-

chapter 6

The numerical solution of

minimization problems

Objectives
Chapter 5 showed the equivalence behA:'een a certain class of PDEs and minimiza-
tion problems. As a consequence solvmg the PDE also solves the minimization
problem and vice versa. Cha~ters 3, ~ were devoted to solving the PDE directly
by finite differences or, after mtegrahon over a volume, by finite volumes. In
this chapter we shall solve the corresponding minimization problem numerically.
Hence the PDE is solved in an indirect way.
The numerical technique that will be applied is the classical Ritz's method based on
expressing the solution as a linear combination of previously chosen functions: the
basis functions. These are in general not related to the problem, but chosen before-
hand. This method itself is not very practical, but combined with a clever choice
of basis functions we arrive at the finite element method (FEM). The FEM is well
suited for unstructured grids and has a strict local character. All information in
one element is used, without considering neighbors. This makes the method very
attractive for computer implementation. For certain types of PDEs, for example
those arising from elasticity and plasticity problems, the FEM is the most popular
method at this moment.
Another way of looking at the FEM, is to consider it as an automatic tool to derive
finite difference formula for unstructured grids. An important advantage of the
FEM is that the treatment of boundary conditions is almost always very natural
and therefore simpler than in classical difference methods.

6.1 Ritz's method

6.1.1 Introduction
Suppose we want to solve the general minimization problem

~J[u]; J[u]= jF(x,y,u,ux,uy)dO, (6.1.1)

()

Where the minimum must be found over a class of functions in the target space I::

I:. = {u sufficiently smooth; uIr = g} . (6.1.2)

111111.........
 Numerical methods in scientific co
Illputing
98
that the solution of this problem is not .
demons trated • sun l
Chapter 5 alrea d Y . minimization problem with one unknown ( ) Pe.
Actually we transformed it to ~uations. € 'thus
deriving the ~ule_r-Lagra;f;;, (6.1.2) is in general only possible if we have a . .
Direct minimization of ( · · finite
ber of unknowns.
num oximating the solution by a linear combinati
This can be achieved ~y app~ x)· on of a
finite fixed set of functions q,, ( .
11

un(x) = r:aj(f'j(x). (6.1.3)

j=l

. f hi tion we assume homogeneous essential boundary

In the remainder o t s sec con-
ditions, i.e. g = 0.
. ( ) (th basis functions) must be chosen such that:
The functions cp; x e '
(f'i(x) E r. for all i .
at m·(x) must be sufficiently smooth, such that (6.1.1) makes sense
This means th ..,, 1 b d d" · '
and also that q,;(x) must satisfy the homogeneous oun ary co~ 1tions. So in fact
the functions q,;(x) span a subspace of r.. Moreo~er: the functions_ <p;(x) should
preferably be linearly independen~ (why?). Now Ritz s method consists of solving
the minimization problem over this subspace.
Since only the ai in (6.1.3) are unknown, this means that the problem reduces to
minimizing over the set a1.--a11:
(6.1.4)

The necessary condition for the existence of a minimum is

aJ[un] = 0, i = l, 2, ... , n . (6.1.5)

aai
(6. 1.5) forms a set of n equations with n unknowns, which under certain conditions,
can be solved uniquely. This produces a solution u11 (x). By increasing the number
of basis functions we hope that u11 (x) converges to the solution u(x) of (6.1.1) and
(6.1.2). It is clear that the choice of the basis functions q>i(x) is essential for the
convergence and especially for the speed of convergence of Ritz's method.
Let us first consider a simple one-dimensional example to show how Ritz's method
behaves in practice. ·

6.1.2 A simple one-dimensional example

Theorem 6.1.1 Let u satisfy the following minimization problem (cf Section 5.1.1)

~J[u] = J{~ (::

0
r- f(x)u(x)} dx, (6.1.6)

r. : {u I u sufficiently smooth; u (O) = O} .

and let un (x) be defined bY (6 ·1·3), then the set of Ritz equations is given by

t fl
._ aj
dcp; dcpj
-;[; dx dx=
fl f(x)q,i(x)dx, i=l,2, ... ,n. (6.1.7)
}- 1 0
0
>
6. The numerical solution of minimization problems 99

Proof
Substitution of (6.1.5) in (6.1.6), (6.1.5) gives

!_ j{1
! (d .Eaj<pj(x))
J=l
2
n
aai 2 dx - f(x)([aj<pj(x))}dx=O . (6.1.8)
0 j=l

Hence ai satisfies (6.1.7)

□
Exercise 6.1.1 Verify Equation (6.1.7)
□
Exercise 6.1.2 Show that the solution of (6.1.6) satisfies the DE

d2u
- dx2 = f (x)' (6.1.9)

with boundary conditions

du
u(O) = 0, dx (I) = 0, (6.1.10)
provided the solution of (6.1.6) is twice differentiable. □
The system of equations (6.1.7) is uniquely solvable if the coefficient matrix S is
non-singular. This system can be written in matrix-vector notation by

Sa= f, (6.1.11)
1
with San (n x n) matrix with elements sij = f 1x if dx,
0

a an (n x 1) vector with elements aj,

1
f an (n x 1) vector with elements f; = Jf (x) cp; (x) dx .
0

There are many possible choices for the basis functions cp;(x), but we shall restrict
ourselves to 2 specific ones.
Theorem 6.1.2 Let the basis functions (f)i(x) be given by

(f)k(x) = sinknx (6.1.12)

then the matrix Sin (6.1.11) has elements

(6.1.13)

and the solution ak satisfies

1
ak = / jf(x) sin(knx) dx. (6.1.14)
k n2
0

□
The basis functions 'Pk(x) are elements of r., since they are analytical functions
and satisfy <t>k(O) = 0. Note that none of them satisfies the natural boundary condi-
tion. Using the orthogonality relations of the cosine we see that the basis functions
<'Pk(x) produce a diagonal matrix S, with diagonal elements (6.1.13)

>
 Numerical methods in scientifi
100 CColllputin
h t E uation (6.1.14) is the result of substituting th g
Exercise 6.1.3 Prove t a q e basis Ju
tions (6.1.12) into (6.1.7) ne-
t the set ak defined by (6.1.14) form the coefficients th CJ
Exercise 6.1.4 Show th ~ t' u(x) using functions sin(krrx) . of e Fourie
expansion of the exact ~o u ion r
Hint: substitute 6.1.3 mto 6.1· 9·
· (x) be given by CJ
Theorem 6.1.3 Let the basis junctions <p;
<pk(x) = i
(6,l.15)
. . by
then the matrix Sin (6.1.11) zs gzven
1 1 1 1 1
4 6 8 10
1 g 6
~ -f2
1
S= 1 i f2 f6...,,. (6,1.16)
6
1 6
fo

0
Exercise 6.1.5 Derive Equation (6.1.16) . 0
s
Matrix in (6.1.16) is a Hilbert matrix. _N?t only ~s this matrix ~ll, it is also very
badly conditioned. Although the matnx 1s non-smgular, numenc~Uy it is not in-
vertible for relatively small values of n (order 10 to 20), on a 16 digits computer.

