Energy Norm Error Estimates

Mats G Larson
[email protected]

Computational Mathematics,
and
Fraunhofer Centre for Industrial Mathematics

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.1

Outline
• Residual based estimates
• Local problems
• Recovery estimates
• Lower bounds
• Refinement techniques
• Adaptive algorithms
• Other sources of error

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.2

Poisson Equation
Strong form. Find u ∈ H01 (Ω) such that

−4u = f in Ω,
u = 0 on ∂Ω.

where f ∈ H −1 (Ω), and Ω is the domain.

Weak form. Find u ∈ H01 (Ω) such that
Z Z
∇u · ∇v dx = f v dx for all v ∈ H01 (Ω).
Ω Ω

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.3

Discretization
We introduce a finite dimensional space Vh ⊂ H01 (Ω).
• The mesh consists of elements K ∈ K.

• hK = diam(K) and h(x) = hK for x ∈ K.

• The functions v ∈ Vh are piecewise polynomials.

Finite element method. Find U ∈ Vh such that

Z Z
∇U · ∇v dx = f v dx for all v ∈ Vh .
Ω Ω

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.4

Galerkin Orthogonality
We subtract the finite element equations from the
weak form and obtain the following equations for the
error e = u − U ,
Z
∇e · ∇v dx = 0 for all v ∈ Vh ,
Ω

i.e. the error e is orthogonal to the space Vh in this

sense. We let πv be the interpolant of v.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.5

Residual Based Estimate
Z
k∇ek2 = ∇e · ∇e dx
ZΩ
= ∇e · ∇(e − πe) dx
Ω
Z
= − 4e(e − πe) dx
Z Ω
= (f + 4U )(e − πe) dx
Ω
R
We need to interpret Ω (f + 4U )v dx.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.6

Residual Based Estimate, cnt
Z XZ
(f + 4U )v dx = (f + 4U )v dx
Ω K∈K K
Z
− 0.5 [n · ∇U ]v ds
∂K\Γ
X
−1
≤ kh(f + 4U )kK kh vkK
K∈K

+ 0.5kh1/2 [n · ∇U ]k∂K\Γ kh−1/2 vk∂K\Γ

we let v = e − πe.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.7

Residual Based Estimate, cnt
Interpolation estimate.
X
kh−1 (e − πe)k2K ≤ Ckek21 ≤ Ck∇ek2,
K∈K

where the last inequality is due to Poincare-Friedrichs

Lemma.
Trace inequality.

kh−1/2 (e − πe)k2∂K ≤ Ckh−1 (e − πe)kK k(e − πe)kK,1

X
kh−1/2 (e − πe)k2∂K ≤ Ck∇ek2
K∈K

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.8

 Residual Based Estimate, cnt
Remember
Z
k∇ek2 = (f + 4U )(e − πe) dx
Ω
XD 1
2
≤ kh(f + 4U )kK + kh−1 e − πek2K
2 2D
K∈K
D 1/2 2 1
+ k0.5 ∗ h [n · ∇U ]k∂K\Γ + kh−1/2 e − πek2∂K\Γ
2 2D
DX
≤ kh(f + 4U )k2K + k0.5 ∗ h1/2 [n · ∇U ]k2∂K\Γ
2
K∈K
1
+ k∇ek2 , D = 2C
2
Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.9
Residual Based Estimate, cnt
We end up with,
X
k∇ek2 ≤ C 2
RK ,
K∈K

2
where RK = h2K kf + 4U k2K + 14 hK k[n · ∇U ]k2∂K\Γ

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.10

Residual Based Estimate, cnt
We can extend this result to the case of variable
coefficient,
−∇ · a∇u = f
in this case we get
√ X
k a∇ek2 ≤ C 2
RK ,
K∈K

where
2
RK = h2K k √1a f +∇·a∇U k2K + 41 hK k √1a [n·a∇U ]k2∂K\Γ

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.11

Residual Based Estimate, cnt
• Simple to implement
• Correct h-dependence of the error
• Unknown constants appear

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.12

Local Problems on Stars
Next we turn to estimates on mesh stars.
Nochetto-Carstensen derives the following
R technique
for the energy norm, let (v, w) = Ω vw dx,

k∇ek2 = (−4e, e)
= (f + 4U, e)
Xn
= (ϕi (f + 4U ), e)
i=1
Xn
= (ϕi ∇Ei , ∇e)Si
i=1

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.13

Local Problems on Stars, cnt
The functions Ei ∈ H 1 (Si ) are defined to solve,
(ϕi ∇Ei , ∇v)Si = (ϕi (f + 4U ), v)Si ,

for all v ∈ H 1 (Si ). This equation is solvable since the

kernel consists of v = C and (ϕi (f + 4U ), C)Si = 0
from Galerkin Orthogonality.
1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1

Figure 1: The mesh star Si .

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.14
Local Problems on Stars, cnt
We get the following estimate,
X
n X
n
k∇ek2 = (ϕi ∇Ei , ∇e)Si = ( ϕi ∇Ei , ∇e)
i=1 i=1

By using the Cauchy-Schwarz inequality we get,

X
n X Xn
k∇ek2 ≤ k ϕi ∇Ei k2 = k ϕi ∇Ei k2K
i=1 K∈K i=1

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.15

Local Problems on Stars, cnt
We perform a similar calculation to get a simpler
version.
k∇ek2 = (−4e, e) = (f + 4U, e)
Xn X
n
= (ϕi (f + 4U ), e) = (∇Ei , ∇e)Si
i=1 i=1

Here Ei is determined by,

(∇Ei , ∇v)Si = (ϕi (f + 4U ), v)Si ,
P Pn
then we get, k∇ek ≤ K∈K k∇ ( i=1 Ei ) k2K .
2

