Numerical Methods for

Partial Differential Equations

Joachim Schöberl

May 12, 2022

2
Contents

1 Introduction 7
1.1 Classification of PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2 Weak formulation of the Poisson Equation . . . . . . . . . . . . . . . . . . 8
1.3 The Finite Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 The abstract theory 13

2.1 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Projection onto subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3 Riesz Representation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Symmetric variational problems . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5 Coercive variational problems . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.5.1 Approximation of coercive variational problems . . . . . . . . . . . 22
2.6 Inf-sup stable variational problems . . . . . . . . . . . . . . . . . . . . . . 23
2.6.1 Approximation of inf-sup stable variational problems . . . . . . . . 25

3 Sobolev Spaces 27
3.1 Generalized derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Trace theorems and their applications . . . . . . . . . . . . . . . . . . . . . 30
3.3.1 The trace space H 1/2 . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Equivalent norms on H 1 and on sub-spaces . . . . . . . . . . . . . . . . . . 39
3.5 Interpolation Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.1 Hilbert space interpolation . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.2 Banach space interpolation . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.3 Operator interpolation . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.5.4 Interpolation of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . 46
3.6 The weak formulation of the Poisson equation . . . . . . . . . . . . . . . . 47
3.6.1 Shift theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4 Finite Element Method 51

4.1 Finite element system assembling . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Finite element error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.3 A posteriori error estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3
4 CONTENTS

4.4 Equilibrated Residual Error Estimates . . . . . . . . . . . . . . . . . . . . 69

4.4.1 General framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4.2 Computation of the lifting kσ ∆ k . . . . . . . . . . . . . . . . . . . . 71
4.5 Non-conforming Finite Element Methods . . . . . . . . . . . . . . . . . . . 73
4.6 hp - Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.6.1 Legendre Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.6.2 Error estimate of the L2 projection . . . . . . . . . . . . . . . . . . 82
4.6.3 Orthogonal polynomials on triangles . . . . . . . . . . . . . . . . . 84
4.6.4 Projection based interpolation . . . . . . . . . . . . . . . . . . . . . 85

5 Linear Equation Solvers 89

5.1 Direct linear equation solvers . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.2 Iterative equation solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.3 Preconditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4 Analysis of the multi-level preconditioner . . . . . . . . . . . . . . . . . . . 114

6 Mixed Methods 119

6.1 Weak formulation of Dirichlet boundary conditions . . . . . . . . . . . . . 119
6.2 A Mixed method for the flux . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3 Abstract theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.4 Analysis of the model problems . . . . . . . . . . . . . . . . . . . . . . . . 124
6.5 Approximation of mixed systems . . . . . . . . . . . . . . . . . . . . . . . 131
6.6 Supplement on mixed methods for the flux : discrete norms, super-
convergence and implementation techniques . . . . . . . . . . . . . . . . . 134
6.6.1 Primal and dual mixed formulations . . . . . . . . . . . . . . . . . 134
6.6.2 Super-convergence of the scalar . . . . . . . . . . . . . . . . . . . . 136
6.6.3 Solution methods for the linear system . . . . . . . . . . . . . . . . 138
6.6.4 Hybridization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

7 Discontinuous Galerkin Methods 141

7.1 Transport equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.1.1 Solvability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.2 Discontinuous Galerkin Discretization . . . . . . . . . . . . . . . . . . . . . 143
7.3 Nitsche’s method for Dirichlet boundary conditions . . . . . . . . . . . . . 146
7.3.1 Nitsche’s method for interface conditions . . . . . . . . . . . . . . . 147
7.4 DG for second order equations . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.4.1 Hybrid DG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.4.2 Bassi-Rebay DG . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
7.4.3 Matching integration rules . . . . . . . . . . . . . . . . . . . . . . . 150
7.4.4 (Hybrid) DG for Stokes and Navier-Stokes . . . . . . . . . . . . . . 150
CONTENTS 5

8 Applications 153
8.1 The Navier Stokes equation . . . . . . . . . . . . . . . . . . . . . . . . . . 153
8.1.1 Proving LBB for the Stokes Equation . . . . . . . . . . . . . . . . . 155
8.1.2 Discrete LBB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.2 Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
8.3 Maxwell equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166

9 Parabolic partial differential equations 173

9.1 Semi-discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
9.2 Time integration methods . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
9.3 Space-time formulation of Parabolic Equations . . . . . . . . . . . . . . . . 179
9.3.1 Solvability of the continuous problem . . . . . . . . . . . . . . . . . 179
9.3.2 A first time-discretization method . . . . . . . . . . . . . . . . . . . 181
9.3.3 Discontinuous Galerkin method . . . . . . . . . . . . . . . . . . . . 181

10 Second order hyperbolic equations: wave equations 185

10.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
10.2 Time-stepping methods for wave equations . . . . . . . . . . . . . . . . . . 186
10.2.1 The Newmark time-stepping method . . . . . . . . . . . . . . . . . 186
10.2.2 Methods for the first order system . . . . . . . . . . . . . . . . . . . 188

11 Hyperbolic Conservation Laws 191

11.1 A little theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
11.1.1 Weak solutions and the Rankine-Hugoniot relation . . . . . . . . . 192
11.1.2 Expansion fans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
11.2 Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
6 CONTENTS
Chapter 1

Introduction

Differential equations are equations for an unknown function involving differential opera-
tors. An ordinary differential equation (ODE) requires differentiation with respect to one
variable. A partial differential equation (PDE) involves partial differentiation with respect
to two or more variables.

1.1 Classification of PDEs

The general form of a linear PDE of second order is: find u : Ω ⊂ Rd → R such that
d d
X ∂ ∂u(x)
X ∂u(x)
− ai,j (x) + bi (x) + c(x)u(x) = f (x). (1.1)
i,j=1
∂xi ∂xj i=1
∂x i

The coefficients ai,j (x), bi (x), c(x) and the right hand side f (x) are given functions. In
addition, certain type of boundary conditions are required. The behavior of the PDE
depends on the type of the differential operator
d d
X ∂ ∂ X ∂
L := ai,j + bi + c.
i,j=1
∂x i ∂x j i=1
∂x i

∂
Replace ∂xi
by si . Then
d
X d
X
si ai,j sj + b i si + c = 0
i,j=1 i=1

describes a quartic shape in Rd . We treat the following cases:

1. In the case of a (positive or negative) definite matrix a = (ai,j ) this is an ellipse,
and the corresponding PDE is called elliptic. A simple example is a = I, b = 0, and
c = 0, i..e.
X ∂ 2u
− = f.
i
∂x2i

7
8 CHAPTER 1. INTRODUCTION

Elliptic PDEs require boundary conditions.

2. If the matrix a is semi-definite, has the one-dimensional kernel span{v}, and b · v 6= 0,

then the shape is a parabola. Thus, the PDE is called parabolic. A simple example
is
d−1 2
X ∂ u ∂u
− 2
+ = f.
i=1
∂x i ∂x d

Often, the distinguished direction corresponds to time. This type of equation requires
boundary conditiosn on the d − 1-dimensional boundary, and initial conditions in the
different direction.

3. If the matrix a has d − 1 positive, and one negative (or vise versa) eigenvalues, then
the shape is a hyperbola. The PDE is called hyperbolic. The simplest one is
d−1 2
X ∂ u ∂ 2u
− + = f.
i=1
∂x2i ∂x2d

Again, the distinguished direction often corresponds to time. Now, two initial con-
ditions are needed.

4. If the matrix a is zero, then the PDE degenerates to the first order PDE
∂u
bi + cu = f.
∂xi
Boundary conditions are needed at a part of the boundary.
These cases behave very differently. We will establish theories for the individual cases.
A more general classicfication, for more positive or negative eigenvalues, and systems of
PDEs is possible. The type of the PDE may also change for different points x.

1.2 Weak formulation of the Poisson Equation

The most elementary and thus most popular PDE is the Poisson equation

−∆u = f in Ω, (1.2)

with the boundary conditions

u = uD on ΓD ,
∂u
∂n
= g on ΓN , (1.3)
∂u
∂n
+ αu = g on ΓR .

The domain Ω is an open and bounded subset of Rd , where the problem dimension d is
usually 1, 2 or 3. For d = 1, the equation is not a PDE, but an ODE. The boundary
1.2. WEAK FORMULATION OF THE POISSON EQUATION 9

Γ := ∂Ω consists of the three non-overlapping parts ΓD , ΓN , and P ΓR . The outer unit

∂2
normal vector is called n. The Laplace differential operator is ∆ := di=1 ∂x 2 , the normal
i
∂ d
:= i=1 ni ∂x∂ i . Given are the functions f , uD and g in
P
derivative at the boundary is ∂n
proper function spaces (e.g., f ∈ L2 (Ω)). We search for the unknown function u, again, in
a proper function space defined later.
The boundary conditions are called

• Dirichlet boundary condition on ΓD . The function value is prescribed,

• Neumann boundary condition on ΓN . The normal derivative is prescribed,

• Robin boundary condition on ΓR . An affine linear relation between the function

value and the normal derivative is prescribed.

Exactly one boundary condition must be specified on each part of the boundary.
We transform equation (1.2) together with the boundary conditions (1.3) into its weak
form. For this, we multiply (1.2) by smooth functions (called test functions) and integrate
over the domain: Z Z
− ∆uv dx = f v dx (1.4)
Ω Ω

We do so for sufficiently
R manyRtest functions v in a proper function space. Next, we apply
Gauss’ theorem Ω div p dx = Γ p · n ds to the function p := ∇u v to obtain
Z Z
div(∇u v) dx = ∇u · n v ds
Ω Γ

From the product rule there follows div(∇uv) = ∆uv + ∇u · ∇v. Together we obtain
Z Z Z
∂u
∇u · ∇v dx − v ds = f v dx.
Ω Γ ∂n Ω

Up to now, we only used the differential equation in the domain. Next, we incorporate
the boundary conditions. The Neumann and Robin b.c. are very natural (and thus are
∂u
called natural boundary conditions). We simply replace ∂n by g and −αu + g on ΓN and
ΓR , respectively. Putting unknown terms to the left, and known terms to the right hand
side, we obtain
Z Z Z Z Z
∂u
∇u · ∇v dx + αuv ds − v ds = f v dx + gv ds.
Ω ΓR ΓD ∂n ΓN +ΓR

Finally, we use the information of the Dirichlet boundary condition. We work brute force
and simple keep the Dirichlet condition in strong sense. At the same time, we only allow
test functions v fulfilling v = 0 on ΓD . We obtain the
10 CHAPTER 1. INTRODUCTION

Weak form of the Poisson equation:

Find u such that u = uD on ΓD and
Z Z Z Z
∇u · ∇v dx + αuv ds = f v dx + gv ds (1.5)
Ω ΓR Ω ΓN +ΓR

∀ v such that v = 0 on ΓD .

We still did not define the function space in which we search for the solution u. A proper
choice is
V := {v ∈ L2 (Ω) : ∇u ∈ [L2 (Ω)]d and u|Γ ∈ L2 (∂Ω)}.
It is a complete space, and, together with the inner product

(u, v)V := (u, v)L2 (Ω) + (∇u, ∇v)L2 (Ω) + (u, v)L2 (Γ)

it is a Hilbert space. Now, we see that f ∈ L2 (Ω) and g ∈ L2 (Γ) is useful. The Dirichlet
b.c. uD must be chosen such that there exists an u ∈ V with u = uD on ΓD . By definition
of the space, all terms are well defined. We will see later, that the problem indeed has a
unique solution in V .

1.3 The Finite Element Method

Now, we are developing a numerical method for approximating the weak form (1.5). For
this, we decompose the domain Ω into triangles T . We call the set T = {T } triangulation.
The set N = {xj } is the set of nodes. By means of this triangulation, we define the finite
element space, Vh :

Vh := {v ∈ C(Ω) : v|T is affine linear ∀ T ∈ T }

This is a sub-space of V . The derivatives (in weak sense, see below) are piecewise constant,
and thus, belong to [L2 (Ω)]2 . The function vh ∈ Vh is uniquely defined by its values v(xj )
in the nodes xj ∈ N . We decompose the set of nodes as

N = ND ∪ Nf ,

where ND are the nodes on the Dirichlet boundary, and Nf are all others (f as free). The
finite element approximation is defined as

Find uh such that uh (x) = uD (x) ∀ x ∈ ND and

Z Z Z Z
∇uh · ∇vh dx + αuh vh ds = f vh dx + gvh ds (1.6)
Ω ΓR ΓN +ΓR

∀ vh ∈ Vh such that vh (x) = 0 ∀ x ∈ ND

1.3. THE FINITE ELEMENT METHOD 11

Now it is time to choose a basis for Vh . The most convenient one is the nodal basis
{ϕi } characterized as
ϕi (xj ) = δi,j . (1.7)
The Kronecker-δ is defined to be 1 for i = j, and 0 else. These are the popular hat
functions. We represent the finite element solution with respect to this basis:
X
uh (x) = ui ϕi (x) (1.8)
P
By the nodal-basis property (1.7) there holds uh (xj ) = i ui ϕi (xj ) = uj . We have to
N
determine the coefficients ui ∈ R , with N = |N |. The ND := |ND | values according to
nodes on ΓD are given explicitely:
uj = uh (xj ) = uD (xj ) ∀ xj ∈ ΓD
The others have to be determined from the variational equation (1.6). It is equivalent to
fulfill (1.6) for the whole space {vh ∈ Vh : vh (xj ) = 0 ∀ xj ∈ ND }, or just for its basis
{ϕi : xi ∈ Nf } associated to the free nodes:

XnZ Z o Z Z
∇ϕi · ∇ϕj dx + αϕi ϕj ds ui = f ϕj dx + gϕj ds (1.9)
i Ω ΓR ΓN +ΓR

∀ ϕj such that xj ∈ Nf
We have inserted the basis expansion (1.8). We define the matrix A = (Aji ) ∈ RN ×N and
the vector f = (fj ) ∈ RN as
Z Z
Aji := ∇ϕi · ∇ϕj dx + αϕi ϕj ds,
Ω ΓR
Z Z
fj := f ϕj dx + gϕj ds.
ΓN +ΓR

According to Dirichlet- and free nodes they are splitted as

   
ADD ADf fD
A= and f= .
Af D Af f ff
Now, we obtain the system of linear equations for u = (ui ) ∈ RN , u = (uD , uf ):
    
I 0 uD uD
= . (1.10)
Af D Af f uf ff
At all, we have N coefficients ui . ND are given explicitely from the Dirichlet values. These
are Nf equations to determine the remaining ones. Using the known uD , we can reformulate
it as symmetric system of equations for uf ∈ RNf :
Af f uf = ff − Af D uD
12 CHAPTER 1. INTRODUCTION
Chapter 2

The abstract theory

In this chapter we develop the abstract framework for variational problems.

2.1 Basic properties

Definition 1. A vector space V is a set with the operations + : V × V → V and
· : R × V → V such that for all u, v ∈ V and λ, µ ∈ R there holds

• u+v =v+u

• (u + v) + w = u + (v + w)

• λ · (u + v) = λ · u + λ · v, (λ + µ) · u = λ · u + µ · u

Examples are Rn , the continuous functions C 0 , or the Lebesgue space L2 .

Definition 2. A normed vector space (V, k · k) is a vector space with the operation k.k :
V → R being a norm, i.e., for u, v ∈ V and λ ∈ R there holds

• ku + vk ≤ kuk + kvk

• kλ uk = |λ| kuk

• kuk = 0 ⇔ u = 0

Examples are (C 0 , k · ksup ), or (C 0 , k · kL2 ).

Definition 3. In a complete normed vector space, Cauchy sequences (un ) ∈ V N converge

to an u ∈ V . A complete normed vector space is called Banach space.

Examples of Banach spaces are (L2 , k · kL2 ), (C 0 , k · ksup ), but not (C 0 , k · kL2 ).
k·kV
Definition 4. The closure of a normed vector-space (W, k · kV ), denoted as W is the
smallest complete space containing W .

13
14 CHAPTER 2. THE ABSTRACT THEORY

k·kL2
Example: C = L2 .
Definition 5. A functional or a linear form l(·) on V is a linear mapping l(·) : V → R.
The canonical norm for linear forms is the dual norm
l(v)
klkV ∗ := sup .
06=v∈V kvk

A linear form l is called bounded if the norm is finite. The vector space of all bounded
linear forms on V is called the dual space V ∗ .
R
An example for a bounded linear form is l(·) : L2 → R : v → v dx.
Definition 6. A bilinear form A(·, ·) on V is a mapping A : V × V → R which is linear
in u and in v. It is called symmetric if A(u, v) = A(v, u) for all u, v ∈ V .
R
Examples are the bilinear form A(u, v) = uv dx on L2 , or A(u, v) := uT Av on Rn , where
A is a (symmetric) matrix.
Definition 7. A symmetric bilinear form A(·, ·) is called an inner product if it satisfies
• A(v, v) ≥ 0 ∀ v ∈ V

• A(v, v) = 0 ⇔ v = 0
Often, is is denoted as (·, ·)A , (·, ·)V , or simply (·, ·).
An example on Rn is uT Av, where A is a symmetric and positive definite matrix.
Definition 8. An inner product space is a vector space V together with an inner product
(·, ·)V .
Lemma 9. Cauchy Schwarz inequality. If A(·, ·) is a symmetric bilinear form such
that A(v, v) ≥ 0 for all v ∈ V , then there holds

A(u, v) ≤ A(u, u)1/2 A(v, v)1/2

Proof: For t ∈ R there holds

0 ≤ A(u − tv, u − tv) = A(u, u) − 2tA(u, v) + t2 A(v, v)

If A(v, v) = 0, then A(u, u) − 2tA(u, v) ≥ 0 for all t ∈ R, which forces A(u, v) = 0, and the
inequality holds trivially. Else, if A(v, v) 6= 0, set t = A(u, v)/A(v, v), and obtain

0 ≤ A(u, u) − A(u, v)2 /A(v, v),

which is equivalent to the statement. 2

1/2
Lemma 10. kvkV := (v, v)V defines a norm on the inner product space (V, (·, ·)V ).
2.1. BASIC PROPERTIES 15

Definition 11. An inner product space (V, (·, ·)V ) which is complete with respect to k · kV
is called a Hilbert space.

Definition 12. A closed subspace S of an Hilbert space V is a subset which is a vector

space, and which is complete with respect to k · kV .

A finite dimensional subspace is always a closed subspace.

Lemma 13. Let T be a continuous linear operator from the Hilbert space V to the Hilbert
space W . The kernel of T , ker T := {v ∈ V : T v = 0} is a closed subspace of V .

Proof: First we observe that ker T is a vector space. Now, let (un ) ∈ ker T N converge
to u ∈ V . Since T is continuous, T un → T u, and thus T u = 0 and u ∈ ker T . 2

Lemma 14. Let S be a subspace (not necessarily closed) of V . Then

S ⊥ := {v ∈ V : (v, w) = 0 ∀ w ∈ S}

is a closed subspace.

The proof is similar to Lemma 13.

Definition 15. Let V and W be vector spaces. A linear operator T : V → W is a linear

mapping from V to W . The operator is called bounded if its operator-norm

kT vkW
kT kV →W := sup
06=v∈V kvkV

is finite.
d
An example is the differential operator on the according space dx
: (C 1 (0, 1), k · ksup +
d
k dx · ksup ) → (C(0, 1), k · ksup ).

Lemma 16. A bounded linear operator is continuous.

Proof. Let vn → v, i.e. kvn − vkV → 0. Then kT vn − T vk ≤ kT kV →W kvn − vkV converges

to 0, i.e. T vn → T v. Thus T is continuous.

Definition 17. A dense subspace S of V is such that every element of V can be ap-
proximated by elements of S, i.e.

∀ε > 0 ∀u ∈ V ∃v ∈ S such that ku − vkV ≤ ε.

Lemma 18 (extension principle). Let S be a dense subspace of the normed space V , and
let W be a complete space. Let T : S → W be a bounded linear operator with respect to the
norm kT kV →W . Then, the operator can be uniquely extended onto V .
16 CHAPTER 2. THE ABSTRACT THEORY

Proof. Let u ∈ V , and let vn be a sequence such that vn → u. Thus, vn is Cauchy. T vn

is a well defined sequence in W . Since T is continuous, T vn is also Cauchy. Since W
is complete, there exists a limit w such that T vn → w. The limit is independent of the
sequence, and thus T u can be defined as the limit w.
Definition 19. A bounded linear operator T : V → W is called compact if for every
bounded sequence (un ) ∈ V N , the sequence (T un ) contains a convergent sub-sequence.
Lemma 20. Let V, W be Hilbert spaces. An operator is compact if and only if there exists
a complete orthogonal system (un ) for (ker T )⊥ and values λn → 0 such that

(un , um )V = δn,m (T un , T um )W = λn δn,m

This is the eigensystem of the operator K : V → V ∗ : u 7→ (T u, T ·)W .

(T v,T v)W
Proof. (sketch) There exists an maximizing element of (v,v)V
. Scale it to kvkV = 1
(T u1 ,T u1 )W
and call it u1 , and λ1 = (u1 ,u1 )V
. Repeat the procedure on the V -complement of u1 to
generate u2 , and so on.

2.2 Projection onto subspaces

In the Euklidean space R2 one can project orthogonally onto a line through the origin, i.e.,
onto a sub-space. The same geometric operation can be defined for closed subspaces of
Hilbert spaces.
Theorem 21. Let S be a closed subspace of the Hilbert space V . Let u ∈ V . Then there
exists a unique closest point u0 ∈ S:

ku − u0 k ≤ ku − vk ∀v ∈ S

There holds
u − u0 ⊥ S
Proof: Let d := inf v∈S ku−vk, and let (vn ) be a minimizing sequence such that ku−vn k → d.
We first check that there holds

kvn − vm k2 = 2 kvn − uk2 + 2 kvm − uk2 − 4 k1/2(vn + vm ) − uk2 .

Since 1/2(vn +vm ) ∈ S, there holds k1/2(vn +vm )−uk ≥ d. We proof that (vn ) is a Cauchy
sequence: Fix ε > 0, choose N ∈ N such that for n > N there holds ku − vn k2 ≤ d2 + ε2 .
Thus for all n, m > N there holds

kvn − vm k2 ≤ 2(d2 + ε2 ) + 2(d2 + ε2 ) − 4d2 = 4ε2 .

Thus, vn converge to some u0 ∈ V . Since S is closed, u0 ∈ S. By continuity of the norm,

ku − u0 k = d.
2.3. RIESZ REPRESENTATION THEOREM 17

Fix some 0 6= w ∈ S, and define ϕ(t) := ku − u0 − tw k2 . ϕ(·) is a convex function, it

| {z }
∈S
takes its unique minimum d at t = 0. Thus

dϕ(t)
0= |t=0 = {2(u − u0 , w) − 2t(w, w)}|t=0 = 2(u − u0 , w)
dt
We obtained u − u0 ⊥S. If there were two minimizers u0 6= u1 , then u0 − u1 = (u0 − u) −
(u1 − u)⊥S and u0 − u1 ∈ S, which implies u0 − u1 = 0, a contradiction. 2

Theorem 21 says that given an u ∈ V , we can uniquely decompose it as

u = u0 + u1 , u0 ∈ S u1 ∈ S ⊥

This allows to define the operators PS : V → S and PS⊥ : V → S ⊥ as

PS u := u0 PS⊥ u := (I − PS )u = u1

Theorem 22. PS and PS⊥ are linear operators.

Definition 23. A linear operator P is called a projection if P 2 = P . A projector is

called orthogonal, if (P u, v) = (u, P v).

Lemma 24. The operators PS and PS⊥ are both orthogonal projectors.

Proof: For u ∈ S there holds PS u = u. Since PS u ∈ S, there holds PS2 u = PS u. It is

orthogonal since

(PS u, v) = (PS u, v − PS v + PS v) = (PS u , v − PS v ) + (PS u, PS v) = (PS u, PS v).

|{z} | {z }
∈S ∈S ⊥

With the same argument there holds (u, PS v) = (PS u, PS v). The co-projector PS⊥ = I − PS
is a projector since
(I − PS )2 = I − 2PS + PS2 = I − PS .
It is orthogonal since ((I − PS )u, v) = (u, v) − (PS u, v) = (u, v) − (u, PS v) = (u, (I − PS )v)
2

2.3 Riesz Representation Theorem

Let u ∈ V . Then, we can define the related continuous linear functional lu (·) ∈ V ∗ by

lu (v) := (u, v)V ∀ v ∈ V.

The opposite is also true:

18 CHAPTER 2. THE ABSTRACT THEORY

Theorem 25. Riesz Representation Theorem. Any continuous linear functional l on

a Hilbert space V can be represented uniquely as

l(v) = (ul , v) (2.1)

for some ul ∈ V . Furthermore, we have

klkV ∗ = kul kV .

Proof: First, we show uniqueness. Assume that u1 6= u2 both fulfill (2.1). This leads to
the contradiction

0 = l(u1 − u2 ) − l(u1 − u2 )
= (u1 , u1 − u2) − (u2 , u1 − u2 ) = ku1 − u2 k2 .

Next, we construct the ul . For this, define S := ker l. This is a closed subspace.
Case 1: S ⊥ = {0}. Then, S = V , i.e., l = 0. So take ul = 0.
Case 2: S ⊥ 6= {0}. Pick some 0 6= z ∈ S ⊥ . There holds l(z) 6= 0 (otherwise, z ∈ S ∩ S ⊥ =
{0}). Now define
l(z)
ul := 2
z ∈ S⊥
kzk
Then

(ul , v) = ( ul , v − l(v)/l(z)z ) + (ul , l(v)/l(z)z)

|{z} | {z }
S⊥ S
2
= l(z)/kzk (z, l(v)/l(z)z)
= l(v)

Finally, we prove klkV ∗ = kul kV :

l(v) (ul , v)V

klkV ∗ = sup = sup ≤ kul kV
06=v∈V kvk v kvkV

and
l(z) l(z)
kul k = 2
kzk = ≤ klkV ∗ .
kzk kzk

2.4 Symmetric variational problems

Take the function space C 1 (Ω), and define the bilinear form
Z Z
A(u, v) := ∇u∇v + uv ds
Ω Γ
2.5. COERCIVE VARIATIONAL PROBLEMS 19

and the linear form Z

f (v) := f v dx
Ω

The bilinear form is non-negative, and A(u, u) = 0 implies u = 0. Thus A(·, ·) is an inner
product, and provides the norm kvkA := A(v, v)1/2 . The normed vector space (C 1 , k.kA ) is
not complete. Define
k.kA
V := C 1 (Ω) ,
which is a Hilbert space per definition. If we can show that there exists a constant c such
that Z
f (v) = f v dx ≤ ckvkA ∀v ∈ V
Ω

then f (.) is a continuous linear functional on V . We will prove this later. In this case, the
Riesz representation theorem tells that there exists an unique u ∈ V such that

A(u, v) = f (v).

This shows that the weak form has a unique solution in V .

Next, take the finite dimensional (⇒ closed) finite element subspace Vh ⊂ V . The finite
element solution uh ∈ Vh was defined by

A(uh , vh ) = f (vh ) ∀ vh ∈ Vh ,

This means
A(u − uh , vh ) = A(u, vh ) − A(uh , vh ) = f (vh ) − f (vh ) = 0
uh is the projection of u onto Vh , i.e.,

ku − uh kA ≤ ku − vh kA ∀ vh ∈ Vh

The error u − uh is orthogonal to Vh .

2.5 Coercive variational problems

In this chapter we discuss variational problems posed in Hilbert spaces. Let V be a Hilbert
space, and let A(·, ·) : V × V → R be a bilinear form which is

• coercive (also known as elliptic)

A(u, u) ≥ α1 kuk2V ∀ u ∈ V, (2.2)

• and continuous
A(u, v) ≤ α2 kukV kvkV ∀ u, v ∈ V, (2.3)
20 CHAPTER 2. THE ABSTRACT THEORY

with bounds α1 and α2 in R+ . It is not necessarily symmetric. Let f (.) : V → R be a

continuous linear form on V , i.e.,

f (v) ≤ kf kV ∗ kvkV .

We are posing the variational problem: find u ∈ V such that

A(u, v) = f (v) ∀ v ∈ V.

Example 26. Diffusion-reaction equation:

Consider the PDE

− div(a(x)∇u) + c(x)u = f in Ω,
with Neumann boundary conditions. Let V be the Hilbert space generated by the inner
product (u, v)V := (u, v)L2 + (∇u, ∇v)L2 . The variational formulation of the PDE involves
the bilinear form Z Z
A(u, v) = (a(x)∇u) · ∇v dx + c(x)uv dx.
Ω Ω

Assume that the coefficients a(x) and c(x) fulfill a(x) ∈ Rd×d , a(x) symmetric and λ1 ≤
λmin (a(x)) ≤ λmax (a(x)) ≤ λ2 , and c(x) such that γ1 ≤ c(x) ≤ γ2 almost everywhere. Then
A(·, ·) is coercive with constant α1 = min{λ1 , γ1 } and α2 = max{λ2 , γ2 }.

Example 27. Diffusion-convection-reaction equation:

The partial differential equation

−∆u + b · ∇u + u = f in Ω

with Dirichlet boundary conditions u = 0 on ∂Ω leads to the bilinear form

Z Z Z
A(u, v) = ∇u ∇v dx + b · ∇u v dx + uv dx.

If div b ≤ 0, what is an important case arising from incompressible flow fields (div b = 0),
then A(·, ·) is coercive and continuous w.r.t. the same norm as above.

Instead of the linear form f (·), we will often write f ∈ V ∗ . The evaluation is written
as the duality product
hf, viV ∗ ×V = f (v).

Lemma 28. A continuous bilinear form A(·, ·) : V × V → R induces a continuous linear

operator A : V → V ∗ via

hAu, vi = A(u, v) ∀ u, v ∈ V.

The operator norm kAkV →V ∗ is bounded by the continuity bound α2 of A(·, ·).
2.5. COERCIVE VARIATIONAL PROBLEMS 21

Proof: For every u ∈ V , A(u, ·) is a bounded linear form on V with norm

A(u, v) α2 kukV kvkV

kA(u, ·)kV ∗ = sup ≤ sup = α2 kukV
v∈V kvkV v∈V kvkV

Thus, we can define the operator A : u ∈ V → A(u, ·) ∈ V ∗ . It is linear, and its operator
norm is bounded by

kAukV ∗ hAu, viV ∗ ×V

kAkV →V ∗ = sup = sup sup
u∈V kukV u∈V v∈V kukV kvkV
A(u, v) α2 kukV kvkV
= sup sup ≤ sup sup = α2 .
u∈V v∈V kukV kvkV u∈V v∈V kukV kvkV

2
Using this notation, we can write the variational problem as operator equation: find
u ∈ V such that
Au = f (in V ∗ ).
Theorem 29 (Banach’s contraction mapping theorem). Given a Banach space V and a
mapping T : V → V , satisfying the Lipschitz condition

kT (v1 ) − T (v2 )k ≤ L kv1 − v2 k ∀ v1 , v2 ∈ V

for a fixed L ∈ [0, 1). Then there exists a unique u ∈ V such that

u = T (u),

i.e. the mapping T has a unique fixed point u. The iteration u1 ∈ V given, compute

uk+1 := T (uk )

converges to u with convergence rate L:

ku − uk+1 k ≤ Lku − uk k

Theorem 30 (Lax Milgram). Given a Hilbert space V , a coercive and continuous bilinear
form A(·, ·), and a continuous linear form f (.). Then there exists a unique u ∈ V solving

A(u, v) = f (v) ∀ v ∈ V.

There holds
kukV ≤ α1−1 kf kV ∗ (2.4)
Proof: Start from the operator equation Au = f . Let JV : V ∗ → V be the Riesz
isomorphism defined by

(JV g, v)V = g(v) ∀ v ∈ V, ∀ g ∈ V ∗ .

22 CHAPTER 2. THE ABSTRACT THEORY

Then the operator equation is equivalent to

JV Au = JV f (in V ),

and to the fixed point equation (with some 0 6= τ ∈ R chosen below)

u = u − τ JV (Au − f ). (2.5)

We will verify that

T (v) := v − τ JV (Av − f )
is a contraction mapping, i.e., kT (v1 )−T (v2 )kV ≤ Lkv1 −v2 kV with some Lipschitz constant
L ∈ [0, 1). Let v1 , v2 ∈ V , and set v = v1 − v2 . Then

kT (v1 ) − T (v2 )k2V = k{v1 − τ JV (Av1 − f )} − {v2 − τ JV (Av2 − f )}k2V

= kv − τ JV Avk2V
= kvk2V − 2τ (JV Av, v)V + τ 2 kJV Avk2V
= kvk2V − 2τ hAv, vi + τ 2 kAvk2V ∗
= kvk2V − 2τ A(v, v) + τ 2 kAvk2V ∗
≤ kvk2V − 2τ α1 kvk2V + τ 2 α22 kvk2V
= (1 − 2τ α1 + τ 2 α22 )kv1 − v2 k2V

Now, we choose τ = α1 /α22 , and obtain a Lipschitz constant

L2 = 1 − α12 /α22 ∈ [0, 1).

Banach’s contraction mapping theorem state that (2.5) has a unique fixed point. Finally,
we obtain the bound (2.4) from

kuk2V ≤ α1−1 A(u, u) = α1−1 f (u) ≤ α1−1 kf kV ∗ kukV ,

and dividing by one factor kuk. 2

2.5.1 Approximation of coercive variational problems

Now, let Vh be a closed subspace of V . We compute the approximation uh ∈ Vh by the
Galerkin method
A(uh , vh ) = f (vh ) ∀ vh ∈ Vh . (2.6)
This variational problem is uniquely solvable by Lax-Milgram, since, (Vh , k.kV ) is a Hilbert
space, and continuity and coercivity on Vh are inherited from the original problem on V .
The next theorem says, that the solution defined by the Galerkin method is, up to a
constant factor, as good as the best possible approximation in the finite dimensional space.
Theorem 31 (Cea). The approximation error of the Galerkin method is quasi optimal

ku − uh kV ≤ α2 /α1 inf ku − vh kV
v∈Vh
2.6. INF-SUP STABLE VARIATIONAL PROBLEMS 23

Proof: A fundamental property is the Galerkin orthogonality

A(u − uh , wh ) = A(u, wh ) − A(uh , wh ) = f (wh ) − f (wh ) = 0 ∀ wh ∈ Vh .

Now, pick an arbitrary vh ∈ Vh , and bound

ku − uh k2V ≤ α1−1 A(u − uh , u − uh )

= α1−1 A(u − uh , u − vh ) + α1−1 A(u − uh , vh − uh )
| {z }
∈Vh
≤ α2 /α1 ku − uh kV ku − vh kV .

Divide one factor ku − uh k. Since vh ∈ Vh was arbitrary, the estimation holds true also for
the infimum in Vh . 2

If A(·, ·) is additionally symmetric, then it is an inner product. In this case, the coer-
civity and continuity properties are equivalent to

α1 kuk2V ≤ A(u, u) ≤ α2 kuk2V ∀ u ∈ V.

The generated norm k.kA is an equivalent norm to k.kV . In the symmetric case, we can
use the orthogonal projection with respect to (., .)A to improve the bounds to

ku − uh k2V ≤ α1−1 ku − uh k2A ≤ α1−1 inf ku − vh k2A ≤ α2 /α1 ku − vh k2V .

vh ∈Vh

The factor in the quasi-optimality estimate is now the square root of the general, non-
symmetric case.

2.6 Inf-sup stable variational problems

The coercivity condition is by no means a necessary condition for a stable solvable system.
A simple, stable problem with non-coercive bilinear form is to choose V = R2 , and the
bilinear form B(u, v) = u1 v1 −u2 v2 . The solution of B(u, v) = f T v is u1 = f1 and u2 = −f2 .
We will follow the convention to call coercive bilinear forms A(·, ·), and the more general
ones B(·, ·).
Let V and W be Hilbert spaces, and B(·, ·) : V × W → R be a continuous bilinear form
with bound
B(u, v) ≤ β2 kukV kvkW ∀ u ∈ V, ∀ v ∈ W. (2.7)
The general condition is the inf-sup condition

B(u, v)
inf sup ≥ β1 . (2.8)
u∈V v∈W
u6=0 v6=0
kukV kvkW
24 CHAPTER 2. THE ABSTRACT THEORY

Define the linear operator B : V → W ∗ by hBu, viW ∗ ×W = B(u, v). The inf-sup
condition can be reformulated as
hBu, vi
sup ≥ β1 kukV , ∀u ∈ V
v∈W kvkW
and, using the definition of the dual norm,

kBukW ∗ ≥ β1 kukV . (2.9)

We immediately obtain that B is one to one, since

Bu = 0 ⇒ u = 0

Lemma 32. Assume that the continuous bilinear form B(·, ·) fulfills the inf-sup condition
(2.8). Then the according operator B has closed range.
Proof: Let Bun be a Cauchy sequence in W ∗ . From (2.9) we conclude that also un is
Cauchy in V . Since V is complete, un converges to some u ∈ V . By continuity of B, the
sequence Bun converges to Bu ∈ W ∗ . 2
∗
The inf-sup condition (2.8) does not imply that B is onto W . To insure that, we can
pose an inf-sup condition the other way around:
B(u, v)
inf sup ≥ β1 . (2.10)
v∈W u∈V
v6=0 u6=0
kukV kvkW

It will be sufficient to state the weaker condition

B(u, v)
sup >0 ∀ v ∈ W. (2.11)
u∈V kukV kvkW
u6=0

Theorem 33. Assume that the continuous bilinear form B(·, ·) fulfills the inf-sup condition
(2.8) and condition (2.11). Then, the variational problem: find u ∈ V such that

B(u, v) = f (v) ∀v ∈ W (2.12)

has a unique solution. The solution depends continuously on the right hand side:

kukV ≤ β1−1 kf kW ∗

Proof: We have to show that the range R(B) = W ∗ . The Hilbert space W ∗ can be split
into the orthogonal, closed subspaces

W ∗ = R(B) ⊕ R(B)⊥ .

Assume that there exists some 0 6= g ∈ R(B)⊥ . This means that

(Bu, g)W ∗ = 0 ∀ u ∈ V.
2.6. INF-SUP STABLE VARIATIONAL PROBLEMS 25

Let vg ∈ W be the Riesz representation of g, i.e., (vg , w)W = g(w) for all w ∈ W . This vg
is in contradiction to the assumption (2.11)
B(u, vg ) (Bu, g)W ∗
sup = sup = 0.
u∈V kukV u∈V kukV
Thus, R(B)⊥ = {0} and R(B) = W ∗ . 2
Example 34. A coercive bilinear form is inf-sup stable.
Example 35. A complex symmetric variational problem:
Consider the complex valued PDE
−∆u + iu = f,
√
with Dirichlet boundary conditions, f ∈ L2 , and i = −1. The weak form for the real
system u = (ur , ui ) ∈ V 2 is
(∇ur , ∇vr )L2 + (ui , vr )L2 = (f, vr ) ∀ vr ∈ V
(2.13)
(ur , vi )L2 − (∇ui , ∇vi )L2 = −(f, vi ) ∀ vi ∈ V
We can add up both lines, and define the large bilinear form B(·, ·) : V 2 × V 2 → R by
B((ur , ui ), (vr , vi )) = (∇ur , ∇vr ) + (ui , vr ) + (ur , vi ) − (∇ui , ∇vi )
With respect to the norm kvkV = (kvk2L2 +k∇vk2L2 )1/2 , the bilinear form is continuous, and
fulfills the inf-sup conditions (exercises !) Thus, the variational formulation: find u ∈ V 2
such that
B(u, v) = (f, vr ) − (f, vi ) ∀v ∈ V 2
is stable solvable.

2.6.1 Approximation of inf-sup stable variational problems

Again, to approximate (2.12), we pick finite dimensional subspaces Vh ⊂ V and Wh ⊂ W ,
and pose the finite dimensional variational problem: find uh ∈ Vh such that
B(uh , vh ) = f (vh ) ∀ vh ∈ Wh .
But now, in contrast to the coercive case, the solvability of the finite dimensional equation
does not follow from the solvability conditions of the original problem on V × W . E.g.,
take the example in R2 above, and choose the subspaces Vh = Wh = span{(1, 1)}.
We have to pose an extra inf-sup condition for the discrete problem:
B(uh , vh )
inf sup ≥ β1h . (2.14)
uh ∈Vh v ∈W
uh 6=0 h h kuh kV kvh kW
vh 6=0

On a finite dimensional space, one to one is equivalent to onto, and we can skip the second
condition.
26 CHAPTER 2. THE ABSTRACT THEORY

Theorem 36. Assume that B(·, ·) is continuous with bound β2 , and B(·, ·) fulfills the
discrete inf-sup condition with bound β1h . Then there holds the quasi-optimal error estimate

ku − uh k ≤ (1 + β2 /β1h ) inf ku − vh k (2.15)

vh ∈Vh

Proof: Again, there holds the Galerkin orthogonality B(u, wh ) = B(uh , wh ) for all wh ∈ Vh .
Again, choose an arbitrary vh ∈ Vh :

ku − uh kV ≤ ku − vh kV + kvh − uh kV
−1 B(vh − uh , wh )
≤ ku − vh kV + β1h sup
wh ∈Wh kwh kV
−1 B(vh − u, wh )
= ku − vh kV + β1h sup
wh ∈Wh kwh kV
−1 β2 kvh − ukV kwh kW
≤ ku − vh kV + β1h sup
wh ∈Wh kwh kW
= (1 + β2 /β1h )ku − vh kV .
Chapter 3

Sobolev Spaces

In this section, we introduce the concept of generalized derivatives, we define families of

normed function spaces, and prove inequalities between them. Let Ω be an open subset of
Rd , either bounded or unbounded.

