i

i
i
i
Chapter 7
Theoretical foundations of
FEM
Using nite element methods, we need to answer these questions:
What is the right functional space V for the solution?
What is the right weak or variational form of a dierential equation.
What kind of basis functions or nite element spaces should we choose?
How accurate is a FEM solution?
We try briey to answer these equations in this Chapter. Remember that nite element
methods are based on integral forms, not in the point-wise sense as in nite dierence
methods. We need to generalize the theory corresponding to the point-wise form to integral
forms.
7.1 Functional space of C
m
()
Functional Space is a set of functions with operations. For example,
C() = C
0
() =
_
u(x), u(x) is continuous on
_
(7.1)
is a linear space that contains all continuous functions on . It is a linear space because
for any real numbers and , we have,
if u1 C(), u2 C(), then u1 + u2 C().
In the notation, is a domain where the functions are dened, for example, = [0, 1].
The functional space with rst order continuous derivatives is dened as
C
1
() =
_
u(x), u(x), u

(x) are continuous on

_
, (7.2)
and similarly
C
m
() =
_
u(x), u(x), u

(x), , u
(m)
are continuous on
_
. (7.3)
141
i
i
i
i
142 Chapter 7. Theoretical foundations of FEM
Obviously, we have,
C
0
C
1
C
m
. (7.4)
Let m go to the innity, we dene
C

() = {u(x), u(x) is indenitely dierentiable on } . (7.5)

For examples, e
x
, sin x and elementary functions are in C

.
7.1.1 Multi-dimensional spaces and multi-index notations
Consider multi-dimensional functions u(x1, x2, , xn), or u(x), x R
n
. We use multi-
index notations to simplify the expressions of partial derivatives. Let = (1, 2, , n),
i 0 be an integer vector in R
n
. For example, if n = 5, = (1, 2, 0, 0, 2) is one of s.
The muilti-index notation is used to express the following partial derivative
D

u(x) =

||
u
x

1
1
x

2
2
x
n
n
, || = 1 + 2 + + n, i 0. (7.6)
Example 7.1. When n = 2 and u(x) = u(x1, x2), all possible D

u are
= (2, 0), D

u =

2
u
x
2
1
,
= (1, 1), D

u =

2
u
x1x2
,
= (0, 2), D

u =

2
u
x
2
2
.
The C
m
space can be dened as
C
m
() =
_
u(x1, x2, , xn), D
||
u are continuous on , || m
_
. (7.7)
That is all possible derivatives up to order m are continuous on .
Example 7.2. When n = 2, m = 3, then u, ux, uy, uxx, uxy, uyy, uxxx, uxxy, uxyy, uyyy
all have to be continuous on if u C
3
(). Note that C
m
() has innite dimensions.
The distance in C
0
() is dened as,
d(u, v) = max
x
|u(x) v(x)| (7.8)
A linear space with distance dened is called a metric space. A distance is an non-negative
function of u and v that satises,
d(u, v) 0, d(u, v) = 0, i u v. (7.9)
d(u, v + w) d(u, v) + d(u, w), triangle inequality. (7.10)
In the expressions above. the word i means if and only if.
i
i
i
i
7.2. Spaces in integral forms L
2
() and L
p
() 143
The Norm in C
0
() is an non-negative function of u dened by
u(x) = d(u, ) = max
x
|u(x)|, (7.11)
where is the zero element. A norm needs to satisfy
u(x) 0, u(x) = 0, i u 0. (7.12)
u(x) = ||u(x), is a number, (7.13)
u(x) +v(x) u(x) +v(x), triangle inequality. (7.14)
A linear space with a norm dened is called a normed space. In C
m
(), the distance and
the norm are dened as
d(u, v) = max
0||m
max
x

u(x) D
||
v(x)

