Department of Mathematics and Computer Science

Convergence of Adaptive Finite Element

Methods for the Poisson Problem
Master Thesis

Submitted by: Ines Perko

Supervisor: Univ.-Prof. Dr. Volker John

Second Reviewer: Dr. Alfonso Caiazzo

Berlin, July 30, 2020

Contents
1 Introduction 1

2 Finite Element Methods 3

2.1 Lebesgue spaces and Sobolev spaces . . . . . . . . . . . . . . . . 3
2.2 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Conforming Finite Element discretization . . . . . . . . . . . . . 10
2.4 A priori estimate of the error in the L2 norm of gradient . . . . . 12
2.5 Neumann and Dirichlet cases . . . . . . . . . . . . . . . . . . . . 16

3 A Posteriori Error Estimators 20

3.1 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Adaptive algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Residual a posteriori error estimates . . . . . . . . . . . . . . . . 21
3.4 Ways for refining the mesh . . . . . . . . . . . . . . . . . . . . . 25

4 Convergence of an adaptive algorithm 30

5 Numerical Studies 34

6 Conclusion and Outlook 44

6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

List of figures 46

List of tables 46

References 47

Appendix A Tools from the Analysis 50

Nomenclature 58
1 Introduction
Many physical problems arising in engineering and science can be mathemat-
ically modeled and expressed through partial differential equations. Since an-
alytical solutions of these equations cannot be explicitly computed in general,
numerical methods for finding approximate solutions, such as the finite element
methods are used. With its roots in the 1940s [Cou43], and further development
through the 1950s to 1970s by engineers and mathematicians, the first book that
focused on the mathematical foundations is [SF73], published in 1973. Nowa-
days, many physical and chemical phenomena apply finite element methods to
solve them since this method easily handles boundary conditions, complex ge-
ometry and because of its clear structure it is possible to develop a software for
applications. From the mathematical side, the function spaces are needed for
the presentation of the finite element method. More precisely, the Lebesgue and
Sobolev spaces presented in this thesis are sufficient. Furthermore, the varia-
tional formulation of the partial differential equation should be obtained. Here,
the model problem is the Poisson equation with Dirichlet-Neumann boundary
conditions, on which the finite element method will be developed. The model
problem is rewritten in its variational formulation and the existence and unique-
ness of the solution are proven. Knowing that the solution exists, an approxima-
tion is done by finite elements. In the late 1970s, the concept of adaptivity was
developed within the framework of finite element methods. The method is called
the adaptive finite element method. Its importance and potential come from
its concept where automatic adjustment is done to improve approximations.
The adaptive method handles practical problems in solid and fluid mechanics,
in porous media flow and semiconductor device simulation [BS08] and provides
the optimal overall accuracy of the numerical approximations, since in these
cases local singularities might be present and therefore the overall accuracy de-
creases. One of the first works that presented the idea of this method based on
the error estimates is [BR78]. In particular, the a posteriori error estimators are
shown as a necessity for adaptivity. They are important for the evaluation of
the reliability of the results as well as controlling the refinement process, which
is a crucial process in this method.
The main goal of the thesis is to show the convergence of adaptive finite
element methods for the Poisson problem. The proof presented here follows
[Ver13], and it is based on the Poisson problem with pure Dirichlet boundary
conditions. The adaptive method is presented through the adaptive algorithm
[Ver13, Algorithm 1.1]. To make this algorithm effective, an error indicator
which supplies the a posteriori error estimate should be chosen along with a
reliable refinement strategy. In the proof here, for the a posteriori error es-
timates, the residual estimates are utilized. They estimate the error of the
numerical solution by a norm of its residual. Specifically, for this algorithm,
the error estimator in the L2 -norm of the gradient is used. The first proof of
the convergence based on a posteriori error estimates was published in [D9̈6] in
1996. Nowadays, these results are improved and a lot of literature is based on
the idea presented in [D9̈6]. Some of the improvements are simultaneous control
of the error indicator and the data oscillation [MNS02].
Chapter 2 provides the mathematical theory required for the development
of the finite element method. Also, the a priori estimate of the error in the
L2 -norm of gradient is introduced. The chapter ends with an inspection of the

1
different values for Dirichlet and Neumann boundary conditions.
Chapter 3 presents the adaptive algorithm along with all the necessary data
for it. Therefore, the residual a posteriori error estimates that provide upper and
lower bounds for the error are introduced as well as the residual a posteriori error
indicator. The marking strategies are presented and the processes of refinement,
coarsening, and smoothing of meshes that obtain the adaptive discretizations
are described.
After having successfully defined data and making an adaptive algorithm
effective, chapter 4 proves the convergence of it. The idea is to show that by
every iteration of the algorithm, the error is reduced at least by some factor.
Chapter 5 describes the numerical results of the simulations of the two ex-
amples which consider the Poisson equation. The properties of the interest
are discussed and the obtained results support the convergence of the adaptive
algorithm. Chapter 6 presents the conclusion and the outlook of the thesis.

2
2 Finite Element Methods
2.1 Lebesgue spaces and Sobolev spaces
This chapter starts with definitions and basic properties of the Lebesgue spaces
and Sobolev spaces following the literature [AF03], [OR76], [Maz85], [BS08],
[Che05], [Gri11], [DD12], and [Ada75].
Let the domain Ω be an open, non-empty set in n-dimensional real Euclidean
space Rn . The standard Lebesgue spaces are denoted with Lp (Ω), whenever
1 ≤ p < ∞. The space Lp (Ω) is the space of Lebesgue measurable functions u
defined on Ω for which
Z
|u(x)|p dx < ∞. (1)
Ω

The norm on Lp (Ω), which is denoted by k · kLp (Ω) , is defined by

Z  p1
kukLp (Ω) = |u(x)|p dx . (2)
Ω

Positivity and homogeneity follow from the definition and the inequality that is
used for the triangle inequality is

ku + vkLp (Ω) ≤ kukLp (Ω) + kvkLp (Ω) .

This inequality is also known as Minkowski’s inequality. Therefore, this is a

normed space. In Lp (Ω) one identifies functions which are equal almost every-
where in Ω. Thus, the elements of Lebesgue spaces are equivalence classes of the
functions u defined on Ω satisfying (1). For the case p = ∞, one gets the space
of essentially bounded Lebesgue measurable functions. This Lebesgue’s space
is denoted by L∞ (Ω). As a norm, one can take its essential supremum, i.e.,

kukL∞ (Ω) = ess sup |u(x)|. (3)

x∈Ω

Another fundamental inequality in Lp (Ω) spaces is Hölder’s inequality.

1 1
Theorem 2.1. (Hölder’s inequality) Let 1 ≤ p, q ≤ ∞ such that p + q = 1.
If u ∈ Lp (Ω) and v ∈ Lq (Ω), then uv ∈ L1 (Ω) and
Z
|u(x)v(x)|dx ≤ kukLp (Ω) kvkLq (Ω) . (4)
Ω

Proof. The proof can be found in [AF03], Theorem 2.16.

One of the basic properties of the Lebesgue spaces Lp (Ω) is that they are
Banach spaces for 1 ≤ p ≤ ∞. For the case p = 2, one gets the space L2 (Ω).
That is a Hilbert space with respect to the inner product
Z
(u, v) = u(x)v(x)dx. (5)
Ω

Hölder’s inequality in L2 (Ω) space is called Cauchy’s or Schwarz’s inequality.

3
 Furthermore, the space C0∞ (Ω) of test functions on Ω should be introduced.
It is the subset of the space C ∞ (Ω) which consists of the functions from C ∞ (Ω)
on Ω with compact support in Ω, i.e.,

C0∞ (Ω) = {ϕ ∈ C ∞ (Ω)|suppϕ is compact in Ω}. (6)

In other words, a test function is a function that has continuous derivatives of

all orders but it will vanish outside of some bounded set thanks to the compact
support. Thus, when Ω is bounded, it is equivalent to saying that u vanishes
in a neighborhood of the boundary Γ of Ω. The space C0∞ (Ω) is contained in
the most of the spaces defined in connection with Ω as C0∞ (Ω) ⊂ Lp (Ω), for
1 ≤ p ≤ ∞.
Corollary 2.1. C0∞ (Ω) is dense in Lp (Ω) if 1 ≤ p < ∞.
Proof. The proof can be found in [AF03], Corollary 2.30.
To define the Sobolev spaces, the weak derivative has to be introduced. For
that, the following function space is needed.
A function u is locally integrable function if the Lebesgue integral
Z
|u|dx < ∞
K

exists for every compact set K ⊂ Ω. The space of all locally integrable functions
is denoted by L1loc (Ω). The space Lp (Ω) is a subset of the space L1loc (Ω) for
1 ≤ p ≤ ∞ and any domain Ω.
Let u and wα be locally integrable on Ω. A function u has a weak derivative
w of order α if wα fulfills
α

Z Z
wα φdx = (−1)|α| uDα φdx (7)
Ω Ω

for every test function φ ∈ C0∞ (Ω). Here, wα is the αth generalized derivative
of the u. Whenever the classical derivative Dα u(x) exists, it is also a weak
derivative of u. Also, wα is unique up to a set of measure zero. The concept of
the weak derivatives is brought where the classical differentiability fails. Thus,
if classical derivatives do not exist, Dα will refer to the weak derivatives.
Sobolev spaces are vector subspaces of different Lebesgue spaces Lp (Ω). The
standard Sobolev space is denoted by W m,p (Ω), where m is any positive integer
and 1 ≤ p ≤ ∞. It contains all functions u ∈ Lp (Ω) which have all weak
derivatives up to order m in Lp (Ω):

W m,p (Ω) = {u : Dα u ∈ Lp (Ω); ∀α such that |α| ≤ m}. (8)

The Sobolev norm on the space W m,p (Ω), for the case 1 ≤ p < ∞ is
 X  p1
kukW m,p (Ω) = kDα ukpLp (Ω) . (9)
|α|≤m

For the case p = ∞, it is defined as

kukW m,∞ (Ω) = max kDα ukL∞ (Ω) . (10)

|α|≤m

4
 One can see that (9) or (10) defines a norm on any vector space of functions
on which the right-hand side takes finite values. Therefore, W m,p (Ω) is a normed
space. The following theorem shows that it is complete. Also, one can notice
that W 0,p (Ω) is equal to Lp (Ω).
Theorem 2.2. W m,p (Ω) is a Banach space.
Proof. This proof can be found in [AF03], Theorem 3.3.
The subspace C ∞ (Ω) ∩ W m,p (Ω) is dense in W m,p (Ω). The stronger density
condition is that C0∞ (Ω) is dense in W m,p (Ω). Whenever the part of the domain
lies on both sides of part of its boundary (as occurs with a slit domain) this
does not hold. Thus, some regularity conditions must hold to obtain this. The
segment condition should be satisfied, i.e., if for every x ∈ Γ there is a neighbor-
hood Ux and a nonzero vector nx such that if z ∈ Ω ∩ Ux then z + tnx ∈ Ω for
0 < t < 1. The vector nx is an inward directed normal to Γ at x. The boundary
Γ satisfying this condition must be (n − 1)-dimensional. With this density con-
dition, one can always approximate an element of W m,p (Ω) by smooth bounded
functions having bounded derivatives of all orders on Ω.
Furthermore, W0m,p (Ω) is a Sobolev space defined as the completion of
∞
C0 (Ω) with respect to the norm k · kW m,p (Ω) . This space can be defined in
another way, but first, one should mention imbeddings of the space W m,p (Ω)
into Banach spaces. The most important of the imbedding properties of the
spaces W m,p (Ω) are gathered together under the theorem called the Sobolev
Imbedding Theorem, even though they are of different types and they can re-
quire different properties and different methods of the proof. One can find more
information about it in the [AF03, Chapter IV].
The question which appears is what does imbedding mean for the elements
of Sobolev spaces. Elements of W m,p (Ω) are equivalence classes of functions
defined everywhere on Ω equal up to sets of measure zero. To start with, one can
j
look at imbeddings into the continuous function spaces, i.e., C j (Ω), CB (Ω) and
j,λ m,p
C (Ω). For example, observe the existence of imbedding W (Ω) → C j (Ω).
That means that each u ∈ W m,p (Ω) (when considered as a function) can be
redefined on a subset of Ω, which has a zero measure. The new, modified
function v ∈ C j (Ω) is produced in such a way that v = u in W m,p (Ω) and it
satisfies
kvkC j (Ω) ≤ KkukW m,p (Ω) ,
with K independent of u.
Also, imbeddings can be interpreted as an inclusion relation. Therefore, it
provides the ordering among them. For nonnegative integers m and k satisfying
k ≤ m it holds that

W m,p (Ω) ⊂ W k,p (Ω), 1 ≤ p ≤ ∞. (11)

When Ω is bounded it holds that

0
W m,p (Ω) ⊂ W m,p (Ω), 1 ≤ p ≤ p0 ≤ ∞. (12)

There exists an operator E called an (m, p)-extension operator for Ω. It maps

W m,p (Ω) into W m,p (Rn ) where Ω has a Lipschitz boundary Γ. As usual, m is a

5
non negative integer and p is a real number in the range 1 ≤ p ≤ ∞. Extension
mapping satisfies Ev|Ω = v for all u ∈ W m,p (Ω) and

kEukW m,p (Rn ) ≤ CkukW m,p (Ω) , (13)

where C is independent of u.
This property does not hold when Ω does not have Lipschitz boundary.
Thus, with the definition of Lipschitz boundary one can now relate Sobolev
spaces defined on Ω to those on Rn . The complementary result is also possible,
for any domain where the restriction permits to see functions in W m,p (Rn ) as
well defined in W m,p (Ω). The Sobolev Imbedding Theorem claims the existence
of imbeddings of W m,p (Ω) in Lq (Ω). If it is proven for Rn and if domain satisfies
stronger regularity conditions, then it must hold for domain Ω as well. Thus, if
W m,p (Rn ) → Lq (Rn ) and u ∈ W m,p (Ω) one can extend the inequality (13) and
get the imbedding W m,p (Ω) → Lq (Ω) through the chain of inequalities:

kukLq (Ω) ≤ kEukLq (Rn ) ≤ C2 kEukW m,p (Rn ) ≤ C2 C1 kukW m,p (Ω) .

Another inequality that should be introduced is Sobolev’s inequality. It

presents the relationship between Sobolev spaces with different indices. This
inequality also tells that any function with suitable weak derivatives can be
interpreted as a continuous and bounded function.
Theorem 2.3. (Sobolev’s inequality) Let Ω ⊂ Rn be a domain with Lipschitz
boundary Γ, let m be a positive integer and let p be a real number in the range
1 ≤ p < ∞ such that
m ≥ n when p = 1,
n
m > when p > 1.
p
Then there is a constant C such that for all u ∈ W m,p (Ω)

kukL∞ (Ω) ≤ CkukW m,p (Ω) .

There is a continuous function in the L∞ (Ω) equivalence class of u.

Proof. This proof can be found in [BS08], Theorem 1.4.6.

