Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
Scientiﬁc Computing, advanced
Discretisation of PDEs,
Finite Element Method
P.W. Hemker
ii P.W. Hemker
Contents
1 Introduction to PDEs 1
1.1 Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Ordinary and partial diﬀerential equations . . . . . . . . . 1
1.1.2 Elliptic, parabolic or hyperbolic equations . . . . . . . . . 4
1.1.3 Conservation laws . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Hyperbolic equations . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.1 Characteristics . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.2 Discontinuous solutions . . . . . . . . . . . . . . . . . . . 10
2 Discretisation Principles 17
2.1 Discrete representation of the solution . . . . . . . . . . . . . . . 17
2.1.1 Discretisation of the domain, solution and equation . . . . 19
2.1.2 Finite diﬀerence approximation . . . . . . . . . . . . . . . 19
2.1.3 Finite element approximation . . . . . . . . . . . . . . . . 21
2.1.4 Finite volume approximation . . . . . . . . . . . . . . . . 21
2.1.5 Spectral methods . . . . . . . . . . . . . . . . . . . . . . . 21
2.1.6 Staggered grid . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2 Techniques for discretisation of PDEs . . . . . . . . . . . . . . . 24
2.2.1 Finite diﬀerence methods . . . . . . . . . . . . . . . . . . 24
2.2.2 Variational method . . . . . . . . . . . . . . . . . . . . . . 26
2.2.3 Weighted residual methods . . . . . . . . . . . . . . . . . 27
2.2.4 Collocation methods . . . . . . . . . . . . . . . . . . . . . 27
2.2.5 Galerkin methods . . . . . . . . . . . . . . . . . . . . . . 28
2.2.6 Box methods = Finite Volume methods . . . . . . . . . . 28
2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1 The diﬀusion equation . . . . . . . . . . . . . . . . . . . . 29
2.3.2 The source term and solution of the system . . . . . . . . 30
2.3.3 The convection equation . . . . . . . . . . . . . . . . . . . 31
2.4 Techniques for timediscretisation . . . . . . . . . . . . . . . . . . 31
2.4.1 Timespace discretisation or semidiscretisation . . . . . . 31
2.4.2 FDM for a linear hyperbolic problem . . . . . . . . . . . . 32
2.4.3 The equivalent diﬀerential equation . . . . . . . . . . . . 35
2.5 Criteria for good discretisation methods . . . . . . . . . . . . . . 42
iii
iv P.W. Hemker
3 Finite Element Methods 45
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Examples of boundary value problems . . . . . . . . . . . . . . . 45
3.2.1 Onedimensional second order . . . . . . . . . . . . . . . . 45
3.2.2 Twodimensional second order . . . . . . . . . . . . . . . 46
3.2.3 Fourth order problems . . . . . . . . . . . . . . . . . . . . 48
3.3 Abstract Elliptic Theory . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.1 Introduction to Functional Analysis . . . . . . . . . . . . 49
3.3.2 The Generalised LaxMilgram Theorem . . . . . . . . . . 56
3.3.3 Distributions and Sobolev Spaces . . . . . . . . . . . . . . 58
3.3.4 The Poincar´eFriedrichs Inequality. . . . . . . . . . . . . . 64
3.3.5 Variational formulations for diﬀerential equations . . . . . 65
3.4 The technique of the ﬁnite element method . . . . . . . . . . . . 69
3.4.1 The principles . . . . . . . . . . . . . . . . . . . . . . . . 69
3.4.2 Piecewise Lagrange Interpolation in one dimension . . . . 70
3.4.3 The construction of the discrete equations . . . . . . . . . 73
3.4.4 Other piecewise polynomial approximations in one dimen
sion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.4.5 Approximation in two dimensions . . . . . . . . . . . . . . 83
3.4.6 Isoparametric elements . . . . . . . . . . . . . . . . . . . . 86
3.5 Error estimates for the ﬁnite element method . . . . . . . . . . . 87
3.5.1 The discrete version of the Generalised LaxMilgram . . . 87
3.5.2 More general error estimate . . . . . . . . . . . . . . . . . 90
3.5.3 The formalisation of the ﬁnite element method . . . . . . 92
3.5.4 Interpolation theory and applications . . . . . . . . . . . 94
3.5.5 Pointwise error estimate and superconvergence . . . . . . 102
3.5.6 The inﬂuence of the quadrature rule . . . . . . . . . . . . 104
References 109
Index 111
Chapter 1
Introduction to PDEs
1.1 Classiﬁcation
In these lectures we study numerical methods for partial diﬀerential equations
(PDEs). In the later part, in particular, we will concentrate on so called ellip
tic equations. These equations are used in many areas of application, as e.g.
mechanics, electromagnetics, etc..
To get insight into the quantitative behaviour of the solutions of the PDEs,
and to determine the solutions with some accuracy, in most cases analytic meth
ods fail and we have to rely on numerical methods. The numerical techniques
to compute (approximations to) the solution are often split in two parts. First,
the PDE (the solution of which is a continuous function or set of functions) is
transformed into a set of linear or nonlinear algebraic equations, usually with a
large number of unknowns. This transformation is called the discretisation pro
cess. The second step is the solution of the large algebraic system of equations,
in order to determine the unknowns that describe (approximately) the solution
of the PDE.
In the present chapter diﬀerent kinds of diﬀerential equations are classiﬁed,
and some elementary properties of the PDEs are explained. In Chapter 2 we
will give a survey of the methods that are used for discretisation, for diﬀerent
kinds of PDEs. In Chapter 3 we study the Finite Element Method for the
discretisation of elliptic equations.
1.1.1 Ordinary and partial diﬀerential equations
An ordinary diﬀerential equation (ODE) implicitly describes a function depend
ing on a single variable and the ODE expresses a relation between the solution
and one or more of its derivatives. The order of the diﬀerential equation is the
order of the highest derivative in the equation. Beside the ODE, usually one
or more additional (initial) conditions are needed to determine the unknown
function uniquely.
1
2 P.W. Hemker
Example 1.1.1
An example of a ﬁrst order ODE is
u(x) = u
(x), u(0) = 1.
An example of a second order ODE is
u(x) = −c
2
u
(x), u(0) = 0, u
(0) = 1.
In many interesting cases a function depends on more independent variables
and a relation is given for the function and its partial derivatives. This relation
describes a partial diﬀerential equation (PDE). In many practical problems the
independent variables represent the time and space coordinates (t, x, y, z). To
determine the unknown function, again additional relations are required: initial
and/or boundary conditions.
Example 1.1.2
Let u(t, x, y, z) denote the mass fraction of a chemical species, m, in a medium
with density ρ(t, x, y, z). In the presence of a velocity ﬁeld v(t, x, y, z) the con
servation of the species m is expressed by
∂
∂t
(ρu) + div (ρuv) = s, (1.1)
where s(t, x, y, z) is a possible source
1
.
If diﬀusion also plays a role, then the diﬀusive ﬂux J
d
is described by Fick’s
law of diﬀusion
2
J
d
= −d grad u,
where d is the diﬀusion coeﬃcient, and the complete equation describing the
behaviour of u(t, x, y, z) reads
∂
∂t
(ρu) + div (ρvu) + div J
d
= s. (1.2)
In this equation we distinguish between the diﬀusive ﬂux, J
d
, and the convective
ﬂux, J
c
= ρuv.
1
The divergence of the vector ﬁeld v is deﬁned by div v = ∇v =
∂v
x
∂x
+
∂v
y
∂y
+
∂v
z
∂z
. An
important relation for the divergence is Gauss’ theorem
Ω
div udΩ =
Γ
u · ndΓ,
where Ω ⊂ R
3
is a bounded volume with surface Γ and n is the outward pointing normal on
Γ.
2
The gradient gradient of a scalar function φ(x, y, z) is the vector ﬁeld deﬁned by
grad φ = ∇φ =
∂φ/∂x
∂φ/∂y
∂φ/∂z
.
Version: September 13, 2004 3
The principle part of a diﬀerential equation is the part (the set of terms)
that contain the highest derivatives. E.g. the principle part of (1.1) is (ρu)
t
+
div (ρuv) and the principle part of (1.2) is (ρu)
t
− div (d grad u). PDEs can
have quite diverse properties. A few of such typical forms are given in the next
example. Most typical properties are related with the shape of the principle
part of the equation.
Example 1.1.3
• With Ω ⊂ R
2
an open domain, the Poisson equation is
u
xx
+ u
yy
= f(x, y), (1.3)
where f(x, y) is a prescribed function. In the special case of f ≡ 0 equation
(1.3) is called Laplace’s equation. The solution of these equations can be
determined if the value of u(x, y) is given at the boundary of Ω. Notice
that the equation (in its threedimensional form) is obtained from (1.2) in
the case of a timeindependent solution with v = 0 and s = 0.
• The diﬀusion equation
u
t
= d u
xx
, (1.4)
describes the onedimensional unsteady case of simple diﬀusion, without
a velocity ﬁeld or a source term.
• A diﬀerent kind of PDE is the wave equation
u
tt
= a
2
u
xx
, (1.5)
where a is a positive real number. It is simple to see that a solution of
this equation is given by
u(x, y) = φ(at −x) + ψ(at + x), (1.6)
where φ and ψ are arbitrary (twice diﬀerentiable) functions.
All the above mentioned examples are diﬀerential equations that are linear
in the dependent variable u. Although most theory available is for such linear
partial diﬀerential equations, many important equations are nonlinear. E.g. the
equation (1.1) or (1.2) would be nonlinear if the velocity ﬁeld v was dependent
on u or the source s was a nonlinear function of u. An important subclass of
the nonlinear diﬀerential equations are the quasilinear equations, i.e. equations
in which the highest derivatives appear linearly.
In general, the dependent variable u in the PDE is a function of time, t, and
the three space coordinates, (x, y, z). In numerical methods we have to choose
how to represent the function u, and often we select values of the independent
variables at which the values of u will be calculated. The fewer the number of
degrees of freedom in the approximation of u(t, x, y, z), the less computational
4 P.W. Hemker
eﬀort is needed to determine an approximation of u(t, x, y, z). Fortunately, not
all problems require all independent variables. By symmetry considerations (i.e.
by a judicious choice of the coordinate system) the number of space dimensions
can often be reduced to two or one.
If a solution is sought that is not timedependent the solution is called steady
and the timevariable can be neglected. The corresponding equations are the
steady equations.
To understand the behaviour of the solution it is important to notice that
in some equations we can distinguish a direction of “ﬂow of information”. For
the solution (1.6) of equation (1.5) we see that the information contained in the
part φ of the solution ﬂows forward (i.e. in the positive xdirection) with speed
a, whereas the information in the part ψ ﬂows backward as time proceeds. Also,
in equation (1.1) the ﬂow of information follows the vector ﬁeld v. However,
no such direction is found in (1.3). For timedependent PDEs such as (1.2) or
(1.4) we can always distinguish a unique direction “forward in time”.