, (7.15)
u(x) = max
0||m
max
x
|D

u| . (7.16)
7.2 Spaces in integral forms L
2
() and L
p
()
In analogue to point-wise spaces C
m
(), we can dene spaces H
m
() called Sobolev spaces
using integral forms. The square-integrable space H
0
() = L
2
() is dened as
L
2
() =
_
u(x),
_

u
2
(x)dx <
_
, (7.17)
corresponding to the point-wise C
0
() space. It is easy to see that
C(0, 1) = C
0
(0, 1) L
2
(0, 1).
Example 7.3. It is easy to verify that u(x) = 1/x
1/4
/ C
0
, but
_
1
0
_
1
x
1/4
_
2
dx =
_
1
0
1

x
dx = 2 < .
Therefore u(x) L
2
(0, 1).
The distance in L
2
() is dened as
d(f, g) =
__

|f g|
2
dx
_
1/2
. (7.18)
Thus L
2
() is a metric space. We say that f g (identical) if d(f, g) = 0. The following
two functions are identical in L
2
(2, 2)
f(x) =
_
0 if x < 0
1 if x 0,
g(x) =
_
0 if x 0
1 if x > 0.
The norm in the L
2
() space is dened as
u
L
2 = u0 =
__

|u|
2
dx
_
1/2
. (7.19)
i
i
i
i
144 Chapter 7. Theoretical foundations of FEM
We can prove that these denitions satisfy the requirements of the properties requirements
for the distance and the norm.
L
2
() is a complete space means that any Cauchy sequence {fn} in L
2
has a limit
in L
2
. In other words, there is a f L
2
() such that
lim
n
fn f
L
2 = 0, or lim
n
fn = f.
A Cauchy sequence is a sequence that satises the property: for any given positive number
, there is an integer N, such that
fn fm
L
2 < , if m N, n N.
A complete normed space is called a Banach space (a Cauchy sequence converges in
terms of the norm). Therefore L
2
is a Banach space.
7.2.1 The inner product in L
2
In R
n
, given any two vectors,
x =
_
_
_
_
_
_
_
_
_
x1
x2
.
.
.
xn
_

_
R
n
, y =
_
_
_
_
_
_
_
_
_
y1
y2
.
.
.
yn
_

_
R
n
,
we know that the inner product is dened as
(x, y) = x
T
y =
n

i=1
xiyi = x1y1 + x2y2 + + xnyn.
Similarly, the inner product in the L
2
space is dened as
(f, g) =
_

f(x)g(x)dx, if f L
2
(), g L
2
(). (7.20)
It satises the requirements of the denition of an inner product:
(f, g) = (g, f),
(f, g) = (f, g) = (f, g) R,
(f, g + w) = (f, g) + (f, w).
for any f L
2
() and g L
2
().
With the inner product, the weak form for the simple model problem
u

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

can be written as
(u

, v

) = (f, v), v H
1
(0, 1),
i
i
i
i
7.2. Spaces in integral forms L
2
() and L
p
() 145
and the minimization forms is
min
vH
1
0
(0,1)
F(v) : F(v) =
1
2
(v

, v

) (f, v).
The norm, distance, and the inner product in L
2
() have the following relations
u0 =
_
(u, u) = d(u, ) =
__

|u|
2
dx
_
1/2
. (7.21)
7.2.2 The Cauchy-Schwartz inequality in L
2
()
For a Hilbert space with a norm (u, v) and its resulted norm u =
_
(u, u), the Cauchy-
Schwartz inequality is the following,
|(u, v)| u0v0. (7.22)
Below we list some examples of the Cauchy-Schwartz inequality corresponding to
inner products in R
n
and L
2
spaces:

i=1
xiyi

_
n

i=1
x
2
i
_
1/2
_
n

i=1
y
2
i
_
1/2
;

i=1
xi

n
_
n

i=1
x
2
i
_
1/2
;

fgdx

__

f
2
dx
_
1/2
__

g
2
dx
_
1/2
;

fdx

__

f
2
dx
_
1/2

V ,
where V is the volume of .
A proof of the Cauchy-Schwartz inequality
Noting that (u, u) = u
2
, we construct a quadratic function of given u and v as follows
f() = (u + v, u + v) = (u, u) + 2(u, v) +
2
(v, v) 0.
The quadratic function is non-negative, and therefore the discriminant of the quadratic
form satises
= b
2
4ac 0, i.e. 4(u, v)
2
4(u, u)(v, v) 0, or (u, v)
2
(u, u)(v, v).
Taking square root from both sides of the last inequality above, we then have the Cauchy-
Schwartz inequality.
A complete Banach space with an inner product dened is called a Hilbert space. So
L
2
() is a Hilbert space (linear space, inner product, complete).
i
i
i
i
146 Chapter 7. Theoretical foundations of FEM
Summary of denitions of spaces
The following diagram shows relations among the denitions of spaces.
Metric Space (distance) Normed Space (norm) Banach space (complete)
Hilbert space (inner product).
7.2.3 The L
p
() spaces
An L
p
() space is dened as
L
p
() =
_
u(x),
_