The question which intuitively arises is how is a function from the Sobolev
space defined on the boundary.
The boundary Γ of an n-dimensional domain Ω can be considered as an
(n − 1)-dimensional manifold. For the case n = 1 one gets zero-dimensional
manifold which consists of distinct points. In that case, Sobolev inequality
can be used to obtain conditions under which point values are well defined for
functions in the Sobolev space. Furthermore, restrictions u|Γ of functions u
from Sobolev spaces to manifolds of dimension n − 1 are used for the higher
dimensional cases. Also, one should use the Lipschitz domain and Lipschitz
boundary. For the smooth boundary and u ∈ W 1,p (Ω), the restriction to the
boundary can be considered as a function in Lp (Γ), whenever 1 ≤ p ≤ ∞, but
it does not guarantee that pointwise values of u on Γ will make sense. If p = 2,
then by the definition, the restriction u|Γ is square integrable on Γ. Using this
property one can define space W0m,p (Ω), that is the subset of W m,p (Ω), as

6
 W0m,p (Ω) = {u ∈ W m,p (Ω) : Dα u|Γ = 0 in L2 (Γ), |α| < m}. (14)
This restriction is often called the trace map. It is denoted by γ and it is
defined on a Lipschitz domain Ω as a continuous linear map γ : W 1,p (Ω) →
Lp (Γ) such that if u ∈ C(Ω) ∩ W 1,p (Ω) its image γu is well-defined function on
Γ.
Another important inequality that should be introduced is Poincaré inequal-
ity. It is also called Friedrichs inequality. In order to do that, the space H01 (Ω)
should be defined first.
For the standard Sobolev space W m,p (Ω), in the case of p = 2, one gets the
space W m,2 (Ω). This Sobolev space is denoted by H m (Ω) and consequently,
space W0m,2 (Ω) is denoted by H0m (Ω). It can be defined as

H m (Ω) = {u : Dα u ∈ L2 (Ω) : ∀α such that |α| ≤ m}. (15)

Also, it is a Hilbert space with the inner product

X X Z
(u, v)H m (Ω) = (Dα u, Dα v)L2 (Ω) = Dα uDα vdx. (16)
|α|≤m |α|≤m Ω

The corresponding norm is

 X  12 1
kukH m (Ω) = kDα uk2L2 (Ω) = [(u, u)H m (Ω) ] 2 < ∞. (17)
|α|≤m

The space H01 (Ω) is the space of all functions u ∈ L2 (Ω) which have all their
first order derivatives in L2 (Ω) and their trace vanishes at the boundary. It is
an important space used in the theory of boundary value problems. Finally, the
Poincaré inequality can be introduced.
Proposition 2.1. (Poincaré inequality) Let Ω ⊂ Rn be a bounded Lipschitz
domain, then there exists a constant CP > 0 such that

kukH 1 (Ω) ≤ Cp k∇ukL2 (Ω) , ∀u ∈ H01 (Ω). (18)

Proof. The proof can be found in [DD12], Proposition 5.28.

Here, the Poincaré inequality is defined on H01 (Ω), but there also exists a
generalized version of it. For that, the term seminorm has to be introduced.
For a non-negative integer m and u ∈ W m,p (Ω), the seminorm on W m,p (Ω)
is defined as
 X  p1
|u|W m,p (Ω) = kDα ukpLp (Ω) , 1 ≤ p < ∞, (19)
|α|=m

|u|W m,∞ (Ω) = max|α|=m kDα ukL∞ (Ω) , p = ∞. (20)

It has all properties of a norm, except that |u|W m,p = 0 does not imply u = 0
in W m,p (Ω). If Ω is bounded then | · |W m,p (Ω) is a norm on W0m,p (Ω) equivalent
to the usual norm k · kW m,p (Ω) .

7
Proposition 2.2. (Generalized Poincaré inequality) Let Ω ⊂ Rn be a
bounded Lipschitz domain, 1 ≤ p < ∞ and let N be a continuous seminorm
on W 1,p (Ω), i.e., a norm on the constants. Then there exists a constant C > 0
that depends on Ω, n, p such that
 Z  p1 
kukW 1,p (Ω) ≤ C |∇u(x)|p dx + N (u) , ∀u ∈ W 1,p (Ω). (21)
Ω

Proof. The proof can be found in [DD12], Proposition 5.55.

The Sobolev spaces W m,p (Ω) can also be defined for negative integers m.
The definition is based on the duality of Banach spaces. The dual space of the
Lebesgue space Lp (Ω) can be easily defined through Hölder’s inequality. Let p
be in the range 1 < p < ∞ and let q be the dual index to p, i.e., 1q + p1 = 1.
Thus, space Lq (Ω) is a dual space of Lp (Ω). The space Lp (Ω) is reflexive if and
only if 1 < p < ∞. The dual of L∞ (Ω) is larger than L1 (Ω). Therefore for
space L1 (Ω) and L∞ (Ω), this does not hold. The dual space of the Sobolev
space W m,p (Ω) is defined as (W0−m,q (Ω))0 where q is the dual index to p. Also,
a negative Sobolev space H −m (Ω) is defined as H −m (Ω) = (H0m (Ω))0 .

2.2 Model problem

A model problem on which the finite element method will be developed is in-
troduced in this section following [Ver13], [BS08], [Che05], and [LB13].
The domain Ω is defined as a connected, bounded and polygonal set in
2-dimensional real Euclidean space R2 . The boundary Γ consists of two dis-
joint parts: ΓD (Dirichlet boundary) and ΓN (Neumann boundary). Functions
f and g belong to the spaces L2 (Ω) and L2 (ΓN ), respectively. The focus is on
linear second order elliptic equations with the Poisson equation as the main
model problem. Thus, the model problem is called the Poisson equation with
Dirichlet-Neumann boundary conditions and it is defined as

−∆u = f in Ω, (22)
u = 0 on ΓD , (23)
n · ∇u = g on ΓN . (24)

Naturally, one wants to find a solution u of a given problem. Sometimes, this

problem can be solved analytically, but usually, it is hard or even impossible to
find u in that way. Therefore, a numerical technique for solving a differential
equation should be used. The first step is to rewrite the differential equation
as a variational equation. Before doing that, the question that arises is how to
define a proper space.
1
Following the notation defined in [Ver13], the test space HD (Ω) is the Sobolev
1
space of functions from H (Ω) whose trace vanishes on the Dirichlet part of the
boundary:
1
HD (Ω) = {φ ∈ H 1 (Ω) : φ = 0 on ΓD }. (25)
1
As mentioned in a previous subsection, the space HD (Ω) is a Hilbert space.

8
 To get a variational formulation of the model problem, one should multiply
1
(22) by a test function v ∈ HD (Ω) and integrate over domain Ω. Furthermore,
using Green’s formula one obtains:
Z Z
f vdx = − ∆uvdx
Ω
Z Ω Z Z
= ∇u · ∇vdx − n · ∇uvds − n · ∇uvds
Ω Γ ΓN
Z Z D
= ∇u · ∇vdx − n · ∇uvds
Ω Γ
Z Z N
= ∇u · ∇vdx − gvds.
Ω ΓN

In the penultimate line, the integral on a Dirichlet part of the boundary vanishes
due to the assumption v = 0 on ΓD .
The solution of the Poisson’s equation is also the solution of the variational
formulation. The converse in usually not true, because the solution of the
variational formulation does not need to be two times differentiable. For this
reason, it is also called the weak formulation. It is called variational because
the function v can vary arbitrarily.
Thus, a variational formulation of the boundary value problem (22)-(24) is:
1
find u ∈ HD (Ω) such that
Z Z Z
a(u, v) = ∇u · ∇v = fv + gv = f (v) (26)
Ω Ω ΓN

1
for all v ∈ HD (Ω).
The following fundamental theorem should be introduced to analyze this
variational problem. The extension of the Riesz representation theorem (Theo-
rem A.1) to non-symmetric bilinear forms is called the Lax-Milgram theorem.
Theorem 2.4. (Lax-Milgram theorem) Given a Hilbert space V , a contin-
uous, coercive bilinear form a(·, ·) and a continuous linear functional f ∈ V 0 ,
there exists a unique u ∈ V such that

a(u, v) = f (v) ∀v ∈ V.

Proof. This proof can be found in [BS08], Theorem 2.7.7.

Finally, the existence and uniqueness of the solution of the given varia-
tional problem (26) is provided by the following corollary and the property that
1
HD (Ω) ⊂ H01 (Ω).
Corollary 2.2. Let H be a Hilbert space, V a subspace of H. Let a(·, ·) be a
coercive, continuous bilinear form on V , not necessarily symmetric. Then there
exists a unique solution of the variational problem: given f ∈ V 0 , find u ∈ V
such that
a(u, v) = f (v) ∀v ∈ V.
Proof. The proof can be found in [BS08], Corollary 2.7.12.

9
2.3 Conforming Finite Element discretization
The finite element method for the numerical solution of partial differential equa-
tions in two dimensions in its simplest form, i.e., conforming finite element
method, will be introduced in this section. The used literature is [Ver13], [BS08],
[Che05], [LB13], and [Cia91].
The finite element method is a process of constructing finite-dimensional
subspaces Vh ⊂ V , which are then called the finite element spaces, consisting
of piecewise polynomials over the finite element partition T h (i.e. mesh). The
finite element method is called conforming because the space Vh is a subspace
of the space V .
To begin with, one should transform the model problem into the variational
formulation and then search for an approximate solution in the space of con-
tinuous piecewise functions. As defined before, f ∈ V 0 . The finite element
approximation of variational problem means to find
uh ∈ Vh such that
a(uh , vh ) = f (vh ) (27)
for all vh ∈ Vh .
It is called the Galerkin method for approximating the solution. If the
bilinear form is symmetric, it is called the Ritz-Galerkin method. For both
symmetric and non-symmetric forms, there exists a unique solution uh that
solves (27). It is because subset Vh is also a Hilbert space, then the existence
and uniqueness are implied by using either the Riesz Representation theorem
or the Lax-Milgram theorem.
Furthermore, it can be shown how problem (27) is solved. The choice of a
proper basis of the space Vh is an important step. Let (wi )N i=1 be the basis of
space Vh , where N is the dimension of Vh . Then the solution of the problem
(27) can be written as
XN
uh = ϕi wi
i=1

where the vector (ϕ1 , ϕ2 , ..ϕN ) is the solution of the linear system
N
X
a(wi , wj )ϕi = f (wj ), 1 ≤ j ≤ N. (28)
i=1

The matrix A whose entries are a(wi , wj ) is called the stiffness matrix and the
vector b where bj = f (wj ) is called the load vector. A stiffness matrix is always
invertible and inherits properties of a bilinear form a(·, ·).
The question which arises is how to define subspace Vh of space V where V is
usually defined as one of the Hilbert spaces H01 (Ω), H 1 (Ω), H02 (Ω)... Therefore,
some properties are established in order to define it.
First of all, one should define triangulation T h over the set Ω. It is a decom-
position of the set Ω into a finite number of subsets K such that the following
conditions are satisfied. To start with, Ω can be defined as Ω = ∪K∈T h K. Fur-
thermore, for each K ∈ T h , the set K is closed and its interior K̊ is non-empty
and connected, and the boundary Γ is Lipschitz-continuous. Also, for each dis-
tinct K1 , K2 ∈ T h it holds K̊1 ∩ K̊2 = ∅. Since straight finite elements will be
considered, i.e., finite elements that are all polyhedra in Rn , n ∈ {2, 3}, another
condition should be added. Any face of any element K1 in the triangulation is

10
either a subset of the boundary Γ or a face of another element K2 in the trian-
gulation. In other words, an intersection of any two elements of triangulation
is either empty or a common face. After the triangulation process is done, one
obtains a finite-dimensional space Vh of functions defined over the set Ω.
Let PK be the space defined as PK = {vh |K : vh ∈ Vh }. The functions
vh ∈ Vh are piecewise polynomials such that for each K ∈ T h the space PK
consists of polynomials. Let {θj }nj=1 be the basis of the space PK , where n =
dim(PK ). The basis functions θj are called the shape functions. By taking a
linear combination of them and coefficients, one obtains a polynomial function
in PK for each polygon K. The linear functionals Φi (·), i = 1, .., n are specifying
the shape functions. A necessary compatibility condition for Φi (·), K and PK
is called unisolvency. Unisolvency is defined in the sense that for any given real
scalars ai , 1 ≤ i ≤ n, there exists a unique function p ∈ PK that satisfies

Φi (p) = ai , 1 ≤ i ≤ n.

Also, calculation of the shape functions is done by solving the linear system

Φi (θj ) = δij , i, j = 1, .., n.

Functionals determine both the local and the global properties of the finite
element space Vh . As mentioned before, functionals specify the shape functions
on each K ∈ T h but also, they can extend their influence on the behaviour of
these shape functions outside of the polygon K, on adjacent polygons or even
on the whole mesh T h .
Furthermore, the computation of the coefficients of the linear system (28)
is preferably done on a reference finite element. Since simplicial finite ele-
ments are discussed, one can consider the reference triangle {x ∈ R2 : x1 ≥
0, x2 ≥ 0, x1 + x2 ≤ 1} or the reference square [0, 1]2 , denoted by K̂. Also,
R1 (K̂) = span{1, x1 , x2 } is defined for the reference triangle or R1 (K̂) =
span{1, x1 , x2 , x1 x2 } for the reference square.
Let FK : K̂ → K be an affine diffeomorphism. Then every element K ∈ T h
is the image of the reference element K̂ under FK . Therefore, one can define set
−1
R1 (K) = {φ ◦ FK : φ ∈ R1 (K̂)}. The lowest order conforming finite element
space associated with T h is

S 1,0 (T h ) = {φ ∈ C(Ω) : φ|K ∈ R1 (K) for all K ∈ T h },

or if a boundary condition is used, then

1,0
SD (T h ) = {φ ∈ S 1,0 (T h ) : φ = 0 on ΓD }.

Finally, the finite element dicretization of a problem (26) can be written as:
1,0
Find uT h ∈ SD (T h ) such that:
Z Z Z
∇uT h · ∇vT h = f vT h + gvT h (29)
Ω Ω ΓN

1,0
holds for all vT h ∈ SD (T h ).
The Lax-Milgram theorem (Theorem 2.4) implies a unique solution of the prob-
lem (29).

