Finite Element Method

H. Kasumba

October 6, 2021

1 Introduction

The finite element method (FEM) is a numerical technique for solving a wide range of complex
physical phenomena, particularly those exhibiting geometrical and material non-linearities (such
as those that are often encountered in the physical and engineering sciences). These problems can
be structural in nature, thermal (or thermo-mechanical), electrical, magnetic, acoustic etc. plus
any combination of. It is used most frequently to tackle problems that aren’t readily amenable to
analytical treatments.

Figure 1:

Physical problems are governed by differential equations which satisfy Boundary conditions Initial
conditions

One variable: Ordinary differential equation (ODE) Multiple independent variables: Partial dif-
ferential equation (PDE)

1
 Figure 2:

2 Weak Formulation and Mathematical equivalences

Consider the problem

−u′′ (x) = f (x), 0 < x < 1, u(0) = 0, u(1) = 0.
R1 R1
Set F (v) = 12 0 (v ′ )2 dx − 0 f v dx.

2.1 Minimization problem

The minimization problem is set as follows: Find u ∈ V such that

F (u) ≤ F (v) for all v ∈ V
Note: If u represents the transverse displacement of a membrane, then F (u) is the total energy.

2.2 Variational formulation

Consider again the problem

−u′′ (x) = f (x)
We multiply this PDE with a test function ϕ and integrate

Z 1 Z 1
′′
(1) −u (x)ϕ(x) dx = f (x)ϕ(x) dx
0 0

2
Integrating by parts the left hand side of (1), we obtain
Z 1 Z 1
′ ′ ′ ′
−u (1)ϕ(1) + u (0)ϕ(0) + u (x)ϕ (x) dx = f (x)ϕ(x) dx
0 0

Next, we choose ϕ(x) such that it has compact support in [0, 1], i.e., ϕ(0) = ϕ(1) = 0. Then, the
variational formulation becomes
Z 1 Z 1
′ ′
(2) u (x)ϕ (x) dx = f (x)ϕ(x) dx
0 0

2.3 Mathematical Equivalences

1. If u is a solution to the strong formulation, then u is a solution to the variational formulation.

Proof. Integrate by parts the strong formulation.

2. If u is a solution to the variational formulation, then u solves the minimization problem.

Proof. Suppose u∗ satisfies the (VP) i.e.,

Z 1 Z 1
∗
(3) ux ϕx dx = f (x)ϕ(x) dx for any ϕ ∈ V and ϕ(0) = ϕ(1) = 0
0 0

Then

F (u) = F (u∗u − u∗ )
Z 1 Z 1
1 ∗
= 2
[(u + w)x ] dx − f (u∗ + w) dx where w = u − u∗
0 2 0
1 1 ∗ 2
Z Z 1
∗
= 2
(u ) + 2ux wx + (wx ) dx − (f u∗ + f w) dx
2 0 x 0
1 1 ∗ 2
Z 1 Z 1
1 1 2
Z Z
∗ ∗
= (u ) − f u dx + ux wx − f w dx + w dx
2 0 x 0 0 2 0 x
| {z } | {z }
=0 ≥0
1 1 2
Z
∗
F (u) = F (u ) + w dx
2 0 x
| {z }
≥0
⇒ F (u∗ ) ≤ F (u) for all u ∈ V

Therefore, u∗ solves the minimization problem.

3. If u solves the minimization problem, then u solves the variational problem.

3
 Proof. u solution to minimization problem implies that

F (u) ≤ F (u + ϵv), ϵ ∈ R, ∀v ∈ V.

The functional
1 1 1 1 1
ϵ2
Z Z Z Z Z
1 ′ 2 ′ ′ ′ 2
g(ϵ) := F (u + ϵv) = (u ) dx + ϵ u v dx + (v ) dx − f u dx − ϵ f v dx.
2 0 0 2 0 0 0

has a minimum at ϵ = 0, implying that g ′ (0) = 0. But

Z 1 Z 1 Z 1
′ ′ ′
g (ϵ) = u v dx − f v dx + ϵ (v ′ )2 dx
0 0 0
R1 R1
Hence, g ′ (0) = 0 implies g ′ (0) = 0
u′ v ′ dx − 0
f v dx = 0. Thus u solves the variational
problem.