|u(x)|
p
dx <
_
. (7.23)
The distance in L
p
() is dened as
d(f, g) =
__

|f g|
p
dx
_
1/p
. (7.24)
An L
p
() space is a metric and complete space, so it is also a Banach space. But it is
not a Hilbert space because there is no corresponding inner product can be dened unless
p = 2.
7.3 The Sobolev spaces related to the weak
derivatives
Similar to C
m
() spaces, we use H
m
() to dene function spaces with derivatives in
integral forms. Such function spaces are part of the Sobolev spaces.
If there is no derivative, then we dene the space
H
0
() = L
2
() =
_
v(x),
_

|v|
2
dx <
_
. (7.25)
7.3.1 The denition of the weak derivatives
We know that if u(x) C
1
(0, 1), then for any function C
1
(0, 1), (0) = (1) = 0, we
have
_
1
0
u

(x)(x) = u

1
0

_
1
0
u(x)

(x)dx =
_
1
0
u(x)

(x)dx. (7.26)
Such a (x) is called a testing function in C
1
0
(0, 1). Therefore we can dene the weak
derivative of u(x) L
2
() = H
0
() as a function v(x) that satises
_

v(x)(x)dx =
_

u(x)

(x)dx (7.27)
for all (x) C
1
0
(), (0) = (1) = 0. If such a function exists, it is called the rst order
weak derivative of u(x) and it is denoted as v(x) = u

(x).
Similarly, we can dene the m-th order weak derivative of u(x) H
0
() as a function
v(x) that satises
_

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

u(x)
(m)
(x)dx (7.28)
i
i
i
i
7.3. The Sobolev spaces related to the weak derivatives 147
for all (x) C
m
0
(). That is
(x) C
m
(), (x) =

(x) = =
(m1)
(x) = 0, if x .
If such a function v(x) exists, it is called the m-th order weak derivative of u(x) and it
is denoted as v(x) = u
(m)
(x). High order weak derivatives are dened from lower order
derivatives.
7.3.2 Denition of the Sobolev spaces H
m
()
Having dened the rst partial derivatives, we dene the Sobolev space H
1
() as
H
1
() =
_
v(x), D
||
v L
2
(), || 1
_
. (7.29)
For example, H
1
(a, b) is dened as
H
1
(a, b) =
_
v(x), a < x < b,
_
b
a
v
2
dx < ;
_
b
a
(v

)
2
dx <
_
.
In two space dimensions, H
1
() is dened as
H
1
() =
_
v(x, y), v L
2
(),
v
x
L
2
(),
v
y
L
2
()
_
.
Similarly, the Sobolev space H
m
() is dened as
H
m
() =
_
v(x), D
||
v L
2
(), || m
_
. (7.30)
7.3.3 The inner product in H
m
spaces
The inner product in H
0
() is the same as that in L
2
(),
(u, v)
H
0
()
= (u, v)0 =
_

uvdx.
The inner product in H
1
(a, b) is dened as
(u, v)
H
1
(a,b)
= (u, v)1 =
_
b
a
_
uv + u

_
dx.
The inner product in H
1
() of two variables is dened as
(u, v)
H
1
()
= (u, v)1 =
__

_
uv +
u
x
v
x
+
u
y
v
y
_
dxdy.
The inner product in H
m
() of general dimensions is
(u, v)
H
m
()
= (u, v)m =
__

||m
(D
||
u(x)) (D

v(x)) dx. (7.31)

