Numerical Methods for Partial Differential

Equations

Volker John

Summer Semester 2013

Contents

1 Some Partial Differential Equations From Physics 3

1.1 The Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The Diffusion Equation . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 The Navier–Stokes Equations . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Classification of Second Order Partial Differential Equations . . . . . 10
1.5 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 Finite Difference Methods for Elliptic Equations 13

2.1 Basics on Finite Differences . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Finite Difference Approximation of the Laplacian in Two Dimensions 15
2.3 The Discrete Maximum Principle for a Finite Difference Operator . 17
2.4 Stability and Convergence of the Finite Difference Approximation of
the Poisson Problem with Dirichlet Boundary Conditions . . . . . . 23
2.5 An Efficient Solver for the Dirichlet Problem in the Rectangle . . . . 25
2.6 A Higher Order Discretizations . . . . . . . . . . . . . . . . . . . . . 29
2.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3 Introduction to Sobolev Spaces 36

3.1 Elementary Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2 Weak Derivative and Distributions . . . . . . . . . . . . . . . . . . . 38
3.3 Lebesgue Spaces and Sobolev Spaces . . . . . . . . . . . . . . . . . . 40
3.4 The Trace of a Function from a Sobolev Space . . . . . . . . . . . . 42
3.5 Sobolev Spaces with Non-Integer and Negative Exponents . . . . . . 44
3.6 Theorem on Equivalent Norms . . . . . . . . . . . . . . . . . . . . . 45
3.7 Some Inequalities in Sobolev Spaces . . . . . . . . . . . . . . . . . . 46
3.8 The Gaussian Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.9 Sobolev Imbedding Theorems . . . . . . . . . . . . . . . . . . . . . . 49

4 The Ritz Method and the Galerkin Method 52

4.1 The Theorems of Riesz and Lax–Milgram . . . . . . . . . . . . . . . 52
4.2 Weak Formulation of Boundary Value Problems . . . . . . . . . . . . 54
4.3 The Ritz Method and the Galerkin Method . . . . . . . . . . . . . . 56

5 Finite Element Methods 60

5.1 Finite Element Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2 Finite Elements on Simplices . . . . . . . . . . . . . . . . . . . . . . 63
5.3 Finite Elements on Parallelepipeds . . . . . . . . . . . . . . . . . . . 69
5.4 Parametric Finite Elements on General d-Dimensional Quadrilaterals 73
5.5 Transform of Integrals . . . . . . . . . . . . . . . . . . . . . . . . . . 73

1
6 Interpolation 76
6.1 Interpolation in Sobolev Spaces by Polynomials . . . . . . . . . . . . 76
6.2 Inverse Estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.3 Interpolation of Non-Smooth Functions . . . . . . . . . . . . . . . . 85

7 Finite Element Methods for Second Order Elliptic Equations 88

7.1 General Convergence Theorems . . . . . . . . . . . . . . . . . . . . . 88
7.2 Linear Finite Element Method on Non-Polyhedral Domains . . . . . 90
7.3 Finite Element Method with the Nonconforming Crouzeix–Raviart
Element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7.4 L2 (Ω) Error Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 98

8 Outlook 101

2
Chapter 1

Some Partial Differential

Equations From Physics

Remark 1.1 Contents. This chapter introduces some partial differential equations
(pde’s) from physics to show the importance of this kind of equations and to moti-
vate the application of numerical methods for their solution. 2

1.1 The Heat Equation

Remark 1.2 Derivation. The derivation of the heat equation follows (Wladimirow,
1972, p. 39). Let x = (x1 , x2 , x3 )T ∈ Ω ⊂ R3 , where Ω is a domain, t ∈ R, and
consider the following physical quantities
• u(t, x) – temperature at time t and at the point x with unit [K],
• ρ(t, x) – density of the considered species with unit [kg/m3 ],
• c(t, x) – specific heat capacity of the species with unit [J/(kg K)] = [W s/(kg K)],
• k(t, x) – thermal conductivity of the species with unit [W/(m K)],
• F (t, x) – intensity of heat sources or sinks with unit [W/m3 ].
Consider the heat equilibrium in an arbitrary volume V ⊂ Ω and in an arbitrary
time interval (t, t + ∆t). First, there are sources or sinks of heat: heat can enter or
leave V through the boundary ∂V , or heat can be produced or absorbed in V . Let
n(x) be the unit outer normal at x ∈ ∂V . Due to Fourier’s law , one finds that the
heat
Z t+∆t Z Z t+∆t Z
∂u
Q1 = k (t, s) ds dt = (k∇u · n) (t, s) ds dt, [J],
t ∂V ∂n t ∂V

enters through ∂V into V . One obtains with partial integration (Gaussian theorem)
Z t+∆t Z
Q1 = ∇ · (k∇u)(t, x) dx dt.
t V

In addition, the heat

Z t+∆t Z
Q2 = F (t, x) dx dt, [W s] = [J],
t V

is produced in V .
Second, a law for the change of the temperature in V has to be derived. Using
a Taylor expansion, on gets that the temperature at x changes in (t, t + ∆t) by
∂u
u(t + ∆t, x) − u(t, x) = (t, x)∆t + O((∆t)2 ).
∂t

3
Now, a linear ansatz is utilized, i.e.,
∂u
u(t + ∆t, x) − u(t, x) = (t, x)∆t.
∂t
With this ansatz, one has that for the change of the temperature in V and for
arbitrary ∆t, the heat
Z t+∆t Z Z t+∆t Z
u(t + ∆t, x) − u(t, x) ∂u
Q3 = cρ dx dt = cρ (t, x) dx dt
t V ∆t t V ∂t
is needed. This heat has to be equal to the heat sources, i.e., it holds Q3 = Q2 + Q1 ,
from what follows that
Z t+∆t Z  
∂u
cρ − ∇ · (k∇u) − F (t, x) dx dt = 0.
t V ∂t
Since the volume V was chosen to be arbitrary and ∆t was arbitrary as well, the
term in the integral has to vanish. One obtains the so-called heat equation
∂u
cρ − ∇ · (k∇u) = F in (0, T ) × Ω.
∂t
At this point of modeling one should check if the equation is dimensionally correct.
One finds that all terms have the unit [W/m3 ].
For a homogeneous species, c, ρ, and k are positive constants. Then, the heat
equation simplifies to
∂u
− ε2 ∆u = f in (0, T ) × Ω, (1.1)
∂t
with ε2 = k/(cρ), [m2 /s] and f = F/(cρ), [K/s]. To obtain a well-posed problem,
(1.1) has to be equipped with an initial condition u(0, x) and appropriate boundary
conditions on (0, T )∂Ω. 2
Remark 1.3 Boundary conditions. For the theory and the numerical simulation
of partial differential equations, the choice of boundary conditions is of utmost
importance. For the heat equation (1.1), one can prescribe the following types of
boundary conditions:
• Dirichlet1 condition: The temperature u(t, x) at a part of the boundary is pre-
scribed
u = g1 on (0, T ) × ∂ΩD
with ∂ΩD ⊂ ∂Ω. In the context of the heat equation, the Dirichlet condition is
also called essential boundary conditions.
• Neumann2 condition: The heat flux is prescribed at a part of the boundary
∂u
−k = g2 on (0, T ) × ∂ΩN
∂n
with ∂ΩN ⊂ ∂Ω. This boundary condition is a so-called natural boundary
condition for the heat equation.
• Mixed boundary condition, Robin3 boundary condition: At the boundary, there
is a heat exchange according to Newton’s law
∂u
k + h(u − uenv ) = 0 on (0, T ) × ∂Ωm ,
∂n
with ∂Ωm ⊂ ∂Ω, the heat exchange coefficient h, [W/(m2 K 2 )], and the temper-
ature of the environment uenv .
1 Johann Peter Gustav Lejeune Dirichlet (1805 –1859)
2 CarlGottfried Neumann (1832 – 1925)
3 Gustave Robin (1855 – 1897)

4
 2

Remark 1.4 The stationary case. An important special case is that the temper-
ature is constant in time u(t, x) = u(x). Then, one obtains the stationary heat
equation
− ε2 ∆u = f in Ω. (1.2)
This equation is called Poisson4 equation. Its homogeneous form, i.e., with f (x) = 0,
is called Laplace5 equation. Solution of the Laplace equation are called harmonic
functions. The Poisson equation is the simplest partial differential equation. The
most part of this lecture will consider numerical methods for solving this equation.
2

Remark 1.5 Another application of the Poisson equation. The stationary distri-
bution of an electric field with charge distribution f (x) satisfies also the Poisson
equation (1.2). 2

Remark 1.6 Non-dimensional equations. The application of numerical methods

relies on equations for functions without physical units, the so-called non-dimensional
equations. Let
• L – a characteristic length scale of the problem, [m],
• U – a characteristic temperature scale of the problem, [K],
• T ∗ – a characteristic time scale of the problem, [s].
If the new coordinates and functions are denoted with a prime, one gets with the
transformations
x u t
x0 = , u0 = , t0 = ∗
L U T
from (1.1) the non-dimensional equations

0 d  0
 0  
∂ 0 ∂t
X ∂ ∂ 0 ∂xi ∂xi T
(U u ) − ε2
0 0 (U u ) = f in 0, × Ω0 ⇐⇒
∂t0 ∂t i=1
∂x i ∂x i ∂xi ∂xi T∗
d
U ∂u0 ε2 U X ∂ 2 u0
 
T
− = f in 0, × Ω0 .
T ∗ ∂t0 L2 i=1 ∂ (x0i )2 T∗

Usually, one denotes the non-dimensional functions like the dimensional functions,
leading to
∂u ε2 T ∗ T∗
 
T
− ∆u = f in 0, × Ω.
∂t L2 U T∗
For the analysis, one sets L = 1m, U = 1K, and T ∗ = 1s which yields
∂u
− ε2 ∆u = f in (0, T ) × Ω, (1.3)
∂t
with a non-dimensional temperature diffusion ε2 and a non-dimensional right hand
side f (t, x).
The same approach can be applied to the stationary equation (1.2) and one gets

− ε2 ∆u = f in Ω, (1.4)

with the non-dimensional temperature diffusion ε2 and the non-dimensional right

hand side f (x). 2
4 Siméon Denis Poisson (1781 – 1840)
5 Pierre Simon Laplace (1749 – 1829)

5
Remark 1.7 A standard approach for solving the instationary equation. The heat
equation (1.3) is an initial value problem with respect to time and a boundary value
problem with respect to space. Numerical methods for solving initial value problems
were topic of Numerical Mathematics 2.
A standard approach for solving the instationary problem consists in using a
so-called one-step θ-scheme for discretizing the temporal derivative. Consider two
consecutive discrete times tn and tn+1 with τ = tn+1 − tn . Then, the application
of a one-step θ-scheme yields for the solution at tn+1
un+1 − un
− θε2 ∆un+1 − (1 − θ)ε2 ∆un = θfn+1 + (1 − θ)fn ,
τ
where the subscript at the functions denotes the time level. This equation is equiv-
alent to

un+1 − τ θε2 ∆un+1 = un + τ (1 − θ)ε2 ∆un + τ θfn+1 + τ (1 − θ)fn . (1.5)

For θ = 0, one obtains the forward Euler scheme, for θ = 0.5 the Crank–Nicolson
scheme (trapezoidal rule), and for θ = 1 the backward Euler scheme.
Given un , (1.5) is a boundary value problem for un+1 . That means, one has to
solve in each discrete time a boundary value problem. For this reason, this lecture
will concentrate on the numerical solution of boundary value problems. 2

Example 1.8 Demonstrations with the code MooNMD John and Matthies
(2004).
• Consider the Poisson equation (1.4) in Ω = (0, 1)2 with ε = 1. The right
hand side and the Dirichlet boundary conditions are chosen such that u(x, y) =
sin(πx) sin(πy) is the prescribed solution, see Figure 1.1 Hence, this solution
satisfies homogeneous Dirichlet boundary conditions. Denote by uh (x, y) the
computed solution, where h indicates the refinement of a mesh in Ω. The errors
obtained on successively refined meshes with the simplest finite element method
are presented in Table 1.1.

Figure 1.1: Solution of the two-dimensional example of Example 1.8.

One can observe in Table 1.1 that ku − uh kL2 (Ω) converges with second order
and k∇(u − uh )kL2 (Ω) converges with first order. A main topic of the numerical
analysis of discretizations for partial differential equations consists in showing
that the computed solution converges to the solution of an appropriate contin-
uous problem in appropriate norms. In addition, to prove a certain order of
convergence (in the asymptotic regime) is of interest.

6
 Table 1.1: Example 1.8, two-dimensional example.
h degrees of freedom ku − uh kL2 (Ω) k∇(u − uh )kL2 (Ω)
1/4 25 8.522e-2 8.391e-1
1/8 81 2.256e-2 4.318e-1
1/16 289 5.726e-3 2.175e-1
1/32 1089 1.437e-3 1.089e-1
1/64 4225 3.596e-4 5.451e-2
1/128 16641 8.993e-5 2.726e-2
1/256 66049 2.248e-5 1.363e-2
1/512 263169 5.621e-6 6.815e-3

• Consider the Poisson equation (1.4) in Ω = (0, 1)3 with ε = 1 and f = 0. At

z = 1 the temperature profile should be u(x, y, 1) = 16x(1 − x)y(1 − y) and
at the opposite wall should be cooled u(x, y, 0) = 0. At all other walls, there
∂u
should be an undisturbed temperature flux ∂n (x, y, z) = 0. A approximation of
the solution computed with a finite element method is presented in Figure 1.2.

Figure 1.2: Contour lines of the solution of the three-dimensional example of Ex-
ample 1.8.

The analytical solution is not known in this example (or it maybe hard to com-
pute). It is important for applications that one obtains, e.g., good visualizations
of the solution or approximate values for quantities of interest. One knows by
the general theory that the computed solution converges to the solution of the
continuous problem in appropriate norms and one hopes that the computed
solution is already sufficiently close.
2

1.2 The Diffusion Equation

Remark 1.9 Derivation. Diffusion is the transport of a species caused by the
movement of particles. Instead of Fourier’s law, Newton’s law for the particle flux
through ∂V per time unit is used

dQ = −D∇u · n ds

with
• u(t, x) – particle density, concentration with unit [mol/m3 ],
• D(t, x) – diffusion coefficient with unit [m2 /s].

7
The derivation of the diffusion equation proceeds in the same way as for the heat
equation. It has the form
∂u
c − ∇ · (D∇u) + qu = F in (0, T ) × Ω, (1.6)
∂t
where
• c(t, x) – is the porosity of the species, [·],
• q(t, x) – is the absorption coefficient of the species with unit [1/s],
• F (t, x) – describes sources and sinks, [mol/(s m3 )].
The porosity and the absorption coefficient are positive functions. To obtain a well
posed problem, an initial condition and boundary conditions are necessary.
If the concentration is constant in time, one obtains

− ∇ · (D∇u) + qu = F in Ω. (1.7)

Hence, the diffusion equation possesses a similar form as the heat equation. 2

1.3 The Navier–Stokes Equations

Remark 1.10 Generalities. The Navier6 –Stokes7 equations are the fundamental
equations of fluid dynamics. In this section, a viscous fluid (with internal friction)
with constant density (incompressible) will be considered. 2

Remark 1.11 Conservation of mass. The first basic principle of the flow of an
incompressible fluid is the conservation of mass. Let V be an arbitrary volume.
Then, the change of fluid in V satisfies
Z Z Z
∂
− ρ dx = ρv · n ds = ∇ · (ρv) dx,
∂t
V ∂V V
| {z } | {z }
change flux through the boundary of V

where
• v(t, x) – velocity (v1 , v2 , v3 )T at time t and at point x with unit [m/s],
• ρ – density of the fluid, [kg/m3 ].
Since V is arbitrary, the terms in the volume integrals have to be the same. One
gets the so-called continuity equation

ρt + ∇ · (ρv) = 0 in (0, T ) × Ω.

Since ρ is constant, one obtains the first equation of the Navier–Stokes equation,
the so-called incompressibility constraint,

∇ · v = 0 in (0, T ) × Ω. (1.8)

Remark 1.12 Conservation of linear momentum. The second equation of the

Navier–Stokes equations represents Newton’s second law of motion

net force = mass × acceleration.

It states that the rate of change of the linear momentum must be equal to the net
force acting on a collection of fluid particles.
6 Claude Louis Marie Henri Navier (1785 - 1836)
7 George Gabriel Stokes (1819 - 1903)

8
 The forces acting on an arbitrary volume V are given by
Z Z Z
FV = −P n ds + S0 n ds + ρg dx ,
∂V ∂V V
| {z } | {z } | {z }
outer pressure friction gravitation

where
• S 0 (t, x) – stress tensor with unit [N/m2 ],
• P (t, x) – the pressure with unit [N/m2 ],
• g(t, x) – standard gravity (directed), [m/s2 ].
The pressure possesses a negative sign since it is directed into V , whereas the stress
acts outwardly.
The integral on ∂V can be transformed into an integral on V with integration
by parts. One obtains the force per unit volume

−∇P + ∇ · S0 + ρg.

On the basis of physical considerations (Landau and Lifschitz, 1966, p. 53), one
uses the following ansatz for the stress tensor
 2 
S0 = η ∇v + ∇vT − (∇ · v)I + ζ(∇ · v)I,
3
where
• η – first order viscosity of the fluid, [kg/(m s)],
• ζ – second order viscosity of the fluid, [kg/(m s)],
• I – unit tensor.
For Newton’s second law of motion one considers the movement of particles with
velocity v(t, x(t)). One obtains the following equation
dv(t, x(t))
−∇P + ∇ · S0 + ρg = ρ
| {z } |{z} | {zdt }
force per unit volume mass per unit volume
acceleration
= ρ (vt + (v · ∇)v) .

The second formula was obtained with the chain rule. The detailed form of the
second term is
 
v1 (v1 )x + v2 (v1 )y + v3 (v1 )z
(v · ∇)v =  v1 (v2 )x + v2 (v2 )y + v3 (v2 )z  .
v1 (v3 )x + v2 (v3 )y + v3 (v3 )z

If both viscosities are constant, one gets

∂v ∇P 1 η 
− ν∆v + (v · ∇)v − =g+ + ζ ∇(∇ · v),
∂t ρ ρ 3
where ν = η/ρ, [m2 /s] is the kinematic viscosity. The second term on the right
hand side vanishes because of the incompressibility constraint (1.8).
One obtains the dimensional Navier–Stokes equations
∂v ∇P
− ν∆v + (v · ∇)v − = g, ∇ · v = 0 in (0, T ) × Ω.
∂t ρ
2

Remark 1.13 Non-dimensional Navier–Stokes equations. The final step in the

modeling process is the derivation of non-dimensional equations. Let

9
 • L – a characteristic length scale of the problem, [m],
• U – a characteristic velocity scale of the problem, [m/s],
• T ∗ – a characteristic time scale of the problem, [s].
Denoting here the old coordinates with a prime, one obtains with the transforma-
tions
x0 v t0
x= , u= , t= ∗
L U T
the non-dimensional equations
L ν
∗
∂t u − ∆u + (u · ∇)u + ∇p = f , ∇ · u = 0 in (0, T ) × Ω,
UT UL
with the redefined pressure and the new right hand side
P Lg
p(t, x) = (t, x), f (t, x) = (t, x).
ρU 2 U2

This equation has two dimensionless characteristic parameters: the Strouhal8 num-
ber St and the Reynolds 9 number Re
L UL
St := , Re := .
UT∗ ν
Setting T ∗ = L/U , one obtains the form of the incompressible Navier–Stokes equa-
tions which can be found in the literature
∂u
− Re−1 ∆u + (u · ∇)u + ∇p = f in (0, T ) × Ω,
∂t
∇·u = 0 in [0, T ) × Ω.
2

Remark 1.14 About the incompressible Navier–Stokes equations. The Navier–

Stokes equations are not yet understood completely. For instance, the existence
of an appropriately defined classical solution for Ω ⊂ R3 is not clear. This problem
is among the so-called millennium problems of mathematics Fefferman (2000) and
its answer is worth one million dollar. Also the numerical methods for solving the
Navier–Stokes equations are by far not developed sufficiently well as it is required
by many applications, e.g. for turbulent flows in weather prediction. 2

Remark 1.15 Slow flows. Am important special case is the case of slow flows
which lead to a stationary (independent of time) flow field. In this case, the first
term in the in the momentum balance equation vanish. In addition, if the flow
is very slow, the nonlinear term can be neglected. One gets the so-called Stokes
equations
−Re−1 ∆u + ∇p = f in Ω,
∇ · u = 0 in Ω.
2

1.4 Classification of Second Order Partial Differ-

ential Equations
Definition 1.16 Quasi-linear and linear second order partial differential
equation. Let Ω ⊂ Rd , d ∈ N. A quasi-linear second order partial differential
8 Čeněk Strouhal (1850 – 1923)
9 Osborne Reynolds (1842 - 1912)

10
equation defined on Ω has the form
d
X
ajk (x)∂j ∂k u + F (x, u, ∂1 u, . . . , ∂d u) = 0 (1.9)
j,k=1

or in nabla notation

∇ · A(x)∇u + F̃ (x, u, ∂1 u, . . . , ∂d u) = 0.

A linear second order partial differential equation has the form

d
X
ajk (x)∂j ∂k u + b(x) · ∇u + c(x)u = F (x).
j,k=1

Remark 1.17 The matrix of the second order operator. If u(x) is sufficiently reg-
ular, then the application of the Schwarz’10 theorem yields ∂j ∂k u(x) = ∂k ∂j u(x). It
follows that equation (1.9) contains the coefficient ∂j ∂k u(x) twice, namely in ajk (x)
and akj (x). For definiteness, one requires that

ajk (x) = akj (x).

Now, one can write the coefficient of the second order derivative with the symmetric
matrix  
a11 (x) · · · a1d (x)
A(x) =  ... .. ..  .

. . 
ad1 (x) · · · add (x)
All eigenvalues of this matrix are real and the classification of quasi-linear second
order partial differential equations is based on these eigenvalues. 2

Definition 1.18 Classification of quasi-linear second order partial differ-

ential equation. On a subset Ω̃ ⊂ Ω let α be the number of positive eigenvalues
of A(x), β be the number of negative eigenvalues, and γ be the multiplicity of the
eigenvalue zero. The quasi-linear second order partial differential equation (1.9) is
said to be of type (α, β, γ) on Ω̃. It is called to be
• elliptic on Ω̃ if it is of type (d, 0, 0) = (0, d, 0),
• hyperbolic on Ω̃, if its type is (d − 1, 1, 0) = (1, d − 1, 0),
• parabolic on Ω̃, if it is of type (d − 1, 0, 1) = (0, d − 1, 1).
In the case of linear partial differential equations, one speaks of a parabolic equation
if in addition to the requirement from above it holds that

rank(A(x), b(x)) = d

in Ω̃. 2

Remark 1.19 Other cases. Definition 1.18 does not cover all possible cases. How-
ever, the other cases are only of little interest in practice. 2

Example 1.20 Types of second order partial differential equations.

• For the Poisson equation (1.4) one has aii = −ε2 < 0 and aij = 0 for i 6= j.
It follows that all eigenvalues of A are negative and the Poisson equation is an
elliptic partial differential equation. The same reasoning can be applied to the
stationary diffusion equation (1.7).
10 Hermann Amandus Schwarz (1843 – 1921)

11
 • In the heat equation (1.3) there is besides the spatial derivatives also the tem-
poral derivative. The derivative in time has to be taken into account in the defi-
nition of the matrix A. Since this derivative is only of first order, one obtains in
A a zero row and a zero column. One has, e.g., aii = −ε2 < 0, i = 2, . . . , d + 1,
a11 = 0, and aij = 0 for i 6= j. It follows that one eigenvalue is zero and
the others have the same sign. The vector of the first order term has the form
b = (1, 0, . . . , 0)T ∈ Rd+1 , where the one comes from the ∂t u(t, x). Now, one
can see immediately that (A, b) possesses full column rank. Hence, (1.3) is a
parabolic partial differential equation.
• An example for a hyperbolic partial differential equation is the wave equation

∂t2 u − ε2 ∆u = f in (0, T ) × Ω.

1.5 Literature
Remark 1.21 Some books about the topic of this class. Books about finite differ-
ence methods are
• Samarskij (1984), classic book, the English version is Samarskii (2001)
• LeVeque (2007)
Much more books can be found about finite element methods
• Ciarlet (2002), classic text,
• Strang and Fix (2008), classic text,
• Braess (2001), very popular book in Germany,
• Brenner and Scott (2008), rather abstract treatment, from the point of view of
functional analysis,
• Ern and Guermond (2004), modern comprehensive book,
• Grossmann and Roos (2007)
• Šolı́n (2006), written by somebody who worked a lot in the implementation of
the methods,
• Goering et al. (2010), introductory text, good for beginners,
• Deuflhard and Weiser (2012), strong emphasis on adaptive methods
• Dziuk (2010).
These lists are not complete.
These lectures notes are based in some parts on lecture notes from Sergej
Rjasanow (Saarbrücken) and Manfred Dobrowolski (Würzburg). 2

12
Chapter 2

Finite Difference Methods

for Elliptic Equations

Remark 2.1 Model problem. The model problem in this chapter is the Poisson
equation with Dirichlet boundary conditions

−∆u = f in Ω,
(2.1)
u = g on ∂Ω,

where Ω ⊂ R2 . This chapter follows in wide parts Samarskij (1984). 2

2.1 Basics on Finite Differences

Remark 2.2 Grid. This section considers the one-dimensional case. Consider the
interval [0, 1] which is decomposed by an equidistant grid

xi = ih, i = 0, . . . , n, h = 1/n, – nodes,

ωh = {xi : i = 0, . . . , n} – grid.

Definition 2.3 Grid function. A vector uh = (u0 , . . . , un )T ∈ Rn+1 which as-

signs every grid point a function value is called grid function. 2

Definition 2.4 Finite differences. Let v(x) be a sufficiently smooth function and
denote by vi = v(xi ), where xi are the nodes of the grid. The following quotients
are called
vi+1 − vi
vx,i = – forward difference,
h
vi − vi−1
vx,i = – backward difference,
h
vi+1 − vi−1
vx̊,i = – central difference,
2h
vi+1 − 2vi + vi−1
vxx,i = – second order difference,
h2
see Figure 2.1. 2

13
 Figure 2.1: Illustration of the finite differences.

Remark 2.5 Some properties of the finite differences. It is (exercise)

1
vx̊,i = (vx,i + vx,i ), vxx,i = (vx,i )x,i .
2
Using the Taylor series expansion for v(x) at the node xi , one gets (exercise)
1
= v 0 (xi ) + hv 00 (xi ) + O h2 ,


vx,i
2
1
= v 0 (xi ) − hv 00 (xi ) + O h2 ,


vx,i
2
= v 0 (xi ) + O h2 ,


vx̊,i
= v 00 (xi ) + O h2 .


vxx,i

Definition 2.6 Consistent difference operator. Let L be a differential oper-

ator. The difference operator Lh : Rn+1 → Rn+1 is called consistent with L of
order k if

max |(Lu)(xi ) − (Lh uh )i | = k(Lu)(xi ) − (Lh uh )i k∞,ωh = O hk

0≤i≤n

for all sufficiently smooth functions u(x). 2

Example 2.7 Consistency orders. The order of consistency measures the quality
of approximation of L by Lh .
d
The difference operators vx,i , vx,i , vx̊,i are consistent to L = dx with order 1, 1,
d2
and 2, respectively. The operator vxx,i is consistent of second order to L = dx 2 , see

Remark 2.5. 2

Example 2.8 Approximation of a more complicated differential operator by differ-

ence operators. Consider the differential operator
 
d du
Lu = k(x) ,
dx dx

14
where k(x) is assumed to be continuously differentiable. Define the difference op-
erator Lh as follows
1 
(Lh uh )i = (aux,i )x,i = a(xi+1 )ux,i (xi+1 ) − a(xi )ux,i (xi )
 h 
1 ui+1 − ui ui − ui−1
= ai+1 − ai ,
h h h

where a is a grid function which has to be determined appropriately. One gets with
the product rule
(Lu)i = k 0 (xi )(u0 )i + k(xi )(u00 )i
and a Taylor series expansion for ui−1 , ui+1

ai+1 − ai 0 ai+1 + ai 00 h(ai+1 − ai ) 000

(u )i + O h2 .


(Lh uh )i = (u )i + (u )i +
h 2 6
Thus, the difference of the differential operator and the difference operator is
   
0 ai+1 − ai 0 ai+1 + ai
(Lu)i − (Lh uh )i = k (xi ) − (u )i + k(xi ) − (u00 )i
h 2
h(ai+1 − ai ) 000
(u )i + O h2 .


− (2.2)
6
In order to define Lh such that it is consistent of second order to L, one has to
satisfy the following two conditions
ai+1 − ai ai+1 + ai
= k 0 (xi ) + O h2 , = k(xi ) + O h2 .



h 2
From the first requirement, it follows that ai+1 − ai = O
(h). Hence, the third term
in the consistency error equation (2.2) is of order O h2 . Possible choices for the
grid function are (exercise)
 
ki + ki−1 h 1/2
ai = , ai = k xi − , ai = (ki ki−1 ) .
2 2

Note that the ’natural’ choice, ai = ki , leads only to first order consistency.
(exercise) 2

2.2 Finite Difference Approximation of the Lapla-

cian in Two Dimensions
Remark 2.9 The five point stencil. The Laplacian in two dimensions is defined by

∂2u ∂2u
∆u(x) = + 2 = ∂x2 u + ∂y2 u = uxx + uyy , x = (x, y).
∂x2 ∂y
The simplest approximation uses for both second order derivatives the second order
differences. One obtains the so-called five point stencil and the approximation
ui+1,j − 2uij + ui−1,j ui,j+1 − 2uij + ui,j−1
∆u ≈ Λu = uxx + uyy = 2
+ , (2.3)
hx h2y

see Figure 2.2. From the consistency order of the second order difference

it follows
immediately that Λu approximates the Laplacian of order O h2x + h2y . 2

15
 Figure 2.2: Five point stencils.

Remark 2.10 The five point stencil on curvilinear boundaries. There is a difficulty
if the five point stencil is used in domains with curvilinear boundaries. The approx-
imation of the second derivative requires three function values in each coordinate
direction

(x − h− +
x , y), (x, y), (x + hx , y),
(x, y − h− +
y ), (x, y), (x, y + hy ),

see Figure 2.3. A guideline of defining the approximation is that the five point
stencil is recovered in the case h− +
x = hx . A possible approximation of this type is

∂2u 1 u(x + h+ u(x, y) − u(x − h−

 
x , y) − u(x, y) x , y)
≈ − (2.4)
∂x2 hx h+x h−
x

−
with hx = (h+ x + hx )/2. Using a Taylor series expansion, one finds that the error
of this approximation is

∂2u 1 u(x + h+ u(x, y) − u(x − h−

 
x , y) − u(x, y) x , y)
− −
∂x2 hx h+x h−
x
3
1 + ∂ u  2
= − (hx − h− x) + O hx .
3 ∂x3
−
For h+
x 6= hx , this approximation is of first order.

