Introduction to variational methods and finite elements

1.2.3. Variational formulations of BVP:

ax − b

x
−b
Problem: Sove ax = b x=
a
Reformulate the problem:
Consider E = 12 ax2 + bx
Find x∗ : E(x∗ ) = min E(x)
x
x

1. Rayleigh-Ritz Method:

Consider a differential equation

Functional
 an ∞ dimension
Au = u= f (x) (1a) vector
u(0) = α u(1) = β (1b)

1 1  2
Consider the functional: E[u] = 2 (u ) + f u dx ← potential energy functional.
0

Claim: If u∗ : E[u∗ ] = min E[u] and u∗ satisfies (1b) then u∗ solves (1).
u

Proof:

Let u = u∗ + εη ←(arbitrary) η(0) = 0 = η(1) and N

u∗
η ∈ C 2 so we have a 1 parameter family of β
functions which are perturbations of u∗ . The function α

1
∗ 1  ∗ 2
E(ε) = E[u + εη] = u + εη  + f (u∗ + εη) dx
2
0
1 
 
E  (0) = (u∗ + εη  )η  + f η dx
ε=0
0
1
= (u∗ ) η  + f η dx
0
1 1   
∗
= u η − u∗ − f η dx = 0
0
0

22
   
⇒ Since η is arbitrary we could choose η = u∗ − f

⇒ u∗ − f = 0 ← Euler -Lagrange equation for E[u].


N −1
How is this useful? Let us assume that U (x) = αk Ψk (x)
k=1
↑ basis functions

Then
1 N −1
2
1  
N

E(α) = αk Ψk (x) + f (x) αk Ψk (x) dx
2
0 k=1 k=1
For a min:
1 N
−1

∂E
0= = αk Ψk · Ψj + f (x)Ψj (x) dx
∂αj
0 k=1


N −1 1 1
= αk Ψk Ψj dx + f (x)Ψj (x) dx, j = 1, . . . , N − 1
k=1 0 0
1 1
⇒ Aα = b where Akj = Ψk Ψj dx bj = − f (x)Ψj (x)dx.
0 0


Example : u = −x2 x
u(x) = (1 − x3 )
u(0) = 0 = u(1) 12

Assume {Ψj } = 1, x, x2 , x3

U3 (x) = c0 + c1 x + c2 x2 + c3 x3
U3 (0) = c0 = 0 u3 (1) = c1 + c2 + c3 = 0 −c2 x + c2 x2 = c3 x + c3 x3
U3 (x) = ax(1 − x) + bx2 (1 − x) −c2 x(1 − x)
= a{x − x2 } +b {x2 ± x3 } c2 = −c1 − c3
Ψ1 Ψ2

1 1 1

4  1
A11 = (1 − 2x)2 dx = 1 − 4x + 4x2 dx = x − 2x2 + x3  =
3  3
0 0 0
1 1 1
2 2 2 4 3
3 4 9 5 
4 20 − 45 + 27 2
A22 = (2x − 3x ) dx = 4x − 12x + 9x dx = x − 3x + x  = =
3 5  15 15
0 0 0
1
1
A12 = A21 = (1 − 2x)(2x − 3x2 ) dx =
6
0
1
1
b1 = + x2 (x − x2 ) dx =
20
0

23
 1
1
b2 = x2 (x2 − x3 )dx =
30
0
⎡ 1 1
⎤⎡ ⎤ ⎡ 1
⎤
3 6 a 20
⎣ ⎦⎣ ⎦=⎣ ⎦⇒a= 1 b=
1
1 2 1 15 6
6 15 b 30
x
U3 (x) = (1 − x)(2 + 5x)
30
Exact:
x
uex (x) = (1 − x3 )
12

Natural and Essential B-C

Notice that the basis functions {Ψi } were required to satisfy the BC u(0) = 0 = u(1). B-C that have
to be forced onto the trial solution are called essential B-C – typically these involve the solution
values and not the derivatives. In the case of derivative B-C it is possible to build the B-C into the
energy functional to be minimized.

u(0) = α u (1) = β
Consider
u = f (x)

Old energy functional Boundary Condition

1
1  2
E[u] = (u ) + f u dx − βu(1)
2
0
1
δE[u] = u δu + f δu dx − βδu(1)
0
1 1  

= u δu − u − f δu dx − βδu(1)
0
0
1

= {u (1) − β}δu(1) − (u − f )δu dx = 0
0
⇒ u = f and u (1) = β but says nothing about u(0)!

↑ Don’t have to enforce BC at this endpoint.