The norm in H
m
() of general dimensions is
u
H
m
()
= um =
_
_
_
_

||m
|D

u(x)|
2
dx
_
_
_
1/2
. (7.32)
i
i
i
i
148 Chapter 7. Theoretical foundations of FEM
Therefore H
m
() is a Hilbert space. A norm can be dened from the inner product. For
example, in H
1
(a, b), the H
1
norm is dened by
u1 =
__
b
a
_
u
2
+u
2
_
dx
_
1/2
.
The distance in H
m
() of general dimensions is then dened as
d(u, v)m = u vm. (7.33)
7.3.4 Relations between C
m
() and H
m
() Sobolev embedding
theorem
In one dimensional space, we have
H
1
() C
0
(), H
2
() C
1
(), , H
1+j
() C
j
().
Theorem 7.1. The Sobolev embedding theorem: If 2m > n , then
H
m+j
C
j
, j = 0, 1, , (7.34)
where n is the dimension of the independent variables of the elements in the Sobolev space.
Example 7.4. In two dimensional space, we have n = 2. The condition 2m > n means
that m > 1. From the embedding theorem, we have
H
2+j
C
j
, j = 0, H
2
C
0
, j = 1, H
3
C
1
, (7.35)
If u(x, y) H
2
, which means u, ux, uy, uxx, uxy, and uyy all belong to L
2
, we can
conclude that u(x, y) is continuous, but ux and uy may not be continuous!
Example 7.5. In three dimensional space, we have n = 2. The condition 2m > n means
that m > 3/2. We have the same result as that in 2D:
H
2+j
C
j
, j = 0, H
2
C
0
, j = 1, H
3
C
1
, (7.36)
We regard the regularity of solutions as the degree of smoothness or lack of it of a
class of problems measured in C
m
or H
m
spaces. If u(x) H
m
or C
m
spaces, then the
larger m is, the smoother of the function is.
7.4 The FEM analysis for the 1D model problem
For the simple 1D model problem
u

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

we know that the weak form is
_
1
0
u

dx =
_
1
0
fvdx or (u

, v

) = (f, v).
i
i
i
i
7.4. The FEM analysis for the 1D model problem 149
Since v is arbitrary, intuitively, we can take v = f or v = u. Then we get
_
1
0
u

dx =
_
1
0
u
2
dx,
_
1
0
fvdx =
_
1
0
f
2
dx.
So both u, u

, f, v, and v

should belong to L
2
(0, 1). That is u H
1
(0, 1) and v H
1
(0, 1).
So the solution is in the H
1
(0, 1) Sobolev space. We should also take v in the same space.
From the Sobolev embedding theorem, we also know that H
1
C
0
, which means the
solution is continuous.
7.4.1 Conforming nite element methods
Denition 7.2. If the nite element space is a subspace of the solution space, then the
nite element method is called a conforming nite space.
For example, the piecewise linear function over a given a triangulation is a conform-
ing nite element space for the model problem. The nite element method is called a
conforming nite element method. We mainly discuss conforming nite element methods
in this book.
If we put the boundary condition together, then we dene the solution space as
H
1
0
(0, 1) =
_
v(x), v(0) = v(1) = 0, v H
1
(0, 1)
_
. (7.37)
When we look for a nite element solution in a nite dimensional space V
h
, V
h
should
be a subspace of H
1
0
(0, 1) for conforming nite element methods. For example, given a
mesh for the 1D model problem, we can dene a nite dimensional space using piecewise
continuous linear functions over the mesh
V
h
=
_
v
h
, v
h
(0) = v
h
(1) = 0, v
h
is continuous piecewise linear, v(h) H
1
(0, 1)
_
.
The nite element solution would be chosen from the nite dimensional space V
h
, a
subspace of H
1
0
(0, 1). If the solution of the weak form is in H
1
0
(0, 1) but not in V
h
space,
then an error is introduced as the result of substituting the solution space with the nite
dimensional space. Nevertheless, the nite element solution is the best approximation in
V
h
in some norm as we will see later.
7.4.2 The FEM analysis for the one dimensional Sturm-Liouville
problems
A one-dimensional Sturm-Liouville problem with Dirichlet boundary conditions at two
ends has the following form,
(p(x)u

(x))

+ q(x)u(x) = f(x), x
l
< x < xr
u(x
l
) = 0, u(xr) = 0,
p(x) pmin > 0, q(x) qmin 0.
(7.38)
The condition on p(x) and q(x) is to guarantee the problem is well-posed, that is, the
weak form has a unique solution. It is convenient to assume that p(x) C(x
l
, xr) and
i
i
i
i
150 Chapter 7. Theoretical foundations of FEM
q(x) C(x
l
, xr). Later we will see that with these conditions along with f(x) L
2
(x
l
, xr),
we can guarantee an unique solution to the weak form of the problem. To derive the weak
form, we multiply a testing function v(x), v(x
l
) = v(xr) = 0 to the both sides of the
equations above and integrate from x
l
to xr to get
_
xr
x
l
_
(p(x)u