11
2.4 A priori estimate of the error in the L2 norm of gra-
dient
The goal of this section is to introduce an a priori error estimate for conforming
finite element discretization, between the solution u of the original problem and
the solution uh of the discrete problem. The used literature is [GB05], [Cia91],
[BS08], [LB13], [ESW05] and [EG04].
An error estimate is an expression that represents an approximation to the
actual unknown error. There are two types of error estimation procedures, a
priori and a posteriori. A posteriori error estimators are discussed in Chapter
3. A priori error estimates are used to get useful information on the asymptotic
behavior of the discretization errors. They involve the unknown solution u, thus
they are not computable.
To start with, the basic equation for error is derived by subtracting a(uh , v)
from the weak formulation. Therefore one gets

a(u, v) − a(uh , v) = f (v) − a(uh , v)

a(u − uh , v) = f (v) − a(uh , v), ∀v ∈ V. (30)

Restricting the test functions to the space Vh and using (27) one obtains the
Galerkin orthogonality property, i.e., for any wh ∈ Vh it follows a(u − uh , wh ) =
0. Furthermore, assume that ku−vh k < ku−uh k with vh = uh +wh and wh 6= 0.
Then, by a straightforward calculation and using the Galerkin orthogonality one
obtains ku − vh k2 = ku − uh k2 + kwh k2 . Since wh 6= 0 and kwh k > 0, it follows
that ku − vh k2 < ku − uh k2 . Therefore, another property that arises is the best
approximation property

a(u − uh , u − uh ) = inf a(u − vh , u − vh ). (31)

vh ∈Vh

p bilinear form a(·, ·) is an inner product associated

As mentioned before, the
with the norm kukV = a(u, u), thus the above mentioned equation (31) can
be written as
ku − uh kV = inf ku − vh kV . (32)
vh ∈Vh

Equation (32) states that the finite element solution is the best approxima-
tion result with respect to k · kV , i.e., there is no better approximation result in
Vh . Because of the Galerkin orthogonality, the error u − uh is orthogonal to the
space Vh , i.e., uh is the projection of the solution u over Vh with the respect to
k · kV .
An upper bound of the best approximation error is the interpolation error.
Thus, it is more convenient to use the interpolation theory.
For introduction, the error estimate is established using the approximation
space P1 of piecewise linear functions.
Given a continuous function u, the Lagrangian piecewise linear interpolant
is denoted by Ih1 u. It satisfies Ih1 u(xi ) = u(xi ) at every vertex xi of the tri-
angulation. It is a well-defined function in Vh . Also, u is assumed to be two
times weakly differentiable. Taking vh = Ih1 u one wants to estimate the norm
of the interpolation error u − Ih1 u. The error can be represented as the sum
of elementwise error bounds which are calculated in the norm defined on every

12
K ∈ T h . The case for the L2 -norm of the gradient is
X
k∇(u − Ih1 u)k2L2 (Ω) = k∇(u − Ih1 u)k2L2 (K) . (33)
K∈T h

Thus, the problem of finding an estimate for the overall error is reduced to the
problem of evaluating local interpolation estimates. The idea is to map the local
interpolation error onto the reference element K̂, apply the estimate on K̂, and
transform back to K. One obtains that through the following bounds. The first
one is
h2
k∇(u − Ih1 u)k2L2 (K) ≤ 2 K k∇(û − Ih1 û)k2L2 (K̂) ,
|K|
h2
where |K| is the area of element K and |K| K
is the triangle aspect ratio. The
second bound is a special case of the Bramble-Hilbert lemma (Lemma A.2), a
lemma that is used as a general estimate for the interpolation error. One obtains

k∇(û − Ih1 û)kL2 (K̂) ≤ CkD2 (û − Ih1 û)kL2 (K̂) = CkD2 ûkL2 (K̂) .

The third bound is

h2K
kD2 ûk2L2 (K̂) ≤ 18h2K kD2 uk2L2 (K) .
|K|

One can find more information and intermediate steps in [ESW05, Chapter 1].
The following interpolation error estimates hold for any K ∈ T h .
Theorem 2.5. The interpolant Ih1 u satisfies the estimates

ku − Ih1 ukL2 (K) ≤ Ch2K kD2 ukL2 (K) ,

kD(u − Ih1 u)kL2 (K) ≤ ChK kD2 ukL2 (K) . (34)

Proof. More details in [LB13], Proposition 3.1.

Finally, with the help of the interpolation error, one finds that the error
satisfies the following theorem.
Theorem 2.6. [LB13][Theorem 4.8.] Let a(·, ·) be an inner product on V . The
finite element solution uh satisfies the estimate
X
k∇(u − uh )k2L2 (Ω) ≤ C h2K kD2 uk2L2 (K) .
K∈T h

Proof. Let vh = Ih1 u in the best approximation result (32). Furthermore, using
the interpolation error estimate (Theorem 2.5), one obtains

k∇(u − uh )k2L2 (Ω) ≤ k∇(u − Ih1 u)k2L2 (Ω)

X
= kD(u − Ih1 u)k2L2 (K)
K∈T h
X
≤ Ch2K kD2 uk2L2 (K) ,
K∈T h

thus the estimate is proven.

13
 Also, for the diameter hK , i.e., the length of the longest edge of K ∈ T h , one
gets hK ≤ h for all triangles K ∈ T h . Therefore one can derive a bound that is
independent of the local mesh width. Hence, from Theorem 2.6, it follows
X
k∇(u − uh )k2L2 (Ω) ≤ Ch2 kD2 uk2L2 (K) = Ch2 kD2 uk2L2 (Ω) . (35)
K∈T h

One can conclude that the gradient of the error tends to zero as the mesh
size h tends to zero and the convergence order is at least 1.
The error estimate (35) was established using approximation space P1 of
piecewise linear functions. For other cases, the following definitions are pre-
sented.
Let P̂ (K̂) be a polynomial space of dimension ns , {Φ̂i }ni=1
s
linear functionals
ns
and {θ̂}i=1 the local basis of P̂ (K̂). The local interpolation operator IK̂ v̂ ∈
P̂ (K̂) is defined as
ns
X
IK̂ v̂ = Φ̂i (v̂)θ̂i . (36)
i=1

The domain of the interpolation operator is V (K̂). Usually, it is assumed to be

of the form C s (K̂) for some integer s ≥ 0.
For all K ∈ T h , one must define a linear bijective mapping ψK : V (K) →
V (K̂) to get T h -based finite elements. In this case, K is defined as K =
−1
FK (K̂) and PK = {ψK (p̂); p̂ ∈ P̂ (K̂)}. Also, functionals are ΦK,i (p) =
−1
Φ̂(ψK (p)), ∀p ∈ PK and basis functions are θK,i = ψK (θ̂i ), 1 ≤ i ≤ ns . There-
fore, the local interpolation operator IK v ∈ PK is
ns
X
IK v = ΦK,i (v)θK,i . (37)
i=1

Due to the linearity of ψK , a property that is important for the analysis of the
interpolation error, it is
ns
X ns
X
IK̂ (ψK (v)) = Φ̂i (ψK (v))θ̂i = ΦK,i (v)ψK (θK,i ) = ψK (IK (v)). (38)
i=1 i=1

The global interpolation operator Ih v can be described elementwise using the

local interpolation operators (37), i.e.,
ns
X
h
∀K ∈ T , (Ih v)|K = IK (v|K ) = ΦK,i (v|K )θK,i .
i=1

Furthermore, one wants to estimate the norm of the interpolation error v − Ih v

with the interpolation operator defined as (37). The diameter of the largest ball
inscribed in K ∈ T h is denoted by ρK . The affine mapping FK is defined as
FK x̂ = BK x̂ + b where BK is a non-singular n × n matrix and b is a n vector.
If k · k is the matrix norm, then it holds
hK −1 h
kBK k ≤ and kBK k ≤ K̂ . (39)
ρK̂ ρK

Mappings ψK : V (K) → V (K̂) are set as ψK (v) = v̂ = v ◦FK . The following

lemma will be used in the proof of Theorem 2.7.

14
Lemma 2.1. Let s ≥ 0 and let 1 ≤ p ≤ ∞. There exists c such that, for all K
and w ∈ W s,p (K),
X 1 X
kDα ŵkLp (K̂) ≤ ckBK ks | det(BK )|− p kDα wkLp (K) ,
|α|=s |α|=s

X 1 X
−1 s
kDα wkLp (K) ≤ ckBK k | det(BK )| p kDα ŵkLp (K̂) ,
|α|=s |α|=s

where ŵ = w ◦ FK . For any positive real x in the case when p = ∞, it is set

1
x± p = 1.

Proof. This proof can be found in [EG04], Lemma 1.101.

Theorem 2.7. [EG04][Theorem 1.103.] Let a family of finite elements be given
by its reference cell K̂, the functionals {Φ̂i } and a space of polynomials P̂ (K̂).
Let 1 ≤ p ≤ ∞ and assume that there exists an integer k such that Pk ⊂ P̂ (K̂) ⊂
W k+1,p (K̂) ⊂ V (K̂). Let FK : K̂ → K be an affine bijective mapping and let
k
IK be the local interpolation operator on K defined in (37). Let 0 ≤ l ≤ k and
W l+1,p (K̂) ⊂ V (K̂) with continuous embedding. Then, setting σK = hρK K
, there
l+1,p
exists C > 0 such that, for all m ∈ {0, ..., l + 1} and ∀K, ∀v ∈ W (K),
X X
kDα (v − IK k
v)kLp (K) ≤ Chl+1−m
K
m
σK kDα vkLp (K) .
|α|=m |α|=l+1

k
Proof. Let IK̂ be the local interpolation operator on K̂ defined in (36). Let
l+1,p
ŵ ∈ W (K̂) and let F : W l+1,p (K̂) → W m,p (K̂) be a linear operator defined
k
as F ŵ = ŵ − IK̂ ŵ. Since W l+1,p (K̂) ⊂ V (K̂) with continuous embedding,
the linear operator F is continuous from W l+1,p (K̂) to W m,p (K̂) for all m ∈
{0, ..., l + 1}. Since l ≤ k, then Pl ⊂ P̂ (K̂) and, therefore, Pl is invariant under
k
IK̂ . This holds because of the property of the local interpolation operator that
k k
P̂ (K̂) is invariant under IK̂ , i.e., ∀p̂ ∈ P̂ (K̂), IK̂ p̂ = p̂. Hence, F vanishes on Pl .
As a consequence, and with using Deny-Lions Lemma (Lemma A.1), it follows
X X
kDα (ŵ − IK̂ k
ŵ)kLp (K̂) = kDα F(ŵ)kLp (K̂)
|α|=m |α|=m
X
= inf kDα F(ŵ + p̂)kLp (K̂)
p̂∈Pl
|α|=m
X
≤ kFkL(W l+1,p (K̂);W m,p (K̂)) inf kDα (ŵ + p̂)kLp (K̂)
p̂∈Pl
|α|≤l+1
X
α
≤ c inf kD (ŵ + p̂)kLp (K̂)
p̂∈Pl
|α|≤l+1
X
≤c kDα ŵkLp (K̂) .
|α|=l+1

Let v ∈ W l+1,p (K) and set v̂ = ψK (v) = v ◦ FK . From the property (38), it

15
 k k
follows [IK v] ◦ FK = IK̂ v̂. Using Lemma 2.1 yields
X 1 X
−1 m
kDα (v − IK
k
v)kLp (K) ≤ ckBK k | det(BK )| p kDα (v̂ − IK̂
k
v̂)kLp (K̂)
|α|=m |α|=m
1 X
−1 m
≤ ckBK k | det(BK )| p kDα v̂kLp (K̂)
|α|=l+1
X
−1 m
≤ ckBK k kBK kl+1 kDα vkLp (K)
|α|=l+1
X
−1 m
≤ c(kBK kkBK k) kBK kl+1−m kDα vkLp (K) .
|α|=l+1

With property (39) one can conclude the proof.

The following theorem is used to establish the error bounds with higher-order
approximation space, i.e., Pm , m ≥ 2.
Theorem 2.8. Using a higher-order finite element approximation space Pm or
Qm with m ≥ 2 leads to the higher-order convergence bound

k∇(u − uh )kL2 (Ω) ≤ Cm hm kDm+1 ukL2 (Ω) .

Proof. More details in [ESW05], Theorem 1.21.

In other words, one gets m-th order convergence as long as the regularity of
the target solution is good enough. Also, remark that kDm+1 ukL2 (Ω) < ∞ if
and only if the (m + 1)st generalized derivatives of u are in L2 (Ω).

2.5 Neumann and Dirichlet cases

The mixed Dirichlet-Neumann boundary conditions are presented in this section
following [EG04] and [BS08].
A model problem (22)-(24), i.e., the Poisson equation with Dirichlet-Neumann
boundary conditions, is considered on a bounded, polygonal, and connected do-
main Ω. As mentioned before, the boundary is Γ = ΓD ∪ ΓN . A Dirichlet
condition is defined on ΓD and a Neumann condition on ΓN . One can obtain
homogeneous Dirichlet condition (23) from the case when the Dirichlet condition
is non-homogeneous. To get that, assume that ΓD is smooth enough. Then it
1 1
is possible to define H 2 (ΓD ). One should remark that every function in H 2 (Γ)
is the trace of a function in H 1 (Ω), which follows from the statement (i) of
1
the Theorem A.3. Furthermore, for all g ∈ H 2 (ΓD ), there exists an extension
1
g̃ ∈ H 2 (Γ) with properties g̃|ΓD = g and kg̃k 12 ≤ ckgk 12 uniformly
H (Γ) H (ΓD )
in g. Using the lifting of g̃ in H 1 (Ω) (Corollary A.1) one can assume that the
Dirichlet condition is homogeneous. Therefore, the boundary conditions are de-
fined as in (23)-(24). Taking the solution and the test function in the Sobolev
1
space HD (Ω), one obtains the weak formulation (26) by multiplying Poisson’s
equation by a test function, integrating over Ω and then integrating by parts.
One concludes:

16
Proposition 2.3. Let ΓD ⊂ Γ, assume meas(ΓD ) > 0, and set ΓN = Γ \ ΓD .
Let g ∈ L2 (ΓN ). If u solves (26), then −∆u = f almost everywhere in Ω, u = 0
almost everywhere on ΓD and n · ∇u = g almost everywhere on ΓN .

Proof. More details in [EG04], Proposition 3.6.

Thus, mixed Dirichlet-Neumann boundary conditions formulate the weak
problem. The equation and the boundary conditions are satisfied almost every-
where by the weak solution.
In the case when a model problem (22)-(24) has homogeneous Dirichlet and
Neumann boundary conditions, the equations (22) and (24) are defined as u =
0 on ΓD and n · ∇u = 0 on ΓN , respectively. Then, the variational problem is
defined as:
Proposition 2.4. Let u ∈ H 2 (Ω) solve Poisson’s equation (22) (this implies
f ∈ L2 (Ω)) with homogenous Dirichlet and Neumann boundary conditions.
1
Then u can be characterized via u ∈ HD (Ω) satisfies a(u, v) = (f, v) for every
1
v ∈ HD (Ω).
Proof. The proof can be found in [BS08], Proposition 5.1.7.
1
One should remark that the boundary term vanishes for v ∈ HD (Ω) because
either v or n · ∇u is zero on any part of the boundary.
Proposition 2.5. [BS08][Proposition 5.1.9] Let f ∈ L2 (Ω) and suppose that
u ∈ H 2 (Ω) solves the variational equation. Then u solves Poisson’s equation
(22) with homogeneous Dirichlet and Neumann boundary conditions.
1
Proof. The Dirichlet boundary condition on u follows since u ∈ HD (Ω). Using
∞ 1
Green’s formula (Definition 16), with v ∈ C0 (Ω) ⊂ HD (Ω) and a(u, v) = (f, v),
one gets
Z Z
(f + ∆u)vdx = (f, v) − ∇u · ∇vdx = (f, v) − a(u, v) = 0.
Ω Ω

Since C0∞ (Ω) is 2

dense in L (Ω) (Corollary 2.1), the differential equation (22) is
2
satisfied in L (Ω). Also, Green’s formula then implies that

Z Z Z
0 = (f, v) − a(u, v) = (−∆u)vdx − ∇u · ∇vdx = n · ∇uvds
Ω Ω Γ

1
for all v ∈ HD (Ω). The Neumann boundary condition on u follows if v|ΓN can be
chosen arbitrarily with v ∈ HD 1
(Ω). Let U and U 0 be defined as in the Definition
11, and let Γ fulfills conditions a) and b) of the Definition 11. In other words,
in a neighbourhood of x, Ω is below the graph of ϕ and the boundary Γ is the
graph of ϕ. Since it is assumed that Γ is Lipschitz boundary and ΓD a closed
subset of it, this means that for any point x ∈ ΓN there is a neighbourhood
U of x such that Γ ∩ U can be written as a graph of a Lipschitz function ϕ.
Therefore, Γ ∩ U is defined as Γ ∩ U = {y = (y 0 , yn ) ∈ U : yn = ϕ(y 0 )}. Also,
one has that

Ω ∩ U = {y = (y 0 , yn ) ∈ U : yn < ϕ(y 0 )}.