Eg. 2:

7x x3
u = −x = f uexact = 2 + −
2 6
u(0) = 2 u (1) = 3

↑ essential ↑ natural

Let: U (x) = 2 + x(a1 + a2 x) = 2 + a1 x + a2 x2

24
 ↓ satisfies homogeneous version of essential BC

N 
N
UN (x) = α + ai Ψi (x) = ai Ψi (x) a0 = 1 Ψ0 (x) = 2
i=1 i=0
1
1  2    
E(a) = ai Ψi + f ai Ψi dx − 3 ai Ψi (1)
2
0
1  
∂E
0= = ai Ψi Ψj + f Ψj dx − 3Ψj (1)
∂aj
0
 
0 = Ψi , Ψj ai + (f, Ψj ) − 3Ψj (1)
0 = Aa + b
1 1 1 1
2 4 3  4
A11 = dx = 1 A21 = A12 = 2x dx = 1 A22 = 4x dx = x  =
3  3
0 0 0 0
1
b1 = (−x) · x dx − 3 = −3 13     
0 1 1 a1 10/3 43 1
⇒ = ⇒ a1 = a2 = −
1 1 4/3 a2 13/4 12 4
b2 = (−x)x2 dx − 3 = −3 14
0
43 x2
U2 (x) = 2 + x−
12 4

x 0 0.2 0.4 0.6 0.8 1

RR 2 2.707 3.393 4.060 4.707 5.333
EX 2 2.699 3.389 4.064 4.715 5.333 ← Built in

2. General calculus of variations:

b
I[u] = F (x, y, y  ) dx y(a) = 0, y(b) = 0
a
b
∂F ∂F
0 = δI = δy +  (δy  ) dx
∂y ∂y
a
b b    
∂F  d ∂F ∂F
= δy  − − δy dx
∂y  dx ∂y  ∂y
a a
 
d ∂F ∂F
⇒ − =0 Euler-Lagrange Eq.
dx ∂y  ∂y

F = p(u )2 + qu2 − 2f u
 
Eg.1 − (pu ) + qu = f ∂F
= 2qu − 2f
u (a) = α u(b) = β ∂u
b ∂F
= 2pu
I[u] = p(u )2 + qu2 − 2f u dx + αu ∂u  
∂ ∂F ∂F 
a

− = 2 (pu ) − qu + f = 0
∂x ∂u ∂u

25
Eg. 2: Sturm-Lioville Eigenvalue problems:

u(a) = 0 = u(b); +(pu ) + qu + λru = 0 (1)

b b
Multiply by v and integrate +(pu ) v + quv dx + λ ruv dx = 0
a a

b
p(u )2 − qu2 dx
a I1
λ = I[u] = =
b I2
ru2 dx
a
δI1 I2 − I1 δI2 1
δλ = = (δI1 − λδI2 )
I22 I2
b b  
2pu δu − 2quδu − λruδu dx −2 (pu ) + qu + λru δu dx
a a
= = =0
b b
ru2 dx ru2 dx
a a
∴ δλ = 0 ⇒ the function u0 which minimizes I[u] is an eigenfunction of (1) and
λ = I[u0 ] is its eigenvalue.

Higher eigenvalues: λn = min I[u]

(u,Ψk )=0
k=1,...,n−1 ←− constrained minimization problem.
3. Method of weighted residuals

What do we do for nonlinear or dissipative problems for which potential energy functionals don’t
exist or cannot be found easily?

Consider a BVP

Lu = f in Ω = (a, b) (1a)

u(a) = α ; u (b) + σu(b) = β (1b)

An approximate solution U won’t in general satisfy (1a) and we associate with U the so-called
residual

r(U ) = LU − f
Note that r(u) = 0
exact ↑

A whole class of methods are obtained by considering various ways to minimize the residual in
some sense, usually:
b
r(U )φi dx = 0
a
↑ test or weight functions

N
where U= ai Ψi (x) ← basis function
i=1

26
1. Collocation: weight functions = δ(x − xi ) basis functions polynomials:

Eg:

u − u = 0
u(0) = 1, u(1) = 0

 i+1 
(A) Let Ψi (x) = x − x which satisfy the homogeneous B-C for the problem
   
U (x) = 1 − x + a1 x2 − x + a2 x3 − x
 
r(U ) = {a1 2 + a2 6x} − a1 (2x − 1) + a2 3x2 − 1 + 1
 
= (3 − 2x) a1 − a2 3x2 − 6x − 1 + 1

With x1 = 0 and x2 = 1 as collocation points we have

      
3 1 a1 −1 3 2
= ⇒ U1 (x) = 1 − x − (x2 − x) − (x3 − x)
1 4 a2 −1 11 11

If we choose x1 and x2 are chosen to be the zeros of the 2nd degree Legendre polynomial
 