The motivation for the discussion of “direction of ﬂow of information” is
that, if such a direction can be identiﬁed, the behaviour of the solution can be
better understood. The knowledge also may be used to reduce the computer
storage and/or the computer time needed to ﬁnd the numerical solution. E.g. in
a timedependent problem as (1.1), for which the solution has to be computed
over a long timeinterval [t
0
, t
2
], the solution at time t = t
1
, with t
0
< t
1
< t
2
,
contains all information to determine the solution for t > t
1
. Having computed
the solution for t ≤ t
1
, it may be eﬃcient from the point of view of computer
resources to disregard all or most information that was obtained about the
solution for t < t
1
(or t < t
1
− τ, for some small τ > 0) before the solution on
(t
1
, t
2
] is determined.
The notion of direction “forward in time” shows also that in timedependent
problems all or part of the sideconditions will be given as “initial conditions”,
that determine the solution (i.e. the state of the physical system) at a later
stage.
1.1.2 Elliptic, parabolic or hyperbolic equations
In the previous section we noticed that diﬀerent PDEs may show a diﬀerent
character. This character is mainly determined by the highest derivatives that
appear in the equation. To characterise some typical classes of PDEs, we con
sider the general linear second order diﬀerential equation in two space dimen
sions:
Lu := a(x, y)u
xx
+ 2 b(x, y)u
xy
+ c(x, y)u
yy
+2 d(x, y)u
x
+ 2 e(x, y)u
y
+ f(x, y)u = g(x, y).
(1.7)
L is a linear diﬀerential operator of order two. With this diﬀerential operator
we associate the characteristic polynomial
P
xy
= a(x, y)ξ
2
+ 2 b(x, y)ξη + c(x, y)η
2
+
+2 d(x, y)ξ + 2 e(x, y)η + f(x, y).
(1.8)
Version: September 13, 2004 5
The second order diﬀerential equation (1.7) is called elliptic at (x, y) if a(x, y)c(x, y)−
b
2
(x, y) > 0; it is hyperbolic at (x, y) if a(x, y)c(x, y) − b
2
(x, y) < 0, and it is
parabolic at (x, y) if a(x, y)c(x, y) − b
2
(x, y) = 0. Thus, this terminology is re
lated with the geometric interpretation of the characteristic polynomial. The
diﬀerential equation is called elliptic, parabolic or hyperbolic (without reference
to a speciﬁc point) if it satisﬁes the corresponding (in)equalities for all (x, y)∈Ω,
where Ω is the domain on which the diﬀerential equation is deﬁned. We see that
the Poisson equation, the diﬀusion equation and the wave equation are elliptic,
parabolic and hyperbolic, respectively.
For an equation of more variables (x
1
, x
2
, ..., x
n
), the general second order
equation reads
Lu :=
n
¸
i,j=1
a
ij
(x)u
x
i
x
j
+
n
¸
i=1
a
i
(x)u
x
i
+ a(x)u. (1.9)
Now, for the principle part of the equation, the characteristic polynomial in
ξ
1
, ξ
2
, ..., ξ
n
is
P
x
(ξ
1
, ..., ξ
n
) =
n
¸
i,j=1
a
ij
(x) ξ
i
ξ
j
. (1.10)
The operator is elliptic if the matrix (a
ij
(x)) is positive or negative deﬁnite; it
is hyperbolic if (a
ij
(x)) has one negative (positive) eigenvalue and n−1 positive
(negative) eigenvalues; it is parabolic if one eigenvalue vanishes and the other
n − 1 have the same sign. It is clear that in more dimensions not all linear
second order PDEs can be classiﬁed as elliptic, parabolic or hyperbolic.
The characterisation of the diﬀerent PDEs is important for distinguishing
the kind of additional conditions that is needed to specify a (unique) solution for
the equation and for understanding the “ﬂow of information” that may appear
in its solution. We show this for the three linear examples given earlier in this
section.
The hyperbolic equation
For the wave equation (1.5) with initial conditions u(0, x) = f(x), u
t
(0, x) =
g(x), the general solution for t ≥ 0 reads
u(t, x) =
1
2
f(x + at) +
1
2
f(x −at) +
1
2a
x+at
x−at
g(ξ) dξ. (1.11)
This means that the solution at (t, x) depends only on the initial solution on
the interval [x −at, x + at].
The parabolic equation
For the linear diﬀusion equation (1.4), with initial condition u(0, x) = f(x),
with x∈R, the solution reads
u(t, x) =
1
2
√
πdt
+∞
−∞
f(ξ) e
−(ξ−x)
2
/(4dt)
dξ. (1.12)
6 P.W. Hemker
This implies that the solution at x∈R for any t > 0 depends on all of the solution
u(0, x) over the interval −∞ < x < +∞.
The elliptic equation
For Laplace’s equation, written in polar coordinates (r, φ), it is known from the
theory of harmonic functions that the solution u(x, y) = ˜ u(r, φ) in the interior
of a circle with radius R, is given by
˜ u(r, φ) =
1
2π
2π
0
R
2
−r
2
R
2
−2Rr cos(θ −φ) + r
2
˜ u(R, θ) dθ. (1.13)
This shows that the behaviour of the solution inside a circle, r = R, is completely
determined by all the values of the solution at the boundary of that circle.
Generally, the function of an elliptic second order equation is determined by its
values at the boundary of its domain of deﬁnition.
1.1.3 Conservation laws
A slight extension of equation (1.1) is the system of equations
∂
∂t
u(t, x, v) + div J(t, x, v)) = s(t, x, v), (1.14)
where v(t, x) is the solution sought, u(t, x, v) is a state vector and J(t, x, v) is
the ﬂux of the state variables
3
; s(t, x, v) is a possible source term.
Equation(1.14) is called a system of conservation laws because the equations
describe the conservation of the quantities u. This is seen by considering an
arbitrary volume Ω and by integration of (1.14) over Ω. Using Gauss’ theorem,
we see that the increase of the amount of u inside the volume Ω should be
caused either by inﬂow over its boundary or by a source described in s(t, x, v).
In natural coordinates, a physical system can often be described with u =
u(v), J = J(v), s = s(v). Then, the dependence on (t, x) may be caused by a
coordinate transformation.
Example 1.1.4 The Euler equations.
The Euler equations of gas dynamics describe the ﬂow of an inviscid nonheat
conducting compressible ﬂuid (a gas). They represent the conservation of mass,
momentum and energy. With ρ(t, x) density, v(t, x) velocity, e(t, x) speciﬁc
energy (temperature) and p(t, x) the presure of the gas, these conservation laws
read respectively
∂
∂t
(ρ) + div (ρv) = 0, (1.15)
∂
∂t
(ρv
i
) + div (ρv
i
v) = −
∂p
∂x
i
, i = 1, 2, 3,
3
Because we consider a system of equations, now each component of J is a vector.
Version: September 13, 2004 7
∂
∂t
(ρe) + div (ρev) = −div (pv),
and p = p(ρ, e, v). For a perfect gas we have the state equation
p = (γ −1)ρ(e −
1
2
[v[
2
). (1.16)
With suitable boundary conditions the variables ρ, v, e and p can be solved from
(1.15) and (1.16). These variables play the role of the vector v in (1.14). Notice
that they are diﬀerent from the conserved quantities u.
1.2 Hyperbolic equations
The second order equation (1.5) allows decomposition into a system of two ﬁrst
order equations. Similar to the factorisation of the characteristic polynomial
P(τ, ξ) = τ
2
− a
2
ξ
2
= (τ + aξ)(τ − aξ) of (1.5), the ﬁrst order operators are
found to be
∂
∂t
+a
∂
∂x
and
∂
∂t
−a
∂
∂x
. The decomposition ﬁrst can be written
4
as
a system
∂u
∂t
= −a
∂p
∂x
,
∂p
∂t
= −a
∂u
∂x
.
(1.17)
If we make the change of dependent variables v = u + p and w = u − p, this
equation (1.17) uncouples and transforms into
∂v
∂t
+ a
∂v
∂x
= 0,
∂w
∂t
−a
∂w
∂x
= 0.
(1.18)
The solution of the system reads v(t, x) = v(x−at), w(t, x) = w(x+at). These
equations show that the quantities u+p (respectively u−p) are constant along
the line x − at = constant (x + at = constant). These lines are called the
characteristics of the diﬀerential equation (or of the diﬀerential operator).
A typical ﬁrst order “oneway” wave equation in R
d
, d = 1, 2, 3, is
∂u
∂t
−v grad u = 0. (1.19)
It shares with (1.18) the property of having a characteristic direction, v, along
which the solution is constant. Therefore also the ﬁrst order equations are called
hyperbolic PDE. To illustrate further the importance of the characteristics, we
consider the inhomogeneous equation
∂u
∂t
−v grad u = s(u). (1.20)
We assume that the vector ﬁeld v(x) is suﬃciently smooth so that it determines
a complete family of characteristics in Ω. Then we will show in Section 1.2.1
that the solution of this PDE is uniquely determined as soon as one value of the
solution is given for each characteristic. Along the characteristic the solution is
now determined by a simple ODE.
4
The minus sign is by convention only.
8 P.W. Hemker
1.2.1 Characteristics
We consider a general ﬁrst order (quasi)linear DE in more dimensions. We
write it down for dimension d = 3; for the one, two or more dimensional case
the treatment is completely analogous. We consider the equation
Pp + Qq + Rr = S , (1.21)
where p = u
x
, q = u
y
, r = u
z
, and P, Q, R, S are functions of (x, y, z, u). I.e.
(x, y, z) are the independent variables and u is the dependent variable.
Let u(x, y, z) be the solution of (1.21), then, by deﬁnition of p, q, and r,
(1.21) is satisﬁed as well as
du = pdx + qdy + rdz.
Or, if A is a point (x, y, z, u) on a solution, then at A we ﬁnd
(p, q, r, −1) ⊥ (P, Q, R, S),
(p, q, r, −1) ⊥ (dx, dy, dz, du),
(p, q, r, −1) ⊥ ( ˙ x, ˙ y, ˙ z, ˙ u),
with ˙ = ∂/∂s, and A(s) is an (arbitrary) curve in a solution surface. Appar
ently, the vector ﬁeld (P, Q, R, S)(x, y, z, u) is tangential to the solution in each
point of the solution. Streamlines in this vector ﬁeld are characteristic solutions.
The projections of these characteristic solutions on the (x, y, z)space are char
acteristics (characteristic direction lines). For a solution along a characteristic
solution line (parametrised by s) we have
(P, Q, R, S)  ( ˙ x, ˙ y, ˙ z, ˙ u), (1.22)
i.e. the two vectors are linearly dependent. It follows that the characteristic
direction can be computed from
P
dx
=
Q
dy
=
R
dz
=
S
du
(1.23)
Further, it follows that, if a value of (x, y, z, u) is given in one point of a char
acteristic, then the characteristic and the solution on it can be found by the
solution of the simple ODE (1.23).