3 The Finite Element Method: 1D Case

We illustrate the FE method for the 1D two-point BVP

−u′′ (x) = f (x), 0 < x < 1, u(0) = 0, u(1) = 0.

The steps to follow are as follows:

1. Construct a variational or weak formulation .

Z 1 Z 1
′ ′
u v dx = f v dx,
0 0

where v(x) is a test function satisfying the boundary conditions, i.e., v(0) = v(1) = 0.

Step 2: We generate a mesh: e.g. a uniform mesh.

ˆ Without loss of generality, let Ω = (0, 1).

ˆ Let τh := partition of Ω into N subintervals Ki = (xi−1 , xi ) of width hi = xi − xi−1 :

a = 0 < x1 < x2 < · · · < xn = 1.

ˆ The mesh will be called uniform if the points xi are equidistant, that is,
1
xi = ih, with h = , 0 ≤ i ≤ n.
n

4
 ˆ The points xi are called vertices or nodes of the mesh.

3. We replace the Hilbert space V on which we pose the variational formulation by a subspace Vh
of finite dimension.

Motivation

Suppose the value of u(x) at nodes xi and xi+1 are known such that

u(xi ) = ui , u(xi+1 ) = ui+1

but the value of u(x) is unknown.

This value can be approximated by uh (x) and using interpolation, we have

ui+1 − ui uh (x) − ui
=
xi+1 − xi x − xi
Re-arranging, we obtain
   
x − xi xi+1 − x
uh (x) = ui+1 + ui
h h
= ui+1 ψi+1 (x) + ui ψi (x),

where  
xi+1 − x
ψi (x) = , ψi (xi ) = 1, ψi (xi+1 ) = 0.
h
Similarly,  
x − xi
ψi+1 (x) = , ψi+1 (xi ) = 0, ψi+1 (xi+1 ) = 1.
h

5
The above functions are known as shape functions. The union of 2 shape functions in neighboring
elements gives a global basis function at a node xi :
 x−x

 h
i−1
x ∈ (xi−1 , xi ),

xi+1 −x
(4) ϕi (x) = h
x ∈ (xi , xi+1 ),


0 otherwise


Note: Supp(ϕi (x)) = Ki ∪ Ki+1 and belong to the set

ˆ Vh = {vh ∈ C([0, 1]) : vh |Kj ∈ P1 , for all Kj ∈ τh , vh (0) = vh (1) = 0}.

ˆ Here, P1 is the space of polynomials of degree ≤ 1.
ˆ Vh1 =continuous, piecewise linear functionals.
ˆ dim(Vh1 ) = N .

step 4: Representation of approximate solution in terms of the basis functions.

Lemma 3.1. The space Vh1 is a subspace of V = H01 (0, 1) and every function uh (x) is defined
uniquely by its values at the vertices (xi )1≤i≤n−1 :
n−1
X
uh (x) = u(xi )ϕi (x), for all x ∈ [0, 1].
i=1

6
 ˆ On assuming the hat basis functions, obviously uh (x) is also a piecewise linear function,
although this is not usually the case for the true solution u(x).

ˆ Next substitute the approximate solution

n−1
X
uh (x) = ui ϕi (x)
i=1

for the exact solution u(x) in the weak form

Z 1 Z 1
′ ′
u v dx = f v dx,
0 0

i.e.,
Z 1 Z 1
(5) u′h v ′ dx = f v dx,
0 0
Z n−1
1X Z 1
(6) ⇒ ui ϕ′i v ′ dx = f v dx
0 i=1 0
n−1
X Z 1 Z 1
(7) ⇒ ui ϕ′i v ′ dx = f v dx
i=1 0 0

Now the test function v(x) is chosen to be ϕ1 , ϕ2 , · · · , ϕn−1 successively, to get the system of linear
equations (noting that further errors are introduced).