1 2 1
U2 (x) = 1 − x − (x − x) − (x3 − x)
4 6

(e − ex )
The exact solution is u(x) = .
(e − 1)

x u(x) U1 (x) U1 − u U2 U2 − u U3 U3 − u
1/4 0.834704 0.837750 3.05×10−3 0.835938 1.23×10−3 0.769965 −6.47 × 10−2
1/2 0.622459 0.636364 1.39 × 10−2 0.625000 2.54 × 10−3 0.506144 −1.16 × 10−1
3/4 0.349932 0.360795 0.351563 1.63 × 10−3 0.269965 −8.00 × 10−2

(B) With different trial functions:

U (x) = 1 − x + a1 sin πx + a2 sin 3πx

r(U ) = −π(π sin πx + cos πx)a1 − 3π(3π sin 3πx + cos 3πx)c2
U3 (x) = 1 − x + 0.017189 sin πx + 0.011045 sin 3πx

using zeros of 2nd degree Legendre polys.

Where to collocate?

What basis fucntions to use?

DeBoor, C. and Swartz, B. , Collocation at Gaussian points, SIAM J. Num. Anal. 10 (1973),
582–606.

27
2. Method of moments: weight functions φi = xi .

Eg:

u + u + x = 0
u(0) = 0 = u(1)
Let U (x) = a1 x(1 − x) + a2 x2 (1 − x) + . . .
r(U ) = x + a1 (−2 + x − x2 ) + a2 (2 − 6x + x2 − x3 )
1 1
r(U ) · 1 dx = 0 and r(U ) · x dx = 0
0 0
⎡ 11 11
⎤⎡ ⎤ ⎡ 1
⎤
6 12 a1 2
⇒ ⎣ ⎦⎣ ⎦=⎣ ⎦
11 19 1
22 a2
20 3
 
122 110
U (x) = x(1 − x) + x
649 649
sin x
uex (x) = −x
sin 1

3. Galerkin method: ‘Expand U (x) and w(x) in terms of the same basis functions.’


N 
N
u(x) = ai Ψi w(x) = bi Ψ i
i=1 i=1

 N  N   N 
  
r ai Ψi bi Ψi dx = 0 ⇒ r ai Ψi Ψj dx = 0
Ω i=1 i=1 Ω i=1

Since the bi are arbitrary.

Eg. u + u + x = 0
u(0) = 0 = u(1)

Let U (x) = a1 x(1 − x) + a2 x2 (1 − x)

r(U ) = x + a1 (−2 + x − x2 ) + a2 (2 − 6x + x2 − x3 )
1 1
    x(1 − x)
r(u) · Ψ1 (x)dx = x + a1 −2 + x − x2 + a2 2 − 6x + x2 − x3 · 2 dx = 0
x (1 − x)
0 0
⎡ 3 3 ⎤⎡ ⎤ ⎡ 1 ⎤
a1 12
⎢ 20 ⎥ ⎣ ⎦ ⇒ a1 = 7 7
⇒ ⎣ 10 ⎦ ⎦=⎣ a2 =
1 369 41
3 13 a2 20
20 105
 
71 7
U (x) = x(1 − x) + x
369 41
sin x
ue (x) = −x
sin 1

28
Relationship between Rayleigh Ritz and Galerkin:

Rayleigh Ritz u + u = f = −x u(0) = 0 = u(1)

1
I[u] = u2 − u2 + 2f u dx
0
1 1       
0 = δI = 2u δu − 2uδu + 2f δu dx I(a) = ai Ψi ai Ψj − ai Ψi aj Ψj
0
0
1 
 1   +2f aj Ψj dx
= u δu − u + u − f δu dx
0
∂I  
0 0 = =2 ai (Ψi , Ψk ) − ai (Ψi , Ψk )
1 ∂ak
+(f, Ψk )}
⇒0 = (u + u − f )δu dx
f = −x
0
1
= r(u)δu dx
0

Rayleigh-Ritz and Galerkin methods are identical for this problem. – true in general for linear
problems but not for nonlinear problems.

From WR to the weak form:

1
WR ⇒ (u + u + x)v dx = 0
0
u(0) = 0 = u(1) v satisfy homogeneous Dirichlet i.e. v(0) = 0 = v(1)

Say we wanted to express u in terms of functions that are not twice differentiable then we integrate
by parts to throw as many derivatives as needed from u to v.

1 1 1


(u + u + x)v dx = u v  + −u v  + uv + xv dx = 0.

0 0 0
 
Now if we let u = ai Ψi and v = bi Ψi we have since v is arbitrary
 
− ai (Ψi , Ψj ) + ai (Ψi , Ψj ) + (x, Ψj ) = 0
i

Aa = b
where
[A]jj = (Ψi , Ψj ) − (Ψi , Ψj ); b = (x, Ψj )
Identical to Rayleigh-Ritz eqs.

