Introduction to the Finite Element Method

Dr. C. S. Jog
Dept. of Mechanical Engineering
Indian Institute of Science
Bangalore-560012
Chapter 1

Calculus of Variations

1.1 Introduction
In 1696, the Swiss mathematician Johann Bernoulli challenged his colleagues
to solve an unresolved issue called the brachistochrone problem which is as
follows: Find the shape of a frictionless chute between points (1) and (2) in
a vertical plane such that a body sliding under the action of its own weight
goes from point (1) to point (2) in the shortest time. Bernoulli originally
specified a deadline of six months, but extended it to a year and half at the
request of Leibniz, one of the leading scholars of the time, and the man who
had, independently of Newton, invented the differential and integral calculus.
The challenge was delivered to Newton at four p.m. on January 29, 1697.
Before leaving for work the next morning, he had invented an entire new
branch of mathematics called the calculus of variations, used it to solve the
brachistochrone problem and sent off the solution, which was published at
Newtons request, anonymously. But the brilliance and originality of the
work betrayed the identity of its author. When Bernoulli saw the solution,
he commented, We recognize the lion by his claw. Newton was then fifty-
five years old. Much of the formulation of this discipline was also developed
by the Swiss mathematician Leonhard Euler.
What is the relevance of the calculus of variations to the finite element
method? For one, the finite element formulation can be derived in a more
direct way from a variational principle than from the corresponding energy
functional. But perhaps more importantly, there are several problems such
as, for example, inelastic deformations, where the governing differential equa-
tions can be cast in variational form but for which no corresponding energy
functional exists. In such cases, developing the finite element formulation
from the variational principle is the only alternative. We start by presenting
the principles of the calculus of variations. The treatment presented here
follows closely the one given in Shames and Dym [2].

1
1.2 Principles of the Calculus of Variations
While dealing with a function y = f (x) in the ordinary calculus, we often
want to determine the values of x for which y is a maximum or minimum. To
establish the conditions of local extrema, we expand f (x) as a Taylor series
about a position x = a:
1 d2 f 
 
df 
f (x) = f (a) + (x a) + 2
(x a)2 + . . . , (1.1)
dx x=a
 2! dx x=a


or alternatively,
1 d2 f 
 
df 
f (x) f (a) = x + (x)2 + . . . ,
dx x=a 2! dx2 x=a
where x = x a. Since x can be positive or negative, the necessary
condition for a local maxima or minima at an interior point (i.e., point not
lying on the boundary of the domain) x = a is

df 
= 0.
dx x=a
The point x = a is called an extremal point.
In the calculus of variations, we are concerned with extremizing (mini-
mizing or maximizing) functionals which, roughly speaking, are functions of
functions. An example of a functional is
Z x2
I= F (x, y, y )dx,
x1

where y denotes the derivative of y with respect to x.1 Note that I is a

function of a function y(x), and x is the independent variable. We now
present examples of some functionals.

1.2.1 Examples of functionals

(i) The Brachistochrone Problem:
This is the problem which was mentioned in the introduction. Referring
to Fig. 1.1, the problem can be stated as: Find the shape, y = f (x), of a
frictionless chute between points 1 and 2 in a vertical plane so that a body
starting from rest and sliding under the action of its own weight goes from 1
to 2 in the shortest possible time.
If v denotes the velocity of the body, and ds denotes an element of length
along the curve, the time for descent is given by
s  2
Z 2 Z 2p 2 2
Z x2
ds dx + dy 1 dy
I= = = 1+ dx.
1 v 1 v x1 v dx
1
We shall follow the convention of denoting derivatives by primes throughout the re-
mainder of these notes.

2
 1 x

y=f(x)
y
g 2

Figure 1.1: The brachistochrone problem.

The velocity v at a height y can be found using the equation of conservation

of energy:
1 2 1
mv1 mgy1 = mv 2 mgy,
2 2
where the sign on the potential energy terms is negative because we have
assumed our y-axis as positive in the direction of the gravitational acceler-
ation. Since we have fixed our origin at point 1, and since the body starts
from rest, y1 = v1 = 0, and we get v = 2gy. Hence, the time for descent is
Z x2 s
1 + (y )2
I= dx. (1.2)
x1 2gy
We shall prove later that the curve y = f (x) which minimizes I is a cycloid.
(ii) The Geodesic Problem:
This problem deals with the determination of a curve on a given surface
g(x, y, z) = 0, having the shortest length between two points 1 and 2 on this
surface. Such curves are called geodesics. Hence, the problem is:
Find the extreme values of the functional
Z x2 p
I= 1 + (y )2 + (z )2 dx, (1.3)
x1

subject to the constraint

g(x, y, z) = 0. (1.4)
Hence, this is a constrained optimization problem.
(iii) The Isoperimetric Problem:
The isoperimetric problem is: Of all the closed non-intersecting curves having
a fixed length L, which curve encloses the greatest area A? From calculus, we
know that the answer can be given in terms of the following contour integral:
1 2
I  
1 y x
I
I =A= (x dy y dx) = x y d, (1.5)
2 2 1
subject to the constraint
2
" 2  2 #1/2
x y
I I
L= ds = + d, (1.6)
1

3
 y
1 ~y(x)
y(x)
~y(x)
2
Figure 1.2: Extremal path and the varied paths.

where is a parameter. We shall show later that, as expected, the curve

which solves the isoperimetric problem is a circle.
Note that each of the above three problems are different in character.
The brachistochrone problem is an unconstrained problem since there are no
constraints while performing the minimization. In the geodesic problem, we
have a constraint on a function of the independent variables (see Eqn. 1.4),
whereas in the isoperimetric problem we have a constraint on a functional
(see Eqn. 1.6).

1.2.2 The first variation

Consider the functional
Z x2
I= F (x, y, y ) dx, (1.7)
x1

where F is a known function. Assume that y(x1 ) and y(x2 ) are given (e.g.,
the starting and ending points in the brachistochrone problem). Let y(x) be
the optimal path which extremizes (minimizes or maximizes) I, and let y(x)
be varied paths as shown in Fig. 1.2. Note that y(x) agrees with y(x) at the
end points y(x1 ) and y(x2 ) since these are known. Our goal is to expand the
functional I in the form of a Taylor series similar to Eqn. 1.1 so that the usual
principles of calculus can be used for obtaining the path which minimizes or
maximizes I. Towards this end, we relate y(x) and y(x) by using a scalar
parameter as follows:
y(x) = y(x) + (x), (1.8)
where (x) is a differentiable function having the requirement (x1 ) = (x2 ) =
0, so that y(x1 ) = y(x1 ) and y(x2 ) = y(x2 ). We see that an infinite number
of paths can be generated by varying . In terms of the varied paths y(x),
the functional in Eqn. 1.7 can be written as
Z x2 Z x2

I=
F (x, y, y )dx = F (x, y + , y + )dx. (1.9)
x1 x1

For = 0, y coincides with the extremizing function y, and I reduces to the

extreme value of the functional. Since I is a function of the scalar parameter

4
, we can expand it as a power series similar to the one in Eqn. 1.1. Thus,
 
 d
I  d 2 
I 2
I = I + + + ..., (1.10)
  
d  d2  2!
 
=0
=0 =0

or, alternatively,
 
 d 
I d2 I 2
I I = I I = + 2 + ....

d  d  2!

=0
=0 =0

The necessary condition for I to be extreme is


dI
= 0,
d 

=0

or in other words,
x2
F y F y
Z   
+ dx = 0.
x1 y y =0

Since y = y + , we have y/ = , y / = , y|=0 = y and y |=0 = y .

Hence, the above equation reduces to
Z x2  
F F
+ dx = 0 (x).
x1 y y

Integrating by parts, we get

Z x2 x Z x2  
F F  2 d F
dx + dx = 0 (x).
x1 y y x1 x1 dx y

Since (x1 ) = (x2 ) = 0, the above equation reduces to

Z x2   
F d F
dx = 0 (x),
x1 y dx y

which implies that  

F d F
= 0. (1.11)
y dx y
Equation 1.11 is the famous Euler-Lagrange equation.

1.2.3 The delta operator

We saw that the process of deriving the Euler-Lagrange equation was quite
straightforward, but cumbersome. In order to make the procedure more
mechanical, we introduce the delta operator. Again referring to Fig. 1.2,
we define the variation of y(x) by

y(x) = y(x) y(x). (1.12)

5
Thus, y is simply the vertical distance between points on different curves at
the same value of x. Since
 
dy dy dy d d(y)
= = (y y) = ,
dx dx dx dx dx

the delta-operator is commutative with the derivative operator. The delta-

operator is also commutative with the integral operator:
Z Z Z Z Z
y dx = y dx y dx = (y y) dx = (y) dx.

From Eqn. 1.12, y = y + y and y = y + y . Assuming the existence of

an extremal function, y(x), and its derivative, y (x), we can expand F as a
Taylor series about the extremal path and its derivative, using increments y
and y . That is
 
F F
y + y +o (y)2 +o (y )2 ,



F (x, y +y, y +y ) = F (x, y, y )+
y y

or, equivalently,
 
F
F
y + y +o (y)2 +o (y )2 .



F (x, y +y, y +y )F (x, y, y ) =
y y
(1.13)
The left hand side of the above equation is the total variation of F , which
we denote as (T ) F . We call the bracketed term on the right-hand side of the
equation as the first variation, (1) F . Thus,

(T ) F = F (x, y + y, y + y ) F (x, y, y ), (1.14)

F F
(1) F = y + y . (1.15)
y y
Integrating Eqn. 1.13 between the limits x1 and x2 , we get
x2 x2 x2  
F F
Z Z Z

F (x, y+y, y +y ) dx F (x, y, y ) dx = y + y dx
x1 x1 x1 y y
+ o (y)2 + o (y )2 .

This, may be written as

Z x2  
F F
I I = y + y dx + o (y)2 + o (y )2 .




x1 y y

Note that I represents the extreme value of the functional occurring at y = y.

We call the term on the left hand side of the above equation as the total
variation of I, and denote it by (T ) I, while we call the first term on the right
as the first variation and denote it by (1) I. Thus,

(T ) I = I I,

6
 x2  
F F
Z
(1)
I= y + y dx.
x1 y y

Writing y as d(y)/dx, and integrating the expression for (1) I by parts, we

get Z x2     x2
(1) F d F F 
I=
y dx +
y  . (1.16)
x1 y dx y y 
x1

Since y(x1 ) = y(x2 ) = 0, the above equation reduces to

Z x2   
(1) F d F
I= y dx,
x1 y dx y
so that we have
(T ) I = (1) I + o (y)2 + o (y )2



Z x2   
F d F 2

2


= y dx + o (y) + o (y ) .
x1 y dx y

In order for I to be a maximum or minimum, (T ) I must have the same

sign for all possible variations y over the interval (x1 , x2 ). Note that y
at any given position x can be either positive or negative. Hence, the only
way (T ) I can have the same sign for all possible variations is if the term in
parenthesis vanishes, which leads us again to the Euler-Lagrange equation
given by Eqn. 1.11. Thus, we can conclude that the a necessary condition
for extremizing I is that the first variation of I vanishes, i.e.,
!

I
(1) I = 0 y.

=0

We are now in a position to solve some example problems.

1.2.4 Examples
In this subsection, we shall solve the brachistochrone and geodesic problems
which were presented earlier.
The Brachistochrone Problem:
We have seen that the functional for this problem is given by Eqn. 1.2. We
thus have
Z x2 s
1 1 + (y )2
I= dx,
2g x1 y
or in other words s
1 + (y )2
F = .
y
Substituting for F in the Euler-Lagrange equation given by Eqn. 1.11, we
get the differential equation

2yy + 1 + (y )2 = 0.

7
To solve the above differential equation, we make the substitution y = u to
get
du 1 + u2
u = ,
dy 2y
or
udu dy
Z Z
2
= .
1+u 2y
Integrating, we get
y(1 + u2 ) = c1 ,
or, alternatively
y(1 + (y )2 ) = c1 .
The above equation can be written as
Z
ydy
x= + x0 .
c1 y

If we substitute y = c1 sin2 ( /2), we get x = c1 ( sin )/2 + x0 . Since

x = y = 0 at = 0, we have x0 = 0. Hence, the solution for the extremizing
path is
c1
x= ( sin ),
2
c1
y = (1 cos ),
2
which is the parametric equation of a cycloid.
Geodesics on a Sphere:
We shall find the geodesics on the surface of a sphere, i.e., curves which
have the shortest length between two given points on the surface of a sphere.
Working with spherical coordinates, the coordinates of a point on the sphere
are given by

x = R sin cos ,
y = R sin sin , (1.17)
z = R cos .

Hence, the length of an element ds on the surface of the sphere is

p
ds = dx2 + dy 2 + dz 2
q
= R d2 + sin2 d2
"  2 #1/2
d
= Rd 1 + sin2 .
d

Since "  2 #1/2

d
Z Z
2
I= ds = R 1 + sin d,
d

8
we have "  2 #1/2
d
F = R 1 + sin2 .
d
Noting that F is a function of but not of , the Euler-Lagrange equation
when integrated gives the solution
cos c1 sin sin sin c1 sin cos = c2 cos ,
where c1 and c2 are constants. Using Eqn. 1.17, the above equation can also
be written as
(cos c1 )y (cos c1 )x = c2 z,
which is nothing but the equation of a plane surface going through the origin.
Thus, given any two points on the sphere, the intersection of the plane passing
through the two points and the origin, with the sphere gives the curve of the
shortest length amongst all curves lying on the surface of the sphere and
joining the two points. Such a curve is known as a Great Circle.
In this example, the constraint that the curve lie on the surface of the
sphere was treated implicitly by using spherical coordinates, and thus did
not have to be considered explicitly. This procedure might not be possible
for all problems where constraints on functions are involved. In such cases,
we have to devise a procedure for handling constraints, both of the functional
and function type. We shall do so in Sections 1.2.6 and 1.2.7.

1.2.5 First variation with several dependent variables

We now extend the results of the previous sections to finding the Euler-
Lagrange equations for a functional with several dependent variables but
only one independent variable. Consider the functional
Z x2
I= F (y1, y2 , ..., yn , y1 , y2 , ..., yn , x) dx
x1

Setting the first variation of the above functional to zero, i.e., (1) I = 0, we
get
Z x2  
F F F F
y1 + + yn + y1 + + yn dx = 0.
x1 y1 yn y1 yn
Since the variations can be arbitrary, choose y2 = y3 = . . . = yn = 0.
Then since y1 is arbitrary, we get the first Euler-Lagrange equation as
 
F d F
= 0.
y1 dx y1
Similarly, by taking y2 as the only nonzero variation, we get the second
Euler-Lagrange equation with the subscript 1 replaced by 2. In all, we get n
Euler-Lagrange equations corresponding to the n dependent variables:
 
F d F
= 0, i = 1, 2, . . . , n. (1.18)
yi dx yi

9
 x1 x2

k1 k2 k1
M M

Figure 1.3: Spring mass system with two degrees of freedom.

To illustrate the above formulation, consider the problem shown in Fig. 1.3.
Our goal is to write the equations of motion of the two mass particles. In
order to do that, we use Hamiltons principle (see Section 6.2) which states
that for a system of particles acted on by conservative forces, the paths taken
from a configuration at time t1 to a configuration at time t2 are those that
extremize the following functional:
Z t2
I= (T V ) dt,
t1

where T is the kinetic energy of the system, while V is the potential energy
of the system. For the problem under consideration, we have
1 1
T = M x21 + M x22 ,
2 2
1 1 1
V = k1 x21 + k2 (x2 x1 )2 + k1 x22 .
2 2 2
Thus,
t2  
1 1 1 2
Z
M x21 + M x22 k1 x1 + k2 (x2 x1 )2 + k1 x22

I= dt.
t1 2 2 2
Note that x1 , x2 are the two dependent variables, and t is the independent
variable in the above equation. From Eqn. 1.18, we have
 
F d F
= 0,
x1 dt x1
 
F d F
= 0.
x2 dt x2
Substituting for F , we get
M x1 + k1 x1 k2 (x2 x1 ) = 0,
M x2 + k1 x2 + k2 (x2 x1 ) = 0,
which are the same as Newtons laws applied to the two masses. By inte-
grating the above equations using the initial conditions x1 (0), x2 (0), x1 (0),
x2 (0), we can find the subsequent motions of the masses x1 (t) and x2 (t).
In this example, we could have more easily employed Newtons laws di-
rectly. However, there are many problems where it is easier to proceed by the
variational approach to arrive at the governing equations of motion. In addi-
tion, variational formulations also yield the appropriate boundary conditions
to be imposed as we shall see in future sections.

10
1.2.6 Functional constraint
In the isoperimetric problem, we saw that we have to minimize a functional
subject to a constraint on a functional. So far, we have dealt only with the
minimization of unconstrained functionals. We now show how to carry out
the extremization process in the presence of a functional constraint.
Consider the problem of minimizing
Z x2
I= F (x, y, y ) dx,
x1

subject to the restriction

Z x2
J= G(x, y, y ) dx = constant.
x1

If we still use a one-parameter family of admissible functions of the form

y = y(x) + y, where y = , then we might no longer be able to satisfy the
constraint Z x2

J= G(x, y, y ) dx = constant. (1.19)
x1

Hence, we now need to introduce additional flexibility into the system by

using a two-parameter family of admissible functions of the form

y(x) = y(x) + y1 + y2 .

The variation y1 is arbitrary, while the variation y2 is such that Eqn. (1.19)
is satisfied. For extremizing I, we set (1) I to zero, while (1) J is zero since
J is a constant. Hence, using the method of Lagrange multipliers, we have

(1) I + (1) J = 0,

where is a constant known as the Lagrange multiplier. We can write the

above condition as
Z x2 
d F

d F
  
F F
y1 + y2 dx = 0,
x1 y dx y y dx y
where F = F +G. We choose such that the second integral term vanishes.
Then we have Z x2 
d F

F
y1 dx = 0,
x1 y dx y
which, by virtue of the arbitrariness of y1 , yields the required Euler-Lagrange
equations:
F d F
 
= 0. (1.20)
y dx y
So far we have considered only one constraint equation. If we have n
constraint equations
Z x2
Gk (x, y, y ) dx = ck k = 1, 2, . . . , n,
x1

11
then we need to use an n + 1-parameter family of varied paths

y = y + y1 + y2 + + yn+1 ,

out of which, say y1 is arbitrary, and the remaining are adjusted so as to

satisfy the n constraint equations. The n Lagrange multipliers are chosen
such that n Z x2 
F d F
X 

yi+1 dx = 0,
i=1 x 1
y dx y

Pn
where F = F + k=1 k Gk . The arbitrariness of y1 finally leads to the
Euler-Lagrange equations
F d F
 
= 0.
y dx y
Finally, for p independent variables, i.e.,
Z x2
I= F (x, y1, y2 , . . . , yp , y1 , y2 , . . . , yp ) dx,
x1

and n constraints
Z x2
Gk (x, y1 , y2, . . . , yp , y1 , y2 , . . . , yp ) dx = ck , k = 1, 2, . . . , n,
x1

the Euler-Lagrange equations are given by

F d F
 
= 0, i = 1, 2, . . . , p,
yi dx yi
where F = F + nk=1 k Gk .
P
We are now in a position to tackle the isoperimetric problem. We are
interested in finding a curve y(x), which for a given length
s
Z 2  2  2
x y
L= + d,
1

encloses the greatest area

2  
1 y x
Z
A= x y d.
2 1
Note that is the independent variable, and x and y are the dependent
variables. Denoting x/ by x and y/ by y, we have
1
F = (xy y x) + (x2 + y 2)1/2 .
2
The Euler-Lagrange equations are
 
y d y x
+ 2 = 0,
2 d 2 (x + y 2 )1/2

12
  
x d x y
+ = 0.
2 d 2 (x2 + y 2 )1/2

Integrating with respect to , we get

x
y = c1 ,
(x2
+ y 2 )1/2
y
x+ 2 = c2 .
(x + y 2 )1/2

Eliminating from the above two equations, we get

(x c2 ) dx + (y c1 ) dy = 0.

Integrating the above equation, we obtain

(x c2 )2 + (y c1 )2 = c23 ,

which is the equation of a circle with center at (c2 , c1 ) and radius c3 . Thus, as
expected, the curve which encloses the maximum area for a given perimeter
is a circle.

1.2.7 Function constraints

We have seen in the geodesic problem that we can have a constraint on
functions of the independent variables. In this subsection, we merely present
the Euler-Lagrange equations for function constraints without going into the
derivation which is quite similar to the derivation for functional constraints
presented in the previous subsection.
Suppose we want to extremize
Z x2
I= F (x, y1 , y2 , . . . , yn , y1 , y2 , . . . , yn ) dx,
x1

with the following m constraints

Gk (x, y1 , y2 . . . , yn , y1 , y2 , . . . , yn ) = 0, k = 1, 2, . . . , m.

Then the Euler-Lagrange equations are

F d F
 
= 0, i = 1, 2, . . . , n,
yi dx yi

where F = F + m
P
k=1 k (x)Gk . Note that the Lagrange multipliers, k (x),
are functions of x, and are not constants as in the case of functional con-
straints.

13
1.2.8 A note on boundary conditions
So far, while extremizing I, we assumed that the value of the extremizing
function y(x) was specified at the endpoints x1 and x2 . This resulted in
y(x1 ) = y(x2 ) = 0, and hence to the elimination of the boundary terms
in Eqn. 1.16. Such a boundary condition where y(x) is specified at a point
is known as a kinematic boundary condition. The question naturally arises:
Are there other types of boundary conditions which are also valid? The
answer is that there are.
Let Z x2
I= F (x, y, y ) dx g0 y|x=x2 ,
x1

where g0 is a constant. Rewrite using Eqn. 1.16 the necessary condition for
extremizing the functional I as
Z x2       
(1) F d F F F 
I= y dx+ g0 y y  = 0 y.
x1 y dx y y x2 y x1
(1.21)
Since the variations y are arbitrary, we can choose variations y such that

y|x=x1 = y|x=x2 = 0.

Then we get
x2   
F d F
Z
y dx = 0,
x1 y dx y
for all y which satisfy y|x=x1 = y|x=x2 = 0. This in turn leads us to the
Euler-Lagrange equations
 
F d F
= 0.
y dx y
Hence, Eqn. 1.21 can be written as
   
F F 
g0 y y = 0 y.
y x2 y x1

Consider variations such that y|x=x2 = 0. From the above equation, we get

F 
y = 0,
y x1

which in turn implies that


F 
=0 or y prescribed at x1 .
y x=x1

We have already encountered the second boundary condition which we have

called the kinematic boundary condition. The first condition is something we
have not encountered so far, and is known as the natural boundary condition.

14
Similar to the above process, by considering variations such that y|x=x1 = 0,
we get the boundary conditions at x = x2 as

F 
= g0 or y prescribed at x2 .
y x=x2

Summarizing, the variational process has resulted in not only yielding the
governing differential equation but also the boundary conditions which can
be specified. We can either specify
1. kinematic boundary conditions at both endpoints,

2. natural boundary boundary conditions at both endpoints, or

3. a natural boundary condition at one of the endpoints and a kinematic

one at the other.
Note that physical considerations rule out specifying both natural and kine-
matic boundary conditions at the same boundary point. To give some exam-
ples, in the boundary value problems of elasticity, the kinematic conditions
would correspond to specified displacements, while the natural boundary con-
ditions would correspond to specified tractions. In problems of heat transfer,
the kinematic and natural boundary conditions would correspond to specified
temperature and specified normal heat flux, respectively.

1.2.9 Functionals involving higher-order derivatives

So far, we have considered only first-order derivatives in the functionals. We
now extend the work of finding the Euler-Lagrange equations and the appro-
priate boundary conditions for functionals involving higher-order derivatives.
Consider the functional
Z x2
I= F (x, y, y , y , y ) dx M0 y |x=x2 V y |x=x2 .
x1

For finding the extremizing function y(x), we set the first variation of I to
zero, i.e.,
Z x2  
(1) F F F F
I= y + y + y + y dxM0 y |x=x2 + V y |x=x2 = 0.
x1 y y y y

Integrating by parts, we get

x2
d3 d2
       
F F d F F
Z
2 + y dx
x1 dx3
y dx y dx y y
  x2      x2
F 
 d F F 

+ V y y
y 
x dx y y 
x1
 2    1   x2
d F d F F
+ y  M0 y |x=x2 = 0 y.

+ 2

dx y dx y y x1

15
Since the variations, y, are arbitrary, the above equation yields the Euler-
Lagrange equation

d3 d2
     
F F d F F
+ = 0,
dx3 y dx2 y dx y y

and the boundary conditions

F
y specified or = V,
y
 
d F F
y specified or = M0 ,
dx y y
d2
   
F d F F
y specified or 2

+ = 0.
dx y dx y y

In each row, the first condition corresponds to the kinematic boundary con-
dition, while the second one corresponds to the natural boundary condition.
We shall consider the extension of the above concepts to more than one in-
dependent variable directly by dealing with problems in elasticity and heat
transfer in three-dimensional space. Before we do that, we consider some
basic concepts in functional analysis. We shall need these concepts in formu-
lating the variational principles and in presenting the abstract formulation.

1.3 Some Concepts from Functional Analysis

If V is a Hilbert space, then we say that L is a linear form on V if L : V ,
i.e., L(v) v V , and if for all u, v V and , ,

L(u + v) = L(u) + L(v).

a(., .) is a bilinear form on V V if a : V V , i.e., a(u, v) for all

u, v V , and if a(., .) is linear in each argument, i.e., for all u, v, w V and
, , we have

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

a(u + v, w) = a(u, w) + a(v, w).

A bilinear form a(., .) on V V is symmetric if

a(u, v) = a(v, u) u, v V.

A symmetric bilinear form a(., .) on V V is said to be a scalar product on

V if
a(v, v) > 0 v V.
The norm associated with the scalar product a(., .) is defined by

kvka = [a(v, v)]1/2 v V.

16
 Now we define some Hilbert spaces which arise naturally in variational
formulations. If I = (a, b) is an interval, then the space of square-integrable
functions is  Z  b
L2 (I) = v: v 2 dx < .
a

the scalar product associated with L2 is

Z b
(v, w)L2(I) = vw dx,
a

and the corresponding norm (called the L2 -norm) is

Z b 1/2
2
kvk2 = v dx = (v, v)1/2 .
a

Note that functions belonging to L2 can be unbounded, e.g., x1/4 lies in

L2 (0, 1), but x1 does not. In general, v(x) = x lies in L2 (0, 1) if < 1/2.
Now we introduce the Hilbert space

H 1 (I) = {v : v L2 (I), v L2 (I)} .

This space is equipped with the scalar product

Z b
(v, w)H 1(I) = (vw + v w ) dx,
a

and the corresponding norm

Z b 1/2
1/2 2 2
kvkH 1 (I) = (v, v)H 1 (I) = (v + (v ) ) dx .
a

For boundary value problems, we also need to define the space

H01 (I) = {v H1 (I) : v(a) = v(b) = 0} .

Extending the above definitions to three dimensions

 Z 
2
L2 () = v : v d < ,

 
1 v
H () = v L2 (); L2 () ,
xi
Z
(v, w)L2 () = vw d,

Z 1/2
2
kvkL2 () = v d ,

Z
(v, w)H 1 () = (vw + v w) d,

17
 Z 1/2
2
kvkH 1 () = (v + v v) d ,

H01 () = {v H1 () : v = 0 on } .

For vector fields, we have

Z 1/2
kvkL2 = (v12 + +v22 v32 ) d
,

Z "X 3 X
3  2 #
vi
kvk2H 1 = kvk2L2 + d.
i=1 j=1
xj

1.4 Variational Principles in Elasticity

In this course, we shall restrict ourself to single-field or irreducible variational
statements, i.e., variational statements which have only one independent so-
lution field. The dependent solution fields are found by pointwise applica-
tion of the governing differential equations and boundary conditions. For
example, in the case of elasticity, one can have a variational principle with
the displacement field as the independent solution field, and the strain and
stress fields as the dependent solution fields, while in the case of heat transfer
problems, we can have temperature as the independent, and heat flux as the
dependent solution field.
We start by presenting the strong form of the governing equations for
linear elasticity. In what follows, scalars are denoted by lightface letters,
while vectors and higher-order tensors are denoted by boldface letters. The
dot symbol is used to denote contraction over one index, while the colon
symbol is used to denote contraction over two indices. When indicial notation
is used summation over repeated indices is assumed with the indices ranging
from 1 to 3. For example, t u = ti ui = t1 u1 + t2 u2 + t3 u3 , (C : )ij = Cijklkl ,
where i and j are free indices, and k and l are dummy indices over which
summation is carried out.