If the solution is prescribed on a suitable (d−1)dimensional manifold (suit
able means that the manifold intersects all characteristics once) then the solution
is apparently uniquely determined.
Example 1.2.1 Inviscid Burgers’ equation.
inviscid Burgers’ equation A wellknown model problem is the inviscid Burgers’
equation
u
t
+ (
1
2
u
2
)
x
= 0 .
Version: September 13, 2004 9
Figure 1.1: The generation of a shock for Burgers’ equation
(See also example (1.2.2)). We notice that
1
dt
=
u
dx
=
0
du
.
The characteristic satisﬁes dx/dt = u and du/dt = 0, so that the solution is
constant along the characteristics, and the characteristics are straight lines. If
the initial solution is increasing in some area, the characteristics will intersect
and no smooth solution will further be possible. This will be considered further
in the next section.
We may study the above arguments more generally in ddimensional space.
Let a solution be prescribed along a (d − 1)dimensional manifold C. Under
what conditions is a solution determined in a layer in the neighbourhood of C?
We easily extend the question to the case of systems of equations. For simplicity
we take d = 3. Let
a
1
u
x
+ b
1
u
y
+ c
1
u
z
+ d
1
v
x
+ e
1
v
y
+ f
1
v
z
+ = g
1
,
a
2
u
x
+ b
2
u
y
+ c
2
u
z
+ d
2
v
x
+ e
2
v
y
+ f
2
v
z
+ = g
2
,
,
(1.24)
or, in short,
¸
j = 1, 2, 3
i = 1, 2, · · · , m
a
kij
∂u
i
∂x
j
= b
k
, k = 1, , m. (1.25)
We introduce a function φ(x
1
, x
2
, x
3
) such that a surface Φ is determined by
φ(x
1
, x
2
, x
3
) = 0. (The surface Φ is our (d − 1)dimensional manifold, with
d=3.) Moreover, we introduce a new set of independent variables (φ
1
, φ
2
, φ
3
)
such that the surface is determined by φ
1
= constant; e.g. φ
1
= 0. (I.e. we
introduce local coordinates around the surface Φ.) Now we have
∂u
i
∂x
j
=
¸
l=1,2,3
∂u
i
∂φ
l
∂φ
l
∂x
j
10 P.W. Hemker
Equation (1.25) becomes
¸
l, j = 1, 2, 3
i = 1, 2, · · · , m
a
kij
∂u
i
∂φ
l
∂φ
l
∂x
j
= b
k
, k = 1, , m, (1.26)
and we get
¸
i=1,···,m
¸
j=1,2,3
a
kij
∂φ
1
∂x
j
¸
¸
∂u
i
∂φ
1
= b
k
−
¸
l = 2, 3
j = 1, 2, 3
i = 1, 2, · · · , m
a
kij
∂u
i
∂φ
l
∂φ
l
∂x
j
. (1.27)
This is a linear mmsystem for the computation of ∂u
i
/∂φ
1
, i.e. the normal
component of u
i
on the surface Φ. For the computation we need (i) the coef
ﬁcients a
kij
; (ii) the righthandsides b
k
; (iii) ∂u
i
/∂φ
l
, l = 2, 3, on the surface
φ
1
= 0.
If the solution u(x, y, z) of the system of PDEs is known in the surface Φ,
then we can also ﬁnd the normal derivatives (and hence the analytic solution in
a layer around Φ), provided that
det
¸
¸
j=1,2,3
a
kij
∂φ
1
∂x
j
¸
= 0. (1.28)
If, on the contrary, det = 0, then in general no smooth solution exists in the
neighbourhood of Φ.
1.2.2 Discontinuous solutions
For simplicity, in this section we ﬁrst consider the onedimensional case. We
have seen that the solution of
u
t
(t, x) + a u
x
(t, x) = 0, (1.29)
with a∈R, and the initial condition
u(0, x) = u
0
(x) , x∈R, (1.30)
is u(t, x) = u
0
(x − at). Also, with a(x) a continuous function, the behaviour
of u(t, x) is easily understood for t > 0. The problem gets considerably more
interesting if the problem is quasilinear, i.e. if we consider
u
t
(t, x) + a(u) u
x
(t, x) = 0, (1.31)
where a(u) is a continuous function depending on the solution. Together with
the initial condition (1.30), it now turns out that not in all cases a smooth so
lution exists for all t > 0, even if u
0
(x) is a smooth function. The reason is that
characteristic lines for the equation (1.31) may intersect. As we have seen, a
Version: September 13, 2004 11
classical solution of (1.31) should be constant on a characteristic line, and dif
ferent characteristics can carry diﬀerent values. At points where characteristics
intersect a shockline is formed and a discontinuity appears. The characteristic
line ends. But this cannot be described by the diﬀerential equation.
There are essentially two ways out of this diﬃculty. The ﬁrst approach is to
generalise the concept of a solution. The class of possible solutions is extended
from the class of diﬀerentiable functions to the class of integrable functions, and
the equation is considered in a weak sense, i.e. it is written in an integral form.
(We see this form later in (1.36).) However, the weak solutions turn out to be
nonunique, (for a given set of initial data), and it remains to characterise the
“physically relevant” weak solutions.
Before we consider this question, we present the second approach. Instead
of (1.31) we now consider the slightly perturbed equation
u
t
(t, x) + a(u) u
x
(t, x) = u
xx
, 0 < < 1. (1.32)
This corresponds with the transition from (1.1) to (1.2) and it means that we
add a small diﬀusion term to the original equation without diﬀusion. Under
reasonable conditions for the initial function u
0
(x) the parabolic equation (1.32)
always has a unique smooth solution u
(t, x) for > 0, and a (nonsmooth) limit
function u
0
(t, x) = lim
→0
u
(t, x) exists.
Now we return to the ﬁrst approach, where we obtain weak solutions of
(1.31). As the functions u
(t, x) converge to such a weak solution, then that
solution is called the physically relevant weak solution of (1.31).
The integral or weak form
5
of the conservation law in divergence form
6
u
t
+ div J(u) = 0, (1.34)
is found by integration over an area R = [t
0
, t
1
] Ω. The integral form reads
[t
0
,t
1
]
Ω
u
t
+ div J(u) dxdt = 0, (1.35)
5
The general principle of weak solution for a diﬀerential equation N(u) = 0 on R
d
is
actually to take an arbitrary φ∈C
∞
0
(R
+
×R
d
), i.e. an arbitrary smooth function with compact
support, and to consider the integral
∞
0
R
d
φ(t, x) N(u)dxdt = 0.
In our case, with N(u) = u
t
+ div J(u), on the half space t ≥ 0, after partial integration this
comes down to
−
∞
0
R
d
uφ
t
+J(u) · grad φ dxdt = −
R
d
φ(0, x) u
0
(x) dx (1.33)
for all φ∈C
∞
0
(R
+
×R
d
). We see that here the requirement of a diﬀerentiable u has disappeared
from the formulation of the equation. By taking a sequence of functions φ∈C
∞
0
(R
+
× R
d
),
that have the characteristic function on [t
0
, t
1
] × Ω as its limit, it can be shown that (1.33)
and (1.36) are equivalent.
6
Notice that (1.19) is written as (1.34) with J(u) = −vu, with div v = 0, because
div(−vu) = −u∇
j
(v
j
u) = −u∇
j
v
j
−v
j
∇
j
u = −vgradu.
12 P.W. Hemker
which can also be written as
Ω
u(t
1
) dΩ −
Ω
u(t
0
) dΩ = −
[t
0
,t
1
]
∂Ω
n J(u) dΓdt. (1.36)
The integral form holds also if there is a (d −1)dimensional manifold on which
u(t, x) is discontinuous. Thus, (1.35) is a generalisation of (1.34).
For d = 2 we immediately see (cf. Figure ??? ) that at a discontinuity D we
have
dx
dt
D
[u] = [J
⊥
(u)] , (1.37)
where [u] denotes the jump of u over the discontinuity. Denoting the propagation
speed of the discontinuity by s, we see
7
s =
J
⊥
(u
R
) −J
⊥
(u
L
)
u
R
−u
L
. (1.38)
This expression for the shock speed is called the RankineHugoniot relation.
Apparently two types of discontinuities may appear in ﬁrst order hyperbolic
equations: linear and nonlinear ones. The ﬁrst type is present in the initial
condition and is carried along a characteristic. It is convected with a speed v
that does not essentially depend on the solution. The second type may appear
even if the initial solution is smooth, and the speed essentially depends the
solution.
Example 1.2.2 Burgers’ equation.
The appearance of a nonlinear discontinuity is often shown for Burgers’ equa
tion. In one space dimension it reads
u
t
+ uu
x
= u
xx
. (1.39)
If the r.h.s. is neglected, i.e. if we take = 0, in a positive solution the top of a
wave moves faster than the bottom. Hence, after a suﬃciently long period this
would lead to a multiple valued function (if that would be possible). Because
such solutions are not allowed, a discontinuity appears of which the speed s is
determined by
J(u) =
u
2
2
, (1.40)
and it follows that the discontinuity moves with a speed s = (u
left
+ u
right
)/2.
7
In more dimensions we get
J
⊥
(u
R
) −J
⊥
(u
L
) = s(u
R
−u
L
)
where J
⊥
denotes the ﬂux vector component in the direction perpendicular to the disconti
nuity.
Version: September 13, 2004 13
t
x
Figure 1.2: A curved shock solution for Burgers’ equation
If we do allow > 0, then the r.h.s. will always have a signiﬁcant value in
those regions the solution is steep and the dissipation will prevent the disconti
nuity to appear
8
.
Example 1.2.3 A discontinuous solution.
As an example (from [17]), the following wave is also a solution to the inviscid
Burgers’ equation:
u(t, x) =
x/t, −
√
t < x <
√
t,
0, otherwise.
(1.43)
This solution (see Figure 1.2) has two shocks, propagating with speeds ±1/2
√
t.
The right shock has left and right states u
L
= 1/
√
t and u
R
= 0, so the
RankineHugoniot condition is satisﬁed. The other shock behaves similarly.
We can use the relation (1.38) to solve explicitly some initial value problems
that are not classically solvable. However, now it appears that these solutions
are not always uniquely determined. We take again the inviscid Burgers’ equa
tion as an example.
8
For some initial conditions the shape of the dissipated shock is easy to determine for
Burgers’ equation. With
u(0, x) =
u
−∞
= q > 0, x < 0,
u
∞
= −q < 0, x > 0,
(1.41)
the steady solution of (1.39) is
u(∞, x) = −q tanh
qx
2
. (1.42)
Notice that the initial condition is already a weak solution of the reduced Burgers’ equation
((1.39) with = 0). A nontrivial steady solution does not appear for u
−∞
< 0 < u
∞
. In
that case, after a ﬁrst dissipative phase, the nonlinear convection will level out all diﬀerences
in u.