29
 What about a natural BC?


u + u + x = 0
(S)
u(0) = α u (1) = β.
1
(u + u + x)v dx = 0 ∀v ∈ H01
0
1 1

u v  − u v  − uv − xv dx = 0

0
0
1
βv(1) − u v  − uv − xv dx = 0
0

Find u ∈ Hα1 such that

1
(W) u v  − uv − xv dx − βv(1) = 0 ∀v ∈ H01
0

The Finite Element method

F.E. Model Problem


u + f = 0
(S)
u(a) = G; u (b) = H

b
 
W Residual u + f v dx = 0 ←− can’t plug in Ni (x) =
a

N
uh (x) = ui Ni (x)
i=0
Ni (x) = δ(x − xi−1 ) − 2δ(x − xi ) + δ(x − xi+1 )
Ni (x) =
We must therefore relax the continuity required of the trial so-
lution by going to a weak formulation of the problem.
 b b

 
u v  + −u v  + f v dx = 0

a a
b b
 
∴ u v dx = Hv(b) + f v dx
a a
a(u, v) = Hv(b) + (f, v) ←− (∗)

⎫
(W) Given f ∈ L2 and constants G, H find ⎪
⎪
⎪
u ∈ {u ∈ H 1 , u(a) = g} = Hg1 such that for all ⎪
⎬
v ∈ H01 = {v ∈ H 1 , v(a) = 0} we have that H k = {u : u(k) ∈ L2 } ← Sobolov space
b   b ⎪
⎪
⎪
⎪
u v dx = Hv(b) + f v dx ⎭
a a

30
Claim: If u is a solution of (S) then usatisfies (W ). Conversely, provided u is sufficiently differen-
tiable then if u satisfies (W ) it follows that u also satisfies (S).

(S) ⇒ (W )( See *)
(W ) ⇒ (S)
b b
u v dx = Hv(b) + f v dx for all v ∈ H01
 

a a
b b b

  
u v  − u v dx = hv(b) + f v dx

a a a
b
(u + f )v dx + {H − u (b)}v(b) = 0
a
v arbitrary ⇒ u + f = 0 u (b) = H

↑ natural BC

Using Finite Element Basis Functions: It is useful in the calculation of the integrals to trans-
form each of the subintervals in turn to the same standard interval – on which we construct cardinal
basis functions

N1 N2

x0 xi x xi+1 x3
hi+1 N1 (ξ) N2 (ξ)

−1 0 1
ξ

⎫
N1 (ξ) = 12 (1 − ξ) ⎬
1
Na (ξ) = (1 + ξ ξa )
⎭ 2
N2 (ξ) = 12 (1 + ξ)

−1 a = 1
ξa =
1 a=2

Transformation:
2 2
ξ(x) = c1 + c2 x ξ(xi ) = c1 + c2 xi = −1 c2 = =
xi+1 − xi hi+1
(xi + xi+1 )
ξ(xi+1 ) = c1 + c2 xi+1 = 1 c1 = −
hi+1

2x − (xi + xi+1 )
ξ(x) =
hi+1
hi+1 ξ + (xi + xi+1 )
Similarly x(ξ) =
2

31
Notice we can express the transformation in terms of Na (ξ):
2

x(ξ) = Na (ξ)xa
a=1
2  
dx dNa dNa ξa
= (ξ)xa = (ξ) =
dξ dξ dξ 2
a=1
2
1 1 hi+1
= ξa xa = (−xi + xi+1 ) =
2 2 2
a=1
dξ 2
=
dx hi+1

Galerkin approximation:


N
Let uh (x) = GN0 (x) + Ni (x)ui
i=1

N
v h (x) = Ni (x)vi ←− arbitrary
i=1
   
a uh , v h = hv h (b) + f, v h
b  ⎛ N ⎞
b 
N   N
⇒ 
ui Ni (x) ⎝  ⎠
vi Nj (x) dx = h vj Nj (b) + f vj Nj (x) dx
a i=1 j=1 a j=1

N
 
vj arbitrary ⇒ ui Ni , Nj = hNj (xN ) + (Nj , f ) j = 1, . . . , N
i=1

Ku = f
stiffness matrix ↑
b
 
Kij = Ni , Nj = Ni Nj dx
i−2 i−1 i i+1 i+2
a
fj = hNj (xN ) + (Nj , f )

Notice :

(a) Kij = Kji Symmetric

(b) Kij is positive definite

    b  
2
ui Kij uj = h h
ui a(Ni , Nj )uj = a u , u = u dx ≥ 0 pos. semi def.
1≤ij≤N ij a

32
  
uh = 0
=⇒ uh = const but since uh (0) = 0

not included

=⇒ uh (x) = ui Ni (x) = 0
↑ form a basis ⇒ ui ≡ 0.
∴ u Ku ≥ 0
T
=0⇒u=0

Tridiagonal system - like finite differences.