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.16

Local Problems on Stars, cnt
We can also solve the equations for Ei on stars of
more layers.
1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

j
Figure 2: One and two layer mesh stars Si j = 1, 2.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.17

Local Problems on Elements
We now use Neumann conditions on the local
problems,

k∇ek2 = (∇e, ∇e)

X
= (∇e, ∇e)K
K∈K
X
= (f + 4U, e)K + (n · ∇e, e)∂K
K∈K
X
= (f + 4U, e)K + (Σn − n · ∇U, e)∂K ,
K∈K

where Σn is a conservative approximation of n · ∇u.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.18
Local Problems on Elements
We define EK ∈ H 1 (K) to solve,
(∇EK , ∇v)K = (f + 4U, v)K + (Σn − n · ∇U, v)∂K ,

for all v ∈ H 1 (K). This equation is solvable if

(f + 4U, C)K + (Σn − n · ∇U, C)∂K = 0,
for a constant C i.e.
Z Z
f+ Σn = 0.
K ∂K

We get theP
following estimate
k∇ek2 ≤ K∈K k∇EK k2 .
Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.19
Local Problems on Elements
The Neumann approach requires
• Computation of equilibrium fluxes Σn

• Solution of local Neumann problems

Some references on computation of equilibrium fluxes

• Ainsworth-Oden

• Baker

• Ladeveze

• Larson-Niklasson

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.20

Local Problems
• No unknown constants
• More efficient
• More complicated
• E needs to be computed numerically
• We get a truth mesh error estimate

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.21

Recovery Estimates
Let GU ∈ Vhd be determined by the equation
(GU, v)Lump = (∇U, v),

for all v ∈ Vhd . Here (·, ·)Lump refers to the lumped

L2 (Ω)-product. Then it holds,
X
k∇ek ≤ C ρ2K ,
K∈K

where ρK = kGU − ∇U kK .

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.22

Lower Bounds
Assume we compute an approximation of the error
E ∈ Vh ⊂ H01 (Ω) such that,
(∇e, ∇v) = (∇E, ∇v),
for all v ∈ Vh . Then we have,
k∇Ek2 = (∇E, ∇E)
= (∇e, ∇E)
≤ k∇ekk∇Ek,

So we get k∇Ek ≤ k∇ek.

Ex: Solve a local Dirichlet problem on each element.
Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.23
Lower Bounds, cnt
For estimates based on local Neumann problems on
stars we obtain an approximation E of the error e. E
is discontinuous on element edges and satisfies,
(∇E, ∇v) = (∇e, ∇v),

for all v ∈ H01 (Ω) by construction.

Assume we can compute a continuous approx. E c of
E. Then we have
2λ(∇E, ∇E c) − λ2 k∇E ck2 ≤ k∇ek2 ,

for all λ ∈ R. See (Diez, Pares and Huerta).

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.24

Lower Bounds, cnt

0 ≤ (∇(e − λE c), ∇(e − λE c))

= k∇ek2 − 2λ(∇e, ∇E c ) + λ2 k∇E ck2
= k∇ek2 − 2λ(∇E, ∇E c) + λ2 k∇E ck2 .

This leads to
k∇ek2 ≥ 2λ(∇E, ∇E c) − λk∇E ck2 .
(∇E,∇E c )
An optimal value λ = k∇E c k2 gives,
c 2
2 |(∇E, ∇E )|
k∇ek ≥
k∇ek2
Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.25
Refinement Techniques
Bisect triangles. Joining the midpoint of the longest
edge with the opposite vertex.

Divide into four parts.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.26

Refinement Techniques, cnt
Red refinement.

Blue refinement.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.27

Refinement Techniques, cnt
Green refinement.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.28

Refinement Techniques, cnt
Bricks.

Hanging nodes may be dealt with,

• By enforcing strong continuity

• By enforcing weak continuity

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.29

Adaptive Algorithms
• Start on initial mesh
• Solve equations to get the solution U
• Calculate the error indicators EK or RK
• Refine elements K with large values of the
indicator
• Go back to step 2 or stop if the error is small
enough

In practice computer memory or time can also be used

as a natural stopping criteria.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.30

Adaptive Algorithms, cnt
There are different ways to choose which elements to
refine. Assume
P that we have an estimate
k∇ek2 ≤ K∈K R2K . Let 0 ≤ θ ≤ 1,
• Refine all elements where RK ≥ θ max RK .
• Let D ⊂ K such that
X X
R2K ≥ θ2 R2K .
k∈D K∈K

Refine elements in D.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.31

Other Sources of Error
A posteriori estimates can be extended to include
• Errors in data

• Errors in solution of algebraic system of

equations
• Errors in solution due to approximation of
geometry
• Quadrature errors

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.32

Error in Data
We study error in data f ,
(∇e, ∇e) = (f + 4U, e)
= (f˜ + 4U, e) + (f − f˜, e)
≤ |(f˜ + 4U, e − πe)| + Ckf − f˜k−1 kek1 ,
so we get
X
k∇ek2 ≤ C 2
RK + Ckf − f˜k2−1 .
K∈K

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.33

Algebraic Error
If the algebraic system of equation is solved
approximately we get,

k∇ek2 = (f + 4U, e − πe) + (f + 4U, πe),

where the second term is the algebraic residual.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.34

Quadrature Error
Finally we consider quadrature error,

k∇ek2 = (f + 4U, e − πe)

+ (∇e, ∇πe) − (∇e, ∇πe)h
= (f + 4U, e − πe)
+ (f, πe) − (f, πe)h
+ (∇U, ∇πe) − (∇U, ∇πe)h ,

where (·, ·)h in the form obtained by quadrature.

Computational Mathematics, and Fraunhofer Centre for Industrial Mathematics – p.35