The dissipated shock moves at a speed a = (u
L
+ u
R
)/2, and its shape is given by
u(t, x) =
u
L
+ u
R
2
−
u
L
−u
R
2
tanh
(u
L
+ u
R
)(x −at)
2
.
14 P.W. Hemker
t
x
u=1
u=1
u=0
u=0
u = (1−x)/(1−t)
x = 1 + (t−1)/2
x = t
t = 1
Figure 1.3: A continuous solution developing a discontinuity
t
t
x x
u = 1 u = 0 u = 0
u = 1
u = x/t
u = 0
Figure 1.4: A discontinuous and a continuous solution solution
Example 1.2.4 A unique solution.
We take Burgers’ equation with initial function
u
0
(x) =
1, x < 0,
1 −x, 0 ≤ x ≤ 1,
0, x > 1.
(1.44)
The solution we ﬁnd reads (see Figure (1.3))
u(t, x) =
1, x < t, or
x < 1 + (t −1)/2,
(1 −x)/(1 −t), t ≤ x ≤ 1,
0, x > 1, or
x > 1 + (t −1)/2.
(1.45)
Example 1.2.5 A nonunique solution.
We take Burgers’ equation with initial function
u
0
(x) =
0, x < 0,
1, x > 0.
(1.46)
The solutions we ﬁnd are (see Figure (1.4)) are a family of discontinuous func
tions for a∈(0, 1)
u(t, x) =
0, x < at/2,
a, at/2 < x < (1 + a)t/2,
1, (1 + a)t/2 < x.
(1.47)
Version: September 13, 2004 15
x
t
u = 1 u = −1
u = −a u = a
2x=−(1+a)t 2x=(1+a)t
Figure 1.5: Another discontinuous solution
and a continuous function
u(t, x) =
0, x < 0,
x/t, 0 < x < t,
1, x > t.
(1.48)
Example 1.2.6 Another nonunique solution.
We may extend the previous example to obtain a more complex nonunique
solution. We again take Burgers’ equation, now with initial function
u
0
(x) =
−1, x < 0,
1, x > 0.
(1.49)
The solution we ﬁnd reads (see Figure (1.5)), with some a∈(0, 1),
u
a
(t, x) =
−1, 2x < −(1 + a)t,
−a, −(1 + a)t < 2x < 0,
+a, 0 < 2x < (1 + a)t,
+1, (1 + a)t < 2x.
(1.50)
To pick out the “physically relevant solution” we have to ﬁnd the solution
u(t, x) = lim
→0
u
(t, x), with u
the solution of
u
t
+ f(u)
x
= u
xx
.
First, however, we formulate a condition that guarantees unicity of the discon
tinuous solution. Such conditions are generally called entropy conditions. Later
we shall see a relationship between these entropy conditions and the physical
relevance of a solution. In the case of a single equation u
t
+ f(u)
x
= 0 with
f
> 0, one can show that there exists a unique solution that satisﬁes the
“entropy” condition [Oleinik] (1.51),
u(t, x + h) −u(x, t)
h
≤
E
t
, h > 0, t > 0, (1.51)
where E is independent
9
of x, t, and h. This condition implies that if we ﬁx
t > 0, and we let x go from −∞ to +∞, then we can only jump down (in one
direction) over the discontinuity.
9
E depends on u
0
∞
, min{f
(u)} and max{f
(u)}. We do not go into details. The
interested reader is referred to Chapter 16 of [23].
16 P.W. Hemker
This entropy condition of Oleinik guarantees uniqueness of the weak solution.
Condition (1.51) says that, at discontinuities, we have u
L
> u
R
. From the
RankineHugoniot condition we have
s =
f(u
L
) −f(u
R
)
u
L
−u
R
= f
(ζ) ,
for some ζ∈[u
R
, u
L
]. We had assumed f
> 0 (i.e. f is convex) and hence we
ﬁnally ﬁnd that Oleiniks entropy condition implies
f
(u
L
) > s > f
(u
R
) . (1.52)
Sometimes this inequality is also called the entropy condition. The inequality
(1.52) means that the characteristic lines on both sides of the discontinuity run
into the discontinuity. The entropy condition (1.51) of Oleinik has not been
extended to systems of conservation laws. However (1.52) has been extended
by Lax [16].
Example 1.2.7 Traﬃc ﬂow (shocks and rarefaction waves).
(This example is taken from [17].) Let 0 ≤ ρ(x, t) ≤ ρ
m
denote car density on
a road, and u(x, t) car speed. Because cars don’t disappear we have
ρ
t
+ (ρu)
x
= 0 .
Let u be given as u = u(ρ), e.g. u(ρ) = u
m
(1 − ρ/ρ
m
). Combining both equa
tions, gives
ρ
t
+ f(ρ)
x
= 0 .
with f(ρ) = ρu
m
(1 −ρ/ρ
m
). We easily derive the characteristic speed
f
(ρ) = u
m
(1 −2ρ/ρ
m
) ,
and the shock speed for a jump in the car density from ρ
L
to ρ
R
:
s =
f(ρ
L
) −f(ρ
R
)
ρ
L
−ρ
R
= u
m
(1 −(ρ
L
+ ρ
R
)/ρ
m
) . (1.53)
The entropy condition implies that a shock must satisfy f
(ρ
L
) > f
(ρ
R
), which
implies ρ
L
< ρ
R
. Now, with initial data
ρ(x, 0) =
ρ
L
, x < 0,
ρ
R
, x > 0,
several situations can be distinguished. With 0 < ρ
L
< ρ
R
< ρ
m
a shock wave
travels with speed (1.53). This shock speed can move in the positive or in the
negative direction. With 0 < ρ
R
< ρ
L
< ρ
m
we rather have a “rarefaction
wave’.
Chapter 2
Discretisation Principles
In many cases it is important to obtain quantitative data about physical, tech
nical or other reallife systems. One may obtain these through measurements.
However, if we have suﬃciently accurate mathematical models for a system, we
can also try to derive quantitative data from these models. Models for technical
problems often take the form of a PDE or a system of PDEs, and although much
analytical theory is available for PDEs, most of these equations do not allow an
explicit solution in closed form. Therefore numerical methods are used to get
insight in their quantitative behaviour.
In practice it is clear that there are many restrictions in gathering reliable
data from reallife measurements. In the numerical modelling of these systems,
limitations of accuracy are caused by possible inaccuracies in the mathematical
model, because the model is often a simpliﬁcation of reality, and also because
there is always a limitation of computing resources.
Numerical mathematics takes the mathematical equations as a starting point
and strives for an eﬃcient and accurate approximation of the the quantitative
data.
2.1 Discrete representation of the solution
In many problems from practice, the solution of a PDE is a steady or time
dependent (vector) function in d space variables (d = 1, 2, 3). Thus the solution
is an element of an inﬁnitelydimensional space, and in general it cannot be
described by a ﬁnite number of real numbers. Because computational capacity
is always ﬁnite, the problem has to be discretised: i.e. the solution has to be
approximated by a function that can be represented by a ﬁnite set of numbers.
Example 2.1.1
Let d = 1, and let the solution of the equation be a simple continuous function
u(x), x∈[a, b] = Ω. A few possible ways to represent this function are:
17
18 P.W. Hemker
Piecewise linear Piecewise polynomial
Figure 2.1: Piecewise polynomial approximations
i) a table of equidistant function values
u
i
≈ u(x
i
), i = 0, 1, 2, , N, x
i
= a + i(b −a)/N;
ii) a table of nonequidistant function values
u
i
≈ u(x
i
), i = 0, 1, 2, , N, a ≤ x
0
≤ x
1
≤ ≤ ≤ x
N
≤ b;
iii) a table of function values and values of derivatives of that function (either
equidistant or nonequidistant)
u
i,k
=
d
dx
k
u(x
i
); k = 0, 1; i = 0, 1, 2, , N.
In all these cases the solution at the intermediate values of x can be approxi
mated by interpolation. This can be done e.g. by a choosing a polynomial, a
spline or a piecewise polynomial as the interpolating function.
Another way to approximate the function is by choosing a priory a ﬁnite
dimensional function space, selecting a basis in this space, and determining the
coeﬃcients of the approximation with respect to that basis. Then we have to
select a family of functions ¦φ
i
(x)[i = 1, , N¦, which forms the basis for the
approximating function space, and we compute coeﬃcients ¦a
i
[i = 1, , N¦ so
that the function
φ(x) =
N
¸
i=1
a
i
φ
i
(x) (2.1)
approximates the function that is to be represented.
Possible choices for ¦φ
i
¦ are, e.g. (i) the polynomials φ
i
(x) = x
i
; (ii) gonio
metric functions ¦φ
i
(x)¦ = ¦1, sin x, cos x, sin(2x), cos(2x), ¦; or (iii) piece
wise polynomials on a partition of the interval [a, b], e.g. piecewise linear or
piecewise parabolic functions.
To obtain a more accurate approximation of the function u(x), we need 
in general a denser set of function values or more coeﬃcients to compute. If
Version: September 13, 2004 19
a partition of nodal points is made, or a partition in subintervals, the largest
distance between two neighbouring points, or the size of the largest interval in
the partition is usually denoted by a parameter h. This parameter is called the
meshsize. For a smaller h the approximation gets (in general) (i) more accurate,
and (ii) more laborious (timeconsuming) (”more expensive”) to compute.
We notice that the classiﬁcation of approximation methods does not yield
disjoint classes of methods. In the representation (2.1) the coeﬃcients may
coincide with function values or values of derivatives. Example (iii) can also
be seen as an interpolation method for a set of function values, given at the
subinterval endpoints. If there is a set of nodal point ¦x
i
¦ and a set of basis
function ¦φ
i
¦, such that they satisfy the relation
φ
i
(x
j
) = δ
i,j
,
where δ
i,j
denotes the Kronecker delta, the same method can be considered as
a pointwise approximation and as a functional approximation as well.
2.1.1 Discretisation of the domain, solution and equation
In the transformation of the PDE into a ﬁnite set of equations, (i.e. in the
discretisation process) it is convenient to distinguish a number of stages:
i) the discretisation of Ω, the domain of deﬁnition of the PDE;
ii) the discretisation of u, the solution (i.e. the selection of the representation
for the solution); and
iii) the discretisation of the equation.
Before we describe the techniques how to discretise the equations, we ﬁrst
describe diﬀerent methods for the discretisation of the domain and of the solu
tion.
To ﬁnd the discrete representation of the solution, one usually selects in the
domain where the PDE is deﬁned, a (regular or irregular) mesh of points, or a
partitioning of the domain in triangles or quadrilaterals (See Figure 2.2 or 2.3).
