
_
w
h
K
d.
Figure 11.2:
Splitting the boundary into four parts as in Fig. 11.2, we get
(11.17) E
h
(u, w
h
) =
Kt
h
_
1,K
_
u,
w
h
x
1
_
+
2,K
_
u,
w
h
x
2
__
,
where for j = 1, 2, we dene
j,K
_
u,
w
h
x
j
_
=
_
K
j
_
_
u + (1 )
2
u
2
K
_
_
_
w
h
x
j

K
_
w
h
x
j
__
d
_
K
j
_
_
u + (1 )
2
u
2
K
_
_
_
w
h
x
j

K
_
w
h
x
j
__
d,
(11.18)
K
being the Q
1
-interpolation operator associated with the values at the 119
four vertices.
Note that (11.17) is the same as (11.15). This is evident provided
we show that the contribution of the terms involving
K
is zero. This is
so because w
h
= 0 on the boundary and on the common boundaries,
K
_
w
h
x
j
_
is linear and equal in value for both adjacent nite elements
since it agrees at the vertices, but occurs with opposite signs as is obvi-
ous from Fig. 11.3 (K
1
1
= K
2
1
).
Figure 11.3:
We also record that
w
h
x
j
K

j
P
K
where
(11.19)
j
P
K
=
_
p
x
j
: K R; p P
K
_
.
We now prove a result analogous to the Bramble-Hilbert lemma
which will help us to estimate that
j,K
s and hence E
h
(u, w
h
).
Theorem 11.2 (BILINEAR LEMMA). Let R
n
be open with Lip-
schitz continuous boundary . Let W be a subspace of W
l+1,q
() such
that P
1
W. Let b be a continuous bilinear form over W
k+1, p
() W
such that
_
_
b(p, w) = 0 for all p P
k
, w W
b(v, q) = 0 for all v W
k+1, p
(), q P
1
.
Then there exists a constant C = C() such that for all v W
k+1, p
120
(), w W,
(11.21) b(v, w) Cb v
k+1, p,
w
l+1,q,
.
Proof. For a given w W, b(, w) : v b(v, w) is a continuous lin-
ear form on W
k+1, p
() vanishing on P
k
. Hence by the Bramble-Hilbert
lemma,
(11.22) b(v, w) Cb(, w)
k+1, p,
v
k+1,q,
.
However for all q P
1
, b(v, w) = b(v, w + q) so that
b(v, w) b v
k+1, p,
w + q
l+1,q,
and hence
b(v, w) b v
k+1, p,
inf
qP
1
w + q
l+1,q,
Cb v
k+1, p,
w
l+1,q,
, by the theorem 6.2,
so that
b(, w)
k+1, p,
Cb w
l+1,q,
.
and substituting in (11.22), we get (11.21), which completes the proof.
We may now prove the theorem on our error estimate and order of
convergence.
Theorem 11.3. For a regular family (t
h
) of triangulations made up of
Adinis rectangles
(11.23) u u
h
h
C hu
3,
.
Proof. Let us rst estimate
j,K
for j = 1, 2. Set
_
_
= u + (1 )
2
u
2
K
H
1
(K), since u H
3
(),
v =
w
h
x
1

1
P
K
.
Dene 121
(11.24)
1,K
(, v) =
_
K
1
(v
K
v)d
_
K
1
(v
K
V)d,
for v
1
P, H
1
(K). If h
2
is the length of K
1
(and K
1
) and h
1
is that
of K
2
(and K
2
) we have by a simple change of variable
(11.25)
1,K
(, v) = h
2
1,

K
( , v),
where

K is the reference nite element. Since P
0
Q
1
which is pre-
served by
K
we have that for all v P
0
, H
1
(

K),
1,

K
( , v) = 0.
Now let P
0
and v

1
P. We wish to show that
1,

K
( , v) = 0:
We may take for

K, the unit square. Since P
0
, its value on

K is
a constant, say, b
0
. Now let v

1
P. Then v is of the form
(11.26) v = a
0
+ a
1
x
1
+ a
2
x
2
+ a
3
x
2
1
+ a
4
x
1
x
2
+ a
5
x
2
2
+ a
6
x
2
1
x
2
+ a
7
x
3
2
.
Taking the values at the four vertices we get
(11.27)
K
v = a
0
+ (a
1
+ a
3
)x
1
+ (a
2
+ a
5
+ a
7
)x
2
+ (a
4
+ a
6
)x
1
x
2
.
Now K
1
is the line x
1
= 1 and K
1
is the line x
1
= 0. Thus
v
K
v
x
1
=0
= (a
5
+ a
7
)x
2
+ a
5
x
2
2
+ a
7
x
3
2
,
v
K
v
x
1
=1
= (a
5
+ a
7
)x
2
+ a
5
x
2
2
+ a
7
x
3
2
.
Hence,
_
K
1
( v
K
v)d =
1
_
0
b
0
((a
5
+ a
7
)x
2
+ a
5
x
2
2
+ a
7
x
3
2
)dx
2
=
_
K
1
( v
K
v)d.
(11.28)
Thus
1,