+ qu
_
vdx =
_
xr
x
l
fvdx
pu

xr
x
l
+
_
xr
x
l
_
pu

+ quv
_
dx =
_
xr
x
l
fvdx

_
xr
x
l
_
pu

+ quv
_
dx =
_
xr
x
l
fvdx, v H
1
0
(x
l
, xr).
7.4.3 The bilinear form
We dene the bilinear form as
a(u, v) =
_
xr
x
l
_
pu

+quv
_
dx. (7.39)
We call it the bilinear form because it is linear for both u and v, that is
a(u + w, v) =
_
xr
x
l
_
p(u

+ w

)v

+ q(u +w)v
_
dx
=
_
xr
x
l
_
pu

+ quv
_
dx +
_
xr
x
l
_
w

+ qwv
_
dx
= a(u, v) + a(w, v),
where and are scalars. Similarly
a(u, v + w) = a(u, v) + a(u, w).
Note that the bilinear form is an inner product dierent from the L
2
and H
1
inner
products in general. If p 1 and q 1, then
a(u, v) = (u, v)1.
Since a(u, v) itself is an inner product, we can dene a corresponding norm under the
conditions (p(x) pmin > 0, q(x) 0). Such a norm is called the energy norm:
ua =
_
a(u, u) =
__
xr
x
l
_
pu
2
+ qu
2
_
dx
_1
2
, (7.40)
where the rst term is called the kinetic energy, and the second term is called the potential
energy in some literature. The Cauchy-Schwartz inequality implies |a(u, v)| uava.
Use the bilinear form, the notations are much simplied. The week form for the
Sturm-Liouville problem is
a(u, v) = f(v), v H
1
0
(x
l
, xr). (7.41)
The minimization form is
min
vH
1
0
(x
l
,xr)
F(v) = min
vH
1
0
(x
l
,xr)
1
2
a(v, v) (f, v). (7.42)
i
i
i
i
7.4. The FEM analysis for the 1D model problem 151
Later we will see that all self-adjoint elliptic dierential equations have this kind of weak
form and minimization from. The nite element method using the Ritz form is the same
as the Galerkin form.
7.4.4 The FEM for the 1D S-L problems using piecewise linear
basis functions in H
1
.
Given any nite dimensional space Vs H
1
0
(x
l
, xr) whose basis are
1(x) H
1
0
, 2(x) H
1
0
, , M(x) H
1
0
(x
l
, xr),
that is,
Vs = span{1, 2, , M}
=
_
vs, vs =
M

i=1
ii
_
H
1
0
(x
l
, xr).
The Galerkin method (using the weak form) is the following. Let the approximate solution
be
us =
M

j=1
jj. (7.43)
The coecients are chosen such that the weak form is satised
a(us, vs) = (f, vs), vs V
h
.
In other words, we enforce the weak form on Vs not in H
1
0
(a, b). This is where the error
comes in.
In a nite element method, we enforce the weak form in the nite dimensional space
Vs instead of the solution space (H
1
0
(x
l
, xr)). This can be done by enforcing the weak
for the basis functions since any element in the space is a linear combination of the basis
functions. Therefore we have
a (us, i) = (f, i), i = 1, 2, , M,
or
a
_
M

j=1
j j, i
_
= (f, i), i = 1, 2, , M,
or
M

j=1
a(j, i) j = (f, i), i = 1, 2, , M.
In the matrix-vector form, AU = F, the system of equations for the coecients is
_
_
_
_
_
_
_
_
_
a(1, 1) a(1, 2) a(1, M)
a(2, 1) a(2, 2) a(2, M)
.
.
.
.
.
.
.
.
.
.
.
.
a(M, 1) a(M, 2) a(M, M)
_

_
_
_
_
_
_
_
_
_
_
1
2
.
.
.
M
_

_
=
_
_
_
_
_
_
_
_
_
(f, 1)
(f, 2)
.
.
.
(f, M)
_

_
.
i
i
i
i
152 Chapter 7. Theoretical foundations of FEM
The linear system of equations has the following nice properties:
A is symmetric. That is {aij} = {aji}, or A = A
T
since a(i, j) = a(j, i). Note
that this is only true for self-adjoint elliptic problems. In one dimensions, the second
order ODE has the form
(pu

+ qu = f.
The following equation
u

+ xu

= f
is not self-adjoint, the weak form for this equation is
(u

, v

) + (u

, v) = (f, v), or (u

, v

) + (u, v

) = (f, v).
We still can use the Galerkin nite element method, but we will have terms like
(

i
, j) which is dierent from (

j
, i). So the coecient matrix for the none self-
adjoint one is not symmetric anymore.
A is positive denite, that is,
1. x
T
Ax > 0, if x = 0.
2. All eigenvalues of A are positive.
Proof: For any = 0, we show that
T
A > 0.