2.1.2 Finite diﬀerence approximation
For the ﬁnite diﬀerence method (FDM), the domain Ω is represented by a ﬁnite
subset of points ¦x
i
¦ = Ω
h
⊂ Ω. These points are the so called “nodal points”
of the grid. This grid is almost always arranged in (uniform or nonuniform)
rectangular manner, see Figure 2.2.
In the ﬁnite diﬀerence method (FDM) the function u is represented by a
set of function values u
i
that approximate u(x
i
). The discrete equations are
obtained by replacing the diﬀerentials in the PDE by ﬁnite diﬀerences in the
discrete equations.
20 P.W. Hemker
Uniform regular grid Non−uniform regular grid
Figure 2.2: Sets of nodal points in rectangular regions
Figure 2.3: Partition of twodimensional domains
Version: September 13, 2004 21
Piecewise constant Piecewise linear Piecewise linear
Piecewise quadratic Hermite
piecewise cubic
Figure 2.4: Examples of nodal points in a triangular domain
2.1.3 Finite element approximation
In the ﬁnite element method (FEM) the domain Ω is partioned in a a ﬁnite set
of elements ¦Ω
i
¦, so that Ω
i
∩ Ω
j
= ∅, for i = j, and ∪Ω
i
= Ω. Usually one
takes for Ω
i
triangles or quadrangles. Then the function is approximated by
u
h
=
¸
a
i
φ
i
(x) ,
where φ
i
are functions that are polynomials on each Ω
i
(i.e. piecewise polyno
mials). Usually the functions φ
i
are continuous polynomials of a low degree.
Further they are constructed so that their support extends only over a small
number of elements.
2.1.4 Finite volume approximation
In this case the domain Ω is also partitioned in a ﬁnite set of volumes (mostly
triangles or quadrangles) so that Ω
i
∩ Ω
j
= ∅, for i = j, and ∪Ω
i
= Ω. Now in
A cell centered (or block centered) ﬁnite volume method (FVM), the function u
is approximated by
u
h
=
¸
a
j
φ
j
(x) , (2.2)
where the (possibly discontinuous) functions φ
i
(x) are deﬁned only on a single
volume. or on the boundaries of the volumes.
If there is a single coeﬃcient a
i
associated with each volume (so that a
i
may
represent u(x) for x∈Ω
i
), then the grid is called a block centered or cell centered.
If a (single) coeﬃcient is associated with each vertex, then the method is called
cell vertex (see ﬁgure 2.5). If the discretisation is constructed on cells (dashed
lined in Figure 2.5) that are located around vertices of a grid (solid lines), then
the mesh is called vertex centered.
2.1.5 Spectral methods
The function u is again approximated by
u
h
=
¸
a
i
φ
i
(x)
22 P.W. Hemker
Cell centered Cell vertex Vertex centered
Figure 2.5: Examples of nodal points in a quadrangular domain
where the functions φ
i
(x) are deﬁned on all of Ω (they have a support that
extends over more than only a small part of Ω) then the methods are called
global or spectral methods. (We see that the diﬀerent types of methods are not
always clearly to distinguish.)
2.1.6 Staggered grid
For systems of equations all methods mentioned above are easy to generalise. All
functions can simply be replaced by vector functions. However, it is also possible
to approximate the diﬀerent (dependent) variables by diﬀerent representations.
One way is to approximate the diﬀerent components of the solution on diﬀerent
point sets. This is the case e.g. with “staggered grids”. We explain this by two
examples.
Example 2.1.2
We consider the Cauchy Riemann equations
(i) u
x
+ v
y
= f
1
(ii) u
y
−v
x
= f
2
on Ω ⊂ R
2
, (2.3)
Ω is a rectangular domain, and Ω is partitioned in a number of equal sub
rectangles. Approximations are computed for u and v. The values of u are
calculated for the midpoints of the vertical edges, the values of v for the mid
points of the horizontal edges. The discrete equations corresponding with equa
tion (2.3(i)) are related with the midpoints of the cells. The discrete equations
corresponding with equation (2.3(ii)) are constructed at the cell vertices.
To understand the character of these equations, we write the CauchyRiemann
equations as
L
u
v
=
∂
∂x
∂
∂y
∂
∂y
−
∂
∂x
u
v
=
f
1
f
2
. (2.4)
We see that
det(L) = −
∂
∂x
2
−
∂
∂y
2
= −∆.
Version: September 13, 2004 23
u u u u u
u u u u u
u u u u u
u
u u
u u
u
v v v v v
v v v v v
v v v v v v
v v v v v v
v
v
u u u u u
u u u u u
u u u u u
u
u u
u u
u
v v v v v
v v v v v
v v v v v v
v v v v v v
v
v
1 1 1 1 1 1
1 1 1 1 1 1
1 1 1 1 1
1
2 2 2 2
2 2 2 2
2
2
1 1 1 1 1 1
1 1 1 1 1 1
1 1 1 1 1
1
2 2 2 2
2 2 2 2
2
2
variables equations
Figure 2.6: A staggered grid for the CauchyRiemann equations
u u u u u
u u u u u
u u u u u
u
u u
u u
u
v v v v v
v v v v v
v v v v v v
v v v v v v
v
v
p p p p p p
p p p p p p
p p p p p p
u u u u u
u u u u u
u u u u u
u
u u
u u
u
v v v v v
v v v v v
v v v v v v
v v v v v v
v
v
p p p p p p
p p p p p p
p p p p p p
Figure 2.7: A staggered grid for Stokes equations
This shows that the system is elliptic and second order. It is also directly seen
by elimination that
1
∆u = div f , or ∆v = −curl f .
The system being second order elliptic, it needs one boundary condition
along all the boundary. We take
u.n
x
+ v.n
y
= v.n = g(x, y) .
In our discrete system, this condition ﬁxes the values of u and v on the boundary.
Now the number of unknowns is 27 (viz. 3 5 for u, and 2 6 for v), and the
number of equations is 28 (3 6 for equation (i), and 2 5 for equation (ii)).
Hence, the system seems to be overdetermined. However, it is easily seen that
the boundary function g(x, y) should satisfy a condition. From u
x
+v
y
= div v =
f
1
, it follows that
Ω
u
x
+ v
y
dΩ =
Ω
div v dΩ =
∂Ω
v.ndΓ =
∂Ω
g(x, y) dΓ,
1
In three dimensions
curl v =
¸
∂v
3
∂x
2
−
∂v
2
∂x
3
∂v
1
∂x
3
−
∂v
3
∂x
1
∂v
2
∂x
1
−
∂v
1
∂x
2
¸
,
in two dimensions curl v =
∂v
2
∂x
1
−
∂v
1
∂x
2
.
24 P.W. Hemker
i.e. the compatibility condition
Ω
f
1
(x, y) Ω =
∂Ω
g(x, y) dΓ.
Similarly it follows from adding all discrete equations (u
R
−u
L
)/h + (u
T
−
u
B
)/h = f
1
for each cell, that in total
h
¸
k
g
k
= h
¸
i
u
R,i
−h
¸
i
u
L,i
+ h
¸
i
v
T,i
−h
¸
i
v
B,i
= h
2
¸
i
f
1,i
, (2.5)
i.e.
¸
k∈boundary
g
k
= h
¸
i∈interior
f
1,i
(2.6)
Here f
1,i
denotes the value of f
1
(x, y) in the ith cell in the interior of Ω, and g
k
denotes the value of g(x, y) in the kth boundary edge element. If the boundary
values satisfy condition (2.6), one (linear) dependence relation exists between
the 28 equations and we get a solvable system of equations.
Example 2.1.3 Stationary Stokes equations
Analogous to the staggered grid for the CauchyRiemann equations, we can
construct one for the stationary Stokes equations,
η∆u −p
x
= 0,
η∆v −p
y
= 0,
u
x
+ v
y
= 0.
This can be written as a system with linear diﬀerential operator L,
L
¸
u
v
p
¸
=
¸
η∆ 0 ∇
1
0 η∆ ∇
2
∇
1
∇
2
0
¸
¸
u
v
p
¸
=
¸
0
0
0
¸
.
The character of the system of equations is determined by the determinant of
the principle part of the diﬀerential operator
detL = η∆ ∇
2
1
+ η∆ ∇
2
2
+ 0 = η∆
2
.
This is a 4th order elliptic operator, for which we need two conditions over all
of the boundary. The operator ∆ is represented by the ﬁnite diﬀerence stencil
1
1 −4 1
1
¸
¸
, the operator ∇
1
by
−1 1
and ∇
2
by
¸
1
−1
.
2.2 Techniques for discretisation of PDEs
2.2.1 Finite diﬀerence methods
For this method, we have already seen examples. The discrete equation is
constructed by replacing diﬀerential operators by diﬀerence operators. The
Version: September 13, 2004 25
Figure 2.8: Diﬀerence approximation and polynomial interpretation
construction of the diﬀerence operators is as follows: in the neighbourhood of a
point x∈Ω the approximation of u(x) is replaced by a truncated Taylor series:
u(x) = a
0
+ a
10
x + a
01
y + a
11
x
2
+ .
With the values u
ij
in the nodal points x
ij
around x, the coeﬃcients a are
expressed in these function values u
ij
. These coeﬃcients a
pq
are associated
with the derivatives of the approximation in x. Now, in the PDE at this point
the diﬀerentials are replaced by diﬀerence approximations. In this way, for each
unknown u
ij
an equation is set up at the point x
ij
.
For a regular grid, in the interior points of the domain the same diﬀerence
approximation can be used. Diﬀerent diﬀerence approximations appear near
the boundary of the domain, or in parts of the domain where the nodal points
are not regularly placed.
In the neighbourhood of x
ij
, the approximation of u(x) is, thus, considered
as a piecewise polynomial (of a low degree). That does not mean that the
complete function u is considered as a polynomial. The interpretation of the
polynomial representation depends on the nodal point where the interpretation
is made.
The (linear or nonlinear) algebraic system of equations that results from the
discretisation process has an important property. By the diﬀerence approxima
tions the value of the unknown function at a particular point is only coupled
with the values at neighbouring points. In this way a “sparse system” of alge
braic equations appears. The corresponding matrix (of the linearisation) only
contains a few nonzero elements in each row; most entries in the matrix vanish.
Whereas the systems that appear as a result of discretisations can become very
large, they mostly have this nice property, by which the solution can often be
found in a relatively eﬃcient
2
manner. As this property is shared by most dis
cretisations of PDEs, special methods are sought (and found) for the eﬃcient
solution of such sparse systems of equations.
2
Compared with a full system of equations of the same dimensions.
26 P.W. Hemker
2.2.2 Variational method
Variational method for diﬀerence methods
An important class of PDEs, the symmetric elliptic equations, can be written as
a minimisation problem. This gives another possibility to construct the discrete
equations. We show this by means of the Poisson equation
∆u = f, on Ω,
u = 0, on Γ.