1.4.1 Strong form of the governing equations

Let be an open domain whose boundary is composed of two open,
disjoint regions, = u t . We consider the following boundary value
problem:
Find the displacements u, stresses , strains , and tractions t, such that

+ b = 0 on , (1.22)
= C( 0 ) + 0 on , (1.23)
1
(u) + (u)t on ,

= (1.24)
2
t = n on , (1.25)
t = t on t , (1.26)

18
 u = u on u , (1.27)
where 0 and 0 are the initial strain and initial stress tensors, n is the out-
ward normal to , and t are the prescribed tractions on t . In a single-field
displacement-based variational formulation which we shall describe shortly,
we enforce Eqns. 1.23-1.25, and Eqn. 1.27 in a strong sense, i.e., at each
point in the domain, while the equilibrium equation given by Eqn. 1.22, and
the prescribed traction boundary condition given by Eqn. 1.26 (which is the
natural boundary condition) are enforced weakly or in an integral sense. For
the subsequent development, we find it useful to define the following function
spaces:
Vu = {v H 1 () : v = 0 on u },
Lu = u + Vu where u H 1 () satisfies u = u on u .
In what follows, we shall use direct tensorial notation. The corresponding
formulation in indicial notation is given in [2].

1.4.2 Principle of virtual work

We obtain the variational formulation using the method of weighted residuals
whereby we set the sum of the weighted residuals based on the equilibrium
equation and the natural boundary condition to zero:
Z Z
v ( + b) d + v (t t) d = 0 v Vu . (1.28)
t

In the above equation, v represent the variation of the displacement field,

i.e., v u. Using the tensor identity
( t v) = v ( ) + v : , (1.29)
we get
Z Z
t
 
( v) v : + v b d + v (t t) d = 0 v Vu .
t

Now applying the divergence theorem to the first term of the above equation,
we get
Z Z Z
t
( v) n d [v : v b] d + v (t t) d = 0 v Vu .
t
(1.30)
Using Eqn. 1.25 and the fact that v = 0 on u , we have
Z Z Z
t
( v) n d = v t d = v t d v Vu .
t

Hence, Eqn. 1.30 simplifies to

Z Z Z
v : d = v b d + v t d v Vu .
t

19
Using the symmetry of the stress tensor, the above equation can be written
as
1
Z Z Z
t

v + (v) : d = v b d + v t d v Vu .
2 t

Using Eqn. 1.24, the single-field displacement-based variational formulation

can be stated as
Find u Lu such that
Z Z Z
(v) : d = v b d + v t d v Vu , (1.31)
t

where is given by Eqn. 1.23. What we have shown is that if u is a solution

to the strong form of the governing equations, it is also a solution of the weak
form of the governing equations given by Eqn. 1.31. Note that the variational
formulation has resulted in a weakening of the differentiability requirement
on u. While in the strong form of the governing equations, we require u
to be twice differentiable, in the variational formulation, u just needs to be
differentiable. This is reflected by the fact that u lies in the space Lu . The
variational formulation presented above is often known as the principle of
virtual work.
We have shown that the governing equations in strong form can be stated
in variational form. The converse if also true provided the solution u is
assumed to be twice differentiable, i.e., a solution u Lu of the variational
formulation is also a solution of the strong form of the governing equations
provided it has sufficient regularity. To demonstrate this, we start with
Eqn. 1.31. A large part of the proof is a reversal of the steps used in the
preceding
R proof. RUsing the symmetry of the stress tensor, we can write

(v) : d
R as v : d. Using the tensor identity given by Eqn. 1.29
to eliminate v : d, we get
Z Z Z Z
t
v b d + v t d ( v) d + v ( ) d = 0 v Vu .
t

Now using the divergence theorem, and the fact that v = 0 on u , we get
Z Z
v ( + b) d + v (t n) d = 0 v Vu . (1.32)
t

Since the variations v are arbitrary, assume them to be zero on t , but not
zero inside the body. Then
Z
v ( + b) d = 0,

which in turn implies that Eqn. 1.22 holds. Now Eqn. 1.32 reduces to
Z
v (t n) d = 0 v Vu .
t

20
Again using the fact that the variations are arbitrary yields Eqn. 1.25.
Note that the principle of virtual work is valid for any arbitrary consti-
tutive relation since we have not made use of Eqn. 1.23 so far. Hence, we
can use the principle of virtual work as stated above for nonlinear elasticity,
plasticity etc., provided the strains are small (since we have used Eqn. 1.24
which is valid only for small displacement gradients). The variational state-
ment for a linear elastic material is obtained by substituting Eqn. 1.23 in
Eqn. 1.31. We get the corresponding problem statement as
Find u Lu such that
Z Z
(v) : C(u) d = v t d +
t
Z
v b + (v) : C0 (v) : 0 d
 
v Vu . (1.33)

For elastic bodies, we can formulate the method of total potential energy
which is equivalent to the variational formulation already presented as we
now show

1.4.3 Principle of minimum potential energy

For a hyperelastic body (not necessarily linear elastic), we have the existence
of a strain-energy density function U such that
U
= , (1.34)

or, in component form,
U
ij = .
ij
As an example, for a linear elastic body with the constitutive equation given
by Eqn. 1.23, we have
1
U = (u) : C(u) + (u) : 0 (u) : C0 .
2
Using Eqn. 1.34, Eqn. 1.31 can be written as
U
Z Z Z
v b d + v t d = : (v) d
t
Z
= (1) U d
Z

= (1) U d v Vu ,

where we have used the fact that the delta-operator and integral commute.
Using the same fact on the left hand side of the above equation, we get
Z Z 
(1)
b u d + t u d = (1) U.
t

21
or, alternatively,
(1) [U V ] = (1) = 0,
where is the potential energy defined as
= U V, (1.35)
with
Z
U= U d,
Z Z
V = b u d + t u d.
t

As an example, the potential energy for a linear elastic material is

1
Z Z Z
= (u) : C(u) d t u d b u d+
2 t
Z
(u) : 0 (u) : C0 d (1.36)
 

We have shown that the principle of virtual work is equivalent to the van-
ishing of the first variation of the potential energy. We now show that the
converse is also true. i.e., (1) = 0 implies the principle of virtual work.
Thus, assume that (1) = (1) [U V ] = 0 holds, or alternatively,
Z Z Z
(1)
U d b v d t v d = 0 v Vu .
t

Using the fact that

U
(1) U = : = : (v),

we get the principle of virtual work given by Eqn. 1.31.
We have shown that the principle of virtual work is equivalent to ex-
tremizing . In fact, more is true. By using the positive-definiteness of the
strain-energy density function, U, one can show that the principle of virtual
work is equivalent to minimizing the potential energy with respect to the
displacement field (see [2]).

1.5 Variational Formulation for Heat Trans-

fer Problems
As another example of how to formulate the variational problem for a given
set of differential equations, we consider the problem of steady state heat
transfer with only the conduction and convection modes (the radiation mode
makes the problem nonlinear). Analogous to the presentation for elasticity,
we first present the strong forms of the governing equations, and then show
how the variational formulations can be derived from it. Note that in contrast
to the elasticity problem, the independent solution field (the temperature) is
a scalar, so that its variation is also a scalar field.

22
1.5.1 Strong form of the governing equations
Let be an open domain whose boundary is composed of two open,
disjoint regions, = T q . The governing equations are
q = Q on ,
q = kT on ,
T = T on T ,
h(T T ) + (kT ) n = q on q ,
where T is the absolute temperature, q is the heat flux, Q is the heat gen-
erated per unit volume, k is a symmetric second-order tensor known as the
thermal conductivity tensor, n is the outward normal to , T is the ab-
solute temperature of the surrounding medium (also known as the ambient
temperature), T is the prescribed temperature on T , and q is the externally
applied heat flux on q . Note that q is taken as positive when heat is supplied
to the body. Typical units of the field variables and material constants in
the SI system are
T : o C, kij : W/(m-o C), Q : W/m3 , q : W/m2 , h : W/(m2 -o C).

For presenting the variational formulation, we define the function spaces

VT = {v H 1 (); v = 0 on T },
LT = T + VT where T H 1 () satisfies T = T on T .

1.5.2 Variational formulation

Similar to the procedure for elasticity, we set the weighted residual of the
governing equation and the natural boundary condition to zero:
Z Z
v [ (kT ) + Q] d+ v [q h(T T ) (kT ) n] d = 0 v VT .
q

Using the identity

v( (kT )) = (vkT ) v (kT ),
the above equation can be written as
Z Z
(vkT ) d [v (kT ) vQ] d+

Z
v [q h(T T ) (kT ) n] d = 0 v VT .
q

Applying the divergence theorem to the first term on the left hand side, and
noting that v = 0 on T , we get
Z Z
[v (kT ) vQ] d + v [q + h(T T )] d = 0 v VT .
q

23
Taking the known terms to the right hand side, we can write the variational
statement as
Find T LT such that
Z Z Z Z
v (kT ) d + vhT d = vQ d + v(q + hT ) d v VT .
q q
(1.37)
We have shown how the variational form can be obtained from the strong
form of the governing equations. By reversing the steps in the above proof,
we can prove the converse, i.e., the governing differential equations can be
obtained starting from the variational formulation, provided the solution has
sufficient regularity. The energy functional corresponding to the weak form
is
Z  
1 1 2
Z
= [T (kT ) QT ] d + hT hT T qT d. (1.38)
2 q 2

One can easily verify that (1) = 0 yields the variational formulation, and
vice versa. By using the positive-definiteness of k, we can show that the vari-
ational formulation is equivalent to minimizing the above energy functional.
We have seen that it is possible to give an energy formulation which is
equivalent to the variational formulation, both in the case of the elasticity
and heat transfer problems. So the following questions arise
1. Is it possible to provide an abstract framework for representing prob-
lems in different areas such as elasticity and heat transfer?
2. Can one find the appropriate energy functional from the variational
statement in a systematic way?
The answer to these questions is that an abstract formulation is possible for
self-adjoint and positive-definite bilinear operators. The elasticity and heat
transfer problems already considered are examples which fall under this cat-
egory. There are several other problems from fluid mechanics and elasticity
which also fall in this category. An abstract formulation provides a unified
treatment of such apparently diverse problems. With the help of the abstract
formulation, one can not only derive the energy functional in a systematic
way, but also prove uniqueness of solutions. In addition, it is also possible
to give an abstract formulation for the finite element method based on such
an abstract formulation which makes it possible to derive valuable informa-
tion such as error estimates without having to treat each of the problems,
whether they be from elasticity, fluid mechanics or heat transfer, individually.
We now discuss the details of such a formulation.

1.6 An Abstract Formulation

Given that V is a Hilbert space (for details on Hilbert spaces, see any book
on functional analysis), suppose the differential operator Lu = f can be

24
expressed in the form
Problem (V): Find u such that

a(u, v) = L(v) v V

where a(., .) is a bilinear operator, and L(.) is a linear operator with the
following properties:
1. a(., .) is symmetric (or self-adjoint).
2. a(., .) is continuous, i.e.,

a(v, w) kvkV kwkV ,

where is a positive constant.

3. a(., .) is V-elliptic (or positive-definite), i.e., there is a constant > 0
such that
a(v, v) kvk2V .

4. L(v) is continuous, i.e., there is a constant such that

L(v) kvk .

Consider the following minimization problem

Problem (M): Find u such that

F (u) = min F (v)

v

where
1
F (v) = a(v, v) L(v).
2
We have the following result relating problems (V) and (M):
Theorem 1.6.1. The minimization problem (M) is equivalent to problem
(V). The solution to Problems (M) and (V) is unique.
Proof. First we prove that M = V. Let v V and . Since u is a
minimum,
F (u) F (u + v) , v V.
Using the notation g() F (u + v), we have

g(0) g() ,

or, alternatively, g (0) = 0. The function g() can be written as

1
g() = a(u + v, u + v) L(u + v)
2
1 2
= a(u, u) + a(u, v) + a(v, u) + a(v, v) L(u) L(v)
2 2 2 2
1 2
= a(u, u) L(u) + a(u, v) L(v) + a(v, v),
2 2
25
where we have used the symmetry of a(., .) in obtaining the last step. Now
g (0) = 0 implies a(u, v) = L(v) for all v V .
Now we prove that V = M. Let v = u + v. Then

1 2
F (v) = a(u, u) L(u) + [a(u, v) L(v)] + a(v, v).
2 2
Since a(u, v) = L(v), the above equation reduces to

2
F (v) = F (u) + a(v, v) F (u).
2
Noting that F (v) = F (u) when = 0, we can write the above equation as

F (u) = min F (v),

vV

which is nothing but Problem (M).

To prove uniqueness, assume that u1 and u2 are two distinct solutions of
problem (V). Then

a(u1 , v) = L(v) v V,
a(u2 , v) = L(v) v V.

Using the fact that a(., .) is a bilinear operator, we get

a(u1 u2 , v) = 0 v V.

Choosing v = u1 u2 , we get

a(u1 u2 , u1 u2 ) = ku1 u2 ka = 0.

By the V-ellipticity condition, this implies that

ku1 u2 kV = 0,

which by the definition of a norm implies that u1 u2 = 0, or that u1 = u2 .

Note that if a(., .) is non-symmetric, we still have uniqueness of solutions,
but no corresponding minimization problem (M). We now consider some
examples to illustrate the above concepts.
Examples
Consider the governing equation of a beam

d2 d2 v
 
EI 2 q = 0,
dx2 dx

with the boundary conditions

d2 v  d2 v 
   
dv  d
v(0) = = 0; EI = M0 ; EI 2  = V.
dx x=0 dx2 x=L dx dx x=L

26
The variational form is obtained as follows:
First define the appropriate function space in which v lies in:

Vu = H 2 (0, L), = 0, = 0 at x = 0 ,


where

H 2 (0, L) = { L2 (0, L), L2 (0, L), L2 (0, L)} .

Multiplying the governing differential equation by a test function and in-

tegrating, and adding the residues of the boundary terms, we get
L
d2 d2 v d2 v
   
  
d
Z
EI 2 q dx + EI 2 M0
0 dx2 dx dx dx
  2
  x=L
d d v 
+ V EI 2  = 0 Vu .
dx dx x=L

Integrating by parts twice, we get

L  L L
d2 v d2 d2 v 2

d d v d
Z
 
EI 2 2 dx + EI 2  EI 2 
0 dx dx dx dx dx dx 0
 2
   0 2
  Z L
d v d d dv 
+ EI 2 M0 + V EI 2 
 = q dx Vu .
dx dx x=L dx dx x=L 0
(1.39)

Since |x=0 = d/dx|x=0 = 0, we can write Eqn. 1.39 as

a(v, ) = L() Vu ,

where
L
d2 v d2
Z
a(v, ) = EI 2 2 dx,
0 dx dx
Z L 
d 
L() = q dx + M0 V |x=L .
0 dx x=L

Clearly a(., .) is symmetric. To prove V-ellipticity, note that

L 2
d2 v
Z 
a(v, v) = EI dx > 0 for v 6= 0.
0 dx2

If a(v, v) = 0, then d2 v/dx2 = 0, which implies that v = c1 x + c2 . For a

simply supported beam or a beam cantilevered at one or both ends, we get
c1 = c2 = 0, leading to v = 0. Hence, a(., .) is symmetric and positive-
definite. The functional corresponding to the variational form is
1
= a(v, v) L(v)
2

27
 L 2 L
d2 v
 
1 dv 
Z Z
= EI dx qv dx M0 + V v|x=L . (1.40)
2 0 dx2 0 dx x=L
Now consider the variational formulation for the elasticity problem given
by Eqn. 1.33. The bilinear operator, a(., .) and the linear operator L(.) are
given by
Z
a(u, v) = (v) : C(u) d,
Z Z
v b + (v) : C0 (v) : 0 d.
 
L(v) = v t d +
t

a(., .) is clearly symmetric. It is positive definite if C is positive definite. The

functional to be minimized is
1
= a(u, u) L(u)
2Z
1
Z Z
= (u) : C(u) d t u d b u d+
2 t
Z
(u) : 0 (u) : C0 d,
 

which is the same expression as given by Eqn 1.36.

Finally, consider the variational formulation for the heat transfer problem
given by Eqn. 1.37. We now have
Z Z
a(T, v) = v (kT ) d + vhT d,
q
Z Z
L(v) = vQ d + v(q + hT ) d.
q

The functional to be minimized is

1
= a(T, T ) L(T )
2Z Z  
1 1 2
= [T (kT ) QT ] d + hT hT T qT d,
2 q 2

which is the same expression as given by Eqn. 1.38.

1.7 Traditional Approximation Techniques

We discuss two popular traditional approximation techniques, namely, the
Rayleigh-Ritz method and the Galerkin method.

1.7.1 Rayleigh-Ritz method

In this method, we find an approximate solution using the weak form of
the governing equations, i.e., using either Problem (V) or Problem (M) in

28
section 1.6. We seek an approximate solution of the form
n
X
u = 0 + cj j ,
j=1

where 0 satisfies the essential boundary condition. The functions j , j =

1, . . . , n are zero on that part of the boundary. For example, if u(x0 ) = u0
where x0 is a boundary point then 0 (x0 ) = u0 and i (x0 ) = 0 for i = 1, . . . , n.
If all the boundary conditions are homogeneous then 0 = 0.
The variation of u is given by v = ni=1 ci i . Substituting for u and v in
P
the variational formulation, we get
n
X n
X n
X
a(0 + cj j , ci i ) = ci L(i ), ci .
j=1 i=1 i=1

Simplifying using the bilinearity of a(., .), we get

n
X n X
X n n
X
ci a(0 , i ) + ci cj a(j , i ) = ci L(i ), ci .
i=1 i=1 j=1 i=1

Choosing first c1 = 1 and the remaining ci as zero, then choosing c2 = 1 and

the remaining ci as zero, and so on, we get
n
X
Aij cj = L(i ) a(0 , i ).
j=1

In matrix form, the above equation can be written as

Ac = f , (1.41)

where
Aij = a(j , i ), (1.42)
is a symmetric matrix (since a(., .) is symmetric), and

fi = L(i ) a(0 , i ). (1.43)

One obtains exactly the same set of equations by using minimization of

the functional I given by
n n
! n
1 X X X
I(u) = a 0 + cj j , 0 + ci i L(0 ) cj L(j )
2 j=1 i=1 j=1
n n n
! n
1 X 1 X X X
= a(0 , 0 ) + a(0 , ci i ) + a cj j , ci i L(0 ) cj L(j ).
2 i=1
2 j=1 i=1 j=1

Now setting the partial derivatives of I with respect to the undetermined

coefficients, i.e.,
I I I
= 0, = 0, . . . , = 0,
c1 c2 cn

29
we get Eqn. 1.41. Note however, that when a(, ., ) is not symmetric, we do not
have the existence of the functional I. Hence, deriving the matrix equations
directly from the variational form (V) is preferred.
The approximation functions i should satisfy the following conditions:
1. They should be such that a(i , j ) should be well-defined.
2. The set i , i = 1, . . . , n should be linearly independent.
3. {i } must be complete, e.g., when i are algebraic polynomials, com-
pleteness requires that the set {i } should contain all terms of the
lowest-order order admissible, and upto the highest order desired.
Example:
We shall consider the same example of a beam that we considered in sec-
tion 1.6. Recall that the variational formulation was

a(v, w) = L(w) w Vu ,

where
L
d2 v d2 w
Z
a(v, w) = EI dx,
0 dx2 dx2
Z L 
dw 
L(w) = qw dx + M0 .
0 dx x=L
Since the specified essential boundary conditions are v = dv/dx = 0 at x = 0,
we can take 0 = 0, and select i such that i (0) = (0) = 0. Choosing
1 = x2 , 2 = x3 and so on, we can write v(x) as
n
X
v(x) = cj j (x); j = xj+1 .
j=1

The governing matrix equations are

Ac = f ,

where A and f are obtained using Eqns. 1.42 and 1.43:

Z L
Aij = EI(i + 1)ixi1 (j + 1)jxj1 dx
0
ij(i + 1)(j + 1)Li+j1
= EI
i+j1
qLi+2
fi = + M0 (i + 1)Li .
i+2
For n = 2, we get

4L 6L2 c1 3 4 2
EI = qL + M0 L .
6L2 12L3 c2 12 3L 3L

30
On solving for c1 and c2 and substituting in the expression for v(x), we get

5qL2 + 12M0 2
 
qL 3
v(x) = x x.
24EI 12EI
For n = 3, we get

2 1 2
4 6L 8L c qL + 2M0
1 3
EI 6L 12L2 3 = 1 3
18L c2 qL + 3M0 L .

4
1
8L2 18L3 28.8L4
c3 5
4
qL + 4M0 L 2

which leads to
qx2 2 2 M0 x2
v(x) = (6L 4Lx + x ) + .
24EI 2EI

1.7.2 Galerkin method

The Galerkin method falls under the category of the method of weighted
residuals. In the method of weighted residuals, the sum of the weighted
residuals of the governing differential equation and the natural boundary
conditions is set to zero, i.e.,
Z Z
(Lu f )w + (B u t)w = 0,

where u is the approximate solution, w is a weighting function, and denotes

the boundary over which the natural boundary condition is applied. In the
Galerkin method, the weighting function is constructed using the same basis
functions as are used for the approximate solution u. Thus, if
n
X
u = 0 + cj j ,
i=1

then w is taken to be of the form

n
X
w= ci .
i=1

If the governing equation permits, one can carry out integration by parts,
and transfer the differentiation from u to as in the Rayleigh-Ritz method.
Note, however, that the weighted residual method is more general than the
Rayleigh-Ritz method since no variational form is required. Also note that
the approximating functions in the Galerkin method are of a higher order
than in the Rayleigh-Ritz method. The Rayleigh-Ritz and Galerkin methods
yield identical results for a certain class of problems, e.g., self-adjoint, second-
order ordinary differential equations with homogeneous or non-homogeneous
boundary conditions or the linear elasticity problem. But, in general, they
can yield different results.

31
 As an example, consider the ordinary differential equation:

d2 y dy
x2 2
+ 2x 6x = 0,
dx dx
with the boundary conditions y(1) = y(2) = 0. The exact solution of the
above equation is
6
y = + 3x 9.
x
In the Galerkin method, we choose an approximate solution in the form of a
polynomial which satisfies the two essential boundary conditions:

y = c1 1 = c1 (x2 3x + 2).

Corresponding to this solution, we have

w = c1 (x2 3x + 2).

The weighted residual statement for the choice c1 = 1 is given by

Z 2 2

2 d y dy
x + 2x 6x 1 dx = 0,
1 dx2 dx

which in turn yields

Z 2
d2 1

2 d1
x 1 2 c1 + 2x1 c1 6x1 dx = 0.
1 dx dx

Solving the above equation, we get c1 = 1.875. To improve the accuracy, we

can take more terms in the expression for y, e.g.,

y = (x 1)(x 2)(c1 + c2 x).

One can compute c1 and c2 using the Galerkin method and confirm that the
Rayleigh-Ritz method also gives the same results.

1.8 Drawbacks of the Traditional Variational

Formulations
We have studied two representative traditional variational methods for find-
ing approximate solutions to differential equations, namely the Rayleigh-Ritz
method and the Galerkin method. However, all such traditional methods suf-
fer from drawbacks which we now discuss. Later, we shall see how the finite
element method overcomes these drawbacks.
The main difficulty in the traditional methods is in constructing the ap-
proximation functions. When the given domain is geometrically complex, the
selection becomes even more difficult. An effective computational method
should have the following features:

32
 1. The method should be convergent.

2. It should be easy to implement, even with complex geometries.

3. The formulation procedure should be independent of the shape of the

domain and the specific form of the boundary conditions.

4. It should involve a systematic procedure which can be easily imple-

mented.

The finite element method meets these requirements. In this method,

we divide the domain into geometrically simple domains. The approxima-
tion functions (also known as basis functions or shape functions) are often
algebraic polynomials that are derived using interpolation theory. Once the
approximation functions have been derived, the method of obtaining the un-
known coefficients is exactly the same as in the Rayleigh-Ritz or Galerkin
method.

33
Chapter 2

Formulation of the Finite

Element Method

In this chapter, we discuss the formulation of the finite element method, first
in an abstract setting, and then applied to problems in elasticity and heat
transfer.

2.1 Steps Involved in the Finite Element Method

The steps involved in a finite element formulation are as follows:

1. Discretize the domain into a collection of elements. Number the nodes

of the elements and generate the geometric properties such as coordi-
nates needed for the problem.

2. Assemble the element level matrices to form the global equation. For
example, for a linear statics problem, we get a matrix equation of the
form
Ku = f .

3. Impose the boundary conditions.

4. Solve the assembled equations.

5. Carry out the postprocessing of the results.

2.2 Abstract Formulation

Before showing how the finite element formulation is carried out for elasticity
or heat transfer, we show how it is formulated in an abstract setting, so that
later it is clear how the various special formulations fit into this more general
framework.
First choose a finite-dimensional subspace Vh of V with dimension m + n,
where m is the prescribed degrees of freedom, and n is the free degrees of

34
freedom to be solved for. Let (1 , 2, . . . , n+m ) be the basis for Vh so that
any v Vh has the unique representation
n+m
X
v= i i , i .
i=1

The discrete analogue of problems (M) and (V) are

Find uh Vh such that

F (uh ) F (v) v Vh .

or, equivalently,
a(uh , v) = L(v) v Vh . (2.1)
The approximate solution, uh , is of the form
m n
(0 )i i +
X X
uh = i i ,
i=1 i=1

and the corresponding variation is of the form

n
X
v= i i .
i=1

Substituting for uh and v in the finite dimensional version of Problem (V)

given by Eqn. 2.1, and successively choosing as (1, 0, 0, . . .), (0, 1, 0, . . .)
etc., we get
n m
(0 )i i , j ) = L(j ) j = 1, n.
X X
a( i i , j ) + a(
i=1 i=1

Using the bilinearity of a(., .), we can write the above equation as
n m
a((0 )i , j )i = L(j ) j = 1, n,
X X
a(i , j )i +
i=1 i=1

or in matrix form as
A = b A,
where is an m 1 vector, is an n 1 vector, and

Aji = a(i , j ) (nn),

Aji = a((0 )i , j ) (nm),
bj = L(j ) (n1).

Note that the stiffness matrix A is symmetric and positive definite since
a(., .) is symmetric and V-elliptic:

t A = a(v, v) kvk2 0 .

35
2.3 Finite Element Formulation for Elasticity
In this section, we consider the finite element formulation for the linear elas-
ticity problem based on Eqn. 1.33. Observe that this variational formulation
involves dealing with fourth-order tensors such as the elasticity tensor, C,
whose matrix representation is given by a 3 3 3 3 matrix. Thus, carry-
ing out the assembly of the stiffness matrix and force vector in a computer
implementation based on Eqn. 1.33 would be quite cumbersome. What we
need is an equivalent form of Eqn. 1.33 which would lead to simpler matrix
manipulations in a computer implementation.
Towards this end, using the symmetry of the strain and stress tensors,
we replace them by vectors such that the strain energy density expression
in terms of the stress and strain components remains the same. This leads
to the use of engineering strain components instead of the tensorial compo-
nents. For example, if we consider two-dimensional plane stress/plane strain
problems, then we have
: = xx xx + xy xy + yx yx + yy yy
= xx xx + 2xy xy + yy yy
= tc c ,

where (with xy = 2xy )

xx xx
c = yy ; c = yy .


xy xy

Note that c is the vector of engineering strain components. The subscript

c on and in the above equation, denotes that the strain and stress com-
ponents are written in column form. The elasticity tensors for an isotropic
material in the case of plane stress/plane strain are 3 3 matrices given by

1 0 1 0
E E
C p. stress = 1 0 ; C p. strain = 1 0 .

2
1 (1 + )(1 2)
0 0 12
0 0 12
2

Note that we now simply have

c = C(c 0c ) + 0c , (2.2)

instead of the earlier relation = C(0 )+ 0 . The virtual work statement

given by Eqn. 1.33 written in terms of c and c is
Find u Lu such that
Z Z
t
[c (v)] Cc (u) d = v t t d +
t

36
 Z
v t b + [c (v)]t C0c [c (v)]t 0c d
 
v Vu . (2.3)

To establish the finite element formulation, we introduce the interpolation

functions or shape functions, N , and write the displacement field, u, as

uh = N u + N u.
In the above equation, N are the shape functions associated with the vector
of unknown displacement degrees of freedom u, while N are the shape func-
tions associated with the vector of known displacement degrees of freedom,
on u . Typically, the prescribed displacement, u, is zero, and then u
u, = 0.
Note that the shape functions are formulated directly over the entire domain,
though for convenience we shall assemble matrices formulated at an element
level in an actual implementation (see Section 2.5 for an example).
Since the displacement is prescribed on u , the variation v there is zero.
Hence, the variation field, v, expressed using the same shape functions as
those used for the displacement field, is
v = N v.
Substituting the displacement and variation fields in the strain-displacement
relation, we get

c (u) = B u + B u,
c (v) = Bv,
where B is known as the strain-displacement matrix. B is the strain-displacement
matrix associated with the shape functions N . The stress tensor is obtained
by substituting for c in Eqn. 2.2:

c = CB u C0 + 0 + C B u.
Let Vh be a finite-dimensional subspace of Vu . Substituting the above
quantities in Eqn. 2.3, the finite element problem can be stated as
Find uh Vh such that
Z  hZ Z
t t t
v B CB d u = v N t d + (N t b + B t C0c B t 0c ) d
t
t
Z i
t
B C B d u v. (2.4)

Since the above equation is valid for all vectors v, we can write the finite
element formulation as
Find u which satisfies
K u = f , (2.5)
where
Z
K= B t CB d,

37
 Z Z Z 
f= t
N t d + t
(N b + B t
C0c B t
0c ) d t
B C B d u.
t

Note that the stiffness matrix, K, is symmetric and positive-definite since

the elasticity matrix, C, is symmetric and positive-definite. Note also that
a nonzero prescribed displacement contributes to the load vector via the last
term in f . The force vector f is known as the work-equivalent load vector,
since it is derived from the virtual work principle. The above way of formu-
lating the load vector is the variationally correct way in which a distributed
load on the actual structure should be converted to point loads at the nodes
in the finite element model. If n is the number of free degrees of freedom
and m is the number of prescribed degrees of freedom in a plane-stress/plane-
strain problem, then the order of the various matrices is as follows: K : nn,
u : n 1, f : n 1, u : m 1, B : 3 m, B : 3 n, C : 3 3, N : 2 n.
It is also possible to derive Eqn. 2.5 using the potential energy functional,
. First we write the expression for given by Eqn. 1.36 in terms of c . We
get
1
Z Z Z
t t
= [c (u)] Cc (u) u t d ut b d+
2 t
Z
[c (u)]t 0c [c (u)]t C0c d. (2.6)
 

Substituting for u in Eqn. 2.6, we get

1
= ut K u ut f . (2.7)
2
We carry out the minimization of with respect to u using indicial notation.
We have
1
= ui Kij uj ui fi ,
2
so that
1 1
= im Kij uj + ui Kij jm im fi
um 2 2
1 1
= Kmj uj + Kim ui fm
2 2
= Kim ui fm . (since K is symmetric)
For minimizing , we set / um = 0 which yields Eqn. 2.5. Note that
by using the variational formulation, we obtained Eqn. 2.5 directly without
having to carry out any differentiation.