Figure 2.3: Sketch to Remark 2.10.

A different way consists in using

∂2u 1 u(x + h+ u(x, y) − u(x − h−

 
x , y) − u(x, y) x , y)
≈ −
∂x2 h̃x h+x h−
x

16
 −
with h̃x = max{h+ x , hx }. However, this approximation possesses only the order
zero, i.e., there is actually no approximation.
Altogether, there is a loss of order of consistency in this situation. 2
Example 2.11 The Dirichlet problem. Consider the Poisson equation which is
equipped with Dirichlet boundary conditions (2.1). First, R2 is decomposed by a
grid with rectangular mesh cells xi = ihx , yj = jhy , hx , hy > 0, i, j ∈ Z. Denote by
wh◦ = {◦} inner nodes, five point stencil completely in Ω,
wh∗ = {∗} inner nodes that are close to the boundary,
γh = {∗} boundary nodes,
ωh = wh◦ ∪ wh∗ inner nodes,
ωh ∪ γh grid,
see Figure 2.4.

Figure 2.4: Different types of nodes in the grid.

The finite difference approximation of problem (2.1) which will be studied in the
following consists in finding a mesh function u(x) such that
−Λu(x) = φ(x) x ∈ wh◦ ,
−Λ∗ u(x) = φ(x) x ∈ wh∗ , (2.5)
u(x) = g(x) x ∈ γh ,
where φ(x) is a grid function that approximates f (x) and Λ∗ is an approximation of
the Laplacian for nodes that are close to the boundary, e.g., defined by (2.4). The
discrete problem is a large sparse linear system of equations. The most important
questions are:
• Which properties possesses the solution of (2.5)?
• Converges the solution of (2.5) to the solution of the Poisson problem and if yes,
with which order?
2

2.3 The Discrete Maximum Principle for a Finite

Difference Operator
Remark 2.12 Contents of this section. Solutions of the Laplace equation, i.e., of
(2.1) with f (x) = 0, fulfill so-called maximum principles. This section shows, that
the finite difference approximation of an operator, where the five point stencil of
the Laplacian is a special case, satisfies a discrete analog of one of the maximum
principles. 2

17
Theorem 2.13 Maximum principles for harmonic functions. Let Ω ⊂ Rd
be a bounded domain and u ∈ C 2 (Ω) ∩ C(Ω) harmonic in Ω, i.e. u(x) solves the
Laplace equation −∆u = 0 in Ω.
• Weak maximum principle. It holds
max u(x) = max u(x).
x∈Ω x∈∂Ω

That means, u(x) takes it maximal value at the boundary.

• Strong maximum principle. If Ω is connected and if the maximum is taken in Ω
(note that Ω is open), i.e., u(x0 ) = maxx∈Ω u(x) for a point x0 ∈ Ω, then u(x)
is constant
u(x) = max u(x) = u(x0 ) ∀ x ∈ Ω.
x∈Ω

Proof: See the literature, e.g., (Evans, 2010, p. 27, Theorem 4) or the class on the
theory of partial differential equations.

Remark 2.14 Interpretation of the maximum principle. The Laplace equation

models the temperature distribution of a heated body without heat sources in Ω.
Then, the weak maximum principle just states that the temperature in the interior
of the body cannot be higher than the highest temperature at the boundary.
There are maximum principles also for more complicated operators than the
Laplacian, e.g., see Evans (2010).
Since the solution of the partial differential equation will be only approximated
by a discretization like a finite difference method, one has to expect that basic
physical properties are satisfied by the numerical solution also only approximately.
However, in applications, it is often very important that such properties are satisfied
exactly. 2
Remark 2.15 The difference equation. In this section, a difference equation of the
form X
a(x)u(x) = b(x, y)u(y) + F (x), x ∈ ωh ∪ γh , (2.6)
y∈S(x)

will be considered. In (2.6), for each node x, the set S(x) is the set of all nodes
on which the sum has to be performed, x 6∈ S(x). That means, a(x) describes the
contribution of the finite difference scheme of a node x to itself and b(x, y) describes
the contributions from the neighbors.
It will be assumed that the grid ωh is connected, i.e., for all xa , xe ∈ ωh exist
x1 , . . . , xm ∈ ωh with x1 ∈ S(xa ), x2 ∈ S(x1 ), . . . , xe ∈ S(xm ). E.g., the situation
depicted in Figure 2.5 is not allowed.

Figure 2.5: Grid which is not allowed in Section 2.3.

It will be assumed that the coefficients a(x) and b(x, y) satisfy the following
conditions:
a(x) > 0, b(x, y) > 0, ∀ x ∈ ωh , ∀ y ∈ S(x),
a(x) = 1, b(x, y) = 0 ∀ x ∈ γh (Dirichlet boundary condition).

18
The values of the Dirichlet boundary condition are incorporated in (2.6) into the
function F (x). 2

Example 2.16 Five point stencil for approximating the Laplacian. Inserting the
approximation of the Laplacian with the five point stencil (2.3) for x = (x, y) ∈ ωh◦
into the scheme (2.6) gives
"
2(h2x + h2y ) 1 1
2 2
u(x, y) = 2
u(x + hx , y) + 2 u(x − hx , y)
hx hy hx hx
#
1 1
+ 2 u(x, y + hy ) + 2 u(x, y − hy ) + φ(x, y).
hy hy

It follows that
2(h2x + h2y )
a(x) = ,
h2x h2y
b(x, y) ∈ {h−2 −2
x , hy },
S(x) = {(x − hx , y), (x + hx , y), (x, y − hy ), (x, y + hy )}.

For inner nodes that are close to the boundary, only the one-dimensional case
(2.4) will be considered for simplicity. Let x + h+ x ∈ γh , then it follows by inserting
(2.4) into (2.6)

u(x − h− u(x + h+
 
1 1 1 x , y) x , y)
+ + − u(x, y) = −
+ +
+φ(x),
hx hx hx hx hx hx hx
| {z }
on γh →F (x)
 
where a(x) = 1
hx
1
h+
+ 1
h−
, b(x, y) = 1
hx h−
und S(x) = {(x − h−
x , y)}. 2
x x x

Remark 2.17 Reformulation of the difference scheme. Scheme (2.6) can be refor-
mulated in the form
X

d(x)u(x) = b(x, y) u(y) − u(x) + F (x) (2.7)
y∈S(x)

2
P
with d(x) = a(x) − y∈S(x) b(x, y).

Example 2.18 Five point stencil for approximating the Laplacian. Using the five
point stencil for approximating the Laplacian, form (2.7) of the scheme is obtained
with
2(h2x + h2y ) 2 2
d(x) = 2 2
− 2 − 2 =0
hx hy hx hy
for x ∈ ωh◦ .
The coefficients a(x) and b(x, y) are the weights of the finite difference stencil
for approximating the Laplacian. A minimal condition for consistency is that this
approximation vanishesP for constant functions. It follows that also for the nodes
x ∈ ωh∗ it is a(x) = y∈S(x) b(x, y). However, as it was shown in Example 2.16,
in this case the contributions from the neighbors on γh are included in the scheme
(2.6) in F (x). Hence, one obtains for nodes that are close to the boundary
X X X
d(x) = b(x, y) − b(x, y) = b(x, y). (2.8)
y∈S(x) y∈S(x),y6∈γh y∈S(x),y∈γh
| {z }
=a(x)

19
In the one-dimensional case, one has, by the definition of hx and with h− +
x = hx ≥ hx ,
 
1 1 1 1 1 2
d(x) = + + − − −
= =
hx hx hx hx hx +
hx hx +
hx hx + h+ +
x hx
2 1
≥ = > 0.
hx hx + hx hx hx hx
2

Lemma 2.19 Discrete maximum principle (DMP). Let u(x) 6= const on ωh

and d(x) ≥ 0 for all x ∈ ωh . Then, it follows from
X

Lh u(x) := d(x)u(x) − b(x, y) u(y) − u(x) ≤ 0 (2.9)
y∈S(x)

(or Lh u(x) ≥ 0, respectively) on ωh that u(x) does not possess a positive maximum
(or negative minimum, respectively) on ωh .

Proof: The proof is performed by contradiction. Let Lh u(x) ≤ 0 for all x ∈ ωh and
assume that u(x) has a positive maximum on ωh at x, i.e., u(x) = maxx∈ωh u(x) > 0.
Then, the idea of the proof consists in showing that with these assumptions there is a node
x̃ ∈ ωh with Lh u(x̃) > 0.
For the node x it holds that
X

Lh u(x) = d(x)u(x) − b(x, y) u(y) − u(x) ≥ d(x)u(x) ≥ 0.
| {z } |
y∈S(x)
{z }
>0 ≤0 by definition of x

Hence, it follows that Lh u(x) = 0 and, in particular, that d(x) = 0. All terms in the sum
are non-positive. Consequently, if the sum should be zero, all terms have to be zero, too.
Since it was assumed that b(x, y) is positive, it must also hold

u(y) = u(x) ∀ y ∈ S(x).

From the assumption u(x) 6= const it follows that there exists a node x̂ ∈ ωh with u(x) >
u(x̂). Because the grid is connected, there is a path x, x1 , . . . , xm , x̂ such that

x1 ∈ S(x), u(x1 ) = u(x),

x2 ∈ S(x1 ), u(x2 ) = u(x1 ) = u(x),
···
x̂ ∈ S(xm ), u(xm ) = u(xm−1 ) = . . . = u(x) > u(x̂).

For the last node xm , for which u(x) has the same value as for x, it holds that


Lh u(xm ) ≥ d(xm ) u(xm ) − b(xm , x̂) u(x̂) − u(xm ) > 0.
| {z } | {z } | {z } | {z }
≥0 >0 >0 <0

Hence, the node xm is the wanted node x̃.

Corollary 2.20 Non-negativity of the grid function. Let u(x) ≥ 0 for x ∈ γh

and Lh u(x) ≥ 0 on ωh . Then, the grid function u(x) is non-negative for all x ∈
ωh ∪ γh .

Proof: Assume there is a node x ∈ ωh with u(x) < 0. Then, the grid function has
a negative minimum on ωh , which is a contradiction to the discrete maximum principle.

Corollary 2.21 Unique solution of the discrete Laplace equation with ho-
mogeneous Dirichlet boundary conditions. The Laplace equation Lh u(x) = 0
possesses only the trivial solution u(x) = 0 for x ∈ ωh ∪ γh .

20
 Proof: The statement of the corollary follows by applying Corollary 2.20 and its
analog for the non-positivity of the grid function if u(x) ≤ 0 for x ∈ γh and Lh u(x) ≤ 0
on ωh . Note that in the definition Lh u(x) = 0 contains also the boundary values, which
are homogeneous Dirichlet.

Corollary 2.22 Comparison lemma. Let

Lh u(x) = f (x) for x ∈ ωh ; u(x) = g(x) for x ∈ γh ,
Lh u(x) = f (x) for x ∈ ωh ; u(x) = g(x) for x ∈ γh ,

with |f (x)| ≤ f (x) and |g(x)| ≤ g(x). Then it is |u(x)| ≤ u(x) for all x ∈ ωh ∪ γh .
Proof: Exercise.

Remark 2.23 Remainder of this section. The remaining corollaries presented in

this section will be applied in the stability proof in Section 2.4. In this proof, the ho-
mogeneous problem (right hand side vanishes) and the problem with homogeneous
Dirichlet boundary conditions will be analyzed separately. 2
Corollary 2.24 Homogeneous problem. For the solution of the problem
Lh u(x) = 0, x ∈ ωh ,
u(x) = g(x), x ∈ γh ,
with d(x) = 0 for all x ∈ ωh◦ , it holds that
kukl∞ (ωh ∪γh ) ≤ kgkl∞ (γh ) .

Proof: Consider the problem

Lh u(x) = 0, x ∈ ωh ,
u(x) = g(x) = const = kgkl∞ (γh ) , x ∈ γh .

It is u(x) = kgkl∞ (γh ) = const, since for inner nodes that are not close to the boundary it
holds that X

Lh u(x) = d(x) u(x) − b(x, y) u(y) − u(x) = 0.
|{z}
y∈S(x)
| {z }
=0 =0
By definition of the problem, Lh vanishes for constant functions. With the same arguments
as in Example 2.18, one can derive the representation (2.8) for inner nodes that are close
to the boundary. Inserting (2.8) into (2.9) and using in addition u(x) = u(y) yields
X X
Lh u(x) = d(x)u(x) = b(x, y)u(x) = b(x, y)u(y).
y∈S(x),y∈γh y∈S(x),y∈γh

This expression is exactly the contribution of the nodes on γh that are included in F (x)
in the scheme (2.6), see also Example 2.16. That means, the finite difference equation is
also satisfied by the nodes that are close to the wall.
Now, the statement of the corollary follows by the application of Corollary 2.22, since
u(x) ≥ |u(x)|.

Corollary 2.25 Problem with homogeneous boundary condition. For the

solution of the problem
Lh u(x) = f (x), x ∈ ωh ,
u(x) = 0, x ∈ γh ,
with d(x) > 0 for all x ∈ ωh , it is
kukl∞ (ωh ∪γh ) ≤ D−1 f l∞ (ω

h)

with D = diag(d(x)) for x ∈ ωh .

21
 Proof: Consider the grid function
f (x) = |f (x)| ≥ f (x) ∀ x ∈ ωh .
From the discrete maximum principle it follows that the solution of the problem
Lh u(x) = f (x), x ∈ ωh ,
u(x) = 0, x ∈ γh ,
is non-negative, i.e., it holds u(x) ≥ 0 for x ∈ ωh ∪ γh . Define the node x by the condition
u(x) = kukl∞ (ωh ∪γh ) .
In x, it is X

Lh u(x) = d(x)u(x) − b(x, y) u(y) − u(x) = |f (x)| ,
| {z } |
y∈S(x)
{z }
>0 ≤0

from what follows that

 
|f (x)| |f (x)|  f (x)  −1
u(x) ≤ ≤ max = max 
  = D f ∞
l (ωh )
.
d(x) x∈ωh d(x) x∈ωh d(x) 

Since u(x) ≤ u(x) for all x ∈ ωh ∪ γh because of Corollary 2.22, the statement of the
corollary is proved.

Corollary 2.26 Another problem with homogeneous boundary condition.

Consider
Lh u(x) = f (x), x ∈ ωh ,
u(x) = 0, x ∈ γh ,
with f (x) = 0 for all x ∈ ωh◦ . With respect to the finite difference scheme it will
be assumed that d(x) = 0 for all x ∈ ωh◦ , and d(x) > 0 for all x ∈ ωh∗ . Then the
following estimate is valid
kukl∞ (ωh ∪γh ) ≤ D+ f l∞ (ω )

h

−1
+
with D = diag(0, d(x) ). The zero entries appear for x ∈ ωh◦ and the entries
d(x)−1 for x ∈ ωh∗ .
Proof: Let f (x) = |f (x)|, x ∈ ωh , and g(x) = 0, x ∈ γh . The solution u(x) is
non-negative, u(x) ≥ 0 for all x ∈ ωh ∪ γh , see the proof of Corollary 2.25. Define x by
u(x) = kukl∞ (ωh ∪γh ) .
One can choose x ∈ ωh∗ ,
because if x ∈ ωh◦ , then it holds that
X

d(x) u(x) − b(x, y) u(y) − u(x) = f (x) = 0,
|{z} | {z } |
y∈S(x)
{z }
=0 >0 ≤0

i.e. u(x) = u(y) for all y ∈ S(x). Let x̂ ∈ ωh∗ and x, x1 , . . . , xm , x̂ be a connection with
xi 6∈ ωh∗ , i = 1, . . . , m. For xm it holds analogously that
u(xm ) = kukl∞ (ωh ∪γh ) = u(y) ∀ y ∈ S(xm ).
Hence, it follow in particular that u(x̂) = kukl∞ (ωh ∪γh ) such that one can choose x = x̂.
It follows that
X

d(x̂) u(x̂) − b(x̂, y) u(y) − u(x̂) = f (x̂).
|{z} | {z } | {z } |
y∈S(x̂)
{z }
>0 =kukl∞ (ω ∪γ ) >0 ≤0
h h

Since all terms in the sum over x ∈ ωh are non-negative, it follows, using also Corol-
lary 2.22, that
f (x̂) f (x̂)
≤ D+ f l∞ (ω ) .

kukl∞ (ωh ∪γh ) ≤ kukl∞ (ωh ∪γh ) ≤ ≤
d(x̂) d(x̂) h

22
2.4 Stability and Convergence of the Finite Differ-
ence Approximation of the Poisson Problem
with Dirichlet Boundary Conditions
Remark 2.27 Decomposition of the solution. A short form to write (2.5) is
Lh u(x) = f (x), x ∈ ωh , u(x) = g(x), x ∈ γh .
The solution of (2.5) can be decomposed into
u(x) = u1 (x) + u2 (x),
with
Lh u1 (x) = f (x), x ∈ ωh , u1 (x) = 0, x ∈ γh (homogeneous boundary cond.),
Lh u2 (x) = 0, x ∈ ωh , u2 (x) = g(x), x ∈ γh (homogeneous right hand side).
2

Stability with Respect to the Boundary Condition

Remark 2.28 Stability with respect to the boundary condition. From Corollary 2.24
it follows that
ku2 kl∞ (ωh ) ≤ kgkl∞ (γh ) . (2.10)
2

Stability with Respect to the Right Hand Side

Remark 2.29 Decomposition of the right hand side. The right hand side will be
decomposed into
f (x) = f ◦ (x) + f ∗ (x)
with
f (x), x ∈ ωh◦

◦
f (x) = , f ∗ (x) = f (x) − f ◦ (x).
0, x ∈ ωh∗
Since the considered finite difference scheme is linear, also the function u1 (x) can
be decomposed into
u1 (x) = u◦1 (x) + u∗1 (x)
with
Lh u◦1 (x) = f ◦ (x), x ∈ ωh , u◦1 (x) = 0, x ∈ γh ,
Lh u∗1 (x) = f ∗ (x), x ∈ ωh , u∗1 (x) = 0, x ∈ γh .
2
Remark 2.30 Estimate for the inner nodes. Let B((0, 0), R) be a circle with center
(0, 0) and radius R, which is chosen such that R ≥ kxk2 for all x ∈ Ω. Consider the
function
u(x) = α R2 − x2 − y 2 with α > 0,

which takes only non-negative values for (x, y) ∈ Ω. Applying the definition of the
five point stencil, it follows that
Λu(x) = −αΛ(x2 + y 2 − R2 )
(x + hx )2 − 2x2 + (x − hx )2 (y + hy )2 − 2y 2 + (y − hy )2
 
= −α +
h2x h2y
= −4α =: −f (x), x ∈ ωh◦ ,

23
and
(x + h+ 2 2
x2 − (x − h− 2
  
∗ 1 x) −x x)
Λ u(x) = −α −
hx h+
x h−
x
!#
+ 2 2 2 − 2
1 (y + hy ) − y y − (y − hy)
+ −
hy h+y h−y
 + − + −
hx + hx h y + hy
= −α + =: −f (x), x ∈ ωh∗ .
hx hy

Hence, u(x) is the solution of the problem

Lh u(x) = f (x),
x ∈ ωh ,
u(x) = α R2 − x2 − y 2 ≥ 0, x ∈ γh .

It is u(x) ≥ 0 for all x ∈ γh . Choosing α = 1

4 kf ◦ kl∞ (ωh ) , one obtains

f (x) = 4α = kf ◦ kl∞ (ωh ) ≥ |f ◦ (x)| , x ∈ ωh◦ ,

f (x) ≥ 0 = |f ◦ (x)| x ∈ ωh∗ .

Now, Lemma 2.22 (Comparison Lemma) can be applied, which leads to

R2 ◦
ku◦1 kl∞ (ωh ) ≤ kukl∞ (ωh ) ≤ αR2 = kf kl∞ (ωh ) . (2.11)
4
One gets the final lower or equal estimate because (0, 0) does not need to belong to
Ω or ωh . 2

Remark 2.31 Estimate for the nodes that are close to the boundary. Corollary 2.26
can be applied to estimate u∗1 (x). For x ∈ ωh◦ it is d(x) = 0, see Example 2.18. For
x ∈ ωh∗ one has
X 1
d(x) = b(x, y) ≥ 2
h
y∈S(x),y∈γh

with h = max{hx , hy }, since all terms are of the form

1 1 1 1
, , , ,
hx hx hx hx hy hy hy h−
+ − +
y

see Example 2.18. One obtains

ku∗1 kl∞ (ωh ) ≤ D+ f ∗ l∞ (ω ≤ h2 kf ∗ kl∞ (ωh ) .

(2.12)
h)

Lemma 2.32 Stability estimate The solution of the discrete Dirichlet problem
(2.5) satisfies

R2
kukl∞ (ωh ∪γh ) ≤ kgkl∞ (γh ) + kφkl∞ (ω◦ ) + h2 kφkl∞ (ω∗ ) (2.13)
4 h h

with R ≥ kxk2 for all x ∈ Ω and h = max{hx , hy }, i.e., the solution u(x) can be
bounded in the norm k·kl∞ (ωh ∪γh ) by the data of the problem.

Proof: The statement of the lemma is obtained by combining the estimates (2.10),
(2.11), and (2.12).

24
Convergence
Theorem 2.33 Convergence. Let u(x) be the solution of the Poisson equation
(2.1) and uh (x) be the finite difference approximation given by the solution of (2.5).
Then, it is
ku − uh kl∞ (ωh ∪γh ) ≤ Ch2
with h = max{hx , hy }.

Proof: The error in the node (xi , yj ) is defined by eij = u(xi , yj ) − uh (xi , yj ). With
−Λu(xi , yj ) = −∆u(xi , yj ) + O h2 = f (xi , yj ) + O h2 ,

one obtains by subtracting the finite difference equation, the following problem for the
error
−Λe(x) = ψ(x), x ∈ wh◦ , ψ(x) = O h2 ,


∗ ∗
−Λ e(x) = ψ(x), x ∈ wh , ψ(x) = O(1),
e(x) = 0, x ∈ γh ,
where ψ(x) is the consistency error, see Section 2.2. Applying the stability estimate (2.13)
to this problem, one obtains immediately

R2
kψkl∞ (ω◦ ) + h2 kψkl∞ (ω∗ ) = O h2 .


kekl∞ (ωh ∪γh ) ≤
4 h h

2.5 An Efficient Solver for the Dirichlet Problem

in the Rectangle
Remark 2.34 Contents of this section. This section considers the Poisson equation
(2.1) in the special case Ω = (0, lx ) × (0, ly ). In this case, a modification of the
difference stencil in a neighborhood of the boundary of the domain is not needed.
The convergence of the finite difference approximation was already established in
Theorem 2.33. Applying this approximation results in a large linear system of
equations Au = f which has to be solved. This section presents an approach for
solving this system in the case of a rectangular domain in an almost optimal way.
2

Remark 2.35 The considered problem and its approximation. The considered con-
tinuous problem consists in solving

−∆u = f in Ω = (0, lx ) × (0, ly ),

u = g on ∂Ω,

and the corresponding discrete problem in solving

−Λu(x) = φ(x), x ∈ ωh ,
u(x) = g(x), x ∈ γh ,

where the discrete Laplacian is of the form (for simplicity of notation, the subscript
h is omitted)
ui+1,j − 2uij + ui−1,j ui,j+1 − 2uij + ui,j−1
Λu = + =: Λx u + Λy u, (2.14)
h2x h2y

with hx = lx /nx , hy = ly /ny , i = 0, . . . , nx , j = 0, . . . , ny , see Figure 2.6. 2

25
 Figure 2.6: Grid for the Dirichlet problem in the rectangular domain.

Remark 2.36 The linear system of equations. The difference scheme (2.14) is
equivalent to a linear system of equations Au = f .
For assembling the matrix and the right hand side of the system, usually a
lexicographical enumeration of the nodes of the grid is used. The nodes are called
enumerated lexicographically if the node (i1 , j1 ) has a smaller number than the node
(i2 , j2 ), if for the corresponding coordinates it is

y1 < y2 or (y1 = y2 ) ∧ (x1 < x2 ).

Using this lexicographical enumeration of the nodes, one obtains for the inner nodes
a system of the form

A = BlockTriDiag(C, B, C) ∈ R(nx −1)(ny −1)×(nx −1)(ny −1) ,

 
1 2 2 1
B = TriDiag − 2 , 2 + 2 , − 2 ∈ R(nx −1)×(nx −1) ,
hx hx hy hx
 
1
C = Diag − 2 ∈ R(nx −1)×(nx −1) ,
hy
x ∈ ωh◦ ,

 φ(x),
g(x ± hx , y)



 φ(x) +

 , i ∈ {1, nx − 1}; j ∈6 {1, ny − 1},
h2x


f = g(x, y ± hy )
φ(x) + , i 6∈ {1, nx − 1}; j ∈ {1, ny − 1},



 h2y

 g(x ± hx , y) g(x, yx ± hy )
 φ(x) + + , i ∈ {1, nx − 1}; j ∈ {1, ny − 1}.


h2x h2y

The last line of the right hand side vector is for inner nodes which are situated in
corner points of ωh◦ . In this approach, the known Dirichlet boundary values are
already substituted into the system and they appear in the right hand side vector.
The matrices B and C possess some modifications for nodes which have a neighbor
on the boundary.
The linear system of equations has the following properties:
• high dimension: N = (nx − 1)(ny − 1) ∼ 103 · · · 107 ,
• sparse: per row and column of the matrix there are only 3, 4, or 5 non-zero
entries,
• symmetric: hence, all eigenvalues are real,
• positive definite: all eigenvalues are positive. It holds that
 
1 1
λmin = λ(1,1) ∼ π 2 2 + 2 = O (1) ,
lx ly
 
1 1
λmax = λ(nx −1,ny −1) ∼ π 2 = O h−2


2
+ 2
(2.15)
hx hy

with h = max{hx , hy }, see Remark 2.37 below.

26
 • high condition number: For the spectral condition number of a symmetric and
positive definite matrix it is
λmax
= O h−2 .


κ2 (A) =
λmin
Since the dimension of the matrix is large, iterative solvers are an appropriate
approach for solving the linear system of equations. The main costs for iterative
solvers are the matrix-vector multiplications (often one per iteration). The cost of
one matrix-vector multiplication is for sparse matrices proportional to the number
of unknowns. Hence, an optimal solver is given if the number of operations for
solving the linear system of equations is proportional to the number of unknowns.
It is known that the number of iterations of many iterative solvers depends on the
condition number of the matrix:
• (damped) Jacobi method, SOR, SSOR. The number of iteration is proportional
to κ2 (A). That means, if the grid is refined once, h → h/2, then the number
of unknowns is increased by around the factor 4 in two dimensions and also
the number of iterations increases by a factor of around 4. Altogether, for one
refinement step, the total costs increase by a factor of around 16.
• (preconditioned)pconjugate gradient (PCG) method. The number of iterations is
proportional to κ2 (A), see the corresponding theorem from the class Numerical
Mathematics II. Hence, the total costs increase by a factor of around 8 if the
grid is refined once.
• multigrid methods. For multigrid methods, the number of iterations is constant.
Hence, the total costs are proportional to the number of unknowns and these
methods are optimal. However, the implementation of multigrid methods is
involved.
2
Remark 2.37 An eigenvalue problem. The derivation of an alternative direct
solver is based on the eigenvalues and eigenvectors of the discrete Laplacian. It is
possible to computed these quantities only in special situations, e.g., if the Poisson
problem with Dirichlet boundary conditions is considered, the domain is rectangu-
lar, and the Laplacian is approximated with the five point stencil.
Consider the following eigenvalue problem
−Λv(x) = λv(x), x ∈ ωh ,
v(x) = 0, x ∈ γh .
The solution of this problem is sought in product form (separation of variables)
(k) (kx ),x (ky ),y
vij = vi vj , k = (kx , ky )T .
It is
(k) (kx ),x (ky ),y (kx ),x (ky ),y (kx ),x (ky ),y
Λvij = Λx vi vj + vi Λy vj = −λk vi vj
with i = 0, . . . , nx , j = 0, . . . , ny refers to the nodes and kx = 1, . . . , nx − 1,
ky = 1, . . . , ny − 1 refers to the eigenvalues. Note that the number of eigenvalues
is equal to the number of inner nodes, i.e. it is (nx − 1)(ny − 1). In this ansatz,
also a splitting of the eigenvalues in a contribution from the x coordinate and a
contribution from the y coordinate is included. From the boundary condition it
follows that
(k ),x (k ),y
v0 x = vn(kxx ),x = v0 y = vn(kyy ),y = 0.
Now, the eigenvalue problem can be split
(kx ),x (ky ),y
Λx vi (x) Λ y vj (y)
(k ),x
+ λkx = − (k ),y
− λky
vi x vj y

27
 (x) (y)
with λk = λkx + λky . Both sides of this equation have to be constant since one of
them depends only on i, i.e., on x, and the other only on j, i.e., on y. The splitting
of λk can be chosen such that the constant is zero. Then, one gets
(kx ),x (x) (kx ),x (ky ),y (y) (ky ),y
Λx vi + λkx vi = 0, Λy vj + λky vj = 0.