(c) Kij = 0 if |i − j| ≥ 2

33
The trick to doing the integrals is to calculate them element by element: so called assembly of the
stiffness matrix:

e−2 e−1 e e+1

Element stiffness matrix

b 
N xe
Kij = Ni Nj dx = Ni (x)Nj (x) dx
a e=1 xe−1


N
e
= kab
e=1
xe
e
kab = Na (x)Nb (x) dx
xe−1
1 1
dNa dNb  dx Na (ξ) = (1 + ξ ξa )
= (ξ)ξ  (x) ξ (x) · dξ 2
dξ dξ dξ ξa
−1 Na (ξ) =
2
1
2 ξa ξb ξa ξb (−1)a+b dξ
=
2
=
2
= · dξ = e = dx xe − xe−1 he
he 2 2 h he
−1
⎡ ⎤
1 −1
1 ⎣ ⎦ and kab
∴ [kab
e
] = e
= 0 if a, b ∈ {e − 1, e}
he
−1 1

Assembly:

FIGURE assume he = h ∀ e i.e. uniform mesh

⎡ ⎤⎡ ⎤
2 −1 u1 ⎡ ⎤
⎢ −1 2 −1 ⎥⎢ ⎥ ⎢
⎢ ⎥⎢ ⎥
1⎢ ⎥⎢ . ⎥ ⎢ ⎥
⎢ .. ⎥⎢ . ⎥ ⎢
=⎢ F ⎥
⎢
h⎢
. ⎥⎢ . ⎥⎥ ⎣ ⎥
.. ⎥⎣ ⎦ ⎦
⎣ . ⎦
−1 1 un

To calculate the force for a general f (x):

b
fj = f (x)Nj (x) dx
a

We assume f (x)  f h (x) = fj Nj (x)
j

The force vectors are also assembled

 
xe

fj = f (x)Na (x) dx
e=1 x
e−1

34
 2
 xe
fae = fb Nb (x)Na (x)dx
b=1 xe−1

Simple Examples

(1) N = 1 N0
N1
(a) x
 1−x
u = 0 f ≡0
ue = G + Hx
u(0) = G u (1) = G
 1
k11 = ·1=1
1 1
0
1
f1 = H · 1 + x · 0 dx = H
0
−u0 + u1 = f1
⇒ u1 = H + G
∴ uh=1 (x) = GN0 (x) + (G + H)N1 (x) = G + Hx

(b)

f (x) = p
u + p = 0 up = Ax2 2A + p = 0
p
ue (x) = α + βx − x2
2
u(0) = G = α
u (x) = β − px u (1) = β − p = H β = (p + H)
 
x2
ue (x) = G + Hx + p x −
2
1
p
f1 = H + x p dx = H +
2
0
p
k11 u1 = H + +G
2
px does not capture exact solution because
⇒ uh (x) = G + Hx +
2 ue ∈ {set of piecewise linear functions on (0, 1)} .
p
however uh (0) = G uh (1) = G + H + = ue (1) exact at nodes
2
(c)

f = qx
 
x x2
ue (x) = G + Hx + q 1−
2 3
1
q
f1 = H + q x2 dx = H +
3
0
q
u1 = H + +G
3

35
 qx
uh (x) = G + Hx +
3
h q
u (1) = G + H + = ue (1) Galerkin exact at nodes
3
1
(2) N = 2 h= 2
(a) f ≡ 0: u0 = G N1 N2
⎡ ⎤
⎡ ⎤⎡ ⎤ 1
1 ⎣
2 −1 u1 ⎢ HN1 (1) + 0 N1 · 0 dx + 2G ⎥
⎦⎣ ⎦=⎢ ⎥
(1/2) ⎣ 1 ⎦
−1 1 u2 HN2 (1) + N2 · 0 dx
x1 x2 = 1
0
⎡ ⎤⎡ ⎤ ⎡ ⎤ h = 1/2
4 −2 u1 2G
⎣ ⎦⎣ ⎦=⎣ ⎦
−2 2 u2 H

H
u1 = G +
2
u2 = G+H
 
h H
u (x) = GN0 + G + N1 (x) + (G + H)N2 (x)
2
= G + Hx

(b) f = p:

⎡ ⎤
1
⎡ ⎤⎡ ⎤
⎢ 2G + p N1 dx ⎥ ⎡ ⎤
4 −2 ⎢ u1 0 ⎥ 2G + p/2
⎣ ⎦⎣ ⎦=⎢
⎢
⎥ ⎣
⎥= ⎦
⎢ 1 ⎥
−2 2 u2 ⎣ ⎦ H + p/4
H + p N2 dx
0
⎡ ⎤ ⎡ ⎤
u1 3p/8 + G + H/2
⎣ ⎦=⎣ ⎦
u2 p/2 + G + H

uh (x) = N0 G + N1 (x) (3p/8 + G + H/2) + N2 (x) (P/2 + G + H)