From these two specific choices for the basis functions we can draw some conclu-
sions with respect to requirements for the basis functions .

6.1.3 Some observations concerning the basis functions

- With respect to the basis function 'Pk (x) defined in (6.1.12) it is clear that the
solution u" (x) converges to the minimization problem, because the Fourier
series is convergent (Exercise 6.1.4).
One can also prove convergence in case of basis functions <pk(x) in (6.1.15),
provided the system of linear equations can be solved.

- Even though the basis functions themselves do not satisfy the natural bound-
ary condition, in the limit the linear combination does in some way, if there
is convergence to the exact solution. In practice ¥x- (x) will be small in some
sense, for n large enough.

- We have seen that with the specific choice (6.1.12) of the basis functions,
the coefficient matrix is diagonal, and therefore the solution of the syStem
of equations is trivial.
This is not a coincidence: these functions form the eigenfunctions of the con-
tinuous eigenvalue problem

d2u (6.1.17)
- dxz =Au; u(O) = 0, u(l) = 0.
These eigenfunctions are orthogonal with respect to the inner product

1
dcp; dcpj
J
0
dxdx dx,
6_The numerical solution of minimization problems 101

which implies that these inner products vanish if i -f. j. Also for more general
problems one can define such an inner p~oduct and again the eigenfunctions
have the sam~ property. ~nfortunately m practice it is almost impossible to
find an analytical expression for the eigenfunctions. Numerical computation
of the eigenfunctions is in general a harder task than solving the system of
equations (6.1.11).

_ On the other hand choosing an arbitrary set of basis functions leads to a full
matrix. Unless the number of basis functions is very small, solution of such
a system is very expensive.
In finite differenc_e meth?ds and finite volume methods we always arrived at
systems of equations with a sparse structure. If we want a sparse matrix in
Ritz's method, it is necessary that most of the integrals

1
d<p; d<pj
J
0
--dx
dx dx

vanish. So the majority of the basis functions must be orthogonal with re-
spect to the inner product defined by these integrals. We shall call such a set
"nearly orthogonal".

- It is obvious that in the limit, the set of basis functions must span the com-
plete space r., otherwise there are elements in r. that can not be represented
as linear combination of basis functions. Besides that, it would be nice if arbi-
trary functions in r. could be approximated accurately with a small number
of basis functions. The basis functions <f'k(x) in (6.1.12) do not satisfy this
property.

Combining all this we come to the following requirements for our set of basis func-
tions:

1 the basis functions must be linearly independent

2 the basis functions must span the complete spacer.

3 the basis functions should be "nearly orthogonal"

4 arbitrary functions in r. must be approximated accurately by a limited num-

ber of basis functions

At first sight it seems very difficult to satisfy all these demands. However, in Sec-
tion 6.2 we shall show how to construct such basis functions by the finite element
method.
We have treated lightly over the convergence of Ritz's method for a good reason:
this is very hard to prove in general. For a specific case of practical importance we
provide a proof: strongly elliptic operators (see Section 5.8.2).

6.1.4 Mathematical theory: convergence of Ritz's method

We consider the convergence of Ritz's method for the specific case of a linear op-
erator satisfying properties (5.9.5)- (5.9.7). In order to do that we need a few tools.
First we recall the definition of a basis for a Hilbert space.

Definition 6.1.1 A family {<p«} E r. is called a basis for the Hilbert space r., if the
following two properties are satisfied.
 Numerical methods in scientific
co.rnputin
102 g

1. Linear independence
f ~;<p; == o implies ~; == 0, i = 1, .. , N .
i=l

2. Completeness ~ d a given E > o, there is a finite linear combination ,r b .

For every u E L, an d h' b' . . O; asrs
. h th t th distance between u an t 1s com mation 1s smaller tha
Junctions sue a e n £.
Infiormula:
:1 {
P.
m } and {~1, t-J2, P. } N
... , t-JN , < 00, sueh that
VE> 0 :J 'Ptr.11 Cf'tr.21 ... , -rtr.N
N
llu - L~;<pir.;llr. < E'
i=l

in which I . llr. is the norm in r..

D

Theorem 6.1.4 The Ritz equations with approximate solution (6.1.3) to solve the mini-
mization problem (5.9.9):
1
minl[u], with J[u] = -2 llulli - (u,f), (6.1.18)
11Er.
are given by
n
[aj(</Ji,Cf'j)L = (f,<p;) i = 1, .. .,n. (6.1.19)
j= l

Exercise 6.1.6 Prove Equation (6.1.19) □

The system of linear equations (6.1.19) has a unique solution if and only if the
coefficient matrix S defined by
(<pi, <p1)L (<p2, <p1)L (q>n,<p1)L
( <pi, <pz)L ( <p2, <p2) L ( <pn, <p2)L
S= (6.1.20)

is non-singular.
S is a Gramm matrix for the set of functions q,1, q,2, .. ., <pn in the space r.. In the
following we assume that {<p;} is a basis for L
Theorem 6.1.5 S defined by (6.1.20) is not singular.
Proof
Suppose there is a non-zero vector« = (a: 1, a: 2, ... , a:n) such that S« = 0.
Then(«, S«) = 0 so
L l)ia:j (<p;, </Jj) L = 0 .
i j
Since the inner product is bilinear this implies
(~a:;<p;, L:>j'Pj)L = 0 or
I j

I ~a:q,;IIL = 0 and because 11-IIL is a norm [:a:cp; == 0 •

I •
I

~y t~e linear indepe~dence of the basis functions, this implies a:; = 0. So S« :::: O
unphes « = 0 and S 1s non-singular. 0
~
umerical solution of minimization problems 103
6. Then
,n 6.1.6 If {<p;} is a basis for E, then approximation (6 .1.3) converges to the solu-
~heore ,f the minimization problem (6.1.18),.
ttOlt UO O;

proofd·ng to (5.9.14),J[u] can be written as

Accor 1
/[u] = ~llu - uolll - ~lluolll, {6.1.21)

where Uo E r. minimizes / [u] over r.. This is a continuous function of the energy
norm (why?), that is
\/£ > 0 :3 i5 > 0 such that
llu - uollL < i5 ⇒ IJ[u] - /[uo]I < £. (6.1.22)
Let u EE"= Span {<pj}?=l minimize /[u] over 1:11 , then
0
/[uo] :S /[uo] :S /[u"], Vu" EE". (6.1.23)
We choose u" using completeness:
:3N 2'. 1, ll'.1, ... , ll'.N such that
"
lluo - u11 IIL < b, with u" = L ll'.jfPj, \/n 2'. N. (6.1.24)
j=l

