JOHANNES KEPLER

T LINZ
UNIVERSITA
N e t z w e r k f u r F o r s c h u n g , L e h r e u n d P r a x i s

High Order Finite Element Methods

for Electromagnetic Field Computation

Dissertation
zur Erlangung des akademischen Grades

Doktorin der Technischen Wissenschaften

Angefertigt am Institut f
ur Numerische Mathematik

Begutachter:

Prof. Dr. Joachim Schoberl

Prof. Dr. Leszek Demkowicz, University of Texas at Austin
Eingereicht von:

Dipl.-Ing. Sabine Zaglmayr

Linz, Juli, 2006

Johannes Kepler Universit

at
A-4040 Linz Altenbergerstrae 69 Internet: http://www.uni-linz.ac.at DVR 0093696

Abstract
This thesis deals with the higher-order Finite Element Method (FEM) for computational
electromagnetics. The hp-version of FEM combines local mesh refinement (h) and local
increase of the polynomial order of the approximation space (p). A key tool in the design and
the analysis of numerical methods for electromagnetic problems is the de Rham Complex
relating the function spaces H 1 (), H(curl, ), H(div, ), and L2 () and their natural
differential operators. For instance, the range of the gradient operator on H 1 () is spanned
by the space of irrotional vector fields in H(curl), and the range of the curl-operator on
H(curl, ) is spanned by the solenoidal vector fields in H(div, ).
The main contribution of this work is a general, unified construction principle for H(curl)and H(div)-conforming finite elements of variable and arbitrary order for various element
topologies suitable for unstructured hybrid meshes. The key point is to respect the de Rham
Complex already in the construction of the finite element basis functions and not, as usual,
only for the definition of the local FE-space. A short outline of the construction is as follows.
The gradient fields of higher-order H 1 -conforming shape functions are H(curl)-conforming
and can be chosen explicitly as shape functions for H(curl). In the next step we extend the
gradient functions to a hierarchical and conforming basis of the desired polynomial space.
An analogous principle is used for the construction of H(div)-conforming basis functions. By
our separate treatment of edge-based, face-based, and cell-based functions, and by including
the corresponding gradient functions, we can establish the local exact sequence property:
the subspaces corresponding to a single edge, a single face or a single cell already form an
exact sequence. A main advantage is that we can choose an arbitrary polynomial order on
each edge, face, and cell without destroying the global exact sequence. Further practical
advantages will be discussed by means of the following two issues.
The main difficulty in the construction of efficient and parameter-robust preconditioners for
electromagnetic problems is indicated by the different scaling of solenoidal and irrotational
fields in the curl-curl problem. Robust Schwarz-type methods for Maxwells equations rely
on a FE-space splitting, which also has to provide a correct splitting of the kernel of the curl
operator. Due to the local exact sequence property this is already satisfied for simple splitting
strategies. Numerical examples illustrate the robustness and performance of the method.
A challenging topic in computational electromagnetics is the Maxwell eigenvalue problem.
For its solution we use the subspace version of the locally optimal preconditioned gradient
method. Since the desired eigenfunctions belong to the orthogonal complement of the gradient functions, we have to perform an orthogonal projection in each iteration step. This
requires the solution of a potential problem, which can be done approximately by a couple

of PCG-iterations. Considering benchmark problems involving highly singular eigensolutions,

we demonstrate the performance of the constructed preconditioners and the eigenvalue solver
in combination with hp-discretization on geometrically refined, anisotropic meshes.

ii

Zusammenfassung
Die vorliegende Arbeit besch
aftigt sich mit der Methode der Finiten Elemente (FEM)
hoherer Ordnung zur Simulation elektromagnetischer Feldprobleme. Die hp-Version der
FEM kombiniert lokale Netzverfeinerung (h) und lokale Erhohung des Polynomgrades
des Approximationsraumes (p). In der analytischen wie auch numerischen Behandlung
elektromagnetischer Probleme spielt die exakte de Rham Folge der Funktionenraume H 1 (),
H(curl, ), H(div, ), L2 () eine wesentliche Rolle: So ist zum Beispiel das Bild des
Gradienten-Operators von H 1 () der Raum der rotationsfreien Funktionen in H(curl, ) und
das Bild des Rotations-Operators von H(curl, ) der Raum der divergenzfreien Funktionen
in H(div, ).
Der wesentliche Beitrag dieser Arbeit ist eine einheitliche Konstruktionsmethode f
ur
H(curl)-konforme und H(div)-konforme Finite Elemente beliebiger und variabler Ordnung
f
ur unterschiedlichen Elementgeometrien auf unstrukturierten hybriden Vernetzungen. Ein
wichtiger Punkt dabei, ist die exakte de Rham Folge bereits in der Konstruktion der Basisfunktionen h
oherer Ordnung zu ber
ucksichtigen und nicht, wie u
blich, nur in der Definition
der globalen diskreten R
aume. Kurz zur Konstruktion: Gradientenfelder von H 1 -konformen
hierarchischen Basisfunktionen h
oherer Ordnung sind H(curl)-konform und konnen daher
explizit als H(curl)-Basisfunktionen gewahlt werden. Im nachsten Schritt werden die
Gradientfunktionen zu einer hierarchischen und konformen Basis f
ur den gew
unschten Polynomraum vervollst
andigt. Das analoge Prinzip wird auch zur Konstruktion H(div)-konformer
Finiter Elemente angewendet. Die hierarchische Konstruktion der Basisfunktionen impliziert
ein nat
urliches Raumsplitting in den globalen Raum der Ansatzfunktionen niedrigster
Ordnung und in lokale Kanten-, Fl
achen- und Zellen-basierte Raume der Ansatzfunktionen
hoherer Ordnung. Durch die spezielle Wahl der Ansatzfunktionen gilt eine exakte de Rham
Folge auch auf den lokalen Teilr
aumen - man spricht von lokalen exakten Folgen. Ein
wesentlicher Vorteil ist, dass der Polynomgrad auf jeder einzelnen Kante, Flache und Zelle
des FE-Netzes beliebig variieren kann, ohne die globale exakte Sequenz zu zerstoren. Weitere
praktische Vorteile werden anhand der folgenden Beispiele genauer diskutiert.
Die Herausforderung in der Konstruktion von effizienten und Parameter-robusten Vorkonditionierern f
ur curl-curl-Probleme liegt in der richtigen Behandlung des nicht-trivialen
Kerns des curl-Operators. Die lokale Zerlegung (lokale exakte Sequenz) des FE-Raumes
hoherer Ordnung garantiert auch eine korrekte Zerlegung des Kerns. Dadurch wird bereits
f
ur einfache Schwarz Vorkonditionierer die notwendige Robustheit im Parameter erzielt.
Numerische Beispiele demonstrieren Robustheit und Performance der Methode.
Die Losung von Maxwell Eigenwertproblemen erfolgt mittels simultaner inexakter inverser

iii

Iteration und deren Beschleunigung durch die vorkonditionierte konjugierte Gradientenmethode (Locally Optimal Block PCG-Methoden). Da die Eigenfunktionen auf dem orthogonalen Komplement der Gradientfunktionen gesucht werden, ist in jedem Iterationsschritt eine orthogonale Projektion erforderlich. Das entspricht der Losung eines Potentialproblems
und kann durch einige PCG-Iterationen naherungsweise durchgef
uhrt werden. Anhand eines Benchmark-Problems mit singularen Eigenfunktionen werden Vorkonditionierer und Eigenwertloser in Verbindung mit hp-Diskretisierung auf geometrisch verfeinerten, anisotropen
Netzen getestet.

iv

Acknowledgements
My special thanks go to my advisor J. Schoberl for supervising and inspiring my work
and for countless time-intensive discussions during the last years. At the same time, I am
greatly indebted to L. Demkowicz for co-refereeing this work and for his constructive remarks.
Special thanks go to H. Egger, C. Pechstein, V. Pillwein, and D. Copeland for numerous
discussions and for proof-reading this work, their comments significantly improved the
presentation. Furthermore, I want to thank all my colleagues, especially those from the
START-Project, from the Institute of Computational Mathematics, and from RICAM for
scientific and social support.
I want to acknowledge the scientific environment and the hostage of the Insitute of Numerical
Mathematics, chaired by U. Langer, and the Radon Institute for Computational Mathematics,
leaded by H.W. Engl.
Financial support by the Austrian Science Fund Fonds zur Forderung der wissenschaftlichen

Forschung in Osterreich
(FWF) through the START Project Y-192 is acknowleged.

vi

Contents
1 Introduction

2 Fundamentals of Electromagnetics
2.1 Maxwells equations . . . . . . . . . . . . . . . . . . .
2.1.1 The fundamental equations . . . . . . . . . . .
2.1.2 Material properties . . . . . . . . . . . . . . . .
2.1.3 Initial, boundary and interface conditions . . .
2.2 Vector and scalar potentials . . . . . . . . . . . . . . .
2.3 Special electromagnetic regimes . . . . . . . . . . . . .
2.3.1 Time-harmonic Maxwell equations . . . . . . .
2.3.2 Magneto-Quasistatic Fields: The Eddy-Current
2.3.3 Static field equations . . . . . . . . . . . . . . .
2.4 The general curl-curl problem . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.

5
5
5
7
8
10
12
12
13
13
14

3 A Variational Framework
3.1 Function Spaces, Trace Operators and Greens formulas . . . . . . . . . . . . .
3.2 Mapping Properties of Differential Operators . . . . . . . . . . . . . . . . . . .
3.2.1 The de Rham Complex and Exact Sequences . . . . . . . . . . . . . . .
3.3 Abstract Variational Problems: Existence and Uniqueness . . . . . . . . . . . .
3.3.1 Coercive variational problems . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2 Mixed formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Variational formulation of electromagnetic problems . . . . . . . . . . . . . . .
3.4.1 The electrostatic problem: A Poisson problem . . . . . . . . . . . . . . .
3.4.2 The magnetostatic problem . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.3 The time-harmonic electromagnetic and magneto-quasi-static problems

17
17
23
25
26
26
26
27
27
28
32

4 The Finite Element Method

4.1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Galerkin Approximation . . . . . . . . . . . . . . . . . .
4.1.2 The Triangulation . . . . . . . . . . . . . . . . . . . . .
4.1.3 The Finite Element . . . . . . . . . . . . . . . . . . . .
4.1.4 The reference element and its transformation to physical
4.1.5 Simplicial elements and barycentric coordinates . . . . .
4.2 Approximation properties of conforming FEM . . . . . . . . . .
4.3 An exact sequence of conforming finite element spaces . . . . .
4.3.1 The classical H 1 -conforming Finite Element Method . .
4.3.2 Low-order H(curl)-conforming Finite Element Methods

35
35
35
36
36
37
38
39
40
40
42

vii

. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
Problem .
. . . . . .
. . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

. . . . . .
. . . . . .
. . . . . .
. . . . . .
elements
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

4.4

4.3.3 Low-order H(div)-conforming Finite Elements Methods

4.3.4 The lowest-order L2 -conforming Finite Element Method
4.3.5 Discrete exact sequences . . . . . . . . . . . . . . . . . .
4.3.6 Element matrices and assembling of FE-matrices . . . .
Commuting Diagram and Interpolation Error Estimates . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

46
49
50
53
55

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

59
60
63
63
68
73
82
89
97
107
108
109
109
110
111
112
113

. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
property
. . . . . .
. . . . . .
. . . . . .
. . . . . .

115
115
116
118
118
119
120
124
125
127
128
129
130
133

5 High Order Finite Elements

5.1 High-order FE-spaces of variable order . . . . . . . . . . . . . . . . . . . .
5.2 Construction of conforming shape functions . . . . . . . . . . . . . . . . .
5.2.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2 The quadrilateral element . . . . . . . . . . . . . . . . . . . . . . .
5.2.3 The triangular element . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.4 The hexahedral element . . . . . . . . . . . . . . . . . . . . . . . .
5.2.5 The prismatic element . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.6 The tetrahedral element . . . . . . . . . . . . . . . . . . . . . . . .
5.2.7 Nedelec elements of the first kind and other incomplete FE-spaces
5.2.8 Elements with anisotropic polynomial order distribution . . . . . .
5.3 Global Finite Element Spaces . . . . . . . . . . . . . . . . . . . . . . . . .
5.3.1 The H 1 -conforming global finite element space . . . . . . . . . . .
5.3.2 The H(curl)-conforming global finite element space . . . . . . . . .
5.3.3 The H(div)-conforming global finite element space . . . . . . . . .
5.3.4 The L2 -conforming global finite element space . . . . . . . . . . .
5.4 The Local Exact Sequence Property . . . . . . . . . . . . . . . . . . . . .
6 Iterative Solvers
6.1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Additive Schwarz Methods (ASM) . . . . . . . . . . . . . . .
6.1.2 A two-level concept . . . . . . . . . . . . . . . . . . . . . . .
6.1.3 Static Condensation . . . . . . . . . . . . . . . . . . . . . . .
6.2 Parameter-Robust Preconditioning for H(curl) . . . . . . . . . . . .
6.2.1 Two Motivating Examples . . . . . . . . . . . . . . . . . . . .
6.2.2 The smoothers of Arnold-Falk-Winther and Hiptmair . . . .
6.2.3 Schwarz methods for parameter-dependent problems . . . . .
6.2.4 Parameter-robust ASM methods and the local exact sequence
6.3 Reduced Nedelec Basis and Special Gauging Strategies . . . . . . . .
6.4 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4.1 The magnetostatic problem . . . . . . . . . . . . . . . . . . .
6.4.2 The magneto quasi-static problem: A practical application .

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

7 The Maxwell Eigenvalue Problem

135
7.1 Formulation of the Maxwell Eigenvalue Problem . . . . . . . . . . . . . . . . . 135
7.2 Preconditioned Eigensolvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.2.1 Preconditioned gradient type methods . . . . . . . . . . . . . . . . . . . 139
7.2.2 Block version of the locally optimal preconditioned gradient method . . 140
7.3 Preconditioned Eigensolvers for the Maxwell Problem . . . . . . . . . . . . . . 141
7.3.1 Exact and inexact projection onto the complement of the kernel function142
7.3.2 A preconditioned eigensolver for the Maxwell problem with inexact projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

viii

7.4

7.3.3 Exploiting the local exact sequence property

Numerical Results . . . . . . . . . . . . . . . . . . .
7.4.1 An h-p-refinement strategy . . . . . . . . . .
7.4.2 The Maxwell EVP on the thick L-Shape . . .
7.4.3 The Maxwell EVP on the Fichera corner . . .

A APPENDIX
A.1 Notations . . . . . . . . . . . . . . . . . .
A.2 Basic Vector Calculus . . . . . . . . . . .
A.3 Some more orthogonal polynomials . . . .
A.3.1 Some Calculus for Scaled Legendre
A.3.2 Some technical things . . . . . . .

.
.
.
.
.

.
.
.
.
.

. . . . . . . .
. . . . . . . .
. . . . . . . .
Polynomials .
. . . . . . . .

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

144
144
145
145
149

.
.
.
.
.

151
. 151
. 151
. 151
. 153
. 153

Bibliography

154

Eidesstattliche Erkl
arung

A1

Curriculum Vitae

A3

ix

Chapter 1

Introduction
State of the Art
Electromagnetic processes are present everywhere in our daily life. Classical applications are
generators, transformers and motors, converting mechanical to electric energy and vice versa.
Wireless communication is based on electromagnetic waves in free space. Here, the design of
antennas is a sophisticated task. A fastly growing application field is optics. Optical fibers
allow the transport of light pulses over much longer distances than achieved by electric signals
through cables. Short light pulses are generated by laser resonators. Optical multiplexers
realized by photonic crystals have obtained much attraction over recent years.
All these applications are rather complex, hence for further technical developments and
optimization a deeper insight into electromagnetic processes is necessary.
Similar to many other physical and technical effects (such as solid and fluid mechanics, heat
transfer, quantum mechanics, geoscience, astrophysics, etc.) electromagnetic phenomena are
modelled by partial differential equations (PDEs). This is the basis for the mathematical
analysis and numerical treatment.
Only for very special problems can the solution of partial differential equations be done
analytically. This calls for numerical discretization techniques. The analysis of partial differential equations is commonly done within a variational framework. In the last fifty years, the
Finite Element Method (FEM) has been established as certainly the most powerful tool in
numerical simulation. This discretization technique is based on the variational formulation
of partial differential equations. The main advantages are its general applicability to linear
and nonlinear PDEs, coupled multi-physics systems, complex geometries, varying material
coefficients and boundary conditions. Furthermore, the method is based on a profound
functional analysis (cf. e.g. Ciarlet [35], Brenner-Scott [31], Braess [27]). In classical
(h-version) finite element methods we obtain convergence by global or local refinement of the
underlying mesh (h-refinement). The polynomial order of approximation on each element is
fixed to a low degree, typically p = 1 or p = 2. The error in the numerical solution decays
algebraically in the number of unknowns.
The p-version of the finite element method (see Babuska-Szabo [87]) allows an increase of
the polynomial order, while keeping the mesh fixed. In case of analytic solutions one obtains
exponential convergence, but in case of lower regularity the convergence rate reduces again
to an algebraic one. Hence, in the presence of singularities, which occur very frequently in

CHAPTER 1. INTRODUCTION

practical problem settings, not much is gained compared with the h-version FEM. However,
by a proper combination of (geometric) h-refinement and local increase of the polynomial
degree p the hp-method exponential convergence can be regained for piecewise analytic
solutions involving singularities, e.g. due to re-entrant corners and edges. This is the
typical situation in practical applications. For pioneering works on hp-version FEM see
Babuska-Guo [10] and Babuska-Suri [11]. We refer also to the books by Schwab [83],
Melenk [68], and Karniadakis [60]. The recent textbook Demkowicz [42] copes also with
the aspects of practical implementation and automatic generation of hp-meshes. For the
successfull application to mechanics see e.g. Szabo et al. [86].
In linear as well as nonlinear time-dependent, time-harmonic, and magneto-static regimes of
Maxwells equations the general curl-curl problem
Z
Z
Z
1
Find u V such that
curl u curl v dx +
u v dx =
j v dx
v V

appears, where u is the magnetic vector potential.

Using standard continuous finite elements in the discretization of electromagnetic problems
fails. In the presence of re-entrant corners and edges, the method may even converge to a
wrong solution (cf. Costabel-Dauge [38]). Moreover, in eigenvalue computation, using
standard elements is one source of spurious (non-physical) eigenvalues, which pollute the
computed spectrum (see Bossavit [26], Boffi et al. [21]).
The natural function space for the solution of the curl-curl problem is the vector-valued space
H(curl), which has less smoothness than H 1 , namely only tangential continuity over material
interfaces. This property goes along with the physical nature of electric and magnetic fields.
de
lec
The classical H(curl)-conforming finite element spaces have been introduced in Ne
[72], [73].
The adaption of hp-methods to electromagnetic problems is not straightforward. Investigations in this direction started only in the last decade, and the numerical analysis is not
complete, especially in 3D. A key tool in the design of numerical methods for Maxwells equations and their numerical analysis is the de Rham Complex (cf. Bossavit [22], [25] and more
recently Arnold et al. [8],[9]), which relates function spaces and their natural differential
operators, and reads in 3D:
id

curl

div

R H 1 () H(curl, ) H(div, ) L2 () {0}.

The sequence is exact in the following sense: the range of an operator in the sequence coincides
with the kernel of the next operator.
The de Rham Complex perfectly fits to electromagnetics: in a variational setting H 1 () is the
natural function space for the electrostatic potential, the magnetic and the electric fields lie
in H(curl, ) and their fluxes belong to H(div, ). For a proper conforming hp-finite element
method, the discrete spaces have to form an analogous exact sequence.
High-order p-version elements for H(curl) are analyzed for constant order p in Monk [69].
Variable order elements are proposed for the first time in Demkowicz-Vardapetyan [45].
The consideration of hp-finite elements, allowing variable order approximation, on the basis
of the de Rham Complex is presented in Demkowicz et al. [44] and Demkowicz [41].
A first general construction strategy for tetrahedral shape functions on unstructured grids
was recently introduced by Ainsworth-Coyle [2] for the whole sequence of H 1 -, H(curl)-,

H(div)- and L2 -conforming spaces (for arbitrary but uniform p). For the formulation of finite
elements in the context of differential forms we refer to Bossavit [24],[25] and Hiptmair [54].
The construction of H(curl)-conforming finite elements is also an active research area in the
engineering community, cf. Lee [66], Webb-Forghani [96], Webb [95], and Sun et al. [85].
Due to the non-trivial, large kernel of the curl-operator - the gradient fields of H 1 () - not
only the convergence analysis but also the iterative solution of discretized Maxwell problems
becomes very challenging. The main difficulty stems from the different scaling of solenoidal
and irrotational fields in the curl-curl problem. This leads to very ill-conditioned system
matrices and standard Schwarz-type preconditioners like multigrid/multilevel techniques yield
only bad convergence behavior. In this context, we want to mention the pioneering works on
robust preconditioning in H(curl) and H(div) by Arnold et al. [7] and Hiptmair [55],
berl [78]. The difficulties can be resolved by a careful choice
revisited and unified in Scho
of the Schwarz smoothers, in particular the space splitting has to respect the kernel of the
curl-operator. Further works on this topic are Toselli [89], Hiptmair-Toselli [58], Beck
et al. [16], and Pasciak-Zhao [75].
A rather complete overview on finite element methods for Maxwells equations can be found
in Monk [70]; for another comprehensive survey we refer to Hiptmair [56]. The topic
of hp-methods for Maxwells equations, including implementational aspects, is covered by
Demkowicz [42].

On this work
In this work we present a general, unified construction principle for H(curl)- and H(div)conforming finite elements of variable and arbitrary order. In order to allow for geometric
h-refinement, we have to consider hybrid meshes, involving hexahedral, tetrahedral, and prismatic elements. The innovation of our framework is to respect the exact de Rham sequence
already in the construction of the FE basis functions. We shortly outline the main points of
the construction for H(curl):
We start with the classical lowest-order Nedelec shape functions. Note that the lowestorder space always has to be treated separately by applying h-version methods, e.g. also
in linear solvers.
We take the gradients of edge-based, face-based and cell-based shape functions of the
higher-order H 1 -conforming FE-space.
Finally, we extend these sets of functions to a conforming basis of the desired polynomial
space.
By our separate treatment of the edge-based, face-based, and cell-based functions, and by
including the corresponding gradient functions, we can establish the following local exact
sequence property: the subspaces corresponding to a single edge, a single face or a single cell
already form an exact sequence.
This construction has several practical advantages:
We can choose an arbitrary polynomial order on each edge, face, and cell independently,
without destroying the global exact sequence property, see Section 5.4.

CHAPTER 1. INTRODUCTION

A correct Schwarz splitting can be constructed by simple strategies, and only the lowestorder space has to be treated globally or by standard h-methods. The parameterrobustness is implied automatically by the local sequence property, see Section 6.2.4.
Since gradients are explicitly available, we can implement gauging strategies by simply skipping the corresponding degrees of freedom (Reduced Basis Gauging). We will
illustrate in numerical tests that this approach tremendously improves the condition
numbers and solving times, see Section 6.4.1.
Discrete differential operators, for instance used within the projection for preconditioned
eigenvalue solvers, can be implemented very easily, see Section 7.3.3.
The full family of finite element shape functions for all sorts of element topologies, covering
also anisotropic polynomial degrees and regular geometric h-refinement towards pre-defined
corners, edges and faces, has been implemented in the open-source software package Netgen/NgSolve
http://www.hpfem.jku.at/.
Other resources for higher-order Maxwell FE-packages are e.g. EMSolve (CASC, Lawrence
Livermore National Lab.), 3Dhp90 by L. Demkowicz (ICES University of Texas Austin,
Rachowicz-Demkowicz [76]), Concepts by P. Frauenfelder (ETH Z
urich).
The thesis is organized as follows.
In Chapter 2, we present the Maxwell equations. We pay special attention to scalar and
vector potential formulations, and consider time-harmonic, quasi-static, and magneto-static
regimes in more detail. All these problems involve the abstract parameter-dependent curlcurl-problem, mentioned before.
The first part of Chapter 3 recalls the natural function spaces and their properties. We
formally introduce the de Rham Complex, which is a guiding principle through the whole
work. We conclude this chapter with presenting variational formulations of the electromagnetic problems introduced in Chapter 2.
Chapter 4 briefly overviews the basic concepts of conforming low-order finite element methods, including the non-standard function spaces H(curl, ) and H(div, ).
Chapter 5 contains the main contribution of this thesis: We present in detail our construction
of high-order FE-shape functions for the space H 1 (), H(curl, ), H(div, () and L2 () and
show that the local exact sequence property, mentioned above, holds.
Chapter 6 deals with parameter-robust Schwarz-type preconditioners for H(curl). We prove
that parameter-robust solvers are obtained also even by simple Schwarz-type smoothers, if
the presented conforming high-order FE-basis is used. In the second part the concept of reduced basis gauging is introduced. Finally, we present numerical tests for magneto-static and
magneto-quasi-static problems, illustrating the benefits of our methods.
Chapter 7 is concerned with the numerical solution of Maxwell eigenvalue problems. We investigate preconditioned eigensolvers and their combination with (in)exact projection methods. The performance of the eigensolver in combination with reduced-basis preconditioners is
demonstrated by the solution of benchmark problems.

Chapter 2

Electromagnetics: Fundamental
equations and formulations
2.1

Maxwells equations

Classical electromagnetics treats electric and magnetic macroscopic phenomena including their
interaction. Electric fields which vary in time cause magnetic fields and vice versa. James
Clark Maxwell described these phenomena in his Treatise on Electricity and Magnetism in
1862.
The classic theory mainly involves the following four time- and space-dependent vector fields:
the electric field intensity denoted by E [V /m],
the magnetic field intensity H [A/m],
the electric displacement field (electric flux) D [As/m2 ],
the magnetic induction field (magnetic flux) B [V s/m2 ].
The sources of electromagnetic fields are electric charges and currents described by
the charge density [As/m3 ]) and
the current density function j [A/m2 ],
where the SI units denotes meter (m), seconds (s), Ampere (A), Volt (V ).

2.1.1

The fundamental equations

The basic relations of electromagnetics are based on experiments and laws by Faraday,
Ampere, and Gau. We start here with integral formulation of the main governing equations, which facilitates a physical interpretation. In the following we refer by A to a surface
and by V to a volume in R3 . The corresponding boundaries are denoted by V with outer
unit normal vector n, and A with unit tangential vector .

CHAPTER 2. FUNDAMENTALS OF ELECTROMAGNETICS

j ind

Figure 2.1: Faradays and Amperes law

Faradays induction law describes how the change (in time) of the magnetic flux through
a surface A induces a voltage in the loop (A) and hence gives rise to an electric field E:
Z
Z
B
n dA +
E ds = 0.
(2.1)
A
A t
Since there exist no magnetic charges (monopoles), the magnetic field is solenoidal (sourcefree). Moreover, magnetic field lines are closed. The magnetic flux B through the surface of
a bounded volume V is conservative, i.e.
Z
B n dA = 0.
(2.2)
V

Amp
eres law states how electric currents through a surface A induce a magentic field
as illustrated in Figure 2.1. The integral of the magnetic field along a closed path (A) is
proportional to the current through the enclosed surface, i.e.,
Z
Z
H ds =
j n dA.
A

Maxwell generalized this law by adding the displacement current density

Z

Gau law

H ds =

D
n dA +
t

D
t ,

which yields

j n dA.

(2.3)

describes how electric charges give rise to an electric field. It has the form
Z
Z
D n dA =
dx.
(2.4)
V

The electric flux D through the boundary of a volume V is proportional to the enclosed
volume charges.
Applying Gau and Stokes theorems
Z
Z
div B dx =
B n dA
V

and

curl H n dA =

H ds

2.1. MAXWELLS EQUATIONS

to the integral equations (2.1)-(2.4) yields the Maxwell equations in the following (classical)
differential form:
B
+ curl E
t
div B
D
curl H
t
div D

= 0,

(2.5a)

= 0,

(2.5b)

= j,

(2.5c)

= .

(2.5d)

The conservation of charges

An important physical property can be derived by taking the divergence of equation (2.5c) in
combination with equation (2.5d), which yields the continuity equation
div j +

= 0.
t

(2.6)

By integration over a volume V and application of Gau theorem we see that this equation
describes the conservation of charges. This can be seen from the Integral formulation, namely
Z
Z

j n dA +
dx = 0,
(2.7)
t V
V
which states that the total charge in a volume V changes according to the net flow of electric
charges across its surface V .

2.1.2

Material properties

The system (2.5) is still undertetemined, i.e., it provides only 8 equations for 12 unknowns.
The gap is closed by including appropriate constitutive laws. First, the magnetic and electric
field intensities are related with the corresponding fluxes by
D = E,

(2.8a)

B = H.

(2.8b)

Furthermore, in conducting materials the electric field induces a conduction current with
density j c , which is given by Ohms law
j = jc + ji

with j c = E,

(2.9)

where j and j i denote the total and the impressed current densities, respectively. As a special
class of conduction currents we want to mention eddy currents, which arise in metallic bodies
if excited by varying magnetic fields.
Hence, the electric and magnetic properties of a medium are characterized by
the electric permittivity [As/V m],
the magnetic permeability [V s/Am],
the electric conductivity [As].

CHAPTER 2. FUNDAMENTALS OF ELECTROMAGNETICS

In general, the three parameters are tensors depending on space and time, and on the
electromagnetic fields themselves. However, in isotropic media they simplify to scalars, and
in so-called linear materials, they are independent of the field intensities. We will consider
only isotropic, linear materials in this thesis, and moreover assume the material parameters
to be time independent.
The values of the parameters in vacuum are
0 8.854 1012 F m1 ,
0 and 0 are further connected by

2.1.3

1
0 0

0 = 4 107 Hm1 ,

0 = 0.

= c, where c denotes the speed of light.

Initial, boundary and interface conditions

Although the number of equations now coincides with the number of unknowns, the system
of differential equations (2.5) is not yet complete. We have to impose initial and boundary conditions as well as interface conditions between different materials where the material
parameters jump. Below we will mainly focus on time-harmonic or static settings, and we
therefore skip the treatment of initial conditions at this point. We only remark that applying
the divergence to Faradays and Amperes law yields

div B(x, t) = 0
t

and

div D(x, t) = div j(x, t).

t

Hence, if the magnetic field is solenoidal (div B = 0) at the initial time, then it is solenoidal
for any time. The second relation together with Gau law (2.5d) yields that the change of
the charge density is given by the electric current density j, see also the continuity equation
(2.6).
Interface conditions
Assume a partition of the domain V R3 into two disjoint domains V1 , V2 such that
V = V 1 V 2 . By := V1 V2 we denote the common interface, and by n we refer to
the unit normal vector pointing from V2 to V1 . In the following we derive the continuity
requirements at interfaces by Gau theorem and Stokes theorem assuming that the involved
functions, domains and surfaces are sufficiently smooth.
From equation (2.2) we obtain
Z
Z
Z
0 =
B n dx +
B 1 n dx +
B 2 n dx
V
V1
V2
Z
Z
=
B 1 n dA +
B 2 n dA
V2
Z V1


=
B n dA,



where B 1 := B |V1 , B 2 := B |V2 , and B n = (B 2 B 1 ) n denotes the jump over the
interface . Since the above formula is valid for arbitrary subsets of V , we obtain that the
normal component of the induction field has to be continuous over the interface, which reads
as


(2.10)
B n = 0,

2.1. MAXWELLS EQUATIONS

where now n denotes a normal vector to the surface (interface) .

A similar argument works for the electric flux density D, but here we also take surface
intoR account. Hence, the total electric charge in V is given by
R
Rcharges SR on the interface

dx
+ S dA, and via (2.5d) we obtain

dx
+

dx
=
V2
V1
V


D n = S .

(2.11)

Next we derive interface conditions for the electric field E. Let the interface be as above
and A denote an arbitrary plane surface intersecting the interface along a line L := A .
Let A1 := A V1 , A2 := A V2 be the two disjoint parts of A such that A1 A2 = A and
A1 A2 = L. Considering Faradays law (2.1) in integral form for A, A1 , A2 , we see that
Z
Z
Z
0 =
E ds +
E 1 1 ds +
E 2 2 ds
A
A1
A2
Z
Z
=
E 1 1 ds +
E 2 2 ds
A2
ZA1


=
E L ds
L

with E 1 := E|A1 and E 2 := E|A2 and L = 1 = 2 . Since A was arbitrary, we conclude

that the tangential components of the electric field have to be continuous over the interface
, which is equivalent to


E n = 0,
(2.12)
where as before n denotes a normal to the surface .

Similar arguments can be applied for the magnetic field H. Taking into account also the
possibility of impressed surface currents j ,i on the interface, we obtain


H n = j ,i .

(2.13)

We conclude this section on interface conditions with some remarks on those components of
the electromagnetic fields that were not considered above: It can be shown easily that the
normal components of fluxes and the tangential components of the field intensities may be
discontinuous when the material parameters jump across the interface. To see this, we assume
for simplicity S = 0 and jS,i = 0 on the interface . Then substituting the constitutive laws
(2.8a), (2.8b) and (2.9) into the above (2.10)-(2.13) yields








E n 6= 0, H n 6= 0, D n 6= 0, and B n 6= 0

 
 
in general, namely if 6= 0 and 6= 0 across the interface .
Boundary conditions

In the sequel we present several frequently used boundary conditions for the normal or the
tangential components of the electromagnetic fields. We focus here on standard boundary conditions of Dirichlet-, Neumann- or Robin-type. For a detailed review of of classical boundary
[18].
conditions, we refer to Bro

10

CHAPTER 2. FUNDAMENTALS OF ELECTROMAGNETICS

Perfect electric conductors (PEC) Let PEC denote a perfectly conducting region with
0 as long as the currents j are bounded.
. Then Ohms law j = E implies E|
PEC
Hence, the interface condition (2.12) justifies to substitute perfectly conducting regions by
the PEC-boundary condition for the electric field, i.e., we assume
En=0

on PEC .

(2.14)

The PEC-wall is in particular suitable for modeling adjacent metallic domains, e.g., metallic
electrodes.
Perfect magnetic conductors (PMC) model materials with very high permeability,
where one can assume a vanishing magnetic field H 1 = 0. The interface condition (2.13)
then implies that an adjacent PMC-region can be substituted by the boundary condition
H n=0

on PMC

(2.15)

Prescribed surface charges S on the boundary are introduce by a condition

D n = S

on

(2.16)

on the normal electric flux, cf. the interface condition (2.11).

Impressed surface currents j ,i yield a condition on the normal component of the magnetic field at the boundary in analogy to the interface condition (2.13), i.e.
H n = j ,i

on .

(2.17)

Impedance boundary conditions It is well-known that in highly conducting materials,

eddy currents are concentrated near the surface. If one is interested in the field intensities
in regions adjacent to highly (but still finitely) conducting materials, the reflection at the
interface can be modeled by impedance boundary conditions. These are Robin-type boundary
conditions relating the magnetic and electric fields in the following form:
H n (E n) n = 0 on ,

(2.18)

where is called the impedance parameter.

2.2

Vector and scalar potentials

Let be a bounded, simply connected domain in R3 . Then due to div B = 0, and by

curl = 0 and div curl v = 0, we know that B can be expressed by a vector potential
A(x, t), i.e.,
B = curl A.
Subsituting into Faradays law (2.5a) yields


A
curl E +
= 0,
t

11

2.2. VECTOR AND SCALAR POTENTIALS

which in turn implies the existence of a scalar potential (x, t) such that
E =

A
.
t

Hence, Amperes law (2.3) can be expressed by

2A

A
+ 2 = j i .
t
t
t
Note, that for any arbitrary scalar function the potentials
curl 1 curl A +

(2.19)

= A +
A
=
t
)
provide the same magnetic and electric fields, since B(A, ) = B(A,
and E(A, ) =

E(A, ).
By choosing a vector potential A such that
Z t

A =A+
dt,
t0

we obtain E = t and curl A = curl A , and the following vector potential formulation of
Maxwells equations
A
2 A
curl 1 curl A +
+
= j i,
(2.20)
t
t2
which is the type of equation whose numerical solution we will investigate in detail in this
theses. Once A is determined, the electric and the magnetic fields can be calculated by

B = curl A and E = A
t .
In analogy to (2.12), the continuity requirements of vector potentials across an interface
between are given by
[A n] = 0.

A perfect electric conducting (PEC) wall is modeled by homogenous essential boundary condition for the vector potential A , cf. (2.14), i.e.,
A n = 0

on PEC .

(2.21)

A PMC wall is described by natural boundary conditions

1 curl A n = 0

on PMC .

(2.22)

on

(2.23)

Compare to (2.15), respectively by

1 curl A n = j ,i

if impressed surface currents j ,i are included.

As already mentioned above, the introduced potentials A and are not unique. Uniqueness
can be enforced by imposing additional conditions, which is called gauging. In case of constant
coefficients the so-called Coulomb-gauge is commonly used, where
div A = 0
is required together with the additional boundary condition A n = 0. Furthermore, the
impressed currents are assumed to be divergence free, i.e. div j i = 0.

12

2.3

CHAPTER 2. FUNDAMENTALS OF ELECTROMAGNETICS

Special electromagnetic regimes

In many practical applications one does not have to deal with the full system of Maxwells
equations (2.5a)-(2.5d). Taking into account the physics of the considered problem, the system
(2.5) can often be simplified. We will distinguish between the time-harmonic case, slowly and
fast varying electromagnetic fields, and the static regime. For a detailed overview of the most
[18], and in the
commonly investigated problem classes, we refer to Van Rienen [92], Bro
special case of high-frequency scattering problems to Monk [70].

2.3.1

Time-harmonic Maxwell equations

The investigation of a time harmonic setting is interesting for two reasons: First, the Fourier
transformation in time can be applied to the nonstationary equations, and a general solution
can be obtained afterwards by superposition of the solutions at the several frequencies. Secondly, if a system is excited with a signal of a single frequency, then a single-frequency analysis
is appropriate. We assume the electromagnetic fields E, D, H and B to be time-harmonic,
i.e. of the form
u(x, t) = Re(
u(x)eit ),
(2.24)
where the hat accounts for complex-valued functions and Re denotes the real part of complex
numbers. To stay consistent we assume also the charge density , as well as the current
density j to be of the form (2.24). For ease of notation, we will omit the hat marker in the
by E and so on. The transformation into the frequency domain
sequel, e.g., we denote E
implies complex-valued fields, but has the advantage that the derivative with respect to time
is replaced by a simple multiplication operator, i.e.,
u
(x, t) iu(x, t).
t
By this transformation, the time-harmonic Maxwell equations can be stated as
curl E(x) + iH(x) = 0,

(2.25a)

div H(x) = 0,

(2.25b)

curl H(x) (i + )E(x) = j i (x),

div E(x) = (x).

(2.25c)
(2.25d)

The continuity equation (2.6) transforms to

i(x) + div j(x) = 0.

(2.26)

Thus we can easily eliminate the charge density in a time-harmonic setting. The vector
potential formulation (2.20) then reads
curl 1 curl A + iA 2 A = j i

(2.27)

where we used E = iA and B = curl A. Furthermore, the mixed impedance boundary

condition (2.18) on R is transformed into a Robin-type condition of the form
1 curl A n + iA n = 0

on R .

2.3. SPECIAL ELECTROMAGNETIC REGIMES

2.3.2

13

Magneto-Quasistatic Fields: The Eddy-Current Problem

A special class of time-dependent electromagnetic problems arises, when at least one of the
electromagnetic fields varies only slowly in time. In so-called quasistatic models the timederivative of the magnetic flux or the electric flux is therefore neglected. We focus here
on magneto-quasistatic problems, which are suitable for low-frequency applications like, e.g.,
electric motors, relays and transformers. In this case, the magnetic induction is the dominant
factor, and the contribution of the displacement currents is negligible in comparison to the
currents, i.e. | D
t | |j|. We can therefore neglect the contribution of the displacement
currents and use Amperes law in its original form to obtain
B
+ curl E = 0,
t
div B = 0,
curl H = j,
div D = .
Note that a coupling between magnetic and electric fields occurs only in conducting materials.
The quasi-static electromagnetic equations in vector potential formulation are given by

A
+ curl 1 curl A = j i
t

(2.28)

A
with B = curl A and E = A
t ; the conduction current is given by j c = t . The time
harmonic magneto-quasistatic vector potential formulation reads

curl 1 curl A + iA = j i .

2.3.3

(2.29)

Static field equations

If the electromagnetic field is generated only by static or uniformly moving charges, we can
assume all fields to be time-independent. We can skip the terms involving time-derivatives in
Faradays and Amperes law, namely (2.5a) and (2.5c).
Electrostatic fields can only occur in non-conducting regions ( = 0). Hence the magnetic
and the electric fields are decoupled and we can deduce two independent systems.
The electrostatic problem

is described by the system

curl E = 0,
div D = ,
D = E.

By introducing the scalar potential such that E = , the electrostatic problem simplifies
to the Poisson-equation
div() = in .
(2.30)
Note that the potential is defined only up to constants. Suitable boundary conditions can
be derived from the boundary conditions for the electric field and its flux. Given surface
charges D n = S are introduced as a natural boundary condition

= S
n

on N .

(2.31)

14

CHAPTER 2. FUNDAMENTALS OF ELECTROMAGNETICS

In case of an adjacent perfect electric conductor (PEC) we require n = 0, which is

satisfied if the potential is constant on each connected part PEC,i of the PEC boundary, i.e.,
= Ui
Here, the constants Ui =

on PEC,i .

(2.32)

E n dA are the applied voltages at the boundary PEC,i .

The magnetostatic problem

simplifies to
curl H = j i ,
div B = 0,
B = H.

Applying the divergence on the first equation shows that impressed currents have to be divergence free, i.e.
div j i = 0,
(2.33)
in order to provide consistency. We again introduce a vector potential A satisfying B =
curl A. Then the magnetostatic vector potential problem reads
curl 1 curl A = j i .

(2.34)

As outlined above, the vector potential A of the magnetostatic problem is only defined up to
gradient functions. Uniqueness can be enforced, e.g., by Coulomb Gauge, where we additionally require
div A = 0 in .
(2.35)
Essential boundary conditions
A n = 0 on B ,

(2.36)

imply that the magnetic flux through the boundary is zero, i.e. B n = 0. Impressed surface
currents (H n = j S ) are introduced by natural boundary conditions
1 curl A n = j S

on H

(2.37)

The homogeneous condition 1 curl A n = 0 models magnetic symmetry planes or adjacent

perfectly magnetic conduction regions. We note that in the case of multiple non-connected
parts of H , one has to pose some further integral conditions for the magnetic flux over
each several part of H or the magnetic voltage between these parts. For details we refer to
[18].
Bro

2.4

The general curl-curl problem

The time-harmonic and magneto-static problems discussed in the previous section have a
common structure, which suggest to treat them - analytically and numerically - in a common framework. Note that similar problems also arise in the solution of Maxwell eigenvalue
problems (see Chapter 7), in the solution of time dependent problems in each time step of
a numerical integration scheme, but also in the solution of nonlinear problems via Newtons
method. The general structure is that of the the following curl-curl problem:

2.4. THE GENERAL curl-curl PROBLEM

15

Problem 2.1. Let R3 a bounded domain with boundary = D N . Find u such

that
curl 1 curl u + u = f
(u n) n = gD

curl u n = gN

in R3 ,

on D ,

on N .

The parameter is defined problem-depending as

= 0 for the magnetostatic problem (2.34),
= i 2 C for the time-harmonic Maxwell equations in the vector potential
formulation (2.27),
= i for quasi-static problems (2.29) in case of low-frequency applications,
= 2 R for high-frequency applications (electromagnetic waves in cavities).
For the sake of simplicity we assume below that the paremeters , , are scalars, i.e. the
underlying material is isotropic.
In order to be able to treat the various problems in this common framework, we have to focus
on numerical methods that are robust with respect to the parameter . This will be one of
the key aspects in our analysis.
In order to show applicability in all regimes of values, we will investigate magneto-static,
time-harmonic quasistatic (low-frequency), and eigenvalue problems in more detail.

16

CHAPTER 2. FUNDAMENTALS OF ELECTROMAGNETICS

Chapter 3

Function Spaces and Variational

Formulations
In the previous chapter we discussed several problems governed by Maxwells equations. The
electrostatic problem resulted in a Poisson-type problem
div() = f

+ b.c.

(3.1)

for the scalar potential . In time-harmonic formulations as well as in time-stepping schemes

we obtained curl-curl problems of the form
curl 1 curl u + u = j

+ b.c.,

(3.2)

with a vector field u, a given current density j, permeability and a problem depending
parameter C.
In this chapter we introduce the natural function spaces for a variational formulation of
the partial differential equations under consideration, i.e., H 1 (), H(curl, ), and H(div, ),
and provide a short overview of definitions and basic results on weak derivatives and trace
operators. We then discuss in some detail properties of the involved differential operators,
namely the gradient, the divergence and the curl operator, and show that the functions spaces
are naturally connected via the differential operators. This leads to the formalism of the de
Rham Complex, which is an important property of the function spaces on the continuous
level that should also be conserved for finite dimensional approximations. Hence de Rham
complexes will play an important role in our construction of finite element spaces as well as
design and analysis of preconditioners and error estimators presented below. Finally, we recall
the basic existence and uniqueness results for variational problems and derive the existence
and uniqueness of solutions to the electromagnetic problems under consideration.