AU = F,

where  
a(ϕ1 , ϕ1 ) a(ϕ1 , ϕ2 ) ··· a(ϕ1 , ϕn−1 )
 a(ϕ2 , ϕ1 ) a(ϕ2 , ϕ2 ) ··· a(ϕ2 , ϕn−1 ) 
A=
 
.. .. .. .. 
 . . . . 
a(ϕn−1 , ϕ1 ) a(ϕn−1 , ϕ2 ) · · · a(ϕn−1 , ϕn−1 )
   
U1 (f, ϕ1 )
Z 1
 U2   (f, ϕ2 ) 
U = ..  , F =   , a(ϕi , ϕj ) = ϕ′i ϕ′j dx.
   
..
 .   .  0
Un−1 (f, ϕn−1 )
Example 3.1. Consider the two point boundary value problem

(P) −u′′ (x) = 1, 0 < x < 1

u(0) = 0, u(1) = 0

(a) Derive the variational formulation of (P). (03 marks)

7
 (b) Divide the interval (0,1) into 3 subintervals of equal length h = 13 and let Vh be the corre-
sponding space of continuous piecewise linear functions vanishing as x = 0 and x = 1.

(i) Use Vh to formulate a finite element method for (P).

(ii) Verify that the stiffness matrix A and the load vector b are given by
   
2 −1 1 1
A=3 , b=
−1 2 3 1

(iii) Discuss the solvability of the system

Ax = b,

where A and b are defined as above.

Solution
Let V := {ϕ ∈ H 1 (0, 1) | ϕ(0) = 0, ϕ(1) = 0}.

(a) Variational formulation

Z 1 Z 1
′ ′
u (x)v(x) dx = v(x) dx + [u′ (x)v(x)]10
Z0 1 Z0 1
⇒ u′ (x)v(x)′ dx = v(x) dx + u′ (1)v(1) − u′ (0)v(0)
Z0 1 Z0 1
⇒ u′ (x)v(x)′ dx = v(x) dx.
0 0

(b) Let ϕi ∈ Vh ⊂ V, i = 1, 2 such that ϕi (xj ) = 1 if i = j and ϕi (xj ) = 0 if i ̸= j. Then we can

express uh as
X2
uh (x) = ui ϕi (x)
i=1

where
 
 3x x ∈ (0, 1/3),  3x − 1 x ∈ (1/3, 2/3),
(8) ϕ1 (x) = 2 − 3x x ∈ (1/3, 2/3), ϕ2 (x) = 3 − 3x x ∈ (2/3, 1),
0 otherwise 0 otherwise
 

and the variational form becomes

X Z 1 Z 1
ui ϕ′i (x)v(x)′ dx = v(x) dx,
Z 1 Z 01 0
Z 1
⇒ u1 ϕ′1 (x)v(x)′ dx + u2 ϕ′2 (x)v(x)′ dx = v(x) dx.
0 0 0

8
 Note
supp(ϕ1 ) = (0, 2/3)
supp(ϕ2 ) = (1/3, 1)
supp(ϕ1 ∩ ϕ2 ) = (1/3, 2/3)

 
 3 x ∈ (0, 1/3),  3 x ∈ (1/3, 2/3),
(9) ϕ′1 (x) = −3 x ∈ (1/3, 2/3), ϕ′2 (x) = −3 x ∈ (2/3, 1),
0 otherwise 0 otherwise
 

Choose v(x) = ϕ1 (x). Then

"Z # "Z #
1/3 Z 2/3 2/3 Z 1
u1 9 dx + 9 dx − u2 9 dx = ϕ1 (x) dx.
0 1/3 1/3 0

1
(10) 6u1 − 3u2 =
3
Next, choose v(x) = ϕ2 (x). Then
"Z # "Z #
2/3 2/3 Z 1 Z 1
−u1 9 dx + u2 9 dx + 9 dx = ϕ2 (x) dx.
1/3 1/3 2/3 0

