Fem Doerfler

Finite Element Methods
Lecture notes by
Prof. Dr. W. Dörfler
Institute for Applied and

Numerical Mathematics
at the Department of Mathematics
Version: 28.02.2022
KIT – University of the State of Baden-Wuerttemberg and National Research Center of the Helmholtz Association www.kit.edu
2 Contents
Contents
1. Variational equations 4
1.1. Modelling with Partial Dierential Equations . . . . . . . . . . . . . . . . . 4
1.2. The Dirichlet boundary value problem . . . . . . . . . . . . . . . . . . . . . 4
1.3. The mixed DirichletNeumann boundary value problem . . . . . . . . . . . 7
1.4. Dirichlet principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2. RitzGalerkin method 9
2.1. The discrete problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2. The Céa lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3. The Strang lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3. Linear nite elements 12

3.1. The one-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.1. The nite element space . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.2. The discrete problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.1.3. Error estimates (1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.1.4. Numerical integration (variational crimes) . . . . . . . . . . . . . . 17
3.1.5. Error estimates (2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.1.6. The Neumann problem . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2. The two-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.1. Triangulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.2.2. The discrete problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.3. Numerical integration . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.4. Piecewise smooth boundary . . . . . . . . . . . . . . . . . . . . . . . 26
3.2.5. Data structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4. Quadratic elements 28
4.1. The one-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2. The two-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3. Error estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.4. Isoparametric elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5. Tensorproduct elements 37
5.1. Discretisation of rectangular domains into rectangles . . . . . . . . . . . . . 37
5.2. Tensor product nite elements . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2.1. The bilinear element . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.2.2. The biquadratic element . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3. The stiness matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.1. The bilinear element . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.2. The biquadratic element . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4. Logical rectangular grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6. Interpolation estimates 40
6.1. The BrambleHilbert lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
6.2. Nodal interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6.3. Nodal interpolation on convex domains . . . . . . . . . . . . . . . . . . . . . 43
6.3.1. Linear elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
6.3.2. Quadratic elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.4. The interpolation operator of Clément . . . . . . . . . . . . . . . . . . . . . 45
6.5. The interpolation operator of Scott/Zhang . . . . . . . . . . . . . . . . . . . 47
2
Contents 3
7. Special topics 48
7.1. Error estimates in the L2 -norm . . . . . . . . . . . . . . . . . . . . . . . . . 48
7.2. The discrete maximum principle . . . . . . . . . . . . . . . . . . . . . . . . 49
7.3. Error estimate for a nonlinear problem . . . . . . . . . . . . . . . . . . . . . 51
7.3.1. Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
7.3.2. Numerical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.3.3. Discrete equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.4. Non-conforming elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
7.5. The condition number of the stiness matrix . . . . . . . . . . . . . . . . . 58
7.6. Eigenvalue problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.7. Adaptive Finite Element Method . . . . . . . . . . . . . . . . . . . . . . . . 62
7.7.1. The residual error estimate . . . . . . . . . . . . . . . . . . . . . . . 62
7.7.2. The adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.7.3. hp nite element method . . . . . . . . . . . . . . . . . . . . . . . . 67
7.7.4. Adaptivity with quantities of interest . . . . . . . . . . . . . . . . . . 68
8. Example of a nite element implementation 68

8.1. Bases on the reference element . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.1.1. One-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.1.2. Two-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.2. Basic matrices on the reference cell . . . . . . . . . . . . . . . . . . . . . . . 69
8.2.1. One-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.2.2. Two-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . 70
8.3. The global nite element discretisation . . . . . . . . . . . . . . . . . . . . . 70
8.3.1. Local space and matrices . . . . . . . . . . . . . . . . . . . . . . . . 70
8.3.2. Global space and matrices . . . . . . . . . . . . . . . . . . . . . . . . 71
9. Saddle Point Problems 72

9.1. The Poisson problem as a rst order system . . . . . . . . . . . . . . . . . . 72
9.2. Minimisation with constraint . . . . . . . . . . . . . . . . . . . . . . . . . . 73
9.3. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
9.3.1. Poisson problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
9.3.2. Stokes problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
9.3.3. Timeharmonic electric elds . . . . . . . . . . . . . . . . . . . . . . . 74
9.4. Operator notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
9.5. Existence of solutions for the saddle point problem . . . . . . . . . . . . . . 75
9.6. Discretisation of saddle point problems . . . . . . . . . . . . . . . . . . . . . 77
9.7. General error estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
9.8. The Fortin operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
9.9. Discretisation of the mixed Poisson problem . . . . . . . . . . . . . . . . . . 80
9.9.1. The primal problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
9.9.2. The dual problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
9.9.3. RaviartThomas elements . . . . . . . . . . . . . . . . . . . . . . . . 81
9.9.4. A Fortin operator for the RaviartThomas element (k = 0) . . . . . 81
A. Stable renement methods for triangulations in R2 84

A.1. RedGreen (Blue) Renement . . . . . . . . . . . . . . . . . . . . . . . . . 84
A.2. Newest node bisection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
A.3. Longest edge bisection. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
B. The discontinuous Galerkin method for the Poisson problem 85
3
4 1 Variational equations
1. Variational equations
1.1. Modelling with Partial Dierential Equations
We refer to the basic physical motivation and examples from the scripts [Dör15, Sect. 2.1,
3] [Dör21b, Sect. 2.2] [Dör21a, Sect. 2.1.1, 2.4.1, 4].
1.2. The Dirichlet boundary value problem

Let Ω ⊂ Rd , a, c, f : Ω → R, uD : ∂Ω → R. We consider the boundary value problem in
divergence form

−∇ · a∇u + cu = f in Ω, (1.1)
D
u=u on ∂Ω. (1.2)
The boundary condition (1.2) is calledDirichlet boundary condition. A special case

is the Poisson problem (a = 1, c = 0)
−∆u = f in Ω, (1.3)
D
u=u on ∂Ω. (1.4)
If in addition f =0 holds, then the solution u to this problem is called harmonic.

Now we will introduce the weak formulation. We multiply (1.1) by v ∈ C ∞ (Ω) with
compact support in Ω (i.e., v∈ Cc∞ (Ω)) and integrate by parts to get
Z Z
{a∇u · ∇v + cuv} = fv for all v ∈ Cc∞ (Ω). (1.5)
Ω Ω
The v 's are called test functions.

Our aim is to show how (1.5) can be used to solve (1.1), (1.2). Thus our problem reads:
Find u with ub∂Ω = uD such that (1.5) holds. (1.6)
It turns out that it is advantageous to reformulate the problem to get homogeneous bound-
ary conditions, i.e. uD = 0. For this we assume that there exists a suciently smooth func-
eD
tion u :Ω→R such that e D b∂Ω = uD .
u Now we formulate the problem for u redened
as u+ e D , this means,
u
Z Z
{a∇u · ∇v + cuv} = u D · ∇v − ce
{f v − a∇e u D v} for all v ∈ Cc∞ (Ω) (1.7)
Ω Ω
and our problem now reads:
Find u:Ω→R with ub∂Ω = 0 such that (1.7) holds. (1.8)
This rst requires to formulate a condition under which this problem is well dened. For
this we seek a linear complete function space where the expressions in (1.7) are well-dened.
Denition 1.1 (Sobolev spaces). For Ω ⊂ Rd , open and bounded with Lipschitz continuous
boundary, we dene
n
W 1,2 (Ω) ≡ H1 (Ω) := v ∈ L2 (Ω) : v is weakly dierentiable with ∇v ∈ L2 (Ω)d ,
o
||v||W 1,2 (Ω) < ∞
([Dör15, Sect. 1.2.2, 1.2.3]), where
1/2
||v||W 1,2 (Ω) = ||v||H1 (Ω) := ||v||2L2 (Ω) + ||∇v||2L2 (Ω)d .
4
1.2 The Dirichlet boundary value problem 5
The following subspace is well-dened [Dör15, Sect. 1.3.3]
W01,2 (Ω) ≡ H01 (Ω) := v ∈ W 1,2 (Ω) : v b∂Ω = 0 .

Both spaces are Hilbert spaces. H1 (Ω) contains C ∞ (Ω) and H01 (Ω) contains Cc∞ (Ω) as a
dense subset.
These complete normed linear function spaces are instances of Sobolev spaces. For
the Lebesgue function spaces Lp and the general Sobolev spaces W m,p , Hm , m ∈ N,
p ∈ [1, ∞], or their generalisation to m ∈ R, see the accompanying script [Dör15, Sect. 1]
and the literature cited therein.
The following abstract result from functional analysis will easily lead to existence und
uniqueness.
Theorem 1.2 (LaxMilgram theorem) . Let V be a Hilbert space and a bilinear form
A:V ×V →R and a linear form F : V → R be given, satisfying the following conditions
for some real constants α0 , αF with 0 < α0 ≤ α1 , 0 ≤ αF :
• A is coercive on V, i.e.,
A[v, v] ≥ α0 ||v||2V for all v ∈ V,
and A is continuous, i.e.,
A[v, w] ≤ α1 ||v||V ||w||V for all v, w ∈ V,
• F is continuous, i.e.,
F [v] ≤ αF ||v||V for all v ∈ V.
Then there exists a unique solution u of the variational equation

A[u, v] = F [v] for all v∈V (1.9)
and it satises the a priori bound

αF
||u||V ≤ .
α0
Proof. This result is for example proved in [Alt16]. To obtain the bound we insert u for v
to get
α0 ||u||2V ≤ A[u, u] = F [u] ≤ αF ||u||V .
Theorem 1.3 . d
Let Ω ⊂ R be open and bounded with Lipschitz continuous
(Existence)
2 ∞
boundary. Let f ∈ L (Ω) and a, c ∈ L (Ω) with a ≥ a0 > 0, c ≥ 0 a.e. in Ω . Assume
e D ∈ H1 (Ω) exists with u
that a function u e D b∂Ω = uD . Let the bilinear form A : H01 (Ω) ×
H01 (Ω) → R and the linear form F : H01 (Ω) → R be dened by
Z

A[w, v] := a∇w · ∇v + cwv ,
ZΩ
F [v] := u D , v].
f v − A[e
Ω
Then there exists a unique function u ∈ H01 (Ω) such that
A[u, v] = F [v] for all v ∈ H01 (Ω) (1.10)
5
Furthermore it holds the a priori bound
||u||H1 (Ω) ≤ C,
with a constant C e D , and Ω . For uD = 0, u is the unique solution

depending on a, c, f, u
1 D
of problem (1.8) in H0 (Ω). For u 6= 0, u + u e D is the unique solution of problem (1.6) in
1
H (Ω).
1/2
V := H01 (Ω) 2
R
Proof. We let with ||v||V := Ω |∇v| . From the following Lemma 1.4
we conclude that this norm is equivalent to ||v||H1 (Ω) on V and thus (V, || . ||V ) is a Hilbert
space. Therefore we have to consider the problem:
Find u∈V such that A[u, v] = F [v] for all v ∈V.

For this we verify the assumptions of Theorem 1.2. By the requirements we easily get the
bounds (using the Cauchy-Schwarz inequality)
A[v, v] ≥ a0 ||∇v||2L2 (Ω)d ,

A[w, v] ≤ max ||a||L∞ (Ω) , ||c||L∞ (Ω) ||w||H1 (Ω) ||v||H1 (Ω) ,
u D , v]
F [v] ≤ ||f ||L2 (Ω) ||v||H1 (Ω) + A[e
for all v, w ∈ V . By the mentioned norm equivalence we have fullled all the require-
ments to apply Theorem 1.2, that guarantees the existence of a unique solution u ∈ V.
Furthermore we have the bound
αF
||u||V ≤ ,
α0
whereα0 and αF can be obtained from the previous estimates. If u solves (1.8), then
e D solves problem
u+u (1.6) by construction.
Lemma 1.4 (Poincaré inequality). Let Ω ⊂ Rd be open, bounded with Lipschitz continuous
boundary. Then there is a constant CP (Ω) > 0 such that the Poincaré inequality
||v||L2 (Ω) ≤ CP (Ω)||∇v||L2 (Ω)d for all v ∈ H01 (Ω)
holds. CP (Ω) is called Poincaré constant.

Remark 1.5.
(1) Let uD = 0. For f ∈ L2 (Ω) let u ∈ H01 (Ω) be the unique solution of (1.8). For a
constant C , depending on a, c, Ω , it holds by Theorem 1.3 that ||u|| 1
H (Ω) ≤ C||f ||L2 (Ω) .
2 1
Thus, the mapping L (Ω) → H0 (Ω) : f 7→ u is linear and continuous.
(2)
D
If u 6= 0, Theorem 1.3 is based on nding any function u e D ∈ H1 (Ω) that extends
D
the boundary data u into Ω . Clearly, if such a function would not exist, our problem
1 D
cannot have a solution in H (Ω)! The precise condition on u will be u ∈ H
D 1/2 (∂Ω)
[Dör15, Sect. 1.3.2]. One shows that there is a linear continuous extension oper-
ator E : H1/2 (∂Ω) → H1 (Ω) (with Ev b∂Ω = v ) and let ue D := EuD .
(3) The function u(x) = Im(log(x1 + ix2 )) is harmonic in the upper half-circle in R2 ,
but its trace on the boundary is discontinuous in x = 0 [Dör15, Sect. 1.3.2]. The
boundary data is not in H
1/2 (∂Ω) [Dör15, Sect. 1.3.1] and u is not in H1 (Ω).
1
(4) For d ≥ 2, functions in H (Ω) need not to be bounded (which means, that we have
1
no continuous embedding of H (Ω) into the space of continuous functions C (Ω))
0
[Dör15, Sect. 1.3.1 (Example)] !
6
1.3 The mixed DirichletNeumann boundary value problem 7
(5) Requiring more regularity on data a, c, f, Ω will lead to more regular solutions, e.g.,
u ∈ H2 (Ω), in which case equation (1.1) holds pointwise a.e. [Dör15, Sect. 2.3.3].
(6) Solutions of (1.8) can in fact be very weak. We consider the equation
−∇·(a∇u) = f on Ω = (−1, 1)2 ,
where a is piecewise constant, a ≈ 100 for x1 x2 < 0 and a≡1 for x1 x2 > 0. Then
u(x) ∼ |x|1/10 for |x| ≈ 0,
and we nd u ∈ H1+1/10 (Ω) [Kel75].
(7) The general linear elliptic equation has the form

−∇ · a∇u + b · ∇u + cu = f in Ω (1.11)
with b ∈ L∞ (Ω)d and c ∈ L∞ (Ω). The weak equation is derived the same way, the
existence theory requires some modications.
1.3. The mixed DirichletNeumann boundary value problem

Let Ω ⊂ Rd be as before and a ∈ L∞ (Ω) : a ≥ a0 > 0 a.e., f ∈ L2 (Ω). Let ˙ N
∂Ω = Γ D ∪Γ
be a disjoint decomposition of ∂Ω , ν the exterior normal eld on ∂Ω ,
N 2 N
and u ∈ L (Γ ).
We consider now the boundary value problem
−∇·(a∇u) = f in Ω, (1.12)
D
u=0 on Γ , (1.13)
a∂ν u = aν · ∇u = uN on Γ N. (1.14)
The imposed boundary condition (1.14) on ΓN is called Neumann boundary condition

and Γ Neumann boundary. Likewise, Γ D is called Dirichlet boundary.
N is called
Note, that the more general boundary condition u = u

D on Γ D can be reduced to (1.13)
as in Section 1.2.
1 (Ω) := {w ∈ H1 (Ω) : w
v ∈ HD
We now test equation (1.12) with bΓ D = 0} to get by
partial integration
Z Z Z
− a∂ν u v + a∇u · ∇v = fv
∂Ω Ω Ω
which is turned into

Z Z Z
a∇u · ∇v = fv + uN v. (1.15)
Ω Ω ΓN
This is again a problem of the form treated in Theorem 1.3:
1 1
Find u ∈ HD (Ω) such that (1.15) holds for all v ∈ HD (Ω). (1.16)
If |Γ D | 6= 0 ∂Ω ), then the Poincaré inequality (Lemma

(with respect to the measure on
Γ D , Γ N [Alt16, Ch. 5.15]. The same technique
1.4) still holds with a constant depending on
1 1
as before yields a unique solution u ∈ HD (Ω) of the equations (1.12)(1.14) .
D
In the case |Γ | = 0 we have Γ
N 1 1
= ∂Ω and thus HD (Ω) = H (Ω). Then we do not have
a unique solution of the equations (1.12)(1.14), since constant functions can be added to
1
For the continuity of F , we need the trace inequality ||v||L2 (∂Ω) ≤ Ctr ||v||H1 (Ω) for all v ∈ H1 (Ω) [Dör15,
Sect. 1.3.2].
7
any solution giving a new solution. Thus we have to introduce some restrictions for the
solution space, for example, we may take the subspace of H1 (Ω) dened by the constraint
1
Z Z
mean(v, Ω) := − v := v=0
Ω |Ω| Ω
and require for the data f, uN the compatibility condition

Z Z
F [1] = 0 ⇐⇒ f+ uN = 0.
Ω ΓN
The replacement of the Poincaré inequality is given in the next lemma. Then we can
proceed as before to prove existence und uniqueness (under the given constraint) of the
Neumann problem [Dör15, Sect. 2.3.5].
Lemma 1.6 (Poincaré inequality with meanvalue zero) . Ω ⊂ Rd be open,

Let bounded
with Lipschitz continuous boundary. Then there is a constant CN (Ω) > 0 such that the
Poincaré inequality with meanvalue zero or PoincaréWirtinger inequality
v − mean(v, Ω) L2 (Ω)
≤ CN (Ω)||∇v||L2 (Ω)d for all v ∈ H1 (Ω)
holds.
1.4. Dirichlet principle

One can derive the variational formulation (1.9) from the Dirichlet principle. Let
Ve = u
e+V be a nonempty(!) ane linear space and dene
E : Ve → R,
1
E(v) := A[v, v] − F [v].
2
We consider the minimisation of this functional on Ve under the assumptions from Theo-
rem 1.2. Since E ∈ C 1 (Ve , R) and E is strictly convex, the minimisation is equivalent to
E 0 (u) = 0, where the linear mapping E 0 (u) is dened as
1
E 0 (u)[v] = lim

E(u + v) − E(u) for all v ∈ V.
→0
Note that with u ∈ Ve and v∈V we have u + v ∈ Ve . This results in
E 0 (u)[v] = A[u, v] − F [v] = 0 for all v ∈ V.
E(v) = Ω {1/2(a|∇v|2 + c|v|2 ) − f v} on the ane

R
In view of our problem (1.5) we take
space Ve = {v ∈ H1 (Ω) : v b∂Ω = uD } = u

e D + H01 (Ω), thus V = H01 (Ω). Hence we obtain
(1.7)
1
for all v ∈ H0 (Ω) as a necessary condition for a minimum. This provides a dierent
view on the problems in Sections 1.2 and 1.3.
Remark 1.7. We started to describe our dierential equation in the strong form (1.1) and
derived a weak form (1.5) by multiplication with suitable test functions. Alternatively, as we
have seen above, the weak form can be derived as a necessary condition for a minimum of
a strictly convex energy E under certain conditions. But, under which conditions are weak
solutions strong solutions? One can show that when u is a weak solution for which contin-
uous second derivatives exist, then u is a strong solution. For this one applies integration
by parts and uses test functions v with small compact support.
8
9
Literature for Section 1. [BS02, Ch. 02, 5] [LT05, Sect. 2.3, 3.5] [Bra07, Ch. II, 1-3]
[Dör15, Sect. 12] [Dör13, Sect. 1].
2. RitzGalerkin method
We will now use the derivations we have considered in the preceding chapter to obtain an
approximation of the solution of our problem. In contrast to the Finite Dierence Method
[Dör21b, Sect. 2] we do not approximate the dierential operators which appear in the
problem formulation, we actually approximate the solution space.
Remark 2.1. The following cases cause problems when using the Finite Dierence Method:
• Complex domains,
• curved boundaries, especially for higher order methods,
• discontinuous a,
• non-regular solutions,
• a symmetric problem is not sucient to yield a symmetric system of equations.
2.1. The discrete problem

We approximate the solution of the variational problem (1.9) on a nite dimensional sub-
space VN ⊂ V , where dim(VN ) = N ∈ N. If we assume VN to be a subspace of V ,
VN ⊂ V , we call this a (V -)conforming ansatz space and thus the resulting method a
(V -) conforming method.
In the following, we consider the case uD = 0 without loss of generality. The inhomoge-
neous case u
D 6= 0 is treated as in Section 1.2.
The discrete problem is now formulated as:
Find uN ∈ VN such that A[uN , vN ] = F [vN ] for all vN ∈ VN . (2.1)
The functions vN are called discretetest functions and VN is the (discrete) test space.
This approach is denoted as RitzGalerkin method.
By the LaxMilgram theorem (Theorem 1.2) we again get that this problem has a unique
solution and the following discrete a priori bound holds:
αF
||uN ||V ≤ .
α0
Note that this bound is independent of N!
Now let B := {ψ} be a basis of VN . Then uN can be uniquely written as
X
uN = αψ ψ
ψ∈B
with coecients αψ ∈ R, ψ ∈ B . Using this as an ansatz and testing with vN = φ, φ ∈ B ,

we get
X
αψ A[ψ, φ] = F [φ] for all φ ∈ B.
ψ∈B
~ ∈ RN ×N ~ = A[ψ, φ]

We then dene the matrix A by A φ,ψ∈B
, the vectors ~u := [αψ ]ψ∈B ,
F~ := F [φ] RN

in and we get
φ∈B
~ u = F~ .
A~
This is a linear system of equations with a symmetric positive denite matrix ~.

A
9
10 2 RitzGalerkin method
Denition 2.2 (Energy norm). Let A be as in Theorem 1.2. Then we dene the energy
norm by
||v||E := A[v, v]1/2 for all v ∈ V.
This is indeed a norm on V that is equivalent to || . ||V since A is assumed to be coercive

and continuous. On VN we have
1/2
~v
||vN ||E = A[vN , vN ]1/2 = ~v · A~ ,
when ~v ∈ RN is the vector that corresponds to the function vN ∈ V N .
2.2. The Céa lemma

Theorem 2.3 (Céa lemma) . Let A and F be as in Theorem 1.2. If u and uN are the
respective unique solutions of (1.9) and (2.1), then
||u − uN ||E = inf ||u − vN ||E .

vN ∈VN
That means that uN is the best approximation in VN to u in the energy norm.
Proof. Let vN ∈ VN be arbitrary. Then we get the Galerkin orthogonality

A[u − uN , vN ] = F [vN ] − F [vN ] = 0.
This yields
||u − uN ||2E = A[u − uN , u − uN ] = A[u − uN , u − vN ]

≤ ||u − uN ||E ||u − vN ||E .
Thus, for all vN ∈ VN ,
||u − uN ||E ≤ ||u − vN ||E .
Taking the inmum over vN ∈ V N , this shows the result but with ≤. The equality holds,
since uN ∈ VN .
Corollary 2.4 (Convergence). Assume that we have a linear continuous operator
PN : V → VN
with the property
N →∞
||v − PN v||V −→ 0 for all v ∈ V.
Then we have convergence of the sequence of discrete solutions {uN }N ∈N of (2.1) towards
the unique exact solution u of (1.9),
N →∞
||u − uN ||V −→ 0.
Proof. We have by coercivity and continuity of A that
√ √
α0 ||u − uN ||V ≤ ||u − uN ||E ≤ ||u − PN u||E ≤ α1 ||u − PN u||V .
Therefore
r
α1 N →∞
||u − uN ||V ≤ ||u − PN u||V −→ 0,
α0
which proves the assertion.
10
2.3 The Strang lemma 11
Remark 2.5 (Quasi-optimal error estimate). In general, estimates are of the form
||u − uN ||V ≤ CS inf ||u − vN ||V .

vN ∈VN
CS is the stability constant of the continuous problem and the inmum gives a bound for
the approximability of u in VN . Compared to a corresponding result in Finite Dierence
Methods ([Dör21b, Thm. 2.22]) we can state here:
Stability and Approximability implies Convergence.
Later, we will construct suitable operators PN explicitly. For practical use we want to
obtain an estimate like, e.g.,
||u − PN u||V ≤ CN −κ ,
with C depending on u and with some κ > 0. For this purpose a certain regularity of u is
necessary, e.g.,
u ∈ V reg ⊂ V.
For example, we will have V = H01 (Ω) and V reg = H2 (Ω) ∩ H01 (Ω). C will then depend on
||u||V reg .
The problem in practice is to compute ~
A and F~ , e.g., in case (1.5),
Z
A[ψ, φ] = {a∇ψ · ∇φ + cψφ},
ZΩ
F [φ] = f φ, for all ψ, φ ∈ B.
Ω
For this we need (in general) approximations for the integrals (for quadrature rules see
Section 3.2.3). This results in perturbed problems
Find uN ∈ VN such that AN [uN , vN ] = FN [vN ] for all vN ∈ V N , (2.2)
where AN and FN are approximations to A, F . In this case the error will be controlled by
the Strang lemma, which will be introduced in the following section.
2.3. The Strang lemma

Theorem 2.6 (Strang lemma) . Let A, F be as in Theorem 1.2 (LaxMilgram). Let AN
be a symmetric positive denite bilinear form, FN a linear form on VN ⊂ V . Assume that
AN further satises
β0 A[vN , vN ] ≤ AN [vN , vN ] for all vN ∈ VN

for some β0 > 0. If u and uN are the respective unique solutions of (1.9) and (2.2), then
(
||u − uN ||E ≤ inf (1 + β0−1 )||u − vN ||E
vN ∈VN
)
A[vN , wN ] − AN [vN , wN ]
+ β0−1 sup
wN ∈VN \{0} ||wN ||E
F [wN ] − FN [wN ]
+ β0−1 sup
wN ∈VN \{0} ||wN ||E
≤ (1 + β0−1 )||u − PN u||E
A[PN u, wN ] − AN [PN u, wN ]
+ β0−1 sup
wN ∈VN \{0} ||wN ||E
F [wN ] − FN [wN ]
+ β0−1 sup .
wN ∈VN \{0} ||wN ||E
11
12 3 Linear nite elements
Proof. Let vN ∈ V N be arbitrary. Then we get
β0 ||uN − vN ||2E = β0 A[uN − vN , uN − vN ]

≤ AN [uN − vN , uN − vN ]
= A[u − vN , uN − vN ] + AN [uN − vN , uN − vN ] − A[u − vN , uN − vN ]
= A[u − vN , uN − vN ] + A[vN , uN − vN ] − AN [vN , uN − vN ]
+ FN [uN − vN ] − F [uN − vN ]
≤ ||u − vN ||E ||uN − vN ||E
A[vN , wN ] − AN [vN , wN ] F [wN ] − FN [wN ]
+ sup + sup ||uN − vN ||E .
wN ∈VN \{0} ||wN ||E wN ∈VN \{0} ||wN ||E
Hence we get a bound for ||uN − vN ||E . We now proceed with
||u − uN ||E ≤ ||u − vN ||E + ||vN − uN ||E

≤ (1 + β0−1 )||u − vN ||E + β0−1 (. . .),
and take the inmum over vN ∈ V N . The second inequality is obtained by choosing
vN = PN u.
Remark 2.7. From this lemma we draw the following conclusions:
• use an approximation AN to A that guarantees coercivity,
• use approximations AN to A and FN to F such that the resulting errors are less than
inf ||u − vN ||E or ||u − PN u||V .

vN ∈VN
Literature for Section 2. [BS02, Ch. 2.8, 10.1] [LT05, Kap. 5] [Bra07, Ch. II, 4, Ch. III,
1] [Dör13, Sect. 2]
3. Linear nite elements

3.1. The one-dimensional case
3.1.1. The nite element space
We recall from Section 1.2 the boundary value problem
−(au0 )0 + cu = f in (0, 1),

u = uD on {0, 1}
for u : (0, 1) → R. (0, 1) stands w.l.o.g. for any bounded interval since we can otherwise
transform the problem to (0, 1). a, c, f : [0, 1] → R, uD (0) = uL , uD (1) = uR given.
Assume again a ≥ a0 > 0 and c ≥ 0.
Denition 3.1 (Basic denitions). We dene the set of nodes Nh by

Nh := x0 , . . . , xn+1 ∈ R : 0 = x0 < . . . < xn+1 = 1
and the set of closed intervals Kh by

Kh := Ki := [xi−1 , xi ] : i = 1, . . . , n + 1 .
12
3.1 The one-dimensional case 13
[0, 1] = n+1 K∈Kh K . The symbol h stands for the meshsize

S S
Then we have i=1 Ki =
function x 7→ h(x) given by

◦
hi := hKi := |xi − xi−1 |, x ∈ Ki

h(x) := 12 (hi+1 + hi ), x = xi for i = 1, . . . , n .

h1 or hn+1 , x = x0 or x = xn+1

Denition 3.2 (Linear nite elements) . The space of linear finite elements on Kh
is
S 1,0 (Kh ) := v ∈ C 0 ([0, 1]) : v bK ∈ P1

for all K ∈ Kh .
Clearly S 1,0 (Kh ) ⊂ H1 ((0, 1)) 2 . A subspace of S 1,0 (Kh ) is S01,0 (Kh ) dened by
S01,0 (Kh ) := v ∈ S 1,0 (Kh ) : v(0) = v(1) = 0 .