3.1

Vector-valued Function Spaces, Trace Operators and

Greens formulas

We start with the definition and basic properties of the operators involved in the various
formulations of Maxwells equations presented so far: Let Rd , d = 2, 3 be a domain. The
gradient operator of a scalar function is defined as



:=
(3.3)
, ...,
x1
xd
17

18

CHAPTER 3. A VARIATIONAL FRAMEWORK

and the divergence of a vector v = (v1 , ..., vd ) is defined as

div v = v :=

d
X
vi
.
xi

(3.4)

i=1

In two dimensions, there are two definitions of the curl-operator. For R2 , the vector curl
operator of a scalar function q is defined as
Curl q := (

q
q T
,
) ,
x2 x1

(3.5)

and the scalar curl operator acting on a vector function v = (v1 , v2 ) as

curl v = v :=

v2
v1

.
x1 x2

(3.6)

For R3 the curl operator of a vector function v = (v1 , v2 , v3 ) is defined as

curl v = v :=

v2
v3 v1
v3 v1
v3

x3 x2 x3 x1 x3 x1

T

(3.7)

We next recall the definition of derivatives in the weak sense:

Definition 3.1 (Generalized differential operators).
1. For w L2 () we call g = w the (generalized) gradient of w, if there holds
Z
Z

3
g v dx = w div v dx v C0 () .

3
2. For u (L2 ())3 we call c = curl u L2 () the (generalized) curl of u, if there
holds
Z
Z
c v dx =
u curl v dx v [C0 ()]3 .

3. For q L2 ()
holds

3

we call d = div q L2 () the (generalized) divergence of q, if there

Z
Z
d v dx = q v dx v C0 ().

The following function spaces will turn out to provide a natural setting for the investigation
of the PDEs discussed above.
Definition 3.2 (Function Spaces). Let us define the spaces
Z
L2 () := {u |
u2 dx < }

H 1 () := { L2 () | [L2 ()]3 }

H(curl, ) := {v [L2 ()]3 | curl v [L2 ()]3 }

H(div, ) := {q [L2 ()]3 | div q L2 ()}

19

3.1. FUNCTION SPACES, TRACE OPERATORS AND GREENS FORMULAS

and the corresponding scalar products by

R
(u, v)0
= uv dx
R
(, )1
= dx + (, )0
R
(u, v)curl = curl u curl v dx + (u, v)0
R
(p, q)div = div p div q dx + (p, q)0

The induced norms are denoted by k k0 , k k1 , k kcurl , and k kdiv .

We only mention that elements in the above spaces have to be understood as equivalence
classes, i.e., two functions are considered to be equal (belong to the same equivalence class)
if they coincide up to a set of zero measure. All the above spaces are Hilbert spaces when
equipped with the corresponding scalar products, in particular they are complete.
For the results of this section we need some conditions on the underlying domain and its
boundary := .
Definition 3.3 (Lipschitz-domain). The boundary of a domain R3 is called Lipschitzcontinuous, if there exist a finite number of domains i , local coordinate systems (i , i , i ),
and Lipschitz-continuous functions b(i , i ) such that
S
i with i = {(i , i , i )|i = b(i , i )} ,
i = {(i , i , i )|i > bi (i , i )}.

is then called a Lipschitz-domain.

The boundary of a Lipschitz-domain can be represented locally by the graph of bi and

lies locally on one side of this graph. We want to mention that this definition allows
domains with corners, but cuts are excluded. We refer to McLean [67] for other examples
of non-Lipschitz domains. Also note that a Lipschitz-boundary enables the definition of an
outer unit vector n almost everywhere on .
For Lipschitz-domains with bounded the following density results hold (see GiraultRaviart [49]):

H 1 () = C ()

kk1

H(curl, ) = C ()

kkcurl

, and H(div, ) = C ()

kkdiv

(3.8)

which allows to easily extend classical properties of C functions also to the above function
spaces.
In the sequel, we recall important trace and extension theorems, and different versions
of Greens formula. Those results will be needed to state the differential equations in a
variational setting and to consider boundary and interface conditions.

Trace theorems for the space H 1 ()

The trace of a smooth function C() is defined pointwise as
(tr u)(x) = u(x) x ,

briefly tr (u) = u| .

Using the density result (3.8) we can extend the classical trace concept to more general
function spaces:

20

CHAPTER 3. A VARIATIONAL FRAMEWORK

Theorem 3.4 (Trace theorem and Greens formula for H 1 ()). Let Rd , d = 2, 3 be a
bounded Lipschitz-domain.
1. The classical trace mapping tr defined on C () can be uniquely extended to a continuous, linear operator
tr : H 1 () H 1/2 ()
kk1/2

where H 1/2 () := C ()
 refers to up to a constant.

with

ktr (v)k1/2  kvk1 v H 1 (),

and kuk21/2 = kuk20 +

|u(x)u(y)|2
|xy|d

dx dy, where

2. There holds the integration by parts formula

Z

u v dx =

u div v dx +

tr (u) v n dx

u H 1 () v H(div, ).
(3.9)

3. Let g H 1/2 (). Then there holds

v H 1 () : tr (v) = g with kvk1  kgk1/2 .
The third statement is an extension theorem, which is important for incorporation of Dirichlet
boundary conditions. Here and below, we use Sobolev spaces of fractional order on manifolds,
which can be defined similarly as in the theorem above, or as traces of certain classes of
functions on ; for their definition we refer in particular to McLean [67].
The following corollary discusses the interface conditions of H 1 -functions.
Corollary 3.5. Let 1 , . . . , m be bounded Lipschitz
S domains and a non-overlapping domain
decomposition of , i.e. i j = for i 6= j and i = with interfaces ij := i j .
Suppose ui := u| H 1 (i ) and trij ui = trij uj . Then
i

u H 1 ()

and

(u)|i = ui .

This result is fundamental for the construction of conforming finite element methods, where
the discrete space is chosen as a subspace of H 1 (). If the finite element functions, which are
defined elementwise, are continuous over the element interfaces, then the global FE-function is
kk1
is equal to the space of H 1 functions
in H 1 (). Note also that the space H01 () := C0 ()
with homogenous Dirichlet boundary conditions, i.e.,
H01 () = {u H 1 () | tr u = 0}.
For Poisson problems with pure Neumann boundary conditions, we will utilize the quotient
space,
Z
H 1 ()/R = {u H 1 () |
u dx = 0}.

21

3.1. FUNCTION SPACES, TRACE OPERATORS AND GREENS FORMULAS

Some results concerning the space H(curl)

d , d = 2, 3:
Let us consider the following tangential traces for vector functions v [C ()]
tr (v)(x) := v(x) (x)
tr (v)(x) := v(x) n(x)

for x and R2 ,
for x and R3 ,
for x and R3 ,

trT (v)(x) := (v(x) n(x)) n(x)

with denoting the tangential vector and n the outer unit normal vector on := . Due
to the density result (3.8) we can extend the trace mapping to a wider class of functions:
Theorem 3.6 (Trace theorem and integration by parts in H(curl)). Let be a bounded
Lipschitz-domain.
1. The classical trace map tr can be extended from [C ()]d to a continuous and linear
map (still denoted by tr )
tr : H(curl, ) [H 1/2 ()]m ,

and

ktr (v)k1/2  kvkcurl v H(curl),

where m = 3 if R3 or m = 1 if R2 .
2. There holds the integration by parts formula
Z
Z
Z
curl u dx =
uCurl dxs tr (u)dx

d
u H(curl, ) H 1 () .

(3.10)

For m = 3 Curl and curl are defined to be the same.

In 3D, the generalized tangential trace operator tr is not surjective onto H 1/2 (), its
range coincides with



H 1/2 (div, ) := v (H 1/2 ())3  v n = 0 a.e.on , div v H 1/2 () .
This results in the following

Theorem 3.7 (Extension theorem for H(curl)). For g H 1/2 (div, ) there exists a v
H(curl, ) with
tr (v) = g

and

kvk2curl  kgk21/2 + k div gk21/2 .

The closure, with respect to the H(curl)-norm, of the space of infinitely differentiable functions
with compact support in is denoted by
H0 (curl) := [C0 ()]3

kkcurl

This space is equal to the space of H(curl) functions with homogenous tangential boundary
conditions, i.e.,
H0 (curl, ) = {u H(curl, ) | (u n) n = 0}.
The following corollary treats the interface conditions for H(curl)-functions.

22

CHAPTER 3. A VARIATIONAL FRAMEWORK

Corollary 3.8.
S Let 1 , . . . , m be a non-overlapping domain decomposition of , i.e. i
j = and i = with interfaces ij = i j . Suppose ui := u| H(curl, i ) and
i
trTi ,ij ui = trTi ,ij uj . Then
u H(curl, )

and

(curl u)|i = curl ui .

Thus conforming finite element spaces for discretization of H(curl) can be constructed by
requiring the tangential components to be continuous over the element interfaces. This ensures
that the resulting global finite element functions are in H(curl, ). Note that the normal
components of H(curl) functions need not to be continuous.
Basic results on the H(div) space
d , d = 2, 3 the normal trace is defined by
For v (C0 ())
(trn(v))(x) := v(x) n(x)

x on := ,

d is dense in
where n denotes the outward unit normal vector on . Utilizing that [C ()]
H(div, ), the trace operator can be extended to the whole H(div) space.
Theorem 3.9 (Trace theorems and integration by parts for H(div, )). Let be a bounded
Lipschitz-domain.
1. The trace map trn can be extended as a bounded, linear mapping (still denoted by trn)
trn : H(div, ) [H 1/2 ()]3 ,

ktrn(v)k1/2  kvkdiv v H(div, ).

and

2. There holds the following version of Greens theorem:

Z
Z

3
div u dx = u dx + htrn(u), i
u H(div, ), H 1 () ,

(3.11)

where h, i denotes the duality product h, i

H 2 H 2

3. Suppose g H 1/2 (). Then there holds the extension theorem:

v H(div, ) :

trnv = g

and

kvkdiv  kgk1/2 .

The closure of arbitrary differentiable functions with compact support in in the H(div)norm is denoted by
H0 (div) := [C0 ()]3

kkcurl

and equals the space of H(div) functions with homogeneous normal boundary condition, i.e.,
H0 (div, ) = {v H(div, ) | trn, (v) = 0}.
The following corollary deals with appropriate interface conditions for H(div)-functions.
Corollary 3.10.
Let 1 , . . . , m be a non-overlapping domain decomposition of , i.e. i
S
j = and i = with interfaces ij = i j . Suppose ui := u| H(div, i ) and
i
ui |ij ni = uj |ij ni . Then
u H(div, )

and

(div u)|i = div ui .

Due to this corollary conformity of finite element functions in H(div) can be guaranteed by
requiring continuity of the normal components across element interfaces.

23

3.2. MAPPING PROPERTIES OF DIFFERENTIAL OPERATORS

3.2

Mapping Properties of Differential Operators

In the previous chapter we introduced vector and scalar potential formulations for Maxwells
equations in their classical, strong form. Now, we want to extend this concept also to the
generalized differentiability setting presented in this chapter. The question on the existence
of potential fields is strongly connected to the characterization of the kernel and the range of
the involved differential operators.
The following notation for the kernel of the differential operators will be used:

ker() := {v H 1 ()  v = 0},

ker(curl) := {v H(curl, )  curl v = 0},

ker(div) := {v H(div, )  div v = 0},

The corresponding range spaces are denoted by H 1 (), curl H(curl, ), div H(div, ).
The well-known identities
curl() = 0,

(3.12a)

div(curl u) = 0,

(3.12b)

which are trivially satisfied for twice continuously differentiable functions, can be generalized
to the Hilbert spaces under consideration in the following way:
H 1 () ker(curl),

curl H(curl, ) ker(div),

div H(div, ) L2 ().

(3.13a)
(3.13b)
(3.13c)

Identities instead of inclusions hold in general only under additional assumptions on

the domain . For the following statements, we assume to be a bounded, simplyconnected Lipschitz-domain (the case of multiply-connected domains is treated in GiraultRaviart [49] and Dautray-Lions [40]. For more general domains, namely pseudo-Lipschitz
domains where also cuts are allowed, we refer to Amrouche et al. [4]).
Assumption 3.11. Let be a bounded, simply-connected Lipschitz-domain.
The nullspace (in H 1 ()) of the gradient operator is the set of constant functions.
Conse
quently, the nullspace can be made trivial by restriction to H01 () or H 1 () R:
ker(, H 1 ()) = R

and

ker(, H01 ()) = ker(, H 1 ()/R) = {0}.

The classical Stokes theorem on the existence of a scalar potential of curl-free functions can
be extended in the following way, cf. [49, Theorem 2.9].
Theorem 3.12 (Existence of scalar potentials). Let u L2 ()d denote a vector field. Then
curl u = 0 in
if and only if there exists a scalar-potential H 1 () such that u = , where is
unique up to an additive constant.
if and only if there exists a unique scalar-potential H 1 ()/R such that u = .

24

CHAPTER 3. A VARIATIONAL FRAMEWORK

In other words, (H 1 ()) = ker(curl).

Theorem 3.13 (Existence of vector potentials, cf. [49, Theorem 3.4, 3.5, and 3.6]). Let

3
u L2 () denote a vector field.
1. u is divergence-free

div u = 0


3
if and only if there exists a vector potential H 1 () such that
u = curl .

Furthermore, can be chosen such that div = 0.

2. Let u ker(div). Then there exists a unique vector potential H(curl, ) such that
curl = u,

div = 0, and n = 0.

3. Let u ker(div) and trn, (u) = 0. Then there exists a unique vector potential
H(curl, ) such that
curl = u,

div = 0, and tr, = n| = 0.

We only mention that in both cases (2. and 3.) the vector potential can be derived as the
solution of a boundary value problem for the Laplace operator. Note that Theorem 3.13
implies
ker(div) = curl(H(curl, )).
A substantial tool within the analysis of the Maxwell equations is the Helmholtz-decomposition:
Every vector field in L2 ()3 can be decomposed into a divergence- and a curl-free function,
which can be specified by means of vector and scalar potentials.
Theorem 3.14 (Helmholtz decomposition). Every vector field u L2 ()3 has an orthogonal
decomposition
u = + curl ,

with H 1 () and H(curl, ).

Here, in H 1 ()/R being the solution of the Laplace-problem = div u ; the

vector potential H(curl, ) is as in Theorem 3.13. For other choices of the vector
and scalar potential in the Helmholtz decomposition we refer to Amrouche et al. [4],
Girault-Raviart [49] or Dautray-Lions [40].
For stating the De Rham Complex we need one last result on the image of the divergence
operator.
Lemma 3.15. The divergence operator is surjective
from H(div, ) onto L2 () as well as
R

from H0 (div, ) onto L2,0 () := {q L2 () q dx = 0}.

Proof. For f L2 () we choose u = L2 () with H01 () solution of the Dirichlet

problem
(div(), ) = (, ) = (f, ) H01 ().

Since div u = f , there holds u H(div). For f L2,0 () we choose u = L2 () with

H 1 () being the solution of the Neumann problem
(div(), ) = (, ) = (f, ) H 1 ().

Since u n| = n| = 0 and div u = f , there holds u H0 (div).

25

3.2. MAPPING PROPERTIES OF DIFFERENTIAL OPERATORS

3.2.1

The de Rham Complex and Exact Sequences

The relation between the functional spaces through their differential operators can be summarized in the de Rham Complex, which reads as

id

curl

div

R H 1 () H(curl, ) H(div, ) L2 () {0}.

For domains

R2

(3.14)

the De Rham Sequence shortens to

id

curl

(3.15)

id

Curl

div

(3.16)

R H 1 () H(curl, ) L2 () {0},
respectively
R H 1 () H(div, ) L2 () {0}.

The main property of the de Rham Complex is the coincidence of ranges and kernels of
consecutive operators.
Corollary 3.16. The de Rham Complexes (3.14)-(3.16) form exact sequences, also knnown
as complete sequences. This means that the range of each operator coincides with the kernel
of the following operator.
In fact the exactness of the sequence summarizes the results of Theorem 3.12, Theorem 3.13,
and Lemma 3.15, namely the coincidence of the following range and kernel spaces:
ker() = R,

(3.17a)
1

ker(Curl) = H (),

(3.17b)

ker(div) = curl(H(curl, )),

(3.17c)

L2 () = div(H(div, )),

(3.17d)

for 3 dimensional spaces, whereas in 2 dimensional spaces there holds

ker() = R,

(3.18a)

ker(Curl) = R,

(3.18b)
1

ker(curl) = H (),

(3.18c)

ker(div) = Curl H (),

(3.18d)

L2 () = curl(H(curl, )),

(3.18e)

L2 () = div(H(div, ).

(3.18f)

Remark 3.17 (De Rham Complex with essential boundary conditions). In case of essential
boundary conditions on , the restricted sequence
id

curl

div

R H01 () H0 (curl, ) H0 (div, ) L2,0 () {0},

(3.19)
R
where L2,0 () := {q L2 ()  q dx = 0}, is still exact.
In the two-dimensional setting the exactness of the shortened de Rham Complexes (3.15) and
(3.16) still holds in the case of essential boundary conditions.
In case that the domain is more complex, i.e. involves holes or if mixed boundary conditions
are imposed, the de Rham Complex need not be exact: the range of an operator is a subset of
the kernel of the following map, but the kernel and the range spaces need not coincide. There is
a low dimensional subspace of kernel functions which cannot be represented as gradient fields
of a scalar potential. The dimension of this space depends on the topological properties of .
The issue of so-called cohomology spaces is treated in Amrouche et. al. [4], Bossavit [25]
and Hiptmair [56].

26

CHAPTER 3. A VARIATIONAL FRAMEWORK

3.3

Abstract Variational Problems: Existence and Uniqueness

In this section, we briefly recall the main existence and uniqueness results for variational
problems. We will then apply these theorems to our classes of problems for Maxwells
equations.
Let V and W denote Hilbert-spaces provided with the scalar products (, )V , (, )W . The
induced norms are denoted by k kV , k kW . By V we refer to the dual space, and the duality
pairing is denoted by h, i. A mapping a(, ) : V W C is called a sesquilinear form if
a(c1 u1 + c2 u1 , v) = c1 a(u1 , v) + c2 a(u2 , v)
a(u, c1 v1 + c2 v1 ) = c1 a(u, v1 ) + c2 a(u, v2 )

c1 , c2 C, u1 , u2 V, v W,
c1 , c2 C, u V, v1 , v2 W,

where c1 denotes the complex conjugate of c1 .

The following properties are at the core of the basic results below:
Definition 3.18. A sesquilinear form is called
1. bounded if
|a(u, w)| kukV kwkW

u V, w W.

2. coercive if V = W and,
> 0 :

3.3.1

|a(u, u)| kuk2V

u V.

Coercive variational problems

We investigate variational problems of the form

Problem 3.19. For given f V , find u V such that
a(u, v) = f (v)

v V.

(3.20)

Existence and uniqueness of a solution is guaranteed by

Theorem 3.20 (Lax-Milgram). Let a : V V C denote a bounded and coercive sesquilinar
form. Then, for any continuous linear form f V , there exists a unique solution u V
satisfying

a(u, v) = f (v)
v V,
and
kukV kf kV ,

where and are as in Definition 3.18.

3.3.2

Mixed formulations

In the context of magnetostatic problems we utilize a mixed formulation of the variational

problem. We consider mixed problems also called saddle-point problems of the following
type:
a(u, v) + b(v, p) = f (v)
v V,
(3.21)
b(u, q)
= g(q)
q W.
The equivalent to the results of the Lax-Milgram Theorem is given by the following result.

3.4. VARIATIONAL FORMULATION OF ELECTROMAGNETIC PROBLEMS

27

Theorem 3.21 (Brezzi). Let V and W be Hilbert-spaces, and the sesquilinear forms a :
V V C and b : V W C fulfill the following properties:
1. The sesquilinear forms are bounded, i.e.
1 0 :
2 0 :

|a(v, w)| 1 kvkV kwkV

|b(v, q)| 2 kvkV kqkW

v, w V,
v V q W.

2. b(, ) satisfies the Babuska-Brezzi condition, i.e.

2 > 0 :

sup
vV

b(v, q)
2 kqkW
kvkV

q W.

(3.22)

3. a(, ) is ker b-coercive, i.e.

1 > 0 |a(v, v)| 1 kvk2V

where ker b := {u V  b(u, q) = 0 q W }.

v ker b,

(3.23)

Then there exists a unique solution (u, p) V W of (3.21) satisfying the a-priori estimates
1
1

1
1+
kgkW ,
kf kV +
1
2
1
1

1
1

1
1+
kf kV + 2 1 +
kgkW .

2
1
1
2

kukV
kpkW

3.4

Variational formulation of electromagnetic problems

We will now utilize the definitions and results of the previous sections to state the electromagnetic problems discussed in Chapter 2 in a variational form, and to prove existence and
uniqueness of solutions. Throughout we assume to be a bounded Lipschitz domain. The
general structure of the derivation of variational formulations is to multiply the partial differential equations by a suitable test function, integrate over , and apply integration by parts
(Greens formula).

3.4.1

The electrostatic problem: A Poisson problem

In the electrostatic regime, the four Maxwell equations can be reduced to a single Poisson
equation for the scalar potential of the electric field intensity E, i.e.,
div() =
=gD

=gN
n

in ,
on D ,
on N ,

= D N ,

It is well-known that subspaces of H 1 () are appropriate for the variational formulation in

this case. By including the Dirichlet boundary conditions in the ansatz space, we arive at


Problem 3.22. Find Hg1D ,D = { H 1 ()  tr = gD } such that
D
Z
Z
Z
1
(x)(x) (x) dx =
(x) (x) dx +
(x) gN (x) (x) ds
H0,D
().

28

CHAPTER 3. A VARIATIONAL FRAMEWORK

RFor bounded permittivity, i.e. 0 < 0 1 almost everywhere, the bilinear form a(, ) :=
dx is
bounded with |a(, )| 1 kkV kkV .

2
coercive, since a(, ) c1
es-Friedrichs inequality
F 0 kk1 owing to Poincar

kk0 cF kk1

1
H0,D
()

where we assumed measRd1 (D ) > 0.

Corollary 3.23. Suppose that the permittivity is bounded by 0 < 0 (x) 1 , the boundary
1
1
data satisfy gD H 2 (D ) with measRd1 (D ) > 0 and gN H 2 (N ), and H 1 ().
Then Theorem 3.20 (Lax-Milgram) guarantees the existence of a unique solution Hg1D ,D ()
for the variational electrostatic potential problem 3.22.
For pure Neumann problems, i.e. N = , the solution of the electrostatic problem is
unique
R only up to constants. Uniqueness can be restored by imposing a gauging condition,
e.g. dx = 0. The electric field E can be recovered by E = . Note that by Theorem
3.12 we have E H(curl) and curl E = 0.

3.4.2

The magnetostatic problem

Next we consider the curl-curl problem (2.34): As it becomes clear from the variational
formulation below, the appropriate spaces (incorporating the boundary conditions (2.36),
(2.37)) are subspaces of H(curl, ).


Problem 3.24. Find u HgD ,D (curl, ) := v H(curl, ) : v n = g D on D such
that
Z
Z
1
curl u curl v dx =
j v dx
v H0,D (curl, ).
(3.25)

Note that we assume the impressed currents j to be consistent here, i.e. to satisfy the
continuity equation
div j = 0 in and j n = 0 on .
(3.26)
R
1 () there
The bilinear form a(u, v) = curl u curl v dx is not coercive, since for all H0,D
holds a(u, ) = 0, while kkcurl = kk0 . Therefore, we cannot apply the Lax-Milgram
Theorem directly. As in the potential problem with pure Neumann boundary conditions
above, uniqueness and coercivity can be restored by appropriate gauging. For simplicity, we
assume for the moment homogeneous boundary conditions (e.g., the boundary conditions are
incorporated in the right hand side by homogenization). The Coulomb gauging condition
div u = 0 with u n = 0 then reads
(u, ) = 0,

W = H 1 ().

Hence we look for a solution u which is orthogonal to gradients. In case of Dirichlet data, we
alternatively set W = H01 ().

3.4. VARIATIONAL FORMULATION OF ELECTROMAGNETIC PROBLEMS

29

Mixed formulation of the magnetostatic problem

We incorporate the gauging condition u H 1 ()
by reformulating the magnetostatic
problem as a mixed variational problem.
Problem 3.25. Find (u, ) in V W := H0 (curl, ) H01 () such that
Z
Z
Z
1
curl u curl v dx +
v dx =
j v dx
v H0,D (curl, ),

Z
u dx
=0
H01 ().

(3.27a)
(3.27b)

We further assume that the permeability

is bounded by 0 < 0 (x) 1R, and denote the
R
involved bilinear forms by a(u, v) := 1 curl u curl v dx and b(v, ) := v dx.

In order to be able to apply Brezzis Theorem we have to prove the ker b-coercivity of a(, ).
For this purpose we require the following theorem.

Theorem 3.26 (Friedrichs inequality for H(curl)). Let be a simply-connected Lipschitz

domain.
1. Suppose v H(curl, ) is orthogonal to gradient functions, i.e. (v, ) = 0 for all
H 1 (). Then
kvk0  k curl vk0 .
2. Suppose v H0 (curl, ) is orthogonal to gradient functions, i.e. (v, ) = 0 for all
H01 (). Then the same estimate holds.
For a proof we refer to Lemma 3.4 and Lemma 3.6 in Girault-Raviart [49].
Applying the H(curl) Friedrichs inequalities, the bilinear forms of the magnetostatic problem
can be shown to satisfy the following properties.:
The bilinear forms a(, ) and b(, ) are bounded, i.e., |a(u, v)| 0 kukV kvkV and by
|b(v, )| kukV kkW .
b(, ) satisfies the Babuska-Brezzi condition, as there is
R
R
()
()
dx ()
v dx
sup

kk0  kk1
kvkcurl
kkcurl
vV
by the choice v = (), kkcurl = kk0 (), and Friedrichs inequality for H 1 ()
( ).
a(, ) is ker b-coercive, i.e. for all u V satisfying (u, ) = 0 H 1 () there holds
()

1
1
a(u, u) 1
1 k curl uk 1 (1 + c F,curl ) kukcurl .

The main step (*) in this estimation is justified by Friedrichs inequality for H(curl)
stated in Theorem 3.26.
Existence and uniqueness for the variational problem in mixed formulation 3.25 now follows
by Brezzis theorem 3.21:

30

CHAPTER 3. A VARIATIONAL FRAMEWORK

Corollary 3.27 (Solution theory of mixed and standard magnetostatic problem). Let the
permeability satisfy 0 < 0 (x) 1 , and the source term j H(curl) satisfy (3.26).
Then
1. Brezzis theorem 3.21 ensures the existence and uniqueness of a solution (u, ) V W
to the magnetostatic problem in mixed form (3.27).
2. the solution of the mixed problem is of the form (u, ) = (u, 0) with u solving the
standard problem (3.25) with div u = 0 and u n = 0.
3. the estimate kukcurl  kjkV holds.
Proof. Let (u, ) V W denote the unique solution (implied by Brezzis theorem) of the
mixed problem (3.27). Testing (3.27a) with gradient functions W RV yields that
W solves the Laplace equation. Having in mind that a(u, )+b(, ) = dx,
we obtain
Z
Z
dx =
j dx

Z
Z
=
div j dx +
j n dx = 0
W.

The source term vanishes since j was assumed to be consistent, i.e. div j = 0, j n = 0.
Since in case of the Neumann problem W = H 1 ()/R and in case of the Dirichlet problem
W = H01 (), we obtain a unique solution = 0 W . Hence, the solution of the mixed
problem (3.27) is of the form (u, = 0).
Substituting (u, = 0) in the mixed problem the first equation (3.27a) yields that u solves
the magnetostatic problem in standard form (3.25). The second equation (3.27b) implies that


we achieve u H 1 () .
While the mixed formulation provides analytic results on existence and uniqueness of solutions
in an elegant manner, a numerical realization has certain drawbacks: first the number of
unknowns is increased by switching from Problem 3.24 to 3.25. Additionally, the resulting
saddle point problems are usually more difficult to solve. In view of a numerical realization
we therefore discuss another regularized version of the magnetostatic problem.
Regularized magnetostatic problem
We choose a regularization parameter > 0 and define the regularized magnetostatic
problem as
Problem 3.28. Find u V s.t.
Z
Z
Z

curl u curl v dx + u v dx =
jv dx

v V.

(3.28)

As we will see, the regularized Problem 3.28 has the following properties:
The Lax-Milgram theory is applicable, i.e. the involved bilinear form is coercive.
The solution u of the regularized problem converges to a solution u solving the standard
problem with 0.

3.4. VARIATIONAL FORMULATION OF ELECTROMAGNETIC PROBLEMS

Let us denote the regularized bilinear form by a (u, v) :=

curl u curl v dx +

1
2
a (u, u) 1
1 (curl u, curl u) + (u, u) min{1 , } kukcurl

31

u v dx.

u H(curl, ). (3.29)

Corollary 3.29. Under the assumptions of Corollary 3.27 and a positive regularization parameter > 0 the bilinearfrom a (u, v) is coercive. Hence, the Lemma of Lax-Milgram (Theorem 3.20) provides the unique solvability of the regularized magnetostatic problem (3.28).
The result follows easily since there holds
1
2
a (u, u) 1
1 (curl u, curl u) + (u, u) min{1 , } kukcurl

u H(curl, ). (3.30)

In order to show that u converges to a uniquely determined solution u of Problem 3.24 as

0, we utilize the mixed formulation. Consider the following mixed formulation of the
regularized problem (3.28): Find (u , ) in V W := H0 (curl, ) H01 () such that
Z
Z
Z
Z
1 curl u curl v dx + u v dx +
v dx =
j v dx
v V, (3.31a)

Z
u dx
=0
W. (3.31b)

Corollary 3.30. Let the assumptions of Corollary 3.27 hold. Then the following assertions
hold:
1. The regularized mixed problem (3.31) has a unique solution (u , = 0) V W


satisfying ku kV  kjkV and u W .

2. If (u , ) denotes the unique solution of the regularized mixed problem, then u solves
the regularized standard problem (3.28).
3. Suppose that (u, = 0) denotes the solution of the original mixed problem (3.27). Then
ku u kV  kjkV .

(3.32)

Proof.
1. Under the assumptions of Corollary 3.27 Brezzis theorem ensures the existence of a
unique solution (u , ) of the mixed problem (3.31).
Testing the first equation (3.31a) on the subspace W V and taking into account the
consistency (3.26) of the source term we achieve
Z
Z

u dx +
dx = 0
W.

Since u is orthogonal on gradients due to (3.31b), we obtain a homogenous Laplace

equation for , i.e. = 0. Since the constants in the inf-sup-condition and in the
ker b-coercivity are independent of the stability estimate is implied by one of Brezzis
theorem.
2. Setting = 0 in the mixed formulations implies that u solves the regularized standard
problem with u (H01 ()) .

32

CHAPTER 3. A VARIATIONAL FRAMEWORK

3. Let (u, ) and (u , ) denote the solutions of the exact and the perturbed mixed problem. Then (u u , ) V W satisfies
R
R
R 1
curl(u u ) curl v dx + v ( ) dx = u v dx v V,

= 0
W,
(u u ) dx
i.e. the exact mixed formulation with consistent right hand-side u . Due to Brezzis
theorem (u u ) is bounded by the source term, i.e. ku u kV ku kV . The
estimate of item 1, namely ku kV  kjkV (where the constant is independent of )
implies the desired estimate.

Remark 3.31. For a numerical realization it will be important to construct solvers for Problem 3.28 that are robust in the parameter , in particular as 0.

3.4.3

The time-harmonic electromagnetic and magneto-quasi-static problems

Now we turn to a variational formulation of the problem

curl 1 curl u + u = j,
1 curl u n = g N

on N ,

u n = gD

on D ,

which includes the time-harmonic eddy-current problem where = i as well as the timeharmonic electromagnetic wave equation in conducting media with = i 2 for > 0
and > 0. In both cases the parameter C\R .
Problem 3.32. Find u V := Hg,D (curl, ) satisfying
Z
Z
Z
Z
1
dx +
dx =
dx +
curl u curl v
uv
j v

The sesquilinear form a(u, v) =

properties:

1 curl u

ds
(n g N ) v

dx +
curl v

v V. (3.33)

dx has the following

v

a(, ) is bounded if the parameters , are bounded, i.e., for all u, v H(curl, ) there
holds
|a(u, v)| k1 k |u|curl |v|curl + kk kuk0 kvk0
 kukcurl kvkcurl .

a(, ) is coercive: For complex = R + iI with I 6= 0 there exits = R + ii C

with || > 0 such that
) + (R R i i )(u, u
)
|||a(u, u)| |Re(a(u, u)| |R 1 (curl u, curl u
 kuk2curl .

The choice = + i Ri1 implies R R i i = 1 and R 1 = 1. Coercivity for the

case = R > 0 was already discussed in the regularized magnetostatic case.
1

Corollary 3.33. Let the boundary data satisfy g D H 2 (D ) and gN H 2 (N ), and

assume j H(curl, ) . Then the Lax-Milgram Theorem implies the unique solvability of the
variational problem (3.33).

3.4. VARIATIONAL FORMULATION OF ELECTROMAGNETIC PROBLEMS

33

Non-conducting regions = 0
In non-conducting regions, the eddy-current problem 3.32 degenerates to the magnetostatic
one and can be treated with similar measures. In particular, we may introduceR a regu.
larized version by either setting 0 < nc c or by adding the regularization term nc u v
We conclude this section by recalling the appropriate function spaces for the various variables:
the electric and magnetic field intensity: E H(curl, ), H H(curl, )
the electric and magnetic fluxes: D H(div, ), B H(div, ).
the scalar potential H 1 () with E = ,
the vector potential A H(curl, ) with B = curl A.

34

CHAPTER 3. A VARIATIONAL FRAMEWORK

Chapter 4

The Finite Element Method

In this chapter, we concentrate on standard finite element methods, namely the h-version
FEM. We will shortly recall the main ingredients for the FEM and recall the main approximation results. We then introduce the classical low-order FE-spaces for H 1 (), H(div, )
and H(curl, ), namely
the low-order scalar elements for H 1 ,
the lowest-order Nedelec elements of first kind and second kind for H(curl), and
the lowest-order Raviart-Thomas elements, which in 3D are also called Raviart-ThomasNedelec elements, and Brezzi-Douglas-Marini elements for H(div).
The low order elements will be used also in the construction of the high order elements in the
next chapter. We will show that by careful construction, the exact sequence property discussed
in Section 3.2.1 carries over to the sequence of finite element spaces, locally (elementwise) as
well as globally.

4.1
4.1.1

Basic Concepts
Galerkin Approximation

Galerkin projection provides a general technique for the construction of discrete approximations to the solution of variational problems
Find u X :

a(u, v) = f (v)

v X.

(4.1)

The infinite-dimensional space X is now replaced by a sequence of finite-dimensional spaces

Xh , which yields the discrete variational problems,
Find uh Xh :

a(uh , vh ) = F (vh )

vh Xh .

(4.2)

In this presentation we investigate conforming methods, where Xh X. The index h typically

refers to a discretization parameter. As we will show below, the discrete solution uh Xh
converges to the solution u X with

h 0 under reasonable assumptions. We expand the
solution uh in terms of a basis (i 1iN of Xh , i.e.,
uh =

N
X
i=1

35

uhi i .

36

CHAPTER 4. THE FINITE ELEMENT METHOD

It suffices to test the variational problem (4.2) only with the basis functions, which yields the
Galerkin system
Au = f .
The system matrix A RN N , load vector f Rn , and solution vector u RN are defined
by






(4.3)
f = f (j ) 1jN , and u = uhi 1iN .
A = a(i , j ) 1i,jN ,
The Finite Element method (FEM) is a special Galerkin method, based on the the following
construction of finite-dimensional subspaces Xh :

The domain is covered by a triangulation Th , i.e., a finite union of non-overlapping

polyhedral elements K.
The space Xh consists of piecewise polynomials, i.e., the restriction of a function vh Xh
onto an element K Th is a polynomial.
Xh has a basis {i } with local supports, i.e., each basis function i is non-zero only on
a few elements.
The latter implies that system matrix A is sparse.

4.1.2

The Triangulation

We assume the domain to be a bounded polygonal or polyhedral domain with Lipschitz

continuous boundary.
Definition 4.1. A triangulation (mesh) Th is a finite non-overlapping subdivision Th =
{Ki }iI of into elements Ki of simple geometry. A triangulation is called regular, if
1. the elements are non-overlapping, i.e. interior(Ki ) interior(Kj ) = if i 6= j,
S
2. the triangulation Th is a covering of , i.e. Ki T Ki = ,

3. the intersection Ki Kj of two different elements (i 6= j) is either empty, or a vertex,

or an edge or a face of both elements. (Hanging nodes are not allowed.)
We will frequently use the following notations
the set of vertices V = {Vi },
the set of faces
F = {Fi },

the set of edges E = {Ei },

the set of cells Th = {Ki }.

The (local) sets of vertices, edges and faces belonging to the element K are denoted by VK ,
EK , and FK .

4.1.3

The Finite Element

Following Brenner-Scott [31], we require three basic definitions for construction of finite
elements. The classical definition is the following, cf., e.g., Ciarlet [35]:
Definition 4.2 (Finite Element). Let
1. the element domain K Rn be a bounded closed set with non-empty interior and piecewise smooth boundary,

37

4.1. BASIC CONCEPTS

2. the space of shape functions PK be a finite-dimensional space of functions on K,

3. the set of nodal variables K = {N1K , N2K , . . . , NkK }, also referred to as degrees of
, where P denotes the dual space of P .
freedom, be a basis for PK
K
K
Then we call the triple (K, PK , K ) a finite element.
The main effort in the construction of a finite element is to verify its unisolvence, i.e. the
basis property of item 3. This can be done by showing the simpler, but equivalent property:
If v PK with NiK v = 0 for i = 1, . . . , k then v 0.
Definition 4.3 (Nodal basis). Let (K, PK , K ) be a finite element. The basis {1 , 2 , . . . , k }
of PK which is dual to the set of nodal variables K , i.e. Ni (j ) = ij , is called the nodal
basis of PK .

4.1.4

The reference element and its transformation to physical elements

A key ingredient in the design, analysis and numerical realization of the FEM is the mapping
trick: The physical (global) finite elements (K, PK , K ) are defined as transfromations of a
of simple shape (e.g. 1D:
P , ), i.e., a reference geometry K
reference (master) element (K,
K
K
segment; 2D: triangle, quadrilateral; 3D: tetrahedron, prism, hexahedron). This construction
has several advantages, e.g., many computations (e.g., numerical integration, derivation) can
be performed a-priori on the reference element.

to physical element K
Figure 4.1: Mapping of reference element K
Suitable transformations

K
K : K

(4.4)

denotes a coordinate system

are continuously differentiable, one-to-one and onto maps. If x
then x = K (
on the reference element K,
x) is the corresponding coordinate system on the
physical element K.
The Jacobian matrix of the transformation K with respect to the reference coordinates and
its determinant are denoted by




K,i
(
xj )
and JK (
x) := det FK (
x) .
FK (
x) :=
j
x
1i,jd

A special
Since K has to be one-to-one and onto, we have JK (
x) does not change sign on K.
but important case of element transformation is the affine linear map, i.e.,
+ bK ,
K (
x) := BK x

BK Rdd , bk Rd

38

CHAPTER 4. THE FINITE ELEMENT METHOD

Using affine mappings ensures

where FK (
x) = BK and JK (
x) = det BK are constant over K.
that polynomials are mapped to polynomials of the same degree, which simplifies the analysis
of the FEM considerably. We will only consider affine transformations in this work.
The local mesh size h(x) for x = K (
x) can be defined by
h(x) := kFK (x)k

and

hK := sup h(x).
xK

Definition 4.4. A triangulation Th is called quasi-uniform, if hK h for all elements K

in Th . Th is called shape-regular, if the condition number of the Jacobian is bounded for all
elements, i.e.,
1
kFK (x)kkFK
(x)k  1
K Th , x K.
(4.5)
For our considerations below we need to know how tangential and normal vectors are mapped
between the reference and the physical elements.
x) and (
Transformation of tangential and normal vectors Let n(
x) denote the outer

normal vector and a tangential vector for x K, respectively. Then the corresponding unit
normal and tangential vector on the physical element K are given by
n(x) =

T
x)
(
x)n(
FK
T
x)k
(
x)n(
kFK

and

(x) =

FK (
x)
(
x)
kFK (
x)
(
x)k

for x = K (
x) K.

4.1.5

Simplicial elements and barycentric coordinates

From now on, we restrict ourselves to a triangulation by simplicial elements, i.e., triangles or
tetrahedra. The simplex K Rd is defined as the convex hull of (d + 1) vertices VK V.
It will turn out to be useful to replace the Euclidian coordinates of the point x Rd by
barycentric coordinates i = i (x) with respect to the (d + 1) vertices Vi VK .

Definition 4.5. Let the simplex K Rd be the convex hull of vertices (Vi )1id+1 . The
barycentric coordinate i (x) with respect to the vertex Vi is defined as the unique linear
polynomial
i P 1 (K) such that i (Vj ) = ij 1 j d + 1.
(4.6)
Pd+1
As an immediate consequence we obtain that i=1 i (x) = 1, x K. Using the elementwise definition (4.6), the barycentric coordinate i (x) associated with the vertex Vi naturally
extends to a global function on with the following properties:
[
K
and
i (Vj ) = ij Vj V.
(4.7)
i C() with supp(i ) =
K:Vi K

Rd is the triangle
For d = 2, the local reference (master) simplex K

:= {
i 1, 0 x
1 + x
2 1}
K
x0x

1 (
2 (
3 (
with
x) = 1 x
1 x
2 ,
x) = x
1 ,
x) = x
2 and for d = 3. We use the tetrahedron

:= {
1 + x
2 + x
3 1}
K
x  0 xi 1, 0 x
1 (
2 (
3 (
4 (
with
x) = 1 x
1 x
2 x
3 ,
x) = x
1 ,
x) = x
2 ,
x) = x
3 .

4.2. APPROXIMATION PROPERTIES OF CONFORMING FEM

4.2

39

Approximation properties of conforming FEM

We assume the variational problem (4.1) to be coercive. Then the Lax-Milgram theorem
states the existence of a unique solution u X to (4.1) as well as the existence of a unique
discrete solution uh Xh to the discrete problem (4.2) (due to Xh X).
The basis of error estimates for coercive problems is Ceas Lemma, which states that the
discretization error ku uh kX is proportional to the approximation error:
Theorem 4.6 (Cea ). Let Xh be a sequence of subspaces of the Hilbert-space X. Suppose
that the variational problem (4.1) fulfills the assumptions of Theorem 3.20 (Lax-Milgram).
Let u X denote the exact solution to (4.1) and uh Xh the solution to the corresponding
discrete problem (4.2).
Then the discretization error can be estimated by the approximation error as follows:
ku uh kX

inf ku vh kX ,
vh Xh

(4.8)

where is the continuity constant and is the coercivity constant of a(, ) on X.

For the time being, we present the classical approximation result for H 1 (), as used e.g. for
the Poisson problem.
Theorem 4.7 (Approximation error in H 1 ()).
1. Let Th a shape-regular triangulation Th of with affine linear element transformations.
2. Let Xh be chosen by the space of continuous, piecewise polynomials oder k:



Xh := vh C()  vh |K P k (K) K Th .

3. Assume a regular solution u belonging to (H k+1 ())3 .

Then the approximation error in the H 1 -norm can be estimated by

inf ku vh kH m ()  hk+1m |u|H k+1 () ,

vh Xh

for m = 0, 1.
Choosing the discrete space Xh by continuous, piecewise polynomials, we achieve convergence
by refining the mesh, i.e. h 0. The order of approximation can be improved by increasing
the polynomial degree k, but as long as we increase k uniformly on the whole mesh, the
benefit is limited by the global regularity of the exact solution u X.
For the time being, this result suffices as motivation for the choice of piecewise polynomial
FE-spaces for approximating the exact solution. Note that in general approximation error
estimates are obtained by estimating the interpolation error, which will be presented in Section
4.4.

40

CHAPTER 4. THE FINITE ELEMENT METHOD

4.3

An exact sequence of conforming finite element spaces

In the sequel we consider the construction of piecewise polynomial, conforming finite element
spaces Xh X of functions over (X usually denotes some Sobolev space), i.e.,



Xh = vh X  vh |K P k (K) K Th ,

and Th denotes a regular triangulation of the domain . We want to construct the spaces Xh
in such a way, that the exact sequence property discussed in Section 3.2.1 carries over from
the continuous spaces to the discrete spaces. In three space dimensions, this means

id

curl

div

R H 1 () H(curl, ) H(div, ) L2 ()
S
S
S
S
id

Wh

Vh

curl

Qh

div

Sh

{0}

(4.9)

{0}.

In the two dimensional setting we consider the corresponding shortened sequences, cf. (3.15)
and (3.16).