The solution of these eigenvalue problems is well known (exercise)

r    
(kx ),x 2 kx πi (x) 4 2 kx π
vi = sin , λkx = 2 sin ,
lx nx hx 2nx
s    
(k ),y 2 ky πj (y) 4 ky π
vj y = sin , λky = 2 sin2 .
ly ny hy 2ny

It follows that the solution of the full eigenvalue problem is

       
(k) 2 kx πi ky πj 4 kx π 4 ky π
vij = p sin sin , λk = 2 sin2 + 2 sin2
lx ly nx ny hx 2nx hy 2ny

with i = 0, . . . , nx , j = 0, . . . , ny and kx = 1, . . . , nx − 1, ky = 1, . . . , ny − 1. Using

a Taylor series expansion, one obtains now the asymptotic behavior of the eigen-
values as given in (2.15). Note that because of the splitting of the eigenvalues into
the directional contributions, the number of individual terms for computing the
eigenvalues is only O (nx + ny ).
Since the matrix corresponding to Λ is symmetric, the eigenvectors are orthog-
onal with respect to the Euclidean vector product. They become orthonormal with
respect to the weighted Euclidean vector product
ny
nx X
X X lx ly
hu, vi = hx hy u(x)v(x) = hx hy uij (x)vij (x), hx = , hy = ,
x∈ωh ∪γh i=0 j=0
nx ny
(2.16)
i.e., then it is
hv (k) , v (m) i = δk,m .
This property can be checked by using the relation
n  
X
2 iπ n
sin = , n > 1.
i=0
n 2

The norm induced by the weighted Euclidean vector product is given by

 1/2
ny
nx X
X
kvkh = hv, vi1/2 = hx hy 2
vij (x) . (2.17)
i=0 j=0

The weights are such that this norm is for constants (almost) independent of the
mesh, i.e.,
 1/2
1/2 nx + 1 ny + 1 1/2
k1kh = (hx hy (nx + 1)(ny + 1)) = lx ly ≈ (lx ly ) .
nx ny
2

Remark 2.38 Solver based on the eigenvalues and eigenvectors. One uses the
ansatz X
f (x) = fk v (k) (x)
k

28
with the Fourier coefficients
nx Xny    
(k) 2hx hy X kx πi ky πj
fk = hf, v i= p fij sin sin , k = (kx , ky ).
lx ly i=0 j=0 nx ny

The solution u(x) of (2.14) is sought in the form

X
u(x) = uk v (k) (x)
k

with unknown coefficients uk . With this ansatz, one obtains

X X
Λu = uk Λv (k) = uk λk v (k) .
k k

Since the eigenfunctions form a basis of the space of the grid functions, a comparison
of the coefficients with the right hand side gives
fk
uk =
λk
or, for each component,
y −1
nx −1 nX    
X fk (k) 2hx hy X fk kx πi ky πj
uij = v = sin sin ,
λk ij
p
k
lx ly k =1 k =1 λk nx ny
x y

i = 0, . . . , nx , j = 0, . . . , ny .
It is possible to implement this approach with the Fast Fourier Transform (FFT)
with

O (nx ny log2 nx + nx ny log2 ny ) = O (N log2 N ) , N = (nx − 1)(ny − 1),

operations. Hence, this method is almost optimal. 2

2.6 A Higher Order Discretizations

Remark 2.39 Contents. The five point stencil is a second order discretization of
the Laplacian. In this section, a discretization of higher order will be studied. In
these studies, only the case of a rectangular domain Ω = (0, lx )×(0, ly ) and Dirichlet
boundary conditions will be considered. 2

Remark 2.40 Derivation of a fourth order approximation. Let u(x) be the solu-
tion of the Poisson equation (2.1) and assume that u(x) is sufficiently smooth. It
is
∂2u
Lu(x) = ∆u(x) = Lx u(x) + Ly u(x), Lα u := .
∂x2α
Let the five point stencil be represented by the following operator

Λu = Λx u + Λy u.

Applying a Taylor series expansion, one finds that

h2x 2 h2y 2
L u + O h4 .


Λu − ∆u = Lx u + (2.18)
12 12 y
From the equation −Lu = f it follows that

L2x u = −Lx f − Lx Ly u, L2y u = −Ly f − Ly Lx u.

29
Inserting these expressions into (2.18) gives

h2x h2y h2x + h2y

Lx Ly u + O h4 .


Λu − ∆u = − Lx f − Ly f − (2.19)
12 12 12
∂4
The operator Lx Ly = ∂x2 ∂y 2 can be approximated as follows
Lx Ly u ≈ Λx Λy u = uxxyy .
The difference operator in this approximation requires nine points, see Figure 2.7
1 
Λx Λy u = ui+1,j+1 − 2ui,j+1 + ui−1,j+1 − 2ui+1,j + 4uij
h2x h2y

−2ui−1,j + ui+1,j−1 − 2ui,j−1 + ui−1,j−1 .

Therefore it is called nine point stencil.

Figure 2.7: The nine point stencil.

One checks, as usual by using a Taylor series expansion, that this approximation
is of second order
Lx Ly u − Λx Λy u = O h2 .

Inserting this expansion into (2.19) and using the partial differential equation shows
that the difference equation
! !
h2x + h2y h2x h2y
− Λ+ Λx Λy u = f + Lx f + Ly f
12 12 12

is a fourth order approximation of the differential equation (2.1). In addition, one

can replace the derivatives of f (x) also by finite differences
Lx f = Λx f + O h2x , Ly f = Λy f + O h2y .

Finally, one obtains a finite difference equation −Λ0 u = φ with

h2x + h2y h2x h2y
Λ0 = Λx + Λy + Λx Λy , φ=f+ Λx f + Λy f.
12 12 12
2
Remark 2.41 On the convergence of the fourth order approximation. The finite
difference problem with the higher order approximation property can be written
with the help of the second order differences. Since the convergence proof is based
on the five point stencil, the following lemma considers this stencil. It will be proved
that one can estimate the values of the grid function by the second order differences.
This result will be used in the convergence proof for the fourth order approximation.
2

30
Lemma 2.42 Embedding lemma. Let

ωh = {(ihx , jhy ) : i = 1, . . . , nx − 1, j = 1, . . . , ny − 1},

and let y be a grid function on ωh ∪γh with y(x) = 0 for x ∈ γh . Then, the following
estimate holds
kykl∞ (ωh ∪γh ) ≤ M kAykh ,
2 2
max{lx ,ly }
with M = √ , A is the matrix obtained by using the five point stencil Λ =
2 lx ly
Λx + Λy for approximating the second derivatives, and the norm on the right hand
side is defined in (2.17).
k
Proof: Let {vij }, k = (kx , ky ), be the orthonormal basis with
   
k 2 kx πi ky πj
vij = p sin sin
lx ly nx ny

which was derived in Remark 2.37. Then, there is a unique representation of the grid
function y = k yk v k and it holds with (2.16)
P

X 1 X 2 2
Ay = yk λk v k , kAyk2h = yk λk .
hx hy
k k

It follows for x ∈ ωh , because of |sin(x)| ≤ 1 for all x ∈ R, that

 
2 X
X  X   X  
k
|y(x)| =  yk v (x) ≤ |yk | v k (x) ≤ |yk | max v k (x) ≤ p |yk | .
     

k

k k
k lx ly k

Applying the Cauchy–Schwarz inequality for sums gives

!2
2 4 X
|y(x)| ≤ |yk |
lx ly
k
!2
4 X 1
= |λk yk |
lx ly λk
k
4 X 2 2X 1
≤ λk yk
lx ly λ2k
k k
4 X 1
= kAyk2h . (2.20)
lx ly λ2k
k

Now, one has to estimate the last sum. It is already known that
   
4 kx π 4 ky π
λk = 2 sin2 + 2 sin2 , kx = 1, . . . , nx − 1, ky = 1, . . . , ny − 1.
hx 2nx hy 2ny
kα π
Setting l = max{lx , ly } and hα = lα /nα , φα = 2nα
∈ (0, π/2), α ∈ {x, y}, leads to
2 2
kx2 π 2 ky2 π 2 kx2 ky2
   
sin φx sin φy 4 2
kx + ky2 .


λk = + ≥4 + ≥
lx2 φx ly2 φy lx2 ly2 l2

In performing this estimate, it was used that the function sin(φ)/φ is monotonically de-
creasing on (0, π/2), see Figure 2.8, and that

sin φ sin(π/2) 2
≥ = ∀ φ ∈ (0, π/2).
φ π/2 π

Let G = {(x, y) : x > 0, y > 0, x2 + y 2 > 1} be the first quadrant of the complex plane

−2
without the part that belongs to the unit circle, see Figure 2.9. The function kx2 + ky2

31
 Figure 2.8: The function sin(φ)/φ.

has its smallest value in the square [kx − 1, kx ] × [ky − 1, ky ] in the point (kx , ky ). Using
the lower estimate of λk , one obtains
X 1 l4 X
−2
≤ kx2 + ky2
λ2k 16
k,k6=(1,1) k,k6=(1,1)
4 Z kx Z ky
l X
−2
= kx2 + ky2 dxdy
16 kx −1 ky −1
k,k6=(1,1)
| {z }
smallest value in square | {z }
=1
kx ky
l4
Z Z
X
−2
= kx2 + ky2 dxdy
16 kx −1 ky −1
k,k6=(1,1)

l4
Z

−2
≤ x2 + y 2 dxdy
16 G
∞ π/2
ρ=∞ !
l4 l4 π ρ2  πl4
Z Z
polar coord. ρ
= 4
dφdρ = −  = .
16 1 0 ρ 16 2 2 ρ=1 64

For performing this computation, one has to exclude ρ → 0.

Figure 2.9: Illustration to the proof of Lemma 2.42.

For λ(1,1) it is
       
4 2 π 4 2 π 4 2 hx π 4 2 hy π
λ(1,1) = sin + sin = sin + sin
h2x 2nx h2y 2ny h2x 2lx h2y 2ly
2
 2   2
 2  
π 2lx hx π π 2ly hy π
= sin2 + 2 sin2
lx2 hx π 2lx ly hy π 2ly
π2 8 π2 8 16
≥ 2 2
+ 2 2 ≥ 2. (2.21)
lx π ly π l

32
For this estimate, the following relations and the monotonicity of sin(x)/x, see Figure 2.8,
were used
 2  2
lα hα π π sin φα sin(π/4) 8
hα ≤ , φ α = ≤ , ≥ = 2.
2 2lα 4 φα π/4 π

Collecting all estimates gives

X 1 X 1 l4 πl4 l4
= λ−2
(1,1) + ≤ + ≤ .
λ2k λk2
256 64 16
k k,k6=(1,1)

Inserting this estimate into (2.20), the final estimate has the form

2 l2
kykl∞ (ωh ∪γh ) ≤ p kAykh =: M kAykh .
lx ly 4

Theorem 2.43 Convergence of the higher order finite difference scheme.

The finite difference scheme

−Λ0 u(x) = φ(x), x ∈ ωh◦ ,

u(x) = g(x), x ∈ γh ,

with
h2x + h2y h2x h2y
Λ0 = Λx + Λy + Λx Λy , φ=f+ Λx f + Λy f,
12 12 12
converges of fourth order.

Proof: Analogously as in the proof of Theorem 2.33, one finds that the following
equation holds for the error e = u(xi , yj ) − uij :

−Λ0 e(x) = ψ(x), ψ = O h4 , x ∈ ωh ,

e(x) = 0, x ∈ γh .

Let Ωh be the vector space of grid functions, which are non-zero only in the interior, i.e.,
at the nodes from ωh , and which vanish on γh . Let Aα y = −Λα y, y ∈ Ωh , α ∈ {x, y}. The
operators Aα : Ωh → Ωh are linear and they have the following properties:
• They are symmetric and positive definite, i.e., Aα = A∗α > 0, where A∗α is the adjoint
(transposed) of Aα , and (Aα u, v) = (u, Aα v), ∀ u, v ∈ Ωh .
(α)
• They are elliptic, i.e., (Aα u, u) ≥ λ1 (u, u), ∀u ∈ Ωh , with
 
(α) 4 πhα 8
λ1 = 2 sin2 ≥ 2,
hα 2lα lα

see (2.21).
(α)
• They are bounded, i.e., it holds (Aα u, u) ≤ λnα −1 (u, u) with
 
(α) 4 kα π 4
λnα −1 = 2 sin2 ≤ 2
hα 2nα hα

and kAα k2 ≤ 4/h2α , since the spectral norm of a symmetric positive definite matrix is
the largest eigenvalue.
• They are commutative, i.e., Ax Ay = Ay Ax .
• It holds Ax Ay = (Ax Ay )∗ .
The error equation on ωh is given by

h2α
Ax e + Ay e − (κx + κy )Ax Ay e = ψ with κα = . (2.22)
12

33
Using the boundedness of the operators, one finds for all v ∈ Ωh that

(κx Ax Ay v + κy Ax Ay v, v) = ((κx Ax ) Ay v, v) + ((κy Ay ) Ax v, v)

h2x 4 h2y 4
≤ 2
(Ay , v) + (Ax v, v)
12 hx 12 h2y
1
= ((Ax + Ay ) v, v) .
3
Now, it follows for all v ∈ Ωh that

((Ax + Ay − (κx + κy ) Ax Ay ) v, v) = ((Ax + Ay ) v, v) − (κx Ax Ay v + κy Ax Ay v, v)

| {z }
=A0
2
≥ ((Ax + Ay ) v, v) ≥ 0.
3
The matrices on both sides of this inequality are symmetric and because the matrix on
the lower estimate is positive definite, also the matrix at the upper estimate is positive
definite. Since matrices commute since the order of applying the finite differences in x and
y direction does not matter. Using these properties, one gets (exercise)

2
(Ax + Ay ) e ≤ A0 e = kψk ,

3 h h
h

where the last equality follows from (2.22). The application of Lemma 2.42 to the error
gives

l2 3l2 A0 e = p
3l2
ke(x)kl∞ (ωh ∪γh ) ≤ p k(Λx + Λy ) ekh ≤ p h
kψkh
2 lx ly 4 lx ly 4 lx ly
3l2
(hx hy (nx + 1)(ny + 1))1/2 kψkl∞ (ωh ∪γh ) = O h4 .


≤ p
4 lx ly

Remark 2.44 On the discrete maximum principle. Reformulation of the finite

difference scheme −Λ0 u = φ in the form studied for the discrete maximum principle
gives for uij
X
a(x)u(x) = b(x, y)u(y) + φ(x),
y∈S(x)
 
2 2 1 2 2

4 5 1 1
a(x) = + 2 − h + hy 2 2 = + 2 > 0,
h2x hy 12 x hx hy 3 h2x hy
 
1 1
2 1 5 1
b(x, y) = 2
− h2x + h2y 2 2 = 2
− 2 , i ± 1, j,
hx 12 hx hy 6 hx hy
(left, right node)
 
1 1 5
b(x, y) = −+ 2 , i, j ± 1, (bottom, top node)
6 h2x hy
 
1 1 1
b(x, y) = + 2 , i ± 1, j ± 1, (other neighbors).
12 h2x hy

Hence, the assumptions for the discrete maximum principle, see Remark 2.15, are
satisfied only if
1 h √
√ < x < 5.
5 hy
Consequently, the ratio of the grid widths has to be bounded and it has to be of
order one. In this case, one speaks of an isotropic grid. 2

34
2.7 Summary
Remark 2.45 Summary.
• Finite difference methods are the simplest approach for discretizing partial differ-
ential equations. The derivatives are just approximated by difference quotients.
• They are very popular in the engineering community.
• One large drawback are the difficulties in approximating domains which are not
of tensor-product type. However, in the engineering communities, a number of
strategies have been developed to deal with this issue in practice.
• Another drawback arises from the point of view of numerical analysis. The
numerical analysis of finite difference methods is mainly based on Taylor series
expansions. For this tool to be applicable, one has to assume a high regularity
of the solution. These assumptions are generally not realistic.
• In Numerical Mathematics, one considers often other schemes then finite dif-
ference methods. However, there are problems, where finite difference methods
can compete with other discretizations, like finite element methods.
2

35
Chapter 3

Introduction to Sobolev
Spaces

Remark 3.1 Contents. Sobolev spaces are the basis of the theory of weak or
variational forms of partial differential equations. A very popular approach for
discretizing partial differential equations, the finite element method, is based on
variational forms. In this chapter, a short introduction into Sobolev spaces will be
given. Recommended literature are the books Adams (1975); Adams and Fournier
(2003) and Evans (2010). 2

3.1 Elementary Inequalities

Lemma 3.2 Inequality for strictly monotonically increasing function. Let
f : R+ ∪ {0} → R be a continuous and strictly monotonically increasing function
with f (0) = 0 and f (x) → ∞ for x → ∞. Then, for all a, b ∈ R+ ∪ {0} it is
Z a Z b
ab ≤ f (x) dx + f −1 (y) dy,
0 0
−1
where f (y) is the inverse of f (x).
Proof: Since f (x) is strictly monotonically increasing, the inverse function exists.
The proof is based on a geometric argument, see Figure 3.1.

Figure 3.1: Sketch to the proof of Lemma 3.2.

Consider the interval (0, a) on the x-axis and the interval (0, b) on the y-axis. Then, the
Ra
area of the corresponding rectangle is given by ab, 0 f (x) dx is the area below the curve,
R b −1
and 0 f (y) dy is the area between the positive y-axis and the curve. From Figure 3.1,
the inequality follows immediately. The equal sign holds only iff f (a) = b.

36
Remark 3.3 Young’s inequality. Young’s inequality
ε 2 1
ab ≤ a + b2 ∀ a, b, ε ∈ R+ (3.1)
2 2ε
follows from Lemma 3.2 with f (x) = εx, f −1 (y) = ε−1 y. It is also possible to
derive this inequality from the binomial theorem. For proving the generalized Young
inequality
εp 1
ab ≤ ap + q bq , ∀ a, b, ε ∈ R+ (3.2)
p qε
with p−1 + q −1 = 1, p, q ∈ (1, ∞), one chooses f (x) = xp−1 , f −1 (y) = y 1/(p−1) and
applies Lemma 3.2 with intervals where the upper bounds are given by εa and ε−1 b.
2

Remark 3.4 Cauchy–Schwarz inequality. The Cauchy1 –Schwarz2 inequality (for

vectors, for sums)
|(x, y)| ≤ kxk2 kyk2 ∀ x, y ∈ Rn , (3.3)
where (·, ·) is the Euclidean product and k·k2 the Euclidean norm, is well known.
One can prove this inequality with the help of Young’s inequality.
First, it is clear that the Cauchy–Schwarz inequality is correct if one of the
vectors is the zero vector. Now, let x, y with kxk2 = kyk2 = 1. One obtains with
the triangle inequality and Young’s inequality (3.1)
 n  n n n
X  X 1X 2 1X 2
|(x, y)| =  xi yi  ≤ |xi | |yi | ≤ |xi | + |yi | = 1.
 

i=1

i=1
2 i=1
2 i=1

Hence, the Cauchy–Schwarz inequality is correct for x, y. Last, one considers arbi-
trary vectors x̃ 6= 0, ỹ 6= 0. Now, one can utilize the homogeneity of the Cauchy–
Schwarz inequality. From the validity of the Cauchy–Schwarz inequality for x and
y, one obtains by a scaling argument
(kx̃k−1 x̃, kỹk−1 ỹ) ≤ 1
 
2 2
| {z } | {z }
x y

Both vectors x, y have the Euclidean norm 1, hence

1
|(x̃, ỹ)| ≤ 1 ⇐⇒ |(x̃, ỹ)| ≤ kx̃k2 kỹk2 .
kx̃k2 kỹk2

The generalized Cauchy–Schwarz inequality or Hölder inequality

n
!1/p n
!1/q
X p
X q
|(x, y)| ≤ |xi | |yi |
i=1 i=1

with p−1 + q −1 = 1, p, q ∈ (1, ∞), can be proved in the same way with the help of
the generalized Young inequality. 2

Definition 3.5 Lebesgue spaces. The space of functions which are Lebesgue
integrable on Ω to the power of p ∈ [1, ∞) is denoted by
 Z 
p p
L (Ω) = f : |f | (x) dx < ∞ ,
Ω
1 Augustin Louis Cauchy (1789 – 1857)
2 Hermann Amandus Schwarz (1843 – 1921)

37
which is equipped with the norm
Z 1/p
p
kf kLp (x) = |f | (x) dx .
Ω

For p = ∞, this space is

L∞ (Ω) = {f : |f (x)| < ∞ almost everywhere in Ω}
with the norm
kf kL∞ (Ω) = ess supx∈Ω |f (x)|.
2
Lemma 3.6 Hölder’s inequality. Let p−1 + q −1 = 1, p, q ∈ [1, ∞]. If u ∈ Lp (Ω)
and v ∈ Lq (Ω), then it is uv ∈ L1 (Ω) and it holds that
kuvkL1 (Ω) ≤ kukLp (Ω) kvkLq (Ω) . (3.4)
If p = q = 2, then this inequality is also known as Cauchy–Schwarz inequality
kuvkL1 (Ω) ≤ kukL2 (Ω) kvkL2 (Ω) . (3.5)
Proof: p, q ∈ (1, ∞). First, one has to show that |uv(x)| can be estimated from
above by an integrable function. Setting in the generalized Young inequality (3.2) ε = 1,
a = |u(x)|, and b = |v(x)| gives
1 1
|u(x)v(x)| ≤ |u(x)|p + |v(x)|q .
p q
Since the right hand side of this inequality is integrable, by assumption, it follows that
uv ∈ L1 (Ω). In addition, Hölder’s inequality is proved for the case kukLp (Ω) = kvkLq (Ω) = 1
using this inequality
Z Z Z
1 1
|u(x)v(x)| dx ≤ |u(x)|p dx + |v(x)|q dx = 1.
Ω p Ω q Ω
The general inequality follows, for the case that both functions do not vanish almost
everywhere, with the same homogeneity argument as used for proving the Cauchy–Schwarz
inequality of sums. In the case that one of the functions vanishes almost everywhere, (3.4)
is trivially satisfied.
p = 1, q = ∞. It is
Z Z
|u(x)v(x)| dx ≤ |u(x)| ess supx∈Ω |v(x)| dx = kukL1 (Ω) kvkL∞ (Ω) .
Ω Ω

3.2 Weak Derivative and Distributions

Remark 3.7 Contents. This section introduces a generalization of the derivative
which is needed for the definition of weak or variational problems. For an introduc-
tion to the topic of this section, see, e.g., Haroske and Triebel (2008)
Let Ω ⊂ Rd be a domain with boundary Γ = ∂Ω, d ∈ N, Ω 6= ∅. A domain is
always an open set. 2
Definition 3.8 The space C0∞ (Ω). The space of infinitely often differentiable real
functions with compact (closed and bounded) support in Ω is denoted by C0∞ (Ω)
C0∞ (Ω) = {v : v ∈ C ∞ (Ω), supp(v) ⊂ Ω},
where
supp(v) = {x ∈ Ω : v(x) 6= 0}.
2

38
Definition 3.9 Convergence in C0∞ (Ω). The sequence of functions {φn (x)}∞ n=1 ,
φn ∈ C0∞ (Ω) for all n, is said to convergence to the zero functions if and only if
a) ∃K ⊂ Ω, K compact (closed and bounded) with supp(φn ) ⊂ K for all n,
b) Dα φn (x) → 0 for n → ∞ on K for all multi-indices α = (α1 , . . . , αd ), |α| =
α1 + . . . + αd .
It is
lim φn (x) = φ(x) ⇐⇒ lim (φn (x) − φ(x)) = 0.
n→∞ n→∞
2

Definition 3.10 Weak derivative. Let f, F ∈ L1loc (Ω). (L1loc (Ω): for each com-
pact subset Ω0 ⊂ Ω it holds
Z
|u(x)| dx < ∞ ∀ u ∈ L1loc (Ω).)
Ω0

If for all functions g ∈ C0∞ (Ω) it holds that

Z Z
F (x)g(x) dx = (−1)|α| f (x)Dα g(x) dx,
Ω Ω

then F (x) is called weak derivative of f (x) with respect to the multi-index α. 2

Remark 3.11 On the weak derivative.

• One uses the same notations for the derivative as in the classical case : F (x) =
Dα f (x).
• If f (x) is classically differentiable on Ω, then the classical derivative is also the
weak derivative.
• The assumptions on f (x) and F (x) are such that the integrals in the definition
of the weak derivative are well defined. In particular, since the test functions
vanish in a neighborhood of the boundary, the behavior of f (x) and F (x) if x
approaches the boundary is not of importance.
• The main aspect of the weak derivative is due to the fact that the (Lebesgue)
integral is not influenced from the values of the functions on a set of (Lebesgue)
measure zero. Hence, the weak derivative is uniquely defined only up to a set
of measure zero. It follows that f (x) might be not classically differentiable on a
set of measure zero, e.g., in a point, but it can still be weakly differentiable.
• The weak derivative is uniquely determined, in the sense described above.
2

Example 3.12 Weak derivative. The weak derivative of the function f (x) = |x| is

 −1 x<0
f 0 (x) = 0 x=0
1 x>0


In x = 0, one can use also any other real number. The proof of this statement
follows directly from the definition and it is left as an exercise. 2

Definition 3.13 Distribution. A continuous linear functional defined on C0∞ (Ω)

0
is called distribution. The set of all distributions is denoted by (C0∞ (Ω)) .
∞ ∞ 0
Let u ∈ C0 (Ω) and ψ ∈ (C0 (Ω)) , then the following notation is used for the
application of the distribution to the function

ψ(u(x)) = hψ, ui ∈ R.

39
Remark 3.14 On distributions. Distributions are a generalization of functions.
They assign each function from C0∞ (Ω) a real number. 2
Example 3.15 Regular distribution. Let u(x) ∈ L1loc (Ω). Then, a distribution is
defined by Z
u(x)φ(x) dx = hψ, φi, ∀φ ∈ C0∞ (Ω).
Ω
This distribution will be identified with u(x) ∈ L1loc (Ω).
Distributions with such an integral representation are called regular, otherwise
they are called singular. 2
Example 3.16 Dirac distribution. Let ξ ∈ Ω fixed, then
hδξ , φi = φ(ξ) ∀ φ ∈ C0∞ (Ω)
defines a singular distribution, the so-called Dirac distribution or δ-distribution. It
is denoted by δξ = δ(x − ξ). 2
0
Definition 3.17 Derivatives of distributions. Let φ ∈ (C0∞ (Ω)) be a distribution.
0
The distribution ψ ∈ (C0∞ (Ω)) is called derivative in the sense of distributions or
distributional derivative of φ if
hψ, ui = (−1)|α| hφ, Dα ui ∀u ∈ C0∞ (Ω),
α = (α1 , . . . , αd ), αj ≥ 0, j = 1, . . . , d, |α| = α1 + . . . + αd . 2
Remark 3.18 On derivatives of distributions. Each distribution has derivatives in
the sense of distributions of arbitrary order.
If the derivative in the sense of distributions Dα u(x) of u(x) ∈ L1loc (Ω) is a
regular distribution, then also the weak derivative of u(x) exists and both derivatives
are identified. 2

3.3 Lebesgue Spaces and Sobolev Spaces

Remark 3.19 On the spaces Lp (Ω). These spaces were introduced in Defini-
tion 3.5.
• The elements of Lp (Ω) are, strictly speaking, equivalence classes of functions
which are different only on a set of Lebesgue measure zero.
• The spaces Lp (Ω) are Banach spaces (complete normed spaces). A space X is
complete, if each so-called Cauchy sequence {un }∞ n=0 ∈ X, i.e., for all ε > 0
there is an index n0 (ε) such that for all i, j > n0 (ε)
kui − uj kX < ε.
converges and the limit is an element of X.
• The space L2 (Ω) becomes a Hilbert spaces with the inner product
Z
(f, g) = f (x)g(x) dx, kf kL2 = (f, f )1/2 , f, g ∈ L2 (Ω).
Ω

• The dual space of a space X is the space of all bounded linear functionals defined
on X. Let Ω be a domain with sufficiently smooth boundary Γ. of the Lebesgue
spaces Lp (Ω), p ∈ [1, ∞], then
0 1 1
(Lp (Ω)) = Lq (Ω) with p, q ∈ (1, ∞), + = 1,
p q

0
L1 (Ω) = L∞ (Ω),
0
(L∞ (Ω)) 6= L1 (Ω).

40
 The spaces L1 (Ω), L∞ (Ω) are not reflexive, i.e., the dual space of the dual space
is not the original space again.
2

Definition 3.20 Sobolev3 spaces. Let k ∈ N ∪ {0} and p ∈ [1, ∞], then the
Sobolev space W k,p (Ω) is defined by

W k,p (Ω) := {u ∈ Lp (Ω) : Dα u ∈ Lp (Ω) ∀ α with |α| ≤ k}.

This space is equipped with the norm

X
kukW k,p (Ω) := kDα ukLp (Ω) . (3.6)
|α|≤k

Remark 3.21 On the spaces W k,p (Ω).

• Definition 3.20 has the following meaning. From u ∈ Lp (Ω), p ∈ [1, ∞), it
follows in particular that u ∈ L1loc (Ω), such that u(x) defines (represents) a
distribution. Then, all derivatives Dα u exist in the sense of distributions. The
0
statement Dα u ∈ Lp (Ω) means that the distribution Dα u ∈ (C0∞ (Ω)) can be
p
represented by a function from L (Ω).
• One can add elements from W k,p (Ω) and one can multiply them with real num-
bers. The result is again a function from W k,p (Ω). With this property, the space
W k,p (Ω) becomes a vector space (linear space). It is straightforward to check
that (3.6) is a norm. (exercise)
• It is Dα u(x) = u(x) for α = (0, . . . , 0) and W 0,p (Ω) = Lp (Ω).
• The spaces W k,p (Ω) are Banach spaces.
• Sobolev spaces have for p ∈ [1, ∞) a countable basis {ϕn (x)}∞ n=1 (Schauder
basis), i.e., each element u(x) can be written in the form
∞
X
u(x) = un ϕn (x), un ∈ R n = 1, . . . , ∞.
n=1

• Sobolev spaces are uniformly convex for p ∈ (1, ∞), i.e., for each ε ∈ (0, 2] (note
that the largest distance in the ball is equal to 2) there is a δ(ε) > 0 such that for
all u, v ∈ W k,p (Ω) with kukW k,p (Ω) = kvkW k,p (Ω) = 1, and ku − vkW k,p (Ω) > ε

it holds that u+v
2
k,p
W (Ω)
≤ 1 − δ(ε), see Figure 3.2 for an illustration.
• Sobolev spaces are reflexive for p ∈ (1, ∞).
• On can show that C ∞ (Ω) is dense in W k,p (Ω), e.g., see (Alt, 1999, Satz 1.21, Satz
2.10) or (Adams, 1975, Lemma 3.15). With this property, one can characterize
the Sobolev spaces W k,p (Ω) as completion of the functions from C ∞ (Ω) with
respect to the norm (3.6). For domains with smooth boundary, one can even
show that C ∞ (Ω) is dense in W k,p (Ω).
• The Sobolev space H k (Ω) = W k,2 (Ω) is a Hilbert space with the inner product
X Z
(u, v)H k (Ω) = Dα u(x)Dα v(x) dx
|α|≤k Ω

and the norm kukH k (Ω) = (u, u)1/2 .