We have
dim S 1,0 (Kh ) = n + 2,

dim S01,0 (Kh ) = n.

Denition 3.3 (Nodal basis). A basis of S 1,0 (Kh ) is given by the set of hat functions
{φi : i = 0, . . . , n + 1}, uniquely dened by
φi (xj ) = δij ,
and called nodal basis. We have explicit formulas for the hat functions:
1
 hi (x − xi−1 ),
 x ∈ Ki
1
φi (x) = hi+1 (xi+1 − x), x ∈ Ki+1 .

0,

else
Hence, for each vh ∈ S 1,0 (Kh ) we have the unique representation
n+1
X
vh (x) = vi φi (x),
i=0
with vi := vh (xi ), whereas for vh ∈ S01,0 (Kh ) the rst and the last summand vanish. Note
that it holds
S01,0 (Kh ) ⊂ H01 ((0, 1)).
3.1.2. The discrete problem
The ansatz for a discrete solution is
n
X
uh (x) = ehD (x)
u + uj φj (x),
j=1
where the extended boundary data could be chosen as
ehD (x) = uL φ0 (x) + uR φn+1 (x) ∈ S 1,0 (Kh ).

u
2
More general for Ω ⊂ Rd : If a function is piecewise smooth and globally continuous in Ω , then it is in
H1 (Ω)
13
The formulation (1.7) is then treated as in Section 2.1, and we obtain from A[uh , φi ] =
F [φi ], for i = 1, . . . , n, the linear system
~ u = ~b,
A~
with
~u = [u1 , . . . , un ] ∈ Rn ,
hZ 1 i
A~= {aφ0j φ0i + cφj φi } ∈ Rn,n ,
0 i,j=1,...,n
hZ 1 i
~b = f φi − Ai,0 uL − Ai,n+1 uR ∈ Rn .
0 i=1,...,n
There are only at most 3 non-zero values per row in ~.

A The matrix is therefore sparse. In
case of the natural ordering, ~
A is tridiagonal [Dör21b, Bsp. 2.14]. So we solve by Gaussian
elimination [DR20, Sect. 4.2.1]. A closer inspection of the matrix entries yields (assuming
hi ≈ h for all i for simplicity)
Z 1
Aii = {a|φ0i |2 + c|φi |2 } > 0,
0
Z 1
i6=j
Aij = {aφ0j φ0i + cφj φi } = −O(1/h) + O(h) < 0
0
for |i − j| = 1 and h small enough. Then ~

A is an M-matrix [Dör21b, Sect. 2.4.2], which
can be solved stably without pivoting. This yields, if e.g. a ≡ 1, c ≡ 0,
~ = tridiag − 1 , 1 + 1 , − 1
h i
A .
hi hi hi+1 hi+1
For a uniform grid we obtain the system matrix we already got with the Finite Dierence
Method [Dör21b, Sect. 2.4.1]! Note the scale 1/h on the left, compared to 1/h2 for the
nite dierence method (compensated by a corresponding dierent scale on the right hand
side vector).
3.1.3. Error estimates (1)
We consider
R1 a≡1 and c ≡ 0, i.e., the Poisson problem. We also assume for the while that
0 f φi is integrated exactly. Then the discrete problem reads: Find uh such that
Z 1 Z 1
u0h vh0 = f vh for all vh ∈ Vh := S01,0 (Kh ).
0 0
Existence of a solution
R 1 u to this problem can be obtained by Theorem 1.3 with V :=
H01 ((0, 1)), ||v||V := ( 0 |v 0 |2 )1/2 , and the splitting u = w + u
ehD (note u
ehD ∈ V ). Since we
used the splitting uh = wh + u ehD in Vh we get by Céa's lemma (Theorem 2.3)
||(w − wh )0 ||L2 ((0,1)) ≤ inf ||(w − vh )0 ||L2 ((0,1)) .

vh ∈Vh
(i.e., the constant is 1 here). Since w − wh = u − uh this gives
||(u − uh )0 ||L2 ((0,1)) ≤ inf ||(u − vh )0 ||L2 ((0,1)) .

uhD +Vh
vh ∈e
It remains to estimate the right hand side. We will explicitly construct an interpolation
operator V → Vh .
14
Lemma 3.4 (Nodal interpolation). The mapping
Ih : H1 ((0, 1)) → S 1,0 (Kh ),

n+1
X
v 7→ v(xi )φi ,
i=0
is linear and continuous with respect to the norm in V and
||(Ih v)0 ||L2 ((0,1)) ≤ ||v 0 ||L2 ((0,1)) for all v ∈ H1 ((0, 1)),
00
||(v − Ih v)[l] ||L2 ((0,1)) ≤ Ch2−l
max ||v ||L2 ((0,1)) for all v ∈ H2 ((0, 1)), l ∈ {0, 1}.
Due to Ih v(xi ) = v(xi ), Ih is called nodal interpolation. Note that

3
||v||L∞ ((0,1)) . ||v||H1 ((0,1)) ,
i.e., v 7→ v(xi ) is a linear continuous mapping in H1 ((0, 1)).

Proof. Take v ∈ H1 ((0, 1)) and x ∈ Ki = [xi−1 , xi ]. Then
1
(Ih v)0 (x) = v(xi−1 )φ0i−1 (x) + v(xi )φ0i (x) =

v(xi ) − v(xi−1 )
hi
Z xi √ Z
1 CS hi 1/2
= v0 ≤ |v 0 |2 dx .
hi xi−1 hi Ki
The rst assertion results after insertion of this into

X
||(Ih v)0 ||2L2 ((0,1)) = ||(Ih v)0 ||2L2 (K) .
K∈Kh
To show continuity of Ih ,
we note that v − Ih v vanishes at the boundary and thus allows a
Poincaré estimate. From ||v − Ih v||L2 ((0,1)) ≤ CP ||(v − Ih v)0 ||L2 ((0,1)) we can derive, with the
lower triangle inequality and the previously obtained estimate, a bound of ||Ih v|| 2
L ((0,1)) in
terms of ||v|| 1 and hence have ||Ih v|| 1 . ||v|| .
H ((0,1)) H ((0,1)) H1 ((0,1))
2
Now let v ∈ C ([0, 1]). Then v − Ih v fullls on all K ∈ Kh
v − Ih v b∂K = 0.
Rolle's theorem yields the existence of ξ∈K such that
(v − Ih v)0 (ξ) = 0.
Therefore, for all x∈K

Z x CS
0
(v − Ih v)00 (t) dt ≤ hK ||v 00 ||L2 (K) ,
p
(v − Ih v) (x) =
ξ
so that
X
||(v − Ih v)0 ||2L2 ((0,1)) ≤ hK ||v 00 ||2L2 (K) hK ≤ h2max ||v 00 ||2L2 ((0,1)) .
K∈Kh
Furthermore,
Z x Z x Z t
0
(v − Ih v)(x) = (v − Ih v) (t) dt = (v − Ih v)00 (s) ds dt
xi−1 xi−1 ξ
3
a . b means a ≤ Cb for some constant C that is not relevant here
15
leads to
CS
Z xi p Z xi 1/2
|(v − Ih v)(x)| ≤ hi |(v − Ih v)00 (s)|2 ds dt
xi−1 xi−1
3/2
= hi ||v 00 ||L2 (Ki )
and after integration and summation
X X
||v − Ih v||2L2 ((0,1)) = ||v − Ih v||2L2 (K) ≤ h4K ||v 00 ||2L2 (K)
K∈Kh K∈Kh
≤ h4max ||v 00 ||2L2 ((0,1)) .
Taking roots and using that C 2 ([0, 1]) is dense in H2 ((0, 1)) gives the assertions.
Theorem 3.5. 2
Let a ≡ 1, c ≡ 0 and f ∈ L ((0, 1)). Then there is a unique solution
u∈ H2 ((0, 1)) ∩ 1
H0 ((0, 1)) of the boundary value problem (1D Poisson problem)
−u00 = f in (0, 1),

u=0 on {0, 1}.
There is also a unique solution uh of the discrete problem on S01,0 (Kh ) and the error obeys
the bound
||(u − uh )0 ||L2 ((0,1)) ≤ Chmax ||u00 ||L2 ((0,1)) = Chmax ||f ||L2 ((0,1)) .
Here, exact integration for the right hand side vector was assumed.
Proof. u ∈ H2 ((0, 1)) is a result of regularity theory (here, in 1D, the solution can be
given explicitly in terms of f ). The only thing left to show is the error bound. But all we
have to do is to combine Céa's lemma (Theorem 2.3) with the previous interpolation result
(Lemma 3.4).
Theorem 3.6. Let now a, c ∈ C

0,1 ([0, 1]) with a ≥ a ≥ a > 0 and c ≥ 0. Then there is
1 0
2 1
a unique solution u ∈ H ((0, 1)) ∩ H0 ((0, 1)) of the boundary value problem
−(au0 )0 + cu = f in (0, 1),

u=0 on {0, 1}.
There is also a unique solution uh of the discrete problem on S01,0 (Kh ) and the error obeys
the bound
||(u − uh )0 ||L2 ((0,1)) ≤ Ca,c hmax ||u00 ||L2 ((0,1)) .
Here, exact integration is assumed for all integrals dening the discrete system.
Proof. For the regularity see e.g. the PDE script [Dör15, Sect. 2.2, 2.3]. We recall that
Céa's lemma (Theorem 2.3) gives for the energy norm
√ √
||v||2E := || av 0 ||2L2 ((0,1)) + || cv||2L2 ((0,1)) .
the bound
||u − uh ||E = inf ||u − vh ||E .

vh ∈S01,0 (Kh )
16
In the inmum expression we choose vh = Ih u and obtain with Lemma 3.4
||u − Ih u||2E ≤ ||a||L∞ ((0,1)) ||(u − Ih u)0 ||2L2 ((0,1)) + ||c||L∞ ((0,1)) ||u − Ih u||2L2 ((0,1))
≤ ||a||L∞ ((0,1)) + h2max ||c||L∞ ((0,1)) h2max ||u00 ||2L2 ((0,1)) .

With the coercivity we nally get
1
||(u − uh )0 ||L2 ((0,1)) ≤ √ ||u − uh ||E ≤ Ca,c ||u00 ||L2 ((0,1)) hmax .
a0
Remark 3.7. These error estimates are optimal with respect to the order of hmax . Consider
−u00 = 1 in (0, 1),

u=0 on {0, 1},
whose solution is u(x) = 1/2 x(1 − x). One can verify that the discrete solution for linear
nite elements on a uniform mesh with meshsize h is simply uh = Ih u and this gives
||(u − uh )0 ||L2 ((0,1)) & h.
3.1.4. Numerical integration (variational crimes)
quadrature rule
R1
Let g ∈ C 0 ([0, 1]). A to compute
0 g(x) dx is a formula
mk
X
Q
b k [g] := ωl g(ξl ),
l=1
Pmk 4
with mk ∈ N, weights ωl ∈ R with l=1 ωl = 1, and nodes ξl ∈ [0, 1] [DR20, Sect. 2] .
Preferably one wants ωl ≥ 0 since this guarantees positivity, i.e., if g ≥ 0, then also
Q
b k [g] ≥ 0. The quadrature rule is said to be consistent of order k, if
Z 1
Q
b k [p] = p(x) dx for all p ∈ Pk .
0
For given consistency order k one has the bound mk ≥ (k + 1)/2, where mk = (k + 1)/2 is
achieved by GaussLegendre quadrature, with ωl > 0 for all l [DR20, Sect. 2.4]. We have
the error estimate
Z 1
b k [g] ≤ Ck ||g [k+1] || 1
g(x) dx − Q L ((0,1))
0
for suciently regular g. On a decomposition Kh we use the composite quadrature rule
Z 1 X Z n+1
X Z xi n+1
X Z 1
g(x) dx = g(x) dx = g(x) dx = hi g(xi−1 + ξhi ) dξ
0 K∈Kh K i=1 xi−1 i=1 0
n+1
X m
X
≈ hi ωl g xi−1 + ξl (xi − xi−1 )
i=1 l=1
X
=: QK
k [g] =: Qk [g].
K∈Kh
4
Often quadrature weights and nodes are given for the reference interval [−1, 1].
17
This gives the error estimate
Z 1 n+1
X
g(x) dx − Qk [g] ≤ Ck hk+1
i ||g [k+1] ||L1 (Ki )
0 i=1
≤ Ck hk+1
max ||g
[k+1]
||L1 ((0,1)) ,
provided g [k+1] ∈ L1 ((0, 1)). With GauÿLegendre quadrature we have
Z 1
g(x) dx − Qk [g] = O(h2m
max ).
0
In order to approximate Fh [ . ] and Ah [ . , . ], we take in the following
Z 1
aφ0j φ0i + cφj φi ≈ Ah [φj , φi ] := Q1 [aφ0j φ0i ] + Q1 [cφj φi ],
0
Z 1
f φi ≈ Fh [φi ] := Q1 [f φi ].
0
3.1.5. Error estimates (2)
Theorem 3.8. 2
Assumptions as in Theorem 3.6, but with a, c, f ∈ C ([0, 1]). Then there
1,0
exists a discrete solution uh ∈ Vh = S0 (Kh ) of Ah [uh , vh ] = Fh [vh ] (for all vh ∈ Vh ) using
the quadrature rule Q1 as stated above. Then there holds the error bound
||u − uh ||E ≤ C(a, c, f, u)hmax .
Proof. We will for simplicity only prove the case c ≡ 0. The proof is an application
of Strang's lemma (Theorem 2.6). For interpolation use Ih from Lemma 3.4. From the
assumptions it follows u ∈ H2 ((0, 1)) and so we get
||u − Ih u||E ≤ CA hmax ||u00 ||L2 ((0,1)) .
Let vh ∈ Vh be arbitrary. Then, with xK being the midpoint of K , and since vh0 is constant
on each K it follows
X Z
A[vh , vh ] = a(x)|vh0 (x)|2 dx
K∈Kh K
a(x)
X Z
= a(xK )|vh0 (x)|2 dx
K a(xK )
K∈Kh
n o X Z
≤ max ||a||L∞ (K) ||1/a||L∞ (K) a(xK )|vh0 (xK )|2 dx
K∈Kh
K∈Kh | K {z }
=QK 0 2
1 [a|vh | ]
n o
= max ||a||L∞ (K) ||1/a||L∞ (K) Q1 [a|vh0 |2 ]
K∈K
| h
| {z }
{z } = Ah [vh ,vh ]
= 1+O(hmax )
=: β0−1 Ah [vh , vh ].
18
Therefore, Ah is coercive and thus uh exists. Now let vh , wh ∈ S01,0 (Kh ), then
X Z
A[vh , wh ] − Ah [vh , wh ] = avh0 wh0 − a(xK )vh0 wh0
K∈Kh K
X Z
vh0 wh0 bK

= a − a(xK )
K∈Kh K
X
. |vh0 wh0 |bK ||a00 ||L∞ (K) h3K
K∈Kh
| {z }
= const
CS X
. ||a00 ||L∞ (K) ||vh0 ||L2 (K) ||wh0 ||L2 (K) h2K
K∈Kh
. max h2K ||a00 ||L∞ (K) ||vh0 ||L2 ((0,1)) ||wh0 ||L2 ((0,1))

K∈Kh
≤ Ca h2max ||vh0 ||L2 ((0,1)) ||wh0 ||L2 ((0,1)) .
Note that we have again used that vh0 , wh0 are constant on each K and the discrete Cauchy-
Schwarz inequality.
Now we apply similar arguments to the integral of f to get
Z 1 X Z
f wh − Fh [wh ] ≤ f wh − QK
1 [f wh ]
0 K∈Kh K
X
. h2K ||(f wh )00 ||L1 (K)
K∈Kh
X Z
00
. h2K |f ||wh | + 2|f 0 ||wh0 | + 0
K∈Kh K
X
. h2K ||f 00 ||L2 (K) ||wh ||L2 (K) + ||f 0 ||L2 (K) ||wh0 ||L2 (K)
K∈Kh
X 1/2
. h4K ||f 0 ||2H1 (K) ||wh ||H1 ((0,1))
K∈Kh
≤ Cf h2max ||wh ||H1 ((0,1))
for f ∈ H2 ((0, 1)). By the Strang lemma (Theorem 2.6), we get for vh = Ih u the bound
||u − uh ||E ≤ (1 + β0−1 )CA ||u00 ||L2 ((0,1)) hmax + β0−1 Ca ||Ih u||E + Cf h2max .

With ||Ih u||E . ||u||E we get the result.

We notice, that here the quadrature error is O(h2max ), and is thus asymptotically much
smaller than the discretisation error that is
R1 O(hmax ) here. If a ∈ C 1 ([0, 1]) only, we would
getO(hmax ) using 0 g − Q1 [g] . ||g 0 ||L1 ((0,1)) . One can even reduce the requirements
on a if exceptional points (e.g. where a is discontinuous) are within the set of nodes.
1
Analogously, if f ∈ H ((0, 1)) only, we would get the bound O(hmax ) for the corresponding
quadrature error. This would also prove our result, the data error being now O(hmax ) like
the discretisation error.
3.1.6. The Neumann problem
We consider the Neumann boundary value problem
−u00 = f in Ω = (0, 1),

0
−u (0) = g0 ,
u0 (1) = g1 ,
19
where g0 , g1 ∈ R. We recall the steps from the treatment of the problem in Section 1.3.
First we write this in variational form
Z 1 Z 1
00
0=− u v− fv
0 0
Z 1 Z 1
= u0 v 0 − [u0 v]10 − fv
0 0
Z 1 Z 1
0 0 0 0
= u v − u (1)v(1) + u (0)v(0) − f v.
0 0
Thus we have the following problem: Find u ∈ H1 (Ω) such that
Z 1 Z 1
0 0
uv = f v + g1 v(1) + g0 v(0) for all v ∈ H1 (Ω). (3.1)
0 0
We note that by the trace theorem v 7→ g0 v(0) and v 7→ g1 v(1) [Dör15, Sect. 1.3.2] are
1
continuous functionals on H (Ω). Inserting v = 1, we see that the problem poses the
following constraint on the data
Z 1 Z 1
0= u0 10 = f + g1 + g0 .
0 0
Clearly, if u would be a solution of the Neumann problem, then u+α for arbitrary α∈R
would be also a solution of the Neumann problem. To x α we impose the additional
condition
Z 1
u = 0.
0
With this normalisation we will obtain a well-posed problem on
n Z 1 o
1
V := v ∈ H ((0, 1)) : v=0
0
with the norm
||v||V = ||v 0 ||L2 ((0,1))
[Dör15, Sect. 1.3.2]. Note that we have Poincaré's inequality (with meanvalue 0, see Lemma
1.6)
||v||L2 ((0,1)) ≤ CPN ||v 0 ||L2 ((0,1)) .
The discrete problem. The discrete problem reads: Find uh ∈ S 0,1 (Kh ) such that
Z 1 Z 1
u0h vh0 dx = fh vh + g1 vh (1) + g0 vh (0) for all vh ∈ S 0,1 (Kh ). (3.2)
0 0
Since 1 ∈ S 0,1 (Kh ), the discrete data also has to satisfy
Z 1
fh + g1 + g0 = 0
0
and if uh would be a solution, so would be uh + α. Thus we formulate the discrete problem

on
n Z 1 o
0,1
Vh := vh ∈ S (Kh ) : vh = 0
0
20
and get in this way Vh ⊂ V , hence a V -conforming method. The proof for existence of a
discrete solution follows also as in the continuous problem. To fulll the data constraint,
we have to construct fh with
Z 1 Z 1
fh = f.
0 0
This can be achieved by rst taking any approximation feh of f and then dening
Z 1
fh := feh − feh − (g1 + g0 ) =: feh − c0 . (3.3)
0
Explicit formulas. We already know the equations on the inner nodes and thus only
consider the rst equation of our linear system
Z 1 Z 1
u0h φ00 = fh φ0 + g1 φ0 (1) + g0 φ0 (0).
0 0
Inserting the relevant components of uh and exploiting φ0 (1) = 0, φ0 (0) = 1, we get

Z h1 Z h1
(u0 φ00 + u1 φ01 )φ00 = fh φ0 + g0 .
0 0
This results in
h1
1 1 1 1 1
Z
u0 − + u1 − = h1 fh h1 + g0 ,
0 h1 h1 h1 2 2
and nally
1 1 1
(u0 − u1 ) = h1 fh h1 + g0 .
h1 2 2
Analogously for the last equation
1 1 1
(−uN −1 + uN ) = hN fh 1 − hN + g1 .
hN 2 2
The linear system for an equidistant mesh reads
      
1 −1 u0 fh,0 g0
−1 2   u1   fh,1   0 
1 .. .. ..   ..  1  ..   .. 
     
= h  +  .

. . .
h  .  2  .   . 
    
 2 −1  uN −1  fh,N −1   0 
−1 1 uN fh,N g1
Consistency. Again we only consider the rst equation, evaluated for the solution u,
1 1 1
u(0) − u(h1 ) − h1 fh h1 − g0
h1 2 2
1 1
= u(0) − u(0) + u0 (0)h1 + u00 (0)h21 + O(h31 )
h1 2
1 1
− h1 fh h1 − g0
2 2
1 1 1 1
= g0 h1 + f (0)h21 + O(h31 ) − h1 f h1 − c0 − g0
h1 2 2 2
1 1 1
= h1 f (0) − f h1 + h1 c0 + O(h21 )
2 2 2
= O(h21 ),
21
R1 R1
since c0 = 0 fe + g0 + g1 = 0 {fe − f } = O(h2max ) if f ∈ W 2,1 (Ω).
Note: Consistency is an important tool to prove convergence of nite dierence methods,
but it is not used in the convergence theory of nite element methods.
Now we can state the following theorem.
Theorem 3.9. The Neumann problem (3.1) has a unique solution u ∈ H2 ((0, 1)) ∩ V . The
discrete solution (3.2), with fh dened as in (3.3), admits a unique solution in Vh . For
f ∈ H2 ((0, 1)) we have the error estimate
||(u − uh )0 ||L2 ((0,1)) ≤ Chmax ,
with a constant C depending on f, g0 , g1 .

Proof. We already dened spaces V and Vh for which the LaxMilgram theorem applies.
2
u ∈ H ((0, 1)) follows directly from the weak equation and ||u00 ||L2 (0,1) is easily seen to be
bounded in terms of ||f ||L2 ((0,1)) + |g0 | + |g1 |. Therefore we only have to check the terms in
Strang's lemma. First we dene a suitable interpolation operator Ieh : V → Vh by
Z 1
Ieh u := Ih u − Ih u,
0
Ih from Lemma 3.4. Then we have, since u ∈ H2 ((0, 1)),
||(u − Ieh u)0 ||L2 ((0,1)) = ||(u − Ih u)0 ||L2 ((0,1)) ≤ Chmax ||u00 ||L2 ((0,1)) .
Therefore we have a bound for the rst term in Strang's lemma. Since the bilinear form
does not have to be approximated, it remains to consider the f -integral. We had by
denition
Z 1
f − fh = f − feh − feh − g1 − g0
0
Z 1
= f − feh + feh + g1 + g0 .
0
R1
Since
0 wh = 0 for wh ∈ V h , the constant correction disappears in
Z 1 Z 1
(f − fh )wh = (f − feh )wh .
0 0
Q1 is dened by feh bK = f (xK ) on K and this gives an O(h2max ) bound as in Section 3.1.5,
exploiting f ∈ H2 ((0, 1)).
3.2. The two-dimensional case

3.2.1. Triangulations
Denition 3.10 (Triangulation) .

Ω ⊂ R2 be a bounded domain with a polygonal
Let
boundary. An (admissible) triangulation of Ω is a decomposition of Ω into a set of
triangles
Th := {T : T nondegenerated closed triangle}
which fullls the following conditions:

S
• Ω= T ∈Th T,
• If T, T 0 ∈ Th and T 6= T 0 , then exactly one of the following statements is true:
22
3.2 The two-dimensional case 23
◦ T ∩ T 0 = ∅,
◦ T ∩ T0 is a single vertex,
◦ T ∩ T0 is a whole common edge.
Admissible refers primarily to the second condition. Further we dene the meshsize
hT := diam(T ) for all T ∈ Th .
An important geometric quantity is the shape constant

hT
σTh := max {σT } := max , (3.4)
T ∈Th T ∈Th 2ρT
where ρT is the radius of the inscribed circle in T. Note that σTh > 1.
Example 3.11 (Meshes). Figure 1 shows some types of meshes.
n+1
1 n
Figure 1: Ω = (0, 1)2 : criss, cross, criss-cross and unstructured mesh.
3.2.2. The discrete problem
Denition 3.12 (Linear elements). The space of linear nite elements on Th is
S 1,0 (Th ) := vh ∈ C 0 (Ω) : vh bT ∈ P1

for all T ∈ Th ,
and the subspace with homogeneous boundary conditions is
S01,0 (Th ) := vh ∈ S 1,0 (Th ) : vh b∂Ω = 0 .