4.3.1

The classical H 1 -conforming Finite Element Method

In this subsection, we introduce the classical lower-order finite elements used for H 1 conforming methods.
Definition 4.8 (The linear H 1 -conforming finite element). The classical lowest-order H 1 conforming finite element for the simplex K Rd (a segment for d = 1, a triangle for d = 2,
and a tetrahedron for d = 3), is defined by
the local space PK = P 1 (K) of dimension dim(P 1 (K)) = d + 1,
the (d + 1) vertex-based degrees of freedom, which amount to point-evaluation at the
vertices, i.e.
NiV : v v(Vi )
Vi VK .
Definition 4.9. The second-order H 1 -conforming finite element on a simplex K Rd , d =
2, 3 is defined by

6
for d = 2,
2
2
the local space PK = P (K) with dim(P (K)) =
10 for d = 3,
two types of degrees of freedom (dofs):
the lowest-order, also denoted as vertex-based dofs:
NiV : v v(Vi ) for 1 i d + 1
one edge-based degree of freedom over each edge E EK (see [42]):
Z
v qe
ds,
NE : v
E s s
R
with edge-bubble qe P02 (E) and E qe2 ds = 1.

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

41

In classical finite element methods, as described, e.g. in Ciarlet [35], so-called Lagrange-type
elements are commonly used, i.e., the degrees of freedom are associated with evaluation
of a function or its derivatives at several nodes located on the element. In view of an
implementation of higher-order methods the above formulation will turn out to be more
natural.
For practical implementation, we first construct a local basis (shape functions) on the reference
By (4.6), the barycentric coordinates form a (nodal) basis for the lowest-order
element K.
1
H -conforming element, i.e., we choose
i (
V (
x) =
x)
for 1 i d + 1,
(4.10)
i

which obviously span

P 1 (K).

As nodal shape functions for the second-order element we use

i (
lowest-order functions: Vi (
x) =
x)
1 i d + 1,

and

e (
x) E = [e1 , e2 ] E.

quadratic edge-bubbles: E
x) =
(
1 x)e2 (

(4.11a)
(4.11b)

we map the functions

Having defined the shape functions locally only on a reference element K,
by a conforming transformation onto the physical element K Th . The following lemma states
that for scalar functions this can be realized by a simple change of variables.
K be a continuously differenLemma 4.10 (H 1 -conforming transformation). Let K : K
be a scalar function. Then the change
tiable, invertible and surjective map , and u
H 1 (K)
of variables
(4.12)
u := u
1
K
implies u H 1 (K) with

u 1
u = FK
K .

(4.13)

The local shape functions on the physical element K are hence defined by
Vi (x) = i (x)
E
(x)

for Vi VK ,

for E = [Ve1 , Ve2 ] EK .

= e1 (x)e2 (x)

(4.14a)
(4.14b)

To obtain the global finite element shape functions in Wh we have to piece the elementby-element defined shape-functions together. This is done by identifying the local degrees of
freedom with the global ones. The standard element pull-back transformation (4.12) preserves
E
V (
the degrees of freedom of H 1 -conforming finite elements, i.e. NiV (u) = N
i u) and Ni (u) =
E (
N
. By identification of the vertex-based and/or edge-based degrees of
i u) for u K = u
freedom corresponding to element interfaces we obtain the global finite element spaces as
follows:
M
Wh,1 :=
span{Vi },
Vi V

Wh,2 :=

Vi V

span{Vi }

Ei E

span{E
i }.

Due to (4.7) barycentric coordinates are continuous across element interfaces. Hence, the
above shape functions (4.14) match continuously at element interfaces. Owing to Corollary
3.5 this implies H 1 -conformity, i.e. for k = 1, 2 :

Wh,k = {w H 1 ()  w| P k (K) K Th }.
K

42

CHAPTER 4. THE FINITE ELEMENT METHOD

4.3.2

Low-order H(curl)-conforming Finite Element Methods

In curl curl problems arising from electromagnetics, one is usually not only interested in
the primal variable, but in the corresponding differential field. For example, in the vector

A and the magnetic field

potential formulations (2.28) the electric field intensity E = t
1
intensity H = curl A are the quantities of interest. It is therefore desirable to approximate
the fluxes with high accuracy. This motivates the use of the following families of H(curl)conforming finite elements:
Nedelec element of second kind of order k, which involves a local FE-space of full polynomial order:
NkII (K) := (P k (K))d .
(4.15)
Nedelec element of first kind of order k, which involves a local space containing P k (K),
enriched (in a minimal way) such that the curl of the involved functions also spans
P k (K):

(4.16)
NkI (K) := P k (K)d {q (Pk+1 (K))d  x q = 0},
where Pk+1 (K) denotes the set of homogenous polynomials of degree k + 1. This setting
implies
(P k (K))d NkI (K) (P k+1 (K))d and (P k+1 (K))d = NkI (K) Pk+2 (K).
In this section, we consider the low-order finite element spaces where the local spaces are either
N0I (K) or N1II (K). Both spaces involve finite element functions having piecewise constant
curl-fields. As we will see below, these spaces are also natural in the sense that they can be
used to build an exact sequence.
Definition 4.11 (Nedelec element of first kind of order 0). The lowest order edge-element on
a triangle, respectively a tetrahedron, K is given by
the local space N0I (K) defined as
for d = 2 :
for d = 3 :

y

a R2 , b R
x


I
N0 (K) := a + b x  a, b R3

N0I (K) :=

a+b

the edge-based degrees of freedom:

Z
N0
v dx
N : v
E

with dim(N0I (K)) = 3,

with dim(N0I (K)) = 6.

for = 1, . . . , |EK |,

i.e. the line integrals of the tangential component over each edge E EK .
The local space N0I (K) lies between the full polynomial spaces of order 0 and 1, i.e.

3

3
P 0 (K) N0I (K) P 1 (K) with curl N0I (K) = (P 0 (K))3 .

We choose the lowest-order shape functions corresponding to the 1-form Whitney element, cf.
Bossavit [25].

43

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

Lemma 4.12. The nodal basis of N0I (K) is realized by the edge-based shape functions
0
N
= 1 2 1 2

for each edge E = [V1 , V2 ] EK . Moreover, the shape functions have constant tangential
trace on the edges of K, i.e.,
0
N
|E =

0
in particular NN0 (N
)=

1
,
|E |

for E , E EK ,

(4.17)

0
N
dx = .



Proof. We reformulate the local space as N0I (K) = Bx + b | B = B T Rdd . Since the
N0
N0
0
shape functions N
are linear, we can express (x) = Ax + a with A = ( ) = 1 :
I
0
2 2 : 1 , which is skew-symmetric. Hence, N
N0 (K). Next we consider a
P|EK |
N
trivial vector field v = =1 c 0 = 0. This implies v |E = 0 on all edges. In combination
with (4.17) we obtain c /|E | = 0, E EK and hence the linear independence of the shape
functions. Due to the counting argument we have dim(N0I (K)) = |EK | and hence
I
0
span{N
: E EK } = N0 (K).

To show (4.17), we consider an edge E = [V1 ,
V2 ] where denotes the associated tangential
vector. Then k |E = |E1 | k1 + k2 and k = 0 on E if k
/ E . We obtain

0
N

|E

1
1
(1 + 2 )|E =
.

|E |
|E |

(4.18)

Concerning the remaining edges, i.e. E 6= E there exists a vertex Vk E s.t. Vk

/ E ,
N
0
which implies k |E = 0 as well as k |E = 0 and hence | = 0. Evaluating the

line-integral of (4.17) over several edges E finally yields the nodality of the basis.
Definition 4.13 (Lowest-order Nedelec elements of second kind). The Nedelec element of
second kind of order p = 1 on a simplex K is defined by

d
the full-order space N1II (K) = P 1 (K) of dimension 6 for d = 2 and 12 for d = 3,

the first and second moment of the tangential components over each edge E EK , i.e.
Z
v qk dx
for k = 0, 1,
Nk : v
E

where {qk } is a basis for P 1 (E ) and q0 = 1.

Lemma 4.14. For an edge E = [V1 , V2 ] EK , = 1, . . . , |EK |, define
0
N
= 1 2 1 2 ,


E
= 1 2

and

(4.19a)
(4.19b)

E
0
Then the set {N
edelec element N1II (K) of
, } of shape functions is a basis for the N
second kind:

44

CHAPTER 4. THE FINITE ELEMENT METHOD

Proof. Following the proof of Lemma 4.12 we obtain the extension of 4.18 as
0
N
|E

E
|E

=
=

1
|E | (1
1
|E | (1

+ 2 )|E =

1
|E |

2 )|E

with (1 2 ) being linear on E . This implies the linear-independence of the shape

functions. Verifying the counting argument dim(P 1 (K))d = 2|EK | concludes the proof.
Since the edge-based shape functions, of the element of second type, are gradient functions,
there obviously holds
curl N0I (K) = curl N1II (K).
H(curl)-conforming element transformation. The conforming transformation of vec onto shape functions defined on a phystorial shape functions on the reference element K
ical element K is more involved than in the scalar case. Note that in order to guarantee
H(curl)-conformity, we want that the degrees of freedom are preserved by the transformation.
Moreover, for establishing a local exact sequence property, we have to ensure that gradient
fields on the reference element are mapped onto gradient fields on the physical element. In
view of (4.13), we suggest the following transformation:
K be a continuously
Lemma 4.15 (H(curl)-conforming transformation). Let K : K

H(curl, K). Then the transformation

differentiable, invertible and surjective map, and u
T
1
u := FK
u
K

(4.20)

1
1
(curlx u) = JK
FK curlx u
K

(4.21)

implies u H(curl, K) with

1
1
curlx u
for K R3 . In two dimensions we obtain instead curlx u = JK
K .

A proof of the lemma can be found in Monk [70] (Lemma 3.57, Corollary 3.58).
In the following we collect some important properties of the transformation above:
1. Tangential traces along edges transform as
T

u
(u ) K |E = FK

T


FK
T
k |E
kFK

= u

1
.
T
k |E
kFK

Taking into account the transformation of line integrals we obtain

Z
Z

T
k
k (

k d
s=N
qk kFK
Ni (u) =
u qk ds =
u
i u),

kF

K
K (E)
E
where qk k := qk . Hence, the degrees of freedom are preserved by the transformation.
1 (K)
are mapped onto gradient fields p H 1 (K), since for all
H
2. Gradient fields p

= w:
u

T
T
= FK
w
= (w) K
u
u K = FK
with w K := w
and w H 1 (K).

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

45

0
N
E
Applying the conforming transformation (4.20) on the shape functions
,
defined on
0 , E on the physical element
we can deduce the shape functions N
the reference element K

K. This can be verified by performing a change of variables x = k (

x) in the shape functions
(4.19) and taking into account the transformation of gradient fields according to Lemma 4.10
as follows:
0
N
(x) = 1 (x) 2 (x) 1 (x) 2 (x)




T
(
(
x) F T

x)
x)
1 (
= FK
2
2 x) 1 (
K


T

x)
(
x)
= FK
1 2 2 1 (

T
0
N
= FK
(
x)
x),
(

T

(

x)
x)
E
(x) = (1 2 )(x) = FK (
1
2
T
E
= FK
(
x)
x).
(

Global shape functions and H(curl)-conforming FE-space In order to obtain global

FE-spaces we identify the degrees of freedom corresponding to the global edges of the mesh,
i.e. we glue the shape functions together:
Vh,0 :=
II
Vh,1
:=

Ei E

Ei E

0
span{N
i },

E
0
span{N
i , i }.

The H(curl)-conformity
of the element can be verified by the following argument. Suppose
P
K
that v =
ci i P (K). The associated finite element is H(curl)-conforming: whenever
the degrees of freedom associated with an interface (an edge E or respectively
a face F )

are zero, the tangential trace of v vanishes on the interface, i.e. v E = 0, or resp.

(v n)F = 0. From the proofs of Lemma 4.12 and Lemma 4.14 we know that the edge-based
shape functions associated with the interfaces are the only shape functions with non-vanishing
tangential trace on the interface. Therefore, the tangential trace on faces (only d = 3) and
edges is realized by the shape functions associated with this interface. Since the global FEspace is obtained by identification of the corresponding coefficients at the interfaces, we obtain
tangential continuity of the global shape functions.
We note that one has to take the global edge-orientation (the orientation of the tangential
vector) into account within the setting of global shape functions. If the global edge has the
opposite orientation than the transformed global one, the sign of the transformed shape
function on the corresponding element has to be changed. This can be done by allowing 1
entries in the connectivity matrix during the assembling process.
Summarizing, we obtain tangential continuity over element interfaces and owing to Corollary
3.8 we obtain two H(curl)-conforming FE-spaces:


v H(curl)()  v |K N0I (K) K Th ,



= v H(curl)()  v |K (P 1 (K))d K Th .

Vh,0 =
II
Vh,1

Since the lowest-order degrees of freedom of H(curl)-conforming elements are associated with
the edges of the mesh, it is common to call the Nedelec elements the edge elements.

46

CHAPTER 4. THE FINITE ELEMENT METHOD

4.3.3

Low-order H(div)-conforming Finite Elements Methods

Due to similar considerations as for the H(curl)-space we introduce the following two types
of H(div)-conforming elements, differing in the approximation of divergence fields:
the Brezzi-Douglas-Marini (BDMk ) element of order k with the local FE-space
BDMk (K) := (P k (K))d ,

(4.22)

the Raviart-Thomas (RT k ) element of order k with the enriched local FE-space
RT k (K) := (P k (K))d x Pk (K)

(4.23)

with div RT k (K) = P k (K). In three dimensions, the Raviart-Thomas element is also
called Raviart-Thomas-Nedelec element.
Note that the spaces RT k (K) and BDMk (K) contain the same divergence-free vectors. This
time, the triangular and tetrahedral cases have to be treated separately.
H(div)-conforming elements in two dimensions The two-dimensional case can be handled by a simple rotation of 90 degrees of the H(curl)-conforming element.
Definition 4.16.
1. The lowest-order Raviart-Thomas element (of order k = 0) on a
triangle K is defined by



the local space RT 0 (K) = a + b x  a R2 , b R with dim(RT 0 (K)) = 3,
the total flux over each edge E EK , i.e.,
Z
v n dx
N : v

= 1, 2, 3.

2. The linear BDM -element (of order k = 1) on a triangle K is defined by

the full-polynomial local space BDM1 (K) = (P 1 (K))2 ,

two degrees of freedom associated with each edge E EK , namely

Z
v n qk dx for k = 0, 1,
N : v
E

with (qk )k=1,2 spanning P 1 (E ).

For a hierarchical definition of BDM-elements of higher order it will be important to choose
q0 = 1. We will assume this also for the linear case. As already mantioned above, the fluxes
of these elements span the same spaces, i.e.,
div RT 0 (K) = div BDM1 (K) = P 0 (K),
and the nodal shape functions can be derived by rotating by 90 degrees the nodal shape
functions of the Nedelec element, which yields
0
RT

= 1 Curl 2 Curl 1 2 ,


= Curl 1 2 ,

for each edge E = [V1 , V2 ] EK .

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

47

H(div)-conforming elements in three dimensions

We consider only the lowest-order element RT 0 (K) here, and leave the treatment of BDMk
elements to the next chapter.
Definition 4.17. The classical, lowest-order, H(div)-conforming element on a tetrahedron
K is defined by



the local space RT 0 (K) = a + b x  a R3 , b R with dim(RT 0 (K)) = 4,

the lowest-order degrees of freedom associated with faces, i.e. the total normal fluxes
over each F FK :
Z
v n dx for 1 4.
N : v
F

For the lowest-order Raviart-Thomas element we obtain

3

3
P 0 (K) RT 0 (K) P 1 (K)
with div RT 0 = P 0 (K).

Since the degrees of freedom are associated with the faces of the mesh, such elements (of
arbitrary order) are often called face elements. The lowest-order face-based shape functions
can be realized by 2-form Whitney-elements as described in Bossavit [25].
Lemma 4.18. Associating with each face F FK the shape function
f
f ,
f
f +
f
f
f +
f
RT 0 =
f

1
2
3
1
3
2
3
2
1

where F = [f1 , f2 , f3 ], yields a nodal basis for RT 0 (K).

RT 0

across the face F is 1 and vanishes across all other faces, i.e.
The flux of

Z
RT 0 n
d

x = ,

(4.24)

(4.25)

if the counter-clockwise rotation of the vertices [f1 , f2 , f3 ] points in the direction of the outer

normal vector n.
on the face F , determined by the vertex-triple [f1 , f2 , f3 ],
Proof. The outer normal vector n

f | where f4 denotes the vertex on the tetrahedron

= f4 /|
can be expressed by n
4
1 +
2 +
3 +
4 = 1 in K
4 = 0 on F the
and
opposite to the face [f1 , f2 , f3 ]. Using

relation |4 | = |F | implies

1

1
f
f
f
RT 0 n
f
f
f

|F =

1
2
3
1
2
3

|F | .
|f4 |

f
f
f = 1 if the counter-clockwise rotation of [f1 , f2 , f3 ] points out of K

Since
1
2
3
(otherwise it is 1), integration over F yields (4.25) for = .
=
Now let f1 denote the vertex opposite to F with outer unit normal vector n

f1 /|f1 |. Then there holds

f
f )
f ) = 0
RT 0 n
f

= (

2
3
1
1

f = 0 on F .
since
1

for 6= ,

48

CHAPTER 4. THE FINITE ELEMENT METHOD

H(div)-conforming transformation
We plan to construct global basis functions for H(div, T ) again by a conforming transformation
The following transformation preserves curlof basis functions on the reference element K.
fields, cf. (4.21):
K be a continuously
Lemma 4.19 (H(div)-conforming transformation). Let K : K
the
H(div, K)
differentiable, invertible and surjective mapping. Given a vector function p
Piola transformation
1
1
FK p
p := JK
K ,
also called the contravariant transformation, implies p H(div, K) with
1
1
div p = JK
divx p
K .

The proof of this result can be found in Monk [70, Lemma 3.59]. The Piola transformation
has the following desired properties:
1. Normal traces transform as






F T n
1
1
n

K
= u
FK u
u n K |F = JK
T
T
|F
|F
JK kFK nk

kFK nk

Taking into account the transformation of surface integrals, i.e.

Z
Z
T

dA,
dA =
nk
|JK | kFK
A

we see that the lowest-order degrees of freedom are invariant with respect to the Piola
transformation:
Z
Z
n
d
p n dx = sign(JK )
p
x.
F

are mapped onto curl fields p curl H(curl, K):

curlx H(curl, K)
2. Curl fields p
1
1
= curl v K .
= JK
p K = JK
FK p
FK curlx v

u
v for scalar functions u
:
=
3. The Piola transformation preserves p
, v H 1 (K)


1
1
v)
u) (F T
u
v = (F T
= JK
FK
FK p
p K = JK
K
K


=
u v K .

4. In the case of an affine linear transformation, the Raviart-Thomas space RT k (K) is

invariant with respect to the Piola transformation (see Brezzi-Fortin [32]).
Applying the Piola transformation to the lowest-order Raviart Thomas shape functions on
the reference element, we obtain the following shape functions on the physical element K:
1 T RT 0
0
FK
1
RT
= JK

with normal flux

= f1 f2 f3 + f2 f3 f1 + f3 f1 f2 .

0
RT
n dx = across the face F .

(4.26)

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

49

Global shape functions and H(div)-conformity Identifying the degrees of freedom

associated with the faces of the mesh leads to the global FE-space
M
0
}.
span{ RT
Qh,0 :=
i
Fi F

The H(div)-conformity of the physical elements associated with RT 0 (K) can be verified
as follows: Let q RT 0 (K) with q n = 0 on the face Fj . In order to verify H(div)
conformity we only have to show that all degrees of freedom
(dofs) associated with this face
P
0
and furthermore
vanish, i.e. NjRT 0 (q) = 0. Now observe that 0 = q = Fi F NiRT 0 (q) RT
i
RT 0
RT 0
RT 0
1
(q) |Fj | .
(q)trFj ( j ) = Nj
0 = trFj (q n) = Nj
The normal flux through the interface of two adjacent elements is determined by the degree
of freedom associated with this face. Since we associated the degrees of freedom explicitly
with the fluxes over the faces, the global shape functions have continuous normal components
by construction. Hence the global FE-space is a subspace of H(div), i.e.

Qh,0 = {q H(div, )  q |K RT 0 (K) K Th }.

4.3.4

The lowest-order L2 -conforming Finite Element Method

In contrast to the finite element spaces discussed previously, the L2 -conforming space requires
no continuity across the element interfaces. In principle, L2 -conforming elements can be
constructed in an analogy to H 1 -conforming elements, but without regarding continuity across
element boundaries. We just sketch the such a construction:
Definition 4.20. The L2 -conforming element of order k on a triangle or tetrahedron K is
defined by
the local space P k (K),
the cell-based degrees of freedom:
NlC

: p

p ql dx

0lk

with (ql )0lk a basis of P k (K)

For a hierarchical construction (higher order) it will be important that {qk } is a hierarchical
basis of P k (K), in particular q0 = 1. The nodal lowest-order shape function can be realized
as
1
i =
Ki Th .
(4.27)
|Ki |

Classically one uses the standard element pull-back transformation (4.12) for scalar functions
to define the basis functions on the physical element K. But in view of the global exact
sequence property we define the L2 -conforming transformation by the conforming transformation of divergence fields defined in Lemma 4.19 as follows (cf. Demkowicz et al. [44]).

K be a continuously difLemma 4.21 (L2 -conforming transformation). Let K : K

we define its
ferentiable, invertible and surjective mapping. Given a function q L2 (K)
L2 -conforming transformation by
1
q = JK
q 1
(4.28)
K .

50

CHAPTER 4. THE FINITE ELEMENT METHOD

Since no continuity across element interfaces is required, the global element space can be
locally, i.e.
M



0
2

(4.29)
span{C
Sh,0 :=
i } = g L () g|K P (K) K Th .
Ki Th

4.3.5

Discrete exact sequences

In order to show the exactness of the sequence (4.9) of the global FE-spaces, we first show
is exact.
that the associated local sequence defined on the reference element K
Exact sequences on the reference element level involving spaces of the first family
we obtain that the following sequence
Lemma 4.22. Concerning the reference tetrahedron K
of local FE-spaces associated with the first kind of low order finite elements is exact:
curl
div

id

N0I (K)
RT 0 (K)
P 0 (K),
R P 1 (K)

(4.30)

as illustrated in Figure 4.2.

id

1+3

div

curl

3+3

{0}
1+0

3+1

Figure 4.2: This sequence for the elements of the first family with vertex-based(black), edgebased(red), face-based(green) and element-based(blue) dofs is exact. The range of an opertor
has the same dimension as the kernel of the following operator.
This can be easily verified by two steps:
1. The range of a differential operator is included in the kernel of the following one:
= P 0 (K)
3 N I (K),

P 1 (K)
0
= P 0 (K)
3 RT 0 (K),

curl N I (K)
0

= P 0 (K).

div RT 0 (K)

ker(curl N I (K)),

hence P 1 (K)
0
ker(div RT I0 (K)),

hence curl N I (K)

0

2. We recall the algebraic property of discrete spaces and linear transformations D that
dim(Xh ) = dim(ker(D)) + dim(range(D)).

(4.31)

A simple counting argument now shows that equality holds instead of only inclusion,
i.e., range(div) = span{1}, range() = ker(curl), and range(div) = ker(curl); cf. Figure
4.2.
In the two-dimensional case we obtain the following shortened sequences, involving the lowestorder elements of first kind:

51

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

the shortened sequences of local lowest

Lemma 4.23. For the triangular reference element K
order FE-spaces (as illustrated in Figure 4.3 and Figure 4.4)
id

curl
0



R P 1 (K)
N0I (K)
P 0 (K)
{0},

and

id

Curl

div

(4.32)

RT 0 (K)
P 0 (K)
{0}
R P 1 (K)

(4.33)

are exact.
This follows with arguments to those above, by using
= P 0 (K)
2 N0 (K)
and curl N0 (K)
= P 0 (K),

N0 (K)
= P 0 (K)
2 RT 0 (K)
and div RT 0 (K)
= P 0 (K),

Curl P 1 (K)
and comparing the dimensions of the ranges, kernels and spaces; cf. Figure 4.3 and Figure
4.4.

id

curl

R
1+2

2+1

1+0

{0}

Figure 4.3: 2-dimensional exact sequence involving lowest-order Nedelec element (of first kind)
and splitting of space into kernel and range of the consecutive operator

Curl

id

div

R
1+2

2+1

1+0

{0}

Figure 4.4: 2-dimensional exact sequence involving the lowest-order Raviart-Thomas element

Remark 4.24. In the proposed sequences of local FE-spaces involving the lowest-order
H(curl)- and H(div)-conforming finite elements of the first kind, the polynomial degree is
only decreased by 1 throughout the whole de Rham Complex. As we will see in the next chapter, this holds true also for corresponding high-order simplicial finite elements of first kind.
Exact sequences property for the second family on the reference element level
Analogous considerations imply exact sequences involving H(curl)- and H(div)-conforming
elements of the second family. In two dimensions the following shortened exact sequences hold.
the following two sequences of local
Lemma 4.25. For the triangular reference element K
spaces involving either the Nedelec element of second kind or the Raviart-Thomas element

52

CHAPTER 4. THE FINITE ELEMENT METHOD

(BDM1 ) of second kind are exact:

id

= P 1 (K)

R P 2 (K)
N1II (K)

curl
0

P 0 (K)
{0},

(4.34)

id
Curl
div
id

= P 1 (K)


R P 2 (K)
BDM1 (K)
P 0 (K)
{0}.

(4.35)

(Compare Figure 4.5 and Figure 4.6.)

id

curl

R
1+5

5+1

1+0

{0}

Figure 4.5: 2-dimensional exact sequence involving the linear Nedelec element of second kind
and splitting into kernel and range of consecutive differential operator

id

Curl

R
1+5

div

5+1

1+0

{0}

Figure 4.6: 2-dimensional exact sequence involving the linear BDM element and splitting into
kernel and range of consecutive differential operator
We observe that in the De Rham sequences, which involve triangular finite elements of the
second family, the polynomial degree decreases by order 2 throughout the sequence. In three
dimensions, the de Rham Complex involves the application of three consecutive differential
operators. Therefore, if one uses spaces of full polynomial degree on tetrahedra, the polynomial
degree lowers by 3 throughout the sequence. This leads to the following sequence involving
Nedelec and Raviart-Thomas spaces of the second kind.
denote the tetrahedral reference element. The following sequence of
Lemma 4.26. Let K
local spaces involving only full polynomial spaces
id

curl

div

(P 1 (K))
3 = BDM1 (K)
P 0 (K)
{0}
(P 2 (K))
3 = N II (K)
R P 3 (K)
2
(4.36)
is exact.
This sequence of spaces is the one involving spaces with lowest degree while still ensuring full
polynomial degree and exactness.
The exactness of the three sequences stated in Lemma 4.25 and Lemma 4.26 can be proved
with the same arguments as above: By construction the range of each operator is a subspace
of the following space. Equality then follows again by a counting argument; cf. Figure 4.5,
4.6, and 4.7).
Remark 4.27. If we use a conforming transformation onto the physical elements due to
Lemma 4.10, Lemma 4.15, or Lemma 4.19 respectively, the exact sequence () on the reference
element implies that the exactness still holds on the local spaces of the physical elements.

53

4.3. AN EXACT SEQUENCE OF CONFORMING FINITE ELEMENT SPACES

id

1
0

curl

1+9

div

9+3

{0}
1+0

3+1

Figure 4.7: Exact sequence involving tetrahedral Nedelec element of the second kind.
Global exact sequence property of discrete FE-spaces
As already outlined above, the use of conforming transformations provides that the exactness
is provided for the local spaces defined on the physical elements. Moreover, due to the
conforming construction the exactness then also holds for the global finite element spaces.
We summarize results from Bossavit [23], Bossavit [25] (Proposition 5.5), and ArnoldFalk-Winther [7]:
Theorem 4.28 (Exactness of global discrete sequence). Let be a simply-connected domain
with connected boundary, and let the physical finite elements be constructed by conforming
transformations (cf. Lemma 4.10, Lemma 4.15, and Lemma 4.19). Then the sequences of the
conforming global FE-spaces corresponding to (4.30)
id

curl

div

curl

div

R Wh,1 Vh,0 Qh,0 Sh,0 {0}

and corresponding to (4.36)
id

II
R Wh,2 Vh,1
Qh,0 Sh,0 {0}

are exact.
In two dimensions, the following shortened sequences corresponding to (4.32) or (4.34)
id

curl

Curl

div

R Wh Vh Sh {0}
and corresponding to (4.33) or (4.35)
id

R Wh Qh Sh {0}
are exact.
Remark 4.29 (Exact sequences involving essential boundary conditions). Essential boundary
conditions can be taken into account in a similar manner as in the continuous case, cf. Remark
3.17. The exactness of the resulting sequences follows then by the same arguments as in the
Neumann case.

4.3.6

Element matrices and assembling of FE-matrices

For an efficient computation, the FE-matrices are assembled from element-matrices AK

, for
all K Th . The assembling of the global matrix is done via a connectivity matrix C which
assigns local to global degrees of freedom.

54

CHAPTER 4. THE FINITE ELEMENT METHOD

The global Galerkin system matrix as well as the global load vector, as defined in (4.3), are
assembled as follows:
A=

CK

KTh

T

AK C K ,

f=

CK

KTh

T

fK

with connectivity matrices

K
Ci,

1
0

if local dof is identified with global dof i,

otherwise.

The sign factors in the entries of the connectivity matrix 1 account for the global edgeorientation (the orientation of the tangential vector) or for the global face-orientation (the
direction of the normal vector) and are necessary to enable conformity. The element matrices
AK , and the element load vector fK are stated below. Note that due to the local support of
the shape functions, the connectivity matrices as well as the system matrix A are sparse.
The calculation of the element stiffness matrix and the element load vector is done by integration on the reference element while taking into account the element transformations. We
shortly outline the procedure for the electromagnetic model problems we have in mind.
Electrostatic problem
1 () of the variational
In view of Section 3.4.1, we are looking for solutions u Wh HD
problem
Z
Z
Z
u v dx +
uv dx =
f v dx v Wh .

For discretization we use H 1 -conforming finite elements. The element matrices and vectors
can be computed by taking into account the transformation of scalar and gradient functions
(cf. Lemma 4.10) in the following way:
AK

=
fK

dx +
T
FK

f dx =

dx

T
FK

x+
J d

J d
x,

f J d
x,

where { } denotes the shape functions on the reference element.

Magnetostatic problem
We recall the variational formulations of the magnetostatic problem, cf. Section 3.4.2, and
hence consider the following model: Find u Vh H(curl, ) such that
Z

curl u curl v dx +

u v dx =

j v dx

v Vh .

4.4. COMMUTING DIAGRAM AND INTERPOLATION ERROR ESTIMATES

55

Utilizing the H(curl)-conforming transformations (cf. Lemma 4.15) the element matrices and
vectors can be computed on the reference element as follows:
Z
Z
K
1
A =
curl curl dx +
dx
K

1 1

FK curl

d
x+
FK curl

FK

T
x,
J d
FK

in three spatial dimensions. In two dimensions we obtain instead

Z
Z

T

T
1 1

FK
AK
=
x.

J
curl

curl

d
x
+
J d

FK

The element vector allows a unique representation for K R3 and R2 :

Z
Z


T
J d
fK =
j dx =
x.

j FK

} denote the shape functions on the reference element, spanning a local H(curl)Here {

conforming FE-space on K.

4.4

Commuting Diagram and Interpolation Error Estimates

The commuting diagram, which links the global and discrete sequences of spaces, plays an
important role in the error analysis of finite element approximations. In view of emphasizing
once more the importance of exact sequences, we shortly sketch some ideas and the main
results of interpolation operators and the connected error estimates. An extensive analysis
on commuting diagrams can be found in Bossavit [25] for nodal interpolation operators,
berl [81] and projection-based interpolation
Clement-type interpolation operators in Scho
in Demkowicz-Buffa [43].
In the error analysis of finite-element methods as well as for implementing boundary conditions
we use interpolation operators which map a (Sobolev) space X into the discrete FE-spaces Xh ,
i.e. X : X Xh . In particular we are interested here in the following choice of spaces X:
W H 1 (), V H(curl, ), Q H(div, ) or S L2 (). A special case of interpolation
operators is realized by the degrees of freedom, namely
X (v) =

N
X

Ni (v)i

i=1

with span{i : i = 1, . . . , N } = X.
One essential drawback of nodal interpolation operators is that that they are not well-defined
on the whole function space H 1 (), H(curl), or H(div), since not all traces, which are involved
in the degrees of freedom, are properly defined on the whole space, e.g. point-evaluation in
2D or 3D, is not well-defined in H 1 ().
Interpolation operators which require less smoothness are Clement type operators, which
involve non-local averaging over certain patches. For instance, the Clement type operator for
H 1 -problems is defined by averaging over the patch of elements associated with one vertex
instead of point-evaluation in this vertex.

56

CHAPTER 4. THE FINITE ELEMENT METHOD

Let us consider a sequence of conforming finite-element spaces (W , V , Q, S) and interpolation operators i . We summarize all the involved operators and spaces in the following
(commuting) diagram:

H 1 ()
S

Wh

curl

H(curl, )
S

Vh

curl

div

H(div, )
S

Qh

div

L2 ()
k

Sh

This diagram is said to commute, if for sufficiently smooth functions the following hold:
V w = W w
Q curl v = curl V v
S div q = div Q q

w W H 1 (),
v V H(curl, ),
q Q H(div, ).

Such relations can be derived for certain choices of interpolation operators via Stokes and
Greens theorems. In the following theorem we collect some results on nodal interpolation
error estimates, which are frequently used in approximation error estimates.
Theorem 4.30 (Nodal interpolation error estimates). The nodal interpolation operators are
well defined on the spaces
3

= H 2 + (),

3 

3

1
1
V = v H 2 + ()  curl v H 2 + () ,



3 
1
Q = v H 2 + ()  div v L2 ()

with > 0.

Let Th denote a regular tetrahedral mesh width meshwidth h.

1. Let Wh be the H 1 -conforming FE-space of order k, and W denote the corresponding

nodal interpolation operator. If w (H s ())3 for 32 + s k + 1, then
kw W wk1  hs1 kwks .
2. Let V denote the nodal interpolation operator corresponding to the FE-space Vh of
Nedelec elements of the first kind and order k. If v (H s ())3 and curl v (H s ())3
for 12 + s k + 1, then there holds




kv V vk0  hs kvks + k curl vks and k curl(v V v)k0  hs k curl vks ,

see Monk [70](Theorem 5.41, Remark 5.42) and Hiptmair [56] (Theorem 3.14, Corollary 3.17).

3. Let Qh denote the FE-space of Raviart-Thomas-Nedelec elements of order k, and Q be

the corresponding nodal interpolation operator. If q (H s ())3 and 21 + s k + 1,
then
kq Q qk0  hs kqks and k div q Q qk0  hs kqks
hold, see Monk [70](Theorem 5.25, Remark 5.26).

4.4. COMMUTING DIAGRAM AND INTERPOLATION ERROR ESTIMATES

57

Similar results also exist if Vh and Qh are chosen out of the second exact sequence, cf. (4.36).
However, there the approximation order of v and q is one order better than that of curl v and
div q, respectively, due to the choice of polynomial spaces.

58

CHAPTER 4. THE FINITE ELEMENT METHOD

Chapter 5

High Order Finite Elements

In the previous chapter we introduced the classical Finite Element Method, namely h-FEM,
where the polynomial degree on all elements is fixed to a uniform order, typically p = 1 or
p = 2. Convergence is achieved by local or global refinement of the underlying mesh Th ,
denoted as h-refinement, and the error in the numerical solution decays algebraically.
The idea of p-version finite element methods is to use a fixed triangulation and obtain convergence by increasing the polynomial order p-refinement. In case of an analytic solution the
p-version FEM yields exponential convergence rates in the energy norm, cf., e.g., BabuskaSzabo-Katz [12]. In case of piecewise analytic solutions with singular behavior near corners,
edges or boundary layers, also the p-version yields only algebraic convergence.
The hp-method now combines the two ideas: Exponential convergence can be regained by
the combination of (geometric) h-refinement and local increase of the polynomial order.
Such a strategy is then called hp-refinement. For a survey on hp-refinements we refer to
ster-Rank [86].
Melenk [68] and Szabo-Du
This chapter is devoted to the construction of conforming hierarchical hp-finite element
approximations for the vector-valued spaces H(curl, ) and H(div, ). As we will see in
our numerical tests, regular geometric h-refinement towards singular corners, singular edges
and/or boundary layers naturally leads to various element topologies, involving in particular
not only simplices.
Our goal is to provide a general strategy for constructing a sequence of hierarchical finite
element spaces allowing for arbitrary and variable polynomial order on a mixture of common
element topologies in one single mesh. A main point in such a construction is to ensure the
exact sequence property of discrete spaces in a more localized sense. To achieve this, we
explicitly use higher-order kernel functions of the natural differential operator (in this space)
within the construction of the FE-basis, see also the previous chapter. For instance, in the
construction of an H(curl)-conforming basis we include gradient functions. The framework
we are going to present, is particularly well suited for a practical implementation, and is
extendable to anisotropic polynomial orders.
The construction of finite element spaces presented below is based on the following contributions:
In Webb [95] an H(curl)-conforming finite element basis is provided by taking the
gradient fields of H 1 -conforming shape functions and then extending the basis to the

59

60

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

full polynomial space. The shape functions are formulated as monomials of the affine
coordinates. To avoid very ill-conditioned matrices Gram-Schmidt orthogonalization is
used, which is conveniently proposed up to polynomial order 3.
A first general hierarchical construction strategy of H(curl)-conforming and H(div)conforming finite elements for arbitrary polynomial order on tetrahedral elements is
presented in Ainsworth-Coyle [2].
In Karniadakis-Sherwin [60] H 1 -conforming spaces which yield a general construction for all common element topologies, are suggested. This is obtained by using an
underlying tensor-product structure also for simplicial elements, based on the work of
Dubiner [46].
The study of the de Rham diagram of hp-finite element spaces, presented in Demkowicz et al. [44] and the implementation of H(curl)-conforming hp-finite elements on
hexahedral meshes, as presented in Rachowicz-Demkowicz [76].
Our construction provides a more general version of the exact sequence property of the
discrete global spaces, which we call the local exact sequence property, and formulate in
Theorem 5.32. This local exact sequence is a key property in our design of cheap and robust
preconditioners for curl-curl problems in Chapter 6.
To avoid confusion,we point out in advance that we distinguish between three levels of the
exact sequences of discrete spaces:
1. the exact sequence of the local FE-spaces defined on the reference element K, which we
use for motivating the choice of the appropriate local spaces,
2. the (discrete) global exact sequence of the global FE-spaces (cf. Corollary 5.34), which
is the classical one,
3. the local exact sequence of the discrete global spaces (cf. Theorem 5.32), which is a
generalization of the global exact sequence. By this we mean that the exact sequence
property even holds for the subspaces of a partially local space splitting.

5.1

High-order FE-spaces of variable order

de
lec [72, 73] is stated
The classical definition of the degrees of freedom as introduced in Ne
for uniform polynomial order p all over the mesh and does not generalize to finite element
spaces with varying polynomial order. One of our aims is to be able to vary the polynomial
order locally, i.e., in the sense of assigning to each edge, face and cell in the mesh an arbitrary
polynomial degree, while still ensuring conformity as well as the exactness of the discrete de
Rham complex.
We start with a discussion of using the degrees of freedom for constructing conforming finite
element spaces with varying polynomial order distribution over the mesh. The following
definition of degrees of freedom, enabling variable order tetrahedral elements and providing
commuting interpolation operators, is due to Demkowicz et al. [44].
We consider a bounded Lipschitz polyhedral domain , which is covered by a regular tetrahedral mesh Th . Let p = (pE1 , . . . pE6 , pF1 , . . . , pF4 , pC ) denote the polynomial degrees corresponding to edges, faces and the cell of the tetrahedral element K Th .

5.1. HIGH-ORDER FE-SPACES OF VARIABLE ORDER

61

H 1 -conforming hp FE-spaces
Let us define the following local finite element space:



Wp(K) := P p(K) := w P pC  w|F P pF (F ) F FK , w|E P pE (E) E EK . (5.1)

In order to make this definition reasonable we enforce a minimum order rule, i.e., for each
face F FK there holds
pC pF pE
edges E on F.
(5.2)
Following Demkowicz et al. [44] we define the hp-degrees of freedom as follows:
Vertex-based degrees of freedom are defined by the point-evaluations
NiV () = (Vi )

Vi VK

Edge-based degrees of freedom for each edge E EK :

Z
vi
E
ds
for {vi }2ipE spanning P0pE (E).
Ni () =
s
s
E
Face-based degrees of freedom for each face F :
Z
NiF () =
F F vi dx
for {vi }1inF , a basis of P0pF (F ),
F

with nF = 21 (pF +1)(pF +2) and the surface gradient is defined as F v := n(v)|F n.
Cell-based degrees of freedom
Z
NC () =
v dx
K

for {vi }1inC , a basis of P0pC (T )

with nC = 16 (pC + 3)(pC + 2)(pC + 1).

This choice defines a unisolvent and H 1 -conforming finite element, see Monk [70]. The
degrees of freedom can be used to define an interpolation operator W : W Wh , which is
well-defined for W = H s () for s > 32 . Under the assumption of uniform order p on all edges,
faces and cells, and for w H s () with s 2 the following interpolation error estimate holds:
kw W wkH 1 ()  hmin{p,s1} p1s kwks .
H(curl)-conforming hp-FE space
In order to obtain H(curl)-conformity, we define the local space such that the tangential traces
on edges and faces belong to polynomial spaces of some given order


3 

Vp(K) := v P pC (K)  tr,F (v) P pF (F ) F FK , tr,E (v) P pE (E) E EK ,
(5.3)
where we denote the tangential traces onto edges as trE, (v) = trE (v ) and onto faces as
trF, (v) = trF (n (v n)). In case of a minimum order rule (5.1), the polynomial space
Vp(K) is well-defined and there holds Wp+1 (K) Vp(K).

62

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Edge-based degrees of freedom:

Z
E
Ni (u) =
u vi ds
E

for {vi }0ipE , a basis of P pE (E).

Face-based degrees of freedom:

Z
F
Ni (u) =
curlF u curlF v i dA
ZF
NiF (u) =
u v i dA

F
(F ) ,
for {v i } s.t. {curlF v i } is a basis of curlF P p0,

Cell-based degrees of freedom:

Z
C
Ni (u) =
curl u curl v i dx
K
Z
NiC (u) =
u v i dx

C
for {v i } s.t. {curl v i } is a basis of curl P p,0
(K) ,

for {v i } a basis of F P0pF +1 (F ) .

F
with curlF v = curl v n and P p0,
(F ) := (P pF (F ))2 H0 (curl, F ).

with

C
P p0,
(K)

for {v i } a basis of P0pC +1 (K) .

:= (P pC (K))3 H0 (curl, K).

These degrees of freedom define an H(curl)-conforming unisolvent finite element. Interpolation error estimates for the corresponding commuting projection-based interpolation
operators are presented in Demkowicz-Buffa [43].
Below, we will explicitly construct edge-base, face-based, and cell-based basis functions, which
fit into this framework.
H(div)-conforming hp-FE space
The H(div)-conforming local space of order p = (pF1 , ..., pF4 , pC ) is defined as a vector-valued
polynomial space, where the normal traces on the faces of the element are polynomials of
order pFE :



(5.4)
Qp(K) = v P pC (K)  (v n)|F P pF F FK .
Again, for the spaces to be well-defined, we have to require that
pF pC

F FK .

The degrees of freedom are defined as follows:

Face-based degrees of freedom:
Z
F
Ni (q) =
q nvi dA
F

Cell -based degrees of freedom:

Z
C
Ni (q) =
div q div v i dx
ZK
NiC (q) =
q v i dx
K

with

C
P pn,0
(K)

for {vi } basis of P pF (F ).

C
for {v i } s.t. {div v i } is a basis of div P pn,0
(K) ,


C +1
for {v i } a basis of curl P p,0
(K) ,

:= P pC (K)3 H0 (div, K).

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

63

The above construction gives rise to an H(div)-conforming unisolvent finite element. Interpolation error estimates for commuting projection-based interpolation operators corresponding
to the above defined degrees of freedom are presented in Demkowicz-Buffa [43].

5.2

Construction of conforming shape functions

In the following we present a general approach for the conforming approximation of the
sequence of spaces, namely, H 1 (), H(curl, ), H(div) and L2 (). Our main concern lies in
enabling arbitrary and non-uniform polynomial order possibly varying for each edge,
face, cell,
a strategy which is applicable for all common elements, such as triangles, quadrilaterals,
hexahedra, tetrahedra and prisms and, therefore, can be used on hybrid meshes (e.g.
coming from geometrically refined meshes)
fulfilling the local exact sequence property as formulated in Theorem 5.32.
While eluding the application of interpolation operators, we are not concerned with the
definition of the degrees of freedom. Instead, we explicitly define the polynomial local finite
element spaces and construct H 1 -,H(curl)-, H(div)-conforming polynomial bases spanning
the local space.
The construction of the H 1 -conforming shape functions is rather familiar. We start with
Vertex-Edge-Face(only in 3d)-Cell based shape functions in tensor-product structure (possibly
degenerated in case of simplicial elements) as introduced by Dubiner [46] and KarniadakisSherwin [60]. In our approach we take care of the exact sequence property already during the
construction of the basis functions, in particular, we use gradient fields of higher-order H 1 functions and curl-fields of higher-order H(curl)-functions to construct Nedelec-Edge-FaceCell (N0 -E-F -C) based, respectively Raviart-Thomas-Face-Cell (RT 0 -F -C) based conforming
finite elements. The idea is visualized in Figure 5.1. We point out in advance, that the lowestorder space will always play a special role and therefore has to be treated separetely.