2

3 Sergei Lvovich Sobolev (1908 – 1989)

41
 Figure 3.2: Illustration of the uniform convexity of Sobolev spaces.

Definition 3.22 The space W0k,p (Ω). The Sobolev space W0k,p (Ω) is defined as
the completion of C0∞ (Ω) in the norm of W k,p (Ω)
k·kW k,p (Ω)
W0k,p (Ω) = C0∞ (Ω) .

3.4 The Trace of a Function from a Sobolev Space

Remark 3.23 Motivation. This class considers boundary value problems for par-
tial differential equations. In the theory of weak or variational solutions, the solu-
tion of the partial differential equation is searched in an appropriate Sobolev space.
Then, for the boundary value problem, this solution has to satisfy the boundary
condition. However, since the boundary of a domain is a manifold of dimension
(d − 1), and consequently it has Lebesgue measure zero, one has to clarify how a
function from a Sobolev space is defined on this manifold. This definition will be
presented in this section. 2

Definition 3.24 Boundary of class C k,α . A bounded domain Ω ⊂ Rd and its

boundary Γ are of class C k,α , 0 ≤ α ≤ 1 if for all x0 ∈ Γ there is a ball B(x0 , r)
with r > 0 and a bijective map ψ : B(x0 , r) → D ⊂ Rd such that
1) ψ (B(x0 , r) ∩ Ω) ⊂ Rd+ ,
2) ψ (B(x0 , r) ∩ Γ) ⊂ ∂Rd+ ,
3) ψ ∈ C k,α (B(x0 , r)), ψ −1 ∈ C k,α (D), are Hölder continuous.
That means, Γ is locally the graph of a function with d − 1 arguments. (A function
u(x) is Hölder continuous if
X X
kukC k,α (Ω) = kDα ukC(Ω) + [Dα u]C 0,α (Ω)
|α|≤k |α|=k

with  
|u(x) − u(y)|
[Dα u]C 0,α (Ω) = sup α
x,y∈Ω |x − y|
is finite.) 2

Remark 3.25 Lipschitz boundary. It will be generally assumed that the boundary
of Ω is of class C 0,1 . That means, the map is Lipschitz4 continuous. Such a boundary
4 Rudolf Otto Sigismund Lipschitz (1832 – 1903)

42
is simply called Lipschitz boundary and the domain is called Lipschitz domain. An
important feature of a Lipschitz boundary is that the outer normal vector is defined
almost everywhere at the boundary and it is almost everywhere continuous. 2
Example 3.26 On Lipschitz domains.
• Domains with Lipschitz boundary are, for example, balls or polygonal domains
in two dimensions where the domain is always on one side of the boundary.
• A domain which is not a Lipschitz domain is a circle with a slit
Ω = {(x, y) : x2 + y 2 < 1} \ {(x, y) : x ≥ 0, y = 0}.
At the slit, the domain is on both sides of the boundary.
• In three dimension, a polyhedral domain is not not necessarily a Lipschitz do-
main. For instance, if the domain is build of two bricks which are laying on each
other like in Figure 3.3, then the boundary is not Lipschitz continuous where
the edge of one brick meets the edge of the other brick.
2

Figure 3.3: Polyhedral domain in three dimensions which is not Lipschitz continuous
(at the corner where the arrow points to).

Theorem 3.27 Trace theorem. Let Ω ⊂ Rd , d ≥ 2, with a Lipschitz boundary.

Then, there is exactly one linear and continuous operator γ : W 1,p (Ω) → Lp (Γ),
p ∈ [1, ∞), which gives for functions u ∈ C(Ω) ∩ W 1,p (Ω) the classical boundary
values
γu(x) = u(x), x ∈ Γ, ∀ u ∈ C(Ω) ∩ W 1,p (Ω),
i.e., γu(x) = u(x)|x∈Γ .
Proof: The proof can be found in the literature, e.g., in Adams (1975); Adams and
Fournier (2003).

Remark 3.28 On the trace. The operator γ is called trace or trace operator.
• Since a linear and continuous operator is bounded, there is a constant C > 0
with
kγukLp (Γ) ≤ C kukW 1,p (Ω) ∀ u ∈ W 1,p (Ω)
or
kγkL(W 1,p (Ω),Lp (Γ)) ≤ C.
• By definition of the trace, one gets for u ∈ C(Ω) the classical boundary values.
By the density of C ∞ (Ω) in W 1,p (Ω) for domains with smooth boundary, it
follows that C(Ω) is also dense in W 1,p (Ω) such that for all u ∈ W 1,p (Ω) there
is a sequence {un }∞
n=1 ∈ C(Ω) with un → u in W
1,p
(Ω). Then, the trace of u is
defined to be γu = limk→∞ (γuk ).

43
 • It is

γu(x) = 0 ∀ u ∈ W01,p (Ω),

γDα u(x) = 0 ∀ u ∈ W0k,p (Ω), |α| ≤ k − 1. (3.7)

3.5 Sobolev Spaces with Non-Integer and Nega-

tive Exponents
Remark 3.29 Motivation. Sobolev spaces with non-integer and negative expo-
nents are important in the theory of variational formulations of partial differential
equations.
Let Ω ⊂ Rd be a domain and p ∈ (1, ∞) mit p−1 + q −1 = 1. 2

Definition 3.30 The space W −k,q (Ω). The space W −k,q (Ω), k ∈ N ∪ {0}, con-
tains distributions which are defined on W k,p (Ω)
0
W −k,q (Ω) = ϕ ∈ (C0∞ (Ω)) : kϕkW −k,q < ∞


with
hϕ, ui
kϕkW −k,q = sup .
u∈C0∞ (Ω),u6=0 kukW k,p (Ω)
2

Remark 3.31 On the spaces W −k,p (Ω).

h i0
• It is W −k,q (Ω) = W0k,p (Ω) , i.e., W −k,q (Ω) can be identified with the dual

0
space of W0k,p (Ω). In particular it is H −1 (Ω) = H01 (Ω) .
• It is

. . . ⊂ W 2,p (Ω) ⊂ W 1,p (Ω) ⊂ Lp (Ω) ⊂ W −1,q (Ω) ⊂ W −2,q (Ω) . . .

Definition 3.32 Sobolev–Slobodeckij space. Let s ∈ R, then the Sobolev–

Slobodeckij or Sobolev space H s (Ω) is defined as follows:
• s ∈ Z. H s (Ω) = W s,2 (Ω).
• s > 0 with s = k + σ, k ∈ N ∪ {0}, σ ∈ (0, 1). The space H s (Ω) contains all
functions u for which the following norm is finite:
2 2 2
kukH s (Ω) = kukH k (Ω) + |u|k+σ ,

with
2
(u, v)H s (Ω) = (u, v)H k + (u, v)k+σ , |u|k+σ = (u, u)k+σ ,
X Z Z (Dα u(x) − Dα u(y)) (Dα v(x) − Dα v(y))
(u, v)k+σ = d+2σ
dxdy,
|α|=k Ω Ω kx − yk2

0 k·k −s
• s < 0. H s (Ω) = H0−s (Ω) with H0−s (Ω) = C0∞ (Ω) H (Ω) .


44
3.6 Theorem on Equivalent Norms
Definition 3.33 Equivalent norms. Two norms k·k1 and k·k2 on the linear space
X are said to be equivalent if there are constants C1 and C2 such that

C1 kuk1 ≤ kuk2 ≤ C2 kuk1 ∀ u ∈ X.

Remark 3.34 On equivalent norms. Many important properties, like continuity

or convergence, do not change if an equivalent norm is considered. 2

Theorem 3.35 Equivalent norms in W k,p (Ω). Let Ω ⊂ Rd be a domain with

Lipschitz boundary Γ, p ∈ [1, ∞], and k ∈ N. Let {fi }li=1 be a system with the
following properties:
1) fi : W k,p (Ω) → R+ ∪ {0} is a semi norm,
2) ∃Ci > 0 with 0 ≤ fi (v) ≤ Ci kvkW k,p (Ω) , ∀ v ∈ W k,p (Ω),
3) fni is a norm on othe polynomials of degree k − 1, i.e., if for v ∈ Pk−1 =
α
P
|α|≤k−1 Cα x it holds that fi (v) = 0, i = 1, . . . , l, then it is v ≡ 0.
Then, the norm k·kW k,p (Ω) defined in (3.6) and the norm

l
!1/p
0
X p
kukW k,p (Ω) := fip (u) + |u|W k,p (Ω) with
i=1
 1/p
X Z p
|u|W k,p (Ω) =  |Dα u(x)| dx
|α|=k Ω

are equivalent.

Remark 3.36 On semi norms. For a semi norm fi (·), one cannot conclude from
fi (v) = 0 that v = 0. The third assumptions however states, that this conclusion
can be drawn for all polynomials up to a certain degree. 2

Example 3.37 Equivalent norms in Sobolev spaces.

• The following norms are equivalent to the standard norm in W 1,p (Ω):
Z p 1/p
0  u dx + |u|p 1,p
 
a) kukW 1,p (Ω) =   W (Ω) ,
Ω
Z p 1/p
0  u ds + |u|p 1,p
 
b) kukW 1,p (Ω) =   W (Ω) ,
Z Γ 1/p
0 p p
c) kukW 1,p (Ω) = |u| ds + |u|W 1,p (Ω) .
Γ

• In W k,p (Ω) it is

k−1  i p !1/p
0
XZ ∂ u
kukW k,p (Ω) =   ds + |u|p k,p
 i W (Ω)
i=0 Γ ∂n

equivalent to the standard norm. Here, n denotes the outer normal on Γ with
knk2 = 1.

45
 • In the case W0k,p (Ω), one does not need the regularity of the boundary. It is
0
kukW k,p (Ω) = |u|W k,p (Ω) ,
0

i.e., in the spaces W0k,p (Ω) the standard semi norm is equivalent to the standard
norm.
In particular, it is for u ∈ H01 (Ω) (k = 1, p = 2)

C1 kukH 1 (Ω) ≤ k∇ukL2 (Ω) ≤ C2 kukH 1 (Ω) .

It follows that there is a constant C > 0 such that

kukL2 (Ω) ≤ C k∇ukL2 (Ω) ∀ u ∈ H01 (Ω). (3.8)

3.7 Some Inequalities in Sobolev Spaces

Remark 3.38 Motivation. This section presents a generalization of the last part of
Example 3.37. It will be shown that for inequalities of type (3.8) it is not necessary
that the trace vanishes on the complete boundary.
Let Ω ⊂ Rd Rbe a bounded domain with boundary Γ and let Γ1 ⊂ Γ with
measRd−1 (Γ1 ) = Γ1 ds > 0.
One considers the space

v ∈ W 1,p (Ω) : v|Γ1 = 0 ⊂ W 1,p (Ω) if Γ1 ⊂ Γ,


V0 =
V0 = W01,p (Ω) if Γ1 = Γ

with p ∈ [1, ∞). 2

Lemma 3.39 Friedrichs5 inequality, Poincaré6 inequality, Poincaré–Fried-

richs inequality. Let p ∈ [1, ∞) and measRd−1 (Γ1 ) > 0. Then it is for all u ∈ V0
Z Z
p p
|u(x)| dx ≤ CP k∇u(x)k2 dx, (3.9)
Ω Ω

where k·k2 is the Euclidean vector norm.

Proof: The inequality will be proved with the theorem on equivalent norms, Theo-
rem 3.35. Let f1 (u) : W 1,p (Ω) → R+ ∪ {0} with
Z 1/p
f1 (u) = |u(s)|p ds .
Γ1

This functions has the following properties:

1) f1 (u) is a semi norm.
2) It is
Z 1/p Z 1/p
p p
0 ≤ f1 (u) = |u(s)| ds ≤ |u(s)| ds
Γ1 Γ
= kukLp (Γ) = kγukLp (Γ) ≤ C kukW 1,p (Ω) .

The last estimate follows from the continuity of the trace operator.
5 Friedrichs
6 Poincaré

46
3) Let v ∈ P0 , i.e., v is a constant. Then, one obtains from
Z 1/p
0 = f1 (v) = |v(s)|p ds = |v| (measRd−1 (Γ1 ))1/p ,
Γ1

that |v| = 0.
Hence, all assumptions of Theorem 3.35 are satisfied. That means, there are two constants
C1 and C2 with
Z Z 1/p
C1 |u(s)|p ds + k∇u(x)kp2 dx ≤ kukW 1,p (Ω) ≤ C2 kuk0W 1,p (Ω) ∀ u ∈ W 1,p (Ω).
Γ1 Ω
| {z }
kuk0 1,p
W (Ω)

In particular, it follows that

Z Z Z Z 
p p p p
|u(x)| dx + k∇u(x)k2 dx ≤ C2 |u(s)| ds + k∇u(x)kp2 dx
Ω Ω Γ1 Ω

or Z Z Z 
|u(x)|p dx ≤ CP |u(s)|p ds + k∇u(x)kp2 dx
Ω Γ1 Ω

with CP = C2p . Since u ∈ V0 vanishes on Γ1 , the statement of the lemma is proved.

Remark 3.40 On the Poincaré–Friedrichs inequality. In the space V0 becomes

|·|W 1,p a norm which is equivalent to k·kW 1,p (Ω) . The classical Poincaré–Friedrichs
inequality is given for Γ1 = Γ and p = 2

kukL2 ≤ CP k∇ukL2 ∀ u ∈ H01 (Ω),

where the constant depends only on the diameter of the domain Ω. 2

Lemma 3.41 Another inequality of Poincaré–Friedrichs type. Let Ω0 ⊂ Ω

0
with measRd (Ω ) = Ω0 dx > 0, then for all u ∈ W 1,p (Ω) it is
R

Z Z p Z 
p   p
|u(x)| dx ≤ C  u(x) dx + k∇u(x)k2 dx .
Ω Ω 0 Ω

Proof: Exercise.

3.8 The Gaussian Theorem

Remark 3.42 Motivation. The Gaussian theorem is the generalization of the in-
tegration by parts from calculus. This operation is very important for the theory
of weak or variational solutions of partial differential equations. One has to study,
under which conditions on the regularity of the domain and of the functions it is
well defined. 2

Theorem 3.43 Gaussian theorem. Let Ω ⊂ Rd , d ≥ 2, be a bounded domain

with Lipschitz boundary Γ. Then, the following identity holds for all u ∈ W 1,1 (Ω)
Z Z
∂i u(x) dx = u(s)ni (s) ds, (3.10)
Ω Γ

where n is the unit outer normal vector on Γ.

47
 Proof: sketch. First of all, one proves the statement for functions from C 1 (Ω). This
proof is somewhat longer and it is referred to the literature, e.g., Evans (2010).
The space C 1 (Ω) is dense in W 1,1 (Ω), see Remark 3.21. Hence, for all u ∈ W 1,1 (Ω)
there is a sequence {un }∞ 1
n=1 ∈ C (Ω) with

lim ku − un kW 1,1 (Ω) = 0

n→∞

and (3.10) holds for all functions un (x). It will be shown that the limit of the left hand
side converges to the left hand side of (3.10) and the limit of the right hand side converges
to the right hand side of (3.10).
From the convergence in k·kW 1,1 (Ω) , one has in particular
Z Z
lim ∂i un (x) dx = ∂i u(x) dx.
n→∞ Ω Ω

On the other hand, the continuity of the trace operator gives

lim ku − un kL1 (Γ) ≤ C lim ku − un kW 1,1 = 0,

n→∞ n→∞

from what follows that Z Z

lim un (s) ds = u(s) ds.
n→∞ Γ Γ
Since for a Lipschitz boundary, the normal n is almost everywhere continuous, it is
Z Z
lim un (s)ni (s) ds = u(s)ni (s) ds.
n→∞ Γ Γ

Thus, the limits lead to (3.10).

Corollary 3.44 Vector field. Let the conditions of Theorem 3.43 on the domain

d
Ω be satisfied and let u ∈ W 1,1 (Ω) be a vector field. Then it is
Z Z
∇ · u(x) dx = u(s) · n(s) ds.
Ω Γ

Proof: The statement follows by adding (3.10) from i = 1 to i = d.

Corollary 3.45 Integration by parts. Let the conditions of Theorem 3.43 on

the domain Ω be satisfied. Consider u ∈ W 1,p (Ω) and v ∈ W 1,q (Ω) with p ∈ (1, ∞)
and p1 + 1q = 1. Then it is
Z Z Z
∂i u(x)v(x) dx = u(s)v(s)ni (s) ds − u(x)∂i v(x) dx.
Ω Γ Ω

Proof: exercise.

Corollary 3.46 First Green7 ’s formula. Let the conditions of Theorem 3.43 on
the domain Ω be satisfied, then it is
Z Z Z
∂u
∇u(x) · ∇v(x) dx = (s)v(s) ds − ∆u(x)v(x) dx
Ω Γ ∂n Ω

for all u ∈ H 2 (Ω) and v ∈ H 1 (Ω).

Proof: From the definition of the Sobolev spaces it follows that the integrals are well
defined. Now, the proof follows the proof of Corollary 3.45, where one has now to sum
over the components.

7 Georg Green (1793 – 1841)

48
Remark 3.47 On the first Green’s formula. The first Green’s formula is the for-
mula of integrating by parts once. The boundary integral can be equivalently writ-
ten in the form Z
∇u(s) · n(s)v(s) ds.
Γ
The formula of integrating by parts twice is called second Green’s formula. 2

Corollary 3.48 Second Green’s formula. Let the conditions of Theorem 3.43
on the domain Ω be satisfied, then one has
Z Z  

∂u ∂v
∆u(x)v(x) − ∆v(x)u(x) dx = (s)v(s) − (s)u(s) ds
Ω Γ ∂n ∂n
for all u, v ∈ H 2 (Ω).

3.9 Sobolev Imbedding Theorems

Remark 3.49 Motivation. This section studies the question which Sobolev spaces
are subspaces of other Sobolev spaces. With this property, called imbedding, it is
possible to estimate the norm of a function in the subspace by the norm in the
larger space. 2

Lemma 3.50 Imbedding of Sobolev spaces with same integration power p

and different orders of the derivative. Let Ω ⊂ Rd be a domain with p ∈ [1, ∞)
and k ≤ m, then it is W m,p (Ω) ⊂ W k,p (Ω).

Proof: The statement of this lemma follows directly from the definition of Sobolev
spaces, see Definition 3.20.

Lemma 3.51 Imbedding of Sobolev spaces with the same order of the
derivative k and different integration powers. Let Ω ⊂ Rd be a bounded
domain, k ≥ 0, and p, q ∈ [1, ∞] with q > p. Then it is W k,q (Ω) ⊂ W k,p (Ω).

Proof: exercise.

Remark 3.52 Imbedding of Sobolev spaces with the same order of the derivative k
and the same integration power p in imbedded domains. Let Ω ⊂ Rd be a domain
with sufficiently smooth boundary Γ, k ≥ 0, and p ∈ [1, ∞]. Then there is a map
E : W k,p (Ω) → W k,p (Rd ), the so-called (simple) extension, with
• Ev|Ω = v,
• kEvkW k,p (Rd ) ≤ C kvkW k,p (Ω) , with C > 0,
e.g., see (Adams, 1975, Chapter IV) for details. Likewise, the natural restriction
e : W k,p (Rd ) → W k,p (Ω) can be defined and it is kevkW k,p (Ω) ≤ kvkW k,p (Rd ) . 2

Theorem 3.53 A Sobolev inequality. Let Ω ⊂ Rd be a bounded domain with

Lipschitz boundary Γ, k ≥ 0, and p ∈ [1, ∞) with
k≥d for p = 1,
k > d/p for p > 1.

Then there is a constant C such that for all u ∈ W k,p (Ω) it follows that u ∈ CB (Ω),
where
CB (Ω) = {v ∈ C(Ω) : v is bounded} ,
and it is
kukCB (Ω) = kukL∞ (Ω) ≤ C kukW k,p (Ω) . (3.11)

49
 Proof: See literature, e.g., Adams (1975); Adams and Fournier (2003).

Remark 3.54 On the Sobolev inequality. The Sobolev inequality states that each
function with sufficiently many weak derivatives (the number depends on the di-
mension of Ω and the integration power) can be considered as a continuous and
bounded function in Ω. One says that W k,p (Ω) is imbedded in CB (Ω). It is


C Ω ⊂ CB (Ω) ⊂ C(Ω).

Consider Ω = (0, 1) and f1 (x) = 1/x and f2 (x) = sin(1/x). Then, f1 ∈ C(Ω),
f1 6∈ CB (Ω) and f2 ∈ CB (Ω), f2 6∈ C(Ω).
Of course, it is possible to apply this theorem to weak derivatives of functions.
Then, one obtains imbeddings like W k,p (Ω) → CB s
(Ω) for (k − s)p > d, p > 1. A
comprehensive overview on imbeddings can be found in Adams (1975); Adams and
Fournier (2003). 2

Example 3.55 H 1 (Ω) in one dimension. Let d = 1 and Ω be a bounded interval.

Then, each function from H 1 (Ω) (k = 1, p = 2) is continuous and bounded in Ω. 2

Example 3.56 H 1 (Ω) in higher dimensions. The functions from H 1 (Ω) are in
general not continuous for d ≥ 2. This property will be shown with the following
example.

Figure 3.4: The function f (x) of Example 3.56 for d = 2.

Let Ω = {x ∈ Rd : kxk2 < 1/2} and f (x) = ln |ln kxk2 |, see Figure 3.4. For
kxk2 < 1/2 it is |ln kxk2 | = − ln kxk2 and one gets for x 6= 0
1 1 xi xi
∂i f (x) = − =− 2 .
ln kxk2 kxk2 kxk2 kxk2 ln kxk2

For p ≤ d, one obtains

p 
 xi p 
   p  d
 ∂f  1   1 
 ∂xi (x) =  kxk   kxk ln kxk  ≤  kxk ln kxk  .
       
2 2 2 2 2
| {z } | {z }
≤1 ≥e

The estimate of the second factor can be obtained, e.g., with a discussion of the
curve. Using now spherical coordinates, ρ = e−t and S d−1 is the unit sphere, yields
1/2
ρd−1
Z Z Z Z
p dx
|∂i f (x)| dx ≤ d d
= d
dρdω
Ω Ω kxk2 |ln kxk2 | S d−1 0 ρd |ln ρ|

1/2 ln 2
Z Z
d−1 dρ dt
= −meas S d−1


= meas S d
< ∞,
0 ρ |ln ρ| ∞ td

because of d ≥ 2.

50
 It follows that ∂i f ∈ Lp (Ω) with p ≤ d. Analogously, one proves that f ∈ Lp (Ω)
with p ≤ d. Altogether, one has f ∈ W 1,p (Ω) with p ≤ d. However, it is f 6∈ L∞ (Ω).
This example shows that the condition k > d/p for p > 1 is sharp.
In particular, it was proved for p = 2 that from f ∈ H 1 (Ω) in general it does
not follow that f ∈ C(Ω). 2

Example 3.57 The assumption of a Lipschitz boundary. Also the assumption that
Ω is a Lipschitz domain is of importance.
Consider Ω = {(x, y) ∈ R2 : 0 < x < 1, |y| < xr , r > 1}, see Figure 3.5 for
r = 2.

Figure 3.5: Domain of Example 3.57.

For u(x, y) = x−ε/p with 0 < ε < r it is

 
−ε/p−1 ε
∂x u = x − = C(ε, p)x−ε/p−1 , ∂y u = 0.
p

It follows that
X Z p
Z
α
|D u(x, y)| dxdy = C(ε, p) x−ε−p dxdy
|α|=1 Ω Ω

xr
!
Z 1 Z
−ε−p
= C(ε, p) x dy dx
0 −xr
Z 1
= C̃(ε, p) x−ε−p+r dx.
0

This value is finite for −ε − p + r > −1 or for p < 1 + r − ε, respectively. If one

chooses r ≥ ε > 0, then it is u ∈ W 1,p (Ω). But for ε > 0 the function u(x) is not
bounded in Ω, i.e., u 6∈ L∞ (Ω).
The unbounded values of the function are compensated in the integration by the
fact that the neighborhood of the singular point (0, 0) possesses a small measure.
2

51
Chapter 4

The Ritz Method and the

Galerkin Method

Remark 4.1 Contents. This chapter studies variational or weak formulations of

boundary value problems of partial differential equations in Hilbert spaces. The
existence and uniqueness of an appropriately defined weak solution will be discussed.
The approximation of this solution with the help of finite-dimensional spaces is
called Ritz method or Galerkin method. Some basic properties of this method will
be proved.
In this chapter, a Hilbert space V will be considered with inner product a(·, ·) :
V × V → R and norm kvkV = a(v, v)1/2 . 2

4.1 The Theorems of Riesz and Lax–Milgram

Theorem 4.2 Representation theorem of Riesz. Let f ∈ V 0 be a continuous
and linear functional, then there is a uniquely determined u ∈ V with

a(u, v) = f (v) ∀ v ∈ V. (4.1)

In addition, u is the unique solution of the variational problem

1
F (v) = a(v, v) − f (v) → min ∀ v ∈ V. (4.2)
2
Proof: First, the existence of a solution u of the variational problem will be proved.
Since f is continuous, it holds

|f (v)| ≤ c kvkV ∀ v ∈ V,

from what follows that

1 1
F (v) ≥ kvk2V − c kvkV ≥ − c2 ,
2 2
where in the last estimate the necessary criterion for a local minimum of the expression of
the first estimate is used. Hence, the function F (·) is bounded from below and

d = inf F (v)
v∈V

exists.
Let {vk }k∈N be a sequence with F (vk ) → d for k → ∞. A straightforward calculation
(parallelogram identity in Hilbert spaces) gives

kvk − vl k2V + kvk + vl k2V = 2 kvk k2V + 2 kvl k2V .

52
Using the linearity of f (·) and d ≤ F (v) for all v ∈ V , one obtains
kvk − vl k2V
vk + vl 2
v + v 
k l
= 2 kvk k2V + 2 kvl k2V − 4 − 4f (vk ) − 4f (vl ) + 8f
2 V 2
v + v 
k l
= 4F (vk ) + 4F (vl ) − 8F
2
≤ 4F (vk ) + 4F (vl ) − 8d → 0
for k, l → ∞. Hence {vk }k∈N is a Cauchy sequence. Because V is a complete space, there
exists a limit u of this sequence with u ∈ V . Because F (·) is continuous, it is F (u) = d
and u is a solution of the variational problem.
In the next step, it will be shown that each solution of the variational problem (4.2)
is also a solution of (4.1). It is
1
Φ(ε) = F (u + εv) = a(u + εv, u + εv) − f (u + εv)
2
1 ε2
= a(u, u) + εa(u, v) + a(v, v) − f (u) − εf (v).
2 2
If u is a minimum of the variational problem, then the function Φ(ε) has a local minimum
at ε = 0. The necessary condition for a local minimum leads to
0 = Φ0 (0) = a(u, v) − f (v) for all v ∈ V.
Finally, the uniqueness of the solution will be proved. It is sufficient to prove the
uniqueness of the solution of the equation (4.1). If the solution of (4.1) is unique, then
the existence of two solutions of the variational problem (4.2) would be a contradiction to
the fact proved in the previous step. Let u1 and u2 be two solutions of the equation (4.1).
Computing the difference of both equations gives
a(u1 − u2 , v) = 0 for all v ∈ V.

This equation holds, in particular, for v = u1 − u2 . Hence, ku1 − u2 kV = 0, such that

u1 = u2 .

Definition 4.3 Bounded bilinear form, coercive bilinear form, V -elliptic

bilinear form. Let b(·, ·) : V × V → R be a bilinear form on the Banach space
V . Then it is bounded if
|b(u, v)| ≤ M kukV kvkV ∀ u, v ∈ V, M > 0, (4.3)
where the constant M is independent of u and v. The bilinear form is coercive or
V -elliptic if
2
b(u, u) ≥ m kukV ∀ u ∈ V, m > 0, (4.4)
where the constant m is independent of u. 2
Remark 4.4 Application to an inner product. Let V be a Hilbert space. Then the
inner product a(·, ·) is a bounded and coercive bilinear form, since by the Cauchy–
Schwarz inequality
|a(u, v)| ≤ kukV kvkV ∀ u, v ∈ V,
2
and obviously a(u, u) = kukV . Hence, the constants can be chosen to be M = 1
and m = 1.
Next, the representation theorem of Riesz will be generalized to the case of
coercive and bounded bilinear forms. 2
Theorem 4.5 Theorem of Lax–Milgram. Let b(·, ·) : V × V → R be a
bounded and coercive bilinear form on the Hilbert space V . Then, for each bounded
linear functional f ∈ V 0 there is exactly one u ∈ V with
b(u, v) = f (v) ∀ v ∈ V. (4.5)

53
 Proof: One defines linear operators T, T 0 : V → V by
a(T u, v) = b(u, v) ∀ v ∈ V, a(T 0 u, v) = b(v, u) ∀ v ∈ V. (4.6)

Since b(u, ·) and b(·, u) are continuous linear functionals on V , it follows from Theorem 4.2
that the elements T u and T 0 u exist and they are defined uniquely. Because the operators
satisfy the relation
a(T u, v) = b(u, v) = a(T 0 v, u) = a(u, T 0 v), (4.7)
T 0 is called adjoint operator of T . Setting v = T u in (4.6) and using the boundedness of
b(·, ·) yields

kT uk2V = a(T u, T u) = b(u, T u) ≤ M kukV kT ukV =⇒ kT ukV ≤ M kukV

for all u ∈ V . Hence, T is bounded. Since T is linear, it follows that T is continuous.

Using the same argument, one shows that T 0 is also bounded and continuous.
Define the bilinear form

d(u, v) := a(T T 0 u, v) = a(T 0 u, T 0 v) ∀ u, v ∈ V,

where (4.7) was used. Hence, this bilinear form is symmetric. Using the coercivity of b(·, ·)
and the Cauchy–Schwarz inequality gives
2
m2 kvk4V ≤ b(v, v)2 = a(T 0 v, v)2 ≤ kvk2V T 0 v V = kvk2V a(T 0 v, T 0 v) = kvk2V d(v, v).