2.4 Finite Element Formulation for Heat-Transfer

Problems
The procedure for deriving the finite element formulation for the heat trans-
fer problem is analogous to that of elasticity. Note that for the purposes of

38
deriving the finite element formulation, we can directly make use of the vari-
ational statement given by Eqn. 1.37 since the thermal conductivity tensor
is a second-order tensor (as opposed to its counterpart, the elasticity tensor,
which is a fourth-order tensor).
The temperature T written in terms of the interpolation functions is

T = N w + N w,

where w is the vector of the unknown temperatures at the nodes, N are the
shape functions associated with the nodes where the temperature is unknown,
is the vector of the prescribed temperatures at the nodes lying on T , and
w
N are the corresponding shape functions. Noting that the variation, v, of
the temperature vanishes on T , we have

v = N v.

The gradient of T is given by

T = B w + B w,

where
Nj
Bij =
xi
Nj
Bij = .
xi
Substituting for these quantities in Eqn. 1.37, and using the arbitrariness of
v, we get
Find w such that
K w = f ,
where
Z Z
t
K= B kB d + hN t N d,
q
Z Z Z 
f= t
N Q d + t
N (q + hT ) d t
B kB d w.
q

Note that K is symmetric and positive definite by virtue of the symmetry and
positive-definiteness of k. Also note that the second term in the stiffness
matrix, K, involves an integral over part of the surface of the domain. There
is no corresponding term in the expression for the stiffness matrix in the
elasticity problem. If n is the number of free degrees of freedom and m
the number of prescribed degrees of freedom, then the order of the various
matrices in a three-dimensional problem is as follows: K : n n, w : n 1,
: m 1, N : 1 m.
f : n 1, N : 1 n, B : 3 n, B : 3 m, w

39
 t
1 2 P 3 4

L L L
Figure 2.1: Bar subjected to uniaxial loading.

L 2L 3L
Figure 2.2: Example of a piecewise linear function.

2.5 A Model Problem

To illustrate the direct formulation of the finite element equations over the
entire domain, we consider the simple one-dimensional problem shown in
Fig. 2.1. A bar of length 3L, with Youngs modulus E and cross-sectional
area A is subjected to a uniform surface force per unit length, t, a uniform
body force, b, along its length, and a point force P at x = L/3. Assuming
that E and A are constant, the governing differential equation is
d2 u
E 2 + b = 0.
dx
The stress-strain and strain-displacement relations are = E and =
du/dx.
We discretize the domain into 3 elements, each of length L, as shown in
Fig. 2.1. We are interested in finding the approximate solution uh in the
space of piecewise linear functions, i.e.,

Vh Space of piecewise linear functions.

An example of a function in Vh is as shown in Fig. 2.2. Note that we need

to satisfy Vh H 1 () or Vh H 2 () corresponding to second-order or
fourth-order boundary value problems. When Vh is the space of piecewise
polynomials then we have

Vh H 1 () Vh C 0 (),

40
 u
Vh

uh
Figure 2.3: Finite element solution as a projection of the exact solution on
Vh .

x j-1 xj x j+1

Figure 2.4: A typical shape function.

Vh H 2 () Vh C 1 (),
where
C 0 () = {v : v is a continuous function on }
 
1 0 v 0
C () = v C () : C () .
xi
The solution, uh , that we seek is the projection of u on Vh as shown in
Fig. 2.3.
To describe any function in Vh , we introduce the basis functions
(
0 if i 6= j
Nj (xi ) =
1 if i = j
A typical basis function (or shape function) is shown in Fig. 2.4. Since there
are 4 nodes in our finite element model, the basis functions associated with
the nodes, {Ni }4i=1 , form a basis of Vh for this particular problem. Hence,
any function uh Vh can be represented as
4
X
uh = Ni ui ,
i=1

where ui = uh (xi ) is the value of the displacement at node xi . The variational

problem is (see Eqn. 1.31)
Find uh which satisfies
Z 3L Z 3L Z 3L
(v) A dx = tv dx + bvA dx + P v|x=L v.
0 0 0

41
Substituting for and , we get the problem statement as
Find uh which satisfies
Z 3L Z 3L Z 3L
dv duh
EA dx = tv dx + bvA dx + P v|x=L v.
0 dx dx 0 0

Note that we need to satisfy

Pthe boundary conditions, u(0) = 0 and v(0) = 0.
4 P4
Hence, substituting uh = i=2 Ni ui and v = i=2 Ni vi , the above problem
statement becomes
Find u such that
Z 3L
dN t dN
 Z 3L Z 3L
t t t t
v EA dx u = v N t dx+ v N t (bA) dx+P v2 v,
0 dx dx 0 0

where
N u v
2 2 2
N = N3 , u = u3 , v = v3 .


N4 u4 v4
Using the fact that v2 , v3 and v4 are arbitrary, we get the usual finite element
equation
K u = f , (2.8)
where
3L 3L 3L
Z
dN2 dN2 dN2 dN3 dN2 dN4
Z Z
dx dx dx

0 dx dx 0 dx dx 0 dx dx
Z 3L 3L 3L

dN3 dN2 dN3 dN3 dN3 dN4
Z Z
K = EA dx dx ,
dx

0 dx dx 0 dx dx 0 dx dx
Z 3L 3L 3L

dN4 dN2 dN4 dN3 dN4 dN4
Z Z
dx dx dx
0 dx dx 0 dx dx 0 dx dx

Z 3L N2 Z 3L N2 P

f= N3 t dx + N3 bA dx + 0 .

0 0
N4 N4 0

Note that
dN2 1 dN3 1 dN4
= 0 x L; = L x 2L; = 0 0 x 2L;
dx L dx L dx
1 1 1
= L x 2L; = 2L x 3L; = 2L x 3L;
L L L
= 0 elsewhere; = 0 elsewhere.

42
Assuming that the tractions and body force, t and b, are constant, the stiff-
ness matrix and force vector are given by

2 1 0
EA
K= 1 2 1

L
0 1 1
(2.9)
2 2 P
tL
bAL

f= 2 + 2 + 0
2 2
1 1 0

Note that the stiffness matrix is banded since any shape function interacts
only with its neighboring shape functions. Thus, for example N2 and N3 are
both nonzero on L x 2L, but N2 and N4 are not both nonzero on any
part of the domain. This leads to the K(1, 2) term being nonzero, but the
K(1, 3) term being zero.
Once Eqn. 2.8 has been solved for u, the strains can be recovered from
the displacement field by using

u
1
u
 
duh dN1 dN2 dN3 dN4 2
= =
dx dx dx dx dx u3



u4
= B u,
where B is the strain-displacement matrix. We get
1
= (u2 u1 ); 0 x L,
L
1
= (u3 u2 ); L x 2L,
L
1
= (u4 u3 ); 2L x 3L.
L
Note that as a result of assuming a piecewise-linear displacement field the
strain field is constant over each element. Thus, in general, it can be dis-
continuous across element boundaries as shown in Fig. 2.5. If we had as-
sumed a piecewise-quadratic displacement field then we would have obtained
a piecewise-linear (again, not necessarily continuous) strain field. The stress
is simply obtained by using = E. Since the strain is just multiplied by
a constant, the qualitative behavior of the stress field is the same as that of
the strain field.
We would like to caution the reader though, that the procedure of directly
working on the entire domain as just shown is not followed while implement-
ing a finite element program. In an actual implementation, the stiffness and

43


x
0 L 2L 3L

Figure 2.5: The strain field is, in general, discontinuous across element
boundaries.

force matrices are formulated at an element level and then assembled to form
the global stiffness matrix and force vector as shown in the following chap-
ter. The intention behind carrying out the computations at the global level
was to provide insights into issues which get obscured otherwise (e.g., the
banded nature of the stiffness matrix), and also to provide a justification for
the element-level formulation which is followed in practice.

44
Chapter 3

Implementation of the Finite

Element Method

We have seen in the previous chapter that the procedure of formulating the
stiffness and load vectors can be quite cumbersome. It is more convenient to
form the element stiffness and load vectors, and assemble them to form the
global stiffness and load vectors.

3.1 Formulation of the Stiffness Matrix and

the Load Vector from Element-Level Ma-
trices
Referring to Fig. 2.4, we see that the restriction of the shape function to an
element is as shown in Fig. 3.1. A shape function associated with a particular
node has a value of one at that node and a value of zero at other nodes in that
element. Though it is trivial to satisfy continuity of the independent field
across an element boundary in a one-dimensional problem, it is quite difficult
to achieve this in two or three-dimensional problems when the elements are
not rectangular or parallelepipeds (try it!). To achieve this objective, it is
helpful to introduce a natural or intrinsic coordinate system. We shall first
carry out this exercise for one-dimensional elements, and later extend it to
higher dimensions.
As shown in Fig. 3.2, the element with endpoint coordinates given by

N1 N2

1 e 2

Figure 3.1: Restriction of a shape function to an element.

45
 x1 x2 =1 =+1

Figure 3.2: Mapping of an element to a master element.

u=N1 u1 + N 2 u 2

u2
u1

1 2

Figure 3.3: Displacement interpolation

(x1 , x2 ) is mapped to a natural coordinate system with endpoint coordinates

given by (1, 1). The shape functions on the element shown in Fig. 3.1
are now formulated in terms of . Let us consider the formulation of N1 .
Since, N1 has to satisfy two boundary conditions, namely N1 (1) = 1 and
N1 (1) = 0 at the end nodes, we assume it to be of the form

N1 () = a + b,

where a and b are constants which are determined from the boundary con-
ditions. We get
1
N1 () = (1 ).
2
Following a similar procedure, we get
1
N2 () = (1 + ).
2
The coordinate x of a point in the element is given by

x = N1 ()x1 + N2 ()x2 ,

where is the natural coordinate of that point. If we interpolate the dis-

placements also in the same way, we get

u = N1 ()u1 + N2 ()u2.

If the same shape functions are used to interpolate both the displacements
and the geometry as above, the mapping is known as an isoparametric map-
ping. Note that u( = 1) = u1 , u( = +1) = u2 with a linear interpolation
in between as shown in Fig. 3.3.

46
 The strain is given by
du du d
= =
dx d dx
1
= (u2 u1 )
x2 x1
1
= (u2 u1 ),
le
where le = (x2 x1 ) is the length of the element. The above equation can
be written as
= B u,
where
1
B= [1 1]
le

u1
u = .
u2

The stress is given by

= E = EB u.
Note that the stresses and strains are constant within an element.
The element stiffness matrix and load vector can be derived using either
the potential energy functional or the virtual work principle. We shall follow
the latter approach. We know from Eqn. 2.3 that the virtual work principle
involves finding u Lu such that
Z Z Z X
t t
[c (v)] c d = v t d + v t b d + Pi vi v Vu .
t

where, now, we have added the contribution of the point loads Pi explicitly
instead of considering it as part of the prescribed tractions t. Writing the
above equation as a summation over the number of elements, we get
XZ X Z Z  X
t
[c (v)] c d = v b d + v t d + Pi vi v Vu ,
e e e e e

Note that while computing the surface integral term in the above equation,
we need to consider only the those elements which have a surface on t . The
reason is that the net contribution of the internal tractions at any shared
edge between two elements to the global load vector is zero since the traction
vector on the edge of one element is equal and opposite to the traction vector
on the adjacent element.
For the one-dimensional problem that we are considering, we have v =
N v, (v) = B v, = EB u, d = Adx = Ale d/2 and d = dx = le d/2.
Substituting these various quantities in the above equation, we get
XZ 1 le X Z 1 Z 1 
t le le
t t
v B EAB d u = t
v N t d + v N bA d +v t {Ni (j )Pj },
t t

e 1 2 e 1 2 1 2

47
where in the last term j is the position at which load Pj is applied.
From the above equation, the element stiffness matrix is given by

1 1 t
Z
(e)
K = B EABle d.
2 1

Assuming that EA is constant over the element, we get

Ee Ae 1 1
K (e) =
le 1 1

The element load vectors due to the traction, body forces and point loads
are
R
Z 1 le 1
(e) 1 N 1 t d
ft = N t tle d = 2 R1 ,
2 1 le 1
N t d
2 1 2
R
Z 1 le 1
(e) 1 N 1 bA d
fb = N t Able d = 2 R1 ,
2 1 le 1
N bA d
2 1 2
P
(e) i N1 (i )Pi
f P = P .
i N2 (i )Pi

If t and b are constant over the element then we get

(e) 1
tle (e) Able 1
ft = ; fb = .
2 1 2 1

On the other hand if the body force varies as b = b0 x = b0 (N1 x1 + N2 x2 ),

then
R
1 R1
(e) b0 le A 1 N1 N1 d 1 N1 N2 d x1
fb =
2 R1 R1
N N d 1 N2 N2 d x2
1 2 1

2 1
b0 le A 3 3 x1
=
2 1 2
x2
3 3

2x
b0 le A 1 + x2
= .
6 x1 + 2x2

48
3.2 Assembly of the Element Stiffness and
Load Vectors into the Global Stiffness and
Load Vectors
Consider the contribution of the second element to the term v t K u in Eqn. 2.4.
We have

2 2 1 1 u2
h iE A
(v t K u)2 = v2 v3
l2 1 1 u3

0 0 0 0 u1

h i 0 E 2 A2 E 2 A2 0

u2

l l
= v1 v2 v3 v4 2 2 .
0 E2l A2 E2l A2 0 u3

2 2
0 0 0 0 u4

Thus, we see that the elements of K (2) occupy the second and third rows
and columns of the global stiffness matrix K. Since two and three are the
global degrees of freedom associated with element 2, we conclude that the
element stiffness matrix should be assembled based on the degrees of freedom
associated with the element. Overlapping elements are added. Similar to the
process for the stiffness matrix, we assemble the global load vector using the
individual element load vectors. For the model 1-d problem considered in
Section 2.5, we get

1 1 0 0

1 2 1 0

EA
K=
L 0 1 2 1


0 0 1 1
(3.1)
1 1 R

2 bAL 2 P

tL
f=
+ 2 + ,

2 2 2 0

1 1 0

where R is the reaction at the wall. On incorporating the boundary con-

ditions into the above matrices as shown in the next section, the above
matrices will reduce to those in Eqn. 2.9. Notice that this procedure of
assembling element-level matrices is considerably simpler compared to the
direct method that we followed in Section 2.5.
The stiffness matrix has the following properties:

49
 1. The dimension of the stiffness matrix is n n where n is the number
of free degrees of freedom.

2. K is symmetric.

3. K is banded since basis functions have compact local support.

4. Using the symmetry and bandedness properties, K is stored in compact

form. For example, one of the ways that the stiffness matrix in Eqn. 3.1
is stored is
1 1

2 1

EA
Kbanded =
le
2 1


1 0
Thus, the storage space is reduced from n n to n nbw , where nbw is
the half-bandwidth of K, and is given by

nbw = Nmax nndf + nndf , (3.2)

where Nmax is the maximum difference between node numbers connect-

ing an element and nndf is the number of degrees of freedom per node. In
our model 1-d problem, nndf = 1, and the maximum difference between
node numbers connecting an element is 1. Hence nbw = 2. Arbitrary
numbering of nodes can result in higher bandwidth and hence increased
storage. For example, if in our model problem, we had numbered the
nodes as (1, 4, 3, 2), then

nbw = max(4 1, 4 3, 3 2) + 1 = 4,

whereas we have seen that with sequential node-numbering we have

seen that nbw = 2.

3.3 Treatment of Boundary Conditions

In the previous section, we have seen how to assemble the element level ma-
trices to form the global stiffness matrix. However, the stiffness matrix thus
obtained is singular since rigid body modes have not been eliminated. We
need to incorporate the boundary conditions which prevent rigid body mo-
tion. We discuss two methods of incorporating the boundary conditions, the
elimination approach and the penalty approach. The elimination approach
is more exact, but is more difficult to implement than the penalty approach.

3.3.1 The elimination approach

Suppose that we have obtained the matrix equations K u = f , and suppose
that it is given that u1 = a1 , where a1 is a specified value. Also suppose that

50
the external force applied and the reaction corresponding to this degree of
freedom are F1 and R1 , respectively. Then we have

K K12 . . . K1n a F + R1
11 1 1
K21 K22 . . . K2n u2 F2

= .
. . . . . . . . . . . . . . . . . . . . . .



Kn1 Kn2 . . . Knn un Fn

Writing out the first equation:

R1 = K11 a1 + K12 u2 + + K1n un F1 (3.3)

Writing out the second equation, we get

K21 a1 + K22 u2 + + K2n un = F2 ,

or, alternatively,

K22 u2 + K23 u3 + + K2n un = F2 K21 a1 ,

Similarly we can write the third equation as

K32 u2 + K33 u3 + + K3n un = F3 K31 a1 ,

and so on. In matrix form, all the equations excluding the first can be written
as

K22 K23 . . . K2n u2 F2 K21 a1

K32 K33 . . . K3n u3 F3 K31 a1

=
.
. . . . . . . . . . . . . . . . . . . . . . . ...



Kn2 Kn3 . . . Knn un Fn Kn1 a1
(n1)(n1) (n1)1 (n1)1

Thus, the reduced K matrix is obtained by eliminating the row and column
corresponding to the specified or support degree of freedom. Note that the
force matrix also gets modified when the prescribed quantity is non-zero.
Since the rigid-body modes have been eliminated, the reduced K matrix is
no longer singular. Hence we can solve for the unknown vector u.
In order to compute the strains the element-level quantities are extracted
from the global u vector. We then use = B u(e) . The stress is then obtained
using the appropriate constitutive relation, e.g., in our model problem =
EB u(e) . In order to compute the reactions, we use Eqn. 3.3 since all the
quantities on the right-hand side of that equation are known. Thus, if we
need to compute the reactions, then we need to store the rows of K which
are eliminated for the purposes of getting the reduced stiffness matrix.
The above procedure can be easily generalized for the case when there
are r prescribed degrees of freedom, say up1 = a1 , up2 = a2 , . . ., upr = ar :

51
 C

a1

Figure 3.4: Penalty approach to handle boundary conditions

Store the p1 , p2 , . . ., pr row of the global stiffness matrix and the load
vector f .
Delete the p1 , p2 , . . ., pr th row and column from the K matrix. The
resulting stiffness matrix size is (n r) (n r). Delete the p1 , p2 , . . .,
pr th row from the load vector. Modify the load components as
X
fi fi Ki,pj aj .
j

Solve K u = f .
Extract the element displacement fields u(e) from u, and determine the
strains and stresses.
Evaluate the reactions at each support degree of freedom:
X
Rp1 = Kp1 1 u1 + Kp1 2 u2 + . . . + Kp1 n un fp1 = Kp1 i ui fp1 ,
i
X
Rp2 = Kp2 1 u1 + Kp2 2 u2 + . . . + Kp2 n un fp2 = Kp2 i ui fp2 ,
i

and so on. In general

X
Rpj = Kpj i ui fpj .
i

One good debugging check is to see that the (vectorial) sum of the
reactions is equal to the load applied on the structure.
In the model problem considered in Section 2.5, the reaction at the wall is
given by
1 1 1
R = EAu2 ALb Lt.
L 2 2

3.3.2 The penalty approach

Again consider a boundary condition of the type u1 = a1 . Now we introduce
a spring with large stiffness, C, to model the support as shown in Fig. 3.4.
One end of the spring is displaced by a1 , while the other end is displaced by
u1 . Hence, the net extension of the spring is u1 a1 . Note that if the spring
is infinitely rigid, then the net extension of the spring would be zero, and we
would get u1 = a1 . The strain energy in the spring is
1
Us = C(u1 a1 )2 .
2
52
Hence, the expression for the potential energy gets modified to
1 1
= ut K u + C(u1 a1 )2 ut f .
2 2
To derive the matrix equations, we follow the usual procedure and set the
partial derivatives of with respect to the degrees of freedom to zero, i.e.,

= 0; = 0; . . . ; = 0.
u1 u2 un
We get
K + C K12 . . . K1n u f + Ca1
11 1 1
K21 K22 . . . K2n u2 f2

= . (3.4)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .


Kn1 Kn2 . . . Knn un fn
Thus, a large number gets added to the first diagonal term of K, and Ca1
gets added to f1 . The reaction force is the force exerted by the spring on the
structure. Since the net extension of the spring is u1 a1 , the reaction force
is
R1 = C(u1 a1 ).
Generalizing the above discussion to the case when we have r prescribed
boundary conditions of the form up1 = a1 , up2 = a2 , . . ., upr = ar , the
procedure of treating the boundary conditions is now as follows:
Modify the stiffness matrix K by adding a large number C to each of
the p1 , p2 , . . ., pr th diagonal elements of K. Also modify the global
load vector f by adding Ca1 to fp1 , Ca2 to fp2 , . . ., Car to fpr .
Solve K u = f .
Extract the element displacement field u(e) from the global displace-
ment vector u using the element connectivity, and determine the strains
and stresses.
Evaluate the reaction forces at each support using the equation
Rpi = C(upi ai ) i = 1, 2, . . . , r.

As already mentioned, the penalty approach is an approximate approach,

and the accuracy of the solution depends on the choice of C. We now give a
rough guideline for choosing C. The first equation in Eqn. 3.4 is
(K11 + C)u1 + K12 u2 + . . . + K1n un = F1 + Ca1 ,
which in turn implies that
 
K11 K12 K1n F1
+ 1 u1 + u2 + . . . + un = + a1 .
C C C C
Thus, if C is large compared to K11 , K12 ,. . ., K1n , then u1 a1 . Hence,
choose C as
C = max |Kij | 104 1 i, j n.

53
3.3.3 Multi-point constraints
Multi-point constraints are constraints involving several displacement degrees
of freedom, as, for example, occurs when a node is constrained to move along
an inclined surface. We consider only linear multi-point constraints in what
follows. Let the constraints be of the form

EC u = Ea, (3.5)

where C is an mn matrix corresponding to m constraint equations, u is the

n1 vector containing the displacement degrees of freedom, and the factor E
(Young modulus) has been introduced to alleviate subsequent ill-conditioning
that can occur. We use a Lagrange muliplier technique to impose these
constraints, and modify the potential energy expression in Eqn. (2.7) to read
1
= u K u u f E [C u a] ,
2
where is a m 1 vector of the Lagrange multipliers corresponding to the
m constraints in Eqn. (3.5). The corresponding variational formulation gets
modified to
Find (u, ) such that
h i
v K u f EC T ] E [C u a = 0 (v, ),

First setting to zero, and using the arbitrariness of v yields

K u EC T = f . (3.6)

Next using the arbitrariness of , we recover the constraint equations given

by Eqn. (3.5). The system of equations given by Eqns. (3.5) and (3.6) can
be written as
K EC T u f
= . (3.7)
EC 0 Ea
The above set of equations is to be solved for (u, ). The Lagrange multi-
pliers can be given the physical interpretation of reaction forces.

3.3.4 Example
To illustrate the elimination and penalty approaches, consider the example
shown in Fig. 3.5. The Youngs modulus of the two bars are given by E1 =
70 109 N/m2 and E2 = 200 109 N/m2 . The cross-sectional areas are
A1 = 2400 mm2 and A2 = 600 mm2 . The applied load is P = 200 103 N.
The element stiffness matrices are (with units N/mm)

3 1 1
70 10 2400
K (1) =
300 1 1

54
 300 mm 400 mm

1 2 P 3

E1 E2

Figure 3.5: Analysis of two coaxial bars

3 1 1
200 10 600
K (2) =
400 1 1

The global stiffness matrix and load vector are

0.56 0.56 0 R1

K = 106 0.56 0.86 0.3 ; f = 200 103 .


0 0.3 0.3 R2

Using the elimination approach, we the 1st and 3rd rows and columns from
K. We get
106 (0.86)u2 = 200 103 ,
which yields u2 = 0.23255 mm. The element stresses are

E1 h i u1
1 = E1 B 1 u(1) = 1 1 = 54.77 MPa,
(x2 x1 ) u2

E1 h i u2
2 = E2 B 2 u(2) = 1 1 = 116.29 MPa.
(x2 x1 ) u3

Note the discontinuity in the stresses. The reactions are

R1 = 106 [0.56(0) 0.56(0.23255) + 0] = 130.23 103 N,

R2 = 106 [0 0.3(0.23255) + 0] = 69.77 103 N.

Note that the sum of the reactions is equal to the applied load.
The results for (u1 , u2 , u3, R1 , R2 ) could directly have been obtained using
Eqns. (3.7), with (R1 , R2 ) playing the role of the Lagrange multipliers.
Now let us carry out the same computations using the penalty approach.
Using the suggested guidelines, we have C = 0.8 1010 . The stiffness matrix
is modified by adding C to the appropriate diagonal terms. The matrix
equations are given by

8600.56 0.56 0 u 0
1
106 0.56 0.86 0.3 u2 = 200 103 .


0 0.3 8600.3 u3 0

55
 N1 N2 N3

=1 e =1 =1 e =1 =1 =1
e

Figure 3.6: Quadratic shape functions: Global and element level pictures

On solving these equations, we get

h i h i
6 6
u1 u2 u3 = 15.14 10 0.23257 8.1127 10 mm.

Notice that u1 , u3 0. The stresses are

6
70 10 3 h i 15.14 10
1 = E1 B 1 u(1) = 1 1 = 54.27 MPa,
300 0.23257

3 h 0.23257
200 10 i
2 = E2 B 2 u(2) = 1 1 = 116.27 MPa.
400 8.1127 10 6

The reactions are

R1 = C u1 = [0.86 101 0] (15.143 106 ) = 130.23 103 N

R2 = C u3 = [0.86 101 0] (8.1127 106 ) = 69.77 103 N.

3.4 Higher-order shape functions

The accuracy of a finite element analysis can be increased by reducing the
typical length of an element h or by increasing the order of the shape func-
tions p (or a combination of the two). As an example, consider the quadratic
shape functions
1
N1 () = (1 ),
2
N2 () = (1 + )(1 ), (3.8)
1
N3 () = (1 + ).
2
The global and element level pictures of the shape function are shown in
Fig. 3.6. In order to obtain the expressions given in Eqn. 3.8, consider the

56
expression for N1 . Since N1 is zero at = 0 and = 1, we assume N1 =
c(1 ). The constant c is determined by using the condition N1 |=1 = 1.
A similar procedure is followed for determining N2 and N3 .
Note that the shape functions formulated above give C 0 continuity for the
function being interpolated, but not, in general C 1 continuity. Such shape
functions are called Lagrange shape functions1 . In general, a function of
degree (n 1) and defined by n values i at corresponding i has the form

= N1 1 + N2 2 + . . . + Nn n ,

where
(2 )(3 )(4 ) . . . (n )
N1 = ,
(2 1 )(3 1 )(4 1 ) . . . (n 1 )
(1 )(3 )(4 ) . . . (n )
N2 = ,
(1 2 )(3 2 )(4 2 ) . . . (n 2 )
...
(1 )(2 )(3 ) . . . (n1 )
Nn = .
(1 n )(2 n )(3 n ) . . . (n1 n )

For n = 3, we can easily verify that we get the expressions in Eqn. 3.8.
For an isoparametric element, we have

u = N1 u1 + N2 u2 + N3 u3 ,
x = N1 x1 + N2 x2 + N3 x3 ,

or, in matrix form

u = N u
x = N x.

Thus, both u and x are piecewise quadratic polynomials. In what follows, we

shall assume that the midnode is located at the center of the element, i.e.,
2x2 = x1 + x3 . Thus, we have
dx 1
= (x3 x1 ) + (x1 + x3 2x2 )
d 2
le
= ,
2
where le is the length of the element. The strain is
du
=
dx
du d
=
d dx
1
For providing C 1 continuity, we use Hermitian shape functions which we shall discuss
later.