5.2.1

Preliminaries

Orthogonal polynomials
We intend to present a general method for constructing linearly independent shape functions
based on tensor products of one-dimensional orthogonal polynomials. Although we mainly
use Legendre-type polynomials later on, also other orthogonal polynomial bases spanning
P p ([1, 1]) respectively P0p ([1, 1]) could be used in our constructions, e.g. Gegenbauer, Jacobi, Hermitian polynomials. The main aspect of the choice of a special polynomial family
is the type of orthogonality, and the possibility of fast and recursive point-evaluation, which
are key ingredients for an efficient numerical implementation. Orthogonality influences the
sparsity and the condition number of the involved element matrices, whereas the fast evaluation procedures can be exploited in fast assembling techniques, i.e. in sum factorization
techniques mentioned in Karniadakis-Sherwin [60] section 4.1.5. For details on orthogonal
polynomials we refer to the classical textbook by Szego [88].

64

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Figure 5.1: Vertex-Edge-Cell-based H 1 -conforming shape functions on triangles and their

gradient fields. Edge-based and face-based gradient fields are used as edge-based and as
subset of face-based H(curl)-conforming shape functions.

Legendre polynomials, denoted by (i )0ip are L2 (1, 1)-orthogonal polynomials spanning P p ([1, 1]). In particular, they satisfy the orthogonality relation
Z 1
2
i (x)j (x) dx =
ij .
2i + 1
1
Due to the three-term recurrence relation
0 (x) = 1,
1 (x) = x,
(n + 1)n+1 (x) = (2n + 1)n (x)x nn1 (x),

n1

efficient and stable point evaluation is possible. The derivatives of Legendre polynomials can
be computed by an analogous 3 term recurrence. If both, the Legendre polynomials and their
derivatives, are needed at the same time, we recommend to use the reccurence


d
d
n (x) =
n+1 (x)
n1 (x)
dx
dx
together with the one for the Legendre polynomials. We only mention, that Legendre polynomials are a special case of Jacobi polynomials (see Appendix A.3).
Integrated Legendre polynomials, denoted by (Ln )2ip , are defined as follows:
Z x
n1 () d
for x [1, 1] and n 2.
Ln (x) :=
1

They are mutually orthogonal with respect to the H 1 -seminorm, i.e.,

Z 1
Li (x)Lj (x) dx = 0 for i 6= j.
1

65

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

0.5

-1

-0.5

0.5

-0.5

-1

Figure 5.2: Legendre polynomials of order 1 to 5

Moreover, the polynomials Ln , n 2 vanish at the interval bounds, i.e. Ln (1) = Ln (1) = 0
and span P0p [1, 1]. Once more, there holds a three-term-recurrence allowing fast pointevaluations:
L1 (x) = x,
L2 (x) = 12 (x2 1),
(n + 1)Ln+1 (x) = (2n 1) x Ln (x) (n 2) Ln1 (x) for n 2.
Here, L1 (x) was added to allow for a general recursion relation. Note that L1 (x) 6=

Rx

1 l0 (y)dy.

0.2
0.1
-1

-0.5

0.5

-0.1
-0.2
-0.3
-0.4
-0.5

Figure 5.3: Integrated Legendre polynomials of polynomial order 2 to 6

An interesting property of the integrated Legendre polynomials is that they are almost orthogonal also with respect to the L2 inner product, i.e.,
Z

Li (x)Lj (x) dx = 0

only for |i j| > 2.

Scaled Legendre-type polynomials For the construction of tensor-product-based shape

functions on triangular faces and tetrahedral cells we will use the Scaled Legendre Polynomials
Sn (x, t) := tn n

x

t

for x [t, t], t (0, 1]

66

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

and the Scaled Integrated Legendre Polynomials

LSn (x, t) := tn L n xt
Rx S
=
t n1 (s, t) ds

for n 2.

Note that the limit in zero is well defined and limt0 Sn (x, t) = 0 for n 2. The scaling
parameter t = 1 yields the Legendre or respectively the Integrated Legendre polynomials.
Thanks to the multiplication with the monomial tn , the functions are free of fractions and stay
polynomial of order n, i.e. nS P n (T ) and LnS P0n (T ) for T = [1, 1] [0, 1]. Moreover,
we obtain the following three-term recurrence relations:
0S (x, t) = 1,
1S (x, t) = x,
S
S (x), for n 2,
(n + 1)n+1 (x, t) = (2n + 1) x nS (x) n t2 n1
for the scaled Legendre polynomials, and
L1S (x, t) = x,
L2S (x, t) = 21 (x2 t2 ),
S (x, t) = (2n 1) x L S (x) (n 2) t2 L S (x) for n 3
(n + 1) Ln+1
n
n1

for the scaled integrated Legendre polynomials, respectively.

The orientation problem
One difference of (hierarchical) higher order methods to the well-known Lagrange-type elements is that the degrees of freedom not longer refer exclusively to point-evaluations. In order
to enforce continuity over element-interfaces (edges and faces), the orientation of edges and
faces becomes important (see Figure 5.4). There are two possibilities to deal with this fact:
either one takes the local edge/face orientation into account in the global assembling process,
or one introduces a globally unique orientation of edges and faces, which is available on the
local-element level.

0.2
0.1
0
-0.1
-0.2
0

1
0.75
0.5
0.25

0.2
0.1
0
-0.1
-0.2
0

1
0.75
0.5
0.25

0.25

0.5
0.75

1 0

0.25

0.5
0.75

1 0

Figure 5.4: Edge-based basis function with correct and converse edge orientation
The first approach is quite simple in 2D, where the problem of locally converse running edges
can be easily resolved by a flip of signs of the related edge-based shape functions during the

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

67

global assembling process. For tetrahedra, however, such a procedure is rather sophisticated:
one has to define two types of reference elements, onto which all global elements can be
rotated. See Ainsworth-Coyle [2] for details.
In this work we address the orientation problem by introducing reference elements which
are parameterized by the global vertex indices of the corresponding physical element. By
introducing a unique edge and face orientation based on the global vertex indices, the
be a reference element
global orientation is also locally available on each edge/face. Let K
of arbitrary shape. On the local element level we assume to have access to the global
vertex-indices VK VK of the local vertices {1, ..., nV } = VK .
We define the edge orientation always pointing from the vertex with the higher global vertex
number to the one with the lower one. This means, if the element-edge E connects the local
vertices 1E and 2E , we achieve the edge-orientation on the element-level as from the local
vertex e1 to the local vertex e2 with
e1 := arg max{E ,E } {vK },
2
1
e2 := arg min{E ,E } {vK },
2

i.e.

veK1 > veK2 .

(5.6)

We refer to an oriented edge by brackets, e.g., E = [e1 , e2 ] EK .

The triangular face orientation of the element-face F including the three local vertices
1F , 2F , 3F is defined in the following way: First we set
f1
f3
f2

:=
:=
:=

arg max{F1 ,F2 ,F3 } {vK },

arg min{F1 ,F2 ,F3 } {vK },
3 mod3 (f1 + f3 ),

i.e.

vfK1 > vfK2 > vfK3 .

(5.7)

Then we refer to an oriented triangular face by F = [f1 , f2 , f3 ] FK .

The unique orientation of a local quadrilateral face F = {1F , 2F , 3F , 4F } FK is defined
similarly: we start at the vertex f1 with maximal global vertex number and then proceed in
the direction of the adjacent vertex which has the highest vertex number, i.e., we set
F = [f1 , f2 , f3 , f4 ]

s.t.

K
f1 := arg max{F1 ,F2 ,F3 ,F4 } {v },
f3 opposite f1 on F,

vf2 > vf4 .

(5.8)

The local vertex f2 is either a horizontal or vertical neighbor of f1 , vice versa for vertex f4 .
Remark 5.1. Taking into account the orientation, is necessary for enforcing conformity
in the construction of global hierarchical high-order finite element spaces. In the following
we construct conforming local high-order shape functions on a reference element, which is
parameterized by the global vertex numbers. However, the presented construction of local
bases is not affected by the choice of orientation.
If one chooses the strategy proposed by Ainworth-Coyle [2], one simply has to exchange
the edge orientations, denoted by [e1 , e2 ], and face orientations, denoted by [f1 , f2 , f3 ] or
[f1 , f2 , f3 , f4 ] by the local edge and face orientation of the considered reference element.

68

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

5.2.2

The quadrilateral element

We define the quadrilateral reference element as Q = [0, 1] [0, 1]. Analogous to the construction of shape functions for simplicial shape functions in terms of barycentric coordinates, we
state a general construction of shape functions in terms of the of the bilinear functions i ,
equal one at the vertex i and zero at all other vertices, and the linear functions i for the
quadrilateral Q as follows:
(0,1)

(1,1)

4
3
11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111

1
(0,0)

2
(1,0)

1
2
3
4

:= (1 x)(1 y)
:= x(1 y)
:= xy
:= (1 x)y

1
2
3
4

:= (1 x) + (1 y)
:= x + (1 y)
:= x + y
:= (1 x) + y

(5.9)

This notation will simplify the construction of the spaces H(curl) and H(div) later on, and
is also helpful in the numerical realization.
The edge E = [e1 , e2 ] pointing from vertex e1 to e2 can be parameterized over the interval
I = [1, 1] by
E = e2 e1 [1, 1].

The unit tangential vector E and the outer unit normal vector nE of the edge E can be
deduced by
1
E = (e2 e1 ) and nE = (e1 + e2 )
2

for E = [e1 , e2 ].

(5.10)

We will later extend functions on the edge E into the domain Q. For this purpose, we define
the linear edge-extension parameter
E = e1 + e2 [0, 1],
which is one on E and zero on the edge opposite of E.
We construct now a sequence of local finite element spaces with uniform polynomial order
k 1 of the following form:
id

curl

(5.11)

id

Curl

div

(5.12)

R Qk+1 (Q) Qk,k+1 (Q) Qk+1,k (Q) Qk (Q) {0}

and

R Qk+1 (Q) Qk+1,k (Q) Qk,k+1 (Q) Qk (Q) {0},

cf. (3.15), (3.16) for the corresponding sequences on the continuous level.
The H 1 -conforming quadrilateral element

We construct the H 1 -conforming space as a Vertex-Edge-Cell (V -E-C) based hierarchical

polynomial space:
For the vertex-based shape functions we choose the lowest order bilinear functions
Vi (x) = i (x),

for i = 1, . . . , 4.

Edge-based shape functions associated with the edge E = [e1 , e2 ] are constructed by

69

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

1. specification of pE 1 linearly independent polynomials on the edge such that {trE (E

i )}i
span P0pE (E). This is achieved by the reparameterization of the edge E by E on
I = [1, 1] and using a basis {Li }2ipE of P0p (I), which need not to be necessarily
Integrated Legendre polynomials, i.e.,
trE (E
i )(x) = Li+2 (E (x))

for 0 i pE 2,

2. polynomial lifting of the edge values into the interior of the element such that E
i
P pE (Q) with vanishing trace on the three remaining edges. For the sake of simplicity,
we will use a linear extension by E , i.e.,




for 0 i pE 2.
E
i (x) = Li+2 E (x) e1 (x) + e2 (x)

We define Cell -based shape functions which vanish on the whole boundary K and span
Q0k (Q). This is achieved by the following tensor-product construction:
C
ij (x) := Li+2 (2x 1) Lj+2 (2y 1)

for 0 i, j pC 2.

Summarizing, we have
Hierarchical H 1 -conforming quadrilateral element of variable order p = ({pE }, pC )

Vertex-based functions
for i = 1, ..., 4 :
Vi = i
Edge-based functions

for m = 1, 2, 3, 4: Suppose the edge Em = [e1 , e2 ].

m
E
= Li+2 (e2 e1 ) (e1 + e2 )
i

for i pEm 2 :
Cell-based functions

C
ij = Li+2 (2x 1) Lj+2 (2y 1)

for 0 i, j pC 2 :

We define the local spaces W V (Q)

:= span((Vm )1m4 ),
m
WpEEmm (Q) := span((E
i )1ipEm 1 ), m = 1, ..., 4,
WpCC (Q)
:= span((C
ij )1i,jpC ),
and, finally the local FE-space on the quadrilateral for variable polynomial degree p =
(pE1 , pE2 , pE3 , pC ) by
Wp(Q) := W V (Q)

4
M

m=1

WpEEmm (Q) WpCC (Q).

(5.13)

Theorem 5.2. The V -E-C-based shape functions presented above define an H 1 -conforming
finite element basis on the quadrilateral Q. In particular, they are linearly independent and
for uniform polynomial order p = pEm = pC there holds Wp(Q) = Qp (Q).
Proof. We consider that a trivial element v can be constructed via a linear combination of
shape functions, i.e.,
v=

4
X
i=1

cVi Vi

4 pEX
m 1
X

m=1

i=1

m Em
+
cE
i i

pX
I 1

i,j=1

C
cC
ij ij = 0.

70

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

To prove linear independence of the basis, we have to show that all coefficients are zero:
We know that the only shape function not vanishing in the vertex Vi is the one associated
with this vertex, i.e. v(Vi ) = 0 implies cVi = 0.
Next we restrict v to each edge Em . Since the cell-based shape functions vanish on all
edges as well as all edge-based
shape functions
associated with Ek 6= Em vanish on Em , we


PpEm 1
Em Em
= 0. Since the edge-based functions are linearly
ci i
obtain trEm (v) = trEm
i=1
m
are zero.
independent and span Pp0E (Em ) on the edge, the coefficients cE
i
Considering traces of the cell-based shape functions along horizontal or vertical lines:
P
P
v(x, y) = i,j cij Li (2x 1)Lj (2y 1) = i ci Li (2x 1) = 0
P
implying ci :=
j cij Lj (2y 1) = 0 for all y [0, 1], we can reduce the analysis of the
linear-independence to the 1D case.
Hence, we obtain linear independence of all shape functions.
For uniform polynomial order p, every shape function lies in Qp (Q) by construction. By a
simple counting argument we obtain that the stated shape functions are a basis of Qp (Q):
P
|Wp (Q)| = |W V (Q)| + 4m=1 |WpEm (Q)| + |WpC (Q)|
= 4 + 4(p 1) + (p 1)2 = (p + 1)2 , and
p
|Q (Q)| = (p + 1)2 .
Remark 5.3.

1. For p = pC pEm the shape functions form a basis for



Wp(Q) = w P p (Q) : trE (w) P pEm (Em ) for m = 1, ..., 4 .

2. The common construction of the shape functions by Integrated-Legendre polynomials

(Li )2ip . is not necessary. In view of conditioning and sparsity of the resulting
FE-matrices one can exchange the family of Integrated polynomials by other bases of
P0p ([1, 1]).
3. For pC pEm , the construction of edge-based shape functions is not limited to the linear
extension. One can choose any lifting of the edge onto P pEm (Q) satisfying vanishing
trace on the remaining edges.
H(curl)-conforming shape functions for quadrilaterals
In view of the exact sequence property, we want to construct an H(curl)-conforming finite
element basis for a local FE-space Vp(Q) such that
Wp+1 (Q) Vp(Q).
For uniform polynomial order p, this leads to the choice Wp+1 (Q) = Qp+1 (Q), and hence
Vp (Q) = Qp+1,p (Q) Qp,p+1 (Q)

with dim Vp (Q) = 2(p + 2)(p + 1).

The lowest-order shape functions are chosen to be shape functions for the lowest-order Nedelec
element, asking for the tangential component to be one on the associated edge, and zero on
all other edges. This can be written in the compact form
1
N0
= (e1 e2 ).(e1 + e2 )
m
2

for Em = [e1 , e2 ].

71

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

The functions lie in Q0,1 (Q) Q1,0 (Q). Since the tangential vector of the edge Em = [e1 , e2 ]
is given by Em = 12 (e2 e1 ), the tangential trace on an edge Ek is
0
trEk ,Ek (N
m ) = (e1 + e2 ) Em Ek =

0
on Ek Em ,
m,k 1 on Ek || Em .

The higher-order edge-based H(curl)-conforming shape functions can be chosen as gradientfields of the corresponding H 1 -conforming edge-based shape functions, i.e. for 0 i pEm 1
Em
m
E
i (x) = i (x)

m
WpEEm +1 (Q).
for E
i
m

Since the trace of a scalar edge-based shape function on the edge Ek fulfills
m
trEk (E
i )=

Li+2
0

on Ek = Em
on Ek , k 6= m

with span{Li+2 }i=0,...,pEm 1 = P0 Em

+1

(Em ),

the tangential components of the gradients satisfy



2Li+2 (E (x)) on Em = Ek
Em

E
m
tr,Ek (i )(x) = E (x) =
0
on Ek , k 6= m
k
m
Qi+1,i+2 (Q) Qi+2,i+1 (Q), and due to the linear independence
Note that there holds E
i
of the Integrated Legendre polynomials we obtain

pEm
m
(Em )/R.
span{tr,Em (E
i ) : 0 i pEm 1} = P

We pursue this strategy also for construction of the cell-based shape functions. The gradients
p+1
2
of cell-based shape functions {C
ij }i,j=0,...,p1 spanning Q0 (Q) form p linearly independent
functions in Qp,p+1 (Q) Qp+1,p (Q) having zero tangential trace on the element boundary Q.
Therefore, we start by choosing

1
C
ij = 2Li+2 (2x 1)Lj+2 (2y 1)ex + 2Li+2 (2x 1)Lj+2 (2y 1)ey .

Having a closer look at the contributing terms to the gradient functions, namely

Li+2 (2x 1)Lj+2 (2y 1)ex and Li+2 (2x 1)Lj+2 (2y 1)ey , we observe that each term on
its own fulfills the requirements of cell-based shape functions, i.e., they lie in the required
polynomial space Qp,p+1 (Q) Qp+1,p (Q) and have zero tangential trace, since either ex E
or Lj+2 (2y 1) = Lj+2 (1) = 0 on any edge E. Moreover, the functions defined by the
two contributions are obviously linearly independent. Therefore, we can add any linearly
independent combination of the two terms to the set of cell-based shape functions. We suggest

2
C
ij = Li+2 (2x 1)Lj+2 (2y 1)ex Li+2 (2x 1)Lj+2 (2y 1)ey .

Finally, also the functions

(Li+2 (2y + 1)ex )0ip1

and

(Li+2 (2x 1)ey )0ip1 ,

have zero tangential trace on the boundary, and they are linearly independent of the other
cell-based shape functions chosen so far, since they are the only ones which are constant either
in x- or in y- direction.
Summarizing we arrive at the following set of shape functions:

72

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Hierarchical quadrilateral H(curl)-element of variable order p = ({pE }, pC )

Edge-based shape functions
for m = 1, 2, 3, 4: edge Em = [e1 , e2 ]
Lowest-order edge shape function


1
0
N
m = 2 e2 e1 (e1 + e2 )

Higher-order edge-based functions (gradient fields)

m
E
= Li+2 (e2 e1 ) (e1 + e2 )
i

for 0 i pEm 1
Cell -based functions
Type 1: (gradient fields):

1
C
ij

for 0 i, j pC 1

= Li+2 (2x 1) Lj+2 (2y 1)

Type 2:
C2
ij

for 0 i, j pC 1

= Li+2 (2x 1) Lj+2 (2y 1) ex Li+2 (2x 1) Lj+2 (2y 1) ey

Type 3:
3
C
i
C3
i+pC

for 0 i pC 1

= Li+2 (2y 1) ex
= Li+2 (2x 1) ey

0
:= span((N
m )1m4 ),
C
VpCC1 (Q) := span((ij1 )0i,jpC 1 ),
3
VpCC3 (Q) := span((C
i )0i2pC 1 ),

(Q)
VpEEm
m
VpCC2 (Q)

We define N0 (Q)

m
:= span((E
i )0ipEm 1 ),
C2
:= span((ij )0i,jpC 1 ),

and the local FE-space for the quadrilateral of variable polynomial degree p
(pE1 , pE2 , pE3 , pC ) by
Vp(Q) := N0 (Q)

4
M

m=1

(Q) VpCC (Q).

VpEEm
m

(5.14)

Theorem 5.4. The N0 -E-C based shape functions defined above, form a local H(curl)conforming finite element basis on the quadrilateral Q. In case of uniform polynomial degree
p = pC = pEm for all edges Em they are a basis for Qp,p+1 (Q) Qp+1,p (Q).
Furthermore, there hold the following relations:
(Q)
WpEEm +1 (Q) = VpEEm
m
m

for all edgesEm , m = 1, ..., 4,

and
WpCC +1 (Q) = VpCC1 (Q) VpCC (Q).
Proof. The linear independence of the pEm edge-based shape functions is implied by their construction as gradient fields of linearly independent polynomials with degree greater than 2.
In combination with the lowest-order Nedelec function their tangential traces span PpEm (Em )
on the edge Em , and are zero on all other edges. Concerning the cell-based functions, we constructed 2(pC +1)(pC +2) linearly independent shape functions in QpC ,pC +1 (Q)QpC +1,pC (Q)
with vanishing trace on all edges.
The hierarchical N0 Em C construction for the tangential components implies the overall
linear independence: assume that v = 0 can be built by an arbitrary linear combination of

73

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

the shape functions, i.e.

4
X
 N0 N0 X Em Em  X C C
v=
+
cij ij = 0
cm m +
ci i
m=1

ij

In the first step we consider the tangential traces over the edges, i.e. tr,Em (v) = 0, which
m
to be zero due to the linear independence of edgeimplies all edge-associated coefficients cE
i
based functions. The remaining contribution to v is then spanned only by cell-based functions,
which are a set of linearly independent polynomials. Hence, all coefficients in the linear
combination have to vanish, and we obtain the linear independence of the shape functions.
By a counting argument we have
P
|Vp (Q)| = |VQN0 | + 4m=1 |VQEm | + |VQI |
= 4 + 4p + 2p(p + 1) = 2(p + 2)(p + 1),

|Qp,p+1 (Q)

Qp+1,p (Q)|

and

= 2(p + 2)(p + 1),

and hence the shape functions span the whole space Qp,p+1 (Q) Qp+1,p .
In case pEm pC for all m = 1, . . . 4 the H(curl)-conforming local FE-space reads




Vp (Q) = Qp+1,p (Q) Qp,p+1 (Q)  trE () E = tr ,Em () P pEm (Em ), m = 1, . . . , 4 .

Remark 5.5 (H(div)-conforming shape functions for the quadrilateral element Q). In two
dimensions the H(div)-conforming shape functions can be generated out of the H(curl)conforming ones with a rotation by 90 degrees. By this, we achieve:
Hierarchical quadrilateral H(div)-element of variable order p = ({pE }, pC )
Edge-based shape functions: for m = 1, 2, 3, 4: edge Em = [e1 , e2 ]


0
RT
= 21 Curl e2 e1 (e1 + e2 )
m


m
0 i pEm 1
= Curl Li+2 (e2 e1 ) (e1 + e2 )
E
i
Cell -based functions:


1
= Curl Li+2 (2x 1) Lj+2 (2y 1)
C
ij

2 Li+2 (2x 1) Lj+2 (2y 1) ey Li+2 (2x 1) Lj+2 (2y 1) ex , 0 i, j pC 1

3
C
i
C3
i+pC

5.2.3

2
C
ij

= Li+2 (2x 1) Lj+2 (2y 1) ey + Li+2 (2x 1) Lj+2 (2y 1) ex ,

= Li+2 (2y 1) ey ,

= Li+2 (2x 1) ex ,

0 i, j pC 1

0 i pC 1

0 i pC 1

The triangular element

The triangular reference element is defined by


T := {(x, y) 0 x, y 1, x + y 1},

which involves the vertices V1 = (0, 0), V2 = (1, 0), and V3 = (0, 1). The shape functions will
be formulated in terms of barycentric coordinates.
y

(0,1)
3
11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
1
(0,0)

2
(1,0)

1 x y,

:=

:= x,

:= y.

(5.15)

74

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

By viewing the triangle as a collapsed quadrilateral, as suggested in Dubiner [46] and

Karniadakis, Sherwin [60], we can construct a tensorial-type basis also for triangles in a
similar manner as for quadrilaterals.
The Duffy transformation
D : Q = [1, 1]2
T
(, )
(x, y)

defined as

x =
y =

1
4 (1
1
2 (1

+ )(1 )
+ )

transforms the quadrilateral Q to the triangle T (see Figure 5.5). Using the inverse of the
Duffy transformation, we can parameterize the triangle T by
1
x
1 = 12 +
[1, 1],
= 2 1y
2
= 2y 1 = 23 1 = 1 21 22 [1, 1],

where, as above 1 , 2 , 3 denote the barycentric coordinates of the triangle.

D1 : (x, y) (, )

(0,1)
3

(1,1)
4

(1,1)
3

1
(0,0)

D : (, ) (x, y)

(1,0)

(1,1)

2
(1,1)

Figure 5.5: Collapsed coordinate system: triangle as degenerated quadrilateral

The horizontal lines (3 ) = = const. in the triangle are then mapped by (1 , 2 ) onto
[1, 1], while = = const. yields a parameterization of the lines radiating out of the top
vertex V3 over [1, 1], see Figure 5.5. Due to the transformation above, we can reparameterize
an edge E = [e1 , e1 ] connecting the vertices e1 , e2 and all lines running parallel to E (bounded
by the two other edges) over the reference interval [1, 1] by
E =

e2 e1
.
e1 + e2

The outer normal vector nE and tangential vector E of an edge E = [e1 , e2 ] with e3 denoting
the opposite vertex can be expressed in barycentric coordinates by
nE = e3 = (e1 + e2 ).
We now construct local finite element spaces on the triangle T according to the sequences
id

curl

id

Curl

div

R Pp+1 (T ) Pp (T ) Pp1 (T ) {0}

respectively
R Pp+1 (T ) Pp (T ) Pp1 (T ) {0},
cf. (3.15), (3.16) in the continuous setting.

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

75

The H 1 -conforming triangular element

In the following we introduce a V -E-C-based hierarchical local finite element basis based on
the collapsed tensor-product representation of the triangle as used in, e.g., Dubiner [46],
Karniadakis-Sherwin [60]. For uniform polynomial order, the basis will span
1
Wp (T ) = P p (T ) with dim Vp (T ) = (p + 1)(p + 2).
2
As in the quadrilateral case, the vertex-based shape functions are chosen as the lowest-order
shape functions spanning P 1 (T ), i.e.,
Vi (x) = i (x)

for i = 1, 2, 3.

For edge-based shape functions we choose the scaled Integrated Legendre polynomials of order
2 i + 1 pEm
S
m
:= Li+2
(e2 e1 , e1 + e2 ) = Li+2
E
i

e2 e1

(e1 + e2 )i+2 .
e1 + e2

(5.16)

W shortly summarize the main properties of these functions:

m
m
are zero in all vertices, i.e. E
The functions E
i (Vk ) = 0, k = 1, . . . , 3.
i
(
Li+2 (E )
if Ek = Em
m
on any edge Ek EK ,
Since trEk (E
i )=
i+2
Li+2 (1)(e1 + e2 )
= 0 if Ek 6= Em
the traces of the edge-based shape functions span P0p (Em ) on the corresponding edge
Em , while vanishing on the other two edges.

The extension of edge functions Li+2 () onto the triangle by multiplication with factor
(e1 + e2 )i+2 corresponds to an extension by ((1 )/2)i+2 in the quadrilateral coordinates (Duffy transformation). Hence we refer to this extension as monomial extension.
Finally we define the cell-based shape functions as
S
C
(i,j) = Li+2 (2 1 , 1 + 2 ) 3 j (23 1)

(5.17)

for 1 i + j pC 3. These functions have following properties:

1
They are bubble functions in P0p (T ). The trace vanishes on T , since 21
+2 = 1 on
the edges opposite vertex 1, 2 and 3 vanishes on the edge opposite to vertex 3.
The monomial extension, which is performed by using Scaled Integrated Legendere
p
polynomials, implies that C
(i,j) P0 (T ).

By the inverse Duffy transformation we can interprete the cell-based shape functions in
terms of the quadrilateral coordinates (, ) which yields




1 i+2 + 1
(5.18)
C
j ().

D
(,
)
=
L
()
i+2
ij
2
2
In particular, on the horizontal lines of the triangle, i.e. 3 = const., the functions can
be written as an Integrated Legendre polynomial, since


1 2
C
P0i+2 ([1, 1])
= cij Li+2
ij |
=const.
1 + 2
with cij = 3 j (23 1) (1 3 )i+2 (confer Figure 5.5).

76

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

On ray lines of the triangle, i.e.

C
ij |

=const.

2 1
1 +2

= const., we obtain

= ci 3 j (23 1) (1 3 )i+2 P j+i+3 ([1, 1])

2
with ci = Li+2 ( 11
+2 ).

Summarizing, we obtain the following set of shape functions.

Hierarchical H 1 -conforming shape functions on the triangular element T
of variable order p = ({pE }, pC )
Vertex-based functions
for i = 1, 2, 3 :
Vi = i
Edge-based functions
for m = 1, 2, 3: edge Em = [e1 , e2 ]
for 0 i pEm 2 :

S
m
E
= Li+2
(e2 e1 , e1 + e2 )
i

Cell -based functions

for 0 i + j pC 3, i, j 1 :

S
C
(i,j) = Li+2 (2 1 , 1 + 2 ) 3 j (23 1)

We define the local block spaces via

W V (T )
:= span((Vi )1m4 ),
Em
m
WpEm (T ) := span((E
i )0ipEm 2 ),
WpCC (T )
:= span((C
i,j )0i+jpC 3 ),

and the local FE-space of variable polynomial order p = (pE1 , pE2 , pE3 , pC )
Wp(T ) := W V (T )

3
M

m=1

WpEEmm (T ) WpCC (T ).

(5.19)

Theorem 5.6. The local V -E-Cshape functions defined above are H 1 -conforming and linearly
independent. For uniform order p = pEm = pC they form a basis for P p (T ).
Proof. It remains to prove the linear independence of the shape functions. We consider the
following arbitrary linear combination:
w=

3
X
i=1

cVi Vi

3 pEX
m 2
X

m=1

i=1

m Em
+
cE
i i

C
cC
ij ij = 0.

i, j 0
i + j pC 3

Restricting w to the vertices Vi yields cVi = 0. On edge Em the only non-zero functions,
except for P
the vertex-based functions, are those associated with this edge. Hence, we obtain
Em
Em
trE (w) = i cE
= 0 for all i due to the linear-independence
i tr E (i ) = 0, which yields ci
Em
of {i } on Em .
Hence, there are only cell-based contributions in the expansion of w left. Thanks to the
tensor-product representation (5.18), the cell-based shape functions are linear-indpendent.
This yields linear-independence of the whole set of shape functions.
The V -E-C-based construction implies H 1 -conformity, e.g. considering an edge E = [e1 , e2 ]:
only the two vertex-based shape functions associated with e1 and e2 and the edge-based
shape functions corresponding to E, have non-zero trace.
In the case of uniform polynomial order all shape functions are in P p (T ) by construction. The
counting argument

77

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

P
|Wp (T )| = |W V (T )| + 3i=1 |WpEEmm (T )| + |WpCC (T )|
= 3 + 3(p 1) + 12 (p 2)(p 1) = 21 (p + 1)(p + 2),
|P p (T )| = 21 (p + 2)(p + 1)

and

verifies that for uniform order p we achieve a local basis of P p (T ).

Jacobi-based shape functions We already mentioned that the construction scheme of the
FE-basis functions is independent of the special choice of the type of the involved orthogonal
polynomials. On the contrary, the conditioning of the involved FE-matrices as well as the
fast evaluation of shape functions depends essentially on the chosen polynomial type. We
can improve the conditioning and sparsity of the FE-matrices by exploiting the orthogonality
within the hierarchical blocks.
As an alternative to the Legendre-type polynomials, let us consider now Jacobi polynomials
(,)
Pn
as defined in Appendix A.3 for the construction of the cell-based shape functions. A
first simple, but motivating suggestion for the basis functions, which is similar to the one
given in Dubiner [46], is
S,(2,2)

C
(i,j) = 1 2 Pi

(2i+5,2)

(1 2 , 1 + 2 ) 3 Pj

(23 1)

for 0 i + j pC 3. Due to the inverse Duffy transformation we obtain

(2,2)

2
(C
(i,j) D)(, ) = (1 )(1 + )(1 ) Pi

()

1
i 1 + (2i+5,2)
Pj
().
2
2

Using this identity in the computation of the L2 -inner product on the triangle we achieve
Z

C
(i,j)

C
(k,l)

dx

1
2

1
C
C
(i,j) D (, ) (k, l) D (, ) (1 )d d
8

(2,2)

(1 )2 (1 + )2 Pi

(2,2)

()Pk

() d

1 i+k+5
(2i+5,2)
(2k+5,2)
)
(1 + )2 Pj
() Pl
() d
2

= cij ik jl

Hence, this set of cell-based functions is L2 -orthogonal.

In case we are interested in sparsity of the stiffness matrix, it is suggested in Beuchler berl [17] to choose
Scho
(2i+1,0)

C
(i,j) = Li+2 (1 2 , 1 + 2 ) Pj+1

(23 1),

(,)
(,)
where we use integrated Jacobi polynomials Pj+1 (x) satisfying Pj
(1) = 0. This choice
yields

C
T
T ((i,j) )

C
(k,l) dx = 0

for |i k| > 2 or |i k + j l| > 2

and therefore, we again obtain sparsity of the stiffness matrix.

78

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Tensor-product representation of cell-based H 1 -conforming shape functions. In

the construction of H(curl)- and H(div)-conforming shape functions we will frequently make
use of the following product representation of H 1 -conforming cell-based shape functions:
C
ij (x, y) = ui (x, y) vj (y)

for i, j 0, i + j pC 3,

(5.20)

with ui P i+2 (T ) vanishing on the edges E1 = [1, 3] and E2 = [2, 3] and vj P j+1 (T )
vanishing on E3 = [1, 2]. Within the family of Legendre-type functions, this can be provided
by
S
ui (x, y) := Li+2
(1 2 , 1 + 2 ), vj (y) := 3 j (23 1),
(5.21)

or equivalently in the Duffy-transformed quadrilateral coordinates




1+


1 i+2
j ().
, vj D(, ) =
ui D(, ) = Li+2 ()
2
2

(5.22)

H(curl)-conforming triangular element

We define the local FE-space Vp with Wp+1 Vp such that, for uniform polynomial order
p,

2
Vp (T ) = P p (T )
with dim(P p (T ))2 = (p + 2)(p + 1)
holds. Note that we have P p+1 (T ) Vp (T ). In the construction of the basis we first choose
gradients of the edge- and cell-based shape functions of Wp+1 as shape functions in the
H(curl)-conforming FE-basis and then extend the shape functions to span the full polynomial
spaces on edges and cells, where the latter is defined by


2

C
P p0,
(T ) := q PCp (T ) : tr,T (q) = 0 .
Let us now start with the construction of shape functions for uniform polynomial order p:
First we choose the low-order shape functions as the lowest-order Nedelec type I functions
0
N
m = e1 e2 e1 e2 on each edge Em = [e1 , e2 ]

0
with Ek N
m |E = mk .
k
The higher order edge-based Nedelec shape functions can be chosen as the gradients of the
H 1 -conforming scalar functions of order p
m
m
for i = 0, . . . , p 1.
= E
E
i
i


m
spans P pEm (Em )/R was already
The verification that trE (E
i ) Em : 0 i pEm 1
done in the quadrilateral case. Moreover, we obtain zero tangential trace on all edges Ek with
k 6= m.
It remains to define appropriate cell-based shape functions: taking the gradients of cell-based
H 1 -conforming shape functions up to order p + 1 yields 12 (p 1) p linearly independent shape
functions in P p0, (T ). In view of the exact sequence property, we therefore choose the first
type of H(curl)-conforming cell-based shape functions as:

C
C,1
(i,j) (x, y) = (i,j) (x, y) = ui (x, y) vj + ui (x, y) vj (y)

(5.23)

for 0 i + j p 2 and i, j 0.
For defining further cell-based shape functions we take the underlying product representation
(5.20) of scalar functions into account. Let us first consider the contributions to the gradient
fields in more detail.

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

79

S ( , + ) and v := l (2 1). Then the set

Lemma 5.7. Let ui := Li+2
2
1 1
2
j
3 j
3
{ui vj , ui vj , 0 i + j p 2} forms p (p + 1) linearly independent functions in P p0, (T ).

Proof.
1
= 1
1. The trace of ui vanishes on the two edges opposite to vertex V1 and V2 , since 12 +
2
there. The trace of vj vanishes on the edge opposite to vertex V3 , since 3 = 0. Hence,
the tangential trace of ui vj and vj ui vanishes on all edges. Since ui P i+2 (T ) and
vj P j+1 (T ), we obtain ui vj , ui vj P p0, (T ).

2. To show linear independence we first consider the two sets of functions separately.
In order to compute gradients on the transformed quadrilateral, we
! need the Jacobian
4
0
1
. This yields:
of the inverse Duffy transformation, i.e. FDT =
2
( + 1) 1 2


xvj (x) ui (x, y) = FDT (,) vj D(, ) ui (x, y)

1
i+2
= 2 lj () 1+
Li+2 ()e2 ,
2
2

which implies that {ui vj }i,j is a set of linearly independent functions.

Concerning {ui vj }i,j we consider the derivative in direction of the horizontal lines,
denoted by H , more precisely:
4
)vj
H ui (x, y) vj (y) = ( 1

= Li+2 ()

ui (D(, ))
1
i+1 1+
2
2 lj ()

(5.24)

which are linearly independent for all 0 i + j p 2.

We can easily verify that the two sets of shape functions are linearly independent by taking into account that vj is constant on horizontal lines, which implies
H vj (y) ui (x, y) = 0.

Due to the linear independence, we can use for the second type of cell-based Nedelec shape
functions any combination of ui vj and ui vj , which is linearly independent of C,1
(i,j) , e.g.,
2
C
ij = ui vj ui vj

for 0 i + j p 2.

Finally, we define a third family of cell-based shape functions which is linearly independent
of the first two ones by


3
C
j = 1 2 1 2 vj

for 0 j p 2.

Here also the lowest-order Nedelec function 1 2 1 2 is involved, and the functions
3
C
j are constant along horizontal lines.
The following table collects the suggested H(curl)-conforming shape functions for the triangular master element T of variable order.

80

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Hierarchical high-order triangular H(curl)-conforming element

of variable polynomial order p = ({pE }, pC )
Edge-based shape functions
for m = 1, 2, 3: edge E m = [e1 , e2 ]
Low-order functions:
N0
= e1 e2 e1 e2
m

Higher order edge-based functions (gradient fields):

for 0 i pEm 1
S
iEm = (Li+2
(e1 e2 , e2 + e1 ))

Cell-based shape functions:

We define ui
vj

S
:= Li+2
(2 1 , 1 + 2 ),
:= 3 j (23 1).

for 0 i + j pC 2, i, j 0 :
Type 1 (Gradient fields):

C,1
(i,j) = (ui vj ) = ui vj + ui vj

Type 2:

C,2
(i,j) = ui vj ui vj

Type 3:

C,3
(0,j) = (1 2 1 2 ) vj

We define the local N0 -E-C-based spaces via

0
V N0 (T ) := span((N
i )1m3 ),
E
(T ) := span((i m )0ipEm 1 ),
VpEEm
m
C
VpC (T )
:= span((C
i,j )0i+jpC 2 )

The local FE-space of variable polynomial order p = (pE1 , pE2 , pE3 , pC ) then is
Vp(T ) := V V (T )

3
M

m=1

(T ) VpCC (T ).
VpEEm
m

Theorem 5.8. The N0 -E-C-based shape functions, summarized in the table above, are
H(curl)-conforming and linearly independent. For uniform polynomial order p = pEm = pC ,
we obtain a local H(curl)-conforming FE-basis for Vp (T ) = (P p (T ))2 .
Furthermore, there holds Wp Vp, and the construction involving gradient functions provides
(T )
as well as
WpCC +1 (T ) VpCC (T ).
WpEEm +1 (T ) = VpEEm
m
m

Proof. The edge-based shape functions are gradients of the corresponding (H 1 ) edge-based
m
(P pEm (T ))2 , their
scalar functions (up to order pEm + 1). Hence, we obtain that E
i
p
tangential trace span PEm (Em )/R on Em , and tangential traces vanishes on Ek , k 6= m. This
implies linear independence within the set of all edge-based functions, including lowest-order
0
Nedelec functions N
m .
Due to Lemma 5.8 the set of cell-based functions of type 1 and type 2 span {ui vj , ui vj } and
N0
pC (T ))2 .
3
are linearly independent. The functions C
j = 3 vj of the third type belong to (P

81

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

N0
3
Moreover, there holds tr,T (C
j ) = 0, since the tangential trace of 3 vanishes on the notassociated edges E1 and E2 and because the tangential trace of vj vanishes on E3 = [V1 , V2 ].
Considering equation (5.24), we observe that along the horizontal lines the function ui vj
0
are at least linear, while vj ui vanishes on horizontal lines (3 = const.). Since N
3 and
vj are constant along the horizontal lines, we obtain linear-independence within the set of
C
cell-based shape functions. Moreover, the set of all cell-based shape functions spans P p0,
(T ).
Since the shape functions associated with an edge Em are linearly independent, and they are
the only functions with non-vanishing trace on the edge Em , and the cell-based functions
are linearly independent, we obtain the linear independence of the whole set of shape functions listed in the table above. The first argument also implies H(curl)-conformity, since the
tangential trace on an edge depends only on shape functions corresponding to the edge.
For uniform polynomial order p, all shape functions belong to (P p (T ))2 by construction. Due
to the linear-independence, a simple counting argument again yields Vp = P p (T ):
P
|Vp (T )| = |V N0 (T )| + 3i=1 |VpEm (T )|
+ |VpC (T )|
= 3 + 3p + (p 1)p + p 1 = (p + 1)(p + 2),
and
p
2
|P (T ) | = (p + 2)(p + 1)

Remark 5.9.
1. Owing to relation (A.7) the gradient of scaled (Integrated) Legendre polynomials can be expressed as
S
ui (x, y) = 2i+1
(2 1 , 1 + 2 ) ex

S
i+1
(2 1 , 1 + 2 ) + iS (2 1 , 1 + 2 ) (1 + 2 )


23 j (23 1) + j (23 1) ey .

vj (x, y) =

ey ,

2. The suggested construction principle for H(curl)-conforming shape functions does not
depend on the special choice of the type of the chosen orthogonal polynomials. We only
need the tensor-product-based construction of the H 1 -functions. For example, using
Jacobi-type shape functions yields C
(i,j) = ui (1 2 , 1 + 2 ) v(i,j) (3 ). Here, we choose
the shape functions of type 3 as C,3
(0,j) = (1 2 1 2 )v(0,j) .

In two dimensions, the H(div)-conforming shape functions are obtained by the rotation by
90 degrees of the H(curl)-conforming ones:
Hierarchical high-order triangular H(div)-conforming element
of variable polynomial order p = ({pE }, pC )
Edge-based shape functions
for m = 1, 2, 3: edge E m with vertices {e1 , e2 }
N0
m
= Curl e1 e2 e1 Curl e2
S
(e1 e2 , e2 + e1 )),
iEm = Curl(Li+2
Cell-based shape functions
1
C
i,j

Curl ui vj + ui Curl vj ,

2
C
i,j
3
C
j

Curl ui vj ui Curl vj ,

with ui :=

(Curl 1 2 1 Curl 2 )vj ,

S
Li+2
(2

for 0 i pEm 1
for 0 i + j pC 2

for 0 i + j pC 2

for 0 j pC 2

1 , 1 + 2 ) and vj := 3 j (23 1).

82

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

5.2.4

The hexahedral element

We define the reference hexahedron as H := [0, 1]3 . Similarly to the quadrilateral element, we
will state the shape functions in terms of the trilinear functions i , equal one at the vertex i
and zero at all other vertices. We also introduce additional linear functionals i associated
with the vertex Vi :
1
2
3
4
5
6
7
8

z
8

(1,1,1)

(1,0,0)
6

7
(1,1,1)

(1,0,1)
1

4
(0,0,0)

y
(0,1,0)

2
(0,0,1)

(1,1,0)

= (1 x)(1 y)(1 z),

= x(1 y)(1 z),
= xy(1 z),
= (1 x)y(1 z),
= (1 x)(1 y)z,
= x(1 y)z,
= xyz,
= (1 x)yz,

1
2
3
4
5
6
7
8

= (1 x) + (1 y) + (1 z),
= x + (1 y) + (1 z),
= x + y + (1 z),
= (1 x) + y + (1 z),
= (1 x) + (1 y) + z,
= x + (1 y) + z,
= x + y + z,
= (1 x) + y + z.