3.1 Generalized derivatives

Let α = (α1 , . . . , αd ) ∈ Nd0 be a multi-index, |α| = αi , and define the classical differential
P
operator for functions in C ∞ (Ω)
  α1  αd
α ∂ ∂
D = ··· .
∂x1 ∂xn

For a function u ∈ C(Ω), the support is defined as

supp{u} := {x ∈ Ω : u(x) 6= 0}.

This is a compact set if and only if it is bounded. We say u has compact support in Ω, if
supp u ⊂ Ω. If Ω is a bounded domain, then u has compact support in Ω if and only if u
vanishes in a neighbourhood of ∂Ω.
The space of smooth functions with compact support is denoted as

D(Ω) := C0∞ (Ω) := {u ∈ C ∞ (Ω) : u has compact support in Ω}. (3.1)

For a smooth function u ∈ C |α| (Ω), there holds the formula of integration by parts
Z Z
α |α|
D uϕ dx = (−1) uDα ϕ dx ∀ ϕ ∈ D(Ω). (3.2)
Ω Ω

The L2 inner product with a function u in C(Ω) defines the linear functional on D
Z
u(ϕ) := hu, ϕiD0 ×D := uϕ dx.
Ω

27
28 CHAPTER 3. SOBOLEV SPACES

We call these functionals in D0 distributions. When u is a function, we identify it with

the generated distribution. The formula (3.2) is valid for functions u ∈ C α . The strong
regularity is needed only on the left hand side. Thus, we use the less demanding right hand
side to extend the definition of differentiation for distributions:

Definition 37. For u ∈ D0 , we define g ∈ D0 to be the generalized derivative Dgα u of u by

hg, ϕiD0 ×D = (−1)|α| hu, Dα ϕiD0 ×D ∀ϕ ∈ D

If u ∈ C α , then Dgα coincides with Dα .

The function space of locally integrable functions on Ω is called

Lloc
1 (Ω) = {u : uK ∈ L1 (K) ∀ compact K ⊂ Ω}.

1/x
It contains functions which can behave very badly near ∂Ω. E.g., ee is in L1loc (0, 1). If Ω
is unbounded, then the constant function 1 is in Lloc
1 , but not in L1 .

Definition 38. For u ∈ Lloc α loc

1 , we call g the weak derivative Dw u, if g ∈ L1 satisfies
Z Z
|α|
g(x)ϕ(x) dx = (−1) u(x)Dα ϕ(x) dx ∀ ϕ ∈ D.
Ω Ω

The weak derivative is more general than the classical derivative, but more restrictive
than the generalized derivative.

Example 39. Let Ω = (−1, 1) and

 
1+x x≤0
u(x) =
1−x x>0

Then,
 
1 x≤0
g(x) =
−1 x>0

is the first generalized derivative Dg1 of u, which is also a weak derivative. The second
generalized derivative h is

hh, ϕi = −2ϕ(0) ∀ϕ ∈ D

It is not a weak derivative.

In the following, we will focus on weak derivatives. Unless it is essential we will skip
the sub-scripts w and g.
3.2. SOBOLEV SPACES 29

3.2 Sobolev spaces

For k ∈ N0 and 1 ≤ p < ∞, we define the Sobolev norms
 1/p
X
kukWpk (Ω) :=  kDα ukpLp  ,
|α|≤k

for k ∈ N0 we set
α
k (Ω) := max kD ukL∞ .
kukW∞
|α|≤k

In both cases, we define the Sobolev spaces via

Wpk (Ω) = {u ∈ Lloc
1 : kukWpk < ∞}

In the previous chapter we have seen the importance of complete spaces. This is the
case for Sobolev spaces:
Theorem 40. The Sobolev space Wpk (Ω) is a Banach space.
Proof: Let vj be a Cauchy sequence with respect to k · kWpk . This implies that Dα vj is a
Cauchy sequence in Lp , and thus converges
R to some v α Rin k.kLp .
We verify that Dα vj → v α implies Ω Dα vj ϕ dx → Ω v α ϕ dx for all ϕ ∈ D. Let K be
the compact support of ϕ. There holds
Z Z
α α
(D vj − v )ϕ dx = (Dα vj − v α )ϕ dx
Ω K
≤ kDα vj − v α kL1 (K) kϕkL∞
≤ kDα vj − v α kLp (K) kϕkL∞ → 0
Finally, we have to check that v α is the weak derivative of v:
Z Z
α
v ϕ dx = lim Dα vj ϕ dx
j→∞ Ω
Z
|α|
= lim (−1) vj Dα ϕ dx =
j→∞ Ω
Z
= (−1)α vDα ϕ dx.
Ω
2
An alternative definition of Sobolev spaces were to take the closure of smooth functions
in the domain, i.e.,
f k := {C ∞ (Ω) : k.kW k ≤ ∞}k.kWpk .
W p p

A third one is to take continuously differentiable functions up to the boundary

k.kW k
cpk := C ∞ (Ω)
W p
.
Under moderate restrictions, these definitions lead to the same spaces:
30 CHAPTER 3. SOBOLEV SPACES

fk = W k .
Theorem 41. Let 1 ≤ p < ∞. Then W p p

Definition 42. The domain Ω has a Lipschitz boundary, ∂Ω, if there exists a collection
of open sets Oi , a positive parameter ε, an integer N and a finite number L, such that for
all x ∈ ∂Ω the ball of radius ε centered at x is contained in some Oi , no more than N of
the sets Oi intersect non-trivially, and each part of the boundary Oi ∩ Ω is a graph of a
Lipschitz function ϕi : Rd−1 → R with Lipschitz norm bounded by L.
Theorem 43. Assume that Ω has a Lipschitz boundary, and let 1 ≤ p < ∞. Then
ck = W k .
Wp p

The case W2k is special, it is a Hilbert space. We denote it by

H k (Ω) := W2k (Ω).
The inner product is X
(u, v)H k := (Dα u, Dα v)L2
|α|≤k

In the following, we will prove most theorems for the Hilbert spaces H k , and state the
general results for Wpk .

3.3 Trace theorems and their applications

We consider boundary values of functions in Sobolev spaces. Clearly, this is not well defined
for H 0 = L2 . But, as we will see, in H 1 and higher order Sobolev spaces, it makes sense
to talk about u|∂Ω . The definition of traces is essential to formulate boundary conditions
of PDEs in weak form.
We start in one dimension. Let u ∈ C 1 ([0, h]) with some h > 0. Then, we can bound
Z h n o0
 x x
u(0) = 1 − u(x)|x=0 = − 1− u(x) dx
h 0 h
Z h
−1  x 0

= − u(x) + 1 − u (x) dx
0 h h
1 x
≤ kukL2 + 1 − ku0 kL2
h L2 h L2
' h−1/2 kukL2 (0,h) + h1/2 ku0 kL2 (0,h) .
This estimate includes the scaling with the interval length h. If we are not interested in the
scaling, we apply Cauchy-Schwarz in R2 , and combine the L2 norm and the H 1 semi-norm
ku0 kL2 to the full H 1 norm and obtain
p q
|u(0)| ≤ h−1/2 + h1/2 kuk2L2 + ku0 k2L2 = c kukH 1 .

Next, we extend the trace operator to the whole Sobolev space H 1 :

3.3. TRACE THEOREMS AND THEIR APPLICATIONS 31

Theorem 44. There is a well defined and continuous trace operator

tr : H 1 ((0, h)) → R

whose restriction to C 1 ([0, h]) coincides with

u → u(0).

Proof: Use that C 1 ([0, h]) is dense in H 1 (0, h). Take a sequence uj in C 1 ([0, h]) con-
verging to u in H 1 -norm. The values uj (0) are Cauchy, and thus converge to an u0 . The
limit is independent of the choice of the sequence uj . This allows to define tr u := u0 . 2
Now, we extend this 1D result to domains in more dimensions. Let Ω be bounded, ∂Ω
be Lipschitz, and consists of M pieces Γi of smoothness C 1 .
We can construct the following covering of a neighbourhood of ∂Ω in Ω: Let Q = (0, 1)2 .
For 1 ≤ i ≤ M , let si ∈ C 1 (Q, Ω) be invertible and such that ks0i kL∞ ≤ c, k(s0i )−1 kL∞ ≤ c,
and det s0i > 0. The domains Si := si (Q) are such that si ((0, 1) × {0}) = Γi , and the
parameterizations match on si ({0, 1} × (0, 1)).

Theorem 45. There exists a well defined and continuous operator

tr : H 1 (Ω) → L2 (∂Ω)

which coincides with u|∂Ω for u ∈ C 1 (Ω).

Proof: Again, we prove that

tr : C 1 (Ω) → L2 (∂Ω) : u → u|∂Ω

is a bounded operator w.r.t. the norms k.kH 1 and L2 , and conclude by density. We use
the partitioning of ∂Ω into the pieces Γi , and transform to the simple square domain
Q = (0, 1)2 . Define the functions ui on Q = (0, 1)2 as

ũi (x̃) = u(si (x̃))

We transfer the L2 norm to the simple domain:

M Z
X
k tr uk2L2 (∂Ω) = u(x)2 dx
i=1 Γi
M Z 1
X ∂si
= u(si (ξ, 0))2 (ξ, 0) dξ
i=1 0
∂ξ
XM Z 1
≤ c ũi (ξ, 0)2 dξ
i=1 0
32 CHAPTER 3. SOBOLEV SPACES

To transform the H 1 -norm, we differentiate with respect to x̃ by applying the chain rule
dũi du dsi
(x̃) = (si (x̃)) (x̃).
dx̃ dx dx̃
du
Solving for dx
is
 −1
du dũi ds
(si (x̃)) = (x̃) (x̃)
dx dx̃ dx̃
The bounds onto s0 and (s0 )−1 imply that

c−1 |∇x u| ≤ |∇x̃ ũ| ≤ c |∇x u|

We start from the right hand side of the stated estimate:

M Z
X
kuk2H 1 (Ω) ≥ |∇x u|2 dx
i=1 Si
XM Z
= |∇x u(si (x̃))|2 det(s0 ) dx̃
i=1 Q
M Z
X
≥ c |∇x̃ ũ(x̃)|2 dx̃
i=1 Q

We got a lower bound for det(s0 ) = (det(s0 )−1 )−1 from the upper bound for (s0 )−1 .
It remains to prove the trace estimate on Q. Here, we apply the previous one dimen-
sional result
Z 1(  2 )
∂u(ξ, η)
|u(ξ, 0)|2 ≤ c u(ξ, η)2 + dη ∀ ξ ∈ (0, 1)
0 ∂η

The result follows from integrating over ξ

( 2 )
Z 1 Z 1 Z 1 
∂u(ξ, η)
|u(ξ, 0)|2 dξ ≤ c u(ξ, η)2 + dη dξ
0 0 0 ∂η
≤ c kuk2H 1 (Q) .

Considering the trace operator from H 1 (Ω) to L2 (∂Ω) is not sharp with respect to the
norms. We will improve the embedding later.

By means of the trace operator we can define the sub-space

H01 (Ω) = {u ∈ H 1 (Ω) : tr u = 0}

3.3. TRACE THEOREMS AND THEIR APPLICATIONS 33

It is a true sub-space, since u = 1 does belong to H 1 , but not to H01 . It is a closed sub-space,
since it is the kernel of a continuous operator.

By means of the trace inequality, one verifies that the linear functional
Z
g(v) := g tr v dx
ΓN

is bounded on H 1 .

Integration by parts
The definition of the trace allows us to perform integration by parts in H 1 :
Z Z Z
∇uϕ dx = − u div ϕ dx + tr u ϕ · n dx ∀ ϕ ∈ [C 1 (Ω)]2
Ω Ω ∂Ω

The definition of the weak derivative (e.g. the weak gradient) looks similar. It allows only
test functions ϕ with compact support in Ω, i.e., having zero boundary values. Only by
choosing a normed space, for which the trace operator is well defined, we can state and
prove integration by parts. Again, the short proof is based on the density of C 1 (Ω) in H 1 .

Sobolev spaces over sub-domains

Let Ω consist of M Lipschitz-continuous sub-domains Ωi such that

• Ω = ∪M
i=1 Ωi

• Ωi ∩ Ωj = ∅ if i 6= j

The interfaces are γij = Ωi ∩ Ωj . The outer normal vector of Ωi is ni .

Theorem 46. Let u ∈ L2 (Ω) such that

• ui := u|Ωi is in H 1 (Ωi ), and gi = ∇ui is its weak gradient

• the traces on common interfaces coincide:

trγij ui = trγij uj

Then u belongs to H 1 (Ω). Its weak gradient g = ∇u fulfills g|Ωi = gi .

Proof: We have to verify that g ∈ L2 (Ω)d , defined by g|Ωi = gi , is the weak gradient of
u, i.e., Z Z
g · ϕ dx = − u div ϕ dx ∀ ϕ ∈ [C0∞ (Ω)]d
Ω Ω
34 CHAPTER 3. SOBOLEV SPACES

We are using Green’s formula on the sub-domains

Z XM Z M Z
X
g · ϕ dx = gi · ϕ dx = ∇ui · ϕ dx
Ω i=1 Ωi i=1 Ωi

XM Z Z
= − ui div ϕ dx + tr ui ϕ · ni ds
i=1 Ωi ∂Ωi
Z XZ 
= − u div ϕ dx + trγij ui ϕ · ni + trγij uj ϕ · nj ds
Ω γij γij
Z
= − u div ϕ dx
Ω
We have used that ϕ = 0 on ∂Ω, and ni = −nj on γij . 2
Applications of this theorem are (conforming nodal) finite element spaces. The parti-
tioning Ωi is the mesh. On each sub-domain, i.e., on each element T , the functions are
polynomials and thus in H 1 (T ). The finite element functions are constructed to be contin-
uous, i.e., the traces match on the interfaces. Thus, the finite element space is a sub-space
of H 1 .

Extension operators
Some estimates are elementary to verify on simple domains such as squares Q. One tech-
nique to transfer these results to general domains is to extend a function u ∈ H 1 (Ω) onto
a larger square Q, apply the result for the square, and restrict the result onto the general
domain Ω. This is now the motivation to study extension operators.
We construct a non-overlapping covering {Si } of a neighbourhood of ∂Ω on both sides.
Let ∂Ω = ∪Γi consist of smooth parts. Let s : (0, 1) × (−1, 1) → Si : (ξ, η) → x be an
invertible function such that
si ((0, 1) × (0, 1)) = Si ∩ Ω
si ((0, 1) × {0}) = Γi
si ((0, 1) × (−1, 0)) = Si \ Ω

−1
Assume that k ds i
k
dx L∞
and k dsdx
i
kL∞ are bounded.
This defines an invertible mapping x → x̂(x) from the inside to the outside by
x̂(x) = si (ξ(x), −η(x)).
The mapping preserve the boundary Γi . The transformations si should be such that x → x̂
is consistent at the interfaces between Si and Sj .
With the flipping operator f : (ξ, η) → (ξ, −η), the mapping is the composite x̂(x) =
si (f (s−1
i )). From that, we obtain the bound
 −1
dx̂ ds ds
≤ .
dx dx dx
3.3. TRACE THEOREMS AND THEIR APPLICATIONS 35

e = Ω ∪ S1 ∪ . . . ∪ SM .
Define the domain Ω
We define the extension operator by
(Eu)(x̂) = u(x) ∀ x ∈ ∪Si
(3.3)
(Eu)(x) = u(x) ∀x ∈ Ω

Theorem 47. The extension operator E : H 1 (Ω) → H 1 (Ω)

e is well defined and bounded
with respect to the norms
kEukL2 (Ω)
e ≤ c kukL2 (Ω)

and
k∇EukL2 (Ω)
e ≤ c k∇ukL2 (Ω)

Proof: Let u ∈ C 1 (Ω). First, we prove the estimates for the individual pieces Si :
Z Z  
2 2 dx̂
Eu(x̂) dx̂ = u(x) det dx ≤ ckuk2L2 (Si ∩Ω)
Si \Ω Si ∩Ω dx
For the derivatives we use
dEu(x̂) du(x(x̂)) du dx
= = .
dx̂ dx̂ dx dx̂
Since dx
dx̂
and ( dx
dx̂
)−1 = dx̂
dx
are bounded, one obtains
|∇x̂ Eu(x̂)| ' |∇x u(x)|,
and Z Z
2
|∇x̂ Eu| dx̂ ≤ c |∇u|2 dx
Si \Ω Si ∩Ω

These estimates prove that E is a bounded operator into H 1 on the sub-domains Si \ Ω.

The construction was such that for u ∈ C 1 (Ω), the extension Eu is continuous across ∂Ω,
and also across the individual Si . By Theorem 46, Eu belongs to H 1 (Ω),
e and
M
X
k∇Euk2L2 (Ω)
e = k∇uk2Ω + k∇uk2Si \Ω ≤ ck∇uk2L2 (Ω) ,
i=1

By density, we get the result for H 1 (Ω). Let uj ∈ C 1 (Ω) → u, than uj is Cauchy, Euj is
Cauchy in H 1 (Ω),
e and thus converges to u ∈ H 1 (Ω).
e

The extension of functions from H01 (Ω) onto larger domains is trivial: Extension by 0
is a bounded operator. One can extend functions from H 1 (Ω) into H01 (Ω),
e and further, to
an arbitrary domain by extension by 0.
For x̂ = si (ξ, −η), ξ, η ∈ (0, 1)2 , define the extension
E0 u(x̂) = (1 − η) u(x)

This extension vanishes at ∂ Ω

e
36 CHAPTER 3. SOBOLEV SPACES

Theorem 48. The extension E0 is an extension from H 1 (Ω) to H01 (Ω).

e It is bounded
w.r.t.
kE0 ukH 1 (Ω)
e ≤ ckukH 1 (Ω)

Proof: Exercises
In this case, it is not possible to bound the gradient term only by gradients. To see this,
take the constant function on Ω. The gradient vanishes, but the extension is not constant.

3.3.1 The trace space H 1/2

The trace operator is continuous from H 1 (Ω) into L2 (∂Ω). But, not every g ∈ L2 (∂Ω) is
a trace of some u ∈ H 1 (Ω). We will motivate why the trace space is the fractional order
Sobolev space H 1/2 (∂Ω).
We introduce a stronger space, such that the trace operator is still continuous, and
onto. Let V = H 1 (Ω), and define the trace space as the range of the trace operator

W = {tr u : u ∈ H 1 (Ω)}

with the norm

k tr ukW = inf kvkV . (3.4)
v∈V
tr u=tr v

This is indeed a norm on W . The trace operator is continuous from V → W with norm 1.

Lemma 49. The space (W, k.kW ) is a Banach space. For all g ∈ W there exists an u ∈ V
such that tr u = g and kukV = kgkW

Proof: The kernel space V0 := {v : tr v = 0} is a closed sub-space of V . If tr u = tr v,

then z := u − v ∈ V0 . We can rewrite

k tr ukW = inf ku − zkV = ku − PV0 ukV ∀u ∈ V

z∈V0

Now, let gn = tr un ∈ W be a Cauchy sequence. This does not imply that un is Cauchy,
but PV0⊥ un is Cauchy in V :

kPV0⊥ (un − um )kV = k tr (un − um )kW .

The PV0⊥ un converge to some u ∈ V0⊥ , and gn converge to g := tr u. 2

The minimizer in (3.4) fulfills

tr u = g and (u, v)V = 0 ∀ v ∈ V0 .

This means that u is the solution of the weak form of the Dirichlet problem

−∆u + u = 0 in Ω
u = g on ∂Ω.
3.3. TRACE THEOREMS AND THEIR APPLICATIONS 37

To give an explicit characterization of the norm k.kW , we introduce Hilbert space

interpolation:
Let V1 ⊂ V0 be two Hilbert spaces, such that V1 is dense in V0 , and the embedding
operator id : V1 → V0 is compact. We can pose the eigen-value problem: Find z ∈ V1 ,
λ ∈ R such that
(z, v)V1 = λ(z, v)V0 ∀ v ∈ V1 .
There exists a sequence of eigen-pairs (zk , λk ) such that λk → ∞. The zk form an or-
thonormal basis in V0 , and an orthogonal basis in V1 .

The converse is also true. If zk is a basis for V0 , and the eigenvalues λk → ∞, then the
embedding V1 ⊂ V0 is compact.

Given u ∈ V0 , it can be expanded in the orthonormal eigen-vector basis:

∞
X
u= uk zk with uk = (u, zk )V0
k=0

The k.kV0 - norm of u is

X X X X
kuk2V0 = ( uk zk , ul zl )V0 = uk ul (zk , zl )V0 = u2k .
k l k,l k

If u ∈ V1 , then
X X X X X
kuk2V1 = ( uk zk , ul zl )V1 = uk ul (zk , zl )V1 = uk ul λk (zk , zl )V0 = u2k λk
k l k,l k,l k

λk u2k is finite. This suggests

P P
The sub-space space V1 consists of all u = uk zk such that k
the definition of the interpolation norm
X
kuk2Vs = (u, zk )2V0 λsk ,
k

and the interpolation space Vs = [V0 , V1 ]s as

Vs = {u ∈ V0 : kukVs < ∞}.

We have been fast with using infinite sums. To make everything

Pn precise, one first works
with finite dimensional sub-spaces {u : ∃n ∈ N and u = k=1 uk zk }, and takes the closure.

In our case, we apply Hilbert space interpolation to H 1 (0, 1) ⊂ L2 (0, 1). The eigen-value
problem is to find zk ∈ H 1 and λk ∈ R such that

(zk , v)L2 + (zk0 , v 0 )L2 = λk (zk , v)L2 ∀ v ∈ H1

38 CHAPTER 3. SOBOLEV SPACES

By definition of the weak derivative, there holds (zk0 )0 = (1 − λk )zk , i.e., z k ∈ H 2 . Since
H 2 ⊂ C 0 , there holds also z ∈ C 2 , and a weak solution is also a solution of the strong form

zk − zk00 = λk zk on (0, 1)
(3.5)
zk0 (0) = zk0 (1) = 0

All solutions, normalized to kzk kL2 = 1, are

z0 = 1 λ0 = 1

and, for k ∈ N, √
zk (x) = 2 cos(kπx)λk = 1 + k 2 π 2 .
√
Indeed, expanding u ∈ L2 in the cos-basis u = u0 + ∞
P
k=1 uk 2 cos(kπx), one has

∞
X
kuk2L2 = (u, zk )2L2
k=0

and
∞
X
kuk2H 1 = (1 + k 2 π 2 )(u, zk )2L2
k=0

Differentiation adds a factor kπ. Hilbert space interpolation allows to define the fractional
order Sobolev norm (s ∈ (0, 1))
∞
X
kuk2H s (0,1) = (1 + k 2 π 2 )s (u, zk )2L2
k=0

We consider the trace tr |E of H 1 ((0, 1)2 ) onto one edge E = (0, 1) × {0}. For g ∈
WE := tr H 1 ((0, 1)2 ), the norm kgkW is defined by

kgkW = kug kH 1 .

Here, ug solves the Dirichlet problem ug |E = g, and (ug , v)H 1 = 0 ∀ v ∈ H 1 such that
trE v = 0.
Since W ⊂ L2 (E), we can expand g in the L2 -orthonormal cosine basis zk
X
g(x) = gn zk (x)

The Dirichlet problems for the zk ,

−∆uk + uk = 0 in Ω
uk = zk on E
∂uk
∂n
= 0 on ∂Ω \ E,
3.4. EQUIVALENT NORMS ON H 1 AND ON SUB-SPACES 39

have the explicit solution

u0 (x, y) = 1
and
√ ekπ(1−y) + e−kπ(1−y)
uk (x, y) = 2 cos(kπx) .
ekπ + e−kπ
The asymptotic is
kuk k2L2 ' (k + 1)−1
and
k∇uk k2L2 ' k
P
Furthermore, the uk are orthogonal in (., .)H 1 . Thus ug = n gn uk has the norm
X X
kug k2H 1 = gn2 kuk k2H 1 ' gn2 (1 + k).

This norm is equivalent to H 1/2 (E).

We have proven that the trace space onto one edge is the interpolation space H 1/2 (E).
This is also true for general domains (Lipschitz, with piecewise smooth boundary).

3.4 Equivalent norms on H 1 and on sub-spaces

The intention is to formulate 2nd order variational problems in the Hilbert space H 1 . We
want to apply the Lax-Milgram theory for continuous and coercive bilinear forms A(., .).
We present techniques to prove coercivity.
The idea is the following. In the norm

kvk2H 1 = kvk2L2 + k∇vk2L2 ,

the k∇ · kL2 -semi-norm is the dominating part up to the constant functions. The L2 norm
is necessary to obtain a norm. We want to replace the L2 norm by some different term
(e.g., the L2 -norm on a part of Ω, or the L2 -norm on ∂Ω), and want to obtain an equivalent
norm.
We formulate an abstract theorem relating a norm k.kV to a semi-norm k.kA . An
equivalent theorem was proven by Tartar.

Theorem 50 (Tartar). Let (V, (., .)V ) and (W, (., .)W ) be Hilbert spaces, such that the
embedding id : V → W is compact. Let A(., .) be a non-negative, symmetric and V -
continuous bilinear form with kernel V0 = {v : A(v, v) = 0}. Assume that

kvk2V ' kvk2W + kvk2A ∀v ∈ V (3.6)

Then there holds

40 CHAPTER 3. SOBOLEV SPACES

1. The kernel V0 is finite dimensional. On the factor space V /V0 , A(., .) is an equivalent
norm to the quotient norm

kukA ' inf ku − vkV ∀u ∈ V (3.7)

v∈V0

2. Let B(., .) be a continuous, non-negative, symmetric bilinear form on V such that

A(., .) + B(., .) is an inner product. Then there holds

kvk2V ' kvk2A + kvk2B ∀v ∈ V

3. Let V1 ⊂ V be a closed sub-space such that V0 ∩ V1 = {0}. Then there holds

kvkV ' kvkA ∀ v ∈ V1

Proof: 1. Assume that V0 is not finite dimensional. Then there exists an (., .)V -orthonormal
sequence uk ∈ V0 . Since the embedding id : V → W is compact, it has a sub-sequence
converging in k.kW . But, since

2 = kuk − ul k2V ' kuk − ul k2W + kuk − ul kA = kuk − ul k2W

for k 6= l, uk is not Cauchy in W . This is a contradiction to an infinite dimensional kernel

space V0 . We prove the equivalence (3.7). To bound the left hand side by the right hand
side, we use that V0 = ker A, and norm equivalence (3.6):

kukA = inf ku − vkA ≤ inf ku − vkV

v∈V0 v∈V0

The quotient norm is equal to kPV0⊥ uk. We have to prove that kPV0⊥ ukV ≤ kPV0⊥ ukA for
all u ∈ V . This follows after proving kukV ≤ kukA for all u ∈ V0⊥ . Assume that this is not
true. I.e., there exists a V -orthogonal sequence (uk ) such that kuk kA ≤ k −1 kuk kV . Extract
a sub-sequence converging in k.kW , and call it uk again. From the norm equivalence (3.6)
there follows
2 = kuk − ul k2V  kuk − ul kW + kuk − ul kA → 0

2. On V0 , k.kB is a norm. Since V0 is finite dimensional, it is equivalent to k.kV , say with

bounds
c1 kvk2V ≤ kvk2B ≤ c2 kvk2V ∀ v ∈ V0

From 1. we know that

c3 kvk2V ≤ kvk2A ≤ c4 kvk2V ∀ v ∈ V0⊥ .

3.4. EQUIVALENT NORMS ON H 1 AND ON SUB-SPACES 41

Now, we bound
kuk2V = kPV0 uk2V + kPV0⊥ uk2V
1
≤ k PV0 u k2B + kPV0⊥ k2V
c1 |{z}
u−PV ⊥ u
0

2  
≤ kukB + c2 kPV0⊥ ukV + kPV0⊥ uk2V
2 2
c1
 
2 2 1 2c2
= kukB + 1+ kPV0⊥ uk2A
c1 c2 c1
2 2
 kukB + kukA
3. Define B(u, v) = (PV⊥1 u, PV⊥1 u)V . Then A(., .) + B(., .) is an inner product: A(u, u) +
B(u, u) = 0 implies that u ∈ V0 and u ∈ V1 , thus u = {0}. From 2. there follows that
A(., .) + B(., .) is equivalent to (., .)V . The result follows from reducing the equivalence to
V1 .
2
1
We want to apply Tartar’s theorem to the case V = H , W = L2 , and kvkA = k∇vkL2 .
The theorem requires that the embedding id : H 1 → L2 is compact. This is indeed true
for bounded domains Ω:
Theorem 51. The embedding of H k → H l for k > l is compact.
We sketch a proof for the embedding H 1 ⊂ L2 . First, prove the compact embedding
H01 (Q) → L2 (Q) for a
square Q, w.l.o.g. set Q = (0, 1)2 . The eigen-value problem: Find
z ∈ H01 (Q) and λ suchthat
(z, v)L2 + (∇z, ∇v)L2 = λ(u, v)L2 ∀ v ∈ H01 (Q)
has eigen-vectors zk,l = sin(kπx)sin(lπy), and eigen-values 1 + k 2 π 2 + l2 π 2 → ∞. The
eigen-vectors are dense in L2 . Thus, the embedding is compact.
On a general domain Ω ⊂ Q, we can extend H 1 (Ω) into H01 (Q), embed H01 (Q) into
L2 (Q), and restrict L2 (Q) onto L2 (Ω). This is the composite of two continuous and a
compact mapping, and thus is compact. 2

The kernel V0 of the semi-norm k∇vk is the constant function.

Theorem 52 (Friedrichs inequality). Let ΓD ⊂ ∂Ω be of positive measure |ΓD |. Let
VD = {v ∈ H 1 (Ω) : trΓD v = 0}. Then
kvkL2  k∇vkL2 ∀ v ∈ VD
Proof: The intersection V0 ∩ VD is trivial {0}. Thus, Theorem 50, 3. implies the
equivalence
kvk2V = kvk2L2 + k∇vk2L2 ' k∇vkL2 .
2
42 CHAPTER 3. SOBOLEV SPACES

Theorem 53 (Poincaré inequality). There holds

Z
kvk2H 1 (Ω)  k∇vk2L2 +( v dx)2
Ω
R R
Proof: B(u, v) := ( Ω u dx)( Ω v dx) is a continuous bilinear form on H 1 , and (∇u, ∇v)+
B(u, v) is an inner product. Thus, Theorem 50, 2. implies the stated equivalence. 2

• Let ω ⊂ Ω have positive measure |ω| in Rd . Then

kuk2H 1 (Ω) ' k∇vk2L2 (Ω) + kvkL2 (ω) ,

• Let γ ⊂ ∂Ω have positive measure |γ| in Rd−1 . Then

kuk2H 1 (Ω) ' k∇vk2L2 (Ω) + kvkL2 (γ) ,

Theorem 54 (Bramble Hilbert lemma). Let U be some Hilbert space, and L : H k → U be

a continuous linear operator such that Lq = 0 for polynomials q ∈ P k−1 . Then there holds
kLvkU ≤ |v|H k .
Proof: The embedding H k → H k−1 P is compact. The V -continuous, symmetric and
non-negative bilinear form A(u, v) = α:|α|=k (∂ u, ∂ v) has the kernel P k−1 . Decompose
α α

kuk2H k = kuk2H k−1 + A(u, u). By Theorem 50, 1, there holds

kukA ' inf ku − vkH k
v∈V0

The same holds for the bilinear-form

A2 (u, v) := (Lu, Lv)U + A(u, v)
Thus
kukA2 ' inf ku − vkH k ∀u ∈ V
v∈V0

Equalizing both implies that

(Lu, Lu)U ≤ kuk2A2 ' kuk2A ∀ u ∈ V,
i.e., the claim.

We will need point evaluation of functions in Sobolev spaces H s . This is possible, we

u ∈ H s implies that u is continuous.
Theorem 55 (Sobolev’s embedding theorem). Let Ω ⊂ Rd with Lipschitz boundary. If
u ∈ H s with s > d/2, then u ∈ L∞ with
kukL∞  kukH s
There is a function in C 0 within the L∞ equivalence class.
3.5. INTERPOLATION SPACES 43

3.5 Interpolation Spaces

3.5.1 Hilbert space interpolation
Let V1 ⊂ V0 be two Hilbert spaces with dense embedding. For simplicity we assume that
the embedding is compact. Then there exists a system of eigenvalues λk and eigenvectors zk
such that
(zk , v)1 = λ2k (zk , v)0 ∀ v ∈ V1 .
The eigenvectors are orthogonal and are normalized such that

(zk , zl )0 = δk,l and (zk , zl )1 = λ2k δk,l .

Eigenvalues are ascening, by compactness there holds λk → ∞.

The set of eigenvectors is a complete system. Thus u ∈ V0 can be expanded as
∞
X
u= uk zk with uk = (u, zk )0 .
k=1

There holds
X
kuk20 = u2k
X
kuk21 = λ2k u2k < ∞ for u ∈ V1 .

For s ∈ (0, 1) we define the interpolation norm

∞
X 1/2
kuks̃ := λ2s 2
k uk (3.8)
k=1

and the interpolation space

Vs := [V0 , V1 ]s := {u ∈ V0 : kuks̃ < ∞}.

There holds
V1 ⊂ Vs ⊂ V0 .
Example: Let V0 = L2 (0, 1) and V1 = H01 (0, 1). Then
√
zk = 2 sin(kπx) and λk = k

3.5.2 Banach space interpolation

We give an alternative definition of interpolation spaces, which is also applicable for Banach
spaces. It is known as Banach space interpolation, K-functional method, real method of
interpolation, or Peetre’s method.
44 CHAPTER 3. SOBOLEV SPACES

Let V1 ⊂ V0 be Banach spaces with dense and continuous embedding. We define the
K-functional K : R+ × V0 → R as
q
K(t, u) := inf ku − v1 k20 + t2 kv1 k21 .
v1 ∈V1

Note that

K(t, u) ≤ kuk0 ,
K(t, u) ≤ t kuk1 for u ∈ V1 .

The decay in t measures the smoothness of u. For s ∈ (0, 1) we define the interpolation
norm as Z ∞ 1/2
−2s 2
kuks := t K(t, u) dt/t (3.9)
0

and the interpolation spaces Vs := {u ∈ V0 : kuks < ∞}.

The K-functional method is more general. If the spaces are Hilbert, then both inter-
polation methods coincide:

Theorem 56. Let V1 ⊂ V0 be Hilbert spaces with compact embedding. Then

kuks = Cs kuks̃ ,
R∞ τ 1−2s
where Cs2 = 0 1+τ 2
dτ .
P
Proof. For u = uk zk we calculate the K-functional as

K(t, u)2 = inf ku − vk20 + t2 kvk21

v∈V1
X
= inf (uk − vk )2 + t2 λ2k vk2
(vk )∈`2
(λk vk )∈`2 k
X
= inf (uk − vk )2 + t2 λ2k vk2 .
vk ∈R
k

The minimum of each summand is taken for

1
vk = uk
1 + t2 λ2k

and its value is

t2 λ2k
2
u2 .
2 k
1 + t λk
Thus
∞
2
X t2 λ2k
K(t, u) = u2
k=1
1 + t2 λ2k k
3.5. INTERPOLATION SPACES 45

and
∞ ∞
t2 λ2k
Z Z X
−2s
kuk2s = t 2
K(t, u) dt/t = u2 dt/t
0 0 k
1 + t2 λ2k k
∞
XZ t2 λ2k
= t−2s 2
u2
2 k
dt/t
k 0 1 + t λk

Substitution τ = λk t gives

XZ ∞
 τ −2s τ 2
kuk2s = u2 dτ /τ
2 k
k 0 λ k 1 + τ
∞
τ 1−2s
X Z
= λ2s
k ku 2
2
dτ
k 0 1 + τ
= Cs2 kuk2s̃

Theorem 57. For u ∈ V1 there holds

kuks  kuk01−s kuks1

Proof: Excercise

3.5.3 Operator interpolation

Let V1 ⊂ V0 and W1 ⊂ W0 with dense embedding.

Theorem 58. Let T : V0 → W0 be a linear operator such that T V1 ⊂ W1 with norms

kT kV0 →W0 ≤ c0 and kT kV1 →W1 ≤ c1 .

Then
T : [V0 , V1 ]s → [W0 , W1 ]s

with norm
kT k[V0 ,V1 ]s →[W0 ,W1 ]s ≤ c01−s cs1

Proof. We use the definition of the interpolation norm, T V1 ⊂ W1 , operator norms and
46 CHAPTER 3. SOBOLEV SPACES

substitution τ = c1 t/c0
Z ∞
kT uk[W0 ,W1 ]s = t−2s KW (t, T u)2 dt/t
Z0 ∞
= t−2s inf {kT u − w1 kW0 + t2 kw1 k2W1 } dt/t
w1 ∈W1
Z0 ∞
≤ t−2s inf {kT u − T v1 kW0 + t2 kT v1 k2W1 } dt/t
v1 ∈V1
Z0 ∞
≤ t−2s inf {c20 ku − v1 kV0 + t2 c21 kv1 k2V1 } dt/t
v1 ∈V1
Z0 ∞ 
c0 τ −2s
≤ inf {c20 ku − v1 k2V0 + c20 τ 2 kv1 k2V1 } dτ /τ
0 c 1 v 1 ∈V1
Z ∞
= c2−2s
0 c2s
1 τ −2s KV (t, u)2 dτ /τ
0
= c2−2s
0 c2s
1 kuk2[V0 ,V1 ]s

3.5.4 Interpolation of Sobolev Spaces

As an example of interpolation spaces we show the following:

Theorem 59. Let Ω be a Lipschitz domain. Then

[L2 (Ω), H 2 (Ω)]1/2 = H 1 (Ω).

Proof. Let Q be a square containing Ω, w.l.o.g. Q = (0, 2π)2 , and zk,l = eikx eily be the
m
trigonometric basis for (complex-valued) periodic Sobolev Spaces Hper (Q). Then
X
kuk2H m ' (k 2 + l2 )m |uk,l |2 ,
k,l

1 0 2
and thus Hper (Q) = [Hper (Q), Hper (Q)]1/2 by Hilbert space interpolation.
Now let E : L2 (Ω) → L2 (Q) be an extension operator such that

E : H m (Ω) → Hper
m
(Q)

is continuous for all m ∈ {0, 1, 2}. Furthermore, let

R : L2 (Q) → L2 (Ω) : u 7→ u|Ω

m
be the restriction operator. Trivially, R : Hper (Q) → H m (Ω) is continuous for m ∈ N0 .
We show that
kukH 1 (Ω) ' kuk[L2 (Ω),H 2 (Ω)]1/2 .
3.6. THE WEAK FORMULATION OF THE POISSON EQUATION 47

Using operator interpolation we get

kukH 1 (Ω) = kREukH 1 (Ω) ≤ kRkkEukH 1 (Q)

' kEuk[L2 (Q),Hper
2 (Q)]
1/2
1/2 1/2
≤ kEkL2 (Ω)→L2 (Q) kEkH 2 (Ω)→Hper
2 (Q) kuk[L2 (Ω),H 2 (Ω)]1/2

' kuk[L2 (Ω),H 2 (Ω)]1/2 ,

and similarly the other way around.

Theorem 60. Let Ω be a Lipschitz domain. Then

[L2 (Ω), H02 (Ω)]1/2 = H01 (Ω).

Proof. Exercise

Literature:

1. J. Bergh and J. Lofstrom. Interpolation spaces. Springer, 1976

2. J. H. Bramble. Multigrid Methods. Chapman and Hall, 1993

3.6 The weak formulation of the Poisson equation

We are now able to give a precise definition of the weak formulation of the Poisson problem
as introduced in Section 1.2, and analyze the existence and uniqueness of a weak solution.
Let Ω be a bounded domain. Its boundary ∂Ω is decomposed as ∂Ω = ΓD ∪ ΓN ∪ ΓR
according to Dirichlet, Neumann and Robin boundary conditions.
Let

• uD ∈ H 1/2 (ΓD ),

• f ∈ L2 (Ω),

• g ∈ L2 (ΓN ∪ ΓR ),

• α ∈ L∞ (ΓD ), α ≥ 0.

Assume that there holds

(a) The Dirichlet part has positive measure |ΓD | > 0,

R
(b) or the Robin term has positive contribution ΓR α dx > 0.
48 CHAPTER 3. SOBOLEV SPACES

Define the Hilbert space

V := H 1 (Ω),
the closed sub-space
V0 = {v : trΓD v = 0},
and the linear manifold
VD = {u ∈ V : trΓD u = uD }.
Define the bilinear form A(., .) : V × V → R
Z Z
A(u, v) = ∇u ∇v dx + αuv ds
Ω ΓR

and the linear form Z Z

f (v) = f v dx + gv dx.
Ω ΓN ∪ΓR

Theorem 61. The weak formulation of the Poisson problem

Find u ∈ VD such that

A(u, v) = f (v) ∀ v ∈ V0 (3.10)

has a unique solution u.

Proof: The bilinear-form A(., .) and the linear-form f (.) are continuous on V . Tartar’s
theorem of equivalent norms proves that A(., .) is coercive on V0 .
Since uD is in the closed range of trΓD , there exists an ũD ∈ VD such that

tr ũD = uD and kũD kV  kuD kH 1/2 (ΓD )

Now, pose the problem: Find z ∈ V0 such that

A(z, v) = f (v) − A(e

uD , v) ∀ v ∈ V0 .