Applying now the boundedness of a(·, ·) and of T 0 yields

2
m2 kvk2V ≤ d(v, v) = a(T 0 v, T 0 v) = T 0 v V ≤ M kvk2V .

(4.8)

Hence, d(·, ·) is also coercive and, since it is symmetric, it defines an inner product on V .
From (4.8) one has that the norm induced by d(v, v)1/2 is equivalent to the norm kvkV .
From Theorem 4.2 it follows that there is a exactly one w ∈ V with

d(w, v) = f (v) ∀ v ∈ V.

Inserting u = T 0 w into (4.5) gives with (4.6)

b(T 0 w, v) = a(T T 0 w, v) = d(w, v) = f (v) ∀ v ∈ V,

hence u = T 0 w is a solution of (4.5).

The uniqueness of the solution is proved analogously as in the symmetric case.

4.2 Weak Formulation of Boundary Value Prob-

lems
Remark 4.6 Model problem. Consider the Poisson equation with homogeneous
Dirichlet boundary conditions
−∆u = f in Ω ⊂ Rd ,
(4.9)
u = 0 on ∂Ω.
2

Definition 4.7 Weak formulation of (4.9). Let f ∈ L2 (Ω). A weak formulation

of (4.9) consists in finding u ∈ V = H01 (Ω) such that

a(u, v) = (f, v) ∀ v ∈ V (4.10)

with Z
a(u, v) = (∇u, ∇v) = ∇u(x) · ∇v(x) dx
Ω

and (·, ·) is the inner product in L2 (Ω). 2

54
Remark 4.8 On the weak formulation.
• The weak formulation is also called variational formulation.
• As usual in mathematics, ’weak’ means that something holds for all appropri-
ately chosen test functions.
• Formally, one obtains the weak formulation by multiplying the strong form of
the equation (4.9) with the test function, by integrating the equation on Ω, and
applying integration by parts. Because of the Dirichlet boundary condition, on
can use as test space H01 (Ω) and therefore the integral on the boundary vanishes.
• The ansatz space for the solution and the test space are defined such that the
arising integrals are well defined.
• The weak formulation reduces the necessary regularity assumptions for the so-
lution by the integration and the transfer of derivatives to the test function.
Whereas the solution of (4.9) has to be in C 2 (Ω), the solution of (4.10) has to
be only in H01 (Ω). The latter assumption is much more realistic for problems
coming from applications.
• The regularity assumption on the right hand side can be relaxed to f ∈ H −1 (Ω).
2

Theorem 4.9 Existence and uniqueness of the weak solution. Let f ∈

L2 (Ω). There is exactly one solution of (4.10).

Proof: Because of the Poincaré inequality (3.9), there is a constant c with

kvkL2 (Ω) ≤ c k∇vkL2 (Ω) ∀ v ∈ H01 (Ω).

It follows for v ∈ H01 (Ω) ⊂ H 1 (Ω) that

 1/2  1/2
kvkH 1 (Ω) = kvk2L2 (Ω) + k∇vk2L2 (Ω) ≤ c k∇vk2L2 (Ω) + k∇vk2L2 (Ω)
≤ C k∇vkL2 (Ω) ≤ C kvkH 1 (Ω) .

Hence, a(·, ·) is an inner product on H01 (Ω) with the induced norm

kvkH 1 (Ω) = a(v, v)1/2 ,

0

which is equivalent to the norm k·kH 1 (Ω) .

Define for f ∈ L2 (Ω) the linear functional
Z
f˜(v) := f (x)v(x) dx ∀ v ∈ H01 (Ω).
Ω

Applying the Cauchy–Schwarz inequality (3.5) and the Poincaré inequality (3.9)
 
˜ 
f (v) = |(f, v)| ≤ kf kL2 (Ω) kvkL2 (Ω) ≤ c kf kL2 (Ω) k∇vkL2 (Ω) = c kf kL2 (Ω) kvkH 1 (Ω)
0

shows that this functional is continuous on H01 (Ω).

Applying the representation theorem
of Riesz, Theorem 4.2, gives the existence and uniqueness of the weak solution of (4.10).
In addition, u(x) solves the variational problem
Z
1
F (v) = k∇vk22 − f (x)v(x) dx → min for all v ∈ H01 (Ω).
2 Ω

Example 4.10 A more general elliptic problem. Consider the problem

−∇ · (A(x)∇u) + c(x)u = f in Ω ⊂ Rd ,
(4.11)
u = 0 on ∂Ω,

55
with A(x) ∈ Rd×d for each point x ∈ Ω. It will be assumed that the coefficients
ai,j (x) and c(x) ≥ 0 are bounded, f ∈ L2 (Ω), and that the matrix (tensor) A(x)
is for all x ∈ Ω uniformly elliptic, i.e., there are positive constants m and M such
that
2 2
m kyk2 ≤ yT A(x)y ≤ M kyk2 ∀ y ∈ Rd , ∀ x ∈ Ω.
The weak form of (4.11) is obtained in the usual way by multiplying (4.11) with
test functions v ∈ H01 (Ω), integrating on Ω, and applying integration by parts: Find
u ∈ H01 (Ω), such that
a(u, v) = f (v) ∀ v ∈ H01 (Ω)
with Z
∇u(x)T A(x)∇v(x) + c(x)u(x)v(x) dx.


a(u, v) =
Ω

This bilinear form is bounded (exercise). The coercivity of the bilinear form is
proved by using the uniform ellipticity of A(x) and the non-negativity of c(x):
Z
a(u, u) = ∇u(x)T A(x)∇u(x) + c(x)u(x)u(x) dx
Ω
Z
2
≥ m∇u(x)T ∇u(x) dx = m kukH 1 (Ω) .
0
Ω

Applying the Theorem of Lax–Milgram, Theorem 4.5, gives the existence and
uniqueness of a weak solution of (4.11).
If the tensor is not symmetric, aij (x) 6= aji (x) for one pair i, j, then the solution
cannot be characterized as the solution of a variational problem. 2

4.3 The Ritz Method and the Galerkin Method

Remark 4.11 Idea of the Ritz method. Let V be a Hilbert space with the inner
product a(·, ·). Consider the problem
1
F (v) = a(v, v) − f (v) → min, (4.12)
2
where f : V → R is a bounded linear functional. As already proved in Theorem
4.2, there is a unique solution u ∈ V of this variational problem which is also the
unique solution of the equation

a(u, v) = f (v) ∀ v ∈ V. (4.13)

For approximating the solution of (4.12) or (4.13) with a numerical method, it

will be assumed that V has a countable orthonormal basis (Schauder basis). Then,
there are finite-dimensional subspaces V1 , V2 , . . . ⊂ V with dimVk = k, which has
the following property: for each u ∈ V and each ε > 0 there is a K ∈ N and a
uk ∈ Vk with
ku − uk kV ≤ ε ∀ k ≥ K. (4.14)
Note that it is not required that there holds an inclusion of the form Vk ⊂ Vk+1 .
The Ritz approximation of (4.12) and (4.13) is defined by: Find uk ∈ Vk with

a(uk , vk ) = f (vk ) ∀ vk ∈ Vk . (4.15)

Lemma 4.12 Existence and uniqueness of a solution of (4.15). There exists

exactly one solution of (4.15).

56
 Proof: Finite-dimensional subspaces of Hilbert spaces are Hilbert spaces as well. For
this reason, one can apply the representation theorem of Riesz, Theorem 4.2, to (4.15)
which gives the statement of the lemma. In addition, the solution of (4.15) solves a
minimization problem on Vk .

Lemma 4.13 Best approximation property. The solution of (4.15) is the best
approximation of u in Vk , i.e., it is

ku − uk kV = inf ku − vk kV . (4.16)
vk ∈Vk

Proof: Since Vk ⊂ V , one can use the test functions from Vk in the weak equation
(4.13). Then, the difference of (4.13) and (4.15) gives the orthogonality, the so-called
Galerkin orthogonality,
a(u − uk , vk ) = 0 ∀ vk ∈ Vk . (4.17)
Hence, the error u − uk is orthogonal to the space Vk : u − uk ⊥ Vk . That means, uk is the
orthogonal projection of u onto Vk with respect of the inner product of V .
Let now wk ∈ Vk be an arbitrary element, then it follows with the Galerkin orthogo-
nality (4.17) and the Cauchy–Schwarz inequality that

ku − uk k2V = a(u − uk , u − uk ) = a(u − uk , u − (uk − wk )) = a(u − uk , u − vk )

| {z }
vk

≤ ku − uk kV ku − vk kV .

Since wk ∈ Vk was arbitrary, also vk ∈ Vk is arbitrary. If ku − uk kV > 0, division by

ku − uk kV gives the statement of the lemma. If ku − uk kV = 0, the statement of the
lemma is trivially true.

Theorem 4.14 Convergence of the Ritz approximation. The Ritz approxi-

mation converges
lim ku − uk kV = 0.
k→∞

Proof: The best approximation property (4.16) and property (4.14) give
ku − uk kV = inf ku − vk kV ≤ ε
vk ∈Vk

for each ε > 0 and k ≥ K(ε). Hence, the convergence is proved.

Remark 4.15 Formulation of the Ritz method as linear system of equations. One
can use an arbitrary basis {φi }ki=1 of Vk for the computation of uk . First of all, the
equation for the Ritz approximation (4.15) is satisfied for all vk ∈ Vk if and only if
it is satisfied for each basis function φi . This statement follows from the linearity
of both sides of the equation with respect to the test function and from the fact
that each function vP k ∈ Vk can be represented as linear combination of the basis
k
functions. Let vk = i=i αi φi , then from (4.15) it follows that
k
X k
X
a(uk , vk ) = αi a(uk , φi ) = αi f (φi ) = f (vk ).
k=1 k=1

This equation is satisfied if a(uk , φi ) = f (φi ), i = 1, . . . , k. On the other hand, if

(4.15) holds then it holds in particular for each basis function φi .
Then, one uses as ansatz for the solution also a linear combination of the basis
functions
Xk
uk = uj φj
j=1

57
with unknown coefficients uj ∈ R. Using as test functions now the basis functions
yields
X k k
X
j
a(u φj , φi ) = a(φj , φi )uj = f (φi ), i = 1, . . . , k.
j=1 j=1
This equation is equivalent to the linear system of equations Au = f , where
k
A = (aij )i,j=1 = a(φj , φi )ki,j=1
is called stiffness matrix. Note that the order of the indices is different for the
entries of the matrix and the arguments of the inner product. The right hand side
is a vector of length k with the entries fi = f (φi ), i = 1, . . . , k.
Using the one-to-one mapping between the coefficient vector (v 1 , . . . , v k )T and
Pk i
the element vk = i=1 v φi , one can show that the matrix A is symmetric and
positive definite (exercise)
A = AT ⇐⇒ a(v, w) = a(w, v) ∀ v, w ∈ Vk ,
T
x Ax > 0 for x 6= 0 ⇐⇒ a(v, v) > 0 ∀ v ∈ Vk , v 6= 0.
2
Remark 4.16 The case of a bounded and coercive bilinear form. If b(·, ·) is bounded
and coercive, but not symmetric, it is possible to approximate the solution of (4.5)
with the same idea as for the Ritz method. In this case, it is called Galerkin method.
The discrete problem consists in finding uk ∈ Vk such that
b(uk , vk ) = f (vk ) ∀ vk ∈ Vk . (4.18)
2
Lemma 4.17 Existence and uniqueness of a solution of (4.18). There is
exactly one solution of (4.18).
Proof: The statement of the lemma follows directly from the Theorem of Lax–
Milgram, Theorem 4.5.

Remark 4.18 On the discrete solution. The discrete solution is not the orthogonal
projection into Vk in the case of a bounded and coercive bilinear form, which is not
the inner product of V . 2
Lemma 4.19 Lemma of Cea, error estimate. Let b : V × V → R be a bounded
and coercive bilinear form on the Hilbert space V and let f ∈ V 0 be a bounded linear
functional. Let u be the solution of (4.5) and uk be the solution of (4.18), then the
following error estimate holds
M
ku − uk kV ≤ inf ku − vk kV , (4.19)
m vk ∈Vk
where the constants M and m are given in (4.3) and (4.4).
Proof: Considering the difference of the continuous equation (4.5) and the discrete
equation (4.18), one obtains the error equation
b(u − uk , vk ) = 0 ∀ vk ∈ Vk ,
which is also called Galerkin orthogonality. With (4.4), the Galerkin orthogonality, and
(4.3) it follows that
1 1
ku − uk k2V ≤ b(u − uk , u − uk ) = b(u − uk , u − vk )
m m
M
≤ ku − uk kV ku − vk kV , ∀ vk ∈ Vk ,
m
from what the statement of the lemma follows immediately.

58
Remark 4.20 On the best approximation error. It follows from estimate (4.19)
that the error is bounded by a multiple of the best approximation error, where
the factor depends on properties of the bilinear form b(·, ·). Thus, concerning error
estimates for concrete finite-dimensional spaces, the study of the best approximation
error will be of importance. 2

Remark 4.21 The corresponding linear system of equations. The corresponding

linear system of equations is derived analogously to the symmetric case. The system
matrix is still positive definite but not symmetric. 2

Remark 4.22 Choice of the basis. The most important issue of the Ritz and
Galerkin method is the choice of the spaces Vk , or more concretely, the choice of
an appropriate basis {φi }ki=1 that spans the space Vk . From the point of view of
numerics, there are the requirements that it should be possible to compute the
entries aij of the stiffness matrix efficiently and that the matrix A should be sparse.
2

59
Chapter 5

Finite Element Methods

5.1 Finite Element Spaces

Remark 5.1 Mesh cells, faces, edges, vertices. A mesh cell K is a compact poly-
hedron in Rd , d ∈ {2, 3}, whose interior is not empty. The boundary ∂K of K
consists of m-dimensional linear manifolds (points, pieces of straight lines, pieces of
planes), 0 ≤ m ≤ d − 1, which are called m-faces. The 0-faces are the vertices of
the mesh cell, the 1-faces are the edges, and the (d − 1)-faces are just called faces.
2

Remark 5.2 Finite dimensional spaces defined on K. Let s ∈ N. Finite element

methods use finite dimensional spaces P (K) ⊂ C s (K) which are defined on K. In
general, P (K) consists of polynomials. The dimension of P (K) will be denoted by
dim P (K) = NK . 2

Example 5.3 The space P (K) = P1 (K).The space consisting of linear polynomials
on a mesh cell K is denoted by P1 (K):
( d
)
X
T
P1 (K) = a0 + ai xi : x = (x1 , . . . , xd ) ∈ K .
i=1

There are d + 1 unknown coefficients ai , i = 0, . . . , d, such that dim P1 (K) = NK =

d + 1. 2

Remark 5.4 Linear functionals defined on P (K). For the definition of finite ele-
ments, linear functional which are defined on P (K) are of importance.
Consider linear and continuous functionals ΦK,1 , . . . , ΦK,NK : C s (K) → R
which are linearly independent. There are different types of functionals which can
be utilized in finite element methods:
• point values: Φ(v) = v(x), x ∈ K,
• point values of a first partial derivative: Φ(v) = ∂i v(x), x ∈ K,
• point values of the normal derivative on a face E of K: Φ(v) = ∇v(x) · nE , nE
is the outward pointing unit normal vector R on E,
1
• integral mean values on K: Φ(v) = |K| K
v(x) dx,
1
R
• integral mean values on faces E: Φ(v) = |E| E
v(s) ds.
The smoothness parameter s has to be chosen in such a way that the functionals
ΦK,1 , . . . , ΦK,NK are continuous. If, e.g., a functional requires the evaluation of a
partial derivative or a normal derivative, then one has to choose at least s = 1. For
the other functionals given above, s = 0 is sufficient. 2

60
Definition 5.5 Unisolvence of P (K) with respect to the functionals ΦK,1 ,
. . . , ΦK,NK . The space P (K) is called unisolvent with respect to the functionals
ΦK,1 , . . . , ΦK,NK if there is for each a ∈ RNK , a = (a1 , . . . , aNK )T , exactly one
p ∈ P (K) with
ΦK,i (p) = ai , 1 ≤ i ≤ NK .
2

Remark 5.6 Local basis. Unisolvence means that for each vector a ∈ RNK , a =
(a1 , . . . , aNK )T , there is exactly one element in P (K) such that ai is the image of
the i-th functional, i = 1, . . . , NK .
Choosing in particular the Cartesian unit vectors for a, then it follows from the
unisolvence that a set {φK,i }N i=1 exists with φK,i ∈ P (K) and
K

ΦK,i (φK,j ) = δij , i, j = 1, . . . , NK .

Consequently, the set {φK,i }N

i=1 forms a basis of P (K). This basis is called local
K

basis. 2

Remark 5.7 Transform of an arbitrary basis to the local basis. If an arbitrary

basis {pi }N
i=1 of P (K) is known, then the local basis can be computed by solving
K

a linear system of equations. To this end, represent the local basis in terms of the
known basis
NK
X
φK,j = cjk pk , cjk ∈ R, j = 1, . . . , NK ,
k=1

with unknown coefficients cjk . Applying the definition of the local basis leads to
the linear system of equations
NK
X
ΦK,i (φK,j ) = cjk aik = δij , i, j = 1, . . . , NK , aik = ΦK,i (pk ).
k=1

Because of the unisolvence, the matrix A = (aij ) is non-singular and the coefficients
cjk are determined uniquely. 2

Example 5.8 Local basis for the space of linear functions on the reference triangle.
Consider the reference triangle K̂ with the vertices (0, 0), (1, 0), and (0, 1). A linear
space on K̂ is spanned by the functions 1, x̂, ŷ. Let the functionals be defined by
the values of the functions in the vertices of the reference triangle. Then, the given
basis is not a local basis because the function 1 does not vanish at the vertices.
Consider first the vertex (0, 0). A linear basis function ax̂ + bŷ + c which has the
value 1 in (0, 0) and which vanishes in the other vertices has to satisfy the following
set of equations     
0 0 1 a 1
 1 0 1  b  =  0 .
0 1 1 c 0
The solution is a = −1, b = −1, c = 1. The two other basis functions of the local
basis are x̂ and ŷ, such that the local basis has the form {1 − x̂ − ŷ, x̂, ŷ}. 2

Remark 5.9 Triangulation, grid, mesh, grid cell. For the definition of global finite
element spaces, a decomposition of the domain Ω into polyhedrons K is needed.
This decomposition is called triangulation T h and the polyhedrons K are called
mesh cells. The union of the polyhedrons is called grid or mesh.
A triangulation is called regular, see the definition in Ciarlet Ciarlet (1978), if:
• It holds Ω = ∪K∈T h K.

61
 • Each mesh cell K ∈ T h is closed and the interior K̊ is non-empty.
• For distinct mesh cells K1 and K2 there holds K̊1 ∩ K̊2 = ∅.
• For each K ∈ T h , the boundary ∂K is Lipschitz-continuous.
• The intersection of two mesh cells is either empty or a common m-face, m ∈
{0, . . . , d − 1}.
2

Remark 5.10 Global and local functionals. Let Φ1 , . . . , ΦN : C s (Ω) → R contin-

uous linear functionals of the same types as given in Remark 5.4. The restriction
of the functionals to C s (K) defines local functionals ΦK,1 , . . . , ΦK,NK , where it is
assumed that the local functionals are unisolvent on P (K). The union of all mesh
cells Kj , for which there is a p ∈ P (Kj ) with Φi (p) 6= 0, will be denoted by ωi . 2

Example 5.11 On subdomains ωi . Consider the two-dimensional case and let Φi

be defined as nodal value of a function in x ∈ K. If x ∈ K̊, then ωi = K. In the
case that x is on a face of K but not in a vertex, then ωi is the union of K and the
other mesh cell whose boundary contains this face. Last, if x is a vertex of K, then
ωi is the union of all mesh cells which possess this vertex, see Figure 5.1. 2

Figure 5.1: Subdomains ωi .

Definition 5.12 Finite element space, global basis. A function v(x) defined
on Ω with v|K ∈ P (K) for all K ∈ T h is called continuous with respect to the
functional Φi : Ω → R if

Φi (v|K1 ) = Φi (v|K2 ), ∀ K1 , K2 ∈ ωi .

The space
n
S = v ∈ L∞ (Ω) : v|K ∈ P (K) and v is continuous with respect to
o
Φi , i = 1, . . . , N

is called finite element space.

The global basis {φj }Nj=1 of S is defined by the condition

φj ∈ S, Φi (φj ) = δij , i, j = 1, . . . , N.

Example 5.13 Piecewise linear global basis function. Figure 5.2 shows a piecewise
linear global basis function in two dimensions. Because of its form, such a function
is called hat function. 2

62
 Figure 5.2: Piecewise linear global basis function (boldface lines), hat function.

Remark 5.14 On global basis functions. A global basis function coincides on each
mesh cell with a local basis function. This property implies the uniqueness of the
global basis functions.
For many finite element spaces it follows from the continuity with respect to
{Φi }N
i=1 , the continuity of the finite element functions themselves. Only in this
case, one can speak of values of finite element functions on m-faces with m < d. 2

Definition 5.15 Parametric finite elements. Let K̂ be a reference mesh cell

with the local space P (K̂), the local functionals Φ̂1 , . . . , Φ̂N̂ , and a class of bijective
mappings {FK : K̂ → K}. A finite element space is called a parametric finite
element space if:
• The images {K} of {FK } form the set of mesh cells.
• The local spaces are given by
n o
−1
P (K) = p : p = p̂ ◦ FK , p̂ ∈ P̂ (K̂) . (5.1)

• The local functionals are defined by

ΦK,i (v(x)) = Φ̂i (v(FK (x̂))) , (5.2)

where x̂ = (x̂1 , . . . , x̂d )T are the coordinates of the reference mesh cell and it
holds x = FK (x̂).
2

Remark 5.16 Motivations for using parametric finite elements. Definition 5.12 of
finite elements spaces is very general. For instance, different types of mesh cells
are allowed. However, as well the finite element theory as the implementation of
finite element methods become much simpler if only parametric finite elements are
considered. 2

5.2 Finite Elements on Simplices

Definition 5.17 d-simplex. A d-simplex K ⊂ Rd is the convex hull of (d + 1)
points a1 , . . . , ad+1 ∈ Rd which form the vertices of K. 2

Remark 5.18 On d-simplices. It will be always assumed that the simplex is not
degenerated, i.e., its d-dimensional measure is positive. This property is equivalent

63
to the non-singularity of the matrix
 
a11 a12 ... a1,d+1
 a21 a22 ... a2,d+1 
 
A =  ... .. .. ..
,
 
 . . . 
 ad1 ad2 ... ad,d+1 
1 1 ... 1

where ai = (a1i , a2i , . . . , adi )T , i = 1, . . . , d + 1.

For d = 2, the simplices are the triangles and for d = 3 they are the tetrahedrons.
2

Definition 5.19 Barycentric coordinates. Since K is the convex hull of the

points {ai }d+1
i=1 , the parametrization of K with a convex combination of the vertices
reads as follows
( d+1 d+1
)
X X
K = x ∈ Rd : x = λi ai , 0 ≤ λi ≤ 1, λi = 1 .
i=1 i=1

The coefficients λ1 , . . . , λd+1 are called barycentric coordinates of x ∈ K. 2

Remark 5.20 On barycentric coordinates. From the definition it follows that the
barycentric coordinates are the solution of the linear system of equations
d+1
X d+1
X
aji λi = xj , 1 ≤ j ≤ d, λi = 1.
i=1 i=1

Since the system matrix is non-singular, see Remark 5.18, the barycentric coordi-
nates are determined uniquely.
The barycentric coordinates of the vertex ai , i = 1, . . . , d + 1, of the simplex is
λi = 1 and λj = 0 if i 6= j. Since λi (aj ) = δij , the barycentric coordinate λi can be
identified with the linear function which has the value 1 in the vertex ai and which
vanishes in all other vertices aj with j 6= i.
The barycenter of the simplex is given by
d+1 d+1
1 X X 1
SK = ai = ai .
d + 1 i=1 i=1
d+1

Hence, its barycentric coordinates are λi = 1/(d + 1), i = 1, . . . , d + 1. 2

Remark 5.21 Simplicial reference mesh cells. A commonly used reference mesh
cell for triangles and tetrahedrons is the unit simplex
( d
)
X
d
K̂ = x̂ ∈ R : x̂i ≤ 1, x̂i ≥ 0, i = 1, . . . , d ,
i=1

see Figure 5.3. The class {FK } of admissible mappings are the bijective affine
mappings
FK x̂ = B x̂ + b, B ∈ Rd×d , det(B) 6= 0, b ∈ Rd .
The images of these mappings generate the set of the non-degenerated simplices
{K} ⊂ Rd . 2

64
 Figure 5.3: The unit simplices in two and three dimensions.

Definition 5.22 Affine family of simplicial finite elements. Given a simpli-

cial reference mesh cell K̂, affine mappings {FK }, and an unisolvent set of func-
tionals on K̂. Using (5.1) and (5.2), one obtains a local finite element space on
each non-degenerated simplex. The set of these local spaces is called affine family
of simplicial finite elements. 2
Definition 5.23 Polynomial space Pk . Let x = (x1 , . . . , xd )T , k ∈ N ∪ {0}, and
α = (α1 , . . . , αd )T . Then, the polynomial space Pk is given by
( d d
)
Y X
αi α
Pk = span xi = x : αi ∈ N ∪ {0} for i = 1, . . . , d, αi ≤ k .
i=1 i=1

2
Remark 5.24 Lagrangian finite elements. In all examples given below, the linear
functionals on the reference mesh cell K̂ are the values of the polynomials with the
same barycentric coordinates as on the general mesh cell K. Finite elements whose
linear functionals are values of the polynomials on certain points in K are called
Lagrangian finite elements. 2
Example 5.25 P0 : piecewise constant finite element. The piecewise constant
finite element space consists of discontinuous functions. The linear functional is the
value of the polynomial in the barycenter of the mesh cell, see Figure 5.4. It is
dim P0 (K) = 1. 2

Figure 5.4: The finite element P0 (K).

Example 5.26 P1 : conforming piecewise linear finite element. This finite element
space is a subspace of C(Ω). The linear functionals are the values of the function
in the vertices of the mesh cells, see Figure 5.5. It follows that dim P1 (K) = d + 1.

65
 Figure 5.5: The finite element P1 (K).

The local basis for the functionals {Φi (v) = v(ai ), i = 1, . . . , d + 1}, is {λi }d+1
i=1
since Φi (λj ) = δij , see Remark 5.20. Since a local basis exists, the functionals are
unisolvent with respect to the polynomial space P1 (K).
Now, it will be shown that the corresponding finite element space consists of
continuous functions. Let K1 , K2 be two mesh cells with the common face E and
let v ∈ P1 (= S). The restriction of vK1 on E is a linear function on E as well as the
restriction of vK2 on E. It has to be shown that both linear functions are identical.
A linear function on the (d − 1)-dimensional face E is uniquely determined with d
linearly independent functionals which are defined on E. These functionals can be
chosen to be the values of the function in the d vertices of E. The functionals in S
are continuous, by the definition of S. Thus, it must hold that both restrictions on
E have the same values in the vertices of E. Hence, it is vK1 |E = vK2 |E and the
functions from P1 are continuous. 2

Example 5.27 P2 : conforming piecewise quadratic finite element. This finite ele-
ment space is also a subspace of C(Ω). It consists of piecewise quadratic functions.
The functionals are the values of the functions in the d + 1 vertices of the mesh cell
and the values of the functions in the centers of the edges,
Pd see Figure 5.6. Since
each vertex is connected to each other vertex, there are i=1 i = d(d + 1)/2 edges.
Hence, it follows that dim P2 (K) = (d + 1)(d + 2)/2.

Figure 5.6: The finite element P2 (K).

The part of the local basis which belongs to the functionals {Φi (v) = v(ai ),
i = 1, . . . , d + 1}, is given by

{φi (λ) = λi (2λi − 1), i = 1, . . . , d + 1}.

Denote the center of the edges between the vertices ai and aj by aij . The corre-

66
sponding part of the local basis is given by

{φij = 4λi λj , i, j = 1, . . . , d + 1, i < j}.

The unisolvence follows from the fact that there exists a local basis. The continuity
of the corresponding finite element space is shown in the same way as for the P1
finite element. The restriction of a quadratic function in a mesh cell to a face E is
a quadratic function on that face. Hence, the function on E is determined uniquely
with d(d + 1)/2 linearly independent functionals on E.
The functions φij are called in two dimensions edge bubble functions. 2

Example 5.28 P3 : conforming piecewise cubic finite element. This finite element
space consists of continuous piecewise cubic functions. It is a subspace of C(Ω).
The functionals in a mesh cell K are defined to be the values in the vertices ((d + 1)
values), two values on each edge (dividing the edge in three parts of equal length)
Pd
(2 i=1 i = d(d + 1) values), and the values in the barycenter of the 2-faces of K,
see Figure 5.7. Each 2-face of K is defined by three vertices. If one considers for
each vertex all possible pairs with other vertices, then each 2-face is counted three
times. Hence, there are (d + 1)(d − 1)d/6 2-faces. The dimension of P3 (K) is given
by
(d − 1)d(d + 1) (d + 1)(d + 2)(d + 3)
dim P3 (K) = (d + 1) + d(d + 1) + = .
6 6

Figure 5.7: The finite element P3 (K).

For the functionals

n
Φi (v) = v(ai ), i = 1, . . . , d + 1, (vertex),
Φiij (v) = v(aiij ), i, j = 1, . . . , d + 1, i 6= j, (point on edge), o
Φijk (v) = v(aijk ), i = 1, . . . , d + 1, i < j < k (point on 2-face) ,

the local basis is given by

(
1
φi (λ) = λi (3λi − 1)(3λi − 2),
2
9
φiij (λ) = λi λj (3λi − 1),
2 )
φijk (λ) = 27λi λj λk .

In two dimensions, the function φijk (λ) is called cell bubble function. 2

67
Example 5.29 Cubic Hermite element. The finite element space is a subspace of
C(Ω), its dimension is (d + 1)(d + 2)(d + 3)/6 and the functionals are the values
of the function in the vertices of the mesh cell ((d + 1) values), the value of the
barycenter at the 2-faces of K ((d + 1)(d − 1)d/6 values), and the partial derivatives
at the vertices (d(d + 1) values), see Figure 5.8. The dimension is the same as for
the P3 element. Hence, the local polynomials can be defined to be cubic.