3p p
= G + Hx + N1 · + N2
8 2
H 3p
uh (1/2) = G + + = uexact (1/2)
2 8
uexact (x) = (G + Hx + p(x − x2 /2)

(c)
f = qx
h 11q q
u N1 (x) + N2 (x)
= G + Hx +
48 3
H 11q
uh (1/2) = G + + = uexact (1/2)
2 48

36
Using a finite difference approximation
N =2

u + f = 0 f =p
u(0) = G u (1) = H
1 2 3
0

U2 − 2U1 + G 1 G
n=1: = −p [2U1 − U2 ] = hp +
h2 h h
U3 − 2U2 + U1 U3 − U1
n=2: = −p = H ⇒ U3 = U1 + 2hH
h2 2h
2U2 − 2U1 − 2hH h
∴ = hp =⇒ U2 − U1 = H + p
h 2
⎡ ⎤⎡ ⎤
2 −1 U1  
1 ⎦ = G/h + hp
∴ ⎣ ⎦⎣ h = 1/2
h H + h2 p
−1 1 U2

Exactly the same equations as the FEM.

Some convergence results for the F.E. model problem:

We consider 
u + f = 0
(1)
u(a) = G u,x (b) = G

(A) The Green’s function for (1)

Find G(x, x ) : G,xx (x, x ) + δ(x − x ) = 0

G(a, x ) = 0 = G,x (b, x )

x 
 0 x < x
δ(x − x )dx =
1 x > x
−∞
G,x + H(x − x ) = c1
= H(x − x )
G + x − x  = c1 x + c2 
x
(i) G,x (b) + H(b − x ) = c1  0 x < x
H(x − x )dx =
(x − x ) x ≥ x
0 + 1 = c1 −∞
(ii) G(a) + a − x  = a + c2
0 + 0 = a + c2
G = x − a
∴ G(x, x ) = (x − a) − x − x  x − a
G = (x − a)

a x b

Note that G ∈ c0 is piecewise linear.

37
Theorem 1: uh (x) the PWL Galerkin approximation is exact at the nodes; i.e., uh (xi ) = uex (xi )

Notation: Let v h = {f ∈ span {N1 , . . . , NN } where Ni ∈ H0 }.

Lemma 1: a(ω h , u − uh ) = 0 ∀ ωh ∈ vh

Proof:
u → classical or strong solution
(W) ⇒ a(u, ω) = (f, ω) + Hω(b) ∀ ω ∈ H0 (∗)
(S) ⇒ (W)
 
v h = ω h = ci Nij ; Ni ∈ H0
(G) ⇒ a(uh , ω h ) = (f, ω h ) + Hω h (b) (∗∗) ∀ω ∈ h
h
v ⊂ H0
∴ (∗) ⇒ a(u, ω h ) = (f, ω h ) + Hω h (b) (∗ ∗ ∗)

Subtract (∗∗) from (∗ ∗ ∗):

a(u − uh , ω h ) = 0 ∀ ωh ∈ vh
a(ω h , u − uh ) = a(u − uh , ω h ) symmetry

Lemma 2: a(G(x − x ), ω) = (δ(x − x ), ω) ∀ ω ∈ H0

Proof:

Gxx + δ(x − x ) = 0
G(a) = 0 = Gx (b)
b  b b

 
0= Gxx + δ(x − x ) ω dx = Gx ω  + −Gx ωx + δω dx ∀ ω ∈ H0
a
0 a

∴ a(G, ω) = (δ(x − x ), ω) ∀ω ∈ H0

Lemma 3: u(x ) − uh (x ) = a(G(x − x ) − ω h , u − uh ) ∀ ωh ∈ vh .

Proof: u − uh ∈ H0
u ∈ Hg1 and uh ∈ v h + {gN0 } ⇒ uh (a) = u(a) = g ⇒ u − uh ∈ H0

∈ H0
   
a G x, x − ω h , u − uh = a(G, u − uh ) − a(ω h , u − uh )
= (δ(x − x ), u − uh ) − 0
L(2) L(1)
= u(x ) − uh (x )

Proof of Theorem: u(xi ) = uh (xi )

 
Let x = xi . Then L(3) ⇒ u(xi ) − uh (xi ) = a G (x, xi ) − ω h , u − uh but x = xi ⇒ G(x, xi ) =
 L(1)
ci Ni (x) ← PWL basis functions on net {x0 , . . . , xN } ⇒ G ∈ v h ⇒ G−ω h ∈ v h =⇒ u(xi )−uh (xi ) =
0 