K
( , v) = 0 for P
0
, v

1
P. 122
Note further that the bilinear form
1,

K
is continuous, for
1,

K
( , v) C
L
2
(

K)
v
L
2
(

K)
.
C
1,

K
v
1,

K
, by the Trace theorem (cf. Th. 2.3).
Thus we may apply the bilinear lemma to the bilinear form
1,

K
with
l = k = 0 to get
(11.29)
1,

K
( , v) C
1,

K
v
1,

K
.
We also have the relations
(11.30)
_
_

1,

K
CB
K
det B
K
1
2
1,K
,
v
1,

K
CB
K
det B
K
1
2
v
1,K
.
Now B
K
C h
K
and det B
K
= meas K/ meas

K C
2
K
, and thus
B
K
det B
K
1
2

C h
K
K
C. Also h
2
h
K
, so that
(11.31)
1,K
(, v) C h
K
u
3,K
w
h
2,K
where
_
_
= u + (1 )
2
u
2
K
,
v =
w
h
x
1
,
and similarly,
(11.32)
2,K
(, v) C h
K
u
3,K
w
h
2,K
.
These inequalities lead us to the estimate
(11.33) f (w
h
) a
h
(u, w
h
) Chu
3,
w
h
h
,
for a regular family of triangulations made up of Adinis rectangles. 123
Thus varying w
h
over V
h
and taking the supremum, we get the estimate
(11.11).
Using (11.10) and (11.11) and substituting in (11.8) we get the re-
quired estimate as given in (11.23). This completes the proof.
Remark 11.3. By the Duality Argument, Lesaint and Lascaux [15] have
proved that u u
h
1,
Ch
2
u
4,
assuming u H
4
(). They have
also got an improved 0(h
2
) convergence order in the
h
norm, when
all the rectangles are equal - a superconvergence result.
We close this section with a brief description of other types of nite
elements used in non-conforming methods.
Example 11.2. The Zienkiewicz triangle (cf. Exercise 4.6) cf. Fig. 11.4.
Figure 11.4:
We get V
h
C
0
() only and hence the method is non-conforming.
It does not always yield convergence. The method works if the sides of
all triangles are parallel to three directions only, as in Fig. 11.5.
Figure 11.5:
This is not so if the number of directions is four, as in Fig. 11.6. 124
(The Union Jack Problem).
Figure 11.6:
Example 11.3. Morleys Triangle (cf. Fig. 11.7).
Figure 11.7:
Here P
K
= P
2
. We always get convergence for regular families, of
course. In fact if u H
4
(), then
(11.34) u u
h
h
Chu
4,
.
What is astonishing is that this nite element is not even of class C
0
.
Example 11.4. Fraeijs de Veubeke triangle. This nite element is again
a triangle. Apart from the values of the polynomials at the vertices and
the mid-points of the sides, we also take the average normal derivative
along the sides. Here the space P
K
, which we will not describe, satises
the inclusion
P
2
P
K
P
3
,
and
K
=
_
_
p(a
i
), 1 i 3; p(a
i j
), 1 i < j 3;
_
K
i
p
d, 1 i 3
_
_
.
The nite element is shown symbolically in Fig. 11.8. 125
Figure 11.8:
Here also the nite element is not of class C
0
in general, but the
method always yields convergence.
References: For general reference on non-conforming methods, see
Strang and Fix [22], for the bilinear lemma, see Ciarlet [29]. For a de-
tailed study of the Zienkiewicz triangle, Moreleys triangle and Fraeijs
de Veubeke triangle, see Lascaux and Lesaint [15]. For a nonconform-
ing method with penalty, see Babuska and Zl amal [26].
Bibliography
[1] BRAMBLE, J.H: THOM
EE, V. 126
Interior maximum norm estimates for some simple nite ele-
ment methods. Revue FranC aise dAutomatique, Informatique,
Recherche Operationelle (RAIRO), R-2, (1974), pp.-5-18.
[2] BRAMBLE, J.H; ZL
AMAL, M.
Triangular elements in the nite element methods, Mathematics of
Computation, 24, (1970), pp. 809-820.
[3] BREZIS, H; STAMPACCHIA, G.
Sur la r egularit e de la solution din equations elliptiques, Bull. Soc.
Math. France, 96(1968), pp. 153-180.
[4] CIARLET, P.G.
Sur l el ement de Clough et Tocher, RAIRO, R-2, (1974), pp. 19-27.
[5] CIARLET, P.G; RAVIART, P.-A.
La M ethodes des