1
(11) −3u1 + 6u2 = .
3
Collecting the above results, we find
   
2 −1 1 1
A=3 , b=
−1 2 3 1
Since     
1 20 −11 −9 −11 693
Det(A) = 16det + 9det = ̸= 0
2 −11 11 0 11 2
Matrix A is invertible and the linear system

Ax = b,

is uniquely solvable.
Example 3.2. Consider the two point boundary value problem

(P) −((1 + x)u′ )′ = 1, 0 < x < 1

u(0) = 0, u′ (1) = 1

9
 (a) Derive the variational formulation of (P). (03 marks)
(b) Divide the interval (0,1) into 3 subintervals of equal length h = 31 and let Vh be the corre-
sponding space of continuous piecewise linear functions vanishing as x = 0.
(i) Use Vh to formulate a finite element method for (P). (04 marks)
(ii) Verify that the stiffness matrix A and the load vector b are given by
   
16 −9 0 1
1 1
A =  −9 20 −11  , b =  1  .
2 3
0 −11 11 7
(15 marks)

Let V := {ϕ ∈ H 1 (0, 1) | ϕ(0) = 0}.

(a) Variational formulation

Z 1 Z 1
′ ′
((1 + x)u (x)v(x) dx = v(x) dx + [(1 + x)u′ (x)v(x)]10
Z0 1 Z0 1
⇒ ((1 + x)u′ (x)v(x)′ dx = v(x) dx + 2u′ (1)v(1) − 2u′ (0)v(0)
Z0 1 Z0 1
⇒ ((1 + x)u′ (x)v(x)′ dx = v(x) dx + 2v(1).
0 0

(b) Let ϕi ∈ Vh ⊂ V, i = 1, 2, 3 such that ϕi (xj ) = 1 if i = j and ϕi (xj ) = 0 if i ̸= j. Then we

can express uh as
Xn
uh (x) = ui ϕi (x)
i=1
and the variational form becomes
X Z 1 Z 1
′ ′
ui (1 + x)ϕi (x)v(x) dx = v(x) dx + 2v(1),
0 0
Z 1 Z 1
′ ′
⇒ u1 (1 + x)ϕ1 (x)v(x) dx + u2 (1 + x)ϕ′2 (x)v(x)′ dx
0 0
Z 1 Z 1
′ ′
+u3 (1 + x)ϕ3 (x)v(x) dx = v(x) dx + 2v(1).
0 0

Note supp(ϕ1 ) = (0, 2/3),supp(ϕ2 ) = (1/3, 1),supp(ϕ3 ) = (2/3, 1) supp(ϕ1 ∩ϕ2 ) = (1/3, 2/3),
supp(ϕ2 ∩ ϕ3 ) = (2/3, 1).
 
 3 x ∈ (0, 1/3),  3 x ∈ (1/3, 2/3),
(12) ϕ′1 (x) = −3 x ∈ (1/3, 2/3), ϕ′2 (x) = −3 x ∈ (2/3, 1),
0 otherwise 0 otherwise
 

3 x ∈ (2/1, 1),
(13) ϕ′3 (x) =
0 otherwise

10
Since v ∈ Vh , choose v(x) = ϕ1 (x). Then
"Z # "Z #
1/3 Z 2/3 2/3
u1 (1 + x)9 dx + (1 + x)9 dx − u2 (1 + x)9 dx + u3 · 0 =
0 1/3 1/3
Z 1
ϕ1 (x) dx + 2ϕ1 (1).
0

Since ϕ1 (1) = 0, we have

9 1
(14) 8u1 − u2 + u3 · 0 =
2 3
Next, choose v(x) = ϕ2 (x). Then
"Z # "Z #
2/3 2/3 Z 1 Z 1
−u1 (1 + x)9 dx − u2 (1 + x)9 dx + (1 + x)9 dx − u3 (1 + x)9 dx =
1/3 1/3 2/3 2/3
Z 1
ϕ2 (x) dx + 2ϕ2 (1).
0