17
The boundary integral over Γ ∩ U can be written as
Z Z p
n · ∇uvds = (n · ∇uv)(y 0 , ϕ(y 0 )) 1 + |∇ϕ(y 0 )|2 dy 0 .
Γ∩U U0

Let w ∈ C0∞ (U 0 ) and set

t
v(y 0 , t + ϕ(y 0 )) = w(y 0 )(1 + ) ∀y 0 = (y1 , y2 , ..., yn−1 ) ∈ U 0 , −ai < t < 0
ai
1
where v is defined to be zero elsewhere. Then it follows that v ∈ HD (Ω) and
Z Z Z p
0= n·∇uvds = n·∇uvds = w(y 0 )n·∇u(y 0 , ϕ(y 0 )) 1 + |∇ϕ(y 0 )|2 dy 0 .
Γ Γ∩U U0
p
The L∞ function 1 + |∇ϕ(y 0 )|2 is bounded below by 1 and w was chosen
arbitrary. Therefore, one concludes that n · ∇u|Γ∩U = 0. Since ΓN is covered
by neighborhoods U , one concludes that the Neumann condition holds on all of
ΓN .
Choose gN ∈ L2 (ΓN ) and gD ∈ H 1 (Ω). Although gD is defined on Ω, only
the trace on ΓD is used, with the same symbol. In other words, there exists a
lifting gD of gD in H 1 (Ω). In the case when a model problem (22)-(24) has a
non-homogeneous Dirichlet and Neumann boundary conditions, the equations
(23) and (24) are defined as u = gD on ΓD and n · ∇u = gN on ΓN , respec-
tively. Define the space Hg1D ,D (Ω) = {φ ∈ H 1 (Ω) : φ = gD on ΓD }. Remark
that the Dirichlet condition u = gD on ΓD is built into the definition of the
space Hg1D ,D (Ω). On the other hand, functions in the space HD 1
(Ω) are zero on
the Dirichlet part of the boundary. Any function u that satisfies Poisson’s equa-
tion (22) with non-homogeneous Dirichlet and Neumann boundary conditions,
is also a solution of the following continuous problem: find u ∈ Hg1D ,D (Ω) such
that
Z
a(u, v) = (∇u, ∇v) = (f, v) + gN (s)v(s)ds,
ΓN
1
for all v ∈ HD (Ω). The solution space and the test space are different in this
variational equation. Since Ω is a polygonal and bounded set, boundaries ΓD
and ΓN also have a polygonal form. Then one can triangulate Ω in a way that the
edges of ΓD and ΓN are composed of side faces of grid cells. Let P1 be the space
of the continuous, piecewise, linear finite elements of this triangulation with the
property vh (Vi ) = gD (Vi ) for all corner points Vi ∈ ΓD . Let P1,0 be the space
of the continuous, piecewise, linear finite elements of this triangulation with the
property vh (Vi ) = 0 for all corner points Vi ∈ ΓD . The finite element method is
defined by: find uh ∈ P1 so that
Z
a(uh , vh ) = (∇uh , ∇vh ) = (f, vh ) + gN (s)vh (s)ds ∀vh ∈ P1,0 .
ΓN

If gN 6= 0, the finite element equation has an additional contribution on the

right-hand side of the system of equations that needs to be assembled. Therefore
only the test functions that don’t disappear on ΓN are affected.

18
 The convergence estimation starts as usual with the error equation, which is
obtained by subtracting the finite element equation from the continuous equa-
tion
(∇(u − uh ), ∇vh ) = 0 ∀vh ∈ P1,0 .
Since Ih (u − uh ) ∈ P1,0 , by applying this function in the error equation, Ih uh =
uh and Cauchy-Schwarz inequality, it follows:

k∇(u − uh )k2L2 (Ω) = (∇(u − uh ), ∇(u − uh ) − ∇(Ih (u − uh )))

= (∇(u − uh ), ∇(u − Ih uh ))
≤ k∇(u − uh )kL2 (Ω) k∇(u − Ih uh )kL2 (Ω) .

Due to the change of the boundary conditions, the solution of the continuous
problem is not in H 2 (Ω). Therefore the interpolation has a reduced order of
convergence. In the case d = 2 and if the change of the boundary condition
takes place on a straight line, the following error estimates are optimal:
1
k∇(u − uh )kL2 (Ω) = c(u)h 2 , ku − uh kL2 (Ω) = c(u)h.

Thus, only half of the order of convergence as with pure Dirichlet problems
exists.

19
3 A Posteriori Error Estimators
3.1 Goals
The used literature in this section is [Ver13], [EG04], [ESW05] and [SS05].
A posteriori error estimates are one type of error estimation procedure. Their
goal is to evaluate the error u − uh in terms of known data only, i.e., the size of
the mesh cells, the approximation solution and the problem data. A posteriori
error estimates are one of the main tools in the adaptive finite element method
which is the main topic of this thesis.
The question which arises is when one should use adaptive methods. The
motivation for using them is the fact that many physical problems of interest
have singularities. Singularities appear from re-entrant corners or domains,
interior or boundary layers, and sharp moving fronts. The overall accuracy of
numerical approximations decreases when local singularities increase. The idea
to fix further decreasing of approximations is to put more grid points where
singularities occur. However, those regions should be carefully identified and
one should find a good balance between refined and unrefined regions. Adaptive
methods use information from earlier computations to locally refine the mesh.
Therefore they automatically adjust themselves to improve approximations.
Another question that appears is how to establish good estimates of the
accuracy of the computed solution. The answer is: with a posteriori error esti-
mates. A priori error estimates are not sufficient in this case. They essentially
provide only asymptotic information. On the other hand, a posteriori error
estimates provide reliable upper and lower bounds for the error. A posteriori
error estimates can be divided into several categories: residual estimates, hier-
archical error estimates, averaging methods,... Here, the focus is on the residual
estimates. They estimate the error of the numerical solution by a norm of its
residual. In particular, a residual a posteriori error estimator ηR,K is used for
this purpose. It is extracted a posteriori from the computed solution. The
adaptive method used in this thesis is based on the residual a posteriori error
estimator ηR,K . Therefore the goals are to define it and to obtain upper and
lower bounds for the error. An important requirement is that ηR,K should be
cheaper to compute than to compute the numerical solution. If ηR,K provides
a global upper bound on the error, then ηR,K is reliable and the accuracy of
the solution is below a tolerance. If ηR,K provides a lower bound for the local
error then the estimator ηR,K is effective when it is used to drive an adaptive
refinement process. An adaptive process will provide successive meshes that are
correctly refined in the presence of the singularities.
Another question is how to correctly refine meshes. The location of the nodes
depends on the geometry of the parental element, therefore on the structure of
the initial coarse triangulation. First, the elements of the mesh have to be
marked with the marking strategy. Then, they are either refined or coarsened
by local refinement or coarsening based on error estimators. One can use the
mesh smoothing process to optimize the placement of nodes. The goal is to
obtain an optimal mesh where the number of unknowns is as small as possible
to keep the error below the tolerance. This can be costly and one should try to
minimize the cost. To sum up, an adaptive mesh-refinement process is presented
through the general adaptive algorithm (Algorithm 3.1).

20
3.2 Adaptive algorithm
This section presents the general adaptive algorithm. The used literature is
[Ver13], [D9̈6] and [SS05].
The outline of the algorithm is, starting from an initial triangulation T h0 ,
a sequence of triangulations T hk is obtained for k = 1, 2, ... until the estimated
error is below the given tolerance  using a minimal amount of work. Following
[Ver13, Algorithm 1.1], a general algorithm for stationary problems is defined
as follows:
Algorithm 3.1 (General adaptive algorithm). Given the data of a partial dif-
ferential equation and a tolerance  as an input values, the goal is to provide a
numerical solution with an error less than .
(1) Construct an initial coarse mesh T h0 representing sufficiently well the ge-
ometry and data of the problem; set k = 0.
(2) Solve the discrete problem associated with T hk .
(3) For every element K ∈ T hk compute the a posteriori error indicator.
(4) If the estimated global error is less than  stop, otherwise decide which el-
ements have to be refined and construct the next mesh T hk+1 . Increase k by 1
and return to step (2).
For points (1) and (2), one must specify a discretization method and a
solution method for the discrete problems. It is assumed that the exact solutions
of the finite dimensional problems can be obtained. For point (3) one estimates a
posteriori error indicator for every element K. The upper estimate presents that
error indicator can be used as a reliable stopping criterion for the algorithm,
and the lower estimate proposes that an unnecessary work can be avoided.
For step (4), having the set of local error estimates for all K, one needs an
algorithm that uses this information to construct the next mesh T hk+1 . An
required algorithm, i.e., refinement strategy, determines which elements have
to be refined or coarsened and how to do that. The way these triangles are
marked with marking strategies influences the efficiency of the whole Algorithm
3.1. Therefore, to make an algorithm operative, one needs to specify all of the
above points.
After providing all the necessary data, the goal is to prove that an adap-
tive algorithm gives a sequence of discrete solutions that converges to the true
solution of the differential equation. Chapter 4 proves the convergence of an
adaptive algorithm.

3.3 Residual a posteriori error estimates

This section introduces one type of a posteriori error estimates called residual
a posteriori error estimates. The residual estimates provide upper and lower
bounds for the error of the discrete solution. The steps for obtaining bounds,
as well as the residual a posteriori error indicator ηR,K are presented. The used
literature in this section is [Ver13], [GB05] and [Che05].
1,0
Let u ∈ HD 1
(Ω) and uT h ∈ SD (T h ) be the solutions of (26) and (29).
The starting point is the error equation (30). Considering the model problem
(22)-(24) and its variational formulation (26), (30) can be rewritten as
Z Z Z Z
1
∇(u − uT h ) · ∇v = fv + gv − ∇uT h · ∇v, ∀v ∈ HD (Ω).
Ω Ω ΓN Ω

21
The right-hand side implicitly defines the residual of uT h . It is an element of
1 1
the dual space of HD (Ω), denoted by R(v), ∀v ∈ HD (Ω).
There are several necessary steps needed to derive the a posteriori error
1,0
estimates. To start with, since SD (T h ) ⊂ HD
1
(Ω), the Galerkin orthogonality
1,0 h
of the error holds. For all wT h ∈ SD (T ) it follows
Z
∇(u − uT h ) · ∇wT h = 0, i.e., R(wT h ) = 0.
Ω

Furthermore, one wants to show the equivalence of the norm of the error
and a dual norm of the residual. Defining the L2 -norm of the gradient of v as
Z
1
k∇vkL2 (Ω) = sup ∇v · ∇w, (40)
w∈HD1 (Ω)\{0} k∇wkL2 (Ω) Ω

it follows
R(w)
k∇(u − uT h )kL2 (Ω) = sup . (41)
1 (Ω)\{0} k∇wkL2 (Ω)
w∈HD

The supremum term in (41) is equal to the norm of the residual in the dual
1 1
space of HD (Ω). Therefore this equality implies that the norm in HD (Ω) of the
1
error is equal to the norm of the residual in the dual space of HD (Ω).
Before proceeding to the next step, the following definitions and notations
are introduced. Let ε be the set of all edges associated with a partition T h .
Furthermore, let εΩ denote the set of all interior edges e and εΓ the set of all
edges on Γ. Similarly, let εΓD be the set of edges e on ΓD and εΓN be the set
of edges e on ΓN . The union of all edges in ε is denoted by Λ. With every edge
e a unit vector ne is associated. For e ∈ εΓ it is the outward unit normal to Γ.
For e ∈ εΩ the direction of it is associated with the definition of jumps across
e. For any piecewise continuous function v it holds

Je (v)(x) = lim v(x − tne ) − lim v(x + tne ), ∀x ∈ e.

t→0+ t→0+

The next step is to establish an L2 -representation of the residual. Residual

estimates contain weighted L2 -norms of element and edge residuals. Element
residuals are defined as r|K = f + ∆uT h on every K ∈ T h . Edge residuals are
defined as
−Je (ne · ∇uT h ), if e ∈ εΩ ,


j|e = g − ne · ∇uT h , if e ∈ εΓN , (42)
if e ∈ εΓD .

0,


22
 In the following lines, nK denotes the unit outward normal to the element K.
Also, remark that ∆uT h = 0 on all triangles. Integration by parts element-wise
and rearranging terms leads to
Z Z (Z Z )
X
R(w) = fw + gw + ∆uT h w − nK · ∇uT h w
Ω ΓN K ∂K
K∈T h
X Z X Z
= (f + ∆uT h )w + (g − ne · ∇uT h )w
K∈T h
K e∈ εΓN e

X Z
− Je (ne · ∇uT h )w
e∈ Ωε e
Z Z
1
= rw + jw, ∀w ∈ HD (Ω). (43)
Ω Λ

The next step is an error estimate for an interpolant IT h w. First of all, the
set of all vertices associated with a partition T h is denoted by N and sets NΩ ,
NΓN are defined for vertices analogously as they are defined for the sets of edges
εΩ , εΓN . The interpolant is defined as
X
IT h w = wKn λn .
n∈NΩ ∪NΓN

Here, wKn is the average of w on Kn , i.e., the average of w on the union of all
elements having n as a vertex. λn ∈ S 1,0 (T h ) is a nodal shape function of a
vertex n ∈ N that gives values λn (n) = 1, λn (x) = 0, ∀x ∈ N \ {n}. Utilizing
the Galerkin orthogonality condition into (43) and choosing wT h = IT h w for
1
w ∈ HD (Ω) results in
X Z XZ
R(w) = r(w − IT h w) + j(w − IT h w). (44)
K∈T h
K e∈ ε e

Furthermore, the shape regularity is another condition that partition T h has

to satisfy, i.e., for any element K, the ratio ρhKk is bounded independently of K.
The shape parameter of T h , defined as σT h = maxK∈T h ρhKk , must be bounded
uniformly with respect to all partitions derived by local or global refinement.
Let Ke n and K e e denote the unions of all elements that share at least a vertex with
a given element K or edge e, respectively. Let c1 and c2 be the interpolation
constants that depend, for the used model problem, on the shape parameter.
According to the interpolation theory, one has
kv − IT h vkL2 (K) ≤ c1 hK k∇vkL2 (K
e ), (45)
n
1
kv − IT h vkL2 (e) ≤ c2 he k∇vkL2 (K
2
e ). (46)
e

Applying the Cauchy–Schwarz inequality element-wise on (44) and using results

of interpolation theory (45, 46) yields
X X
R(w) ≤ krkL2 (K) kw − IT h wkL2 (K) + kjkL2 (e) kw − IT h wkL2 (e) (47)
K∈T h e∈ ε
X X 1
≤ krkL2 (K) c1 hK k∇wkL2 (K + kjkL2 (e) c2 he2 k∇wkL2 (K
e ) . (48)
e n) e
K∈T h e∈ ε