These vertex-based functions allow us to construct edge-based and face-based shape functions
in a unified manner.
Each edge E = [e1 , e2 ] can be parameterized by
E := (e2 e1 ) over [1, 1].

The tangential vector associated with the edge E is then given by E = 12 (e2 e1 ).
The edge extension parameter E := e1 + e2 is one on the edge E and zero on all
other edges parallel to E.
Each face F = [f1 , f2 , f3 , f4 ], where the vertices f3 and f1 are not connected by an edge,
can be parameterized via




F , F := f1 f2 , f1 f4 over [1, 1] [1, 1].
The linear face extension parameter F = f1 + f2 + f3 + f4 is equal to 1 on the
corresponding face F and zero on the opposite face. The outer normal vector of F can
be computed via nF = F .

Remark 5.10. The auxiliary functions corresponding to the vertices on an edge or a face,
provide all necessary information for constructing the corresponding edge-based or face-based
shape functions. We only have to state them for one edge and one face, and then apply the
same construction for all edges and faces. This can also be utilized in a numerical implementation. Since the same strategy is used for tetrahedra and prisms, we can easily ensure
conformity on hybrid meshes.
We denote the tensor-product polynomial space on the hexahedron H by



Qp1 ,p2 ,p3 (H) := q1 (x) q2 (y) q3 (z)  qi P pi ([0, 1]), i = 1, 2, 3 ,
p

p,p,p

Q (H) := Q

(H).

(5.25)
(5.26)

For a uniform polynomial order p we try to establish a local exact sequence

p+1,p,p

Q
(H)
Qp,p+1,p+1 (H)
div
0
curl
R Qp+1,p+1,p+1 (H) Qp+1,p,p+1 (H) Qp,p+1,p (H) Qp,p,p (H) {0}.
Qp,p,p+1 (H)
Qp+1,p+1,p (H)
(5.27)
id

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

83

H 1 -conforming hexahedral element

We aim to construct an H 1 -conforming hierarchical V -E-F -C basis, which in case of uniform
polynomial order p spans Wp = Qp (H). Extending the quadrilateral shape functions by linear
extension into the interior of the hexahedron H and adding tensor-product shape functions
spanning Qp0C (H) yields the following collection:
Hierarchical hexahedral H 1 -element of variable order p = (pEm , pFm , pC )
Vertex-based functions

for i = 1, ..., 8 :

Vi = i

Edge-based functions
for m = 1, . . . , 12 : Suppose the edge Em = [e1 , e2 ].
m
E
= Li+2 (e1 e2 ) (e1 + e2 )
i

for 0 i pEm 2

Face-based functions
for m = 1, ...6: Suppose the face Fm = [f1 , f2 , f3 , f4 ].

m
F
for 0 i, j pFm 2
(i,j) = Li+2 (F ) Lj+2 (F ) F
P4
with F := =0 f and (F , F ) := (f1 f2 , f1 f4 )

Cell-based functions
for 0 i, j, k pC 2

C
(i,j,k) = Li+2 (2x 1)Lj+2 (2y 1)Lk+2 (2z 1)

We define the lowest-order space as W V (H) := span (Vi )1i8 and denote the span of
p
edge-based (associated with Em ) by WEEmm (H), the span of face-based (associated with Fm )
p
pFm
(H), and the span of cell-based functions by WCFm (H). The local FE-space
functions by WFm
spanned by the whole set of shape functions on H up to order p = ({pEm }, {pFm }, pC ) is given
by
12
6
M
M
p
pFm
Wp(H) = W V (H)
WEEmm (P)
(H) WCpC (H).
W Fm
m=1

m=1

Theorem 5.11. The shape functions stated in the table above are linearly independent and
H 1 -conforming. For uniform polynomial order p = pEm = pFk = pC , they form a basis of
Qp (H).
Proof. The vertex functions are in Q1 (H) and linearly independent, since Vi (Vj ) = i,j .
The edge-extension (e1 +e2 ) is independent of E = e2 e1 , and hence the edge-based shape
p
m
QpEm (H), and their traces on the corresponding edge Em span P0 Em (Em )
functions E
i
while being zero on all other edges.
The face-extension F is constant on planes parallel to the face Fm , hence the face based
functions are in QpFm (H). Obviously, their traces on Fm span P pFm (Fm ), while being zero on
all other faces.
Finally, the cell-based functions, defined using the tensor products, span Qp0c (H).
To show linear independence and H 1 -conformity, we assume that the trivial function 0 = w
Wp can be assembled with the basis functions, i.e.
X X
X
X X
X
Fm Fm
C
m E
+
cC

+
cE
w=
cVi Vi +
m
ijk (i,j,k) = 0.
ij
i
(i,j)
i

Em EH

Fm FH i,j

i,j,k

Restricting w successively to the vertices, then to the edges, then to faces we obtain cVi = 0,
1
m
= 0, cFijm = 0 and finally cC
cE
i
ijk = 0 for all i, j, k. The H -conformity follows with similar

84

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

arguments as in the quadrilateral case by the special construction, since, e.g. on a face the
only non-zero shape functions are the ones associated with the four involved vertices, the
four involved edges and the face itself.
In case of uniform polynomial order p a counting argument shows
P
P
|Wp(H)| = |W V | + 1m=1 2|WpEm | + 6m=1 |WpFm | + |WpC |
= 8 + 12(p 1) + 6(p 1)2 + (p 1)3 = (p + 1)3 ,
p
|Q (H)| = (p + 1)3 ,

and

and hence the shape functions span the whole Qp (H).

H(curl)-conforming shape functions for hexahedral elements
Since we want to ensure that Wp+1 Vp, we choose for uniform polynomial degree p the
finite element space corresponding to the following part of the sequence (5.27)

Qp+1 (H) Qp,p+1,p+1 (H) Qp+1,p,p+1 (H) Qp+1,p+1,p (H).

We choose lowest-order, higher-order edge-based and face-based shape functions in analogy
to the quadrilateral case. The extension onto P
the hexahedron requires linear extension by
multiplication with E := e1 + e2 and F := 4i=1 fi , and consideration of the edge- and
face-orientations similar to the H 1 -conforming case.
The construction of cell-based shape functions again involves several types: as first class we
choose the gradient functions of the H 1 -conforming shape functions:


C,
C,1
(i,j,k) = (i,j,k) = Li+2 (2x 1)Lj+2 (2y 1)Lk+2 (2z 1)
= 2i+1 (2x 1)Lj+2 (2y 1)Lk+2 (2z 1)ex
+2Li+2 (2x 1)j+1 (2y 1)Lk+2 (2z 1)ey
+2Li+2 (2x 1)Lj+2 (2y 1)k+1 (2z 1)ez ,

for 0 i, j, k pC 2.
The three terms in the representation above form a set of 3(p 1) linearly independent
functions with vanishing tangential trace on H, since for any face either ex , ey or ez is
normal to the considered face or one of the integrated Legendre polynomials is evaluated at
1.
Hence, we can choose arbitrary linearly independent combinations of the three summands
to obtain linearly independent cell-based shape functions. In particular, we set C,2
(i,j,k) =
C,1
C,2
1
1 1
diag( 12 , 21 , 12 )C,1
(i,j,k) and pC +(i,j,k) = diag( 2 , 2 , 2 )(i,j,k) for 0 i, j, k pC 2.
Finally, the following 3(pC 1)2 functions are linearly independent of any of the functions
above:
C,3
(i,j,k) = Lj+2 (2y 1) Lk+2 (2z 1) ex ,

C,3
(i,j,k) = Li+2 (2x 1) Lk+2 (2z 1) ey ,

C,3
p +(i,j,k) = Li+2 (2x 1)Lj+2 (2y 1) ez .
C

The vanishing of the tangential components on all faces of the hexahedron is shown by the
same argument as given above for the cell-based functions of the second type. Summarizing,
we obtain

85

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

Hierarchical H(curl)-conforming shape functions on the hexahedron H

with variable polynomial order p = ({pEm }, {pFm }, pC )
Edge-based shape functions
for edge Em , m = 1, ..., 12 with local edge-ordering Em = {e1 , e2 }
Lowest-order N0 -function

1
0
N
m = 2 (e1 e2 ) (e1 + e2 )

Higher order edge-based functions (gradient fields)

for 0 i pEm 1


m
E
= Li+2 (e1 e2 )(e1 + e2 )
i
Face-based functions
for faces Fm , m = 1, ...6 with local face-vertex ordering Fm = {f1 , f2 , f3 , f4 }
P4
We define F := =0 f and (F , F ) := (f1 f2 , f1 f4 ).
for 0 i, j pFm 1

Type 1 (gradient fields):

Fm ,1
(i,j)
= Li+2 (F ) Lj+2 (F ) F

Type 2:

Fm ,2
(i,j)

Type 3:

Fm ,3
(0,j)

= Lj+2 (F ) F F

Fm ,3
(i,0)

Li+2 (F ) Lj+2 (F ) F Li+2 (F ) Lj+2 (F ) F

= Li+2 (F ) F F

Cell-based functions
for 0 i, j, k pC 1

Type 1 (gradient fields)

C,1
= Li+2 (2x 1) Lj+2 (2y 1) Lk+2 (2z 1)
ijk

Type 2:

C,2
(i,j,k)

C,2
p +(i,j,k)
C

Type 3:

C,3
(0,j,k)
C,3
(i,0,k)
C,3
(i,j,0)

= diag(1, 1, 1) C,1
(i,j,k)

= diag(1, 1, 1) C,1
(i,j,k)
= Lj+2 (2y 1) Lk+2 (2z 1) ex

= Li+2 (2x 1)Lk+2 (2z 1) ey

= Li+2 (2x 1)Lj+2 (2y 1) ez

0
We define the lowest-order space as V N0 (H) := span (N
m )1m12 , and denote the span of
pEm
(H, the span of face-based functions associated with
edge-based associated with Em by VEm
p
pFm
the face Fm by VFm (H), and the span of cell-based functions by VC Fm (H). The local FEspace spanned by the whole set of shape functions on H up to order p = ({pEm }, {pFm }, pC )
is denoted by
Vp(H) := VN0 (H)

12
M

m=1

Em
(H)
VEm

6
M

m=1

VFmFm (H) VCpC (H).

86

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Theorem 5.12. The shape functions listed in the table above are H(curl)-conforming and
linearly independent. Moreover, there holds
(H) 1 m 12
W V (H) V N0 (H), WpEEm +1 (H) = VpEEm
m
m

WpFFm +1 (H) VpFFmm (H) 1 m 6, and WpCC +1 (H) VpCC (H)

m

For uniform polynomial order p = pEm = pFm = pC we obtain an H(curl)-conforming basis

spanning
Vp (H) = Qp,p+1,p+1 (H) Qp+1,p,p+1 (H) Qp+1,p+1,p (H).
Proof. The linear independence of the edge-based and face-based shape functions can be
deduced from the linear independence of the quadrilateral case. The linear independence of
cell based functions follows from their definition.
Suppose that 0 = v Wp(H) has a representation
v=

0
c0m N
m +

pEm 1

Em EH

Em
m
+
cE
i i

Fm
X nX

cFi m Fi m +

Fm FH i=1

i=0

nC
X

C
cC
i i ,

i=1

C
where span{Fi m } = VpFm (H) and span{C
i } = Vp (H).
To prove linear independence of the whole set of stated shape functions we set v = 0. Rem
= 0, i, since the edge-based shape
striction to edges yields tr,Em (v) = 0. This implies cE
i
functions are the only ones with
non-vanishing
tangential
trace on Em . By restriction to the
PnFm Fm Fm
faces we obtain tr,Fm (v) = i=1 ci i = 0, and due to the linear independence of facebased functions cFi m = 0, since the remaining cell-based part consists of linearly independent
functions, we obtain cC
i = 0, i.
k
= cC
H(curl)-conformity: If all dofs associated with a face Fm vanish, i.e. cFi m = cE
i
k = 0
for Ek on Fm , then due to the N0 -E-F -C-based construction tr,Fm (v) = 0. An analogue
statement holds for edges.

Finally, in case of uniform polynomial order p, all shape functions belong to Qp,p+1,p+1 (H)
Qp+1,p,p+1 (H) Qp+1,p+1,p (H). Due to their linear independence the following counting argument concludes the proof:
|Vp (H)| = |V N0 | +

P1

Em
m=1 2|Vp |

= 12 + 12p + 12p(p + 1) +

P6

Fm
C
m=1 |Vp | + |Vp |
3p(p + 1)2 = 3(p + 2)2 (p

+ 1),

and

|Qp,p+1,p+1 Qp+1,p,p+1 Qp+1,p+1,p | = 3(p + 2)2 (p + 1).

H(div)-conforming shape functions for hexahedral elements

In view of the sequence (5.27), we intend to construct a basis for local H(div)-conforming
FE-space in such a way that, for uniform polynomial order p, it spans
Qp = Qp+1,p,p (H) Qp,p+1,p (H) Qp,p,p+1 (H).

87

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

The lowest-order Raviart-Thomas function corresponding to the quadrilateral face Fm =

[f1 , f2 , f3 , f4 ] are given by
0
RT
= nFm F = F F .
m

Since the linear face extension F := f1 + f2 + f3 + f4 is independent of (F , F ), there

0
0
holds RT
Q1,0,0 Q0,1,0 Q0,0,1 (H), and additionally trFk ( RT
nFk ) = m,k .
m
m
The face-based shape functions can be chosen as curls of face-based H(curl)-based functions: Since trn,Fm (curl()) = n curl((tr,Fm ()) = curlFm ((tr,Fm ()(F , F )). Hence,
if q i (F , F ) = tr,Fm (i ) span Qp,p+1 /R Qp+1,p /R then curl(q i ) span Qp,p /R. The curl of
the gradient functions is zero, i.e. there remains (p + 1)2 1 linearly independent curl-fields,
whose normal traces span Qp,p (Fm )/R on the corresponding face Fm .
The first choice of cell-based shape functions are the curl-fields of cell-based H(curl)conforming shape functions of order pC . The curl-fields of type 2 and type 3 H 1 -conforming
shape functions, are constructed by linear combination of the following functions:
Li+2 (2x 1) j+1 (2y 1) k+1 (2z 1) ex ,
i+1 (2x 1) Lj+2 (2y 1) k+1 (2z 1) ey ,
i+1 (2x 1) j+1 (2y 1) Lk+2 (2z 1) ez ,
and
Li+2 (2x 1) j+1 (2y 1) ex ,
Li+2 (2x 1) k+1 (2z 1) ex ,

i+1 (2x 1) Lj+2 (2y 1) ey ,

Lj+2 (2y 1) k+1 (2z 1) ey ,

i+1 (2x 1) Lk+2 (2z 1) ez ,

j+2 (2y 1) Lk+2 (2z 1) ez ,

for 0 i, j, k pC 1. All these functions are linearly independent and are in Q(pC +1,pC ,pC )
Q(pC ,pC +1,pC ) Q(pC ,pC ,pC +1) (H). Moreover, their normal traces vanish on H, since on a face
F either ei nF or the integrated Legendre polynomial is evaluated at 1. Hence, we can
choose a set of cell-based functions involving 2p2C +3pC curl-fields and further p2C +3pC linearly
independent functions which are linearly independent from the curl-fields. Finally, the set can
be further extended by functions
Li+2 (2x 1)ex ,

Lj+2 (2y 1) ey ,

Lk+2 (2z 1) ez .

Summarizing, we obtain
Hierarchical H(div)-conforming shape functions on hexahedra
with variable polynomial order p = ({pFm }, pC )
Face-based functions
for faces Fm , m = 1, ...6 with local face-vertex ordering Fm = [f1 , f2 , f3 , f4 ]
P4
We define F := =0 f and (F , F ) := (f1 f2 , f1 f4 ).
Lowest-order Raviart-Thomas RT 0 function
0
RT
= F F
m

Higher-order face-based functions (divergence-free)

for 0 i, j pFm 1
m
F
(i,j)

m
F
p

Fm +(0,j)

m
F
pFm +(i,0)

Fm ,2
curl (i,j)

curl

Fm ,3
curl (0,j)

curl

Fm ,3
curl (i,0)

curl

Li+2 (F ) Lj+2 (F ) Li+2 (F ) Lj+2 (F )

Lj+2 (F ) F F


Li+2 (F ) F F

88

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Cell-based functions
for 0 i, j, k pC 1

(, , ) := (2x 1, 2y 1, 2z 1).

Type 1: (divergence-free)
C,1
(i,j,k)
C,1
pC +(i,j,k)
C,1
2pC +(0,j,k)
C,1
2pe l+(i,0,k)
C,1
2pC +(i,j,0)

curl C,2
(i,j,k)

4Li+2 () k () k () ex 4i ()j () Lk+2 () ez

curl C,2
pC +(i,j,k)
curl C,3
(0,j,k)
curl C,3
(i,0,k)
curl C,3
(i,j,0)

4i () Lj+2 ()k () ey 4i () j ()Lk+2 () ez

2Lj+2 () k+1 () ey 2j+1 () Lk+2 () ez

2i+1 () Lk+2 () ez 2Li+2 () k+1 () ex

2Li+2 () j+1 () ex 2i+1 () Lj+2 () ey

=
=
=

Type 2:
C,2
(i,j,k)

= Li+2 () j () k () ex + i ()Lj+2 () k () ey

C,2
p +(0,j,k)

= Lj+2 () k+1 () ey + j+1 () Lk+2 () ez

C,2
pC +(i,0,k)
C,2
pC +(i,j,0)

= i+1 () Lk+2 () ez + Li+2 () k+1 () ex

= Li+2 () j+1 () ex + i+1 () Lj+2 () ey

Type 3:
C,3
(i,0,0)

= Li+2 () ex

C,3
(0,j,0)
C,3
(0,0,k)

= Lj+2 () ey
= Lk+2 () ez

0
We define QRT 0 (H) := span{( RT
: 1 m 6)} and denote the span of face-based
m
(associated with Fm ) and the span of cell-based shape functions by QFpFmm (H) and QC
pC (H)
respectively, and define
L
Qp(H) := QRT 0 (H) 6m=1 QFpFmm (H) QC
pC (H).

Theorem 5.13. The shape functions collected in the above table are linearly independent and
H(div)-conforming. Moreover,
curl VpFFmm (H) QFpFmm (H), m = 1, . . . , 6, and curl VpCC (H) QC
pC (H).
For uniform polynomial order p we obtain an H(div)-conforming basis spanning
Qp (H) = Qp+1,p,p (H) Qp,p+1,p (H) Qp,p,p+1 (H).

Proof. The linear independence of face-based shape functions is implied by their choice as
curl-fields of a set of linearly independent functions. Moreover, we already discussed the
linear independence of the set of cell-based shape functions. Since the normal traces of the
p2Fm face-based shape functions (including the lowest-order Raviart-Thomas functions) span
QpFm QpFm (Fm ), while the cell-based functions have zero tangential trace, we obtain linear
independence of all shape functions.
H(div)-conformity is implied by the RT 0 -F -C-based construction: all shape functions not
associated to the face Fm have zero tangential trace on the face. Hence, the normal trace on
Fm is only affected by the degrees of freedom associated with the face.

89

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

For uniform order p, all shape functions belong to Qp+1,p,p (H) Qp,p+1,p (H) Qp,p,p+1 (H) by
construction. A comparison of the dimensions of the spaces under considerations completes
the proof:
|Qp (H)| = |QRT (H)| + 6 |QFp m (H)| + |QC
p (H)|
2
3
2
= 6 + 6(p + 2p) + 3p + 6p + 3p
= 3(p + 2)(p + 1)2
Qp+1,p,p (H) Qp,p+1,p (H) Qp,p,p+1 (H) = 3(p + 2)(p + 1)2

5.2.5

The prismatic element

We define the reference prism as the tensor product

P := T [0, 1] = {(x, y, z) : 0 x + y 1, 0 x, y, z 1}.
In this section, we provide a general formulation of shape functions corresponding to this
tensor-product representation. We introduce the barycentric coordinates i corresponding to
the triangle T and z corresponding to the segment Iz := [0, 1] as follows:
z
(1,0,0)
4

(1,1,1)
6

(1,0,1)
5

1
(0,0,0)

2
(0,0,1)

3
(0,1,0)

1
2
3
4
5
6

= 1 x y,
= x,
= y,
= 1 x y,
= x,
= y,

1
2
3
4
5
6

= 1 z,
= 1 z,
= 1 z,
= z,
= z,
= z.

The parametrization of horizonal edges and the corresponding tangential vectors are computed
as for the tetrahedral element, whereas the parametrization of the vertical edges and the
computation of the associated tangential vector follows the construction for the hexahedral
element. Thus, the normal component on a quadrilateral face [f1 , f2 , f3 , f4 ] can be expressed
by
1
nF = F
(5.28)
2
with F = f1 + f2 + f3 + f4 , and on a triangular face [f1 , f2 , f3 ] by
nF = f1 .

(5.29)

Corresponding to the tensor-product domain P = T [0, 1], we introduce the polynomial

space,


Rp1 ,p2 (P) := q1 (x, y) q2 (z) : q1 P p1 (T ), q2 Qp2 ([0, 1]) .
The sequence of polynomial spaces corresponding to the exact sequence (3.14) on the continuous level we intend to construct, is given by

Rp,p (P)
Rp,p+1 (P)
div
0
curl

Rp,p (P)
Rp1,p (P) {0}.
R Rp+1,p+1 (P) Rp,p+1 (P)
Rp1,p+1 (P)
Rp+1,p (P)
id

(5.30)

90

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

H 1 -conforming shape functions on the prismatic element P

As for the previous topologies, we construct the vertex-, edge- and face-based shape functions
such that the traces on quadrilateral or triangular faces, respectively, coincide with the cellbased functions of the quadrilateral Q or the triangle T , respectively. This approach will in
particular allow for hybrid meshes and for a locally varying polynomial degree.
On the quadrilateral faces we use linear extension (in terms of i and i ) of edge-based shape
functions, and on triangular faces we agree on a monomial edge-extension provided by scaled
(integrated) Legendre polynomials (in terms of i ). The construction of cell-based shape
functions is then simple, by considering the tensor-product shape of the prismatic element.
Since the cell based shape functions shall span


p
p
WpCC := q1 (x, y) q2 (z) : q1 (x, y) P0 C,1 (T ), q2 (z) Q0C,2 ([0, 1]) ,

with pC := (pC,1 , pC,2 ), we construct them as products of cell-based triangular shape functions
and cell-based shape functions of the segment Iz = [0, 1]:
C
C
C
C
ijk (x, y, z) = ij (x, y) Lk+2 (z) with ij (x, y) WpC (T ), 0 k pC2 2.

Putting the various sets of shape functions together we get the following representation:
Hierarchical H 1 -conforming shape functions for prismatic elements
of variable order p = (pEm , pFm , pC )
Vertex-based functions:
Vi = i i

for i = 1, . . . , 6 :
Edge-based functions

for m = 1, . . . 6: horizontal edge Em = [e1 , e2 ]

for 0 i pEm 2 :

S
m
E
= Li+2
(e2 e1 , e1 + e2 )e1
i

for m = 7, 8, 9 : vertical edge Em = [e1 , e2 ]

for 0 i pEm 2 :

m
E
= Li+2 (2e1 1)(e1 + e2 )
i

Face-based functions
for m = 1, 2 : triangular face Fm = [f1 , f2 , f3 ]
for 0 i + j pFm 3 :

S
m
F
(i,j) = Li+2 (f1 f2 , f1 + f2 ) f3 j (2f3 1) f1

for m = 3, 4, 5: quadrilateral face Fm = [f1 , f2 , f3 , f4 ]

with horizontal edge
for 0 i, j pFm 2 :

[f1 , f2 ],

i.e.

f2

f2
f4

if f1 = f2
else

S
m

F
(i,j) = Li+2 (f2 f1 , f1 + f2 ) Lj+2 (2f1 1)

Cell-based functions
for 0 i + j pC 3, 0 k pC 2 :

S
C
(i,j,k) = Li+2 (1 2 , 1 + 2 ) 3 j (23 1) Lk+2 (21 1)

91

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

We define the lowest-order space as W V (P) := span (Vi )1i6 and denote the span of
p
edge-based (associated with Em ) by WEEmm (P), the span of face-based (associated with Fm )
p
pFm
(P), and the span of cell-based functions by WCFm (P). The local FE-space
functions by WFm
spanned by the whole set of shape functions on P up to order p = ({pEm }, {pFm }, pC ) is given
by
9
5
M
M
p
pFm
Wp(P) = W V (P)
WEEmm (P)
(P) WCpC (P).
W Fm
m=1

m=1

Theorem 5.14. The stated shape functions are H 1 -conforming and linearly independent. For
uniform polynomial order p the V -E-F -C-based set of shape functions spans the local FE-space
Wp (P) = Rp,p (P)
and in the case of the minimum rule, i.e. pEm pFm pC :


p
Wpmin (P) = w P pC (P)  trEm (w) PmEm (Em ) 1 m 9,
trFm (w) P pFm (Fm ) 1 m 2,

trFm (w) QpFm ,pFm (Fm ) 3 m 5

Proof. Vertex-, edge-, face-based shape functions are obtained by linear and monomial extension of H 1 -conforming shape functions on the quadrilateral or the triangular element. The
trace of the various functions on the associated face remains unchanged under the extension.
Hence, linear independence and the full span of the trace on the associated edges and faces
are implied by the considerations for the quadrilateral and triangular elements.
Concerning the cell-based shape functions, the construction through tensor-product of cellbased triangular shape functions (having zero tangential trace on all quadrilateral faces) with
bubble functions (having zero tangential traces on the remaining triangular faces) in vertical
z-direction implies two properties, namely zero tangential trace on P, as well as linear independence within the set of cell-based shape functions.
By the V -E-F -C construction we can easily deduce the linear independence of the whole set
of stated shape functions as well as the H 1 -conformity in a similar manner as for the hexahedral element.
Finally, for uniform polynomial order p the counting argument
P
P
|Wp (P)| = |W V (P)| + 9|WpEm (P)| + 2m=1 |WpFm (P)| + 5m=3 |WpFm (P)| + |WpC (P)|
= 6 + 9(p 1) + (p 2)(p 1) + 3(p 1)2 + 12 (p 2)(p 1)2
= 21 (p + 1)2 (p + 2),
and
|Rp,p (P)| = |P p (T )| |Qp ([0, 1])| = 21 (p + 1)2 (p + 2)
completes the proof.
H(curl)-conforming shape functions on the prismatic element P
In view of the exact sequence property, we are interested in establishing the following inclusion:

Wp+1 (P) = Rp+1,p+1 (P) Rp,p+1 (P) Rp+1,p (P) Rp+1,p := Vp (P).
Concerning lowest-order shape functions we have to distinguish between horizontal and ver0
tical edges here. The aim is to ensure tr,Ek (N
Em ) = m,k , and the extension onto the faces

92

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

should coincide with the extension of the lowest-order shape functions on triangular faces of
tetrahedra and quadrilateral faces of hexahedra. On horizontal edges Em = [e1 , e2 ] this is
provided by
0
N
m = (e1 e2 e1 e2 ) e1

whereas on vertical edges Em = [e1 , e2 ] we choose

0
N
m = e1 e1 .

The edge-based shape functions up to order pEm can be chosen as gradients of H 1 -conforming
edge-based shape functions of order pEm + 1.
The face-based and cell-based shape functions can be defined similarly as in three types of
shape functions above.
The following table summarizes H(curl)-conforming shape functions of variable polynomial
order:
H(curl)-conforming shape functions for prismatic elements
of variable polynomial order p = ({pEm }, pFm , pC )
Lowest-order N
ed
elec functions
for m=1,...,6: horizontal edge Em = [e1 , e2 ]
for m=7,8,9: vertical edge Em = [e1 , e2 ]

0
N
m

0
N
m

= e1 e1

Higher-order edge-based functions

for m = 1, ..., 6 : horizontal edge Em = [e1 , e2 ]
m
E
for 0 i pEm 1:
i
for m = 7, 8, 9 : vertical edge Em = [e1 , e2 ]
for 0 i pEm 1:

(e1 e2 e1 e2 ) e1

S
= Li+2
(e2 e1 , e1 + e2 ) e1

m
E
i

= Li+2 (2E 1) e1

Face-based functions
for m = 1, 2 : triangular face Fm = [f1 , f2 , f3 ]
We define ui
vj

S
:= Li+2
(f2 f1 , f1 + f2 )
:= f3 j (2f3 1)

for 0 i + j pFm 2:

Type 1: (gradient fields)

Fm ,1
(i,j)

Type 2:

Fm ,2
(i,j)

Type 3:

Fm ,3
(0,j)

s.t.

= ui vj f1

m
,F
(i,j) = ui vj f1 .

ui vj ui vj f1

f1 f2 f1 f2 vj f1

for m = 3, 4, 5: quadrilateral face Fm = [f1 , f2 , f3 , f4 ] with horizontal edge [f1 , f2 ],



1
f2 = f2
f2 if f1 = f2

.
, and =
i.e. f2 =
1 else
f4 else
We define ui
wj

S
:= Li+2
(f2 f1 , f1 + f2 )
:= Lj+2 (2f1 1)

,Fm
= ui wj .
s.t. (i,j)

for 0 i, j pFm 1:

Fm ,1
Type 1:(gradient fields) (i,j)

Type 2:
Type 3:

Fm ,2
(i,j)
Fm ,3
(0,j)

= (ui wj )

= ui wj ui wj .

(f1 f2 f1 f2 ) wj

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

93

Cell-based functions
We define ui
vj
wk

S
:= Li+2
(2 1 , 1 + 2 )
:= 3 j (23 1)
:= Lk+2 (21 1)

for 0 i + j pC 2, 0 k pC 1 :
Type 1:(gradient fields) C,1
(i,j,k)
C,2
(i,j,k)

Type 2:

C,2
p +(i,j,k)
C

C,3
(i,j,0)
C,3
(0,j,k)

Type 3:

= ui vj wk
= ui vj wk

= ui vj wk
= ui vj ez
=

1 2 1 2 vj wk

0
We define the lowest-order space as V N0 (P) := span (N
m )1m9 , and denote the span of
pEm
(P), the span of face-based functions associated with
edge-based associated with Em by VEm
p
pFm
the face Fm by VFm (P), and the span of cell-based functions by VC Fm (P). The local FEspace spanned by the whole set of shape functions on P, up to order p = ({pEm }, {pFm }, pC )
is denoted by
Vp(P) := VN0 (P)

9
M

m=1

Em
(P)
VEm

5
M

m=1

VFmFm (P) VCpC (P).

Theorem 5.15. The basis functions defined above are linearly independent and H(curl)conforming. There holds
WpEEm

+1 (P)

VpEEm (P) for m = 1, . . . , 9, WpFFm

m

and WpCC +1 (P)

+1 (P)
VpCC (P).

VpFFm (P) for m = 1, . . . , 5,

m

Moreover, for uniform polynomial order p the set of shape functions span
Vp (P) = Rp,p+1 (P) Rp,p+1 (P) Rp+1,p (P).
Proof. We already verified the linear independence of the set of edge-based shape functions
as well as of the set of face-based shape functions. It remains to show linear independence
of the cell-based shape functions. Due to the tensor-product-based construction this can be
deduced from the linear independence of the H 1 -conforming and H(curl)-conforming shape
functions for the triangular element as follows.
First, we consider the functions in M1 = {ui vj wk , ui vj wk wk + (1 2 1 2 )wk },
which are built as products of H(curl)-conforming cell-based triangular shape functions
Tij (x, y) of Theorem 5.6 and linearly independent functions wk (z). Hence, we obtain linear independence of the functions in M1 .
Next consider the set M2 = {ui vj wk = ui vj wk ez , ui vj ez }, which again consists of linearly
independent functions due to the tensor-product construction, and the linear independence
of ui vj , cf. Theorem 5.6. Since ez = 0 for M1 , we obtain linear independence of the
functions
0
{ui vj wk , ui vj wk , ui vj wk , N
[1,2] (x, y) wk , ui vj ez , }

which have vanishing tangential trace on P . This implies the linear independence of the
complete set of cell-based shape functions.

94

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Thanks to the N0 -E-F -C-based construction, linear independence of all shape functions as
well as the H(curl)-conformity follows. Moreover, for uniform polynomial order we conclude
again with a simple counting argument:
|Vp (P)| = 9(p + 1) + 2(p + 1)(p 1) + 3(2p2 + 2p)
+ 23 (p + 1)p(p 1)
and
= 21 (p + 1)(p + 2)(3p + 7),


1
p,p+1
p,p+1
p+1,p
2
| R
R
R
(P)| = (p + 1)(p + 2) + 2 (p + 1)(p + 2)(p + 3)
= 12 (p + 1)(p + 2)(3p + 7).

Remark 5.16. In the case of the minimum order rule, i.e., pEm pFm pC , the shape
functions span the following local FE-space


Vpmin (P) = v X pC (P)  tr,Em (v) P pEm (Em )
tr,Fm (v)

tr,Fm (v)

(P pFm (F

))2

m
,p
p
F
F
Q m m +1

with X p (P) := Rp,p+1 (P) Rp,p+1 (P) Rp+1,p (P).

1 m 9,

1 m 2,



QpFm +1,pFm (Fm ) 3 m 5 ,

H(div)-conforming shape functions for the prismatic element P

Similarly to the previous cases, the lowest order Raviart-Thomas shape functions can be
chosen as lowest order shape functions by linear extension of the element normal vector into
the interior of the element:
0
RT
= f1 n = f1 f1 for triangular faces and
m

0
RT
= 21 F n = 41 F F for quadrilateral faces,
m

due to the representation of the normal vector stated in (5.28) and (5.29).
The higher-order face-based shape functions can be chosen as the curl-fields of the set of
H(curl)-conforming face-based shape functions of VpFFm +1 (P) which are linearly independent
m
from gradient functions. Since c1 curlF (q 1 ) + c2 curlF (q 2 ) = 0 implies s.t. c1 q 1 + c2 q 2 +
= 0, the linear independence of the proposed curl-field shape functions follows directly
from the linear independence of the H(curl)-conforming shape functions, and their linear
independence from the gradient fields. Moreover, the normal trace of the constructed functions
span QpFm (Fm )/R on quadrilateral faces, and P pFm (Fm )/R on triangular faces, respectively,
which can be seen by a comparison of the dimension of the involved spaces.
The first set of cell-based functions is provided by the curl-fields of H(curl)-conforming shape
functions. Considering the terms involved by these curl-fields in more detail, we can extract
the following linearly independent cell-based shape functions, which are already extended to
span the full set of cell-based functions.
S ( , + ), v = v (y) := (2 1),
Lemma 5.17. Let ui = ui (x, y) := Li+2
2
1 1
2
j
j
3 j
3
wk = wk (z) := Lk+2 (21 1) and N0 (x, y) := 1 2 2 1 , and let us define the following

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

95

sets of polynomials:


M1 := ui vj wk , curl(N0 (1 , 2 )vj )wk , ez wk 0i+jp,0i,j,kp

0
= {curl(x,y) (ui (x,y) vj ) wk ez , curl(x,y) (N
[1,2] (x, y) vj ) wk ez , wk ez
0
with N
[1,2] (x, y) := (x,y) 1 2 (x,y) 1 2 ,


M2 := wk ui vj , N0 (1 , 2 ) wk vj , ez ui vj , N0 (1 , 2 ) ez vj 0i+jp2,kp1 ,


M3 := wk vj ui , vj ez ui 0i+jp2,kp1

= wk (z) vj (y) ui ex , vj (y)ex },
The union of M1 M2 M3 defines a set of linearly independent vector-fields with vanishing
normal trace on P. Moreover, there holds

3
P p (P) H0 (div, P) = span M1 span M2 span M3 .

3
Proof. Due to the inclusion in P p (P) as well as the vanishing of tangential traces on P,
we only show linear independence of M1 M2 M3 .
1. First, we show linear independence of functions in M1 : Since these functions are
constructed as the products of curl-fields of H(curl)-conforming triangular-face shape
functions, whose linear independence we had proved in the course of the construction
of face-based shape functions, and a set of linearly independent functions in the
z-component, their linear independence follows directly.

2. The functions of M3 are linearly independent by construction.

3. Multiplying the vector fields of M3 with ey yields (informally written) {x ui , (1 y)}
{wk , 1} vj yielding linear independence.
4. We observe that the vector-fields from each set Mi are linearly independent, when
restricted to the i-th component. Hence, we obtain the linear independence of M1
M2 M3 .
Observing that dim(M1 )+dim(M2 )+dim(M3 ) = 21 (3p2 2)(p+1) = dim(P p0, (P)) completes
the proof.
Hierarchical H(div)-conforming shape functions for prismatic elements
of variable order p = (pFm , pC )
Lowest-order Raviart-Thomas functions
for m = 1, 2 : triangular face Fm = [f1 , f2 , f3 ]:
for m = 3, . . . , 5 : quadrilateral face Fm = [f1 , f2 , f3 ]:
P4
Set F = i=1 fi .

0
RT
m

= f1 f1

0
RT
m

= 41 F F

96

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Face-based functions (div-free)

for m = 1, 2 : triangular face Fm = [f1 , f2 , f3 ]
We define ui
vj

S
:= Li+2
(f2 f1 , f1 + f2 )
:= f3 j (2f3 1) f1

for 0 i + j pFm 2:

curl,(Fm ,2)

m
F
(i,j)

= curl (i,j)

pFm ,1+(0,j)

= curl p

= 2vj ui

= vj f1 f2 f2 f1 + 2f2 f1 vj

curl,(Fm ,3)
Fm +(0,j)

Fm

m
,F
(i,j) = ui vj .

s.t.

for m = 3, 4, 5: quadrilateral face Fm = [f1 , f2 , f3 , f4 ] with horizontal edge [f1 , f2 ],



1
f2 = f2
f2 if f1 = f2

.
, and =
i.e. f2 =
1 else
f4 else
We define ui
wk

m
s.t. ,F
(i,j) = ui vj .

S
:= Li+2
(f2 f1 , f1 + f2 )
:= Lk+2 (2f1 1)

for 0 i, k pFm 1:
curl,(F ,2)
m
F
= curl (i,k) m = 2 wk ui
(i,k)
curl,(Fm ,3)

m
F
p

= curl (0,j)

Fm +(0,j)

= curl (f1 f2 f1 f2 ) Lk+2 (2f1 1)

Cell-based functions
We define ui
vj
wk

S
(2 1 , 1 + 2 )
:= Li+2
:= 3 j (23 1)
:= Lk+2 (21 1)

for 0 i + j pC 2, 0 k pC 1
Type 1:(div-free) C,1
(i,j,k)

Type 2:

Type 3:

,C
(i,j,k) = ui vj wk .

s.t.

curl,(C,1)

curl (i,j,k)

curl,(C,2)
curl (i,j,k)
curl,(C,3)
curl (0,j,k)
curl,(C,3)
curl (i,j,0)

C,1
pC +(i,j,k)
C,1
2p +(0,j,k)
C
C,1
2pC +(i,j,0)

C,2
(i,j,k)

= ui vj wk

= wk ui vj ui vj wk
= ui vj wk vj wk ui
= 2vj wk ez N0 (1 , 2 ) (vj wk )
= (ui vj ) ez

C,2
pC +(0,j,k)
C,2
pC +(i,j,0)

= N0 (1 , 2 ) vj wk

C,3
(0,0,k)

= wk ez

C,3
(0,j,0)

= N0 (x, y) ez vj

= ui ez vj

with N0 (1 , 2 ) := (1 2 1 2 ).

0
We define the lowest-order space as QRT 0 (P) := span ( RT
m )1m5 , and denote the span
p
of face-based functions associated with Fm by QFFmm (P), and the span of cell-based functions
pFm
by QC (P). The local FE-element space is given by
Qp(P) := QRT 0 (P)

5
M

m=1

QFFmm (P) QCFm (P).

Theorem 5.18. The set of shape functions defined above is linearly independent and H(div)conforming. For uniform polynomial order p the shape functions are a basis of the local

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

97

FE-space
Qp = Rp,p (P) Rp,p (P) Rp1,p+1 (P).
Proof. The linear independence of lower- and higher-order face based functions, which span
either QpFm +1,pFm QpFm ,pFm +1 (Fm ) if Fm is a quadrilateral or P pFm (Fm ) if Fm is a triangular
face, has already been shown.
Since the cell-based shape functions are linearly independent combinations of linearly independent polynomial vector fields, cf. Lemma 5.17, also the set of cell-based functions is linearly
independent.
The RT 0 -F -C-based construction implies the linear independence of all shape functions as
well as their H(div)-conformity.
In case of uniform polynomial order p a comparison of the dimension of the involved spaces
|Qp (P)| = 5 + 2( 12 (p 1)(p + 2)) + 3(p2 + 2p) + 12 (3p2 2)(p + 1)

|(Rp,p (P))2 Rp1,p+1 (P)| = (p + 1)2 (p + 2) + 12 p(p + 1)(p + 2) = 21 (p + 1)(p + 2)(3p + 2)

completes the proof.
Remark 5.19. In the case of the minimum order rule, i.e. pEm pFm pC , the shape
functions span the following local FE-space


Vpmin (P) = v X pC (P)  trn,Fm (v) (P pFm 1 (Fm ))2 1 m 2,



trn,Fm (v) QpFm +1,pFm QpFm ,pFm +1 (Fm ) 3 m 5 ,
with X pC (P) := RpC ,pC (P) RpC ,pC (P) RpC 1,pC +1 (P).

In the end, we now consider the tetrahedron.

5.2.6

The tetrahedral element

We define the reference tetrahedron as




T = (x, y, z)  x, y, z 0, 0 x + y + z 1 ,

which is the convex hull of the four vertices V1 = (0, 0, 0), V2 = (1, 0, 0), V3 = (0, 1, 0), and
V4 = (1, 1, 1). Similarly as for the triangle, the shape functions can be formulated in terms of
the barycentric coordinates i , which are defined as follows:
z

(1,0,0)
4

1
(0,0,0)

3
(0,1,0)

1
2
3
4

= 1 x y z,
= x,
= y,
= z.

2
(0,0,1)

The outer normal vector on the face Fm = [f1 , f2 , f3 ] is given by

nFm = o = F

with F := 1 + 2 + 3 ,

where o is the coordinate corresponding to the vertex opposite to the face Fm .

98

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

To obtain a tensorial basis also for tetrahedral elements, we utilize the 3-dimensional version
of the Duffy transformation, i.e.,
D:

H
(, , )

T
(x, y, z)

with

x =
y =
z =

1
8 (1
1
4 (1
1
2 (1

+ )(1 )(1 )
+ )(1 )
.
+ )

The tetrahedron T is then interpreted as a collapsed hexahedron H. The inverse map is given
by
1
x
= 2 1yz
1 = 22 +
[1, 1],
1
y
= 2 1z
1
= 2z 1

(14 )
= 23 1
=
4
= 24 1,

3 (1 +2 )
1 +2 +3

[1, 1],
[1, 1].

We only mention that in Karniadakis-Sherwin [60] the collapsing procedure is done stepwise from the hexahedra to a prism, then to a pyramid, and finally to the tetrahedron by
succesively applying the 2-dimensional Duffy transformation (quadrilaterals to triangles).
For the space of FE-shape functions we choose


P p (T ) := xi y j z k : 0 i + j + k p, i, j, k 0 ,
(5.31)
and we set to define the local finite element spaces in such a way that in the case of uniform
polynomial order they correspond to the sequence
id

curl

div

R P p+1 (T ) P p (T ) P p1 (T ) P p2 (T ) {0}.

(5.32)

The H 1 -conforming tetrahedral element

As for the triangle, the vertex-based shape functions can be chosen as the barycentric coordinates, i.e., Vi = i for i = 1, 2, 3, 4, satisfying Vi (Vj ) = ij . As already seen in the low order
construction, the vertex functions span P 1 (T ).
The edge-based shape functions associated with Em = [e1 , e2 ] are also chosen similarly as for
triangles, i.e.
S
m
= Li+2
(e2 e1 , e2 + e1 )
E
i

for 0 i pEm 2.

m
This yields trEm (E
i ) = Li+2 (e2 e1 ) = Li+2 (E ), where E = e2 e1 is the edgeparameterization over the interval [1, 1]. The scaled Legendre-polynomials provide a monomial extension of the edge values into the interior of the tetrahedron by involving a multiplim
P i+2 (T ), and also guarantees zero tangential
cation with (e1 + e2 )i+2 . This ensures E
i
p
traces on all other edges. Moreover, the set of edge-based shape functions spans P0 Em (Em )
on the associated edge.
The face-based functions are chosen to coincide with the cell-based shape functions of the
triangular element T on the face Fm and use an appropriate extension to the domain. In
particular, we set

S
m
= Li+2
(f1 f2 , f1 + f2 ) f3 jS (2f3 F , F ),
F(i,j)

where we used the face extension parameter F = f1 + f2 + f3 . This construction immediately implies linear independence and furthermore,


p
m
) : i + j pFm 3, i, j 0 = P0 Fm (Fm ).
span trFm (F(i,j)

99

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

Concerning the extension of face values into the interior of the element, the scaled Legendre
polynomial provides a monomial extension by jF (as illustrated in Figure 5.6) as follows:
S ( , + )
m
F(i,j)
= Li+2
f1
f2
f3 j
f1
f2

2f3 F
F

jF .