57
 u

x x

Figure 3.7: Variation of the displacement and stress field

u
2 h i 1
= dN1 dN2 dN3
u2
(x3 x1 ) d d d
u3
= B u(e) ,
where
2 h 12 1+2
i
B= 2 2 .
le 2

The stress is given by

= EB u(e) .
The qualitative variation of the displacement and stress is shown in Fig. 3.7.
Note that a quadratic variation of the displacement implies that the stresses
and strains can vary linearly within the element though they can still be
discontinuous across element boundaries as shown.
Assuming that E and A are constant, the element stiffness matrix is given
by
Z
(e)
K = B t CB d
Zxe2
= B t EBA dx
x1
EAle 1 t
Z
= B B d
2 1

7 8 1
EA
= 8 16 8 . (3.9)

3le
1 8 7

Note that the size of the element stiffness matrix is ndof ndof , where ndof is
the number of element degrees of freedom.
The element force vector due to a constant body force b is
Z x2
(e)
f = N t bA dx
x1
Z 1
le
= N t bA d
1 2

58
 5

cm
0
3

10
x 2

cm
1

50
Figure 3.8: Analysis of a rotating bar

1
bAle
Z
= N t d
2
1
1
bAle
= 4 .

6
1

Similarly, the element-force vector due to a constant force per unit length t
is
1
(e) l t
e
ft = 4 .
6
1

3.5 Examples
Example 1:
A rod of length 150 cm rotates about a point with a constant angular velocity
as shown in Fig. 3.8. The density, Youngs modulus, cross sectional area, and
angular velocity are given by = 7500 kg/m3 , E = 60 GPa, A = 4104 m2 ,
and = 20c /s. Assume u = u(x)ex , and the loading to be only due to the
centrifugal body force (given by b = x 2 ). Formulate the element stiffness
matrix, and the consistent load vector for a quadratic (3-node) element. Then
using the two-element (5-node) model shown in the figure, determine the
displacements and the stress distribution in the rod using two quadratic
elements. Plot the stress distribution as a function of x. What do you
observe about the stress distribution at node 3?
Solution:

59
The body force is given by

x
h i 1
b = x 2 = 7500(400) N1 N2 N3 x2 .


x3

Hence, the element force vector is

Z
(e)
f b = N t bA dx

Z 1 N1 h il
x
1
e
= 1.2 103 N2 N1 N2 N3 d x2

1 2
N3 x3

2
Z 1 N1 N1 N2 N1 N3
1
x
= 0.6 103 le N1 N2 N22 N2 N3 d x2

1
2
N1 N3 N2 N3 N3 x3

4 2 1 x
1
= 40le 2 16 2 x2


1 2 4 x3

4x1 + 2x2 x3

= 40le 2x1 + 16x2 + 2x3


x1 + 2x2 + 4x3

For element 1, we have x1 = 0, x2 = 0.25, x3 = 0.5, le = 0.5, while for

element 2, we have x1 = 0.5, x2 = 1.0, x3 = 1.4, le = 1.0, which yields

0 100

(1) (2)
fb = 100 ; fb = 800 .


50 300

The global force vector obtained after assembling the two load vectors is

0

100


f = 150 .


800


300

60
As a check, we verify that the total load obtained by summing the entries
in the load vector (1350 N) is the same as obtained from the continuum
problem, i.e., Z 1.5
1.2 103 x dx = 1350 N.
0
The element stiffness matrix is given by Eqn. 3.9. For the two elements,
we have

14 16 2 7 8 1

(1) 7 (2) 7
K = 0.8 10 16 32 16 ; K = 0.8 10 8 16 8 .


2 16 14 1 8 7

Assembling the two stiffness matrices, we get

14 16 2 0 0

16 32 16 0 0


7
K = 0.8 10 2 16 21 8 1


0 0 8 16 8


0 0 1 8 7

Incorporating the boundary condition u1 = 0, we get

32 16 0 0 u 100
2
16 21 8 1 u3 150

= 1.25 107
0 8 16 8 u4 800


0 1 8 7 u5 300

Solving, we get u1 = 0, u2 = 1.39323 105 m, u3 = 2.7083 105 m,

u4 = 4.791667 105 m, u5 = 5.625 105 m. The reaction is
R1 = 0.8 107 [16u2 + 2u3 ] = 1350 N.
The stresses in the two elements are
(1) = Ebu(1)

u
2E h 12 i 1
= 2 2 12 u2

l1 2
u3

u
2E h 12 i 3
(2) = 2 2 12 u4 .

l2 2
u5

61


3437500 3375000
3281250
3250000
3250000 3000000

3062500 1750000
1875000
250000

x x
(a) (b)

Figure 3.9: Stress field in rotating bar: (a) Finite element & (b) Exact
solutions

The exact solution is

2 x3
 
2
u= L x ,
2E 3
1
= 2 (L2 x2 ).
2
Note that the FEM displacement solution is exact at the nodes! The variation
of the stress field as per the finite element solution and the exact solution
are shown in Fig. 3.9.
Example 2:
Problems with thermal stresses can be treated as initial strain problems. For
example, for a one-dimensional problem, we can take 0 = T . In general,
the load vector is given by
Z
Fth = B tu C0c d.

For a one-dimensional linear bar element, with constant E, T and A, and

with B = l1e [1 1], we have

Z 1
1 1
Fth = ET Ale d
2le 1 1

1
= EAT . (3.10)
1

We shall use this load vector to solve the statically indeterminate problem
shown in Fig. 3.10. The material and geometric properties are E1 = 70 GPa,
A1 = 900 mm2 , 1 = 23 106 /o C, E2 = 200 GPa, A2 = 1200 mm2 , 2 =
11.7106 /o C. The point load and change in temperature are P = 3105 N
and T = 40o C.
The element stiffness matrices are

1 1 3 1 1
E1 A1 = 70 10 900
K (1) = ,
l1 1 1 200 1 1

62
 T = 40 o C

E1, A1, 1 P E 2 , A2 , 2

200 mm 300 mm

Figure 3.10: Stepped bar subjected to thermal loading

E2 A2 1 1 3 1 1
K (2) = = 200 10 1200 .
l2 1 1 300 1 1

The global stiffness matrix after assembling the above matrices is

315 315 0

K = 103 315 1115 800 N/mm.


0 800 800

The global load vector due to the thermal loads (obtained using Eqn. 3.10)
and the point load P is given by

57.96 + R1

3
f = 10 245.64 .


112.32 + R2

Incorporating the boundary conditions u1 = u3 = 0, we get

(1115 103 )u2 = 245.64 103 ,

which on solving yields u2 = 0.22 mm. The global displacement vector is,
thus, [0 0.22 0]t mm. The stress in the bars is

1 = E1 B 1 u(1) E1 1 T = 12.60 MPa,

2 = E2 B 2 u(2) E2 2 T = 240.27 MPa.

Note from the above example that statically indeterminate problems can be
handled in a routine way by the finite element method.

3.6 Truss structures

Since elements of a truss have various orientations, we introduce local and
global degrees of freedom for each member as shown. The local coordinate
system is along the length of the member, while the global one is along the

63
 u2
6 10
u4
5 9
(x 2, y2 )
u3
8 u1 le
2 4 12
u2
1 3 7
11

(x 1, y1 )
u1

Figure 3.11: Global and local degrees of freedom for a truss member.

global coordinate axes. Thus, for example, each member of a two-dimensional

truss of length le , and oriented at an angle has the global degrees of freedom
u1 , u2 , u3 and u4 , and the local degrees of freedom u1 and u2 as shown in
Fig. 3.11. Note that

u1 = u1l + u2 m,
u2 = u3l + u4 m,

where l = cos and m = sin are the direction cosines which are given by
x2 x1 y2 y1 p
l= ; m= ; le = (x1 x2 )2 + (y1 y2 )2 .
le le
In matrix form, we have
u
1
u u

1 = Q 2 , (3.11)
u2 u3


u4
where
l m 0 0
Q= .
0 0 l m
In the local coordinate system, the element stiffness is given by

Ee Ae 1 1
K = .
le 1 1

The element stiffness matrix in global coordinates is derived using the fact
that by virtue of the tensorial nature of the quantities involved, the virtual
work contribution is the same irrespective of the coordinate system that we
work in. Thus,
(v )t K u = v t (Qt K Q)u,
leading to

K (e) = Qt K Q

64

l2 lm l2 lm

2 2
lm m lm m

Ee Ae
= .
le l 2
lm l 2
lm


2 2
lm m lm m

One gets the same result by equating the strain energy in the local and global
coordinate systems.
The stress in a member is given by

Ee h i u1
= 1 1 Ee e T
le u2

u
1
u2
i
Ee h
= l m l m
Ee e T.
le u3

u4

The thermal load vector is obtained by equating the virtual work expres-
sions in the local and global coordinate systems. We get
Z 1
t t t le
(v ) f th = (v) Q B t Ee e T Ae d,
1 2

leading us to the expression

l

m

t t

f th = Ee e T Ae le Q B = Ee e T Ae
.
l


m

It is easy to extend the above discussion for two-dimensional trusses to

three-dimensional ones. The transformation matrix between the local and
global coordinate systems is now given by

l m n 0 0 0
Q= ,
0 0 0 l m n

where the direction cosines, l, m, and n, are

x2 x1 y2 y1 z2 z1
l= ; m= ; n= .
le le le

65
3.7 One-Dimensional Heat Transfer Problems
Since the procedure for formulating the stiffness and load vectors is exactly
analogous to the elasticity problem, we shall illustrate it directly by means
of an example.
Problem: Find the temperature distribution in a bar of length L subject
to the following boundary conditions:
1. T (0) = T1 ; T (L) = T2 .

2. k dT = q; k dT
 

dx x=0 dx
+ h(T T ) x=L = 0.

3. T (0) = T1 ; k dT 
dx x=L
= q.
Solution: If we assume linear elements, then
T = N1 w1 + N2 w2 ,
where N1 = (1 )/2 and N2 = (1 + )/2. Thus,

dT h i w 1 w
= dN 1 dN2 = B 1
dx dx dx
w2 w2
1. For the boundary conditions given, we see that = T . Thus,

Z 1
l kA 1 1
K (e) = kB t B A d = ,
1 2 l 1 1

where l denotes the length of the element. If we divide the domain into
3 elements of equal length (l = L/3), then the finite element equations
prior to incorporating the boundary conditions are

1 1 0 0 T1 1 q
1
1 2 1 0 w2 QAl 2 0

kA
= + A ,
l 0 1 2 1 w3
2 2

0


0 0 1 1 T2 1 q4
where q1 and q4 are the unknown heat fluxes at the two ends. The
equations for w2 and w3 can be written as

kA 2 1 w2 1 kA T1
= QAl + ,
l 1 2 w3 1 l T2

while those for q1 and q4 can be written as

T
1
q 1 1 0 0 w2 Ql 1

1 = k .
q4 l 0 0 1 1 w3
2 1

T2

66
 Solving for w2 and w3 , we get
QL2 1
w2 = + (2T1 + T2 ),
9k 3
QL2 1
w3 = + (T1 + 2T2 ),
9k 3
which on substituting in the equations for q1 and q4 yields
1 k
q1 = QL + (T1 T2 ),
2 3
1 k
q4 = QL + (T2 T1 ).
2 3
The exact solution is
QL2 x  x 2
 
1
T (x) = + (T2 T1 )x + T1 ,
2k L L L
 
dT QL 2x k
q(x) = k = 1 (T2 T1 ).
dx 2 L L
Note that the solution for the temperature field is exact at the nodes.
However, since we have assumed a linear interpolation for the temper-
ature, while the real variation is quadratic, there is an error in the
finite element solution at points other than the nodes. Since the exact
solution is quadratic, we can get the exact solution simply using one
quadratic element. We now have

7 8 1 T 1 q
kA 1 AQl 1
8 16 8 w2 = 4 + 0 ,

3l 6
1 8 7 T2 1 q3
where l = L. The solution is
1 QL2 1
w2 = + (T1 + T2 ).
8 k 2
The heat flux at the ends is
1 k
q1 = QL + (T1 T2 ),
2 L
1 k
q3 = QL + (T2 T1 ).
2 L
2. The two ends of the bar constitute q . The finite element equations
are

1 1 0 0 w 1 Aq
1
1 2 1 0 w2 AQl 2 0

kA
= +
l
0 1 2 1 w3
2 2 0


hl
0 0 1 1 + k w4 1 AhT

67
 On solving, we get
QL + q L(QL + 2q)
w1 = + + T ,
h 2k
QL + q 2L(2QL + 3q)
w2 = + + T ,
h 9k
QL + q L(5QL + 6q)
w3 = + + T ,
h 18k
QL + q
w4 = + T .
h
The exact solution is
QL2 x2
   
2k qL k x
T (x) = 1+ + 1+ + T .
2k hL L2 k hL L
Note again that the temperature solution is exact at the nodes.
3. The right end of the bar constitutes q . The finite element equations
are
1 1 0 0 T 1 Aq
1 1
1 2 1 0 w2 AQl 2 0

kA = + ,
l 0 1 2 1 w3
2 2 0


0 0 1 1 w4 1 Aq
which on incorporating the boundary conditions becomes

2 1 0 w2 2 0 T1
kA

AQl kA

1 2 1 w3 = 2 + 0 + 0 .
l 2 l
0 1 1 w4 1 Aq 0
On solving, we get
L(5QL + 6q)
w2 = + T1 ,
18k
2L(2QL + 3q)
w3 = + T1 ,
9k
L(QL + 2q)
w4 = + T1 .
2k
The flux at the left end is
k 1
q1 = (T1 w2 ) Ql = (QL + q).
l 2
The exact solution is
L2 Q 2x x2
   
qx
T (x) = 2 + 2 + T1 ,
k 2 L L L
dT h  x i
q(x) = k = QL 1 + q .
dx L
Note that the temperature solution is exact at the nodes.

68
3.8 Beams and Frames: An Introduction to
C 1 elements
We shall consider only the classical or slender beam theory here, and not the
Timoshenko beam theory which includes the effect of shear deformation. We
restrict our attention to beams with symmetric cross-sections.
Assuming that the x axis is directed along the length, and the y-axis is
oriented downward with y = 0 at the neutral plane, we have
My
xx = = Exx .
I
The remaining stress components are assumed to be zero. The governing
differential equation for the displacement is

d2 v
EI = M(x),
dx2
where M(x) is the bending moment along the beam.
The expression for the potential energy is
1
Z Z Z X
t
= c d t u d b u d Pi u i
2 t
1 L M2
Z Z  Z X X
2
= 2
y dA dx qv dx p i vi Mi vi
2 0 EI A
Z L  2 2 Z L
1 d v X X
= EI dx qv dx P i vi Mi vi .
2 0 dx2 0

The variational formulation is obtained by setting the first variation of to

zero. We get
Z L Z L
d2 v d2 X X
EI 2 2 dx q Pi i Mi i = 0 .
0 dx dx 0

The variational formulation can also be obtained from the governing equa-
tions
dM dV d2 v
= V ; = q; EI 2 = M,
dx dx dx
where V is the shear force. The above equations lead to the governing equa-
tion
d2 d2 v
 
EI 2 q = 0.
dx2 dx
Multiplying this governing equation by a test function , and integrating
by parts leads us to the variational formulation as has already been demon-
strated in Section 1.6.
In order that d2 v/dx2 is well-defined, we need v C 1 , i.e, both v and

v must be continuous. Now we formulate Hermite shape functions which

69
 u1 u3

=1 u =+1 u
2
=0 4

Figure 3.12: Degrees of freedom for a beam.

Slope=0

N1
1 N2

Slope=0 Slope=0
Slope=1

Slope=0

N3 Slope=0 Slope=1
1
Slope=0 N4

Figure 3.13: Hermite shape functions.

ensure that v C 1 . Let u1 , u2 , u3 and u4 be the local degrees of freedom as

shown in Fig. 3.12, where
  
dv  dv 
(u1 , u2 , u3, u4 ) v1 , , v2 , .
dx 1 dx 2

The vertical displacement v is given by

v() = N1 v1 + N2 v1 + N3 v2 + N4 v2 ,

where  
dv  dv 
v1 = ; v2 = .
d =1 d =+1
In order that v(1) = v1 , v (1) = v1 , v(+1) = v2 and v (+1) = v2 , the
shape functions should be as shown in Fig. 3.13. We assume each of the Ni s
to be of the form
Ni () = ai + bi + ci 2 + di 3 ,
and find the constants ai -di based on the boundary conditions that each shape
function has to satisfy. For example, N1 (1) = 1, N1 (1) = 0, N1 (+1) = 0,
and N1 (+1) = 0. We get
1 1
N1 = (1 )2(2 + ) = (2 3 + 3 ),
4 4

70
 1 1
N2 = (1 )2(1 + ) = (1 2 + 3 ),
4 4
1 1
N3 = (1 + )2 (2 ) = (2 + 3 3 ),
4 4
1 1
N4 = (1 + )2 ( 1) = (1 + 2 + 3 ).
4 4
We assume that the coordinate interpolation is still given by
1 1
x = (1 )x1 + (1 + )x2 ,
2 2
leading to dx = le d/2. Using the chain rule

dv dv dx le dv
= = .
d dx d 2 dx

Since (dv/dx)1 = u2 and (dv/dx)2 = u4 , we have

le le
v() = N1 u1 + N2 u2 + N3 u3 + N4 u4.
2 2
In matrix form, the above equation can be written as

v = N u(e) , (3.12)

where
h i
l l
N N1 2 N2 N3 2 N4 ,
e e

h it
u(e) u1 u2 u3 u4 .

Using the chain rule again, we get

d2 v 4 d2 v
= .
dx2 le2 d 2

Using the same interpolation for the variations and substituting in the vari-
ational formulation, we get
Z 1
16EI d2 N t d2 N le
Z 1
t t t le
X X
4 2 d 2 2
d u = N q d+ P i (xi )+ Mi (xi ).
1 l e d 1 2

Since is arbitrary, we get the familiar equation

(e)
K (e) u(e) = f ,

where
1
le
Z
(e)
K = B t (EI)B d
1 2

71
 1
le
Z
(e)
f = N tq d + {Ni (j )Pj } + {Ni (j )Mj },
1 2

with B given by

4 d2 N 4 h3 le
i
B= 2 = 2 2 (3 1) 23 le
(3 + 1) .
le d 2 le 4 4

If EI is assumed to be constant over the element, then we get

12 6le 12 6le

2 2
6le 4le 6le 2le

(e) EI
K = 3
le 12

6le 12 6le


6le 2le2 6le 4le2

Assuming that q is uniform over the element, and that there are no point
loads and moments, the element load vector is

qle
2
qle2
12
f (e)

= .
qle
2


2
ql12e

The element level matrices are assembled into the global ones as usual,
and we solve the system of equations K u = f . The moment and shear force
are found as follows:
d2 v
M = EI
dx2
d2 N
= EI 2 u(e)
dx
4EI d2 N (e)
= 2 u
le d 2
EI
= 2 [6u1 + (3 1)le u2 6u3 + (3 + 1)le u4 ] ,
le
dM
V =
dx
2 dM
=
le d
6EI
= 3 [2u1 + le u2 2u3 + le u4 ] .
le

As an example, consider the beam shown in Fig. 3.14. The Youngs

modulus and moment of inertia are E = 200 GPa and I = 4 106 m4 . The

72
 12 kN/m

1m 1m

Figure 3.14: Beam subjected to distributed load.

element stiffness matrices are

12 6 12 6

6 4 6 2

K (1) = K (2) 5
= 8 10
12 6 12 6


6 2 6 4

The global stiffness matrix is

12 6 12 6 0 0

6 4 6 2 0 0


12 6 24 0 12 6

K = 8 105

.
6 2 0 8 6 2


0 0 12 6 12 6


0 0 6 2 6 4

The global load vector is

R1

R2



R3 + 6000

f =
.
1000


R5 + 6000


1000

The prescribed degrees of freedom are u1 = u2 = u3 = u5 = 0. The equations

for the unknowns u4 and u6 are

8 2 u 1000
8 105 4 = ,
2 4 u6 1000

73
which on solving yields u4 = 2.679 104 radians and u6 = 4.464
104 radians. The reactions are given by

R1 = 8 105 [6u4 ] = 1285.92 N,

R2 = 8 105 [2u4 ] = 428.64 N-m,
R3 = 8 105 [6u6 ] 6000 = 8142.72 N,
R5 = 8 105 [6u4 6u6 ] 6000 = 5143.2 N.

The positive sign on R1 indicates that the reaction at the wall is directed
downwards, while the negative signs on R3 and R5 indicate that the reactions
at the roller supports are directed upwards. The positive sign on R2 indicates
that the moment is clockwise at the wall.
The deflection at the midpoint of the second element is

v = N u(2) 

=0

0

iu4
h
= 0.5 0.5(0.25) 0.5 0.5(0.25)
0


u6
= 0.0893 mm.

The moment and shear force at this point are

EI
M= [le u4 + le u6 ]
le2
EI
= [u6 u4 ],
le
V = 12EI(u4 + u6 ).

The shear force in element 1 is

V = 6EI u4 = R1 ,

as expected.

3.9 Plane Frames

The element-level matrices for a frame element are derived using the element-
level matrices for the bar and beam elements in a manner analogous to the
derivation of the truss element in Section 3.6. For the plane frame element
shown in Fig. 3.15, the relation between the displacement degrees of freedom
in the local and global coordinate systems is

u = Qu,

74
 u5
u5 u4
u4
u 6 (u6 )

u2
u2
u1
u1
u 3 (u3)

Figure 3.15: Frame element.

where
l m 0 0 0 0

m l 0 0 0 0


0 0 1 0 0 0

Q= .
0 0 0 l m 0


0 0 0 m l 0


0 0 0 0 0 1
Similar to the truss element, the element stiffness matrix in global coordinates
is given by
K (e) = Qt K Q,
where
EA
l
0 0 EA
l
0 0

12EI 6EI
0 0 12EI 6EI

l3 l2 l3 l2


6EI 4EI
0 0 6EI 2EI

l2 l l2 l
K =

.
EA EA
l 0 0 0 0

l
0 12EI 6EI 0 12EI
6EI

l3 l2 l3 l2


6EI 2EI
0 l2 l
0 6EI
l2
4EI
l

The load vector in global coordinates is

f (e) = Qt f ,

where f is the load vector in the local coordinate system.

75
3.10 A Note on Equation Solving
We have seen that our finite element equations are a system of algebraic
equations of the form
K u = f .
There are two types of methods available for solving these equations, namely
(i) Direct methods which are based on some variant of Gauss elimination,
(ii) Iterative methods.
In the direct methods, K is decomposed into a lower triangular matrix,
L with unit diagonals, and an upper triangular matrix U , i.e.,

K = LU .

The solution to the equations is obtained by solving the equations

Lx = f and U u = x.

Thus,

x1 = f1
i1
X (3.13)
xi = fi Lij xj i = 2, . . . , n.
j=1

xn
un = ,
Unn
(3.14)
" n
#
1 X
ui = xi Uij xj i = n 1, n 2, . . . , 1.
Uii j=i+1

Eqn. 3.13 is called forward elimination, while Eqn. 3.14 is called back sub-
stitution. The actual decomposition K = LU is carried out using some
variation of Gauss elimination.
Once the L-U decomposition is computed, several solutions for different
load vectors can be computed using Eqns. 3.13 and 3.14. This process is
called resolution since it is not necessary to recompute L and U . Note
that the L-U decomposition is very costly compared to the resolution (o (n3 )
versus o (n2 )).
Several strategies exist for taking advantage of the banded and symmetric
form of the stiffness matrix. Even further storage reductions are possible by
using a profile of the K matrix. We then need to use a profile equation
solving program. See [4] for details of such solvers and of iterative techniques
such as the Gauss-Seidel method.

76
Chapter 4

Modeling of Two-Dimensional
Problems

In this chapter, we shall consider problems which can be modeled as two-

dimensional problems such as plane stress, plane strain or axisymmetric prob-
lems. We shall first consider the constant strain triangle (CST) element, and
later quadrilateral and higher-order triangular elements.

4.1 Constant Strain Triangle Element (CST

element)
The domain is discretized into triangular elements as shown in Fig. 4.1. The
corners of the elements are known as nodes1 . Note that the curved portion
of the body is approximated by piecewise linear segments corresponding to
the edges of the elements.
Since we are modeling a two-dimensional domain, each node has two de-
grees of freedom. Thus, the degrees of freedom corresponding to node j
(before incorporating the boundary conditions) are u2j1 and u2j The geo-
metric input data which is used by a finite element code for computing the
element level matrices is the set of nodal coordinates and the connectivity
1
Nodes need not always be located at the corners of an element as we have seen even
in the case of a quadratic one-dimensional element, but in the CST element they are.

u6
4 5 6
8 7
(x 3, y3 ) u5
8 7 4 1
2
3 1 u2 u4
u1 u3
(x 1, y1 ) (x 2, y2 )

Figure 4.1: Discretization of a two-dimensional domain; a typical element

77
 =0
N
=0

=0
=0
=0
1 =0
3
1
1
=1 =1
2

Figure 4.2: Shape functions at a node; restriction of shape function to an

element

data. The connectivity of a given element is the node numbers associated

with that element. For example, in Fig. 4.1, the connectivity of element 1 is
(1, 2, 5), that of element 4 is (1, 5, 4) that of element 8 is (3, 4, 8) and so on.
The shape functions are hat functions as shown in Fig. 4.2 assuming a
value of one at the node associated with it and zero elsewhere. Due to this
property they can be represented by area coordinates as shown in Fig. 4.3.
Mathematically, they are given by
A1 A2 A3
N1 = ; N2 = ; N3 = . (4.1)
A A A
Note that N1 + N2 + N3 = 1. We define natural coordinates as
A1 A2
= ; = . (4.2)
A A
With this definition, we have = 0 on side 2-3, and = 1 at point 1, and
= 0 on side 1-3, and = 1 at point 2, as shown in Fig. 4.3. From Eqns. 4.1
and 4.2, we get
N1 = ; N2 = ; N3 = 1 . (4.3)
For developing an isoparametric formulation, we use the same shape func-
tions for interpolating both, the displacements and the geometry. We thus

78
 =0
3

A2 A1
=1
A3 2
1 =0

=1
Figure 4.3: Representation of shape functions using area coordinates

have
u = N1 u1 + N2 u3 + N3 u5 ,
(4.4)
v = N1 u2 + N2 u4 + N3 u6 ,

or, alternatively,
u = N u,
where

N1 0 N2 0 N3 0
N =
0 N1 0 N2 0 N3
h it
u = u1 u2 u3 u4 u5 u6 .

Using the same interpolation functions for the geometry, we have

x = N1 x1 + N2 x2 + N3 x3 ,
(4.5)
y = N1 y1 + N2 y2 + N3 y3 .

In a sub-parametric formulation, the geometry is represented by lower-

order elements than those used to approximate the displacements, e.g., Euler-
Bernoulli beam element where Hermite shape functions are used to approx-
imate the deflection while the geometry is interpolated using linear interpo-
lation functions.
In superparametric formulations, the geometry is represented by higher-
order interpolations than the displacement. Such a formulation is seldom
used in practice.
Using Eqn. 4.3, Eqns. 4.4 and 4.5 can be written as

u = (u1 u5 ) + (u3 u5 ) + u5 ,

79
 v = (u2 u6 ) + (u4 u6 ) + u6 ,
x = (x1 x3 ) + (x2 x3 ) + x3 ,
y = (y1 y3 ) + (y2 y3 ) + y3 ,

We now turn to the evaluation of the B matrix. Using the chain rule, we
have
u u x u y
= + ,
x y
u u x u y
= + ,
x y

or, in matrix form

u x y u
= x
u x y u
y

u
=J x ,
u
y

where (with xij xi xj )

x y
x13 y13
J=
= ,
x y

x23 y23

is the Jacobian. Since

det J = x13 y23 x23 y13
is twice the area of the triangle, it is nonzero, and hence the Jacobian matrix
is invertible. Thus,
u u
x = J 1 .
u u
y

where
1 y23 y13
J 1 =
det J x23 x13

If the points 1, 2 and 3 are ordered in a counterclockwise manner, det J is

positive in sign.
As an example, let the coordinates of the nodes of an element be given
by (1.5, 2), (7, 3.5) and (4, 7), and let P be a point inside the triangle with
coordinates (3.85, 4.8). Then the natural coordinates of P are obtained using
3.85 = 1.5N1 + 7N2 + 4N3 ,

80
 4.8 = 2N1 + 3.5N2 + 7N3 ,
1 = N1 + N2 + N3 .

Solving the above set of equations, we get N1 = = 0.3 and N2 = = 0.2,

N3 = 0.5. The Jacobian matrix is

x13 y13 2.5 5.0
J = = ,
x23 y23 3.0 3.5

with det J = 23.75.