Continuity of /[u] gives 1/[u"] - /[uo] I < £. Note that£ > 0 is arbitrary, for which
b > Oand N ~ 1 exist, and hence /[u"] ➔ J[uo] as n ➔ oo.
The Squeeze Theorem is applied to (6.1 .23) to conclude that
J[u0] ➔ J[uo] as n ➔ oo. (6.1 .25)
Equation {6.1.21) finally implies llua - uoll ➔ 0 as n ➔ 00, □

6.2 The finite element method in R1

6.2.1 Introduction
Ritz's method can be used to solve the minimization problem and therefore also
the corresponding PDE. The main issue in Ritz' method is the choice of the basis
functions. In Section 6.1.3 we have formulated a number of properties the basis
functions should satisfy, in order to get an attractive solution method.
We derive a construction technique that creates basis functions satisfying all these
properties. The key to this method is the subdivision of the region n into sub-
parts (elements) and an element-wise polynomial approximation of the unknown
function.
First we demonstrate this construction in JR1, subsequently it will be extended to
]R2,

6.2.2 The Poisson equation in R 1

As first example we consider Poisson's equation in one dimension:
d2u
- dx2 = J(x)'
u(O) = 0, (6.2.1)

du(l) =0.
dx
 Numerical methods in scientific
104 cornputing

Exercise 6.2.1 Show that the solution of (6.2.V satisfies the m inimization problem

J{½(::)2 -
1
mml[u]; J[u] = u(x)f(x)} dx .
(6.2..2.)
uEE O
r. : {u sufficient! y smooth ; u ( 0) = 0}

0
.. irement implies that the integral in (6.2.2) makes sen
The smoo thness reql,..U.A se.
The s stem of Ritz equations is given by (6.1 .7). . . .
y tru t the basis functions we subd1v1de the mterval [O l] into b·
In ord er to cons c . . ' su in
elements as shown m Figure 6 .1. -
terva ls ek = Xk - 1, k
[ X ] the '
ek en
e1 )( )( )( )(
)( )(
>< Xn
Xk-1 Xk
Xo XJ
X =l
x=O
Figure 6.1 : Subdivision of the interval (0, 1] in elements.

The solution u is approximated by a pie~ewi~e (l?we_r ord~r) p~lynomial defined

element-wise. The most simple approxunation 1s p1ecew1se linear per element.
Figure 6.2 shows a typical approximation a_~fa function u_ by a piecewise linear
polynomial. Note that the boundary condition u(O) = 0 1s already satisfied by
a(O) = 0.

"1

"n- 1

.\'.
Xo xn-1 n

Figure 6.2: Approximation of u(x) by a(x).

Exercise 6.2.2 Let a be a piecewise linear approximation of u. Then a does not belong to
C1 (0, 1). Let f(x) be a continuous function.
Show that the integral in (6.2.2) remains finite when u is replaced by a. D

The linear interpolation polynomial of the function u ( x) over the element ek is

defined by
uk(x) = X - Xk u(xk-1) + X - Xk-1 u(xk). (6.2.3)
Xk-1 - Xk Xk - Xk-1

Exercise 6.2.3 Show that Formula (6.2.3) is indeed the linear interpolation polynomial.
□

Formally spea~g it is not correct to use u(xk) since u(x) is unknown. It wo~d
be ~tter ~o use u ( xk) • However, as long as there is no confusion possible, we will
orrut the tilde.
6_The numerical solution of minimization problems
-
105

We define linear Lagrangian polynomials lk(x)

(6.2.4)

and write (6.2.3) as

(6.2.5)
Uk denotes u(xk)-
Clearly ek-l (x) and fk(x) are linear on ek and are defined by the relations

f;(xi) = 6;;; i,j = k-1,k. (6.2.6)

6;; is the Kronecker delta, defined by

6i;=0 if i=/=j (6.2.7)

6;; = 1 if i = j. (6.2.8)
These relations define E;(x) uniquely (why?).
From (6.2.5) it is clear that u(x) is a linear function of u0,u1, ... ,un so that we can
write
11
u(x) = LUfP;(x). (6.2.9)
j=O
The function <p;(x) consist of piecewise linear Lagrangian polynomials and may be
considered as a generalized Lagrangian polynomials defined over the whole re-
gion 0.
A typical <pi (x) has been sketched in Figure 6.3.

Figure 6.3: Example of a typical generalized Lagrangian polynomial.

<p; is found taking all coefficients uk = 0 (i =I= k) and u; = l.

Exercise 6.2.4 Sketch the basis functions <po (x) and <pn (x). □

Note that <p;(x) is only non-zero in the elements that contain the node x;.
It is immediately clear that q,;(x) is defined by the following rules:

a. <p;(x) is linear in each element.

(6.2.10)
b. <p;(x;) = bij·
Since uo = 0, (6.2.9) can be written as
n
a(x) = L Uj<fJj(x) . (6.2.11)
j=l

The basis function cpo(x) will be used for non-homogeneous boundary conditions
(see Section 6.2.4).
 106
N1111wrk11l 111~/hods 111 sc1t,, ti ,
1
~
co,,, Pt111,
Irr ,

• ,,;,tis/1111/ ~ri,I is 11st•d (.,·, 1 1 - \' • _ '11

• ,; • IIIii I 1111 111 ' ,' • . . 1 - 1I) '/•
Thtorem 6.~1 ~IIJl/ltl, c , ·11r 1/r,· h1sis ji111ct11111s d,t111l·rl by (6 .2. W) lc11,ls I . Ir~• syN/'
,rR,·1-~ c'lflllllll>IIS• (6. 1.7) ru
I '}
u '''t'}ll//Oluf,t 11,
s~tm, ift'lf 1111/rcms: 's
2 - 1 II I
li
- 1 2 - 1 0
2 -1
"2 h
- 1
,, ::::

(6.2.12)
0 - 1 2 - :I
11,,
-1 1 J,,
1
· II Jir =
Wlf J/(r)11>·(l·)
· Tl . cl.r.
0

. 6.2•S Prove Theorem 6.2.1.