Denition 3.13 (Nodal basis). We dene

N h := x ∈ Ω : x is vertex of a T ∈ Th ,
Nh∂ := N h ∩ ∂Ω, (boundary nodes)
Nh := N h \ Nh∂ . (interior nodes)
Again we can dene the nodal basis {φp }p∈N h uniquely by
φp (q) = δpq for all p, q ∈ N h .
According to this denition each vh ∈ S 1,0 (Th ) is given by
X
vh (x) = vp φp (x)
p∈N h
and each vh ∈ S01,0 (Th ) is given by
X
vh (x) = vp φp (x),
p∈Nh
23
where vp := vh (p). We may again identify vh with a vector ~v ∈ R|N h | for a given enumer-
ation. The linear elliptic boundary value problem from Section 1.2 reads: Find u ∈ H01 (Ω)
such that
Z Z
u D · ∇v − ce
u D v}.

a∇u · ∇v + cuv = {f v − a∇e
Ω Ω
By the Galerkin procedure we get the system
~ u = ~b,
A~
with
Z
Apq := {a∇φq · ∇φp + cφq φp },
ZΩ
bp := uhD · ∇φp − ce
{f φp − a∇e uhD φp },
Ω
where we dene ehD ∈ S 1,0 (Th )

u by
ehD = Ph u
u eD
for some suitable operator Ph : Hk (Ω) → S 1,0 (Th ), k ≥ 1.

b
p′′
b
p
b
p′
Figure 2: Considered mesh with corresponding nodes.
Example 3.14. Consider the case of the Poisson problem
−∆u = f in Ω = (0, 1)2 ,

and the approximation on a criss mesh (see Figure 1). We compute the matrix elements
explicitly (see Figure 2):
Z 1 1 √ 2 2 1
2
∇φp · ∇φp = 22
h 4+2 h2 = 4,
Ω h h 2
Z 1 √2 √2 1
∇φp · ∇φp0 = − h2 2 = −1,
h h 2 2
ZΩ
∇φp · ∇φp00 = 0.
Ω
Thus the row of ~

A that belongs to node p ∈ Nh is

... − 1 ... − 1 ... 4 ... − 1 ... − 1 ... ,
which is the 5-pt stencil known from nite dierence methods [Dör21b, Bsp. 2.13] (in an
arbitrary numbering).
Theorem 3.15. Consider problem (1.6) for a, c ∈ C 0,1 (Ω), a ≥ a0 > 0, c ≥ 0, f ∈ L2 (Ω)
2
on a polygonally bounded and convex domain Ω ⊂ R with an admissible triangulation Th .
Assume that there is a function u
D 2
e ∈ H (Ω) with u e D b∂Ω = uD . Then the solution u is in
H2 (Ω) and if the integrals are evaluated exactly, the discrete problem is solvable on S 1,0 (Th )
and it holds
||u − uh ||E = O(hmax ).
24
Proof. We formulate the problem as in (1.8) for u ∈ V = H01 (Ω) and uh ∈ Vh = S01,0 (Th ).
We follow the proof of the Strang lemma because of the term with the boundary data.
Since all other integrals are exactly computed we obtain β0 = 1 and
uD − u
A[e ehD , wh ]
||u − uh ||E ≤ 2 inf ||u − vh ||E + sup
vh ∈Vh wh ∈Vh \{0} ||wh ||E
uD − u
≤ 2 inf ||u − vh ||E + ||e ehD ||E .
vh ∈Vh
Rewriting the solutions as eD

u+u and ehD
uh + u we obtain
e D ) − vh ||E + 2||e
||u − uh ||E ≤ 2 inf ||(u − u uD − u
ehD ||E .
vh ∈Vh
The assumptions imply u ∈ H2 (Ω). Then the results follow with the interpolation results
from Section 6: There exist operators Ph : V → V h with
||v − Ph v||E ≤ Chmax
for all v ∈ H2 (Ω) ∩ H01 (Ω).
3.2.3. Numerical integration
Let T ∈ Th and {ξ l }l ⊂ T and {ωl }l ⊂ R (preferably ωl > 0). Then the corresponding
quadrature rule is given by
Z mk
X
v ≈ |T | ωl v(ξ l ) =: QTk [v] for all v ∈ C 0 (T ),
T l=1
where, as usual, the consistency order is dened by the highest degree k≥1 such that
Z
v = QTk [v] for all v ∈ Pk
T
and m k ∈ N.
Example 3.16 (Quadrature formulae on triangles). We consider the cases k ∈ {1, 2}. Let
xT be the barycentre of T and xE the midpoint of the edge E ⊂ ∂T . Then we use the
quadrature rules
QT1 [v] := |T |v(xT ) (m1 = 1),

1 X
QT2 [v] := |T | v(xE ) (m2 = 3)
3
E⊂∂T
for v ∈ C 0 (T ). Without proof (see Theorem 6.4) we state

(
Ch2T ||∇2 v||L1 (T ) , for k = 1,
Z
v − QTk [v] ≤
T Ch3T ||∇3 v||L1 (T ) , for k = 2,
for all v ∈ C ∞ (T ).
Theorem 3.17. uD = 0, a, c, f ∈ C 2 (Ω), a ≥ a0 > 0, c ≥ 0,
Consider problem (1.6) for
2
on a polygonally bounded and convex domain Ω ⊂ R with an admissible triangulation Th .
2
By these requirements the solution u is in H (Ω). For the integrals use quadrature rules of
1,0
consistency order 1 (i.e., Q1 ). Then the discrete problem is solvable on Vh = S0 (Th ) and
it holds
||u − uh ||E = O(hmax ).
25
Proof. We follow the proof of Theorem 3.8. The proofs of the estimates are actually the
same, since |∇uh | is constant on each triangle. Consider for example the F -term in the
Strang lemma. We rst get, for arbitrary vh ∈ V h ,
Z Z
T 2
f vh − Q1 [f vh ] ≤ ChT ∇2 (f vh )
T T
≤ Ch2T ||∇f ||H1 (T ) ||vh ||H1 (T ) .
Hence, after summation,
X
Fh [vh ] − F [vh ] ≤ C h2T ||∇f ||H1 (T ) ||vh ||H1 (T )
T ∈Th
CS X 1/2
. h4T ||∇f ||2H1 (T ) ||vh ||H1 (Ω) .
T ∈Th | {z }
≤ C||vh ||E
Collecting the results yields
||u − uh ||E ≤ (1 + β0−1 ) inf ||u − vh ||E + β0−1 Ca ||Ph u||E + Cf h2max .

vh ∈Vh
By assumptions u ∈ H2 (Ω). Then the results follows with suitable interpolation results
from Section 6.
3.2.4. Piecewise smooth boundary
Let Ω ⊂ R2 be bounded. Assume that ∂Ω is piecewise smooth and that Th is a triangulation

as in Denition 3.10, but with
[
Ω h := T.
T ∈Th
Notice, that in general Ω 6= Ωh and neither Ω ⊂ Ωh nor Ω ⊃ Ωh holds. Concerning the

boundary nodes of the triangulation, we assume that they are in ∂Ω ,
Nh∂ = N h \ Nh ⊂ ∂Ω. (3.5)
We can set up a discrete problem, here the Poisson problem, by

Z Z Z
∇uh · ∇vh = fh vh − uhD · ∇vh
∇e for all vh ∈ Vh = S01,0 (Th ),
Ωh Ωh Ωh
but we have to construct fh and u ehD in a suitable way. The problem is in general noncon-
1,0 (T ) and S 1,0 1 1
forming, since S h 0 (Th ) will not be subspaces of H (Ω) and H0 (Ω) respectively.
1,0
But if Ω is convex, then Ωh ⊂ Ω and if we extend each function in S0 (Th ) by zero
1,0 1 1,0 (T ) to
outside Ωh , we get S0 (Th ) ⊂ H0 (Ω). Further we can extend each function in S h
∂ 1,0 1
Ω \ Ωh (since it is a polynomial by our choice of Nh ) and also get S (Th ) ⊂ H (Ω). This
will yield the known result
uD − u
||u − uh ||E . inf ||u − vh ||E + ||e ehD ||E
vh ∈Vh
in case of exact integration. Note that the inuence of the boundary approximation is now
hidden in the construction of the interpolation operators. For smooth (say, C 2 ) boundaries
this approximation error will be O(hmax ). The non-convex case is quite technical and will
not be considered here.
26
Figure 3: Assignment of the dierent notations to a triangle. Vertices and neighbours of

triangle Tk .
3.2.5. Data structures
Consider linear elements in 2D. We assign numbers to triangles and vertices:
T1 , . . . , TNτ , x1 , . . . , xNν .
With respect to triangle k we will nd vertices xNk,1 , xNk,2 , xNk,3 and neighbouring triangles
TMk,1 , TMk,2 , TMk,3 as in Figure 3. We need several lists for organising the mesh:
coord[i, :] = [xi,1 , xi,2 ], (coordinates)
nodes[k, :] = [Nk,1 , Nk,2 , Nk,3 ], (nodes)
neigh[k, :] = [Mk,1 , Mk,2 , Mk,3 ], (neighbours)
where the reference is always in mathematical positive direction (counter-clockwise). The

lth neighbour is opposite to the lth vertex. A negative value is assigned, if no neighbour
exists.
4 3
b b
3
5
4 b
2
1
b b
1 2
Figure 4: Triangulation with numbering of vertices and triangles.
Example 3.18 (Data structure). For the mesh given in Figure 4, we get the lists:
coord = [0, 0; 1, 0; 1, 1; 0, 1; 0.5, 0.5],

nodes = [1, 2, 5; 2, 3, 5; 3, 4, 5; 4, 1, 5],
neigh = [2, 4, −1; 3, 1, −1; 4, 2, −1; 1, 3, −1].
The matrix is set up according to the numbering of the vertices. The are several possi-
bilities to handle the matrix:
• One may store the matrix (but never N × N !). Compute ~

A as
1: function A = setSystemMatrix(Th )
~
2: ~ = 0;
A
3: for all T ∈ Th do
4: for all i ∈ nodes[T, : ] do
5: for all j ∈ nodes[T, : ] do
27
28 4 Quadratic elements
6: Aij = Aij + Ah [φj , φi ]bT ;
• The matrix can be stored in the compressed row storage +

[BBC 93], so store
S = [. . . , Aii , Aii1 , . . . , Aiim , . . . ],

| {z }
ith row of A
~
J = [. . . , i, i1 , . . . , im , . . . ],
I = [. . . , Ii , . . . ].
For given i we nd the second indices in J at positions Ii , . . . , Ii+1 − 1 and the eld
entries SIi ,. . . ,SIi+1 −1 are the matrix entries Ai,JIi ,. . . ,Ai,JIi+1 −1 .
• We can omit the storing of ~

A and instead implement the routine w ~ v:
~ = A~
1: function w~ = MatVecMult(Th ,vh )
2: ~ = ~0;
w
3: for all T ∈ Th do
4: for all i ∈ nodes[T, : ] do
5: wi = wi + Ah [vh , φi ]bT ;
The systems of equations then have to be solved by iterative methods, especially

Krylov space methods [DR20, Sect. 4.2.3/4] that can work solely with this routine.
Such methods are called matrix free.
Literature for Section 3. [BS02, Ch. 3.1, 3.2] [Cia79] [LT05, Kap. 5] [Bra07, Ch. II, 4,
Ch. III, 1] [Dör13, Sect. 3]
4. Quadratic elements
4.1. The one-dimensional case
Denition 4.1 (Quadratic elements). Let Kh be a discretisation of Ω = (0, 1) as in Section
3.1.1 and
n o
S 2,0 (Kh ) := vh ∈ C 0 (Ω) : vh bK ∈ P2 for all K ∈ Kh .
is the nite element space of quadratic finite elements. The subspace with homoge-
2,0
neous boundary conditions is S0 (Kh ) := S 2,0 (Kh ) ∩ H01 (Ω).
Denition 4.2 (Nodal basis for quadratic elements). For the given discretisation of (0, 1)
as in Section 3.1.1 we dene by
(1)
Nh := Nh ,
the set of vertices. Now we enlarge this set by the cells' barycentres to get
(2) (1)
n 1 o
Nh := Nh ∪ xi−1/2 := (xi−1 + xi ) : i = 1, . . . , n + 1 .
2
On this set of nodes we uniquely dene the nodal basis of S 2,0 (Kh ) by
(2)
φp (q) = δpq for all p, q ∈ Nh .
As a consequence we have
dim S02,0 (Kh ) = 2n + 1.

28
φp φq b grid points
| midpoints
b b b b
| | |
p q
Figure 5: The two dierent types of basis functions, φq is called quadratic element
bubble function.
(2)
A convenient numbering of the nodes Nh is the natural numbering {pi : i = 1, . . . , 2n+
1} where odd indices refer to vertices and even indices refer to barycentres. The system
matrix ~
A for the problem in Section 1.2 then has the following structure:
• If i is odd, the corresponding row for this node has 5 entries.
• If i is even, the corresponding row for this node has 3 entries.
Thus the matrix for the natural ordering is pentadiagonal and we can use Gaussian elimi-
nation for bandmatrices. Since it is still positive and symmetric, the elimination is stable
without pivoting at a cost of O(n) operations.
Hierarchical basis. For the hierarchical basis we keep the already known linear basis
functions and add, elementwise, the quadratic surplus, i.e., the quadratic element bub-
ble functions (Figure 6). This yields a dierent basis which is not nodal anymore, since
the amplitudes of the bubble functions are not the function values, see Figure 6. The
representation of a function vh ∈ S02,0 (Kh ) in this basis is
X X
vh = vh (p)φp + vh,q φq ,
p vertex q barycentre
where
1
vh (q) = vh (pL ) + vh (pR ) + cq vh,q .
2
(2) (1)
Here pL is the left and pR is the right neighbour vertex of q ∈ N h \ Nh and cq = φq (q)
(which maybe 1, but not necessarily). If Ih is the nodal interpolation, we nd
1
vh,q = v(q) − (Ih v)(q) . (4.1)
cq
b b b
| | | |
n+1 1 n+2 2 n+3 3 n+4
Figure 6: Linear basis functions plus quadratic bubble functions.
We rst take the vertices in their natural numbering and then append the barycentres
in their natural numbering. Now the system matrix looks as follows:
• For p 6= q , p, q ≥ n + 1 (both midpoints):
Z 1
aφ0p φ0q = 0.
0
29
• For p ≤ n, q > n (p vertex, q midpoint):

Z 1 Z Z
0 0 0 0 0
aφp φq = aφp φq = φp bsupp(φq ) aφ0q
0 supp(φq ) supp(φq )
Z
a∈C 1
= −φ0p bsupp(φq ) a0 φq .
supp(φq )
Thus the system matrix ~ ∈ R2n+1,2n+1 has the structure

A
" #
~ lin A
A ~ cou
~=
A ~ qua ,
~ †cou D
A
where
~ lin ∈ Rn,n is the tridiagonal matrix from linear nite elements,

• A
~ qua ∈ Rn+1,n+1 is diagonal,
• D
~ cou ∈ Rn,n+1 is the coupling part that has in general 2 non-zeros per row,
• A but
vanishes in the special case when a is piecewise constant wrt the triangulation.
Now we distinguish between two cases:
• If a is piecewise constant, then we have to solve
~ lin ~ulin = ~blin ,

A
~ qua ~uqua = ~bqua ,
D
where we use a P2 -consistent quadrature rule for ~b = [~blin , ~bqua ].

• If a is
1
piecewise in C , then the coupling entries are of the order h smaller than the
entries in the diagonal block. Therefore the block-diagonal part is a good precondi-
tioner for iterative solvers [DR20, Kap. 4.4.1]. Therefore we choose
" #
~ −1
~ = Alin
B
0
0 ~ −1
Dqua
and perform for example the iteration
~ u k − ~b).
~ A~
~u k+1 = ~u k − B(
In each step we eectively solve a tridiagonal linear system
~ lin d~lin = f~lin

A
by Gaussian elimination. Alternatively, ~

B can be used as a preconditioner in the
CG-method [DR20, Kap. 4.4.3].
4.2. The two-dimensional case

Denition
S 4.3 (Nodal basis for quadratic elements) . Let Ω be polygonally bounded and
Ω= T ∈Th T for an admissible triangulation Th . Dene
S 2,0 (Th ) := vh ∈ C 0 (Ω) : vh bT ∈ P2 for all T ∈ Th ,

S02,0 (Th ) := vh ∈ S 2,0 (Th ) : vh b∂Ω = 0 ,

(2) (1)
Nh := Nh ∪ x ∈ Ω : x midpoint of an edge .
Again dene uniquely the nodal basis by
(2)
φp (q) = δpq for all p, q ∈ Nh .
30
4.3 Error estimates 31
For the criss-triangulation of a rectangle with n elements per dimension (Figure 1) we

have ∼ n2 vertices, ∼ 2n2 triangles, and ∼ 3n2 edges (modulo O(n) for large n). This
yields
dim S02,0 (Th ) = {inner

vertices} ∪ {inner edges of Th }
2 2 2
∼ n + 3n = 4n
(a general result of Euler connects the number of vertices V, edges E and faces F in a
planar graph (without holes) by V − E + F = 1).
The following structure of the system matrix results on this mesh:
• The basis function corresponding to a vertex p couples with itself, with the basis
functions of the 6 adjacent vertices and with 12 edge-based basis functions, which
results in 19 possible nonzero entries in the row belonging to p.
• The edge-based basis function corresponding to an edge-centre q couples with it-
self, with the basis functions of the 4 adjacent vertices and with 4 edge-based basis
functions, which results in 9 entries in the row belonging to q.
The hierarchical concept applies analogously to the 1D-case, we take the linear nodal
basis and add the edge based functions. Decoupling now does not happen, since supp(φq )
(q an edge-centre) does not lie completely inside one triangle.
Remark 4.4 (Data structure). To adjust the data structure from Section 3.2.5, we enlarge
the nodes array by the centres of the edges in positions 46, where the j -th edge-centre is
opposite the j -th vertex.
4.3. Error estimates

Theorem 4.5. uD = 0, a, c, f ∈ C 3 (Ω), a ≥ a0 > 0, c ≥ 0
Consider problem (1.6) for
2
on a polygonally bounded and convex domain Ω ⊂ R with an admissible triangulation
2,0
Th . We discretise the problem on S0 (Th ), assuming that the quadrature rules are at least
of consistency order 2. If hmax is small enough, the discrete solution uh exists and, if
u ∈ H3 (Ω), it holds
||u − uh ||E = O(h2max ).
Proof. The assertion is proved by applying Strang's lemma. For simplicity, we omit the
c, f -integrals, they are estimated with the same technique. For arbitrary choice of T ∈ Th ,
q ∈ P21 and vh , wh ∈ Vh := S02,0 (Th ) we get, thanks to the required consistency,
Z
(a∇vh − q) · ∇wh − QT2 (a∇vh − q) · ∇wh

A[vh , wh ]bT − Ah [vh , wh ]bT =
T
2
. hT ||a∇vh − q||L∞ (T ) ||∇wh ||L∞ (T ) .
Note that we have to use the L∞ (T )-norm since QT2 uses point evaluations. We can now
2
take the inmum over q ∈ P1 and with the BrambleHilbert lemma (see Section 6), an
−1 3
inverse estimate ||∇wh ||L∞ (T ) . hT ||∇wh || 2
L (T ) and ∇ vh = 0, we proceed with
A[vh , wh ]bT − Ah [vh , wh ]bT . h4T ||∇2 (a∇vh )||L∞ (T ) h−1 T ||∇wh ||L2 (T )

. h2T ||a||W 2,∞ (T ) ||∇2 vh ||L2 (T ) + ||∇vh ||L2 (T ) ||∇wh ||L2 (T ) .
31
Applying Strang's lemma with vh = Ph u, for some suitable continuous operator Ph : V →

Vh , gives
1
sup A[Ph u, wh ] − Ah [Ph u, wh ]
wh ∈Vh \{0} ||w h ||E
X 1/2
. h4T ||a||2W 2,∞ (T ) ||∇2 Ph u||2L2 (T ) + ||∇Ph u||2L2 (T )
T ∈Th
≤ Ca h2max ||∇Ph u||H1 (Ω) .
We now show the coercivity of the discrete form Ah . We use the local bound on the
error A − Ah bT from the beginning (with wh = vh ), an inverse estimate (Lemma 4.6) and
1
coercivity of A on H0 (Ω) to get
X
A[vh , vh ] = Ah [vh , vh ] + A[vh , vh ]bT − Ah [vh , vh ]bT
T ∈Th
X
≤ Ah [vh , vh ] + C||a||W 2,∞ (Ω) h2T ||∇vh ||H1 (T ) ||∇vh ||L2 (T )
T ∈Th | {z }
C
≤ ||vh ||
hT H1 (T )
≤ Ah [vh , vh ] + Chmax ||vh ||H1 (Ω) ||∇vh ||L2 (Ω)

| {z } | {z }
≤ C||vh ||E ≤ C||vh ||E
≤ Ah [vh , vh ] + Chmax A[vh , vh ].
1
By choosing hmax small enough, such that Chmax ≤ 2 , we get the result
1
A[vh , vh ] ≤ Ah [vh , vh ],
2
1
hence β0 = 2.
In summary we have
1
||u − uh ||E ≤ 2||u − Ph u||E + 2Ca h2max ||∇Ph u||H1 (Ω) .
2
We made use of an interpolation operator Ph : H3 (Ω) ∩ H01 (Ω) → S02,0 (Th ) such that for
l = 0, 1, 2
||∇l Ph u||L2 (Ω) . ||∇l u||L2 (Ω)
and for l = 0, 1
||∇l (u − Ph u)||L2 (Ω) . hmax

3−l
||∇3 u||L2 (Ω) .
For this we refer to Section 6.
Lemma 4.6 (Inverse estimate) . Let T ∈ Th , p, q ∈ [1, ∞] : q < p and k, l, m ∈ N0 : k <

l ≤ m. Then we have for all vh ∈ Pm the inverse estimates
1/p−1/q
||vh ||Lp (T ) ≤ ChT ||vh ||Lq (T ) ,
||∇l vh ||L2 (T ) ≤ Chk−l k
T ||∇ vh ||L2 (T ) .
The constant C depends on p, q , resp., k, l, m, and σT .
32
4.4 Isoparametric elements 33
Proof. We assume diam(T ) = 1, which is possible by scaling with hT . We consider both

||vh ||Lq (T ) ||vh ||Lq (T ) as norms for the vector ~v , the coecients of vh in the local nite
and
element basis on T . By equivalency of all norms on nite dimensional spaces, we get the
asserted bound for hT = 1. Now we scale back to T with diameter hT .
k
For the second inequality we note that ||∇ vh || 2
2
L (T ) is a quadratic form in the coecients
of vh in the local nite element basis on T and the same holds for l instead of k . We want
to estimate the quotient of the two quadratic forms
||∇l vh ||2L2 (T )
Q := .
||∇k vh ||2L2 (T )
This quotient is well-dened for k ≤ l, since
||∇k vh ||2L2 (T ) = 0 =⇒ ||∇l vh ||2L2 (T ) = 0.
Thus Q is bounded on the compact set ||∇k vh ||2L2 (T ) = 1 (after extraction of its kernel)
Q ≤ C(k, l, m, T ).
Now we scale back to get the desired result.
4.4. Isoparametric elements

Let ∂Ω C 2 . We need to improve the boundary approximation: For linear
be smooth, say
elements we have ||∇(u−uh )|| 2
L (Ω) = O(hmax ) for piecewise linear boundary approximation
(at least for convex domains, Section 3.2.4). If we want to have a second order approxi-
mation with higher order elements, we will also have to approximate the boundary with
higher accuracy.
So we map the reference simplex
Σ2 := (λ1 , λ2 ) ∈ R2 : λ1 + λ2 ≤ 1, λ1 , λ2 ≥ 0

onto a triangle T by the ane linear mapping
F T : Σ2 → T,
(λ1 , λ2 ) 7→ p0 + λ1 (p1 − p0 ) + λ2 (p2 − p0 ) = (1 − λ1 − λ2 ) p0 + λ1 p1 + λ2 p2 , (4.2)
| {z }
=:λ0
where p0 , p1 , p2 are the corners of T and λ0 , λ1 , λ2 are the barycentric coordinates .

We can write F T in the form (λ = [λ1 ; λ2 ])

λ1
p1 p0 , p2 p0 + p0 ,

F T (λ) = − −
λ2
thus FT has the dierential
∇F T = ∇λ F T = p1 − p0 , p2 − p0 ∈ R2,2 .

We transform integrals over T = FT (Σ2 ) into integrals over the reference triangle
Z Z
v(x)w(x) dx = v(F T (λ))w(F T (λ)) det(∇F T (λ)) dλ.
T Σ2
Further, with
(∇v) ◦ F T = (∇F T )−† ∇λ (v ◦ F T )
33
(the gradients are interpreted as column vectors!) it holds

Z Z
∇v · ∇w dx = (∇F T )−† ∇λ (v ◦ F T ) · (∇F T )−† ∇λ (w ◦ F T )| det(∇F T )| dλ
T Σ2
Z
= ∇λ (v ◦ F T ) · (∇F T )−1 (∇F T )−† ∇λ (w ◦ F T )| det(∇F T )| dλ.
Σ2
Since F T is ane linear, det(∇F T ) is the area of the parallelogram spanned by the rows
of ∇F T , thus
det(∇F T ) = 2|T |.
Now we extend this idea to approximate curved boundaries. We choose FT to be of the

same polynomial order as the local nite element space, which leads to so called isopara-
metric elements. This will be done here explicitly only for quadratic elements. For
interior triangles (i.e., T ⊂ Ω) we proceed as before, we do not need to treat them in a
special way. At the boundary we assume, that (3.5) holds and T ◦ ⊂ Ω. This holds for
convex domains. Additionally, at most one edge of T should be a boundary edge, otherwise
we rene such boundary triangles until this constraint is fullled. For such a T we proceed
as follows:
(1) Assume that the triangulation is so ne, that each boundary arc between two boundary
vertices is a graph over the corresponding edge.
(2) Take the edge facing the boundary, denote the midpoint of this edge by p and the
normal unit vector to this edge by e, pointing towards the boundary.
(3) Represent ∂Ω ∂Ω = {x : g(x) = 0} for some function g .

for this edge locally by
(4) Find t > 0 such that g(p + te) = 0. Now q := p + te denes a point on ∂Ω .
(5) Dene the curved triangle Te from T by replacing the mentioned edge with the boundary
arc that is the quadratic interpolant determined by the two edge points and q .
Figure 7: The dierent transformations of the standard simplex.
In other words, we dene F

eT by the unique quadratic mapping, which maps the corners
of Σ2 onto the corresponding corners of Te Σ2
and the midpoints of the edges of onto the
midpoints of edges of T
e (Figure 7). With notations as in (4.2) we can express F
eT in terms
of the hierarchical quadratic basis (see Section 4.1) as
e T : Σ2 → Te,
F
1
(λ1 , λ2 ) 7→ λ0 p0 + λ1 p1 + λ2 p2 + 4λ1 λ2 q − (p1 + p2 ) .
2
34
4.4 Isoparametric elements 35
Thus eT ≡ FT
F for interior triangles T . Now we dene the nite element space on the
approximated domain e h := S
Ω e T (Σ2 ) by
F
T ∈Th
n o
Vh := vh ∈ C 0 (Ω
e h ) : vh ◦ F
e T ∈ P2 (Σ2 ) for all T ∈ Th .
Note that vh is not anymore in P2 on boundary triangles due to the quadratic mapping,
but it is still on interior triangles. The bilinear form then reads
Z X Z
Ah [vh , wh ] = ∇vh · ∇wh = ∇vh · ∇wh
Ω
eh
T ∈Th Te
X Z
= ∇λ (vh ◦ F e −1 ∇λ F
e T ) · ∇λ F e −t ∇λ (wh ◦ F
e ) | det(∇λ F
e ) | dλ,
Σ2 | {z } T T | {z T } | {z T }
T ∈Th
| {z }
∈P1 rational functions ∈P1 ∈P2
which means that we have to integrate by a higher order quadrature rule.