23
Using the above estimates, the Cauchy-Schwarz inequality for sums and the
properties of the shape regularity leads to
( ) 12
X X
R(w) ≤ max{c1 , c2 }σT h k∇wkL2 (Ω) h2K krk2L2 (K) + he kjk2L2 (e) .
K∈T h e∈ ε
Finally, defining c∗ = max{c1 , c2 }σT h and combining the above estimates
with (41) one can bound the error k∇(u − uT h )kL2 (Ω) from above:
( ) 12
X X
k∇(u − uT h )kL2 (Ω) ≤ c∗ h2K krk2L2 (K) + he kjk2L2 (e) . (49)
K∈T h e∈ ε
The upper bounds are global concerning the domain. The right-hand side
of equation can be used as an a posteriori error indicator. To evaluate integrals
for equation (49), one should approximate integrals by appropriate quadrature
formulas or f and g should be approximated by simpler polynomial functions.
Both approaches are often equivalent for generating a posteriori estimators.
Furthermore, one wants to bound the error k∇(u − uT h )kL2 (Ω) from below.
The lower bounds are local because the error indicator assigned to an element is
bounded by the error on the given element and neighboring elements. To start
with, one should replace functions f and g by their mean values
Z Z
1 1
fK = f and ge = g.
|K| K he e
The next important step for deriving the a posteriori error estimates is in-
troducing the local cut-off functions and inverse estimates for them. One can
see definitions and intermediate steps for obtaining a lower bound in [Ver13,
Chapter 1]. Let c∗ be a constant that only depends on the shape parameter of
T h and Ke a union of all elements sharing an edge with K. The final results,
presented in [Ver13, Theorem 1.5], are a posteriori error indicator
(
1 X
ηR,K = h2K kfK + ∆uT h k2L2 (K) + he kJe (ne · ∇uT h )k2L2 (e)
2
ε
e∈ K,Ω
) 12
X
+ he kge − ne · ∇uT h k2L2 (e) , (50)
ε
e∈ K,ΓN

and the estimates

(
X X
k∇(u − uT h )kL2 (Ω) ≤ c∗ 2
ηR,K + h2K kf − fK k2L2 (K)
K∈T h K∈T h

) 12
X
+ he kg − ge k2L2 (e) , (51)
ε
e∈ ΓN

24
and
(
X
ηR,K ≤ c∗ k∇(u − uT h )k2L2 (Ke ) + h2K 0 kf − fK 0 k2L2 (K 0 )
K 0 ⊂Ke
) 12
X
+ he kg − ge k2L2 (e) , (52)
ε
e∈ K,ΓN

The upper bound in the estimate (51) shows that in the case when an equation
where error indicator ηR,K is less than tolerance  implies that the true error
is also less than the tolerance up to the multiplicative constant c∗ . The error
indicator is then called reliable. On the other hand, by providing a lower bound
in the estimate (52), the error indicator is efficient, i.e., an equation where error
indicator ηR,K is greater than tolerance  implies that the true error is also
greater than the tolerance up to the multiplicative constant c∗ .

3.4 Ways for refining the mesh

This section shows how to obtain adaptive discretizations using the following
processes: refinement, coarsening, and smoothing of meshes. The key ingredi-
ent is an a posteriori error indicator. The processes are introduced using the
literature [Ver13], [D9̈6], [Ver94], [CM10], [SS05], [BS97] and [Che05].
To start with, the idea of the mesh refinement is the following. Given an error
indicator for all K ∈ T hk , one needs an algorithm that uses this information
to construct the next mesh T hk+1 . There are two steps to obtain this task.
First, the marking strategy is a problem of selecting elements to be refined.
In the second step, the refinement rules determine the construction of a new
subdivision. Considering Algorithm 3.1, one wants to make the number of
iterations as small as possible since a discrete problem has to be solved in every
iteration. Thus, the marking strategy should select sufficiently many mesh
elements for refinement in each iteration, but not too many elements than is
needed to reduce the error below the prescribed tolerance. Among all marking
strategies, it is heuristically assumed that the mesh is optimal when the local
error is equally distributed for all elements of the mesh. That means that the
local error indicators are equally distributed since the true error is not known.
Furthermore, elements that will be marked for refinement are elements with a
large local error indicator. On the other hand, elements with a very small local
error indicator are unchanged or can be coarsened.
One of the most used marking strategies in adaptive finite element methods
is the maximum strategy. Define β as a threshold where 0 < β < 1. Let
ηmax = maxK∈T hk ηR,K and mark K if ηR,K ≥ βηmax . Then, marked K should
be placed in a set Te hk , a subset of marked elements that should be refined.
When β is close to 0, one refines globally and an unnecessary amount of degrees
of freedom is obtained. When β is close to 1 and the error is not equally
distributed, one marks a small number of triangles. Thus, to fix this, one should
do more iterations, which is costly. Therefore this strategy won’t be efficient
so β is typically set at the value 0.5. This strategy is cheap and often yields
satisfactory results.
Another widely used marking strategy is the equilibrium strategy that is
presented with the following algorithm.

25
Algorithm 3.2. [Ver13, Algorithm 2.2] Given a partition T hk , error indicators
ηR,K for the elements K ∈ T hk , and a threshold β ∈ (0, 1). The goal is to seek
for a subset Te hk of marked elements that P
should be refined.
2
. Set T hk = 0 and Te hk = ∅.
P
(1) Compute ΘT hk = K∈T hk ηR,K
(2) If T hk ≥ βΘT hk return Te hk ; stop. Otherwise go to step (3).
P
(3) Compute ηemax = maxK∈T hk \Tehk ηR,K .
(4) For all elements K ∈ T hk \ Te hk check whether ηR,K = ηemax . If this is
the case, put K in Te hk and add ηR,K
2
P
to T hk . Otherwise skip K. When all
elements have been checked, return to step (2).

At the end of the algorithm, the set Te hk satisfies:

X X
2 2
ηR,K ≥β ηR,K .
hk
K∈T h
e k K∈T

A small value of β leads to a small set Te hk and a large value of β leads to a

large set Te hk , i.e., almost all elements are marked.
After the decision which elements should be refined with marking strategies,
one has to decide based on the element geometry, how to perform the refinement.
The refinement process has two stages. First, one wants to get a subset Te hk
of T hk . For that, so-called red-refinement is used (regular refinement). In the
second stage, one uses a red-green-blue-refinement of further elements to avoid
hanging nodes (irregular refinement). Hanging nodes are vertices where the
triangulation condition that two triangles share at most a common edge or a
common vertex is violated.
Given K ∈ T hk and edges E1 , E2 , E3 of K where E1 is the longest edge,
a red-refinement of K is derived by dividing K into four new sub-triangles.
That is obtained by joining the midpoints of its edges. They are similar to the
parent element, thus they have the same angles and the shape parameter of the
elements doesn’t change. A blue-refinement of K is derived by dividing K into
three sub-triangles by joining the midpoint of E1 with the opposite vertex and
the midpoint of E2 or E3 . A green-refinement of K is derived by dividing K
into two sub-triangles by joining the midpoint of the longest edge E1 with the
opposite vertex. The refinements are illustrated in Figure 1.

Figure 1: Red, blue and green refinement [CM10, Figure 6.1.]

Too acute or too obtuse triangles and hanging nodes are avoided using the
following rules [Ver94].
(1) A triangle having three hanging nodes uses red-refinement.
(2) A triangle having two hanging nodes uses blue-refinement, if one of them
lies on the longest edge of the triangle; otherwise it uses red-refinement.

26
(3) A triangle having one hanging node uses green-refinement, if the hanging
node lies on its longest edge; otherwise uses blue-refinement.
The rules (2) and (3) can lead to new hanging nodes.
An alternative strategy to the red-refinement is the marked-edge bisection.
Then one bisects triangles only by dividing a marked edge. The following rules
for it are established. The coarsest mesh is formed in a way that the longest
edge of any element is also the longest edge of the adjacent element except if it
is a boundary edge. Only the longest edge of every element in the coarsest mesh
is marked. The element is bisected by connecting the midpoint of its marked
edge with the opposite vertex. Then, its two unmarked edges are marked edges
of the two new triangles. Figure 2 shows the process, where marked edges are
labeled with •.

Figure 2: Marked edge bisection [Ver13, Figure 2.5.]

The next used technique is the mesh coarsening. It is the inverse process
of refinement. As in the refinement process, the main goal for coarsening is
to equally distribute the local errors. The idea for the coarsening process is
to gather all elements that were created during the refinement such that their
parents build a corresponding refinement patch. After the coarsening process,
refinement edges are back at their original position on parent elements. Fur-
thermore, an element will be coarsened if all involved neighbor elements are
marked for coarsening. In contrast, during refinement, it is possible that by
bisecting an element, an unmarked element is also refined in order to keep the
mesh conforming. Therefore, in the adaptive method, this assures that if the
local error indicator is not small enough no element is coarsened. On the other
hand, marked elements with a large local error indicator are refined. One should
remark that after the coarsening process, the local error should not be larger
than the tolerance used for refinement. If it is, then in the next iteration the el-
ements would be refined again and the desired results might never be obtained.
The negative aspect of coarsening is that additional errors can be produced and
some information can be lost during the process. The suggestion to avoid the
loss of information is to delay the mesh coarsening until the final iteration of
the adaptive procedure. Furthermore, one can use the marking strategies de-
fined above to mark mesh elements for coarsening. The following algorithm is
a modification of the maximum strategy defined for element’s marking and it is
best suited for the marked edge bisection.

27
 Let T hk be a current partition, K an element of it and n ∈ N a vertex of
it. K has refinement level l if it is obtained by subdividing an element of the
coarsest partition l times. If there exists a vertex of K which is not a vertex of
its parent triangle K 0 , then it is called the refinement vertex of K. Finally, a
vertex n ∈ N and patch Kn are resolvable if n is the refinement vertex of all
elements in Kn and if all elements in Kn have the same refinement level.
Algorithm 3.3. [Ver13, Algorithm 2.4] Given a partition T hk , error indicators
ηR,K for all elements K ∈ T hk , and parameters 0 < β1 < β2 < 1. The goal
is to find subsets Tc and Tr of elements that should be coarsened and refined,
respectively.
(1) Set Tc = ∅, Tr = ∅ and compute ηmax = maxK∈T hk ηR,K
(2) For all K ∈ T hk check whether ηR,K ≥ β2 ηmax . If this is the case, put K
in Tr .
(3) For all vertices n ∈ N check whether n is resolvable. If this is the case and
if maxK⊂Kn ηR,K ≤ β1 ηmax , put all elements contained in Kn into Tc .
One can notice that this algorithm simultaneously refines and coarsen ele-
ments of the current partition and in that way, it constructs the partition of the
next level.
The next technique is the mesh smoothing. This method doesn’t change
the topology in the triangulation as the above two methods do and that is
the critical ability to the success of an adaptive method. Therefore, a mesh
smoothing algorithm cannot be the basis of an effective adaptive method. In-
stead, it complements mesh-refinement methods. After creating a mesh with
mesh-refinement methods where a proper density of mesh vertices and topology
is obtained, one uses a smoothing algorithm to improve the quality of a mesh.
That means that the vertices move a little but the number of elements and the
adjacency stay unchanged. The result of a vertex movement is error reduction.
Even small movements of the vertices can result in a high error reduction. This
method can be used when further refinement is impossible. In this section, the
focus is on triangular meshes, i.e., all partitions consist of triangles.
The mesh-smoothing algorithms follow a strategy similar to Gauss-Seidel
which optimizes a quality function q. A quality function q assigns a non-negative
number to every element. One can see the different choices for q and their prop-
erties in [BS97]. A larger value of q shows a better quality. Given a triangula-
tion T hk one wants to obtain a new improved triangulation Te hk with the same
topology as T hk such that the error is minimized. One seeks the solution of the
optimization problem
min q(K)
e > min q(K).
e Tehk
K∈ K∈T hk

Therefore, several iterations of a strategy similar to Gauss-Seidel should be

performed. There one should go through the vertices and locally optimize the
position of a single vertex while having all others fixed. In other words, for
every vertex n ∈ T hk , fix the vertices of the boundary of Kn and find a new
vertex n
e inside Kn such that
min q(K)
e > min q(K).
K⊂K
e ñ K⊂Kn

This is the local optimization problem and its solution depends on the choice
of the quality function q.

28
 Now that techniques for refining the mesh are introduced, the following ex-
ample from [Ver13, Page 3] and [Che05, Example 6.1.] is presented to show the
potential of the adaptive refinement method. The domain Ω is defined as a cir-
cular segment with radius one, angle 32 π and center at the origin. The problem
(22)-(24) is considered on Ω. Thus, a function u is harmonic in the interior of
Ω, it vanishes on the straight parts of the boundary ∂Ω and it has a normal
derivative 23 sin( 23 γ) on the curved part of the boundary. In terms of polar co-
2
ordinates, one gets that the exact solution is u = r 3 sin( 32 γ). Furthermore, one
calculates the Ritz projections uT h of u onto the spaces of continuous piecewise
linear finite elements. Those spaces are associated with the two triangulations
shown in Figure 3.

Figure 3: Uniform and adaptive triangulations [Che05, Figure 6.8.]

The left triangulation represents the triangulation obtained by uniform re-

finement and the right one is obtained by an adaptive refinement of meshes.
The right triangulation is constructed from an initial triangulation T h0 by us-
ing Algorithm 3.1. based on the error estimator ηR,K . Specifically, the adaptive
refinement strategy of Algorithm 3.1 is used. A triangle K ∈ T hk is refined if
ηR,K ≥ 0.5 maxK 0 ∈T hk ηR,K 0 . The right triangulation in Figure 3 is obtained
such that a triangle K ∈ T hk is divided into four smaller triangles by connecting
the midpoints of its edges if ηR,K ≥ 0.5 maxK 0 ∈T hk ηR,K 0 . The midpoint of an
edge having its two endpoints on ∂Ω is projected onto ∂Ω. The left triangu-
lation is also constructed from initial triangulation T h0 which is composed of
three right-angled isosceles triangles with short edges of unit length. The left
triangulation in Figure 3 is obtained by five uniform refinements of T h0 . In each
refinement step, as described above, every triangle K ∈ T hk is divided into four
smaller triangles by connecting the midpoints of its edges. Again, the midpoint
of an edge having its two endpoints on ∂Ω is projected onto ∂Ω. For the final
result, there are 3072 triangles and 1552 unknowns in uniform refinement, while
there are 298 triangles and 143 unknowns in adaptive refinement. Also, the
k|∇(u−uT h )|kL2 (Ω)
relative error k|∇u|kL2 (Ω) is lower with adaptive refinement. Therefore,
one can see the advantages of the adaptive refinement strategy.

29
4 Convergence of an adaptive algorithm
This section presents the proof of the convergence of the adaptive Algorithm
3.1 following the proof from [Ver13, Chapter 1.14]. Further used literature is
[MNS02], [D9̈6] and [Beb03].
One considers the problem (22)-(24) with homogeneous Neumann boundary
ΓN , thus (24) is defined as n · ∇u = 0 on ΓN . As mentioned before, Algorithm
3.1 uses the residual error indicator ηR,K , and a suitable marking strategy in step
(4). After providing all the necessary data, the goal is to prove that an algorithm
gives a sequence of discrete solutions that converge to the true solution of the
differential equation. Also, some remarks on refinement strategies are made, as
well as how to apply them to obtain the results.
The idea is to show convergence when the right-hand side is general and not
restricted as piecewise constant. To start with, the procedure for a piecewise
constant right-hand side is presented, therefore, assume that f is piecewise
constant on all partitions. Let T h1 be a triangulation of Ω, and define its subset
Te h1 in the marking strategy of Algorithm 3.2. Let T h2 be a refinement of T h1
1,0 1,0
satisfying SD (T h1 ) ⊂ SD (T h2 ), i.e, the finite element spaces are nested. This
property is crucial for error reduction and a sequence of nested spaces can be
obtained by using refinement by bisection. Assume that

each element of Te h1 , and also each of its faces, contains a node of (53)
h2
T in its interior.