This equation can be written as the minimisation problem: ﬁnd a continuous
function u such that u = 0 on Γ, and
J(u) =
Ω
∂u
∂x
2
+
∂u
∂y
2
+ 2fu dΩ
exists and is minimal. (Exists: u
2
x
+u
2
y
and fu should be integrable functions.)
Now we can set up a discrete analogue for this minimisation problem [5].
On a regular square grid (i.e. Ω a rectangle, and square meshes with a meshsize
h) we can write: ﬁnd u
h
= ¦u
ij
¦ such that u
ij
= 0 on Γ, and
J
h
(u
h
) =
¸
ij
u
i+1,j
−u
i,j
h
2
+
u
i,j+1
−u
i,j
h
2
+ 2f
ij
u
i,j
is minimal. The minimisation of this quadratic functional in the unknowns ¦u
ij
¦
leads to linear system in ¦u
ij
¦. This system also yields a diﬀerence equation for
each internal point of the grid.
The variational method for coeﬃcients
For problems that can be written as minimisation problems, we can use the
variational method also if we seek an approximation in the form
u
h
(x, y) =
¸
i
a
i
φ
i
(x, y) .
We take the same example (2.2.2), so that with φ
i
(x, y) = 0 on Γ. Now we
compute the coeﬃcients a
i
such that
Ω
¸
i
a
i
∂
∂x
φ
i
(x, y)
¸
2
+
¸
i
a
i
∂
∂y
φ
i
(x, y)
¸
2
+ 2f(x, y)
¸
i
a
i
φ
i
(x, y) dΩ
attains its minimum. This minimisation problem also leads to a linear system,
now in the unknown coeﬃcients (a
i
). The same technique is applicable on
irregular grids.
Version: September 13, 2004 27
2.2.3 Weighted residual methods
A quite general discretisation technique is based on the following consideration.
Let the PDE be given by
^
Ω
(u) = s
Ω
, on Ω,
^
Γ
(u) = s
Γ
, on Γ,
(2.7)
then, for any function w we have
Ω
w
Ω
(^
Ω
(u) −s
Ω
) dΩ +
Γ
w
Γ
(^
Γ
(u) −s
Γ
) dΓ = 0. (2.8)
Or, in an even more general notation, let the equation be ^(u) = s and let
s, ^(u)∈V , with V a Banach space, then for an arbitrary linear functional
l
w
∈V
the equation
3
l
w
(^(u) −s) = 0
is satisﬁed. Now, let S be a set of functions in which the solution u is sought:
^ : S → V , then the PDE can be formulated as: ﬁnd u∈S such that
l
w
(^(u) −s) = 0, for all l
w
∈V
.
In a weighted residual method an ndimensional subspace S
h
⊂ S is selected
and an ndimensional subspace V
D
h
⊂ V
. The discretisation now reads: ﬁnd
u
h
∈S
h
such that
l
w
(^(u
h
) −s) = 0 for all l
w
∈V
D
h
.
In this way, the discretisation yields an n nsystem of equations.
Of course, there are many diﬀerent applications of the weighted residual
principle. We will treat the most important.
2.2.4 Collocation methods
Collocation methods are weighted residual methods with the functionals l
w
in
the space V
D
h
that is spanned by
¦l
x
i
[ x
i
∈Ω; i = 1, , N¦ .
Here l
x
i
is deﬁned by
l
x
i
(f) = f(x
i
), for f∈V.
This means that the discrete equations are
^(u
h
)(x
i
) = s(x
i
), for i = 1, , N.
This means that the approximate solution u
h
should satisfy the original PDE
precisely in n given points, where n = dim(S
h
).
3
The dual space of a Banach space V (denoted by V
) is the family of all bounded linear
functionals deﬁned on V , furnished with the obvious addition and scalar multiplication.
28 P.W. Hemker
2.2.5 Galerkin methods
When the functionals l are deﬁned by integrals with integrable weight functions,
we obtain “Galerkin”methods. Typically l
w
∈V
D
h
is deﬁned by
l
w
(f) =
Ω
w(x) f(x) dΩ.
The space V
D
h
is now deﬁned by n linearly independent functions w
i
(x).
If we choose ¦w
i
(x)¦
i=1,2,···,N
, such that V
h
:= span (w
i
) = S
h
then the
methods are called BubnovGalerkin methods. An obvious choice is the w
i
= φ
i
.
Then the discrete equations read
Ω
φ
j
^(
¸
a
i
φ
i
) dΩ =
Ω
φ
j
s dΩ.
If we take V
h
:= span (w
i
) = S
h
, the method is called a PetrovGalerkin
method.
Almost all usual Finite Elements methods are of the (Bubnov) Galerkin
type. When applied to symmetric positive deﬁnite elliptic PDEs, these methods
can also be regarded as variational methods.
2.2.6 Box methods = Finite Volume methods
Box methods are a special kind of Galerkin methods. Here the domain Ω is
partitioned in a number of volumes ¦Ω
i
¦
i=1,···,N
I.e. Ω
i
∩ Ω
j
= ∅, if i = j, and
∪Ω
i
= Ω. The weight functions are now the characteristic functions on these
Ω
i
,
w
i
(x) = 1, if x∈Ω
i
,
w
i
(x) = 0, if x∈Ω
i
,
and the discrete equations read
Ω
i
^
¸
a
i
φ
i
dΩ
i
=
Ω
i
s dΩ
i
.
The volumes Ω
i
are called “boxes”, “cells” or “control volumes”.
Box methods are particularly popular for the discretisation of equations in
divergence form
div J(u) = s(u), on Ω.
Then the discrete equation gets the form
∂Ω
i
J(u
h
) ndΓ
i
=
Ω
i
s(u
h
) dΩ
i
, for i = 1, , N. (2.9)
If the computation of J(u
h
) at an interface between two cells is made consis
tently for both neighbouring cells, then the discretisation satisﬁes a discrete
conservation law. This means that for an arbitrary union of cells G = ∪Ω
i
hold:
G
div J(u
h
) dΩ =
G
s(u
h
) dΩ.
Version: September 13, 2004 29
x x x
i−1 i i+1
x x
w e
Figure 2.9: Nodal points and boxes on the real line
2.3 Examples
2.3.1 The diﬀusion equation
To show the principle of discretisation, in this section we discretise a steady
onedimensional diﬀusion equation, with a sourceterm s and a variable diﬀusion
coeﬃcient k,
−
d
dx
k(x)
d
dx
u
= s(x, u), on Ω ⊂ R, (2.10)
together with a boundary condition (e.g. u = 0, on ∂Ω). This may be seen
as the description of heatconduction through a bar, where u(x) denotes the
temperature, k(x) the heat diﬀusion coeﬃcient, and s(x, u) the heat generation
per unit length, by some source.
We divide Ω in (unequal) boxes Ω
i
and we choose for each Ω
i
a nodal point
x
i
(see Figure 2.9).
We may choose the boxwalls halfway the nodal points, or we may put the
nodal points in the midpoint of the boxes. (Such choice is not relevant at the
moment.)
The discrete approximation is represented by the function values at the nodal
points u
i
≈ u(x
i
). Integration of the equation over Ω
i
yields
x
e
x
w
s(x, u)dx = −
x
e
x
w
(ku
x
)
x
dx = − k(x)u
x
(x)[
x
e
x
w
. (2.11)
To set up the discrete equations we need an approximation of u(x) over
Ω
i
, and approximations u
x
(x
e
) and u
x
(x
w
). Starting from u
i
≈ u(x
i
) we can
approximate u, e.g. by
i) taking a constant value u
i
on Ω
i
;
ii) taking a piecewise linear approximation between the nodal points.
With (i) we cannot approximate the r.h.s. of (2.11), but with (ii) we can. Then
we obtain
k(x
e
)
u
i+1
−u
i
x
i+1
−x
i
−k(x
w
)
u
i
−u
i−1
x
i
−x
i−1
= −
x
e
x
w
s(x, u
h
) dx = −s
i
(u
h
) (x
e
−x
w
) ,
(2.12)
where s
i
is the mean value of s(x, u) over the volume Ω
i
.
30 P.W. Hemker
2.3.2 The source term and solution of the system
If s is independent of u, or if s depends linearly on u, then (2.12) is a linear
system of equations that can be solved immediately (as soon as the values of
the endpoints u
0
and u
n
are given) if we take s
i
(u
h
) = s(u
i
).
On the other hand, if s = s(u) is a nonlinear function of u, then (i) s
i
(u
h
)
should be approximated, and (ii) the equation has to be solved iteratively. The
source term s
i
can either be approximated by a constant s
i
:= s
i
(u
(n)
h
), where
the previous approximation u
(n)
h
is used directly to obtain a (hopefully better)
approximation u
(n+1)
h
by (2.12), or by linearisation, i.e. taking a linear approx
imation of s as s
i
= s
C
i
(u
(n)
h
) + s
L
i
(u
(n)
h
)(u
(n+1)
i
−u
(n)
i
. In this case we have to
adapt (2.12) slightly.
The diﬀerence between (i) and (ii) is that in (i) s
i
depends completely on the
previous approximation, whereas in (ii) linear dependence of s on u is taken into
account; s
C
i
and s
L
i
are constants, that have to be determined for each nodal
point. This makes that the linear system to solve is changed:
k(x
e
)
x
i+1
−x
i
u
i+1
−
k(x
e
)
x
i+1
−x
i
+
k(x
w
)
x
i
−x
i−1
−(x
e
−x
w
)s
L
i
(u
(n)
h
)
u
i
+
+
k(x
w
)
x
i
−x
i−1
u
i−1
= −(x
e
−x
w
)s
C
i
(u
(n)
h
)
(2.13)
where s
L
i
and s
C
i
are such that s
L
i
(u
h
)u
i
+ s
C
i
(u
h
) = s
i
(u
h
). Here, the process
reads: ﬁrst make an initial estimate for u
h
, then determine s
C
i
(u
h
), and s
L
i
(u
h
),
and solve the linear system (2.13). This process is repeated until convergence
is reached.
The ﬁnal solution of the two processes (i) and (ii) is the same if
s
i
(u
h
) = s
C
i
(u
h
) + s
L
i
(u
h
) u
i
.
An advantage of the alternative (ii) is that, by a judicious choice of s
L
i
we may
expect that the iterative process (ii) converges faster than process (i). The
processes are identical if we take s
L
i
= 0 for all i. However, if s is a linear
function of u, and if we know s
L
(u) = ds/du, then process (ii) may converge in
one step.
We see that by (2.12) or (2.13) u
i
is only coupled with u
i−1
nd u
i+1
, the
values in the neighbouring cells. Hence the linear system is tridiagonal. This
band structure of the matrix makes that the linear system can be solved very
eﬃciently.