Now we want to extend functions in Vh onto the whole of Ω. For inner triangles let
ψT := F T . For boundary triangles ψT shall map Σ2 onto the triangle which is bounded
by the exact boundary, see Figure 7. However, ψT should be ane linear on the straight
edges in order to guarantee global continuity of our local denitions across inner edges. By
construction we have
[
Ω= ψT (Σ2 ).
T ∈Th
For T ∈ Th let
e T ◦ ψ −1 : ψT (Σ2 ) → Te = F
φT := F e T (Σ2 ),
T
and then
φ: Ω → Ω
eh,
φ(x) := φT (x) for x ∈ Tb := ψT (Σ2 ).
For arbitrary vh ∈ V h we then dene
vbh : Ω → R,

vbh (x) := vh φ(x) for x ∈ Ω.
In order to impose the discrete problem, we have to explain the discrete right hand side
Ω , we will use f ◦φ−1 : Ω
R
Ω f w. Since it is unclear whether f can be dened outside eh → R
instead. Our discrete problem is thus to nd uh ∈ Vh such that
Z
Ah [uh , wh ] = (f ◦ φ−1 )wh for all wh ∈ V h .
Ω
eh
Theorem 4.7. Let Ω ⊂ R2 be a bounded domain with piecewise C 3 -boundary. Let Th be an

admissible triangulation with piecewise quadratic boundary approximation as above, where
singular points coincide with boundary vertices. Let u and uh ∈ Vh be the solutions to the
D
continuous and discrete Poisson problem with u = 0 and f ∈ H1 (Ω). With u bh = uh ◦ φ
we have the error estimate
bh )||L2 (Ω) ≤ Ch2max ,

||∇(u − u
assuming u ∈ H3 (Ω) ∩ H01 (Ω), exact integration and a suciently small hmax .
35
Proof. Let

Vb h := vh ◦ φ : vh ∈ Vh
be the space of exact elements and let zbh ∈ Vb h be the solution of the variational problem
A[b
zh, w
bh ] = A[u, w
bh ] for all w
bh ∈ Vb h ,
where u is the solution of the continuous problem. zbh exists by the standard theory, since
V h ⊂ H01 (Ω). For all w
b bh ∈ Vb h we have
A[b
zh − u
bh , w
bh ] = A[u, w
bh ] − A[b
uh , w
bh ]
Z Z
= fwbh − (f ◦ φ−1 )wh + Ah [uh , wh ] − A[b
uh , w
bh ],
Ω Ω
eh
where we have assumed exact integration to dene uh . Existence of uh follows from coer-
civity of Ah that is proved below. We have further, by optimality of zbh in Vb h ,
||∇(u − u
bh )||L2 (Ω) = ||u − u
bh ||E
≤ ||u − zbh ||E + ||b
zh − u
bh ||E
A[b
zh − ubh , w
bh ]
= inf ||u − vbh ||E + sup .
bh ∈Vb h
v bh ∈Vb h
w \{0}
||w
bh ||E
The second term can be estimated using the result obtained above
A[b
zh − u bh , w
bh ] Ah [uh , wh ] − A[b uh , w
bh ]
sup ≤ sup
bh ∈Vb h \{0}
w
|| w
b ||
h E bh ∈Vb h \{0}
w
|| w
b ||
h E
R R
Ω fw bh − Ω f w bh | det(∇φ)|
+ sup .
wb ∈Vb \{0}
||w
bh ||E
h h
We only estimate the rst term on the right, the same ideas apply to the remaining terms.
If we transform Ah to an integral over Ω, we can split this error term as
Z n o
Ah [uh , wh ] − A[b
uh , w
bh ] = ∇φ−† ∇buh · ∇φ−† ∇wbh | det(∇φ)| − ∇b uh · ∇w
bh
ZΩ
∇φ−† − Id ∇b uh · ∇φ−† ∇w

= bh | det(∇φ)|
Ω
Z
uh · | det(∇φ)|∇φ−† − Id ∇w

+ ∇b bh .
Ω
By assumptions on ∂Ω we can derive the local bounds
(
0, if T is an interior triangle
||φT − id||L∞ (Tb) = .
O(h3T ), if T is a boundary triangle
Using this and an inverse estimate (Lemma 4.6) we obtain
C
||∇φT − Id||L∞ (Tb) ≤ ||φT − id||L∞ (Tb) = O(h2T ),
hT
||∇φ−† −† † †
T − Id||L∞ (Te) = ||∇φT (Id − ∇φT ) ||L∞ (Te) ≤ C||(Id − ∇φT ) ||L∞ (Tb)
= C||Id − ∇φT ||L∞ (Tb) = O(h2T )
36
37
and from this
det(∇φT ) = det(Id + ∇φT − Id)

= det(Id) + det 0 (Id)[∇φT − Id] + O ||∇φT − Id||2L∞ (Tb)

= 1 + O(h2T ),
and
|| det(∇φT )∇φ−† 2
T − Id||L∞ (Te) = O(hT ).
This yields
X
Ah [uh , wh ] − A[b
uh , w
bh ] ≤ C h2T ||∇b
uh ||L2 (Tb) ||∇w
bh ||L2 (Tb)
T ∈Th
≤ Ch2max ||b
uh ||E ||w
bh ||E
The coercivity of Ah follows from that of A and the bound for A − Ah as in the proof of
Theorem 4.5. Here we need hmax to be suciently small. So we obtain
bh ||E ≤ 2 inf ||u − vbh ||E + Ch2max ||b

||u − u uh ||E .
bh ∈Vb h
v
To estimate the inmum we need interpolation results which will be derived later in Section
6. A bound for ||b
uh ||E follows from the bound for ||uh ||E . The f -term will also result in
O(h2max ).
Literature for Section 4. [BS02, Ch. 3.1-3.2, 10.4] [Cia79] [LT05, Kap. 5] [Bra07, Ch. II,
5, Ch. III, 2] [Dör13, Sect. 3.2]
5. Tensorproduct elements
5.1. Discretisation of rectangular domains into rectangles
Let Ω = (0, 1)2 . We now choose a decomposition Rh
[
Ω= R,
R∈Rh
that consists of closed rectangles R. Under the conditions we formulated for an admissible
triangulation we are limited to domains composed of rectangles. Moreover, any choice of
a small element of size h1 × h2 would enforce a whole column with meshsize h1 and a
whole row with meshsize h2 . In order to get rectangles of quite dierent size inside the
decomposition one has to use so called hanging nodes, i.e., nodes which lie on edges.
These nodes do not carry any degree of freedom, the function values in hanging nodes are
always dened by interpolation of correct order from their neighbours (compare [Dör21b,
Sect. 2.6.1]). Moreover, we have to take care, if we rene towards a point, because we can
produce more hanging nodes than degrees of freedom, see Figure 8 on the left. On the
right we present a possibility to avoid this issue. Note that hanging nodes increase the
work, but do not improve the approximation properties.
37
38 5 Tensorproduct elements
bc b bc b b
bc b bc b b b
b bc b b bc
rs rs
Figure 8: Renement towards a point with two renement strategies.
5.2. Tensor product nite elements

5.2.1. The bilinear element
Let {φb0 , φb1 } be the linear nodal nite element basis on [0, 1]. On the reference rectangle
b := [0, 1]2 we dene
R
ψbk (ξ1 , ξ2 ) := φbi (ξ1 )φbj (ξ2 ),
where k = 2j + i + 1 ∈ {1, 2, 3, 4}. These functions are bilinear since

ψbk ∈ span ξ1 , ξ2 , ξ1 ξ2 .
On a general rectangle R with corners p1 , . . . , p4 these functions correspond to the nodal
basis {ψ1 , . . . , ψ4 } dened uniquely by
ψpk (pl ) = δkl .

We dene ψp globally this way on all rectangles. These functions are globally continuous
and locally supported with
[
supp(ψp ) = R ∈ Rh : p ∈ R .
The resulting spaces of ansatz functions are
Q(1,1),0 := vh ∈ C 0 (Ω) : vh bR ∈ P1 ⊗ P1 for all R ∈ Rh ⊂ H1 (Ω),

(1,1),0
:= vh ∈ Q(1,1),0 : vh b∂Ω = 0 ⊂ H01 (Ω).

Q0
Since Q(1,1),0 contains the linear polynomial space (but not the quadratic one), we will get
convergence as for linear nite elements.
5.2.2. The biquadratic element
Let {φb0 , φb 1 , φb1 } be the quadratic nodal basis on [0, 1], φb 1 being the bubble function. On
2 2
R
b we dene the biquadratic basis {ψb1 , . . . , ψb9 } by all combinations
1
ψbk (ξ1 , ξ2 ) = φbi (ξ1 )φbj (ξ2 ), i, j ∈ 0, , 1 .
2
The support of the quadratic basis functions is the same as for the bilinear basis functions
except for the bubble functions, where i = j = 1/2, which are supported on one rectangle
only. One could also use a hierarchical polynomial basis, by instead using the linear nodal
functions for φb0 and φb1 . In this way we get the biquadratic spaces
Q(2,2),0 := vh ∈ C 0 (Ω) : vh bR ∈ P2 ⊗ P2 for all R ∈ Rh ⊂ H1 (Ω),

(2,2),0
:= vh ∈ Q(2,2),0 : vh b∂Ω = 0 ⊂ H01 (Ω).

Q0
Since Q(2,2),0 contains the quadratic polynomial space (but not the cubic one), we will get
convergence as for quadratic nite elements.
38
5.3 The stiness matrix 39
5.3. The stiness matrix

5.3.1. The bilinear element
We consider b = [0, 1]2

R and for p1 = (0, 0), p2 = (1, 0) the corresponding basis functions
ψb1 (ξ1 , ξ2 ) = φb0 (ξ1 )φb0 (ξ2 ),

ψb2 (ξ1 , ξ2 ) = φb1 (ξ1 )φb0 (ξ2 ).
For the matrix element S12 of the stiness matrix we nd

Z Z

∇ψb1 · ∇ψb2 = ∂1 ψb1 ∂1 ψb2 + ∂2 ψb1 ∂2 ψb2
R
b R
b
Z 1 Z 1
0 0
= φ0 (ξ1 )φ1 (ξ1 ) dξ1
b b φb0 (ξ2 )φb0 (ξ2 ) dξ2
0 0
Z 1 Z 1
+ φ0 (ξ1 )φ1 (ξ1 ) dξ1
b b φb00 (ξ2 )φb00 (ξ2 ) dξ2
0 0
(1) (1) (1) (1)
= S01 M00 + S00 M01 .
~ (1)
S and ~ (1)
M are the 1D-stiness and 1D-mass matrix, respectively. In 2D we nally nd
~ (2) = S
S ~ (1) ⊗ M
~ (1) + M
~ (1) ⊗ S
~ (1) ,
Rb b R b Rb Rb R
where we have dened for A, B ∈ R2×2 the Kronecker product

AB11 AB12
A ⊗ B := ∈ R4×4 .
AB21 AB22
The resulting system matrix is again sparse and has 9 non-zero entries per row.
5.3.2. The biquadratic element
We follow the same lines as in the previous subsection and with ~ (1) , M
S ~ (1) ∈ R3×3 we get
Rb R b
S ~ (1) ⊗ M
~ (2) = S ~ (1) + M
~ (1) ⊗ S
~ (1) ∈ R9×9 .
Rb bR b R b R
b R
Remark 5.1 (Sparsity). We consider the sparsity of the biquadratic element. We nd the
following nonzero entries in the row (independent of the ordering) corresponding to a
vertex: 9 vertices, 12 edges and 4 bubbles, i.e., 25 entries.
edge node: 6 vertices, 7 edges and 2 bubbles, i.e., 15 entries.
bubble: 4 vertices, 4 edges and 1 bubble, i.e., 9 entries.
One way to order the unknowns is to do it elementwise, while following the elements in a
lexicographic way. Another ordering is to rst list all nodes corresponding to vertices (v),
then all nodes corresponding to edges (e), at last all nodes corresponding to the volume
bubbles (b):
†
~u = ~uv ~ue ~ub .
Then the stiness matrix is of the form

 
S~vv S
~ve S
~vb
~=
S S~ev S
~ee S
~eb 
,
S~bv S
~be S
~bb
39
40 6 Interpolation estimates
where Sbb is diagonal! Since the last equation of the corresponding linear system reads
~bv ~uv + S
S ~be ~ue + S
~bb ~ub = F~b ,
we can easily calculate
~ −1 F~b − S
~bv ~uv − S
~be ~ue .

~ub = S bb
This can be explicitly inserted into the rst two equations to get a system solely for [~uv , ~ue ].
This procedure is called static condensation.
5.4. Logical rectangular grids

Similarly to the triangular elements we can improve the boundary approximation with rect-
angular elements by appropriate mappings, see the isoparametric elements. Furthermore,
one can also work with meshes that are global images from a discretised rectangle. This is
used e.g. for the simulation of ow of air around an air plane.
Literature for Section 5. [BS02, Ch. 3.5] [Bra07, Ch. II, 5]
6. Interpolation estimates
6.1. The BrambleHilbert lemma
In this section we assume D ⊂ Rd to be bounded and ∂D ∈ C 0,1 . The following prepares
the BrambleHilbert lemma (Thm. 6.4).
Lemma 6.1. For all v ∈ Hm (D), m ∈ N, there is a unique polynomial p ∈ Pm−1 such that
Z
∇ν (v − p) = 0 for all ν ∈ Nd0 , |ν| ≤ m − 1.
D
Proof. We use multi-index notation with | . | := | . |1 to denote a p ∈ Pm−1 by
m−1
X X
p(x) = cµ xµ .
l=0 |µ|=l
Inserting this ansatz for all ν ∈ Nd0 with |ν| ≤ m − 1 we get
m−1
X X Z Z
ν µ
cµ ∇ x = ∇ν v.
l=0 |µ|=l D D
| {z }
=:Aνµ
This is a linear system with a square matrix ~ = [Aνµ ]νµ .

A To prove the lemma we have
to show that ~
A is injective. Thus it remains to show that the homogeneous problem has
only the trivial solution. For all ν with|ν| = m − 1
Z
∇ν p(x) ∼ cν =⇒ cν = 0 =⇒ cν = 0,
D
consequently p ∈ Pm−2 . Inductively going down with |ν| we get p ≡ 0.
Denition 6.2. For m ∈ N0 and v ∈ Hm (D) we dene

X
||∇m v||2L2 (D) := ||∇ν v||2L2 (D) .
|ν|=l
40
6.1 The BrambleHilbert lemma 41
Lemma 6.3. For m ∈ N there exists a constant C = C(D, m) such that for all l ∈
{0, . . . , m − 1} and all v ∈ Hm (D) holds
||∇l v||L2 (D) ≤ C ||∇m v||L2 (D) + ||v||L2 (D) .

Proof. We choose v ∈ Hm (D) and p as in Lemma 6.1 and obtain iteratively, using Lemmata
1.6 and 6.1,
||v − p||L2 (D) ≤ C1 ||∇(v − p)||L2 (D) ≤ · · · ≤ Cm ||∇m v||L2 (D) .
Especially, it holds
||∇l (v − p)||L2 (D) ≤ Cm ||∇m v||L2 (D) for all l ∈ {0, . . . , m − 1}. (6.1)
For l=0 we additionally have
−||v||L2 (D) + ||p||L2 (D) ≤ ||v − p||L2 (D) ≤ C||∇m v||L2 (D) ,
which provides a bound for ||p||L2 (D) . Together with an inverse estimate (Lemma 4.6) we
get
||∇l v||L2 (D) ≤ ||∇l (v − p)||L2 (D) + ||∇l p||L2 (D) ≤ ||∇l (v − p)||L2 (D) + C||p||L2 (D)
. ||∇m v||L2 (D) + ||v||L2 (D) .
Theorem 6.4 (BrambleHilbert lemma). Let V = Hm (D) for some m ∈ N, D as before.

LetG : V → R≥0 be a subadditive and continuous mapping, i.e., for all v, w ∈ V and
some K > 0 holds
G(v + w) ≤ G(v) + G(w),

G(v) ≤ K||v||V .
If
G(p) = 0 for all p ∈ Pm−1 ,
then there is a constant C = C(D, m) such that
G(v) ≤ C||∇m v||L2 (D) for all v ∈ V.
Proof. Let v∈V and choose p according to Lemma 6.1. Then we have
G(v) = G(v − p + p) ≤ G(v − p) + G(p)

| {z }
=0
m
X 1/2
≤ K||v − p||Hm (D) = K ||∇l (v − p)||2L2 (D)
l=0
(6.1)
≤ C||∇m v||L2 (D) .
Remark 6.5. Note that one can extract from the proof
G(v) ≤ G(v − p) for all p ∈ Pm−1 ,
and thus
G(v) ≤ inf G(v − p) ≤ K inf ||v − p||Hm (D) .

p∈Pm−1 p∈Pm−1
41
6.2. Nodal interpolation

Theorem 6.6 (Linear elements)
S . Let Ω be as in the previous section with d ≤ 3, ∂Ω
polygonal and Ω = Ωh = T ∈Th T , where Th is an admissible triangulation. We dene the
continuous interpolation operator
Ih : H2 (Ω) → S 1,0 (Th ),

X
Ih v(x) := v(p)φp (x).
p∈N h
Note that
v ∈ H2 (Ω) ∩ H01 (Ω) =⇒ Ih v ∈ S01,0 (Th ).
Then we have for all v ∈ H2 (Ω)
||v − Ih v||L2 (Ω) + hmax ||∇(v − Ih v)||L2 (Ω) ≤ Ch2max ||∇2 v||L2 (Ω) .
Proof. For 1D we have already proved this result in Lemma 3.4. Here Ih is well-dened
since
||v||L∞ (Ω) ≤ C||v||H2 (Ω)
holds for d ≤ 3 by Sobolev's embedding theorem [Dör15, Sect. 1.3.1]. Now we x a T ∈ Th

and scale T to Tb with diam(Tb) = 1. Dene for all vb ∈ H2 (Tb)
G(b
b v ) := ||b
v − Ibh vb||L2 (Tb) ,
where Ibh is the interpolation operator on Tb. For this mapping we have sublinearity and
we have the properties
G(b
b v) = 0 for all vb ∈ P1 ,
G(b
b v ) ≤ C||b
v || L∞ (Tb)
≤ C||b
v ||H2 (Tb) .
Thus we can apply the BrambleHilbert lemma, Theorem 6.4, to get
v − Ibh vb||L2 (Tb) ≤ C||∇2 vb||L2 (Tb) .

||b
Transforming back to T yields
||v − Ih v||L2 (T ) ≤ Ch2T ||∇2 v||L2 (T ) .
Analogously one proves
||∇(v − Ih v)||L2 (T ) ≤ ChT ||∇2 v||L2 (T ) .
By summation over T ∈ Th we get
X X
||v − Ih v||2L2 (T ) ≤ C h4T ||∇2 v||2L2 (Ω) ,
T ∈Th T ∈Th
which shows the L2 -estimate. The H1 -estimate is proved likewise.
42
6.3 Nodal interpolation on convex domains 43
Theorem 6.7 (Higher order elements). We consider now S k,0 (Th ) for k ≥ 1. Assume that
Ih is an interpolation operator depending on nodal function evaluations and with
Ih v bT = v bT for all T ∈ Th , v ∈ Pk .
Further, Ih is assumed to be continuous as a mapping
Ih : Hm (Ω) → S k,0 (Th ).
Then, for l ∈ {0, . . . , m}, m ≤ k + 1,
||∇l (v − Ih v)||L2 (T ) ≤ Chm−l

T ||∇m v||L2 (T ) for all v ∈ Hm (T ).
Summing up over T ∈ Th yields
||∇l (v − Ih v)||L2 (Ω) ≤ Chm−l m

max ||∇ v||L2 (Ω) for all v ∈ Hm (Ω).
Note, that continuity of Ih requires the Sobolev embedding Hm (Ω) ,→ L∞ (Ω) [Dör15,
Sect. 1.3.1 (Example)], so it has to hold
d
m> .
2
For example, for d=1 this only holds for m = 1, while for d=2 we need m > 1. The
regularity requirements increase with the space dimension d.
Proof. As for Theorem 6.6.
6.3. Nodal interpolation on convex domains

Let now Ω ⊂ R2 be convex and bounded, ∂Ω ∈ C 2 . Further let Th be an admissible
∂
triangulation with Nh ⊂ ∂Ω and
[
Ωh = T ⊂ Ω.
T ∈Th
6.3.1. Linear elements
First we consider linear elements, so
Vh := S01,0 (Th ).
We extend Vh to a function space on Ω by extending vh ∈ V h by 0 outside Ωh , hence we

can state
Vh ⊂ H01 (Ω).
Let ψT be the mapping from T toψT (T ) = Tb, where Tb is the adaption to the exact
boundary (see Section 4.4 and
2 1
Figure 7). For v ∈ H (Ω) ∩ H0 (Ω) let Ih v be the nodal
interpolant in Vh . We only have to consider the boundary triangles. For such a boundary
triangle T holds
||v − Ih v||2L2 (Tb) = ||v − Ih v||2L2 (T ) + ||v − Ih v||2L2 (Tb\T )

≤ Ch4T ||∇2 v||2L2 (T ) + ||v||2L2 (Tb\T ) .
In order to estimate the norm on Tb \ T one applies Poincaré's inequality
||v||L2 (Tb\T ) ≤ Ch2T ||∇v||L2 (Tb\T ) .
43
To show this, rst reect this domain at the secant, then apply the Poincaré estimate
(Lemma 1.4) to the lens-shaped domain. Note that the estimate actually holds with the
smallest diameter of this domain, that is here O(h2T ) by the C 2 -regularity. Summing up
yields
X
||v − Ih v||2L2 (Ω) ≤ Ch4max ||∇2 v||2L2 (Ωh ) + C h4T ||∇v||2L2 (Tb\T )
T near ∂Ω
≤ Ch4max ||∇v||2H1 (Ω) .
For the gradient we get
||∇(v − Ih v)||2L2 (Tb) = ||∇(v − Ih v)||2L2 (T ) + ||∇v||2L2 (Tb\T )

≤ Ch2T ||∇2 v||2L2 (T ) + ||∇v||2L2 (Tb\T ) .
Thus
||∇(v − Ih v)||L2 (Ω) ≤ Chmax ||∇2 v||L2 (Ωh ) + ||∇v||L2 (Ω\Ωh ) ,

| {z }
≤ Chmax ||∇v||
H1 (Ω)
where the last inequality will be proved in the next section.
6.3.2. Quadratic elements
Now we proceed with the case of quadratic elements
Vh := S02,0 (Th ),
without improved boundary approximation. Then for v ∈ H3 (Ω) ∩ H01 (Ω)
||v − Ih v||L2 (Ω) ≤ Ch3max ||∇3 v||L2 (Ωh ) + Ch2max ||∇v||L2 (Ω\Ωh )
| {z }
≤ Chmax ||∇v||
H1 (Ω)
(the last inequality is proved below). For the gradient we get
||∇(v − Ih v)||L2 (Ω) ≤ Ch2max ||∇3 v||L2 (Ωh ) + Chmax ||∇v||H1 (Ω) .
3/2
The second term can be improved to be O(hmax ), which is the best that can be achieved.
Thus
3/2
||∇(v − Ih v)||L2 (Ω) = O(hmax ),
but not O(h2max )! Thus, the optimal result can only be achieved if we use quadratic
boundary approximation.
It remains to prove an estimate for ||∇v||L2 (Ω\Ωh ) that we used two times before. Let
{γt }t∈[0,τ ] for some τ >0 be a family of smooth curves such that
• γt is a dieomorphism mapping onto ∂Ω ,

• x ∈ Ω : x ∈ γt for some t ∈ [0, τ ] ⊃ Ω \ Ωh is a neighbourhood of ∂Ω ,
• γt ∩ γt0 = ∅ for t 6= t0 .
In the last two conditions we used for simplicity γt for range(γt ). Such a family can be
found if the domain Ω is bounded and hmax is small enough with
τ ≤ Ch2max
44
6.4 The interpolation operator of Clément 45
for some constant C depending on the curvature of ∂Ω . The trace theorem for w ∈ H1 (Ω)
yields for t ∈ [0, τ ]
Z
|w|2 ≤ C||w||2H1 (Ω)
γt
[Dör15, Sect. 1.3.2]. We integrate over t ∈ [0, τ ] to get

Z τZ
||w||2L2 (Ω\Ωh ) ≤ |w(x)|2 dx dt ≤ Cτ ||w||2H1 (Ω)
0
γt
≤ Chmax ||w||2H1 (Ω) .
2
Now take w = ∇v to get the result.
6.4. The interpolation operator of Clément

Aim: Find an interpolation that also works when the pointwise evaluation is not continuous.
Let Ω ⊂ R2 be a polygonally bounded domain, Th an admissible triangulation with
[
Ω = Ωh = T,
T ∈Th
the set of nodes denoted by N h, the corresponding nodal basis {φp }p∈N h and Sp :=
supp(φp ) for p ∈ N h .
2
For v ∈ L (Ω) we dene for all p ∈ Nh polynomials γp ∈ Pk , k ∈ N, uniquely by
||v − γp ||L2 (Sp ) = min ||v − w||L2 (Sp ) .

w∈Pk
Now we dene the interpolation operator of Clément Πk by [Clé75]
Πk : L2 (Ω) → S k,0 (Th ),

X
Πk v := γp (p)φp .
p∈N h
Forv ∈ H01 (Ω) we do in general not have Πk v ∈ H01 (Ω) because γp need not to vanish
in p ∈ ∂Ω . Instead, we dene for v ∈ H01 (Ω)
X
Πk0 v := γp (p)φp .
p∈Nh
Theorem 6.8 (Clément interpolation) . For Πk as dened above one obtains for all v∈
Hm (Ω) and all m ∈ {0, . . . , k + 1} and l ∈ {0, . . . , m}
||∇l (v − Πk v)||L2 (Ω) ≤ Chm−l m

max ||∇ v||L2 (Ω) .
In fact it holds
||∇l (v − Πk v)||L2 (T ) ≤ Chm−l

T ||∇m v||L2 (ST ) for all T ∈ Th ,
where
[
ST := T 0 ∈ Th : T ∩ T 0 6= ∅ .
The corresponding results hold for Πk0 and v ∈ H01 (Ω) ∩ Hm (Ω).
45
Proof. Fix any T ∈ Th and dene

NT := p ∈ N h : p ∈ T .
Then x p ∈ NT . We can write

X X
Πk v bT = γp (p)φp = γp + γp (p) − γp (p) φp ,
p∈NT p∈NT
since γp ∈ span{φp }p∈NT = Pk (T ). Thus
X
||∇l (v − Πk v)||L2 (T ) ≤ ||∇l (v − γp )||L2 (T ) + |γp (p) − γp (p)| ||∇l φp ||L2 (T )
p∈NT
X
≤ Chm−l m
Sp ||∇ v||L2 (Sp ) +C |T |1/2 h−l
T ||γp − γp ||L∞ (T ) ,
p∈NT
by using the inverse estimate together with the fact, that ||φp ||L2 (T ) ∼ |T |1/2 . The rst
term was estimated by applying the BrambleHilbert lemma, Theorem 6.4, to G(v) :=
||∇l (v − γp )||L2 (Sp ) . Now we proceed with
1
||γp − γp ||L∞ (T ) ≤ C|T |− 2 ||γp − γp ||L2 (T )
1

≤ C|T |− 2 ||γp − v||L2 (T ) + ||v − γp ||L2 (T )
1

≤ C|T |− 2 ||γp − v||L2 (Sp ) + ||v − γp ||L2 (Sp )
Thm. 6.4 1

≤ C|T |− 2 hm m m m
Sp ||∇ v||L2 (Sp ) + hSp ||∇ v||L2 (Sp ) .
Now we have to assume some regularity of the triangulation, namely that the triangles do
not get too acute-angled, i.e., there is a constant σ>0 such that for all T ∈ Th
hT
1≤ = σT ≤ σTh ≤ σ,
2ρT
hT1
recall the denition (3.4). This gives for all T1 , T2 ⊂ ST : hT2 ≤ Cσ and thus
max{hSp , hSp } ≤ Cσ hT .
Thus we can conclude
||∇l (v − Πk v)||L2 (T ) ≤ Chm−l