The right hand side is the evaluation of the continuous linear form f (.) − A(e
uD , .) on
V0 . Due to Lax-Milgram, there exists a unique solution z. Then, u := ueD + z solves (3.10).
The choice of ueD is not unique, but, the constructed u is unique. 2

3.6.1 Shift theorems

Let us restrict to Dirichlet boundary conditions uD = 0 on the whole boundary. The
variational problem: Find u ∈ V0 such that

A(u, v) = f (v) ∀ v ∈ V0
3.6. THE WEAK FORMULATION OF THE POISSON EQUATION 49

is well defined for all f ∈ V0∗ , and, due to Lax-Milgram there holds

kukV0 ≤ ckf kV0∗ .

Vice versa, the bilinear-form defines the linear functional A(u, .) with norm

kA(u, .)kV0∗ ≤ ckukV0

This dual space is called H −1 :

H −1 := [H01 (Ω)]∗

Since H01 ⊂ L2 , there is L2 ⊂ H −1 (Ω). All negative spaces are defined as H −s (Ω) :=
[H0s ]∗ (Ω), for s ∈ R+ . There holds

. . . H02 ⊂ H01 ⊂ L2 ⊂ H −1 ⊂ H −2 . . .

The solution operator of the weak formulation is smoothing twice. The statements of
shift theorem are that for s > 0, the solution operator maps also

f ∈ H −1+s → u ∈ H 1+s

with norm bounds

kukH 1+s  kf kH −1+s .
In this case, we call the problem H 1+s - regular.

Theorem 62 (Shift theorem).

(a) Assume that Ω is convex. Then, the Dirichlet problem is H 2 regular.

(b) Let s ≥ 2. Assume that ∂Ω ∈ C s . Then, the Dirichlet problem is H s -regular.

We give a proof of (a) for the square (0, π)2 by Fourier series. Let

VN = span{sin(kx) sin(ly) : 1 ≤ k, l ≤ N }
PN
For an u = k,l=1 ukl sin(kx) sin(ly) ∈ VN , there holds

kuk2H 2 = kuk2L2 + k∂x uk2L2 + k∂y uk2L2 + k∂x2 uk2L2 + k∂x ∂y uk2L2 + k∂y2 uk2
N
X
' (1 + k 2 + l2 + k 4 + k 2 l2 + l4 )u2kl
k,l=1
N
X
' (k 4 + l4 )u2kl ,
k,l=1
50 CHAPTER 3. SOBOLEV SPACES

and, for f = −∆u,

N
X N
X
k− ∆uk2L2 = (k +2
l2 )2 u2kl ' (k 4 + l4 )u2kl .
k,l=1 k,l=1

Thus we have kukH 2 ' k∆ukL2 = kf kL2 for u ∈ VN . The rest requires a closure argument:
There is {−∆v : v ∈ VN } = VN , and VN is dense in L2 . 2
2
Indeed, on non-smooth non-convex domains, the H -regularity is not true. Take the
sector of the unit-disc

Ω = {(r cos φ, r sin φ) : 0 < r < 1, 0 < φ < ω}

with ω ∈ (π, 2π). Set β = π/ω < 1. The function

u = (1 − r2 )rβ sin(φβ)

is in H01 , and fulfills ∆u = −(4β + 4)rβ sin(φβ) ∈ L2 . Thus u is the solution of a Dirichlet
problem. But u 6∈ H 2 .
On non-convex domains one can specify the regularity in terms of weighted Sobolev
spaces. Let Ω be a polygonal domain containing M vertices Vi . Let ωi be the interior angle
at Vi . If the vertex belongs to a non-convex corner (ωi > π), then choose some
π
βi ∈ (1 − , 1)
ω
Define Y
w(x) = |x − Vi |βi
non-convex
Vertices Vi

Theorem 63. If f is such that wf ∈ L2 . Then f ∈ H −1 , and the solution u of the

Dirichlet problem fulfills
kwD2 ukL2  kwf kL2 .
Chapter 4

Finite Element Method

Ciarlet’s definition of a finite element is:

Definition 64 (Finite element). A finite element is a triple (T, VT , ΨT ), where
1. T is a bounded set

2. VT is function space on T of finite dimension NT

3. ΨT = {ψT1 , . . . , ψTNT } is a set of linearly independent functionals on VT .

The nodal basis {ϕ1T . . . ϕN
T } for VT is the basis dual to ΨT , i.e.,
T

ψTi (ϕjT ) = δij

Barycentric coordinates are useful to express the nodal basis functions.

Finite elements with point evaluation functionals are called Lagrange finite elements,
elements using also derivatives are called Hermite finite elements.
Usual function spaces on T ⊂ R2 are

P p := span{xi y j : 0 ≤ i, 0 ≤ j, i + j ≤ p}
Qp := span{xi y j : 0 ≤ i ≤ p, 0 ≤ j ≤ p}

Examples for finite elements are

• A linear line segment

• A quadratic line segment

• A Hermite line segment

• A constant triangle

• A linear triangle

• A non-conforming triangle

51
52 CHAPTER 4. FINITE ELEMENT METHOD

• A Morley triangle

• A Raviart-Thomas triangle

The local nodal interpolation operator defined for functions v ∈ C m (T ) is

NT
X
IT v := ψTα (v)ϕαT
α=1

It is a projection.
Two finite elements (T, VT , ΨT ) and (Tb, VTb , ΨTb ) are called equivalent if there exists an
invertible function F such that

• T = F (Tb)

• VT = {v̂ ◦ F −1 : v̂ ∈ VTb }

• ΨT = {ψiT : VT → R : v → ψiT̂ (v ◦ F )}

Two elements are called affine equivalent, if F is an affine-linear function.

Lagrangian finite elements defined above are equivalent. The Hermite elements are not
equivalent.
Two finite elements are called interpolation equivalent if there holds

IT (v) ◦ F = ITb (v ◦ F )

Lemma 65. Equivalent elements are interpolation equivalent

The Hermite elements define above are also interpolation equivalent.

A regular triangulation T = {T1 , . . . , TM } of a domain Ω is the subdivision of a domain
Ω into closed triangles Ti such that Ω = ∪Ti and Ti ∩ Tj is

• either empty

• or an common edge of Ti and Tj

• or Ti = Tj in the case i = j.

In a wider sense, a triangulation may consist of different element shapes such as segments,
triangles, quadrilaterals, tetrahedra, hexhedra, prisms, pyramids.
A finite element complex {(T, VT , ΨT )} is a set of finite elements defined on the geo-
metric elements of the triangulation T .
It is convenient to construct finite element complexes such that all its finite elements
are affine equivalent to one reference finite element (Tb, V̂T , Ψ̂T ). The transformation FT is
such that T = FT (Tb).
Examples: linear reference line segment on (0, 1).
4.1. FINITE ELEMENT SYSTEM ASSEMBLING 53

The finite element complex allows the definition of the global interpolation operator for
m
C -smooth functions by
IT v|T = IT vT ∀T ∈ T
The finite element space is

VT := {v = IT w : w ∈ C m (Ω)}

We say that VT has regularity r if VT ⊂ C r . If VT 6= C 0 , the regularity is defined as −1.

Examples:

• The P 1 - triangle with vertex nodes leads to regularity 0.

• The P 1 - triangle with edge midpoint nodes leads to regularity −1.

• The P 0 - triangle leads to regularity −1.

For smooth functions, functionals ψT,α and ψTe,α̃ sitting in the same location are equiv-
alent. The set of global functionals Ψ = {ψ1 , . . . , ψN } is the linearly independent set of
functionals containing all (equivalence classes of) local functionals.
The connectivity matrix CT ∈ RN ×NT is defined such that the local functionals are
derived from the global ones by
ΨT (u) = CTt Ψ(u)
Examples in 1D and 2D
The nodal basis for the global finite element space is the basis in VT dual to the global
functionals ψj , i.e.,
ψj (ϕi ) = δij
There holds
NT
X
ϕi |T = IT ϕi = ψTα (ϕi )ϕαT
α=1
NT
X
= (CTt ψ(ϕi ))α ϕαT
α=1
NT
X NT
X
= (CTt ei )α ϕαT = CT,iα ϕαT
α=1 α=1

4.1 Finite element system assembling

As a first step, we assume there are no Dirichlet boundary conditions. The finite element
problem is
Find uh ∈ VT such that : A(uh , vh ) = f (vh ) ∀ vh ∈ VT (4.1)
54 CHAPTER 4. FINITE ELEMENT METHOD

The nodal basis and the dual functionals provides the one to one relation between RN and
VT :
XN
N
R 3 u ↔ uh ∈ VT with uh = ϕi ui and ui = ψi (uh ).
i=1
Using the nodal basis expansion of uh in (4.1), and testing only with the set of basis
functions, one has

XN
A( ui ϕi , ϕj ) = f (ϕj ) ∀j = 1...N
i=1
With
Aji = A(ϕi , ϕj ) and f j = f (ϕj ),
one obtains the linear system of equations
Au = f
The preferred way to compute the matrix A and vector f is a sum over element contribu-
tions. The restrictions of the bilinear and linear form to the elements are
Z Z
AT (u, v) = ∇u · ∇v dx + αuv ds
T ∂Ω∩T

and Z Z
fT (v) = f v dx + gv ds
T ∂Ω∩T
Then X X
A(u, v) = AT (u, v) f (v) = fT (v)
T ∈T T ∈T
On each element, one defines the NT ×NT element matrix and element vector in terms
of the local basis on T :
AT,αβ = AT (ϕTβ , ϕTα ) f T,α = f T (ϕTα )
Then, the global matrix and the global vector are
X
A= CT AT CTt
T ∈T

and X
f= CT f T
T ∈T
Namely,
X X X
f i = f (ϕi ) = fT (ϕi |T ) = fT ( CT,iα ϕαT )
T ∈T T ∈T α
XX XX
= CT,iα fT (ϕαT ) = CT,iα f α
T ∈T α T ∈T α
4.2. FINITE ELEMENT ERROR ANALYSIS 55

and
X X X X
Aji = A(ϕi |T , ϕj |T ) = A( CT,iα ϕαT , CT,jβ ϕβT )
T ∈T T ∈T α β
X XX
= CT,iα AT,αβ CT,jβ
T ∈T α β

On the elements T , the integrands are smooth functions. Thus, numerical integration
rules can be applied.

In the case of Dirichlet boundary conditions, let γD ⊂ {1, . . . , N } correspond to the

vertices xi at the Dirichlet boundary, and γf = {1, . . . N } \ γD .
We have the equations
X X
Aji ui + Aji ui = fj ∀ j ∈ γf
i∈γD i∈γf

Inserting ui = uD (xD ) for i ∈ γi results in the reduced system

X X
Aji ui = fj − Aji uD (xi )
i∈γf i∈γD

An alternative approach is to approximate Dirichlet boundary conditions by Robin b.c.,

∂u
∂n
+ αu = αuD , with large parameter α.

4.2 Finite element error analysis

Let u be the solution of the variational problem, and uh its Galerkin approximation in the
finite element sub-space Vh . Cea’s Lemma bounds the finite element error u − uh by the
best approximation error

ku − uh kV ≤ C inf ku − vh kV .
vh ∈Vh

The constant factor C is the ratio of the continuity bound and the coercivity bound of the
bilinear form A(., .).
Provided that the solution u is sufficiently smooth, we can take the finite element
interpolant to bound the best approximation error:

inf ku − vh kV ≤ ku − IT ukV
v∈Vh

In the following, we will bound the interpolation error.

56 CHAPTER 4. FINITE ELEMENT METHOD

Lemma 66. Let Tb and T be d-dimensional domains related by the invertible affine linear
transformation FT : Tb → T
FT (x) = a + Bx,
where a ∈ Rd and B is a regular matrix in Rd×d . Then there holds:

ku ◦ FT kL2 (Tb) = (det B)−1/2 kukL2 (T ) (4.2)

d d  
∂ ∂ X X ∂ ∂
... (u ◦ FT ) = ... ... u ◦ FT Bjm ,im . . . Bj1 ,i1 (4.3)
∂xim ∂xi1 j =1 j =1
∂x j m ∂x j 1
m 1

|u ◦ F |H m (Tb)  (det B)−1/2 kBkm |u|H m (T ) (4.4)

Proof: Transformation of integrals, chain rule. 2

We define the diameter of the element T

hT = diam T

A triangulation is called shape regular, if all its elements fulfill

|T |  h2T

with a “good” constant ∼ 1. If one studies convergence, one considers families of triangu-
lations with decreasing element sizes hT . In that case, the family of triangulations is called
shape regular, if there is a common constant C such that all elements of all triangulations
fulfill |T | ≥ Ch2T .

Lemma 67. Let FT = a + Bx be the mapping from the reference triangle to the triangle
T . Let |T |  h2T . Then there holds

kBT k ' hT
kBT−1 k ' h−1
T

The following lemma is the basis for the error estimate. This lemma is the main
application for the Bramble Hilbert lemma. Sometimes, it is called the Bramble Hilbert
lemma itself:

Lemma 68. Let (T, VT , ΨT ) be a finite element such that the element space VT contains
polynomials up to order P k . Then there holds

kv − IT vkH 1 ≤ C|v|H m ∀ v ∈ H m (T )

for all m > d/2, m ≥ 1, and m ≤ k + 1.

4.2. FINITE ELEMENT ERROR ANALYSIS 57

Proof: First, we prove that id − IT is a bounded operator from H m to H 1 :

X
kv − IT vkH 1 ≤ kvkH 1 + kIT vkH 1 = kvkH 1 + k ψα (v)ϕα kH 1
α
X
≤ kvkH 1 + kϕα kH 1 |ψα (v)|
α
 kvkH m

The last step used that for H m , with m > d/2, point evaluation is continuous. Now, let
v ∈ P k (T ). Since P k ⊂ VT , and IT is a projection on VT , there holds v − IT v = 0. The
Bramble Hilbert Lemma applied for U = H 1 and L = id − IT proves the result. 2
To bound the finite element interpolation error, we will transform functions from the
elements T to the reference element Tb.

Theorem 69. Let T be a shape regular triangulation of Ω. Let VT be a C 0 -regular finite

element space such that all local spaces contain P 1 . Then there holds
( )1/2
X
kv − IT vkL2 (Ω)  h4T |v|2H 2 (T ) ∀ v ∈ H 2 (Ω)
T ∈T

( )1/2
X
|v − IT v|H 1 (Ω) ≤ h2T |v|2H 2 (T ) ∀ v ∈ H 2 (Ω)
T ∈T

Proof: We prove the H 1 estimate, the L2 one follows the same lines. The interpolation
error on each element is transformed to the interpolation error on one reference element:
X
|v − IT v|2H 1 (Ω) = |(id − IT )vT |2H 1 (T )
T ∈T
X
 (det BT ) kBT−1 k2 |(id − IT )vT ◦ FT |2H 1 (Tb)
T ∈T
X
= (det BT )kBT−1 k2 k(id − ITb )(vT ◦ FT )k2H 1 (Tb)
T ∈T

On the reference element Tb we apply the Bramble-Hilbert lemma. Then, we transform

back to the individual elements:
X
|v − IT v|2H 1 (Ω)  (det BT )kBT−1 k2 |vT ◦ FT |2H 2 (Tb)
T ∈T
X
 (det BT ) kBT−1 k2 (det BT−1 ) kBT k4 |vT |2H 2 (T )
T ∈T
X
' h2T kvk2H 2 (T ) .
T ∈T
58 CHAPTER 4. FINITE ELEMENT METHOD

A triangulation is called quasi − uniform, if all elements are essentially of the same
size, i.e., there exists one global h such that

h ' hT ∀T ∈ T .

On a quasi-uniform mesh, there hold the interpolation error estimates

ku − IT ukL2 (Ω)  h2 |u|H 2

|u − IT u|H 1 (Ω)  h |u|H 2

We are interested in the rate of the error in terms of the mesh-size h.

Theorem 70 (Finite element error estimate). Assume that
• the solution u of the weak bvp is in H 2 ,

• the triangulation T is quasi-uniform of mesh-size h,

• the element spaces contain P 1 .

Then, the finite element error is bounded by

ku − uh kH 1  h |u|H 2

Error estimates in L2 -norm

The above theorem bounds the error in the L2 -norm of the function, and the L2 -norm of
the derivatives with the same rate in terms of h. This is obtained by the natural norm of
the variational formulation.
The interpolation error suggests a faster convergence in the weaker norm L2 . Under
certain circumstances, the finite element error measured in L2 also decays faster. The
considered variational problem is

Find u ∈ V : A(u, v) = f (v) ∀ v ∈ V.

We define the dual problem as

Find w ∈ V : A(v, w) = f (v) ∀ v ∈ V.

In the case of a symmetric bilinear form, the primal and the dual problem coincide.
Theorem 71 (Aubin-Nitsche). Assume that
• the dual weak bvp is H 2 regular

• the triangulation T is quasi-uniform of mesh-size h,

4.2. FINITE ELEMENT ERROR ANALYSIS 59

• the element spaces contain P 1 .

Then, there holds the L2 -error estimate
ku − uh kL2  h2 |u|H 2
Proof: Solve the dual problem with the error u − uh as right hand side:
Find w ∈ V : A(v, w) = (u − uh , v)L2 ∀ v ∈ V.
Since the dual problem is H 2 regular, there holds w ∈ H 2 , and kwkH 2  ku − uh kL2 .
Choose the test function v := u − uh to obtain the squared norm
A(u − uh , w) = (u − uh , u − uh )L2 .
Using the Galerkin orthogonality A(u − uh , vh ) = 0 for all vh ∈ Vh , we can insert IT w:
ku − uh k2L2 = A(u − uh , w − IT w).
Next we use continuity of A(., .) and the interpolation error estimates:
ku − uh k2L2  ku − uh kH 1 kw − IT wkH 1  ku − uh kH 1 h |w|H 2 .
From H 2 regularity:
ku − uh k2L2  hku − uh kH 1 ku − uh kL2 ,
and, after dividing one factor
ku − uh kL2  hku − uh kH 1  h2 |u|H 2 .
2

Approximation of Dirichlet boundary conditions

Till now, we have neglected Dirichlet boundary conditions. In this case, the continuous
problem is
Find u ∈ VD : A(u, v) = f (v) ∀ v ∈ V0 ,
where
VD = {v ∈ H 1 : trΓD v = uD } and V0 = {v ∈ H 1 : trΓD v = 0}.
The finite element problem is
Find uh ∈ VhD : A(uh , vh ) = f (vh ) ∀ vh ∈ Vh0 ,
where
VhD = {IT v : v ∈ VD } and Vh0 = {IT v : v ∈ V0 }.
The definition of VhD coincides with {vh ∈ Vh : vh (xi ) = uD (xi ) ∀ vertices xi on ΓD }.
There holds Vh0 ⊂ V0 , but, in general, there does not hold VhD ⊂ VD .
60 CHAPTER 4. FINITE ELEMENT METHOD

Theorem 72 (Error estimate for Dirichlet boundary conditions). Assume that

• A(., .) is coercive on Vh0 :

A(vh , vh ) ≥ α1 kvh k2V ∀ vh ∈ Vh0

• A(., .) is continuous on V :

A(u, v) ≤ α2 kukV kvkV ∀ u, v ∈ V

Then there holds the finite element error estimate

ku − uh kH 1  h|u|H 2

Proof: To make use of the coercivity of A(., .), we need an element in Vh0 . There holds
Galerkin orthogonality A(u − uh , vh ) = 0 ∀ vh ∈ Vh0 :

ku − uh k2V = ku − Ih u + Ih u − uh k2V ≤ 2 ku − Ih uk2V + 2 kIh u − uh k2V

2
≤ 2 ku − Ih uk2V + A(Ih u − uh , Ih u − uh )
α1
2 2
≤ 2ku − Ih uk2V + A(Ih u − u, Ih u − uh ) + A(u − uh , Ih u − uh )
α1 α1
2α 2
≤ 2 ku − Ih uk2V + kIh u − ukkIh u − uh k + 0
α1
2α2
≤ 2 ku − Ih uk2V + kIh u − uk(kIh u − uk + ku − uh k)
α1
2α2 2α2
= (2 + ) ku − Ih uk2V + ku − Ih ukV ku − uh kV
α1 α1
2α2
Next, we apply ab ≤ 12 a2 + 21 b2 for a = α1
ku − Ih ukV and b = ku − uh kV :

2α2 α22 1
ku − uh k2V ≤ (2 + )ku − Ih ukV + 2 2 ku − Ih uk2V + ku − uh k2V
2
α1 α1 2

Moving the term 12 ku − uh k to the left, we obtain

ku − uh k2V  ku − Ih uk2V  h |u|H 2

High order elements

One can obtain faster convergence, if the solution is smooth, and elements of higher order
are used:
Theorem 73. Assume that
4.2. FINITE ELEMENT ERROR ANALYSIS 61

• the solution is smooth: u ∈ H m for m ≥ 2

• all element spaces VT contain polynomials P p for p ≥ 1
• the mesh is quasi-uniform
Then there holds
h−1 ku − Ih ukL2 + ku − Ih ukH 1  hmin{m−1,p} kukH m
The proof is analogous to the case m = 2 and p = 1. The constants in the estimates
depend on the Sobolev index m and on the polynomial order p. Nodal interpolation is
instable (i.e., the constant grow with p) for increasing order p. There exist better choices
to bound the best approximation error.

Graded meshes around vertex singularities

On non-convex meshes domains, the solution is in general not in H 2 , but in some weighted
Sobolev space. The information of the weight can be used to construct proper locally
refined meshes.
On a sector domain with a non-convex corner of angle ω > π, the solution is bounded
in the weighted Sobolev norm
krβ D2 ukL2 ≤ C,
with β = ωπ . One may choose a mesh such that

hT ' hrTβ , ∀T ∈ T
where rT is the distance of the center of the element to the singular corner, and h ∈ R+ is
a global mesh size parameter.
We bound the interpolation error:
X X
ku − IT uk2H 1  h2T |u|H 2 (T ) ' h2 |rβ D2 u|L2 (T )
T ∈T T ∈T
2
' h kr D β 2
uk2L2 (Ω)  C h2
The number of elements in the domain can be roughly estimated by the integral over
the density of elements. The density is number of elements per unit volume, i.e., the inverse
of the area of the element:
Z Z Z
−2 −2β −2
Nel ' −1
|T | dx = h r dx = h r−2β dx ' Ch−2
Ω Ω

In two dimensions, and β ∈ (0, 1), the integral is finite.

Combining the two estimates, one obtains a relation between the error and the number
of elements:
ku − IT uk2V  Nel−1
This is the same order of convergence as in the H 2 regular case !
62 CHAPTER 4. FINITE ELEMENT METHOD

4.3 A posteriori error estimates

We will derive methods to estimate the error of the computed finite element approximation.
Such a posteriori error estimates may use the finite element solution uh , and input data
such as the source term f .
η(uh , f )
An error estimator is called reliable, if it is an upper bound for the error, i.e., there
exists a constant C1 such that

ku − uh kV ≤ C1 η(uh , f ) (4.5)

An error estimator is efficient, if it is a lower bound for the error, i.e., there exists a
constant C2 such that
ku − uh kV ≥ C2 η(uh , f ). (4.6)
The constants may depend on the domain, and the shape of the triangles, but may not
depend on the source term f , or the (unknown) solution u.

One use of the a posteriori error estimator is to know the accuracy of the finite element
approximation. A second one is to guide the construction of a new mesh to improve the
accuracy of a new finite element approximation.
The usual error estimators are defined as sum over element contributions:
X
η 2 (uh , f ) = ηT2 (uh , f )
T ∈T

The local contributions should correspond to the local error. For the common error
estimators there hold the local efficiency estimates

ku − uh kH 1 (ωT ) ≥ C2 ηT (uh , f ).

The patch ωT contains T and all its neighbor elements.

In the following, we consider the Poisson equation −∆u = f with homogenous Dirichlet
boundary conditions u = 0 on ∂Ω. We choose piecewise linear finite elements on triangles.

The Zienkiewicz Zhu error estimator

The simplest a posteriori error estimator is the one by Zienkiewicz and Zhu, the so called
ZZ error estimator.
The error is measured in the H 1 -semi norm:

k∇u − ∇uh kL2

4.3. A POSTERIORI ERROR ESTIMATES 63

Define the gradient p = ∇u and the discrete gradient ph = ∇uh . The discrete gradient
ph is a constant on each element. Let p̃h be the p.w. linear and continuous finite element
function obtained by averaging the element values of ph in the vertices:
1 X
p̃h (xi ) = ph|T for all vertices xi
|{T : xi ∈ T }| T :x ∈T
i

The hope is that the averaged gradient is a much better approximation to the true gradient,
i.e.,
kp − p̃h kL2 ≤ α kp − ph kL2 (4.7)
holds with a small constant α  1. This property is known as super-convergence.It is
indeed true on (locally) uniform meshes, and smoothness assumptions onto the source
term f .
The ZZ error estimator replaces the true gradient in the error p − ph by the good
approximation p̃h :
η(uh ) = kp̃h − ph kL2 (Ω)
If the super-convergence property (4.7) is fulfilled, than the ZZ error estimator is reli-
able:

k∇u − ∇uh kL2 = kp − ph kL2 ≤ kph − peh kL2 + kp − peh kL2

≤ kph − peh kL2 + αkp − ph kL2 ,

and
1
k∇u − ∇uh kL2 ≤ kph − peh kL2 .
1−α
It is also efficient, a similar short application of the triangle inequality.
There is a rigorous analysis of the ZZ error estimator, e.g., by showing equivalence to
the following residual error estimator.

The residual error estimator

The idea is to compute the residual of the Poisson equation

f + ∆ uh ,

in the natural norm H −1 . The classical ∆-operator cannot be applied to uh , since the first
derivatives, ∇uh , are non-continuous across element boundaries. One can compute the
residuals on the elements
f|T + ∆ uh|T ∀T ∈ T ,
and one can also compute the violation of the continuity of the gradients on the edge
E = T1 ∩ T2 . We define the normal-jump term
 
∂uh ∂uh ∂uh
:= |T1 + |T .
∂n ∂n1 ∂n2 2
64 CHAPTER 4. FINITE ELEMENT METHOD

The residual error estimator is

X
η res (uh , f )2 := ηTres (uh , f )2
T

with the element contributions

  2
X ∂uh
ηTres (uh , f )2 := h2T kf + ∆uh k2L2 (T ) + hE .
E:E⊂T
∂n L2 (E)
E⊂Ω

The scaling with hT corresponds to the natural H −1 norm of the residual.

To show the reliability of the residual error estimator, we need a new quasi-interpolation
operator, the Clément- operator Πh . In contrast to the interpolation operator, this operator
is well defined for functions in L2 .
We define the vertex patch of all elements connected with the vertex x
[
ωx = T,
T :x∈T

the edge patch consisting of all elements connected with the edge E
[
ωE = T,
T :E∩T 6=∅

and the element patch consisting of the element T and all its neighbors
[
ωT = T 0.
T 0 :T ∩T 0 6=∅

The nodal interpolation operator Ih was defined as

X
Ih v = v(xi )ϕi ,
xi ∈V

where ϕi are the nodal basis functions. Now, we replace the nodal value v(xi ) by a local
mean value.
Definition 74 (Clément quasi-interpolation operator). For each vertex x,let v ωx be the
mean value of v on the patch ωx , i.e.,
Z
ωx 1
v = v dx.
|ωx | ωx
The Clément operator is X
Πh v := v ωxi ϕi .
xi ∈V

In the case of homogeneous Dirichlet boundary values, the sum contains only inner vertices.
4.3. A POSTERIORI ERROR ESTIMATES 65

Theorem 75. The Clément operator satisfies the following continuity and approximation
estimates:

k∇Πh vkL2 (T )  k∇vkL2 (ωT )

kv − Πh vkL2 (T )  hT k∇vkL2 (ωT )
1/2
kv − Πh vkL2 (E)  hE k∇vkL2 (ωE )

bT of dimension ' 1. The quasi-interpolation

Proof: First, choose a reference patch ω
operator is bounded on H 1 (ωT ):

kv − Πh vkL2 (Tb) + k∇(v − Πh v)kL2 (Tb)  kvkH 1 (bωT ) (4.8)

If v is constant on ωT , then the mean values in the vertices take the same values, and
also (Πh v)|T is the same constant. The constant function (on ωT ) is in the kernel of
kv − Πh vkH 1 (T ) . Due to the Bramble-Hilbert lemma, we can replace the norm on the right
hand side of (4.8) by the semi-norm:

kv − Πh vkL2 (Tb) + k∇(v − Πh v)kL2 (Tb)  k∇vkL2 (bωT ) (4.9)

The rest follows from scaling. Let F : x → hx scale the reference patch ω
bT to the actual
patch ωT . Then

kv − Πh vkL2 (T ) + h k∇(v − Πh v)kL2 (T )  h k∇vkL2 (ωT )

The estimate for the edge term is similar. One needs the scaling of integrals from the
reference edge E
b to E:
1/2
kvkL2 (E) = hE kv ◦ F kL2 (Ê)

Theorem 76. The residual error estimator is reliable:

ku − uh k  η res (uh , f )

Proof: From the coercivity of A(., .) we get

A(u − uh , u − uh ) A(u − uh , v)
ku − uh kH 1  ≤ sup .
ku − uh kH 1 06=v∈H 1 kvkH 1

The Galerkin orthogonality A(u − uh , vh ) = 0 for all vh ∈ Vh allows to insert the Clément
interpolant in the numerator. It is well defined for v ∈ H 1 :
A(u − uh , v − Πh v)
ku − uh kH 1 ≤ sup .
06=v∈H 1 kvkH 1
66 CHAPTER 4. FINITE ELEMENT METHOD

We use that the true solution u fulfills A(u, v) = f (v), and insert the definitions of A(., .)
and f (.):

A(u − uh , v − Πh v) = f (v − Πh v) − A(uh , v − Πh v)
Z Z
= f (v − Πh v) dx − ∇uh ∇(v − Πh v) dx
Ω Z Ω
X XZ
= f (v − Πh v) dx − ∇uh ∇(v − Πh v) dx
T ∈T T T ∈T T

On each T , the finite element function uh is a polynomial. This allows integration by parts
on each element:
XZ X Z Z
∂uh

A(u − uh , v − Πh v) = f (v − Πh v) dx − − ∆uh (v − Πh v) dx + (v − Πh v) ds
T ∈T T T ∈T T ∂T ∂n

All inner edges E have contributions from normal derivatives from their two adjacent
triangles TE,1 and TE,2 . On boundary edges, v − Πh v vanishes.

A(u − uh , v − Πh v)
XZ X Z  ∂uh ∂uh

= (f + ∆uh )(v − Πh v) dx + |T + |T (v − Πh v) ds
T T E E ∂n E,1 ∂n E,2
XZ X Z  ∂uh 
= (f + ∆uh )(v − Πh v) dx + (v − Πh v) ds
T T E E ∂n

Applying Cauchy-Schwarz first on L2 (T ) and L2 (E), and then in Rn :

A(u − uh , v − Πh v)
X X  ∂uh 
≤ kf + ∆uh kL2 (T ) kv − Πh vkL2 (T ) + kv − Πh vkL2 (E)
T E
∂n L2 (E)
X X 1/2  ∂uh  −1/2
−1
= hT kf + ∆uh kL2 (T ) hT kv − Πh vkL2 (T ) + hE hE kv − Πh vkL2 (E)
T E
∂n L2 (E)
( )1/2 ( )1/2
X X
≤ h2T kf + ∆uh k2L2 (T ) h−2 2
T kv − Πh vkL2 (T ) +
T T
( )1/2 ( )1/2
  2
X ∂uh X
+ hE h−1 2
E kv − Πh vkL2 (E)
E
∂n L2 (E) E

We apply the approximation estimates of the Clément operator, and use that only a
bounded number of patches are overlapping:
X X
h−2
T kv − Π h vk2
L2 (T )  k∇vk2L2 (ωT )  k∇vk2L2 (Ω) ,
T T
4.3. A POSTERIORI ERROR ESTIMATES 67

and similar for the edges

X
h−1 2 2
E kv − Πh vkL2 (E) ≤ k∇vkL2 (Ω) .
E

Combining the steps above we observe

A(u − uh , v − Πh v)
ku − uh kV  sup
v∈H 1 kvk1H
nP  ∂uh  2
o1/2
2 2
P
T hT kf + ∆u h k L2 (T ) + E hE ∂n L2 (E)
k∇vkL2 (Ω)
 sup
V ∈H 1 kvkH 1
( )1/2
  2
X
2 2
X ∂uh
≤ hT kf + ∆uh kL2 (T ) + hE ,
T E
∂n L2 (E)

what is the reliability of the error estimator η res (uh , f )

Theorem 77. If the source term f is piecewise polynomial on the mesh, then the error
estimator η res is efficient:
ku − uh kV  η res (uh , f )

Goal driven error estimates

The above error estimators estimate the error in the energy norm V . Some applications
require to compute certain values (such as point values, average values, line integrals, fluxes
through surfaces, ...). These values are descibed by linear functionals b : V → R. We want
to design a method such that the error in this goal, i.e.,

b(u) − b(uh )

is small. The technique is to solve additionally the dual problem, where the right hand
side is the goal functional:

Find w ∈ V : A(v, w) = b(v) ∀ v ∈ V.

Usually, one cannot solve the dual problem either, and one applies a Galerkin method also
for the dual problem:

Find wh ∈ Vh : A(vh , wh ) = b(vh ) ∀ vh ∈ Vh .

In the case of point values, the solution of the dual problem is the Green function (which
is not in H 1 ). The error in the goal is

b(u − uh ) = A(u − uh , w) = A(u − uh , w − wh ).

68 CHAPTER 4. FINITE ELEMENT METHOD

A rigorous upper bound for the error in the goal is obtained by using continuity of the
bilinear-form, and energy error estimates η 1 and η 2 for the primal and dual problem,
respectively:

|b(u − uh )|  ku − uh kV kw − wh kV  η 1 (uh , f ) η 2 (wh , b).

A good heuristic is the following (unfortunately, not correct) estimate

X X
b(u − uh ) = A(u − uh , w − wh )  ku − uh kH 1 (T ) kw − wh kH 1 (T )  ηT1 (uh , f ) ηT2 (wh , b)
T ∈T T
(4.10)
The last step would require a local reliability estimate. But, this is not true.
We can interpret (4.10) that way: The local estimators ηT2 (wh ) provide a way for
weighting the primal local estimators according to the desired goal.

Mesh refinement algorithms

A posteriori error estimates are used to control recursive mesh refinement:

Start with initial mesh T

Loop
compute fe solution uh on T
compute error estimator ηT (uh , f )
if η ≤ tolerance then stop
refine elements with large ηT to obtain a new mesh

The mesh refinement algorithm has to take care of

• generating a sequence of regular meshes
• generating a sequence of shape regular meshes
Red-Green Refinement:
A marked element is split into four equivalent elements (called red refinement):

But, the obtained mesh is not regular. To avoid such irregular nodes, also neighboring
elements must be split (called green closure):
4.4. EQUILIBRATED RESIDUAL ERROR ESTIMATES 69

If one continues to refine that way, the shape of the elements may get worse and worse:

A solution is that elements of the green closure will not be further refined. Instead, remove
the green closure, and replace it by red refinement.

Marked edge bisection:

Each triangle has one marked edge. The triangle is only refined by cutting from the middle
of the marked edge to the opposite vertex. The marked edges of the new triangles are the
edges of the old triangle.
If there occurs an irregular node, then also the neighbor triangle must be refined.

To ensure finite termination, one has to avoid cycles in the initial mesh. This can be
obtained by first sorting the edges (e.g., by length), end then, always choose the largest
edges as marked edge.

Both of these refinement algorithms are also possible in 3D.

4.4 Equilibrated Residual Error Estimates

4.4.1 General framework
Equilibrated residual error estimators provide upper bounds for the discretization error
in energy norm without any generic constant. We consider the standard problem: find
u ∈ V := H01 (Ω) such that
Z Z
λ∇u · ∇v = fv ∀v ∈ V
Ω Ω
70 CHAPTER 4. FINITE ELEMENT METHOD

The left hand side defines the bilinear-form A(·, ·), the right hand side the linear-form f (·).
We define a finite element sub-space Vh ⊂ V of order k, and the finite element solution

find uh ∈ Vh : A(uh , vh ) = f (vh ) ∀ vh ∈ Vh .

We assume that f is element-wise polynomial of order k −1, and λ is element-wise constant

and positive.

The residual r(·) ∈ V ∗ is

r(v) = f (v) − A(uh , v) v∈V

Since
A(u − uh , v) r(v)
ku − uh kA = sup = sup ,
v∈V kvkA v∈V kvkA

we aim in estimating krk in the norm dual to k · kA , which is essentially the H −1 -norm.
In general, the direct evaluation of this norm is not feasible. Using the structure of the
problem, we can represent the residual as
XZ XZ
r(v) = rT v + rE v,
T ∈T T E∈E E

where rT and rE are given given as

 
∂uh
rT = fT + div λT ∇uh|T and rE = λ
∂n E
The element-residual rT is a polynomial of order k − 1 on the element T , and the edge
residual (the normal jump) is a polynomial of order k − 1 on the edge E.
The residual error estimator estimates the residual in terms of weighted L2 -norms:
X h2 X hE
T
krk2 ' η res (uh , f )2 := krT k2L2 (T ) + krE k2L2 (E)
T
λT E
λE

Here, λE is some averaging of the coefficients on the two elements containing the edge E.
The equivalence holds with constants depending on the shape of elements, the relative
jump of the coefficient, and the polynomial order k.
The equilibrated residual error estimator η er is defined in terms of the same data rT
and rE . It satisfies

ku − uh kA ≤ η er reliable with constant 1

ku − uh kA ≥ c η er efficient with a generic constant c

The lower bound depends on the shape of elements and the coefficient λ, but is robust
with respect to the polynomial order k.
4.4. EQUILIBRATED RESIDUAL ERROR ESTIMATES 71

The main idea is the following: Instead of calculating the H −1 -norm of r, we compute
a lifting σ ∆ such that div σ ∆ = r, and calculate the L2 -norm of σ ∆ . Since r is not a regular
function, the equation must be posed in distributional form:
Z
σ ∆ · ∇ϕ = −r(ϕ) ∀ϕ ∈ V
Ω

Then, the residual can be estimated without envolving any generic constant:
R ∆
r(v) σ · ∇v
krkA∗ = sup = sup
v∈V kvkA v kvkA
qR qR
R −1/2 ∆ 1/2 −1 ∆
λ |σ | 2 λ|∇v|2
λ σ · λ ∇v
= sup ≤ sup = kσ ∆ kL2 ,1/λ
v kvkA v kvkA
The norm kσ ∆ k := λ−1 |σ ∆ |2 can be evaluated easily.
R

Remark: The flux-postprocessing σ := λ∇uh + σ ∆ provides a flux σ ∈ H(div) such

that div σ = f , i.e. the flux is in exact equilibrium with the source f . Thus the name.

4.4.2 Computation of the lifting kσ ∆ k

The residual is a functional of the form
X X
r(v) = (rT , v)L2 (T ) + (rE , v)L2 (E) ,
T E

where rT and rE are polynomials of order k − 1. We search for σ ∆ which is element-wise

a vector-valued polynomial of order k, and not continuous across edges. Element-wise
integration by parts gives
Z XZ XZ
σ · ∇ϕ = − div σ|T ϕ + [σ · n]E ϕ.
Ω T T E E

Thus div σ = r in distributional sense reads as

div σ|T = rT and [σ · n]E = −rE
for all elements T and edges E. We could now pose the problem
min kσkL2 ,1/λ
σ∈P k (T )2
div σ=r

We minimize the weighted-L2 norm since we want to find the smallest possible upper bound
for the error. This is already a computable approach. But, the problem is global, and its
solution is of comparable cost as the solution of the original finite element system. The
existence of a σ such that div σ = r also needs a proof.
We want to localize the construction of the flux. Local problems are associated with
vertex-patches ωV = ∪T :V ∈T T . We proceed in two steps:
72 CHAPTER 4. FINITE ELEMENT METHOD

rV
P
1. localization of the residual: r = V

2. local liftings: find σ V such that div σ V = rV on the vertex patch

Then, for σ := σ V there holds div σ = r

P
The localization is given by multiplication of the P 1 vertex basis functions (hat-
functions) φV :
rV (v) := r(φV v)
Since V φV = 1, there holds rV (·) = r(·). The localized residual has the same structure
P P
of element and edge terms:
X X
rV (v) = (rTV , v)L2 (T ) + (rEV , v)L2 (E) ,
T ⊂ωV E⊂ωV

with
rTV = φV rT and rEV = φV rE
The local residual vanishes on constants on the patch:

rV (1) = r(φV 1) = A(u − uh , φV ) = 0

The last equality follows from the Galerkin-orthogonality.

We give an explicit construction of the lifting σ V in terms of the Brezzi-Douglas-Marini
(BDM) element. The k th order BDM element on a triangle is given by VT = [P k ]2 and the
degrees of freedom:
R
(i) E σ · n qi with qi a basis for P k (E)
R
(ii) T div σ qi with qi a basis for P k−1 (T ) ∩ L02 (T )

(iii) T σ · curl qi with qi a basis for P0k+1 (T )

R

Exercise: Show that these dofs are unisolvent. Count dimensions, and prove that [∀i :
ψi (σ) = 0] ⇒ σ = 0.
Now, we give an explicit construction of equilibrated fluxes on a vertex patch. Label
elements T1 , T2 , . . . Tn in a counter-clock-wise order. Edge Ei is the common edge between
triangle Ti−1 and Ti (with identifying T0 = Tn ). We define σ by specifying the dofs of the
BDM element:

1. Start on T1 . We set σn = −rEV 1 on edge E1 . On the edgeR on the Rpatch-boundary we

V
set σn = 0, and on E2 we set σn = const such that ∂T1 σn = T1 rT . We use the
R R
dofs of type (ii) to specify T div σ q = T rTV q ∀q ∈ P k−1 ∩ L02 (T ). Together with get
div σ = rT . Dofs of type (iii) are not needed, and set 0. There holds
Z Z Z Z Z
V V
σn = rT − σn = rT1 + rEV 1
E2 T1 E1 T1 E1
4.5. NON-CONFORMING FINITE ELEMENT METHODS 73

2. Continue with element T2 . On edge E2 common with T1 set σn such that [σ · n]E2 =
rE2 . Otherwise, proceed as on T1 . Thus
Z Z Z Z Z
V V V
σn = rT1 + rE1 + rT2 + rEV 2
E3 T1 E1 T2 E2

3. Continue to element Tn . Observe that on Tn :

Z Xn Z n Z
X
V
σn = rTi + rEV i = 0,
E1 i=1 Ti i=1 Ei

which follows from rV (1) = 0. Thus, also [σ · n]E1 = rEV 1 is satisfied.