In particular, for every 1-face, i.e., edge, this condition can be interpreted in a
way that the midpoint of every edge of every element in Te h1 is a node of an
element in T h2 . One can fulfill the special refinement (53) by applying either
two steps of the red-refinement, or three steps of the marked edge bisection to
every element in Te h1 . Define by u1 and u2 the unique solutions of problem
(29) that correspond to the partitions T h1 and T h2 , respectively. The goal is
to show that there exists a constant 0 < α < 1 such that

k∇(u − u2 )k2L2 (Ω) ≤ α2 k∇(u − u1 )k2L2 (Ω) .

This means that by every iteration of Algorithm 3.1, the error is reduced at least
by the factor α. Then Algorithm 3.1 converges when f is piecewise constant on
the coarsest partition T h0 . To obtain α some estimates and inequalities should
be presented.
Lemma 4.1. [MNS02, Lemma 4.1] If T h2 is a local refinement of T h1 , such
1,0 1,0
that SD (T h1 ) ⊂ SD (T h2 ), the following relation holds:

k∇(u − u2 )k2L2 (Ω) = k∇(u − u1 )k2L2 (Ω) − k∇(u1 − u2 )k2L2 (Ω) . (54)

Proof. By Galerkin orthogonality, u2 − u1 is perpendicular to u − u2 . Therefore,

since u−u1 = (u−u2 )+(u2 −u1 ), (54) follows from the Pythagoras theorem.

As mentioned in the last section, using Algorithm 3.2 to obtain Te h1 of T h1 ,

where β ∈ (0, 1), one gets that Te h1 satisfies
X X
2 2
ηR,K ≥β ηR,K . (55)
h K∈T h1
K∈Te 1

30
 1
The terms hK kf − fK kL2 (K) and he2 kg − ge kL2 (e) in equation (51) are called
data oscillations. Since f is piecewise constant on T h1 , the term hK kf −
1
fK kL2 (K) vanishes. The term he2 kg − ge kL2 (e) vanishes because the Neumann
boundary condition is homogeneous. Therefore, equation (51) implies
2 X
k∇(u − u1 )k2L2 (Ω) ≤ c∗ 2
ηR,K . (56)
K∈T h1

Using estimates (55) and (56), one can conclude that

2
c∗ X
k∇(u − u1 )k2L2 (Ω) ≤ 2
ηR,K . (57)
β
eh1
K∈T
The special refinement (53) is required to obtain further results. One can
find intermediate steps in [Ver13, Chapter 1]. The final estimate is
X
2
ηR,K ≤ c2∗ k∇(u2 − u1 )k2L2 (Ω) . (58)
K∈Te h1

Furthermore, estimates (57) and (58) imply

1 X 2 β
− k∇(u2 − u1 )k2L2 (Ω) ≤ − ηR,K ≤ − 2 ∗2 k∇(u − u1 )k2L2 (Ω) . (59)
c2∗ c∗ c
eh1
K∈T

Using the estimates (54) and (59), one obtains that

k∇(u − u2 )k2L2 (Ω) = k∇(u − u1 )k2L2 (Ω) − k∇(u1 − u2 )k2L2 (Ω)
β
≤ (1 − )k∇(u − u1 )k2L2 (Ω) .
c2∗ c∗2
q
β
Thus, α is defined as α = 1− c2∗ c∗2
, and the above-mentioned statement
implies that Algorithm 3.1 converges when f is piecewise constant. Also, one
should remark that α depends on the parameter β and the shape parameter of
T h0 .
The procedure for the general right-hand side is presented in the following,
thus let f ∈ L2 (Ω) be an arbitrary function. An approximation by piecewise
constants is required for resolving the right-hand side. Given an arbitrary par-
tition T h , let fT h be the L2 -projection of f onto the space
P of piecewise constant
functions corresponding to T h . It is defined as fT h = K∈T h fK χK , where χK
denotes the characteristic function of the set K. fT h is the L2 (Ω) best approx-
imation to f by piecewise constants dependent on T h . Let u and u e denote the
unique solutions of problem (26) with right-hand sides f and fT h , respectively.
Denote by uT h and u eT h the unique solutions of the discrete problem (29) with
right-hand sides f and fT h , respectively. uT h is the best approximation to u in
1,0
SD (T h ) with respect to the norm k∇ · kL2 (Ω) , thus one has
k∇(u − uT h )kL2 (Ω) ≤ k∇(u − u
eT h )kL2 (Ω) . (60)
Applying the triangle inequality on the right-hand side of the inequality (60)
implies
k∇(u − u
eT h )kL2 (Ω) ≤ k∇(u − u
e)kL2 (Ω) + k∇(e
u−u
eT h )kL2 (Ω) .

31
Therefore, it holds

k∇(u − uT h )kL2 (Ω) ≤ k∇(u − u

e)kL2 (Ω) + k∇(e
u−u
eT h )kL2 (Ω) . (61)

For the term k∇(e u−u eT h )kL2 (Ω) one can apply the procedure for the piecewise
constant right-hand sides shown before. For the term k∇(u − u e)kL2 (Ω) , one
wants to show again, that every iteration of the modified Algorithm 3.1 reduces
k∇(u − ue)kL2 (Ω) by some factor. Through this section, it is shown that in order
to control the term k∇(u − u e)kL2 (Ω) one should control data oscillation. To
connect this term with hK kf − fK kL2 (K) , one should first define v = u − u e and
apply it into equation (40). Therefore, one obtains
Z
1
k∇(u − u e)kL2 (Ω) = sup ∇(u − u
e) · ∇w.
w∈HD 1 (Ω)\{0} k∇wkL2 (Ω) Ω

Since u and u e are solutions to the problem (26) with right-hand sides f and
1
fT h , respectively, one gets that for every w ∈ HD (Ω),
Z Z
∇(u − ue) · ∇w = (f − fT h )w.
Ω Ω

Let wK be defined as the integral mean value of w over K. Because of the

definition of fT h as L2 -projection one can conclude that
Z Z X Z
(f − fT h )w = (f − fT h )(w − wT h ) = (f − fK )(w − wK ).
Ω Ω K
K∈T h

Therefore, combining previous equalities, it is obtained

Z X Z
∇(u − ue) · ∇w = (f − fK )(w − wK ).
Ω K
K∈T h

Now, using the Cauchy–Schwarz inequality on each of the terms in the last sum,
it follows
Z X
∇(u − u
e) · ∇w ≤ kf − fK kL2 (K) kw − wK kL2 (K) .
Ω
K∈T h

The modification of the classical Poincaré inequality, Theorem A.4, will be

used for the next step. It is practical to know an explicit expression for the
constant Cp in (18). In the case of convex domains, this constant is πd , where d
is the diameter of Ω. Here, every element K is convex and hK is the diameter
of K. One can check more details in [Ver13, Chapter 3], and conclude that for
all elements it holds
hK
kw − wK kL2 (K) ≤ k∇wkL2 (K) .
π
Then by combining the previous expressions, bounding the last terms with the
modified Poincaré inequality, and applying a weighted Cauchy–Schwarz inequal-
ity, one gets
1n X 2 o 12
k∇(u − u
e)kL2 (Ω) ≤ hK kf − fK k2L2 (K) .
π h
K∈T

32
 Therefore, in the case of general right-hand sides f , the right-hand side of
this estimate, i.e., data oscillation, should be controlled. To obtain that, one
2
should change Algorithms 3.1 and 3.2. First, in Algorithm 3.2, ηR,K is replaced
2 2 h1 h2 e h1
with hK kf − fK kL2 (K) . Let T , T , T be defined as in the beginning of
the section and let β ∈ (0, 1). With the changed Algorithm 3.2, a subset Te h1 of
T h1 satisfies
X X
h2K kf − fK k2L2 (K) ≥ β h2K kf − fK k2L2 (K) .
eh1
K∈T K∈T h1

Assume that

every element K ∈ Te h1 is the union of elements K̂ ∈ T h2 such that

1
hK̂ ≤ hK .
2

One can obtain this special condition by applying to every element in Te h1 either
two steps of the red refinement or three steps of the marked edge bisection.
The set T h2 can
S S be split into two disjoint subsets T
hB
and T hA such that
K̂∈T hA K̂ = K∈Teh1 K. One has:
X X X
h2K kf − fK k2L2 (K) = h2K̂ kf − fK̂ k2L2 (K̂) + h2K kf − fK k2L2 (K)
K∈T h2 K̂∈T hA K∈T hB

1 X X
≤ h2K kf − fK k2L2 (K) + h2K kf − fK k2L2 (K)
4
K∈
e K∈T h1
T h1 \T h1
e
1 X X
= h2K kf − fK k2L2 (K) + h2K kf − fK k2L2 (K)
4
K∈T h1
K∈T h1
e
X
− h2K kf − fK k2L2 (K)
K∈T
e h1
X 3 X
= h2K kf − fK k2L2 (K) − h2K kf − fK k2L2 (K)
4
K∈T h1 K∈T h1
e
3β X
≤ (1 − ) h2K kf − fK k2L2 (K) .
4
K∈T h1

Furthermore, Algorithm 3.1 should be changed. To do this, step (2) is left out
and in step (3), the a posteriori error estimate is replaced by hK kf − fK kL2 (K) .
Thus, by every iteration, k∇(u − u e)kL2 (Ω) is reduced at least by the factor
q
3β
1− 4 .
To conclude, given a partition T h and any fixed tolerance , this algorithm
provides k∇(u − u e)kL2 (Ω) ≤ 2 after finitely many iterations. Using the adaptive
algorithm described at the beginning of this section with partition T h as a
starting point, after finitely many iterations, one obtains a refined partition
0
T h with k∇(e u−u eT h0 )kL2 (Ω) ≤ 2 . From (61) one can conclude that k∇(u −
uT h0 )kL2 (Ω) ≤ .

33
5 Numerical Studies
This section presents the results of simulations of two numerical examples. Both
examples inspect the Poisson equation (22). In example 5.1, a known solution
with a boundary layer is studied. In example 5.2, a known solution with a circu-
lar internal layer is studied. The goal is to show the convergence of the adaptive
Algorithm 3.1, defined in chapter 3. The used literature is [ABR17], [Joh00],
[JS08], [JM04], [JMT97] and [WBA+ 17]. All simulations were performed with
the program package ParMooN (Parallel Mathematics and object-oriented Nu-
merics) [WBA+ 17], used for the numerical solution of elliptic and parabolic par-
tial differential equations. It is a modernized version of its predecessor MooNMD
[JM04]. The program flow shown in Figure 4 is utilized.

Figure 4: The program flow [Joh00, Figure 1.]

The computation starts on an initial grid, i.e., level 0. The uniform grid
refinement is applied until the appropriate level is reached, then the adaptive
grid refinement starts. This level has to be found by numerical tests. In the
examples presented here, the value of ref inement max n unif orm steps is set
to 2. Then the adaptive refinement iterations are run. It was sufficient to
run 20 iterations (ref inement max n adaptive steps) in each example to get
enough data for the evaluation of results. An error estimator is the next thing
that one has to choose. In this case, it is the residual error estimator which
is introduced in section 3.3. After its computation, every mesh cell K has a
number ηR,K . As mentioned before, these numbers influence which mesh cells
should be refined or coarsened. The coarsening process is not important for the
stationary problems. The maximum strategy is used as a marking strategy and
one can see more details in section 3.4. Let ηmax be a reference value which is
previously defined as ηmax = maxK∈T h ηR,K and let β ∈ (0, 1) be a threshold.
Then, a mesh cell K will be refined if ηR,K ≥ β ηmax . In the examples run

34
here, this value is set as β = 0.5. After every adaptive refinement step, a
sufficient increase in the number of degrees of freedom should occur. In this
way, the expensive assembling and solving a discrete system with only a few
new degrees of freedom is prevented. In Figure 4 one can see that during the
program flow, the marking criterion is relaxed to allow more mesh cells to be
marked for refinement. To obtain that, a minimal amount of mesh cells that
have to be marked for refinement per iteration is set. This value is defined as
min ref and in the examples here is set as min ref = 0.1(= 10%). In the case
when not sufficiently many mesh cells are marked, β is set as β = β/1.1 and
the mesh cells are marked again. Usually, the value of β has to decrease several
times before a sufficient number of cells are marked for refinement.
Example 5.1. [ABR17, Example 6.1] (A known two dimensional solution with
a boundary layer) Let  = 10−3 , Ω = (0, 1)2 , and Dirichlet boundary conditions
and the right-hand side f such that the exact solution of (22) is given by
−(1−x) 1
 e  − e−  
u(x, y) = y(1 − y) x − 1 .
1 − e− 
A numerical solution is presented in Figure 5.a).

(a) Solution of Example 4.1. (b) Iteration 15

Figure 5: A known two dimensional solution with a boundary layer

Level Mesh cells Min mesh cells Marked mesh cells

0 32 3 4
1 57 5 6
4 361 36 42
10 14771 1477 5377
15 425262 42526 92566

Table 1: Table with information to the number of mesh cells

The development of the mesh through iterations of the Algorithm 3.1 is

shown in Figure 6: the top-left (a) is the initial triangulation, top-right (b) is
iteration 1 , bottom-left (c) is iteration 4 and bottom-right (d) is iteration 10. In

35
addition, Figure 5.b) presents iteration 15. One can see that the refinement pro-
cess is concentrated around the boundary layer and the approximate solutions
are nonnegative, as well as the true solution. One can find refinement results
of the mentioned iterations 0, 1, 4, 10 and 15 in Table 1. In the second column
the number of mesh cells is shown. In the third column the minimal amount of
mesh cells that have to change is shown. The program also provides the num-
ber of cells which are marked for refinement, and that is presented in the fourth
column. Therefore, the above-mentioned threshold of required 10% of mesh
cells as a minimal amount of mesh cells that must be marked for refinement, is
satisfied.

(a) Initial triangulation (b) Iteration 1

(c) Iteration 4 (d) Iteration 10

Figure 6: Development of adaptive meshes through iterations of Algorithm 2.1.

Furthermore, the value of ηmax decreases with every iteration. As an ex-

ample, for iteration 0, the value is ηmax = 414.405 while for iteration 17 it is
ηmax = 0.0006. Therefore, the value of ηR,K also decreases. One can see values
of ηmax per iteration in Figure 7.
Figure 8 shows the increasing number of degrees of freedom, while the error
k∇(u − uh )kL2 (Ω) decreases. These results are obtained after 20 iterations. One
can see that the error increases until iteration 2 and then it starts to decrease.
Once when the mesh is sufficiently refined, the error decreases with the optimal
rate. Therefore, the adaptive Algorithm 3.1 converges. With respect to the
number of degrees of freedom, the order of convergence is the first order.

36
 Figure 7: Decreasing values of ηmax through iterations

Figure 8: The behavior of the error k∇(u − uh )kL2 (Ω) for adaptive refinement

Figure 9 shows the error of the function in the L2 -norm using the data
of 20 iterations. In the log-log plot, the optimal decrease of ku − uh kL2 (Ω) is
represented with a red dotted line with factor −2. One can see, that the error
decreases optimally as the number of degrees of freedom increases. Thus, it can
be concluded, that it is a second order convergence with respect to the L2 norm
of the error.