T ||∇m v||L2 (ST ) ,
S
since p∈NT Sp = ST . Summing up over T gives
X X
||∇l (v − Πk v)||2L2 (Ω) = ||∇l (v − Πk v)||2L2 (T ) ≤ Ch2(m−l)
max ||∇m v||2L2 (ST )
T ∈Th T ∈Th
X
≤ Ch2(m−l)
max δ ||∇ m
v||2L2 (T ) ,
T ∈Th
where we used
δT := {T 0 ∈ Th : T ⊂ ST 0 } , δ := max δT .
T ∈Th
δ is bounded by a constant depending on σ. Thus
||∇l (v − Πk v)||L2 (Ω) ≤ Chm−l m

max ||∇ v||L2 (Ω) .
46
6.5 The interpolation operator of Scott/Zhang 47
It remains to show the estimate for homogeneous Dirichlet boundary conditions. Let
v ∈ Hm (Ω) ∩ H01 (Ω). We only have to estimate
||∇l (Πk − Πk0 )v||L2 (T ) ,
since the triangle inequality will give the desired result. By construction, (Πk − Πk0 )v
only has non-zero values at the boundary. Let T ∈ Th be a boundary triangle and E the
corresponding boundary edge where v vanishes. Then for all γ ∈ Pk
||γ||2L∞ (E) ≤ Cσ |E|−1 ||γ||2L2 (E) ≤ Cσ h−1 2

T ||γ − v||L2 (E)

≤ Ch−1
T h −1
T ||v − γ|| 2
2
L (T ) + h 2
T ||∇(v − γ)|| 2
2
L (T )
where we used again the trace theorem [Dör15, Sect. 1.3.2]. Finally,
X
||∇l (Πk − Πk0 )v||2L2 (T ) ≤ C |γp (p)|2 ||∇l φp ||2L2 (T )
p∈Nh∂
X n o
≤ Ch−2 −2l
T |T |hT ||v − γp ||2L2 (T ) + h2T ||∇(v − γp )||2L2 (T )
p∈Nh∂
X n 2(m−1)
o
≤ Ch−2l
T h2m m 2 2
Sp ||∇ v||L2 (Sp ) + hSp hSp ||∇m v||2L2 (Sp )
p∈Nh∂
≤ Ch2m−2l
T δ||∇m v||2L2 (Sp ) .
6.5. The interpolation operator of Scott/Zhang

We will shortly introduce the interpolation operator of Scott and Zhang [SZ90] in case
Ω = Ωh . It is an operator
e 0 : Hm (Ω) → S k,0 (Th ),

Π k 0 0
fork ≥ 1 and 1 ≤ m ≤ k + 1. For the construction x for each p ∈ Nh an edgeEp that

∂
contains p, and for p ∈ Nh we additionally assume Ep ⊂ ∂Ω . We only consider k = 1:
Z
Λp [v] := ψp∗ v,
Ep
X
e 01 v :=
Π Λp [v]φp .
p∈N h
This mapping is well-dened by the trace theorem [Dör15, Sect. 1.3.2] and ψp∗ : Ep → R is
a dual basis function, dened by
Z
ψp∗ φq = δpq for all p, q ∈ N h .
Ep
If v ∈ H01 (Ω), then we have Λp [v] = 0 p ∈ Nh∂ . One obtains

for the same results as for the
Clément operator from the previous section (for m ≥ 1).
Literature for Section 6. [BS02, Ch. 4] [Bra07, Ch. II, 6] [LT05, Kap. 5.3]
47
48 7 Special topics
7. Special topics
7.1. Error estimates in the L2 -norm
Theorem 7.1. Let Ω ⊂ Rd
be an open and bounded domain with polygonal boundary.
Assume that the Poisson problem is H -regular on Ω , i.e., for each g ∈ L (Ω) there is
2 2
2 1
a solution v ∈ H (Ω) ∩ H0 (Ω) of
−∆v = g in Ω,
v=0 on ∂Ω,
and there is a constant C∗ = C∗ (Ω) such that
||v||H2 (Ω) ≤ C∗ ||g||L2 (Ω) .
Now let f ∈ L2 (Ω) be the right hand side and u ∈ H2 (Ω) ∩ H01 (Ω) the solution of the
corresponding Poisson problem. Let Th be an admissible triangulation of Ω and uh the
1,0
discrete solution in S0 (Th ), where we assumed exact integration. Then we have
||u − uh ||L2 (Ω) ≤ Ch2max ||f ||L2 (Ω)
for some C = C(Ω, σTh ).
Proof. The method of proof is called the AubinNitsche trick or duality trick. We
know that u and uh exist and dene w∈ H2 (Ω) ∩ H01 (Ω) to be the solution of
−∆w = u − uh in Ω,
w=0 on ∂Ω,
where we notice u − uh ∈ H01 (Ω). By assumption, we can conclude for w
||w||H2 (Ω) ≤ C∗ ||u − uh ||L2 (Ω) .
Then we have with the error estimate for uh and wh := Π10 w (from Section 6.4)
Z
||u − uh ||2L2 (Ω) = (u − uh )(−∆w)
Ω
Z
= ∇(u − uh ) · ∇(w − wh ) (Galerkin orthogonality)
Ω
≤ ||∇(u − uh )||L2 (Ω) ||∇(w − wh )||L2 (Ω)
≤ CI2 h2max ||∇2 u||L2 (Ω) ||∇2 w||L2 (Ω)
≤ CI2 C∗2 h2max ||f ||L2 (Ω) ||u − uh ||L2 (Ω) .
Remark 7.2. Ω ⊂ R2 is convex, then the Poisson problem is H2 -regular. Otherwise

If
one can work with W
2,p -regularity for some p ∈ [1, 2) to get hκp with κ ∈ [1, 2) or work
max p
2
with norms of the form ||dΩ ∇ u|| 2 where dΩ is non-negative and has zeros of rst order
L (Ω)
in the corner points.
Literature for Section 7.1. [BS02, Ch. 5.4] [LT05, Kap. 5.4] [Bra07, Ch. II, 7]
48
7.2 The discrete maximum principle 49
7.2. The discrete maximum principle

Lemma 7.3. Let Th be a triangulation of Ω ⊂ R2 andT ∈ Th with the nodes x1 , x2 , x3
and corresponding linear nite element basis functions φ1 , φ2 , φ3 . Then
1
Z
∇φi · ∇φj = − cot(αk ),
T 2
for [i j k] a cyclic permutation of [1 2 3] and αk the angle in xk .
Proof. Denote by ν i the outer unit normal vector of the edge opposite to xi and by hi the
distance between xi and the opposing edge. Then we have
1 i
∇φi = − ν,
hi
and thus, for A12 ,
1
Z
∇φ1 · ∇φ2 = ν 1 · ν 2 |T |
T h1 h2
1
=− cos(α3 )|T |.
h1 h2
We also have
2|T | = |(x1 − x3 ) × (x2 − x3 )|

= |x1 − x3 | |x2 − x3 | sin(α3 ),
or
2|T | = |x2 − x3 |h1 = |x1 − x3 |h2 .
Multiplying both options yields
4|T |2 = |x1 − x3 | |x2 − x3 |h1 h2 ,
which results in
h1 h2
|T | =
2 sin(α3 )
and thus the claim is proved by cyclic permutation.
Lemma 7.4. Let xi , xj

be vertices of the triangulation Th , that are endpoints of the edge
0
E . Let T, T be the triangles, that share E . The angles opposite to E are called αE T and
T 0
αE . Then, for corresponding linear nite element basis functions φi , φj it holds
T + αT ) 0
1 sin(αE
Z
E
Aij = ∇φi · ∇φj = − T T0)
.
T ∪T 0 2 sin(αE ) sin(αE
Proof. The previous lemma stated
1 T T0

Aij = − cot(αE ) + cot(αE ) .
2
Now use the addition theorem for sin.
49
50 7 Special topics
Theorem 7.5 (Discrete maximum principle) . Let Th be a triangulation where for each
inner edge E holds
T T 0
αE + αE < π,
using the notation form before, and for E ⊂ ∂Ω

T π
αE < .
2
Then the linear nite element stiness matrix ~
A (for −∆) is an M-Matrix [Dör21b,
Def. 2.16]. Especially, for the solution uh of the Poisson problem with right hand side f
and f ≤0 we get
max uh (x) ≤ max uh (x).

x∈Ω x∈∂Ω
Proof. The method of proof is as in [Dör21b, Lem. 2.18]. We show that ~

A is weakly
diagonal dominant with Aii > 0 and Aij ≤ 0 for i 6= j . It is clear, that the diagonal is
strictly positive. For i 6= j we have by assumption
T T 0
Aij ∼ − sin(αE + αE ) < 0.
For E at the boundary we get
T
Aij ∼ − cot(αE )<0
when
T < π/2.
αE If xi ∈ N h , we get for the sum over the i-th row
X Z X
Aij = ∇φi · ∇ φj = 0.
j Ω j
| {z }
≡1
For interior points xi with neighbour xj on the boundary we have

X X
Aij = − Aij > 0.
j interior j boundary
Thus ~
A is weakly diagonal dominant [Dör21b, Def. 2.16]. Irreducibility follows, since
Aij 6 0
= if xi and xj are neighbours. Thus ~
A is inverse monotone, see the above cited
Lemma.
0
T + αT ≤ π T ≤ π
Note: The condition on the angles can often be relaxed to αE E and αE 2
if the matrix stays irreducible. For example the rst three meshes in Figure 1 would also
provide an M-Matrix. Such meshes are called to be of weakly acute type.
Remark 7.6.
∞
(1) One can now show L -stability as for Finite Dierence Methods [Dör21b, Sect. 2.4.3].
∞
Optimal L -estimates cannot in general be derived from this, since the stencils that
result from Finite Element discretisations are in general not consistent to optimal
order.
(2) Deriving L∞ -error bounds for the Finite Element Method is complicated. The idea is
0
to solve for x ∈ Ω the Poisson problem for the right hand side δx0 (Dirac-function)
and homogeneous Dirichlet boundary condition. In the weak form the problem is to
1
nd w ∈ H0 (Ω) such that
a(w, z) = δx0 (z) for all z ∈ H01 (Ω).
50
7.3 Error estimate for a nonlinear problem 51
Then for
e := u − uh
we nd
e(x0 ) = δx0 (e) = a(e, w).
Now one would like to proceed with the AubinNitsche trick (Section 7.1). However,
/ H01 (Ω)0 . As a remedy one uses the regularized
this does not work since δx0 ∈
Dirac measure: take η > 0 and dene
Z
δη,x0 (z) := − z.
Bη (x0 )
Literature for Section 7.2. [BS02, Ch. 8.18.4]
7.3. Error estimate for a nonlinear problem

7.3.1. Analysis
0,1
Letf ∈ Cloc (R). We want to nd u ∈ V := H01 (Ω), with Ω ⊂ Rd bounded by a polygonal
2 12
R
boundary and ||v||V := (
Ω |∇v| ) , such that
−∆u = f (u) in Ω,
u≡0 on ∂Ω.
The weak form reads
Z Z
∇u · ∇v = f (u)v for all v ∈ V.
Ω Ω
The question however is, whether
Z
Fw [v] = f (w)v
Ω
is, for arbitrary w ∈V, an element of the dual space V 0 = H−1 (Ω) := H01 (Ω)0 .
The one-dimensional case. Let w ∈ BR (0) ⊂ V for some R > 0. Then by a Sobolev
embedding [Dör15, Sect. 1.3.1 (Example)]
||w||L∞ (Ω) ≤ C||w||V ≤ CR =: K
and thus for all v∈V

1
Fw [v] ≤ sup |f (z)||Ω| 2 ||v||L2 (Ω) ,
|z|≤K | {z }
≤CP ||v||V
1
hence ||Fw ||H−1 (Ω) ≤ CP |Ω| 2 ||f ||L∞ (−K,K) .
51
52 7 Special topics
The higher dimensional case. Now it holds V 6⊂ L∞ (Ω). We need a bound for the
nonlinear term, so we assume that f fullls a growth condition
|f (z)| ≤ C |z|p + 1 ,

(7.1)
for suitable p ≥ 1. Then we get by Hölder's inequality (q ∈ [1, ∞), 1/q 0 = 1 − 1/q )
Z Z 0
1/q0 Z 1/q 1
Fw [v] ≤ C (|w|p + 1)|v| ≤ C |w|pq |v|q + |Ω| 2 ||v||L2 (Ω) .
Ω Ω Ω
By Sobolev's embedding theorem [Dör15, Sect. 1.3.1] we have the bound
||v||Lq (Ω) ≤ C||v||V
for arbitrary q ∈ [1, ∞) for d=2 and
2d
q≤
d−2
for d > 2. In case d = 2 we can verify for arbitrary p, q ∈ [1, ∞) that Fw is bounded in H−1
0
for w in a bounded set in V . In case d = 3 we take q = 6 and thus q = 6/5. Therefore,
with p = 5 we get
1
Fw [v] ≤ C ||w||5L6 (Ω) ||v||L6 (Ω) + |Ω| 2 ||v||L2 (Ω) ≤ C(||w||V )||v||V .

Note that we can exploit (7.1) for p=5 whenever it holds for p ≤ 5.
The second condition on f is the monotonicity, i.e., the one-sided bound
f (z1 ) − f (z2 ) (z1 − z2 ) ≤ λ|z1 − z2 |2

for all z1 , z2 ∈ R (7.2)
for some λ < CP−2 (compare [Dör21b, Thm. 1.7]). Especially for z = z1 and z2 = 0 we have
f (z) − f (0) z ≤ λ|z|2 .

Now we dene
A : V → V 0,
Z

A(w)[v] := ∇w · ∇v − f (w)v .
Ω
So we seek u∈V such that
A(u)[v] = 0 for all v∈V
or
A(u) = 0 in V 0.
The problem of unique existence is treated with the theory of monotone operators (non-
linear LaxMilgram) [R·º04, Ch. 3]. The main point to show existence is the coercivity
of A, stated as
A(v)[v] → ∞ for ||v||V → ∞.
52
7.3 Error estimate for a nonlinear problem 53
This is veried as follows

Z
|∇v|2 − f (v)v

A(v)[v] =
ZΩ
|∇v|2 − f (v) − f (0) (v − 0) − f (0)v

=
Ω
(7.2) 1
≥ ||v||2V − λ ||v||2L2 (Ω) −|f (0)||Ω| 2 ||v||L2 (Ω)
| {z } | {z }
2 ||v||2
≤CP ≤CP ||v||V
V
1

≥ (1 − λCP2 )||v||V − |f (0)|CP |Ω| 2 ||v||V → ∞ for ||v||V → ∞.
The last step needs λ < CP−2 , otherwise C0 := 1 − λCP2 would be negative. Existence of a
solution follows by Brouwer's xed point theorem [R·º04, Ch. 1.2]. As a byresult one gets
the following a priori bound for any solution
CP 1
||u||V ≤ R0 := |Ω| 2 |f (0)|,
C0
since we found A(v)[v] > 0 for ||v||V > R0 .
To get uniqueness, we again use the monotonicity. Let u1 , u2 be two solutions of A(u) =
0, then we get with testfunction w = u1 − u2
Z
|∇(u1 − u2 )|2 − (f (u1 ) − f (u2 ))(u1 − u2 )

0 = A(u1 )[u1 − u2 ] − A(u2 )[u1 − u2 ] =
Ω
≥ ||u1 − u2 ||2V − λ||u1 − u2 ||2L2 (Ω) ≥ (1 − λCP2 )||u1 − u2 ||2V = C0 ||u1 − u2 ||2V .
Since C0 > 0 by assumption, we conclude u1 = u2 .

The required monotonicity is necessary, since we may consider for g ∈ L2 (Ω)
−∆u = λ1 u + g in Ω,
u=0 on ∂Ω.
Then λ1 = CP−2 is the rst eigenvalue of −∆ on Ω (with homogeneous boundary condi-

tions), which will lead to non-uniqueness (g = 0) or non-solvability (unless g is orthogonal
to the rst eigenspace).
7.3.2. Numerical Analysis
We formulate the discrete problem as

Z Z
∇uh · ∇vh = fh (uh )vh for all vh ∈ Vh ⊂ V.
Ω Ω
Here we choose Vh := S01,0 (Th ). Existence of a discrete solution is proved as before.

For the following we will assume that f induces a Lipschitz continuous mapping on V.
This means that there exists a constant C > 0 such that for all v, w ∈ V
||f (w) − f (v)||L2 (Ω) ≤ C||w − v||V .
This is fullled under the assumption f ∈ C 0,1 (R), e.g., |f 0 (z)| ≤ L for all z ∈ R, because
||f (w) − f (v)||L2 (Ω) ≤ L||w − v||L2 (Ω) ≤ LCP ||w − v||V . (7.3)
Although this looks quite restrictive, it is applicable in many cases: if u turns out to be
bounded, say |u| ≤ K , then it suces to use the Lipschitz bound for f on [−K, K].
53
54 7 Special topics
Now let u and uh be the solutions of the continuous, resp., discrete equation and let e :=
u − uh . Subtracting both equations (for testfunction vh ) yields the Galerkin orthogonality
Z n
o
∇e · ∇vh − f (u) − fh (uh ) vh = 0 for all vh ∈ Vh .
Ω
Then
Z Z Z
2
|∇e| = ∇e · ∇(u − uh ) = ∇e · ∇(u − vh + vh − uh )
Ω ZΩ Ω
n o
= ∇e · ∇(u − vh ) + f (u) − fh (uh ) (vh − uh )
ZΩ n

= ∇e · ∇(u − vh ) + f (u) − fh (uh ) (vh − u)
Ω
o
+ f (u) − f (uh ) (u − uh ) + f (uh ) − fh (uh ) (u − uh )
and thus
Z n o
|∇e|2 − f (u) − f (uh ) (u − uh )

Ω Z Z n

= ∇e · ∇(u − vh ) + − f (u) − f (uh ) (u − vh )
Ω Ω
o
− f (uh ) − fh (uh ) (u − vh ) + f (uh ) − fh (uh ) (u − uh ) .
Similar to the coercivity proof in the last paragraph the left hand side can estimated from
below as
Z n o
uh ||2V |∇e|2 − f (u) − f (uh ) (u − uh ) .

C0 ||u − ≤
Ω
5
Assuming ||f (uh ) − fh (uh )||L2 (Ω) ≤ δ and with the Lipschitz bound (7.3) we arrive at
C0 ||e||2V ≤ ||e||V ||u − vh ||V + L||e||L2 (Ω) ||u − vh ||L2 (Ω)

+ ||f (uh ) − fh (uh )||L2 (Ω) ||u − vh ||L2 (Ω) + ||e||L2 (Ω)

≤ ||u − vh ||V + LCP2 ||u − vh ||V + δCP ||e||V + δCP ||u − vh ||V
1 1 2
≤ C0 ||e||2V + (1 + LCP2 )||u − vh ||V + δCP + δCP ||u − vh ||V .
2 2C0
We absorb the rst term on the right and get
1 2 CP
||e||2V ≤ 2 ||u − vh ||V + LCP2 ||u − vh ||V + δCP + 2δ ||u − vh ||V ,
C0 C0
which yields, bounding the last summand like (δCP )2 + 1/C02 ||u − vh ||2V ,
1
||e||V ≤ (2 + LCP2 )||u − vh ||V + δCP ,
C0
and nally

||e||V ≤ C inf ||u − vh ||V + δ .
vh ∈Vh
5
Note: For a, b, > 0 holds 2ab ≤ a2 + 1/ b2 .
54
7.4 Non-conforming elements 55
7.3.3. Discrete equations
The discrete problem is now
hZ n oi
~ u) = S~
G(~ ~ u − F~ (~u) = ∇uh · ∇φp − fh (uh )φp = 0.
Ω p∈Nh
This can be solved with Newton's method that converges quadratically if the initial guess
is close enough [DR20, Sect. 6.5]). We need the derivative of ~
G given by
hZ n oi
~ 0 ~ ~ ~ ~ 0 ~
G (~u)[d] = S d − F (~u)[d] = ∇dh · ∇φp − fh0 (uh )dh φp =0
Ω p∈Nh
and to solve in each step the linear equation
G ~ = −G(~
~ 0 (~u)[d] ~ u)
for d~. The Newton update then is ~u new = ~u + d~. A more robust version uses the update
~unew ~
= ~u +sd, where s is chosen in (0, 1] in every step (with some criteria) and s approaches
1 to guarantee quadratic convergence.
Note that computing F~ (~u) and F~ 0 (~u) needs repeated quadrature which is costly. A
simplication can be achieved if we use the approximations (in case of a nodal basis)
X
f (uh ) ≈ fh (uh ) = f (uh (p))φp
p∈Nh
because then we get
~ u) = S~
G(~ ~u − M ~ u − f~(~u),
~ [f (uh (p))]p∈N = S~
h
where ~
M is the mass matrix that we need to compute only once. Accordingly, the derivative
is also simplied to
~ 0 (~u) = S
~ −M
~ diag f 0 (u) .

G
Literature for Section 7.3. [HB02, Ch. 19] [DH02, Ch. 4.2] [BS02, Ch. 8.7] [R·º04,
Ch. 13]
7.4. Non-conforming elements

Before, we required linear piecewise polynomials to be continuous to have Vh ⊂ V . We
want to derive a result for one non-conforming case, the case that the elements are not
(globally) continuous anymore. Here we only consider the linear case. The idea is to relax
the continuity condition in a suitable way. Let E be an edge in the triangulation Th of Ω,
i.e., E ∈ Eh , and T1 , T2 the triangles with E = T1 ∩ T2 . Let nE be one xed normal to E.
Then we dene the jump of vh on E by
[vh ]E (x) := lim vh (x + snE ) − vh (x − snE ) .

s→0+
As an example we might dene the orientation of an edge as pointing from the vertex with
the lower number to the one with the higher number in our vertex list. The normal on E
might then point into the mathematically positive 90◦ rotated edge direction.
We now require as a replacement for continuity
Z
[vh ]E = 0. (7.4)
E
55
56 7 Special topics
6
The non-conforming nite element space is now dened by
S01,nc := vh ∈ L2 (Ω) : vh bT ∈ P1 for

n
all T ∈ Th , (7.4) holds for all inner edges,
Z o
otherwise vh = 0 .
E
We have (here comes the non-conformity)
(
C 0 (Ω)
S01,nc 6⊂ .
H01 (Ω)
Remark 7.7. Note that we have for any function wh that is linear on E
Z
wh = |E|wh (xE ),
E
where xE is the midpoint of the edge E. Thus, (7.4) implies that functions in S 1,nc are in
general only continuous in the centres of edges between adjacent triangles.
Without loss of generality we consider the Poisson problem with homogeneous Dirichlet
boundary conditions. First we dene the bilinear form. To this end let V := H01 (Ω) and
Vh := S01,nc . Then
X Z
Ah [v, w] := ∇v · ∇w for all v, w ∈ V + Vh ,
T ∈Th T
especially,
Ah [v, w] = A[v, w] for all v, w ∈ V.

We dene the broken norm on Vh + V by
p
||v||h := Ah [v, v].
This is indeed a norm on Vh : Assume ||vh ||h = 0. Then ∇vh bT = 0, which implies that
vh ≡ cT on T. The weak continuity (of Remark 7.7) suces to deduce vh ≡ c on Ω and
due to the boundary conditions we can conclude vh ≡ 0.
The discrete problem reads
Z
Ah [uh , vh ] = F [vh ] := f vh for all vh ∈ V h .
Ω
We now dene a basis of Vh and will then derive an error estimate with respect to || . ||h .
For this we dene the nodal sets
N h := x ∈ Ω : x = xE

for some edge E⊂Ω ,
Nh := N h ∩ Ω.
We get the nodal basis {φx }x∈N h uniquely by
φx (y) = δxy for all x, y ∈ N h ,

as usual. Then we can dene some linear and continuous interpolation operator by
Πh : V → Vh ,
X Z
Πh v := − v φxE for all v ∈ V.
E∈Eh E
Linearity is clear, the continuity on V is due to the trace theorem [Dör15, Sect. 1.3.2].
Note that Πh = Id on Vh .
6
Note that this denition is ambiguous for points at edges and vertices which is no harm since we dene
an L2 (Ω)-function. If needed, we will take limits for those points from adjacent triangles.
56
7.4 Non-conforming elements 57
Error estimate
Theorem 7.8. Let Ω ⊂ R2 be bounded and H2 -regular (Theorem 7.1). Let u ∈ H2 (Ω) ∩ V
be the solution of
A[u, v] = F [v] for all v ∈ V,
and uh the solution of
Ah [uh , vh ] = F [vh ] for all vh ∈ V h .
Then, assuming exact integration, it holds
||u − uh ||h ≤ Chmax ||u||H2 (Ω) .
The constant C depends on σ Th only.
Proof. The proof is similar to that of the Strang lemma (Theorem 2.6). For u we dene
the continuous linear form vh 7→ Ah (u, vh ) and solve
Ah (e
uh , vh ) = Ah (u, vh ) for all vh ∈ V h
for u
eh ∈ Vh . This is possible since Ah is coercive and continuous with respect to (Vh , || · ||h )
and thus
||u − u
eh ||h = inf ||u − vh ||h
vh ∈Vh
by denition of || · ||h . Therefore
Ah [e
uh − uh , vh ]
||u − uh ||h ≤ ||u − u
eh ||h + ||e
uh − uh ||h = inf ||u − vh ||h + sup
vh ∈Vh vh ∈Vh ||vh ||h
Ah [u − uh , vh ]
= inf ||u − vh ||h + sup .
vh ∈Vh vh ∈Vh ||vh ||h
Since the conforming linear nite element space is a subspace of Vh we have
inf ||u − vh ||h ≤ Chmax ||∇2 u||L2 (Ω) .

vh ∈Vh
For the second term we get, with some arbitrary numbers rE , sE ∈ R for E ∈ Eh (note,
7
that the trace of ∇u u ∈ H2 (Ω) [Dör15, Sect.
exists since 1.3.2])
X Z
Ah [u − uh , wh ] = ∇(u − uh ) · ∇wh
T ∈Th T
X nZ Z Z o
= ∂n uwh − ∆uwh − f wh
T ∈Th ∂T T T
X Z X Z
= ∂n u[wh ]E = ∂n u[wh − sE ]E
E∈Eh E E∈Eh E
(7.4) X Z
= (∂n u − rE )[wh − sE ]E
E∈Eh E
X
≤ ||∂n u − rE ||L2 (E) ||[wh − sE ]E ||L2 (E) .
E∈Eh
7
Alternatively, one can choose rE = −E ∂n u, sE = −E wh and use Lemma 1.6 later.
R R
57
58 7 Special topics
Let ΩE consist of the triangles that meet at E. Take the inma over rE ∈ R and sE ∈ R
and get by the trace theorem [Dör15, Sect. 1.3.2] and the BrambleHilbert lemma, Theorem
6.4,
1/2
inf ||∂n u − rE ||L2 (E) ≤ ChE ||∇2 u||L2 (ΩE ) ,
rE ∈R
1/2
inf ||wh − sE ||L2 (E) ≤ ChE ||wh ||h,ΩE .
sE ∈R
Thus
X
Ah [u − uh , wh ] ≤ C hE ||∇2 u||L2 (ΩE ) ||wh ||h,ΩE
E∈Eh
≤ Chmax ||∇2 u||L2 (Ω) ||wh ||h .
Remark 7.9 (Discontinuous Galerkin method). It is possible to do without any continuity

Q
condition and dene Vh T ∈Th PmT (T ) with dierent
to be a piecewise polynomial space
orders mT . The bilinear form and the norm will contain the jump-terms [ . ]E as above.
This method is called discontinuous Galerkin method [CDGH17].
Literature for Section 7.4. [BS02, Ch. 10.1, 10.3] [Bra07, Ch. III, 1]
7.5. The condition number of the stiness matrix