38
Notes:

1. True no matter how large or small N is.

2. Not true for F.D. equations

3. For this problem it is as if we knew the exact solution and fed this information in at the
nodes. Thus the only error in the Galerkin approximation is the interpolation error:

h2  h2
∴ ||u − uh ||∞ ≤ ||u ||∞ = ||f ||∞ since u + f = 0.
8 8

4. It is not possible to use this technique for all problems as it relied on the fact that G ∈ v h which
is not necessarily true for all problems. For further analysis see Strang and Fix, p. 39–51.

||f  ||∞ = max sup |u (x)| <∞

i,j x∈(x ,x )
i j


N
Theorem: (Error in PWL approximation) Let f ∈ P C 2,∞ (0, 1) then ||f − fi Ni (x)|| ≤
i=0
1 2 
8 h ||f ||∞

Proof:
f (x) fi+1

Σfi Ni (x)
fi for any 0 ≤ i ≤ N

xi xi+1

Let w(x) = (x − xi )(x − xi+1 ) = x2 − (xi + xi+1 )x + xi xi+1

      
xi + xi+1 xi + xi+1 xi+1 − xi xi − xi+1
w = 2x − (xi + xi+1 ) = 0 x = w =
2 2 2 2

N
e(x) = f (x) − Ni (x)fi
i=0

Claim: For each x ∈ [xi , xi+1 ] ∃ ξx ∈ [xi , xi+1 ] : e(x) = 12 f  (ξx )w(x)

Proof:
1) x = xi , xi+1 any ξx suffices.
2) Choose an arbitrary x = x̃. Choose λ such that

θ(x̃) = e(x̃) − λw(x̃) = 0

39
 0 0 0
x x θ
then θ(x) has 3 zeros on [xi , xi+1 ] and
θ = 0
Rolle ⇒ ∃ ξx ∈ [xi , xi+1 ] : θ (ξx ) = 0 0 0 0
θ” = 0
0

But

θ (ξx ) = f  (ξx ) − 2λ = 0

1 
∴λ = f (ξx )
2
1  1
∴ max |e(x)| ≤ ||f ||∞ max |w(x)| ≤ h2 ||f  ||∞
x∈[xi ,xi+1 ] 2 x∈[xi ,xi+1 ] 8


Theorem 2: Assume u ∈ c1 [a, b]. Then there exists at least one point c ∈ (a, b) at which uh,x (c) =
uexjx (c).

Proof:

MVT ⇒ ∃c ∈ (xi , xi+1 ) : u(xi+1 ) = u(xi ) + (xi+1 − xi )u,x (c)

|| ||
h
u (xi+1 ) uh (xi ) + (xi+1 − xi )u,x (c)
uh (xi+1 ) − uh (xi )
∴ u,x (c) = = uh,x (x)
(xi+1 − xi )
1 1 uh (xi+1 ) − uh (xi )
Ni (x) = − 
Ni+1 (x) = ∴ uh,x (x) =
xi+1 − xi xi+1 − xi (xi+1 − xi )

The Barlow points:

If we don’t know c then for linear elements uh,x at the midpoints are optimally accurate:

Assume u ∈ C 3 (a, b) and expand u about x = α ∈ [xi , xi+1 ]:

1 1
u(xi+1 ) = u(α) + (xi+1 − α)u (α) + (xi+1 − α)2 u (α) + (xi+1 − α)3 u (c1 )
2 3!
1 1
u(xi ) = u(α) + (xi − α)u (α) + (xi − α)2 u (α) + (xi − α)3 u(3) (c2 )
2 3!
Subtract and divide by xi+1 − xi = hi+1

u(xi+1 ) − u(xi ) 1 (xi+1 − xi ) 

= u (α) + (xi+1 + xi − 2α) u (α) + O(h2i )
xi+1 − xi 2 xi+1 − xi

uh (xi+1 ) − uh (xi ) 1
If α = xi then uh,x (xi ) = = u (xi ) + hi+1 u (xi ) + O(h2i )
xi+1 − xi 2
 
xi+1 + xi xi+1 + xi
If we choose α = then uh,x (x) = u + O(h2i )
2 2

Conclusion:
For linear elements the midpoints are superconvergent with respect to derivatives, i.e. most accu-
rate derivatives are calculated there. These are called the BARLOW points. (See Barlow, Int. J.

40
of Num. Meth. Eng., p. 243–251, 1976.