In the construction of (2.12) we have used the simplest possible representa
tion for u
h
that still allows an approximation of (ku
x
). Other representations
are possible as well. Apparently, it is not necessary to be consistent in the
choice for u
h
, (i.e. in the assumptions for the approximation of u) when the
diﬀerent parts in the diﬀerential operator are discretised. E.g. we could take
s
i
= s(u
i
) piecewise constant, whereas a piecewise linear approximation was
used to determine u
x
.
Version: September 13, 2004 31
2.3.3 The convection equation
Now we consider the problem
∂
∂x
(v(x)u(x)) = s(u, x), (2.14)
where v(x) is a given function. As a boundary condition we use e.g. u = 0 at
the left boundary of the interval. Integration, analogous to (2.11) gives
x
e
x
w
s(u, x)dx =
x
e
x
w
(vu)
x
dx = v(x)u(x)[
x
e
x
w
. (2.15)
We take the same grid as in (2.3.1) and, again, a piecewise linear approximation
for u
h
. This yields the discrete equation
v(x
e
)
u
i+1
+ u
i
2
−v(x
w
)
u
i
+ u
i−1
2
= s
i
(u
h
) (x
e
−x
w
) . (2.16)
For a constant v(x) = v, this leads to
v
2
(u
i+1
−u
i−1
) = s
i
(u
h
) (x
e
−x
w
) . (2.17)
Here we see that an important problem appears. In the simplest case of a
constant v and s, we see that in equation (2.16) u
i+1
is coupled with u
i−1
only.
The approximations of u at the even nodes are coupled to each other, and also
the values at the odd nodes, but there is no coupling between the odd and
the even nodes. If the problem is deﬁned for 0 ≤ i ≤ N, and the boundary
condition is given at the inﬂow boundary x
0
(i = 0), then there is no boundary
condition to determine the values at the odd nodes. Because the linear system
is singular there is a family of possible solutions. The homogeneous equation
allows nontrivial solutions. This is also called instability: the discrete solution
is not bounded by the rhs of the discrete equation.
Notice that the technique for the discretisation of the equations in the in
terior domain cannot be used at the outﬂow boundary. There a diﬀerent dis
cretisation method should be used, because no value u
n
is available. In fact it
is this particular discretisation at the outﬂow boundary that determines the be
haviour of the discrete solution at the odd points. We consider this completely
intolerable.
2.4 Techniques for timediscretisation
2.4.1 Timespace discretisation or semidiscretisation
We have seen that (because the timedirection is unique) it makes sense to treat
the timedependence of a problem diﬀerent from the spacedependence. The dis
crete solution is determined in a number of subsequent timesteps. Usually one
uses the same timesteps for all nodal points in a grid. Thus, in a 2dimensional
32 P.W. Hemker
ﬁnite diﬀerence discretisation, the solutions are represented by ¦u
n
ij
¦, where
(i, j) denotes the nodal point x
ij
and n the timecoordinate t
n
.
By the special character of the timevariable (the future depends on the past,
and not the past on the future) the values ¦u
n
ij
¦ are computed on basis of the
known values ¦u
k
ij
¦, k = n −1, n −2, .
Another possibility for discretisation of timedependent PDEs is to defer
the timediscretisation and to apply semidiscretisation: the solution of the
equations is represented by ¦u
ij
(t)¦. The value of the solution at the point
x
ij
(or another coeﬃcient used for the spacediscretisation) is considered as a
function in time. Then, the semidiscretised PDE leads to a very large system
of ODEs for the unknown functions ¦u
ij
(t)¦, as e.g. in
d
dt
u
ij
(t) = F
ij
(¦u
kl
(t)¦
kl
), ∀i, j.
For the solution of these equations any suitable technique for the solution of
ODEs can be used.
2.4.2 FDM for a linear hyperbolic problem
In this section we will show and analyse a number of characteristic problems
that are encountered in the treatment of timedependent problems. We do this
by means of the simple convection equation
∂u
∂t
+
∂
∂x
f(u) = 0. (2.18)
As a grid we simply take a set of equidistant nodal point ¦x
i
¦. Halfway the
nodal points we situate the cell interfaces of the cell Ω
i
at x
i±
1
2
. Integration of
(2.18) over Ω
i
yields
∂
∂t
Ω
i
udx + f(u(x))
x=x
i+
1
2
x=x
i−
1
2
= 0 . (2.19)
We represent the discrete approximation by ¦u
i
¦, with u
i
the mean value of u
over the cell Ω
i
. Then we arrive at the equation
(x
i+
1
2
−x
i−
1
2
)
∂
∂t
u
i
(t) + f
x
i+
1
2
−f
x
i−
1
2
= 0 . (2.20)
Here, f
x
i±
1
2
denotes an approximation for f(u(x
i±
1
2
)).
For the space discretisation we use upwind, downwind and central ﬂux com
putations. To construct a system of equations for the variables ¦u
i
¦ we still
have to decide how we relate f(u(x
i±
1
2
)) to the values of ¦u
i
¦. As in section
2.3.3, one possible choice is
f
x
i+
1
2
=
1
2
(f(u
i
) + f(u
i+1
)),
Version: September 13, 2004 33
the central approximation. Other choices are e.g.
f
x
i+
1
2
= f(u
i
) or f
x
i+
1
2
= f(u
i+1
),
respectively the backward and the forward discretisation
4
.
For convenience we will combine the three choices in one formula, with a
parameter w. After some analysis we make a choice for w. We write
f
i+
1
2
= wf(u
i
) + (1 −w)f(u
i+1
), i = 1, 2, 3, . (2.21)
Now we have completed our semidiscretisation:
∂
∂t
u
i
(t) =
w
∆x
f(u
i−1
) +
1 −2w
∆x
f(u
i
) −
1 −w
∆x
f(u
i+1
), (2.22)
where ∆x = x
i+
1
2
−x
i−
1
2
. We write this equation brieﬂy as
∂
∂t
u
h
(t) = g
h
(u
h
(t)). (2.23)
We realise the time discretisation of (2.23) simply by timestepping with
steps ∆t. This can be done e.g. by the forward Euler method:
1
∆t
u
h
(t
n+1
) −u
h
(t
n
)
= g
h
(u
h
(t
n
)). (2.24)
In this case the new values u
h
(t
n+1
) are expressed explicitly in the old values
u
h
(t
n
).
An alternative possibility is the backward Euler method:
1
∆t
u
h
(t
n+1
) −u
h
(t
n
)
= g
h
(u
h
(t
n+1
)), (2.25)
or also the CranckNicolson method
1
∆t
u
h
(t
n+1
) −u
h
(t
n
)
= ( g
h
(u
h
(t
n
)) + g
h
(u
h
(t
n+1
)) )/2. (2.26)
These two methods are “implicit” timediscretisation methods. Here we have
to solve a system of equations to compute the new values u
h
(t
n+1
).
We summarise the three possibilities, again, in one parametrised formula:
1
∆t
u
h
(t
n+1
) −u
h
(t
n
)
= (1 −θ)g
h
(u
h
(t
n
)) + θg
h
(u
h
(t
n+1
)) . (2.27)
The fully discretised system reads:
u
n+1
i
−u
n
i
=
∆t
∆x
θ ¦wf
n+1
i−1
+ (1 −2w)f
n+1
i
−(1 −w)f
n+1
i+1
¦+
∆t
∆x
(1 −θ) ¦wf
n
i−1
+ (1 −2w)f
n
i
−(1 −w)f
n
i+1
¦,
(2.28)
4
With ∂f/∂u > 0, the approximation f
x
i+
1
2
= f(u
i
) is called upwind approximation,
because the ﬂow of information is from left to right and for the approximation of f
x
i+
1
2
the
upstream value u
i
is used. The other approximation is called downstream approximation.
34 P.W. Hemker
where f
k
j
represents f(u
k
j
).
In general, to compute u
n+1
i
from u
n
i
, we have to solve a system of nonlinear
equations. Because it is diﬃcult to analyse nonlinear equations we simplify the
problem by linearising it. We set
df
du
= f
constant, and we combine a few
parameters in a single: λ = f
∆t/∆x. Now equation (2.28) becomes
u
n+1
i
−u
n
i
= λθ ¦wu
n+1
i−1
+ (1 −2w)u
n+1
i
−(1 −w)u
n+1
i+1
¦+
λ(1 −θ) ¦wu
n
i−1
+ (1 −2w)u
n
i
−(1 −w)u
n
i+1
¦ .
(2.29)
Having simpliﬁed the equation so far, it is simple to solve it also analytically.
We will use this knowledge about the analytic solution to check if our numerical
techniques yield reliable results.
The equation now is u
t
+ f
u
x
= 0, and the solution is ψ(x − f
t) for an
arbitrary initial solution ψ(x) = u(0, x). In time the solution doesn’t really
change, it doesn’t grow or shrink, it only moves with the velocity f
. With the
diﬀerent possible choices for the numerical method, we can see which properties
of the true solution are inherited by the numerical counterparts.
In order to study the behaviour of the solution of the discrete system, we
consider the behaviour in time of the approximate solution in the domain Ω =
(−∞, +∞), for an initial function u(t
0
, x) = e
jωx
. We use the notation j :=
√
−1 in order not to interfere with the index i previously used. (We refer to a
textbook in Fourier analysis to understand the relevance of this choice.) The
discrete initial function, then, is
u
0
i
= e
jωhi
.
Using equation (2.29) we see immediately that the solution at a later stage is
given by
u
n
i
= γ
n
e
jωhi
, (2.30)
where γ is a complex number that satisﬁes
γ −1 = λθ ¦wγe
−jωh
+ (1 −2w)γ −(1 −w)γe
+jωh
¦+
λ(1 −θ) ¦we
−jωh
+ (1 −2w) −(1 −w)e
+jωh
¦ .
(2.31)
This can also be written as
γ =
1 + λ(1 −θ)¦(1 −2w) + we
−jωh
−(1 −w)e
+jωh
¦
1 −λθ¦(1 −2w) + we
−jωh
−(1 −w)e
+jωh
¦
(2.32)
=
1 + λ(1 −θ)¦(1 −2w)(1 −cos(ωh)) −j sin(ωh)¦
1 −λθ¦(1 −2w)(1 −cos(ωh)) −j sin(ωh)¦
;
γ is the ampliﬁcation factor for the timeintegration: [u
n
i
[ = [γ[
n
. The solutions
increases for t → ∞ if [γ[ > 1, and it decreases for [γ[ < 1. We see that
[γ[
2
=
¦1 + λ(1 −θ)(1 −2w)(1 −cos(ωh))¦
2
+ λ
2
(1 −θ)
2
sin
2
(ωh)
¦1 −λθ(1 −2w)(1 −cos(ωh))¦
2
+ λ
2
θ
2
sin
2
(ωh)
(2.33)
Now we want to study the eﬀect of the parameter w, that was introduced in
the space discretisation, and of θ that was introduced in the time discretisation.