11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
0000011
11111
00000
11111
0000000
11111
00000
11111
00
0000011
11111
00000
11111
00
F
0000011
11111
000
111
00000
11111
m
00
00
11
0000011
11111
000
111
00000
11111
00
11
00
11
00000
11111
000
111
00000
11111
00
11
00
11
00000
11111
000
111
00000
11111
00
11
00
11
00000
11111
000
111
00000
11111
00
00
11
0000011
11111
000
111
00000
11111
00
00
11
0000011
11111
000
111
00000
11111
00
0000011
11111
000
111
00000
11111
00
0000011
11111
00000
11111
00
0000011
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
F =const.,
00000
11111
00000
11111
00000
11111
00000 F=const.
11111

Figure 5.6: Monomial face extension

Finally, we propose cell-based functions in collapsed tensor-product-structure of the form
C
(i,j,k) (x, y, z) = ui (x, y, z) vj (y, z) wk (z)
with factors


S
ui (x, y, z) := Li+2
1 2 , 1 + 2 ,


S 23 (1 4 )
,
vj (x, y, z) = vj (y, z) := 3 j
1 4
wk (x, y, z) = wk (z) := 4 k (24 1),

(5.33a)
(5.33b)
(5.33c)

for i + j + k pC 4, 0 i, j, k. In terms of the hexahedral coordinates (, , ) [1, 1]3

(see figure 5.7) this amounts to

 

1 i+2 1 i+2
ui D (, , ) = Li+2 ()
,
2
2




1 j+1
1+
j ()
,
vj D (, , ) =
2
2


1+
wk D (, , ) =
k ().
2

(5.34a)
(5.34b)
(5.34c)

The linear independence of the cell-based functions now follows easily, and carries over onto
the tetrahedron. Moreover, since there holds (ui D)(1, , ) = 0, (vj D)(, 1, ) = 0, as
well as (wk D)(, , 1) = 0, we obtain vanishing trace on T .
Summarizing, we define the following set of shape functions:

100

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Figure 5.7: Degenerated tensor-product on the tetrahedron: iso-planes where = const.,

= const., or = const.

H 1 -conforming shape functions for the tetrahedral element T

of variable polynomial order p = ({pEm }, {pFm }, pC )
Vertex-based functions:

Vi = i

for i = 1, 2, 3, 4

Edge-based functions:
for m = 1, . . . , 6: edge Em = [e1 , e2 ]
for 0 i pEm 2
S
m
E
= Li+2
(e1 e2 , e1 + e2 )
i
Face-based functions:
for face Fm = [f1 , f2 , f3 ], m=1,...,4
We define F := f1 + f2 + f3 .
for 0 i + j pFm 3,
S
S
m
F
(i,j) = Li+1 (f1 f2 , f1 + f2 ) f3 j (2f3 F , F )
Cell -based function:
for 0 i + j + k pC 4

S
S
C
(i,j,k) = Li+2 (1 2 , 1 + 2 ) f3 j (23 (1 4 ), 1 4 ) f4 k (24 1)

We define the lowest-order space as W V (T ) := span (Vi )1i4 , and denote the span of edgep
based functions associated with Em by WEEmm (T ), the span of face-based functions associated
p
pFm
(T ), and the span of cell-based functions by WCFm (T ).
with Fm by WFm
The local FE-space spanned by the whole set of shape functions on T up to order p =
({pEm }, {pFm }, pC ) is given by
V

Wp(T ) := W (T )

6
M

m=1

p
WEEmm (T

4
M

m=1

Fm
(T ) WCpC (T ).
W Fm

Theorem 5.20. The shape functions collected in the table above are linearly independent and

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

101

H 1 -conforming. For uniform polynomial order p, the shape functions span the local FE-space
Wp (T ) = P p (T ).
Proof. We have already discussed the linear independence of the set of edge-based, facebased, and of cell-based shape functions. The linear independence of the whole set of shape
functions can be verified with similar arguments as in the hexahedral case, i.e., by consecutive
consideration of function traces on edges, faces and cells. The same arguments imply the
H 1 -conformity. Comparing the dimensions of the involved spaces, namely
|Wp (T )| = 4 + 6(p 1) + 2(p 2)(p 1) + 16 (p 3)(p 2)(p 1)
= 61 (p + 3)(p + 2)(p + 1),
and
1
p
|P (T )| = 6 (p + 3)(p + 2)(p + 1)
implies that Wp (T ) = P p (T ) in case of uniform polynomial order p.
Remark 5.21.
1. In the case of the minimum order condition (5.2), the shape functions
span the local space


p
Wpmin (T ) = w P pC (T )  trEm (w) PmEm (Em ) 1 m 6,
.
(5.35)
trFm (w) P pFm (Fm ) 1 m 4
2. The Legendre-type polynomials used in the construction of shape functions can be replaced by other families of orthogonal polynomials, e.g., Jacobi-type polynomials can be
used to improve the condition number and sparsity of the resulting FE-matrices. In the
case of Jacobi polynomials, the polynomial degrees of the factors are not chosen independently, i.e. one obtains ui , v(i,j) , w(i,j,k) . Still, the construction principle of H(curl)- and
H(div)-conforming shape functions, which we present in the following for Legendre-type
polynomials carries over also to other classes of polynomials.

The H(curl)-conforming tetrahedral element

Motivated by the previous constructions, we choose for the lowest-order edge-based shape
functions the lowest-order Nedelec shape functions of the first kind.
The H(curl)-conforming edge-based shape functions are chosen as the gradient fields of edgebased H 1 -conforming shape functions up to order pEm + 1. Hence, by Theorem 5.20 we get
linear independence and obtain that the tangential traces on the associated edge Em span
P pEm (Em )/R, whereas they vanish on all other edges.
The face-based shape functions can be derived similarly as for the H(curl)-conforming triangular elements, namely by involving summands of gradient fields of face-based H 1 -conforming
functions as follows:
ui vj + ui vj ,

ui vj ui vj ,

(f1 f2 f1 f2 ) vj

S ( , + ), v := S ( , ) for 0 i, j, i + j p
with ui := Li+2
j
F
F
Fm 2.
f1
f2
f1
f2
f3 j
f3
On the associated face Fm = [f1 , f2 , f3 ], the tangential traces of these functions coincide
with the cell-based H(curl)-conforming shape functions for the triangular element up to order
pFm . Hence, we obtain linear independence of the face-based functions from the corresponding considerations on triangles. Moreover, by construction, the above functions belong to

102

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

3
pFm
(Fm ) on the corresponding face, whereas the
P pFm (T ) . Their tangential traces span P ,0
traces vanish on Fk 6= Fm , since there either ui = 0 or vj = 0.
Concerning cell-based shape functions, we first consider the gradients of the scalar cell-based
shape functions
C,1
(i,j,k) = ui vj wk + ui vj wk + ui vj wk

0 i + j + k pC 3,

where the factors ui , vj , wk are chosen according to (5.33). The set of gradient fields can be
enriched in the following way:
S ( , + ), v := S (2 (1 ), w :=
Lemma 5.22. Suppose ui := Li+2
2
1 1
2
j
3 j
3
f4
k
0
4 k (24 1), and N
=
(

).
Then
the
set
1 2
1
2
[1,2]
0
VpC (T ) := span{ui vj wk , ui vj wk , ui vj wk , N
[1,2] vj wk : 0 i, j, k, i + j + k p 3}

3
contains 12 (p + 1)(p 1)(p 2) linearly independent functions in P p (T ) with vanishing
tangential trace on T .
0
Proof.
1. The trace of ui as well as the tangential trace N
[1,2] vanishes on the faces opposite
vertex V1 and vertex V2 . Moreover, the trace of vj vanishes on the face opposite V3 , and
wk is zero on the face opposite V4 . Hence, the tangential trace of VpC (T ) vanishes
on the whole of T .

2. We can easily show linear independence by utilizing the tensor-product construction:

(a) The vector fields {(i,j,k) = ui vj wk }ijk are linearly independent:

Since wk only depends on z, the linear independence can be deduced analogously
to the linear independence of the H 1 -conforming cell-based shape functions.
0
(b) The vector fields {ui vj wk , ui vj wk , N
[1,2] vj wk } are linearly independent: We
4 ), where 4 is constant and hence also wk is
restrict the vector fields to planes Tz (
constant. The tangential trace on the triangular faces Tz (4 ) (with n = 4 ) of
0
the vector fields ui vj wk , vj ui wk and N
[1,2] vj wk coincides (up to scaling and
a constant) with the cell-based shape functions of the H(curl)-triangular element
T . Hence, by Theorem 5.6 and Theorem 5.8, we obtain linear independence of
0
ui vj wk , vj ui wk and N
[1,2] vj wk for all i, j.
Considering

=
v

1
i,j,k cijk ui vj

0
wk + c2ijk vj ui wk + +c4jk N
[1,2] vj wk = 0

we obtain by restriction to planes Tz (4 ) that there holds

clij =

X
k

clijk wk = 0

for l = 1, 2, 4, i, j

and for any 4 (0, 1). Due to the linear independence of wk in the z-component
we obtain clijk = 0 for l = 1, 2, 4.

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

103

(c) Let the trivial function v = 0 be constructed by a combination of basis functions,

i.e.,
X
0
v=
c1ijk ui vj wk + c2ijk vj ui wk + c3ijk vj ui wk + c4jk N
[1,2] vj wk = 0.
i,j,k

By considering the tangential components of v on planes Tz (4 ) (where Tz wk =

0), we obtain clijk = 0 for all l = 1, 2, 4 due to the linear independence of (b). The
remaining part of v is a linear combination of linearly independent vector fields.
Hence, c3ijk = 0, which implies the overall linear independence.
The proposed basis functions are summarized in the following table:
H(curl)-conforming shape functions for the tetrahedron
T


of variable order p = {pEm }, {pFm }, pC

Edge-based functions

for m = 1, . . . 6: edge Em = [e1 , e2 ]

0
N
m = e1 e2 e1 e2

Lowest-order Nedelec function:

Higher-order edge-based functions (gradient fields):

for 0 i pEm 1 :
Face-based functions

S
m
E
= Li+2
(e1 e2 , e1 + e2 )
i

for m = 1, . . . , 4: face Fm = [f1 , f2 , f3 ]

We define ui
vj

S
(f1 f2 , f1 + f2 ),
:= Li+2
S
:= 3 j (2f3 F , F ),

for 0 i + j pFm 2:
Type 1 (gradient fields):

m, 1
F
(i,j)

Type 2:

Fm ,2
(i,j)

Type 3:

Fm ,3
(0,j)

= ui vj

= ui vj ui vj
=

f1 f2 f1 f2 vj

Cell-based functions
S
We define ui := Li+2
(1 2 , 1 + 2 ),
vj := 3 jS (23 (1 4 ), 1 4 ),
wk := 4 k (24 1),
for 0 i + j + k pC 3:

s.t. Fm = ui vj .

s.t. Fm = ui vj wk .

Type 1 (gradient fields):

C,1
(i,j,k)

= C
(i,j,k) = ui vj wk

Type 2:

C,2
(i,j,k)

= ui vj wk ui vi wk + ui vj wk

C,2
p +(i,j,k)
C

Type 3:

C,3
(0,j,k)

= ui vj wk + ui vi wk ui vj wk


= 1 2 1 2 vj wk

0
We define the lowest-order space as VN0 (T ) := span (N
m )1m6 , and denote the span
pEm
(T ), the span of face-based functions
of edge-based functions associated with Em by VEm
p
pFm
associated with Fm by VFm (T ), and the span of cell-based functions by VC Fm (T ).

104

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

The local FE-space spanned by the whole set of shape functions on T up to order p =
({pEm }, {pFm }, pC ) is given by
Vp(T ) := VN0 (T )

6
M

pEm
(T
VEm

m=1

4
M

m=1

VFmFm (T ) VC Fm (T ).

Theorem 5.23. The N0 -E-F -C-based shape functions presented in the table above, are
linearly independent and H(curl)-conforming.
Moreover, they contain gradient fields
Wp+1 (T ) Vp(T ), i.e.,
(T )
WpEEm +1 (T ) = VpEEm
m
m

WpCC +1 (T

VpCC (T

).

m = 1, . . . , 6,

WpFFm +1 (T ) VpFFmm (T )
m

m = 1, . . . 4,

For uniform polynomial order p the shape functions span the local FE-space

3

Vp(T ) = P p (T )
with P p+1 (T ) P p (T ).

Proof. The edge-based shape functions (gradient fields) are linearly independent due to Theorem 5.20. Their tangential traces span P pEm (Em ) and vanish on all other edges Ek 6= Em .
The face-based shape functions are linearly independent and their tangential traces on Fm
pFm
(Fm ) with zero tangential trace on all edges and all faces Fk 6= Fm .
span P ,0
Due to Lemma 5.22 we obtain linear independence of the cell-based shape functions and the
vanishing of their tangential trace on T .
By consecutive consideration of tangential traces on edges and faces we can deduce the linear
independence of the whole set of shape functions. The H(curl)-conformity follows with the
same arguments.

3
Finally, for uniform polynomial order p all shape functions belong to P p (T ) by construction. Comparing the dimensions of the involved spaces, i.e.,

1
1
|Vp(T )| = 6(p + 1) + 4(p 1)(p + 1) + (p 2)(p 1)(p + 1) = (p + 1)(p + 2)(p + 3)
2
2

3
1
| P p (T ) | = (p + 1)(p + 2)(p + 3)
2

3
we conclude that the shape functions form a basis for Vp (T ) = P p (T ) .

Remark 5.24. If the minimum order condition (5.2) holds, the stated shape functions span
the local space


3 
Vpmin (T ) = v P pC (T )  tr,Em (v) P pEm (Em ) 1 m 6,

2

(5.36)
tr,Fk (v) P pFk (Fk ) 1 k 4 .
H(div)-conforming shape functions for tetrahedral elements

We want to construct the local FE-space of order p in such a way that

curl Vp(T ) Qp(T ).
This requires that the normal trace (on faces) of the local space spans P pFm 1 (Fm ), and

3
the cell-based space spans P pC 1 (T ) H0 (div, T ). In view of the sequence of polynomial
spaces (5.32), we choose the FE-space by
Qp (T ) = P p1 (T ))3

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

105

in case of uniform polynomial order p. We remind the reader that the classical defintion
of the H(div)-conforming local space of order p on the tetrahedral element, coincides with

3
P p (T ) and corresponds to Qp+1 (T ) in our notation.

The lowest-order shape functions are provided by the lowest-order Raviart-Thomas functions
with constant normal trace on the corresponding face and zero trace normal on all other faces,
cf. Lemma 4.18.
The higher-order face-based H(div)-conforming shape functions can be chosen as the curlfields of the face-based H(curl)-conforming shape functions except for gradient fields. This
yields 21 (pF 1)(pF + 2) linearly independent shape functions, where the normal traces span
P pFm 1 (Fm )/R on Fm , while vanishing on all other faces.
We choose the curl-fields of H(curl)-conforming cell-based shape functions as subset of the
cell-based shape functions:
curl C,2
(i,j,k)
curl C,2
pC +(i,j)
curl C,3
(0,j,k)

= 2ui vj wk 2ui vj wk
= 2 wk ui vj 2ui vj wk




= 1 2 1 2 vj wk 21 2 vj wk

The main properties of these sets of functions are summarized in the following lemma.
S ( , + ), v := S (2 (1 ), w := (2 1),
Lemma 5.25. Let ui := Li+2
2
1 1
2
j
3 j
3
4 k
4
f4
k
N0
RT 0
and [1,2] = (1 2 1 2 ) and [1,2,3] := 1 2 3 + 2 3 1 + 3 1 2 .
We use the notation



RT 0
0

M1 := ui vk wk , curl( N
[1,2] vj ) wk , [1,2,3] wk i, j, k 0, i + j + k p 3



0

M2 := wk ui vj , N
[1,2] wk vj i, j, k 0, i + j + k p 3



M3 := vj wk ui  i, j, k 0, i + j + k p 3 .

Then the functions in M1 M2 M3 are linearly independent. Moreover, there holds

P p1
n,0 (T ) = span(M1 ) span(M2 ) span(M3 ).

Proof. By construction, all involved functions belong to P p (T ) and have vanishing normal
trace on T . We show linear independence by looking at linear combinations q = c1 1 +
c2 2 + c3 3 = 0 with i Mi .
1. Restricting q to the normal trace corresponding to planes Tz where z = const. yields
trn,Tz (q) = c1 trn,Tz ( 1 ) = 0, since trn,Tz ( 3 ) = trn,Tz ( 2 ) = 0.
We observe that up to scaling, the normal traces of the functions in M1 correspond to
the face-based H(div)-conforming shape functions on the triangle/face [1, 2, 3]. Since
the vector-fields in M1 are constructed as products of (scaled, but linearly independent)
face-based H(div)-conforming shape functions on planes z = const., and a set of linearly
independent polynomials wk in z, we obtain linear independence of the functions in M1 .
Furthermore, we obtain c1 = 0 in the linear combination of q.
2. Next, we restrict q to planes T , where = const. Since there holds vj wk = 0

j
0
on T , we obtain trn,T (q) = c2 2 = c2,1 x ui wz + c2,2 N
[1,2] ex ui wz (1 z) = 0. Here,

106

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

linear independence follows as in the proof of Lemma 5.7 by replacing vj with wk and
0
utilizing the fact that x ui is at least linear in x, while N
[1,2] ex is constant in x. Hence,
we obtain c2 = 0.
3. We conclude by showing the linear independece of M3 :
Taking the scalar product with ex yields (vj wk ui ) ex = y vj wk ui . The linear
independence can be shown by forming partial derivatives and following the proof of
the linear independence of H 1 -conforming cell-based shape functions. Hence, we obtain
c3 =0.
Since the shape function in M1 M2 M3 are linearly independent, counting dim(P pn,0 (T )) =
1
2
2 p(p p 2) concludes the proof.
We summarize the shape functions in the following table:
H(div)-conforming shape functions for the tetrahedron T
of variable order p = ({pFm }, pC )
Face-based functions
for m = 1, . . . , 4 face Fm = [f1 , f2 , f3 ]
Raviart-Thomas functions
0
RT
= f1 f2 f3 + f2 f3 f1 + f3 f1 f2
m

Higher-order face based functions (divergence-free)

We define ui
vj

S
(f1 f2 , f1 + f2 ),
:= Li+2
:= 3 jS (2f3 F , F ),

s.t. ,Fm = ui vj .

for 0 i + j pFm 2
m
F
(i,j)

m
F
pFm +(0,j)

Fm ,2
curl (i,j)

Fm ,3
curl (0,j)

= 2ui vj
=

f1 f2 f1 f2

Cell-based functions
S
(1 2 , 1 + 2 ),
We define ui := Li+2
vj := 3 jS (23 (1 4 ), 1 4 ),
wk := 4 k (24 1),

vj 2f1 f2 vj

s.t. ,Fm = ui vj wk .

for 0 i + j + k pC 3

Type 1: (div-free)
C,1
(i,j,k)

C,1
p +(i,j,k)

C,1
2p +(0,j,k)

curl C,2
(i,j,k) = 2ui vj wk 2ui vj wk

curl C,2
p +(i,j) = 2 wk ui vj 2ui vj wk
C

curl C,3
(,j,k)

= 1 2 1 2 vj wk 21 2 vj wk

rrcl

107

5.2. CONSTRUCTION OF CONFORMING SHAPE FUNCTIONS

Type 2:

Type 3:

C,2
(i,j,k)

= ui vj wk

C,2
(0,j,k)

C,3
(0,0,k)

RT 0
wk
= [1,2,3]

(1 2 1 2 ) wk vj

0
with RT
[1,2,3] := 1 2 3 + 2 3 1 + 3 1 2 .

0
We define the lowest-order space as QRT 0 (T ) := span ( RT
m )1m4 , and denote the span
p
of face-based (associated with Fm ) functions by QFFmm (T ), the span of cell-based functions by
p
QCFm (T ) and the local FE-element space by
L
p
p
Qp(T ) := QRT 0 (T ) 4m=1 QFFmm (T ) QCFm (T ).

Theorem 5.26. The RT 0 -F -C-based shape functions listed in the table above, are linearly
independent and H(div)-conforming. Moreover, they include gradient fields, i.e.,
curl VpFFmm (T ) QFpFmm (T ) m = 1, . . . 4 and curl VpCC (T ) QC
pC (T ).
For uniform polynomial order p the shape functions span the local FE-space

3

curl
Qp(T ) = P p1 (T )
with P p+1 (T ) P p+1 (T ) P p1 (T ).

Proof. We already have shown the linear independence of the face-based shape functions,
since they are curl-fields of a set of linearly independent functions which are above all linearly independent to gradient fields. In combination with the lowest-order shape functions,
their tangential traces span P pFm 1 (Fm ) on the associated face Fm , while vanishing on the
remaining faces.
The cell-based shape functions are constructed as linearly independent combinations of linearly independent cell-based vector-fields, cf. 5.25. Hence, they are linearly independent and
pC 1
span P0,n
(P).
Due to the hierarchical RT 0 -F -C constructions, we can deduce linear indpendence of the
whole set of shape functions. In the case of uniform polynomial order p, we conclude by
counting the dimensions of the involved spaces:
1
1
|Qp (T )| = 4 + 2(p 1)(p + 2) + p(p2 p 2) = p(p + 1)(p + 2),
2
2
1
|(P p1 (T ))3 | = p(p + 1)(p + 2).
2

and

Remark 5.27. If the minimum order condition (5.2) holds, the shape functions span the local
space



3 
(5.37)
Qpmin (T ) = q P pC 1 (T )  trn,Fm (q) P pFm 1 (Fm ) 1 m 4 .

5.2.7

N
ed
elec elements of the first kind and other incomplete FE-spaces

In the case of simplicial elements, the presented H(curl)- and H(div)-conforming finite elements are Nedelec elements of second kind or Brezzi-Douglas-Marini elements, resectively. As
outlined in Chapter 4, the flux fields of these spaces are approximated with one degree less

108

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

than the primal variables. The same approximation order is provided by Nedelec elements of
the first kind and Raviart-Thomas-Nedelec elements, respectively.
Due to the explicit usage of differential fields in the construction of basis functions we can
easily extend our concepts also to the higher-order Nedelec elements of the first kind and
higher-order Raviart-Thomas-Nedelec elements. In the following we shortly summarize some
results for the three-dimensional case; the two-dimensional case follows easily then.
We denote the span of H(curl)-conforming face-based shape functions of first, second, and
third type associated with the face Fm by VpFFmm,1 (K), VpFFmm,2 (K), VpFFmm,3 (K), and the span of
C,2
H(curl)-conforming cell-based shape of first, second and third type by VpC,1
C (K), VpC (K),
VpC,3
edelec element of first kind of order p is now defined by
C (K). The local space of N
M
M
VpFFmm,1 (K) VpFFm ,2+1 (K) VpFFm ,3+1 (K)
(K)

VpEEm
NpI (K) := VN0 (K)
m
m

Fm FK

Em EK

(K) VpC,2
(K) VpC,3
(K),
VpC,1
C
C +1
C +1

and the basis functions are chosen in a similar manner as outlined above.
We note that the reduced spaces do not exactly coincide with the classically defined Nedelec
elements of first kind which are rotational invariant. In neither of both cases the resulting
FE-basis is orthogonal on the reduced gradient fields. In both variants the spaces of the
curl-fields coincide as well as the orders of approximation obtained for the functions and its
curl.
Similarly, we denote the span of H(div)-conforming cell-based shape of first, second and
C,2
C,3
third type by QC,1
pC (K), QpC (K), QpC (K), and define the local space of the Raviart-ThomasNedelec element of order p as follows:
M
C,3
(K) QC,2
QFpFmm (K) QpC,1
RT p(K) := QRT 0 (K)
pC +1 (K) QpC +1 (K).
C
Fm FK

Again the choice of basis functions is very similar to the one outlined in detail above.
In the case of uniform polynomial order p, these definitions are related to the corresponding
classical elements, if we take into account that on simplicial elements the classical H(div)conforming space of order p 1 refers to the H(div)-conforming space p + 1 in our notation.
Note that since the differential fields are explicitly realized in the FE-basis, we can achieve any
order of approximation for the function and for its differential field, separately, by defining
incomplete spaces as presented above. For instance, this will be used in gauging strategies
in Chapter 6, where we reduce all higher-order gradient shape functions in the H(curl)conforming FE-basis.

5.2.8

Elements with anisotropic polynomial order distribution

In applications, one may be interesting in highly-anisotropic elements. These arise, e.g., from
geometric h-refinement, when handling thin structures within the computational domain or
plates and shells. In such cases, there is no need to support full polynomial degree in all
directions. Hence, one is interested in allowing variable polynomial order, in particular for
tensor-product elements, i.e., on hexahedra and prisms. Since the tensor product structure
was explicitly used in the construction the shape functions, such anisotropic polynomial distributions can easily be realized.

109

5.3. GLOBAL FINITE ELEMENT SPACES

First, each quadrilateral face is assigned two polynomial degrees pF = (pF,1 , pF,2 ), and each
hexahedral cell may have three polynomial degrees pC = (pC,x , pC,y , pC,z ). Then the definition
of conforming shape functions can easily be extended to anisotropic polynomial order by
using different ranges for the indices corresponding to different directions, i.e., 1 i pF1 ,
1 j pF2 on quadrilateral faces, and 1 i pCx , 1 j pCy , 1 k pCz for hexahedral
cells. Prismatic elements are treated similarly.
Note that assigning two polynomial degrees pF = (pF,1 , pF,2 ) to a quadrilateral face, the fixed
face orientation establishes automatically the conformity on the global level.

5.3

Global Finite Element Spaces

In this section we show that, based on the local finite element spaces defined on reference
elements, the construction of H 1 -, H(curl)-, and H(div)-conforming global finite element
spaces on a triangulation Th is straightforward. We assume that the triangulation is regular,
cf. Section 4.1.2, but may involve a mixture of element topologies, i.e., hexahedra, tetrahedra,
and prisms in a three-dimensional setting, respectively quadrilaterals and triangles in the
two-dimensional case. We will only treat the three-dimensional case in detail, but the same
constructions and results also hold for the simpler two-dimensional case. Moreover, we assume
K satisfy the properties as defined in Section 4.1.4, and
that the element mappings K : K
utilize the notation introduced there. By restricting ourselves to affine element mappings, the
polynomial degree of shape functions is not increased by the mapping to physical coordinates.
In the course of this section we mark all quantities concerning the reference element by the
hat marker (), and use the following notation for local patches:
V :=

KTh : V VK

K,

E :=

KTh : EEK

K,

F :=

K,

(5.38)

KTh : F FK

which are associated with a vertex V V, an edge E E, a face F F, respectively. Each

edge E E, face F F, and cell K Th in the mesh is assigned a polynomial degree pE , pF ,
respectively pC ; the polynomial degrees are collected in a vector p.
The construction of global basis functions in the physical domains works as follows: first we
define the shape functions locally (elementwise) by conforming transformation of functions
on the reference domain. The global functions, which have support only on local patches, are
then defined by taking the union of the contributing element shape functions. As we will see,
global conformity is established easily due to our special constructions.

5.3.1

The H 1 -conforming global finite element space

Shape functions defined on a physical element K Th are defined by the H 1 -conforming

onto the reference element K.
Due to Lemma 4.10
mapping of shape functions Wh,p(K)
we have
:= 1
K .
We associate the vertices, edges, and faces of each physical element with the corresponding
entities in the global mesh, and we identify the mapped shape functions (with support on
the physical element K) with the restriction of the global basis functions (with support on a
patch). Such an identification is meaningful, since the vertices, edges, and faces on interelement boundaries coincide for adjacent physical elements. Here the global orientation of edges

110

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

and faces on the reference element level plays an important role. We arrive at a global space
M
M
M
(5.39)
WpKK (Th )
WpF (Th )
Wh,p(Th ) := Wh,1 (Th )
WpEE (Th )
EE

F F

KTh

with lowest-order, edge-based, face-based, and cell-based spaces as follows



Wh,1 (Th ) := span Vi : Vi V ,


WpEEm (Th ) := span E
having support on E ,
i : 0 i pE 2

 F
F
having support on F ,
WpF (Th )
:= span i : 0 i < NH 1 (Fm , pF )


(K,
p
)
having
support on K,
WpKC (Th )
:= span C
:
0

i
<
N
C
H1
i

where NH 1 (F, pF ) and NH 1 (K, pC ) denote the number of H 1 -conforming degrees of freedom
associated with the face F and to the cell K, respectively. These depend on the polynomial
degree as well as on the topology of the face or cell. The number of degrees of freedom on the
edges is uniquely determined by the polynomial degree pE .
Theorem 5.28. The space Wh,p(Th ) defined in (5.39) is an H 1 ()-conforming finite element
space.
Proof. The shape functions on reference elements are constructed in such a way that they
coincide at vertices, on edges and on faces even after transformation to physical coordinates.
Since the shape functions associated with an interfaces between elements (a vertex, an edge,
or a face) are the only ones with non-vanishing trace on the interface, and the orientation
is treated globally (cf. Section 5.2.1), the global basis functions match each other on interelement boundaries.

5.3.2

The H(curl)-conforming global finite element space

We define the restriction of a global basis function onto a physical element K Th through
defined on the reference
Vh,p(K)
the H(curl)-conforming mapping of the shape function

element K. According to Lemma 4.15 we have

T
1

:= FK
K .

The identification of topological entities on the local and global levels works as in the H 1 case.
By numbering the mapped shape functions sequentially on each edge, face, and cell we arrive
at the following representation of the global space:
M
M
M
(5.40)
VpKK (Th ).
VpF (Th )
VpEE (Th )
Vh,p(Th ) := Vh,N0 (Th )
EE

F F

KTh

The lowest-order, edge-based, face-based and cell-based spaces are defined as follows
 0

Vh,1 (Th ) := span N
: E E ,
E


VpEE (Th ) := span E
having support on E ,
i : 0 i pE 1
 F

F
VpF (Th ) := span i : 0 i < NH(curl) (F, pF )
having support on F ,
 C

K
VpC (Th ) := span i : 0 i < NH(curl) (K, pC )
having support on K,

where NH(curl) (F, pF ) and NH(curl) (K, pC ) denote the number of H(curl)-conforming degrees
of freedom associated with the face F and with the cell K, respectively, depending on the
local polynomial degrees and the topology of the face or cell.

111

5.3. GLOBAL FINITE ELEMENT SPACES

Theorem 5.29. The space Vh,p(Th ) is H(curl)-conforming, i.e., a subspace of H(curl). There
holds
Wh,p+1 (Th ) Vh,p(Th )
(5.41)
in particular
WpEE +1 (Th )

VpEE (Th ),

Wh,1 (Th ) Vh,N0 (Th ),

WpFF +1 (Th )

VpFF (Th ),

(5.42a)
WpCC +1 (Th )

VpCC (Th ),

(5.42b)

for all E E, for all F F and for all K K even for non-uniform polynomial degrees.
Proof. The shape functions for all reference topologies have been constructed in such a way
their tangential traces on edges and faces coincide. Moreover, only shape functions associated
with the interface of elements contribute to the tangential trace on common edges and/or
the common face. Due to the use of global orientations for edges and faces, the orientation
matches also locally on the interfaces which ensures tangential continuity of the mapped basis
functions across inter-element boundaries.
The properties (5.42) are easily verified by the construction of the basis functions: The lowestorder space was already considered in Chapter 4. The explicit usage of gradient functions
within the construction of the finite-element basis on the reference element implies that the
inclusions hold on reference elements. By using H(curl)-conforming transformations for the
construction of the shape functions on physical elements, we ensure that the gradient fields
on the reference element are mapped onto the gradient fields on the physical element, which
yields
WpEE +1 (Th ) VpEE (Th ) : =

WpFF +1 (Th ) VpFF (Th ) : =

WpCC +1 (Th ) VpCC (Th ) : =

for all edges E E, all faces F F, and all cells K Th . Hence, Wh,1 (Th )
Vh,N0 (Th ), WpEE +1 (Th ) VpEE (Th ) and WpEE +1 (Th ) VpEE (Th ). The counting argument




dim WpEE +1 (Th ) = dim VpEE (Th ) = pEm yields the coincidence of the edge-based spaces.
Finally, (5.42) implies (5.41).
Remark 5.30. If one does not use gradient fields explicitly in the construction of H(curl)conforming shape functions, the following may be observed: The inclusion (5.41) concerning
the global spaces only holds in the case when the polynomial degrees on edges, faces and cells
satisfy the minimum order rule
F F E E on F :
K Th F FK :

pE pF ,

pF pK .

(5.43)

The space splitting of higher-order gradients (5.42b) does not hold in general, not even in case
of a uniform polynomial degree.

5.3.3

The H(div)-conforming global finite element space

The restriction of global basis functions onto a physical element K Th is defined via the
Piola transformation
1
1 ,
:= JK
FK
K

112

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Qh,p(K)
from the reference element K.
Numbering the global basis
of shape functions
functions sequentially on each edge, face and cell yields the global space
M
M
(5.44)
QK
QFpF (Th )
Qh,p(Th ) := Qh,RT 0 (Th )
pC (Th )
F F

KTh

with the lowest-order, edge-based, face-based and cell-based spaces defined as follows


0
Qh,1 (Th ) := span RT
: F F ,
F


QFpFm (Th ) := span Fi : 1 i NH(div) (Fm , pFm )
having support on F ,
 C

K
QpC (Th ) := span i : 1 i NH(div) (K, pC )
having support on K,

where NH(div) (Fm , pFm ) and NH(div) (K, pC ) denote the number of H(div)-conforming degrees
of freedom associated with the face Fm and to the cell K, respectively.
Theorem 5.31. The global space Qh,p(Th ) is an H(div)-conforming finite element space.
Moreover, there holds
curl Vh,p(Th ) Qh,p(Th ).
(5.45)
In particular
curl Vh,N0 (Th ) Qh,RT 0 (Th ),
curl VpFF (Th )

QFpF (Th ),

curl VpCC (Th )

(5.46a)
QC
pC (Th )

(5.46b)

for all F F and for all K K.

Proof. We follow the proof of Theorem 5.29: The normal traces of the local shape functions
on faces coincide for all types of the reference topologies. Due to the H(div)-conforming
transformation and the global orientation of local edges and local faces, we obtain continuity
of the normal traces on inter-element boundaries.
Formuly (5.46) was already considered for lowest-order space in Chapter 4. The explicit
usage of gradient functions within the construction of the finite element basis on the reference
element implies that the inclusions 5.46 hold on the reference element. Similar statements for
the physical elements follows again from the construction using the conforming transformation,
i.e., we obtain
VpFF (Th ) QFpF (Th ) : = curl

VpCC (Th ) QC
pC (Th ) : = curl
for all faces F F and all cells K Th . This implies the inclusions of the various local curl
spaces stated (5.46b). Counting the dimensions yields dim | curl VpFF (Th )| = dim |QFpF (Th )|.
Finally, (5.46) implies (5.45), which completes the proof.
Observations similar to those in Remark 5.30 also hold for the H(div)-conforming spaces.

5.3.4

The L2 -conforming global finite element space

For the sake of completeness, we also address the L2 -conforming spaces briefly. The L2 conforming global FE-space is defined by divergence-fields of the proposed H(div)-conforming

113

5.4. THE LOCAL EXACT SEQUENCE PROPERTY

basis functions. In particular, we use constants on the elements as lowest-order basis functions,
and construct the higher-order cell-based functions as divergence-fields
0 = 1,

)
i := div(
i

QC (K)
being linear-independent to curl-fields. We use the transformation of
for
i
p
divergence-fields according to Lemma 4.19 to define the basis function on the physical
1
element K: = JK
1
K .
In the case of L2 -conforming finite element spaces there are no continuity requirements across
inter-element boundaries, hence, there is no need to associate any global degrees of freedom
with the boundary of the element. This implies the following global FE-space





M
div QK
Sh,p(Th ) = div Qh,p(Th ) = div Qh,RT 0 (Th )
p (Th ) ,

(5.47)

KK

which coincides with the classical L2 -conforming global FE-spaces used within the de Rham
sequence. By the exact sequence property of
the lowest-order global FE-spaces, cf.
Theorem

= div QC
4.28, we obtain Sh,0 (Th ) = div Qh,RT 0 (Th ) . Moreover, SpCC (K)
pC (K) can be
and a counting argument.
easily verified by the definition of SpCC (K)

5.4

The Local Exact Sequence Property

The global finite element spaces were constructed as direct sums of the lowest-order space and
spaces having only local support on edge-patches, on face-patches or single elements. Due to
the special construction of the finite element spaces we obtain exact sequences on a finer level.
The following theorem summarizes the results of Theorem 5.28 and Theorem 5.29.
Theorem 5.32 (The local exact sequence property). Let denote a simply-connected domain
with connected boundary, Th be a regular mesh, and let pE , pF , and pC denote the polynomial
orders associated with edges E E, faces F F, and cells K Th . For ease of notation we
collect the polynomial degrees in a vector p.
If the sequence of global lowest-order spaces (cf. Theorem 4.28)) is exact, then the sequence
of conforming finite element spaces provided by (5.39), (5.40), (5.44), and (5.47) satisfies the
following local exact sequence property:

curl

curl

{0}

curl

QFpF (Th )

curl

QK
pK (Th )

SpKK (Th ) {0}

div

V (T )/R V N0 (T ) QRT 0 (T ) S
Wh,1
h
h
h
h, 0 (Th ) {0}
h
h

WpEE +1 (Th )

VpEE (Th )

WpFF +1 (Th )

VpFF (Th )

WpKK +1 (Th )

VpKK (Th )

E E
div

div

{0}

F F
0

K Th .

Proof. Due to Theorem 5.28 and Theorem 5.29 we only have to verify that the divergence
K
operator is surjective from QK
pK (Th ) onto SpK (Th ), which follows by counting the space dimensions.

114

CHAPTER 5. HIGH ORDER FINITE ELEMENTS

Remark 5.33. The local exact sequence property is also fulfilled if we use Nedelec elements of
the first kind and Raviart-Thomas-Nedelec finite elements, or finite elements with anisotropic
polynomial order distribution.
A general set of H(curl)-conforming and H(div)-conforming shape functions, which does explicitly use differential fields in the construction of the local basis, or at least not provide the
Then the sequence
analogue of the local exact sequence property on the reference element K.
(5.48) of global finite element spaces is only exact, if the polynomial order distribution all over
the mesh satisfies the minimum order rule (5.43).
Corollary 5.34 (Exactness of the global discrete sequence). The local exact sequence property
implies that the sequence of the constructed conforming global FE-spaces is exact, i.e.
id

curl

div

R Wh,p+1 (Th ) Vh,p(Th ) Qh,p(Th ) Sh,p(Th ) {0}.

(5.48)

Assuming the local exact sequence property is one of main issues of this thesis. Besides the
advantage of arbitrary and variable polynomial order distribution over the mesh, we will rely
on this property in the course of designing robust preconditioners for curl curl problems as
well as reduced basis approaches in the next chapter.

Chapter 6

Iterative Solvers for

Electromagnetic Problems
This chapter is devoted to stable and efficient solution strategies for linear systems arising from
H(curl)-conforming discretizations for a general curl-curl-problem curl 1 curl u+u = f . In
general, these linear systems are of huge dimension and ill-conditioned. Hence, direct solving
is out of question and even impossible due to memory and time resources. We have to use
iterative solvers involving preconditioning.
Unfortunately, standard solution techniques, which yield successive, stable and efficient linear solvers for H 1 ()-problems, usually fail on curl-curl problems. Even classical multigrid
techniques (Hackbusch [52], Bramble-Zhang [30]), which provide one of the most efficient
solvers for elliptic problems, result in poor convergence rates. Moreover, standard solvers
deteriorate for small L2 -coefficients 0 < 1.
The problem can be traced back to the large kernel of the curl-operator. On irrotational
fields the curl-curl problem reduces to the identity operator, while acting as second-order
differential operator on solenoidal fields. By the choice of appropriate smoothers, respecting
the Helmholtz decomposition, multigrid techniques regain their efficiency. We refer to the
pioneering works by Hiptmair [59], [55], and Arnold-Falk-Winther [6], [7].
In the following, we focus on the design of parameter()-robust preconditioners, which in
addition also provide appropriate smoothers for multigrid schemes. By means of an additive Schwarz framework the parameter-robustness is guaranteed, if the local decomposition
of the FE-space also represents a correct splitting of the kernel of the curl-operator. We
exploit the local exact sequence property to ensure a correct space splitting even for simple
decompositions.

6.1

Basic concepts of iterative solvers

We consider a finite element space V , an associated FE-basis {i : 1 i N } and the

corresponding finite dimensional variational problem:
Find u V such that
a(u, v) = f (v)
v V.
(6.1)
Due to the Galerkin isomorphism (cf. Chapter 4) the discrete variational problem is equivalent
to the linear system: Find u RN such that
A u = f,

115

(6.2)

116

CHAPTER 6. ITERATIVE SOLVERS

for A RN N and f RN defined by Aij := a(i , j ), f i := f (i ), and u =

PN

i=1 ui i .

For a detailed exposition on iterative solvers and preconditioning, we refer to textbooks as

Braess [27], Grossmann-Roos [51] and Hackbusch [53]. In the following, we briefly recall
some of the main aspects.
The convergence rate of iterative solvers, such as Richardsons method and the conjugate gradient (CG) method, mainly depends on the the condition number (A ) := kA kkA 1 k of the
matrix A . Hence, these solvers can be tremendously improved by means of a preconditioned
system C 1 A .
Suppose A and C to be symmetric, positive definite matrices. A good preconditioner C to a
matrix A must satisfy following two properties:
the matrix-vector product w = C 1 d can be computed fast,
C is spectrally equivalent to A , i.e.
1 uT C u uT A u 2 uT C u
with a small condition number (C 1 A)

2
1

(6.3)

of the preconditioned matrix.

A customary and efficient method for solving symmetric positive definite linear systems is
theq
preconditioned conjugate gradient method (PCG). The iteration number
behaves then as


(C 1 A )1
.
O (C 1 A ) in the case of a fixed error bound. The convergence rate is 1
(C

A +1

In case of non-symmetric and indefinite problems one uses Krylov-based subspace methods,
such as the generalized minimal residual method (GMRES) or the stabilized bi-orthogonal
gradient method (BI-CGSTAB). We refer to standard textbooks on iterative solvers for sparse
linear systems as e.g. Saad [77], van der Vorst [91].

6.1.1

Additive Schwarz Methods (ASM)

The additive Schwarz technique provides an abstract framework for the design of preconditioners based on the splitting of the underlying finite element space. For a detailed description
we refer to textbooks such as Smith et al. [84] and Toselli-Widlund [90]. The short
overview below follows the latter one.
We consider the finite dimensional variational problem (6.1) with a positive definite bilinear
form a : V V R. This is equivalent to the linear operator equation
Au = F in V
with A : V V defined by hAu, vi = a(u, v) and F V defined by hF, vi = f (v), where
h, i denotes the duality product in V V .
We consider a family of finite element spaces {Vi } := {Vi , i = 0, . . . , M } and prolongation
operators
Ri : Vi V.
(6.4)

The corresponding adjoint operators will be denoted by RiT . We assume that the whole space
can be decomposed in prolongated local spaces:
V =

M
X
i=0

Ri Vi .

(6.5)

117

6.1. BASIC CONCEPTS

This splitting need not be unique.

On each space Vi we introduce a symmetric, positive definite local bilinear form
ci (, ) : Vi Vi R.

(6.6)

In operator notation, we use Ci : Vi Vi , which is defined by hCi ui , vi i = ci (ui , vi ) for

ui , vi Vi . The bilinear form ci (, ) should provide a local approximation of a(, ).
Moreover, we define the subspace solution operator Ti : V Vi satisfying
ci (Ti u, vi ) = a(u, Ri vi ).

(6.7)

The solution operator equals to Ti = Ci1 RiT A. Due to the symmetry and the positivedefiniteness of the original bilinear form and the local ones, the operator Ri Ti : V V is
self-adjoint with respect to a(, ).
We define the global preconditioner C : V V by its inverse. In an additive Schwarz
framework this is defined by the additive operator
M
X

C 1 =

Ri Ci1 RiT .

(6.8)

i=0

This yields the preconditioned operator of the form

T := C 1 A =

M
X

Ri Ti .

(6.9)

i=0

A natural choice for the local bilinear form (6.6) is given by the restriction of the original
bilinear form to the local spaces:
ci (ui , vi ) := a(Ri ui , Ri ui ), and Ci = RiT ARi , respectively.

(6.10)

This means that the local problems on Vi are solved exactly, and the operator Ti provides
the a(, )-orthogonal projection onto the subspace Ri Vi V .
The following theorem plays a central role in additive Schwarz theory. A proof can be found
in Zhang [97] (Lemma 4.1).
Theorem 6.1. The additive Schwarz preconditioner is well defined by its inverse. Moreover,
1

the norm generated by the additive Schwarz preconditioner kukC := (Cu, u)V2 is equal to the
splitting norm, i.e the quadratic forms satisfy
(Cu, u)V =
u=

inf
P

u V.