We dene the condition number of a regular matrix ~ ∈ RN ×N
A with respect to a
submultiplicative matrix norm || . || by
~ := ||A||||
κ(A) ~ A ~ −1 ||.
Recall that the condition number describes the sensitivity of the solution of a system of
equations with respect to data and appears in the convergence estimate of iterative methods
[DR20, Sect. 3.3.2, 3.4.1, 4.4.2, 4.4.4].
Let Ω be a polygonal domain in Rd , Th a decomposition into simplices and Vh ⊂ H01 (Ω)
the linear nite element space (for simplicity) with nodal basis {φi }i∈Nh . We want to know
the condition number ~

κ(A) for the reduced stiness matrix
Z
Aij = ∇φi · ∇φj for all i, j ∈ Nh .
Ω
Furthermore, we assume that we have an almost uniform triangulation, i.e., it holds
max hT ≤ C min hT
T ∈Th T ∈Th
for some constant that only depends on σTh . Then we dene the average meshsize h, the
vectornorm
X N 1/2
|~v |h := hd |vi |2
i=1
and the corresponding matrix norm
~ v |h
|A~
~ :=
||A|| sup .
h
~v ∈RN \{0} |~
v |h
58
7.5 The condition number of the stiness matrix 59
Lemma 7.10. For vh ∈ V h and its vector representation ~v holds
|~v |h ∼ ||vh ||L2 (Ω) ,
where ∼ means that there are estimates from both sides with constants independent of h.
Proof. Fix T ∈ Th and rescale it to Tb with hTb = 1. The local nite element space is a
three-dimensional space with the two equivalent norms
3
X 1/2 Z 1/2
2
|vi | ∼ v h |2
|b .
i=1 Tb
The constant will only depend on Tb. We scale back to T and obtain
3
X 1/2 1 Z 1/2
2
|vi | ∼ |vh |2 .
|T | T
i=1
Now we sum this over all triangles and since |T | ∼ hd , the denition of the norm gives the
desired estimates with constants depending only on σTh .
Recalling the inverse estimate (Lemma 4.6)
||∇vh ||L2 (Ω) ≤ Ch−1 ||vh ||L2 (Ω) ,
we nd
Z
w ~v =
~ · A~ ∇wh · ∇vh ≤ ||∇wh ||L2 (Ω) ||∇vh ||L2 (Ω)
Ω
≤ Ch−2 ||wh ||L2 (Ω) ||vh ||L2 (Ω) ≤ Ch−2 |w|
~ h |~v |h .
If we set w ~ v,
~ := A~ we get
|A~ ~ v · A~
~ v |2 ≤ hd A~ ~ v ≤ Chd−2 |A~
~ v |h |~v |h
h
which yields
~ v |h ≤ Chd−2 |~v |h .
|A~
This leads to the bound
~ ≤ Chd−2 .
||A|| h
It remains to show an estimate for the inverse of ~

A in terms of h.
We start with the Poincaré inequality (Lemma 1.4) on Vh ⊂ H01 (Ω) that states ||vh ||L2 (Ω) ≤
CP ||∇vh ||L2 (Ω) . By Lemma 7.10 this implies
~ v · ~v . h−d |A~
|~v |2h . ||vh ||2L2 (Ω) . ||∇vh ||2L2 (Ω) . A~ ~ v |h |~v |h
hence
~ v |h .
|~v |h ≤ Ch−d |A~
Taking especially ~ −1 w
~v = A ~ we conclude
~ −1 || ≤ Ch−d .
||A h
59
60 7 Special topics
Combining the two results gives
~ ≤ Ch−2 .
κ(A)
The condition number actually is of the order h−2 . Since ~

A is symmetric one has
~ = λmax (A)/λ
κ(A) ~ ~
min (A). For example, for the rst two meshes in Figure 1 and lexi-
cographic ordering of the unknowns, we have the same matrix ~
A (up to scaling) as for the
nite dierence method for which we have explicit formulas for the eigenvalues [Dör21b,
Lem. 2.18] that give ~ = 4/h2 .
κ(A) On more general grids we dene ~ v )1/2
|~v |h := (~v · M~ and
obtain ~ from the eigenvalues of the generalised eigenvalue
κ(A) problem A~ ~ v = λM~~ v.
~ is getting large if h gets small, it is necessary to use
Since κ(A) preconditioners ([DR20,
Sect. 2.4.3] [Dör21a, Sect. 2.1.4]). Using multilevel methods one can get essentially h-
independent convergence rates (up to log(h)-factors) [BS02, Ch. 67] [Bra07, Ch. VI].
Literature for Section 7.5. [BS02, Ch. 9.59.6]
7.6. Eigenvalue problems

Let Ω be an open and polygonally bounded domain that is H2 -regular (see Theorem 7.1).
The Laplace eigenvalue problem is to nd u ∈ H2 (Ω) ∩ H01 (Ω) \ {0} and λ ∈ R such that
−∆u = λu in Ω,
u=0 on ∂Ω.
In the weak formulation this can be written as
A[u, v] = λ(u, v)Ω for all v ∈ V := H01 (Ω)
with ( . , . )Ω denoting the L2 (Ω) scalar product. The ambiguity of u is resolved by the
normalisation ||u||L2 (Ω) = 1.
The theoretical result (see [Dör15, Sect. 2.6]) is that there exists a sequence of numbers
{λk }k∈N ⊂ R>0 such that
0 < λ1 < λ2 ≤ · · · ≤ λk ≤ . . . , λk → ∞ (k → ∞),
and corresponding eigenfunctions uk as required. The eigenfunctions are orthonormal

(uk , ul )Ω = δkl . Note that only nitely many λk , for k > 2, can be equal while the
eigenspace for λ1 is always one-dimensional.
After discretisation with the nite element method we obtain the generalised eigenvalue
problem
~ u = λM~
A~ ~u
on RN with stiness matrix ~

A and mass matrix ~ , both symmetric and positive matri-
M
ces. Thus we get a set of discrete eigenvalues {λh,k }k=1,...,N ⊂ R>0 and corresponding
eigenfunctions uh,k , resp., eigenvectors ~uk .
Our aim is to estimate the errors λh,k − λk and uh,k − uk .
Theorem 7.11 (Error estimate for nite element eigenvalues) . For all k≥1 there exists
a constant C such that for h := hmax ≤ (1/Cλk )1/2
0 ≤ λh,k − λk ≤ Ch2 λk .
60
7.6 Eigenvalue problems 61
Proof. [LT05, p. 90] We recall the minimax-characterisation
||∇v||2L2 (Ω)
λk = inf sup .
Wk ≤ H1
0 (Ω), v∈Wk \{0} ||v||2L2 (Ω)
dim(Wk ) = k
For λh,k this holds likewise with Wh,k ≤ RN . We now dene the Ritz projection
Ph : H01 (Ω) → Vh by
A[Ph u, vh ] = A[u, vh ] for all vh ∈ Vh .
Ph u is uniquely dened by Theorem 1.2 and it holds ||∇Ph u||L2 (Ω) ≤ ||∇u||L2 (Ω) and, with
Theorem 2.3, ||∇(u − Ph u)||L2 (Ω) ≤ Ch for u ∈ H2 (Ω).

With Ek := span{u1 , . . . , uk } and Eh,k := Ph Ek we obtain
||∇vh ||2L2 (Ω) ||∇Ph v||2L2 (Ω) ||∇v||2L2 (Ω)

λh,k ≤ max = max ≤ max .
vh ∈Eh,k \{0} ||vh ||2L2 (Ω) v∈Ek \{0} ||Ph v||2L2 (Ω) v∈Ek \{0} ||Ph v||2L2 (Ω)
For the denominator we get with the L2 -error bound and H2 -regularity, using Theorem
7.1,
||Ph v||L2 (Ω) ≥ ||v||L2 (Ω) − ||v − Ph v||L2 (Ω) ≥ ||v||L2 (Ω) − Ch2 ||∇2 v||L2 (Ω)
≥ ||v||L2 (Ω) − Ch2 ||∆v||L2 (Ω) ≥ ||v||L2 (Ω) − Ch2 λk ||v||L2 (Ω)
= (1 − Ch2 λk )||v||L2 (Ω) .
The last inequality is an inverse estimate on Ek which is easily obtained by writing v =

k
1 − Ch2 λk > 0 we
P
j=1 rj uj . For suciently small h, such that obtain
1 ||∇v||2L2 (Ω)
λh,k ≤ max ≤ (1 + Ch2 λk )λk .
1 − Ch2 λk v∈Ek \{0} ||v||2 2
L (Ω)
Theorem 7.12 (Error estimate for the eigenfunctions). Let u1 and uh,1 be the eigenfunc-
tions to the rst eigenvalues λ1 , λh,1 respectively. Then there is a constant C > 0 such
that
||u1 − uh,1 ||L2 (Ω) ≤ Ch2 ,

||∇(u1 − uh,1 )||L2 (Ω) ≤ Ch.
Proof. [LT05, p. 92] By expansion into the discrete eigenbasis we have
N
X
Ph u1 = aj uh,j
j=1
with coecients
aj = (Ph u1 , uh,j )Ω for j = 1, . . . , N.
From
λh,j aj = λh,j (Ph u1 , uh,j )Ω = (∇Ph u1 , ∇uh,j )Ω = (∇u1 , ∇uh,j )Ω = λ1 (u1 , uh,j )Ω
61
62 7 Special topics
we deduce

(λh,j − λ1 )aj = λ1 (u1 , uh,j )Ω − (Ph u1 , uh,j )Ω = λ1 (u1 − Ph u1 , uh,j )Ω .
For j≥2 we know by Theorem 7.11 that λh,j − λ1 ≥ λ2 − λ1 and hence we get the bound
N N
X X λ1 2
||Ph u1 − a1 uh,1 ||2L2 (Ω) = a2j ≤ (u1 − Ph u1 , uh,j )2Ω
λh,j − λ1
j=2 j=2
λ1 2
≤ ||u1 − Ph u1 ||2L2 (Ω) . h4
λ2 − λ1
and further
||u1 − uh,1 ||L2 (Ω) ≤ ||u1 − a1 uh,1 ||L2 (Ω) + ||(a1 − 1)uh,1 ||L2 (Ω)
≤ ||u1 − Ph u1 ||L2 (Ω) + ||Ph u1 − a1 u1 ||L2 (Ω) + ||(a1 − 1)uh,1 ||L2 (Ω)
≤ Ch2 + |1 − a1 |.
Finally we note that
|1 − a1 | = ||u1 ||L2 (Ω) − ||a1 uh,1 ||L2 (Ω)

≤ ||u1 − a1 uh,1 ||L2 (Ω) ≤ ||u1 − Ph uh,1 ||L2 (Ω) + ||Ph u1 − a1 uh,1 ||L2 (Ω) . h2 .
This yields the required L2 -bound. For the gradient we have
||∇(u1 − uh,1 )||2L2 (Ω) = ||∇u1 ||2L2 (Ω) + ||∇uh,1 ||2L2 (Ω) − 2(∇u1 , ∇uh,1 )Ω
= λ1 + λ1,h − 2λ1 (u1 , uh,1 )Ω
= λ1 + λ1,h + λ1 ||u1 − uh,1 ||2L2 (Ω) − 2λ1
= λ1,h − λ1 + λ1 ||u1 − uh,1 ||2L2 (Ω) . h2 .
Remark 7.13.
• The results are also valid if we consider the more general elliptic problem from Section
1.2.
• For increasing eigenvalues λk , the solution will be more and more oscillating. A
condition that h needs to be small in terms of λk has to be expected.
Literature for Section 7.6. [LT05, Kap. 6]
7.7. Adaptive Finite Element Method

7.7.1. The residual error estimate
Consider a domain Ω ⊂ R2 with a corner of angle α ∈ (π, 2π) (e.g., see [DR20, Fig. 4.2]).
In general, we have the following error estimate for linear nite elements for the Poisson
problem on a domain with such a single corner
||u − uh ||E ≤ Chγ ,

π
with γ := α . Note that γ ∈ (1/2, 1) in this situation since u may not be in H2 (Ω). In the
extreme case of a slitted disk we get γ ≈ 1/2. In the case of Remark 1.5 (6), a problem
with discontinuous coecient a, we would even nd γ = 0.1, so that we would need
h . 10−10
62
7.7 Adaptive Finite Element Method 63
to get a reasonable solution. On uniform grids this is not feasible, so we try to do better
with non-uniform grids. It has been shown, that the construction of such grids can be
done automatically, whereby the grid is quasi-optimal and the relation between the error
and the number of unknowns N is as in the fully regular case. We note, that the relation
err ∼h turns, due to N ∼ h−d (in Rd ) into err ∼ N −1/d . In terms of a given tolerance Tol
−d
for the error this means we need N ∼ Tol unknowns, but for the reduced regularity we
need N ∼ Tol−d/γ !
We consider as a model problem the Poisson problem: Let Ω ⊂ Rd be polygonally
bounded, Th a triangulation with Ω = Ωh . Seek u ∈ V := H01 (Ω) such that
Z Z
∇u · ∇v = fv for all v ∈ V.
Ω Ω
If we use nite elements of order p, we get the a priori error estimate
||∇(u − uh )||L2 (Ω) ≤ Chpmax ||∇p+1 u||L2 (Ω) .
This estimate is certainly of theoretical interest, however, the practical problem is that we
have no access to the values C and ||∇p+1 u||L2 (Ω) in order to get an explicit error bound.
Moreover, as we have seen, u may even not be in H2 (Ω). For this we will establish an a
posteriori error estimate, i.e., an error bound that will not depend on u but on the actual
numerical solution uh .
To derive this result, we exploit the weak formulations, where in the discrete case we
assume f is approximated by a piecewise polynomial fh . Thereby we get, for the error
e := u − uh and some xed v ∈ V ,
Z Z

∇e · ∇v = f v − ∇uh · ∇(v − vh ) − fh vh
Ω ZΩ

= − ∇uh · ∇(v − vh ) + (f − fh )v + fh (v − vh )
ZΩ
X X Z
= (fh + ∆uh )(v − vh ) − [∂ u ] (v − vh ) (7.5)
T | {z } ◦ E
| n{zh E}
T ∈Th =:rT E∈Eh
=:RE (Sect. 7.4)
Z
+ (f − fh )v
Ω
X X
≤ ||rT ||L2 (T ) ||v − vh ||L2 (T ) + ||RE ||L2 (E) ||v − vh ||L2 (E) (7.6)
T ∈Th E∈Eh◦
+ ||f − fh ||L2 (Ω) CP ||∇v||L2 (Ω)
and choosing the Clément interpolant vh := Πh v (Section 6.4), we continue with
1/2
X X
≤ ||rT ||L2 (T ) ChT ||∇v||L2 (ST ) + ||RE ||L2 (E) ChE ||∇v||L2 (SE )
T ∈Th E∈Eh◦
+ CP ||f − fh ||L2 (Ω) ||∇v||L2 (Ω)

X 1/2 X 1/2
≤ C h2T ||rT ||2L2 (T ) +C hE ||RE ||2L2 (E)
T ∈Th E∈Eh◦

+ CP ||f − fh ||L2 (Ω) ||∇v||L2 (Ω) .
63
64 7 Special topics
This yields the upper a posteriori error estimate

1
Z
||∇e||L2 (Ω) ≤ sup ∇e · ∇v
v∈V \{0} ||∇v||L2 (Ω) Ω
X n X o1/2
2 2 2
≤C hT ||rT ||L2 (T ) + hE ||RE ||L2 (E)
T ∈Th E⊂∂T
+ CP ||f − fh ||L2 (Ω)
(with RE := 0 for Eh∂ ), since the right hand side can assumed to be computable. The
constant C depends on properties of the mesh and can be estimated. CP is bounded by
the smallest diameter of Ω .
Remark 7.14.
• (7.5) is called error representation formula.
• RE is independent of the chosen orientation.
• C is only dependent on p and σTh .
Now we also want to have a lower bound. With the local barycentric coordinates we
dene an element bubble on T by
βT = λ1 λ2 λ2
and extend it by 0 onto Ω. We choose
v := βT rT and vh = 0
in the error representation formula (7.5) to get

Z Z Z
∇e · ∇(βT rT ) = (f − fh )βT rT + βT rT2 ,
T T T
since βT vanishes on the edges and outside of T. This gives, using inverse estimates,
Z Z
c rT2 ≤ βT rT2
T T
≤ ||f − fh ||L2 (T ) ||βT ||L∞ (T ) ||rT ||L2 (T ) + ||∇e||L2 (T ) ||∇(βT rT )||L2 (T )
| {z } | {z }
≤1 ≤Ch−1
T ||βT rT ||L2 (T )
C
≤ ||f − fh ||L2 (T ) + ||∇e||L2 (T ) ||rT ||L2 (T ) .
hT
Therefore

hT ||rT ||L2 (T ) ≤ C ||∇e||L2 (T ) + hT ||f − fh ||L2 (T ) .
Now take
v := θE RE and vh = 0
in the error representation formula (7.5), where RE is extended on Ω constantly on lines

perpendicular to E and θE is the quadratic edge bubble for E with supp(θE ) = ΩE (Section
4.2). We then get
Z Z X Z Z
2
∇e · ∇(θE RE ) = (f − fh )θE RE + rT θE RE + θE R E
ΩE ΩE T ⊂ΩE T E
64
and derive from this the bound, again using inverse estimates,
Z Z
2 2
c RE ≤ θE RE
E E
C X 1/2
≤ ||∇e||L2 (ΩE ) ||RE ||L2 (ΩE ) + ||rT ||2L2 (T ) ||RE ||L2 (ΩE )
hE
T ⊂ΩE
+ ||f − fh ||L2 (ΩE ) ||RE ||L2 (ΩE ) .
We now observe that
1/2
||RE ||L2 (ΩE ) ≤ ChE ||RE ||L2 (E)
and get
X 1/2
1/2
hE ||RE ||L2 (E) ≤ C||∇e||L2 (ΩE ) + hE ||rT ||L2 (ΩE ) + C h2T ||rT ||2L2 (T )
T ⊂ΩE
X 1/2
+C h2T ||f − fh ||2L2 (T )
T ⊂ΩE
!
X 1/2
≤ C ||∇e||L2 (ΩE ) + h2T ||f − fh ||2L2 (T )
T ⊂ΩE

= C ||∇e||L2 (ΩE ) + ||h(f − fh )||L2 (ΩE ) ,
where we used the bound on ||rT ||L2 (ΩE ) , that was derived in the step before, and using h
dening a piecewise constant meshsize function with value hT on T.
Finally we get the lower a posteriori error estimate

1/2
X
hT ||rT ||L2 (T ) + hE ||RE ||L2 (E) ≤ C ||∇e||L2 (ST ) + ||h(f − fh )||L2 (ST ) ,
E⊂∂T \∂Ω
with ST from Theorem 6.8.
7.7.2. The adaptive algorithm
Based on the previous results we dene the local error estimator

X
ηT2 := c1 h2T ||rT ||2L2 (T ) + c2 hE ||RE ||2L2 (E) ,
E⊂∂T \∂Ω
for some constants c1 , c2 that depend on the mesh and the polynomial degree of the nite
element method. The residual error estimate is then given by
X
ηh2 := ηT2 .
T ∈Th
Then we have the bounds (by suitable choice of c1 , c2 )
||∇e||L2 (Ω) ≤ ηh + CP ||f − fh ||L2 (Ω)
and, for all T ∈ Th ,

ηT ≤ C ||∇e||L2 (ST ) + ||h(f − fh )||L2 (ST ) ,
65
66 7 Special topics
or, summing up over T ∈ Th ,

ηh ≤ C ||∇e||L2 (Ω) + ||h(f − fh )||L2 (Ω) .
Note, that the constant depends only on p and σTh . This result holds also on Ω ⊂ Rd when
T denotes simplices and E the corresponding faces.
We assume that we are given a coarse triangulation T0 (macro-triangulation) and a
tolerance Tol > 0. The following algorithm is intended to deliver us a triangulation Th
p,0
and a solution uh on Vh = S0 (Th ) with ||u − uh || 1 ≤ Tol. The decisive point is that
H (Ω) 0
one renes in each step the grid locally on behalf of the information provided by the error
indicators ηT .
Algorithm 1 Adaptive Algorithm

1: function [uk , Tk ] = AdaptiveAlgorithm(T0 , data, Tol)
2: for k = 0, 1, . . . do
3: Set up system of equations on Vk (Tk );
4: Solve foruk ∈ Vk ;
5: Estimate {ηT }T ∈T
k
(needs uk , datak ); compute ηk ;
6: STOP if ηk < Tol;
7: Mark a set of triangles Ak ⊂ Tk ;
8: Rene Tk + Ak → Tk+1 ;
There are several strategies to mark elements, for example:
(1) T ∈ Ak if and only if ηT ≥ 21 ηk , which means, that the local estimated error on T has
half of the contribution to the global estimated error (maximum strategy ).
(2) For some θ ∈ (0, 1) choose Ak to be a minimal set of triangles such that
X
ηT2 ≥ θ2 ηk2
T ∈Ak
(xed energy fraction strategy ).
To rene triangles one has to obey two important issues: the new triangulations should
be admissible (to obtain a new conforming nite element space), the renement should stay
local if necessary (locality ), and the shape constant σ Th should stay bounded independently
on the number of subsequent renement steps (geometric stability ).
A method that meets these requirements is the newest vertex bisection strategy. 3 sub-
sequent bisections of a triangle T will yield an interior new node inside T on the rened
triangulation (i.e., the corresponding new basis function will have compact support in-
side T ), see Figure 9. This is called interior node property. A short description of
renement methods in R2 is given in Appendix A.
Theorem 7.15 (Convergence of the adaptive nite element algorithm). Let u∈V be the
solution of the Poisson problem and TH be an admissible triangulation. Let uH ∈ VH be the
solution of the discrete Poisson problem. We now perform one step of the adaptive algorithm
described in Algorithm 1 to get uh and Th using the xed energy fraction marking strategy
with some θ ∈ (0, 1), and a geometrically stable renement algorithm with the interior new
node property (e.g., 3 subsequent newest vertex bisections). Furthermore, assume that we
also have ||f − fh ||L2 (Ω) ≤ µηH ( data saturation) in each step (e.g., by further renement).
Hereby, µ is suciently small, only depending on p, σTh . Then there is a positive constant
c = c(p, σTh ), such that
p
||u − uh ||H1 (Ω) ≤ 1 − cθ2 ||u − uH ||H1 (Ω)
0 0
66
P1’ Pnew
P1 Pnew
Figure 9: Left: Bisection of interior triangles. Middle: Bisection of a boundary triangle.

Right: Renement with interior node property (3 bisections).
as long as ηH > Tol. Hence, the adaptive algorithm gives a solution satisfying the tolerance
in nitely many steps.
Proof. [Dör96] [MNS02].
Figure 10: Results of an adaptive loop in case of a corner domain. Left: Convergence
history on uniform meshes err ∼ N −1/3 and on adaptive meshes err ∼ N −1/2
for linear nite elements. Right: Resulting adaptive mesh.
Optimality. The question now is whether this algorithm provides us with the best possi-
ble, in view of number of unknowns, triangulation for the given problem. A given answer
is: if Bs is the class of functions in H1 (Ω) that can be best approximated of a set of
rened grids starting from T0 (e.g., by newest node bisections) of tolerance Tol with a
complexity O(N −1/s )
N → ∞), then this algorithm will produce a mesh with this
(for
s
complexity if the solution to the problem is in B (see [Ste07]). In this sense this algorithm
is asymptotically optimal.
For linear nite elements and d = 2 this means: The optimal algorithm provides us with
−2
a solution of tolerance Tol on a mesh with N = O(Tol ) vertices, independent of the
regularity of u beyond H1 !
7.7.3. hp nite element method
The method so far will work for mesh-renement (h-method) for xed polynomial degree.
If one is interested in variable polynomial degrees, one faces the problem to decide whether
to rene the mesh (for xed p) or increase the polynomial degree for xed h. The optimal
choice of h and p will lead to exponential (instead of algebraic) decay of the error with
the number of unknowns. For the error estimation one may also use the residual error
estimator, however, the explicit dependence of the error indicator from the polynomial
degree is important [Sch98].
67
68 8 Example of a nite element implementation
7.7.4. Adaptivity with quantities of interest
The previous considerations worked with the error in the energy norm, however, we might
as well ask for other norms, like the L2 -norm, or even an output value E[u] for some linear
continuous mapping E: H1 (Ω) → R, on the given solution u. Using the duality technique,
see Section 7.1 or Remark 7.6, it is possible to derive an error representation formula, that
allows to derive an upper error bound for |E[u − uh ]|, a posteriori up to higher order.
Literature for Section 7.7. [Sch98] [BS02, Ch. 9.19.4] [HR03] [MNS02] [LT05, Kap. 5.5]
[Bra07, Ch. III, 8] [Ver96] [AO00]
8. Example of a nite element implementation

8.1. Bases on the reference element
In this section we will introduce two types of bases for polynomial spaces, one that is useful
for evaluations and one that is useful in connection with integration.
8.1.1. One-dimensional case
As the reference element we take Σ1 := (−1, 1). For given polynomial degree p ≥ 0 the
reference nodes are −1 = ξ1 < ξ2 < · · · < ξp+1 = 1. The Lagrange basis of Pp , or
nodal basis, on this set of nodes is denoted by {φi }i=1,...,p+1 ⊂ Pp and uniquely dened by
φi (ξj ) := δij [DR20, Sect. 1.1]. Thus each uh ∈ Pp has the representation
Np
X
uh (ξ) = uh (ξj )φj (ξ),
j=1
where we let Np := p + 1.
We let {Pi }i∈N be a family of L2 -orthonormal polynomials on Σ1 , with deg(Pi ) = i − 1
so that also Pp = span{Pi : i = 1, . . . , Np }. This basis is called modal basis . Then there is
a uniquely dened set of coecients {αi }i=1,...,Np such that
Np Np
X X
uh (ξ) = uh (ξj )φj (ξ) = αj Pj (ξ).
j=1 j=1
Inserting ξ = ξi gives
Np Np
X X
uh (ξi ) = αj Pj (ξi ) =: Vij αj = (Vα)i .
j=1 j=1
The matrix V = [Pj (ξi )]i,j=1,...,Np is called Vandermonde matrix [DR20, Sect. 1.1]. Thus,
dening the nodal vector u = [u(ξ1 ), . . . , u(ξNp )] and the modal vector α = [α1 , . . . , αNp ]
we nd the connection
u = Vα ⇐⇒ α = V −1 u.
Because of
Np Np Np
Vij† φj (ξ)
X X X
Pi (ξ) = Pi (ξj )φj (ξ) = Vji φj (ξ) = ⇐⇒ P (ξ) = V † Φ(ξ),
j=1 j=1 j=1
we have the reciprocal relation
Np
Vij−† Pj (ξ)
X
φi (ξ) = ⇐⇒ Φ(ξ) = V −† P (ξ).
j=1
68
8.2 Basic matrices on the reference cell 69
8.1.2. Two-dimensional case
Let Σ2 be the reference simplex {(ξ1 , ξ2 ) : 0 ≤ ξ1 , ξ2 , ξ1 + ξ2 ≤ 1} or {(ξ1 , ξ2 ) : −1 ≤

ξ1 , ξ2 , ξ1 + ξ2 ≤ 0}. For p ∈ N we let {ξ i }i=1,...,Np , with Np := 1/2 (p + 1)(p + 2), be an
unisolvent pointset in Σ2 and {φi }i=1,...,Np the corresponding nodal basis, see Figure 11.left.
Further let {Pi }i=1,...,Np be a basis of Pp = Pp (Σ2 ), the set of polynomials in two variables
of order p. Note that dim(Pp ) = Np . With the Vandermonde matrix V = V

2D
= [Pj (ξ i )]ij
we can relate the nodal coecients to the modal coecients precisely as in Section 8.1.1.
Remark 8.1. The choice of the local mesh and the basis in Pp are crucial, since we know
that monomials on a uniform mesh lead to exponentially increasing condition numbers for
the Vandermonde matrix in p [DR20, Sect. 1.1].
In [HW08] the following settings were chosen: In one space dimension the polynomial
are the Legendre polynomials and the local grid points are the GauÿLobatto points [HW08,
Sect. 3.1].
In two space dimensions the polynomial are Jacobi polynomials and the local grid points
result from an optimisation approach that gives moderate condition numbers up to p = 15
[HW08, Sect. 6.1, Fig. 6.6].
8.2. Basic matrices on the reference cell

8.2.1. One-dimensional case
Constant coecients. The mass matrix on the reference cell is given by
Z p+1 Z p+1
−† −† −† −1
X X
Mij = φj φi = Vik Vjm Pk Pm = Vik Vkj = (VV † )−1
ij
Σ1 k,m=1 Σ1 k=1
or
M = (VV † )−1 .
Now let
Z p+1 Z
X p+1
X p+1
X
Bij = φ0j φi = φ0j (ξk )φk φi = φ0j (ξk )Mik =: Mik Dkj = (MD)ij ,
Σ1 k=1 Σ1 k=1 k=1
with the dierentiation matrix elements Dkj = φ0j (ξk ). From this we obtain for the stiness
matrix
Z p+1 Z
X p+1
X
Sij = φ0j φ0i = φ0i (ξk )φk φ0j = Dki Bkj = (D † MD)ij
Σ1 k=1 Σ1 k=1
or
S = D † MD.
To compute D , we can use Φ0 (ξ) = V −† P 0 (ξ) to get D † = V −† D †P or D = D P V −1 , where

the
0
coecients of D P are given by (D P )kj = Pj (ξk ). This yields
S = D † MD = V −† D †P MD P V −1 = V −† D †P V −† V −1 D P V −1
= (V −1 D P V −1 )† V −1 D P V −1 .
Thus we can compute M and S from the basic data [Pj (ξk )]j,k , [Pj0 (ξk )]j,k that can be
computed beforehand.
69
70 8 Example of a nite element implementation
Variable coecients. In case of variable coecients in the mass and stiness matrix we
use local quadrature formulas for the resulting matrix M.
8.2.2. Two-dimensional case
We obtain as in the one-dimensional case M = (VV † )−1 , but now with V = V 2D . Then
we dene the two dierentiation matrices
D 1 = ∂1 φj (ξ i ) ij , D 2 = ∂2 φj (ξ i ) ij .