Figure 5.8: The cubic Hermite element.

This finite element does not define an affine family in the strict sense, because
the functionals for the partial derivatives Φ̂i (v̂) = ∂i v̂(0) on the reference cell are
mapped to the functionals Φi (v) = ∂ti v(a), where a = FK (0) and ti are the di-
rections of edges which are adjacent to a, i.e., a is an end point of this edge. This
property suffices to control all first derivatives. On has to take care of this property
in the implementation of this finite element.
Because of this property, one can use the derivatives in the direction of the edges
as functionals
Φi (v) = v(ai ), (vertices)
Φij (v) = ∇v(ai ) · (aj − ai ), i, j = 1, . . . , d − 1, i 6= j, (directional derivative)
Φijk (v) = v(aijk ), i < j < k, (2-faces)

with the corresponding local basis

−2λ3i + 3λ2i − 7λi j<k,j6=i,k6=i λj λk ,

P
φi (λ) =
φij (λ) = λi λj (2λi − λj − 1),
φijk (λ) = 27λi λj λk .

The proof of the unisolvence can be found in the literature.

Here, the continuity of the functions will be shown only for d = 2. Let K1 , K2
be two mesh cells with the common edge E and the unit tangential vector t. Let
V1 , V2 be the end points of E. The restriction v|K1 , v|K2 to E satisfy four conditions

v|K1 (Vi ) = v|K2 (Vi ), ∂t v|K1 (Vi ) = ∂t v|K2 (Vi ), i = 1, 2.

Since both restrictions are cubic polynomials and four conditions have to be satis-
fied, their values coincide on E.
The cubic Hermite finite element possesses an advantage in comparison with the
P3 finite element. For d = 2, it holds for a regular triangulation T h that

#(K) ≈ 2#(V ), #(E) ≈ 2#(V ),

where #(·) denotes the number of triangles, nodes, and edges, respectively. Hence,
the dimension of P3 is approximately 7#(V ), whereas the dimension of the cubic

68
Hermite element is approximately 5#(V ). This difference comes from the fact that
both spaces are different. The elements of both spaces are continuous functions,
but for the functions of the cubic Hermite finite element, in addition, the first
derivatives are continuous at the nodes. That means, these two spaces are different
finite element spaces whose degree of the local polynomial space is the same (cubic).
One can see at this example the importance of the functionals for the definition of
the global finite element space. 2
Example 5.30 P1nc : nonconforming linear finite element, Crouzeix–Raviart finite
element Crouzeix and Raviart (1973). This finite element consists of piecewise
linear but discontinuous functions. The functionals are given by the values of the
functions in the barycenters of the faces such that dim P1nc (K) = (d + 1). It follows
from the definition of the finite element space, Definition 5.12, that the functions
from P1nc are continuous in the barycenter of the faces
P1nc = v ∈ L2 (Ω) : v|K ∈ P1 (K), v(x) is continuous at the barycenter


of all faces . (5.3)
Equivalently, the functionals can be defined to be the integral mean values on the
faces and then the global space is defined to be
(
P1nc = v ∈ L2 (Ω) : v|K ∈ P1 (K),

Z Z )
0
v|K ds = v|K 0 ds ∀ E ∈ E(K) ∩ E(K ) , (5.4)
E E

where E(K) is the set of all (d − 1) dimensional faces of K.

Figure 5.9: The finite element P1nc .

For the description of this finite element, one defines the functionals by
Φi (v) = v(ai−1,i+1 ) for d = 2, Φi (v) = v(ai−2,i−1,i+1 ) for d = 3,
where the points are the barycenters of the faces with the vertices that correspond
to the indices. This system is unisolvent with the local basis
φi (λ) = 1 − dλi , i = 1, . . . , d + 1.
2

5.3 Finite Elements on Parallelepipeds

Remark 5.31 Reference mesh cells, reference map. On can find in the literature
two reference cells: the unit cube [0, 1]d and the large unit cube [−1, 1]d . It does

69
not matter which reference cell is chosen. Here, the large unit cube will be used:
K̂ = [−1, 1]d . The class of admissible reference maps {FK } consists of bijective
affine mappings of the form

FK x̂ = B x̂ + b, B ∈ Rd×d , b ∈ Rd .

If B is a diagonal matrix, then K̂ is mapped to d-rectangles.

The class of mesh cells which are obtained in this way is not sufficient to tri-
angulate general domains. If one wants to use more general mesh cells than par-
allelepipeds, then the class of admissible reference maps has to be enlarged, see
Section 5.4. 2

Definition 5.32 Polynomial space Qk . Let x = (x1 , . . . , xd )T and denote by

α = (α1 , . . . , αd )T a multi-index. Then, the polynomial space Qk is given by
( d )
Y
αi α
Qk = span xi = x : 0 ≤ αi ≤ k for i = 1, . . . , d .
i=1

Example 5.33 Q1 vs. P1 . The space Q1 consists of all polynomials which are
d-linear. Let d = 2, then it is

Q1 = span{1, x, y, xy},

whereas
P1 = span{1, x, y}.
2

Remark 5.34 Finite elements on d-rectangles. For simplicity of presentation, the

examples below consider d-rectangles. In this case, the finite elements are just tensor
products of one-dimensional finite elements. In particular, the basis functions can
be written as products of one-dimensional basis functions. 2

Example 5.35 Q0 : piecewise constant finite element. Similarly to the P0 space,

the space Q0 consists of piecewise constant, discontinuous functions. The functional
is the value of the function in the barycenter of the mesh cell K and it holds
dim Q0 (K) = 1. 2

Example 5.36 Q1 : conforming piecewise d-linear finite element. This finite ele-
ment space is a subspace of C(Ω). The functionals are the values of the function in
the vertices of the mesh cell, see Figure 5.10. Hence, it is dim Q1 (K) = 2d .
The one-dimensional local basis functions, which will be used for the tensor
product, are given by
1 1
φ̂1 (x̂) = (1 − x̂), φ̂2 (x̂) = (1 + x̂).
2 2
With these functions, e.g., the basis functions in two dimensions are computed by

φ̂1 (x̂)φ̂1 (ŷ), φ̂1 (x̂)φ̂2 (ŷ), φ̂2 (x̂)φ̂1 (ŷ), φ̂2 (x̂)φ̂2 (ŷ).

The continuity of the functions of the finite element space Q1 is proved in the
same way as for simplicial finite elements. It is used that the restriction of a function
from Qk (K) to a face E is a function from the space Qk (E), k ≥ 1. 2

70
 Figure 5.10: The finite element Q1 .

Figure 5.11: The finite element Q2 .

Example 5.37 Q2 : conforming piecewise d-quadratic finite element. It holds

that Q2 ⊂ C(Ω). The functionals in one dimension are the values of the function
at both ends of the interval and in the center of the interval, see Figure 5.11. In d
dimensions, they are the corresponding values of the tensor product of the intervals.
It follows that dim Q2 (K) = 3d .
The one-dimensional basis function on the reference interval are defined by
1 1
φ̂1 (x̂) = − x̂(1 − x̂), φ̂2 (x̂) = (1 − x̂)(1 + x̂), φ̂3 (x̂) = (1 + x̂)x̂.
2 2
Qd
The basis function i=1 φ̂2 (x̂i ) is called cell bubble function. 2

Example 5.38 Q3 : conforming piecewise d-quadratic finite element. This finite

element space is a subspace of C(Ω). The functionals on the reference interval are
given by the values at the end of the interval and the values at the points x̂ = −1/3,
x̂ = 1/3. In multiple dimensions, it is the corresponding tensor product, see Figure
5.12. The dimension of the local space is dim Q3 (K) = 4d .
The one-dimensional basis functions in the reference interval are given by
1
φ̂1 (x̂) = − (3x̂ + 1)(3x̂ − 1)(x̂ − 1),
16
9
φ̂2 (x̂) = (x̂ + 1)(3x̂ − 1)(x̂ − 1),
16
9
φ̂3 (x̂) = − (x̂ + 1)(3x̂ + 1)(x̂ − 1),
16
1
φ̂4 (x̂) = (3x̂ + 1)(3x̂ − 1)(x̂ + 1).
16

71
 Figure 5.12: The finite element Q3 .

2
Example 5.39 Qrot 1 : rotated nonconforming element of lowest order, Rannacher–
Turek element Rannacher and Turek (1992): This finite element space is a gen-
eralization of the P1nc finite element to quadrilateral and hexahedral mesh cells. It
consists of discontinuous functions which are continuous at the barycenter of the
faces. The dimension of the local finite element space is dim Qrot 1 (K) = 2d. The
space on the reference mesh cell is defined by
 
p̂ : p̂ ∈ span{1, x̂, ŷ, x̂2 − ŷ 2 }

Qrot
1 K̂ = for d = 2,
 
2 2 2 2

Qrot
1 K̂ = p̂ : p̂ ∈ span{1, x̂, ŷ, ẑ, x̂ − ŷ , ŷ − ẑ } for d = 3.

Note that the transformed space

−1
Qrot rot
1 (K) = {p = p̂ ◦ FK , p̂ ∈ Q1 (K̂)}

contains polynomials of the form ax2 − by 2 , where a, b depend on FK .

Figure 5.13: The finite element Qrot

1 .

For d = 2, the local basis on the reference cell is given by

3 1 1
φ1 (x̂, ŷ) = − (x̂2 − ŷ 2 ) − ŷ + ,
8 2 4
3 2 2 1 1
φ2 (x̂, ŷ) = (x̂ − ŷ ) + x̂ + ,
8 2 4
3 2 2 1 1
φ3 (x̂, ŷ) = − (x̂ − ŷ ) + ŷ + ,
8 2 4
3 2 2 1 1
φ4 (x̂, ŷ) = (x̂ − ŷ ) − x̂ + .
8 2 4

72
 Analogously to the Crouzeix–Raviart finite element, the functionals can be de-
fined as point values of the functions in the barycenters of the faces, see Figure 5.13,
or as integral mean values of the functions at the faces. Consequently, the finite
element spaces are defined in the same way as (5.3) or (5.4), with P1nc (K) replaced
by Qrot
1 (K).
In the code MooNMD John and Matthies (2004), the mean value oriented
Qrot
1 finite element space is implemented fro two dimensions and the point value
oriented Qrot
1 finite element space for three dimensions. For d = 3, the integrals on
the faces of mesh cells, whose equality is required in the mean value oriented Qrot 1
finite element space, involve a weighting function which depends on the particular
mesh cell K. The computation of these weighting functions for all mesh cells is an
additional computational overhead. For this reason, Schieweck (Schieweck, 1997,
p. 21) suggested to use for d = 3 the simpler point value oriented form of the Qrot 1
finite element. 2

5.4 Parametric Finite Elements on General d-Di-

mensional Quadrilaterals
Remark 5.40 Parametric mappings. The image of an affine mapping of the refer-
ence mesh cell K̂ = [−1, 1]d , d ∈ {2, 3}, is a parallelepiped. If one wants to consider
finite elements on general q-quadrilaterals, then the class of admissible reference
maps has to be enlarged.
The simplest parametric finite element on quadrilaterals in two dimensions uses
bilinear mappings. Let K̂ = [−1, 1]2 and let
 1   
FK (x̂) a11 + a12 x̂ + a13 ŷ + a14 x̂ŷ i
FK (x̂) = 2 = , FK ∈ Q1 , i = 1, 2,
FK (x̂) a21 + a22 x̂ + a23 ŷ + a24 x̂ŷ

be a bilinear mapping from K̂ on the class of admissible quadrilaterals. A quadri-

lateral K is called admissible if
• the length of all edges of K is larger than zero,
• the interior angles of K are smaller than π, i.e. K is convex.
This class contains, e.g., trapezoids and rhombi. 2

Remark 5.41 Parametric finite element functions. The functions of the local
−1
space P (K) on the mesh cell K are defined by p = p̂ ◦ FK . These functions
are in general rational functions. However, using d-linear mappings, then the re-
striction of FK on an edge of K̂ is an affine map. For instance, in the case of the
Q1 finite element, the functions on K are linear functions on each edge of K for
this reason. It follows that the functions of the corresponding finite element space
are continuous, see Example 5.26. 2

5.5 Transform of Integrals

Remark 5.42 Motivation. The transform of integrals from the reference mesh
cell to mesh cells of the grid and vice versa is used as well for analysis as for the
implementation of finite element methods. This section provides an overview of the
most important formulae for transforms.
Let K̂ ⊂ Rd be the reference mesh cell, K be an arbitrary mesh cell, and
FK : K̂ → K with x = FK (x̂) be the reference map. It is assumed that the
reference map is a continuous differentiable one-to-one map. The inverse map is

73
 −1
denoted by FK : K → K̂. For the integral transforms, the derivatives (Jacobians)
−1
of FK and FK are needed
∂xi −1 ∂ x̂i
DFK (x̂)ij = , DFK (x)ij = , i, j = 1, . . . , d.
∂ x̂j ∂xj
2
Remark 5.43 Integral with a function without derivatives. This integral trans-
forms with the standard rule of integral transforms
Z Z
v(x) dx = v̂(x̂) |det DFK (x̂)| dx̂, (5.5)
K K̂

where v̂(x̂) = v(FK (x̂)). 2

Remark 5.44 Transform of derivatives. Using the chain rule, one obtains
d
∂v X ∂v̂ ∂ x̂j 
−1

T 
(x) = (x̂) = ∇x̂ v̂(x̂) · DFK (x)
∂xi j=1
∂ x̂j ∂xi i

T 
−1
= ∇x̂ v̂(x̂) · DFK (FK (x̂)) , (5.6)
i
d
∂v̂ X ∂v ∂xj 
T

(x̂) = (x) = ∇v(x) · (DFK (x̂))
∂ x̂ j=1
∂xj ∂ x̂i i

T 
−1
= ∇v(x) · DFK (FK (x)) . (5.7)
i
The index i denotes the i-th row of a matrix. Derivatives on the reference mesh cell
are marked with a symbol on the operator. 2
Remark 5.45 Integrals with a gradients. Using the rule for transforming integrals
and (5.6) gives
Z
b(x) · ∇v(x) dx
K
Z h i
−1 T


= b (FK (x̂)) · DFK (FK (x̂)) ∇x̂ v̂(x̂) |det DFK (x̂)| dx̂. (5.8)
K̂
Similarly, one obtains
Z
∇v(x) · ∇w(x) dx
K
Z h i h i
−1 T −1 T



= DFK (FK (x̂)) ∇x̂ v̂(x̂) · DFK (FK (x̂)) ∇x̂ ŵ(x̂)
K̂
× |det DFK (x̂)| dx̂. (5.9)
2
Remark 5.46 Integral with the divergence. Integrals of the following type are
important for the Navier–Stokes equations
Z Z X d
∂vi
∇ · v(x)q(x) dx = (x)q(x) dx
K K i=1 ∂x i
Z X d h
−1

T  i
= DFK (FK (x̂)) · ∇x̂ v̂i (x̂) q̂(x̂) |det DFK (x̂)| dx̂
K̂ i=1 i
Z h i
−1

T
= DFK (FK (x̂)) : Dx̂ v(x̂) q̂(x̂) |det DFK (x̂)| dx̂. (5.10)
K̂

In the derivation, (5.6) was used. 2

74
Example 5.47 Affine transform. The most important class of reference maps are
affine transforms
x = B x̂ + b, B ∈ Rd×d , b ∈ Rd ,
where the invertible matrix B and the vector b are constants. It follows that

x̂ = B −1 (x − b) = B −1 x − B −1 b.

In this case, there are

−1
DFK = B, DFK = B −1 , det DFK = det(B).

One obtains for the integral transforms from (5.5), (5.8), (5.9), and (5.10)
Z Z
v(x) dx = |det(B)| v̂(x̂) dx̂, (5.11)
K K̂
Z Z
b(x) · ∇v(x) dx = |det(B)| b (FK (x̂)) · B −T ∇x̂ v̂(x̂) dx̂, (5.12)
Z K ZK̂

∇v(x) · ∇w(x) dx = |det(B)| B −T ∇x̂ v̂(x̂) · B −T ∇x̂ ŵ(x̂) dx̂, (5.13)

K K̂
Z Z
 −T 
∇ · v(x)q(x) dx = |det(B)| B : Dx̂ v(x̂) q̂(x̂) dx̂. (5.14)
K K̂

75
Chapter 6

Interpolation

Remark 6.1 Motivation. Variational forms of partial differential equations use

functions in Sobolev spaces. The solution of these equations shall be approximated
with the Ritz method in finite dimensional spaces, the finite element spaces. The
best possible approximation of an arbitrary function from the Sobolev space by a
finite element function is a factor in the upper bound for the finite element error,
e.g., see the Lemma of Cea, estimate (4.19).
This section studies the approximation quality of finite element spaces. Esti-
mates are proved for interpolants of functions. Interpolation estimates are of course
upper bounds for the best approximation error and they can serve as factors in
finite element error estimates. 2

6.1 Interpolation in Sobolev Spaces by Polynomi-

als
Lemma 6.2 Unique determination of a polynomial with integral condi-
tions. Let Ω be a bounded domain in Rd with Lipschitz boundary. Let m ∈ N ∪ {0}
be given and let for all derivatives with multi-index α, |α| ≤ m, a value aα ∈ R be
given. Then, there is a uniquely determined polynomial p ∈ Pm (Ω) such that
Z
∂α p(x) dx = aα , |α| ≤ m. (6.1)
Ω

Proof: Let p ∈ Pm (Ω) be an arbitrary polynomial. It has the form

X
p(x) = bβ xβ .
|β|≤m

Inserting this representation into (6.1) leads to a linear system of equations M b = a with
Z
M = (Mαβ ), Mαβ = ∂α xβ dx, b = (bβ ), a = (aα ),
Ω

for |α| , |β| ≤ m. Since M is a squared matrix, the linear system of equations possesses a
unique solution if and only if M is non-singular.
The proof is performed by contradiction. Assume that M is singular. Then there
exists a non-trivial solution of the homogeneous system. That means, there is a polynomial
q ∈ Pm (Ω) \ {0} with Z
∂α q(x) dx = 0 for all |α| ≤ m.
Ω

The polynomial q(x) has the representation q(x) = |β|≤m cβ xβ . Now, one can choose a
P

cβ 6= 0 with maximal value |β|. Then, it is ∂β q(x) = Ccβ = const 6= 0, where C > 0 comes

76
from the differentiation rule for polynomials, which is a contradiction to the vanishing of
the integral for ∂β q(x).

Remark 6.3 To Lemma 6.2. Lemma 6.2 states that a polynomial is uniquely
determined if a condition on the integral on Ω is prescribed for each derivative. 2
Lemma 6.4 Poincaré-type inequality. Denote by Dk v(x), k ∈ N ∪ {0}, the
total derivative of order k of a function v(x), e.g., for k = 1 the gradient of v(x).
Let Ω be convex and be included into a ball of radius R. Let k, l ∈ N ∪ {0} with
k ≤ l and let p ∈ R with p ∈ [1, ∞]. Assume that v ∈ W l,p (Ω) satisfies
Z
∂α v(x) dx = 0 for all |α| ≤ l − 1,
Ω

then it holds the estimate

k l−k l

D v p
L (Ω)
≤ CR D v Lp (Ω)
,

where the constant C does not depend on Ω and on v(x).

Proof: There is nothing to prove if k = l. In addition, it suffices to prove the lemma
for k = 0 and l = 1, since the general case follows by applying the result to ∂α v(x). Only
the case p < ∞ will be discussed here in detail.
Since Ω is assumed to be convex, the integral mean value theorem can be written in
the form Z 1
v(x) − v(y) = ∇v(tx + (1 − t)y) · (x − y) dt, x, y ∈ Ω.
0
Integration with respect to y yields
Z Z Z Z 1
v(x) dy − v(y) dy = ∇v(tx + (1 − t)y) · (x − y) dt dy.
Ω Ω Ω 0

It follows from the assumption that the second integral on the left hand side vanishes.
Hence, one gets
Z Z 1
1
v(x) = ∇v(tx + (1 − t)y) · (x − y) dt dy.
|Ω| Ω 0
Now, taking the absolute value on both sides, using that the absolute value of an integral is
estimated from above by the integral of the absolute value, applying the Cauchy–Schwarz
inequality for vectors and the estimate kx − yk2 ≤ 2R yields
Z Z 1 
1  
|v(x)| = ∇v(tx + (1 − t)y) · (x − y) dt dy 
|Ω|  Ω 0 
Z Z 1
1
≤ |∇v(tx + (1 − t)y) · (x − y)| dt dy
|Ω| Ω 0
Z Z 1
2R
≤ k∇v(tx + (1 − t)y)k2 dt dy. (6.2)
|Ω| Ω 0
Then (6.2) is raised to the power p and then integrated with respect to x. One obtains
with Hölder’s inequality (3.4), with p−1 + q −1 = 1 =⇒ p/q − p = p(1/q − 1) = −1, that
Z Z Z 1 p
CRp
Z
|v(x)|p dx ≤ k∇v(tx + (1 − t)y)k 2 dt dy dx
Ω |Ω|p Ω Ω 0
" Z Z 1 p/q
CRp
Z
q
≤ 1 dt dy
|Ω|p Ω
| Ω 0 {z }
|Ω|p/q
Z Z 1 #
× k∇v(tx + (1 − t)y)kp2 dt dy dx
Ω 0
1
CRp
Z Z Z 
= k∇v(tx + (1 − t)y)kp2 dt dy dx.
|Ω| Ω Ω 0

77
Applying the theorem of Fubini allows the commutation of the integration

CRp 1
Z Z Z Z 
|v(x)|p dx ≤ k∇v(tx + (1 − t)y)kp2 dy dx dt.
Ω |Ω| 0 Ω Ω

Using the integral mean value theorem in one dimension gives that there is a t0 ∈ [0, 1],
such that
CRp
Z Z Z 
|v(x)|p dx ≤ k∇v(t0 x + (1 − t0 )y)kp2 dy dx.
Ω |Ω| Ω Ω

The function k∇v(x)kp2 will be extended to Rd by zero and the extension will be also
denoted by k∇v(x)kp2 . Then, it is

CRp
Z Z Z 
p p
|v(x)| dx ≤ k∇v(t0 x + (1 − t0 )y)k2 dy dx. (6.3)
Ω |Ω| Ω Rd

Let t0 ∈ [0, 1/2]. Since the domain of integration is Rd , a substitution of variables

t0 x + (1 − t0 )y = z can be applied and leads to
Z Z
1
k∇v(t0 x + (1 − t0 )y)kp2 dy = k∇v(z)kp2 dz ≤ 2 k∇vkpLp (Ω) ,
Rd 1 − t0 Rd
since 1/(1 − t0 ) ≤ 2. Inserting this expression into (6.3) gives
Z
|v(x)|p dx ≤ 2CRp k∇vkpLp (Ω) .
Ω

If t0 > 1/2 then one changes the roles of x and y, applies the theorem of Fubini to
change the sequence of integration, and uses the same arguments.
The estimate for the case p = ∞ is also based on (6.2).

Remark 6.5 On Lemma 6.4. The Lemma 6.4 proves an inequality of Poincaré-
type. It says that it is possible to estimate the Lp (Ω) norm of a lower derivative of
a function v(x) by the same norm of a higher derivative if the integral mean values
of some lower derivatives vanish.
An important application of Lemma 6.4 is in the proof of the Bramble–Hilbert
lemma. The Bramble–Hilbert lemma considers a continuous linear functional which
is defined on a Sobolev space and which vanishes for all polynomials of degree less
or equal than m. It states that the value of the functional can be estimated by the
Lebesgue norm of the (m + 1)th total derivative of the functions from this Sobolev
space. 2

Theorem 6.6 Bramble–Hilbert lemma. Let m ∈ N ∪ {0}, m ≥ 0, p ∈ [1, ∞],

and F : W m+1,p (Ω) → R be a continuous linear functional, and let the conditions
of Lemma 6.2 and 6.4 be satisfied. Let

F (p) = 0 ∀ p ∈ Pm (Ω),

then there is a constant C(Ω), which is independent of v(x) and F , such that

|F (v)| ≤ C(Ω) Dm+1 v Lp (Ω) ∀ v ∈ W m+1,p (Ω).

Proof: Let v ∈ W m+1,p (Ω). It follows from Lemma 6.2 that there is a polynomial
from Pm (Ω) with Z
∂α (v + p)(x) dx = 0 for |α| ≤ m.
Ω
Lemma 6.4 gives, with l = m + 1 and considering each term in k·kW m+1,p (Ω) individually,
the estimate

kv + pkW m+1,p (Ω) ≤ C(Ω) Dm+1 (v + p) Lp (Ω) = C(Ω) Dm+1 v Lp (Ω) .

78
From the vanishing of F for p ∈ Pm (Ω) and the continuity of F it follows that

|F (v)| = |F (v + p)| ≤ c kv + pkW m+1,p (Ω) ≤ C(Ω) Dm+1 v Lp (Ω) .

Remark 6.7 Strategy for estimating the interpolation error. The Bramble–Hilbert
lemma will be used for estimating the interpolation error for an affine family of
finite elements. The strategy is as follows:
• Show first the estimate on the reference mesh cell K̂.
• Transform the estimate on an arbitrary mesh cell K to the reference mesh cell K̂.
• Apply the estimate on K̂.
• Transform back to K.
One has to study what happens if the transforms are applied to the estimate. 2

Remark 6.8 Assumptions, definition of the interpolant. Let K̂ ⊂ Rd , d ∈ {2, 3},

be a reference mesh cell (compact polyhedron), P̂ (K̂) a polynomial space of dimen-
sion N , and Φ̂1 , . . . , Φ̂N : C s (K̂) → R continuous linear functionals. It will be
assumed that the space P̂ (K̂) is unisolvent with respect to these functionals. Then,
there is a local basis φ̂1 , . . . , φ̂N ∈ P̂ (K̂).
Consider v̂ ∈ C s (K̂), then the interpolant IK̂ v̂ ∈ P̂ (K̂) is defined by
N
X
IK̂ v̂(x̂) = Φ̂i (v̂)φ̂i (x̂).
i=1

The operator IK̂ is a continuous and linear operator from C s (K̂) to P̂ (K̂). From
the linearity it follows that IK̂ is the identity on P̂ (K̂)

IK̂ p̂ = p̂ ∀ p̂ ∈ P̂ (K̂).

Example 6.9 Interpolation operators.

• Let K̂ ⊂ Rd be an arbitrary reference cell, P̂ (K̂) = P0 (K̂), and
Z
1
Φ̂(v̂) =  
  v̂(x̂) dx̂.
K̂  K̂

The functional Φ̂ is continuous on C 0 (K̂) since

 
K̂ 
Z  
 1
 
Φ̂(v̂) ≤   |v̂(x̂)| dx̂ ≤   max |v̂(x̂)| = kv̂kC 0 (K̂) .

K̂  K̂ K̂  x̂∈K̂

For the constant function 1 ∈ P0 (K̂) it is Φ̂(1) = 1 6= 0. Hence, {φ̂} = {1} is

the local basis and the space is unisolvent with respect to Φ̂. The operator
Z
1
IK̂ v̂(x̂) = Φ̂(v̂)φ̂(x̂) =  
  v̂(x̂) dx̂
K̂  K̂

is an integral mean value operator, i.e., each continuous function on K̂ will be

approximated by a constant function whose value equals the integral mean value,
see Figure 6.1

79
Figure 6.1: Interpolation of x2 in [−1, 1] by a P0 function with the integral mean
value and with the value of the function at x0 = 0.

• It is possible to define Φ̂(v̂) = v̂(x̂0 ) for an arbitrary point x̂0 ∈ K̂. This
functional is also linear and continuous in C 0 (K̂). The interpolation operator IK̂
defined in this way interpolates each continuous function by a constant function
whose value is equal to the value of the function at x̂0 , see also Figure 6.1.
Interpolation operators which are defined by using values of functions, are called
Lagrangian interpolation operators.
This example demonstrates that the interpolation operator IK̂ depends on P̂ (K̂)
and on the functionals Φ̂i . 2

Theorem 6.10 Interpolation error estimate on a reference mesh cell. Let

Pm (K̂) ⊂ P̂ (K̂) and p ∈ [1, ∞] with (m + 1 − s)p > d. Then there is a constant C
that is independent of v̂(x̂) such that

≤ C Dm+1 v̂ Lp (K̂) ∀ v̂ ∈ W m+1,p (K̂).

v̂ − I v̂ m+1,p (6.4)
K̂ W (K̂)

Proof: Because of the Sobolev imbedding, Theorem 3.53, (λ = 0, j = s, m (of Sobolev

imbedding) = m + 1 − s) it holds that

W m+1,p (K̂) → C s (K̂)

if (m + 1 − s)p > d. That means, the interpolation operator is well defined in W m+1,p (K̂).
From the identity of the interpolation operator in Pm (K̂), the triangle inequality, the
boundedness of the interpolation operator (it is a linear and continuous operator mapping
C s (K̂) → P̂ (K̂) ⊂ W m+1,p (K̂)), and the Sobolev imbedding, one obtains for q̂ ∈ Pm (K̂)

kv̂ − IK̂ v̂kW m+1,p (K̂) = kv̂ + q̂ − IK̂ (v̂ + q̂)kW m+1,p (K̂)
≤ kv̂ + q̂kW m+1,p (K̂) + kIK̂ (v̂ + q̂)kW m+1,p (K̂)
≤ kv̂ + q̂kW m+1,p (K̂) + c kv̂ + q̂kC s (K̂)
≤ c kv̂ + q̂kW m+1,p (K̂) .

Choosing q̂(x̂) in Lemma 6.2 such that

Z
∂α (v̂ + q̂) dx̂ = 0 ∀ |α| ≤ m,
K̂

the assumptions of Lemma 6.4 are satisfied. It follows that

kv̂ + q̂kW m+1,p (K̂) ≤ c Dm+1 (v̂ + q̂) Lp (K̂) = c Dm+1 v̂ Lp (K̂) .

Remark 6.11 On Theorem 6.10.