Exercise CJ
. 6.2.6 Compare !i·ystem (6.2.12) with the system obtained by the FDM •
Exercise CJ

6_2_3 Numerical integration

The ng. ht-han d-side vector of (6.2.12) contains

. l · over a function !( x)
an integral
In I one can not compute such an mtegra1ana ytica 11 y, so a numerical ·
genera
. ti·on 1•s req,.;....,d
. t· h l ap.
proXIma ,u... · Since we are mtegra mg over eac e ement separate y an
obvious choice is to use a numerical rule based on the same element. 1
Well-known integration rules are for example
Xt

mid-point rule: j g(x) dx ~ (xk - xk-1)g(xk-112), (6.2.13)

trapezoid rule:
Xt-1
(6.2.14)

Simpson's rule:

(6.2.15)
All these rules can be written in the general form:
Xk

j g(x) dx ~ t wkg(vk), (6.2.16)

xk-1 k=1

with r the number of quadrature points,

wk the weights, and
vk quadrature points.

Exercise 6.2.7 Give r, wk and vkfor the midpoint rule, the trapezoid rule and Simpson's
~ □
th
Anoth er class of integration rules of the shape (6.2.16) are the Gaussian rules. These
me ods are characterized by the fact that integration points and weights are cho-
th
sen such at the highest order of accuracy is reached with a particular number of
-
6.
'[he nwnerical solution of minimization problems 107

. t gration points. Weights and integration points of Gaussian integration rules

: ; be found in various text books, like for example [50] and [37].
/; in (6.2.12) is defined as
1 X; Xj + J

f
Q
J(x)cp;(x) dx = f f (x)<pi(x)
X j- 1
dx + f f (x)cp;(x)
X;
dx. (6.2.17)

The integrand g(x) in (6.2.17) is defined by /(x)<p;(x). We could use every possible
integration rule of type (6.2.16) to compute (6.2.17).
We consider integration over the element [xk- 1, xkl• In the Finite Element Method,
one represents the numerical solution in terms of a linear combination of basis
functions. For the case of linear basis functions, one approximates the function
g(x) by linear interpolation, that is
g(x) ~ g(xk-1)'Pk- 1(x) + g(xk)cpk(x), (6.2.18)

over the interval {xk- 1, xkl- Subsequently, integration over {xk-l, xd gives
Xk- 1 Xk - 1

Jg(x)dx ~ Jg(xk- 1)<f'k- 1(x) + g(xk)<pk(x) dx (6.2.19)

xk - 1 xk - 1

= g(xk - 1) f <f)k- 1(x) dx + g(xk) J <pk(x) dx. (6.2.20)

Using linearity of the basis functions and the relation <p;(xj) = Oij, we get

f
Xk - 1

g(x) dx ~ xk -tk- 1 (g(xk - 1) + g(xk)). (6.2.21)

Xk

Similar rules are derived for higher order basis functions with more quadrature
points. Imagine that one integrates over interval {xk- 1, Xk+ml, l, m 2:: 0, l · m i=- 0.
Let this interval contain nodes xk- 1, xk- 1+1, ... , xk+m, then using the basis func-
tions <f)k- 1, <f)k- 1+1' ... , <f)k+m one can write the following interpolating approxima-
tion for g( x)
k+m
g(x) ~ [, g(xp)<pp(x). (6.2.22)
p=k-1
This interpolation is substituted into the integral over g(x)

Xk + m k+m Xk+m

j g(x) dx ~ [, g(xp) J <pp(x) dx. (6.2.23)

Xk-1 p=k-l Xk-1

This type of quadrature based on interpolation on the FEM basis functions is called
Newton-Cotes rule.
Theorem 6.2.2 The Newton-Cotes rule applied to (6.2.12), the right-hand side vector can
be written as

h (6.2.24)
108 Numerical methods in sc·
ientific co .....
"'Putin g
Exercise 6.2.8 Prove Theorem 6.2.2.

Note that the Newton-Cotes rule applied .

for a linear
. .
approximatio
.
.
ntn]RI .
ti
tical to the trapezoid rule. lf a quadratic approximation 1s used this 1 . •side
ru e is id n-
to Simpson's rule. entica1
Not only the type of interpolation, also the type of integration rule inf!
accuracy of the solution. This subject will be considered in Section 8.?. uences the

6.2.4 Boundary conditions

In our example (6.2.1) we have seen how homogeneous boundary conditio
to be treated. In summary: ns hact
_ homogeneous natural b~~d~ry _conditions pose no problem at all. The
an implicit part of the muurruzation problem. Yare

- homogeneous essen~al b?undary c?ndition~ fix the parameters on the bo

ary. The corresponding interpolation ~ctions are not used as basis UOd-
tions In this way all basis functions satisfy the essential boundary func-
. · condi-
tions.
Non-homogeneous boundary conditions require only a small adaptation u, h
· examp Ie (6 .2 .1) w1·th non-homogeneous bound
· by extendmg
demonstrate this · Yve s all
conditions: ary
d2u
J(x) ,
dx2
u(O) = a,
(6.2.25)
du (l) = b.
dx
The solution of (6.2.25) satisfies the minimization problem
1
minJ[u] =
uEE
j{-21 (ddu ) X
2
- u(x)J(x)} dx - bu(l), (6.2.26)
0
r. : { u I u sufficient!y smooth; u (O) = a}
Exercise 6.2.9 Show that the solution of (6.2 .25) satisfies the minimization problem (6.2.26).
D

In order to apply Ritz's method we define

n n
u(x) = L UfPj(x) = L Uj<pj(x) + uo<po(x) (6.2.27)
j=O j=l

~ga~ we use the linear Lagrangian polynomials £i(x) as basis functions, so (fli(x)
1s defined by (6.2.10). Now it is clear that u 0 = a (why?).
If we use the approximation (6.2.27), the Ritz equations corresponding to (6.2.26)
are equal to

~uj J
1 d
(/Ji d<pj
dxdxdx=
J
1
d ·d
1
fq,idx-uoj d<p1 d<po dx+bq,i(l) (6.2.28)
J-1 O X X
0 0
i = 1, 2, .. ., n.
~ •
_The numerical solution of minimization problems 109
6
1 -1 Uo /o/2
-1 2 -1 U1
-1 2 -1 Ii
U2
1
=h
h
h
-1 2 -1 Un-1 fn - 1
-1 1 Un fn/2 + b/h

Figure 6.4: System of equation before applying essential boundary conditions.