Since ϕ2 (1) = 0, we have

9 11 1
(15) − u1 + 10u2 − u3 = .
2 2 3
Finally, set v(x) = ϕ3 (x). Then
Z 1 Z 1
u1 · 0 − u2 (1 + x)9 dx + u3 (1 + x)9 dx =
2/3 2/3
Z 1
ϕ3 (x) dx + 2ϕ3 (1).
0

Since ϕ3 (1) = 1, we have

11 11 1
(16) 0 · u1 − u2 + u3 = + 2.
2 2 3
Collecting the above results, we find
  
16 −9 0 1
1 1
A =  −9 20 −11  , b=  1 
2 3
0 −11 11 7

11
4 Implementation of FEM

weak formulation . Z 1 Z 1
′ ′
u v dx = f v dx,
0 0
Discretization
n−1
X Z 1 Z 1
(17) ui ϕ′i ϕ′j dx = f ϕj dx
i=1 0 0

Note: Z 1 Nk Z
X
ϕ′i ϕ′j dx = ϕ′i ϕ′j dx
0 k=1 ek

In element ek we have two local basis functions ϕi and ϕi+1

Figure 3:

given by
xi+1 − x 1
ϕi (x) = , ϕi (x)′ = −
h h
x − xi 1
ϕi+1 (x) = , ϕi+1 (x)′ =
h h
The local stiffness matrix is given by
" R xi+1 ′ 2 R xi+1 ′ ′ #
x i
(ϕ i ) dx x i
ϕi ϕi+1 dx
(18) Kie = R xi+1 R xi+1 ′ 2
′ ′
ϕi ϕi+1 dx xi (ϕi+1 ) dx
" x1i
− h1
#
h
(19) =
− h1 h1
Local load vector.
Assume f = 1, then
" R xi+1 #
xi
ϕi dx
(20) Fie = R xi+1
xi
ϕi+1 dx
h
" #
2
(21) =
h
2

12
4.1 Implementation of the sum

Consider the case of 3 sub-intervals.

4.1.1 Global Stiffness Matrix

Define an empty global stiffness matrix S of size 4 × 4.

 
0 0 0 0
 0 0 0 0 
S=  0 0 0 0


0 0 0 0

In element 1
S([1, 2], [1, 2]) = S([1, 2], [1, 2]) + K1e
 
1 −1 0 0
1  −1 1 0 0 
S=  
h 0 0 0 0 
0 0 0 0
In element 2
S([2, 3], [2, 3]) = S([2, 3], [2, 3]) + K2e
The global stiffness matrix becomes
 
1 −1 0 0
1  −1 2 −1 0 
S=  
h  0 −1 1 0 
0 0 0 0

In element 3
S([3, 4], [3, 4]) = S([3, 4], [3, 4]) + K3e
The global stiffness matrix becomes
 
1 −1 0 0
1  −1 2 −1 0 
S=  
h  0 −1 2 −1 
0 0 −1 1

13
4.1.2 Global Load Vector

Assume f = 1, then
h
" #
2
(22) Fie =
h
2

Then        
0 1 1 1
 0  h 1  h  1+1
⇒ h 2 
  
 0 ⇒ 2
F =   ⇒ 
0  2 1  2 1+1 

0 0 0 1
The entire system reads
    
1 −1 0 0 u1 1
1 −1 2 −1 0   u2

 1 2 
=  
h  0 −1 2 −1   u3  h 2 
0 0 −1 1 u4 1

Since, u1 = u4 = 0, we re-modify the above system to obtain

    
1 0 0 0 u1 0
1
 −1 2 −1 0 

 u2 
= h  2 

h  0 −1 2 −1   u3  2  2 
0 0 0 1 u4 0

14