To find the strain-displacement matrix, note that
u

x
v
c =

y

u v
+
y x

y23 (u1 u5 ) y13 (u3 u5 )
1
= x23 (u2 u6 ) + x13 (u4 u6 ) .

det J


x23 (u1 u5 ) + x13 (u3 u5 ) + y23 (u2 u6 ) y13 (u4 u6 )

Now using the fact that y13 = y31 , y12 = y21 and so on, we get

y23 0 y31 0 y12 0
1
B= 0 x32 0 x13 0 x21

det J
x32 y23 x13 y31 x21 y12

Since B is a matrix of constants which are independent of the position, the

strain is constant over the element. This is as expected since the displacement
field is a linear combination of (1, x, y) leading to constant strain over the
element. The stiffness matrix is
Z
K= B t CB d
Z
= B t CBt dA
Ae
t
= tB CBAe ,

where the last step follows from the fact that B and C are constant over the
element. Using Eqn. 3.2, the half-bandwidth is given by
 
nbw = 2 max me + 1 .
1ene

81
In the above equation me is given by
me = max{|i1 i2 | , |i2 i3 | , |i1 i3 |}
where i1 , i2 and i3 are the node numbers of an element.
The consistent force vector due to body forces is given by
Z
fb = N t bte dA.
Ae

Assuming that the body force is constant, we get

N1 0

0 N1


N2 0 bx
Z
f b = te dA
Ae 0 N2 by


N3 0


0 N3

N b
1 x
N1 by


N2 bx
Z
= te dA
N2 by


N3 bx


N3 by

b
x
by


b

te Ae
x ,
=
3 by


bx


by
R
where the last step follows from either the physical interpretation of Ni dA
as the volume of a tetrahedron of unit height over a triangular base (Volume=one-
third of area of base multiplied by height), or by carrying out the integration
explicitly as
Z Z 1 Z 1
N1 dA = N1 det J dd
0 0
Z 1 Z 1
= 2Ae dd
0 0

82
 u4
u3
=1 2
u2
l
u6
1 u1
=0 3 u5
=1
=0

Figure 4.4: Computation of the load vector due to traction

Ae
= .
3
The contribution to the load vector due to the tractions acting on an edge
of the element is given by

Nt
1 x
N1 ty


N t
Z Z
t
2 x
f t = N tte dA = te ds.
N2 ty


N3 tx


N3 ty

As an example consider the case shown in Fig. 4.4. On edge 1-2, we have

x = x1 + x2 ,
y = y1 + y2 ,

leading to

dx = dx1 + dx2 ,
dy = dy1 + dy2 .

Since N3 = 0 on edge 1-2, we have + = 1, and hence d = d. Thus,

p p
ds = dx2 + dy 2 = (x1 x2 )2 + (y1 y2 )2 d = l d.

83
 4
10000 N
3 8
2 4
7

2
30 mm
1
2 6
1 3 5
1
60 mm
Figure 4.5: Example problem

Therefore R
1
t N l d
x R0 1 e
1
ty 0 N1 le d

R
tx 01 N2 le d

f t = te
R1
.
ty 0 N2 le d

R
1
tx 0 N3 le d

R
1
ty 0 N3 le d
Using the fact that one edge 1-2, N1 = , N2 = = 1 and N3 = 0, we get

t
x
ty


t

te l
x .
ft =
2 ty


0


0

Loads due to thermal effects are calculated by substituting 0 = [T, T, 0]t

in the case of plane stress, and 0 = (1 + ) [T, T, 0]t in the case of
plane strain in the expression
Z
f th = te B t C0 dA.
Ae

As an example, consider the problem shown in Fig. 4.5. The thickness,

Young modulus and Poisson ratio are t = 10 mm, E = 70 109 Pa and
= 1/3. The connectivity of the elements is

11 3 2

84
 2 3 4 2.

The coordinates of the nodes 1-4 are (0.0, 0.0), (0.0, 0.03), (0.06, 0.0) and
(0.06, 0.03). For element 1, we have x13 = 0, y23 = 0.03, x23 = 0.06,
y13 = 0.03, y12 = 0, x21 = 0.06, det J = 0.06(0.03) = 0.0018. Thus,

0.03 0 0.03 0 0 0
(1) 1
B = 0 0.06 0 0 0 0.06 .

0.0018
0.06 0.03 0 0.03 0.06 0

For element 2, we have x13 = 0.06, y23 = 0, x23 = 0.06, y13 = 0.03,
y12 = 0.02, x21 = 0, det J = 0.06(0.03) = 0.0018, yielding

0 0 0.03 0 0.03 0
(2) 1
B = 0 0.06 0 0.06 0 0 .

0.0018
0.06 0 0.06 0.03 0 0.03

The elasticity matrix is given by

1 1/3 0
9 70 109
C= 1/3 1 0

8
0 0 1/3

The matrices B t CB for the two elements are given by

2.1 1.2003 0.9 0.6 1.2 0.6

1.2003 3.9 0.6 0.3 0.6 3.6


0.9 0.6 0.9 0 0 0.6

t 13
[B CB]1 = 2.43 10 ,
0.6 0.3 0 0.3 0.6 0


1.2 0.6 0 0.6 1.2 0


0.6 3.6 0.6 0 0 3.6

1.2 0 1.2 0.6 0 0.6

0 3.6 0.6 3.6 0.6 0


1.2 0.6 2.1 1.2 0.9 0.6

t 13

[B CB]2 = 2.43 10 .
0.6 3.6 1.2 3.9 0.6 0.3


0 0.6 0.9 0.6 0.9 0


0.6 0 0.6 0.3 0 0.3

85
The global degrees of freedom associated with the two elements are (1, 2, 5, 6, 3, 4)
and (5, 6, 7, 8, 3, 4). Assembling the element level matrices into the global
stiffness matrix and incorporating the boundary conditions, we get

2.1 0 1.2 0.6 u5 0

0 3.9 0.6 3.6 u6 0

7

9 2.43 10 =
.
1.2 0.6 2.1 1.2 u7 0


0.6 3.6 1.2 3.9 u8 10000

Solving the above set of equations, we get u5 = 2.1105 m, u6 = 1.208

104 m, u7 = 3.123 105 m, u8 = 1.361 104 m, and R1 = 20000.0 N,
R2 = 10690.0 N, R3 = 20000.0 N, R4 = 690.42 N. Note that R1 + R3 0
and R2 + R4 10000 N as expected.
The stresses are given by

u
1
u2

6.3
u

(1) (1)
5 6
= CB = 4.375 10 2.1 ,

u6
12.08
u3


u4

u
5
u6

6.3
u

(2) 7
(2)
= CB = 4.375 106 6.07 .

u8
3.167
u3


u4

4.2 The Four-Node Quadrilateral Element (Bi-

linear Element)
An easy way to ensure compatibility of the field variable across the element
edges and also simplify computations of the element matrices is to map the
element into a master element as shown in Fig. 4.6. The shape functions
are now formulated in terms of and . As an example consider finding the
expression for N1 . N1 should have a value of one at node 1 and zero at the
other nodes. We assume an expression for N1 of the form

N1 = c(1 )(1 ),

86
 u6 (1, 1) (1, 1)

u8 u5

u7

u2 u4
u1 (1, 1) (1, 1)
u3

Figure 4.6: The bilinear element: Mapping the actual element into a master
element

which ensures that it is zero on the edges 2-3 and 3-4. The constant c is
determined by using the fact that N1 is one at (, ) = (1, 1). A similar
procedure is followed for finding N2 , N3 and N4 . We get
1
N1 = (1 )(1 ),
4
1
N2 = (1 + )(1 ),
4
1
N3 = (1 + )(1 + ),
4
1
N4 = (1 )(1 + ).
4
A general way of representing the above functions is
1
Ni = (1 + i )(1 + i ),
4
where (i , i ) are the coordinates of node i.
The displacement field can be written as
u = N1 u1 + N2 u3 + N3 u5 + N4 u7 ,
v = N1 u2 + N2 u4 + N3 u6 + N4 u8 ,
or alternatively in matrix form as u = N u, where

N1 0 N2 0 N3 0 N4 0
N = .
0 N1 0 N2 0 N3 0 N4

In an isoparametric formulation, the geometry is also interpolated using the

same shape functions, i.e.,
x = N1 x1 + N2 x2 + N3 x3 + N4 x4 ,
v = N1 y1 + N2 y2 + N3 y3 + N4 y4 .
By using the chain rule, we get

f f
=J x ,
f f
y

87
where

x y
J11 J12
J =
x y
J21 J22

x1 y1

1 (1 ) (1 ) (1 + ) (1 + ) x2 y2

= .
4 (1 ) (1 + ) (1 + ) (1 ) x3 y3


x4 y4

Thus,
f f
11 12
x = ,
f f
y
21 22

where
J22 J12 J21 J11
11 = ; 12 = ; 21 = ; 22 = ,
|J | |J | |J| |J |
where |J | det J .
The strain-displacement matrix is found as follows:
We have (using the comma notation for denoting partial derivatives)

u,x u
xx 1 0 0 0 ,x
u,y u


c = yy = 0 0 0 1
= R1 ,y . (4.6)
v,x

v,x

xy 0 1 1 0
v,y v,y

Next, we have

u,x 11 12 0 0 u, u,

u,y 21 22 0 0 u, u

= = R2 , . (4.7)
v,x 0 0 11 12 v, v,


v,y 0 0 21 22 v, v,

88
Finally, we have

u
1
u

2
u N 0 N2, 0 N3, 0 N4, 0 u3

, 1,
u, N1, 0 N2, 0 N3, 0 N4, 0 u4

=
v, 0 N1, 0 N2, 0 N3, 0 N4, u5


v, 0 N1, 0 N2, 0 N3, 0 N4, u6


u7


u8

(1 ) 0 (1 ) 0 (1 + ) 0 (1 + ) 0

(1 ) 0 (1 + ) 0 (1 + ) 0 (1 ) 0

1
= u
4

0 (1 ) 0 (1 ) 0 (1 + ) 0 (1 + )


0 (1 ) 0 (1 + ) 0 (1 + ) 0 (1 )
= R3 u. (4.8)

From Eqns. 4.6, 4.7 and 4.8, we get

c = R1 R2 R3 u,

thus yielding
B = R1 R2 R 3 .
The stiffness matrix is given by
Z Z 1 Z 1
t
K= B CB d = B t CBt |J | d d.
1 1

Since the above integral is complicated to evaluate in closed form, we use a

numerical technique such as Gaussian quadrature (see Section 4.4).
The consistent load vector is given by
Z Z 1Z 1
(e) t
f = N t d + N t bt |J | d d.
t 1 1

The body force term is most easily computed using Gaussian quadrature. As
an example, consider the computation of the consistent load vector due to
a uniform traction on the edge 2-3 in Fig. 4.6. Since x = N2 x2 + N3 x3 and
y = N2 y2 + N3 y3 with N2 = (1 )/2 and N3 = (1 + )/2 along this edge,
we have dx = (x3 x2 ) d/2 and dy = (y3 y2 ) d/2 leading to
p tl23
d = t ds = t dx2 + dy 2 = d,
2
89
 =0

=1 3
=1/2
5 2 6
=1/2 5 =1
3
1 4
6 =0 2
4
=1 =1/2 =0
1

Figure 4.7: Six-node triangular element

p
where l23 = (x2 x3 )2 + (y2 y3 )2 is the length of the side 2-3. Hence,

0 0

0 0


1
0

2

Z 1 1
0 t
2 x l23
f (e) = t d,
1
1+
0

t 2
2 y

0 1+

2


0 0


0 0
tl23 h it
= 0 0 tx ty tx ty 0 0 ,
2
which agrees with the intuitively expected answer.

4.3 Higher-Order Elements

We now consider higher-order elements such as the 6-node triangle (quadratic
triangle), and the 8-node and 9-node quadrilateral elements. Note that all
these elements can model curved boundaries more accurately than the linear
elements.

4.3.1 Six-node triangular element

The master element is as shown in Fig. 4.7. The shape functions are
N1 = (2 1), N4 = 4,
N2 = (2 1), N5 = 4,
N3 = (2 1), N6 = 4,
where = (1 ). N1 is obtained by assuming it to be of the form
c(2 1) (so that it is zero on the lines = 0 and = 1/2) and then finding
c so that N1 = 1 at = 1. The other shape functions are obtained similarly.

90
 3 4 7 3
7 1++=0
1+=0

4 6
8 6

8 1=0
2 1+=0
5 1 5 2
1

Figure 4.8: Eight-node quadrilateral element

The displacements are interpolated using u = 6i=1 Ni ui . The geometry

P
is interpolated using the same shape functions in an isoparametric formula-
tion. The B matrix is found in the same way as for the four-node quadrilat-
eral element. The stiffness matrix is given by
Z 1 Z 1
(e)
K = B t CBt |J| d d.
0 0

4.3.2 Eight-node quadrilateral (serendipity) element

The mapping is as shown in Fig. 4.8. The shape function N1 is found by
assuming it to be of the form

c(1 )(1 )(1 + + ),

so that it is zero at all nodes except node 1. Since N1 = 1 at (1, 1), we

get c = 1/4. Similarly, N5 is found assuming it to be of the form

k(1 )(1 )(1 + ),

and then determining k. The final expressions for the shape functions are
1 1
N1 = (1 )(1 )(1 + + ), N5 = (1 2 )(1 ),
4 2
1 1
N2 = (1 + )(1 )(1 + ), N6 = (1 + )(1 2 ),
4 2
1 1
N3 = (1 + )(1 + )(1 ), N7 = (1 2 )(1 + ),
4 2
1 1
N4 = (1 )(1 + )(1 + ), N8 = (1 )(1 2 ).
4 2
The stiffness matrix is given by
Z 1Z 1
(e)
K = B t CBt |J | d d,
1 1

where B is formulated in the same way as for the four-node quadrilateral

element.

91
 4 7 3
3
7

4 6 8 9 6
9

8
2 1 5 2
5
1

Figure 4.9: Nine-node quadrilateral element

4.3.3 Nine-node quadrilateral element

The shape functions for the nine node-quadrilateral element shown in Fig. 4.9
are found by using products of quadratic shape functions in one-dimension
along the and directions. The shape functions are given by
N1 = L1 ()L1 (), N5 = L2 ()L1 (),
N2 = L3 ()L1 (), N6 = L3 ()L2 (),
N3 = L3 ()L3 (), N7 = L2 ()L3 (),
N4 = L1 ()L3 (), N8 = L1 ()L2 (),
N9 = L2 ()L2 (),
where
1 1
L1 () = (1 ), L1 () = (1 ),
2 2
L2 () = (1 + )(1 ), L2 () = (1 + )(1 ),
1 1
L3 () = (1 + ), L3 () = (1 + ).
2 2
The stiffness matrix is formed in the same way as for the eight-node quadri-
lateral element.

4.3.4 Work-equivalent loads for the eight-node and nine-

node quadrilateral elements
As an example of the computation of work-equivalent loads, consider that
a linearly varying pressure is applied to edge 2-3 in Fig. 4.8 (or Fig. 4.9;
the same work-equivalent loads are obtained in either case). Assume that
node 6 is located midway between nodes 2 and 3, i.e., x6 = (x2 + x3 )/2
y6 = (y2 + y3 )/2. Let the magnitude of the pressure be pi at node 2 and pj
at node 3. Since edge 2-3 is mapped into the edge = 1, we have
1 1
t = (1 )pi + (1 + )pj .
2 2
Also
1
N2 |=1 = (1 ),
2
92
 1
N3 |=1 = (1 + ),
2
N6 |=1 = 1 2 .
The element of area is given by
p tl23
d = t ds = t dx2 + dy 2 = d.
2
The elements of the consistent load vector corresponding to the degrees of
freedom associated with edge 2-3 are
Z
f = N t t d

(1 )[pi + pj (pi pj )]

0



Z 1 2
2(1 )[pi + pj (pi pj )]

tl23
= d
8 1 0



(1 + )[pi + pj (pi pj )]


0

4pi
3
0


8(pi +pj )
tl23 3

=
8 0


4pj
3

0

pi l23
6
0



2(pi +pj )l23
= t 6


0



pj l23
6

In the special case when pi = pj = p, the total load P = ptl23 is distributed

as (P/6, 4P/6, P/6) at the three nodes 2, 6 and 3, something which is not
intuitively obvious.

4.4 Gauss Quadrature

Gauss quadrature is an approximate method of computing integrals. We first
consider the gauss quadrature for one-dimensional integrals

93
4.4.1 One-dimensional problems
Consider an integral of the form
Z 1
I= f () d.
1

We approximate it as

I = w1 f (1 ) + w2 f (2 ) + . . . + wn f (n ), (4.9)

where w1 , w2 ,...,wn are the weights, and 1 , 2 ,..., n are the sampling points
or Gauss points. We select the Gauss points and weights such that Eqn. 4.9
provides an exact answer for polynomials of as large a degree as possible. The
idea is that if n-point formula is exact for polynomials of as high a degree as
possible, then the formula will work well even if f () is not a polynomial.
One-point formula
We approximate the integral just using the first term, i.e.,
Z 1
f () d w1 f (1 ).
1

The two parameters w1 and 1 are found such that the integral is exact when
f () is of the form a0 + a1 . Hence, we get
Z 1
(a0 + a1 ) d = w1 (a0 + a1 1 ),
1

or, alternatively,
2a0 = w1 a0 + w1 1 a1 .
Comparing the coefficients of a0 and a1 , we get w1 = 2 and 1 = 0. Thus,
the one-point approximation formula for any arbitrary f () is
Z 1
f () d 2f (0).
1

Two-point formula:
Now we have the approximation
Z 1
f () d = w1 f (1 ) + w2 f (2 ).
1

Since there are four parameters, w1 , 1 , w2 and 2 , we can integrate a cubic

polynomial exactly. Thus, we determine these four parameters such that
Z 1
(a0 + a1 + a2 2 + a3 3 ) d = w1 (a0 + a1 1 + a2 12 + a3 13 )+
1
w2 (a0 + a1 2 + a2 22 + a3 23 ).

94
Comparing the coefficients of a0 , a1 , a2 and a3 , we get

w1 + w2 = 2,
w1 1 + w2 2 = 0,
2
w1 12 + w2 22 = ,
3
w1 13 + w2 23 = 0.

Solving the above equations, we get w1 = w2 = 1, and 1 = 2 = 1/ 3.
Thus, for any arbitrary function f (), we have
Z 1    
1 1
f () d f +f .
1 3 3
In general, an n-point Gauss-quadrature formula will provide an exact
answer if f () is a polynomial of order (2n 1) or less.

Sampling points and weights for Gaussian quadrature

Order n Location i Weight wi

1 0 2
2 13 1
5 5 8
3 0.6, 0 , ,
9 9 9
q
4 3+27 1.2 1
2
1
6 1.2
q
327 1.2 1
2
+ 1
6 1.2

Note that the Gauss points are located symmetrically with respect to the
origin,and that symmetrically placed points have the same weight.
Example:
Consider the approximate evaluation of
Z 1 
x 2 1
I= 3e + x + dx = 8.8165.
1 x+2

Using a one-point quadrature,we get I = 2f (0) = 7.0. Using a two-point

quadrature, we get I = f (1/ 3) + f (1/ 3) = 8.7857. Note that the result
converges to the exact solution.

4.4.2 Two-dimensional Gauss quadrature

An integral with two independent variables such as
Z 1Z 1
I= f (, ) d d,
1 1

95


( 13 , 13 ) ( 13 , 13 )

4 3

1 2
( 13 , 13 ) ( 13 , 13 )

Figure 4.10: Gauss points for a 2 2 integration rule

is approximated as
" n #
Z 1 X
I= wi f (i , ) d
1 i=1
n
" n #
X X
= wj wi f (i , j )
j=1 i=1
n
XX n
= wi wj f (i , j ).
i=1 j=1

As an example, consider the evaluation of the stiffness matrix given by

Z 1Z 1
(e)
K = B t CB |J| t d d.
1 1

Let (, ) = t |J | (B t CB)ij . For a four-node quadrilateral element, if we

use a 2 2 integration rule (see Fig. 4.10), we get

Kij = w12 (1 , 1 ) + w1 w2 (1, 2 ) + w2 w1 (2 , 1 ) + w22 (2 , 2 )

1 1 1 1 1 1 1 1
= ( , ) + ( , ) + ( , ) + ( , ).
3 3 3 3 3 3 3 3

4.4.3 Order of integration for quadrilateral elements

We first calculate the required integration order for a four-node rectangular
element. Let the coordinates of the four nodes be (x1 , y1 ), (x2 , y1 ), (x2 , y2)
and (x1 , y2 ). Then

x = N1 x1 + N2 x2 + N3 x2 + N4 x1
1 1
= (1 )x1 + (1 + )x2
2 2
1 1
y = (1 )y1 + (1 + )y2 .
2 2
96
The Jacobian is

x y 1
(x x1 ) 0
J=
= 2 2 = constant.
x y 1

0 (y
2 2
y1 )

Hence, the determinant of the Jacobian, |J | is also constant at all points in

the element. Therefore,

B t CB |J| = f ( 2 , 2 , ).

Since the highest order of the polynomial in one direction is 2, the order of
Gaussian integration which is needed is 2 2 (2n 1 2 = n = 2). In
a similar way, the required integration order for evaluating the mass matrix
or the load vector can be assessed.
If the element is distorted, i.e., it is not rectangular or a parralelogram,
then the Jacobian matrix is not a constant over the element. In such a case,
the integrand in the stiffness matrix is a ratio of polynomials which in general
is not a polynomial. Full numerical integration is the order of integration
that gives exact matrices (i.e., same as analytical integration) when the ele-
ments are undistorted. Using this integration order for a distorted element
will not yield the exactly integrated element matrices. However, the numer-
ical integration errors are small if the geometric distortions are small. If the
element is highly distorted (a situation which should be avoided as far as
possible), one should use a higher order of quadrature than the one used in
full integration. The recommended full gauss numerical integration orders
for the 4, 8, 9 and 16 node isoparametric quadrilateral elements are 2 2,
3 3, 3 3 and 4 4, respectively.
If an integration order lower than the full order is used it is called re-
duced integration. Using reduced integration is, in general, unreliable. It
can result in an instability known variously as a mechanism, kinematic mode,
hourglass mode or zero-energy mode. For example, the nodal displacement
in the setup shown in Fig. 4.11 where 2 2 integration is used to compute
the stiffness matrix are very large due to the zero-energy mode. An eigen-
value analysis of the element stiffness matrices for a plane [solid] element
should yield three [six] zero eigenvalues corresponding to the three [six] rigid
body modes (2 translations and a rotation [3 translations and 3 rotations]).
However, in the problem shown above, the element stiffness computed using
a 2 2 integration rule has one spurious zero-energy mode, i.e., four zero
eigenvalues, due to the use of reduced integration.

4.4.4 Numerical integration for triangular elements

Let (, , ) be a function to be integrated. Then
n
1X
Z
|J | d d wi |J|i i , (4.10)
2 i=1

97
 P

Figure 4.11: An example of zero-energy modes due to reduced integration

where wi , i and |J |i are the values of the weight, function and determinant
of the Jacobian at the sampling point. The factor 1/2 occurs in the formula
because for anP undistorted triangle
R of unit area, |J | = 2 throughout the
triangle. Since wi = 1, we get dA = 1 when = 1. The area coordinates
and the weights for various orders of integration are presented in the following
table.

Area coordinates and weights for triangular elements

Order n Area coordinates Weight wi

1 1 1
1 , ,
3 3 3
1
2 1 1 1
2 , ,
3 6 6 3
1 2 1 1
, ,
6 3 6 3
1 1 2 1
, ,
6 6 3 3

2 1 1
, ,0 1 (4.11)
2 2 3

0, 21 , 21 1
3
1
2
, 0, 21 1
3
1 1 1
3 , ,
3 3 3
0.5625
3 1 1
, ,
5 5 5
0.5208333..
1 3 1
, ,
5 5 5
0.5208333..
1 1 3
, ,
5 5 5
0.5208333..

The location of the Gauss points and the associated weights in Table (4.11)
are found as follows. If f (, ) = a0 + a1 + a2 , then using Eqn. (4.10), we
get
1
a0 + (a1 + a2 ) = w1 (a0 + a1 1 + a2 1 ),
3
which yields w1 = 1 and 1 = 1 = 1/3. If f (, ) = a0 + a1 + a2 + a3 2 +

98
a4 + a5 2 , then using Eqn. (4.10), we get

1 1 1
a0 + (a1 +a2 )+ (a3 +a5 )+ a4 = w1 (a0 +a1 1 +a2 1 +a3 12 +a4 1 1 +a5 12 )
3 6 12
+ w2 (a0 + a1 2 + a2 2 + a3 22 + a4 2 2 + a5 22 ).

However, the Gauss points are not symmetrically located! If we choose a

3-point quadrature rule, i.e.,

1 1 1
a0 + (a1 +a2 )+ (a3 +a5 )+ a4 = w1 (a0 +a1 1 +a2 1 +a3 12 +a4 1 1 +a5 12 )
3 6 12
+w2 (a0 +a1 2 +a2 2 +a3 22 +a4 2 2 +a5 22 )+w3 (a0 +a1 3 +a2 3 +a3 32 +a4 3 3 +a5 32 ),

then we have more unknowns than equations. The Gauss points can now be
chosen to be symmetrically located, but as seen in Table (4.11), the choice
is not unique.
We have seen that for a three-node triangle, the stiffness matrix can be
constructed explicitly, and there is no need of numerical integration. For a
quadratic triangle,
B t CB |J| = f ( 2 , , 2).
Hence, a three point integration is exact when the triangle is an equilateral
triangle. As in the case of the quadrilateral elements, the integration formula
is approximate when the triangular element is distorted.

4.4.5 Stress Computation

If stresses are directly computed using

c = C(B u 0c ) + 0c ,

then the results obtained can be quite inaccurate since the process of dif-
ferentiation results in a loss of accuracy. For isoparametric elements, the
stresses are most accurate at Gauss points of a quadrature rule one order
lower than that required for full integration of the element stiffness. Such
points are known as Barlow points.
Consider the examples shown in Fig. 4.12. In example (a), the displace-
ments are of the form u = a, and v = 0, where a is a positive constant.
Thus, the shear strain xy is proportional to () yielding the variation shown
in the figure. Since the exact shear strain is zero, we see that the finite ele-
ment solution matches the exact solution at = 0 and = 0, which is also
the Gauss point location for a order-1 integration rule. Similarly, as shown
in Fig. 4.12b, for an 8-node (or even 9-node) element, the Barlow points (or
superconvergent points) are the Gauss-point locations for the 2 2 inte-
gration rule ( = = 1/ 3). These conclusions are true for rectangular
elements. For distorted elements, Gauss points may not be optimal locations,
but, nonetheless, they remain good choices.
To obtain more accurate stresses than those obtained from direct compu-
tation, the strategy used is to extrapolate the stress values from the Gauss

99
 xy
P Exact
2L
0 L
P

(a)

xy Exact

V x

0 L 2L

(b)

Figure 4.12: Stress variation within an element

=1

r=1 =1
s=1

Figure 4.13: Extrapolation of stress

point to the nodes, and then average the values obtained from all the el-
ements sharing that node to obtain the nodal stresses. An example of an
extrapolation strategy is shown in Fig. 4.13. Since the element shown is a
four-node element, the 2 2 Gauss points are located at = = 1/ 3.
Let r and s be a scaled version of the - system such
that they have the
sameorigin but with r = s = 1 when = = 1/ 3. In other words, let
r = 3 and s = 3. The extrapolated stress components are given by (
is a typical stress component)
4
X
= Ni i ,
i=1

where i are the values of the stress components at the Gauss points, and
1
Ni = (1 r)(1 s).
4
100
Thus, the stress components
at point A in Fig. 4.13 are obtained by substi-
tuting r = s = 3 in the above formula. The four values obtained at each
node are averaged, and this mean value is taken as the nodal value. The
improved stress field is given by
n
X
= Ni (mean )i ,
i=1

where n is the number of nodes, and Ni are the displacement shape functions.

4.5 Heat transfer Problems

The procedure for formulating the various matrices is similar to that for
elasticity. We shall just consider the details for the CST element. The
temperature is interpolated as

T = N1 T1 + N2 T2 + N3 T3 ,

where N1 = , N2 = and N3 = 1 . Thus, the temperature gradient

is
T T
= J x ,
T T
y

where (with xij xi xj )

x13 y13
J = .
x23 y23

Inverting the above relation, we get

T T
y y13
x = 1 23
T det J x23 x13 T
y

T1
1 y23 y13 1 0 1
= T2


det J x23 x13 0 1 1
T3

T
1
= B T2 ,


T3

where
1 y23 y31 y12
B= .
det J x32 x13 x21

101
Assuming that K is constant and isotropic over each element, i.e., K = kI,
the volume term of the stiffness matrix is
Z
(e)
K 1 = ke B t Bt dA = tke Ae B t B.

Assuming that convection takes place across edge 2-3 ( = 0) of the

element, the shape function matrix can be written as
h i
N = 0 1 .

The surface area element is given by t ds = tl23 d. Hence, the surface term
of the stiffness matrix is
Z 1
(e)
K2 = hN t N tl23 d
0

0 0 0
htl23
= 0 2 1 .

6
0 1 2

The load term due to the heat source Q is

Z
f = N t Qt dA.
(1)

We get

1
tQe Ae
f (1) = 1 if Q is constant,

3
1

2Q1 + Q2 + Q3
tAe
= Q1 + 2Q2 + Q3 if Q = N1 Q1 + N2 Q2 + N3 Q3 ,

12
Q1 + Q2 + 2Q3

0

= tQ0 N0 (0 , 0 ) = tQ0 0 if Q is a point source.