This explicit construction proves the existence of an equilibrated flux. Instead of this
explicit construction, one may solve a local constrained optimization problem

min kσkL2 ,λ−1

σ V :div σ V =rV

This applies also for 3D. Furthoer notes

• mixed boundary conditions are possible
• the efficiency for the h-FEM is shown by scaling arguments, and equivalence to the
residual error estimator
• efficiency is also proven to be robust with respect to polynomial order k, examples
show overestimation less than 1.5

Literature:
1. D. Braess and J. Schöberl. Equilibrated Residual Error Estimator for Maxwell’s
Equations. Mathematics of Computation, Vol 77(262), 651-672, 2008
2. D. Braess, V. Pillwein and J. Schöberl: Equilibrated Residual Error Estimates are
p-Robust. Computer Methods in Applied Mechanics and Engineering. Vol 198,
1189-1197, 2009

4.5 Non-conforming Finite Element Methods

In a conforming finite element method, one chooses a sub-space Vh ⊂ V , and defines the
finite element approximation as

Find uh ∈ Vh : A(uh , vh ) = f (vh ) ∀ vh ∈ V h

For reasons of simpler implementation, or even of higher accuracy, the conforming frame-
work is often violated. Examples are:
74 CHAPTER 4. FINITE ELEMENT METHOD

• The finite element space Vh is not a sub-space of V = H m . Examples are the non-
conforming P 1 triangle, and the Morley element for approximation of H 2 .
• The Dirichlet boundary conditions are interpolated in the boundary vertices.
• The curved domain is approximated by straight sided elements
• The bilinear-form and the linear-form are approximated by inexact numerical inte-
gration
The lemmas by Strang are the extension of Cea’s lemma to the non-conforming setting.

The First Lemma of Strang

In the first step, let Vh ⊂ V , but the bilinear-form and the linear-form are replaced by
mesh-dependent forms
Ah (., .) : Vh × Vh → R
and
fh (.) : Vh → R.
We do not assume that Ah and fh are defined on V . We assume that the bilinear-forms
Ah are uniformly coercive, i.e., there exists an α1 independent of the mesh-size such that

Ah (vh , vh ) ≥ α1 kvh k2V ∀ vh ∈ Vh

The finite element problem is defined as

Find uh ∈ Vh : Ah (uh , vh ) = fh (vh ) ∀ vh ∈ V h

Lemma 78 (First Lemma of Strang). Assume that

• A(., .) is continuous on V
• Ah (., .) is uniformly coercive
Then there holds
 
|A(vh , wh ) − Ah (vh , wh )|
ku − uh k  inf ku − vh k + sup
vh ∈Vh wh ∈Vh kwh k
f (wh ) − fh (wh )
+ sup
wh ∈Vh kwh k
Proof: Choose an arbitrary vh ∈ Vh , and set wh := uh − vh . We use the uniform
coercivity, and the definitions of u and uh :

α1 kuh − vh k2V ≤ Ah (uh − vh , uh − vh ) = Ah (uh − vh , wh )

= A(u − vh , wh ) + [A(vh , wh ) − Ah (vh , wh )] + [Ah (uh , wh ) − A(u, wh )]
= A(u − vh , wh ) + [A(vh , wh ) − Ah (vh , wh )] + [fh (wh ) − f (wh )]
4.5. NON-CONFORMING FINITE ELEMENT METHODS 75

Divide by kuh − vh k = kwh k, and use the continuity of A(., .):

|A(vh , wh ) − Ah (vh , wh )| |f (wh ) − fh (wh )|

kuh − vh k  ku − vh k + + (4.11)
kwh k kwh k

Using the triangle inequality, the error ku − uh k is bounded by

ku − uh k ≤ inf ku − vh k + kvh − uh k
vh ∈Vh

The combination with (4.11) proves the result. 2

Example: Lumping of the L2 bilinear-form:

Define the H 1 - bilinear-form
Z Z
A(u, v) = ∇u · ∇v + uv dx,
Ω Ω

and perform Galerkin discretization with P 1 triangles. The second term leads to a non-
diagonal matrix. The vertex integration rule
3
|T | X
Z
v dx ≈ v(xT,α )
T 3 α=1
R
is exact for v ∈ P 1 . We apply this integration rule for the term uv dx:
Z X |T | X3
Ah (u, v) = ∇u · ∇v + u(xT,α )v(xT,α )
T ∈T
3 α=1

The bilinear form is now defined only for u, v ∈ Vh . The integration is not exact, since
uv ∈ P 2 on each triangle.
Inserting the nodal basis ϕi , we obtain a diagonal matrix for the second term:

1 for xi = xj = xT,α
ϕi (xT,α )ϕj (xT,α ) =
0 else

To apply the first lemma of Strang, we have to verify the uniform coercivity
3
X |T | X XZ
2
|vh (xT,α )| ≥ α1 |vh |2 dx ∀ vh ∈ Vh , (4.12)
T
3 α=1 T T

which is done by transformation to the reference element. The consistency error can be
estimated by
3
|T | X
Z
| uh vh dx − uh (xα )vh (xα )|  h2T k∇uh kL2 (T ) k∇vh kL2 (T ) (4.13)
T 3 α=1
76 CHAPTER 4. FINITE ELEMENT METHOD

Summation over the elements give

A(uh , vh ) − Ah (uh , vh )  h2 kuh kH 1 (Ω) kvh kH 1 (Ω)
The first lemma of Strang proves that this modification of the bilinear-form preserves the
order of the discretization error:
 
|A(vh , wh ) − Ah (vh , wh )|
ku − uh kH 1  inf ku − vh kH 1 + sup
vh ∈Vh wh ∈Vh kwh kH 1
|A(Ih u, wh ) − Ah (Ih u, wh )|
 ku − Ih ukH 1 + sup
wh ∈Vh kwh kH 1
2
h kIh ukH 1 kwh kH 1
 h kukH 2 + sup
wh ∈Vh kwh kH 1
 h kukH 2
A diagonal L2 matrix has some advantages:
• It avoids oscillations in boundary layers (exercises!)
• In explicit time integration methods for parabolic or hyperbolic problems, one has to
solve linear equations with the L2 -matrix. This becomes cheap for diagonal matrices.

The Second Lemma of Strang

In the following, we will also skip the requirement Vh ⊂ V . Thus, the norm k.kV cannot
be used on Vh , and it will be replaced by mesh-dependent norms k.kh . These norms must
be defined for V + Vh . As well, the mesh-dependent forms Ah (., .) and fh (.) are defined on
V + Vh . We assume
• uniform coercivity:
Ah (vh , vh ) ≥ α1 kvh k2h ∀ vh ∈ Vh

• continuity:
Ah (u, vh ) ≤ α2 kukh kvh kh ∀ u ∈ V + Vh , ∀ vh ∈ Vh

The error can now be measured only in the discrete norm ku − uh kVh .
Lemma 79. Under the above assumptions there holds
|Ah (u, wh ) − fh (wh )|
ku − uh kh  inf ku − vh kh + sup (4.14)
vh ∈Vh wh ∈Vh kwh kh
Remark: The first term in (4.14) is the approximation error, the second one is called
consistency error.
Proof: Let vh ∈ Vh . Again, set wh = uh − vh , and use the Vh -coercivity:
α1 kuh − vh k2h ≤ Ah (uh − vh , uh − vh ) = Ah (uh − vh , wh )
= Ah (u − vh , wh ) + [fh (wh ) − Ah (u, wh )]
4.5. NON-CONFORMING FINITE ELEMENT METHODS 77

Again, divide by kuh − vh k, and use continuity of Ah (., .):

Ah (u, wh ) − fh (wh )
kuh − vh kh  ku − vh kh +
kwh kh

The rest follows from the triangle inequality. 2

The non-conforming P 1 triangle

The non-conforming P 1 triangle is also called the Crouzeix-Raviart element.
The finite element space generated by the non-conforming P 1 element is

Vhnc := {v ∈ L2 : v|T ∈ P 1 (T ), and v is continuous in edge mid-points}

The functions in Vhnc are not continuous across edges, and thus, Vhnc is not a sub-space
of H 1 . We have to extend the bilinear-form and the norm in the following way:
XZ
Ah (u, v) = ∇u∇v dx ∀ u, v ∈ V + Vhnc
T ∈T T

and X
kvk2h := k∇vk2L2 (T ) ∀ v ∈ V + Vhnc
T ∈T

We consider the Dirichlet-problem with u = 0 on ΓD .

We will apply the second lemma of Strang.
The continuous P 1 finite element space Vhc is a sub-space of Vhnc . Let Ih : H 2 → Vhc be
the nodal interpolation operator.
To bound the approximation term in (4.14), we use the inclusion Vhc ⊂ Vhnc :

inf ku − vh kh ≤ ku − Ih ukH 1  h kukH 2

vh ∈Vhnc

We have to bound the consistency term

r(wh ) = Ah (u, wh ) − f (wh )

XZ XZ
= ∇u∇wh − f wh dx
T T T T
XZ ∂u XZ
= wh ds − (∆u + f ) wh ds
T ∂T ∂n T T
XZ ∂u
= wh ds
T ∂T ∂n

Let E be an edge of the triangle T . Define the mean value wh E . If E is an inner edge,
then the mean value on the corresponding edge of the neighbor element is the same. The
78 CHAPTER 4. FINITE ELEMENT METHOD

∂u
normal derivative ∂n on the neighbor element is (up to the sign) the same. If E is an edge
on the Dirichlet boundary, then the mean value is 0. This allows to subtract edge mean
values:
X X Z ∂u
r(wh ) = (wh − wh E ) ds
T E⊂T E
∂n

Since E wh − wh E ds = 0, we may insert the constant function ∂I∂nh u on each edge:

R

X X Z  ∂u ∂Ih u 
r(wh ) = − (wh − wh E ) ds
T E⊂T E
∂n ∂n

Apply Cauchy-Schwarz on L2 (E):

XX
r(wh ) = k∇(u − Ih u)kL2 (E) kwh − wh E kL2 (E)
T E⊂T

To estimate these terms, we transform to the reference element Tb, where we apply the
Bramble Hilbert lemma. Let T = FT (Tb), and set

b = u ◦ FT
u bh = wh ◦ FT
w

There hold the scaling estimates

|wh |H 1 (T ) ' |w
bh |H 1 (Tb)
1/2 E
kwh − wh E kL2 (E) ' hE kw
b
bh − w
bh kL2 (E)
b

|u|H 2 (T ) ' h−1

T |b
u|H 2 (Tb)
−1/2
k∇(u − Ih u)kL2 (E) ' hE k∇(b
u − Ibh u
b)kL2 (E)

On the reference element, we apply the Bramble Hilbert lemma, once for wh , and once for
u. The linear operator
E
L : H 1 (Tb) → L2 (E)
b
bh → w
b :w bh − w
bh

is bounded on H 1 (Tb) (trace theorem), and Lw = 0 for w ∈ P0 , thus

E
b
kw
bh − w
bh kL2 (E)
b  |w
bh |H 1 (Tb)

Similar for the term in u: There is k∇(u − Ih u)kL2 (E)  kukH 2 (T ) , and u − Ih u vanishes for
u ∈ P 1.
Rescaling to the element T leads to

kwh − wh E kL2 (E)  h1/2 |wh |H 1 (T )

k∇(u − Ih u)kL2 (E)  h1/2 |u|H 2 (T )
4.6. HP - FINITE ELEMENTS 79

This bounds the consistency term

X
r(wh )  h |u|H 2 (T ) |wh |H 1 (T )  h kukH 2 (Ω) kwh kh .
T

The second lemma of Strang gives the error estimate

ku − uh k  h kukH 2

There are several applications where the non-conforming P 1 triangle is of advantage:

• The L2 matrix is diagonal (exercises)

• It can be used for the approximation of problems in fluid dynamics described by the
Navier Stokes equations (see later).

• The finite element matrix has exactly 5 non-zero entries in each row associated with
inner edges. That allows simplifications in the matrix generation code.

4.6 hp - Finite Elements

Let Vhp be a p-th order finite element sub-space of H 1 . By scaling and Bramble-Hilbert
technique one obtains the best-approxiamtion error estimate

inf ku − vhp kH 1 ≤ chm−1 kukH m

vhp ∈Vhp

for m ≤ p + 1. The constant c depends on the order p. If m is fixed, we do obtain reduction

of the approximation error as we increase p. Next we develop methods to obtain so called
p-version error estimates
 m−1
h
inf ku − vhp kH 1 ≤ c p
kukH m ,
vhp ∈Vhp

where c is independent of h and p. This estimate proves also convergence of the p-version
finite element method: One may fix the mesh, and increase the order p.
A detailed analyis of local H m norms allows an optimal balance of mesh-size h and
polynomial order p. This hp-version leads to exponential convergence
α
inf ku − vhp kH 1 ≤ ce−N ,
vhp ∈Vhp

where N is the number of unknowns.

We will prove the p-version estimate, but not the hp-result.
80 CHAPTER 4. FINITE ELEMENT METHOD

4.6.1 Legendre Polynomials

Orthogonal polynomials are important to construct stable basis functions for the p-FEM,
as well as for error estimates.
Let Πn denote the space of polynomials up to order n. We write πn for a generic
polynomial in Πn , with a different value any time it appears.
Definition of Legendre polynomials via Rodrigues’ formula:

1 dn 2
Pn (x) := (x − 1)n .
2n n! dxn

It is a polynomial of degree n. The first few Legendre polynomials are

P0 (x) = 1
P1 (x) = x
P2 (x) = 23 x2 − 1
2

Pn is even if n is even, and Pn is odd if n is odd. Since (x2 − 1)n = x2n − nx2n−2 + π2n−4
(with proper modification for small n) we have

1 (2n)! n n (2n − 2)! n−2

Pn (x) = x − n x + πn−4 (4.15)
2n n! n! 2 n! (n − 2)!

Lemma 80. There holds

Z 1
2
Pn (x)Pm (x) dx = δn,m . (4.16)
−1 2n + 1

Proof. W.l.o.g. let n ≤ m. Multiple integration by parts gives

1 1
dn 2 m
Z Z
n+m n d
2 n!m! Pn (x)Pm (x) dx = (x − 1) (x2 − 1)n dx
−1 −1 dxn dxm
 1
1
dn+1 2 m−1  dn 2 m−1
Z
n d n d
2 m 2

m
= (x − 1) (x − 1) +  (x − 1) (x − 1)
−1 dxn+1 dxm−1  dxn
|dx
m−1
{z }

=0 for x∈{−1,1} −1
= ·Z· ·
1
dn+m 2
= n+m
(x − 1)n (x2 − 1)m dx
−1 dx

For n < m, the first factor of the integrand vanishes, and we have orthogonality. For
4.6. HP - FINITE ELEMENTS 81

n = m this equals
Z 1 Z 1
2n 2
2 (n!) kPn k2L2 = 2 n
(2n)!(x − 1) dx = (2n)! (x − 1)n (x + 1)n
−1 −1
Z 1
n
= −(2n)! (x − 1)n+1 (x + 1)n−1
−1 n + 1
Z 1
n(n − 1)
= (2n)! (x − 1)n+2 (x + 1)n−2 = ...
−1 (n + 1)(n + 2)
Z 1
n! 1
= (2n)! (x − 1)2n dx = (n!)2 22n+1 ,
2n(2n − 1) · · · (n + 1) −1 2n + 1

which proves the scaling.

Next we prove the 3-term recurrency, which can be used for efficient evaluation.

Lemma 81. There holds

(n + 1)Pn+1 (x) = (2n + 1)xPn (x) − nPn−1 (x). (4.17)

Proof. Set r(x) = (n + 1)Pn+1 (x) − (2n + 1)xPn (x) + nPn−1 (x). Using (4.15), we see that
the leading coefficients cancel, and thus r ∈ Πn−2 . From Lemma 80 we get for any q ∈ Πn−2
Z 1 Z 1 Z 1 Z 1
r(x)q(x) dx = (n + 1) Pn+1 q − (2n + 1) Pn xq +n Pn−1 q = 0,
−1 −1 −1 −1
|{z}
∈Πn−1

and thus r = 0.

Lemma 82. Legendre polynomials satisfy the Sturm-Liouville differential equation

d 2 d 
(x − 1) Pn (x) = n(n + 1)Pn (x)
dx dx

Proof. Both sides are in Πn . We compare leading coefficients, for this set Pn = an xn + πn−2
(with an = 2n1n! (2n)!
n!
).
 
d 2 d n
lhs = (x − 1) (an x + πn−2 )
dx dx
d
(x2 − 1)(an nxn−1 + πn−3 )
 
=
dx
d 
an nxn+1 + πn−1

=
dx
= n(n + 1)an xn + πn−2 ,
82 CHAPTER 4. FINITE ELEMENT METHOD

and we get the same leading coefficient for rhs. Furthermore, for q ∈ Πn−1 there holds
Z 1 Z 1
1
(x2 − 1)Pn0 q 0 dx + (x2 − 1)Pn0 q −1

lhs q = −
−1 −1 | {z }
=0
Z 1

0 1
Pn (x2 − 1)q 0 dx − Pn (x2 − 1)q 0 −1 = 0,

=
−1 | {z }
∈Πn−1

and the same for the rhs. Thus the identity is proven.
Lemma 82 implies that the Legendre polynomials are also orthogonal w.r.t.
(u0 , v 0 )L2 ,1−x2 , i.e.
Z 1
(1 − x2 )Pn0 Pm0 = n(n + 1)kPn k2L2 δn,m
−1

4.6.2 Error estimate of the L2 projection

Since polynomials are dense in L2 (−1, 1), we get
∞
X
u= an Pn
n=0

with the generalized Fourier coefficients

(u, Pn )L2
an = ,
kPn k2L2

and ∞
X
kuk2L2 = a2n kPn k2 .
n=0
Π
Let PL2p denote the L2 -projection onto Πp . There holds
p
Π
X
PL2p u = an P n
n=0

The projection error is

∞
X
Π
ku − PL2p uk2L2 = a2n kPn k2
n=p+1

Lemma 83. The L2 -projection error satisfies

Π 1
ku − PL2p ukL2 (−1,1) ≤ p |u|H 1 (−1,1) (4.18)
(p + 1)(p + 2)
4.6. HP - FINITE ELEMENTS 83

Proof. Since Pn are orthogonal also w.r.t. (u0 , v 0 )L2 ,1−x2 , there holds
X
ku0 k2L2 ,1−x2 = a2n kPn0 k21−x2 ,
n∈N

provided that u is in H 1 . The projection error satisfies

Π
X X 1
ku − PL2p uk2L2 = a2n kPn k2L2 = a2n kP 0 k2 2
n>p n>p
n(n + 1) n 1−x
1 X 1 X
≤ a2n kPn0 k21−x2 ≤ a2 kP 0 k2 2
(p + 1)(p + 2) n>p
(p + 1)(p + 2) n∈N n n 1−x
1
= ku0 k21−x2
(p + 1)(p + 2)

Finally, the result follows from

Z 1 Z 1
0 2
2
(1 − x )(u ) dx ≤ (u0 )2 dx.
−1 −1
84 CHAPTER 4. FINITE ELEMENT METHOD

Similar as in Lemma 82 on shows also

dm  2 m d
m 
(x − 1) P n (x) = (n + m)(n + m − 1) . . . (n − m + 1)Pn (x)
dxm dxm
for m ≤ n, and, as in Lemma 83
s
Π (p − m + 1)!
ku − PL2p ukL2 ≤ |u|H m .
(p + m + 1)!

4.6.3 Orthogonal polynomials on triangles

Orthogonal polynomials on tensor product elements are simply constructed by tensoriza-
tion. Orthogonal polynomials on simplicial elements are more advanced. They are based
on Jacobi polynomials:
For α, β > −1, Jacobi polynomials are defined by
(−1)n 1 dn
Pn(α,β) (x) := n 2 n


w(x)(1 − x )
2 n! w(x) dxn
with the weight function
w(x) = (1 − x)α (1 + x)β .
Jacobi polynomials are orthogonal w.r.t. the weighted inner product
Z 1
2α+β+1 Γ(n + α + 1) Γ(n + β + 1)
w(x)Pn(α,β) (x)Pm(α,β) (x)dx = δn,m .
−1 2n + α + β + 1 n!Γ(n + α + β + 1)
(0,0)
Note that Pn = Pn .
Define the unit-triangle T with vertices (−1, 0), (1, 0) and (0, 1).
Lemma 84 (Dubiner basis). The functions
(2i+1,0)
x
(1 − y)i Pj


ϕi,j (x, y) := Pi 1−y (2y − 1) i+j ≤p
form an L2 (T )-orthogonal basis for Πp (T ).
x
Proof. Note that ϕi,j ∈ Πi+j . Substitution ξ = 1−y leads to
Z
ϕij (x, y)ϕkl (x, y) d(x, y) =
T
Z 1 Z 1−y
(2i+1,0) (2k+1,0)
x
(1 − y)i Pj x
(1 − y)k Pl



= Pi 1−y (2y − 1)Pk 1−y (2y − 1) dxdy
0 −1+y
Z 1 Z 1
(2i+1,0) (2k+1,0)
= Pi (ξ)Pk (ξ)(1 − y)i+k+1 Pj (2y − 1)Pl (2y − 1) dξdy
0 −1
Z 1
2 (2i+1,0) (2i+1,0)
= δi,k kPi kL2 (1 − y)2i+1 Pj (2y − 1)Pl (2y − 1) dy
0
= Cij δi,k δj,l
4.6. HP - FINITE ELEMENTS 85

4.6.4 Projection based interpolation

By means of the orthogonal polynomials one shows approximation error estimates of the
form
inf ku − qkH k (T ) ≤ cpk−m |u|H m (T ) m ≥ k,
q∈Πp (T )

with c 6= c(p), easily in 1D and tensor product elements, and also on n-dimensional sim-
plices [Braess+Schwab: Approximation on simplices with respect to weighted Sobolev
norms, J. Approximation Theory 103, 329-337 (2000)].
But, an interpolation operator to an H 1 -conforming finite element space has to satisfy
continuity constraints across element boundaries. We show that we get the same rate of
convergence under these constraints.

The 1D case
Let I = (−1, 1). We define the operator Ip : H 1 (I) → Πp such that

Ip u(x) = u(x) x ∈ {−1, 1} (4.19)

Z Z
0 0
(Ip u) q = u0 q 0 ∀ q ∈ Πp,0 (I), (4.20)
I I

where Πp,0 (D) := {q ∈ Πp (D) : q = 0 on ∂D}. This procedure is exactly a p-version

Galerkin-method for the Dirichlet problem. Since boundary values are preserved, the
interpolation operator produces a globally continuous function. The operator Ip is a kind
of mixture of interpolation and projection, thus the term projection based interpolation
introduced by Demkowicz has been established.

Lemma 85 (Commuting diagram). There holds

Π
ΠL2p−1 u0 = (Ip u)0

Proof. The range of both sides is Πp−1 . We have to show that (Ip u)0 is indeed the L2 -
projection of u0 , i.e. Z Z
(Ip u) q = u0 q
0
∀ q ∈ Πp−1 .
I I
0
R
This holds since {q : q ∈ Πp,0 } = {q ∈ Πp−1 : q = 0} and (4.20), and
Z 1 Z 1
0
(Ip u) 1 = (Ip u)(1) − (Ip u)(−1) = u(1) − u(−1) = u0 1.
−1 −1

The H 1 -error estimate follows directly from the commuting diagram property:
Π c
|u − Ip u|H 1 (I) = ku0 − (Ip u)0 kL2 = ku0 − PL2p−1 u0 kL2 ≤ |u0 |H m−1 .
pm−1
86 CHAPTER 4. FINITE ELEMENT METHOD

By the Aubin-Nitsche technique one obtains an extra p for the L2 -error:

1 1
ku − Ip ukL2 (I)  |u − Ip u|H 1 (I) ≤ m |u|H m (I)
p p

One also gets for q ∈ Πp

|u − Ip u|H 1 = |u − q − Ip (u − q)|1 ≤ ||Id − Ip ||H 1 →H 1 |u − q|H 1

Projection based interpolation on triangles

We define the operator Ip : H 2 (T ) → Πp (T ) as follows:

Ip u(x) = u(x) ∀ vertices x (4.21)

Z Z
∂τ (Ip u)∂τ q = ∂τ u∂τ q ∀ edges E, ∀ q ∈ Πp,0 (E) (4.22)
E
Z ZE

∇(Ip u)∇q = ∇u∇q ∀ q ∈ Πp,0 (T ) (4.23)

T T

Note that Ip u on the edge E depends only on u|E , and thus the interpolant is continuous
across neighbouring elements.

Lemma 86. Let v ∈ C(∂T ) such that v|E ∈ Πp (E). Then there exists an extension
ṽ ∈ Πp (T ) such that ṽ|∂T = v and

|ṽ|H 1 ≤ c |v|H 1/2 (∂T ) ,

where c is independent of p.

Major steps have been shown in exercises 5.2 and 6.6. Note that the minimal-norm
extension ṽ is the solution of the Dirichlet problem, i.e.
Z
∇ṽ∇w = 0 ∀ w ∈ Πp,0
T

Theorem 87 (error estimate). There holds

X 1 1
|u − Ip u|H 1  inf |u − q|H 1 (T ) + √ inf |u − q|H 1 (E)  m−1 |u|H m
q∈Πp
E⊂∂T
p q∈Πp (E) p

for u ∈ H m , m ≥ 2.

Proof. Let up be the | · |H 1 best approximation to u, i.e.

Z Z
∇up ∇v = ∇u∇v ∀ v ∈ Πp ,
T T
4.6. HP - FINITE ELEMENTS 87
R R
and, for uniqueness, mean values are preserved: T
up = T
u. There holds
c
|u − up |H 1 ≤ |u|H m .
pm−1
We apply the triangle inequality:

|u − Ip u|H 1 ≤ |u − up |H 1 + |up − Ip u|H 1

Since Z Z Z
∇up ∇v = ∇u∇v = ∇Ip u∇v ∀ v ∈ Πp,0 (T ),
T T T
we have that
up − Ip u ⊥H 1 Πp,0 ,
i.e. up − Ip u is solution of the Dirichlet problem with boundary values (up − Ip u)|∂T .
Lemma 86 implies that

|up − Ip u|H 1 (T )  |up − Ip u|H 1/2 (∂T )

We insert an u on the boundary to obtain

|u − Ip u|H 1  |u − up |H 1 (T ) + |up − Ip u|H 1/2 (∂T )

≤ |u − up |H 1 (T ) + |up − u|H 1/2 (∂T ) + |u − Ip u|H 1/2 (∂T )
1/2 1/2
≤ |u − up |H 1 (T ) + ku − Ip ukL2 (∂T ) |u − Ip u|H 1 (∂T ) .

In the last step we used that H 1/2 (∂T ) = [L2 , H 1 ]1/2 (i.e. the interpolation space). Next,
we observe that Ip restricted to one edge E is exactly the 1D operator. Using Aubin-Nitsche
we get

|u − Ip u|H 1  |u − up |H 1 (T ) + p−1/2 ku − Ip ukH 1 (∂T )

X
 |u − up |H 1 (T ) + p1−m |u|H m−1/2 (E)
E
 p1−m |u|H m (T )

In the last step we used the trace theorem.

88 CHAPTER 4. FINITE ELEMENT METHOD
Chapter 5

Linear Equation Solvers

The finite element method, or other discretization schemes, lead to linear systems of equa-
tions
Au = f.
The matrices are typically
• of large dimension N (104 − 108 unknowns)
• and sparse, i.e., there are only a few non-zero elements per row.
A matrix entry Aij is non-zero, if there exists a finite element connected with both
degrees of freedom i and j.
A 1D model problem: Dirichlet problem on the interval. A uniform grid with n ele-
ments. The matrix is
 
2 −1
 −1 2 −1 
 
A=
 .. .. .. 
,
. . . 
 
 −1 2 −1 
−1 2 (n−1)×(n−1)

A 2D model problem: Dirichlet problem on a unit-square. A uniform grid with 2n2

triangles. The unknowns are enumerated lexicographically:
h

2
(n−1)

1 2 n−1

89
90 CHAPTER 5. LINEAR EQUATION SOLVERS

The FEM - matrix of dimension N = (n − 1)2 is

   
D −I 4 −1
 −I D −I   −1 4 −1 
   
A=
 .. .. .. 
with D=
 .. .. .. 
,
. . .  . . . 
   
 −I D −I   −1 4 −1 
−I D −1 4 (n−1)×(n−1)

and the (n − 1) × (n − 1) identity matrix I.

5.1 Direct linear equation solvers

Direct solvers are factorization methods such as LU -decomposition, or Cholesky factoriza-
tion. They require in general O(N 3 ) = O(n6 ) operations, and O(N 2 ) = O(n4 ) memory. A
fast machine can perform about 109 operations per second1 . This corresponds to

n ∼N time memory
10 102 1 ms 80 kB
100 104 16 min 800 MB
1000 106 30 years 8 TB

A band-matrix of (one-sided) band-width b is a matrix with

Aij = 0 for |i − j| > b

The LU -factorization maintains the band-width. L and U are triangular factors of band-
width b. A banded factorization method costs O(N b2 ) operations, and O(N b) memory.
For the 1D example, the band-with is 1. Time and memory are O(n). For the 2D example,
the band width is O(n). The time complexity is O(n4 ), the memory complexity is O(n3 ).
This corresponds to

n time memory
10 10 µs 8 kB
100 0.1 s 8 MB
1000 16 min 8 GB

1
time of writing was 2003
5.1. DIRECT LINEAR EQUATION SOLVERS 91

Block-elimination methods
By splitting the unknowns into two groups, we rewrite the equation Au = f as a block
system     
A11 A12 u1 f1
= .
A21 A22 u2 f2
First, expressing u1 from the first row gives

u1 = A−1
11 (f1 − A12 u2 ),

and the Schur-complement equation to determine u2

(A22 − A21 A−1 A )u = f2 − A21 A−1

11 f1 .
| {z 11 12} 2
=:S

This block-factorization is used in sub-structuring algorithms: Decompose the domain into

n n
m×m sub-domains, each one containing 2 m ×m triangles. Split the unknowns into interior
(I), and coupling (C) unknowns.
m=2

The interior ones corresponding to different sub-domains have no connection in the matrix.
The block matrix is
 
AI,1 AIC,1

AI AIC
  ... .. 
=
 . 
ACI AC

 AI,m2 AIC,m2 
ACI,1 · · · ACI,m2 AC

Factorizing the block-diagonal interior block AI splits into m2 independent factorization

problems. If one uses a banded factorization, the costs are
 n 4 n4
m2 = 2
m m
Computing the Schur complement
m 2
X
S = AC − ACI A−1
I AIC = AC − ACI,i A−1
I,i AIC,i
i=1
92 CHAPTER 5. LINEAR EQUATION SOLVERS

is of the same cost. The Schur complement is of size mn, and has band-width n. Thus,
the factorization costs O(mn3 ). The total costs are of order
n4
2
+ mn3
m
Equilibrating both terms lead to the optimal number of sub-domains m = n1/3 , and to the
asymptotic costs
n3.33
If a parallel computer is used, the factorization of AI and the computation of Schur com-
plements can be performed in parallel.
The hierarchical sub-structuring algorithm, known as nested dissection, eliminates in-
terior unknowns hierarchically:

Let n = 2L . On level l, with 1 ≤ l ≤ L, one has 4l sub-domains. Each sub-domain has

O(2L−l ) unknowns. The factorization of the inner blocks on level l costs
4l (2L−l )3 = 23L−l
Forming the Schur-complement is of the same cost. The sum over all levels is
L  
X
3L−l 3L 1 1
2 =2 + + . . . ≈ 23L
l=1
2 4
The factorization costs are O(n3 ). Storing the matrices on each level costs
4l (2L−l )2 = 22L .
The total memory is O(L × 22L ) = O(n2 log n).
This corresponds to
n time memory
10 1 µs 3 kB
100 1 ms 500 kB
1000 1 s 150 MB

A corresponding sparse factorization algorithm for matrices arising from unstructured

meshes is based on minimum degree ordering. Successively, the unknowns with the least
connections in the matrix graph are eliminated.
In 2D, a direct method with optimal ordering is very efficient. In 3D, the situation is
worse for the direct solver. There holds N = n3 , time complexity = O(N 2 ), and memory
= O(N 1.33 ).
5.2. ITERATIVE EQUATION SOLVERS 93

5.2 Iterative equation solvers

Iterative equation solvers improve the accuracy of approximative solution by an successive
process. This requires in general much less memory, and, depending on the problem and
on the method, may be (much) faster.

The Richardson iteration

A simple iterative method is the preconditioned Richardson iteration (also known as simple
iteration, or Picard iteration):

start with arbitrary u0

for k = 0, 1, . . . convergence
dk = f − Auk
wk = C −1 dk
uk+1 = uk + τ wk

Here, τ is a damping parameter which may be necessary to ensure convergence. The

matrix C is called a preconditioner. It should fulfill

1. C is a good approximation to A

2. the matrix-vector multiplication w = C −1 d should be cheap

A simple choice is C = diag A, the Jacobi preconditioner. The application of C −1 is cheap.

The quality of the approximation C ≈ A will be estimated below. The optimal choice for
the first criterion would be C = A. But, of course, w = C −1 d is in general not cheap.
Combining the steps, the iteration can be written as

uk+1 = uk + τ C −1 (f − Auk )

Let u be the solution of the equation Au = f . We are interested in the behavior of the
error uk − u:

uk+1 − u = uk − u + τ C −1 (f − Auk )
= uk − u + τ C −1 (Au − Auk )
= (I − τ C −1 A)(uk − u)

We call the matrix

M = I − τ C −1 A
the iteration matrix. The error transition can be estimated by

kuk+1 − uk ≤ kM k kuk − uk.

94 CHAPTER 5. LINEAR EQUATION SOLVERS

The matrix norm is the associated matrix norm to some vector norm. If ρ := kM k < 1,
then the error is reduced. The error after k steps is

kuk − uk ≤ ρk ku0 − uk

To reduce the error by a factor ε (e.g., ε = 10−8 ), one needs

log ε
Nits =
log ρ

iterations.
We will focus on the symmetric (A = AT ) and positive definite (uT Au > 0 for u 6=
0) case (short: SPD). Then it makes sense to choose symmetric and positive definite
preconditioners C = C T . Eigenvalue decomposition allows a sharp analysis. Pose the
generalized eigenvalue problem
Az = λCz.
Let (λi , zi ) be the set of eigen-pairs. The spectrum is σ{C −1 A} = {λi }. The eigen-vectors
zi are normalized to
kzi kC = 1
The eigenvalues can are bounded from below and from above by the Rayleigh quotient:

v T Av v T Av
min ≤ λ i ≤ max
v v T Cv v v T Cv
The ratio of largest to smallest eigen-value is the relative spectral condition number

λN
κ=
λ1

We will establish the spectral bounds

γ1 v T Cv ≤ v T Av ≤ γ2 v T Cv ∀ v ∈ RN ,

which allow to bound the eigenvalues

λi ∈ [γ1 , γ2 ],

and the condition number κ ≤ γγ12 .

P
A vector v can be expressed in terms of the eigen-vector basis zi as v = vi ei . There
holds
X
kvk2C = vi2
X
kvk2A = λi vi2
5.2. ITERATIVE EQUATION SOLVERS 95

Lemma 88. The iteration matrix M can be bounded in A-norm and in C-norm:

kM kA ≤ sup |1 − τ λ|
λ∈[γ1,γ2 ]

kM kC ≤ sup |1 − τ λ|
λ∈[γ1,γ2 ]
P
Proof: Express v = vi zi . Then
X X
M v = (I − τ C −1 A)v = vi (I − τ C −1 A)zi = vi (1 − τ λi )zi

The norm is
X
kM vk2A = λi vi2 (1 − τ λi )2
X
≤ sup(1 − τ λi )2 λi vi2
i
≤ sup (1 − τ λ)2 kvk2A
λ∈[γ1 ,γ2 ]

and thus
kM vkA
kM kA = sup ≤ sup |1 − τ λ|.
v∈Rn kvkA λ∈[γ1 ,γ2 ]

The proof is equivalent for kM kC . 2

1−τγ

1
1−τγ 1
γ2
γ
γ1
1−τγ2

The optimal choice of the relaxation parameter τ is such that

1 − τ γ1 = −(1 − τ γ2 ),

i.e.,
2
τ=
γ1 + γ2
The convergence factor is
γ2 − γ1
1 − τ γ1 = .
γ2 + γ1
Assume we knew sharp spectral bounds γ1 = λ1 and γ2 = λN . Then the convergence factor
is
κ−1 2
kM k = ≈1−
κ+1 κ
96 CHAPTER 5. LINEAR EQUATION SOLVERS

The number of iterations to reduce the error by a factor ε is

log ε log ε κ
Nits = ≈ = log ε−1
log ρ −2/κ 2
Take the 1D stiffness matrix of dimension N × N :
 
2 −1
 −1 2 −1 
 
A=
 .. .. .. ,

. . .
 
 −1 2 −1 
−1 2

and the trivial preconditioner C = I. The eigen-vectors zi and eigen-values λi are

 
ijπ
zi = sin
N + 1 j=1,...N
iπ
λi = 2 − 2 cos( )
N +1
The extremal eigenvalues are
π π2
λ1 = 2 − 2 cos( )≈ 2
N +1 N
Nπ π2
λN = 2 − 2 cos( ≈ 4 − 2.
N +1 N
The optimal damping is
2 1
τ= = ,
λ1 + λN 2
and the convergence factor is
2λ1 2π 2
kM k ≈ 1 − ≈1− 2
λN N
The number of iterations is
Nits ' log ε−1 N 2

For the 2D model problem with N = (n − 1)2 , the condition number behaves like

κ ' n2 .

The costs to achieve a relative accuracy ε are

Nits × Costs-per-iteration ' log ε−1 n2 N ' log ε−1 n4

The costs per digit are comparable to the band-factorization. The memory requirement is
optimal O(N ).
5.2. ITERATIVE EQUATION SOLVERS 97

The gradient method

It is not always feasible to find the optimal relaxation parameter τ a priori. The gradient
method is a modification to the Richardson method to find automatically the optimal
relaxation parameter τ :
The first steps are identic:

dk = f − Auk wk = C −1 dk

Now, perform the update

uk+1 = uk + τ wk
such that the error is minimal in energy norm:

Find τ such that ku − uk+1 kA = min!

Although the error cannot be computed, this minimization is possible:

ku − uk+1 k2A = ku − uk − τ wk k2A

= (u − uk )T A(u − uk ) − 2τ (u − uk )T Awk + τ 2 (wk )T Awk

This is a convex function in τ . It takes its minimum at

0 = 2(u − uk )T Awk + 2τopt (wk )T Awk ,

i.e.,
wk A(u − uk ) w k dk
τopt = =
(wk )T Awk (wk )T Awk
Since the gradient method gives optimal error reduction in energy norm, its convergence
rate can be estimated by the Richardson iteration with optimal choice of the relaxation
parameter:
κ−1
ku − uk+1 kA ≤ ku − uk kA
κ+1

The Chebyshev method

We have found the optimal choice of the relaxation parameter for one step of the iteration.
If we perform m iterations, the overall rate of convergence can be improved by choosing
variable relaxation parameters τ1 , . . . τm .
The m-step iteration matrix is

M = Mm . . . M2 M1 = (I − τm C −1 A) . . . (I − τ1 C −1 A).

By diagonalization, the A-norm and C-norm are bounded by

kM k ≤ max |(1 − τ1 λ) . . . (1 − τm λ)|

λ∈[γ1 ,γN ]
98 CHAPTER 5. LINEAR EQUATION SOLVERS

The goal is to optimize τ1 , . . . τm :

min max |(1 − τ1 λ) . . . (1 − τm λ)|
τ1 ,...τm λ∈[γ1 ,γN ]

This is a polynomial in λ, of order m, and p(0) = 1:

min max |p(λ)|. (5.1)
p∈P m λ∈[γ1 ,γN ]
p(0)=1

This optimization problem can be solved explicitely by means of Chebyshev polynomials.

These are the polynomials defined by

cos(m arccos(x)) |x| ≤ 1
Tm (x) =
cosh(m arccosh(x)) |x| > 1
The Tm fulfill the recurrence relation
T0 (x) = 1
T1 (x) = x
Tm+1 (x) = 2xTm (x) − Tm−1 (x)
The Tm fulfill also
1h √ √ i
Tm (x) = (x + x2 − 1)m + (x + x2 − 1)−m
2
The optimum of (5.1) is
 
2x−γ1 −γ2
Tm γ2 −γ1

2x − γ1 − γ2

p(x) =   = Cm Tm
Tm −γ1 −γ2 γ2 − γ1
γ2 −γ1

The numerator is bounded by 1 for the range γ1 ≤ x ≤ γ2 . The factor Cm can be

computed as √ √
2cm γ2 − γ1
Cm = with c= √ √
1 + c2m γ2 + γ1
Using the condition number we have
2
c≈1− √ ,
κ
and
2
Cm ≈ (1 − √ )m
κ
Now, an error reduction by a factor of ε can be achieved in
√
Nits ≈ log ε−1 κ
steps. The original method by choosing m different relaxation parameters τk is not a good
choice, since
5.2. ITERATIVE EQUATION SOLVERS 99

• it is not numerically stable

• one has to know a priori the number of iterations
The recurrence relation for the Chebyshev polynomials leads to a practicable iterative
method called Chebyshev iteration.