T
A =
T
(A) =
M

i=1
i
M

j=1
aijj
=
M

i=1
i
M

j=1
a(i, j)j
=
M

i=1
i
M

j=1
a(i, jj)
=
M

i=1
i a
_
i,
M

j=1
jj
_
= a
_
M

i=1
ii,
M

j=1
jj
_
= a (vs, vs) = vs
2
a
> 0
since vs =

M
i=1
ii = 0 because is a nonzero vector and i are linear independent.
7.4.5 The local stiness matrix and load vector using the hat
basis functions
The local stiness matrix using hat basis function is still 2 by 2 matrix that has the form
K
e
i
=
_
_
_
x
i+1
x
i
p(

i
)
2
dx
_
x
i+1
x
i
p

i+1
dx
_
x
i+1
x
i
p

i+1

i
dx
_
x
i+1
x
i
p(

i+1
)
2
dx
_
_
+
_
_
_
x
i+1
x
i
q
2
i
dx
_
x
i+1
x
i
qii+1dx
_
x
i+1
x
i
qi+1idx
_
x
i+1
x
i
q
2
i+1
dx
_
_
i
i
i
i
7.5. Error analysis for FEM 153
and the load vector is
F
e
i
=
_
_
_
x
i+1
x
i
fidx
_
x
i+1
x
i
fi+1dx
_
_
.
The global stiness matrix and load vector can be assembled by element by element ap-
proach.
7.5 Error analysis for FEM
Error analysis for nite element methods usually include two parts. One is error estimates
for a given nite element space. The other one is the convergence analysis ( a limiting
process) that shows the nite element solution converges the true solution to the weak
form in some norms as the mesh size h approaches zero. We rst recall some notations
and set-ups.
1. We are given a weak form a(u, v) = L(v) and a space V , which usually has innite
dimensions. The problem is to nd a u, u V , such that the weak form is satised
for any v V . We call u the solution of the weak form.
2. We denote V
h
as a nite dimensional and subspace of V , i.e. V
h
V (condition for
a conforming FEM). It does not have to depend on h though.
3. We denote u
h
as the solution of the weak form in the subspace V
h
. That is,
a(u
h
, v
h
) = L(v
h
) is true for any v
h
V
h
.
4. We dene e
h
= u(x) u
h
(x) as the global error. We want to nd a sharp upper
bound for e
h
using certain norms.
Note that error arises when we substitute the solution space with a nite dimensional
space. In other words, the weak form is satised only in the sub-space V
h
, not in the
solution space V . However, we can prove that the solution that satises the weak form in
the sub-space V
h
is the best approximation to the exact solution u in the nite element
space in the energy norm.
Theorem 7.3.
1. u
h
is the projection of u onto V
h
in terms of energy inner product, that is
u u
h
V
h
, or u u
h
i, i = 1, 2, , M, (7.44)
a(u u
h
, v
h
) = 0, v
h
V
h
, or a(u u
h
, i) = 0, i = 1, 2, , M,(7.45)
assuming that {i} are the basis functions.
2. Best approximation in the energy norm:
u u
h
a u v
h
a, v
h
V
h
.
i
i
i
i
154 Chapter 7. Theoretical foundations of FEM
Proof:
a(u, v) = (f, v), v V,
a(u, v
h
) = (f, v
h
), v
h
V
h
since V
h
V,
a(u
h
, v
h
) = (f, v
h
), v
h
V
h
, since u
h
is the solution in V
h
,
subtract a(u u
h
, v
h
) = 0, or a(e
h
, v
h
) = 0, v
h
V
h
.
Now we prove that u
h
is the best approximation in V
h
.
u v
h