El ements Finis pour les Probl` ems aux Limites
Elliptiques, (to appear).
Maximum principle and uniform convergence for the nite ele-
ment method, Computer Methods in Applied Mechanics and En-
gineering, 2, (1973), pp. 17-31.
129
130 BIBLIOGRAPHY
General Lagrange and Hermite interpolation in R
n
with applica-
tions to nite element methods, Arch. Rational Mech. Anal. 46,
(1972), pp. 177-199.
[8] CIARLET, P.G; WAGSCHAL, C.
Multipoint Taylor formulas and applications to the nite element
method, Numer. Math., 17, (1971), pp. 84-100.
[9] CIAVALDINI, J.F; NEDELEC, J.C.
Sur l el ement de Fraeijs de Veubeke et Sander, RAIRO, R-2,
(1974), pp. 29-45.
[10] DUVAUT, G; LIONS, J.-L.
Les In equations en M ecanique et en Physique, Dunod, Paris
(1972).
[11] FALK, R.S.
Error estimates for the approximation of a class of variational in-
equalities, Math. Comput. (to appear).
[12] FALK, R.S. 127
Approximation of an elliptic boundary value problem with unilat-
eral constraints, RAIRO, (to appear).
[13] HABER, S.
Numerical evaluation of multiple integrals, SIAM Review, 12,
(1970), pp. 481-526.
[14] LANDAU, L; LIFCHITZ, E.
Th eorie de lElasticit e, Mir, Moscow (1967).
[15] LASCAUX, P; LESAINT, P.
Some non-conforming nite elements for the plate bending prob-
lem, RAIRO, (to appear).
BIBLIOGRAPHY 131
[16] LEWY, H; STAMPACCHIA, G.
On the regularity of the solution of a variational inequality, Comm.
Pure Appl. Math., 22, (1969), pp. 153-188.
[17] LIONS, J.-L; MAGENES, E.
Probl` emes aux Limites non Homogenes et Applications. Vol. 1,
Dunod, Paris, (1968).
[18] LIONS, J.-L; STAMPACCHIA, G.
Variational inequalities, Comm. Pure Appl. Math., 20, (1967).
pp. 493-519.
[19] MOSCO, U; STRANG, G.
One-sided approximation and variational inequalities, Bull. Amer.
Math. Soc. 80 (1974), pp. 308-312.
[20] NE
CAS, J.
Les M ethodes Directes en Theorie des Equations Elliptiques, Mas-
son, Paris, (1967).
[21] STRANG, G.
Approximation in the nite element method, Number. Math.,
19(1972), pp. 81-98.
[22] STRANG, G; FIX, G.J.
An Analysis of the Finite Element Method, Prentice-Hall, Engle-
wood Clis, (1973).
[23]

ZENI
SEK, A.
Interpolation polynomials on the triangle, Numer. Math., 15,
(1970), pp. 283-296.
[24]