Exercise 6.2.10
a. Derive (6.2 .28).
b. Why is i = 0 not part of (6.2.28)?
c. Which of the functions cp;(x) is non-zero in x = 1?

From Formula 6.2.28 it will be clear that the non-homogeneous essential boundary
condition gives a contribution to the right-hand side. To compute this contribution
we first build the matrix and right-hand side as if there are no essential bound-
ary conditions (see Figure 6.4). Following Exercise 6.2.10 row 1 (corresponding to
<p; = cpo), must be removed. Since the matrix must be square also column 1 must
be removed. This is done by multiplying this column by the given value uo and
subtracting it from the right-hand side as sketched in Figure 6.5.

*-!, Uo
··.)- -:,._._) .·· ..·· ..·· .. ·· ..·· ..·· ..·· . .· .- -· Y-0· .· . .. ·· .· .. . ·· .fo/2· .. ·· ..··..· .·· 't:~olh
·-r:-i'~ 2 -1 u1 Ji -uo / h
1 . 1 2 -1 U2 /2 0
Ti .. =h
-1 2 -1 Un-1 fn-1
-1 1 Un fn/2 + b/h 0

Figure 6.5: Remove row 1. Multiply column 1 by uo and put it into the right-hand
side.
The result of this operation is in Figure 6.6. The inhomogeneous natural boundary

2 -1 Ut ft -1
-1 2 -1 u2 h 0
1 UQ
=h h
h
-1 2 -1 Un-1 fn-1
-1 1 Un fn/2+ b/h 0

Figure 6.6: System of equations after application of essential boundary conditions.

condition also contributes to the right-hand side. This contribution is an immedi-

ate consequence of the minimization problem.
110
Numerical methods in scienfific ~
co.rnputing
Element matrices and element vectors
6.2.5
ct the matrix in (6.2.12) it was necessary to evaluat th .
In order to cons tru . . 1ement-wise. e e mt
.m ( .l .ll} . Since m·(x)
.,. , is defined
. mane manner, the natu egrals
ra1 wa
6
do this is by an element-wise way. y to

1 1 d<p; d<pj dx
dx dx
= tJ
k=l
dcp; dcpj dx .
dx dx (6.2.29)
0 ek

So instead of computing the left-hand side for all i and j, one might first compute
all integrals d d

J 'Pj dx ,
(f) i
dx dx (6.2.30)

for all i and j and add these integrals afterwards to get (6.2.29) . This seems ave
complicated way to compute the integrals. However at most~ of the integrals~
(6.2.30) are different from zero (Why?). We store these four integrals in a small
matrix, the element matrix:
J d"akx- 1 ~
]
dpk -1 dpk - 1 dx dx
5ek _
J
ek
dx dx
ek
dx
(6.2.31)
- [ J ~d'Pk-1
ax dx dx J ax
~~ dx
dx
.
ek ek

In the same way we create the element vector:

(6.2.32)

Once all element matrices and vectors are computed, it is a matter of addition to
compute the large matrix Sand the large right-hand side F. The main advantage
of this approach is that all information of the minimization problem, the type of
approximation in the element as well as the numerical integration rule applied, is
stored locally.
To create the large matrix it is sufficient to know which unknowns are present in
the element and to which entries the entries of the element matrix must be added.
This is called the topology of the problem. The same holds for the large vector on
the right-hand side.
~ is .a_ ~ig adv~tage of the FEM. Once the region is subdivided into e_lemen:i~
it IS sufficient to give a generic algorithm for the contributions of an arbitrary
. E ·ally for
ement. . Ther~ IS no need to worry about neighboring elements. speci
more-dimensional unstructured grids, this is very attractive.

6.2.6 Assembly of the large matrix and vector

We have seen that a11 inforrnation
. for the FEM is stored m. element ma trices, ele-
ct
ment vectors ~d problem topology. The question is now: how can V-:e coi::ge
the l~rge matrix an~ vector from this information. The process of creating~ reuse
m?~.an~ vector is known as assembly. To demonstrate this process w
IlUIUIIUz~tion problem (6.2.2). . Fig-
We consider the su~division of the region [O, l] into 4 elements as shown 0 fcoin
ure 6.7. Element ei is defined by ei = [xi-t, Xi] and the nodes are numbered
 ..
6_The numerical solution of minimization problems 111

)( )( e.3 )(
Xo
Figure 6.7: Subdivision of [0, 1] into 4 elements, and corresponding numbering of
nodes and elements.

to 4. The unknowns have in this special case exactly the same numbering, where
we know that uo = 0, so that the real unknowns are numbered from 1 to 4. In first
instance the large matrix has size (5 x 5) and the large vector (5 x 1). The actual
essential boundary condition is eliminated afterwards. The problem topology of
this case is very simple; each element contains two unknowns.

e1: (0,1),
e2: (1,2),
(6.2.33)
e3: (2,3),
e4: (3,4) .
In the first step the large matrix and vector are cleared:

1 2 3 4 5
-i -i -i -i -i
0 0 0 0 0 t- 1 0
0 0 0 0 0 t- 2 0 (6.2.34)
50 = 0 0 0 0 0 t- 3 fl = 0
0 0 0 0 0 t- 4 0
0 0 0 0 0 t- 5 0

The element matrix for an arbitrary element ek has shape

5ek = 1
Xk - Xk - l
[ 1 -1
-l l
l. (6.2.35)

We apply the Newton-Cotes rule, to obtain the element vector

e Xk - Xk-l [ f(xk-1) ] (6.2.36)

fk= 2 f (xk) .
For the sake of simplicity we assume an equidistant grid with step size
Xk - Xk-l = h.
Adding the first element matrix and right-hand side to (6.2.34) gives

1 -1 0 0 0 f (xo)
-1 1 0 0 0 h f(x1)
51 = ! 0 0 0 0 0 f =
1
z 0 (6.2.37)
h O O 000 0
0 0 0 0 0 O

Next we add 5e2 and Jt2 to 5 1 and /1:

1 -1 0 0 0 f(xo)
-1 2 -1 0 0 2f(x1)
h (6.2.38)
O -1 1 0 0 ¥=-2 J(x2)
O O O O 0 0
O O O O 0 0
 Numerical methods in scientifi'
econ..
112 ·•IJ)ii~
g
d ·ves·
. g this process for e3 an e4 gI .
Repea tin
1 -1 0 0 0
-1 2 -1 0 0
1 0 -1 2 -1 0 f =f
4
= -2h
S == 54 == h O O -l 2 -1 . (6,2.39)
0 0 0 -1 1
xpression as (6.2.12) and (6.2.24).
This is of co_ur~e ~e sa~e e == 0 as described in Figure 6.6 the matnx S
' , and th

_ j,
After the ehrrunatton o uo

right-han: s~d; I[, b~t~~ ~ -l f ~

= [ ~;I;~l j- (.
6240
:
-1 1 f(x4)

This construction seems very long-winded, especially for such a simple one di
mensional problem. However, it is very well s_u~t~d f~r computer implementation~
All one needs is a topology formed by a subd1v1s10n m elements, as well as a
cedure to compute an element matrix and element vector for an arbitrary elernpro-
. 1· d h
The rest is a matter of book keeping. H?~ comp icate t e mesh may be, the as-
ent.
sembly process is always the same. All finite element codes work according to this
principle.