These matrices can be obtained from the dierentiation matrices of the polynomials by
D 1 = D P ,1 V −1 and D 2 = D P ,2 V −1 . For the stiness matrix we now easily nd
S = D †1 MD 1 + D †2 MD 2 .
In case of variable coecients we proceed as before and use local quadrature formulas for
the resulting matrix M.
8.3. The global nite element discretisation

The local discretisation will follow from the discretisation on the reference cell via the
mappings F T : Σd → T for T ∈ Th (Section 4.4). In the following we will assume that the
mappings F T are ane, i.e., we assume Ωh = Ω .
Figure 11: Left: Reference simplex with local numbering (p = 2). Right: Global (non-
correlated) numbering in a mesh (p = 2). The nodes 1,4,6 and 7,8,9 may have
dierent values but share the same coordinates.
8.3.1. Local space and matrices
On the reference cell we dene a xed numbering of the nodes for a given polynomial degree
p (Figure 11.left). For T ∈ Th we let i, j be the global node numbers of nodes located in
T . Note that here the numbering is cell-wise and not correlated to neighbouring cells, that
is, the given nodes i, j exist precisely in one cell T (Figure 11.right) and scalar products
between basis functions are only performed on T . The nodes i, j correspond to local node
numbers i0 , j0 in the reference cell Σd , d ∈ {1, 2}, via the mapping F T .
The aim of this section is to show how we can compute mass and stiness matrix (for
constant coecients) using those on the reference element.
For the mass matrix on T we have simply
Z Z
Mij = φj φi = JT φbj0 φbi0 .
T Σd
70
8.3 The global nite element discretisation 71
Here, with (.)

c we denote quantities on the reference cell and JT = det(F 0T ) = |T |/|Σd |.
DG
Thus for the uncorrelated mass matrix MT on T we nd
MDG
T = JT M.
c
We use the superscript 'DG' since this type of matrices appears in the Discontinuous
Galerkin Method, where the local polynomial spaces are independent of each other.
To integrate for the right hand side we use
Z Np
Z X Np
X
Mij f (xj ) = MDG

Fi = f φi ≈ f (xj )φj (x)φi (x) dx = T [f (xj )]j i .
T T j=1 j=1
For the stiness matrix we get after applying the transformation

Z Z
Sij = ∇φj · ∇φi = JT (F 0T )−† ∇φbj0 · (F 0T )−† ∇φbi0 .
T Σd
For d=1 this simply gives
1 b
S DG
T = S.
JT
For d=2 we get, for some coecients βT,kl , k, l ∈ {1, 2}, the representation
2 Z 2
b †M
X X
Sij = βT,kl ∂k φbi0 ∂l φbj0 = βT,kl (D k
cD
b l )i ,j ,
0 0
k,l=1 Σd k,l=1
or
2
b †M
X
S DG
T = βT,kl D k
cD
b l.
k,l=1
Again, in case of variable coecients we use local quadrature formulas for the resulting
matrix M.
8.3.2. Global space and matrices
The DG-space consists of Nc cells with Np degrees of freedom in each. Note that nodes at
inner cell borders share the same evaluation point in space. The previous section has now
led to an equation
ADG uDG = F DG ,
which is a block-diagonal system of equations that can, due to the lacking interactions
between the cells, not dene an approximation to the solution of our PDE. In the DG-
context, one includes additional coupling terms by integrating the jumps over the cell
boundaries that are penalised to enforce the coupling [CDGH17]. To switch to the nite
element context of H1 -conforming spaces (that requires continuity on cell boundaries), we
will identify all nodes with a common point and therefore introduce the embedding matrix
E c : VhCG → VhDG that simply distributes the value in a FEM-node to all its identications
in the DG-space. Then we make the ansatz uDG = E c uCG and multiply the resulting
†
system with E c to get
ACG uCG := E †c ADG E c uCG = E †c F DG =: F CG .
71
72 9 Saddle Point Problems
ACG is the nite element system matrix which uses exact integration for constant coe-
cients and uses a special rule for variable coecients. However, this treatment ts to the
error estimates of Section 3.2.3. For the right hand side we may collect a vector of function
evaluations in the FEM-nodes f CG and let f DG := E c f CG . Thus we have
F CG = E †c F DG = E †c MDG f DG = E †c MDG E c f CG =: MCG f CG .

This leads to the equation
ACG uCG = MCG f CG

that is the discretisation of the partial dierential equation. If we have Dirichlet boundary
conditions, we have now to use the embedding E cb so that uCG = E cb uCG CG
red + ucb and
CG CG
derive the system to compute ured . If we want to compute ured from u
CG , we obtain
CG † CG CG †
ured = E cb (u − ucb ) since E cb E cb is the identity on the reduced space. See [Dör21a,
Sect. 2.1.3.2] for the Finite Dierence method.
For given uCG we get uDG through uDG = E c uCG . If, however, uDG is given, we
multiply this relation by E †c to get
E †c E c uCG = E †c uDG .
E †c E c is a diagonal matrix Dc , whose diagonal element counts the number of identied

points. So we get
uCG = D −1 † DG
c Ecu .
Note that D −1 † † −1 †
c E c = (E c E c ) E c is the pseudo-inverse to Ec [DR16, Sect. 2.3].
Literature for Section 8. [HW08, Ch. 3.13.2, 6.16.3]
9. Saddle Point Problems

9.1. The Poisson problem as a rst order system
In order to write the Poisson problem from Section 1.2 with homogeneous boundary con-
ditions as a system of rst order we rewrite −∆u = −∇ · ∇u = f into
σ − ∇u = 0 in Ω,
−∇ · σ = f in Ω,
with σ : Ω → Rd . We get a weak formulation by multiplying the equations with appropriate
test functions
Z
τ ∈ C ∞ (Ω)d ,

σ · τ − ∇u · τ = 0 for all
Ω Z Z
σ · ∇v = fv for all v ∈ C0∞ (Ω).
Ω Ω
Here we integrated the second equation by parts (using thereby that we want to solve
a Dirichlet problem) to avoid derivatives on σ, τ . On the other hand, we may avoid
derivatives on u, v and integrate by parts in a dierent way to get
Z
∞ d

σ · τ + u∇ · τ = 0 for all τ ∈ C (Ω) ,
Ω Z Z
∞
∇·σv = − fv for all v ∈ C (Ω).
Ω Ω
Here we used the required homogeneous Dirichlet condition for u. This type of formulation
is called mixed problem or mixed formulation .
72
9.2 Minimisation with constraint 73
9.2. Minimisation with constraint

Let V be a Hilbert space and a : V × V → R, F : V → R be as in Section 1.2. Then
1
E(v) := a[v, v] − F [v]
2
denes a strictly convex mapping E :V →R and it has therefore a unique minimiser u
that is uniquely dened by
a[u, v] = F [v] for all v ∈ V.
This is the Dirichlet principle from Section 1.4. We now assume that we have in addition a
Hilbert space M and a bilinear form b : V ×M → R. We seek a solution of the minimisation
problem for E , but with the constraint

u ∈ Z := v ∈ V : b[v, q] = 0 for all q ∈ M .
To characterise this minimum we introduce the Lagrange function
L:V ×M →R
1
L(v, q) := a[v, v] − F [v] + b[v, q]
2
and seek for its extremal points. Any such extremal point [u, p] ∈ V × M will satisfy
!
0 = ∂1 L(u, p)[v] = a[u, v] − F [v] + b[v, p],
!
0 = ∂2 L(u, p)[q] = b[u, q].
p is called Lagrange parameter . With a slight generalisation we aim to nd, for given F as
above and continuous G : M → R, a solution [u, p] ∈ V × M of the equations
a[u, v] + b[v, p] = F [v] for all v ∈ V, (9.1)
b[u, q] = G[q] for all q ∈ M. (9.2)
The Lagrange function is not denite even if a is non-negative and the characterisation of
the solution is
L(u, p) = sup inf L(v, q)

q∈M v∈V
or
L(u, q) ≤ L(u, p) ≤ L(v, p) for all v ∈ V, q ∈ M.
[u, p] is called a saddle point of L. A problem of the form (9.1)(9.2) is called saddle point
problem .
9.3. Examples
9.3.1. Poisson problem
A rst order equation for the Poisson problem has been introduced in Section 9.1. We
repeat the weak formulation, but now with the denition of appropriate function spaces.
The rst formulation will give: Seek [σ, u] ∈ V × M such that
Z
τ ∈ V := L2 (Ω)d ,

σ · τ − ∇u · τ = 0 for all (9.3)
Ω Z Z
σ · ∇v = fv for all v ∈ M := H01 (Ω). (9.4)
Ω Ω
73
The norms are given by ||τ ||V := ||τ ||L2 (Ω)d and ||v||M := ||∇v||L2 (Ω)d . We now identify the
forms
Z Z
a[σ, τ ] = σ · τ, b[τ , v] = − τ · ∇v.
Ω Ω
The second formulation will lead to the problem: Seek [σ, u] ∈ V × M such that
Z

σ · τ + u∇ · τ = 0 for all τ ∈ V := H(div, Ω), (9.5)
Ω Z Z
∇·σv = − fv for all v ∈ M := L2 (Ω). (9.6)
Ω Ω
Here we have dened
H(div, Ω) := τ ∈ L2 (Ω)d : ∇ · τ ∈ L2 (Ω)

with norms ||τ ||2V := ||τ ||2L2 (Ω)d + ||∇ · τ ||2L2 (Ω) and ||v||M := ||v||L2 (Ω) . We identify the forms
Z Z
a[σ, τ ] = σ · τ, b[τ , v] = ∇ · τ v.
Ω Ω
The rst formulation is called primal problem , while the second formulation is called
dual problem .
9.3.2. Stokes problem
For a stationary viscous ow (without turbulent eects) inside a domain Ω ⊂ Rd (d ≥ 2),
the uid velocity u:Ω→ Rd and the pressure p:Ω→R satisfy the system
−∆u + ∇p = f in Ω,
∇·u=0 in Ω,
called Stokes equation, where f : Ω → Rd is an external force density. Requiring homoge-

neous boundary values u = 0 on ∂Ω , we derive from this the variational equations
Z Z
1 d

∇u : ∇v − p∇ · v = f ·v for all v ∈ V := H0 (Ω) ,
Ω Z Ω
2
∇ · uq = 0 for all q ∈ M := L0 (Ω)
Ω
with ||v||V := ||∇v||L2 (Ω)d,d and ||q||M := ||q||L2 (Ω) . We identify the forms
Z Z
a[u, v] = ∇u : ∇v, b[v, q] = − ∇ · v q.
Ω Ω
9.3.3. Timeharmonic electric elds
From Maxwell's equations we derive in the stationary case the following equations for
the electric eld u : Ω ⊂ R3 → R3
1
∇× ∇×u = εr f in Ω,
µr
∇ · (εr u) = 0 in Ω,
and given data functions εr , µr : Ω → R>0 and f : Ω → R3 . This system is accomplished

with suitable boundary conditions. In the rst equation one can add a term εr ∇p on the
74
9.4 Operator notation 75
left to get a system of saddle point structure (but usually p turns out to be 0 afterwards).
Here we write
1
Z Z
∇×u · ∇×v + εr v · ∇p = εr f · v for all v ∈ V := H0 (curl, Ω),
Ω µr Z Ω
εr u · ∇q = 0 for all q ∈ M := H01 (Ω).

Ω
Here we dened
H(curl, Ω) := v ∈ L2 (Ω)3 : ∇×v ∈ L2 (Ω)3 ,

H0 (curl, Ω) := v ∈ H(curl, Ω) : n × v = 0 on ∂Ω
with norms ||v||2V := ||v||2L2 (Ω)3 + ||∇×v||2L2 (Ω)3 and ||q||M := ||q||L2 (Ω) and we identify the
forms
1
Z Z
a[u, v] = ∇×u · ∇×v, b[v, q] = εr v · ∇q.
Ω µr Ω
9.4. Operator notation

We dene operators A:V →V and B:V →M by
(Av, w)V := a[v, w] for all v, w ∈ V,

(Bv, q)M := b[v, q] for all v ∈ V, q ∈ M.
Note that Λa := ||A||IL(V,V ) and Λb := ||B||IL(V,M ) are bounded by the respective bounds of
the forms a, b. Using these operators the problem (9.1)(9.2) for u and p reads 8
Au + B † p = F in V, (9.7)
Bu = G in M. (9.8)
Here we used the adjoint B† : M → V of B, dened by
(Bv, q)M = (v, B † q)V for all v ∈ V, q ∈ M.
9.5. Existence of solutions for the saddle point problem

Theorem 9.1 (Existence for saddle point problems 1). Let Z := ker(B) and PZ : V → Z
the orthogonal projection with respect to ( . , . )V , PZ ⊥ = Id−PZ , i.e., V = Z ⊕Z ⊥ . Assume
in addition that
(i) PZ APZ : Z → Z is continuously invertible

(ii) B : Z ⊥ → Rg(B) = M is continuously invertible
with
1 1
λa := , λb := .
||(PZ APZ )−1 ||IL(V,V ) ||(B † )−1 ||IL(Z ⊥ ,M )
Then, there is a unique solution [u, p] ∈ V × M of the saddle point problem (9.7)(9.8) and
1 1 Λa
||u||V ≤ ||F ||V + 1+ ||G||M ,
λa λa λb
1 Λa
||p||M ≤ 1+ ||F ||V + Λb ||G||M .
λb λa
8
If V is a Hilbert space, then each functional F [] in V 0 can be identied with F ∈ V by F [v] = (F, v)V
for all v ∈ V .
75
Proof. Since B is continuous, Z is closed and therefore PZ is continuous. Writing
u = PZ u + (I − PZ )u = PZ u + PZ ⊥ u = uZ + uZ ⊥
and inserting this into (9.7)(9.8) yields
PZ A(uZ + uZ ⊥ ) = PZ F,
PZ ⊥ A(uZ + uZ ⊥ ) + PZ ⊥ B † p = PZ ⊥ F,
BPZ ⊥ uZ ⊥ = BuZ ⊥ = G.
Here we used PZ B † = (BPZ )† = 0 and BuZ = 0. We can write this as a system

    
PZ APZ PZ APZ ⊥ 0 uZ PZ F
 PZ ⊥ APZ PZ ⊥ APZ ⊥ PZ ⊥ B †   uZ ⊥  =  PZ ⊥ F 
0 BPZ ⊥ 0 p G
that we solve by elimination. By assumption (ii) we get from the third equation
uZ ⊥ = (BPZ ⊥ )−1 G ∈ Z ⊥ .
Using this in the rst equation yields, together with assumption (i),
uZ = (PZ APZ )−1 PZ (F − AuZ ⊥ ).
From assumption (ii) we also get that B † = (BPZ ⊥ )† = PZ ⊥ B † : M → Z ⊥ is invertible

and thus the second equation gives
p = (PZ ⊥ B † )−1 PZ ⊥ (F − Au) ∈ M.
The a priori estimates follow easily from these representations.
This rst version is useful due to its transparent proof. The second version is formulated
in terms of weak forms and is more useful for the practical application.
Theorem 9.2 (Existence for saddle point problems 2) . The saddle point problem (9.1)
(9.2) has a unique solution [u, p] ∈ V × M for all F ∈V0 andG ∈ M 0 , if there are positive
numbers α, β such that
A[v, v] b[v, q]
inf ≥ α, inf sup ≥ β. (9.9)
v∈Z\{0} ||v||2V q∈M \{0} v∈V \{0} ||v||V ||q||M
The estimates in Theorem 9.1 then hold with λa = α and λb = β . The second inequality in
(9.9) is called inf-sup condition or (Ladyshenskaja)BabushkaBrezzi condition, in short
(L)BB condition.
Proof. We use the assumptions to verify those of Theorem 9.1. The assertion on PZ APZ
is clear by the asserted lower bound for A. We have to show that the inf-sup condition
implies the assumption (ii) on B in Theorem 9.1. Assume that the inf-sup condition (9.9)
holds. We dene linear continuous mappings B : V → M and B † : M → V as in Section
9.4. Since BPZ = 0 we consider B : Z⊥ → M and by B † = (BPZ ⊥ )† = PZ ⊥ B † we see
B† : M → Z ⊥ . The inf-sup condition gives
b[v, p]
||B † p||V = sup ≥ β||p||M . (9.10)
v∈V \{0} ||v||V
We conclude that B † is injective. Since B † is continuous, RB := Rg(B † ) is closed in

Z⊥ ⊂ V. If RB Z ⊥ , then take v ∈ Z ⊥ ∩ R⊥ †
B \ {0} and get b[v, q] = (v, B q)V = 0 for
76
9.6 Discretisation of saddle point problems 77
allq ∈ M . This would prove v ∈ Z which contradictsv ∈ Z ⊥ \ {0}. Hence, RB = Z ⊥

†
and thus B : M → Z
⊥ is bijective. † −1 : Z ⊥ → M exists and is
As a consequence (B )
continuous. In the inequality (9.10)
† −1 v for arbitrary v ∈ Z ⊥ to
we can now let p = (B )
get
||v||V ≥ β||(B † )−1 v||M for all v ∈ Z⊥
and this shows
1
||(B † )−1 ||IL(Z ⊥ ,M ) ≤
β
The bound follows likewise for ||B −1 ||IL(M,Z ⊥ ) because ||B −1 ||IL(M,Z ⊥ ) = ||(B † )−1 ||IL(Z ⊥ ,M ) .
Example 9.3. We study the Stokes problem (Section 9.3.2) with homogeneous boundary
conditions
−∆u + ∇p = f in Ω,
∇·u=0 in Ω,
u=0 on ∂Ω.
Note that A V := H01 (Ω)d and thus on V h ⊂ V

is coercive on
R , b satises the inf-sup
2 2
condition (9.9) on V × M , where M := L0 (Ω) := {q ∈ L (Ω) : Ω q = 0}.
This is seen as follows: It can be proved that for all q ∈ M there exist a v q ∈ V such
that ∇ · v q = q in Ω and ||v q ||V ≤ C||q||M for some C > 0. We then use
|q|2
R R R
Ω ∇·vq Ω ∇ · vq q 1
sup ≥ = Ω ≥ ||q||M
v∈V \{0} ||v||V ||v q ||V ||v q ||V C
to prove the inf-sup condition.
9.6. Discretisation of saddle point problems

Now let Vh ⊂ V , Mh ⊂ M be nite dimensional approximations to V, M (conforming
ansatz spaces). The discrete problem now reads
a[uh , vh ] + b[vh , ph ] = F [vh ] for all vh ∈ Vh , (9.11)
b[uh , qh ] = G[qh ] for all qh ∈ Mh , (9.12)
or, in terms of nite dimensional operators,
Ah Bh†

uh fh
= .
Bh 0 ph gh
As in the continuous case we dene Zh := ker(Bh ).
Theorem 9.4 (Existence for discrete saddle point problems) . Under the conditions of
Theorem 9.2 (for the discrete setting Vh , Zh , Mh ), there exists a unique solution of the
discrete problem (9.11)(9.12) with modied λa , λb (but same Λa , Λb ).
Note: The proof is as for Theorem 9.2 and allows λa , λb to be h-dependent. Discrete
stability however requires h-independent bounds, but this does not automatically follow
from Theorem 9.2!
77
Example 9.5 (Invalid inf-sup-condition) . The crucial point for any choice of spaces Vh
and Mh is the inf-sup condition on V h × Mh . We recall Example 9.3. If we let V h =
S01,0 (Th )d and Mh = S 1,0 (Th ) ∩ L20 (Ω), then this pair does not satisfy this condition for
2 2
the Stokes problem! Indeed, for Ω = (0, 1) ⊂ R and Th being one of the two left-most
triangulations as in Figure 1 but with n = 3, we nd dim(V h ) = 2, dim(Mh ) = 9 − 1 = 8.
†
Thus, B h : Mh → V h cannot be injective and therefore there exists qh ∈ Mh \ {0} with
0 = B †h qh · v h = b[v h , qh ]
for allv h ∈ V h , and therefore the inf-sup condition cannot hold. Also, the choice Mh =
{vh ∈ L20 (Ω) : vh bT ∈ P0 for all T ∈ Th } does not work. A typical sign of instability is the
checkerboard instability [Bra07, Ch. III, 7].
Examples of stable discretisations are the TaylorHood element and the Mini element
[Bra07, Ch. III, 7].
9.7. General error estimate

Theorem 9.6 (A priori error estimate for the saddle point problem) . Assume that the
requirements of coercivity and continuity and the inf-sup conditions hold for the continuous
and discrete problem. Then, both the solutions u of (9.1)(9.2) and uh of (9.11)(9.12)
exist and
Λa Λb Λb
||u − uh ||V ≤ 1 + 1+ inf ||u − vh ||V + inf ||p − qh ||M ,
λa,h λb,h h h
v ∈V λa,h qh ∈Mh
Λb Λa
||p − ph ||M ≤ 1+ inf ||p − qh ||M + ||u − uh ||V .
λb,h qh ∈Mh λb,h
Proof. Here we let G = 0. Zh 6⊂ Z may hold9 , we
Since have to proceed as for the Strang
lemma (Theorem 2.6). Let u ∈ Z and uh ∈ Zh be the respective solutions and veh ∈ Zh
arbitrary. Then, using coercivity of a on Zh , we get
1 a[e
v h − uh , wh ]
||u − uh ||V ≤ ||u − veh ||V + sup
λa,h wh ∈Zh \{0} ||wh ||V
and inserting ±u in the last term leads to
Λa 1 a[u − uh , wh ]
||u − uh ||V ≤ 1 + ||u − veh ||V + sup .
λa,h λa,h wh ∈Zh \{0} ||wh ||V
For wh ∈ Z h we can write, for arbitrary q h ∈ Mh ,
a[u − uh , wh ] = a[u, wh ] − F [wh ] = −b[wh , p]
= −b[wh , p − qh ]
so that
a[u − uh , wh ] ≤ Λb ||p − qh ||M ||wh ||V

and therefore
Λa Λb
||u − uh ||V ≤ 1 + ||u − veh ||V + ||p − qh ||M .
λa,h λa,h
By taking the inmum over veh ∈ Zh and qh ∈ Mh we get
Λa Λb
||u − uh ||V ≤ 1 + inf ||u − veh ||V + inf ||p − qh ||M .
λa,h veh ∈Zh λa,h qh ∈Mh
We want to get rid of the restriction veh ∈ Zh 10 . For this let vh ∈ V h arbitrary and dene
9
We called the method conforming since Vh ⊂ V and Mh ⊂ M holds, however, one might call the method
non-conforming since Zh may not be a subspace of Z . For example, the condition Zh ⊂ Z would
require high polynomial degree ansatz spaces in case of Stokes' equation.
10
If a would be coercive on Vh , we were done here.
78
9.8 The Fortin operator 79
wh ∈ Zh⊥ by
b[wh , qh ] = b[u − vh , qh ] = b[PZ ⊥ (u − vh ), qh ] for all q h ∈ Mh .
wh is well-dened, since by assumption Bh† : Mh → Zh⊥ and thus Bh : Zh⊥ → Mh is

invertible and it holds
Λb
||wh ||V ≤ ||u − vh ||V
λb,h
and from
b[wh + vh , qh ] = b[u, qh ] = 0
we get wh + v h ∈ Z h and
Λb
||u − (vh + wh )||V ≤ ||u − vh ||V + ||wh ||V ≤ 1 + ||u − vh ||V .
λb,h
Thus
Λb
inf ||u − veh ||V ≤ ||u − (vh + wh )||V ≤ (1 + )||u − vh ||V .
eh ∈Zh
v λb,h
Now taking the inmum over vh ∈ V h gives the required estimate for u − uh . To prove the
bound for ||p − ph ||M , we take arbitrary vh ∈ Vh to get

b[vh , p − ph ] = F [vh ] − a[u, vh ] − F [vh ] − a[uh , vh ]
= −a[u − uh , vh ].
By the inf-sup condition (and inserting ±p)
b[vh , qh − ph ]
λb,h ||qh − ph ||M ≤ sup
vh ∈Vh \{0} ||vh ||V
− a[u − uh , vh ] + b[vh , qh − p]
≤ sup
vh ∈Vh \{0} ||vh ||V
≤ Λa ||u − uh ||V + Λb ||p − qh ||M .
Thus
Λb Λa
||p − ph ||M ≤ ||p − qh ||M + ||qh − ph ||M ≤ 1 + ||p − qh ||M + ||u − uh ||V
λb,h λb,h
which yields the required estimate.
The problem that we have to solve is: when does the inf-sup condition hold? The
following gives a quite general characterisation.
9.8. The Fortin operator

Theorem 9.7 (Fortin operator). Let b satisfy the inf-sup condition on V × M . Then, the
conforming discretisation on Vh × Mh satises the inf-sup condition if, and only if, there
exists an operator Πh : V → Vh such that
||Πh v||V ≤ C||v||V ,

and b[Πh v, qh ] = b[v, qh ] for all v ∈ V, qh ∈ Mh .
Πh is called Fortin operator.
79
Proof. ⇒ For all v, w we let a[v, w] = (v, w)V (the scalar product on V ). Let v be given
and dene F [w] := a[v, w] and G[q] := b[v, q] for all w ∈ V , q ∈ M . Clearly, a is continuous
and
0 0
coercive on V and F ∈ V , G ∈ M . We consider the problem
a[e
v , w] + b[w, pe] = F [w] for all w ∈ V,
b[e
v , q] = G[q] for all q∈M
dening ve ∈ V and pe ∈ M . This problem is well-posed by Theorem 9.2 and thus the only
solution is [e
v ; pe] = [v; 0]. Since the discrete pair satises the inf-sup condition, we have a
unique solution [vh ; ph ] ∈ Vh × Mh of the corresponding discrete problem and
||v − vh ||V + ||ph ||M ≤ C(||F ||V 0 + ||G||M 0 ) ≤ C||v||V

for some C > 0. We set Πh v := vh and have
b[Πh v, qh ] = G[qh ] = b[v, qh ] for all q h ∈ Mh

and
||Πh v||V ≤ ||v||V + ||Πh v − v||V ≤ C||v||V .