80
 • One can construct examples which show that the Sobolev imbedding is not valid
if (m + 1 − s)p > d is not satisfied. In the case (m + 1 − s)p ≤ d, the statement
of Theorem 6.10 is not true.
Consider the interpolation of continuous functions (s = 0) with piecewise linear
elements (m = 1) in Sobolev spaces that are also Hilbert spaces (p = 2). Then
(m+1−s)p = 4 and it follows that the theorem can be applied only for d ∈ {2, 3}.
For piecewise constant finite elements, the statement of the theorem is true only
for d = 1.
• The theorem requires only that Pm (K̂) ⊂ P̂ (K̂). This requirement does not
exclude that P̂ (K̂) contains polynomials of higher degree, too. However, this
property is not utilized and also not needed if the other assumptions of the
theorem are satisfied.
2

Remark 6.12 Assumptions on the triangulation. For deriving the interpolation

error estimate for arbitrary mesh cells K, and finally for the finite element space,
one has to study the properties of the affine mapping from K to K̂ and of the back
mapping.
Consider an affine family of finite elements whose mesh cells are generated by
affine mappings
FK x̂ = B x̂ + b,
where B is a non-singular d × d matrix and b is a d vector.
Let hK be the diameter of K = FK (K̂), i.e., the largest distance of two points
that are contained in K. The images {K = FK (K̂)} are assumed to satisfy the
following conditions:
• K ⊂ Rd is contained in a ball of radius CR hK ,
−1
• K contains a ball of radius CR hK ,
where the constant CR is independent of K. Hence, it follows for all K that
radius of circumcircle 2
≤ CR .
radius of inscribed circle
A triangulation with this property is called a quasi-uniform triangulation. 2

Lemma 6.13 Estimates of matrix norms. For each matrix norm k·k one has
the estimates
kBk ≤ chK , B −1 ≤ ch−1

K ,

where the constants depend on the matrix norm and on CR .

Proof: Since K̂ is a Lipschitz domain with polyhedral boundary, it contains a ball
B(x̂0 , r) with x̂0 ∈ K̂ and some r > 0. Hence, x̂0 + ŷ ∈ K̂ for all kŷk2 = r. It follows that
the images
x0 = B x̂0 + b, x = B(x̂0 + ŷ) + b = x0 + B ŷ
are contained in K. Since the triangulation is assumed to be quasi-uniform, one obtains
for all ŷ
kB ŷk2 = kx − x0 k2 ≤ CR hK .
Now, it holds for the spectral norm that
kBẑk2 1 CR
kBk2 = sup = sup kBẑk2 ≤ hK .
ẑ6=0 kẑk2 r kẑk2 =r r

An estimate of this form, with a possible different constant, holds also for all other matrix
norms since all matrix norms
are equivalent.
The estimate for B −1 proceeds in the same way with interchanging the roles of K
and K̂.

81
Theorem 6.14 Local interpolation estimate. Let an affine family of finite
elements be given by its reference cell K̂, the functionals {Φ̂i }, and a space of
polynomials P̂ (K̂). Let all assumptions of Theorem 6.10 be satisfied. Then, for all
v ∈ W m+1,p (K) there is a constant C, which is independent of v(x) such that

≤ Chm+1−k
k m+1
D (v − IK v) p D v Lp (K) , k ≤ m + 1. (6.5)
L (K) K

Proof: The idea of the proof consists in transforming left hand side of (6.5) to the
reference cell, using the interpolation estimate on the reference cell and transforming back.
(−1)
i). Denote the elements of the matrices B and B −1 by bij and bij , respectively.
Since kBk∞ = maxi,j |bij | is also a matrix norm, it holds that
 
|bij | ≤ ChK , bij  ≤ Ch−1
 (−1) 
K . (6.6)

Using element-wise estimates for the matrix B (Leibniz formula for determinants), one
obtains
|det B| ≤ ChdK , det B −1  ≤ Ch−d
 
K . (6.7)
ii). The next step consists in proving that the transformed interpolation operator is
equal to the natural interpolation operator on K. The latter one is given by
N
X
IK v = ΦK,i (v)φK,i , (6.8)
i=1

where {φK,i } is the basis of the space

−1
P (K) = {p : K → R : p = p̂ ◦ FK , p̂ ∈ P̂ (K̂)},

which satisfies ΦK,i (φK,j ) = δij . The functionals are defined by

ΦK,i (v) = Φ̂i (v ◦ FK )

−1
Hence, it follows with v = φ̂j ◦ FK from the condition on the local basis on K̂ that
−1
ΦK,i (φ̂j ◦ FK ) = Φ̂i (φ̂j ) = δij ,
−1
i.e., the local basis on K is given by φK,j = φ̂j ◦ FK . Using (6.8), one gets
N N N
!
X X −1
X
IK̂ v̂ = Φ̂i (v̂)φ̂i = ΦK,i (v̂ ◦ φK,i ◦ FK =
FK ) ΦK,i (v)φK,i ◦ FK
i=1 i=1
| {z } i=1
=v
= IK v ◦ FK .

Hence, IK̂ v̂ is transformed correctly.

iii). One obtains with the chain rule
d d
∂v(x) X ∂v̂(x̂) (−1) ∂v̂(x̂) X ∂v(x)
= b , = bji .
∂xi j=1
∂ x̂j ji ∂ x̂i j=1
∂xj

It follows with (6.6) that (with each derivative one obtains an additional factor of B or
B −1 , respectively)

k −k k k k k
Dx v(x) ≤ ChK Dx̂ v̂(x̂) , Dx̂ v̂(x̂) ≤ ChK Dx v(x) .

2 2 2 2

One gets with (6.7)

Z p Z p Z p
k −kp k −kp+d k
Dx v(x) dx ≤ ChK |det B| Dx̂ v̂(x̂) dx̂ ≤ ChK Dx̂ v̂(x̂) dx̂

K 2 K̂ 2 K̂ 2

and
Z p Z p Z p
k kp −1 k kp−d k
 
Dx̂ v̂(x̂) dx̂ ≤ ChK det B  Dx v(x) dx ≤ ChK Dx v(x) dx.

K̂ 2 K 2 K 2

82
Using now the interpolation estimate on the reference cell (6.4) yields
p p
k
≤ C Dx̂m+1 v̂ Lp (K̂) , 0 ≤ k ≤ m + 1.

Dx̂ (v̂ − IK̂ v̂)

Lp (K̂)

It follows that
p p
k −kp+d k
Dx (v − IK v) p ≤ ChK Dx̂ (v̂ − IK̂ v̂) p

L (K) L (K̂)
−kp+d m+1 p

≤ ChK Dx̂ v̂ Lp (K̂)
(m+1−k)p m+1 p

≤ ChK Dx v Lp (K) .

Taking the p-th root proves the statement of the theorem.

Remark 6.15 On estimate (6.5).

• Note that the power of hK does not depend on p and d.
• Consider a quasi-uniform triangulation and define

h = max {hK }.
K∈T h

Then, one obtains by summing over all mesh cells an interpolation estimate for
the global finite element space
 1/p
k X p
Dk (v − IK v) p 

D (v − Ih v) p = 
L (Ω) L (K)
K∈T h
 1/p
X (m+1−k)p m+1 p
≤  chK D v Lp (K) 
K∈T h

≤ ch(m+1−k) Dm+1 v Lp (Ω) .

(6.9)

For linear finite elements P1 (m = 1) it is, in particular,

kv − Ih vkLp (Ω) ≤ ch2 D2 v Lp (Ω) , k∇(v − Ih v)kLp (Ω) ≤ ch D2 v Lp (Ω) ,

if v ∈ W 2,p (Ω).
2

Corollary 6.16 Finite element error estimate. Let u(x) be the solution of the
model problem (4.9) with u ∈ H m+1 (Ω) and let uh (x) be the solution of the corre-
sponding finite element problem. Consider a family of quasi-uniform triangulations
and let the finite element spaces V h contain polynomials of degree m. Then, the
following finite element error estimate holds
∇(u − uh ) 2 ≤ chm Dm+1 u L2 (Ω) = chm |u|H m+1 (Ω) .

L (Ω)
(6.10)

Proof: The statement follows by combining Lemma 4.13 (for V = H01 (Ω)) and (6.9)

h
∇(u − u ) ≤ inf ∇(u − v h ) ≤ k∇(u − Ih u)kL2 (Ω) ≤ chm |u|H m+1 (Ω) .

L2 (Ω) v h ∈V h L2 (Ω)

Remark 6.17 To (6.10). Note that Lemma 4.13 provides only information about
the error in the norm on the left-hand side of (6.10), but not in other norms. 2

83
6.2 Inverse Estimate
Remark 6.18 On inverse estimates. The approach for proving interpolation error
estimates can be uses also to prove so-called inverse estimates. In contrast to
interpolation error estimates, a norm of a higher order derivative of a finite element
function will be estimated by a norm of a lower order derivative of this function.
One obtains as penalty a factor with negative powers of the diameter of the mesh
cell. 2

Theorem 6.19 Inverse estimate. Let 0 ≤ k ≤ l be natural numbers and let

p, q ∈ [1, ∞]. Then there is a constant Cinv , which depends only on k, l, p, q, K̂, P̂ (K̂)
such that
l h (k−l)−d(p−1 −q −1 )
Dk v h p ∀ v h ∈ P (K).

D v q
L (K)
≤ Cinv hK L (K)
(6.11)

Proof: In the first step, (6.11) is shown for hK̂ = 1 and k = 0 on the reference mesh
cell. Since all norms are equivalent in finite dimensional spaces, one obtains

l h
D v̂ q ≤ v̂ h ≤ C v̂ h ∀ v̂ h ∈ P̂ (K̂).

L (K̂) l,q
W (K̂) p L (K̂)

If k > 0, then one sets

n o
P̃ (K̂) = ∂α v̂ h : v̂ h ∈ P̂ (K̂), |α| = k ,

which is also a space consisting of polynomials. The application of the first estimate of
the proof to P̃ (K̂) gives
X   X
l h l−k
D v̂ q = ∂α v̂ h ≤C h
α v̂

D q
∂ p
L (K̂) L (K̂) L (K̂)
|α|=k |α|=k

= C Dk v̂ h .

Lp (K̂)

This estimate is transformed to an arbitrary mesh cell K analogously as for the interpo-
lation error estimates. From the estimates for the transformations, one obtains

l h −l+d/q l h −l+d/q k h
D v q ≤ ChK D v̂ q ≤ ChK D v̂ p
L (K) L (K̂) L (K̂)

k−l+d/q−d/p k h
≤ Cinv hK D v .
Lp (K)

Remark 6.20 On the proof. The crucial point in the proof was the equivalence of
all norms in finite dimensional spaces. Such a property does not exist in infinite
dimensional spaces. 2

Corollary 6.21 Global inverse estimate. Let p = q and let T h be a regular

triangulation of Ω, then
l h k−l k h

D v p,h
L (Ω)
≤ C inv h D v Lp,h (Ω)
,

where  1/p
X p
k·kLp,h (Ω) =  k·kLp (K)  .
K∈T h

Remark 6.22 On k·kLp,h (Ω) . The cell wise definition of the norm is important
for l ≥ 2 since in this case finite element functions generally do not possess the
regularity for the global norm to be well defined. It is also important for l ≥ 1 and
non-conforming finite element functions. 2

84
6.3 Interpolation of Non-Smooth Functions
Remark 6.23 Motivation. The interpolation theory of Section 6.1 requires that
the interpolation operator is continuous on the Sobolev space to which the function
belongs that should be interpolated. But if one, e.g., wants to interpolate discon-
tinuous functions with continuous, piecewise linear elements, then Section 6.1 does
not provide estimates.
A simple remedy seems to be first to apply some smoothing operator to the
function to be interpolated and then to interpolate the smoothed function. However,
this approach leads to difficulties at the boundary of Ω and it will not be considered
further.
There are two often used interpolation operators for non-smooth functions. The
interpolation operator of Clément (1975) is defined for functions from L1 (Ω) and it
can be generalized to more or less all finite elements. The interpolation operator
of Scott and Zhang (1990) is more special. It has the advantage that it preserves
homogeneous Dirichlet boundary conditions. Here, only the interpolation operator
of Clément, for linear finite elements, will be considered.
Let T h be a regular triangulation of the polyhedral domain Ω ⊂ Rd , d ∈ {2, 3},
with simplicies K. Denote by P1 the space of continuous, piecewise linear finite
elements on T h . 2

Remark 6.24 Construction of the interpolation Operator of Clément. For each

vertex Vi of the triangulation, the union of all grid cells which possess Vi as vertex
will be denoted by ωi , see Figure 5.1.
The interpolation operator of Clément is defined with the help of local L2 (ωi )
projections. Let v ∈ L1 (Ω) and let P1 (ωi ) be the space of continuous piecewise
linear finite elements on ωi . Then, the local L2 (ωi ) projection of v ∈ L1 (ωi ) is the
solution pi ∈ P1 (ωi ) of
Z
(v − pi )(x)q(x) dx = 0 ∀ q ∈ P1 (ωi ) (6.12)
ωi

or equivalently of
(v − pi , q)L2 (ωi ) = 0 ∀ q ∈ P1 (ωi ).
Then, the Clément interpolation operator is defined by
N
X
h
PCle v(x) = pi (Vi )φhi (x), (6.13)
i=1

where {φhi }Ni=1 is the standard basis of the global finite element space P1 . Since
h h
PCle v(x) is a linear combination of basis functions of P1 , it defines a map PCle :
1
L (Ω) → P1 . 2

Theorem 6.25 Interpolation estimate. Let k, l ∈ N ∪ {0} and q ∈ R with

k ≤ l ≤ 2, 1 ≤ q ≤ ∞ and let ωK be the union of all subdomains ωi that contain
the mesh cell K, see Figure 6.2. Then it holds for all v ∈ W l,q (ωK ) the estimate
k h
≤ Chl−k Dl v q

D (v − PCle v) q , (6.14)
L (K) L (ωK )

with h = diam(ωK ), where the constant C is independent of v(x) and h.

Proof: The statement of the lemma is obvious in the case k = l = 2 since it is

D2 PCle
h
v(x)|K = 0.

85
 Figure 6.2: A subdomain ωK .

Let k ∈ {0, 1}. Because the L2 (Ω) projection gives an element with best approximation,
one gets with (6.12)
h
PCle p = p in K ∀ p ∈ P1 (ωK ). (6.15)
h
One says that PCle is a consistent operator.
h
The next step consists in proving the stability of PCle . One obtains with the inverse
inequality (6.11)

kpkL∞ (ωi ) ≤ ch−d/2 kpkL2 (ωi ) for all p ∈ P1 (ωi ).

The inverse inequality and definition (6.12) of the local L2 projection with the test function
q = pi gives

kpi k2L∞ (ωi ) ≤ ch−d kpi k2L2 (ωi ) ≤ ch−d kvkL1 (ωi ) kpi kL∞ (ωi ) .

Dividing by kpi kL∞ (ωi ) and applying Hölder’s inequality, one obtains for p−1 = 1 − q −1
(exercise)

|pi (Vi )| ≤ ch−d/q kvkLq (ωi ) (6.16)

for all Vi ∈ K. From the regularity of the triangulation, it follows for the basis functions
that (inverse estimate)
≤ ch−k , k = 0, 1.
k
D φi ∞ (6.17)
L (K)
h
Using the triangle inequality, combining (6.16) and (6.17) yields the stability of PCle
X
k h
D PCle v ≤ |pi (Vi )| Dk φi

Lq (K) Lq (K)
Vi ∈K
X
≤ c h−d/q kvkLq (ωi ) Dk φi k1kLq (K)

L∞ (K)
Vi ∈K
X
≤ c h−d/q kvkLq (ωi ) h−k hd/q
Vi ∈K

= ch−k kvkLq (ωK ) . (6.18)

The remainder of the proof follows the proof of the interpolation error estimate for the
polynomial interpolation, Theorem 6.10, apart from the fact that a reference cell is not
used for the Clément interpolation operator. Using Lemma 6.2 and 6.4, one can find a
polynomial p ∈ P1 (ωK ) with

j
D (v − p) ≤ chl−j Dl v , 0 ≤ j ≤ l ≤ 2. (6.19)

Lq (ωK ) Lq (ωK )

86
With (6.15), the triangle inequality, k·kLq (K) ≤ k·kLq (ωK ) , (6.18), and (6.19), one obtains
   
k h k h h
D v − PCle v = D v − p + PCle p − PCle v q

Lq (K) L (K)

k k h
≤ D (v − p) q + D PCle (v − p)

L (K) Lq (K)

k −k
≤ D (v − p) q + ch kv − pkLq (ωK )

L (ωK )

≤ chl−k Dl v + ch−k hl Dl v

Lq (ωK ) Lq (ωK )

l−k l
= ch D v .
Lq (ωK )

Remark 6.26 Uniform meshes.

• If all mesh cells in ωK are of the same size, then one can replace h by hK in the
interpolation error estimate (6.14). This property is given in many cases.
• If one assumes that the number of mesh cells in ωK is bounded uniformly for all
considered triangulations, the global interpolation estimate
k h
v) Lq (Ω) ≤ Chl−k Dl v Lq (Ω) , 0 ≤ k ≤ l ≤ 2,

D (v − PCle

follows directly from (6.14).

2

87
Chapter 7

Finite Element Methods for

Second Order Elliptic
Equations

7.1 General Convergence Theorems

Remark 7.1 Motivation. In Section 5.1, non-conforming finite element methods
have been introduced, i.e., methods where the finite element space V h is not a sub-
space of V , which is the space in the definition of the continuous variational problem.
The property V h 6⊂ V is given for the Crouzeix–Raviart and the Rannacher–Turek
element. Another case of non-conformity is given if the domain does not possess a
polyhedral boundary and one has to apply some approximation of the boundary.
For non-conforming methods, the finite element approach is not longer a Ritz
method. Hence, the convergence proof from Theorem 4.14 cannot be applied in
this case. The abstract convergence theorem, which will be proved in this section,
allows the numerical analysis of complex finite element methods. 2

Remark 7.2 Notations, Assumptions. Let {h > 0} be a set of mesh widths and
let S h , V h normed spaces of functions which are defined on domains {Ωh ⊂ Rd }.
It will be assumed that the space S h has a finite dimension and that S h and V h
possess a common norm k·kh . In the application of the abstract theory, S h will be a
finite element space and V h is defined such that the restriction and/or extension of
the solution of the continuous problem to Ωh is contained in V h . Strictly speaking,
the modified solution of the continuous problem does not solve the given problem
any longer. Hence, it is consequent that the continuous problem does not appear
explicitly in the abstract theory.
Given the bilinear forms

ah : S h × S h → R,
ãh : (S h + V h ) × (S h + V h ) → R.

Let the bilinear form ah be regular in the sense that there is a constant m > 0,
h h h h
h is independent of h, such that for each v ∈ S there is a w ∈ S with
which
w = 1 such that
h
m v h h ≤ ah (v h , wh ).

(7.1)
This condition is equivalent to the requirement that the stiffness matrix A with the
entries aij = ah (φj , φi ), where {φi } is a basis of S h , is uniformly non-singular, i.e.,

88
its regularity is independent of h. For the second bilinear form, only its boundedness
will be assumed

ãh (u, v) ≤ M kukh kvkh ∀ u, v ∈ S h + V h . (7.2)

Let the linear functionals {f h (·)} : S h → R be given. Then, the following

discrete problems will be considered: Find uh ∈ S h with

ah (uh , v h ) = f h (v h ) ∀ v h ∈ S h . (7.3)

Because the stiffness matrix is assumed to be non-singular, there is a unique solution

of (7.3). 2

Theorem 7.3 Abstract finite element error estimate. Let the conditions
(7.1) and (7.2) be satisfied and let uh be the solution of (7.3). Then the following
error estimate holds for each ũ ∈ V h
(  h h h )
ã (v , w ) − ah (v h , wh )
h h

ũ − u ≤ c inf ũ − v + sup
h v h ∈S h h
wh ∈S h kwh kh
 h 
ã (ũ, wh ) − f h (wh )
+c sup (7.4)
wh ∈S h kwh kh

with c = c(m, M ).

Proof: Because of (7.1) there is for each v h ∈ S h a wh ∈ S h with wh h = 1 and

m uh − v h ≤ ah (uh − v h , wh ).

h

Using the definition of uh from (7.3), one obtains

m uh − v h ≤ f h (wh ) − ah (v h , wh ) + ãh (v h , wh ) + ãh (ũ − v h , wh ) − ãh (ũ, wh ).

h

From (7.2) and wh h = 1 it follows that

ãh (ũ − v h , wh ) ≤ M ũ − v h .

h

Rearranging the terms appropriately and using wh / wh h h = 1 gives
 h h h 
ã (v , w ) − ah (v h , wh )
m uh − v h ≤ h
M ũ − v + sup

h h wh ∈S h kwh kh
 h h h h

ã (ũ, w ) − f (w )
+ sup .
wh ∈S h kwh kh

Applying the triangle inequality

h h h h
ũ − u ≤ ũ − v + u − v

h h h

and inserting the estimate from above gives (7.4).

Remark 7.4 To Theorem 7.3. An important special case of this theorem is the
case that the stiffness matrix is uniformly positive definite, i.e., the condition
2
m v h h ≤ ah (v h , v h ) ∀ v h ∈ S h (7.5)

is satisfied. Dividing (7.5) by v h h reveals that condition (7.1) is implied by (7.5).

89
 If the continuous problem is also defined with the bilinear form ãh (·, ·), then
 h h h 
ã (v , w ) − ah (v h , wh )
sup
wh ∈S h kwh kh
can be considered as consistency error of the bilinear forms and the term
 h 
ã (ũ, wh ) − f h (wh )
sup
wh ∈S h kwh kh
as consistency error of the right-hand sides. 2

Theorem 7.5 First Strang lemma Let S h be a conform finite element space,
i.e., S h ⊂ V , with k·kh = k·kV and let the space V h be independent of h. Consider
a continuous problem of the form

ãh (u, v) = f (v) ∀ v ∈ V,

then the following error estimate holds.

(  h h h )
ã (v , w ) − ah (v h , wh )
h h

u − u ≤ c inf u − v + sup
V v h ∈S h V
wh ∈S h kwh kV
 
f (wh ) − f h (wh )
+c sup .
wh ∈S h kwh kV
Proof: The statement of this theorem follows directly from Theorem 7.3.

7.2 Linear Finite Element Method on Non-Poly-

hedral Domains
Remark 7.6 The continuous problem. The abstract theory will be applied to the
linear finite element method for the solution of second order elliptic partial differ-
ential equations.
Let Ω ⊂ Rd , d ∈ {2, 3}, be a bounded domain with Lipschitz boundary, which
does not need to be polyhedral. Let

Lu = f in Ω, u=0 on ∂Ω, (7.6)

where the operator is given by

Lu = −∇ · (A∇u)

with
A(x) = (aij (x))di,j=1 , aij ∈ W 1,p (Ω), p > d, (7.7)
It will be assumed that there are two positive real numbers m, M such that
2 2
m kξk2 ≤ ξ T A(x)ξ ≤ M kξk2 ∀ ξ ∈ Rd , x ∈ Ω. (7.8)

From the Sobolev inequality it follows that aij ∈ L∞ (Ω). With

Z
a(u, v) = (A(x)∇u(x)) · ∇v(x) dx
Ω

and the Cauchy–Schwarz inequality, one obtains

Z
|a(u, v)| ≤ kAkL∞ (Ω) |∇u(x) · ∇v(x)| dx ≤ c k∇ukL2 (Ω) k∇vkL2 (Ω)
Ω

90
for all u, v ∈ H01 (Ω). In addition, it follows that
2
m k∇ukL2 (Ω) ≤ a(u, u) ∀ u ∈ H01 (Ω).
Hence, the bilinear form is bounded and elliptic. Using the Theorem of Lax–
Milgram, Theorem 4.5, it follows that there es a unique weak solution u ∈ H01 (Ω)
of (7.6) with
a(u, v) = f (v) ∀ v ∈ H01 (Ω).
2
Remark 7.7 The finite element problem. Let T h be a regular triangulation con-
sisting of simplices {K} such that the vertices of the simplices belong to Ω, see
Figure 7.1, and define Ωh = ∪K∈T h K.

Figure 7.1: Approximation of the boundary by the finite element mesh.

The space of continuous and piecewise linear functions that vanish at the bound-
ary of Ωh will be denoted by P1 . It will be assumed that for the data of the problem
aij (x), f (x) there exist extensions ãij (x), f˜(x) to a larger domain Ω̃ ⊃ Ωh with

kãij kW 1,p (Ω̃) ≤ c kaij kW 1,p (Ω) , f˜ ≤ c kf kL2 (Ω) . (7.9)

L2 (Ω̃)

In addition, it will be assumed that the coefficients ãij (x) satisfy the ellipticity
condition (7.8) on Ω̃.
Obviously, f (x) can be continued simply by zero. The extensions of aij (x) have
to be weakly differentiable. It is possible to show that such extensions exist, see the
literature.
The finite element method is defined as follows: Find uh ∈ P1 with
ah (uh , v h ) = f h (v h ) ∀ v h ∈ P1 ,
where
Z   Z
ah (uh , v h ) = Ã(x)∇uh (x) · ∇v h (x) dx, f h (v h ) = f˜(x)v h (x) dx.
Ωh Ωh

In practice, it might be hard to apply the method in this form. From the
existence of the extension operators for aij (x) it is not yet clear how to compute
them. On the other hand, in practice often the coefficients aij (x) are constant or
at least piecewise constant. In these case, the extension is trivial. As remedy in the
general case, one can use quadrature rules whose nodes are situated within Ω, see
the literature. 2
Remark 7.8 Goal of the analysis, further assumptions. The goal consists in prov-
ing the linear convergence of the finite element method in k·kh = k·kH 1 (Ωh ) . In the
analysis, one has to pay attention to the fact that in general neither holds Ωh ⊂ Ω
nor Ω ⊂ Ωh . It will be assumed that there is an extension ũ ∈ H 2 (Ω̃) of u(x) with
kũkH 2 (Ω̃) ≤ c kukH 2 (Ω) . (7.10)

91
In addition, it will be assumed that Ωh is a sufficiently good approximation of Ω in
the following sense
max dist(x, ∂Ω) ≤ ch2 . (7.11)
x∈∂Ωh

One can show that (7.11) is satisfied for d = 2 if the boundary of Ω is piecewise
C 2 and the corners of Ω are vertices of the triangulation. In this case, one can
rotate the coordinate system locally such that the distance between ∂Ω and ∂Ωh
can be represented as the error of a one-dimensional interpolation problem with
continuous, piecewise linear finite elements. Using error estimates for this kind of
problem, e.g., see Goering et al. (2010), one can estimate the error by ch2 . For
three-dimensional domains, with piecewise C ∞ boundary, one needs in addition a
smoothness assumption for the edges. 2

Lemma 7.9 Estimate of a function on the difference of the domains. Let

the condition (7.11) be satisfied. Then, for all v ∈ W 1,1 (Ω) it holds the estimate
Z Z
|v(x)| dx ≤ ch2 (|v(x)| + k∇v(x)k2 ) dx, (7.12)
Ωs Ω

where Ωs is the set Ω \ Ωh or Ωh \ Ω.

Proof: At the beginning, a one-dimensional estimate will be shown. Let f ∈ C 1 ([0, 1]),
then one obtains with the fundamental theorem of calculus
Z x
f (x) = f 0 (ξ) dξ + f (y) ∀ x, y ∈ [0, 1].
y

It follows that Z 1  0 
|f (x)| ≤ f (ξ) dξ + |f (y)| .
0
Integrating this inequality with respect to y in [0, 1] and with respect to x in [0, a] with
a ∈ (0, 1] yields
Z a Z 1 Z 1 Z 1
 0 
|f (x)| + f 0 (x) dx. (7.13)
 

|f (x)| dx ≤ a f (ξ) dξ + a |f (y)| dy = a
0 0 0 0

h
Consider the case Ωs = Ω \ Ω . Since Ω has a Lipschitz boundary, it can be shown
that ∂Ω can be covered with a finite number of open sets U1 , . . . , UN . After a rotation of
the coordinate system, one can represent ∂Ω ∩ Ui as a Lipschitz continuous function gi (y0 )
of (d − 1) arguments y0 = (y1 , . . . , yd−1 ) ∈ Ui0 ⊂ Rd−1 .
In the next step of the proof, sets will be constructed whose union covers the difference
Ω \ Ωh . Let

Si,σ = (y0 , yd ) : gi (y0 ) − σ < yd < gi (y0 ), y0 ∈ Ui0 , i = 1, . . . , N,



see Figure 7.2. Then, using (7.11) it is (Ω \ Ωh ) ∩ Ui ⊂ Si,c1 h2 , where c1 depends on gi (y0 )
but not on h. In addition, there is a σ0 such that Si,σ0 ⊂ Ω for all i.
The transform of (7.13) to the interval [0, σ0 ] gives for sufficiently small h, such that
c1 h2 ≤ 1,
Z c 1 h2 Z σ0
|f (x)| dx ≤ ch2 |f (x)| + f 0 (x) dx.
 

0 0

For v ∈ C 1 (Ω), one applies this estimate to the rotated function v(y0 , x)
Z Z Z c 1 h2
v(y0 , x) dx dy0
 
|v(y)| dy =
Si,c h2 Ui0 0
1
Z Z σ0
ch2 ∂x v(y0 , x) + v(y0 , x) dx dy0
   

≤
Ui0 0
Z
≤ ch2 (|∂yd v(y)| + |v(y)|) dy,
Ω

92
 Figure 7.2: Si,σ .

where in the first step the theorem of Fubini was used. Taking the sum over i proves the
lemma for functions from C 1 (Ω). Since C 1 (Ω) is dense in W 1,1 (Ω), the statement of the
lemma holds also for v ∈ W 1,1 (Ω).
The case Ωs = Ωh \ Ω is proved analogously.

Theorem 7.10 Error estimate, linear convergence. Let the assumptions (7.7),
(7.8), (7.9), (7.10), and (7.11) be satisfied. Then, it holds the error estimate
ũ − uh 1 h ≤ ch kuk 2 ,

H (Ω ) H (Ω)

where c does not depend on u, f , and h.

Proof: For proving the error estimate, the abstract error estimate, Theorem 7.3, is
used with S h = P1 , V h = H 1 (Ωh ), k·kh = k·kH 1 (Ωh ) , and
Z  
ah (u, v) = ãh (u, v) = Ã(x)∇u(x) · ∇v(x) dx.
Ωh

h h
With this choice of a (·, ·) and ã (·, ·), the middle term in the abstract error estimate
(7.4) vanishes. Setting in the abstract error estimate v h = Ih ũ, one obtains with the
interpolation error estimate (6.5) and (7.10)

kũ − Ih ũkH 1 (Ωh ) ≤ ch D2 ũ L2 (Ωh ) ≤ ch kukH 2 (Ω) .