6.2.7 Boundary conditions and assembly

We would like to apply the same procedure as in Section 6.2.6, even in the case of
non-homogeneous boundary conditions. To that end we consider the DE (6.2.25)
with corresponding minimization problem (6.2.26).
In the right-hand side of (6.2.28) we see two extra terms compared to (6.1.11):
1
d((J;d(f)o
-uo
J
0
- - - dx + b((J1·(l).
dx dx (6.2.41)

The integral in the first of these terms is already present in the element matrix of
element e1:

sei =[ £~ ~ dx £Wi dx ]
(6.2.42)
f i-~dx JWWdx ·

~y::::s~~:s::;::
fir~t term in (6_2~~1).
e1 e1

of sei and ~ultiply the remaining part of the first column

term of th~ right-hand side vector, then we get precisely the
This step can easily be perfo db . . . .
marked as a point 'th ~e Ya finite element program, provided point O15
So, even if the first:~ 0~s:~~?al bounda~. condition.
15
a 2 x 2 matrIX· f all not used, It IS conceptually simpler always to create
or e1ements e
The term -bcp(l) onl infl
J k·
also in this case it is tt uences the element vector in the last element. However,
vector. e er not to worry a\,out boundary conditions in the element
In order to create this e tr
x a term · thi
case a point element) co . . we introduce an extra boundary element (in 5
meant to incorporate th' tns1sting of 1 point (x = 1) only. This element is solely
e ennbcpi(l)
~
fhe numerical solution of minimization problems 113
6.

Exercise 6.2.11 Show that the element matrix and element vector for the boundary con-
dition du I
- -b
dx (x=1) -
aregiven by:
(6.2.43)

The elimination of the essential boundary conditions can be described in the fol-
lowing formulae.
Suppose we renumber the unknowns such that we have first all non-prescribed un-
}<nowns (u;) (also called degrees of freedom) and subsequently all unknowns given
by the essential boundary conditions (ub).
The system of equations can be written as:

[~:; t~ l[:; l l = [ ::
(6.2.44)

Since ub is given (6.2.44) can be reduced to

(6.2.45)
and this is the actual system to be solved.
The last set of equations in (6.2.44) contains also some useful information. Suppose
that ub is not given, but that the flux (natural boundary condition) is prescribed.
In that case the last equation would be:

sb;u; +sbbub = £b +b, (6.2.46)

where b is the given flux (see 6.2.43)).
The consequence is that if ub is given and u; has been solved from (6.2.44), the flux
can be approximated by
(6.2.47)
This term is also known under the name reaction force. It represents the flux through
the boundary with essential boundary conditions.
In the next Section we shall extend our example to two-dimensions.

6.2.8 Periodical boundary conditions

Consider the Poisson equation with periodical boundary conditions.

_ ( ) du(O) _ du(l) (6.2.48)

u(O) - u 1 , dx - dx .

Theorem 6.2.3 The minimization problem corresponding to (6.2.48) is given by

1 2
rninJ[u] =
uEE
j{ !2 (du)
dx
- f(x)u(x)} dx, (6.2.49)
0
r.: {u I usufficientlysmooth; u(O) = u(l)}.
Exercise 6.2.12 Prove Theorem 6.2.3. □
 Numerical methods in scientifi
114 C COJl1pllfin
g
. . da® = ~x is a natural boundary conditin.-.
t the boundary condition x
Note tha v 11 for
...... ;nimization problem. t
this u=• Ri ethod we se
In order to apply the tz m
n
u" (x) = L a;cp;(x)'
j=O (6.2.SQ)

unknowns in the first and last node are identifi

the
•th -
w1 ao - n·. a Henced t element are couple d to each other, which is pre . ed · By
doing so, the first an 1as d.ti. c1se1y the
. lb undary con 1 ons.
idea of period ica o
. C ute the matrix and right-hand side for the solution of 6_2_49 .
Exercise 6.2.13 . omp Cotes quadrature. using
linear basis functwns and Newton 0

6.2.9 The structure of finite element packages

. tions it has been made clear that the finite element method.
In thlle ~redv1of us setcomati·zation As a consequence a lot of (commercial) packag1s
we suite or au · k . . es
r the last decades. Most pac ages subdivide the fini•t
have been deve1oPed Ove e
element process in three steps.
• Preprocessing: usually the mesh generation

• Solving: the actual FEM

• Postprocessing: showing the results
The solve part consists globally of the following steps:
Read input and mesh
Compute the structure of the large matrix from the topology
Clear large matrix and vector
for all elements (including boundary elements) do
Compute element matrix and vector
Add element matrix to large matrix
Add element vector to large vector
end for
Apply essential boundary conditions
Solve system of equations
Write results for postprocessing
The crucial step is the computation of element matrix and vector. In fact this part
defines the actual differential equation and type of approximation.
In general one uses preprogrammed finite element subroutines to compute ele-
ment matrix and vector, however, it is also possible that the user supplies his own
element matrix and vector. In this way she may use the general concept of the
FEM, and still solve her own specific problem.

6.3 The finite element method in R 2

6.3.1 The Poisson equation in ]R2
"!e ha~e demo~trated the FEM for the one-dimensional Poisson equati~n usin~
linear interpolations. And even that simple equation showed many of the issues
th~ FEM. In this section we shall extended that example to R 2 . More general cases
will be the subject of Chapter 7.
 6_ The numerical solution of minimization problems 115

Consider Poisson's equation defined on a bounded region nc JR2 with boundary

r === r 1 u r2 u f3 .
-~u = f XE !1
I
(6.3.1)
with boundary conditions

U = g1 ( X) , XE f 1
au
an = g2 (X) , XE f 2 (6.3.2)
au
a:u + an = g3(x), x E f3 (a: ~ 0) .