Ci (ui , ui )
R i ui

(6.11)

ui Vi

In particular, this lemma applies for deducing spectral bounds 1 , 2 of the preconditioned
problem, i.e.
Ci (ui , ui ).
(6.12)
Ci (ui , ui ) A(u, u) 2
inf
1
inf
P
P
u=

Ri ui

ui Vi

u=

Ri ui

ui Vi

If only a finite number of subspaces Ri Vi overlap, we obtain 2 = O(1).

118

CHAPTER 6. ITERATIVE SOLVERS

In practice, one prefers the multiplicative Schwarz method (MSM) instead of the additive
one, see Bramble-Zhang [30]. There one defines T := I M
i=0 (I Ti ) instead of (6.9),
but uses the symmetric operator T T . Although one can only show that the symmetric
multiplicative Schwarz method is not worse than the additive one, in most cases it leads to
faster convergence. The additive version corresponds to Block-Jacobi iterations, whereas the
multiplicative one to Block-Gauss-Seidel iterations.
A prominent class of Schwarz preconditioners is provided by multigrid/multilevel preconditioners. These techniques have become very popular in the last decades, since
they lead to optimal convergence rates with respect to the degrees of freedom. Since in
electromagnetics the keypoint is the choice of the correct Schwarz smoother on each level,
we permit ourselves to refer to the literature for the issue of multigrid techniques in the sequel.

6.1.2

A two-level concept for linear equations originating from hp-FEM

discretization

We proceed with establishing a hierarchical two-level concept for the splitting of finite element
spaces. As a start, we introduce the method for H 1 () problems. An analog concept will be
used later for solving H(curl)-problems.
The hierarchical V -E-F -C-based construction of the conforming FE-spaces (cf. Chapter 5)
naturally provides a simple space splitting for the FE-space W H 1 () :
X
X
X
WpKK +1 (Th )
(6.13)
WpFF +1 (Th ) +
WpEE +1 (Th ) +
W = W1,h (Th ) +
EE

F F

KTh

By this, we introduce the following two-level concept for solving the discretized equations:
The lowest-order FE-subspace W1,h plays a special role. We use it as the coarse space,
which carries the global information.
On the coarse level we can exploit the full range of well-developed h-version preconditioners, e.g. multigrid techniques. If the coarse system is of modest size, i.e. if the
underlying mesh is sufficiently coarse, we can also use an exact solver.
On the fine level (the high-order FE-space) we use a Block-Jacobi (ASM) or Gauss-Seidel
(MSM) smoother according to the hierarchic V -E-F -C-based splitting.
The V -E-F -C-based splitting (6.13) provides a natural, simple decomposition. The quality
of the preconditioner can be improved by bigger overlaps of the above local spaces at the
berl et al. [82] we suggest various overlaps
cost of higher complexity. For instance, in Scho
combined with some variants of adding monomials of the lowest-order shape functions and
achieve h and p robust condition numbers for H 1 () problems.
Further on, we can improve the condition number by static condensation of the cell-based
degrees of freedom.

6.1.3

Static Condensation

Static condensation is a customary tool in solving linear equations arising from the hierarchical
hp finite element discretization (cf. the textbooks of Schwab [83] and Melenk [68]). The
main idea is to exploit the local support of the cell-based shape functions. In the following,

6.2. PARAMETER-ROBUST PRECONDITIONING FOR H(curl)

119

we classify two sets of degrees of freedom, or two sets of shape functions, respectively: the
cell-based ones and the remaining ones. The latter we call exterior (ex) degrees of freedom
(exterior shape functions). We can rewrite the linear system (6.2) as


Aex ex Aex C
AC ex ACC



uex
uC

f ex
fC

(6.14)

The local support of the cell-based global FE-functions is only the associated element itself.
Hence, the cell-based degrees of freedom corresponding to two different elements do not couple.
The cell-based submatrix ACC of the stiffness matrix is block-diagonal:


CC
for Ki Th and n = |Th |.
ACC = diag ACC
K1 , . . . , AKn

If we perform the Schur-complement with respect to the submatrix ACC within the linear
system (6.14), we can extract the condensed system
e ex = fe
Au

(6.15)

e := Aex ex Aex C ACC 1 AC ex and fe := f ex Aex C ACC 1 f C .
with A
The cell-based degrees of freedom can eventually be calculated via

1 C


uC = ACC
f AC ex uex .

In a practical implementation, the Schur complement is computed locally on the element

level and only the condensed element matrices are assembled into the global matrix.
The application of static condensation provides two main advantages:
The resulting global condensed linear system is of much smaller dimension: the condensed element matrices are of dimension O(pd1 ) compared to O(pd ) in case of the
original element matrices.
In general, the condensed system matrix is better conditioned. Providing the Schur
complement means an orthogonalization of the external shape functions to the cellbased ones (cf. Karniadakis-Sherwin [60] (Subsection 4.2.3)).
Equipped with the main concepts of preconditioning we turn now to iterative solvers and
preconditioners for H(curl)-problems.

6.2

Parameter-Robust Preconditioning for H(curl)

We consider a conforming finite element space V H0,D (curl, ). By this, we define a finite
dimensional variational problem:
Find u V such that
a (u, v) = f (v)
v V
(6.16)
involving the bilinear form

a (u, v) :=

curl u curl v + u v dx

(6.17)

120

CHAPTER 6. ITERATIVE SOLVERS

and the linear form

f (v) :=

j v dx.

The impressed current density j is assumed to be divergence-free, i.e. div j = 0. Note, that
the restriction to homogenous Dirichlet and homogenous Neumann data is done only for the
sake of simplicity to avoid the technical treatment of boundary integrals and inhomogeneous
Dirichlet data.
Due to the Galerkin isomorphism (cf. Chapter 4) the discrete variational problem (6.16) is
equivalent to the linear system in RN N of the form
with A := K + M RN N
R
R
R
with K ij := curl i curl j dx, M ij := i j dx, and fi := f i dx
for {i : 1 i N } denoting the finite element basis of V .
Au = f

(6.18)

The parameter is problem-dependent, see the discussion in Section 2.4. We call a preconditioner C of A robust in , if the spectral bounds 1 , 2 in
1 C(u, u) a (u, u) 2 C(u, u)
are independent of the parameter .
In the following, we are mainly concerned with the design of -robust preconditioners for
positive and possibly small 1. This corresponds to the following list of problems of
interest:
the magnetostatic problems and non-conducting regions of magneto-quasi-static problems, where the parameter originates from a small regularization parameter (cf. Subsection 3.4.2 and Subsection 6.4.1)
the magneto-quasi-static problem ( = i), where we need good preconditioners for
A = K + ||M (cf. Subsection 6.4.2),
the Maxwell eigenvalue problem, where we need a -robust preconditioners for
A = K + M with > 0 in each iteration step (cf. Chapter 7).
First, we analyze why classical preconditioners fail on the H(curl)-problems. Next, we introduce the successful smoothers of Arnold-Falk-Winther and Hiptmair, which can be chosen
for solving the coarse grid system in a correct and robust way. We use an abstract result
berl [80] on robust preconditioning for parameter dependent problems and reforof Scho
mulate it for H(curl)-problems. We verify that the local exact sequence property implies
parameter-robustness even for simple Schwarz preconditioners.

6.2.1

Two Motivating Examples

We consider the magnetostatic problem with small regularization parameter 0 < 1 on

the lowest-order FE-space V = VhN0 (Th ).
Find u VhN0 (Th ) such that
Z
Z
Z
curl u curl v dx + u v dx =
j v dx
v VhN0 (Th ).
(6.19)

6.2. PARAMETER-ROBUST PRECONDITIONING FOR H(curl)

121

We refer to the parameter-dependent bilinear form with a (, ). In the following, we state

two simple Schwarz preconditioners and their asymptotic behavior as gets small:
1. The standard diagonal Jacobi-preconditioner is provided by the simple space splitting
X
0
span{N
(6.20)
VhN0 (Th ) =
E }.
EE

This yields a parameter-dependent condition number even for the preconditioned system, namely
1
(Cdiag
A ) = O(h2 1 ).
(6.21)
2. The Arnold-Falk-Winther block-Jacobi preconditioner is provided by a space decomposition with respect to the vertex patches (as visualized in the figure for the 2D case):

VhN0 (Th ) =

vV

0
span{N
E : E E so that v on E}.

(6.22)

This choice provides a parameter-independent condition number

1

2
(CAF
W A ) = O(h ).

(6.23)

A numerical experiment
We consider problem (6.19) on the unit cube = [0, 1]3 , which is covered with a mesh of 12
tetrahedra. In Figure 6.1 we illustrate the condition number of the preconditioned system
1
matrix Cdiag
A and C 1 AAF W for varying parameters from 1 to 106 and uniform mesh
refinements in h.
10000
kappa=1e-2
kappa=1e-4
kappa=1e-6

cond(C-1A)

1000

100

10

1
1

10

100

(dofs)1/3

Figure 6.1: Condition numbers of regularized magnetostatic problem, applying a standard

Gauss-Seidel preconditioner (left) and an AFW-block Gauss-Seidel preconditioner (right, robust) for varying regularization parameters with uniform h-refinement.
We observe that in practice, the condition number of the standard Gauss-Seidel-preconditioner
already suffers from modest values of . The AFW smoother is apparently robust in .

122

CHAPTER 6. ITERATIVE SOLVERS

Analysis of the two ASM preconditioners

For analyzing the parameter-dependency of the discussed preconditioners, we have to provide
some inverse estimates first:
Remark 6.2 (Some inverse estimates). We assume a shape-regular triangulation of Rd
(cf. Chapter 4) so that
T
kFK
k  h1
K ,

T
kFK
k  hK ,

and

|JK | hdK .

(6.24)

0
Let {i } and {N
E }, respectively, refer to the conforming nodal FE-basis functions of Wh,1 (Th )
N0
and Vh (Th ), respectively. Assuming conforming transformations we can estimate norms of
the finite element basis functions using (6.24) by

d2
0 2
kN
E kL2 (K)  hK ,

ki k2L2 (K)  hdK ,

d4
0 2
k curl N
E kL2 (K)  hK .

ki k2L2 (K)  hd2

K ,

Concerning the H 1 -conforming and H(curl)-conforming degrees of freedom, we obtain inverse

estimates by conforming transformation to the reference element (hat marker), then performing the estimate and afterwards conforming back-transformation onto the physical element.
Let K be a physical element sharing either the vertex Vi or the edge E.
2



d
2
w(Vi )2 = w(
V )  kwk
2L (K)
 hK kwkL2 (K) ,
2
Z
2  Z
2




2d
2



E d
uE ds = 
s  k
u
uE k2L (K)
 hK kuE kL2 (K) .

2

Moreover, we will need the discrete analog of the Friedrichs inequality (cf. Theorem 3.26) as
stated in Monk [70] (Lemma 7.20)
Lemma 6.3 (Discrete Friedrichs inequality for H(curl)). Let be a simply-connected Lipschitz domain. Suppose v Vh,p (Th ) H(curl, ) is orthogonal to discrete gradient functions,
i.e. (v, ) = 0 for all Wh,p+1 (Th ) H 1 (). Then there exists a constant cF > 0
independent of h such that
kvk0 cF k curl vk0 .
Proof of the condition number estimate (6.21) for the Jacobi-preconditioner.
The quadratic form of the splitting norm corresponding to (6.20) evaluates to the unique
splitting
Z
X
0
uE with uE =
u ds N
u=
E .
E

EE

Let K be an element sharing an edge E. Using the inverse estimates from Remark 6.2, we
obtain
Z
2


N0 2
2
0 2
kuE kA ,K =
u ds
k curl N
E kL2 (K) + kE kL2 (K)
E

()


 h2 + kuk2L2 (K) 

h2

+ 1 k curl uk2L2 (K) + kuk2L2 (K) .

The notation for energy norm k kA ,K indicates that the integrals in the bilinear form a (, )
are evaluated on the physical element K.

6.2. PARAMETER-ROBUST PRECONDITIONING FOR H(curl)

123

In H 1 -problems we can get rid of the -dependency within the step () by Friedrichs
inequality. But in H(curl) Friedrichs inequality only holds for the subspace orthogonal to
gradients.
Summing up over all edges, and taking into account that only a finite number of edge patches
overlap, we obtain
X




1
kuE k2A  2 + 1 k curl uk2L2 (K) + kuk2L2 (K) .
Cdiag (u, u) =
h
EE

Proof of the condition number estimate (6.23) for the AFW-block Jacobi-preconditioner.
The key of the AFW-subspace splitting is that it provides also a correct splitting of gradient
fields in Wh,1 (Th ).
We consider the Helmholtz-decomposition (cf. Theorem 3.14) of u VhN0 (Th ):
u = q + w with w Wh,1 (Th ) H 1 (),

v Wh,1 (Th ) : (q, v)0 = 0.

(6.25)

and analyze the solenoidal and the irrotational field, separately.

Concerning the decomposition of gradient field w, we consider first the
Pdecomposition of w
with respect to the H 1 -conforming shape functions {i }, namely w = vi V w(vi )i . Since
i Vi , we obtain a decomposition of the gradient field w in Vi :
X
wi
with wi = w(vi )i Vi .
w =
vi V

We consider now a physical element K, sharing vertex vi , and use the inverse inequality from
Remark 6.2, to obtain
kwi k2A ,K = kwi k2L2 (K)  h2 kwk2L2 (K) .
Summing up we obtain
X

vi V

()

kwi k2A  h2 kwk2L2 ()  h2 kuk2L2 ()  h2 kuk2A .

(6.26)

Note that we have used Friedrichs inequality for H 1 () in () and have exploited the
L2 -orthogonality of the decomposition in the sense of kuk20 = kqk20 + kwk20 .
Now, we consider the solenoidal part q (Wh , 1) with curl q = curl u. Since the decompo0
sition into Nedelec basis functions N
provides a finer decomposition of the AFW-splitting,
i
we can use the estimate from the proof of the Jacobi-preconditioner. But here, we can use
Friedrichs inequality
kqk0  k curl qk = k curl uk.
(6.27)
in the estimate (), which yields
X


kq i k2A  h2 + kqk2L2 ()  h2 k curl uk20 + kuk20
vi V

 h2 + 1 k curl uk20 + kuk20  h2 + 1 kuk2A .

(6.28)

124

CHAPTER 6. ITERATIVE SOLVERS

Concluding, we consider the decomposition ui = q i + wi and combine the two estimates

(6.26) and (6.28). This yields
inf
P

u=

ui

kui k2A  h2 + 1 kuk2A .

The AFW-decomposition provides a correct space splitting of gradient fields in the following
sense
X
Wh,1 Vi = Wh,1 .
vi V

The h2 -dependency of condition number can be tackled by multigrid methods.

6.2.2

The smoothers of Arnold-Falk-Winther and Hiptmair

The two preceeding examples demonstrate the importance and necessity of applying an
appropriate smoother for H(curl)-problems. The operator scales differently on the solenoidal
and the irrotational fields. Appropriate smoothers can be designed by means of a Schwarz
splitting of the underlying Nedelec FE-space, which respects the Helmholtz-decomposition.
Only then can associated multigrid/multilevel methods work efficiently.
We assume an H(curl)-conforming Nedelec space Vh,p (Th ) of uniform order p, and an appropriate H 1 -conforming FE-space Wh,p+1 (Th ), which satisfy the global exact sequence property,
i.e. Wh,p+1 (Th ) Vh,p (Th ).
Up to the authors knowledge, there are two established and commonly used techniques for
designing appropriate smoothers which can be defined using the discussed space splittings
within the Schwarz framework:
The Arnold-Falk-Winther (AFW) smoother is based on an overlapping space splitting
according to the vertex patches:
Vh,p =

vV

Vv

with

Vv := {u V : supp(u) v },

where v denotes the vertex patch as defined in (5.38) (cf.

Winther [7]).

(6.29)

Arnold-Falk-

The space splitting suggested by Hiptmair explicitly respects the Helmholtz decomposition:
X
X
W v
(6.30)
VE +
Vh,p =
EE

vV

with the subspaces W v := {w Wh,p+1 : supp{w} v } corresponding to the vertexpatch v as defined in (5.38) and V E := {v Vh,p : supp{v} E } corresponding to
the edge-patch E .
The smoothing iteration corresponding to the explicit splitting of gradient fields can be
performed by a solver on the scalar potential problem. For a detailed description see
Hiptmair-Toselli [58].

6.2. PARAMETER-ROBUST PRECONDITIONING FOR H(curl)

125

Unfortunately, we observe that as the polynomial degree p increases, these local spaces can
grow quite large. For instance, the suggested vertex-patch space involved in the AFW-splitting
is spanned by the following shape functions

M
M
M
 N0 
C
F
E
v
V =

Em E : v on Em

span Em
VpEm
m

Fm F : v on Fm

VpFmm

VpC .

KTh : v on K

At this point, we can exploit the local exact sequence property of the shape functions discussed
in Chapter 5, which ensures parameter-robust preconditioning even for simple space splittings.
Toward this goal, we will need a more general result on Schwarz preconditioners for parameterdependent problems.

6.2.3

Schwarz methods for parameter-dependent problems

berl [80] (Theorem

In the following, we present a key result stated and proved in Scho
4.2), which provides a general framework for defining robust Schwarz preconditioners driven
by parameter dependent problems in structural mechanics. In advance, we remark that we
present a simplified version of this theorem with Q = L2 () and k kQ,0 = k kc = k
k0 on curl V , cooresponding to the problem setting in H(curl)).

Theorem 6.4 (Robust one-level ASM preconditioners for parameter-dependent problems).

We consider a finite-dimensional Hilbert space V . Let A : V V R be a parameterdependent coercive bilinear form
A (u, v) = a(u, v) + 1 (u, v)L2

(6.31)

with
an energy norm, which for = 1 is equivalent to the norm of the underlying space, i.e.
k kA1 k kV ,
a continuous linear operator

(6.32)

: V L2 ()

with a non-trivial nullspace V0 := ker and that satisfies the inverse estimate
q V v V :

kvkV c3 (h)kqk0 .

(6.33)

Let {Vi } be a local space splitting with finite overlap such that the kernel V0 can be split locally
into kernel functions, i.e.
V =

M
X

Vi

and

i=0

V0 =

M
X
i=0

(Vi V0 ).

(6.34)

Furthermore, we assume that the local splittings of the functions u V and u0 V0 can be
estimated in k kV as follows
u=

inf
P

u0 =

M
X

u0,i i=0

u0,i Vi V0

kui k2V c1 (h)kuk2V ,

(6.35)

ku0,i k2a c2 (h)ku0 k2V .

(6.36)

ui i=0

ui Vi

inf
P

M
X

126

CHAPTER 6. ITERATIVE SOLVERS

Then the additive Schwarz preconditioner C built on the space splitting {Vi } fulfills
(c1 (h) + c2 (h)c3 (h)2 )1 C(u, u)  A (u, u) No C(u, u)

u Vh,p(Th )

(6.37)

with bounds c1 (h) and c2 (h) independent of the parameter 0 <  1. The constant No denotes
the maximal number of overlapping subspaces.
The main achievement of this theorem is that for a subspace splitting which provides (6.34)
we obtain parameter-robustness and that the spectral bounds of the preconditioned problem
hold with the estimates of the splitting norms, separately for u and the kernel functions u0
for = 1.
We set := curl. Then (6.33) holds by choosing v such that q = curl v with (v, w) = 0, w
and using the discrete Friedrichs inequality for H(curl)
kvk20 cF (p)k curl vk20

v (V ker(curl)) .

(6.38)

Hence, we can directly deduce a result on parameter-robust preconditioners for the variational
problem (6.20) in H(curl).
Corollary 6.5 (Parameter-robust ASM preconditioner for H(curl)).
Let Vh,p(Th ) H(curl, ) denote a Nedelec FE-space of order p and Wh,p+1 (Th ) H 1 ()
an appropriate scalar FE-space of order p + 1 with ker(curl, Vh,p(Th )) = Wh,p+1 (Th ) and
Wh,p+1 (Th ) Vh,p(Th ).
We consider following subspace splitting of the FE-spaces, assuming finite overlap:
Wh,p(Th ) =

M
W
X

Wj

and

Vh,p(Th ) =

j=0

M
X

Vi .

i=0

If
Wj Vi : Wj Vi

(6.39)

and if the local splitting of the functions in Vh,p and the involved gradient fields provide the
estimates

u=

inf
P

ui i=0

u i Vi

inf
P

w=

M
X

wj j=0

wj Wj

kui k2H(curl) c1 (h, p)kuk2H(curl) ,

(6.40)

MW kwj k20 c2 (h, p)kwk20 ,

(6.41)

then the additive Schwarz preconditioner C built on the space splitting {Vi }, applied to the
parameter-dependent curl-curl-problem (6.16), is robust in :
(c1 (h, p) + c2 (h, p)(1 + cF )2 )1 C(v, v)  A (v, v) No C(v, v)

v V.

(6.42)

with cF the constant from the discrete Friedrichs inequality (6.38) is independent of h and p
(see Demkowicz-Buffa [43] and Gopalakrishnan-Demkowicz [50]).

6.2. PARAMETER-ROBUST PRECONDITIONING FOR H(curl)

6.2.4

127

Parameter-robust ASM methods and the local exact sequence property

In the following, we apply Corollary 6.5 first for general H 1 -conforming and H(curl)conforming FE-spaces satisfying the local exact sequence property and next for FE-spaces
based on the finite element shape functions presented in Chapter 5.
Corollary 6.6.
1. Let the FE-spaces Wh,p+1 (Th ) H 1 () and Vh,p(Th ) H(curl, ) satisfy the local exact
sequence property (Theorem 5.32). Then any space splitting {Vi } of the Nedelec space
Vh,p(Th )} built on the (possibly finer) N0 -E-F -C-based splitting
Vh,p = VhN0 (Th ) +

EE

VpEE +1 (Th ) +

F F

VpFF +1 (Th ) +

VpKK +1 (Th )

(6.43)

KTh

fulfills (6.39) and specifies a parameter-robust ASM preconditioner.

2. Let Wh,p+1 (Th ) H 1 () be defined by (5.39) and Vh,p(Th ) H(curl, ) defined by
(5.39), i.e. the Nedelec FE-basis explicitly contains the gradients of the
Phigher-order basis functions of the scalar FE-basis. Then any Nedelec space splitting i Vi = Vh,p(Th )}
which is built on a two-level concept, where the lowest-order space VhN0 (Th ) is either
solved exactly or is split corresponding to the Arnold-Falk-Winther splitting (6.29) or
to the one of Hitpmair (6.30) fulfills (6.39) of Corrollary 6.5 and implies a parameterrobust ASM preconditioner.
Of course, the choice of the local subspaces influences the condition number through h and p
dependency. But, a splitting based on the FE-basis, suggested in Chapter 5, automatically
respects the Helmholtz-decomposition for higher-order FE-subspaces. To achieve parameterrobustness we only have to care about the correct splitting of lowest-order space within the
design of appropriate smoothers.
The following numerical experiment verifies the -robustness using the H(curl)-conforming
FE-basis of Chapter 5.
Numerical experiments
R
R
We consider A (u, v) = curl u curl v dx + u v dx on the unit cube = [0, 1]3 , which
is covered by a triangulation of 12 tetrahedra. We use a Schwarz preconditioner with the
N0 -E-F -C-based space splitting (6.43). On the lowest-order space VhN0 (Th ) we use a direct
solver.
We run the experiment for two parameters = 1 and = 106 . Alternatively, we apply static
condensation of the cell-based degrees of freedom.
Due to the local exact sequence property the space splitting recovers the gradient fields in the
sense of (6.39). As stated by Corollary 6.5, we can observe in Figure 6.2 that the condition
number is not affected by the choice of the parameter .

128

CHAPTER 6. ITERATIVE SOLVERS

800

800

-6

kappa=10 -6
stat.cond., kappa=10

700

700

600

600

500

500
kappa

kappa

kappa=1
stat.cond., kappa=1

400

400

300

300

200

200

100

100

0
0

10

12

10

12

Figure 6.2: Robustness of condition numbers of parameter-dependent curl curl problem on

the unit cube for = 1 (left) and = 106 (right)

6.3

Reduced N
ed
elec Basis and Special Gauging Strategies

In this section, we review the gauging strategy of magnetostatic problems, which also apply in
non-conducting regions of magneto-quasi-static problems. We start with the initial variational
formulation: Find u V H(curl) such that
Z
Z
curl u curl v dx =
jv dx, v V.

The solution of the curl curl problem is determined up to gradients. In the following, we
suggest a practical method for gauging high-order gradient functions.
The idea is simple and practical convenient, if we assume the H(curl)-conforming finite element
basis, suggested in Chapter 5, which explicitly involves (higher-order) gradient functions:
The higher-order gradients can be gauged by locally skipping the corresponding degrees
of freedom and basis functions in the higher-order edge-based, face-based and cell-based
finite element subspaces.
On the lowest-order subspace, we follow the former strategy of adding a regularization
term.
The reduced N
ed
elec basis Following Subsection 5.2.7 we introduce the reduced local
FE-space on the element K as
Vpred (K) := V N0 (K)

Fm FK

(K) VpC,3
(K).
VpFFmm,2 (K) VpFFmm,3 (K) VpC,2
C
C

(6.44)

red (T ).
and the corresponding reduced global finite space Vhp
h
Here, we have skipped the higher-order gradient functions spanning the local subspaces
(K), VpFFmm,1 (K), and V C,1 (K)). We remark that, in general, the reduced finite element
VpEEm
m
space is not orthogonal to higher-order gradient fields. Moreover, the skipping of gradient
functions can be done separately on each edge, face and cell, hence we can perform reduced
basis gauging only on single subdomains as required for magneto-quasi-static problems involving subdomains corresponding to non-conducting materials.

129

6.4. NUMERICAL RESULTS

Gauging on the lowest-order space Gauging with respect to lowest-order shape functions is performed by adding a zero-order term with small regularization parameter. Hence,
in practice, it is sufficient to apply only a low-order integration formula within the integration
of the element mass matrix denoted by (, )L2 (),red .
Note, that for gauging in lowest-order spaces, the spanning tree technique, which is based on a
graph theoretic treatment (cf. Bossavit [25], Section 5.3), provides an alternative. But, since
we are equipped with -robust preconditioners gauging by regularization is more suitable, in
particular from the implementational point of view.
The discrete variational problem on the reduced FE-space Hence, for magnetostatic
problem or for non-conducting regions of magneto-quasi-static problems we state the discrete
problem in the reduced FE-basis with a regularization parameter 0 < 1 as follows.
red (T ) such that
Find u Vh,p
h
Z

curl u curl v dx + (u, v)L2 (),red =

jv dx,

red
v Vh,p
(Th ).

(6.45)

We remark that the resulting magnetic induction field B = curl u, in which we are mainly
interested, is unchanged by the choice of the full FE-basis or the reduced FE-basis.
We observe two main advantages in using the reduced FE-basis:
More than one third of degrees of freedom can be saved (for p 1), which results in a
decrease of the non-zero entries in the system matrix to 55 %.
As to be observed in the following numerical experiments, the condition numbers highly
improves within the reduced FE-basis setting.
Numerical experiment Again, we consider the unit cube covered with a mesh of 6 tetra1
hedra. In Figure 6.3 we present the condition numbers of the preconditioned system Cred
Ared
computed in the reduced bases to the ones computed in the full FE-basis. In both cases,
we use N0 -E-F -C-based multiplicative Schwarz preconditioners using a direct solver on the
coarse level. We run our tests for uniform p-refinement and choose the regularization term
by = 106 . Optionally, we use static condensation of the cell-based degrees of freedom. We
observe a considerable improvement of the condition numbers in case of the reduced basis
gauging.

6.4

Numerical Results

In the following, we are concerned with the numerical evidence of -robust preconditioning
and reduced basis gauging within more involved problems.
We choose a two-level Schwarz solver analogous to the one stated in (6.1.2), which is based
on the local N0 -E-F -C-based splitting and uses either exact solvers on the coarse level, i.e.
on the space of lowest-order Nedelec functions, or h-version multigrid methods with ArnoldFalk-Winther smoothers, if the problem size of the coarse system increases.

130

CHAPTER 6. ITERATIVE SOLVERS

1000
full basis
reduced basis
full basis, stat. cond.
reduced basis, stat. cond.

-1

cond(A C)

100

10

1
0

6
p

10

12

Figure 6.3: Condition numbers using full basis vs. reduced basis on unit cube with/without
static condensation

6.4.1

The magnetostatic problem

We consider the magnetostatic boundary value problem

curl 1 curl A = j,
where j denotes the given current density with div j = 0, and A is the vector potential of the
magnetic flux B = curl A. For gauging, we add a small zero-order term (with = 106 1 )
to the operator. In order to show the performance of the constructed shape functions we
compute the magnetic field induced by a cylindrical coil.
berl [79]) contains 2029 curved tetrahedral
The mesh generated by NETGEN (cf. Scho
elements. Figures 6.4 and 6.5, show the magnetic field lines and the absolute value of the
magnetic flux simulated with H(curl)-elements of uniform degree p = 6.
We have chosen various uniform polynomial degrees p, and compared the number of unknowns
(dofs), the condition numbers of the preconditioned system, the required iteration numbers
of the PCG for an error reduction by 109 , and the required computation time on a Dual
Processor Intel Xeon 64Bit 2,8 GHz machine. We run the experiments once keeping the
gradient shape functions, and a second time skipping them. The computed B-field is the
same for both versions. One can observe a considerable improvement of the required solver
time in the Table 6.1.
We run the experiments a third and a fourth time, but extend the solution algorithm by
static condensation of the cell-based degrees of freedom. Again, one observes a considerable
improvement, when using the reduced basis (cf. Table 6.2).

6.4. NUMERICAL RESULTS

131

Figure 6.4: Magnetic field induced by the coil, order p=6.

Figure 6.5: Absolute value of the magnetic flux B = curl A (zoom to coil), order p=6.

132

CHAPTER 6. ITERATIVE SOLVERS

p
2
2
3
3
4
4
5
5
6
6
7
7
8
8

dofs
19707
10676
50820
29084
104350
61744
186384
112714
303009
186052
460312
285816
664380
416064

grads
yes
no
yes
no
yes
no
yes
no
yes
no
yes
no
yes
no

(C 1 A)
16.99
6.64
46.17
15.48
102.07
26.27
210.34
43.52
406.11
67.48
751.06
95.08
1033.05
128.61

iter
31
23
49
34
69
44
96
54
127
63
171
72
201
82

solvertime
1.5 s
0.4 s
6.6 s
1.6 s
27.8 s
5.5 s
94.8 s
16.7 s
314.0 s
43.5 s
923.3 s
106.6 s
2080.8 s
272.4 s

Table 6.1: Performance of solver: full vs. reduced basis (no static condensation)

p
2
2
3
3
4
4
5
5
6
6
7
7
8
8

dofs
19707
10676
50820
29084
104350
61744
186384
112714
303009
186052
460312
285816
664380
416064

grads
yes
no
yes
no
yes
no
yes
no
yes
no
yes
no
yes
no

(C 1 A)
16.99
6.64
42.81
13.79
79.86
21.01
135.70
27.45
207.02
33.33
295.79
38.63
398.03
43.38

iter
31
23
46
32
62
39
78
44
91
48
103
51
114
54

solvertime
1.2 s
0.4 s
5.2 s
1.3 s
20.2 s
3.2 s
51.5 s
6.9 s
120.3 s
13.5 s
244.5 s
24.2 s
430.1 s
41.7 s

Table 6.2: Performance of solver: full vs. reduced basis with static condensation

133

6.4. NUMERICAL RESULTS

6.4.2

The magneto quasi-static problem: A practical application

The intention of this section is to verify the applicability of the constructed FE-basis, the
parameter-robust preconditioners and the reduced basis gauging (in non-conducting regions)
in a practical industrial application. We consider the industrial problem setting of computing
the eddy currents within a bus bar, a common device in transformers.
We consider the magneto-quasi-static problems, i.e., Problem 3.32 Find u V :=
H0,D (curl, ) satisfying
Z
Z
Z
1
dx +
dx =
dx
curl u curl v
u v
jv
v V.
(6.46)

with = i with given frequency and given conductivity coefficient 0.

In non-conducting domains we perform gauging by a small regularization parameter
= 1 .
A solver strategy for the magneto-quasi-static problem
We briefly outline the chosen strategy for solving the magneto-quasi-static system:
In non-conducting domains, e.g. air, where = due to regularization or 1 we
can apply the reduced basis gauging strategy. We simply skip the higher-order gradients
in the FE-basis on edges, faces and cells belonging to non-conducting domains.
For solving A u = f with K + M we use a complex PCG method, i.e. PCG with
complex-valued matrices and a complex-valued search parameter.
The preconditioner for A = K + M is chosen to be a -robust preconditioner of the
system A = K + ||M .
A detailed description and justification of the solution algorithm provided by the second and
the third item can be found in Bachinger et al. [14], [13]. Concerning the solution of
symmetric complex-valued systems we further refer to Clemens-van Rienen-Weiland [36]
and Clemens-Weiland [37].
Numerical Simulation of Eddy-currents in a bus bar
We want to simulate the eddy currents which are originated in a bus bar of a power transformer.
Since we are only interested in simulating the eddy currents in the bus bar we separate it
from the transformer by cutting the bars. Hence, we have to construct artificially a current
distribution in the non-resolved domain, which models the current distribution of the original
problem. This can be done by solving a scalar Neumann problem where the Neumann data
is specified by the prescribed normal currents on the cutting planes of the bars and vanishing
on the remaining boudary. This results in a current density j in the outer domain, which we
use in the right hand side of the magneto-quasi-static problem (6.46). Note that the obtained
current density is not divergence-free.
Figure 6.6 shows the eddy-currents in the bus bar of a transformer for polynomial order p = 3.
In the air domain we use the reduced basis with polynomial order p = 3. The mesh consists

134

CHAPTER 6. ITERATIVE SOLVERS

Figure 6.6: Eddy currents in a bus bar

p
2
2
3
3

partially reduced basis

no
yes
no
yes

dofs
249297
193502
645932
503601

solver time
968 s
768 s
4824 s
3682 s

Table 6.3: Comparison of degrees of freedom and solver times by using a partially reduced
basis for the bus bar eddy-current problem
of 26094 tetrahedral elements.
In Table 6.3 we compare the numbers of degrees of freedom and the solution time of the
magneto-quasi-static problem running the experiment once using the full FE-basis and a
second time using the partially reduced basis gauging in the non-conducting outer domain.

Chapter 7

The Maxwell Eigenvalue Problem

The numerical solution of eigenvalue problems is in general more challenging than the one of
boundary value problems. Eigenvalue problems for Maxwells equations involve even further
difficulties, in particular due to large kernels of the involved operators, e.g., the curl-operator.
The discretization with H(curl)-conforming finite elements is highly important in order
to avoid spurious (non-physical) eigenvalues within the computed spectrum. Moreover,
the involved H 1 -conforming and H(curl)-conforming FE-spaces have to satisfy the exact
sequence property.
The numerical analysis of Maxwell eigenvalue problems (cf. e.g. Boffi et al [20], Monk et
al. [71], Caorsi et al. [34], Boffi et al. [21],[19], and Kikuchi [61]) and preconditioned
eigensolvers (cf. e.g. Knyazev-Neymeyr [64], Knyazev [62], and Neymeyr [74]) is
very challenging and subject of ongoing research. A rigorous analysis is available only for
subproblems, and many approaches are justified mainly by numerical evidence.
The aim of this section is to present a reasonable approach for attacking Maxwell eigenvalue
problems. We will see that the local exact sequence property, which holds for our construction
of conforming FE-spaces, has certain advantages in the solution of the eigenvalue problems,
in particular in efficient preconditioning. We will finally demonstrate the performance of the
proposed algorithm by solving a benchmark problem involving highly-singular eigensolutions.
For the numerical solution we will also use anisotropic hp-discretizations and geometric mesh
refinement.

7.1

Formulation of the Maxwell Eigenvalue Problem

An important problem in electromagnetics is the simulation of bounded cavities, which are

characterized by
a bounded domain filled with non-conducting dielectric materials ( = 0, = 0) with
strictly positive, piecewise constants and ,
surrounding perfectly conducting materials, e.g. metallic surfaces.
By reflection at the perfectly conducting walls, standing waves can be excited within the cavity.
Therefore, even with homogenous boundary conditions and no sources in form of charges or
currents, non-trivial solutions arise at special, so-called resonant frequencies. The problem

135

136

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

of determining resonant frequencies 6= 0 and the corresponding eigenmodes satisfying the

time-harmonic Maxwell equations is the following:
Find the electric field E, the magnetic field H and resonant frequencies > 0 satisfying
curl E + iH = 0,

(7.1a)

curl H iE = 0,

(7.1b)

div(H) = 0,

(7.1c)

div(E) = 0.

(7.1d)

For simplicity, we assume the material parameters and to be strictly positive scalars.
The surrounding perfect electric conductor yields homogenous Dirichlet boundary condition
for the tangential components of the electric field, which in turn corresponds to homogenous
Neumann condition for the magnetic field. By eliminating either the magnetic or the electric
field, we can reduce the system (7.1) to one of the following problems:
Find resonant frequencies 6= 0 and the related elec-

The electric eigenvalue problem:

tric fields E such that

curl 1 curl E = 2 E

in ,

(7.2a)

div E = 0

in ,

(7.2b)

En=0

on .

(7.2c)

The magnetic eigenvalue problem:

that

Find 6= 0 and the related magnetic fields H such

curl 1 curl H = 2 H
div H = 0
1

curl H n) n = 0

in ,

(7.3a)

in ,

(7.3b)

on .

(7.3c)

In the following, we focus only on the electric formulation, but similar considerations also
apply to the magnetic case.
Variational formulation of the eigenvalue problem:
Problem 7.1. Find 6= 0 and u H0 (curl, ) such that
Z
Z
1
curl u curl v dx = u vdx
v H0 (curl, ).

(7.4)

Note that this formulation does not explicitly involve the divergence conditions (7.2b) and
(7.3b). But testing (7.4) with gradient fields v = H 1 () yields (u, )0 . Hence,
eigenvectors corresponding to non-zero eigenvalues are L2 -orthogonal to gradient fields.
We summarize the main properties of the electric eigenvalue problem (cf. Monk [70] Theorem
4.18):
1. The eigenvalue = 0 (which is excluded in our definition of the problem) corresponds
to the infinite space of gradient functions H 1 ().

137

7.1. FORMULATION OF THE MAXWELL EIGENVALUE PROBLEM

2. The discrete set of eigenvalues i 6= 0 correspond to eigenfunctions ui with div ui = 0

solving the eigenvalue problem (7.4). The eigenvalues form an increasing sequence
0 < 1 2 ,
where every i > 0 has finite multiplicity.

3
We want to mention that H(curl, ) is not compactly embedded in L2 () , which makes
the statement on the behavior of the

3 spectrum difficult. On the other hand, the embedding
of H0 (curl, ) ker(div) in L2 () is compact, and standard results apply. This already
indicates that kernel functions will play an important role in the consideration of Maxwell
eigenvalue problems.
Galerkin approximation of the eigenvalue problem: We choose an H(curl)conforming FE-space Vh,p(Th ) H0 (curl, ), and formulate the discrete Maxwell eigenvalue
problem as follows:
Find h > 0 and uh Vh,p(Th ) such that
Z
Z
1
uh v h dx
curl uh curl v h dx = h

v Vh,p(Th ).

(7.5)

There is an ongoing debate on how to involve the divergence condition in the solution
of the eigenvalue problem. Here, we address this point by excluding the zero eigenvalues
within the definition of the eigenvalue problem, and projection onto the complement of
gradient fields. Alternatively, one of the following approaches could be used: utilizing a
mixed-formulation as introduced in Boffi [21]; introduction of Lagrange-Multiplier, cf.
Kikuchi [61]; adding a term which penalizes the divergence as suggested in Wang et al. [93].
Before we discuss numerical schemes in more detail, we make some brief remarks concerning
the convergence analysis of FE-discretizations of Maxwell eigenvalue problems, which is still
fairly incomplete:
A Galerkin approximation is called spectrally correct (cf. Caorsi et al. [34]), if all discrete
eigenvalues of (7.5) converge to exact eigenvalues of (7.5), and vice versa, all exact eigenvalues
are approximated by discrete ones respecting their multiplicity.
By the discrete compactness of Galerkin approximations, a convergence proof for h-version
FEM, including edge-elements of general order, is given in Monk et al. [71]. Concerning general hp-discretizations in three dimensions, the required discrete compactness of FEapproximations is not fully proved yet, up to our knowledge.
The algebraic Maxwell eigenvalue problem: By means of the Galerkin isomorphism
we deduce the generalized algebraic eigenvalue problem equivalent to (7.5):
Find > 0 and u RN s.t.

Au = M u.

(7.6)

The matrices are definedPby Aij = (curl i , curl j )0 and Mij = (i , j )0 , and the solution
vector u is given by u = ui i , where {i : 1 i N } is a FE-basis of Vh,p(Th ). We obtain
that A RN N is symmetric positive semi definite but has a large nullspace associated with
the gradient fields, and M RN N is symmetric positive definite.

138

7.2

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

Preconditioned Eigensolvers

System matrices originating from Finite Element discretization of second-order partial differential equations are typically of huge size and ill-conditioned. Due to the large dimension,
classical eigensolvers which rely on factorizations of the system matrices can therefore not be
applied. On the other hand, also classical matrix-free iterative schemes cannot be used for
an efficient solution, due to the ill-conditioned matrices. We refer to Bai et al. [15] for an
overview of the full range of classical and iterative eigensolvers.
In recent years, preconditioned eigensolvers have attracted significant interest, particularly
for the solution of elliptic eigenvalue problems; we refer to Dyakonov [47], Knyazev [62]
and Neymeyr [74] in this context. Note however, that even for symmetric positive definite
generalized eigenvalue problems, sharp non-asymptotic convergence estimates are only available for the simplest iterative scheme, i.e., for the preconditioned inverse iteration, which will
be defined below.
In the following, we focus on gradient type preconditioned eigensolvers, which offer some
advantages:
The implementation can be matrix-free, since they only use matrix-vector products of
the system matrices or the preconditioner.
They involve and can exploit the same preconditioners as used for solving the corresponding linear systems.
The iterative schemes are simple and easy to implement.
We start with considering symmetric positive-definite eigenvalue problems originating, e.g.,
from self-adjoint elliptic partial differential equations, and then turn back to the Maxwell
problems we have in mind.
The symmetric positive definite generalized eigenvalue problem: Let A and M be
real, symmetric, positive definite N -by-N matrices. We consider the following (algebraic)
eigenvalue problem:
Find an eigenvalue R and a non-zero eigenvector u RN such that
Au = M u.

(7.7)

We briefly recall some basic results:

1. The generalized eigenvalue problem (7.7) admits N positive eigenvalues i with arbitrary
multiplicity such that
0 < 1 2 N .
(7.8)
The corresponding eigenvectors ui can be chosen to be mutually M -orthogonal, i.e.
uTi M uj = ij , with uTi Auj = i ij .
2. All stationary points ui of the Rayleigh quotient
(u) :=

uT Au
uT M u

for u 6= 0,

are eigenvectors, and i = (ui ) the corresponding eigenvalues.

(7.9)

7.2. PRECONDITIONED EIGENSOLVERS

7.2.1

139

Preconditioned gradient type methods

Gradient type methods for eigenvalue problems are based

on the idea of minimizing the
Rayleigh-quotient. Since (u) = uT 2M u Au (u)M u) , the resulting iterative scheme is
given by


un+1 = un Aun (un )M un .
(7.10)

The optimal a-priori choice for is such
that the Rayleigh quotient is minimized, i.e. opt :=
arg min >0 un (Aun (un )M un ) .
We assume that our matrices arise from FE-discretization of second order differential operators, and hence we have to deal with the case that A is ill-conditioned by implementing an
appropriate preconditioner. Let C RN N be a symmetric positive definite preconditioner
for the matrix A satisfying
1 uT Cu uT Au 2 uT Cu,
with spectral bounds 1 , 2 (0, 1]. A preconditioned gradient type method then has the form


ui+1 = ui C 1 Aui (ui )M ui .
(7.11)

We shortly comment on the choice of the scaling parameter :

The preconditioned inverse iteration is defined by the choice := 1.

The preconditioned steepest descent method is defined by choosing the optimal scaling
parameter opt in the following way. We define the new iterate un+1 as the minimizer of
the Rayleigh-quotient on the two-dimensional subspace



i
i
1
i
span u , C
A (u )M u .
(7.12)
This can be done via the Rayleigh-Ritz subspace method, which requires the solution
of the following 2-by-2 projected eigenvalue problem:
Find the smallest Ritz-value min > 0 and the corresponding Ritz-vector
ymin R2 solving


X T AXy = X T M Xy
with X = ui , C 1 (A (ui )M )ui RN 2 .
(7.13)
By setting ui+1 = Xymin we obtain an optimal subspace iteration in the sense that we
implicitly chose the optimal parameter = opt in (7.11).
The following estimate for the preconditioned steepest descent method has been obtained in
Knyazev-Neymeyr [65] via the ones for preconditioned inverse iteration.
Theorem 7.2. Let 1 ... N denote the the discrete spectrum of the generalized EVP
(7.7) and assume that (ui ) [k , k+1 ). Then (ui+1 ) computed by iterate (7.11) of the
steepest descent method satisfies either (ui+1 ) < k , or (ui+1 ) [k , (ui )). In the latter
case, there holds
i
(ui+1 ) k
2 (u ) k

q
k+1 (ui+1 )
k+1 (ui )
with convergence factor


k
2


.
1
q := 1
k+1
C 1 A + 1

140

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

In the following, we expand the local space (7.12) in order to improve the iterative scheme.
The locally optimal preconditioned conjugate gradient method was introduced in
Knyazev [63]. Motivated by the preconditioned conjugate gradient method for solving linear
systems, Knyazev suggested to enlarge the local space (7.12) with the approximate eigenvector
from the previous iteration. The new eigenvector approximation ui+1 is chosen to minimize
the Rayleigh-quotient within the three-dimensional subspace



i
i1 i
1
i
span u , u , C
A (u )M u .