(B) Higher order elements:

Lagrange elements:
We use the Lagrange polynomials to construct the basis functions:

(
d+1 )(
d
Lda (ξ) = (ξ − ξb ) (ξa − ξb )
b=1 b=1
b=a b=a

Note: Lda (ξc ) = δac just the right properties

d = 1: FIGURE
(ξ − 1) 1
L11 (ξ) = = (1 − ξ) = N11 (ξ)
(−1 − 1) 2

(ξ − (−1)) 1
L12 (ξ) = = (1 + ξ) = N21 (ξ)
1 − (−1) 2

d = 2: FIGURE

(ξ − 0) (ξ − 1) 1 N12 (ξ)
L21 (ξ) = = ξ(ξ − 1) = N12 (ξ)
(−1 − 0) (−1 − 1) 2
−1 0 1

(ξ − (−1))(ξ − 1) N22 (ξ)

L22 (ξ) = = (1 − ξ 2 ) = N22 (ξ)
(0 − (−1))(0 − 1)
−1 0 1

(ξ − (−1))ξ 1 N32 (ξ)

L23 (ξ) = = ξ(1 + ξ) = N32 (ξ)
(1 − (−1))(1 − 0) 2
−1 0 1

3
 1 1 1 1
Note: Nk2 (ξ) = ξ 2 − ξ + 1 − ξ 2 + ξ + ξ 2 = 1
2 2 2 2
k=1

Can represent a constant function (or rigid body motion) exactly.

41
d=3:
1
L31 (ξ) = (1 − ξ)(9ξ 2 − 1) = N13 (ξ)
16
9
L32 (ξ) = (3ξ − 1)(ξ 2 − 1) = N23 (ξ)
16 ξ1 ξ2 ξ3 ξ4

9 −1 −1/3 1/3 1
L33 (ξ) = − (3ξ + 1)(ξ 2 − 1) = N33 (ξ)
16
1
L34 (ξ) = (1 + ξ)(9ξ 2 − 1) = N43 (ξ)
16

42
Example:

u + f = 0 xi−1 1/2(xi + xi−1 ) xi


u(0) = G u (1) = H xm

N
uh = ui Ni (x) ξ= 2
− xm )
hi (x
i=0
dξ 2
u = NT u dx = hi

(W) ⇒ a(uh , v h ) = Hv h (b) + (f, v h )

xe
e
kab = Na (x)Nb (x) dx
xe−1
1
2 2 h
= Na (ξ) · Nb (ξ) · dξ
h h 2
−1
1  2 1 1

2 1 1 2 4 3  7
e
k11 = (2ξ − 1) dξ = 2
4ξ − 4ξ + 1 dξ = ξ + ξ  =
h 2 2h 2h 3 0 3h
−1 −1
1 1 
2 1 2 2 2 ξ 3 1 8
e
k12 = (2ξ − 1)(−2ξ)dξ = − 2ξ − ξ dξ = − · 4  = −
h 2 h h 3 0 3h
−1 −1
1 1  1
2 1 1 2 2 1 4ξ 3 1
e
k13 = (2ξ − 1) (2ξ + 1)dξ = + ·2 4ξ − 1dξ = −ξ =
h 2 2 4h h 3 0 3h
−1 0
1 
e 2 2 8 ξ 3 1 16
k22 = (−2ξ) dξ = 2  =
h h 3 0 3h
−1
⎡ ⎤
7 −8 1
1 ⎣ −8 16 −8 ⎦
[k e ] =
3h
48 −8 7
 2

Let f = p :⇒ ue (x) = G + Hx + p x − x2

For 1 element h = 1
0 2
u0 u1 u2
⎡ ⎤⎡ ⎤ ⎡ 8 2 ⎤
16 −8 u1 3G + 3 ·p
1 ⎣ ⎦⎣ ⎦=⎣ ⎦
3 p
−8 7 u2 H+ 6 − G
3
xe 1  1
h ξ3 h
f1 = p · N1 (x)dx = (1 − ξ 2 )dξ = h ξ − =2 p
2 3 0 3
xe−1 −1
1 1
h 1 2 h ξ 3  h
f2 = H+ (ξ + ξ )dξ = H + p 2  = H + p
2 2 4 3 6
−1 0

43
 ⎡ ⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
u1 7/48 1/6 8G + 2p 7G/6 − G/6 + 7/24p + H/2 + 1/12p
∴ ⎣ ⎦=⎣ ⎦⎣ ⎦=⎣ ⎦
u2 1/6 1/3 −G + 3H + 1/2p 4G/3 + 1/3p − G/3 + H + p/16
⎡ ⎤
G + 3/8p + H/2
N1 (x) = 4x(1 − x)
=⎣ ⎦
N2 (x) = x(2x − 1)
G + 1/2p + H
 
3
∴ uh (x) = GN0 (x) + G + H/2 + p N1 (x) + (G + p/2 + H) N2 (x)
8
 
3 p
= G + (H/2N1 (x) + HN2 (x)) + pN1 (x) + N2 (x)
8 2
   
1  3   1
= G+H 4x − 4x2 + 2x2 − x + p 4x − 4x2 + (2x2 − x)
2 8 2
p 2 2

= G + Hx + 3x − 3x + 2x − x
2
p 
= G + Hx + 2x − x2
2 
x2
= G + Hx + p x − the exact solution
2

44