The conjugate gradient method

The conjugate gradient algorithm automatically finds the optimal relaxation parameters
for the best k-step approximation.
Let p0 , p1 , . . . be a finite sequence of A-orthogonal vectors, and set
Vk = span{p0 , . . . pk−1 }
We want to approximate the solution u in the linear manifold u0 + Vk :
min ku − vkA
v∈u0 +Vk

We represent uk as
k−1
X
uk = u0 + α l pl
l=0
The optimality criteria are
k−1
X
0 = (u − uk , pj )A = (u − u0 − αl pl , pj )A 0 ≤ j < k.
l=0

The coefficients αl follow from the A-orthogonality:

(u − u0 )T Apl (f − Au0 )T pl
αl = =
pTl Apl pTl Apl
The αl are computable, since the A-inner product was chosen. The best approximations
can be computed recursively:
uk+1 = uk + αk pk
Since uk − u0 ∈ Vk , and pk ⊥A Vk , there holds
(f − Auk )T pk
αk = .
pTk Apk

Any k-step simple iteration approximates the solution uk in the manifold

u0 + Kk (d0 )
with the Krylov space
Kk (d0 ) = {C −1 d0 , C −1 AC −1 d0 , . . . , C −1 (AC −1 )k−1 d0 }.
Here, d0 = f − Au0 is the initial residual. The conjugate gradient method computes an
A-orthogonal basis of the Krylov-space. The term conjugate is equivalent to A-orthogonal.
100 CHAPTER 5. LINEAR EQUATION SOLVERS

Conjugate Gradient Algorithm:

Choose u0 , compute d0 = f − Au0 , set p0 = C −1 d0 .
for k = 0, 1, 2, . . . compute

dTk pk
αk =
pTk Apk
uk+1 = u k + α k pk
dk+1 = dk − αk Apk
dT C −1 Apk
βk = − k+1T
pk Apk
−1
pk+1 = C dk+1 + βk pk

Remark 89. In exact arithmetic, the conjugate gradient algorithm terminates at a finite
number of steps k ≤ N .

Theorem 90. The conjugate gradient algorithm fulfills for k ≤ k

1. The sequence pk is A-orthogonal. It spans the Krylov-space Kk (d0 )

2. The uk minimizes
min ku − vkA
v∈u0 +Kk (d0 )

3. There holds the orthogonality

dTk pl = 0 ∀l < k

Proof: Per induction in k. We assume

pTk Apl = 0 ∀l<k

dTk pl = 0 ∀l<k

This is obvious for k = 0. We prove the property for k + 1: For l < k there holds

dTk+1 pl = (dk − αk Apk )T pl = dTk pl − αk pTk Apl = 0

per induction. For l = k there is

dTk pk T
dTk+1 pk = (dk − αk Apk )T pk = dTk pk − p Apk = 0.
pTk Apk k

Next, prove the A-orthogonality of the pk . For l < k we have

(pk+1 , pl )A = (C −1 dk+1 + βk pk , pl )A
= dTk+1 C −1 Apl
5.2. ITERATIVE EQUATION SOLVERS 101

There is
C −1 Apl ∈ span{p0 , . . . pk },
and dTk+1 pj = 0 for j ≤ k. For l = k there is

(pk+1 , pk )A = (C −1 dk+1 + βk pk , pk )A
−1 dTk+1 C −1 Apk T
= (C dk+1 , pk )A − pk Apk = 0
pTk Apk
2
The coefficients αk and βk should be computed by the equivalent, and numerically more
stable expressions
dTk C −1 dk dTk+1 C −1 dk+1
αk = T βk = .
pk Apk dTk C −1 dk

Theorem 91. The conjugate gradient iteration converges with the rate
√ k
κ−1
ku − uk kA ≤ √
κ+1

Proof: The conjugate gradient gives the best approximation in the Krylov space. Thus,
it can be bounded by the Chebyshev method leading to that rate.

The conjugate gradient iteration is stopped as soon as a convergence criterion is fulfilled.

Ideally, one wants to reduce the error in the energy norm by a factor ε:

ku − uk kA ≤ ε ku − u0 kA

But, the energy error cannot be computed. We rewrite

ku − uk k2A = kA−1 (f − Auk )k2A = kA−1 dk k2A = dTk A−1 dk

If C is a good approximation to A, then also C −1 is one to A−1 . The error can be

approximated by
dTk C −1 dk .
This scalar is needed in the conjugate gradient iteration, nevertheless.

For solving the 2D model problem with C = I, the time complexity is

√
log ε−1 N κ = log ε−1 n3

The costs for one digit are comparable to the recursive sub-structuring algorithm. In 3D,
the conjugate gradient method has better time complexity.
102 CHAPTER 5. LINEAR EQUATION SOLVERS

5.3 Preconditioning
In the following, let the symmetric and positive definite matrix A arise from the finite
element discretization of the H 1 -elliptic and continuous bilinear-form A(., .). We construct
preconditioners C such that the preconditioning action

w = C −1 × d

is efficiently computable, and estimate the spectral bounds

γ1 uT Cu ≤ uT Au ≤ γ2 uT Cu ∀ u ∈ RN

The analysis of the preconditioner is performed in the finite element framework. For
this, define the Galerkin isomorphism
X
G : RN → Vh : u → u = ui ϕi ,

where ϕi are the fe basis functions. Its dual is

G∗ : Vh∗ → RN : d(·) → (d(ϕi ))i=1,...,N .

To distinguish vectors and the corresponding finite element functions, we write vectors
u ∈ RN with underlines (when necessary).
The evaluation of the quadratic form is

uT Au = A(Gu, Gu) ' kGuk2H 1

The Jacobi Preconditioner

The Jacobi preconditioner C is
C = diag A.
The preconditioning action is written as
N
X
C −1 × d = ei (eTi Aei )−1 eTi d
i=1

Here, ei is the ith unit-vector. Thus, eTi Aei gives the ith diagonal element Aii of the matrix,
which is
Aii = A(ϕi , ϕi ) ' kϕi k2H 1 .
The quadratic form generated by the preconditioner is
N
X N
X
T
u Cu = u2i kϕi k2A ' u2i kϕi k2H 1
i=1 i=1
5.3. PRECONDITIONING 103

Theorem 92. Let h be the minimal mesh-size of a shape-regular triangulation. Then there
holds
h2 uT Cu  uT Au  uT Cu (5.2)
Proof: We start to prove the right inequality
X X
uT Au = k ui ϕi k2A  uT Cu = u2i kϕi k2A .
i i

We define the interaction matrix O with entries


1 A(ϕi , ϕj ) 6= 0
Oij =
0 else
On a shape regular mesh, only a (small) finite number of basis functions have overlapping
support. Thus, O has a small number of entries 1 per row. There holds
X XX
k ui ϕi k2A = ui uj A(ϕi , ϕj )
i i j
XX
= ui uj Oij A(ϕi , ϕj )
i j
XX
≤ (ui kϕi kA )Oij (uj kϕj kA )
i j
X
≤ ρ(O) (ui kϕi kA )2
i
T
= ρ(O)u Cu.
T
The spectral radius ρ(O) = maxx∈RN xkxkOx2 is bounded by the (small) finite row-sum norm

of O.
The other estimate is proven element by element. Note that
X X
uT Au ' kuk2H 1 (Ω) = k ui ϕi k2H 1 (T )
T i

and X XX
uT Cu ' kui ϕi k2H 1 (Ω) = kui ϕi k2H 1 (T ) .
i T i
We prove the inequality for each individual element. The triangle T has diameterP hT . On
T , we expand u in terms of the element shape functions ϕα , namely u|T = 3α=1 uα ϕα .
We transform to the reference element Tb:
X X X
k uα ϕα k2H 1 (T ) = k uα ϕα k2L2 (T ) + k∇ uα ϕα k2L2 (T )
α α α
X X
' h2T k bα k2L2 (Tb)
uα ϕ + k∇ bα k2L2 (Tb)
uα ϕ
α α
X
≥ h2T k bα k2L2 (Tb)
uα ϕ
α
104 CHAPTER 5. LINEAR EQUATION SOLVERS

and
X X X
kuα ϕα k2H 1 (T ) = kuα ϕα k2L2 (T ) + k∇uα ϕα k2L2 (T )
α α α
X X
' h2T bα k2L2 (Tb)
kuα ϕ + bα k2L2 (Tb)
k∇uα ϕ
α α
X
 bα k2L2 (Tb)
kuα ϕ
α

nP o1/2
bα k2L are norms on R3 . Since all
P
Both, (u)α → k bα kL2 (Tb) and u →
α uα ϕ α kuα ϕ 2 (T )
b
norms in R3 are equivalent, we have
X X
kuα ϕα k2H 1 (T )  h−2
T k uα ϕα k2H 1 (T ) . (5.3)
α α

By summing over all elements and choosing h = min hT , we have proven the left inequality
of (5.2). 2

Remark: Inequality (5.3) is sharp. To prove this, choose uα = 1.

Block-Jacobi preconditioners
Instead of choosing the diagonal, one can choose a block-diagonal of A, e.g.,
• In the case of systems of PDEs, choose blocks consisting of all degrees of freedom
sitting in one vertex. E.g., mechanical deformations (ux , uy , uz ).

• For high order elements, choose blocks consisting of all degrees of freedom associated
to the edges (faces, inner) of the elements.

• On anisotropic tensor product meshes, choose blocks consisting of unknowns in the

short direction

• Domain decomposition methods: Choose blocks consisting of the unknowns in a

sub-domain
Decompose the unknowns into M blocks, the block i has dimension Ni . Define the
rectangular embedding matrices

Ei ∈ RN ×Ni i = 1, . . . , M.

Ei consists of Ni unit vectors corresponding to the unknowns in the block i. Each u ∈ RN

can be uniquely written as
M
X
u= Ei ui with ui ∈ RNi
i=1
5.3. PRECONDITIONING 105

The diagonal blocks are

Ai = EiT AEi i = 1, . . . , M.
The block Jacobi preconditioner is
M
X
C −1 × d = Ei A−1 T
i Ei d
i=1

The quadratic form induced by C can be written as

X X
uT Cu = uTi Ai ui = kGEi ui k2A
i i
P
where u = Ei ui .
Example: Discretize the unit interval I = (0, 1) into n elements of approximate size
h ' 1/n. Split the unknowns into two blocks, the left half and the right half, and define
the corresponding block-Jacobi preconditioner.
Set
I = I1 ∪ Tn/2 ∪ I2 ,
with I1 = (0, xn/2 ), Tn/2 = [xn/2 , xn/2+1 ], and I2 = (xn/2+1 , 1). Decompose
u = E1 u1 + E2 u2 .
The corresponding finite element functions are ui = GEi ui . There holds

 Gu(x) x ∈ I1
u1 (x) = linear x∈T ,
0 x ∈ I2


and u2 vice versa. The quadratic form is

X X
uT Cu = ui Ai ui = kGEi ui k2A
i i

Evaluation gives
ku1 k2A = ku1 k2H 1 (I1 ) + ku1 k2H 1 (T )
' ku1 k2H 1 (I1 ) + h−1 |u(xn/2 )|2
 kuk2H 1 (I) + h−1 kuk2H 1 (I) (trace theorem)
' h−1 kuk2A ,
and thus X
uT Cu = kui k2A  h−1 kuk2A ' h−1 uT Au.
i
d
The situation is the same in R .
Exercise: Sub-divide the interval I into M sub-domains of approximative size H ≈ 1/M .
What are the sprectral bounds of the block-Jacobi preconditioner ?
106 CHAPTER 5. LINEAR EQUATION SOLVERS

Additive Schwarz preconditioners

The next generalization is an overlapping block Jacobi preconditioner. For i = 1, . . . , M
let Ei ∈ RN ×Ni be rectangular matrices such that each u ∈ RN can be (not necessarily
uniquely) written as
XM
u= Ei ui with ui ∈ RNi
i=1

Again, the overlapping block-Jacobi preconditioning action is

M
X
C −1
×d= Ei A−1 T
i Ei d
i=1

Example: Choose the unit-interval problem from above. The block 1 contains all nodes in
(0, 3/4), and the block 2 contains nodes in (1/4, 1). The blocks overlap, the decomposition
is not unique.
The columns of the matrices Ei are not necessarily unit-vectors, but are linearly in-
dependent. In this general setting, the preconditioner is called Additive Schwarz precon-
ditioner. The following lemma gives a useful representation of the quadratic form. It
was proven in similar forms by many authors (Nepomnyaschikh, Lions, Dryja+Widlund,
Zhang, Xu, Oswald, Griebel, ...) and is called also Lemma of many fathers, or Lions’
Lemma:
Lemma 93 (Additive Schwarz lemma). There holds
M
X
uT Cu = inf uTi Ai ui
ui ∈RNi
P
u= Ei ui i=1

Proof: The right hand side is a constrained minimization problem of a convex function.
The feasible set is non-empty, the CMP has a unique solution. It is solved by means of
Lagrange multipliers. Define the Lagrange-function (with Lagrange multipliers λ ∈ RN ):
X X
L((ui ), λ) = uTi Aui + λT (u − Ei ui ).

Its stationary point (a saddle point) is the solution of the CMP:

0 = ∇ui L((ui ), λ) = 2Ai ui + EiT λ

X
0 = ∇λ L((ui ), λ) = u − Ei ui

The first line gives

1
ui = A−1 E T λ.
2 i i
Use it in the second line to obtain
1X 1 −1
0=u− Ei A−1
i Ei λ = u − C λ,
2 2
5.3. PRECONDITIONING 107

i.e., λ = 2Cu, and

ui = A−1 T
i Ei Cu.
The minimal value is
X X
uTi Ai ui = uT CEi A−1 −1 T
i Ai Ai Ei Cu
X
= uT CEi A−1 T
i Ei Cu

= uT CC −1 Cu = uT Cu
2
Next, we rewrite the additive Schwarz iteration matrix
M
X
I − τC −1
A=I −τ Ei A−1 T
i Ei A
i=1

in the fe framework. Let

Vi = GEi RNi ⊂ Vh
be the sub-space corresponding to the range of Ei , and define the A-orthogonal projection
Pi : Vh → Vi : A(Pi u, vi ) = A(u, vi ) ∀ vi ∈ Vi
Lemma 94. Set u = Gu, the application of the iteration matrix is û = (I − τ C −1 A)u, and
set û = Gû. Then there holds
M
!
X
û = I − τ Pi u.
i=1

Proof: Let wi = A−1 T

i Ei Au. Then
û = u − τ GEi wi .
There holds wi := GEi wi ∈ Vi , and
A(GEi wi , GEi v i ) = v Ti EiT AEi wi
= v Ti Ai wi = v Ti EiT Au
= A(Gu, GEi v i ) ∀ vi ∈ RNi ,
i.e., wi = Pi u. 2
The additive Schwarz preconditioner is defined by the space splitting
M
X
V = Vi
i=1

P
If the spaces Vi are A-orthogonal, then i Pi = I, and (with τ = 1), and the iteration
matrix is M = 0.
The reformulation of the additive Schwarz lemma 93 in the finite element framework is
108 CHAPTER 5. LINEAR EQUATION SOLVERS

Lemma 95 (Additive Schwarz lemma). Let u = Gu. There holds

M
X
T
u Cu = inf kui k2A
∈Vi
uiP
u= ui i=1

Example: Let Z 1 Z
0 0
A(u, v) = uv +ε uv dx
0

with 0 ≤ ε  1. The bilinear-form is H 1 -elliptic and continuous, but the bounds depend
on the parameter ε. Let CJ be the Jacobi preconditioner. The proof of Theorem 92 shows
that
εh2 uT CJ u  uT Au  uT CJ u.
The non-robust lower bound is sharp: Take u = (1, . . . , 1)T .
The solution is to add the additional sub-space

V0 = span{1} = GE0 R1

to the AS preconditioner (with E0 ∈ RN ×1 consisting of 1-entries). The preconditioning

action is
C −1 × d = diag{A}−1 d + E0 (E0T AE0 )−1 E0T d.
The spectral bounds are robust in ε:

h2 uT Cu  uT Au  uT Cu,

namely
M
X
uT Cu = inf kui k2A
ui ∈Vi
u= M
P
i=0
0 ui
 
 M
X 
= inf ku0 k2A + inf kui k2A
u0 ∈V0  ui ∈Vi 
u−u0 = M
P
i=1
1 ui

 inf ku0 k2A + h−2 ku − u0 k2H 1

u0 ∈V0

The last stepR was the result of the Jacobi preconditioner applied to (u, v)H 1 . Finally, we
1
choose u0 = 0 u dx to obtain

uT Cu  ku0 k2A + h−2 ku − u0 k2H 1

 ε ku0 k2L2 + h−2 k∇(u − u0 )k2L2
 ε kuk2L2 + h−2 k∇uk2L2
= h−2 kuk2A
5.3. PRECONDITIONING 109

Overlapping domain decomposition preconditioning

Let Ω = ∪Mi=1 Ωi be a decomposition of Ω into M sub-domains of diameter H. Let Ωi be

e
such that
Ωi ⊂ Ω
ei e i \ ∂Ω, ∂Ωi }  H,
dist{∂ Ω

and only a small number of Ω e i are overlapping. Choose a finite element mesh of mesh
size h ≤ H, and the finite element space is Vh . The overlapping domain decomposition
preconditioner is the additive Schwarz preconditioner defined by the sub-space splitting
X
Vh = Vi with Vi = Vh ∩ H01 (Ω
e i ).

The bilinear-form A(., .) is H 1 -elliptic and continuous. The implementation takes the sub-
matrices of A with nodes inside the enlarged sub-domains Ω e i.

Lemma 96. The overlapping domain decomposition preconditioner fulfills the spectral es-
timates
H 2 uT Cu  uAu  uT Cu.

Proof: The upperP bound is generic. For the lower bound, we construct an explicit
decomposition u = ui .
There exists a partition of unity {ψi } such that

M
X
0 ≤ ψi ≤ 1, supp{ψi } ⊂ Ω
e i, ψi = 1
i=1

and
k∇ψi kL∞  H −1 .

Let Πh : L2 → Vh be a Clément-type quasi-interpolation operator such that Πh is a

projection on Vh , and

kΠh vkL2  kvkL2 , and k∇Πh vkL2  k∇vkL2 .

For given u ∈ Vh , we choose the decomposition

ui = Πh (ψi u).

Indeed ui ∈ Vi is a decomposition of u ∈ Vh :
X X X 
ui = Πh (ψi u) = Πh ( ψi )u = Πh u = u
110 CHAPTER 5. LINEAR EQUATION SOLVERS

The lower bound follows from

X
uT Cu = inf
P kvi k2A
u= vi
i
X X
≤ kui k2A  kui k2H 1
i i
X
= kΠh (ψi u)k2H 1
i
X
 kψi uk2H 1
i
Xn o
= kψi uk2L2 (Ωe i ) + k∇(ψi u)k2L2 (Ωe i )
i
Xn o
 kψi uk2L2 (Ωe i ) + k(∇ψi )uk2L2 (Ωe i ) + kψi ∇uk2L2 (Ωe i )
i
Xn o
2 −2 2 2
 kukL2 (Ωe i ) + H kukL2 (Ωe i ) + k∇ukL2 (Ωe i )
i
 kuk2L2 (Ω) + H −2 kuk2L2 (Ω) + k∇uk2L2 (Ω)
 H −2 kuk2A .
2

Overlapping DD preconditioning with coarse grid correction

The local DD preconditioner above gets worse, if the number of sub-domains increases. In
the limit, if H ' h, the DD preconditioner is comparable to the Jacobi preconditioner.
To overcome this degeneration, we add one more subspace. Let TH be a coarse mesh of
mesh-size H, and Th is the fine mesh generated by sub-division of TH . Let VH be the finite
element space on TH . The sub-domains of the domain decomposition are of the same size
as the coarse grid.
The sub-space decomposition is
M
X
Vh = VH + Vi .
i=1

Let GH : RNH → VH be the Galerkin isomorphism on the coarse grid, i.e.,

NH
X
GH uH = uH,i ϕH
i
i=1

The coarse space fulfills VH ⊂ Vh . Thus, every coarse grid basis ϕH

i can be written as
linear combination of fine grid basis functions ϕhj :
N
X
ϕH
i = EH,ji ϕhj .
j=1
5.3. PRECONDITIONING 111

Example:

VH Vh

The first basis function ϕH

1 is
1 h
ϕH h
1 = ϕ1 + ϕ2
2
The whole matrix is  
1
 1/2 1/2 
 
EH = 
 1 .

 1/2 1/2 
1
There holds
GH uH = Gh EH uH .
Proof:
NH
X Nh
NH X
X
GH uH = uH,i ϕH
i = uH,i EH,ji ϕhj
i=1 i=1 j=1
Nh
X
= ϕhj (EH uH )j = GEuH
j=1

The matrix EH transforms the coefficients uH w.r.t. the coarse grid basis to the coefficients
uh = EH uH w.r.t. the fine grid basis. It is called prolongation matrix.
The DD preconditioner with coarse grid correction is
X
C −1 × d = Ei A−1 T T −1 T
i Ei d + EH (EH AEH ) EH d
i

The first part is the local DD preconditioner from above. The second part is the coarse
T
grid correction step. The matrix EH (called restriction matrix) transfers the defect d from
the fine grid to a defect vector on the coarse grid. Then, the coarse grid problem with
T
matrix EH AEH is solved. Finally, the result is prolongated to the fine grid.
T
The matrix AH := EH AEH is the Galerkin matrix w.r.t. the coarse grid basis:

AH,ij = eTj EH
T
AEH ei = A(Gh EH ei , Gh EH ej )
= A(GH ei , GH ej ) = A(ϕH H
i , ϕj ).
112 CHAPTER 5. LINEAR EQUATION SOLVERS

Theorem 97. The overlapping domain decomposition preconditioner with coarse grid sys-
tem fulfills the optimal spectral estimates

uT Cu  uT Au  uT Cu.

Proof: The quadratic form generated by the preconditioner is

M
X
T
u Cu = inf kuH k2A + kui k2A .
uH ∈VH ,ui ∈Vi
i=1
P
u=uH + ui

Again, the upper bound uT Au  uT Cu follows from the finite overlap of the spaces
VH , V1 , . . . VM . To prove the lower bound, we come up with an explicit decomposition.
We split the minimization into two parts:
M
X
uT Cu = inf inf kuH k2A + kui k2A (5.4)
uH ∈VH ui ∈VP
i
u−uH = ui i=1

In the analysis of the DD precondition without coarse grid system we have observed that
M
X
inf kui k2A  H −2 ku − uH k2L2 + k∇(u − uH )k2L2
ui ∈VP
i
u−uH = ui i=1

Using this in (5.4) gives

uT Cu  kuH k2A + H −2 ku − uH k2L2 + k∇(u − uH )k2L2


inf
uH ∈VH

k∇uH k2L2 + H −2 ku − uH k2L2 + k∇uk2L2


 inf
uH ∈VH

To continue, we introduce a Clément operator ΠH : H 1 → VH being continuous in the

1
H -semi-norm, and approximating in L2 -norm:

k∇ΠH uk2L2 + H −2 ku − ΠH uk2L2  k∇uk2L2

Choosing now uH := ΠH u in the minimization problem we obtain the result:

uT Cu  k∇ΠH uk2A + H −2 ku − ΠH uk2L2 + k∇uk2L2

 k∇uk2 ' kuk2A

2
The inverse factor H −2 we have to pay for the local decomposition could be compensated
by the approximation on the coarse grid.

The costs for the setup depend on the underlying direct solver for the coarse grid
problem and the local problems. Let the factorization step have time complexity N α . Let
5.3. PRECONDITIONING 113

N be the number of unknowns at the fine grid, and M the number of sub-domains. Then
the costs to factor the coarse grid problem and the M local problems are of order
 α
α N
M +M
M

Equilibrating both terms gives the optimal choice of number of sub-domains

α
M = N 2α−1 ,

and the asymptotic costs

α2
N 2α−1 .
Example: A Cholesky factorization using bandwidth optimization for 2D problems has
time complexity N 2 . The optimal choice is M = N 2/3 , leading to the costs of

N 4/3 .

Multi-level preconditioners
The preconditioner above uses two grids, the fine one where the equations are solved,
and an artificial coarse grid. Instead of two grids, one can use a whole hierarchy of grids
T0 , T1 , . . . , TL = T . The according finite element spaces are

V0 ⊂ V1 ⊂ . . . ⊂ VL = Vh .

Let El be the prolongation matrix from level l to the finest level L. Define

Al = ElT AEl and Dl = diag{Al }.

Then, the multi-level preconditioner is

L
X
C −1
= E0 A−1 T
0 E0 + El Dl−1 ElT
l=1

The setup, and the application of the preconditioner takes O(N ) operations. One can
show that the multi-level preconditioner fulfills optimal spectral bounds

uT Cu  uT Au  uT Cu.

An iterative method with multi-level preconditioning solves the matrix equation Au = f

of size N with O(N ) operations !
114 CHAPTER 5. LINEAR EQUATION SOLVERS

5.4 Analysis of the multi-level preconditioner

We want to solve a finite element system on VL := Vh ⊂ H 1 . To define the multi-level
preconditioner C = CL , we use also finite element spaces on coarser meshes T0 , T1 , . . . TL :

V0 ⊂ V1 ⊂ . . . ⊂ VL

Assume hl ' 2−l . Let {ϕl,i : 1 ≤ i ≤ Nl } be the hat-basis for Vl , with Nl = dim Vl . Let Al
be the finite element matrix on Vl .
El ∈ RNl ×Nl−1 is the prolongation matrix from level l − 1 to level l.
The multi-level preconditioner is defined recursively:

C0−1 := A−1
0
Cl−1 := (diag Al )−1 + El Cl−1
−1 T
El 1 ≤ l ≤ L.

The computational complexity of one application of CL−1 is O(N ) operations.

(An extended version of) the Additive Schwarz Lemma allows to rewrite
Nl
X
kul k2Cl = infPN kul−1 k2Cl−1 + kul,i k2A
ul =ul−1 + l u
i=1 l,i i=1
ul−1 ∈Vl−1 ,ul,i ∈span{ϕl,i }
XX
= inf
P Nl
kul,i k2A + ku0 k2A
u=u0 + L
P
l=1 i=1 ul,i l i

Reordering the minimization we obtain

L
X Nl
X
kuk2CL = inf ku0 k2A + inf
P kul,i k2A
u= L ul ul = ul,i
P
l=0 l=1 i=1
ul ∈Vl

L
X
' inf ku0 k2A + h−2 2
l kul kL2
u= L ul
P
l=0 l=1
ul ∈Vl

Lemma 98 (simple analysis).

1
C  A  LC
L
Proof. A  LC followsPLfrom maximal overlap of spaces and the inverse estimate k∇ul kL2 
h−1
l ku k
l L2 . Let u = l=0 ul be an arbitrary decomposition:

L
X L
X L
X
ul k2A kul k2A ku0 k2A h−2 2


k ≤ (L + 1) L + l ku l k L2 .
l=0 l=0 l=1

Since the estimate holds for any decompositon, it also holds for the infimum.
5.4. ANALYSIS OF THE MULTI-LEVEL PRECONDITIONER 115

To show C  LA we come up with an explicit decomposition of u ∈ VL . Let Πl : L2 →

Vl be a Clément-type operator which is a projection and satisfies

kΠl ukH 1 + h−1

l ku − Πl ukL2  kukH 1 ∀ u ∈ H 1.

Define

u0 := Π0 u
ul := Πl u − Πl−1 u 1 ≤ l ≤ L.
PL
Then u = l=0 ul and
L
X
kuk2C  kΠ0 uk2A + h−2 2 2 2
l kΠl u − Πl−1 ukL2  L kukH 1 ≈ L kukA
l=1

We have bound each of the L + 1 terms by the H 1 -norm of u, thus the factor L.
Next we show an improved estimate leading to the optimal condition number
κ(C −1 A)  1, independent of the number of refinement levels:
Lemma 99. There holds
CAC
PL
Proof. We show A  C. Let u = l=0 ul an arbitrary decomposition. First, we split up
the coarsest level:
L
X
kuk2A ≤ ku0 k2A +k ul k2A
l=1

Next we show the estimate

|l−k|
A(ul , vk ) ≤ 2− 2 h−1 −1
l kul khk kvk k ∀ ul ∈ Vl , vk ∈ Vk

We assume l ≤ k. We perform integration by parts on the level-l triangles, and apply

Cauchy-Schwarz and scaling techniques:
XZ
A(ul , vk ) = ∇ul ∇vk
T ∈Tl T
XZ Z
∂ul
≤ −∆ul vk + vk
T T
| {z } ∂T ∂n
=0
X ∂ul
≤ kvk k∂Tl
T
∂n ∂Tl
−3/2 −1/2
≤ hl kul kL2 hk kvk kL2
p
= hk /hl h−1 kul kL2 h−1
k kvk kL2
| {z } l
'2−|k−l|/2
116 CHAPTER 5. LINEAR EQUATION SOLVERS

We define the overlap - matrix O ∈ RL×L as

Okl = 2−|k−l|/2 .

Then
L
X N
X X
k ul k2A = A(ul , uk )  Okl h−1 −1
k kuk kL2 hl kul kL2
l=1 l,k=1 l,k
L
X
≤ ρ(O) h−2 2
l kul kL2
l=1

The spectral radius ρ(O) can be estimated by the row-sum-norm, which is bounded by a
convergent geometric sequence
L ∞
X
−|k−l|/2
X √ −k 2
2 ≤2 2 ≤ √ .
k=1 k=0
1− 2

Since the decomposition was arbitrary, the estimate holds for the minimal decomposition.

Now we show C  A. We procede similar as above. Let Πl : L2 → Vl be an Clément-

type operator such that

kΠl ukL2 ≤ kukL2 ∀ u ∈ L2

ku − Πl ukL2  hl kukH 2 ∀ u ∈ H 2 .
2

We define u0 = Π0 u and ul = Πl u − Πl−1 u. We obtain the 2 estimates

h−2 2 −2 2
l kul kL2  hl kukL2 ,
h−2 2 2
l kul kL2  hl kukH 2 .

The idea of the proof is that H 1 is the interpolation space [L2 , H 2 ]1/2 . We define the
K-functional
K(t, u)2 = u=uinf+u ku0 k2L2 + t2 ku2 k2H 2 .

0 2
u0 ∈L2 ,u2 ∈H 2

Combining the 2 estimates above we get

h−2 2 −2 2 2
l kul kL2  hl K (hl , u)

Thus, the sum over L levels is

L
X L
X L
X
h−2 2
l kul kL2 ≤ h−2 2 2
l K (hl , u) ' 2l K 2 (2−l , u)
l=1 l=1 l=1
5.4. ANALYSIS OF THE MULTI-LEVEL PRECONDITIONER 117

Next we use that K(s, .) ' K(t, .) for t ≤ s ≤ 2t and replace the sum by an integral, and
substitute t := 2−l , dt ' −2−l dl = −tdl:
L
X Z L+1
h−2 2
l kul kL2 ≤ 2l K 2 (2−l , u)dl
l=1 l=1
Z 1
dt
' t−1 K 2 (t, u)
−L−1 t
Z2 ∞
dt
 t−1 K 2 (t, u)
0 t
= kuk2[L2 ,H 2 ]1/2 ' kuk2H 1

An intuitive explanation of the proof is that different terms of the sum Ll=1 h−2
P
l kΠl u −
2 1
Πl−1 ukL2 are dominated by different frequency components of u. The squared H -norm is
the sum over H 1 -norms of the individual frequency components.
118 CHAPTER 5. LINEAR EQUATION SOLVERS
Chapter 6

Mixed Methods

A mixed method is a variational formulation involving two function spaces, and a bilinear-
form of a special saddle point structure. Usually, it is obtained from variational problems
with constraints.

6.1 Weak formulation of Dirichlet boundary condi-

tions
We start with the Poisson problem

−∆u = f in Ω, (6.1)

and boundary conditions

u = uD on ΓD ,
∂u
∂n
= 0 on ΓN .
In contrast to the earlier method, we multiply equation (6.1) with test functions v ∈ H 1
(without imposing Dirichlet constraints), and integrate by parts. Using the Neumann
boundary conditions, we obtain
Z Z Z
∂u
∇u∇v dx − v ds = f v dx
Ω ΓD ∂n Ω

∂u
The normal derivative ∂n
is not known on ΓD . We simply call it −λ:

∂u
λ := −
∂n
To pose the Dirichlet boundary condition, we multiply u = uD by sufficiently many test
functions, and integrate over ΓD :
Z Z
uµ ds = uD µ ds ∀µ ∈ ?
ΓD ΓD

119
120 CHAPTER 6. MIXED METHODS

Combining both equations, we get the system of equations: Find u ∈ V = H 1 (Ω) and
λ ∈ Q =? such that
R R R
Ω
∇u · ∇v dx + ΓD vλ ds = f v dx ∀ v ∈ V,
R R (6.2)
ΓD
uµ ds = ΓD
uD µ ds ∀ µ ∈ Q.

A similar formulation can be obtained for interface conditions.

6.2 A Mixed method for the flux

We start from the second order pde

div(a∇u) = f in Ω,

and boundary conditions

u = uD on ΓD
∂u
a = g on ΓN
∂n
Next, we introduce the flux variable σ := a∇u to rewrite the equations as: Find u and σ
such that

a−1 σ − ∇u = 0, (6.3)
div σ = −f, (6.4)

and boundary conditions

u = uD on ΓD
σ·n = g on ΓN .

We want to derive a variational formulation for the system of equations. For that, we
multiply the first equations by vector-valued test functions τ , the second equation by test
functions v, and integrate:
R −1 R
Ω
(a σ) · τ dx − Ω τ · ∇u dx = 0 ∀τ
R R
Ω
div σ v dx = − f v dx ∀v

We would like to have the second term of the first equation of the same structure as the
first term in the second equation. This can be obtained by integration by parts applied to
either one of them. The interesting case is to integrate by parts in the first line to obtain:
Z Z Z Z
−1
(a σ) · τ dx + div τ u dx − τn u ds − τn u ds = 0.
Ω Ω ΓD ΓN
6.3. ABSTRACT THEORY 121

Here, we make use of the boundary conditions. On the Dirichlet boundary, we know
u = uD , and use that in the equation. The Neumann boundary condition σ · n = g must
be put into the approximation space, it becomes an essential boundary condition. Thus,
it is enough to choose test functions of the sub-space fulfilling τ · n = 0. The problem is
now the following. The space V will be fixed later. Find σ ∈ V, σn = g on ΓN , and u ∈ Q
such that
R −1 R R
Ω
R (a σ) · τ dx + Ω
div τ u dx = R uD τn ds
ΓD
∀ τ, τn = 0 on ΓN
Ω
div σ v dx = − f v dx ∀v

The derivatives are put onto the flux unknown σ (and its test function τ ). We don’t have
to derive the primal unknown u. This will give us better approximation for the fluxes than
for the scalar. That is one of the reasons to use this mixed method.

6.3 Abstract theory

A mixed variational formulation involves two Hilbert spaces V and Q, bilinear-forms

a(u, v) : V × V → R,
b(u, q) : V × Q → R,

and continuous linear-forms

f (v) : V → R,
g(q) : Q → R.

The problem is to find u ∈ V and p ∈ Q such that

a(u, v) + b(v, p) = f (v) ∀ v ∈ V,
(6.5)
b(u, q) = g(q) ∀ q ∈ Q.

The two examples from above are of this form.

Instead of considering this as a system of equations, one can look at the mixed method
as one variational problem on the product spaces V × Q. For this, simply add both lines,
and search for (u, p) ∈ V × Q such that

a(u, v) + b(u, q) + b(v, p) = f (v) + g(q) ∀ (v, q) ∈ V × Q.

Define the big bilinear-form B(., .) : (V × Q) × (V × Q) → R as

B((u, p), (v, q)) = a(u, v) + b(u, q) + b(v, p),

to write the whole system as single variational problem

Find (u, p) ∈ V × Q : B((u, p), (v, q)) = f (v) + g(q) ∀ (v, q) ∈ V × Q

122 CHAPTER 6. MIXED METHODS

By the Riesz-representation theorem, we can define operators:

A : V → V : u → Au : (Au, v)V = a(u, v) ∀ v ∈ V
B : V → Q : u → Bu : (Bu, q)Q = b(u, q) ∀ q ∈ Q
B ∗ : Q → V : p → B ∗ p : (B ∗ p, v)V = b(v, p) ∀ v ∈ V.
By means of these operators, we can write the mixed variational problem as operator
equation
Au + B ∗ p = JV f,
(6.6)
Bu = JQ g.
Here, we used the Riesz-isomorphisms JV : V ∗ → V and JQ : Q∗ → Q.

In the interesting examples, the operator B has a large kernel:

V0 := {v : Bv = 0}
Lemma 100. Assume that B ∗ Q is closed in V . Then there holds the V -orthogonal decom-
position
V = V0 + B ∗ Q
Proof: There holds
V0 = {v : Bv = 0}
= {v : (Bv, q)Q = 0 ∀ q ∈ Q}
= {v : (v, B ∗ q)V = 0 ∀ q ∈ Q}.
This means, V0 is the V -orthogonal complement to B ∗ Q. 2
Now, we will give conditions to ensure a unique solution of a mixed problem:
Theorem 101 (Brezzi’s theorem). Assume that a(., .) and b(., .) are continuous bilinear-
forms
a(u, v) ≤ α2 kukV kvkV ∀ u, v ∈ V, (6.7)
b(u, q) ≤ β2 kukV kqkQ ∀ u ∈ V, ∀ q ∈ Q. (6.8)
Assume there holds coercivity of a(., .) on the kernel,i.e.,
a(u, u) ≥ α1 kuk2V ∀ u ∈ V0 , (6.9)
and there holds the LBB (Ladyshenskaja-Babuška-Brezzi) condition
b(u, q)
sup ≥ β1 kqkQ ∀ q ∈ Q. (6.10)
u∈V kukV
Then, the mixed problem is uniquely solvable. The solution fulfills the stability estimate
kukV + kpkQ ≤ c{kf kV ∗ + kgkQ∗ },
with the constant c depending on α1 , α2 , β1 , β2 .
6.3. ABSTRACT THEORY 123

Proof: The big bilinear-form B(., .) is continuous

B((u, p), (v, q))  (kuk + kpk) (kvk + kqk).

We prove that it fulfills the inf − sup condition

B((u, p), (v, q))
inf sup ≥ β.
v,q u,p (kvkV + kqkQ )(kukV + kpkQ )

Then, we use Theorem 33 (by Babuška-Aziz) to conclude continuous solvability.

To prove the inf − sup-condition, we choose arbitrary v ∈ V and q ∈ Q. We will
construct u ∈ V and p ∈ Q such that

kukV + kpkQ  kvkV + kqkQ

and
B((u, p), (v, q)) = kvk2V + kqk2Q .
First, we use (6.10) to choose u1 ∈ V such that

b(u1 , q) = kqk2Q and ku1 kV ≤ 2β1−1 kqkQ .

Next, we solve a problem on the kernel:

Find u0 ∈ V0 : a(u0 , w0 ) = (v, w0 )V − a(u1 , w0 ) ∀ w0 ∈ V0

Due to assumption (6.9), the left hand side is a coercive bilinear-form on V0 . The right
hand side is a continuous linear-form. By Lax-Milgram, the problem has a unique solution
fulfilling
ku0 kV  kvkV + ku1 kV
We set
u = u0 + u1 .
By the Riesz-isomorphism, we define a z ∈ V such that

(z, w)V = (v, w)V − a(u, w) ∀w ∈ V

By construction, it fulfills z⊥V V0 . The LBB condition implies

b(v, p) (v, B ∗ p)V
kpkQ ≤ β1−1 sup = β1−1 sup = β1−1 kB ∗ pkV ,
v kvkV v kvkV
and thus B ∗ Q is closed, and z ∈ B ∗ Q. Take the p ∈ Q such that

z = B ∗ p.

It fulfills
kpkQ ≤ β1−1 kzkV  kvkV + kqkQ
124 CHAPTER 6. MIXED METHODS

Concluding, we have constructed u and p such that

kukV + kpkQ  kvkV + kqkQ ,

and

B((u, p), (v, q)) = a(u, v) + b(v, p) + b(u, q)

= a(u, v) + (z, v)V + b(u, q)
= a(u, v) + (v, v)V − a(u, v) + b(u, q)
= kvk2V + b(u1 , q)
= kvk2V + kqk2Q .

6.4 Analysis of the model problems

Now, we apply the abstract framework to the two model problems.

