Norms and spaces

Definition:
The space of all square
L2 (Ω) = integrable funcions defined in
the domain
Ω
is a finite number
f ∈ L2 ( Ω ) ⇒ ∫Ω
f 2 dx not infinity

(2,2)
L2 norm of f Example compute
2
f L2 ( Ω )
f ( x, y ) = x + y
2 2
2

(∫ )
f L2 ( Ω )
= ∫ ∫ ( x + y 2 ) 2 dxdy
1/ 2 (0,0)
2
f = f dx
0 0
2
L2 ( Ω ) Ω =∫ (
8
3
)
+ 4 y 2 + 2 y 4 dy = 144
5 12
0 f L2 ( Ω )
= 5
 Norms and spaces

Definition:
H 1 (Ω) = {u : u , u x , u y ∈ L2 (Ω)} Ω
The function and its first derivatives are
square integrable

Example
Def:
f ( x, y ) = x + y 2
(2,2) ∇v = vx
2
+ vy
2
L2 ( Ω ) L2 (Ω) L2 ( Ω )
2 2
2
f L2 ( Ω )
= ∫ ∫ ( x + y 2 ) 2 dxdy = 144
5 2 2
0 0 ∇f L2 ( Ω )
= fx L2 ( Ω )
+ fy = 4 + 643
L2 ( Ω )
2 2
2
fx = ∫ ∫ (1) 2 dxdy = 4 (0,0)
L2 ( Ω )
0 0 Def: v = v 2L ( Ω ) + ∇v 2L ( Ω )
H 1 (Ω) 2 2
2 2
2
fy = ∫ ∫ (2 y ) 2 dxdy = 64
f ∈ H 1 (Ω)
L2 ( Ω )
0 0
3
f H 1 (Ω)
= 28
3
+ 4+
64
3
 Norms and spaces

Def:
2 2
Energy norm
∇v L2 ( Ω )
= vx L2 ( Ω )
+ vy
L2 ( Ω )

∇f L2 ( Ω )
= fx
2
L2 (Ω)
+ fy
2

L2 ( Ω )
= 4 + 643 ||| v |||= ∇v L2 ( Ω )

Example f ( x, y ) = x + y 2 Example
(2,2) Compute the energy
2
2 2 norm of the function
f L2 ( Ω )
= ∫ ∫ ( x + y 2 ) 2 dxdy = 28
3
0 0

2 2
v( x, y ) = x 2 + 3 xy
2
fx L2 ( Ω )
= ∫ ∫ (1) 2 dxdy = 4 (0,0)
0 0

2 2
2
fy = ∫ ∫ (2 y ) 2 dxdy = 64
3
L2 ( Ω )
0 0
Norms and spaces

Definition: H 01 (Ω) = {u : u ∈ H 1 (Ω), u = 0 on ∂Ω}

Example (2,2)

Which of the following

belongs to H 01 (Ω) Ω
v( x, y ) = x 2 + 3 xy (0,0)
u ( x, y ) = x( x − 2) y ( y − 2)
 Norms and spaces

Triangle inequality (Real numbers) a+b ≤ a + b

Triangle inequality

u+v L2 ( Ω )
≤ u L2 ( Ω )
+ v L2 ( Ω )

Cauchy-Schwartz inequality Cauchy-Schwartz inequality

∫ uv ≤ u L2
v L2 ∫
Ω
uv ≤ ||| u ||| ||| v |||
Ω
 Norms and spaces

Definition:

v ∈ C 0 (Ω), 
Vh =  
v
 K ∈ P 1 ( K ) for all triangles K 
the space of all continuous
piecewise linear polynomials

Definition: Relations:

Vh , 0 = {v ∈ Vh : v = 0 on ∂Ω} Vh H 1 (Ω)
the space of all continuous
piecewise linear polynomials Vh , 0 H 01 (Ω)
which vanishes on the boundary
 Mesh size h

local mesh size mesh size Example:

the length of the

h = max hK h = max{hK 1 ,  , hK 7 }
hK = K
longest edge on K h=4

Example: (0,4) (2,4) (3,4) (4,4)

5 4
hK 1 = 8 hK 3 = 4 K6 K5
K7
hK 2 = 8 hK 4 = 4 K1
(0,2) 1
K4
hK 5 = 8 hK 6 = 5 K2
hK 7 = 8
K3
2 3
(0,0) (4,0)
A Priori Error Estimates
− ∆u = f in Ω
u=0 on ∂Ω a (u , v) = ∫ ∇u ⋅ ∇v
Ω

Find u ∈ H 01 (Ω) such that Find uh ∈ Vh , 0 such that

a (u , v) = F (v) ∀v ∈ H 01 (Ω) a (uh , vh ) = F (vh ) ∀vh ∈ Vh , 0

u uh

e = u − uh = ? in what sense the error

e becomes small
A Priori Error Estimates
variational formulation finite element method