1 0 0

Assuming that the edge 2-3 constitutes part of q , and assuming that (q +
hT ) = constant, the load term due to the prescribed heat flux and convec-
tion is
Z 1
(2)
f = N t (q + hT )tl23 d
0

102
 T = 180o C 5 T = 180o C 4

3
3
h = 50 W/m2o C q = 0 2 0.3 m
0.6 m q = 0 T = 25o C 1 q = 0
1 2
0.4 m
o
q = 0
T = 180 C

0.4 m

Figure 4.14: Example problem

0
1
= (q + hT )tl23 1 .

2
1

As an example, consider the problem shown in Fig. 4.14 where the do-
mains conductivity is K = kI, where k = 1.5 W/m o C. Due to the symme-
try of the domain and boundary conditions, we can model only the upper half
of the domain as shown. Due to symmetry, the heat flux at the bottom edge
of the finite element analysis domain is q = T
n
= 0. On the right edge, since
T
there is no prescribed flux, we have n = h(T T ). The connectivity is
given by

11 2 3
25 1 3
35 3 4

The B matrices are

1 0.15 0.15 0
B (1) = ,
0.06 0 0.4 0.4

1 0.15 0.15 0.3
B (2) = ,
0.12 0.4 0.4 0

1 0.15 0 0.15
B (3) = .
0.06 0 0.4 0.4

The stiffness matrix due to the conductivity part is

(e)
K T = kAB t B,

while that due to the convection part (assuming that the convection is taking

103
place at edge 2-3 of the element) is

0 0 0
(e) hl23
Kh = 0 2 1 ,

6
0 1 2

The global stiffness matrix is

1.42125 0.28125 0.28125 0 0.86

0.28125 4.78125 0.75 0 0



K = 0.28125 0.75 9.5625 0.75 0.28125 .


0 0 0.75 4.78125 0.28125



0.86 0 0.28125 0.28125 1.42125

The load vector due to convection is

0
(e) hT l23
fh = 1 ,

2
1

The global load vector is

0

1


f = 93.75 2


1


0
The prescribed degrees of freedom are T4 = T5 = 180. Solving for T1 , T2 and
T3 we get h i h i
T1 T2 T3 = 124.5 34.0 45.4 .
Note that a large number of nodes are required to capture accurately the
large temperature gradient along the edge 2-4 (T2 = 34, T4 = 180).

4.6 Axisymmetric Problems

We now consider problems with axisymmetric geometry and axisymmetric
loading. Such problems can be treated as two-dimensional problems since
there is no dependence of any of the variables on . We denote the com-
ponents (ur , u , uz ) by (u, v, w). Because of the symmetry about the z-axis,

104
 12
z 11
10
u6
9 u5
8
7 u2 u4

5 6 u1 u3
4

r 1 2 3

Figure 4.15: Meshing for an axisymmetric problem

v = 0. The engineering strains are

u

rr r
w
zz z
= u w .
c =
rz z + r

u
r

If [rr , zz , rz , ]t , the stress-strain relations are

c = C(c 0c ),

where (with f = /(1 ) and g = (1 2)/(2(1 )))

1 f 0 f

f 1 0 f

(1 )E
C= .
(1 + )(1 2)
0 0 g 0


f f 0 1

For thermal loading 0c = T [1, 1, 0, 1]t.

The meshing for an axisymmetric problem using CST elements, and the
degrees of freedom for a single element are shown in Fig. 4.15. The displace-
ment is given by

u = N u

N1 0 N2 0 N3 0 h it
= u1 u2 u3 u4 u5 u6 .
0 N1 0 N2 0 N3

For an isoparametric element,

r = r1 + r2 + (1 )r3 ,
z = z1 + z2 + (1 )z3 .

105
In the usual manner, we have

u u w w
r =J 1
; r =J 1
,
u u w w
z z

where (with rij ri rj and zij zi zj )

r z
r13 z13
J=
=
r z

r23 z23

Note that det J = r13 z23 r23 z13 = 2A. The engineering strain vector is
given by
c = B u,
where
z23 z31 z12
0 0 0
|J | |J| |J |
r32 r13 r21
0 0 0

|J | |J| |J |

B=
r32
|J | z23 r13 z31 r21 z12
|J | |J| |J| |J | |J |


N1 N2 N3
r
0 r
0 r
0
The element stiffness matrix is given by
Z
K = 2 B t CBr drdz
(e)

Z 1 Z 1
= 2 B t CBr(2Ae ) dd
0 0
Z 1 Z 1
= 4Ae B t CBr dd,
0 0

where r = r1 + r2 + (1 )r3 . The element force vectors are (assuming

that the traction acts on edge 2-3 of the element)
Z 1 Z 1 Z 1
(e) t
f = 2 N br(2Ae ) dd + 2 N t trl23 d
0 0 0

N1 rbr 0

N1 rbz 0


Z 1 Z 1 Z 1
N2 rbr N2 rtr

= 4Ae dd + 2l23 d,
N2 rbz 0 N2 rtz

0 0

N3 rbr N3 rtr


N3 rbz N3 rtz

where r = r2 + (1 )r3 in the second integral.

106
 Since the 2 factor is common in both, the stiffness and the force matrices,
one can cancel it out, or alternatively, carry out the formulation using a one
radian sector in the direction. An example of a axisymmetric problem is a
rotating flywheel with br = r 2 and bz = g.
For the four-node bilinear element, the strain-displacement matrix is
found in the same manner as in the plane stress/plane strain case. We have

u
r
rr
u

z

zz

= H w

, (4.12)
rz
r
w
z

u

where
1 0 0 0 0

0 0 0 1 0

H = .
0 1 1 0 0


0 0 0 0 1r
Also
u u
r
u u
z

w w
r = A , (4.13)

w w
z

u u
where
J 1 0 0

A= 0 1
0 .

J

0 0 1
Finally,
u

u


w
= Qu, (4.14)

w


u

107
where

N 0 N2, 0 N3, 0 N4, 0
1,
N1, 0 N2, 0 N3, 0 N4, 0


Q = 0 N1, 0 N2, 0 N3, 0 N4, .


0 N1, 0 N2, 0 N3, 0 N4,


N1 0 N2 0 N3 0 N4 0

From Eqns. 4.12-4.14, we get

B = HAQ.

The stiffness matrix is given by

Z 1Z 1
(e)
K = 2 B t CB |J | r dd,
1 1

where r = N1 r1 + N2 r2 + N3 r3 + N4 r4 . Assuming that the tractions act on

edge 2-3, the element load vector is
Z 1Z 1 Z 1
(e) t
f = 2 N br |J | dd + 2l23 [N t tr]=1 d.
1 1 1

Some terms in the K matrix are of the form 1/r. With Gauss quadrature,
these terms remain finite since there are no Gauss points at r = 0. For stress
computation, the indeterminate form = u/r = 0/0 arises for points on
the z-axis. Hence, it is convenient to calculate the stresses at the Gauss
points and extrapolate them to the axis. Another way is to use the fact that
= rr at r = 0. Thus, for r = 0, replace the row of the B matrix by
the rr row. The radial displacement u = 0 has to be prescribed at all nodes
that lie on the z-axis. Rigid body motion is restrained by prescribing w on
a single nodal circle.

108
Chapter 5

Convergence of the Finite

Element Method

In this chapter, we discuss some of the important theoretical properties of

the finite element method. Recall that Vh is a subspace of the Hilbert space
V in which the finite element solution lies, and that h is the generic element
size.

5.1 Some Properties of the Finite Element

Method
The abstract formulation of the finite element method is
Find uh Vh such that
a(uh , vh ) = L(vh ) vh Vh .
Assume a(., .) to be symmetric. Then we have the following properties:
Property 1 (Error equation): Let the error between the exact solution and
the finite element solution be eh , i.e., eh = u uh . Then
a(eh , vh ) = 0 vh Vh .
Proof: Since Vh V , we have
a(u, vh ) = L(vh ) vh Vh ,
a(uh , vh ) = L(vh ) vh Vh .
Subtracting the second equation from the first and using the bilinearity of
a(., .), we get the desired result. This result can be interpreted as saying that
the error is orthogonal to all elements of Vh in a(., .). Thus, as Vh increases,
with the larger space containing the smaller space, the solution accuracy
increases continuously.
Property 2:
a(uh , uh ) a(u, u).
Proof:
a(u, u) = a(uh + eh , uh + eh )

109
 = a(uh , uh ) + a(eh , eh ) + 2a(uh , eh )
= a(uh , uh ) + a(eh , eh ), (5.1)

since a(uh , eh ) = 0 by property 1 with vh = uh . Since a(eh , eh ) > 0 for

any eh 6= 0, we get the desired relation. The physical interpretation of
this statement is that the strain energy corresponding to the finite element
solution is always smaller than or equal to the strain energy corresponding
to the exact solution, or in other words, the finite element approximation
makes the structure stiffer than it actually is.
Property 3:
a(eh , eh ) a(u vh , u vh ) vh Vh .
Proof: Since a(eh , wh ) = 0 for all wh Vh ,

a(u uh + wh , u uh + wh ) = a(eh , eh ) + a(wh , wh ),

which in turn implies that

a(eh , eh ) a(eh + wh , eh + wh ).

Choose wh = uh vh to get the desired result. Thus, the finite element

solution uh is chosen from all possible displacement patterns vh Vh such
that the energy distance between u and the elements in Vh is minimized.
Thus, the solution uh is the projection of u onto Vh as shown in Fig. 2.3.

5.2 Superconvergence of the Finite Element

Method in One-dimensional Problems
We show that in certain types of one-dimensional problems, the finite element
solution is exact at the nodes. In two or three dimensions however, we do
not have this property. We first introduce the Dirac-Delta function y (x) =
(y x) which is defined by means of the equation
Z 1
w(x)y (x) dx = w(y).
0

Integration of y (x) yields the Heaviside (or step function)

(
0 if x < y
Hy (x) = H(x y) = .
1 if x > y

The integral of Hy (x) is the Macaulay bracket:

(
0 if x y
hx yi = .
x y if x > y

The pictorial representation of the three functions discussed above is shown

in Fig. 5.1.

110
 y( x) H(x-y) x-y

1 (1-y)

y x y x y 1 x

Figure 5.1: Dirac-Delta, Heaviside and Macaulay bracket functions

Consider the one-dimensional problem

d2 u
+ f = 0 on [0, 1], (5.2)
dx2
with the boundary conditions u(0) = u(1) = 0. The variational formulation
is Z 1 Z 1
du dv
dx = f v dx v H01 (I). (5.3)
0 dx dx 0

The solution to Eqn. 5.2 when f = y (x) (unit point load at x = y) is known
as the Greens function (denoted by G):

d2 G
+ y (x) = 0 on [0, 1], (5.4)
dx2
with G(0) = G(1) = 0. Integrating Eqn. 5.4, we get

G(x) + hx yi = c1 x + c2 ,

where c1 and c2 are determined from the boundary conditions on G. We get

c1 = (1 y) and c2 = 0, thus yielding
(
(1 y)x if 0 x y
G(x) =
y(1 x) if y x 1.

Note that G is piecewise linear and hence belongs to Vh when y is a nodal

point. Substituting u = G and f = y (x) in Eqn. 5.3, we get
Z 1 Z 1
dG dv
a(G, v) = = y (x)v dx = v(y) v H01 (I).
0 dx dx 0

The error in the finite element solution at a nodal point xi is

e(xi ) = u(xi ) uh (xi )

Z 1
= (u uh )xi (x) dx
0
Z 1
d dG
= (u uh ) dx (since u uh V )
0 dx dx
= a(eh , vh ) (letting vh = G)

111
 = 0,

where the last step follows from the error equation. Note that we have used
the fact that G belongs to Vh in the penultimate step. Thus,

uh (xi ) = u(xi ) i = 1, 2, . . . , n,

and we get the result that the finite element solution is exact at the nodes.
In some other kinds of one-dimensional problems, and in two and three-
dimensional problems, we do not, in general, have this property since the
Greens function does not necessarily belong to Vh .

5.3 Convergence of the Finite Element Solu-

tions to the Exact Solution
We define convergence as

a(u uh , u uh ) 0 as h 0,

or, equivalently, using Eqn. 5.1 as

a(uh , uh ) a(u, u) as h 0.

Physically, the above statement means that the strain energy calculated by
the finite element solution converges to the exact strain energy.

5.3.1 Criteria for Monotonic Convergence

For monotonic convergence, the elements must be complete and compatible.
If these conditions are fulfilled, the accuracy of the solution will increase
continuously as the mesh is refined. Mesh refinement is performed by subdi-
viding a previously used element into two or more elements. Thus, the old
mesh is embedded in the new mesh.
Completeness means that the displacement functions of the element must
be able to represent rigid body displacements and constant strains. Rigid-
body displacements are those displacement modes that the element must
undergo as a rigid body without stresses being developed in it. As an ex-
ample, consider the shaded element in Fig. 5.2. This element should remain
stress-free even after the application of the shown distributed load.
The straining and rigid-body modes can be found by solving the eigen-
value problem
K = .
The zero eigenvalues correspond to either rigid-body modes or spurious modes
(such as those generated by using reduced integration).
An element should also be able to model a constant strain state because
in the limit of mesh refinement, the strain in each element approaches a
constant value, and any complex variation of strain can be approximated.

112
 11
00
00
11
00
11
00
11
00
11

111
000
000
111
000
111
000
111
000
111
Figure 5.2: Element undergoing rigid-body displacement.

For example, a plane stress element should be able to represent constant xx ,

xy and yy conditions.
The second requirement for monotonic convergence, viz. compatibility,
means that the displacements within an element and across element bound-
aries must be continuous. Physically, compatibility ensures that no gaps
occur between elements when the assemblage is loaded. For example, in a
3-D element, the u, v and w displacements must be continuous, in a beam
element v and dv/dx must be continuous, and in a plate element based on
the Kirchoff theory w, w/x, w/y must be continuous.
In general if the potential = () contains derivatives of through
order m, then

1. Within each element, the assumed field must contain a complete poly-
nomial of degree m,

2. Across element boundaries, must be continuous through order m 1.

A polynomial is complete if it is of high-enough degree and if no terms are

omitted. For example, in one-dimension, a complete polynomial of degree
2 is a0 + a1 x + a2 x2 . In two-dimensions, a polynomial is of degree n if it
contains a term of the form xl y m , where l and m are nonnegative integers,
and l + m = n, and is complete if all combinations of l and m for which
l + m n are included. For example, a complete polynomial of degree 2 in
two-dimensions is

a0 + a1 x + a2 y + a3 x2 + a4 xy + a5 y 2.

113
 (bilinear) 1
(9 node) Q2 Q1
x y
CST
2 2
x xy y
quadratic triangle
3 2 2 3
x xy xy y
cubic triangle
4 3 2 2 3 4
x xy xy xy y

Figure 5.3: Pascal triangle.

In two dimensions, a complete polynomial of degree n contains (n+1)(n+

2)/2 terms which can be seen from the Pascal triangle shown in Fig. 5.3.
Geometric isotropy requires that there are no preferred directions. Hence,
the displacement expansions for u and v in a two-dimensional problem should
be the same, and should include terms symmetric about the axis of the Pascal
triangle. As an example

u = a1 + a2 x + a3 y + a4 x2 ,
v = a5 + a6 x + a7 y + a8 y 2 ,

is not geometrically isotropic, whereas

u = a1 + a2 x + a3 y + a4 xy,
v = a5 + a6 x + a7 y + a8 xy,

is geometrically isotropic. Similar remarks apply for three-dimensions.

5.4 Rate of Convergence

The rate of convergence depends on the order of the polynomials used in the
displacement assumptions. Let us assume that we employ elements with com-
plete polynomials of degree k, and that the exact solution u to our elasticity
problem is smooth in the sense that u H k+1(), so that kukH k+1 () < ,
where
Z "X 3 3 X 3  2 Xk+1 X 3 X  n ui 2
#
2 2
X ui
kukH k+1 () = ui + + d.
i=1 i=1 j=1
xj n=2 i=1 r+s+t=n
xr1 xs2 xt3

Then the rate of convergence of uh to u is given by

ku uh kH 1 () chk kukH k+1 () , (5.5)

where c is a constant independent of h, but dependent on the material prop-

erties. From Eqn. 5.5, we can say that the rate of convergence is o hk .

114
The dependence of c on the material properties is detrimental when (almost)
incompressible materials are considered because c becomes very large, and
the order of convergence k results in good accuracy only at very small (im-
practical) values of h. The constant c depends on the type of element being
used. For example, the CST and Q4 elements which both belong to the
k = 1 group give different magnitudes of error for a given h, while the order
with which the error decreases as the mesh is refined is the same. Eqn. 5.5
gives in essence an error estimate for the displacement gradient (and hence
for the stresses and strains) because the primary contribution to the H 1 ()
norm is due to errors in the derivatives of displacements. The error in the
displacements is

ku uh kL2 () chk+1 kukH k+1 () .

Thus, the order of convergence for the displacements is one order higher than
for the strains.
These results are intuitively reasonable. If we think in terms of a Taylor
series analysis, a finite element of dimension h with a complete displacement
field expansion of order k can represent displacement variations upto that
order exactly. Hence, the local error in
representing arbitrary displacements
with a uniform mesh should be o hk+1 . The stresses are obtained

by differ-
entiating the displacements, and hence have error of o hk .
In the above convergence study, it is assumed that uniform discretiza-
tions are used (e.g., square elements in 2-D), and that the exact solution is
smooth. If the solution is not smooth, and a uniform mesh is used, the order
of convergence decreases. In practice, graded meshes are used with small
elements in the areas of stress concentration and larger elements away from
those regions. In general, a refined mesh is required in places where acute
changes in geometry, boundary conditions, loading, material properties or
solution occur. A good mesh is one for which the error is almost equal for
all the elements, and below a certain tolerance.
The method just discussed where we increase the number of elements
keeping the interpolation functions fixed is known as h-adaptivity. An al-
ternative method of increasing the accuracy of the solution is the p method
where the order of the interpolation functions in high-error elements is in-
creased using hierarchical shape functions.
A very high rate of convergence in the solution can be obtained if we
increase the number of elements and at the same time increase the order of
interpolation functions. This approach is known as the h-p method. In all
of the above approaches, the mesh grading should be such that the error in
each element is almost the same.
Mesh distortion and numerical integration will not reduce the order of
convergence provided the distortion is reasonable and full numerical integra-
tion is used.

115
 (a) (b)

L
1 2 B
L

incompatibility C
(c)
(d)

Figure 5.4: Compatibility of an element assemblage.

5.5 Convergence of Isoparametric Elements

As we have seen, the two requirements for convergence are compatibility and
completeness. To investigate the compatibility of an element assemblage, we
need to consider each face between adjacent elements. For compatibility, it is
necessary that the coordinates and the displacements of the elements at the
face must be the same. For example, consider the cases shown in Fig. 5.4.
For the bilinear element shown
P in (a), the coordinates
P and displacements
are interpolated as x = N x
i i i and u = N u
i i i . At the shared edge
= 1, we have x . Since, a linear interpolation between two points
can be carried out only in one way, we have compatibility of the geometry
and displacements along the edge. Similarly, for the serendipity element
shown in (b), we have compatibility along the shared edge = 1, since
a quadratic interpolation through three points is unique. In contrast, the
element assemblage shown in (c) is incompatible, since the geometry and
displacements can vary quadratically along the edge in element 1, but only
linearly in element 2. Compatibility can be achieved as shown in (d) by
imposing the constraint equations uB = (uA + uC )/2 and vB = (vA + vC )/2.
Mesh grading can be achieved easily using isoparametric elements. A
one-to-two layer transition is shown in Fig. 5.5.
We now analyze the criterion P of completeness. First, we prove that in
an
P isoparametric
P formulation, P i Ni = 1. Observe that the relations x =
i Ni xi , y = i Ni yi and z = i Ni zi can be used regardless of the location

116
 Figure 5.5: Mesh grading using isoparametric elements.

of the coordinate system. Let xi = Xi + h. Then

X X X X X
x= Ni xi = Ni Xi + h Ni = X + h Ni = x h + h Ni ,
i i i i i
P
which implies that i Ni = 1 since h is nonzero.
Completeness requires that rigid body displacements and constant strain
states should be possible, i.e., the lowest possible interpolation should be
u = a1 + b1 x + c1 y + d1 z,
v = a2 + b2 x + c2 y + d2 z, (5.6)
w = a3 + b3 x + c3 y + d3 z.
The nodal displacements are
ui = a1 + b1 xi + c1 yi + d1 zi ,
vi = a2 + b2 xi + c2 yi + d2 zi , (5.7)
wi = a3 + b3 xi + c3 yi + d3 zi ,
at each node i. The test of completeness is that the displacement field
given by Eqn. 5.6 should be obtained within thePelement when P the nodal
displacements
P are given by Eqn. 5.7. Since u = i Ni ui , v = i Ni vi and
w = i Ni wi , we have
X X X X
u = a1 Ni + b1 Ni xi + c1 Ni yi + d1 Ni zi ,
i i i i
X X X X
v = a2 Ni + b2 Ni xi + c2 Ni yi + d2 Ni zi , (5.8)
i i i i
X X X X
w = a3 Ni + b3 Ni xi + c3 Ni yi + d3 Ni zi .
i i i i
P P P
For an isoparametric formulation
P i Ni xi = x, i Ni yi = y and i Ni zi = z.
Also, we have proved that i Ni = 1. Thus, Eqn. 5.8 agrees with Eqn. 5.6,
and completeness is established.

5.6 Effect of Element Distortion on the Order

of Convergence
We saw that the rate of convergence of uh to u is given by

117
 hm
hm

m
m
Aspect ratio distortion Parallelogram distortion

hm

m
m
hm
Angular distortion
Curved-edge distortion

Midnode distortion

Figure 5.6: Types of element distortion.

118
 10
P
P 2
B
C A
A
xB=0.096 vA =0.091 xC=0.727 vA =0.682

C A
xC=0.301 vA =0.494

Figure 5.7: Effect of mesh distortion in bilinear elements.

ku uh kH 1 () chk kukH k+1 () , (5.9)

where k is the order of polynomial, and c is a constant independent of h but

dependent on material properties. We introduce a measure of mesh regularity
hm
m = ,
m
where hm is the largest dimension of the element, and m is the diameter
of the largest circle or sphere that can be inscribed in m (see Fig. 5.6). A
sequence of meshes is regular if m 0 , where 0 is a fixed positive value.
We also require that for each element, the ratio of the largest to the smallest
side lengths is smaller than a reasonable positive number.
Since all the distortions (except the midnode distortion which is used
only in fracture mechanics problems) shown in Fig. 5.6 are used in mesh
designs (to achieve grading), we need to study the effect of these distortions
on the order of convergence. The rate of convergence given by Eqn. 5.9 is
maintained if the distortion compared to the size of the element is small.
When element sizes are large, geometric distortions can affect the pre-
dictive capabilities to a significant degree as shown in Fig. 5.7 and 5.8. The
values shown in these figures are finite element solutions which have been
normalized with respect to the exact solution. One can see from the results
that geometric distortion makes the results less accurate. From Fig. 5.8,
we see that the 8-node and 9-node results are comparable for undistorted
elements, but that the 9-node element is vastly superior when the elements
are distorted. In fact, the 8-node element even predicts a wrong sign for the
stress at point B with 2 2 integration.
An ideal element is compact, straight-sided and has equal corner angles.
Hence, one should try to minimize the distortion of elements to as large an
extent as possible. The value of the determinant of the Jacobian matrix at
the Gauss points is a good indicator of the amount of distortion. We need
|J | > 0 at all the Gauss points. The loss of predictive capability is due to
the element no longer being able to represent the same order of polynomials

119
 P P
B B
B

B vA B vA
B vA
2x2 1.0 0.968 0.051 0.362 -0.048 0.430
8-node 0.048 0.161 0.050 0.221
3x3 1.129 0.93
1.006 1.125 1.109 0.958 0.955
9-node 2x2 1.00
0.737
3x3 1.141 0.954 0.687 0.791 0.705

Figure 5.8: Effect of mesh distortion in 8-node and 9-node quadrilateral

elements.

a a u u

b M M

u u

Figure 5.9: Parasitic shear in a bilinear element.

in the physical coordinates x, y and z after the geometric distortion as it did

without the distortion. For example, the 8-node element is able to represent
only the terms 1, x and y after geometric distortion, while before distortion,
it can represent 1, x, y, x2 , xy, y 2 , x2 y and xy 2 . Similarly, the 9-node
element can represent only 1, x, y, x2 , xy and y 2 after distortion, while it
can represent 1, x, y, x2 , xy, y 2, x2 y, xy 2 and x2 y 2 without distortion.

5.7 Incompatible Elements and the Patch Test

Bilinear elements are attractive because of their simplicity. Unfortunately,
they are too stiff in bending. Consider a bilinear element subjected to pure
bending as shown in Fig. 5.9. The finite element solution is

u = u; v = 0,

leading to the strains

u u
xx = ; yy = 0; xy = .
a b
The exact displacements and strains are
au bu
u = u; v = (1 2) + (1 2 ) ,
2b 2a
u u
xx = ; yy = ; xy = 0.
a a
120
Comparing the finite element solution with the exact solution, we see that
there is a spurious shear strain xy . For large a/b, the mesh locks.
To overcome the problem of parasitic shear, we improve the Q4 element
by adding additional degrees of freedom:
X
u= Ni ui + (1 2 )a1 + (1 2 )a2 ,
i
X
v= Ni vi + (1 2 )a3 + (1 2 )a4 ,
i

where Ni are the usual bilinear functions. The above element represents
bending exactly, but now the compatibility of displacements at the element
edges is lost. Such elements are called incompatible or nonconforming ele-
ments.
The results obtained using incompatible elements are of high quality.
However, the exact potential energy of the system is not necessarily an upper
bound on the calculated potential energy. Hence, monotonic convergence is
not ensured. But we can establish conditions so that there is at least non-
monotonic convergence. The conditions are
The element completeness condition must always be satisfied, i.e., it
should be able to represent rigid body modes and the constant strain
states
As the finite element mesh is refined, each element should approach a
constant strain condition. Hence, a assemblage of incompatible finite
elements should be able to represent constant strain conditions. Note
that this condition is on an assemblage of elements and not on a single
element.
To test the second condition, the patch test has been proposed. In this test
a patch of elements is subjected to nodal loads consistent with constant xx ,
yy and xy in a two-dimensional problem (or all the six stress components in
a three-dimensional one). If for any patch of elements, the element stresses
actually represent the constant stress conditions, then the element passes the
patch test. If an incompatible element passes the patch-test then convergence
is ensured (although it may not be monotonic). In the case of a compatible
element, we know that the convergence is monotonic. In such a case, the
patch test can be used as a debugging aid.
Two examples of the patch test are shown in Fig. 5.10. The roller supports
are provided since a fixed support would prevent the Poisson contraction in
the y-direction. Actually, we need to test all possible patches. But testing
just one patch is what is practiced. This practice is generally found to be
quite reliable.
One can design variations of the patch test. For example, if an element
fails to display the constant stress state in a patch of elements, but displays
the constant stress state as the mesh is repeatedly subdivided, then a weak
patch test can be designed with a larger number of elements in the assem-
blage being tested. Similarly a higher-order patch test can be designed for

121
 P
P/2

2P

P/2
P

Figure 5.10: Two examples of the patch test.

a 9-node element subjected to pure bending to test whether it gives a linear

stress distribution within the element.

122
Chapter 6

Eigenvalue and
Time-Dependent Problems

6.1 Eigenvalue problems

An eigenvalue problem deals with the determination of the values of the
parameter such that the equation
A(u) = B(u),
where A(u) and B(u) are differential operators, has nontrivial solutions. The
values of are called the eigenvalues and the associated functions u are called
the eigenfunctions. As an example, in the equation
d2 u
= u,
dx2
denotes the square of the frequency of vibration .
In structural problems, eigenvalues denote natural frequencies or buck-
ling loads. In fluid mechanics and heat transfer, eigenvalue problems arise in
connection with the determination of the homogeneous part of the solution.
Eigenvalues denote the amplitudes of the Fourier components making up the
solution. Eigenvalues are also useful in determining the stability character-
istics of temporal schemes.
Consider the axial motion of a bar
2u
 
u
2 E = b, (6.1)
t x x
where b is the axial force per unit length. The natural (i.e., when b = 0)
axial oscillations of the bar are periodic and can be determined by assuming
a solution of the form

u(x, t) = U(x)eit , i = 1, (6.2)
where is the natural frequency and U(x) is the mode shape. Substituting
Eqn. 6.2 into the homogeneous form of Eqn. 6.1, we get
  
2 d dU
U E eit = 0,
dx dx

123
or  
d dU
E U = 0, (6.3)
dx dx
where 2 . Eqn. 6.3 is an eigenvalue problem which involves finding
and U.
For constant E and , the solution is given by
r r

U(x) = c1 sin x + c2 cos x,
E E
yielding " r r #

u(x, t) = c1 sin x + c2 cos x eit .
E E
The constants c1 , c2 and are determined from the initial and boundary
conditions.
Now consider the parabolic partial differential equation
 
T T
c k = Q. (6.4)
t x x
We seek a homogeneous solution to the above equation in the form

T = U(x)et . (6.5)

Substituting Eqn. 6.5 in the homogeneous part of Eqn. 6.4, we get

 
d U
k cU = 0,
dx x
which is of the same form as Eqn. 6.3. Hence, we obtain
" r r #
c c
T = et c1 sin x + c2 cos x .
k k

Similarly, the transverse free vibration of a beam using the Euler-Bernoulli

beam theory is governed by
2v 2 2v
 
A 2 + 2 EI 2 = 0.
t x x
Assuming the solution to be of the form

v = U(x)eit ,

where is the frequency of vibration, we get

d2 d2 U
 
EI 2 = AU,
dx2 dx
which is an eigenvalue problem.