37
Figure 9: The behavior of the error ku − uh kL2 (Ω) with respect to the number
of degrees of freedom

Figure 10: CPU time in seconds through 15 iterations

Furthermore, the number of degrees of freedom increases with every iteration

along with the computation time, especially when the number of degrees of
freedom is more than 105 . As an example, the iteration 1 was executed in 0.001
seconds while the iteration 15 needed 593.837 seconds to be executed. One can
see the increasing values of CPU time in Figure 10.

38
Example 5.2. [JS08, Example 7.2] (Circular internal layer) In this example,
the right-hand side is chosen such that the solution of the Poisson problem (22)
is

u(x, y) = 16 sin(π)x(1 − x)y(1 − y)

1
!
1 arctan[2− 2 (0.252 − (x − 0.5)2 − (y − 0.5)2 )]
× + ,
2 π

with  = 10−3 . The problem is considered in Ω = (0, 1)2 . A numerical solution

is presented in Figure 11. One can see that it is a hump. √The parameter 
controls the steepness of the circular internal layer, while  represents the
thickness of the layer.

Figure 11: Solution of the example 4.2

Simulations are performed on the adaptively refined grids starting with the
same initial grid as in the previous example. Therefore, Figure 12 shows an
initial grid with 32 mesh cells.

Figure 12: Initial grid

When Algorithm 3.1 with parameters β = 0.5 and min ref = 0.1 of the max-
imum marking strategy is applied to the problem, a sequence of grids through
iterations is obtained and it is shown in Figure 13. Iteration 1 in 13.(a) has 64
mesh cells, iteration 4 in 13.(b) has 668 cells, iteration 7 in 13.(c) has 7052 cells
and iteration 10 in 13.(d) has 52448 cells.

39
 (a) Iteration 1 (b) Iteration 4

(c) Iteration 7 (d) Iteration 10

Figure 13: Adaptive refinement of meshes

Figure 14: Increasing number of mesh cells through iterations

40
 In Figure 14 one can see the number of mesh cells, which increases exponen-
tially with every iteration, shown with the green line. The blue line presents
the marked mesh cells. The red dotted line shows a threshold of minimum mesh
cells that have to be marked for refinement. Therefore, for min ref = 0.1, that
number is obtained.

Figure 15: Decreasing values of ηmax

Figure 16: The decrease of the error k∇(u − uh )kL2 (Ω) with respect to the
degrees of freedom

41
 As expected, the value of ηmax that directs the refinement process, decreases
(Figure 15). One can conclude that the value of ηR,K also decreases.
Figure 16 presents the decreasing values of the error k∇(u − uh )kL2 (Ω) while
the values of degrees of freedom increase. The error decreases towards 0. There-
fore, as in the previous example, one can conclude that the adaptive Algorithm
3.1 converges. The error decreases optimally following the red dotted line, which
presents the optimal decrease of the error. The order of convergence is the first
order. The data from 20 iterations is used.

Figure 17: The decrease of the error ku − uh kL2 (Ω) with respect to the degrees
of freedom

In Figure 17, one can see that the error ku − uh kL2 (Ω) also decreases as
the number of degrees of freedom increases. In that case, the second order
convergence is obtained.

42
 Figure 18: The increase of CPU time

As in the previous example, the computation time increases with increasing

the number of degrees of freedom. In Figure 18, one can see results after 13
iterations and significant growth after the number of degrees of freedom exceed
105 .

43
6 Conclusion and Outlook
6.1 Conclusion
This thesis discusses the convergence of the adaptive finite element method for
the Poisson problem. It is proved that an adaptive method based on a residual
error estimates and a refinement strategy converges to the solution of the Poisson
equation. This result is supported by the numerical results of simulations of the
Poisson equation for two different examples.
The conforming finite element method is based on the variational form.
Variational forms use functions in Sobolev spaces. The Poisson equation with
Dirichlet-Neumann boundary conditions is rewritten in its variational form, and
then the solution is approximated with the Ritz-Galerkin method in finite ele-
ment spaces.
It is concluded that a priori error estimates are not sufficient for adaptive
methods since they provide only the asymptotic behavior of the discretization
errors, while a posteriori error estimates provide reliable upper and lower bounds
for the error. The steps for obtaining these bounds are presented as well as an
a posteriori error indicator. Having a residual a posteriori error indicator, an
adaptive algorithm and refinement strategies are operative. The idea is to refine
elements that give a large contribution to the estimated error.
The convergence is proved for cases when the right-hand side is piecewise
constant and when the right-hand side is general. It is proved that by every
iteration of the adaptive algorithm, the error in the L2 -norm of gradient is
reduced by some fixed factor. Thus the adaptive algorithm converges.
The results from simulations support the theory. It is shown how the values
of error in L2 -norm of gradient decrease while the number of degrees of freedom
increases and that the order of convergence is the first order. The examples
presented here show the adaptive mesh refinement accurately. They have a
boundary layer and an internal layer, which is where singularities appear. Those
regions should be refined more, therefore more grid points are placed there. One
can see that in the examples, the refinement process is concentrated around the
boundary layer and the internal layer, where more grid points are placed.

6.2 Outlook
From the viewpoint of the computations, several additional studies would be
possible. The choice of the coarsest grid directs the course of the refinement
process. Also, a different level as a starting level for adaptive grid refinement,
after the uniform grid refinement, would show a difference. Furthermore, an-
other strategy could be used as a marking strategy to obtain better results. In
the examples here the maximum strategy is used as a marking strategy. If max-
imum marking strategy changes the values of β and min ref the computations
would change a lot. Also, to have an efficient adaptive algorithm for the solution
of stationary problems, after an adaptive refinement step, a sufficient increase
in the number of degrees of freedom is needed. As it is shown, the increas-
ing number of degrees of freedom requires more computation time. Therefore,
some balance and optimality between the number of degrees of freedom and
computation time should be obtained.
From the mathematical viewpoint, the next step would be to try to use

44
different a posteriori error estimates to show the convergence of the adaptive
algorithm since the refinement strategies work with them. Also, since the con-
vergence is shown in the case with pure Dirichlet boundary conditions, the case
with more complicated boundary conditions could be investigated. Eventually,
the more complicated partial differential equations could be studied.

45
List of Figures
1 Red, blue and green refinement [CM10, Figure 6.1.] . . . . . . . . 26
2 Marked edge bisection [Ver13, Figure 2.5.] . . . . . . . . . . . . . 27
3 Uniform and adaptive triangulations [Che05, Figure 6.8.] . . . . . 29
4 The program flow [Joh00, Figure 1.] . . . . . . . . . . . . . . . . 34
5 A known two dimensional solution with a boundary layer . . . . 35
6 Development of adaptive meshes through iterations of Algorithm
2.1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
7 Decreasing values of ηmax through iterations . . . . . . . . . . . . 37
8 The behavior of the error k∇(u − uh )kL2 (Ω) for adaptive refinement 37
9 The behavior of the error ku − uh kL2 (Ω) with respect to the num-
ber of degrees of freedom . . . . . . . . . . . . . . . . . . . . . . 38
10 CPU time in seconds through 15 iterations . . . . . . . . . . . . 38
11 Solution of the example 4.2 . . . . . . . . . . . . . . . . . . . . . 39
12 Initial grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
13 Adaptive refinement of meshes . . . . . . . . . . . . . . . . . . . 40
14 Increasing number of mesh cells through iterations . . . . . . . . 40
15 Decreasing values of ηmax . . . . . . . . . . . . . . . . . . . . . . 41
16 The decrease of the error k∇(u − uh )kL2 (Ω) with respect to the
degrees of freedom . . . . . . . . . . . . . . . . . . . . . . . . . . 41
17 The decrease of the error ku−uh kL2 (Ω) with respect to the degrees
of freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
18 The increase of CPU time . . . . . . . . . . . . . . . . . . . . . . 43

List of Tables
1 Table with information to the number of mesh cells . . . . . . . . 35

46
References
[ABR17] Alejandro Allendes, Gabriel R. Barrenechea, and Richard Rankin.
Fully computable error estimation of a nonlinear, positivity-
preserving discretization of the convection-diffusion-reaction equa-
tion. SIAM Journal on Scientific Computing, 39(5):A1903–A1927,
9 2017. Acceptance date: 23/03/17.
[Ada75] Robert A. Adams. Sobolev spaces. Academic Press [A subsidiary of
Harcourt Brace Jovanovich, Publishers], New York-London, 1975.
Pure and Applied Mathematics, Vol. 65.
[AF03] Robert A. Adams and John J. F. Fournier. Sobolev spaces, vol-
ume 140 of Pure and Applied Mathematics (Amsterdam). Else-
vier/Academic Press, Amsterdam, second edition, 2003.
[Beb03] M. Bebendorf. A note on the Poincaré inequality for convex do-
mains. Z. Anal. Anwendungen, 22(4):751–756, 2003.
[BR78] I. Babuška and W. C. Rheinboldt. Error estimates for adaptive
finite element computations. SIAM J. Numer. Anal., 15(4):736–
754, 1978.
[BS97] Randolph E. Bank and R. Kent Smith. Mesh smoothing using a
posteriori error estimates. SIAM J. Numer. Anal., 34(3):979–997,
1997.
[BS08] Susanne C. Brenner and L. Ridgway Scott. The mathematical theory
of finite element methods, volume 15 of Texts in Applied Mathemat-
ics. Springer, New York, third edition, 2008.
[Che05] Zhangxin Chen. Finite element methods and their applications. Sci-
entific Computation. Springer-Verlag, Berlin, 2005.
[Cia91] P. G. Ciarlet. Basic error estimates for elliptic problems. In Hand-
book of numerical analysis, Vol. II, Handb. Numer. Anal., II, pages
17–351. North-Holland, Amsterdam, 1991.
[CM10] C. Carstensen and C. Merdon. Estimator competition for Poisson
problems. J. Comput. Math., 28(3):309–330, 2010.
[Cou43] R. Courant. Variational methods for the solution of problems of
equilibrium and vibrations. Bull. Amer. Math. Soc., 49:1–23, 1943.
[D9̈6] Willy Dörfler. A convergent adaptive algorithm for Poisson’s equa-
tion. SIAM J. Numer. Anal., 33(3):1106–1124, 1996.
[DD12] Françoise Demengel and Gilbert Demengel. Functional spaces for
the theory of elliptic partial differential equations. Universitext.
Springer, London; EDP Sciences, Les Ulis, 2012. Translated from
the 2007 French original by Reinie Erné.
[EG04] Alexandre Ern and Jean-Luc Guermond. Theory and practice
of finite elements, volume 159 of Applied Mathematical Sciences.
Springer-Verlag, New York, 2004.

47
[ESW05] Howard C. Elman, David J. Silvester, and Andrew J. Wathen. Finite
elements and fast iterative solvers: with applications in incompress-
ible fluid dynamics. Numerical Mathematics and Scientific Compu-
tation. Oxford University Press, New York, 2005.
[Far16] Douglas Farenick. Fundamentals of functional analysis. Universi-
text. Springer, Cham, 2016.
[GB05] Thomas Grätsch and Klaus-Jürgen Bathe. A posteriori error esti-
mation techniques in practical finite element analysis. Comput. &
Structures, 83(4-5):235–265, 2005.
[Gri11] Pierre Grisvard. Elliptic problems in nonsmooth domains, volume 69
of Classics in Applied Mathematics. Society for Industrial and Ap-
plied Mathematics (SIAM), Philadelphia, PA, 2011. Reprint of the
1985 original [ MR0775683], With a foreword by Susanne C. Bren-
ner.
[Han05] Weimin Han. A posteriori error analysis via duality theory, volume 8
of Advances in Mechanics and Mathematics. Springer-Verlag, New
York, 2005. With applications in modeling and numerical approxi-
mations.

[JM04] Volker John and Gunar Matthies. MooNMD—a program package

based on mapped finite element methods. Comput. Vis. Sci., 6(2-
3):163–169, 2004.
[JMT97] V. John, J. M. Maubach, and L. Tobiska. Nonconforming
streamline-diffusion-finite-element-methods for convection-diffusion
problems. Numer. Math., 78(2):165–188, 1997.
[Joh00] Volker John. A numerical study of a posteriori error estimators
for convection-diffusion equations. Comput. Methods Appl. Mech.
Engrg., 190(5-7):757–781, 2000.

[JS08] Volker John and Ellen Schmeyer. Finite element methods for time-
dependent convection-diffusion-reaction equations with small diffu-
sion. Comput. Methods Appl. Mech. Engrg., 198(3-4):475–494, 2008.
[LB13] Mats G. Larson and Fredrik Bengzon. The finite element method:
theory, implementation, and applications, volume 10 of Texts in
Computational Science and Engineering. Springer, Heidelberg,
2013.
[Maz85] Vladimir G. Maz’ja. Sobolev spaces. Springer Series in Soviet Math-
ematics. Springer-Verlag, Berlin, 1985. Translated from the Russian
by T. O. Shaposhnikova.

[MNS02] Pedro Morin, Ricardo H. Nochetto, and Kunibert G. Siebert. Con-

vergence of adaptive finite element methods. SIAM Rev., 44(4):631–
658 (2003), 2002. Revised reprint of “Data oscillation and conver-
gence of adaptive FEM” [SIAM J. Numer. Anal. 38 (2000), no. 2,
466–488 (electronic); MR1770058 (2001g:65157)].

48
[OR76] J. T. Oden and J. N. Reddy. An introduction to the mathematical
theory of finite elements. Wiley-Interscience [John Wiley & Sons],
New York-London-Sydney, 1976. Pure and Applied Mathematics.

[SF73] Gilbert Strang and George J. Fix. An analysis of the finite element
method. Prentice-Hall, Inc., Englewood Cliffs, N. J., 1973. Prentice-
Hall Series in Automatic Computation.
[SS05] Alfred Schmidt and Kunibert G. Siebert. Design of adaptive finite
element software, volume 42 of Lecture Notes in Computational Sci-
ence and Engineering. Springer-Verlag, Berlin, 2005. The finite
element toolbox ALBERTA, With 1 CD-ROM (Unix/Linux).
[Ver94] R. Verfürth. A posteriori error estimation and adaptive mesh-
refinement techniques. In Proceedings of the Fifth International
Congress on Computational and Applied Mathematics (Leuven,
1992), volume 50, pages 67–83, 1994.
[Ver13] Rüdiger Verfürth. A posteriori error estimation techniques for finite
element methods. Numerical Mathematics and Scientific Computa-
tion. Oxford University Press, Oxford, 2013.

[WBA+ 17] Ulrich Wilbrandt, Clemens Bartsch, Naveed Ahmed, Najib Alia,
Felix Anker, Laura Blank, Alfonso Caiazzo, Sashikumaar Ganesan,
Swetlana Giere, Gunar Matthies, Raviteja Meesala, Abdus Shamim,
Jagannath Venkatesan, and Volker John. ParMooN—A modernized
program package based on mapped finite elements. Comput. Math.
Appl., 74(1):74–88, 2017.

49
A Tools from the Analysis
This chapter consists of the necessary mathematical results, definitions, and
notations from functional analysis needed for this thesis.

Notation. The notation Γ stands for the boundary of Ω. The notation Ω

stands for the closure of the Ω.