Weak formulation of Dirichlet boundary conditions

The problem is well posed for the spaces

V = H 1 (Ω) and Q = H −1/2 (ΓD )

Remember, H −1/2 (ΓD ) is the dual to H 1/2 (ΓD ). The later one is the trace space of H 1 (Ω),
the norm fulfills
kuD kH 1/2 (ΓD ) ' inf1 kwkH 1 (Ω) .
w∈H
tr w=uD

The bilinear-forms are

Z
a(u, v) = ∇u ∇v dx
Ω
b(u, λ) = hλ, tr uiH −1/2 ×H 1/2
R
To be precise, the integral ΓD λu dx is extended to the duality product hλ, ui. For regular
functions (λ ∈ L2 (ΓD )), we can write the L2 -inner product.
Theorem 102. The mixed problem (6.2) has a unique solution u ∈ H 1 (Ω) and λ ∈
H −1/2 (ΓD ).
Proof: The spaces V and Q, and the bilinear-forms a(., .) and b(., .) fulfill the assump-
tions of Theorem 101. The kernel space V0 is
Z
V0 = {u : uµ dx = 0 ∀ µ ∈ L2 (ΓD )} = {u : trΓD u = 0}
ΓD
6.4. ANALYSIS OF THE MODEL PROBLEMS 125

The continuity of a(., .) on V is clear. It is not coercive on V , but, due to Friedrichs

inequality, it is coercive on V0 .
The bilinear-form b(., .) is continuous on V × Q:

b(u, µ) = hµ, tr uiH −1/2 ×H 1/2 ≤ kµkH −1/2 k tr ukH 1/2 (ΓD )  kµkQ kukH 1 = kµkQ kukV

The LBB - condition of b(., .) follows more or less from the definition of norms:

hq, ui
kqkQ = sup
u∈H 1/2 kukH 1/2
hq, ui
' sup
u∈H 1/2 inf w∈H 1 (Ω) kwkH 1 (Ω)
tr w=u

hq, ui
= sup sup
u∈H 1/2 w∈H 1 (Ω) kwkH 1
tr w=u

hq, tr wi b(w, q)
= sup = sup
w∈H 1 kwkH 1 w∈V kwkV

Mixed method for the fluxes

This mixed method requires the function space H(div, Ω):

Definition 103. A measurable function g is called the weak divergence of σ on Ω ⊂ Rd if

there holds Z Z
g ϕ dx = − σ · ∇ϕ dx ∀ ϕ ∈ C0∞ (Ω)
Ω Ω

The function space H(div) is defined as

H(div, Ω) := {σ ∈ [L2 (Ω)]d : div σ ∈ L2 },

its norms is
1/2
kσkH(div) = kσk2L2 + k div σk2L2


The mixed method is formulated on the spaces

V = H(div) Q = L2

The bilinear-forms are

Z
a(σ, τ ) = a−1 στ dx ∀ σ, τ ∈ V
Z
b(σ, v) = div σ v dx ∀ σ ∈ V, ∀ v ∈ Q

We assume that the symmetric matrix a ∈ Rd×d and its inverse a−1 are bounded.
126 CHAPTER 6. MIXED METHODS

Theorem 104. The mixed problem for the fluxes is well posed.

Proof: We check the conditions of the theorem of Brezzi: The bilinear-forms are
bounded, namely
Z
a(σ, τ ) = a−1 στ dx ≤ ka−1 kL∞ kσkL2 kτ kL2  kσkV kτ kV

and Z
b(σ, v) = div σv dx ≤ k div σkL2 kvkL2 ≤ kσkV kvkQ .

The kernel space V0 = {τ : b(τ, v) = 0 ∀ v ∈ Q} is

V0 = {τ ∈ H(div) : div τ = 0}

There holds the kernel-ellipticity of a(., .). Let τ ∈ V0 . Then

Z Z
T −1 −1
a(τ, τ ) = τ a τ dx ≥ inf λmin (a ) |τ |2 dx  kτ k2L2 = kτ k2H(div)
x∈Ω

We are left to verify the LBB condition

R
div σ v dx
sup  kvkL2 ∀ v ∈ L2 . (6.11)
σ∈H(div) kσkH(div)

For given v ∈ L2 , we will construct a flux σ satisfying the inequality. For this, we solve
the artificial Poisson problem −∆ϕ = v with Dirichlet boundary conditions ϕ = 0 on ∂Ω.
The solution satisfies k∇ϕkL2  kvkL2 . Set σ = −∇ϕ. There holds div σ = v. Its norm is

kσk2H(div) = kσk2L2 + k div σk2L2 = k∇ϕk2L2 + kvk2L2  kvk2L2 .

Using it in (6.11), we get the result

R R 2
div σ v dx v dx
=  kvkL2 .
kσkH(div) kσkH(div)

The function space H(div)

The mixed formulation has motivated the definition of the function space H(div). Now,
we will study some properties of this space. We will also construct finite elements for
the approximation of functions in H(div). In Section 3.3.1, we have investigated traces of
functions in H 1 . Now, we apply similar techniques to the space H(div). Again, the proofs
are based on the density of smooth functions.
For a function in H 1 , the boundary values are well defined by the trace operator. For
a vector-valued function in H(div), only the normal-component is well defined on the
boundary:
6.4. ANALYSIS OF THE MODEL PROBLEMS 127

Theorem 105. There exists a normal-trace operator

trn : H(div) → H −1/2 (∂Ω)

such that for σ ∈ H(div) ∩ [C(Ω)]d it coincides with its normal component

trn σ = σ · n on ∂Ω.

Proof: For smooth functions, the trace operator gives the normal component on the
boundary. We have to verify that this operator is bounded as operator from H(div) to
H −1/2 (∂Ω). Then, by density, we can extend the trace operator to H(div). Let σ ∈
H(div) ∩ [C 1 (Ω)]d :
R R
∂Ω
σ · n ϕ ds σ · n tr ϕ ds
k trn σkH −1/2 = sup ' sup ∂Ω
ϕ∈H 1/2 (∂Ω) kϕkH 1/2 ϕ∈H 1 (Ω) kϕkH 1
R R
(σ tr ϕ) · n ds div(σϕ) dx
= sup ∂Ω = sup Ω
ϕ∈H 1 (Ω) kϕkH 1 ϕ∈H 1 (Ω) kϕkH 1
R R
(div σ)ϕ dx + Ω σ · ∇ϕ dx k div σkL2 kϕkL2 + kσkL2 k∇ϕkL2
= sup Ω ≤ sup
ϕ∈H 1 (Ω) kϕkH 1 ϕ∈H 1 (Ω) kϕkH 1
1/2
≤ kσk2L2 + k div σk2L2

= kσkH(div)

Lemma 106. There holds integration by parts

Z Z
σ · ∇ϕ dx + (div σ)ϕ dx = htrn σ, tr ϕiH −1/2 ×H 1/2
Ω Ω

for all σ ∈ H(div) and ϕ ∈ H 1 (Ω).

Proof: By density of smooth functions, and continuity of the trace operators.
Now, let Ω1 , . . . ΩM be a non-overlapping partitioning of Ω. In Section 3.3.1, we have
proven that functions which are in H 1 (Ωi ), and which are continuous across the boundaries
γij = Ωi ∩ Ωj , are in H 1 (Ω). A similar property holds for functions in H(div).
Theorem 107. Let σ ∈ [L2 (Ω)]d such that
•
σ|Ωi ∈ H(div, Ωi )

•
trn,i σ|Ωi = − trn,j σ|Ωj on γij .

Then σ ∈ H(div, Ω), and

(div σ)|Ωi = div (σ|Ωi ).
128 CHAPTER 6. MIXED METHODS

The proof follows the lines of Theorem 46.

We want to compute with functions in H(div). For this, we need finite elements for
this space. The characterization by sub-domains allows the definition of finite element
sub-spaces of H(div). Let T = {T } be a triangulation of Ω. One family of elements are
the BDM (Brezzi-Douglas-Marini) elements. The space is

Vh = {σ ∈ [L2 ]2 : σ|T ∈ [P k ]d , σ · n continuous across edges}.

This finite element space is larger than the piece-wise polynomial H 1 -finite element space
of the same order. The finite element functions can have non-continuous tangential com-
ponents across edges.

The cheapest element for H(div) is the lowest order Raviart-Thomas element RT0.
The finite element (T, VT , {ψi }) is defined by the space of shape functions VT , and linear
functionals ψi . The element space is
    
a x
VT = +c : a, b, c ∈ R ,
b y
the linear functionals are the integrals of the normal components on the three edges of the
triangle Z
ψi (σ) = σ · n ds i = 1, 2, 3
ei

The three functionals are linearly independent on VT . This means, for each choice of
σ1 , σ2 , σ3 , there exists three unique numbers a, b, c ∈ R such that
a  
x
σ= +c .
b y
satisfies ψi (σ) = σi .
Exercise: Compute the shape functions for the RT0 - reference triangle.

The global finite element functions are defined as follows. Given one value σi for each
edge ei of the triangulation. The corresponding RT0 finite element function σ is defined
by Z
σ|T ∈ VT and σ|T · nei ds = σi
ei

for all edges ei ⊂ T and all triangles T ∈ T .

We have to verify that this construction gives a function in H(div). For each ele-
ment, σ|T is a linear polynomial, and thus in H(div, T ). The normal components must be
continuous. By construction, there holds
Z Z
σ|T,i · n ds = σ|T,j · n ds
e e
6.4. ANALYSIS OF THE MODEL PROBLEMS 129

for the edge e = Ti ∩ Tj . The normal component is continuous since σ · ne is constant on

an edge: Points (x, y) on the edge e fulfill xnx + yny is constant. There holds
     
a x nx
σ · ne = +c · = anx + bny + c(xnx + yny ) = constant
b y ny

The global RT0-basis functions ϕRTi are associated to the edges, and satisfy
Z
ϕRT
j · ne ds = δij ∀ i, j = 1, . . . Nedges
ei

By this basis, we can define the RT - interpolation operator

X Z 
RT
Ih σ = σ · ne ds ϕRT
i
edges ei ei

It is a projection on Vh . The interpolation operator preserves the divergence in mean:

Lemma 108. The RT0 - interpolation operator satisfies
Z Z
div Ih σ dx = div σdx
T T
for all triangles T ∈ T .
Let Ph be the L2 projection onto piece-wise constant finite element functions. This
R Let Qh = {q
is: R ∈ L2 : q|T = const ∀T ∈ T }. Then Ph p is defined by Ph p ∈ Qh and
P
Ω h
p qh dx = Ω
P p qh dx ∀ qh ∈ Qh . This is equivalent to Ph p satisfies Ph p ∈ Qh and
Z Z
Ph p dx = p dx ∀T ∈ T .
T T
The Raviart-Thomas finite elements are piecewise linear. Thus, the divergence is piece-
wise constant. From div Ih σ ∈ Qh and Lemma 108 there follows
div Ih σ = Ph div σ.
This relation is known as commuting diagram property:
div
H(div) −→ L2
 

yIh

yPh (6.12)
div
VhRT −→ Qh

The analysis of the approximation error is based on the transformation to the reference
element. For H 1 finite elements, interpolation on the element T is equivalent to interpo-
lation on the reference element Tb, i.e., (Ih v) ◦ FT = Iˆh (v ◦ FT ). This is not true for the
RH(div) elements:
R The transformation F changes the direction of the normal vector. Thus
e
σ · n ds 6
= ê
σ̂ · n̂ ds.
The Piola transformation is the remedy:
130 CHAPTER 6. MIXED METHODS

Definition 109 (Piola Transformation). Let F : Tb → T be the mapping from the reference
element Tb to the element T . Let σ̂ ∈ L2 (Tb). Then, the Piola transformation

σ = P{σ̂}

is defined by
σ(F (x̂)) = (det F 0 )−1 F 0 σ̂(x̂).
The Piola transformation satisfies:
Lemma 110. Let σ̂ ∈ H(div, Tb), and σ = P{σ̂}. Then there holds

(div σ)(F (x̂)) = (det F 0 )−1 div σ̂

Let ê be an edge of the reference element, and e = F (ê). Then

Z Z
σ · n ds = σ̂ · n̂ ds
e ê

b ∈ C0∞ (Tb), and ϕ(F (x̂)) = ϕ(x).

Proof: Let ϕ b Then there holds
Z Z
div σ ϕ dx = σ · ∇ϕ dx
T ZT
(det F 0 )−1 F 0 σ · (F 0 )−T ∇ϕ̂ (det F 0 ) dx̂
   
=
ZT
b
Z
= σ̂∇ϕ̂ dx̂ = div σ̂ ϕ̂ dx̂
ZTb Tb

= (det F 0 )−1 (div σ̂)ϕ dx.

T

Since C0∞ is dense in L2 (T ), there follows the first claim. To prove the second one, we show
that Z Z
(σ · n)ϕ ds = (σ̂ · n)ϕ̂ dx
e ê
∞
holds for all ϕ ∈ C (T ), ϕ = 0 on ∂T \ e. Then, let ϕ → 1 on the edge e:
Z Z Z Z
(σ · n)ϕ ds = div(σϕ) dx = div(σ̂ ϕ̂) dx̂ = (σ̂ · n̂)ϕ̂ dŝ.
e T Tb ê

2
Lemma 111. The Raviart-Thomas triangle T and the Raviart-Thomas reference triangle
are interpolation equivalent:
IhRT P{σ̂} = P{IbhRT σ̂}

R Proof: The element spaces are equivalent, i.e., VT = P{VTb }, and the functionals ψi (σ) =
e
σ · n ds are preserved by the Piola transformation.
6.5. APPROXIMATION OF MIXED SYSTEMS 131

Theorem 112. The Raviart-Thomas interpolation operator satisfies the approximation

properties

kσ − IhRT σkL2 (Ω)  h k∇σkL2 (Ω)

k div σ − div IhRT σkL2 (Ω)  h k∇ div σkL2 (Ω)

Proof: Transformation to the reference element, using that the interpolation preserves
constant polynomials, and the Bramble Hilbert lemma. The estimate for the divergence
uses the commuting diagram property

k div(I − IhRT )σkL2 = k(I − Ph ) div σkL2  h k∇ div σkL2

6.5 Approximation of mixed systems

We apply a Galerkin-approximation for the mixed system. For this, we choose (finite
element) sub-spaces Vh ⊂ V and Qh ⊂ Q, and define the Galerkin approximation (uh , ph ) ∈
Vh × Qh by

B((uh , ph ), (vh , qh )) = f (vh ) + g(qh ) ∀ vh ∈ Vh ∀ qh ∈ Qh .

Theorem 113. Assume that the finite element spaces fulfill the discrete stability condition
B((u, p), (v, q))
inf sup ≥ β. (6.13)
v∈Vh ,q∈Qh u∈Vh ,p∈Qh (kvkV + kqkQ )(kukV + kpkQ )

Then the discretization error is bounded by the best-approximation error

ku − uh kV + kp − ph kQ  inf {ku − vh kV + kp − qh kQ }
vh ∈Vh ,qh ∈Qh

Proof: Theorem 36 applied to the big system B((u, p), (v, q)). 2

The stability on the continuous level V × Q does not imply the discrete stability !
Usually, one checks the conditions of Brezzi on the discrete level to prove stability of B(., .)
on the discrete levels. The continuity of a(., .) and b(., .) are inherited from the continuous
levels. The stability conditions have to be checked separately. The discrete kernel ellipticity

a(vh , vh )  kvh k2V ∀ vh ∈ V0h = {vh ∈ Vh : b(vh , qh ) = 0 ∀ qh ∈ Qh }, (6.14)

and the discrete LBB condition

b(uh , qh )
sup  kqh kQ ∀ qh ∈ Qh . (6.15)
uh ∈Vh kuh kV
132 CHAPTER 6. MIXED METHODS

The discrete LBB condition is posed for less dual variables qh in Qh ⊂ Q, but the space in
the supremum is also smaller. It does not follow from the LBB condition on the continuous
levels.
There is a canonical technique to derive the discrete LBB condition from the continuous
one:

Lemma 114. Assume there exists a quasi-interpolation operator

Πh : V → Vh

which is continuous
kΠh vkV  kvkV ∀ v ∈ V,
and which satisfies
b(Πh v, qh ) = b(v, qh ) ∀ qh ∈ Qh .
Then, the continuous LBB condition implies the discrete one.

Proof: For all ph ∈ Qh there holds

b(vh , ph ) b(Πh v, ph ) b(v, ph )

sup ≥ sup  sup  kph kQ
vh ∈Vh kvh kV v∈V kΠh vkV v∈V kvkV

Approximation of the mixed method for the flux

Choose the pair of finite element spaces, the Raviart Thomas spaces

Vh = {v ∈ H(div) : v|T ∈ VTRT } ⊂ V = H(div)

and the space of piece-wise constants

Qh = {q ∈ L2 : q|T ∈ P 0 } ⊂ Q = L2 .

Pose the discrete mixed problem: Find (σh , uh ) ∈ Vh × Qh such that

R −1 R R
RΩ (a σh ) · τ h dx + Ω
div τh uh dx = R uD τn ds
ΓD
∀ τh ∈ Vh
(6.16)
Ω
div σ v
h h dx = − Ω
f vh dx ∀ vh ∈ Qh .

Lemma 115 (Discrete Stability). The discrete mixed variational problem (6.16) is well
posed.

Proof: By Brezzi’s theorem. Continuity of the bilinear-form and the linear-form follow
from the continuous level. We prove the kernel ellipticity: Since

div Vh ⊂ Qh ,
6.5. APPROXIMATION OF MIXED SYSTEMS 133

there holds Z
div σh qh dx = 0 ∀ qh ∈ Qh ⇒ div σh = 0,

and thus V0h ⊂ V0 . In this special case, the discrete kernel ellipticity is simple the restriction
of the continuous one to V0h . We are left with the discrete LBB condition. We would like
to apply Lemma 114. The quasi-interpolation operator is the Raviar-Thomas interpolation
operator IhRT . The abstract condition
b(IhRT σ, vh ) = b(σ, vh ) vh ∈ Qh
reads as Z Z
div IhRT σ dx = div σdx,
T T
which was proven in Lemma 108. But, the interpolation operator is not continuous on
H(div). The edge-integrals are not well defined on H(div). We have to include the sub-
space [H 1 ]d ⊂ H(div). There holds
kIhRT σkH(div)  kσkH 1 ∀ σ ∈ [H 1 ]d ,
and the stronger LBB condition (see Section on Stokes below)
(div σ, v)L2
sup ≥ βkvkL2 ∀ v ∈ L2 .
σ∈[H 1 ]d kσkH 1

We follow the proof of Lemma 114: For all vh ∈ Qh there holds

b(σh , vh ) (div IhRT σ, vh ) (div σ, vh )
sup ≥ sup RT
 sup  kvh kL2 .
σh ∈Vh kσh kV σ∈[H 1 ]d kIh σkV σ∈[H 1 ]d kσkH 1

Brezzi’s theorem now proves that the discrete problem is well posed, i.e., it fulfills the
discrete inf-sup condition.
2
Theorem 116 (A priori estimate). The mixed finite element method for the flux satisfies
the error estimates
kσ − σh kL2 + k div(σ − σh )kL2 + ku − uh kL2  h (kσkH 1 + kukH 1 + kf kH 1 ) (6.17)
Proof: By discrete stability, one can bound the discretization error by the best approx-
imation error

kσ − σh kH(div) + ku − uh kL2  inf kσ − τh kH(div) + ku − vh kL2 .
τh ∈Vh
vh ∈Qh

The best approximation error is bounded by the interpolation error. The first term is
(using the commuting diagram property and div σ = f )
inf {kσ − τh kL2 + k div(σ − τh )kL2 } ≤ kσ−IhRT σkL2 +k(I−P 0 ) div σkL2  h (kσkH 1 + kf kH 1 ) .
τh ∈Vh
134 CHAPTER 6. MIXED METHODS

The second term is

inf ku − vh kL2 ≤ ku − P 0 ukL2  h kukH 1 .

vh ∈Qh

The smoothness requirements onto the solution of (6.17) are fulfilled for problems on
convex domains, and smooth (constant) coefficients a. There holds kukH 2  kf kL2 . Since
σ = a∇u, there follows kσkH 1  kf kL2 . The mixed method requires more smoothness
onto the right hand side data, f ∈ H 1 . It can be reduced to H 1 on sub-domains, what is
a realistic assumption. On non-convex domains, u is in general not in H 2 (and σ not in
H 1 ). Again, weighted Sobolev spaces can be used to prove similar estimates on properly
refined meshes.

Approximation of the mixed method for Dirichlet boundary conditions

A possibility is to choose continuous and piece-wise linear finite element spaces on the
domain and on the boundary

Vh = {v ∈ C(Ω) : v|T ∈ P 1 ∀ T },

Qh = {µ ∈ C(∂Ω) : µ|E ∈ P 1 ∀ E ⊂ ∂Ω}.

Theorem 117. The discrete mixed method is well posed.

Proof: Exercises.

6.6 Supplement on mixed methods for the flux : dis-

crete norms, super-convergence and implementa-
tion techniques
6.6.1 Primal and dual mixed formulations
A mixed method for the flux can be posed either in the so called primal form: find σ ∈
V = [L2 ]2 , u ∈ H 1 with u = uD on ΓD such that
R −1 R
Ω
(a σ) · τ dx − Ω τ · ∇u dx = 0 ∀ τ,
R R R
− Ω σ · ∇v dx = − f v dx − ΓN gv ds ∀ v, v = 0 on ΓD ,

or in the so called dual mixed form: find σ ∈ V = H(div), u ∈ L2 with σ · n = g on ΓN

R −1 R R
Ω
(a σ) · τ dx + Ω div τ u dx = ΓD uD τn ds ∀ τ, τn = 0 on ΓN
R R
Ω
div σ v dx = − f v dx ∀ v.
6.6. SUPPLEMENT ON MIXED METHODS FOR THE FLUX : DISCRETE NORMS, SUPER-CONVER

Both are formally equivalent: If the solutions are smooth enough for integration by parts,
both solutions are the same. In both cases, the big-B bilinear-form is inf − sup stable with
respect to the corresponding norms.
The natural discretization for the primal-mixed formulation uses standard H 1 -finite
elements of order k for u, and discontinuous L2 elements of order k − 1 for σ. Here,
the discrete Brezzi conditions are trivial. The dual one requires Raviart-Thomas (RT)
or Brezzi-Douglas-Marini (BDM) elements
R for σ,
R and L2 elements for u. This pairing
delivers locally exact conservation ( ∂T σn = − T f ). In particular this property makes
the method interesting by itself, but often this scheme is a part of a more complex problem
(e.g. Navier-Stokes equations).
Our plan is as follows: We want to use the dual finite element method, but analyze
it in a primal - like setting. Since Qh is no sub-space of H 1 , we have to use a discrete
counterpart of the H 1 -norm:

kτ k2Vh := kτ k2L2
X X X
kvk2Qh := kvk2H 1 ,h := k∇vk2L2 (T ) + 1
h
k[v]k2L2 (E) + 1
h
kvk2L2 (E)
T E⊂Ω E⊂ΓD

The factor h1 provides correct scaling: If we transform an element patch to the reference
patch, the jump term scales like the H 1 -semi-norm. This norm is called discrete H 1 -norm,
or DG-norm (as it is essential for Discontinuous Galerkin methods discussed later).
There holds a discrete Friedrichs inequality

kvkL2  kvkH 1 ,h .

Theorem 118. The dual-mixed discrete problem satisfies Brezzi’s conditions with respect
to L2 and discrete H 1 -norms.

Proof. The a(., .) bilinear-form is continuous and coercive on (Vh , k · kL2 ). Now we show
continuity of b(., .) on the finite element spaces. We integrate by parts on the elements,
and rearrange boundary terms:
Z XZ
b(σh , vh ) = div σh vh = div σh vh
Ω T
X Z ZT
= − σh · ∇vh + σh · n vh
T T ∂T
X Z XZ X Z X Z
= − σh · ∇vh + σh nE [v] + σh nE v + σh nE v
T E E E
| {z }
T E⊂Ω E⊂ΓD E⊂ΓN =0

The jump term is defined as [v](x) = limt→0+ v(x + tnE ) − v(x − tnE ). Thus, σh nE [v] is
independent of the direction of the normal vector. Next we apply Cauchy-Schwarz, and
136 CHAPTER 6. MIXED METHODS

use that hkσh · nk2L2 (E)  kσh kL2 (T ) for some E ⊂ T (scaling and equivalence of norms on
finite dimensional spaces):
X X X
b(σh , vh ) ≤ kσh kL2 (T ) k∇vh kL2 (T ) + h1/2 kσh kL2 (E) h−1/2 k[vh ]kL2 (E) + ...
T E⊂Ω E⊂ΓD
 kσh kL2 (Ω) kvh kH 1 ,h
The linear-forms are continuous with norms h−1/2 kuD kL2 (ΓD ) and kf kL2 (Ω) , respectively.
Finally, we show the LBB - condition: Given an vh ∈ Qh , we define σh as follows:
σh · nE = h1 [vh ] on E ⊂ Ω
1
σh · nE = h vh on E ⊂ ΓD
Zσh · nE = 0 Z on E ⊂ ΓN
σh · q = − ∇vh · q ∀ q ∈ [P k−1 ]2 .
T T

This definition mimics σ = −∇v. This construction is allowed by the definition of Raviart-
Thomas finite elements. Thus we get
X Z X X
2
b(σh , vh ) = − σh ∇vh + h
1
k[vh ]kL 2 (E) + 1
h
kvh k2L2 (E)
|{z}
T ∈[P k−1 ]2 E⊂Ω E⊂ΓD
XZ X X
= ∇vh · ∇vh + 1
h
k[vh ]k2L2 (E) + 1
h
kvh kL2 (E)
T E⊂Ω E⊂ΓD

= kvh k2H 1 ,h
By scaling arguments we see that kσh kL2  kvh kH 1 ,h . Thus we got σh such that
b(σh , vh )
 kvh kH 1 ,h ,
kσh kL2
and we have constructed the candidate for the LBB condition.

6.6.2 Super-convergence of the scalar

Typically, the discretization error of mixed methods depend on best-approximation errors
in all variables:
kσ − σh kL2 + ku − uh kH 1 ,h  inf kσ − τh kL2 + ku − vh kH 1 ,h
τh ,vh

By the usual Bramble-Hilbert and scaling arguments we see that (using the element-wise
L2 -projection Ph ):
ku − Ph ukH 1 ,h  hk kukH 1+k k ≥ 0.
In the lowest order case (k = 0) we don’t get any convergence !!
But, we can show error estimates for σ in terms of approximability for σ only. Further-
more, we can perform a local postprocessing to improve also the scalar variable.
6.6. SUPPLEMENT ON MIXED METHODS FOR THE FLUX : DISCRETE NORMS, SUPER-CONVER

Theorem 119. There holds

kσ − σh kL2 + kPh u − uh k  kσ − Ih σkL2 ,
where Ih is the canonical RT interpolation operator satisfying the commuting diagram prop-
erty div Ih = Ph div.
Proof. As usual for a priori estimates, we apply stability of the discrete problem, and use
the Galerkin orthogonality:
B((Ih σ − σh , Ph u − uh ), (τh , vh ))
kIh σ − σh kL2 + kPh u − uh kH 1 ,h  sup (6.18)
τh ,vh kτh kL2 + kvh kH 1 ,h
B((Ih σ − σ, Ph u − u), (τh , vh ))
= sup (6.19)
τh ,vh kτh kL2 + kvh kH 1 ,h
R −1
RNow we elaborate Ron the terms of B((Ih σ − σ, Ph u − u), (τh , vh )) = a (Ih σ − σ) · τh +
div(Ih σ − σ)vh + div τh (Ph u − u): For the first one we use Cauchy-Schwarz, and bounds
for the coefficient a:
Z
a−1 (Ih σ − σ) · τh  kσ − Ih σkL2 kτh kL2

For the second one we use the commuting diagram, and orthogonality:
Z Z
div(Ih σ − σ)vh = (Ph − Id) div σ vh = 0
| {z } |{z}
∈Q⊥ ∈Qh
h

For the third one we use that div Vh ⊂ Qh :

Z
div τh (Ph u − u)
| {z } | {z }
∈Qh Q⊥
h

Thus, the right hand side of equation (6.19) can be estimated by kσ − Ih σkL2 . Finally, an
application of the triangle inequality proves the result.
Remark 120. This technique applies for BDM elements as well as for RT. Both satisfy
div Vh = Qh , and the commuting diagram. For RTk elements, i.e. [P k ]2 ⊂ RTk ⊂ [P k+1 ]2
as well as BDMk elements, i.e. BDMk = [P k ]d we get the error estimate
kσ − Ih σkL2  hk+1 kσkH k
with k ≥ 0 for RT and k ≥ 1 for BDM.
Remark 121. The scalar variable shows super-convergence: A filtered error, i.e. Ph u − uh
is of higer order than the error u − uh itself: One order for RT and two orders for BDM
We can apply a local post-processing to compute the scalar part with higher accuracy:
We use the equation σ = a∇u, and the good error estimates for σ and Ph u. We set
eh := P k+1 , and solve a local problem on every element:
Q
min ka∇ṽh − σh k2L2 (T ),a−1
ṽh ∈Q
Rh
e
R
T vh = T ṽh
138 CHAPTER 6. MIXED METHODS

6.6.3 Solution methods for the linear system

The finite element discretization leads to the linear system for the coefficient vectors (called
σ and u again):     
A Bt σ 0
=
B 0 u −f
This matrix is indefinite, it has dim Vh positive and dim Qh negative eigenvalues. This
causes difficulties for the linear equation solver.
The first possibility is a direct solver, which must (in contrast to positive definite
systems) apply Pivot strategies.
A second possibility is block-elimination: eliminate σ from the first equation. The
regularity of A follows from L2 -coercivity of a(., .):

σ = −A−1 B t u

and insert it into the second equation:

−BA−1 B t u = −f

Thanks to the LBB-condition, B has full rank, and thus the Schur complement matrix is
regular. Since B is the discretization of the div-operator, B t of the negative gradient, and
A of a(x)−1 I, the equation can be interpreted as a discretization of

div a∇u = −f

This approach is not feasible, since A−1 is not a sparse matrix anymore.
One can use extensions of the conjugate gradient (CG) method for symmetric but
indefinit matrices (e.g. MINRES). Here, preconditioners are important. Typically, for
block-systems one uses block-diagonal preconditioners to rewrite the system as
! !
e−1 0 e−1 0
T
  
G V A B σ GV 0
e−1 = e−1
0 G Q
B 0 u 0 G Q
−f

where G
eV and G
eQ are approximations to the Gramien matrices in Vh and Qh :

GV,ij = (ϕσi , ϕσj )V and GQ,ij = (ϕui , ϕuj )Q

One can either choose the H(div)-L2 , or the [L2 ]2 -H 1 setting, which leads to different kind
of preconditioners. Here, the later one is much simpler. This is a good motivation for
considering the alternative framework.

6.6.4 Hybridization
Hybridization is a technique to derive a new variational formulation which obtains the same
solution, but its system matrix is positive definit. For this, we break the normal-continuity
6.6. SUPPLEMENT ON MIXED METHODS FOR THE FLUX : DISCRETE NORMS, SUPER-CONVER

of the flux functions, and re-inforce it via extra equations. We obtain new variables living
on the element-edges (or faces in 3D).
We start from the first equation a−1 σ − ∇u = 0, multiply with element-wise discontin-
uous test-functions τ , and integrate by parts on the individual elements:
Z XZ Z
−1
a στ + u div τ − u τn = 0
Ω T T ∂T

We now introduce the new unknown variable û which is indeed the restriction of u onto
the mesh skeletonQu|∪E .
We set V := T H(div, T ) and Q := L2 (Ω) × E H 1/2 (E), and pose the so called
Q
hybrid problem: find σ ∈ V and (u, û) ∈ Q such that
R −1 P R P R
Ω
(a σ) · τ dx + T T div τ u dx + T ∂T û τn = 0 ∀τ ∈ V
P R R
T T div σ v dx = − f v dx ∀ v ∈ L2 (Ω)
P R
∀ v̂ ∈ E H 1/2 (E).
Q
T ∂T v̂ σn = 0

The last equation can be rearranged edge by edge:

XZ X Z Y
[σn ] v̂ + σn v̂ = 0 ∀ v̂ ∈ H 1/2 (E),
E⊂Ω E E⊂∂Ω E E

which implies normal-continuity of σ. Dirichlet/Neumann boundary conditions are posed

now for the skeleton variable û.
This system is discretized by discontinuous RT /BDM elements for σ, piecewise poly-
nomials on elements T for u, and piecewise polynomials on edges for û such that the order
matches with σ · n.

P system is well-posed with respect to the norms kσkV := kσkL2 and

1. This discrete
ku, ûk2 = T k∇uk2L2 (T ) + h1 ku − ûk2∂T . Similar proof as for Theorem 118.

2. The components σh and uh of the solution of the hybrid problem correspond to the
solution of the mixed method.

3. Since the σh is discontinuous across elements, the arising matrix A is block-diagonal.

Now it is cheap to form the Schur complement
u f 
−1 T
−BA B =
û 0

This is now a system with a positive definite matrix for unknowns in the elements and
on the edges. Since the matrix block for the element-variables is still block-diagonal,
they can be locally eliminated, and only the skeleton variables are remaining. In the
lowest order case, the matrix is the same as for the non-conforming P 1 element.
140 CHAPTER 6. MIXED METHODS
Chapter 7

Discontinuous Galerkin Methods

Discontinuous Galerkin (DG) methods approximate the solution with piecewise functions
(polynomials), which are discontinuous across element interfaces. Advantages are

• block-diagonal mass matrices which allow cheap explicit time-stepping

• upwind techniques for dominant convection

• coupling of non-matching meshes

• more flexibility for stable mixed methods

DG methods require more unknowns, and also have a denser stiffness matrix. The last
disadvantage can be overcome by hybrid DG methods (HDG).

7.1 Transport equation

We consider the first order equation

div(bu) = f on Ω,

where b is the given wind, and f is the given source. Boundary conditions are specified

u = uD on Γin ,

where the inflow boundary is

Γin = {x ∈ ∂Ω : b · n < 0},

and the outflow boundary Γout = ∂Ω \ Γin .

The instationary transport equation
∂u
+ div(bu) = f on Ω × (0, T )
∂t

141
142 CHAPTER 7. DISCONTINUOUS GALERKIN METHODS

with initial conditions u = u0 for t = 0 can be considered as stationary transport equation

in space-time:
divx,t (b̃u) = f,
where b̃ = (b, 1). The inflow boundary consists now of the lateral boundary Γin ×(0, T ) and
the bottom boundary Ω × {0}, which is an inflow boundary according to (b, 1) · (0, −1) < 0.
The equation in conservative form leads to a conservation principle. Let V be an
arbitrary control volume. From the Gauß theorem we get
Z Z
b · nu = f
∂V V

The total outflow is in balance with the production inside V .

For stability, we assume div b = 0. This is a realistic assumption, since the wind is
often the solution of the incompressible Navier Stokes equation.
A variational formulation is
Z Z
div(bu)v = f v ∀v

If we set v = u, and use

div(bu)u = 12 div(bu2 )
We obtain Z Z Z
1 2
div(bu)u = 2
bn u = fu
∂Ω
For f = 0 we obtain Z Z
2
|bn |u = |bn |u2 .
Γout Γin

This inflow-outflow isometry is a stability argument. For time dependent problems (with
bn = 0 on ∂Ω), it ensures the conservation of L2 -norm in time.

7.1.1 Solvability
We assume b ∈ L∞ with div b = 0. We consider the problem: find u ∈ V , u = uD on Γin
and
B(u, v) = f (v) ∀v ∈ W
with Z Z
B(u, v) = div(bu)v and f (v) = fv

The space V shall be defined by the (semi)-norm

kukV = kb∇ukL2

Depending on b, it is a norm for {u = 0 on Γin }. Roughly speaking, if every point in Ω

can be reached by a finite trajectory along b, then kukL2  kukV , by Friedrichs’ inequality.
7.2. DISCONTINUOUS GALERKIN DISCRETIZATION 143

It does not hold if b has vortices. Then k · kV is only a semi-norm. Note, for space-time
problems b̃ cannot have vortices. For theory, we will assume that

kukL2  kukV ∀ u = 0 on Γin

The test space is

W = L2
The forms B(., .) and f (.) are continuous. The inf − sup conditions is trivial:
R
B(u, v) (b∇u)2
sup ≥ = kb∇ukL2
v kvkL2 |{z} kb∇uk
v:=b∇u

7.2 Discontinuous Galerkin Discretization

A DG method is a combination of finite volume methods and finite element methods. We
start with a triangulation {T }. On every element we multiply the equation by a test-
function: Z Z
b∇uv = fv ∀v
T T
We integrate by parts: Z Z Z
− bu∇v + bn uv = fv
T ∂T T
On the element-boundary we replace bn u by its up-wind limit bn uup . On the element inflow
boundary ∂Tin = {x ∈ ∂T : bnT < 0}, the upwind value is the value from the neighbour
element, while on the element outflow boundary it is the value from the current element T .
For elements on the domain inflow boundary, the upwind value is taken as the boundary
value uD . For the continuous solution there holds
Z Z Z
up
− bu∇v + bn u v = fv
T ∂T T

Now we integrate back:

Z Z Z
up
b∇uv + bn (u − u)v = fv
T ∂T T

On the outflow boundary, the boundary integral cancels out, on the inflow boundary we
can write it as a jump term [u] = uup − u:
Z Z Z
b∇uv + bn [u]v = fv
T ∂Tin T

We define the DG bilinear-form

X Z Z 
DG
B (u, v) = b∇uv + bn [u]v .
T T ∂Tin
144 CHAPTER 7. DISCONTINUOUS GALERKIN METHODS

The true solution is consistent with

B DG (u, v) = f (v) ∀ v piece-wise continuous.

We define DG finite element spaces:

Vh := Wh := {v ∈ L2 : v|T ∈ P k }

The DG formulation is: find uh ∈ Vh such that

B DG (uh , vh ) = f (vh ) ∀ vh ∈ Wh

For the discontinuous space, the jump-term is important. If we use continous spaces, the
jump-term disappears. The discrete norms are defined as
X X
kuh k2Vh := kb∇uk2L2 (T ) + 1
h
kbn [u]k2L2 (E)
T E
kvh kWh = kvh kL2

The part with the jump-term mimics the derivative as kind of finite difference term across
edges.
We prove solvability of the discrete problem by showing a discrete inf − sup condition.
But, in general, one order in h is lost due to a mesh-dependent inf − sup constant. This
factor shows up in the general error estimate by consistency and stability. It can be avoided
in 1D, and on special meshes.
Theorem 122. There holds the discrete inf − sup condition
B(uh , vh )
sup  h kuh kVh
vh kvh kWh

Proof. We take two different test-functions: v1 = uh and v2 := b · ∇T u, and combine them

properly. The second test-function would not be possible in the standard C 0 finite element
space.
There holds (dropping sub-scripts h):
XZ Z
B(uh , v1 ) = B(uh , uh ) = b∇u u + bn [u]u
T T ∂Tin
X1 Z Z
= bn u 2 + bn [u]u
T
2 ∂T ∂Tin

We reorder the terms edge-by-edge. On the edge E we get contributions from two
elements: On the inflow-boundary of the down-wind element we get
Z Z Z
1 d 2 u d d 1 d 2 u d


2
b n (u ) + b n (u − u )u = |b n | 2
(u ) − u u
E E E
7.2. DISCONTINUOUS GALERKIN DISCRETIZATION 145

We used that b · n is negative on the inflow boundary. From the up-wind element we get
on its outflow-boundary: Z
1
|b | (uu )2
2 n
E
Summing up, we have the square
Z
1
|b |(uu
2 n
− ud )2 ,
E

and summing over elements we get the non-negative term

XZ
1
B(uh , uh ) = 2 |bn |[u]2 .
E E

An extra treatment of edges on the whole domain boundary gives that the jump must be
replaced by the function values on ∂Ω.
We plug in the second test function v2 :
XZ Z
2
B(uh , v2 ) = (b∇u) + bn [u] b∇u
T T ∂Tin

We use Young’s inequality to bound the second term from below:

XZ 1 γ
B(uh , v2 ) ≥ (b∇u)2 − kbn [u]k2∂L2 (Tin ) − kb∇uk2L2 (∂Tin )
T T 2γ 2

By the choice γ ' h and a inverse trace inequality (which needs smoothness assumptions
onto b) we can bound the last term by the first one on the right hand. Thus
XZ X1
B(uh , v2 )  (b∇u)2 − k[u]k2L2 (E),bn
T T E
h

Finally, we set
1
vh = v1 + v2
h
to obtain XZ X1
B(uh , vh )  (b∇u)2 + k[u]k2L2 (E),bn ' kuh k2Vh .
T T E
h
But, for this choice we get
kvh kWh  h−1 kuh kVh ,
and thus the h-dependent inf − sup-constant.
146 CHAPTER 7. DISCONTINUOUS GALERKIN METHODS

7.3 Nitsche’s method for Dirichlet boundary condi-

tions
We build in Dirichlet b.c. in a weak sense. In constrast to a mixed method, we obtain a
positive definit matrix.
We consider the equation

−∆u = f and u = uD on ∂Ω.