124
 The study of buckling also leads to an eigenvalue problem. The equation
of equilibrium of a beam subjected to an axial force P is

d2 d2 v d2 v
 
EI + P = 0,
dx2 dx2 dx2

which is an eigenvalue problem with = P as the eigenvalue, which rep-

resents the buckling load. The smallest value of is the critical buckling
load.

6.2 Hamiltons principle for dynamics prob-

lems
The equation of motion is given by

2u
= + b.
t2
The virtual work principal is derived in the usual way. Starting from

2u
Z   Z
+ b 2 v d + (t t) v d = 0 v.
t t

Integrating by parts, we get

2u
Z Z Z Z
2 v d + bv d + tv d : (v) d = 0 v. (6.6)
t t

We shall use Eqn. 6.6 to develop the finite element formulation. Just for the
sake of completeness, however, we shall derive Hamiltons principle which we
had used in Section 1.2.5.
Assuming the existence of a strain-energy density function W for the
body so that
W
= ,

the above expression can be written as

2u
Z
2 v d + (1) = 0, (6.7)
t
where Z Z Z
= W d b u d t u d.
t

Now we enforce Eqn. 6.7 which holds at every instant of time t in a weak
sense with respect to time:
Z t2 Z  2  Z t2
u
2 v d dt + (1) dt = 0,
t1 t t1

125
which on interchanging the order of integration yields
Z Z t2 2  Z t2
u (1)
2 v dt d + dt = 0.
t1 t t1

Integrating by parts, we get

t Z Z t2 Z t2
u  2 u v
Z
(1)
v dt + dt = 0. (6.8)
t t1 t1 t t t1

u
Assuming that u is prescribed or t
= 0 at t1 and t2 , Eqn. 6.8 reduces to
t2 t2
(1) u u
Z Z Z
d dt + (1) dt = 0.
t1 2 t t t1

Assuming to be constant, the above equation can be written as

Z t2
(1)
(T ) dt = 0,
t1

where
u u
Z
T = d
2 t t
is the kinetic energy of the body at time t. Alternatively, we can write the
above equation as Z t2
(1)
L dt = 0,
t1

where L = T , which is nothing but Hamiltons principle.

6.2.1 Finite element formulation

As mentioned, we use Eqn. 6.6 to develop the finite element formulation.
Discretizing the space variables (but not the time variable; if only partial
discretization is carried out, the process is known as semi-discretization, if
time is also discretized, then it is known as full-discretization), we get

u = N u + N u,
c (u) = B u + B u,
v = N v, c (v) = B v, (6.9)
u = N u
+ N u,
c (u) = CB u C0c + 0c + C B u,

where the shape function N are functions of the space variables, and the
vectors u are functions of time. It may appear from our semi-discrete ap-
proximate form given by Eqn. 6.9 that this approximation is valid only when
the exact solution is separable in x and t. This is not so. Even nonseparable
solutions can be approximated by our separable approximation. Substituting
Eqn. 6.9 in Eqn. 6.6, we get
+ K u = f ,
M u (6.10)

126
where
Z
M= N t N d,
Z
K= B t CB d,
Z h i Z
f= t t 0 t 0 t t

N b + B Cc B c B C B u N N u d + N t t d.
t

Since any system is not perfectly elastic, the effects of energy dissipation
have to be incorporated in Eqn. (6.10). These effects are incorporated by
modifying Eqn. (6.10) to
+ C u
M u + K u = f , (6.11)

where C is the damping matrix. Eqn. 6.11 is solved subject to the initial

conditions u(0) = u0 and u(0) 0.
= u
The damping matrix, C, is positive definite. To see this in the purely
mechanical theory where thermal effects are ignored, consider the first law
of thermodynamics
Z  
d 1
Z Z
v v + e d = t v d + b v d.
dt 2 t

If one considers the loading to be periodic with period T , then

Z t1 +T Z  Z   t1 +T
1
Z
t v d + b v d dt = v v + e d
t1 t 2 t1
Z t1 +T
= e d
t1
> 0,

since the kinetic energy has the same value at times t1 and t1 + T , and since
the internal energy increases due to dissipation (the contribution to e due to
the elastic strain energy, however, is the same at times t1 and t1 + T ). Now
taking the dot product of the finite element equations
 
= M u
f u + C u
+ K u u
1 d  
+ u K u + u
C u.

= u M u
2 dt
Integrating the above equation between the limits t1 and t1 + T and using
the fact that the kinetic and strain energies are the same at times t1 and
t1 + T , we get Z t1 +T Z t1 +T
C u
u dt = dt > 0.
f u
t1 t1
t1 and T , we conclude
Since the above equation holds for all choices of u,
that C is positive definite.

127
 There is no systematic technique to generate the damping matrix unlike
the mass and stiffness matrices. A simple but popular model that is usually
used is Rayleigh or proportional damping in which we assume

C = M + K,

where and are constants which are determined experimentally. We shall

discuss more about the damping matrix in Section 6.4.1.

6.2.2 Free vibrations

Consider the special case of undamped harmonic motion.
+ K u = 0.
M u (6.12)

Assume a solution for u of the form u = ueit , where u is the set of constant
values at the nodes called as the modal vector, and is known as the natural
frequency. Substituting this solution in Eqn. 6.12, we get

(K 2 M )u = 0.

For a nontrivial solution to the above equation, it is necessary that the de-
terminant of the coefficients of u be zero, i.e.,

det(K 2 M ) = 0.

The eigenvalues are real. To see this, let and the corresponding mode
shape be complex-valued. Then

u K u = 2 u M u,
u K u = ( 2 ) u M u ,

where denotes complex conjugate. Subtracting the first equation from the
second, and using the symmetry of K and M , we get

u M u[ 2 ( 2 ) ] = 0.

Since M = N T N d, it follows that u M u = u u d > 0,

R R

where u = N u. Thus,
2 = ( 2) ,
i.e., 2 is real-valued. Since now uK u = 2 uM u, and since M is positive
definite and K is positive semi-definite, we have 2 0. If the order of K
and M is n, then there are n natural frequencies and correspondingly n
natural mode shapes or eigenvectors.
Assuming that all eigenvalues are distinct, the eigenvectors are orthogonal
in the following sense:

ui K uj = 0, i 6= j,
ui M uj = 0, i 6= j.

128
The proof is as follows. For the i th and j th modes, we have

K ui = i2 M ui ,
K uj = j2 M uj .

Hence,

utj K ui = i2 utj M ui , (6.13)

uti K uj = j2 uti M uj . (6.14)

Since K and M are symmetric matrices, we have utj K ui = uti K uj and

utj M ui = uti M uj . Subtracting Eqn. 6.14 from Eqn. 6.13, we get

0 = (i2 j2 )utj M ui .

For i 6= j, i 6= j , we get
utj M ui = 0.
Hence, from Eqn. 6.13 utj K ui = 0. For i = j, we get

uti M ui = EM ,
uti K ui = i2 EM ,

where EM is a constant. Usually, u is normalized so that EM = 1. Then, we

have

uti M ui = 1,
uti K ui = i2 .

6.2.3 The Rayleigh quotient

The Rayleigh quotient is defined as

v t Kv
Q(v) = ,
vtM v
where v is an arbitrary vector. One of the important properties of the
Rayleigh quotient is
12 Q(v) n2 ,
where the frequencies 1 , 2 ,..., n are assumed to be ordered such that

1 2 . . . n .

The proof is as follows. Any vector v can be expressed in terms of the

normalized eigenvector ui as
n
X
v= i ui .
i=1

129
Substituting this expansion in the expression for Q(v), and using the orthog-
onality property of u, we get

12 12 + 22 22 + + n2 n2
Q(v) =
2 + 22 + + n2
1  2  2
2 2 2 2 n
1 + 2 1 + + n 1
= 12
12 + 22 + + n2

12 .

Similarly,
 2  2
2 1 2
1 n
+ 22 n
++ n2
Q(v) = n2
12 + 22 + + n2

n2 .

Thus, we get 12 Q(v) n2 . Choosing v = 1 u1 and v = n un , we get

12 = minn Q(v),
v
n2 = maxn Q(v).
v

The maximum eigenvalue can be estimated from the result

hn n2 max(max
2
)e ,
e

2
where (max )e is the maximum eigenvalue of element e found by solving the
eigenvalue problem for that element:

(K (e) e M (e) )u(e) = 0.

6.2.4 Consistent mass matrices for various element types

In this section, we shall evaluate the consistent element mass matrix given
by Z
(e)
M = N t N d, (6.15)
e

for the various element types that we have studied so far. We shall assume
that = constant.

One-dimensional bar element

The shape function matrix for a linear element is given by
 
1 1+
N= .
2 2

130
Hence, the consistent element mass matrix is
Z
(e)
M = e N t N A dx
e
e Ae le 1 t
Z
= N N d
2 1

e Ae le 2 1
= .
6 1 2

For a quadratic bar element, we get

4 2 1
e Ae le
M (e) = 2 16 2 .

30
1 2 4

Truss element
If u1 -u4 represent the degrees of freedom of a truss element, then

u
1
u N 0 N2 0 u2

= 1 ,
v 0 N1 0 N2 u3


u4

thus yielding
N1 0 N2 0
N = ,
0 N1 0 N2
where N1 = (1 )/2 and N2 = (1 + )2. Substituting in Eqn. 6.15, we get

2 0 1 0

0 2 0 1

e Ae le
M (e) = .
6 1
0 2 0


0 1 0 2

Beam element
Using the Hermite shape function matrix, i.e.,
 
le le
N = N1 N2 N3 N4 ,
2 2

131
where
1
N1 = (2 3 + 3),
4
1
N2 = (1 2 + 3 ),
4
1
N3 = (2 + 3 3),
4
1
N4 = (1 + 2 + 3 ).
4
Substituting in Eqn. 6.15, we get
Z 1
(e) e Ae le
M = N tN d
1 2

156 22le 54 13le

2 2
22le 4le 13le 3le

e Ae le
=
420 54 13le 156 22le


2 2
13le 3le 22le 4le

Frame element
In the body coordinate system, the mass matrix is a combination of the bar
element and the beam element:

2a 0 0 a 0 0

0 156b 22le b 0 54b 13le b


2 2
0 22l b 4l b 0 13l b 3l b

(e)
e e e e
Mb =
a 0 0 2a 0 0


0 54b 13le b 0 156b 22le b


2 2
0 13le b 3le b 0 22le b 4le b
where a = e Ae le /6 and b = e Ae le /420. In the global coordinate system,
the element mass matrix is given by
(e)
M (e) = Lt M b L,
where
l m 0 0 0 0

m l 0 0 0 0


0 0 1 0 0 0

L=
.
0 0 0 l m 0


0 0 0 m l 0


0 0 0 0 0 1

132
CST element
The shape function matrix is given by

N1 0 N2 0 N3 0
N = ,
0 N1 0 N2 0 N3

where N1 = = A1 /A, N2 = = A2 /A and N3 = 1 = A3 /A. The

mass matrix is

2 0 1 0 1 0

0 2 0 1 0 1


1 0 2 0 1 0

te Ae
Z
(e) t
M = e te N N dA = .
A 12 0 1

0 2 0 1


1 0 1 0 2 0


0 1 0 1 0 2

Axisymmetric triangular element

The mass matrix is
Z
M = 2e N t N r dA
(e)

ZA
= 2e (N1 r1 + N2 r2 + N3 r3 )N t N dA
A

4r1
+ 2r 0 2r r33 0 2r r2
0
3 3
4r1
0 + 2r 0 2r r33 0 2r r2

3 3


2r r33 0 4r2
+ 2r 0 2r r1
0

e Ae 3 3

= ,
10 0

2r r33 0 4r2
+ 2r 0 2r r1
3 3


2r r32 0 2r r31 0 4r3
+ 2r 0

3


0 2r r32 0 2r r31 0 4r3
3
+ 2r

where r = (r1 + r2 + r3 )/3.

Q4 element
The shape function matrix is given by

N1 0 N2 0 N3 0 N4 0
N = ,
0 N1 0 N2 0 N3 0 N4

133
 2
A=(1+ /40) cm 2

A=1 cm 2

100 cm 80 cm

Figure 6.1: Calculation of the natural frequencies of a structure

where Ni = (1 + i )(1 + i )/4. The mass matrix is

Z 1Z 1
(e)
M = e te N t N (det J ) d d.
1 1

The integration has to be performed numerically. Usually the same order of

integration as for the stiffness matrix is appropriate.
Note: The sum of the elements of the mass matrix for the bar element
is equal to the mass of the element. For the truss, CST, axisymmetric and
Q4 elements the sum of the elements of the mass matrix is equal to twice
the mass of the element since the linear momentum equation is a vectorial
equation with components along each of the coordinate axes.

Example
Consider the structure shown in Fig. 6.1 with the material properties E =
60 109 Pa and = 7500 kg/m3 . Using two linear elements to model the
structure, the mass and stiffness matrices are

0.0025 0.00125 0

2
M = 10 0.00125 0.0073 0.0042


0 0.0042 0.0128

6 6 0

6
K = 10 6 38.5 32.5


0 32.5 32.5

We have u1 = 0. Hence, the eigenvalue problem corresponding to u2 and u3

is solved using
 
 38.5 (7.3 107 ) 32.5 (4.2 107 )
 
  = 0.
7 6
32.5 (4.2 10 ) 32.5 (1.28 10 ) 
 

8 6
On solving the
above equation, we get 1 = 1.3035
10 and 2 = 1.97 10 .
Thus, 1 = 1 = 11417.3 rad/s and 2 = 2 = 1404.854 rad/s. The

134
 0.0006169
0.000555

0.000763
100
180
0.001175

Figure 6.2: Mode shapes

frequency in Hertz is obtained by dividing by 2. We get f1 = 1817 Hz

and f2 = 223.6 Hz. Substituting 1 in the eigenvalue problem, we get

0.56659 0.87249 u 0
108 2 = ,
0.87249 1.2835 u3 0

which on solving yields u2 = 1.5399u3 . As mentioned, the normalization of

the mode shape is done with respect to the mass matrix, i.e., ut M u = 1, or

0.25 0.125 0 0
h i
0 1.5399u3 u3 0.125 0.73 0.42 1.5399u3 = 1.



0 0.42 1.28 u3

On solving the above equation, we get u3 = 0.76304 m. Thus, u2 =

1.175 m. Similarly, corresponding to 2 , we get u2 = 0.5549 m and u3 =
0.6171 m. The two mode shapes are shown in Fig. 6.2.

6.3 Unsteady Heat Conduction

In this section, we consider the transient heat conduction equation (compare
with the steady state treatment in Section 1.5)
T
c = (kT ) + Q in ,
t
where c is the specific heat (J/kgC) and is the density, subject to the
boundary conditions

T = T on T ,
h(T T ) + (kT ) n = q on q ,

and subject to the initial condition

T (x) = T0 (x) at t = 0.

The weak form of this equation is given by

Find T LT such that

135
 T
Z Z Z Z
t
c v d + v kT d + vhT d = vQ d+
t q
Z
v(q + hT ) d v VT ,
q

subject to the initial condition T (x, 0) = T0 (x). The finite element formu-
lation is carried out similar to that in the elasticity case. Discretizing the
temperature as X
T = N (x)u(t) + N (x)u(t), (6.16)
and substituting in the variational formulation, we get

M u + Ku = f
subject to u(0) = u0 ,

where (with T = Bu)

Z
M= cN t N d,
Z Z
t
K= B KB d + hN t N d,
q
Z Z Z  Z
t t t
f= N Q d + N (q + hT ) d B kB d u cN t N u
d,
q

Note the similarity of the form of M with the form of the mass matrix
(also denoted by M ) in the dynamics formulation. In particular it is sym-
metric and positive definite. In practice, diagonal or lumped mass matrices
are used instead of the consistent mass matrix derived above, because they
lead to economical time-integration schemes (so called explicit methods).
The eigenvalue problem can be derived by assuming

u = uet ,

and substituting in M u + Ku = 0, to get

(K M )u = 0,

which is similar in form to the eigenvalue problem for the dynamics problem.
Example:
Consider a plane wall initially at uniform temperature T0 , which is suddenly
exposed to a fluid at temperature T . The governing equation is

2T T
2
= ,
x t
where = k/(c) is the diffusion coefficient. The initial condition is T (x, 0) =
T0 . Consider two sets of boundary conditions:

1. T (0, t) = T , T (L, t) = T for t > 0.

136
 
2. T (0, t) = T , kA T
x
+ hA(T T )x=L = 0.
To find the solution, we define some nondimensional parameters:
x t T T
x = ; t = 2 ; u = .
L L T0 T
The differential equations and boundary conditions in terms of these param-
eters are (dropping the bars on x and t for notational convenience)
2 u u
+ = 0,
x2 t
1. u(0, t) = 0, u(1, t) = 0, u(x, 0) = 1.

2. u(0, t) = 0, u
x
+ Hux=1 = 0, u(x, 0) = 1, where H = hL/k.
Assume u = et U(x). Substituting in the governing differential equation,
we get
d2 u
2 U = 0.
dx
Solving the above equation, we get
h i t
u = c1 sin x + c2 cos x e .

1. For the first set of boundary conditions, we get c2 = 0, = n. Thus,
the exact solution is
X
2 2
u= [cn sin nx] en t .
n=1

The constants cn can be determined from the initial condition, and the
orthogonality of the sine functions:

X
1= cn sin nx
n=1

For a mesh of 2 linear elements, we have

1 1 0 2 1 0 Q
1



2 1 2 1 1 4 1 u = 0 ,


12
0 1 1 0 1 2 Q2

subject to the boundary conditions u(0) = u(1) = 0. Solving, we get

4
(4 )u2 = 0,
12
or 1 = 12.
For a mesh of one quadratic element, we get
16 16
= 0,
3 30
or 1 = 10. Summarizing, we have

137
 for 2 linear elements, 1 = 12.
for 1 quadratic element, 1 = 10.
exact, 1 = 2 .

The corresponding eigenfunctions are

for 2 linear elements,

(
x
h
0 x 0.5
U(x) = x
1 h
0.5 x 1

for 1 quadratic element,

x x
U(x) = 4 1 .
h h
exact, Un (x) = sin nx.

2. Assuming
H = 1, the second set of boundary conditions yields Un (x) =
sin n x, where n satisfies
p p
1 + n cot n = 0.

For a mesh of two linear elements, we get

4 2 4 1 u2 0
= .
2 3 12 1 2 u3 0

On solving, we get 1 = 4.49, 2 = 36.65. The first eigenmode is

U (1) = (0, 0.6881, 0.7256). The exact eigenvalues are 1 = 4.1159 and
2 = 24.1393.
Note that as the mesh is refined, not only do we increase the number of
eigenvalues, but also improve the accuracy of the preceding ones. Note also
that the convergence of the numerical eigenvalues to the exact ones is from
above, i.e., the finite element solution provides an upper bound to the exact
eigenvalues because the finite element system is stiffer resulting in higher
eigenvalues or frequencies.

6.3.1 Algorithms for parabolic problems

In this section, we discuss algorithms for solving the governing equation

M u + Ku = f , (6.17)

subject to the initial condition u(0) = u0 . First we state the exact solution
to these set of equations. Multiplying Eqn. (6.17) by eM Kt M 1 , we get
1

d h M 1 Kt i
u = eM Kt M 1 f .
1
e
dt
138
Integrating this equation either between the limits [0, t] or [t, t + t], we get
Z t
M 1 Kt
eM K(t) M 1 f () d,
1
u(t) = e u0 + (6.18a)
0
Z t+t
M 1 Kt
eM K(tt) M 1 f () d. (6.18b)
1
u(t + t) = e u(t) +
t

The trapezoidal family of methods to solve Eqn. 6.17 is given by

M v n+1 + Kun+1 = f n+1 , (6.19)

M v n + Kun = f n , (6.20)
un+1 un
= (1 )v n + v n+1 . (6.21)
t
where un and v n are approximations to u(tn ) and u(tn ), f n+1 = f (tn+1 ),
t is the time step and [0, 1].

Method

0 Forward differences; forward Euler

1
2
Trapezoidal rule; midpoint rule; Crank-Nicolson
1 Backward differences; backward Euler

The problem is to determine un+1 and v n+1 given un and v n . The initial
value v 0 is determined from

M v 0 = f 0 Ku0 .

By multiplying Eqn. (6.19) by , Eqn. (6.20) by 1 , adding the two

resulting equations, and using Eqn. (6.21), we get

(M + tK)un+1 (M (1 )tK)un = tf n+ , (6.22)

where f n+ = f n+1 + (1 )f n . After obtaining un+1 from Eqn. (6.22),

v n+1 is recovered in the post-processing step using Eqn. (6.21).
Note

1. When = 0, the method is said to be explicit. In this case, Eqn. (6.22)

reduces to
M un+1 = tf n + (M tK)un .
If M is lumped (i.e., diagonal), the solution for un+1 can be found
without the necessity of equation solving.

2. If 6= 0, the method is said to be implicit. In this case, a system of

equations needs to be solved. If t is constant, only one factorization
of (M + tK) needs to be carried out at t = 0.

139
 We now show how a form close to Eqn. (6.22) can be obtained as an
approximation to the exact solution given by Eqn. (6.18b). In the derivation,
we will use the property that eA+B = eA eB if and only if A and B commute.
First multiply Eqn. (6.18b) by eM Kt to get
1

Z t+t
M 1 Kt (1)M 1 Kt
eM K(t(1)t)
M 1 f () d.
1
e un+1 = e un +
t

Make the substitution h = t (1 )t in the integral to get

Z t
M 1 Kt (1)M 1 Kt
eM Kh
M 1 f (h) dh. (6.23)
1
e un+1 = e un +
(1)t

Assume next that f is approximated by a linear interpolation between the

values at t and t + t, i.e.,
1  
f (h) = (t h)f n + [h + (1 )t]f n+1 . (6.24)
t
Since t and h are small, we approximate the exponential matrices as

eM Kt
I + tM 1 K,
1

e(1)M Kt
I (1 )tM 1 K,
1

eM Kh
1
I,

where in the last approximation, we have retained only the first term to get
a linear approximation in t in the final result. Substituting Eqn. (6.24) and
the above approximations into Eqn. (6.23), and carrying out the integration,
we get
1
I + tM 1 K un+1 = I (1 )tM 1 K un + tM 1 f n + f n+1 .
     
2
Multiplying the above equation by M , we obtain a form close to Eqn. (6.22).

6.3.2 Analysis of the generalized trapezoidal method

An algorithm is convergent if for tn fixed and t = tn /n

un u(tn ) as t 0.

To establish the convergence of an algorithm, two additional notions must

be considered: stability and consistency. We shall prove that stability and
consistency implies convergence. We use the modal approach (also known
as spectral or Fourier analysis) in which the problem is decomposed into n
uncoupled scalar equations. The first step in the analysis is to perform the
reduction to single-degree-of-freedom (SDOF) form.

140
Modal reduction to SDOF form
The essential property used in reducing to SDOF form is the orthogonality
of the eigenvectors of the associated eigenvalue problem

(K i M )i = 0; i = 1, 2, . . . , n

where n is the number of degrees of freedom and the i are assumed to be

ordered such that 0 1 . . . n . By the orthonormality properties
proved in Section 6.2.2, we have
ti M j = ij ,
(6.25)
ti Kj = i ij ,

with no sum on i. The eigenvectors i |ni=1 constitute a basis for n , i.e., any
element of n can be written as a linear combination of i |ni=1 . Therefore,
n
X
u(t) = uj (t)j ,
j=1
n (6.26)
X
u(t) = uj (t)j ,
j=1

where

ui (t) = ti M u(t),
ui (t) = ti M u(t).

Substituting Eqn. 6.26 in Eqn. 6.17, and premultiplying by ti yields

n
X
(uj ti M j + uj ti Kj ) = ti f i = 1, 2, . . . , n.
j=1

Using Eqns. 6.25, we get

ui + i ui = ti f (t), i = 1, 2, . . . , n.

We shall denote ti f (t) by fi . The initial condition is

ui (0) = ti M u(0) = ti M u0 .

We shall denote ti M u0 by u0(i) . Omitting the subscript i, the typical modal

initial-value problem is
u + u = f (6.27)
subject to u(0) = u0 .
Similar to the decomposition of the semi-discrete equation, we now de-
compose the generalized trapezoidal algorithm using
n
X
un = un(j) j ,
j=1

141
 M u + Ku = f
temporal (M + tK)un+1 = (M (1 )tK)un + tf n+
u(0) = u0
discretization u0 given

modal modal
decomposition decomposition

u + u = f
temporal (1 + t)un+1 = (1 (1 )t)un + tfn+
u(0) = u0
discretization u0 given

Figure 6.3: Temporal discretization and modal decomposition for parabolic

problems

n
X
un+1 = un+1(j) j ,
j=1

where

un(i) = ti M un ,
un+1(i) = ti M un+1 .

Substituting for un and un+1 in Eqn. 6.22, and premultiplying by ti , we get

n
X
un+1(j) ti (M + tK)j un(j)ti (M (1 )tK)j = tti f n+ ,
 
j=1

where f n+ = (1 )f n + n+1 . Using Eqn. 6.25, we get the modal equation

as
(1 + ti )un+1(i) = (1 (1 )ti )un(i) + tfn+(i) ,
subject to the given initial condition u0(i) = ti M u0 . Omitting the subscript
i, we get
(1 + t)un+1 = (1 (1 )t)un + tfn+ , (6.28)
with u0 given.
Note that applying the trapezoidal rule to the SDOF equation given
by Eqn. 6.27 also results in the temporally discretized SDOF problem (see
Fig. 6.3).
The convergence of the original problem is equivalent to the convergence
of the SDOF problem, i.e.,

un u(tn ) un(i) ui (tn ) for i = 1, 2, . . . , n.

142
The proof is as follows. The errors in the original and SDOF system at time
tn are given by

e(tn ) = u(tn ) un ,
ei (tn ) = ui (tn ) un(i) .

Hence,
n
X
t
e (tn )M e(tn ) = (ei (tn )i )t M (ej (tn )j )
i,j=1
X n
= ei (tn )ej (tn )ti M j
i,j=1
X n
= ei (tn )ej (tn )ij
i,j=1
X n
= [ei (tn )]2 .
i=1

Hence,
et (tn )M e(tn ) 0 |ei (tn )| 0 i.
Since M is assumed to be positive definite,

et (tn )M e(tn ) 0 e(tn ) 0,

and the theorem is proved.

Stability
If u now denotes a perturbation of the solution, then noting that Eqn. (6.27)
is linear, we get
u + u = 0.
Since the solution of this equation is given by cet , we have u(tn ) = cetn
and u(tn+1 ) = cetn+1 . Thus,

u(tn+1 ) = e(tn+1 tn ) u(tn ).

For the solution to remain stable, i.e., the condition that the perturbation
not blow up, we need
|u(tn+1 )| < |u(tn )| , > 0,
(6.29)
u(tn+1 ) = u(tn ), = 0.
These are the conditions that we wish to mimic in the temporally discrete
case.
If un and un+1 denotes the perturbations in the finite element solution at
times tn and tn+1 , then using Eqn. (6.28), we have

(1 + t)un+1 = (1 (1 )t)un .

143
Note that (1 + t) > 0 since and t are both greater than zero. We
can write the above equation as
un+1 = Aun , (6.30)
where
1 (1 )t
A= ,
1 + t
is the amplification factor. To mimic Eqn. 6.29, we need
|un+1| < |un | , > 0,
un+1 = un , = 0.
From the definition of A, we see that the second condition is automatically
satisfied. The first condition implies |A| < 1, or 1 < A < 1, i.e.,
1 (1 )t
1 < < 1.
1 + t
The right-hand inequality is automatically satisfied. The left-hand inequality
is satisfied whenever 0.5. If < 0.5, then
2
t < .
1 2
An algorithm for which stability imposes a time-step restriction is called
conditionally stable. An algorithm for which there is no time-step restriction
imposed by stability is called unconditionally stable. Even if the amplifica-
tion factor, A, is slightly greater than 1, disastrous growth can occur. For
example, for |A| = 1.01 and n = 1000, we have
un
= An = 1.011000 = 2.09 104 .
u0
Note the following points:
1. In the conditionally stable case, the stability condition
2
t < ,
(1 2)
must hold for all modes i , i = 1, 2, . . . , n. Hence, the greatest i , i.e.,
n imposes the most stringent restriction on the time step, i.e.,
2
t < .
(1 2)n
In heat transfer problems, it can be shown that n = o (1/h2 ), where
h is the typical element size. Thus, we need t < ch2 , where c is
some constant. This, is a very severe constraint on the step size; thus,
unconditionally stable algorithms are generally preferred.
2. For = 0.5 and t 1, A 1. Thus, high modal components
will behave like (1)n (sawtooth pattern). These spurious modes are
filtered out by reporting step-to-step averages , (un+1 + un )/2.