ZENI
SEK, A.
Polynomial approximation on tetrahedrons in the nite element
method, J. Approximation Theory, 7, (1973), pp. 334-351.
132 BIBLIOGRAPHY
[25] ZL
AMAL, M. 128
On the nite element method, Numer. Math., 12, (1968), pp. 394-
409.
[26] BABU
SKA, I.; ZL
AMAL, M.
Nonconforming elements in the nite element method with
penalty, SIAM J. Numer. Anal., 10(1973), pp. 863-875.
[27] BRAMBLE, J.H.; HILBERT, S.R.
Estimation of linear functionals on Sobolev spaces with applica-
tion to Fourier transforms and spline interpolation, SIAM J. Numer.
Anal., 7, (1970), pp. 113-124.
[28] BRAMBLE, J.H.; HILBERT, S.R.
Bounds for a class of linear functionals with applications to Her-
mite interpolation, Numer. Math., 16, (1971), pp. 362-369.
[29] CIARLET, P.G.
Conforming and nonconforming nite element methods for solv-
ing the plate problem, in Conference on the Numerical Solution
of Dierential Equations (G.A. WATSON, Editor), pp. 21-31,
Lecture Notes in Mathematics, Vol. 363, Springer-Verlag, Berlin
(1974).
[30] CIARLET, P.G.; RAVIART, P.-A.
Interpolation theory over curved elements, with applications to -
nite element methods, Comp. Methods in Appl. Mech. and Engi-
neering, 1, (1972), pp. 217-249.
[31] CIARLET, P.G.; RAVIART, P.-A.
The combined eect of curved boundaries and numerical integra-
tion in isoparametric nite element methods, in The Mathemati-
cal Foundations of the Finite Element Method with Applications to
Partial Dierential Equations (A.K. AZIZ, Editor), pp. 409-474,
Academic Press, New York, (1972).
BIBLIOGRAPHY 133
[32] ZL
AMAL, M.
A nite element procedure of the second order of accuracy, Numer.
Math., 16, (1970), pp. 394-402.
134 BIBLIOGRAPHY
Additional References 129
Books (Additional References)
BABUSKA, I; AZIZ, A.K. First part of The Mathematical Founda-
tions of the Finite Element Method with applications to Partial
Dierential Equatios (Symposium, University of Maryland, Bal-
timore, June-26-30, 1972) (Edited by A.K. AZIZ) Academic
Press, New York, (1972).
NORRIE, D.H; DE VRIES, G. The Finite Element Method - Funda-
mentals and Applications, Academic Press, New York, (1973).
ODEN, J.T. Finite Elements of Nonlinear Continue, McGraw Hill,
New York, (1972).
ZIENKIEWICZ, O.C. The nite Element Method in Engineering Sci-
ence, McGraw Hill, London, (1971).
Journals (where articles dealing with the nite element method may
be regularly found).
Revue Francaise dAutomatique, Informatique, Recherche Opera-
tionelle (RAIRO), Red Series
Computer Methods in Applied Mechanics and Engineering.
International Journal for Numerical Methods in Engineering.
Mathematics of Computation.
SIAM Journal on Numerical Analysis
Numerische Mathematik.
Proceedings of Conferences totally or partially devoted to the nite
element method:
The Mathematics of Finite Elements and Applications, (Conference,
Brunel University, 18-20 April, 1972), Edited by J.R. WHITEMAN,
Academic Press, London, 1973.
Conference on the Numerical Solution of Dierential Equations,
(Conference, Dundee, 03-06 July 1973), Edited by G.A. WATSON, Lec-
ture Notes in Mathematics, Vol. 363, Springer-Verlag, New York, 1974.
The Mathematical Foundations of the Finite Element Method with 130
BIBLIOGRAPHY 135
Applications to Partial Dierential Equations, (Symposium, Univer-
sity of Maryland, Baltimore, 26-30 June. 1972), Edited by A.K. AZIA,
Academic Press. New York, 1972.
Computing Methods in Applied Science and Engineering, Vol. 1
and 2 (International Symposium IRIA LABORIA, Versailles, 17-21 De-
cember, 1973) Edited by R. GLOWINSKI and J.L. LIONS, Lecture notes
in Computer Science, Vol., 10 and 11, Springer-Verlag, New York, 1974.
Topics in Numerical Analysis (Royal Irish Academy Conference on
Numerical Analysis, Dublin, 1972), Edited by J.J.H. MILLER, Aca-
demic Press, New York, 1973.
Mathematical Aspects of Finite Elements in Partial Dierential
Equations (Symposium, 01-03 April, 1974, Mathematics Research Cen-
ter, The University of Wisconsin, Madison), Edited by C. de BOOR, Aca-
demic Press, New York, 1974.
Proceedings of the Second Conference on the Mathematics of Finite
Elements and Applications, (Brunel University, April 07-10, 1975, to
appear.)

Theory of Elasticity

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Theory of Elasticity

Caricato da

Copyright:

Formati disponibili

Lectures on

The Finite Element Method

, the dual of V. Then the problem consists in nding an

V given by f f is linear. Now,

V. Since this solution is the projection of

denotes the norm in V

(u)(v) = 0 for all v K. In particular if K = V, J

. If (P) is the prob-

. Thus we have a map A : V V

(); supp(v) is a compact subset of

(), then its restric-

() equipped with the norm

. It is called the trace operator.

v. In fact if v is a smooth function then tr

denotes the tangential derivative.

() such that a 0 a.e.

f v dx, for all v H

v = 0. Hence the integral over vanishes.

= 0 on , and we may interpret this problem as the classical prob-

() satisfy the condition that for some

(cf. Fig. 4.3), we

. But these two agree at the nodes b

4. Examples of Finite Elements 47

imply that all the 6 coecients of p

. The polynomial r(t) =

and we also have r(b) = r(b

is also a polynomial (of degree 5) and

. This is the same thing as saying

along which t is an abscissa

are polynomials of degree 4

(K). The set is said to be

(K) P is called the P-interpolation

with appropriate number of repetitions from the basis, we may write,

, the dual of V. For, if g H, dene g V

since g(v) Cg v. If g(v) = 0 for all v V, then (g, v) = 0 for all

) a(v, ) = (g, v) for all v V,

). This is called the adjoint problem of (P). The existence

) are proved in an identical manner.

E ( ) = 0, i.e., again we need polynomial invariance, an idea already

given by (8.21). For second-order problems, for which the norm is

norms are equivalent on P

and H, then H will be identied with a subspace of V

is a linear map. We now pass on to an abstract error

= 0 on this line and since

= 0 on . Let us have two nodes

, the vertices of a triangle lying on with mid-point c. On this line

will be a polynomial in of degree 4 on this line. Hence for v = 0

) = 0. Thus the only free or unknown parameters

Potrebbero piacerti anche