A diffusion coefficient, or mixed boundary conditions are possible as well. We multiply

with testfunctions, integrate and integrate by parts:
Z Z Z
∇u∇v − ∂n uv = f v ∀v
Ω

We do not restrict test functions to v = 0. To obtain a symmetric bilinear-form, we add a

consistent term
Z Z Z Z Z
∇u∇v − ∂n uv − ∂n vu = f v − ∂n vuD ∀v
Ω ∂Ω ∂Ω ∂Ω

Finally, to obtain stability (as proven below), we add the so called stabilization term
Z Z Z Z Z Z Z
α α
∇u∇v − ∂n uv − ∂n vu + uv = f v − ∂n vuD + uD v ∀v
Ω ∂Ω ∂Ω h ∂Ω h
These are the forms of Nitsche’s method:
Z Z Z Z
α
A(u, v) = ∇u∇v − ∂n uv − ∂n vu + uv
h ∂Ω
ZΩ Z ∂Ω Z∂Ω
α
f (v) = fv − ∂n vuD + uD v
Ω ∂Ω h ∂Ω
A(., .) is not defined for u, v ∈ H 1 , but it requires also well defined normal derivatives.
This is satisfied for the flux ∇u ∈ H(div) of the solution, and finite element test functions
v.
We define the Nitsche norm:

kuk21,h := k∇uk2L2 + h1 kuk2L2 (∂Ω)

Lemma 123. If α = O(p2 ) is chosen sufficiently large, then A(., .) is elliptic on the finite
element space:
A(uh , uh )  kuh k21,h ∀ uh ∈ Vh
Proof. On one element there holds the inverse trace inequality
p2
kuh k2L2 (∂T ) ≤ c kuh k2L2 ∀ uh ∈ P p (T )
h
7.3. NITSCHE’S METHOD FOR DIRICHLET BOUNDARY CONDITIONS 147

The h-factor is shown by transformation to the reference element, the p-factor (polynomial
order) is proven by expansion in terms of orthogonal polynomials. Using the element-wise
estimate for all edges on the domain boundary, we obtain

p2
kuh k2L2 (∂Ω) ≤ c kuk2L2 (Ω) (7.1)
h
Evaluating the bilinear-form, and applying Young’s inequality for the mixed term we get
Z
2 α
A(uh , uh ) = k∇uh kL2 − 2 ∂n uu + kuk2L2 (∂Ω)
∂Ω h
α
≥ k∇uh k2L2 − γ1 kn · ∇uh k2L2 (∂Ω) − γkuk2L2 (∂Ω) + kuk2L2 (∂Ω)
h
The inverse trace inequality applied to ∇uh gives

p2
kn · ∇uh k∂Ω ≤ k∇uh k∂Ω ≤ c k∇uh kΩ
h
By choosing
p2 α
γ>c and γ≤
h h
we can absorb the negative terms into the positive ones. Therefore it is necessary to choose

α > cp2

For the error analysis we apply the discrete stability and consistency:

A(uh − Ih u, vh )
kuh − Ih uk1,h  sup
vh kvh k1,h
A(u − Ih u, vh )
= sup
vh kvh k1,h

We cannot argue with continuity of A(., .) on H 1 (which is not true), but we can estimate
the interpolation error u − Ih u for all four terms of A(uh − Ih u, vh ).

7.3.1 Nitsche’s method for interface conditions

We give now a variational formulation for interface conditions u1 = u2 , ∂n1 u1 + ∂n2 u2 = 0
on the interface γ separating Ω1 and Ω2 . Boundary conditions on the outer boundary are
treated as usual. Integration by parts on the sub-domains leads to
Z Z Z Z Z
∇u∇v − ∂n1 u1 v1 + ∇u∇v − ∂n2 u2 v2 = f v
Ω1 γ Ω2 γ
148 CHAPTER 7. DISCONTINUOUS GALERKIN METHODS

We define the mean value

{∂n1 u} = 21 (∂n1 u1 + ∂n2 u2 )
and jump
[v] = v1 − v2 .
Using continuity of the normal flux (taking the orientation into account) we get
XZ Z Z
∇u∇v − {∂n1 u}[v] = f v.
i Ωi γ

Note that both terms, mean of normal derivative and jump, change sign if we exchange
the enumeration of sub-domains.
We procede as before and add consistent symmetry and stabilization terms:
XZ Z Z
α
Z Z
∇u∇v − {∂n1 u}[v] − {∂n1 v}[u] + [u][v] = f v.
i Ωi γ γ h γ
The variational formulation is consistent on the solution, and elliptic on Vh , which is proven
as before. This approach is an alternative to the mixed method (mortar method), since it
leads to positive definite matrices (called also gluing method).

7.4 DG for second order equations

Nitsche’s method for interface conditions can be applied element-by-element. This is the
(independently developed) Discontinuous Galerkin (DG) method. Precisely, it’s called
SIP-DG (symmetric interior penalty) DG:
X Z Z Z Z 
1 1 α
A(u, v) = ∇u∇v − 2 ∂n u[v] − 2 ∂n v[u] + h [u][v]
T T ∂T ∂T ∂T

(and proper treatment of integrals on the domain boundary). The factor 12 is coming from
splitting the consistent terms to the two elements on the edge.
Convergence analysis similar to Nitsche’s method.
Beside the SIP-DG, also different version are in use: The NIP-DG (non-symmetric
interior penalty) DG:
X Z Z Z Z 
1 1 α
A(u, v) = ∇u∇v − 2 ∂n u[v] + 2 ∂n v[u] + h [u][v]
T T ∂T ∂T ∂T

The term formerly responsible for symmetry is added with a different sign. The variational
problem is still consistent on the true solution. The advantage of the NIP-DG is that
X
A(u, u) = k∇uk2L2 (T ) + αh k[u]k2L2 (∂T ) ,
T
7.4. DG FOR SECOND ORDER EQUATIONS 149

i.e. A(., .) is elliptic in any case α > 0. The disadvantage is that A(., .) is not consistent
for the dual problem, i.e. the Aubin-Nitsche trick cannot be applied. It is popular for
convection-diffusion problems, where the bi-form is non-symmetric anyway. The IIP-DG
(incomplete) skips the third term completely. Advantages are not known to the author.

7.4.1 Hybrid DG
One disadvantage of DG - methods is that the number of degrees of freedom is much
higher than a continuous Galerkin method on the same mesh. Even worse, the number of
non-zero entries per row in the system matrix is higher. The second disadvantage can be
overcome by hybrid DG methods: One adds additional variables û, v̂ on the inter-element
facets (edges in 2D, faces in 3D). The derivation is very similar:
XZ Z X
∇u∇v − ∂n uv = fv ∀ v ∈ P k (T ), ∀ T
T T ∂T T

P R
Using continuity of the normal flux, we may add T ∂T ∂n uv̂ with a single-valued test-
function on the facets:
XZ Z X
∇u∇v − ∂n u(v − v̂) = fv ∀ v ∈ P k (T ), ∀ T
T T ∂T T

Again, we smuggle in consistent terms for symmetry and coercivity:

XZ Z Z
α
Z X
∇u∇v− ∂n u(v−v̂)− ∂n v(u−û)+ (u−û)(v−v̂) = fv ∀ v ∈ P k (T ), ∀ T
T T ∂T ∂T h ∂T T

The jump between neighbouring elements is now replaced by the difference of element-
values and facet values. The natural norm is
X
ku, ûk2 = k∇uk2 + h1 ku − ûk∂T
T

The HDG methods allows for static condensation of internal variables which results in a
global system for the edge-unknowns, only.
The lowest order method uses P 1 (T ) and P 1 (E), and we get O(h) convergence. When
comparing with the non-conforming P 1 -method, HDG has more unknowns on the edges,
but the same order of convergence. The Lehrenfeld-trick is to smuggle in a projector:
XZ Z Z
α
Z X
∇u∇v− ∂n u(v−v̂)− ∂n v(u−û)+ P k−1 (u−û)(v−v̂) = fv ∀ v ∈ P k (T ), ∀ T
T T ∂T ∂T h ∂T T

This allows to reduce the order on edges by one, while maintaining the order of convergence.
150 CHAPTER 7. DISCONTINUOUS GALERKIN METHODS

7.4.2 Bassi-Rebay DG
One disadvantage of IP-DG is the necessary penalty term with α sufficiently large. For well-
shaped meshes α = 5p2 is usually enough. But, for real problems the element deformation
may become large, and then a fixed α is not feasible. Setting α too large has a negative
effect for iterative solvers.
An alternative is to replace the penalty term by
([u], n · σh )2L2 (∂T )
k[u]k2BR := sup
σh ∈[P k−1 ]d kσh k2L2 (T )

It can be implemented by solving a local problem with L2 -bilinear-form.

In the coercivity proof, the bad term is now estimated as
R
n · σh [uh ]
Z
n · ∇uh [uh ] ≤ sup ∂T k∇uh k ≤ k∇uh k k[u]kBR
∂T σh kσh k

The BR-norm scales like the IP - norm (in h and p), but the (typically unknown) constant
in the inverse trace inequality can be avoided.

7.4.3 Matching integration rules

Another method to avoid guessing the sufficiently large α is to use integration rules, such
that the integration points for the boundary integral are a sub-set of the integration points
of the volume term. Now, Young’s inequality can be applied for the numerical integrals.
The αh factor is now replaced by the largest relative scaling of weights for the boundary
integrals and volume integrals. The pro is the simplicity, the con is the need of numerical
integration rules which need more points.

7.4.4 (Hybrid) DG for Stokes and Navier-Stokes

DG or HDG methods allow the construction of numercal methods for incompressible flows,
which obtain exactly divergence free discrete velocities. We discretize Stokes’s equation as
follows:
Vh = BDM k Qh = P k−1,dc ,
and the bilinear-forms
a(uh , vh ) = aDG (uh , vh )
and Z
b(uh , qh ) = div uh qh .

Since the space Vh is not conforming for H 1 , the DG - technique is applied. The b(., .)
bilinear-form is well defined for the H(div)-conforming finite element space Vh . There holds

div Vh = Qh ,
7.4. DG FOR SECOND ORDER EQUATIONS 151

and thus the discrete divergence free condition

Z
div uh qh = 0 ∀qh

implies Z
k div uh k2L2 = div uh div uh = 0.
| {z }
∈Qh

Hybridizing the method leads to facet variables for the tangential components, only. This
method can be applied to the Navier Stokes equations. Here, the exact divergence-free
discrete solution leads to a stable method for the nonlinear transport term (References:
Master thesis Christoph Lehrenfeld: HDG for Navier Stokes, h-version LBB, Master thesis
Philip Lederer: p-robust LBB for triangular elements).
152 CHAPTER 7. DISCONTINUOUS GALERKIN METHODS
Chapter 8

Applications

We investigate numerical methods for equations describing real life problems.

8.1 The Navier Stokes equation

The Navier Stokes equation describe the flow of a fluid (such as water or air). The in-
compressible Navier Stokes equation models incompressible fluids (such as water). The
stationary N.-St. equation models a solution in steady state (no change in time).
The field variables are the fluid velocity u = (ux , uy , uz ), and the pressure p. Conserva-
tion of momentum is
−ν∆u + ρ(u · ∇)u − ∇p = f
The first term describes friction of the fluid (ν is called viscosity). The second one arises
from conservation of momentum of moving particles. It is called the convective term (ρ is
the density). The source term f models forces, mainly gravity. The incompressibility of
the fluid is described by
div u = 0.
Different types of boundary conditions onto u and p are possible.

The Navier Stokes equation is nonlinear. In general, no unique solution is guaranteed.

The common approach to find a solution is the so called Oseen iteration: Given uk , find
the next iterate (uk+1 , pk+1 ) by solving

−ν∆uk+1 + ρ(uk · ∇)uk+1 − ∇pk+1 = f

div uk+1 = 0.

Under reasonable conditions, this Oseen equation is uniquely solvable. Since uk is the
solution of the old step, it satisfies div uk = 0. Furthermore, we assume that the velocity
uk is bounded in L∞ -norm.

153
154 CHAPTER 8. APPLICATIONS

From now on, we continue to investigate the Oseen equation. Given a vector-field
w = (wx , wy , wz ) ∈ [L∞ ]3 such that div w = 0. Find u and p such that

−∆u + (w · ∇)u − ∇p = f
div u = 0.

We have removed the viscosity by rescaling the equation. The factor ρ/ν is incorporated
into the vector-field w.

As usual, we go over to the weak formulation: Find u ∈ V = [H 1 ]3 and p ∈ Q = L2

such that
R R R
{∇u ∇v
R + (w · ∇)u v} dx + div v p dx = f v dx ∀v ∈ V
(8.1)
div u q dx = 0 ∀ q ∈ Q.

This variational problem is a mixed formulation. It satisfies the conditions of Brezzi:

The bilinear forms are
Z
a(u, v) = {∇u∇v + (w · ∇)uv} dx,
Z
b(u, q) = div u q dx.

Both forms are continuous. The form a(., .) is non-symmetric. In a(., .), the x, y, and
z components of u and v are independent. To investigate a(., .), it is enough to consider
scalar bilinear-forms. We define the inflow and outflow boundaries

Γi = {x ∈ ∂Ω : w · n < 0},
Γo = {x ∈ ∂Ω : w · n ≥ 0}.

If we pose Dirichlet boundary conditions on Γi , then a(., .) is coercive (see example 27, and
exercises). The ratio of the continuity bound and the coercivity bound depends on the
norm of the convection w. With increasing w, the problem is getting worse.
The form b(., .) satisfies the LBB condition:
R
div u q dx
sup  kqkL2 ∀ q ∈ L2 .
1 ]3
u∈[H0,D kukH 1

1
In the case of (partial) Dirichlet boundary conditions (H0,D = {u : u = 0 on ΓD }),
this condition Ris very nontrivial to prove. If there are only Dirichlet b.c., one has to use
Q = L02 = {q : Ω q dx = 0}.
Under these conditions, Brezzi’s theorem proves a unique solution of the Oseen equa-
tion.
8.1. THE NAVIER STOKES EQUATION 155

Finite elements for Navier-Stokes equation

We want to approximate the Oseen equation by a Galerkin method: Find uh ∈ Vh and
ph ∈ Qh such that
R R R
{∇uh ∇v h + (w · ∇)uh vh } dx + div vh p h dx = f vh dx ∀ vh ∈ Vh
R (8.2)
div uh qh dx = 0 ∀ qh ∈ Q.

To obtain convergence uh → u and ph → p, it is important to choose proper approxi-

mation spaces Vh and Qh . Using the simplest elements, namely continuous and piece-wise
linear elements for Vh ⊂ [H 1 ]3 , and piece-wise constants for Qh ⊂ L2 does not work. The
discrete LBB condition is not fulfilled: In 2D, thereR are asymptotically twice as many
triangles than vertices, i.e., dim Vh ≈ dim Qh , and div uh qh dx = 0 ∀ qh ∈ Qh implies
uh ≈ 0.

The simplest spaces which lead to convergence are the non-conforming P1 element for
the velocities, and piece-wise constant elements for the pressure. The arguments are
• There are unknowns on the edges to construct a Fortin operator satisfying
Z Z
u · n ds = (Ih u) · n ds,
e e

and thus proving the discrete LBB condition.

• The error due to the non-conforming space Vh 6⊂ V is of the same order as the
approximation error (see Section 4.5).

8.1.1 Proving LBB for the Stokes Equation

Stability of the continuous equation
We consider Stokes equation: find u ∈ [H01 ]d and p ∈ L02 such that
R R R
∇u · ∇v + div v p = fv ∀ v ∈ [H01 ]d
R (8.3)
div u q = 0 ∀ q ∈ L02 .

Solvability follows from Brezzi’s theorem. The only non-trivial part is the LBB condition:
R
div v p
sup ≥ βkpkL2 ∀ p ∈ L02
v∈[H01 ]d kvkH 1

We sketch two different proofs:

Proof 1: The LBB condition becomes simple if we skip the Dirichlet conditions:
R
div v p
sup ≥ βkpkL2 ∀ p ∈ L2
v∈[H 1 ]d kvkH 1
156 CHAPTER 8. APPLICATIONS

Take p ∈ L2 (Ω), extend it by 0 to L2 (Rd ). Now compute a right-inverse of div via Fourier
transform:

p̂(ξ) = F(p)
−iξ
û(ξ) = p̂(ξ)
|ξ|2
u(x) = F −1 (û)

Since div u = p translates to iξ · û = p̂, we found a right-inverse to the divergence. Fur-

thermore, |u|H 1 (Ω) = kiξ ûkL2  kp̂kL2 = kpkL2 . We restrict this u to Ω. The L2 -part of
kukH 1 follows from the Poincare inequality after subtracting the mean value.
The technical part is to ensure Dirichlet - boundary conditions. One can build an
extension operator E from L2 (Rd \ Ω) onto Rd , which commutes with the div-operator:
div Eu = E p div u, and sets
uf inal := u − Eu
This u satisfies u = 0 on ∂Ω. Since div u = p = 0 outside of Ω, the correction did not
change the divergence inside Ω. A self-contained proof is given in J. Bramble: A proof of
the inf-sup condition for the Stokes equation on Lipschitz domains, Mathematical Models
and Methods in Applied Sciences, Vol 13, pp 361-371 (2003).
Proof 2: Directly construct a right-inverse for the div-operator via integration. We
assume that Ω is star-shaped w.r.t. ω, and a ∈ ω. Extend p by 0 to L2 (Rd ):
Z ∞
ua (x) := −(x − a) td−1 p(a + t(x − a)) dt x 6= a
1
R
and ua (a) = 0. If Ω p = 0, then div ua = p. Furthermore, u = 0 outside Ω. Next, we
average over star-points in ω: Z
1
u := ua da
|ω| ω
There is still div u = p. Now, one can show that kukH 1  kpkL2 . See M. Costabel
and A.McIntosh: On Bogovskiĭ and regularized Poincaré integral operators for de Rham
complexes on Lipschitz domains, Mathematische Zeitschrift 265, 297-320 (2010).

8.1.2 Discrete LBB

Now, we turn to the discrete system posed on Vh ⊂ V and Qh ⊂ Q. The discrete LBB
condition follows from the continuous one by construction of a Fortin operator (Lemma 99).

Elements with discontinuous pressure

The simplest pair is the non-conforming P 1 element, and constant pressure:

Vh = P 1,nc , Qh = P 0,dc
8.1. THE NAVIER STOKES EQUATION 157

We have to extend the V -norm and forms by the sum over element-wise norms and forms.
The Fortin-operator IF : V → Vh is defined via
Z Z
IF u = u ∀ edges E
E E

It is continuous from H 1 to broken H 1 (via mapping), and satisfies

Z Z Z Z
div(IF u) = (IF u) · n = u·n= div u,
T ∂T ∂T T

and thus
bh (IF u, qh ) = b(u, qh ) ∀ q h ∈ Qh
The error estimate follows similar as in the second Lemma by Strang:

ku − uh kH 1 ,nc + kp − ph kL2
P R
T ∇u∇wh + p div wh − f wh
 inf ku − vh kH 1 ,nc + kp − qh kL2 + sup
vh ,qh wh kwh kH 1 ,nc
 ch (kukH 2 + kpkH 1 )

This convergence rate O(h) is considered to be optimal for these elements.

Next we consider
Vh = P 2 Qh = P 0,dc
We would like to define the Fortin operator similar as before:
IF u(V ) = u(V ) ∀ vertices V
R R
I u = Eu
E F
∀ edges E

But, the vertex evaluation is not allowed in H 1 . We proceed now in two steps: First
approximate u in the finite element space via a Clément operator Πh :

u1h := Πh u,

and modify this u1h via a correction term:

uh := IF u := u1h + IF2 (u − u1h )

The correction operator IF2 is defined as

IF2 u(V ) = 0 ∀ vertices V,

R 2 R
I u = Eu
E F
∀ edges E.

It preserves edge-integrals and thus satisfies b(u − IF2 u, qh ) = 0 ∀ u ∀ qh . Furthermore, it is

continuous with respect to

kIF2 ukH 1  kukH 1 + h−1 kukL2

158 CHAPTER 8. APPLICATIONS

Thus, the combined operator IF is continuous:

kIF ukH 1  kΠh ukH 1 + kIF2 (u − Πh u)kH1

 kukH 1 + ku − Πh ukH 1 + h−1 ku − Πh ukL2
 kukH 1

It also satisfies the constraints:

b(u − IF u, qh ) = b(u − Πh u − IF2 (u − Πh u), qh ) = b((Id − IF2 )(u − Πh u), qh ) = 0

Error estimates are

ku − uh kH 1 + kp − ph kL2  inf ku − vh kH 1 + kp − qh kL2 = O(h)

vh ,qh

Although we approximate uh with P 2 -elements, the bad approximation of p leads to first

order convergence, only. This element is considered to be sub-optimal.
A solution is the the pairing

Vh = P 2+ Qh = P 1,nc ,

where P 2+ is the second order space enriched with cubic bubbles:

P 2+ (T ) = {vh ∈ H 1 : vh|T ∈ P 3 (T ), vh|E ∈ P 2 (E)}

It leads to second order convergence. Since the costs of a method depend mainly on the
coupling dofs, the price for the additional bubble is low.

Elements with continuous pressure

Although the pressure p is only in L2 , we may approximate it with continuous elements.
The so called mini-element is

Vh = P 1+ Qh = P 1,cont ,

where P 1+ is P 1 enriched by the cubic bubble. The continuous pressure allows integration
by parts: Z Z
div u qh = − u∇qh

The gradient of qh is element-wise constant. We thus construct a Fortin-operator preserving

element-wise mean values. Again, we use the Clément operator and a correction operator:

uh := IF u = Πh u + IF2 (u − Πh u)

The correction is now defined as

2
R IF2 u = 0R on ∪ E
I u = Tu
T F
∀ elements T.
8.2. ELASTICITY 159

It satisfies b(u − IF2 u, qh ) = 0 ∀ u ∀ qh , and, as above:

kIF2 ukH 1  kukH 1 + h−1 kukL2
Thus, the combined operator is a Fortin operator. This method is O(h) convergent.
Another (essentially) stable pair is P 2 × P 1,cont (the popular Taylor Hood element). Its
analysis is more involved. It requires the additional assumption that no two edge of one
element are on the domain boundary. Its convergence rate is O(h2 ).

8.2 Elasticity
We start with a one-dimensional model. Take a beam which is loaded by a force density f
in longitudinal (x) direction. We are interested in the displacement u(x) in x direction.
The variables are
• The strain ε: It describes the elongation. Take two points x and y on the beam.
After deformation, their distance is y + u(y) − (x + u(x)). The relative elongation of
the beam is
{y + u(y) − (x + u(x))} − (y − x) u(y) − u(x)
= .
y−x y−x
In the limit y → x, this is u0 . We define the strain ε as
ε = u0 .

• The stress σ: It describes internal forces. If we cut the piece (x, y) out of the beam,
we have to apply forces at x and y to keep that piece in equilibrium. This force is
called stress σ. Equilibrium is
Z y
σ(y) − σ(x) + f (s) ds = 0,
x
or
σ 0 = −f
Hook’s law postulates a linear relation between the strain and the stress:
σ = Eε.
Combining the three equations
ε = u0 σ = Eε σ 0 = −f
leads to the second order equation for the displacement u:
−(Eu0 )0 = f.
Boundary conditions are
• Dirichlet b.c.: Prescribe the displacement at the boundary
• Neumann b.c: Prescibe the stress at the boundary
160 CHAPTER 8. APPLICATIONS

Elasticity in more dimensions

We want to compute the deformation of the body Ω ⊂ Rd .
• The body is loaded with a volume force density f : Ω → Rd .

• The displacement is described by a the vector-valued function

u : Ω → Rd .

• The strain ε becomes a symmetric tensor in Rd×d . The elongation in the direction of
the unit-vector n is
nT εn.
The (linearized!) relation between the displacement u and the strain is now
 
1 ∂ui ∂uj
εij = + ,
2 ∂xj ∂xi
or, in compact form
1
ε = ε(u) = ∇u + (∇u)T .
2
If the displacement is a pure translation (u = const), then the strain vanishes. Also,
if the displacement is a linearized (!) rotation, (in two dimensions u = (ux , uy ) =
(y, −x), the strain vanishes. We call these deformations the rigid body motions:
    
2D a1 y
R = +b : a1 , a2 , b ∈ R
a2 −x
R3D = a + b × x : a, b ∈ R3


• The stress becomes a tensor σ ∈ Rd×d . Consider the part V ⊂ Ω. To keep V in

equilibrium, on has to apply the surface force density σn at ∂V :
Z Z
σn ds + f dx = 0.
∂V V

Apply Gauss theorem to obtain the differential form

div σ = −f.

The div-operator is applied for each row of σ. A further hypothesis, equilibrium of

angular momentum, implies that σ is symmetric.

• Hook’s law is now a relation between two second order tensors:

X
σij = Dijkl εkl ,
kl
8.2. ELASTICITY 161

in short
σ = Dε,
where D is a fourth order tensor. For an isotropic material (same properties in all
directions), the matrial law has the special structure

σ = 2µ ε + λ tr{ε} I.

The two parameters µ and λ are called Lamé’s parameters. The trace tr is defined
as tr{ε} = di=1 εii .
P

Collecting the equations

ε = ε(u) σ = Dε div σ = −f

leads to
− div D ε(u) = f.

Multiplication with test-functions v : Ω → Rd , and integrating by parts leads to

Z Z
D ε(u) : ∇v dx = f · v dx ∀v
Ω
P
The operator ’:’ is the inner product for matrices, A : B = ij Aij Bij . Next, we use that
σ = Dε(u) is symmetric. Thus, σ : ∇v = σ : (∇v)T = σ : 21 {∇v + (∇v)T }.
1
The equations of elasticity in weak form read as: Find u ∈ V = [H0,D (Ω)]d such that
Z Z
Dε(u) : ε(v) dx = f · v dx ∀ v ∈ V.
Ω Ω

Displacement (Dirichlet) boundary conditions (u = uD at ΓD ) are essential b.c., and

are put into the space V . Neumann boundary conditions
R (natural b.c.) model surface
forces sigman = g, and lead to the additional term ΓN g · v ds on the right hand side.

The bilinear-form in the case of an isotropic material reads as

Z
2µ ε(u) : ε(v) + λ div u div v dx.

We assume a positive definite material law

Dε : ε  ε : ε ∀ symmetric ε ∈ Rd×d

Theorem 124. Assume that the Dirichlet boundary ΓD has positive measure. Then the
equations of elasticity are well posed in [H 1 ]d .
162 CHAPTER 8. APPLICATIONS

Proof: Continuity of the bilinear-form and the linear-form are clear. Ellipticity of the
bilinear-form follows from the positive definite matrial law, and the (non-trivial) Korn
inequality Z
ε(u) : ε(v) dx  kuk2H 1 (Ω) 1
∀ u ∈ [H0,D ]d
Ω
The Lax-Milgram theorem proves a unique solution u. 2

The discretization of the elasticity problem is straight forward. Take a finite dimensional
sub-space Vh ⊂ V , and perform Galerkin projection. One may use the ’standard’ nodal
finite elements for each component.

Structural mechanics
Many engineering applications involve thin structures (walls of a building, body of a car,
...). On thin structures, the standard approach has a problem: One observed that the
simulation results get worse as the thickness decreases. The explanation is that the constant
in Korn’s inequality gets small for thin structures. To understand and overcome this
problem, we go over to beam, plate and shell models.

We consider a thin (t  1) two-dimensional body

Ω = I × (−t/2, t/2) with I = (0, 1)
The goal is to derive a system of one-dimensional equations to describe the two-dimensional
deformation. This we obtain by a semi-discretization. Define
My
(  Mx
)
u x (x, y) X
i i
X
i i
VeM = ∈ V : ux (x, y) = ux (x)y , uy (x, y) = uy (x)y .
uy (x, y) i=0 i=0

This function space on Ω ⊂ R2 is isomorph to a one-dimensional function space with values

in RMx +My +2 . We perform semi-discretization by searching for ũ ∈ VeM such that
A(ũ, ṽ) = f (ṽ) ∀ ṽ ∈ VeM .
As Mx , My → ∞, VeM → V , and we obtain convergence ũ → u.

The lowest order (qualitative) good approximating semi-discrete space is to set Mx = 1

and My = 0. This is  
U (x) − β(x)y
V =
e
w(x)
Evaluating the bilinear-form (of an isotropic material) leads to
  !! Z 1
U − yβ Ũ − y β̃
A , = (2µ + λ)t U 0 Ũ 0 dx +
w w̃ 0

t3 1 0 0 t 1 0
Z Z
(2µ + λ) β β̃ + 2µ (w − β)(w̃0 − β̃) dx
12 0 2 0
8.2. ELASTICITY 163

The meaning of the three functions is as follows. The function U (x) is the average
(over the cross section) longitudinal displacement, w(x) is the vertical displacement. The
function β is the linearized rotation of the normal vector.

We assume that the load f (x, y) does not depend on y. Then, the linear form is
! Z 1 Z 1
Ũ − y β̃
f =t fx Ũ dx + t fy w̃ dx
w̃ 0 0

The semi-discretization in this space leads to two decoupled problems. The first one
describes the longitudinal displacement: Find U ∈ H 1 (I) such that
Z 1 Z 1
0 0
(2µ + λ)t U Ũ dx = t fx Ũ dx ∀ Ũ ∈ H 1 (I).
0 0

The small thickness parameter t cancels out. It is a simple second order problem for the
longitudinal displacement.
The second problems involves the 1D functions w and β: Find (w, β) ∈ V =? such that

t3 1 0 0
Z Z 1 Z 1
0 0
(2µ + λ) β β̃ dx + µt (w − β)(w̃ − β̃) dx = t fy w̃ dx ∀ (w̃, β̃) ∈ V
12 0 0 0

The first term models bending. The derivative of the rotation β is (approximative) the
curvature of the deformed beam. The second one is called the shear term: For thin beams,
the angle β ≈ tan β is approximatively w0 . This term measures the difference w0 − β. This
second problem is called the Timoshenko beam model.

For simplification, we skip the parameters µ and λ, and the constants. We rescale the
equation by dividing by t3 : Find (w, β) such that
Z Z Z
1
β β̃ dx + 2 (w − β)(w̃ − β̃) dx = t−2 f w̃ dx.
0 0 0 0
(8.4)
t
This scaling in t is natural. With t → 0, and a force density f ∼ t2 , the deformation
converges to a limit. We define the scaled force density

f˜ = t−2 f

In principle, this is a well posed problem in [H 1 ]2 :

Lemma 125. Assume boundary conditions w(0) = β(0) = 0. The bilinear-form
A((w, β), (w̃, β̃)) of (8.4) is continuous

A((w, β), (w̃, β̃))  t−2 (kwkH 1 + kβkH 1 )(kw̃kH 1 + kβ̃kH 1 )

and coercive
A((w, β), (w, β)) ≥ kwk2H 1 + kβk2H 1
164 CHAPTER 8. APPLICATIONS

Proof: ...
As the thickness t becomes small, the ratio of the continuity and coercivity bounds
becomes large ! This ratio occurs in the error estimates, and indicates problems. Really,
numerical computations show bad convergence R for small thickness t.
The large coefficient in front of the term (w − β)(w̃0 − β̃) forces the difference w0 − β
0

to be small. If we use piece-wise linear finite elements for w and β, then wh0 is a piece-wise
constant function, and βh is continuous. If wh0 − βh ≈ 0, then βh must be a constant
function !

The idea is to weaken the term with the large coefficient. We plug in the projection P 0
into piece-wise constant functions: Find (wh , βh ) such that
Z Z Z
1
βh β̃h dx + 2 P (wh − βh ) P (w̃h − β̃h ) dx = f˜w̃h dx.
0 0 0 0 0 0
(8.5)
t
Now, there are finite element functions wh and βh fulfilling P 0 (wh0 − βh ) ≈ 0.
In the engineering community there are many such tricks to modify the bilinear-form.
Our goal is to understand and analyze the obtained method.

Again, the key is a mixed method. Start from equation (8.4) and introduce a new
variable
p = t−2 (w0 − β). (8.6)
Using the new variable in (8.4), and formulating the definition (8.6) of p in weak form leads
to the bigger system: Find (w, β) ∈ V and p ∈ Q such that

f˜w̃ dx
R 0 0 R 0 R
β β̃ dx + (w̃ − β̃)p dx = ∀ (w̃, β̃) ∈ V
R 0 R (8.7)
(w − β)p̃ dx − t2 p p̃ dx = 0 ∀ p̃ ∈ Q.

This is a mixed formulation of the abstract structure: Find u ∈ V and p ∈ Q such that
a(u, v) + b(v, p) = f (v) ∀ v ∈ V,
(8.8)
b(u, q) − t2 c(p, q) = 0 ∀ q ∈ Q.

The big advantage now is that the parameter t does not occur in the denominator, and
the limit t → 0 can be performed.
This is a family of well posed problems.
Theorem 126 (extended Brezzi). Assume that the assumptions of Theorem 101 are true.
Furthermore, assume that
a(u, u) ≥ 0,
and c(p, q) is a symmetric, continuous and non-negative bilinear-form. Then, the big form

B((u, p), (v, q)) = a(u, v) + b(u, q) + b(v, p) − t2 c(p, q)

is continuous and stable uniformly in t ∈ [0, 1].

8.2. ELASTICITY 165

We check Brezzi’s condition for the beam model. The spaces are V = [H 1 ]2 and Q = L2 .
Continuity of the bilinear-forms a(., .), b(., .), and c(., .) is clear. The LBB condition is
R 0
(w − β)q dx
sup  kqkL2
w,β kwkH 1 + kβkH 1

We construct a candidate for the supremum:

Z x
w(x) = q(s) ds and β=0
0

Then
(w0 − β)q dx
R R 2
q dx
 = kqkL2
kwkH 1 + kβkH 1 kw0 k
Finally, we have to check kernel ellipticity. The kernel is

V0 = {(w, β) : β = w0 }.

On V0 there holds

kwk1H 1 + kβk2H 1  kw0 k2 + kβk2H 1 = kβk2L2 + kβk2H 1

 kβ 0 kL2 = a((w, β), (w, β))

The lowest order finite element discretization of the mixed system is to choose contin-
uous and piece-wise linear elements for wh and βh , and piecewise constants for ph . The
discrete problem reads as: Find (wh , βh ) ∈ Vh and ph ∈ Qh such that

f˜w̃h dx
R 0 0 R 0 R
βh β̃h dx + (w̃h − βh )ph dx = ∀ (wh , βh ) ∈ Vh
R 0 R (8.9)
(wh − βh )p̃h dx − t2 ph p̃h dx = 0 ∀ p̃h ∈ Qh .

This is a inf-sup stable system on the discrete spaces Vh and Qh . This means, we obtain
the uniform a priori error estimate

k(w − wh , β − βh )kH1 + kp − ph kL2  inf k(w − w̃h , β − β̃h )kH1 + kp − p̃h kL2
w̃h ,β̃h ,p̃h
 h {kwkH 2 + kβkH 2 + kpkH 1 }

The required regularity is realistic.

The second equation of the discrete mixed system (8.9) states that

ph = t−2 P 0 (wh0 − βh )

If we insert this observation into the first row, we obtain exactly the discretization method
(8.5) ! Here, the mixed formulation is a tool for analyzing a non-standard (primal) dis-
cretization method. Both formulations are equivalent. They produce exactly the same
finite element functions. The mixed formulation is the key for the error estimates.
166 CHAPTER 8. APPLICATIONS

The two pictures below show simulations of a Timoshenko beam. It is fixed at the left
end, the load density is constant one. We compute the vertical deformation w(1) at the
right boundary. We vary the thickness t between 10−1 and 10−3 . The left pictures shows
the result of a standard conforming method, the right picture shows the results of the
method using the projection. As the thickness decreases, the standard method becomes
worse. Unless h is less than t, the results are completely wrong ! The improved method
converges uniformly well with respect to t:
0.12 0.12
t=1e-1 t=1e-1
t=1e-2 t=1e-2
t=1e-3 t=1e-3
0.1 0.1

0.08 0.08
w(1)

w(1)
0.06 0.06

0.04 0.04

0.02 0.02

0 0
1 10 100 1000 10000 1 10 100 1000
Elements Elements

8.3 Maxwell equations

Maxwell equations describe electro-magnetic fields. We consider the special case of sta-
tionary magnetic fields. Maxwell equations are three-dimensional.
A magnetic field is caused by an electric current. We suppose that a current density

j ∈ [L2 (Ω)]3

is given. (Stationary) currents do not have sources, i.e., div j = 0.

The involved (unknown) fields are
• The magnetic flux B (in German: Induktion). The flux is free of sources, i.e.,

div B = 0.

• The magnetic field intensity H (in German: magnetische Feldstärke). The field is
related to the current density by Henry’s law:
Z Z
j · n ds = H · τ ds ∀ Surfaces S
S ∂S

By Stokes´ Theorem, one can derive Henry’s law in differential form:

curl H = j

The differential operator is curl = rot = ∇×. Both fields are related by a material law.
The coefficient µ is called permeability:

B = µH
8.3. MAXWELL EQUATIONS 167

The coefficient µ is 103 to 104 times larger in iron (and other ferro-magnetic metals) as
in most other media (air). In a larger range, the function B(H) is also highly non-linear.
Collecting the equations we have

div B = 0 B = µH curl H = j (8.10)

In principle, Maxwell equations are valid in the whole R3 . For simulation, we have to
truncate the domain and have to introduce artificial boundary conditions.
The picture below shows the magnetic field caused by a tangential current density in a
coil:

Compare these equations to the diffusion equation − div a∇u = f . Here, we could
introduce new unknowns g = ∇u and σ = ag. On simply connected domains, g is a
gradient field if and only if curl g = 0. We could reformulate the equations as: Find vector
fields g and σ such that

curl g = 0 σ = ag div σ = −f.

The system of magnetostatic equations looks similar. Only, the right hand side data is
applied to the curl-equation, instead of the div-equation. In a similar way as curl g = 0
allows to introduce a scalar field u such that g = ∇u, div B = 0 allows to introduce a
vector potential A such that
B = curl A.
Inserting the vector-potential into the equations (8.10), one obtains the second order equa-
tion
curl µ−1 curl A = j. (8.11)
The two original fields B and H can be obtained from the vector potential A.
168 CHAPTER 8. APPLICATIONS

The vector-potential A is not uniquely defined by (8.11). One may add a gradient field
to A, and the equation is still true. To obtain a unique solution, the so called Coloumb-
Gauging can be applied:
div A = 0. (8.12)
As usual, we go over to the weak form. Equations (8.11) and (8.12) together become:
Find A such that Z Z
−1
µ curl A curl v dx = j · v dx ∀ v ∈?
Ω Ω
and Z
A · ∇ψ dx = 0.
Ω
We want to choose the same space for A and the according test functions v. But, then
we have more equations than unknowns. The system is still solvable, since we have made
the assumption div j = 0, and thus j is in the range of the curl- operator. To obtain a
symmetric system, we add a new scalar variable ϕ. The problem is now: Find A ∈ V =?
and ϕ ∈ Q = H 1 /R such that
R −1 R R
µ curl A · curl v dx + ∇ϕ · v dx = j · v dx ∀v ∈ V
R (8.13)
A · ∇ψ dx = 0 ∀ψ ∈ Q

The proper space V is the H(curl):

H(curl) = {v ∈ [L2 (Ω)]3 : curl v ∈ [L2 (Ω)]3 }

Again, the differential operator curl is understood in the weak sense. The canonical norm
is
1/2
kvkH(curl) = kvk2L2 + k curl vk2L2

.
Similar to H 1 and H(div), there exists a trace operator for H(curl). Now, only the
tangential components of the boundary values are well defined:
Theorem 127 (Trace theorem). There exists a tangential trace operator trτ v : H(curl) →
W (∂Ω) such that
trτ v = (v|∂Ω )τ
for smooth functions v ∈ [C(Ω)]3 .
Theorem 128. Let Ω = ∪Ωi . Assume that u|Ωi ∈ H(curl, Ωi ), and the tangential traces
are continuous across the interfaces γij . Then u ∈ H(curl, Ω).
The theorems are according to the ones we have proven for H(div). But, the proofs (in
R3 ) are more involved.

The gradient operator ∇ relates the space H 1 and H(curl):

∇ : H 1 → H(curl)
8.3. MAXWELL EQUATIONS 169

Furthermore, the kernel space

H 0 (curl) = {v ∈ H(curl) : curl v = 0}

is exactly the range of the gradient:

H 0 (curl) = ∇ H 1

Theorem 129. The mixed system (8.13) is a well posed problem on H(curl) × H 1 /R.

Proof: The bilinear-forms

Z
a(A, v) = µ−1 curl A · curl v dx

and Z
b(v, ϕ) = v · ∇ϕ dx

are continuous w.r.t. the norms of V = H(curl) and Q = H 1 /R.

The LBB-condition in this case is trivial. Choose v = ∇ϕ:

k∇ϕk2L2
R R
v∇ϕ dx ∇ϕ · ∇ϕ dx
sup ≥ = = k∇ϕkL2 ' kϕkQ
v∈H(curl) kvkH(curl) k∇ϕkH(curl) k∇ϕkL2

The difficult part is the kernel coercivity of a(., .). The norm involves also the L2 -norm,
while the bilinear-form only involves the semi-norm k curl vkL2 . Coercivity cannot hold on
the whole V : Take a gradient function ∇ψ. On the kernel, the L2 -norm is bounded by the
semi-norm:
kvkL2  k curl vk ∀ v ∈ V0 ,
where Z
V0 = {v ∈ H(curl) : v∇ϕ dx = 0 ∀ ϕ ∈ H 1 }

This is a Friedrichs-like inequality.

Finite elements in H(curl)

We construct finite elements in three dimensions. The trace theorem implies that functions
in H(curl) have continuous tangential components across element boundaries (=faces).

We design tetrahedral finite elements. The pragmatic approach is to choose the element
space as VT = P 1 , and choose the degrees of freedom as the tangential component along
the edges in the end-points of the edges. The dimension of the space is 3 × dim{P 1 } =
3 × 4 = 12, the degrees of freedom are 2 per edge, i.e., 2 × 6 = 12. They are also linearly
independent. In each face, the tangential component has 2 components, and is linear.
Thus, the tangential component has dimension 6. These 6 values are defined by the 6
170 CHAPTER 8. APPLICATIONS

degrees of freedom of the 3 edges in the face. Neighboring elements share this 6 degrees of
freedom in the face, and thus have the same tangential component.

There is a cheaper element, called Nédélec, or edge-element. It has the same accuracy
for the curl-part (the B-field) as the P 1 -element. It is similar to the Raviart-Thomas
element. It contains all constants, and some linear polynomials. All 3 components are
defined in common. The element space is

VT = {a + b × x : a, b ∈ R3 }.