144
Convergence
The temporally discrete model problem can be written as

un+1 Aun Ln = 0, (6.31)

where
tfn+
Ln = .
1 + t
If we replace un and un+1 by the exact values, we get

u(tn+1 ) Au(tn ) Ln = t (tn ), (6.32)

where (tn ) is the local truncation error. If

| (t)| ctk t [0, T ],

where c is a constant independent of t, and k > 0, then the algorithm

defined by Eqn. 6.31 is consistent, and k is called the order of accuracy or
rate of convergence.
The generalized trapezoidal methods are consistent. For [0, 1], k = 1,
except = 0.5, when k = 2. The expression for is

= (1 2)o (t) + o t2 .

Thus, the trapezoidal rule ( = 0.5) is the only member of the family of
methods that is second-order accurate.
We have the following result:
Theorem: Consider Eqns. 6.31 and 6.32. Let tn be fixed (n, and hence
t = tn /n are allowed to vary). Assume that the following conditions hold
1. |A| 1 (stability)

2. | (t)| ctk , t [0, T ], k > 0 (consistency).

Then e(tn ) 0 as t 0.
Proof: Subtract Eqn. 6.32 from 6.31 to get

e(tn+1 ) = Ae(tn ) t (tn ). (6.33)

Form Eqn. 6.33 (known as the error equation) it follows that

e(tn ) = Ae(tn1 ) t (tn1 ).

Thus,

e(tn+1 ) = A2 e(tn1 ) tA (tn1 ) t (tn )

= A3 e(tn2 ) tA2 (tn2 ) tA (tn1 ) t (tn )
Xn
n+1
= A e(0) t Ai (tni ).
i=0

145
The first term on the right vanishes since e(0) = u0 u(0) = 0. Thus,
 n1 
X 
|e(tn )| = t  Ai (tn1i )
 
 
i=0
n1
X
t |A|i | (tn1i )|
i=0
n1
X
t | (tn1i )| (by stability)
i=0
tn max | (t)| (t = tn /n)
tn ctk (consistency)
Hence, e(tn ) 0 as
t 0, and furthermore the rate of convergence is k,
i.e., e(tn ) = o tk . The maximal rate of convergence is 2.
The theorem just proved is a particular example of the Lax equivalence
theorem which states that consistency plus stability is necessary and sufficient
for convergence.

6.3.3 Modal Analysis

Modal analysis is an alternative technique to the step-by-step integration
procedure already described. The steps involved are
1. Solve the eigenvalue problem
(K i M )i = 0 i = 1, 2, . . . , neq ,
where 0 1 2 . . . n and
ti M j = ij
ti Kj = i ij (no sum on i).
for the pairs {i , i }, 1 i nmodes , where nmodes neq is the desired
number of modes to participate in the subsequent calculations. Two
important factors to be kept in mind are
(a) i and i should be good approximations of the exact eigenvalues
and eigenvectors.
(b) The spatial variation of u and f must be adequately represented
by the expansion upto nmodes .
2. Solve the ordinary differential equations
ui + i ui = fi 1 i nmodes ,
where ui (0) = u0(i) . Solutions to the above equations can be obtained
easily if f (t) is a simple function of time, or by a step-by-step method
using a very small time step. Since the problem in step 2 is of a scalar
nature, the computational cost in step 2 is negligible compared to step
1.

146
 3. The approximation to u(t) is synthesized from the modes, viz.,
nX
modes

u(t) ui (t)i .
i=1

The calculation in the above equation must be performed for each time
t at which the solution is required.

To summarize, the features of the step-by-step method are

1. they are easily coded

2. they are more efficient for short-time calculations

3. they are generalizable to nonlinear problems

while those of the modal methods are

1. they are efficient if many analyses of the same configuration are re-
quired.

2. for long-time calculations

3. if only a small number of modes are participating in the solution.

6.4 Algorithms for Hyperbolic Problems

We have seen in Section 6.2.1 that the semi-discrete equation of motion is
+ C u
M u + K u = f , (6.34)

where M is the mass matrix, C is the viscous damping matrix, K is the

stiffness matrix, and u, u, and u represent the displacement, velocity and
acceleration vectors. The initial conditions are given to be u(0) = u0 and
v(0) = v 0 . The exact solution can be found by converting Eqn. (6.34) into
two sets of first-order equations. Let v = u, so that Eqn. (6.34) can be
written as M v + C v + K u = f . These two sets of first-order equations can
be written as
M u + K u = f ,
where

I 0 0 I u 0
M = , K = , u = , f = .
0 M K C v f

From Eqns. (6.18), it directly follows that

Z t
1
M Kt
1 1
u(t) = e u0 + eM K(t) M f () d, (6.35a)
0

147
 Z t+t
M
1
Kt
1 1
u(t + t) = e u(t) + eM K(tt)
M f () d, (6.35b)
t

where
1 0 I u0
M K = , u0 = .
M 1 K M 1 C v0
We now discuss the modal analysis and direct methods.

6.4.1 Modal analysis

Modal analysis is preferred for problems of structural dynamics, since fre-
quencies and mode shapes usually need to be computed for design purposes.
Hence, the most costly constituent in modal analysis, which is the solution
of the eigenproblem is effectively taken care of. If one were to calculate all
the eigenvectors, then the modal analysis method would be prohibitively ex-
pensive. The reason why, often, modal analysis is cost-effective is that in
many problems, only a small number of low-frequency modes participate in
the structural response. Hence, only the first few eigenvectors need to be
computed. Thus, modal analysis effectively damps out the higher modes.
Let the normalized eigenvectors i in the eigenvalue problem

Ki = i2 M i

be arranged as columns in the matrix . Then as discussed in Section 6.2.2,

t M = I,
t K = ,

where is a matrix with the eigenvectors i2 along the diagonal and all
other elements zero. Since the vectors i form a basis, we can write the
displacement vector, u as
u = u.
Since is independent of time, we can write Eqn. (6.34) as

M u + Cu + Ku = f .

If we premultiply the above equation by t , then we get

u + t Cu + u = f ,

where f = t f . A considerable simplification, namely, uncoupling of the

above equations takes place if we assume the damping matrix to be orthog-
onal, i.e.,
t C = ,
where is a diagonal matrix. Now one has a set of n uncoupled ordinary
differential equations
u + u + u = f ,

148
which are to be solved subject to the initial conditions u(0) = t M u0 and
u(0) = t M u 0 . Based on the analogy with a sping-mass-damper system,
each diagonal element of is conventionally written as 2i i , where i > 0
is known as the modal damping parameter. The damping ratio = 1 marks
the transition between oscillatory and nonoscillatory response, and is hence
known as the critical damping ratio. Damping ratios are found experimen-
tally by observing the vibratory response of the structure. Typically, ranges
between 0.005 to 0.15.
If the modal damping parameters (and hence ) is known, then the damp-
ing matrix can be constructed explicitly as

C = M t M .

One can easily verify from the above equation that t C = . However,
since computing the matrix is quite an expensive proposition, one usually
assumes Rayleigh or proportional damping, i.e.,

C = M + K,

where and are determined as follows. Premultiplying and postmultiplying

the above equation by t and respectively, we get

+ i2 = 2i i i = 1, 2, ..., n. (6.36)

Obviously the above equation cannot hold, in general, for all values of i
since there are only two undetermined constants and . Hence, the above
equations are solved for two given damping ratios (which are determined
experimentally) corresponding to two distinct natural frequencies 1 and 2 .
We get

21 2 (1 2 2 1 )
= ,
22 12
2(2 2 1 1 )
= .
22 12

Once and are determined, the remaining modal damping parameters are
determined from Eqn. (6.36), though these will in general, not match with
the actual damping parameters of the system. Note that in this approach C
is banded, whereas in other approaches, it might be fully populated.
One usually chooses 1 to be the lowest natural frequency and 2 to be
the maximum frequency of interest in the loading or response. Choosing 1
to be the lowest natural frequency guarantees that C is positive definite. To
see this, let a be some arbitrary vector. We want to prove that at Ca > 0.
Since the columns of act as a basis, we can express a as b, where b is
another vector. Thus, if b (b1 , b2 , ..., bn ) then

at Ca = bt t (M + K)b
= bt (I + )b

149
 = ( + 12)b21 + + ( + n2 )b2n
> 0,

where the last step follows from the fact that

+ n2 + n1
2
+ 12 = 21 1 > 0.

Now we describe the modal analysis method. The steps are

Solve the eigenvalue problem

(K i2 M )i = 0 i = 1, 2, . . . , neq ,

where 0 1 2 . . . n and

ti M j = ij
ti Kj = i2 ij (no sum on i).

for the pairs {i , i }, 1 i nmodes , where nmodes neq is the desired

number of modes to participate in the subsequent calculations.

Solve the ordinary differential equations

ui + 2i i ui + i2 ui = fi 1 i nmodes ,

where ui (0) = u0(i) and ui (0) = u0(i) . Solutions to the above equations
can be obtained easily if f (t) is a simple function of time, or by a step-
by-step method using a very small time step. Since the problem in step
2 is of a scalar nature, the computational cost in step 2 is negligible
compared to step 1.

The approximation to u(t) is synthesized from the modes, viz.,

nX
modes

u(t) ui (t)i .
i=1

The calculation in the above equation must be performed for each time
t at which the solution is required.

6.4.2 Direct methods

We now describe an energy-momentum conserving method for the direct
solution of Eqn. (6.34), which is identical to the trapezoidal scheme in the
Newmark method. We first ignore damping (C = 0). The balance of linear
and angular momenta, and the balance of energy (in the context of linear
elasticity) are given by

d
Z Z Z
v dV = t dS + b dV,
dt V S V

150
 d
Z Z Z
(X v) dV = X t dS + X b dV,
dt V S V
d vv
Z h i Z Z
+ W () dV = t v dS + b v dV,
dt V 2 S V

where W () = : C/2 is the strain-energy density function. In deriving

the last equation, we have used the fact that D in the context of linear
elasticity, so that : D = W/t. Thus, in the absence of any displace-
ment constraints and in the absence of loading, i.e., when t = b = 0, the
linear momentum, angular momentum, and total energy (i.e., kinetic+strain
energy) are conserved, i.e.,
Z
v dV = constant,
V
Z
(X v) dV = constant,
V
vv
Z h i
+ W () dV = constant,
V 2
at all times.
We want to devise a numerical strategy that mimics the above continuum
behavior. Let un , v n and f n denote the nodal displacement, velocity and
load vectors, respectively, at time tn , and let t = tn+1 tn . We propose the
following time-stepping strategy over the time-interval [tn , tn+1 ]:
   
v n+1 v n un + un+1 f + f n+1
M +K = n , (6.37)
t 2 2
where
un+1 un v n + v n+1
= . (6.38)
t 2
In the absence of loading, i.e., when f n = f n+1 = 0, Eqn. (6.37) can be
written as
   
v n+1 v n (un ) + (un+1 )
Z Z
u dV + (u ) : C dV = 0 u .
V t V 2
(6.39)
We now make special choices of u in Eqn. (6.39) in order to prove the
conservation properties of the algorithm. First choose u = c over the entire
domain, where c is a constant vector. This choice is permissible since the
entire boundary is assumed to be free of displacement constraints. For this
choice, (u ) = 0, so that we get
Z  
v n+1 v n
c dV = 0.
V t
Since c is arbitrary, we get
Z Z
v n+1 dV = v n dV,
V V

151
which proves that the linear momentum is conserved.
Now choose u = c X over the entire domain, where c is a constant
vector. If W is the skew tensor whose axial vector is c, then we have u =
W X, so that u = W , again resulting in (u ) = 0. Using the property
(p q) r = p (q r), Eqn. (6.39) reduces to
Z
c X (v n+1 v n ) dV = 0,
V

which, by virtue of the arbitrariness of c, leads to

Z Z
X v n+1 dV = X v n dV,
V V

i.e., the angular momentum is conserved.

Finally to show that the total energy is conserved, choose u = un+1 un ,
so that (u ) = n+1 n . Eqn. (6.39) reduces to
   
v n+1 v n n+1 + n
Z Z
(un+1 un ) dV + (n+1 n ) : C dV = 0.
V t V 2

Using Eqn. (6.38), the above equation reduces to

1 1
Z Z
(v n+1 + v n ) (v n+1 v n ) dV + (n+1 n ) : C(n+1 + n ) dV = 0,
2 V 2 V
which, by virtue of the symmetry of the material constitutive tensor C, re-
duces further to
1 1
Z Z
[v n+1 v n+1 v n v n ] dV + [n+1 : Cn+1 n : Cn ] dV = 0.
2 V 2 V
(6.40)
Thus, the total energy is conserved:

[K.E. + Strain energy]n+1 = [K.E. + Strain energy]n .

Note that the choices for u made above fall within the finite element space
for u, and hence the conservation properties hold in the fully-discrete setting.
By eliminating v n+1 from Eqns. (6.37) and (6.38), we get
 
2M K 2 2 1 1 
+ un+1 = 2 M un + M v n K un + f n + f n+1 . (6.41)
t2 2 t t 2 2
Using the initial displacement and velocity vectors, u0 and v 0 , we first solve
for the nodal displacement vector ut1 at time t1 using Eqn. (6.41). Next,
we substitute for ut1 in Eqn. (6.38) to find v t1 . Using ut1 and v t1 , we find
(ut2 , v t2 ), and so on. Thus, starting from the initial displacement and velocity
fields, we march forward in time, until the time instant at which the response
is desired is reached.
Now consider a nonzero damping matrix C, which is assumed to be
positive semi-definite (e.g., C = M + K). Using the approximation

152
 (un+1 un )/t in Eqn. (6.34), we now get instead of Eqn. (6.40) the
u
relation
1 1
Z Z
[v n+1 v n+1 v n v n ] dV + [n+1 : Cn+1 n : Cn ] dV
2 V 2 V
1
+ (un+1 un ) C(un+1 un ) = 0,
t
which by virtue of the positive definiteness of C shows that

[K.E. + Strain energy]n+1 [K.E. + Strain energy]n .

Thus, in the presence of damping, similar to the continuum dynamics, the

linear and angular momenta are conserved, and the total energy in non-
increasing. Instead of Eqn. (6.41), we now have
 
2M C K 2 2 1 1 1 
+ + un+1 = 2 M un + M v n + C un K un + f n + f n+1 .
t2 t 2 t t t 2 2
(6.42)
Since the energy is non-increasing, the above time-stepping scheme is uncon-
ditionally stable, i.e., there are no restrictions on the time step t . If t is
chosen to be a constant, then the matrix on the left hand side of Eqn. (6.42)
can be decomposed right at the outset, and one merely needs to use back-
substitution for all the subsequent time steps, making the whole process
extremely efficient.
Another direct method for solving Eqn. (6.34) is the Newmark method
which we now describe. If un , v n and an are approximations to u(tn ), u(tn )
and u(tn ) respectively, then the Newmark scheme is given by

M an+1 + Cv n+1 + Kun+1 = f n+1 , (6.43)

t2
un+1 = un + tv n + [(1 2)an + 2an+1 ] , (6.44)
2
v n+1 = v n + t [(1 )an + an+1 ] . (6.45)

The constants, and , determine the stability and accuracy of the algo-
rithm. The implementation to find un+1 , v n+1 and an+1 given un , v n and
an is described next.

6.4.3 Implementation
Define predictors
t2
un+1 = un + tv n + (1 2)an , (6.46)
2
v n+1 = v n + (1 )tan . (6.47)

Using the above relations, Eqns. 6.44 and 6.45 can be written as

un+1 = un+1 + t2 an+1 , (6.48)

153
 v n+1 = v n+1 + tan+1 . (6.49)

To start the process a0 is calculated from

M a0 = f Cv 0 Ku0 ,

or specified directly. The following relation obtained by substituting Eqns. 6.48

and 6.49 in Eqn. 6.43 allows us to solve for an+1 :

M + tC + t2 K an+1 = f n+1 C v n+1 K un+1 .

After finding an+1 , we use Eqns. 6.48 and 6.49 to obtain un+1 and v n+1 ,
respectively.

6.4.4 Convergence Analysis

The convergence analysis is similar to that for parabolic problems. The steps
are

1. Reduction to SDOF model

2. Definition of a suitable notion of stability which is shown to hold under

certain circumstances

3. Determination of the order of accuracy

4. Use of (ii) and (iii) to prove convergence.

If we assume Rayleigh damping, i.e.,

C = aM + bK,

then the symmetry of M and K make the decomposition shown in Fig. 6.4
possible. The parameter in Fig. 6.4 is the undamped frequency of vibration,
while is known as the damping ratio, and is given by
1 a 
= + b .
2
The SDOF model problem can be written as

y n+1 = Ay n + Ln ,

where y n [un , v n ]t , and A is the amplification matrix. The stability con-

ditions for the Newmark method are
Unconditional:
2 0.5.
Conditional:

0.5; < ; t crit ,
2

154
 M u + C u + Ku = f M an+1 + Cv n+1 + Kun+1 = f n+1
temporal un+1 = un + tv n + t2
[(1 2)an + 2an+1 ]
u(0) = u0 2
discretization v n+1 = v n + t [(1 )an + an+1 ]
u(0) = v 0
u0 , v 0 given

modal modal
decomposition decomposition

u + 2 u + 2 u = f an+1 + 2vn+1 + 2 un+1 = fn+1

temporal un+1 = un + tvn + t2
[(1 2)an + 2an+1 ]
u(0) = u0 2
discretization vn+1 = vn + t [(1 )an + an+1 ]
u(0) = v0
u0 , v0 given

Figure 6.4: Temporal discretization and modal decomposition for hyperbolic

problems

where  1/2
( 0.5) + 2
+ 2( 0.5)2
crit =

,
2

is known as the critical sampling frequency. The stability condition must be
satisfied for each mode. Hence, the maximum natural frequency n is critical
for determining the time step. We have
crit
t
n
As usual n is bounded by the maximum frequency of the individual elements.
Observe that when = 0.5, viscous damping has no effect on the stability,
since we have  0.5
crit = .
2
When > 0.5, the effect of viscous damping is to increase the critical time
step of conditionally stable Newmark methods. Thus, the undamped critical
frequency (/2 )0.5 serves as a conservative estimate when an estimate
of the modal damping coefficient is not available.

155
6.4.5 Special cases of the Newmark method
Method Type Stability condition Order of
accuracy
1 1
Average acceleration implicit 4 2
unconditional 2
(trapezoidal rule)
1 1

Linear acceleration implicit 6 2
crit = 2 3 2
1 1

Fox-Goodwin implicit 12 2
crit = 6 2
(royal road)
1
Central difference explicit 0 2
crit = 2 2

Note: For the central difference method M and C need to be diagonal.

Stability is based on the undamped case in which = 0. Second-order
accuracy is achieved if and only if = 0.5. In elastic wave propagation
problems, the time step restriction is not too restrictive. Hence, the central
difference method is preferred.

6.4.6 Finding the A matrix

We have
t2
un+1 = un + tvn + [(1 2)an + 2an+1] , (6.50)
2
vn+1 = vn + t [(1 )an + an+1 ] , (6.51)
an + 2vn + 2 un = fn , (6.52)
an+1 + 2vn+1 + 2 un+1 = fn+1 (6.53)
Eliminate an and an+1 from Eqns. 6.50 and 6.51, and substitute in Eqns. 6.52
and 6.53 to get
A = A1
1 A2 ,

t2
1 2
[(1 2)fn + 2fn+1]
Ln = A1 ,
t [(1 )fn + fn+1 ]
where

1 + t 2 2t2
A1 =
2
t 1 + 2t

2
1 t2 (1 2) 2 t [1 t(1 2)]
A2 =
2
t(1 ) 1 2t(1 )

6.4.7 Convergence and stability

The local truncation error is defined by
y(tn+1 ) = Ay(tn ) + Ln + t (tn ),

156
where y(tn ) = [u(tn ) u(tn )]t . Using an analysis similar to that for parabolic
problems, we get
(t) = o tk


t [0, T ],
where k = 2 if = 0.5 and k = 1 otherwise.
The stability of the Newmark methods is determined by the properties
of the amplification matrix. The error equation is given by
n1
X
n
e(tn ) = A e(0) t Ai (tn1i ).
i=0

Once again, stability plus consistency implies convergence. The rate of con-
vergence is k.
Let i (A) denote the eigenvalues of A. The modulus of i (A) is written
as  1/2
|i (A)| = i (A)i (A) ,
where i (A) is the complex conjugate of i (A). The spectral radius of A,
(A) is defined by
(A) = max |i (A)| .
i

The stability condition is needed to prevent the amplification of An as n

becomes large. The conditions for stability are

1. (A) 1

2. The eigenvalues of A of multiplicity greater than 1 are strictly less than

1 in modulus.

The above conditions define a spectrally stable A. Since the stability condi-
tion depends only on the eigenvalues of A, it can be expressed in terms of the
principal invariants of A. Thus, is given by the solution to 2 I1 +I2 = 0,
i.e., p
I1 I12 4I2
= ,
2
where

I1 = tr A = A11 + A22 ,
I2 = det A = A11 A22 A12 A21 .

It is necessary to have > 0.5 to introduce high-frequency dissipation, since

viscous damping damps only an intermediate band of frequencies without
significant effect on the all important high modes. However, > 0.5 results
in a drop to first-order accuracy. It is necessary to dissipate the higher modes
since they are artifacts of the discretization process, and not representative
of the behavior of the governing partial differential equations.

157
6.4.8 Superconvergence in one-dimensional problems
Consider the one-dimensional wave equation

2u E 2u 2
2 u
= = c ,
t2 x2 x2
p
where c = E/ is the characteristic velocity or the speed of sound. t =
h/c is called the characteristic time step. It is equal to the transit time for a
wave moving at speed c to traverse one element.
The element stiffness and mass matrices are given by

EA 1 1 0.5 r r
K (e) = ; M (e) = Ah .
h 1 1 r 0.5 r

Particular values of r yield the commonly used mass matrices, e.g.,

1. consistent mass (r = 1/6)

h 2 1
M (e) =
6 1 2

2. lumped mass (r = 0)
h 1 0
M (e) =
2 0 1

3. higher-order mass (r = 1/12)

h 5 1
M (e) =
12 1 5

If t is equal to the characteristic time step given by h/c, and = r, = 0.5

and = 0, then approx = exact , i.e., the errors introduced by the finite
element discretization, the particular mass matrix and temporal algorithm
all cancel to yield the exact results at the nodes (superconvergence) no
matter how few elements are employed.
The characteristic time step also represents the stability limit (i.e., the
critical time step) of the lumped mass, central difference method. Hence,
it is considered advantageous to compute at a time step as close to critical
as possible. Also matched methods ( = r) are preferred, e.g., central
difference and lumped mass ( = r = 0).

158
6.4.9 Time estimates for some simple finite elements
Consider the two-node linear rod element. If we assume a lumped mass
matrix then

EA 1 1 Ah 1 0
K (e) = ; M (e) =
h 1 1 2 0 1

Computing the eigenvalues using det(K (e) 2 M (e) ) = 0, i.e.

2c2 2 2c2
2 h2
det h 2
= 0,
2c 2c2 2
h2 h

p
where c = E/. The above equation yields

4c2
2 = or = 0.
h2
Thus, max = 2c/h. For = 0, = 0.5,
2 h
tcritical = = = tcharacteristic .
max c

For a consistent mass matrix max = 2 3c/h, which results in a reduced
critical time step:
h
t .
3c
In general, consistent-mass matrices tend to yield smaller critical time steps
than lumped-mass matrices.
For the heat conduction problem with = 0, the condition t 2(1
2) implies that
2
t .
max
2
Since max = max = 4k/ch2 , we get
ch2
t .
2k
For a 3-node quadratic element,

1 0 0 7 8 1
(e) hA EA
M lumped = 0 4 0 ; K (e) = 8 16 8 ,

6 h
0 0 1 1 8 7

we get max = 2 6c/h, yielding
h
t .
6c

159
Thus, the allowable time step is about 0.4082 that for linear elements with
lumped mass.
For a four-node quadrilateral element

max cd g, (6.54)
p
where cd = ( + 2)/ is the dilatational wave velocity, and are Lame
constants, and g is a geometric parameter. For example, for a quadrilateral
2 4
4 XX
g= 2 Bij Bij ,
A i=1 j=1

where A is the area, and

1 y2 y4 y3 y1 y4 y2 y1 y3
Bij = .
2 x4 x2 x1 x3 x2 x4 x3 x1

For the central-difference method, Eqn. 6.54 leads to

2
t .
cd g

For a rectangular element with sides h1 and h2 , the above formula reduces
to
1
t  1/2 .
1 1
cd h 2 + h 2
1 2

160
Appendix A

Stiffness and mass matrices for

the Timoshenko beam element

In what follows, A, I, J, and L denote the area of cross-section, second

moment of the cross-sectional area, polar moment of inertia, and length of
the beam, and E, G, C, , k denote the Young modulus, shear modulus,
torsional rigidity, density and shear correction factor (usually taken to be
5/6), respectively. We assume all geometric and material properties to be
uniform along the length of the beam element, though, of course, they can
differ from element to element.
Following the notation of Bathe [1], let (w1 , x1 , y1 , w2 , x2 , y2 ), and
(w1 , x1 , y1 , w2 , x2 , y2 , w3 , x3 , y3 ) be the degrees of freedom associated with a
linear and quadratic beam element, respectively, and let = EI/(GAk) and
= C/(GAk).
Linear beam element:
The mass matrix for a beam along the x-axis is given by

2A 0 0 A 0 0

0 2J 0 0 J 0


0 0 2I 0 0 I

L
M= .
6
A 0 0 2A 0 0


0 J 0 0 2J 0


0 0 I 0 0 2I

For a beam along the y-axis, interchange I and J in the above matrix.

161
 The stiffness matrices for beams along the x and y-axes are given by

1 1 1 1
0 0
L 2 L 2


0 0 0 L 0

L
1
2 0 L4 + L 1
2
0 L
4

L
K x = GAk 1
,
1 1 1
0 0

L 2 L 2



0 0 0 0
L L
12 0 L
4

L
1
2
0 L
4
+
L

1 1
L 2
0 L1 1
2
0


1 L
+
0 21 L

0
2 4 L 4 L


0 0 L
0 0 L
K y = GAk .

1 12 0 1
12 0
L L

1 L
0 21 L
+
0

2 4 L 4 L

0 0 L 0 0
L

Quadratic element:
The mass matrix for a beam along the x-axis is

4A 0 0 2A 0 0 A 0 0

0 4J 0 0 2J 0 0 J 0


0 0 4I 0 0 2I 0 0 I


2A 0 0 16A 0 0 2A 0 0

L
M= 0 2J 0 0 16J 0 0 2J 0 .

30
0 0 2I 0 0 16I 0 0 2I


A 0 0 2A 0 0 4A 0 0



0 J 0 0 2J 0 0 4J 0

0 0 I 0 0 2I 0 0 4I

The mass matrix for a beam along the y-axis is obtained by interchanging I
and J in the above equation.

162
 The stiffness matrices for beams along the x and y-axes are given by

7 1 8 2 1 1
0 0 0
3L 2 3L 3 3L 6
7 8
0 0 0 3L 0 0 0

3L 3L


1 7 L 2 8 L 1 L
2 0 3L
+9 3
0 3L + 9 6 0 3L
18

8 2 16 8 2
3L 0 0 0 0

3 3L 3L 3


K x = GAk 0 3L 8 16 8
0 0 0 0 3L 0 ,

3L
2 8 L 16 4L 2 8 L
3 0 3L + 9 0 0 + 9 0 3L + 9

3L 3
1 1 8 2 7 1
0 6 3L 0 0

3L 3 3L 2


8 7
0 3L
0 0 3L 0 0 3L
0

1 L 2 8 L 1 7 L
6
0 3L
18 3 0 3L + 9 2
0 3L
+9

7 1 8 2 1 1
0 0 0
3L 2 3L 3 3L 6
1 7 L 2 8 L 1 L
+9 0 3 3L + 9 0 18 0

2 3L 6 3L

7 8
0 0 0 0 0 0

3L 3L 3L


8
3L 23 0 16
0 0 3L 8 2
0

3L 3
2 8 L 16 4L 2 8 L
K y = GAk 3 3L + 9 0 0 + 9 0 3 3L + 9 0 .

3L
8 16 8
0 0 3L 0 0 0 0

3L 3L


1 1 8
0 3L 23 0 7
12 0

3L 6 3L

1 L 2 8 L 1 7 L
6 18 0 3L + 9 0 2 +9 0

3L 3 3L
8 7
0 0 3L
0 0 3L 0 0 3L

163
Bibliography

[1] Bathe, K. J., Finite Element Procedures, Prentice-Hall, New Jersey,

1997.

[2] Shames, I. H. and C. Dym, Energy and Finite Element Methods in Struc-
tural Mechanics, Hemisphere Publishing Co., 1985.

[3] Hughes, T. J. R., Linear, Static and Dynamic Finite Element Analysis,
Prentice Hall, 1987.

[4] Zienkiewicz, O. C. and R. L. Taylor, The Finite-Element Method, Vol.

1. McGraw-Hill International, 1989

164