(7.14)

It remains to estimate the last term of (7.4).

93
 The regularity and the boundedness of ah (·, ·) can be proved easily using the ellipticity
and the boundedness of the coefficients ãij (x).
The estimate of the last term of (7.4) starts with integration by parts
Z   Z
ah (ũ, wh ) = Ã(x)∇ũ(x) · ∇wh (x) dx = g(x)wh (x) dx
Ωh Ωh

with g(x) = −∇ · (Ã∇ũ)(x). Because of g(x) = f˜(x) = f (x) in Ω it is

Z  
h h h h
a (ũ, w ) − f (w ) = g(x) − f˜(x) wh (x) dx.
Ωh \Ω

Using the extension of wh (x) by zero on Ω \ Ωh , one obtains with (7.12), and noting that
in general Ωh 6⊂ Ω,
Z Z   
 h 2
2 
 h 2
 
2 h
w (x) dx ≤ ch ∇w (x) + w (x) dx

Ωh \Ω Ω 2
Z  2   
 h 2
2
≤ ch2 h
∇w (x) + w (x) dx = ch2 wh .

Ωh 2 H 1 (Ωh )

Applying the Cauchy–Schwarz inequality and the triangle inequality yields

 
a (ũ, w ) − f (w ) ≤ g − f˜ 2 h
 h h h h  h
w 2 h
L (Ω \Ω) L (Ω \Ω)
 
˜ h
≤ ch kgkL2 (Ω̃) + f w 1 h ,
2 L (Ω̃) H (Ω )

where Ω̃ was introduced in Remark 7.7. Now, a bound for kgkL2 (Ω̃) is needed. Using the
product rule and the triangle inequality, one gets
d
∂ 2 ũ
X  
∇ · (Ã∇ũ) 2 ≤ ãij + ∇ · Ã · ∇ũ .

∂xi ∂xj

L (Ω̃) L2 (Ω̃)
i,j=1 L2 (Ω̃)

1,p ∞
of the Sobolev imbedding W (Ω̃) → L (Ω̃) for p > d, Theorem 3.53, it follows
Because
that à ≤ c. One obtains for the first term

∞ L (Ω̃)
d
∂ 2 ũ
X
≤ c D2 ũ L2 (Ω̃) .

ãij

∂xi ∂xj


i,j=1 L2 (Ω̃)

The estimate of the second term uses Hölders inequality (exercise)

  2
∇ · Ã · ∇ũ 2 ≤ ∇ · Ã k∇ũk2L2p/(p−2) (Ω̃) ≤ c k∇ũk2L2p/(p−2) (Ω̃) .

L (Ω̃) p L (Ω̃)

Using a Sobolev inequality, e.g., see Adams (1975), one obtains the estimate

k∇ũkL2p/(p−2) (Ω̃) ≤ c kũkH 2 (Ω̃) .

Inserting all estimates, one obtains with (7.9) and (7.10)

   
a (ũ, w ) − f (w ) ≤ ch kũkH 2 (Ω̃) + f˜ 2
 h h h h  h
w 1 h
L (Ω̃) H (Ω )
 
h
≤ ch kukH 2 (Ω) + kf kL2 (Ω) w
H 1 (Ωh )

h
≤ ch kukH 2 (Ω) w .
H 1 (Ωh )

In the final step of this estimate, one uses the representation of f (x) from (7.6), for which
one can perform estimates that are analog to the estimates of g(x).
The proof of the linear convergence is finished by using (7.4), (7.14), and the last
estimate.

94
7.3 Finite Element Method with the Nonconform-
ing Crouzeix–Raviart Element
Remark 7.11 Assumptions and discrete problem. The nonconforming Crouzeix–
Raviart finite element P1nc was introduced in Example 5.30. To simplify the pre-
sentation, it will be restricted here on the two-dimensional case. In addition, to
avoid the estimate of the error coming from approximating the domain, it will be
assumed that Ω is a convex domain with polygonal boundary.
Let T h be a regular triangulation of Ω with triangles. Let P1nc (nc – non con-
forming) be denote the finite element space of piecewise linear functions which are
continuous at the midpoints of the edges. This space is nonconforming if it is ap-
plied for the discretization of a second order elliptic equation since the continuous
problem is given in H01 (Ω) and the functions of H01 (Ω) do not possess jumps. The
functions of P1nc have generally jumps, see Figure 7.3, and they are not weakly
differentiable. In addition, the space is also nonconforming with respect to the
boundary condition, which is not satisfied exactly. The functions from P1nc vanish
in the midpoint of the edges at the boundary. However, in the other points at the
boundary, their value is generally not equal to zero.

Figure 7.3: Function from P1nc .

The bilinear form

Z
a(u, v) = (A(x)∇u(x)) · ∇v(x) dx
Ω

will be extended to H01 (Ω) + P1nc by

X Z
ah (u, v) = (A(x)∇u(x)) · ∇v(x) dx ∀ u, v ∈ H01 (Ω) + P1nc .
K
K∈T h

Then the nonconforming finite element method is given by: Find uh ∈ P1nc with

ah (uh , v h ) = (f, v h ) ∀ v h ∈ P1nc .

The goal of this section consists in proving the linear convergence with respect

1/2
to h in the energy norm k·kh = ah (·, ·) . It will be assumed that the solution of
the continuous problem (7.6) is smooth, i.e., that u ∈ H 2 (Ω), that f ∈ L2 (Ω), and
that the coefficients aij (x) are weakly differentiable with bounded derivatives. 2

Remark 7.12 The error equation. The first step of proving an error estimate
consists in deriving an equation for the error. To this end, multiply the continuous

95
problem (7.6) with a test function from v h ∈ P1nc , integrate the product on Ω, and
apply integration by parts on each triangle. This approach gives
X Z
h
(f, v ) = − ∇ · (A(x)∇u(x)) v h (x) dx
K
K∈T h
X Z
= (A(x)∇u(x)) · ∇v h (x) dx
h K
K∈T
X Z
− (A(s)∇u(s)) · nK (s)v h (s) ds
h ∂K
K∈T
X Z
h h
= a (u, v ) − (A(s)∇u(s)) · nK (s)v h (s) ds,
h ∂K
K∈T

where nK is the unit outer normal at the edges of the triangles. Subtracting the
finite element equation, one obtains
X Z
ah (u − uh , v h ) = (A(s)∇u(s)) · nK (s)v h (s) ds ∀ v h ∈ P1nc . (7.15)
∂K
K∈T h

Lemma 7.13 Estimate of the right-hand side of the error equation (7.15).
Assume that u ∈ H 2 (Ω) and aij ∈ W 1,∞ (Ω), then it is
 
 X Z 
A(s)∇u(s) · nK (s)v (s) ds ≤ ch kukH 2 (Ω) v h h .
h
 


K∈T h ∂K 

Proof: Every edge of the triangulation which is in Ω appears exactly twice in the
boundary integrals on ∂K. The corresponding unit normals possess opposite signs. One
can choose for each edge one unit normal and then one can write the integrals in the form
X Z h h
i XZ h i
(A(s)∇u(s)) · nE (s)v (s) ds = (A(s)∇u(s)) · nE (s) v h  (s) ds,
  
E E E E
E E
 
where the sum is taken over all edges {E}. Here, v h  E denotes the jump of v h
 h
h i
 h v |K1 (s) − v h |K2 (s) s ∈ E ⊂ Ω,
v  (s) =
E v h (s) s ∈ E ⊂ ∂Ω,

where nE is directed from K1 to K2 or it is the outer normal on ∂Ω. For writing the
integrals in this form, it was used that ∇u(s), A(s), and nE (s) are almost everywhere
continuous, such that these functions can be written as factor in front of the jumps.
nc
Because of the continuity condition for the functions from  P1 and the homogeneous
Dirichlet boundary condition, it is for all v h ∈ P1nc that v h  E (P ) = 0 for the midpoints P
of all edges. From the linearity of the functions on the edges, it follows that
Z h i
 h
v  (s) ds = 0 ∀ E. (7.16)
E E

Let E be an arbitrary edge in Ω which belongs to the triangles K1 and K2 . The next
goal consists in proving the estimate
Z h i 
 (A(s)∇u(s)) · nE (s) v h  (s) ds
 
E
 
E
 
≤ ch kukH 2 (K1 ) ∇v h + ∇v h . (7.17)

2 2
L (K1 ) L (K2 )

96
  
To this end, one uses a reference configuration K̂1 , K̂2 , Ê , where K̂1 is the unit triangle
and K̂2 is the triangle which one obtains by reflecting the unit triangle at the y-axis. The
common edge Ê is the interval (0, 1) on the y-axis. The unit normal on Ê will be chosen
to be the Cartesian unit vector ex , see Figure 7.4. This reference configuration can be
transformed to (K1 , K2 , E) by a map which is continuous and on both triangles K̂i affine.
For this map one, can prove the same properties for the transform as proved in Chapter 6.

Figure 7.4: Reference configuration.

Using (7.16), the Cauchy–Schwarz inequality, and the trace theorem, one obtains for
an arbitrary constant α ∈ R
Z   h i Z    h i
Â(ŝ)∇û(ŝ) · ex v̂1h  dŝ = Â(ŝ)∇û(ŝ) · ex − α v̂1h  dŝ
   
Ê Ê Ê Ê
  h i
 h
≤ c Â∇û · ex − α (7.18)
.

1
v̂1 
H (K̂1 ) Ê L2 (Ê)

In particular, one can choose α such that

Z   
Â(ŝ)∇û(ŝ) · ex − α dŝ = 0.
Ê

2
The L (Ω) term in the first factor of the right-hand side of (7.18) can be bounded using
the estimate from Lemma 6.4 for k = 0 and l = 1
 
Â∇û · ex − α 1

H (K̂1 )
      

≤ c Â∇û · ex − α + ∇ Â∇û · ex − α

L2 (K̂1 ) L2 (K̂1 )

  
≤ c ∇ Â∇û · ex − α

L2 (K̂1 )
  
= c ∇ Â∇û · ex .

L2 (K̂1 )

To estimate the second factor, in the first step, the trace theorem is applied
h i  
 h
≤ c v̂ h + v̂ h

v̂1 
Ê L2 (Ê) H 1 (K̂1 ) H 1 (K̂2 )
 
≤ c ∇v̂ h + ∇v̂ h .

2 L (K̂1 ) 2
L (K̂2 )

The second estimate follows from the equivalence of all norms in finite dimensional spaces.
To apply this argument, one has to prove that the terms in the last line are in fact norms.
Let the terms in the last line be zero, then it follows that v̂ h = c1 in K̂1 and v̂ h = c2 in K̂2 .
h
Because
 v̂ is continuous in the midpoint of Ê, one finds that c1 = c2 and consequently
that v̂ h  Ê = 0. Hence, also the left hand side of the estimate is zero. It follows that the
nc
right-hand
 h  side of this estimate defines a norm in the quotient space of P1 with respect
to v̂ Ê = 0.
 

97
 Altogether, one obtains for the reference configuration
Z   h i 
  h 
 Â(ŝ)∇u(ŝ) · e x v̂1  dŝ
Ê
 
Ê
    
h
≤ c ∇ Â∇u · ex + ∇v̂ h .

2
∇v̂ 2
L (K̂1 ) L (K̂1 ) L2 (K̂2 )

This estimate has to be transformed to the triple (K1 , K2 , E). In this step, one gets for
the integral on the edge the factor c (ch for ∇ and ch−1 for dŝ). For the product of the
norms on the right-hand side, one obtains the factor ch (ch for the first factor and c for
the second factor). In addition, one uses that A(s) and all first order derivatives of A(s)
are bounded to estimated the first term on the right-hand side (exercise). In summary,
(7.17) is proved.
The statement of the lemma follows by summing over all edges and by applying on the
right-hand side the Cauchy–Schwarz inequality.

Theorem 7.14 Finite element error estimate. Let the assumptions of Lemma
7.13 be satisfied, then it holds the following error estimate

u − uh 2 ≤ ch kuk 2 u − uh + ch2 kuk2 2 .

h H (Ω) h H (Ω)

Proof: Applying Lemma 7.13, it follows from the error equation (7.15) that
 
 h h h  h
a (u − u , v ) ≤ ch kukH 2 (Ω) v ∀ v h ∈ P1nc .
h

Let Ih : H01 (Ω) → P1nc be an interpolation operator with optimal interpolation order
in k·kh . Then, one obtains with the Cauchy–Schwarz inequality, the triangle inequality,
and the interpolation estimate
2
h
u − u = ah (u − uh , u − uh ) = ah (u − uh , u − Ih u) + ah (u − uh , Ih u − uh )

h
 
≤ ah (u − uh , u − Ih u) + ch kukH 2 (Ω) Ih u − uh
 
h
 
≤ u − uh ku − Ih ukh + ch kukH 2 (Ω) kIh u − ukh + u − uh

h
 h 
≤ ch u − uh kukH 2 (Ω) + ch kukH 2 (Ω) h kukH 2 (Ω) + u − uh .

h h

Remark 7.15 To the error estimate. If h is sufficiently small, than the second
term of the error estimate is of higher order and this term can be absorbed into the
constant of the first term. One obtains the asymptotic error estimate
u − uh ≤ ch kuk 2 .

h H (Ω)

7.4 L2 (Ω) Error Estimates

Remark 7.16 Motivation. A method is called quasi-optimal in a given norm, if
the order of the method is the same as the optimal approximation order. Already
for one dimension, one can show that at most linear convergence in H 1 (Ω) can be
achieved for the best approximation in P1 . This statement can be already verified
with the function v(x) = x2 . Hence, all considered methods so far are quasi-optimal
in the energy norm.
However, the best approximation error in L2 (Ω) is of one order higher than the
best approximation error in H 1 (Ω). A natural question is if finite element methods

98
converge also of higher order with respect to the error in L2 (Ω) than with respect
to the error in the energy norm.
In this section it will be shown that one can obtain for finite element methods
a higher order of convergence in L2 (Ω) than in H 1 (Ω). However, there are more
restrictive assumptions to prove this property in comparison with the convergence
prove for the energy norm. 2
Remark 7.17 Model problem. Let Ω ⊂ Rd , d ∈ {2, 3}, be a convex polyhedral
domain with Lipschitz boundary. The model problem has the form
− ∆u = f in Ω, u = 0 on ∂Ω. (7.19)
For proving an error estimate in L2 (Ω), the regularity of the solution of (7.19) plays
an essential role. 2
Definition 7.18 m-regular differential operator. Let L be a second order
differential operator. This operator is called m-regular, m ≥ 2, if for all f ∈
H m−2 (Ω) the solutions of Lu = f in Ω, u = 0 on ∂Ω, are in the space H m (Ω) and
the following estimate holds
kukH m (Ω) ≤ c kf kH m−2 (Ω) + c kukH 1 (Ω) . (7.20)
2
Remark 7.19 On the m-regularity.
• The definition is formulated in a way that it can be applied also if the solution
of the problem is not unique.
• For the Laplacian, the term kukH 1 (Ω) can be estimated by kf kL2 (Ω) such that
with (7.20) one obtains (exercise)
kukH 2 (Ω) ≤ c kf kL2 (Ω) .
• Many regularity results can be found in the literature. Loosely speaking, they
say that regularity is given if the data of the problem (coefficients of the oper-
ator, boundary of the domain) are sufficiently regular. For instance, an elliptic
operator in divergence form (∆ = ∇ · ∇) is 2-regular if the coefficients are from
W 1,p (Ω), p ≥ 1, and if ∂Ω is a C 2 boundary. Another important result is the
2-regularity of the Laplacian on a convex domain. A comprehensive overview
on regularity results can be found in Grisvard (1985).
2
Remark 7.20 Variational form and finite element formulation of the model prob-
lem. The variational form of (7.19) is: Find u ∈ H01 (Ω) with
(∇u, ∇v) = (f, v) ∀ v ∈ H01 (Ω).
The P1 finite element space, with zero boundary conditions, will be used for the
discretization. Then, the finite element problem reads as follows: Find uh ∈ P1
such that
(∇uh , ∇v h ) = (f, v h ) ∀ v h ∈ P1 . (7.21)
2
Theorem 7.21 Finite element error estimates. Let u(x) be the solution of
(7.19), let (7.19) be 2-regular, and let uh (x) be the solution of (7.21). Then, the
following error estimates hold
∇(u − uh ) 2

L (Ω)
≤ ch kf kL2 (Ω) ,
h
≤ ch2 kf kL2 (Ω) .

u − u 2
L (Ω)

99
 Proof: With the error estimate in H 1 (Ω), Corollary 6.16, and the 2-regularity, one
obtains
h
∇(u − u ) ≤ ch kukH 2 (Ω) ≤ ch kf kL2 (Ω) .

L2 (Ω)

For proving the L2 (Ω) error estimate, let w ∈ H01 (Ω) be the unique solution of the
so-called dual problem

(∇v, ∇w) = (u − uh , v) ∀ v ∈ H01 (Ω).

For a symmetric differential operator, the dual problem has the same form like the original
(primal) problem. Hence, the dual problem is also 2-regular and it holds the estimate

kwkH 2 (Ω) ≤ c u − uh .

2 L (Ω)

For performing the error estimate, the Galerkin orthogonality of the error is utilized

(∇(u − uh ), ∇v h ) = (∇u, ∇v h ) − (∇uh , ∇v h ) = (f, v h ) − (f, v h ) = 0

for all v h ∈ P1 . Now, the error u − uh is used as test function v in the dual problem. Let
Ih w be the interpolant of w in P1 . Using the Galerkin orthogonality, the interpolation
estimate, and the regularity of w, one obtains
2
h
u − u 2 = (∇(u − uh ), ∇w) = (∇(u − uh ), ∇(w − Ih w))

L (Ω)

≤ ∇(u − uh ) k∇(w − Ih w)kL2 (Ω)

L2 (Ω)

≤ ch kwkH 2 (Ω) ∇(u − uh )

L2 (Ω)

h h
≤ ch u − u ∇(u − u ) .

L2 (Ω) L2 (Ω)

Finally, division by u − uh L2 (Ω) and the application of the already known error estimate

for ∇(u − uh ) L2 (Ω) are used for completing the proof of the theorem.

100
Chapter 8

Outlook

Remark 8.1 Outlook to forthcoming classes. This class provided an introduction

to numerical methods for solving partial differential equations and the numerical
analysis of these methods. There are many further aspects that will be covered in
forthcoming classes.
Further aspects for elliptic problems.
• Adaptive methods and a posteriori error estimators. It will be shown how it
is possible to estimate the error of the computed solution only using known
quantities and in which ways one can decide where it makes sense to refine the
mesh and where not. (Numerical Mathematics IV)
• Multigrid methods. Multigrid methods are for certain classes of problems opti-
mal solvers. (probably Numerical Mathematics IV)
• Numerical analysis of problems with other boundary conditions or taking into
account quadrature rules.
Time-dependent problems. As mentioned in Remark 1.7, standard approaches
for the numerical solution of time-dependent problems are based on solving station-
ary problems in each discrete time.
• The numerical analysis of discretizations of time-dependent problems has some
new aspects, but also many tools from the analysis of steady-state problems are
used. (Numerical Mathematics IV)
Convection-diffusion equations. Convection-diffusion equations are of impor-
tance in many applications. Generally, the convection (first order differential oper-
ator) dominates the diffusion (second order differential operator).
• In the convection-dominated regime, the Galerkin method as presented in this
class does not work. One needs new ideas for discretizations and these new
discretizations create new challenges for the numerical analysis. (Numerical
Mathematics IV)
Problems with more than one unknown function. The fundamental equation of
fluid dynamics, the Navier–Stokes equations, Section 1.3, belong to this class.
• It will turn out that the discretization of the Navier–Stokes equations requires
special care in the choice of the finite element spaces. The numerical analysis
becomes rather involved. (special class)
2

101
Bibliography

Adams, R. A., 1975: Sobolev spaces. Academic Press [A subsidiary of Harcourt

Brace Jovanovich, Publishers], New York-London, xviii+268 pp., pure and Ap-
plied Mathematics, Vol. 65.
Adams, R. A. and J. J. F. Fournier, 2003: Sobolev spaces, Pure and Applied Math-
ematics (Amsterdam), Vol. 140. 2d ed., Elsevier/Academic Press, Amsterdam,
xiv+305 pp.
Alt, H., 1999: Lineare Funktionalanalysis. Eine anwendungsorientierte Einführung.
3d ed., Springer Berlin.
Braess, D., 2001: Finite elements. 2d ed., Cambridge University Press, Cambridge,
xviii+352 pp., theory, fast solvers, and applications in solid mechanics, Translated
from the 1992 German edition by Larry L. Schumaker.
Brenner, S. C. and L. R. Scott, 2008: The mathematical theory of finite ele-
ment methods, Texts in Applied Mathematics, Vol. 15. 3d ed., Springer, New
York, xviii+397 pp., doi:10.1007/978-0-387-75934-0, URL http://dx.doi.org/
10.1007/978-0-387-75934-0.
Ciarlet, P. G., 1978: The finite element method for elliptic problems. North-Holland
Publishing Co., Amsterdam, xix+530 pp., studies in Mathematics and its Appli-
cations, Vol. 4.
Ciarlet, P. G., 2002: The finite element method for elliptic problems, Classics
in Applied Mathematics, Vol. 40. Society for Industrial and Applied Mathe-
matics (SIAM), Philadelphia, PA, xxviii+530 pp., doi:10.1137/1.9780898719208,
URL http://dx.doi.org/10.1137/1.9780898719208, reprint of the 1978 orig-
inal [North-Holland, Amsterdam; MR0520174 (58 #25001)].
Clément, P., 1975: Approximation by finite element functions using local regulariza-
tion. Rev. Française Automat. Informat. Recherche Opérationnelle Sér. RAIRO
Analyse Numérique, 9 (R-2), 77–84.
Crouzeix, M. and P.-A. Raviart, 1973: Conforming and nonconforming finite el-
ement methods for solving the stationary Stokes equations. I. Rev. Française
Automat. Informat. Recherche Opérationnelle Sér. Rouge, 7 (R-3), 33–75.
Deuflhard, P. and M. Weiser, 2012: Adaptive numerical solution of PDEs. de
Gruyter Textbook, Walter de Gruyter & Co., Berlin, xii+421 pp.
Dziuk, G., 2010: Theorie und Numerik partieller Differentialgleichungen. Walter de
Gruyter GmbH & Co. KG, Berlin, x+319 pp., doi:10.1515/9783110214819, URL
http://dx.doi.org/10.1515/9783110214819.
Ern, A. and J.-L. Guermond, 2004: Theory and practice of finite elements, Applied
Mathematical Sciences, Vol. 159. Springer-Verlag, New York, xiv+524 pp.

102
Evans, L. C., 2010: Partial differential equations, Graduate Studies in Mathematics,
Vol. 19. 2d ed., American Mathematical Society, Providence, RI, xxii+749 pp.
Fefferman, C., 2000: Existence & smoothness of the Navier-Stokes equa-
tions. Http://www.claymath.org/millennium/Navier-Stokes Equations/,
http://www.claymath.org/millennium/Navier-Stokes Equations/.
Goering, H., R. Hans-Görg, and L. Tobiska, 2010: Die Finite-Elemente-Methode fr
Anfänger. 4th ed., Wiley-VCH, Berlin, ix + 219 pp.
Grisvard, P., 1985: Elliptic problems in nonsmooth domains, Monographs and Stud-
ies in Mathematics, Vol. 24. Pitman (Advanced Publishing Program), Boston,
MA, xiv+410 pp.
Grossmann, C. and H.-G. Roos, 2007: Numerical treatment of partial dif-
ferential equations. Universitext, Springer, Berlin, xii+591 pp., doi:10.1007/
978-3-540-71584-9, URL http://dx.doi.org/10.1007/978-3-540-71584-9,
translated and revised from the 3rd (2005) German edition by Martin Stynes.
Haroske, D. D. and H. Triebel, 2008: Distributions, Sobolev spaces, elliptic equa-
tions. EMS Textbooks in Mathematics, European Mathematical Society (EMS),
Zürich, x+294 pp.
John, V. and G. Matthies, 2004: MooNMD—a program package based on mapped
finite element methods. Comput. Vis. Sci., 6 (2-3), 163–169, doi:10.1007/
s00791-003-0120-1, URL http://dx.doi.org/10.1007/s00791-003-0120-1.
Landau, L. and E. Lifschitz, 1966: Lehrbuch der theoretischen Physik, Vol. VI,
Hydrodynamik. Akademie-Verlag Berlin.
LeVeque, R. J., 2007: Finite difference methods for ordinary and partial differential
equations. Society for Industrial and Applied Mathematics (SIAM), Philadelphia,
PA, xvi+341 pp., doi:10.1137/1.9780898717839, URL http://dx.doi.org/10.
1137/1.9780898717839, steady-state and time-dependent problems.
Rannacher, R. and S. Turek, 1992: Simple nonconforming quadrilateral Stokes
element. Numer. Methods Partial Differential Equations, 8 (2), 97–111, doi:10.
1002/num.1690080202, URL http://dx.doi.org/10.1002/num.1690080202.
Samarskii, A. A., 2001: The theory of difference schemes, Monographs and Text-
books in Pure and Applied Mathematics, Vol. 240. Marcel Dekker Inc., New York,
xviii+761 pp., doi:10.1201/9780203908518, URL http://dx.doi.org/10.1201/
9780203908518.
Samarskij, A., 1984: Theorie der Differenzenverfahren, Mathematik und ihre An-
wendungen in Physik und Technik, Vol. 40. Akademische Verlagsgesellschaft
Geest & Portig K.-G., Leipzig.
Schieweck, F., 1997: Parallele Lösung der stationären inkompressiblen Navier-
Stokes Gleichungen. Otto-von-Guericke-Universität Magdeburg, Fakultät für
Mathematik, habilitation.
Scott, L. R. and S. Zhang, 1990: Finite element interpolation of nonsmooth func-
tions satisfying boundary conditions. Math. Comp., 54 (190), 483–493, doi:
10.2307/2008497, URL http://dx.doi.org/10.2307/2008497.
Šolı́n, P., 2006: Partial differential equations and the finite element method. Pure
and Applied Mathematics (New York), Wiley-Interscience [John Wiley & Sons],
Hoboken, NJ, xviii+472 pp.

103
Strang, G. and G. Fix, 2008: An analysis of the finite element method. 2d ed.,
Wellesley-Cambridge Press, Wellesley, MA, x+402 pp.
Wladimirow, W. S., 1972: Gleichungen der mathematischen Physik. VEB Deutscher
Verlag der Wissenschaften, Berlin.

104
Index

L2 projection edge, 60
local, 85 error estimate, 83
δ-distribution, 40 essential boundary condition, 4
estimate
adjoint operator, 54 inverse, 84
affine mapping, 64 Euler scheme
backward, 6
backward difference, 13
forward, 6
barycenter, 64
barycentric coordinates, 64 face, 60
basis Fast Fourier Transform, 29
global, 62 finite element
local, 61 P0 , 65
bilinear form P1 , 65
V -elliptic, 53 P1nc , 69
bounded, 53 P2 , 66
coercive, 53 P3 , 67
boundary Q0 , 70
Lipschitz, 42 Q1 , 70
of class C k,α , 42 Qrot
1 , 72
boundary condition, 4 Q2 , 71
Bramble–Hilbert lemma, 78 Q3 , 71
bubble function Crouzeix–Raviart, 69, 95
cell, 67, 71 Lagrangian, 65
edge, 67 parametric, 63
Rannacher–Turek, 72
Cauchy sequence, 40
simplicial, 65
Cauchy–Schwarz inequality, 37
finite element space, 62
central difference, 13
unisolvent, 61
comparison lemma, 21
five point stencil, 15
consistency error, 90
formulation
consistent difference operator, 14
variational, 55
continuity equation, 8
weak, 54
continuous function
forward difference, 13
with respect to functional, 62
Fourier’s law, 3
Crank–Nicolson scheme, 6
functional
diameter of a mesh cell, 81 global, 62
diffusion equation, 8 local, 62
Dirac distribution, 40
Dirichlet condition, 4 Galerkin method, 58
discrete maximum principle, 17, 34 Galerkin orthogonality, 57, 58
distribution, 39 global basis, 62
regular, 40 Green’s formula
singular, 40 first, 48

105
 second, 49 energy, 95
grid, 13, 17, 61 equivalent, 45
grid function, 13
operator
harmonic function, 5 m-regular, 99
hat function, 62
heat equation, 4 parametric finite element, 63
partial differential equation
incompressible Navier–Stokes equa- elliptic, 11
tions, 10 hyperbolic, 11
inequality linear, 11
Cauchy–Schwarz, 38 parabolic, 11
Friedrichs, 46 quasi-linear, 11
Hölder’s, 38 second order, 11
Poincaré, 46 PCG, 27
Sobolev, 49 Poisson equation, 5
interpolant, 79 problem
interpolation estimate, 82 dual, 100
interpolation operator
Clément, 85 reference triangle, 61
Lagrangian, 80 Ritz approximation, 56
inverse estimate, 84 Robin boundary condition, 4
isotropic grid, 34
second order difference, 13
Jacobi method, 27 simplex, 63
jump, 96 unit, 64
SOR, 27
Laplace equation, 5 space
lemma Pk , 65
Bramble–Hilbert, 78 Qk , 70
Cea, 58 Lebesgue, 37
Strang, 90 Sobolev, 41
lexicographical enumeration, 26 Sobolev–Slobodeckij, 44
local basis, 61 stationary heat equation, 5
stiffness matrix, 58
mapping support, 38
affine, 64
parametric, 73 theorem
maximum principle Lax–Milgram, 53
discrete, 17 Riesz, 52
mesh, 61 trace, 42, 43
mesh cell, 60 triangulation, 61
reference, 64, 69 quasi-uniform, 81
mesh cells, 61 regular, 61
method
quasi-optimal, 98 unisolvence, 61
mixed boundary condition, 4 unit cube, 69
multigrid method, 27
vertex, 60
natural boundary condition, 4
Neumann condition, 4 wave equation, 12
Newton’s law, 4 weak derivative, 39
nine point stencil, 30
Young’s inequality, 37
node, 13
norm

106