These are 6 coefficients. For each of the 6 edges of a tetrahedron, one chooses the integral
of the tangential component along the edge
Z
ψEi (u) = u · τEi ds.
Ei

Lemma 130. The basis function ϕEi associated with the edge Ei is

ϕEi = λEi1 ∇λEi2 − ∇λEi2 λEi1 ,

where Ei1 and Ei2 are the two vertex numbers of the edge, and λ1 , . . . λ4 are the vertex shape
functions.
Proof:
• These functions are in VT
• If i 6= j, then ψEj (ϕEi ) = 0.
• ψEi (ϕEi ) = 1
Thus, edge elements belong to H(curl). Next, we will see that they have also very
interesting properties.

The de’Rham complex

The spaces H 1 , H(curl), H(div), and L2 form a sequence:
∇ curl div
H 1 −→ H(curl) −→ H(div) −→ L2
Since ∇H 1 ⊂ [L2 ]3 , and curl ∇ = 0, the gradients of H 1 functions belong to H(curl).
Similar, since curl H(curl) ⊂ [L2 ]3 , and div curl = 0, the curls of H(curl) functions belong
to H(div).
The sequence is a complete sequence. This means that the kernel of the right differential
operator is exactly the range of the left one (on simply connected domains). We have used
this property already in the analysis of the mixed system.
The same property holds on the discrete level: Let
8.3. MAXWELL EQUATIONS 171

Wh be the nodal finite element sub-space of H 1

Vh be the Nédélec (edge) finite element sub-space of H(curl)
Qh be the Raviart-Thomas (face) finite element sub-space of H(div)
Sh be the piece-wise constant finite element sub-space of L2
Theorem 131. The finite element spaces form a complete sequence
∇ curl div
Wh −→ Vh −→ Qh −→ Sh

Now, we discretize the mixed formulation (8.13) by choosing edge-finite elements for
H(curl), and nodal finite elements for H 1 : Find Ah ∈ Vh and ϕh ∈ Wh such that
R −1 R R
µ curl Ah · curl vh dx + ∇ϕh · vh dx = j · vh dx ∀ vh ∈ Vh
R (8.14)
Ah · ∇ψh dx = 0 ∀ ψh ∈ Wh

The stability follows (roughly) from the discrete sequence property. The verification of the
LBB condition is the same as on the continuous level. The kernel of the a(., .)- form are
the discrete gradients, the kernel of the b(., .)-form is orthogonal to the gradients. This
implies solvability. The discrete kernel-coercivity (with h-independent constants) is true
(nontrivial).

The complete sequences on the continuous level and on the discrete level are con-
nected in the de’Rham complex: Choose the canonical interpolation operators (vertex-
interpolation I W , edge-interpolation I V , face-interpolation I Q , L2 -projection I S ). This
relates the continuous level to the discrete level:
∇ curl div
H 1 −→ H(curl) −→ H(div) −→ L2
   
 W  V  Q  S
yI yI yI yI (8.15)
∇ curl div
Wh −→ Vh −→ Qh −→ Sh .

Theorem 132. The diagram (8.15) commutes:

I V ∇ = ∇I W I Q curl = curl I V I S div = div I Q

Proof: We prove the first part. Note that the ranges of both, ∇I W and I V ∇, are in
Vh . Two functions in Vh coincide if and only if all functionals coincide. It remains to prove
that Z Z
(∇I w) · τ ds = (I V ∇w) · τ ds
W
E E
V
Per definition of the interpolation operator I there holds
Z Z
V
(I ∇w) · τ ds = ∇w · τ ds
E E
172 CHAPTER 8. APPLICATIONS

Integrating the tangential derivative gives the difference

Z Z
∂w
∇w · τ ds = ds = w(E 2 ) − w(E 1 )
E E ∂τ
Starting with the left term, and using the property of the nodal interpolation operator, we
obtain Z
(∇I W w) · τ ds = (I W w)(E 2 ) − (I W w)(E 1 ) = w(E 2 ) − w(E 1 ).
E

We have already proven the commutativity of the H(div) − L2 part of the diagram. The
middle one involves Stokes´ theorem. 2
This is the key for interpolation error estimates. E.g., in H(curl) there holds

ku − I V uk2H(curl) = ku − I V uk2L2 + k curl(I − I V )uk2L2

= ku − I V uk2L2 + k(I − I Q ) curl uk2L2
 h2 kuk2H 1 + h2 k curl uk2H 1

Since the estimates for the L2 -term and the curl-term are separate, one can also scale each
of them by an arbitrary coefficient.

The sequence is also compatible with transformations. Let F : Tb → T be an (element)

transformation. Choose

w(F (x)) = ŵ(x)

v(F (x)) = (F 0 )−T v̂(x) (covariant transformation)
q(F (x)) = (det F 0 )−1 (F 0 )q(x) (Piola-transformation)
s(F (x)) = (det F 0 )−1 ŝ(x)

Then

v̂ = ∇ŵ ⇒ v = ∇ w
q̂ = curl v̂ ⇒ q = curl v
ŝ = div q̂ ⇒ s = div q

Using these transformation rules, the implementation of matrix assembling for H(curl)-
equations is very similar to the assembling for H 1 problems (mapping to reference element).
Chapter 9

Parabolic partial differential

equations

PDEs involving first order derivatives in time, and an elliptic differential operator in space,
are called parabolic PDEs. For example, time dependent heat flow is described by a
parabolic PDE.
Let Ω ⊂ Rd , and Q = Ω × (0, T ). Consider the initial-boundary value problem
∂u(x, t)
− div(a(x)∇x u(x, t)) = f (x, t) (x, t) ∈ Q,
∂t
with boundary conditions

u(x, t) = uD (x, t) (x, t) ∈ ΓD × (0, T ),

∂u
a(x) = g(x, t) (x, t) ∈ ΓN × (0, T ),
∂n
and initial conditions
u(x, 0) = u0 (x) x ∈ Ω.

1
Weak formulation in space: Find u : [0, T ] → H0,D (Ω) such that
Z Z Z Z
∂t u(x, t)v(x) dx + a∇u(x, t) · ∇v(x, t) dx = f (x, t)v(x, t) dx + g(x, t)v(x, t) dx
Ω Ω Ω ΓN

1
∀ v ∈ H0,D , t ∈ (0, T ]

In abstract form: Find u : [0, T ] → V s.t.

(u0 (t), v)L2 + a(u(t), v) = hf (t), vi ∀ v ∈ V, t ∈ (0, T ]

In operator form (with hAu, vi = a(u, v)):

u0 (t) + Au(t) = f (t) ∈V∗

173
174 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

Function spaces:

X = L2 ((0, T ), V ) X ∗ = L2 ((0, T ), V ∗ )
with norms
Z T 1/2 Z T 1/2
kvkX = kv(t)k2V dt kvkX ∗ = kv(t)k2V ∗ dt
0 0

Definition 133. Let u ∈ L2 ((0, T ), V ). It has a weak derivative w ∈ L2 ((0, T ), V ∗ ) if

Z T Z T
ϕ(t) hw, viV ∗ ×V dt = − ϕ0 (t)(u, v)L2 dt ∀ v ∈ V, ∀ ϕ ∈ C0∞ (0, T )
0 0

Definition 134.

H 1 ((0, T ), V ; L2 ) = {v ∈ L2 ((0, T ), V ) : v 0 ∈ L2 ((0, T ), V ∗ )}

with norm
kvk2H 1 = kvk2X + kv 0 k2X ∗ .

This space is a one-dimensional Sobolev space with range in a Hilbert space.

Theorem 135 (Trace theorem). Point evaluation is continuous:

max kv(t)kL2  kvkH 1

t∈[0,T ]

This allows the formulation of the initial value u(0) = u0 .

Theorem 136. Assume that a(., .) is coercive

a(u, u) ≥ µ1 kuk2V ∀u ∈ V

and continuous
a(u, v) ≤ µ2 kukV kvkV ∀ u, v ∈ V.
Then, the parabolic problem has a unique solution depending continuously on the right hand
side and the initial conditions:

kukH 1 ((0,T ),V ;L2 )  ku0 kL2 + kf kL2 ((0,T ),V ∗ .

We only prove stability: Choose test functions v = u(t):

(u0 (t), u(t))L2 + a(u(t), u(t)) = hf (t), u(t)i

Use that
d
ku(t)k2L2 = 2(u0 (t), u(t))L2 ,
dt
9.1. SEMI-DISCRETIZATION 175

and integrate the equation over (0, T ):

Z T
1
ku(T )k2L2 − ku0 k2L2 = hf (s), u(s)i − a(u(s), u(s)) ds
2 0
Z T
≤ kf (s)kV ∗ ku(s)kV − µ1 ku(s)k2V ds
0
≤ kf kX ∗ kukX − µ1 kuk2X

Since ku(T )k ≥ 0, one has

1
µ1 kuk2X − kf kX ∗ kukX ≤ ku0 kL2
2
Solving the quadatic inequality, one obtains the bound
1 n q o
kukX ≤ kf kX ∗ + kf k2X ∗ + 2µ1 ku0 k2L2
2µ1
The bound ku0 kL2 ((0,T ),V ∗ ) follows from u0 (t) = f (t) − Au(t).

9.1 Semi-discretization
We start with a discretization in space. Choose a (finite element) sub-space Vh ⊂ V . The
Galerkin discretiztaion is: Find u : [0, T ] → Vh such that

(u0h (t), vh )L2 + a(uh (t), vh ) = hf (t), vh i ∀ vh ∈ Vh , ∀ t ∈ (0, T ],

and initial conditions

(uh (0), vh )L2 = (u0 , vh )L2 ∀ vh ∈ Vh .

Choose a basis {ϕ1 , . . . ϕN } of Vh . Expand the solution w.r.t. this basis:

N
X
uh (x, t) = ui (t)ϕi (x),
i=1

and choose test functions v = ϕj . With the matrices

M = ((ϕj , ϕi )L2 )i,j=1,...,N A = (a(ϕj , ϕi ))i,j=1,...,N ,

and the t-dependent vector

f (t) = (hf (t), ϕj i)i=1,...,N ,
one obtains the system of ordinary differential equations (ODEs)

M u0 (t) + Au(t) = f (t), u(0) = u0

176 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

In general, the (mass) matrix M is non-diagonal. In the case of the (inexact) vertex
integration rules, or non-conforming P1 -elements, M is a diagonal matrix. Then, this ODE
can be efficiently reduced to explicit form

u0 (t) + M −1 Au(t) = f (t)

Theorem 137. There holds the error estimate

ku − uh kH 1 ((0,T ),V ;L2 )  k(I − Rh )ukH 1 ((0,T ),V ;L2 ) ,

where Rh is the Ritz projector

Rh : V → Vh : a(Rh u, vh ) = a(u, vh ) ∀ u ∈ V, ∀ vh ∈ Vh .

Proof: The error is split into two parts:

u(t) − uh (t) = u(t) − Rh u(t) + Rh u(t) − uh (t)

| {z } | {z }
ρ(t) Θh

The first part, u(t) − Rh u(t) is the elliptic discretization error, which can be bounded by
Cea’s lemma. To bound the second term, we use the properties for the continuous and the
discrete formulation:

hf, vh i = (u0 , vh ) + a(u, vh ) = (u0 , vh ) + a(Rh u, vh )

= (u0h , vh ) + a(uh , vh ),

i.e.,
(u0 − u0h , vh ) + a(Rh u − uh , vh ) = 0,
or
(Rh u0 − u0h , vh ) + a(Rh u − uh , vh ) = (Rh u0 − u0 , vh ).
With the abbreviations from above we obtain the discrete parabolic equation for Θh :

(Θ0h , vh ) + a(Θh , vh ) = (ρ0 , vh )

Θh (0) = (I − Rh )u(0).

The stability estimate, and the trace theorem bounds

kΘh kH 1 ((0,T ),V ;L2 )  k(I − Rh )u(0)kL2 (Ω) + kρ0 kL2 ((0,T ),V ∗ )
 k(I − Rh )ukH 1 ((0,T ),V ;L2 )

2
9.2. TIME INTEGRATION METHODS 177

9.2 Time integration methods

Next, we discuss methods for solving the system of ODEs:
M u0 (t) + Au(t) = f (t) (9.1)
u(0) = u0
We focus on simple time integration rules and the specific properties arising from the
space-discretization of parabolic PDEs. Let
0 = t0 < t1 < tm = T,
a partitioning of the interval [0, T ]. Define τj = tj+1 − tj . Integrating (9.1) over the
intervalls leads to
Z tj+1 Z tj+1
M {u(tj+1 ) − u(tj )} + Au(s) ds = f (s) ds.
tj tj

Next, we replace the integrals by numerical integration rules. The left-sided rectangle
rule leads to
M {u(tj+1 ) − u(tj )} + τj Au(tj ) = τj f (tj )
With the notation uj = u(tj ), this leads to the sequence of linear equations
M uj+1 = M uj + τj (fj − Auj )
In the case of a diagoal M -matrix, this is an explicit formulae for the new time step !
Using the right-sided rectangle rule leads to
M {uj+1 − uj } + τj Auj+1 = τj fj+1 ,
or
(M + τj A)uj+1 = M uj + τj fj+1 .
In case of the right-side rule, a linear system must be solve in any case. Thus, this method
is called an implicit time integration method. These two special cases are called the explicit
Euler method, and the implicit Euler method. A third simple choice is the trapezoidal rule
leading to
τj τj
(M + A)uj+1 = M uj + (fj + fj+1 − Auj )
2 2
It is also an implcit method. Since the trapezoidal integration rule is more accurate, we
expect a more accurate method for approximating the ODE.
All single-step time integration methods can be written in the form
uj+1 = Gj (uj , fj ),
where Gj is linear in both arguments and shall be continuous with bounds
kGj (uj , fj )kM ≤ L kuj kM + τj l kfj kM −1 ,
with L ≥ 1.
178 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

Lemma 138. The time integration method fulfills the stability estimate
j−1
X
j j
kuj kM ≤ L ku0 kM + lL τi kfi kM −1 (9.2)
i=0

The explicit Euler method is written as

uj+1 = (I − τ M −1 A)u + τ M −1 fj ,
and has bounds
τ
L = max{1, τ λmax (M −1 A) − 1} ' max{1, }
h2
l = 1
If τ > h2 , the powers Lj become very large. This means that the explicit Euler method
becomes instable. Thus, for the explicit Euler method, the time-step τ must not be greater
than ch2 .
The implicit Euler method is written as
uj+1 = (M + τ A)−1 M uj + τ (M + τ A)−1 fj ,
and has the bounds
L = 1
l = 1
The method is stable for any time-step τ . Such a method is called A-stable.
Lemma 139. The time discretization error ej := u(tj ) − uj of the implicit Euler method
satisfies the difference equation
M {ej+1 − ej } + τ Aej+1 = dj ,
where the dj satisfy
Z tj+1
dj = {f (s) − Au(s)} ds − τj {f (tj+1 ) − Au(tj+1 )}.
tj

Lemma 140. The error of the integration rule can be estimated by

kdj k  τ k(f − Au)0 kL∞ = τ ku00 kL∞
Convergence of the time-discretization method follows from stability plus approxima-
tion:
Theorem 141. The error of the implicit Euler method satisfies
j
X
ku(tj ) − uj kM ≤ τ kdj k  τ ku00 kL∞ (0,T )
i=0

The trapezoidal rule is A-stable, too. It is based on a more accurate integration rule,
and leads to second order convergence O(τ 2 ). Convergence of higher order can be obtained
by Runge-Kutta methods.
9.3. SPACE-TIME FORMULATION OF PARABOLIC EQUATIONS 179

9.3 Space-time formulation of Parabolic Equations

In the previous section we have discretized in space to obtain an ordinary differential
equation, which is solved by some time-stepping method. This approach is known as
method of lines. Now we formulate a space-time variational problem. This is discretized
in time and space by a (discontinuous) Galerkin method. We obtain time-slabs which are
solved one after another. This approach is more flexible, since it allows to use different
meshes in space on different time-slabs.

9.3.1 Solvability of the continuous problem

Let V ⊂ H be Hilbert spaces, typically H = L2 (Ω) and V = H 1 (Ω). Duality is defined
with respect to H. For t ∈ (0, T ) we define the familiy A(t) : V → V ∗ of uniformely
continuous and elliptic operators:

(a) hA(t)u, ui ≥ α1 kuk2V

(b) hA(t)u, vi ≤ α2 kukV kvkV

We assume that hA(t)u, vi is integrable with respect to time. We do not assume that A(t)
is symmetric. We consider the parabolic equation: Find u : [0, T ] → V such that

u0 + Au = f ∀ t ∈ (0, T )
u(0) = u0

We define X = {v ∈ L2 (V ) : v 0 ∈ L2 (V ∗ )} and Y = L2 (V ), with its dual Y ∗ = L2 (V ∗ ). A

variational formulation is: Find u ∈ X such that
Z T Z T
0
hu + Au, vi = hf, vi ∀v ∈ Y (9.3)
0 0
(u(0), v0 )H = (u0 , v0 ) ∀v0 ∈ H (9.4)

Adding up both equations leads to the variational problem B(u, v) = f (v) with the
bilinear-form B(., .) : X × (Y × H) → R:
Z T
B(u, (v, v0 )) = hu0 + Au, vi + (u(0), v0 )H
0

and the linear-form f : Y × H → R:

Z T
f (v, v0 ) = hf, vi + (u0 , v0 )H
0

We assume that f ∈ Y ∗ and u0 ∈ H

Theorem 142 (Lions). Problem (9.3)-(9.4) is uniquely solvable.

180 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

Proof. We apply the theorem by Babuška-Aziz. We observe that all forms are continuous
(trace-theorem). We have to verify both inf-sup conditions.
First, we show
B(u, v)
inf sup ≥β>0 (9.5)
u∈X (v,v0 )∈Y ×H kukX k(v, v0 )kY ×H

We fix some u ∈ X and set (with A−T the inverse of the adjoint operator)
v := A−T u0 + u
v0 := u(0)
and obtain
Z
B(u, v) = u0 + Au, A−T u0 + u dt + (u(0), u(0))H
Z
= A−1 u0 , u0 + hAu, ui + hu0 , ui + hu, u0 i dt + ku0 k2H
Z
d
= A−1 u0 , u0 + hAu, ui + kuk2H + ku0 k2H
dt
≥ α2−1 ku0 k2L2 (V ∗ ) + α1 kuk2L2 (V ) + ku(T )k2H
 kuk2X
Since k(v, v0 )kY ×H  kukX the first inf − sup-condition is proven. For the other one, we
show
∀ 0 6= (v, v0 ) ∈ Y × H ∃ u ∈ X : B(u, v) > 0 (9.6)
We fix some v, v0 . We define u by solving the parabolic equation
u0 + γLu = AT v, u(0) = v0 ,
where L is a symmetric, constant-in-time, continuous and elliptic operator on V . The
parameter γ > 0, γ = O(1) will be fixed later. The equation has a unique solution, which
can be constructed by spectral theory. If (v, v0 ) 6= 0, then also u 6= 0.
Z
B(u, v) = u0 + Au, A−T (u0 + γLu) + kv0 k2H
Z
= u0 , A−T u0 + hu, u0 i + u0 , A−T γLu + γ hu, Lui dt + kv0 k2H
Z
≥ 1
α2
ku0 k2V ∗ + 1 d
2 dt
kuk2H − ku0 kV ∗ kA−T γLukV + γ hu, Lui + kv0 k2H

The second term is integrated in time, and we apply Young’s inequality for the negative
term:
Z
B(u, v) ≥ 1
α2
ku0 k2V ∗ − 2α1 2 ku0 k2V ∗ − α22 kA−T γLuk2V + γ hu, Lui + 12 kv0 k2H + 12 kv(T )k2H
Z
2
≥ 1
2α2
ku0 k2V ∗ − α22γ kA−T Lk2V →V kuk2V + γ hu, Lui + 12 ku(0)k2H
9.3. SPACE-TIME FORMULATION OF PARABOLIC EQUATIONS 181

α2 γ 2
We fix now γ sufficiently small such that 2
kA−T γLk2V →V kuk2V ≤ γ hu, Lui to obtain

B(u, v)  ku0 k2L2 (V ∗ ) + ku0 k2H > 0.

A similar proof of Lions’s theorem is found in Ern + Guermond.

9.3.2 A first time-discretization method

We discretize in time, but keep the spatial function space infinit dimensional. A first
reasonable attempt is to use Xh = P 1 (V ), and Yh = P 0,disc (V ). Evaluation of B(., .) leads
to
Xn Z tj
B(uh , vh ) = hu0h + Auh , vh i + (uh (0), vh (0))
j=1 tj−1

Xn
τj
= huj − uj−1 , vj i + 2
hA(uj−1 + uj ), vj i + (u0 , v0 )
j=1

Here, the time derivative evaluates to finite differences of point values in tj . Since uj ∈ V ,
the duality pairs coincide with inner products in H. Thus, for every time-step we get the
equation
τ
uj − uj−1 + A(uj + uj−1 ) = τ fj
2
This is the trapezoidal method (Crank-Nicolson). From numerics for odes we remember it
is A-stable, but not L-stable. We cannot prove a discrete inf − sup condition.

9.3.3 Discontinuous Galerkin method

We give an alternative, formally equivalent variational formulation for the parabolic equa-
tion by integration by parts in time
Z Z
0
− hu, v i + hAu, vi + (u(T ), v(T ))H − (u(0), v(0))H = hf, vi

Now, we plug in the given initial condition u(0) = u0 :

Z Z
0
− hu, v i + hAu, vi + (u(T ), v(T ))H = hf, vi + (u0 , v(0))H

The higher H 1 -regularity is now put onto the test-space, which validates point-evaluation
at t = 0 and t = T . The trial-space is now only L2 , which gives no meaning for u(T ).
There are two possible remedies, either to introduce a new variable for u(T ), or, to restrict
the test space:
182 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

1. Find u ∈ L2 (V ), uT ∈ H such that

Z Z
0
− hu, v i+hAu, vi+(uT , v(T ))H = hf, vi+(u0 , v(0))H ∀ v ∈ L2 (V ), v 0 ∈ L2 (V ∗ )
(9.7)

2. Find u ∈ L2 (V ) such that

Z Z
0
− hu, v i + hAu, vi = hf, vi + (u0 , v(0))H ∀ v ∈ L2 (V ), v 0 ∈ L2 (V ∗ ), v(T ) = 0
(9.8)

Both problems are well posed (continuity and inf − sup conditions, exercise). Now, the
initial condition was converted from an essential to a natural boundary condition.
Next, we integrate back, but we do not substitute the initial condition back:
Z Z
0
hu + Au, vi + (u(0), v(0))H = hf, vi + (u0 , v(0))

The initial condition is again a part of the variational formulation. Note that this for-
mulation is fulfilled for u ∈ H 1 , and smooth enough test functions providing the trace
v(0).
This technique to formulate initial conditions is used in the Discontinuous Galerkin
(DG) method. For every time-slab (tj−1 , tj ) we define a parabolic equation, where the
initial value is the end value of the previous time-slab.
Here, we first define a mesh T = {t0 , t1 , . . . tn }, and then the mesh-dependent formula-
tion:
Z tj Z tj
0 + +
hu + Au, vi+(u(tj−1 ), v(tj−1 ))H = hf, vi+(u(t− +
j−1 ), v(tj−1 ))H ∀ j ∈ {1, . . . n}
tj−1 tj−1

(with the notation u(t−0 ) := u0 ). By using left and right sided limits, we get the u from the
current time-slab, and the end-value from the previous time-slab, respectively. The vari-
ational formulation is valid for the solution u ∈ H 1 , and piece-wise regular test-functions
on the time-intervals.
The bilinear-form is defined as
X n Z tj
B(u, v) = hu0 + Au, vi + ([u]tj−1 , v(t+
j−1 ))H
j=1 tj−1

−
where the jump is defined as [u]tj = u(t+ +
j ) − u(tj ), and the special case [u]t0 = u(t0 ). The
solution satisfies
Z
B(u, v) = hf, vi + (u0 , v(0)) ∀ p.w. smooth v
9.3. SPACE-TIME FORMULATION OF PARABOLIC EQUATIONS 183

The bilinear-form is defined for discontinuous trial and discontinuous test functions. It
allows to define
Xh = Yh = P k,dc (V )
Let us elaborate the case of piece-wise constants in time:

τ hAuj , vj i + (uj − uj−1 , vj ) = τ hfj , vj i

which leads to the implicit Euler method

uj − uj−1
+ Auj = fj
τ
The imlicit Euler method is A and L-stable.
We define the mesh-dependent norms
X
kuk2Xh = kuk2L2 (tj−1 ,tj ,V ) + ku0 k2L2 (tj−1 ,tj ,V ∗ ) + 1
tj −tj−1
k[u]tj−1 k2V ∗
j
X
kvk2Yh = kvk2L2 (tj−1 ,tj ,V )
j

Since v|[tj−1 ,tj ] is a polynomial, we can bound

c
kv(t+ 2
j−1 )kV ≤ kvk2L2 (tj−1 ,tj ,V ) ,
tj − tj−1
where the constant c deteriors with the polynomial degree. Thus, the bilinear-form is well
defined and continuous on Xh × Yh .
Theorem 143. The discrete problem is inf − sup stable on Xh × Yh
Proof. We mimic the first inf − sup condition in Theorem 142, where we have set v =
u + A−T u0 . We give the proof for the lowest order case (k=0)

vh := uh + γ
tj −tj−1
A(tj−1 )−1 [uh ]j−1 ,

with γ = O(1) to be fixed later. Thanks to the discontinuous test-space, this is a valid
test-function.
In the following we skip the subsripts h, and set τ = tj − tj−1 . There holds

kvkYh  kukXh .

n Z
X tj X
B(uh , vh ) = Au, u + γτ A−T
j−1 [u]j−1 + ([u]j−1 , u + γτ A−1
j−1 [u]j−1 )H
j=1 tj−1 j
Z XZ X X
= hAu, ui + Au, γτ A−T [uj ] + (uj − uj−1 , uj )H + γ
τ
k[u]tj−1 kA−1
j j j
184 CHAPTER 9. PARABOLIC PARTIAL DIFFERENTIAL EQUATIONS

The second term is split by Young’s inequality:

Z Z Z
γ2
γ 1
tj Au, τ Aj−1 [u]j−1 ≤ 1
2
hAu, ui + 2τ 2
kA−1 [u]kA
tj−1

Thus, for sufficiently small γ it can be absorbed into the first and last term.
We reorder the summation of the third term:

(u1 , u1 )2 − (u1 , u2 ) + (u2 , u2 )2 − (u2 , u3 ) + . . .

= 21 ku1 k2H + 21 ku1 − u2 k2H + . . .

Thus, we got (for piecewise constants in time):

X 1
B(uh , vh ) ≥ kuk2L2 ,V + k[u]tj k2H + k[u]k2V ∗  kuh k2Xh
τ

By stability, we get for the discrete error

Bh (Ih u − uh , vh )
kIh u − uh kXh  sup
vh kvh kYh )
Bh (Ih u − u, vh )
= sup
vh kvh kY )
P R 0 h
hu + Au − (Ih u)0 + AIh u, vh i + j ([u] − [Ih u], vh )H
P
j
= sup
vh kvh kL2 (V )
 ...

where the convergence rate depends as usual on the regulariy of the exact solution
Chapter 10

Second order hyperbolic equations:

wave equations

We consider equations second order in time

ü + Au = f

with initial conditions

u(0) = u0 and u̇(0) = v0 ,
with a symmetric, elliptic operator A.

10.1 Examples
• scalar wave equation (acoustic waves)

∂ 2u
− ∆u = f
∂t2

• electromagnetic wave equation:

∂H
µ = − curl E
∂t
∂E
ε = curl H
∂t
with the magnetic field H and the electric field E, and material parameters perme-
ability µ and permittivity ε. By differentiating the first equation in space, and the
second one in time, we obtain

∂ 2E 1
ε 2
+ curl curl E = 0
∂t µ

185
186CHAPTER 10. SECOND ORDER HYPERBOLIC EQUATIONS: WAVE EQUATIONS

• elastic waves: We consider the hyperelastic elastic energy

Z
J(u) = W (C(u)) − f u
Ω

A body is in equilibrium, if J = 0. If not, then J 0 ∈ V ∗ acts as an accelerating force.

0

Newton’s law is
ρü = −J 0 (u),
in variational form Z
ρüv + hJ 0 (u), vi = 0 ∀v

In non-linear elasticity we have J 0 (u) = div P −f , where P is the first Piola-Kirchhoff

stress tensor. In linearized elasticity we obtain
Z Z Z
ρüv + Dε(u) : ε(v) = f v

We observe conservation of energy in the following sense for elasticity, and similar for
the other cases. We define the kinetic energy as 21 ku̇k2ρ and the potential energy as J(u).
Then
d 1
2
ku̇k2ρ + J(u) = (u̇, ü)ρ + hJ 0 (u), u̇i = 0
dt
For the linear equation set J(u) = 21 hAu, ui − hf, ui

10.2 Time-stepping methods for wave equations

We consider the method of lines, where we first discretize in space, and then apply some
time-stepping method for the ODE. In principal, one can reduce the second order ODE to a
first order system, and apply some Runge-Kutta method for it. This will in general require
the solution of linear systems of twice the size. In addition, the structure (symmetric and
positive definite) may be lost, which makes it difficult to solve.
We consider two approaches specially taylored for wave equations.
(a) for the second order equation
(b) for first order systems

10.2.1 The Newmark time-stepping method

We consider the ordinary differential equation

M ü + Ku = f

We consider single-step methods: From given state un ≈ u(tn ) and velocity u̇n ≈ u̇(tn ) we
compute un+1 and u̇n+1 . The acceleration ün = M −1 (fn − Kun ) follows from the equation.
10.2. TIME-STEPPING METHODS FOR WAVE EQUATIONS 187

The Newmark method is based on a Taylor expansion for u and u̇, where second order
derivatives are approximated from old and new accelerations. The real parameters β and
γ will be fixed later, τ is the time-step:

un+1 = un + τ u̇n + τ 2 ( 21 − β)ün + β ün+1

 
(10.1)
u̇n+1 = u̇n + τ [(1 − γ)ün + γ ün+1 ] (10.2)

Inserting the formula for un+1 into M ü + Ku = f at time tn+1 we obtain

M ün+1 + K un + τ u̇n + τ 2 ( 21 − β)ün + β ün+1 = fn+1

 

Now we keep unknows left and put known variables to the right:

M + βτ 2 K ün+1 = fn+1 − K un + τ u̇n + τ 2 ( 12 − β)ün

 

The Newmark method requires to solve one linear system with the spd matrix M + τ 2 βK,
for which efficient direct or iterative methods are available. After computing the new
acceleration, the new state un+1 and velocity u̇n+1 are computed from the explicit formulas
(10.1) and (10.2).
The Newmark method satisfies a discrete energy conservation. See [Steen Krenk: ”En-
ergy conservation in Newmark based time integration algorithms” in Compute methods in
applied mechanics and engineering, 2006, pp 6110-6124] for the calculations and various
extensions:
1 1 T
n+1
2
u̇M u̇ + 2
u K eq u n
= −(γ − 21 )(un+1 − un )Keq (un+1 − un )

where
Keq = K + (β − 12 γ)τ 2 KM −1 K,
and the notation [E]ba := E(b) − E(a). Here, the right hand side f is skipped. From this,
we get the conservation of a modified energy with the so called equivalent stiffness matrix
Keq . Depending on the parameter γ we get
• γ = 21 : conservation
• γ > 12 : damping
• γ < 12 : growth of energy (unstable)
If Keq is positive definite, then this conservation proves stability. This is unconditionally
true if β ≥ 21 γ (the method is called unconditionally stable). If β < 12 γ, the allowed time
step is limited by
1
τ 2 ≤ λmax (M −1 K)−1 1 .
2
γ−β
For second order problems we have λmax (M −1 K) ' h−2 , and thus τ  h which is a
reasonable choice also for accuracy.
Choices for β and γ of particular interests are:
188CHAPTER 10. SECOND ORDER HYPERBOLIC EQUATIONS: WAVE EQUATIONS

• γ = 12 , β = 41 : unconditionally stable, conservation of original energy (Keq = K)

• γ = 12 , β = 0: conditionally stable. We have to solve

τ2
M ün+1 = fn+1 − K(un + τn u̇n + 2 n
ü )

which is explicit iff M is cheaply invertible (mass lumping, DG).

10.2.2 Methods for the first order system

We reduce the wave equation
ü − ∆u = f
Rt
to a first order system of pdes. We introduce σ = 0 ∇u. Then

σ̇ = ∇u
u̇ − div σ = f˜
Rt
with the integrated source f˜ = 0 f . In the following we skip the source f .
A mixed variational formulation in H(div) × L2 , for given initial conditions u(0) and
σ(0), is:

(σ̇, τ ) = −(u, div τ ) ∀τ

(u̇, v) = (v, div σ) ∀v

After space discretization we obtain the ode system

     
Mσ 0 σ̇ 0 −B T σ
=
0 Mu u̇ B 0 u

We get a similar struture from the Maxwell system:

(µḢ, H̃) = (curl E, H̃) ∀ H̃

(εĖ, Ẽ) = −(curl Ẽ, H) ∀ Ẽ

Conservation of energy is now seen from

d 1 T
σ Mσ σ + 12 uT Mu u = σ T Mσ σ̇ + uT Mu u̇ = −σ T B T u + uT Bσ = 0

2
dt
A basis transformation with M 1/2 leads to the transformed system (the transformed B
is called B again):     
σ̇ 0 −B T σ
=
u̇ B 0 u
The matrix is skew-symmetric, and thus the eigenvalues are imaginary. They are contained
in i [−ρ(B), ρ(B)], where the spectral radious ρ(B) ' h−1 for the first order operator. Using
10.2. TIME-STEPPING METHODS FOR WAVE EQUATIONS 189

Runge-Kutta methods, we need methods such that i [−τ ρ(B), τ ρ(B)] is in the stability
region. For large systems, explicit methods (M cheaply invertible !) are often preferred.
While the stability region for the explicit Euler and improved Euler method do not include
an interval on the imaginary axis, the RK4 method does.
Methods taylored for the skew-symmetric (Hamiltonian) structure are symplectic meth-
ods: The symplectic Euler method is
σn+1 − σn
Mσ = −B T un
τ
un+1 − un
Mu = Bσn+1
τ
For updating the second variable, the new value of the first variable is used. For the
analysis, we can reduce the large system to 2 × 2 systems, where β are singular values of
−1/2 −1/2
Mσ BMu :
σ̇ = −βu u̇ = βσ
The symplectic Euler method can be written as
     
σn+1 1 0 1 −τ β σn
=
un+1 τβ 1 0 1 un
|  {z 
}
1 −τ β
T=  
τ β 1 − (τ β)2
√
The eigenvalues of T satisfy λ1 λ2 = det(T ) = 1, and iff τ β ≤ 2 they are conjugate
complex, and thus |λ1 | = |λ2 | = 1. Thus, the discrete solution is oscillating without
damping or growth.
Again, diagonal mass matrices Mu and Mσ render explicit methods efficient.
190CHAPTER 10. SECOND ORDER HYPERBOLIC EQUATIONS: WAVE EQUATIONS
Chapter 11

Hyperbolic Conservation Laws

We consider the equation

∂u
+ div f (u) = 0
∂t
in space dimension n, with the state u ∈ Rm , and the flux f : Rm → Rm×n . We need initial
conditions u(x, 0) = u0 (x), and proper boundary conditions.
Examples:
• Transport equation m = 1, n ∈ {1, 2, 3, . . .}.
f (u) = bT u
with b ∈ Rn the given wind.
• Burgers’ equation m = 1, n = 1
f (u) = 21 u2
Burgers’ equation is a typical model problem to study effects of non-linear conserva-
tion laws.
• Wave equation in Rn : u = (p, v1 , . . . vn ), here m = n + 1:
 
u1 · · · un
 p 
f (p, u) = 
 
 ... 

p

• Euler equations (model for compressible flows) in Rn : m = n + 2, state u =

(ρ, m1 , . . . , mn , E), with density ρ, momentum m = ρv, and energy. The flux function
is  
ρv
f =  ρv ⊗ v + pI 
(E + p)v
with the internal energy e = E/ρ − 12 |v|2 (proportional to temperature), and state
equation p = p(ρ, e). Equations are conservation of mass, momentum and energy.

191
192 CHAPTER 11. HYPERBOLIC CONSERVATION LAWS

11.1 A little theory

Set n = 1, m = 1. We assume that f is convex, i.e. f 0 is strictly monotone increasing. For
linear fluxes f = bu, the solution is the traveling wave
u(x, t) = u0 (x − bt).
It is constant along the characteristic lines x(t) = x0 + bt.
For smooth fluxes f , the solution is constant along characteristic lines x(t) = x0 +
f 0 (u0 (x0 ))t:
u(x(t), t) = u0 (x0 )
proof:
d ∂u dx ∂u ∂u ∂u
0= u(x(t), t) = + = + f 0 (u) = ft + (f (u))x
dt ∂t dt ∂x ∂t ∂x
The smooth solution exists as long as characteristic lines don’t intersect.
Example: Burgers equation. The velocity of the characteristic is f 0 (u) = ( 12 u2 )0 = u,
i.e. the solution itself.

11.1.1 Weak solutions and the Rankine-Hugoniot relation

If characteristic lines intersect, the solution forms a shock. the Rankine-Hugoniot relation
is a equation for the speed of the shock.
We assume the solution is piecewise smooth. To have meaningful discontinuous solu-
tions, we have to consider weak solutions in space-time:
Z
uϕt + f (u)∇ϕ = 0 ∀ϕ ∈ C0∞
Ω×(0,T )

(initial condition is skipped here, easily covered by non-vanishing test-functions for t = 0).
The weak form states that for F = (f, u) there holds
divx,t F = 0
in weak senses. Thus, F ∈ H(div). This requires that F · n is continuous across disconti-
nuities. Let s(t) the position of the shock. The normal vector satisfies
n ∼ (1, −s0 )
Thus, [F · n] = 0 reads as
f (ul ) − ul s0 = f (ur ) − ur s0
and we get the Rankine-Hugoniot relation
f (ul ) − f (ur )
s0 =
ul − ur
Example Burgers: The speed of the shocks is
1 2
u
2 l
− 12 u2r ul + ur
s0 = =
ul − ur 2
11.1. A LITTLE THEORY 193

11.1.2 Expansion fans

Assume u(x+) > u(x−), then, due to convexity of the flux there is also f 0 (u(x+)) >
f 0 (u(x−)). The speed on the right is higher than on the left. Here, all monotone increasing
functions (between u(x−) and u(x+)), constant along lines x + f 0 (u)t are weak solutions.
Another conditions is necessary to pick the meaningful physical solution. Two choices
are
Viscosity solutions. Consider the regularized equation

uεt + f (uε )x − εuεxx = 0

The limit (if existent) limε→0 uε is called viscosity solution.

Entropy solutions. We define some quantity E(u) called entropy, where, for physical
reasons, the total amount should not increase:
Z
d
E(u) ≤ 0
dt Ω
To localize it, we define the entropy flux F such that

F 0 = E 0f 0

If the solution is smooth, then

E(u)t + F (u)x = E 0 ut + F 0 ux = E 0 (ut + f 0 ux ) = 0

Thus Z Z Z Z
d
E(u) = E(u)t = − F (u)x = − F (u) · n
dt Ω Ω Ω ∂Ω
No entropy changes for smooth solutions with isolated boundary. But, this is not true for
discontinuous solutions.
We pose the entropy decrease E(u)t + F (u)x ≤ 0 in weak sense:
Z
− E(u)ϕt + F (u)∇ϕ ≤ 0 ∀ ϕ ∈ C0∞ , ϕ ≥ 0
Ω×(0,T )

Similar to the Rankine-Hugoniot relation we integrate back on smooth regions in space-

time
X Z Z
[E(u)]s0 − [F (u)] ϕ ≤ 0



E(u)t + div F (u) ϕ − ∀ϕ ≥ 0
(Ω×(0,T ))i
| {z } γ
=0

Example Burgers: Choose the entropy E(u) = u2 . Then

Z Z
0 0 2
F (u) = E f = 2uu = u3
3
194 CHAPTER 11. HYPERBOLIC CONSERVATION LAWS

Calculating
ur + ul 2 3
[E]s0 − [F ] = (u2r − u2l ) + (ur − u3l ) = . . .
2 3
2
(ur − ul )
= −(ur − ul )
6
Now, posing the non-negative condition for [E]s0 − [F ] we allow jumps only for ur < ul .

11.2 Numerical Methods

The natural methods for conservation laws are finite volume / discontinuous Galerkin
methods: Z Z
∂u
v − f (u)∇v + g(ul , ur )v = 0 ∀ T ∀ v ∈ P k (T )
T ∂t ∂T
Here, g is the numerical flux on the element boundary, calculated from left- and right sided
states. For continuous u = ul = ur , it satisfies

g(u, u) = f (u)n

Otherwise, up-wind like fluxes (many different choices !) are used.

Finite volume methods can be designed such that entropy is non-increasing, often the
calculations are technical. They are also used to prove the existence of solutions.
Higher order methods do not satisfy the maximum principle, which may lead to prob-
lems for non-linear equations (Euler: divide by ρ). Here, limiters are used: If the solution
produces oscillations, it is smoothed (somehow). E.g., switch back to a finite volume
method.
Recent development (by Guermond+Pasquetti+Popov) is the so-called entropy viscos-
ity method: If the entropy relation is violated, artificial viscosity is switched on. Ideally,
this happens only close to shocks.
Space-time methods include special mesh-generation related to the finite speed of prop-
agation (front tracking methods, tent-pitching methods). [Gopalakrishnan, Schöberl, Win-
tersteiger 2016, master thesis Wintersteiger].