This Πh has the required properties.
⇐ Let p h ∈ Mh with ||ph ||M = 1. Then, there exists β>0 and v∈V with ||v||V = 1
such that
b[v, ph ] ≥ β.
Thus, with Πh as above,
|b[wh , ph ]| b[Πh v, ph ] b[v, ph ]

sup ≥ =
wh ∈Vh \{0} ||wh ||V ||Πh v||V ||Πh v||V
b[v, ph ] 1 1
≥ ≥ β = β||ph ||M .
C||v||V C C
This shows the required bound.
9.9. Discretisation of the mixed Poisson problem

We apply the preceding theory to the case of the Poisson problem in mixed form, Section
9.1 and the given spaces V, M and forms a, b. To prove existence of solutions we will verify
the conditions of Theorem 9.2.
9.9.1. The primal problem
Clearly, a[τ , τ ] = ||τ ||2V , so that a is coercive on the whole V . In order to prove the inf-sup
condition, we dene for given v ∈ M the function τ v := −∇v ∈ V and get
b[τ , v] b[τ v , v] ||v||2M
sup ≥ = = ||v||M .
τ ∈V ||τ ||V ||τ v ||V ||v||M
This proves the stability with β = 1.
For the discretization spaces we choose, for k ≥ 1,
Vhk := {τ h ∈ L (Ω) : τ h bT ∈ Pdk−1
2 d
for all T ∈ Th },
Mhk := {vh ∈ H01 (Ω) : vh b T ∈ Pk for all T ∈ Th }.
Since we have ∇Mhk ⊂ Vhk , the stability proof works in the same way as above. The
error estimate now follows from best approximation results, for example in case k = 1:
||σ − σ h ||L2 (Ω)d . hmax ||∇σ||L2 (Ω)d,d (using the BrambleHilbert Theorem 6.4) and ||∇(u −
uh )||L2 (Ω) . hmax ||∇2 u||L2 (Ω)d,d (Section 6.2). Thus the error is O(hmax ) in the norm on
V × M if u ∈ H2 (Ω).
80
9.9 Discretisation of the mixed Poisson problem 81
9.9.2. The dual problem
Clearly, a[τ , τ ] = ||τ ||2V if ∇ · τ = 0, so that a is coercive on ker(B)! In order to prove the
inf-sup condition, we dene for w ∈ C0∞ (Rd ) the function τ w by
Z x1
τw,1 (x) = w(s, x2 , . . . , xd ) ds,
−∞
τw,j (x) = 0 for j = 2, . . . , d.
Then ∇ · τ w = w, hence τ w ∈ H(div, Ω), and ||τ w ||V . ||w||L2 (Ω)
b[τ , w] b[τ w , w] ||w||2M

sup ≥ & = ||w||M .
τ ∈V \{0} ||τ ||V ||τ w ||V ||w||M
For arbitrary v∈M we now choose a sequence of such functions w that converge to v in
L2 (Ω) to get the same bound for v. This proves the stability with some β > 0.
9.9.3. RaviartThomas elements
For the discretization spaces we dene for k≥0
Vhk := RTk (Th ) := τ h ∈ L2 (Ω)d : τ h bT = a + βx, a ∈ Pdk , β ∈ Pk ,

for all T ∈ Th
τ h · nE is continuous on all E ∈ Eh ,
Mhk := vh ∈ L2 (Ω) : vh bT ∈ Pk

for all T ∈ Th .
By denition one can prove that Vhk ⊂ V = H(div, Ω), that means Vhk is V -conforming.
Clearly, Mh is M -conforming. RTk (Th ) is called RaviartThomas element .
In order to show error estimates we need to prove the discrete stability and the approx-
imation property of the spaces Vh and Mh . For the rst, it suces to construct a Fortin
operator, see Section 9.8, which is done in Section 9.9.4 for k = 0.
Having proved stability, we can conclude that the error measured in the norm of V ×M
is bounded by the best possible approximation error, i. e., in terms of
inf ||σ − τ h ||H(div,Ω) , inf ||u − vh ||L2 (Ω) .

τ h ∈Vh vh ∈Mh
The BrambleHilbert lemma gives
inf ||u − vh ||L2 (Ω) ≤ ||u − Πh0 u||L2 (Ω) ≤ Chmax ||∇u||L2 (Ω)2
vh ∈Mh
and with Πh0 and Π div

h as in Section 9.9.4,
inf ||σ − τ h ||H(div,Ω) ≤ ||σ − Π div div

h σ||L2 (Ω)2 + ||∇ · (σ − Π h σ)||L2 (Ω)
τ h ∈Vh
≤ Chmax ||∇σ||L2 (Ω)2,2 + ||∇ · σ − Π 0h ∇ · σ||L2 (Ω)

≤ Chmax ||∇σ||L2 (Ω)2,2 + ||∇(∇ · σ)||L2 (Ω)2 .
9.9.4. A Fortin operator for the RaviartThomas element (k = 0)

Let Th be an admissible triangulation of a domain Ω = Ωh ⊂ R2 . For an edge E we let nE
be a xed normal on E. For a vector eld τ on Ω for which the following is well-dened,
we let
Z
ΠE τ := τ · nE .
E
81
Since this applies to functions in the RaviartThomas space Vh = Vh0 , we will dene a basis
of it by
ΠE 0 τ E := δEE 0 .
Thus we dene the projection Π div 1 2

h : H (Ω) → Vh by
X
Π div
h τ := ΠE 0 τ τ E 0 .
E 0 ∈Eh
We nd the relation

Z Z X X
ΠE Π div
h τ = Π div
h τ · nE = ΠE 0 τ τ E 0 · nE = ΠE 0 τ ΠE τ E 0
E E E 0 ∈E E 0 ∈Eh
h
= ΠE τ .
For any triangle T ∈ Th we now dene
Z
ΠT v := v
T
for integrable v : Ω → R. Then we dene a basis {χT }T ∈Th of Mh by the conditions
ΠT 0 χT := δT T 0
and a projection Πh0 : L2 (Ω) → Mh by
X
Πh0 v := ΠT 0 v χT 0 .
T 0 ∈Th
Here we nd for all T ∈ Th the relation

Z Z X X
0 0
ΠT Πh v = Πh v = ΠT 0 v χT 0 = ΠT 0 v ΠT χT 0 = ΠT v.
T T T 0 ∈T T 0 ∈Th
h
In the following we want to show that Π div

h is a Fortin operator.
Now let T ∈ Th σET ∈ {−1, 1} be such that n∂T = σET nE on E ∈ Eh , E ⊂ ∂T .
and let
1 2
Then we have for τ ∈ H (Ω) by previously proved relations
Z X Z X Z X
ΠT (∇ · τ ) = ∇·τ = τ · n∂T = σET τ · nE = σET ΠE τ
T E⊂∂T E E⊂∂T E E⊂∂T
X Z
σET ΠE Π div ∇ · (Π div div

= h τ = ··· = h τ ) = ΠT ∇ · (Π h τ ) .
E⊂∂T T
By summation over T ∈ Th
X X
Πh0 (∇ · τ ) = ΠT ∇ · (Π div 0 div

ΠT (∇ · τ )χT = h τ ) χT = Πh ∇ · (Π h τ )
T ∈Th T ∈Th
=∇· (Π div
h τ ),
since ∇ · (Π div
h τ) is piecewise constant. Thus we found the commutation relation
Πh0 (∇ · τ ) = ∇ · (Π div
h τ) for all τ ∈ H1 (Ω)2 . (9.13)
82
9.9 Discretisation of the mixed Poisson problem 83
Since χT is piecewise constant we nd

Z Z
(Π div ∇ · (Π div

b[τ , χT ] = ∇ · τ χT = ΠT (∇ · τ ) χT = ΠT ∇ · h τ) χT = h τ ) χT
T T
= b[Π div
h τ , χT ].
Thus it holds
b[τ , qh ] = b[Π div

h τ , qh ] for all qh ∈ M h
and we thus proved that Π div

h is a Fortin operator if continuous. We omit the continuity
and the approximation result.
By (9.13) the following diagram is commutative
div
H1 (Ω)2 L2 (Ω)
Π div
h Πh0
k div
RTh Mhk
Literature for Section 9. [BS02, Ch. 12] [Bra07, Ch. III, 4-7] [LT05, Kap. 5.7].
83
84 A Stable renement methods for triangulations in R2
A. Stable renement methods for triangulations in R2

The aim of this section is to describe some methods that return a (rened) regular mesh
for a given regular mesh Th and a set Ah of marked triangles to be rened. Such a method
is called geometrically stable if repeated application of the method leads to a sequence of
meshes such that the corresponding sequence of shape functions σh is uniformly bounded.
Note that the bound for σh enters almost all of our previous estimates. The following
methods will fulll this requirement.
A.1. RedGreen (Blue) Renement

[Ver96, Sect. 4.1]. In case T ∈ Ah we will use red renement as shown in Figure 12, left.
Note that all new triangles have the same angles as their origin triangle. For the neighbours
that are not in Ah , we enforce bisection to remove the hanging node, regardless of the size
of the new angles. These so called green renement will be coarsened before such a triangle
will be rened in a subsequent step. If the opposite angle that is divided by bisection will
be very small, one might use bisection methods that avoid these (like the one in the next
subsection). In this context, this renement is called blue renement , Figure 12, right.
Triangles rened in this way need not to be removed in later steps.
Figure 12: Left: Red renement. Right: Blue renement of the right triangle.
A.2. Newest node bisection.

[Bän91]. We store the nodes P1 , P2 , P3 of triangle T in a counter-clockwise sense nodes
[T, : ] = [P1 ,P2 ,P3 ]. If T is marked, we plan to bisect T from P1 towards the opposite
edge as in Figure 13. The two new triangles must have again oriented nodes-entries with
Pnew in the rst place. However, the neighbouring triangle has to be added to Ah if Pnew is
a hanging node. In a practical implementation one actually only renes simple situations
• two triangle share the marked edge,
• the marked edge is a boundary edge,
as seen in Figure 13. It can be shown that Algorithm 2, that is based on these ideas, stops
after nitely many steps and results in a locally rened mesh. The renement method is
geometrically stable if the macro triangulation T0 has a correct distribution of marked
edges. We only have to avoid that all edges emanating from one vertex are marked edges.
A newest node bisection with m-times repeated bisection is as follows. The marking
procedure assigns to each T m(T ) ∈ N. Then Ah := {T ∈ Th : m(T ) > 0}.
a number
After each bisection, we decrease m(T ) until we obtain 0, see Algorithm 2.
A.3. Longest edge bisection.

Rene always towards the longest edge in T and only in one of the situations as in Figure
13. For that work on a copy of the coordinate eld with a random distortion (to assure that
the longest edge in T is unique almost sure). The important question is now whether this
ends after nitely many steps in one of the mentioned situations? The answer is armative
84
85
Algorithm 2 Repeated newest node bisection

1: while any m(T ) > 0 do
2: for T with m(T ) > 0 do
3: identify neighbour T0 opposite P1 ;
4: situation Figure 13 left/middle ?;
5: Yes :
6: bisect T (and T 0 );
7: m(T ) := max{m(T ) − 1, 0};
8: 0 0
(m(T ) := max{m(T ) − 1, 0});
9: No :
10: m(T 0 ) := max{m(T 0 ), 1};
P1’ Pnew
P1 Pnew
Figure 13: Left: Bisection of interior triangles. Middle: Bisection of a boundary triangle.
Right: Renement with interior node property (3 bisections).
since following these choices the length of the largest edge increases and thus will end at
an edge of locally maximal length or at the boundary. The algorithm is like Algorithm 2,
but a special ordering in the nodes-eld is not necessary.
Literature for Section A. [Ver96] [AO00] [Bän91]
B. The discontinuous Galerkin method for the Poisson

problem
We again consider the Poisson problem
−∆u = f in Ω,
(B.1)
u=0 on ∂Ω
for some f ∈ L2 (Ω), Ω ⊂ Rd bounded with polygonal boundary.

We suppose that we have a triangulation Th of Ω such that Ω = Ωh . For k ∈ N0 we
dene the approximation space by
Vh = Vhk = vh : Ω → R : vh bT ∈ Pk (T )

for all T ∈ Th .
In this case, there is no condition relating vh on a triangle T 0.

T to any dierent
Each vh b is dierentiable but vh is not even weakly dierentiable globally on Ω . There-
T
fore we introduce a special notation for the piecewise gradient ∇h by
∇h vh (x) = ∇vh (x) for x ∈ T ◦.
Note that ∇h v = ∇v for v ∈ H1 (Ω).
85
86 B The discontinuous Galerkin method for the Poisson problem
To state a weak formulation for our problem we start considering
Z X Z
0
a [vh , wh ] := ∇h vh · ∇h wh = ∇vh · ∇wh for all vh , wh ∈ Vh .
Ω T ∈Th T
In order to nd out whether this is reasonable, we will check consistency, i.e., we examine
what happens if we insert the continuous solution of (B.1). To follow this idea, we will
rst reformulate our form using integration by parts. We will get
Z X Z
0
a [vh , wh ] = ∇vh · ∇h wh = ∇vh · ∇wh
Ω T ∈Th T
X Z X Z
=− ∆vh wh + ∂ n v h wh .
T ∈Th T T ∈Th ∂T
The idea is to write the second sum in terms of edges. We recall from Section 7.4 the
denition of a jump [[vh ]]E on an inner edge E ∈ E i , while we use {{vh }}E for the mean value
[[vh ]]E (x) = lim vh (x + snE ) − vh (x − snE ) ,

s→0+
1
lim vh (x + snE ) + vh (x − snE )

{{vh }}E (x) =
2 s→0+
for x ∈ E. On boundary edges E ∈ Eb we use [[vh ]]E = {{vh }}E = vh for convenience. With
this denition we nd
[[(∂n vh )wh ]]E = {{vh }}E [[wh ]]E + [[vh ]]E {{wh }}E
and thus
X Z X Z X Z
a0 [vh , wh ] = − ∆vh wh + {{∂n vh }}E [[wh ]]E + [[∂n vh ]]E {{wh }}E .
T ∈Th T E E
E∈Eh E∈Ehi
Now we assume that the solution u of (B.1) is in H2 (Ω) and since a0 [u, wh ] is also
well-dened we get
Z X Z
0
a [u, wh ] = f wh + {{∂n u}}E [[wh ]]E for all wh ∈ V h .
Ω E∈Eh E
Thus we dene a new form

Z X Z
1
a [vh , wh ] := ∇ h v h · ∇ h wh − {{∂n vh }}E [[wh ]]E for all vh , wh ∈ Vh .
Ω E∈Eh E
a1 [u, wh ] =
R
which is now consistent since it holds
Ω f wh for all wh ∈ V h .
The form is not symmetric although the original problem is symmetric. This is a de-
ciency since then the discrete problem is not symmetric and this is unfavourable in view
of the linear solvers. However, we can simply symmetrise this form by
Z X Z n o
sym
a [vh , wh ] := ∇h vh · ∇h wh − {{∂n vh }}E [[wh ]]E + [[∂n vh ]]E {{wh }}E .
Ω E∈Eh E
Note, that this form is still consistent in the previously stated sense because the new term
vanishes for H2 -functions in place of vh .
86
87
In order to discuss stability we want to have coercivity for the bilinear form. However,
only get
Z X Z
sym 2
a [vh , vh ] = |∇h vh | − 2 {{∂n vh }}E [[vh ]]E
Ω E∈Eh E
which will not show a denite sign. This might be understandable since no form of con-
tinuity is required. To enforce continuity in the limit of ner meshes we add a penalty
term
X η Z
sh [vh , wh ] = [[vh ]]E [[wh ]]E ,
|E| E
E∈Eh
for some h-independent η > 0, that penalises the discontinuities over the edges. Again,
this additional term does not change consistency and symmetry.
So in summary we dene the bilinear form
asip [vh , wh ] := asym [vh , wh ] + sh [vh , wh ] for all vh , wh ∈ Vh ,
where 'sip' refers to symmetric interior penalty method. The discrete problem we want to
solve is nding uh ∈ Vh such that
Z
sip
a [uh , wh ] = f wh for all wh ∈ V h .
Ω
The analysis will of this problem will be in the norm
1/2
|||vh |||sip = ||∇h vh ||2L2 (Ω) + sh [vh , vh ]2 .
The derivation of lower and upper bounds for asip is more involved than for continuous
nite elements. As nal results we can however state that we get
|||u − uh |||sip ≤ Chk ||∇k+1 u||L2 (Ω)
and with the duality trick (Section 7.1)
|||u − uh |||L2 (Ω) ≤ Chk+1 ||∇k+1 u||L2 (Ω)
if u shows the requested regularity.
Literature for Section B. [DE12, Ch. 4.2] [CDGH17]
87
88 References
References
[Alt16] H. W. Alt. Linear Functional Analysis. An Application-Oriented Introduction.
Springer, Berlin, 2016.
[AO00] M. Ainsworth and J. T. Oden. A Posteriori Error Estimation in Finite Element

Analysis. John Wiley, New York, 2000.
[Bän91] E. Bänsch. Local mesh renement in 2 and 3 dimensions. Impact Comput. Sci.
Engrg., 3:181191, 1991.
+
[BBC 93] R. Barrett, M. W. Berry, T. F. Chan, J. Demmel, J. Donato, J. Dongarra,
V. Eijkhout, R. Pozo, C. Romine, and H. van der Vorst. Templates for the
Solution of Linear Systems: Building Blocks for Iterative Methods. SIAM,
Philadelphia, 1993.
[Bra07] D. Braess. Finite Elemente. Theorie, schnelle Löser und Anwendungen in der
Elastizitätstheorie. Springer, Berlin, 4 edition, 2007.
[BS02] S. C. Brenner and L. R. Scott. The mathematical theory of nite element

methods, volume 15 of Texts in Applied Mathematics. Springer, New York, 2
edition, 2002.
[CDGH17] A. Cangiani, Z. Dong, E. H. Georgoulis, and P. Houston. hp-Version Discon-

tinuous Galerkin Method on Polygoal and Polyhedral Meshes. Springer Briefs
in Mathematics. Springer, Cham, 2017.
[Cia79] P. G. Ciarlet. The nite element method for elliptic problems. Studies in
Mathematics and its Applications. North-Holland, Amsterdam, 1979.
[Clé75] P. Clément. Approximation by nite element functions using local regulariza-

tions. RAIRO Anal. Numér., 2:7784, 1975.
[DE12] Daniele DiPietro and Alexandre Ern. Mathematical Aspects of Discontinu-

ous Galerkin Methods, volume 69 of Mathematics and Applications. Springer,
Berlin, 2012.
[DH02] P. Deuhard and A. Hohmann. Numerische Mathematik 1: Eine algorithmisch

orientierte Einführung. de Gruyter, Berlin, 3 edition, 2002.
[Dör96] W. Dörer. A convergent adaptive algorithm for Poisson's equation. SIAM J.

Numer. Anal., 33(3):11061124, 1996.
[Dör13] W. Dörer. Numerical methods for partial dierential equations. hp nite

element methods (1d), 2013. Lecture Notes, Karlsruhe Institute of Technology.
[Dör15] W. Dörer. Numerical methods for partial dierential equations. Some theo-
retical background on linear elliptic equations, 2015. Lecture Notes, Karlsruhe
Institute of Technology.
[Dör21a] W. Dörer. Einführung in das Wissenschaftliche Rechnen, 2021. Vor-

lesungsskript, Karlsruher Institut für Technologie.
[Dör21b] W. Dörer. Numerische Methoden für Dierentialgleichungen, 2021. Vor-

lesungsskript, Karlsruher Institut für Technologie.
[DR16] W. Dörer and A. Rieder. Numerische Mathematik 12, 2016. Vorlesungsskript,

Karlsruher Institut für Technologie.
88
References 89
[DR20] W. Dörer and A. Rieder. Numerische Mathematik 12, 2020. Vorlesungsskript,

Karlsruher Institut für Technologie.
[HB02] M. Hanke-Bourgeois. Grundlagen der numerischen Mathematik und des wis-

senschaftlichen Rechnens. Mathematische Leitfäden. Teubner, Stuttgart, 2002.
[HR03] V. Heuveline and R. Rannacher. Duality-based adaptivity in the hp-nite ele-

ment method. J. Numer. Math., 11:95113, 2003.
[HW08] J. S. Hesthaven and T. Warburton. Nodal discontinuous Galerkin methods.

Springer, 2008.
[Kel75] R. B. Kellogg. On the Poisson equation with intersecting interfaces. Applicable

Analysis, 4:101129, 1975.
[LT05] S. Larsson and V. Thomée. Partielle Dierentialgleichungen und numerische

Methoden. Springer, Berlin, 2005.
[MNS02] P. Morin, R. H. Nochetto, and K. G. Siebert. Convergence of adaptive nite

element methods. SIAM Review, 44:631658, 2002.
[R·º04] M. R·ºi£ka. Nichtlineare Funktionalanalysis. Eine Einführung. Springer,

Berlin, Heidelberg, 2004.
[Sch98] C. Schwab. p- and hp-nite element methods. Theory and applications in solid
and uid mechanics. Clarendon Press, Oxford, 1998.
[Ste07] R. Stevenson. Optimality of a standard adaptive nite element method. Found.

Comput. Math., 7:245269, 2007.
[SZ90] L. R. Scott and S. Zhang. Finite element interpolation of non-smooth functions

satisfying boundary conditions. Math. Comput., 54:483493, 1990.
[Ver96] R. Verfürth. A Review of A Posteriori Error Estimation and Adaptive Mesh-

Renement Techniques. Wiley-Teubner, Chichester, 1996.
89

Fem Doerfler

Uploaded by

Copyright:

Available Formats

Fem Doerfler

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fem Doerfler

Uploaded by

Copyright:

Available Formats

Finite Element Methods

Prof. Dr. W. Dörfler

Institute for Applied and

3. Linear nite elements 12

8. Example of a nite element implementation 68

9. Saddle Point Problems 72

A. Stable renement methods for triangulations in R2 84

B. The discontinuous Galerkin method for the Poisson problem 85

1.2. The Dirichlet boundary value problem

The boundary condition (1.2) is calledDirichlet boundary condition. A special case

If in addition f =0 holds, then the solution u to this problem is called harmonic.

The v 's are called test functions.

Find u with ub∂Ω = uD such that (1.5) holds. (1.6)

and our problem now reads:

Find u:Ω→R with ub∂Ω = 0 such that (1.7) holds. (1.8)

([Dör15, Sect. 1.2.2, 1.2.3]), where

The following subspace is well-dened [Dör15, Sect. 1.3.3]

W01,2 (Ω) ≡ H01 (Ω) := v ∈ W 1,2 (Ω) : v b∂Ω = 0 .

A[v, v] ≥ α0 ||v||2V for all v ∈ V,

and A is continuous, i.e.,

A[v, w] ≤ α1 ||v||V ||w||V for all v, w ∈ V,

F [v] ≤ αF ||v||V for all v ∈ V.

Then there exists a unique solution u of the variational equation

and it satises the a priori bound

α0 ||u||2V ≤ A[u, u] = F [u] ≤ αF ||u||V .

Then there exists a unique function u ∈ H01 (Ω) such that

A[u, v] = F [v] for all v ∈ H01 (Ω) (1.10)

Furthermore it holds the a priori bound

with a constant C e D , and Ω . For uD = 0, u is the unique solution

Find u∈V such that A[u, v] = F [v] for all v ∈V.

A[v, v] ≥ a0 ||∇v||2L2 (Ω)d ,

||v||L2 (Ω) ≤ CP (Ω)||∇v||L2 (Ω)d for all v ∈ H01 (Ω)

holds. CP (Ω) is called Poincaré constant.

[Dör15, Sect. 1.3.1 (Example)] !

−∇·(a∇u) = f on Ω = (−1, 1)2 ,

u(x) ∼ |x|1/10 for |x| ≈ 0,

and we nd u ∈ H1+1/10 (Ω) [Kel75].

(7) The general linear elliptic equation has the form

1.3. The mixed DirichletNeumann boundary value problem

The imposed boundary condition (1.14) on ΓN is called Neumann boundary condition

Note, that the more general boundary condition u = u

which is turned into

This is again a problem of the form treated in Theorem 1.3:

If |Γ D | 6= 0 ∂Ω ), then the Poincaré inequality (Lemma

and require for the data f, uN the compatibility condition

Lemma 1.6 (Poincaré inequality with meanvalue zero) . Ω ⊂ Rd be open,

1.4. Dirichlet principle

Note that with u ∈ Ve and v∈V we have u + v ∈ Ve . This results in

E 0 (u)[v] = A[u, v] − F [v] = 0 for all v ∈ V.

E(v) = Ω {1/2(a|∇v|2 + c|v|2 ) − f v} on the ane

space Ve = {v ∈ H1 (Ω) : v b∂Ω = uD } = u

• curved boundaries, especially for higher order methods,

• a symmetric problem is not sucient to yield a symmetric system of equations.

2.1. The discrete problem

Find uN ∈ VN such that A[uN , vN ] = F [vN ] for all vN ∈ VN . (2.1)

with coecients αψ ∈ R, ψ ∈ B . Using this as an ansatz and testing with vN = φ, φ ∈ B ,

This is a linear system of equations with a symmetric positive denite matrix ~.

||v||E := A[v, v]1/2 for all v ∈ V.

This is indeed a norm on V that is equivalent to || . ||V since A is assumed to be coercive

3. Linear nite elements 12

8. Example of a nite element implementation 68

A. Stable renement methods for triangulations in R2 84

The following subspace is well-dened [Dör15, Sect. 1.3.3]

and it satises the a priori bound

and we nd u ∈ H1+1/10 (Ω) [Kel75].

1.3. The mixed DirichletNeumann boundary value problem

Note that with u ∈ Ve and v∈V we have u + v ∈ Ve . This results in

E(v) = Ω {1/2(a|∇v|2 + c|v|2 ) − f v} on the ane

• a symmetric problem is not sucient to yield a symmetric system of equations.

with coecients αψ ∈ R, ψ ∈ B . Using this as an ansatz and testing with vN = φ, φ ∈ B ,

This is a linear system of equations with a symmetric positive denite matrix ~.

3. Linear nite elements

Denition 3.1 (Basic denitions). We dene the set of nodes Nh by

The rst assertion results after insertion of this into

In the inmum expression we choose vh = Ih u and obtain with Lemma 3.4

With the coercivity we nally get

3.1.4. Numerical integration (variational crimes)

for suciently regular g. On a decomposition Kh we use the composite quadrature rule

provided g [k+1] ∈ L1 ((0, 1)). With GauÿLegendre quadrature we have