Since the previous iterate ui1 and the current iterate ui get closer and closer as the iteration
proceeds, the Gram-matrices of the Rayleigh-Ritz eigenvalue problem can get more and more
ill-conditioned. We address this fact by utilizing an auxiliary vector pi instead of the previous
iterate ui1 : Let (min , ymin = (y1 , y2 , y3 )T ) R+ R3 denote the minimal Ritz-pair of


X T AXy = X T M Xy
with X = ui , wi , pi .
(7.14)
Then, we choose

pi+1 = diag(0, 1, 1)Xymin = ui+1 y1 ui .

This approach yields the same local space, i.e.,





i

i
i+1
i1 i
1
i
i i
1
i
u
span u , u , C
A (u )M u = span p , u , C
A (u )M u . (7.15)

Numerical evidence, cf. Knyazev [63], suggests that the ill-conditioning of the Gram-matrices
is partially resolved in this way. Otherwise, we apply an A-orthogonalization step first. Concerning initial values for the iteration, we start with some vector u0 and perform the steepest
descent method to obtain an initial vector for p1 .
We want to note that the only convergence result for the LOPCG method is that the method
converges at least as fast as the preconditioned gradient methods (7.11), i.e., the bounds of
Theorem 7.2 apply. However, besides the Jacobi-Davidson method, the block version of the
LOPCG algorithm is accepted to be one of the most efficient preconditioned eigenvalue solvers
for symmetric-positive definite generalized EVPs arising from FEM-discretization. We refer
to Knyazev [63] and Arbenz-Geus [5] for detailed numerical results.

7.2.2

Block version of the locally optimal preconditioned gradient method

Block methods (also called subspace iterations) enable us to compute groups of eigenvalues
within a special part of the spectrum simultaneously, e.g., the smallest eigenvalues and corresponding eigenvectors. The Rayleigh-Ritz method can easily be extended to a multiple-vector
iterative scheme. The main advantages of iterating sets of vectors simultaneously are the following. First, multiple eigenvalues can be calculated in a correct and stable way. Secondly, we
can expect much faster convergence, since the error estimates of subspace methods involve the
ratio i /m+1 (m is the subspace dimension and i m) instead of 1 /2 , see e.g. Bramble
et al. [28].
The basic algorithm for the LOBPCG method
Starting values:
A set of m 1 initial vectors U 0 = [u01 , ..., u0m ] RN m .

141

7.3. PRECONDITIONED EIGENSOLVERS FOR THE MAXWELL PROBLEM

Perform one step of (7.13) to obtain U 1 = Xymin and P 1 = diag(0, 1)Xymin RN n .

LOBPCG-iteration: for i = 1, , m
1. Compute the preconditioned residuals
W i = C 1 (AU i M U i i )


with i = diag((ui1 ), . . . , (uiN ) .
2. Solve the 3m-by-3m Rayleigh-Ritz-problem with y R3m
X T AXy = X T M Xy

with X = [U i , W i , P i ].

3. Take the m smallest Ritz-Pairs (1 , Xy1 ), ..., (m , Xym ) and set

U i+1 = [Xy1 , ..., Xym ]

7.3

and



P i+1 = diag(0, 1, 1)Xy1 , ..., diag(0, 1, 1)Xym .

Preconditioned Eigensolvers for the Maxwell Problem

We intend to apply the LOBPCG method to the solution of the Maxwell eigenvalue problem
(7.6). Note that, similar to iterative linear system solvers, the large kernel of the curl-operator
makes the problem more difficult:
1. The curl-curl system matrix A is only positive-semidefinite. Apparently, we cannot
apply the algorithms presented above directly.
Solution: We shift the spectrum of the problem by some positive value and arrive at
the following positive-definite eigenvalue problem:




A + M u = + M u.

(7.16)

The corresponding gradient type iteration reads

1
ui+1 = ui CA+M
A + M ((ui ) + )M ui


1
= ui CA+M
A (ui )M ui .

(7.17)

Au
denoting the Rayleigh-quotient of the original problem.
with (u) = uuT M
uT
Due to the results of Chapter 6, we already have -robust preconditioners for A + M
at hand.

2. Iterative scheme, like the gradient type methods, tend to recover the eigenvectors corresponding to the small eigenvalues first. This affects the positive-semidefinite EVP (7.6)
as well as the shifted problem (7.16).
Solution: By projecting the iterates ui+1 into the complement of the kernel functions,
this problem can be resolved, cf. Hiptmair-Neymeyr [57] for the preconditioned inverse iteration.
For the remainder of this section we investigate a combination of the LOBPCG method with
(in)exact projection onto the complement of the kernel functions.

142

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

7.3.1

Exact and inexact projection onto the complement of the kernel function

We leave the algebraic context for a moment, and first consider the projection operators for the
discrete FE-spaces in the operator framework. We want to construct the projection operator


of Vh,p(Th ) onto the complement of the discrete gradient functions Wh,p+1 (Th ) .

Lemma 7.3 (Exact Projection). We consider the finite element spaces V := Vh,p(Th ) and
W := Wh,p+1 (Th ) with W V as introduced in Chapter 5. Let B : W W denote
the gradient operator, MV : V V denote natural embedding operator and K : W W
denote the Laplace-operator.
Then the L2 -orthogonal projection operator

Pker(A)
: V W

onto the complement of the gradient functions is given by

1

B MV u
Pker(A)
u = I B K

u Vh,p.

Proof. We consider the L2 -orthogonal projection onto the kernel of the curl-operator
Pker(A) : V W
The action of the projection operator on an element u V can be written in variational form:
For u V find a function w W such that Pker(A) u = w, i.e.,
(w, )0 = (u, )0

W.

(7.18)

This is equivalent to
M u, i
hK w, iW W = hMV u, B iV V = hB
V
W W .

Setting Pker(A)
u = u B w yields the stated result.

For the preconditioned eigensolver with projection, we have to apply this projection for each
iterate in order to avoid that the approximated eigenvector converges to a kernel vector. As
can be seen from above, the exact L2 -orthogonal projection requires the solution of a Poisson
problem in each iteration step, which usually is much too expensive. We propose to use an
inexact projection instead, requiring only an approximative inverse of the Laplace-operator;
for details on this approach we refer to Hiptmair-Neymeyr [57].
Definition 7.4 (Inexact Projection). Let C denote a scaled preconditioner for the Laplace
operator K such that kK k kC k. We define the inexact projection operator Pker(A)
:

V V (corresponding to the exact projection operator Pker(A) ) by

1

B MV u
Pker(A)
u := I B C

Here we used the notation of the previous lemma.

u V.

7.3. PRECONDITIONED EIGENSOLVERS FOR THE MAXWELL PROBLEM

7.3.2

143

A preconditioned eigensolver for the Maxwell problem with inexact

projection

We return now to the matrix notation, i.e., the algebraic version of the Maxwell eigenvalue
problem (7.6), and use the notation introduced for the discrete operators in the previous
section now also for the matrix notation. Hence, the (in)exact projection operator can be
defined like above.
The inexact projection is realized by repeated application (k 1) of the inexact projector

Pker(A)
. The basic iteration step of the inexact inverse iteration method with inexact projection
is then given by

k


1

ui+1 = Pker(A)
I CA+M
(A + (ui )M ) ui



k
1
1
(A + (ui )M ) ui ,
(7.19)
= I B C
B MV
I CA+M

where MV denotes the mass matrix with respect to the H(curl)-conforming FE-space V and
B denotes the matrix corresponding to the discrete gradient operator, i.e., the representation
of gradient fields within the basis of V . Note that in general, the generation of this matrix
is non-trivial. However, as we will see below, due to the special choice of basis functions,
assembly is rather simple.
We observe that preconditioned gradient methods with inexact projections involve both a
curl-curl-preconditioner CA+M as well as a Laplace preconditioner C .
The basic algorithm: LOBPCG with inexact projection
Initial setting:

k 0

Choose initial vectors U 0 = Pker(A)

U RN m .

k

Xymin
Perform one step of (7.13) to obtain U 1 = Pker(A)


k

and P 1 = Pker(A)
diag(0, 1)Xymin RN m .
The (i + 1)-iteration:

1
(AU i M U i i ) with i =
1. Compute the preconditioned
residuals W i = CA+M


diag((ui1 , . . . uiN )

2. Solve the 3m-by-3m Rayleigh-Ritz-problem with y R3m

X T AXy = X T M Xy

with X = [U i , W i , P i ].

3. Take the m smallest Ritz-Pairs (1 , Xy1 ), ..., (m , Xym ) and set

k

[Xy1 , ..., Xym ],

U i+1 = Pker(A)

k 


diag(0, 1, 1)Xy1 , ..., diag(0, 1, 1)Xym .

P i+1 = Pker(A)

Concerning the practical implementation, the repeated application of the inexact projection
after each iteration is performed by the application of k steps of a preconditioned iterative
solver in the potential space. The scaling of the Poisson-preconditioner required in 7.4 is
provided by using a multiplicative Schwarz preconditioner C .

144

7.3.3

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

Exploiting the local exact sequence property

The discrete gradient matrix

By means of the local exact sequence property, the discrete gradient matrix is block-diagonal
with respect to the low-order, single edge-based, face-based and cell-based blocks. Assuming
a reordering of the degrees of freedom corresponding to edges, faces and cells we obtain


B = diag B, 0 , B, EE , B, F F , B, CC ,
(7.20)

where B, EE , B, F F , B, CC are again block-diagonal corresponding to each single edge,

each single face and each single cell of the triangulation. Since higher-order gradient fields are
explicitly used in the construction of the new H(curl)-conforming FE-basis, stated in Chapter
5, several blocks further simplify to the identity matrix I:
B, EE = I,

B, F F = diag diag(I, 0, 0), . . . , diag(I, 0, 0) ,

B, CC = diag diag(I, 0, 0), . . . , diag(I, 0, 0) .

The lowest-order discrete gradient operator

1


1
B,0 i,j =

(7.21)
(7.22)
(7.23)

B, 0 is given by

if Ei = [e1 , e2 ] and j = e2
if Ei = [e1 , e2 ] and j = e1
else

which can be seen by considering the lowest-orderPdegrees of freedom

P of the Nedelec basis:
For a w Wh,0 (Th ) H 1 () there holds w = Vj V wj j = E wi i with wj = w(Vj )
Re
and wi = e12 w dx = w(Ve2 ) w(Ve1 ) = we2 we2 .

Remark 7.5. In case of a general H 1 -conforming and H(curl)-conforming finite element

basis, the calculation of the discrete gradient operator requires the inversion of the element
mass matrix on each element.
Review of the applied H(curl) preconditioner

The best choice for preconditioning the residual Aui (ui )M ui , would be to apply the pseudoinverse A , which would naturally include also the projection to the complement of the kernel.
1
Note that application of the approximative inverse CA+M
may introduce artificial kernel
red
components. This can be suppressed by applying the reduced basis preconditioner CA+M
instead. Due to our special construction of the FE-basis (cf. Section (6.3)), the reduced
space preconditioner can be realized easily. Although we cannot give a rigorous mathematical
justification that the reduced basis preconditioner is a better approximation for the pseudoinverse, this choice is at least as good the original one. The reduced basis preconditioner is
however typically much less ill-conditioned and its application requires much less numerical
effort.

7.4

Numerical Results

In the following we verify the performance of our chosen finite element basis in combination
with the Locally Optimal Block PCG method with (in)exact projection by means of various

145

7.4. NUMERICAL RESULTS

benchmark problems. We consider the electric eigenvalue problem, Problem 7.1, with homogenous Dirichlet conditions and constant material parameters = 1 and = 1:
Find 6= 0 and u H0 (curl, ) such that
Z
Z
1
curl u curl v dx = u vdx
v H0 (curl, ).
(7.24)

On the webpage of Monique Dauge [39] a convenient range of benchmark problems for
above Maxwell eigenvalue problem is provided. The focus lies on considering geometries with
re-entrant corners and edges, where a part of the spectrum is conjectured to correspond to
highly-singular eigenfunctions.
We consider the following two domains:
The thick L-shape = 2 (0, 1) with 2 := (1, 1)2 \ (1, 0), we obtain singularities
due to the re-entrant corner. For this setting benchmark results are available on
Dauge [39].
The Fichera corner = (1, 1)3 \(1, 0)3 , where all three re-entrant edges as well as the
re-entrant corner produce singularities. There are no benchmark results yet available,
but we will compare our results with the ones of Costabel-Dauge, stated on the webpage
Dauge [39], and Bramble et al. [29].
In the context of computing Maxwell eigenvalue problems on polyhedral domains we also
refer to Frauenfelder [48] and Ainsworth et al. [3] contributing numerical results on
the Maxwell eigenvalue computation on polyhedral domains.

7.4.1

An h-p-refinement strategy

In order to resolve singularities due to re-entrant corners and edges we use the following hprefinement strategy, which is based on marking re-entrant corners and edges a priori in the
geometry as singular.
Our h-p-strategy is the following:
k steps of geometric h-refinement with a progression factor 0 < q 0.5 are performed
towards apriori marked faces, edges and corners. The strategy should become clear
considering the example of the thick L-shape domain, illustrated in Figure 7.1. The
optimal progression factor q depends on the solution, but a commonly used choice is
q = 0.17 (see e.g. Melenk [68]).
Anisotropic distribution of polynomial degrees (p): On each element the level of actual
i is known, i.e. 0 rxi k. Let pmin denote the initial
refinement rxi in each direction x
i
polynomial order. Then we choose the degree pxi = pmin + (k rxi ) in the direction x
of the element.

7.4.2

The Maxwell EVP on the thick L-Shape

The eigenvalues on the thick L-shape can be deduced by addressing to its tensor-product
structure, i.e. the 2 (0, 1), where 2 denotes the two-dimensional L-shaped domain.

146

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

Figure 7.1: Geometric h-refinement towards the re-entrant edge: original prismatic mesh
(left), level 1 (center) and level 2 (right); involving prisms (blue) and hexahedra (yellow).
The eigenvalues on the tensorized L-shape are realized by the sum of a non-zero Neumann
eigenvalue of the 2D L-shape and a Dirichlet eigenvalue on the L-shape domain or the sum of a
Dirichlet eigenvalue on 2 and a non-zero Neumann eigenvalue on (0, 1). Accurate numerical
solutions for the two-dimensional L-shape are available. The following benchmark results for
the 8 smallest eigenvalues on the thick L-shape are due to Dauge [39]:
1 = 9.63972384472
4 = 15.1972519265
7 = 19.7392088022

2 = 11.3452262252
5 = 19.5093282458
8 = 19.7392088022

3 = 13.4036357679
6 = 19.7392088022

(7.25)

The first, the second and the fifth eigenvalue correspond to the first Dirichlet or Neumann
eigensolutions on the two-dimensional L-shape 2 , which has a strong unbounded singularity
in the corner. The third and the fourth eigenvalue refer to an eigenvector, which is in
H 1 (2 ). The eigenvalues number 6, 7, and 8 are related to an analytic eigenvector on the 2
dimensional domain.
We first run the eigenvalue computation for uniform p-refinement on a fixed mesh, namely
the first level of h-refinement depicted in Figure 7.1 with progression factor q = 0.17. The

log-log-plot in Figure 7.2 illustrates the convergence history of the relative errors |i|i| i |
i with respect to the degrees of freedom (dofs).
of the computed 8 smallest eigenvalues
As expected when using a p-method, we observe only algebraic convergence rates for the
eigenvalues of rank 1,2, and 5.

In the second test series we repeat the eigenvalue calculations with uniform p-refinement on
fixed meshes generated by various levels of geometric h-refinements (k = 1, . . . 5, q = 0.17).
The convergence history of the mean of the relative errors of the 8 smallest eigenvalues is
presented in Figure 7.3. Finally, we combine the geometric mesh refinement with anisotropic
p-refinement as described above. The envelope of the convergence curve in Figure 7.4 corresponds to the hp-refined eigenvalue computation with refinement factor q = 0.25. We obtain
the conjectured exponential convergence in the number of unknowns even in the presence of
eigenvalues corresponding to highly-singular eigenfunctions.

147

7.4. NUMERICAL RESULTS

1
lambda1
lambda2
lambda3
lambda4
lambda5
lambda6
lambda7
lambda8

0.1
0.01
0.001

rel. error

1e-04
1e-05
1e-06
1e-07
1e-08
1e-09
1e-10
100

1000

10000

100000

NDofs

Figure 7.2: L-Shape: Average relative errors of the 8 smallest eigenvalues. 1 level of geometric
refinement. Uniform p refinement.

0.1
ref-level 1, uniform p
ref-level 2, uniform p
ref-level 3, uniform p
ref-level 4, uniform p
ref-level 5, uniform p

0.01

0.001

rel. error

1e-04

1e-05

1e-06

1e-07

1e-08

1e-09
100

1000

10000

100000

NDofs

Figure 7.3: L-Shape: Average relative errors of the 8 smallest eigenvalues. Uniform p refinement on fixed geometric refined meshes.

148

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

0.1
ref-level 1, uniform p
ref-level 2, uniform p
ref-level 3, uniform p
ref-level 4, uniform p
ref-level 5, uniform p
ref-ref 1-10, var p=0,..,lev

0.01

0.001

rel. error

1e-04

1e-05

1e-06

1e-07

1e-08

1e-09

1e-10
100

1000

10000

100000

NDofs

Figure 7.4: L-Shape: Average relative error of 8 the smallest eigenvalues. Exponential convergence: Combination of geometric h-refinement and p-anisotropic order distribution

149

7.4. NUMERICAL RESULTS

7.4.3

The Maxwell EVP on the Fichera corner

As already mentioned no benchmark results are available for the Fichera corner. In the
following we contribute with our results. First, we cite the results available on the website
Dauge [39] and in Bramble et al. [29] in Table 7.1.
j
1
2
3
4
5
6
7
8

j (Dauge [39])

3.31380523357
5.88634994607
5.88634994619
10.6945143272
10.6945143276
10.7005804029
12.3345495948
12.3345495949

reliable digits(?)
1
3
3
4
4
2
3
3

conject. ev
3.2??
5.88?
5.88?
10.694
10.694
10.7??
12.32?
12.32?

j
1
2
3
4
5
6
7
8

j (Bramble [29])

3.23432
5.88267
5.88371
10.6789
10.6832
10.6945
12.23653
12.23723

Table 7.1: Results by Dauge and Costabel from Dauge [39] with conjectured eigenvalue
and hopefully reliable digits (left table) and results corresponding to Bramble-KolevPasciak [29] (right table)
Table 7.2 presents the results of our computation of the 8 smallest eigenvalues using the
LOBPCG method with inexact projection in following problem setting:
We started with an initial mesh consisting of 12 tetrahedra and performed 3 levels of geometric
h-refinement. The refined mesh contains hexahedra, tetrahedra and prisms and is depicted in
Table 7.2.
We use the anisotropic polynomial order distribution as described at the beginning of this
section. The degrees on each element vary from p = 3 to p = 6 for the H(curl)-conforming
space, which yields 53982 degrees of freedom, and from p = 4 to p = 7 for the corresponding
potential space H 1 (), which involves 19318 degrees of freedom. The total solution time on a
Dual Processor Intel Xeon 64Bit 2,8 GHz is 6460 seconds. The number of LOBPCG iterations
is 14. Note that one has to provide an H 1 -preconditioner with sufficiently smooth blocks to
avoid zero eigenvectors in the solution. The iteration numbers and results are independent
from the choice of the shift parameter in the H(curl)-preconditioner. We used the reduced
basis preconditioner, since as already verified in the last chapter it is better conditioned and
faster in its application compared to the one in the full basis.
For visualization we rerun the numerical test for our problem setting with homogenous Neumann boundary conditions and present the computed eigenfunctions corresponding to the
first and the seventh eigenvalue in in Figure 7.5.

150

CHAPTER 7. THE MAXWELL EIGENVALUE PROBLEM

j
1
2
3
4
5
6
7
8

3.21999388904
5.88044248619
5.88045528405
10.6856632462
10.6936955486
10.6937289163
12.3168796291
12.3176900965

Table 7.2: The 8 smallest eigenvalues on Fichera corner, computed on left mesh with
anisotropic order of p = 3, . . . 6 (NgSolve). On the left hybrid mesh by 3 levels of hp-refinement
with hexahedra (yellow), prisms (blue), tetrahedra (red).

Figure 7.5: Absolute value of first (left) and seventh (right) eigenfunction (only for visualization computed with homogenous Neumann data)

Appendix A

APPENDIX
A.1

Notations

We use the symbol  for referring to up to a constant.

Polynomial spaces
Let K R3 .
Pp (K)

...

Pep (K)

...

Qp (K)

A.2

...

space of polynomials
of maximal
 total degree p, i.e.
P
Pp (K) = { pi+j+kp cijk xi y j z k  cijk R}.

set of homogenous polynomialsof exact degree p, i.e.

P
Pep (K) = { i+j+k=p cijk xi y j z k  cijk R}.
space of polynomials
of maximal degree p, i.e.
Pp
Qp (K) = { 0i,j,kp cijk xi y j z k  cijk R}.

Basic Vector Calculus

Let denote a scalar function in x, while v is a vector-valued function.

div(v) = v + div v

curl(v) = curl v + v

(A.1)
(A.2)

div(curl v) = 0

(A.3)

curl() = 0

(A.4)

trn,Fm (curl()) = n curl() = n curl(
+ n n)

= n + n n = n ( )
= curlFm ( )

A.3

Some more orthogonal polynomials

(,)

Jacobi-polynomials , denoted by Pn
, are orthogonal polynomials associated with the

weight function (1 x) (1 + x) (for > 1, > 1) in the interval [1, 1]. This means
151

152

APPENDIX A. APPENDIX

they fulfill

(,) (,)
(1 x) (1 + x) Pm
Pn
dx = 0 for m 6= n.

They can be defined by the Rodrigues formula

n


(1)n
+n
+n

d
(1

x)
(1
+
x)
.
(1

x)
(1
+
x)
2n n!
dxn
Fast point-evaluation up to order p is established by the three-term recurrence formula

Pn(,) (x) =

(,)

= 1,

(,)

P0
P1

(,)
Pn+1 (x)

1
2

( ) + 12 ( + + 2)x,

= (an +

(,)
bn x)Pn
(x)

(A.5)

(,)
cn Pn1 (x),

where the coefficients are given as

(2n + + + 1)(2 2 )
,
2(n + 1)(n + + + 1)(2n + + )
(2n + + + 1)(2n + + + 2)
,
=
2(n + 1)(n + + + 1)
(n + )(n + )(2n + + + 2)
=
.
(n + 1)(n + + + 1)(2n + + )

an =
bn
cn

The Legendre polynomials are a special case of the Jacobi ones. Choosing = = 0 yields
n (x) = Pn(0,0) (x) for x [1, 1].
Integrated Legendre polynomials are related to Jacobi polynomials by the relation
1
(1,1)
(1 x2 )Pn2 (x)
for x [1, 1], n 2.
Ln (x) =
2(n 1)
Scaled Jacobi polynomials

are defined as
PnS, (,) (x, t) := Pn(,)

We can compute them by the recurrence formula

S, (,)

Pn+1

x
n
t .
t
S, (,)

(x, t) = (an t + bn x)PnS, (,) (x, t) + cn t2 Pn1

(x, t)

(A.6)

with an , bn , cn as defined in (A.5). We remark that the recurrence formula is free of divisions
by t.
Choosing = = 0 yields the scaled Legendre polynomials
nS (x) = PnS, (0,0) (x).
Jacobi-based bubble functions We can define Jacobi-based polynomials spanning
Pp0 ([1, 1]) by the set of bubble functions
(2,2)

j = (1 x)(1 + x)Pj2 (x)

is L2 -orthogonal in the interval [1, 1] in the sense of
Z 1
i j dx = 0 for i 6= j.
1

153

A.3. SOME MORE ORTHOGONAL POLYNOMIALS

1.5
1
0.5

-1

-0.5

0.5
-0.5
-1

Figure A.1: Jacobi-based bubble functions j of order 2 to 7

A.3.1

Some Calculus for Scaled Legendre Polynomials

By the relation
x Ln (x) n Ln (x) = Ln1 (x)
x n (x) n n (x) = n1 (x)

we can show that

LnS (s, t)

= Ln1

 
 
s n1
s n1
t
t + Ln
t
s
t
t

(A.7)

S
S
= n2
(s, t) t t + n1
(s, t) s

holds. The same holds for Legendre polynomials

nS (s, t)

A.3.2

n1

 
s
n1
s n1

t
s
t
t + n
t
t

(A.8)

= (n1 ),S (s, t) t t + (n ),S (s, t) s.

Some technical things

2
2
2 1
1
(1 + ) 2 2
x(, , ) =
1 1
2
2
(1 + ) 1
1

0
2
1

2
(1 + ) 1

0
0

S
Li+2
(2x 1 + y + z, 1 y z)

2
0
S
(2 1 , 1 + 2 ) 1
= iS (2 1 , 1 + 2 ) (1 + 2 ) 1 + i+1
1
1


vj (y, z) = 3 jS (3 1 2 , 1 4 )

0
= 3 (j1 )S (3 1 2 , 1 4 ) (1 4 ) 0
1

0
0
+3 (j )S (3 1 2 ), 1 4 ) 2 + jS (3 1 2 , 1 4 ) 1
0
1


wk (z) = 4 k (24 1)


= k (2z 1) + 2k (2z 1) ez

ui (x, y, z)

154

APPENDIX A. APPENDIX


2
0

i+1 1
i+1
ui (D(, , )) =
i () 1 + i+1 () 1 1
2
2
1
1



0
0




1+ 1 j1

0
2
+ j ()
vj (D(, , )) =
j1 ()
+ j ()
2
2
1
1


(wk )(D(, , )) = k () + 2k () ez

1 j
2

0
1
0

Bibliography
[1] M. Ainsworth. A hierarchical domain decomposition preconditioner for h-p finite element approximation on locally refined meshes. SIAM Journal on Scientific Computing,
17(6):13951413, 1996.
[2] M. Ainsworth and J. Coyle. Hierarchic finite element bases on unstructured tetrahedral
meshes. International Journal of Numerical Methods in Engineering, 58(14):21032130,
2003.
[3] M. Ainsworth, J. Coyle, P. Ledger, and K. Morgan. Computation of Maxwell eigenvalues
using higher order edge elements in three dimensions. IEEE Transactions on Magnetics,
pages 21492153, 2003.
[4] C. Amrouche, C. Bernardi, M. Dauge, and V. Girault. Vector potentials in threedimensional non-smooth domains. Mathematical Methods in the Applied Sciences, 21:823
864, 1998.
[5] P. Arbenz and R. Geus. A comparison of solvers for large eigenvalue problems occuring
in the design of resonant cavities. Numerical Linear Algebra with Applications, 6(1):316,
1999.
[6] D.N. Arnold, R.S. Falk, and R. Winther. Preconditioning in H(div) and applications.
Mathematics of Computation, 66(219):957984, July 1997.
[7] D.N. Arnold, R.S. Falk, and R. Winther. Multigrid in H(div) and H(curl). Numerische
Mathematik, 85:197218, 2000.
[8] D.N. Arnold, R.S. Falk, and R. Winther. Finite element exterior calculus, homological
techniques, and applications. Acta Numerica, 15:1155, 2006.
[9] D.N. Arnold, R.S. Falk, and R. Winther. Differential complexes and stability of finite
element methods. I. the de Rham complex. In Proceedings of the IMA workshop on
Compatible Spatial Discretizations for PDE, to appear.
[10] I. Babuska and B.Q. Guo. Approximation properties of the h-p version of the finite
element method. Computational Methods of Applied Mechanical Engineering, 133:319
346, 1996.
[11] I. Babuska and M. Suri. The p and h-p versions of the finite element method, basic
principles and properties. SIAM Review, 36(4):578632, 1994.
[12] I. Babuska, B.A. Szabo, and I.N.Katz. The p-version of the finite element method. SIAM
Journal on Numerical Analysis, 18(3):515545., 1981.

155

156

BIBLIOGRAPHY

[13] F. Bachinger, U. Langer, and J. Schoberl. Efficient solvers for nonlinear time-periodic
eddy current problems. SFB-Report 2004-16, Johannes Kepler University Linz, 2004.
submitted to Computing and Visualization in Science.
[14] F. Bachinger, J. Schoberl, and U. Langer. Numerical analysis of nonlinear multiharmonic
eddy current problems. Numerische Mathematik, 100(4):593616, 2005.
[15] Z. Bai, J. Demmel, J. Dongarra, A. Ruhe, and H. van der Vorst, editors. Templates for
the Solution of Algebraic Eigenvalue Problems: A Practical Guide. SIAM, Philadelphia,
2000. online available: http://www.cs.utk.edu/ dongarra/etemplates/index.html.
[16] R. Beck, P. Deuflhard, R. Hiptmair, R.H.W. Hoppe, and B. Wohlmuth. Adaptive multilevel methods for edge element discretizations of Maxwells equations. Surveys Math.
Industry, 8(34):271312, 1999.
[17] S. Beuchler and J. Sch
oberl. New shape functions for triangular p-fem using integrated
jacobi polynomials. Numerische Mathematik, 103:339366, 2006.
[18] O. Bro. Numerische Aspekte von Potentialformulierungen in der Elektrodynamik. Habilitationsschrift, TU Graz, 1992.
[19] D. Boffi. Discrete compactness and Fortin operator for edge elements. Numerische
Mathematik, 87:229246, 2000.
[20] D. Boffi, M. Costabel, D. Monique, and L. Demkowicz. Discrete compactness for the
hp version of rectangular edge finite elements. SIAM Journal on Numerical Analysis,
44(3):9791004, 2006.
[21] D. Boffi, P. Fernandes, L. Gastaldi, and I. Perugia. Computational models of electromagnetic resonators: Analysis of edge element approximation. SIAM Journal on Numerical
Analysis, 36:2641290, 1999.
[22] A. Bossavit. Mixed finite elements and the complex of Whitney forms. In J. Whiteman, editor, The Mathematics of Finite Elements and Applications VI, pages 137144.
Academic Press, 1988.
[23] A. Bossavit. A rationale for edge elements in 3d field computations. IEEE Transactions
on Magnetics, 24(1):7479, 1988.
[24] A. Bossavit. Whitney forms: a class of finite elements for three-dimensional computations
in electromagnetism. IEE Proceedings, 135(8):493500, 1988.
[25] A. Bossavit. Computational Electromagnetism: Variational formulation, complementary,
edge elements. Academic Press Inc., San Diego, CA, 1989.
[26] A. Bossavit. Solving Maxwell equations in a closed cavity, and the question of spurious
modes. IEEE Transactions on Magnetics, 26(2):702 705, 1990.
[27] D. Braess. Finite Elemente: Theorie, schnelle L
oser und Anwendungen in der Elastizit
atstheorie. Springer Verlag Berlin Heidelberg, 3rd edition edition, 2003.
[28] J.H. Bramble, A.V. Knyazev, and J.E. Pasciak. A subspace preconditioning algorithm
for eigenvector/eigenvalue computation. Advances in Computational Mathematics, 6:159
189, 1996.

BIBLIOGRAPHY

157

[29] J.H. Bramble, T.V. Kolev, and Pasciak J.E. The approximation of the Maxwell eigenvalue
problem using a least-squares method. Mathematics of Computation, 74(252):15751598,
2005.
[30] J.H. Bramble and X. Zhang. The analysis of multigrid methods. In P.G. Ciarlet and J.L.
Lions, editors, Handbook of Numerical Analysis, volume 7, pages 173415. North-Holland,
2000.
[31] S.C. Brenner and L.R. Scott. The Mathematical Theory of Finite Element Methods,
volume 15 of Texts in Applied Mathematics. Springer-Verlag New York, 2nd edition,
2002.
[32] F. Brezzi and M. Fortin. Mixed and Hybrid Finite Element Methods. Springer Verlag,
Berlin, 1991.
[33] A. Buffa and P. Jr. Ciarlet. On traces for functional spaces related to Maxwells equation.
Part I: An integration by parts formula in Lipschitz polyhedra. Mathematical Methods
in the Applied Sciences, 24(1):930, 2001.
[34] S. Caorsi, P. Fernandes, and M. Raffetto. On the convergence of galerkin finite element
approximations of electromagnetic eigenproblems. SIAM Journal on Numerical Analysis,
38(2):580607, 2000.
[35] P.G. Ciarlet. The Finite Element Method for Elliptic Problems. North Holland, Amsterdam, 1978.
[36] M. Clemens, U. van Rienen, and T. Weiland. Comparison of krylov-type methods for
complex linear systems applied to high-voltage problems. IEEE Transactions on Magnetics, 34(5):33353338, 1998.
[37] M. Clemens and T. Weiland. Iterative methods for the solution of very large complexsymmetric linear systems of equations in electrodynamics. In Copper Mountain Conference on Iterative Methods, volume 2, 1996. available via MGNET.
[38] M. Costabel and M. Dauge. Singularities of electromagnetic fields in polyhedral domains.
Archive for Rational Mechanics and Analysis, 151(3):221 276, 2000.
[39] Monique Dauge.
Benchmark computations for Maxwell equations for
the
approximation
of
highly
singular
solutions.
http://perso.univrennes1.fr/monique.dauge/benchmax.html.
[40] R. Dautray and J.-L. Lions. Spectral Theory and Applications, volume 3 of Mathematical
Analysis and Numerical Methods for Science and Technology. Springer-Verlag, 1990.
[41] L. Demkowicz. Edge finite elements of variable order for Maxwells equations - a discussion. Technical report, TICAM, 2000. TICAM Report 00-32.
[42] L. Demkowicz. Computing with hp Finite Elements. I.One- and Two-Dimensional Elliptic
and Maxwell Problems. CRC Press, Taylor and Francis, 2006. in press.
[43] L. Demkowicz and A. Buffa. H 1 , H(curl)- and H(div)-conforming projection-based interpolation in three dimensions. quasi-optimal p-interpolation estimates. Comput. Meth.
Appl. Mech. Engrg., 194(2-5):267296, 2005.

158

BIBLIOGRAPHY

[44] L. Demkowicz, P. Monk, L. Vardapetyan, and W. Rachowicz. De Rham diagram for hp

finite element spaces. Computers and Mathematics with Aplications, 39(7-8):2938, 2000.
[45] L. Demkowicz and L. Vardapetyan. hp-adaptive finite elements in electromagnetics.
Computers Methods in Applied Mechanics and Engineering, 169:331344, 1999.
[46] M. Dubiner. Spectral methods on triangles and other domains. Journal of Scientific
Computing, 6(4):345390, 1991.
[47] E. Dyakonov. Optimization in Solving Elliptic Problems. CRC Press, 1996.
[48] Philipp Frauenfelder. hp-Finite element methods on anisotropically, locally refined meshes
in three dimensions with stochastic data. Mathematische wissenschaften, Eidgenossische
Technische Hochschule ETH Z
urich, 2004.
[49] V. Girault and P.A. Raviart. Finite Element Methods for Navier-Stokes Equations: Theory and Algorithms. Springer Series in Computational Mathematics. Springer-Verlag,
Berlin-Heidelberg, 1986.
[50] J. Gopalakrishnan and L. Demkowicz. Quasioptimality of some spectral mixed methods.
Journal on Computational and Applied Mathematics, 167:163182, 2004.
[51] C. Grossmann and H.G. Roos. Numerik partieller Differentialgleichungen. Teubner, 1994.
[52] W. Hackbusch. Multi-Grid Methods and Applications. Springer, Berlin, Heidelberg, New
York, 1985.
[53] W. Hackbusch. Iterative Solution of Large Sparse System of Equations. Springer-Verlag,
1994.
[54] R. Hiptmair. Canonical construction of finite elements. Mathematics of Computation,
68:13251346, 1990.
[55] R. Hiptmair. Multigrid method for Maxwells equations. SIAM Journal of Numerical
Analysis, 36(1):204255, 1999.
[56] R. Hiptmair. Finite elements in computational electromagnetism. Acta Numerica, pages
237339, 2002.
[57] R. Hiptmair and K. Neymeyr. Multilevel method for mixed eigenproblems. SIAM Journal
of Scientific Computing, 23(6):21412164, 2002.
[58] R. Hiptmair and A. Toselli. Overlapping Schwarz methods for vector valued elliptic
problems in three dimensions. In Parallel solution of PDEs, IMA Volume in Mathematics
and its Applications. Springer, Berlin-Heidelberg-New York, 1997.
[59] R. Hitpmair. Multigrid method for H(div) in three dimensions. Electronic Transactions
on Numerical Analysis, 6:777, 1997.
[60] G.E. Karniadakis and S.J. Sherwin. Spectral/hp Element Methods for CFD. Numerical
Mathematics and Scientific Computation. Oxford University Press, New York, Oxford,
1999.

BIBLIOGRAPHY

159

[61] F. Kikuchi. Mixed and penalty formulations for finite element analysis of an eigenvalue
problem in elctromagnetism. Computer Methods in Applied Mechanics and Engineering,
pages 509521, 1987.
[62] A.V. Knyazev. Preconditioned eigensolvers - an oxymoron? Electronic Transactions on
Numerical Analysis, 7:104123, 1998.
[63] A.V. Knyazev. Toward the optimal preconditioned eigensolver: Locally optimal block
preconditioned conjugate gradient method. SIAM Journal on Scientific Computing,
23(2):517541, 2001.
[64] A.V. Knyazev and K. Neymeyr. Efficient solution of symmetric eigenvalue problems using
multigrid preconditioners in the locally optimal block preconditioned gradient method.
Electronical Transaction of Numerical Analysis, 15:3855, 2003.
[65] A.V. Knyazev and K. Neymeyr. A geometric theory for preconditioned inverse iteration.
iii: A short and sharp convergence estimate for generalized eigenvalue problems. Linear
Algebra and its Application, 358:95114, 2003.
[66] J.F. Lee, D.K. Sun, and Z.J. Cendes. Tangential vector finite elements for electromagnetic
field computation. IEEE Transactions on Magnetics, 27(5):4032 4035, 1991.
[67] W. McLean. Strongly Elliptic Systems and Boundary Integral Equations. Cambridge
University Press, 2000.
[68] J.M. Melenk. hp-Finite Element Methods for Singular Perturbations, volume 1796 of
Lecture notes in in mathematics. Springer Berlin Heidelberg, 2002.
[69] P. Monk. On the p- and hp-extension of Nedelecs curl-conforming elements. Journal of
Computational and Applied Mathematics, 53(1):117137, 1994.
[70] P. Monk. Finite Element Methods for Maxwells Equations. Numerical Mathematics and
Scientific Computation. The Clarendon Press Oxford University Press, New York, 2003.
[71] P. Monk, L. Demkowicz, and Vardapetyan L. Discrete compactness and the approximation of Maxwells equations in R3. Mathematics of Computation, 70:507523, 2001.
[72] J.C. Nedelec. Mixed finite elements in R3 . Numerische Mathematik, 35:315341, 1980.
[73] J.C. Nedelec. A new family of mixed finite elements in R3 . Numerische Mathematik,
50:5781, 1986.
[74] K. Neymeyr. A hierarchy of preconditioned eigensolvers for elliptic differential operators,
2001. Habilitation.
[75] J. Pasciak and J. Zhao. Overlapping Schwarz methods in H(curl) on nonconvex domains.
East West Journal on Numerical Analysis, 10:221234, 2002.
[76] W. Rachowicz and L. Demkowicz. An hp-adaptive finite element method for electromagnetics. part 2: A 3d implementation. International Journal for Numerical Methods in
Engineering, 53:147180, 2002.
[77] Y. Saad. Iterative Methods for Sparse Linear Systems. PWS, Boston, 1996.

160

BIBLIOGRAPHY

[78] J. Sch
oberl. A multilevel decomposition result in H(curl). In TUDelft, editor, Proceedings
of the 8th European Multigrid Conference EMG 2005, Netherlands. to appear.
[79] J. Schoberl. NETGEN - An advancing front 2D/3D-mesh generator based on abstract
rules. Computing and Visualization in Science, 1:4152, 1997.
[80] J. Sch
oberl. Robust Multigrid Methods for Parameter Dependent Problems. PhD thesis,
Institute for Computational Mathematics, J.K. University Linz, 1999.
[81] J. Schoberl. A posteriori error estimates for Maxwell equations. Mathematics of Computation, submitted.
[82] J. Schoberl, J. Melenk, Pechstein C., and Zaglmayr S. Additive Schwarz preconditioning for p-version triangular and tetrahedral finite elements. IMA Journal of Numerical
Analysis, 2005. submitted.
[83] Ch. Schwab. p- and hp- Finite Element Methods: Theory and Applications in Solid and
Fluid Mechanics. Numerical Mathematics and Scientific Computation. Oxford University
Press, Oxford, 1998.
[84] B.F. Smith, P.E. Bjrstad, and W.D. Gropp. Domain decomposition: Parallel multilevel
algorithms for elliptic partial differential equations. Cambridge University Press, 1996.
[85] D.-K. Sun, J.-F. Lee, and C. Zoltan. Construction of nearly orthogonal nedelec bases
for rapid convergence with multilevel preconditioned solvers. SIAM Journal on Scientific
Computing, 23(4):10531076, 2001.
[86] B. Szabo, A. D
uster, and E. Rank. The p-version of the finite element method. In E. Stein,
R. de Borst, and J.R.T. Hughes, editors, Encyclopedia Of Computational Mechanics,
volume 1 Fundamentals, chapter 5, pages 119139. John Wiley and Sons, Chichester,
2004.
[87] B. Szabo and Babuska I. Finite Element Analysis. J. Wiley and Sons, New York, 1991.
[88] G. Szego. Orthogonal Polynomials, volume 23 of Colloquium Publications. American
Mathematical Society, Rrovidence, Rhode Island, 3 edition, 1974.
[89] A. Toselli. Overlapping Schwarz methods for Maxwells equations in three dimensions.
Numerische Mathematik, 86(4):733752, 2000.
[90] A. Toselli and O. Widlund. Domain Decomposition Methods - Algorithms and Theory,
volume 34 of Series in Computational Mathematics. Springer, 2005.
[91] H. van der Vorst. Iterative Krylov Methods for Large Linear Systems. Cambridge University Press, 2003.
[92] U. van Rienen. Numerical Methods in Computational Electrodynamics: Linear Systems
in Practical Applications, volume 12 of Lecture notes in computational science and engineering. Springer-Verlag Berlin Heidelberg, 2001.
[93] Y. Wang, P. Monk, and Szabo B. Computing cavity modes using the p version of the
finite element method. IEEE Transaction on Magnetics, (32):19341940, 1996.

BIBLIOGRAPHY

161

[94] T.C. Warburton, S.J. Sherwin, and G.E. Karniadakis. Basis functions for triangular and
quadrilateral high-order elements. SIAM Journal on Scientific Computing, 20(5):1671
1695, 1999.
[95] J.P. Webb. Hierarchical vector basis functions of arbitrary order for triangular and tetrahedral finite elements. IEEE Transactions on Antennas and Propagation, 47(8):1495
1498, August 1999.
[96] J.P. Webb and Forghani B. Hierarchical scalar and vector tetrahedra. IEEE Transactions
on Magnetics, 29(2):14951498, March 1993.
[97] X. Zhang. Multilevel Schwarz methods. Numerische Mathematik, 63:521539, 1992.

162

BIBLIOGRAPHY

Eidesstattliche Erkl
arung
Ich, Sabine Zaglmayr, erklare an Eides statt, dass ich die vorliegende Dissertation selbstandig
und ohne fremde Hilfe verfasst, andere als die angegebenen Quellen und Hilfsmittel nicht
benutzt bzw. die wortlich oder sinngema entnommenen Stellen als solche kenntlich gemacht
habe.

Linz, Juli 2006

Sabine Zaglmayr

A1

A2

Curriculum Vitae
Personal Data
Name:

Sabine Zaglmayr

Nationality:

Austria

Date of Birth:

11. 5. 1978

Place of Birth:

Sch
arding, Austria

Education
1984 - 1988

Elementary School St.Marienkirchen

1988 - 1996

Comprehensive Secondary School Scharding

1996 - 2002

Studies in Technical Mathemtics,

Branch of study: Industrial Mathematics,
Johannes Kepler University Linz

since 2002

PhD studies at Johannes Kepler University Linz

Professional Career
since 2002

Research assistant at FWF-START Project Y-192

3D hp-Finite Elements: Fast Solvers and Adaptivity

since 2004

hosted at J. Radon Institute for Computational and Applied Mathematics,

Austrian Academy of Sciences

Awards
2004

Masters thesis awarded with the Hansjorg Wacker Memorial Prize,

European Consortium for Mathematics in Industry (ECMI)

A3