2
a
= a(u v
h
, u v
h
)
= a(u u
h
+ u
h
v
h
, u u
h
+u
h
v
h
), Let w
h
= u
h
v
h
V
h
,
= a(u u
h
+ w
h
, u u
h
+ w
h
)
= a(u u
h
, u u
h
+ w
h
) + a(w
h
, u u
h
+ w
h
)
= a(u u
h
, u u
h
) + a(u u
h
, w
h
) + a(w
h
, u u
h
) + a(w
h
, w
h
)
= u u
h

2
a
+ 0 + 0 +w
h

2
a
, since, a(e
h
, u
h
) = 0
u u
h

2
a
.
That is
u u
h
a u v
h
a.
Example: For the Sturm-Liouville problem, we have
u u
h

2
a

_
b
a
_
p(u

h
)
2
+ q(u u
h
)
2
_
dx
pmax
_
b
a
(u

h
)
2
dx + qmax
_
b
a
(u u
h
)
2
dx
max{pmax, qmax}
_
b
a
_
(u

h
)
2
+ (u u
h
)
2
_
dx
= Cu u
h

2
1
,
where C = max{pmax, qmax}. So we obtain
u u
h
a

Cu u
h
1,
u u
h
a u v
h
a

Cu v
h
1.
7.5.1 Interpolation functions and error estimates
Usually the solution is unknown, in order to get the error estimate, we need to choose
a special v

h
V
h
, which is very close to the solution. Then we use the error estimate
u u
h
a u v

h
a. Usually we choose the piecewise interpolation function. That is
another reason we choose the nite element space as piecewise linear, quadratic, or cubic
functions over the given triangulation.
i
i
i
i
7.5. Error analysis for FEM 155
Linear piecewise interpolation function of one dimension
Given a triangulation x0, x1, x2, , xM, the linear piecewise interpolation function is de-
ned as uI (x):
uI (x) =
x xi
xi1 xi
u(xi1) +
x xi1
xi xi1
u(xi), xi1 x xi.
It is obvious that uI (x) V
h
, where V
h
H
1
, the set of continuous piecewise linear
functions. Therefore
u u
h
a u uI a.
Since u(x) is unknown, uI(x) is unknown as well. However we know the upper bound of
the interpolation functions.
Theorem 7.4. Given a function u(x) C
2
[a, b] and a triangulation x0, x1, x2, , xM.
The continuous piecewise linear function uI has the following error estimates
u uI = max
x[a,b]
|u(x) uI(x)|
h
2
8
u

, (7.46)
u

(x) u

I
(x)
L
2
(a,b)
h

b au

. (7.47)
Proof: Let e
h
= u(x) uI(x), then
e
h
(xi1) = e
h
(xi) = 0.
From the Rolles theorem, there must be at least one point zi between xi1 and xi, that
is
e
h

(zi) = 0.
Therefore
e
h

(x) =
_
x
z
i
e
h

(t)dt
=
_
x
z
i
_
u

(t) u

I
(t)
_
dt
=
_
x
z
i
u

(t)dt.
We obtain the following error estimate
| e
h

(x)|
_
x
z
i
|u

(t)|dt u

_
x
z
i
dt u

h,
and
e
h

L
(
a,b)
= e
h

0
_
u

_
b
a
h
2
dt
_
1
2

b au

h.
i
i
i
i
156 Chapter 7. Theoretical foundations of FEM
Thus we have proved the second equality. To prove the rst one, we can assume that
xi1 + h/2 zi xi, otherwise we can use the other half interval. We use the Taylor
expansions to get:
e
h
(x) = e
h
(zi + x zi), assume xi1 x xi,
= e
h
(zi) + e
h

(zi)(x zi) +
1
2
e
h

()(x zi)
2
, xi1 xi,
= e
h
(zi) +
1
2
e
h

()(x zi)
2
,
Take x = xi, we have
0 = e
h
(xi) = e
h
(zi) +
1
2
e
h

()(xi zi)
2
,
e
h
(zi) =
1
2
e
h

()(xi zi)
2
,
| e
h
(zi)|
1
2
u

(x zi)
2

h
2
8
u

.
Note that the largest value of e
h
(x) has to be one of zis where the derivative is zero.
7.5.2 Error estimates of the nite element methods using the
interpolation function.
For one dimensional Sturm-Liouville problem, we have
Theorem 7.5.
u u
h
a Chu