Find u ∈ H 01 (Ω) such that Find uh ∈ Vh , 0 such that

a (u , v) = F (v) ∀v ∈ H 01 (Ω) a (uh , vh ) = F (vh ) ∀vh ∈ Vh , 0

∫ ∫
1
∇ u ⋅ ∇ v = fv ∀v ∈ H (Ω)
Ω Ω
0
∫ ∇u
Ω
h ⋅ ∇vh = ∫ fvh ∀vh ∈ Vh , 0
Ω

e = u − uh = ?
Galerkin Orthogonality
Theorem 1 (Galerkin Orthogonality).
The finite element approximation uh, satisfies the orthogonality

Ω
∫ ∇(u − u h ) ⋅∇vh = 0, for all vh ∈ Vh , 0

Proof: from the variational formulation we have: variational formulation

∫ ∇u
Ω
h ⋅ ∇vh = F (vh ) ∀vh ∈ Vh , 0 ∫ ∇u h ⋅ ∇vh = ∫ fvh ∀vh ∈ Vh , 0
Ω Ω

∫ ∇u ⋅ ∇v
Ω
h = F (vh ) ∀vh ∈ Vh , 0 Vh , 0 ⊂ H 01 (Ω)

finite element method

subtract
∫ ∫
1
∇ u ⋅ ∇ v = fv ∀v ∈ H 0 (Ω)
Ω
∫ ∇(u − u h ) ⋅ ∇vh = 0 ∀vh ∈ Vh , 0 Ω Ω
Best Approximation
Theorem 2 (Best Approximation).
The finite element approximation uh, satisfies

||| u − uh ||| ≤ ||| u − vh |||, for all vh ∈ Vh , 0

Proof: for any vh ∈ Vh , 0 we have

||| u − u h |||2 = ∫ ∇(u − uh ) ⋅ ∇(u − uh )

Ω This shows that the finite
element solution uh is the
= ∫ ∇(u − uh ) ⋅ ∇(u − vh + vh − uh ) closest of all functions in Vh
Ω to the exact solution u when
measuring distance using
= ∫ ∇(u − uh ) ⋅ ∇(u − vh ) + ∫ ∇(u − uh ) ⋅ ∇(vh − uh ) the energy norm.
Ω Ω

= ∫ ∇(u − uh ) ⋅ ∇(u − vh )
Ω

≤ ||| u − uh ||| ||| u − vh |||

 Energy Norm Error
Theorem 3 (error depends on meshsize).
The finite element approximation uh, satisfies the a priori error estimate
2
||| u − uh ||| ≤ 2
∑ ChK
2
K
2
Du 2
L (K )
This shows that how
the error depends on
the mesh size.
with a constant C independent of hK

Proof: Start from the best approximation Remark

2
||| u − uh ||| ≤ ||| u − v ||| 2
||| u − u h ||| ≤ ∑ Ch
K
2
K Du 2
L2 ( K )

choosing v= πu 2
≤ Ch 2 ∑ D 2u
L2 ( K )
||| u − uh ||| ≤ ||| u − π u |||
2 2 K

2 2 2
≤ Ch D u
∑ ∇(u − π u )
2
= L2 ( K )
L2 ( Ω )
K
||| u − u h ||| ≤ Ch D 2u
2 L2 ( Ω )
≤ ∑ Ch
K
2
K Du 2
L2 ( K )
the gradient of the error tends to zero as
the maximum mesh size h tend to zero.
Error Depends on h
Theorem (Poincare Inequality). For any function v ∈ H 01 (Ω)

v L2 ( Ω )
≤ C ∇v L2 ( Ω )
= C ||| v |||

Remark
||| u − u h ||| ≤ Ch D 2u
L2 ( Ω )
These constants
C’s are different
u − uh L2 ( Ω )
≤ C ||| u − uh |||

≤ Ch D 2u
L2 ( Ω )

u-uh 2
L (Ω)
≤ Ch D 2u
L2 ( Ω )
 L2 error is h2
we expect that the L2 error to be h2 and not h

Theorem 4 (The L2-error)

The finite element approximation uh, satisfies the a priori error estimate

u − uh L2 ( Ω )
≤ Ch 2 D 2u
L2 ( Ω )
with a constant C independent of h
The proof uses a well-known technique called Nitsche’s trick,
Main problem Dual problem or adjoint problem
− ∆u = f in Ω − ∆φ = e in Ω
u=0 on ∂Ω φ =0 on ∂Ω
 Continuous Piecewise Linear Interpolation

Definition: Let u ∈ C 0 (Ω)

we define its continuous piecewise linear
interpolant π u ∈ V by
n
h
u
π u = ∑ u ( N k )ϕ k
k =1

Remark:
π u approximates u by taking on the
same values in the nodes Ni.