Let α = (α1 , α2 , ..., αn ) be called a multi-index defined as an n-tuple of nonneg-

ative integers
Pn αj . Let xα denote the monomial xα 1 αn
1 · · · xn , which has degree
∂ α α1 αn
|α| = j=1 αj . If Dj = ∂xj then D = D1 · · · Dn denotes a differential oper-
ator of order |α|. Note that D(0,..,0) u is equal to u.

Laplacian operator ∆ is defined as

∂2p ∂2p
∆p = 2 + .
∂x1 ∂x22

Definition 1. (Essentially bounded) Let (X, Σ, µ) be a measurable space. The

essential supremum of a measurable function u : X → C is the quantity
( )
\
ess sup u = sup |λ| : λ ∈ u(E) .
E∈Σ, µ(E C )=0

If the essential supremum of u is finite, then u is essentially bounded. This

definition can be found in [Far16, Def. 5.46].
Definition 2. (Banach space) A normed linear space (V, k·k) is called a Banach
space if it is complete with respect to the metric induced by the norm k · k. This
definition can be found in [BS08, Def. 1.1.7].
Definition 3. (Space C m (Ω)) Let m be any nonnegative integer and let C m (Ω)
denote the vector space consisting of all functions u which, together with all their
partial derivatives Dα u of orders |α| ≤ m, are continuous on Ω. This definition
is taken from [AF03, Def. 1.26].
Definition 4. (Space C ∞ (Ω)) The space C ∞ (Ω) is the space of infinitely dif-
ferentiable functions on Ω, as defined in [Maz85, Page 1.].
m
Definition 5. (Space of bounded, continuous functions) The space CB (Ω) con-
m α
sists of all functions u ∈ C (Ω) for which D u is bounded on Ω for 0 ≤ |α| ≤ m.
It is a Banach space with the norm:

kukCBm (Ω) = max sup |Dα u(x)|.

0≤α≤m x∈Ω

The definition can be found in [AF03, Def. 1.27].

Definition 6. (Space of bounded, uniformly continuous functions)
If u ∈ C(Ω) is bounded and uniformly continuous on Ω, then it possesses a
unique, bounded, continuous extension to the Ω. The vector space C m (Ω)

50
consists of all functions u ∈ C m (Ω) for which Dα u is bounded and uniformly
continuous on Ω for 0 ≤ |α| ≤ m. It has the norm:

kukC m (Ω) = max sup |Dα u(x)|.

0≤α≤m x∈Ω

m
C m (Ω) is a closed subspace of CB (Ω), thus it is a Banach space. This definition
can be found in [AF03, Def. 1.28].
Definition 7. (Space of Hölder continuous functions) Let 0 < λ ≤ 1 and let
C m,λ (Ω) denote the space consisting of functions u for which Dα u, whenever
0 ≤ α ≤ m, satisfies a Hölder condition of exponent λ in Ω. That means that
there exists a constant C such that

|Dα u(x) − Dα u(y)| ≤ C|x − y|λ , x, y ∈ Ω.

The space C m,λ (Ω) is a subspace of C m (Ω), thus it is a Banach space with the
norm:
|Dα u(x) − Dα u(y)|
kukC m,λ (Ω) = kukC m (Ω) + max sup .
0≤|α|≤m x,y∈Ω,x6=y |x − y|λ

The definition can be found in [AF03, Def. 1.29].

Definition 8. (Hilbert space) Let X be a vector space and (·, ·)X a functional
defined on X × X. A functional (·, ·)X is called an inner product on X if for
every x, y ∈ X and a, b ∈ C it holds:

i) (x, y)X = (y, x)X ,

ii) (ax + by, z)X = a(x, z)X + b(y, z)X ,
iii) (x, x)X = 0 if and only if x = 0.

A space X is called an inner product space p if it is equipped with an inner

product. A norm on X is defined as kxkX = (x, x)X . If X is a Banach space
under this norm, then X is called a Hilbert space. This definition is taken from
[AF03, Def. 1.10].
Definition 9. (Dual space) The dual space denoted by B 0 is a set of bounded
linear functionals on a Banach space B as defined in [Che05, Page 25].
Definition 10. (Support of a function) If u is a function defined on Ω, then
the support of u is defined as the set

supp(u) = {x ∈ Ω : u(x) 6= 0}.

The definition is taken from [AF03, Def. 1.3].

Definition 11. (Lipschitz domain, Lipschitz boundary) An open subset Ω ⊂ Rn
is called Lipschitz domain and Γ is called Lipschitz boundary if for every x ∈ Γ
there exists a neighbourhood U of x in Rn and new orthogonal coordinates
{y1 , ..., yn } such that
a) U is an hypercube in the new coordinates:

U = {(y1 , ..., yn )| − ai < yi < ai , 1 ≤ i ≤ n}

51
b) there exists a Lipschitz continuous function ϕ, defined in

U 0 = {(y1 , ..., yn−1 )| − ai < yi < ai , 1 ≤ i ≤ n − 1},

such that

an
|ϕ(y 0 )| ≤ for every y 0 = (y1 , ..., yn−1 ) ∈ U 0 ,
2
Ω ∩ U = {y = (y 0 , yn ) ∈ U | yn < ϕ(y 0 )},
Γ ∩ U = {y = (y 0 , yn ) ∈ U | yn = ϕ(y 0 )}.

The definition can be found in [Gri11, Def. 1.2.1.1]. Only domains with Lip-
schitz boundary are considered in this thesis.

The following definitions can be found in [Che05, Page 26-27].

Definition 12. (Bilinear form) Let V be a Hilbert space with the scalar product
(·, ·) and the corresponding norm k · kV , then a(·, ·) : V × V → R is a bilinear
form, if it satisfies:

i) a(u, αv + βw) = αa(u, v) + βa(u, w),

ii) a(αu + βv, w) = αa(u, w) + βa(v, w),

for α, β ∈ R, and u, v, w ∈ V .
Definition 13. A bilinear form a(·, ·) is symmetric for all u, v ∈ V , if

a(u, v) = a(v, u).

Definition 14. A bilinear form a(·, ·) is called continuous or bounded in the

norm k · kV , if there exists a constant M > 0 such that

|a(u, v)| ≤ M kukV kvkV

for every u, v ∈ V.
Definition 15. A bilinear form a(·, ·) is called coercive or V -elliptic, if there
exists a constant m > 0 such that

|a(u, u)| ≥ mkuk2V

for every u ∈ V.
Definition 16. (Green’s formula) Let Ω be a domain in R2 , Γ boundary and
n an exterior unit normal. Green’s formula is defined as
Z Z Z
−∆u · vdx = ∇u · ∇vdx − n · ∇uvds.
Ω Ω Γ

Green’s formula also holds in higher dimensional spaces. This definition can be
found in [LB13, Def. 4.4].
Definition 17. (Dirichlet boundary condition) The Dirichlet boundary condi-
tion provides the value of the solution at the boundary, as defined in [LB13,
2.2.2.1]. Also, it is called an essential condition, because it appears explicitly in
the variational formulation i.e., in the definition of the space V [BS08, Page 3].

52
Definition 18. (Neumann boundary condition) The Neumann boundary condi-
tion provides the value of the solution derivative at the boundary, as defined in
[LB13, 2.2.2.1]. Also, it is called a natural condition, because it is incorporated
∂u
implicitly. It is denoted by ∂n , i.e. a derivative of u in the direction normal to
∂u
the boundary Γ. It can be equivalently written as ∂n = n · ∇u [BS08, Page 3,
Page 129].
Theorem A.1. (Riesz Representation Theorem) Let V be a Hilbert space with
inner product (·, ·). Every continuous linear form f (·) on V can be uniquely
represented as
f (v) = (u, v)
for some u ∈ V.
Proof. The proof can be found in [LB13], Theorem 7.1.
Theorem A.2. Let V be a Hilbert space, and let a(·, ·) be a continuous, sym-
metric, coercive bilinear form on V . Then the variational problem is equivalent
to the minimization problem: find u ∈ V such that F (u) = minv∈V F (v) where
F (v) = 21 a(v, v) − f (v).
Proof. This proof can be found in [LB13], Theorem 7.2.
Definition 19. (Diameter hK ) A diameter hK of an arbitraty element K ∈ T h
is defined as
hK = diam(K) = max{|x − y| : x, y ∈ K}.
For a triangulation T h , we denote h = maxK∈T h hK a mesh parameter. It is
a measure of how refined the mesh is. The smaller h is, the finner the mesh.
This definition can be found in [Han05, Page 40]. A diameter of every element
or edge F ∈ T h ∪ ε is denoted by hF [Ver13, Page 5].
Lemma A.1. (Deny-Lions) Let 1 ≤ p ≤ +∞ and l ≥ 0. Let Ω be a connected
bounded open set having the (1, p)-extension property. There exists c > 0 such
that
X X
∀v ∈ W l+1,p (Ω), inf kDα (v + π)kLp (Ω) ≤ c kDα vkLp (Ω) .
π∈Pl
|α|≤l+1 |α|=l+1

Proof. This proof can be found in [EG04], Lemma B.67.

Lemma A.2. (Bramble-Hilbert) Assume the hypotheses of the Deny-Lions
Lemma hold. Then, there is c > 0 such that, for all f ∈ (W k+1,p (Ω))0 van-
ishing on Pk ,
X
∀v ∈ W k+1,p (Ω), |f (v)| ≤ ckf k(W k+1,p (Ω))0 kDα vkLp (Ω) .
|α|=k+1

Proof. This proof can be found in [EG04], Lemma B.68.

Definition 20. (Fractional Sobolev spaces) For 0 < s < 1 and 1 ≤ p < +∞,
the Sobolev space with fractional exponent is defined as
( )
s,p p u(x) − u(y) p
W (Ω) = u ∈ L (Ω) : d ∈ L (Ω × Ω) .
kx − yks+ p

53
This definition can be found in [EG04, Def. B.30]. Using mappings, it is possible
to define W s,p (∂Ω) whenever ∂Ω is a smooth manifold.

Theorem A.3. (Trace Theorem 1) Let 1 ≤ p < +∞ and Ω be a Lipschitz

bounded open set. Then,
1
,p
(i) γ0 : W 1,p (Ω) → W p0 (∂Ω) is surjective.
(ii) The kernel of γ0 is W01,p (Ω).

Proof. More details in [EG04], Theorem B.52.

Corollary A.1. Let 1 ≤ p < +∞ and Ω be a Lipschitz bounded open set.
1
,p
Then, there exists a constant c such that, for all g ∈ W p0 (∂Ω), there exists
ug ∈ W 1,p (Ω) satisfying γ0 (ug ) = g and kug kW 1,p (Ω) ≤ ckgk 10 ,p . The
W p (∂Ω)
function ug is said to be a lifting of g in W 1,p (Ω).

Proof. More details in [EG04], Corollary B.53.

Theorem A.4. Let Ω ⊂ Rn be a convex domain with diameter d. Then
d
kukL2 (Ω) ≤ k∇ukL2 (Ω)
π
for all u ∈ H 1 (Ω) satisfying
R
Ω
u(x)dx = 0.
Proof. This proof can be found in [Beb03], Theorem 3.2.

54
Nomenclature
(·, ·) inner product of L2 (Ω), page 3
(·, ·)H m (Ω) inner product of H m (Ω), page 7
ne unit normal vector to edge e, page 21

nK unit outward normal to the element K, page 22

∆ Laplacian operator, page 8
δij Kronecker delta, page 11

ηmax maximum value of ηR,K , K ∈ T h , page 24

ηR,K residual a posteriori error indicator, page 23
Γ boundary, page 4
γ trace map, page 7

ΓD Dirichlet boundary, page 8

ΓN Neumann boundary, page 8
K̂ reference triangle or reference square, page 11
P̂ (K̂) polynomial space, page 14

Λ all edges in ε, page 21

λn nodal shape function of a vertex n, page 22
k·kC j (Ω) norm of C j (Ω), page 5

k·kH m (Ω) norm of H m (Ω), page 7

k·kLp norm of Lp (Ω), page 3

k·kL∞ norm of L∞ (Ω), page 3

k·kW m,∞ (Ω) norm of W m,∞ (Ω), page 4
k·kW m,p (Ω) norm of W m,p (Ω), page 4

Je (·) jump across edge e, page 21

N set of all vertices associated with T h , page 22
NΓN vertices on ΓN , page 22
NΩ interior vertices, page 22

R(v) residual of uT h , page 21

Th mesh, page 10
ε edges corresponding to T h , page 21

55
εΓ edges on Γ, page 21
εΓD edges on ΓD , page 21
εΓN edges on ΓN , page 21
εΩ interior edges, page 21
∇ nabla or gradient operator, page 7
Ω domain, page 3
Φn (·) linear functionals, page 11
ψK mapping from V (K) to V (K̂), page 14
ρK diameter of the largest ball inscribed in K, page 14
σT h shape parameter of T h , page 22
θn shape functions, page 11
g̃ lifting of g, page 16
ϕ Lipschitz function, page 17
Te h subset of marked elements for mesh refinement, page 25
K
ee union of all elements sharing a vertex with e, page 22
K
en union of all elements that share a vertex with K, page 22
A stiffness matrix, page 10
a(·, ·) bilinear form, page 50
b load vector, page 10
BK Jacobian matrix, page 14
C j (Ω) space of bounded, uniformly continuous functions, page 48
j
CB (Ω) space of bounded continuous functions, page 48
C ∞ (Ω) space of infinitely differentiable functions, page 48
C0∞ (Ω) function space, page 4
C j,λ (Ω) space of Hölder continuous functions, page 49
Dα derivative, page 4
E (m, p)-extension operator, page 5
FK affine diffeomorphism, page 11
fK mean value of function f , page 23
fT h L2 -projection of f onto the space of piecewise constant functions corre-
sponding to T h , page 30

56
ge mean value of function g, page 23
h maximum value of hK , K ∈ T h , page 51

H01 (Ω) Sobolev space, page 7

1
HD (Ω) Sobolev space, page 8
Hg1D ,D (Ω) Sobolev space, page 18

H m (Ω) Sobolev space, page 7

1
H 2 (Γ) trace of a function in H 1 (Ω), page 16
he diameter of e, page 51
hK diameter of K, page 51

Ih global interpolation operator, page 14

Ih1 Lagrangian piecewise linear interpolation operator, page 12
IK local interpolation operator, page 14

IK̂ local interpolation operator, page 14

1,0
IT h 1
interpolation operator that maps HD (Ω) to SD (T h ), page 22
j edge residuals, page 22

Ke a union of all elements sharing an edge with K, page 23

Kn union of all elements having n as a vertex, page 22
Lp (Ω) Lebesgue space, page 3
L∞ (Ω) Lebesgue space, page 3

L1loc (Ω) space of locally integrable functions, page 4

n outward unit normal vector, page 8
nx inward normal to Γ at x, page 5
PK space spanned by restrictions of vh to K, page 11

Pm m-th order approximation space, page 16

q quality function, page 27
r element residuals, page 22

R1 (K) first order polynomials on element K, page 11

S 1,0 (T h ) lowest order conforming finite element space corresponding to T h ,
page 11
1,0
SD (T h ) lowest order conforming finite element space corresponding to T h van-
ishing on ΓD , page 11

57
Ux neighborhood of x ∈ Γ, page 5
V (K̂) domain of the interpolation operator, page 14

W m,p (Ω) Sobolev space, page 4

W0m,p (Ω) Sobolev space, page 6
wK integral mean value of w over K, page 31

wKn is the average of w on Kn , page 22

58