(7.48)
u u
h
1

Chu

, (7.49)
where C and

C are two constants.
Proof:
u u
h

2
a
u uI
2
a

_
b
a
_
p(u

I
)
2
+q(u uI )
2
_
dx
max {pmax, qmax}
_
b
a
_
(u

I
)
2
+ (u uI)
2
_
dx
max {pmax, qmax} u

_
b
a
_
h
2
+ h
4
/64
_
dx
Ch
2
The second inequality is obtained because a and 1 are equivalent meaning that
cva v1 Cva, cv1 va

Cv1.
i
i
i
i
7.6. Exercises 157
7.5.3 Error estimates in pointwise norm
For one dimensional Sturm-Liouville problem, we can easily prove the following (over-
estimate though).
Theorem 7.6.
u u
h
Chu

(7.50)
u

h
Does not make sense (7.51)
where C is a constants. Since u

h
is discontinuous at nodal points, the innity norm makes
no sense because it can only be dened for continuous functions.
Proof:
e
h
(x) = u(x) u
h
(x) =
_
x
a
e

h
(t)dt
|e
h
(x)|
_
b
a
|e

h
(t)|dt

__
b
a
|e

h
|
2
dt
_
1/2
__
b
a
1 dt
_
1/2

b a
__
b
a
p
pmin
|e

h
|
2
dt
_
1/2

b a
pmin
e
h
a

b a
pmin
e
h
a
Chu

.
Remark 7.1. Actually we can prove a better inequality:
u u
h
Ch
2
u

.
The nite element method is second order accurate.
7.6 Exercises
1. Take n = 3, the number of variables, describe the Sobolev space H
3
(), i.e. m = 3,
in terms of L
2
() using all the terms but not the multi-index notation. Also explain
the inner product, the norm, the Schwartz inequality, the distance, and the Sobolev
embedding theorem in this space using the multi-index notation when applicable.
2. Consider the function v(x) = |x|

on = (1, 1) with R. For what values of

is v H
0
()? (Consider negative as well). For what values is v H
1
()? in
H
m
()? For what values of is v C
m
()?
i
i
i
i
158 Chapter 7. Theoretical foundations of FEM
Hint: Generally
|x|

=
_
x

if x 0
(x)

if x < 0.
However, if = 2k, where k is a non-negative integer, then
|x|

=
_
x
2k
if = 2k, k > 0,
1 if = 0.
Also
lim
x0
|x|

=
_

_
0 if > 0
1 if = 0
if < 0,
_
1
1
|x|

dx =
_
_
_
2
+ 1
if > 1,
if 1.
3. Are each of the following statements true or false? Justify your answers.
(a) If u H
2
(0, 1) then u

and u

are both continuous functions.

(b) If u(x, y) H
2
(), then u(x, y) may not have continuous partial derivatives
u
x
and
u
y
? Does u(x, y) have rst and second order weak derivatives? Is
u(x, y) continuous in ?
4. Consider the Sturm-Liouville problem:

_
p(x)u(x)

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

u(0) +u

(0) = , u

() = u
b
,
where
0 < pmin p(x) pmax < , 0 qmin q(x) qmax < q.
(a) Derive the weak form for the problem. Dene a bilinear form a(u, v) and a
linear form L(v) to simplify the weak form. What is the energy norm?
(b) What kind of restrictions should we have for , , and in order that the
weak form has a solution?
(c) Determine the space where the solution resides under the weak form.
(d) If we look for a nite element solution in a nite dimensional space V
h
using a
conforming nite element method, should V
h
be a subspace of C
0
, C
1
, C
2
?
(e) Given a triangulation x0 = 0 < x1 < x2 < xM1 < xM = , if the nite
dimensional space is generated by the hat functions, what kind of structure do
the local and global stiness matrix and the load vector have? Is the resulting
linear system of equations formed by the global stiness matrix and the load
vector symmetric, positive denite, and banded?
i
i
i
i
7.6. Exercises 159
5. Extra Credit: Consider the two-point boundary value problem
u

(x) = f(x), a < x < b, u(a) = u(b) = 0.

Let u
h
(x) be the nite element solution using the piecewise linear space (in H
1
0
(a, b))
spanned by a mesh {xi}. Show that
u u
h
Ch
2
,
where C is a constant. Hint: First show that u
h
uI a = 0, where uI(x) is the
interpolation function in V
h
.
i
i
i
i
160 Chapter 7. Theoretical foundations of FEM