πu
∇(u − πu ) ≤ Ch D u 2
 L2 error is h2
Theorem 4 (The L2-error)
The finite element approximation uh, satisfies the a priori error estimate
2 2
u − uh
L2 ( Ω )
≤ Ch D u
L2 ( Ω )
with a constant C independent of h

Proof: let ϕ be the solution Cauchy-Schwartz inequality 2

e ≤ Ch 2 e D 2u
of the dual problem
2
− ∆φ = e in Ω e = ∫ ∇e∇(φ − πφ )
Ω
Dividing by ∥e∥
φ =0 on ∂Ω ≤ ∇e ∇(φ − πφ ) e ≤ Ch 2 D 2u
Multiplying by e and integrating
≤ ∇e ch D 2φ
using Green’s formula
u − uh ≤ Ch 2 D 2u
≤ ∇e Ch ∆φ
∫ e = − ∫ e∆φ
2 L2 ( Ω ) L2 ( Ω )

( )
Ω Ω
≤ Ch D 2u Ch ∆φ
= ∫ ∇e∇φ − ∫ en ⋅ ∇φ
Ω ∂Ω ≤ (Ch D u ) Ch e
2

= ∫ ∇e∇φ ≤ Ch 2 e D 2u
Ω

= ∫ ∇e∇(φ − πφ )
Ω
Error norms
Theorem (Energy Norm)
The finite element approximation uh, satisfies the a priori error estimate

||| u − u h ||| ≤ Ch D 2u
L2 ( Ω )

with a constant C independent of hK

Theorem (L2 Norm)

The finite element approximation uh, satisfies the a priori error estimate
2 2
u − uh 2
L (Ω)
≤ Ch D u
L2 ( Ω )
with a constant C independent of h
Calcuate the L2 error
∑
2 2
u − uh L2 ( Ω )
= u − uh L2 ( K )
K

= ∑ ∫ (u − u
K K
h )2

= ∑ (u ( x , y ) − u h ( x , y ) )2
⋅| K |
K

||| u − u h |||2 = ∑ ||| u − u

K
h |||2K

= ∑∫ x x ∫ y y
(u
K K
− uh ) 2
+ (u − uh ) 2

= (
∑ x
u ( x , y ) − uhx ( x , y ) )2
⋅ | K | + (u y ( x , y ) − uh y ( x , y ) )2
⋅| K |
K
Calcuate the L2 error
− ∆u = f in Ω u − uh ≤ Ch 2 D 2u
u=0 on ∂Ω 2
L (Ω) L2 ( Ω )

Ω = (−1,1) × (−1,1)
f ( x, y ) = −2( x 2 + y 2 − 2)
Exact solution
u ( x, y ) = ( x 2 − 1)( y 2 − 1)
h1 = 0.2 h2 = 0.1

u − uh1 ≈ Ch1α u − u h 2 ≈ Ch2α

u − uh 2 h2α u − uh 2 h 
α
≈ α ≈  2 
u − uh1 h1 u − u h1  h1 

 e2   
α ≈ ln  / ln h2 
 h 
 e1   1 
function [l2error] = compute_error(p, t, uh, u, u_x, u_y)
% calculates the error
function [p,e,t,uh] = solveE(hmax)
nt = size(t, 2); % number of triangles % Poisson's equation on a square [-1,1]X[-1,1] is solved
np = size(p,2); % and the resulting finite element solution is stored in
t1 = t(1,:); t2 = t(2,:); t3 = t(3,:); % the vector uh.
x1 = p(1, t1); x2 = p(1, t2); x3 = p(1, t3); %
y1 = p(2, t1); y2 = p(2, t2); y3 = p(2, t3);
g = 'squareg'; % domain
xc = (x1 + x2 + x3)/3; % x-coord of element midpoints b = 'squareb1'; % Dirichlet data
yc = (y1 + y2 + y3)/3; % y-coord of element midpoints f = '-2*((x.^2 - 1) + (y.^2 - 1))'; % right hand side
exact = feval(u, xc, yc); % exact sol at the midpoints
uhc = pdeintrp(p, t, uh); % FE sol at the midpoints [p, e, t] = initmesh(g,'Hmax',hmax); % triangulation
l2error2 =0; uh = assempde(b, p, e, t, 1, 0, f); % solve pde
for K = 1:nt
loc2glob = t(1:3,K); u = inline('(x.^2 - 1).*(y.^2 - 1)', 'x', 'y'); % exact sol
x = p(1,loc2glob); y = p(2,loc2glob); u_x = inline('2*x.*(y.^2 - 1)', 'x', 'y'); % u_x exact
area = polyarea(x,y); u_y = inline('2*y.*(x.^2 - 1)', 'x', 'y'); % u_y exact
loc_er = (exact(loc2glob) - uhc(loc2glob)).^2;
l2error2 = l2error2 + sum(loc_er)*area h1=0.2;
end [p1,e1,t1,uh1] = solveE(h1);
l2error = sqrt(l2error2); [l2error1] = compute_error(p1, t1, uh1, u, u_x, u_y)

[exact' , uhc', abs(exact-uhc)']

h2=0.1;
[p2,e2,t2,uh2] = solveE(h2);
[l2error2] = compute_error(p2, t2, uh2, u, u_x, u_y)

log(l2error1/l2error2)/log(h1/h2)
Calcuate the energy norm of the error

Exercise
a) Write matlab file to compute the energy norm of the error

function [energy_error] = energy_error(p, t, uh, u, u_x, u_y)

b) Write matlab file to compute the energy norm of the error for
two different meshes with meshsizes h1 and h2 and verify the rate
of convergence

||| u − u h ||| ≤ Ch D 2u
L2 ( Ω )