Figure 6.8: Region O with boundary r = f 1 U r 2 U r 3.

The minimization problem corresponding to (6.3.1), (6.3.2) is given by

minf[u] (6.3.3)
uE!:

with

J[u] = j{;/'vu/ 2 -uf}dO- j g2udf- j g3udf+~j a:u 2 df

o G G G
and .E = {sufficiently smooth/ u = g1 /r1 }.
Exercise 6.3.1 Prove that the PDE formulation (6.3.1) together with (6.3.2) is 'equiva-
lent' to the minimization form (6.3.3). □

To provide a general framework we first apply Ritz's method formally.

First we choose a set of basis functions rp;(x) E .Eo with

l:o = {u I ulr1 = O}. (6.3.4)

Next we choose an arbitrary but known function UB that satisfies

UB(x) = g1 (x), XE f 1 . (6.3.5)

The solution u(x) is approximated by a finite dimensional subset of .E:

n
un (x) =[ Uj(f)j(x) + uB(x) (6.3.6)
j=l
 Numerical methods in scien . .
tilic cofh
~
·••Pui-'llJg
116

Figure 6.9: Subdivision in triangles.

have u" (x) r.. The set of Ritz equations to approximate the
;;::Znweproblem E
(6.3.3) by (6.3.6) is given by:
..
InJ.ni.

t Uj{j ('ilq,; · 'ilq,j) dfl +I "'Mi df} = Jfq,; d[l +Jg2q,; t df

j=l O f3 fl f2

j g q,; d[ - j 'il q,; •'il

3
UB dfl - j •q,;uB df . (6_3_~
[3 0 f3

Exercise 6.3.2 Derive (6.3.7). D

The next step is to provide FEM basis functions. To this end we subdivide the
region into elements and define a polynomial approximation on each element.

6.3.2 Linear elements in JR 2

The extension of the linear line element in 1R1 is the triangle in lR.2. Figure 6.9
shows a typical subdivision of a region into triangles. In order to construct a linear
polynomial on each triangle we need 3 parameters. A natural choice is to use the
function values in the three vertices of the triangle (Figure 6.10).

Figure 6.10: Linear triangle with nodal points.

. added benefi t Of making

This has the
boundanes.
. the approximation continuous across eIement

Following the s d' basis

functions . h ame procedure as in lR 1, it will be clear that the correspon mg
cp, ave the properties:
1) ( ) (6.3.8)
Cf'ix is linear per triangle,
2) 'Pi(Xj) == Jij.
- 6_The numerical solution of minfotization problems 117

A typical basis function is sketched in Figure 6.11.

Figure 6.11: Sketch of a typical linear basis function .

(6.3.8) defines the basis functions implicitly. In order to compute the integral
in (6.3.7), it is necessary to have an explicit expression per element. Consider the
triangle in Figure 6.10).
A linear polynomia 1is defined by
q>;(x) =IX;+ ~;X + 1 ;y. (6.3.9)
(6.3 .8) defines 3 equations for each i to compute the parameters ixi , ~i, 1 ;. Substitu-
tion of (6.3.8) in (6.3.9) leads to the following system of linear equations:

~
[ 1 :X3~ ~~ ] [ ;,1: ;:,2
Y3
;~ ] = [
13
~0 ~0 ~1 ] . (6.3.10)

Exercise 6.3.3 Verify (6.3.10) D

The system of equations (6.3 .10) has a solution if the coefficient determinant f!l (see
(6.3.11)) does not vanish
1 X1 Y1
t!l = 1 X2 Y2 (6.3.11)
1 X3 Y3
t::. in (6.3.11) can be expressed as
t::. = (x2 - xi) (y3 - y1 ) - (y2 - Y1)(x3 - x1), (6.3.12)
which is twice the area of the triangle in Figure 6.10, as will be shown in Section
8.2.
Exercise 6.3.4 Prove (6.3.12) .
Hint: subtract thefirst row from the second and tile tllird row. □
If the orientation of the nodes is counterclockwise t::. is positive, otherwise it is
negative.
Exercise 6.3.4 shows that the system is regular as long as the area of the triangle
differs from 0.
The solution of system of equations (6.3.10) is given by
1 1 1
~1 = -x(Y2 -y3), ~2 = -x(y3 -y1), ~3 = ~(Y1 -y2) 1

')'1 = -/j1 (x 3 - x2), ')'2 = -/j1 (x, - X3),

1 .
Ll
.
')'3 = A(.l; 2 - xi), (6.3.13)
a; =1- ~jXj - ')';y;.
 Numerical methods in seientificc0 .......
.,
11lplJ1-;_

118 . of I6.3.10) .
is the solutwn ""8
3 131
. e 3.5 S/1olV t/1nt _( 6· or a, ~1 and , 1 and subtract the first e
.f i .0
fxerc1s 6 ·
. . r r,, ,latc t1,c eqaat1ons 'f-' h k
t this process Jor the ot er un nowns. quat, n fro., lh
11
H111 1• r /lird one. Repea '
0
second and I . ts to evaluate the integrals in formula (6 3 (J

NoW we t,ave all . ingre 6 2d5ienwe only need to compu t e the element mat· ·.7) . As w
. s,cuon . . ' rix •nd e
hav• seen ,11 ele-
n1ent vector. Jl sider the case that a, gz and g3 are all equal to z
,e sha con .h L t h 11 ero s
first o a " . te rals ill (6.3.7) varus . a er on we s a pay attention;0 o that
f U
.1 ll boundary 111 g . . _,10rnogeneous boundary problems. these

~ ~~: ~~~
boun ary iJ1tegra 1.s U1f uuthe linear triangle correspon d'mg to (6.3.7) is gi
d
T1'• e\enient n,atnx or S'• [ ~: ] ' ven by
6 14
S31 S32 S33 ( .3· l

With Sii :::: J'V (f)i · 'V (f)j dO.

□
Ck
Exercise 6.3.6 Show tl,,at (6.3.14) is the element matrix corresponding to (6.3.7).

From (6.3.9)- (6.3.14) it follows that

S;j ===
,~1
T(~i~ j + ,i'Yj) .
(6.3.15)

The element vector for the Jinear triangle corresponding to (6.3. 7) is given by:

(6.3.16)

(6.3.17)
with
f; === j J(x)(f);(x) dO.
ek
□
Exercise 6.3.7 Verify (6.3.16) and (6.3.1?).

We shall have to evaluate (6.3.17) numerically.

6.3.3 Numerical integration in Rn

Numerical
consider theintegration
mo in 1R1 has been the subject of Section 6.2.3. In . we
2 this section
.m R • Integration re generalhcase Of 111 . over triangles in R or tetrahedro ns
· tegrabon
In JR.23 and R3 we cOverd ot_er ~es of elements will be the sub1·ect of Section 8.7.
trapezoidal rule ors·an enve mt · rules of the same type as rnid-po111t • ru}e '

~~
egration
impson's
exact Y· Besides that lo th 1e, Yintegrating polynomials of a certalll degt ee
b . .
Gaussianl integratio~ ru'i tnangles and tetrahedrons it is possible to conSlfllcl
ous text books. (See for es. eights and integration points can be found in ntUJler·
example [501).
Definition 6•3·1 A Slmplex
. in JR" 15· the convex hull of n + 1 points in R".
A simplex in R1 . . 0
zs an interval' in ]R.2 a triangle
. and in JR3 a tetrahedron.
~
. fhe numerical solution of minimization problems 119
6

11
ext theorems gives a general formula for integration of powers of linear basis
fhe . ns over simplices. It is very useful.
ft,1J1C 0 0
2
theorem 6.3.1 Let S be a triangle in R and let I),, be the determinant defined by
1 X1 Y1
I),,= 1 X2 Y2 , (6.3.18)
1 X3 Y3

·th x xz x3 the vertices of S.

Wl 11 '

Let A;(x) be the linear basis functions over S defined by

A;(x) linear
A;(xj) = b;j i,j,= 1,2,3. (6.3.19)

Then the following general integration rule holds:

for all m; 2:: 0.

Proof: See Holand and Bell (1969)[20], page 84.

Exercise 6.3.8 Use Theorem 6.3.1 to show that

!s A; = ~
6'
(6.3.20)

D
This theorem can be extended to n dimensions:

Theorem 6.3.2 Let S be a simplex in lR n and let

1
1
X1,1
X1,2
X2,1
X2,2
/j,

Xn,1
Xn,2

Xn+l,n
I
be the determinant defined by

,
(6.3.21)

1 Xn+l,1 Xn+l,2

with x1, x2, · · · , Xn + l the vertices of S, and Xi,j the /h component of x;.
Let A; (x) be the linear basis functions over S defined by

,\; (x) linear

,\i(xi) = c5ij i, j, = l, 2, · .. , n + 1 . (6.3.22)

Then the following general integration rule holds:

Js ,\ m1 ,\ m2 ... ,\mn+l
1 2 n+l
dO = m1 !m2! ... mn+l ! \[j,I '
([:m;+n)!
i

for all m; 2:: o.

 · sa·en1-.c
Numerical methods m
~ CC{)
120 l?ipll~
.. 6 3 2 to show that
1 T1 1eore"1 · ·
Exercise 6.3.9 App y
x2 h
j Ai dx = 2.
X\

~-(x) be defin ed as in (6.3.22). Then

Theorem 6.3.3 Let ' •
11 + 1
L Ai(x) = 1.
i= l

. 6 310 Prove Theorem 6.3.3.

Exercise • ·
D
Exercise 6.3.11 Show that Js dO = ~ -
Hint: use Theorem 6.3.2. 0
• 12 Find the midnoi11t rule fo r a triangle and a tdmhedro 11 .
Exercise 6•3• r . . · I , I ·, . . •
. ti 1e 1111
mt: . 'dpo 1·1.1t rule is a one po111 t mtegra
•
t1011 r11
.
t t wt 1:, e:rnct fo r l111t'ar rioly,
. r
.
lllnllnl~
H
Determine this point by in tegrnti11g the l111ear basis fi mct,011s. 0
2
Exerci · the Newton-Cotes ru le for a triangle i11 lR. witli 1; 111•11r IL
. Prove that
•se 6.313 .
,.1s1s
Junction s is given by

f g(x) dO = \~\ (g(x,) +g(x2) +g(x3)) . (6.3.25)

s
Hint: use Theorem 6.3.2. 0

Exercise 6.3.14 Show that if the Newton-Cotes rule is applied to (6.3.16), (6.3.1i), tlrt•
element vector is given by
f'• = I~\ [ j(::\ I. (6,126)
l 1(x 1 3
)
0

6.3.4 Boundary conditions

The way in which essential boundary conditions are treated is independent of the
dimension of the space. With respect to natural boundary conditions we fo\\ow
a similar approach as in 1R1 . In that case we introduced point elements to trcnt
the _extra term bcp;(l). In equation (6.3.7) we find four boundary integrals, three of
which are related to natural boundary conditions.

JtXq>;q>; df I
(6.3.27)

f3
(6.3,28)
Jg2cp; df,
f2
(6.3,29)
Jg3cp; df.
f3
►
e umerical solution of minimizatio n ~bl
6. Th 11 e~
121
;.,ce we use linear triangles we actually ap .·
5J,J• • 7 proxtmate th b
lines. In Section 8. we shall return to the conse e .oundary by straight
quences of this approximation.
for the moment we assume that the bounda . .
undary lines of the subdivision. Of course 1-t~ is ex~ctly given by the straight
bo is possible t O dd th
the integrals (6.3.27)-(6.3.29) to all element m tri a e contribution
of . a ces and v t h
to boundary triangles that have a side in comm "th ec ors t at correspond
. thi . on wi r 2 or r F
·onal point of view sis not so desirable be hi 3· rom a computa-
ti cause t s mean th .
have the same type of element matrix and ele t s at not ~ll triangles
cussion in JR 1 it is natural to introduce extra linme; vecto~. So following our dis-
the integrals in (6.3.27)-(6.3.29) These lin ele ements JUS t for the computation
of . . . . · e e ements (also c U d b d
rnents) are implicitly defined by the boundary a d th b .. ~ e . o1:11' ary ele-
for example Figure 6.12. n e SU division m triangles, see

~ : . :~. ..... . ,~:=---~~ · :::'--~--

>:-- ~ .. ! : .. . : .. : .... : . . . . . : : : : : : : : .

Figure 6.12: Subdivision in triangles ahd line elements.

A typical line element is sketched in Figure 6.13.

Figure 6.13: Example of a linear line element.

Only two base functions differ from zero on this element (why?), so the element
matrix must have size (2 x 2) and the element vector (2 x 1). The element matrix
for the boundary elements along r3 is given by

with
Sij = JIX(f)i<f}j df . (6.3.30)
ek

The element vector for the r3 boundary elements are defined by

f'• =[ j~ 1, f; = Ig3tp; df
e

D
Exercise 6.3.15 Give the line element matrices and vectors along f 2·