Further we can see what is the eﬀect of the parameter λ.
Version: September 13, 2004 35
• If w =
1
2
(i.e. for central discretisation in space) equation(2.33) reduces to
[γ[
2
=
1 + λ
2
(1 −θ)
2
sin
2
(ωh)
1 + λ
2
θ
2
sin
2
(ωh)
, (2.34)
and we ﬁnd
[γ[ < 1 ⇔ (1 −θ)
2
≤ θ
2
⇔ θ ≥
1
2
Depending on θ we can draw the following conclusion:
i) θ =
1
2
⇔ [γ[ = 1. This implies neutral stability (CrankNicolson).
ii) θ = 0 ⇔ [γ[ > 1. This means unconditional instability (forward Euler).
iii) θ = 1 ⇔ [γ[ < 1. This means unconditional stability (backward Euler).
• If w = 1 (i.e. for “backward” discretisation in space)
stability ⇔ [γ[ ≤ 1 ⇔ λ(λ(1 − 2θ) −1) ≤ 0. Depending on the sign of f
we ﬁnd:
– f
> 0 ⇔ λ > 0 ⇔ stability for: λ(1 −2θ) ≤ 1.
This implies stability for θ ≥
1
2
. If θ <
1
2
stability may depend on λ;
stability is obtained for λ ≤
1
1−2θ
and we obtain stability only for λ
(i.e. the timestep) small enough.
– f
< 0 ⇔ λ < 0 ⇔ stability for: λ(1 −2θ) ≥ 1.
This implies instability for θ∈[0,
1
2
]. If θ >
1
2
then stability is guar
anteed for [λ[ ≥
1
2θ−1
and we obtain stability only for λ large (!)
enough.
• If w = 0 (i.e. for “forward” discretisation in space)
stability ⇔ [γ[ ≤ 1 ⇔ λ(λ(1 −2θ) + 1) ≤ 0
– f
> 0 ⇔ λ > 0 ⇔ stability for λ(1 −2θ) ≤ 1.
This implies instability for θ∈[0,
1
2
]; if θ >
1
2
then get stability for
[λ[ ≥
1
2θ−1
, i.e. only for λ large(!) enough.
– f
< 0 ⇔ λ < 0; this implies stability for θ ≥
1
2
; if θ∈[0,
1
2
] we get
stability for [λ[ ≤
1
1−2θ
, i.e. we obtain stability for λ small enough.
We recognise a symmetry If f
> 0 then w = 1 corresponds with upwind dis
cretisation and w = 0 with downwind; with f
< 0 it is the other way. We
notice that the upwind discretisation is unconditionally stable when θ ≥
1
2
. For
θ <
1
2
it is stable if the timestep is small enough. The downwind discretisation
is unstable for all reasonable cases.
2.4.3 The equivalent diﬀerential equation
In this section we study the local discretisation error (i.e. the amount to which
the true solution at the nodal points does not satisfy the discrete equation
5
)
5
For the diﬀerential operator L : X →Y and its discrete operator L
h
: X
h
→Y
h
, and the
restriction operators R
h
: X →X
h
and R
h
: Y →Y
h
, we remember the deﬁnitions for global
36 P.W. Hemker
for (2.28). We consider again the linear problem
∂
∂t
u + f
∂
∂x
u = 0, (2.35)
and its discretisation equation (2.28). We can write (2.29) also as
u
n+1
i
−u
n
i
= −λθ¦
1
2
(u
n+1
i+1
−u
n+1
i−1
) + (
1
2
−w)(u
n+1
i+1
−2u
n+1
i
+ u
n+1
i−1
)¦+
−λ(1 −θ)¦
1
2
(u
n
i+1
−u
n
i−1
) + (
1
2
−w)(u
n
i+1
−2u
n
i
+ u
n
i−1
)¦.
(2.36)
In order to compute the local truncation error, we expand the function u(x, t)
in a Taylor series around the point x = x
i
, and t
= t
n
+ θ∆t. We denote
u
:= u(x
i
, t
). We obtain
(i) u
n
i
= u
−θ∆t
∂u
∂t
+
1
2
θ
2
∆t
2
∂
2
u
∂t
2
+O(∆t
3
);
(ii) u
n+1
i
= u
+ (1 −θ)∆t
∂u
∂t
+
1
2
(1 −θ)
2
∆t
2
∂
2
u
∂t
2
+O(∆t
3
);
(iii) u
i−1
= u
i
−∆x
∂u
∂x
i
+
1
2
∆x
2
∂
2
u
∂x
2
i
+O(∆x
3
);
(iv) u
i+1
= u
i
+ ∆x
∂u
∂x
i
+
1
2
∆x
2
∂
2
u
∂x
2
i
+O(∆x
3
);
(v)
1
2
(u
i+1
−u
i−1
) = +∆x
∂u
∂x
i
+O(∆x
3
);
(vi) (u
i+1
−2u
i
+ u
i−1
) = ∆x
2
∂
2
u
∂x
2
i
+O(∆x
3
);
(vii)
∂u
∂x
n
i
=
∂u
∂x
−θ ∆t
∂
2
u
∂x∂t
+O(∆t
2
);
(viii)
∂u
∂x
n+1
i
=
∂u
∂x
+ (1 −θ) ∆t
∂
2
u
∂x∂t
+O(∆t
2
);
(ix) (1 −θ)
∂u
∂x
n
i
+ θ
∂u
∂x
n+1
i
=
∂u
∂x
+O(∆t
2
);
(x) (1 −θ)
∂
2
u
∂x
2
n
i
+ θ
∂
2
u
∂x
2
n+1
i
=
∂
2
u
∂x
2
+O(∆t
2
).
(2.37)
Substitution of this in (2.36) yields for the lefthandside
u
n+1
i
−u
n
i
= ∆t
∂u
∂t
+
1
2
(1 −2θ)∆t
2
∂
2
u
∂t
2
+O(∆t
3
) =
and for the righthandside
−λθ
∆x
∂u
∂x
n+1
i
+ (
1
2
−w)∆x
2
∂
2
u
∂x
2
n+1
i
+O(∆x
3
)
−λ(1 −θ)
∆x
∂u
∂x
n
i
+ (
1
2
−w)∆x
2
∂
2
u
∂x
2
n
i
+O(∆x
3
)
¸
= −λ
∆x
∂u
∂x
+O(∆t
2
)
+ (
1
2
−w)∆x
2
∂
2
u
∂x
2
+O(∆t
2
)
¸
+O(∆x
3
) .
discretisation error
u(x
i
) −u
i
 = (R
h
u)
i
−u
i
,
and for the local discretisation error
(L
h
R
h
u −R
h
Lu)
i
.
Version: September 13, 2004 37
Making use of the parameter λ = f
∆t/∆x we get the expression for the trun
cation error (in equation (2.36) τ
(u) = rhs − lhs)
τ
(u)
∆t
= −
∂u
∂t
−(
1
2
−θ)∆t
∂
2
u
∂t
2
+O(∆t
2
)
−f
∂u
∂x
+O(∆t
2
) + (
1
2
−w)∆x
∂
2
u
∂x
2
+O(∆x
2
)
= (θ −
1
2
)∆t
∂
2
u
∂t
2
+ f
(w −
1
2
)∆x
∂
2
u
∂x
2
+O(∆x
2
) +O(∆t
2
) .
This means that the method is ﬁrst order consistent in time and space. If θ =
1
2
it is second order in time, and if w =
1
2
it is second order in space.
Diﬀerentiating the PDE (2.35) we ﬁnd that u
tt
= (f
)
2
u
xx
, so that the
principal term in the truncation error can be written as
f
¸
(θ −
1
2
)f
∆t + (w −
1
2
)∆x
u
xx
= (2.38)
f
∆x
¸
(θ −
1
2
)λ + (w −
1
2
)
u
xx
=
num
u
xx
. (2.39)
So, we may see the diﬀerence scheme as a second order discretisation of a mod
iﬁed equation (or the equivalent diﬀerential equation)
u
t
= f
∆x
¸
(θ −
1
2
)λ + (w −
1
2
)
u
xx
−f
u
x
. (2.40)
For this reason
num
is also called the numerical diﬀusion of the diﬀerence
scheme.
Remark:
Notice that the stability of the diﬀerence scheme as studied in Section 2.4.2
corresponds exactly with a positive numerical diﬀusion!
The numerical diﬀusion, however, can cause problems in a convection diﬀu
sion problem if the diﬀusion coeﬃcient in the equation is (much) smaller than
the numerical diﬀusion. The numerical diﬀusion may then overrule the ‘physi
cal’ diﬀusion, and the discrete solutions are much more smeared than the true
solution of the diﬀerential equation.
If we make computations with one of the above schemes, we notice:
(i) the numerical diﬀusion can have a signiﬁcant eﬀect;
(ii) the second order scheme (without numerical diﬀusion) shows an “overshoot”,
i.e in the numerical solutions maxima or minima appear that do not exist in the
true solution;
(iii) unstable scheme are useless.
38 P.W. Hemker
Example 2.4.1
We solve the equation
u
t
= 0.01 u
xx
−u
x
on the domain t ≥ 0, x ≥ 0, with initial condition u = 0 at t = 0, and boundary
condition u = 1 at x = 0 (at the “inﬂow” boundary). We discretise the term
u
xx
by
u
xx
≈
1
2
(u
n+1
i+1
−2u
n+1
i
+ u
n+1
i−1
) +
1
2
(u
n
i+1
−2u
n
i
+ u
n
i−1
) .
The other terms are discretised by (2.29), and we solve the problem for diﬀerent
values of w and θ. In the Figures 2.10 and 2.11 we show the results for t = 1
(and in Figure 2.12 for t = 0.5). We take a ﬁxed ∆x = 0.01 and either ∆t = 0.05
(λ = 0.5) or ∆t = 0.1 (λ = 1.0). For comparison the true solution is also shown,
as well as the solution of the modiﬁed equation
u
t
= (0.01 +
num
)u
xx
−u
x
.
Version: September 13, 2004 39
∆t = 0.1 ∆t = 0.05
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
θ = 1,
num
= 0.1 θ = 1,
num
= 0.075
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o o o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o o o
θ = 0.5,
num
= 0.05 θ = 0.5,
num
= 0.05
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o o o o
o
o
o
o
o
o
o
o o o o o o o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o o o o
θ = 0,
num
= 0.0 θ = 0,
num
= 0.025
Figure 2.10: Numerical solution of the convection diﬀusion equation
In all ﬁgures the numerical solution is given for t = 1; The true sulution is
given by the solid line; the dashed line represents the solution of the modiﬁed
equation. The dotted line is the numerical approximation; obtained with
upwind space discretisation: w = 1.0, ∆x = 0.1.
40 P.W. Hemker
∆t = 0.1 ∆t = 0.05
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o
o o o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o o
o
o
o
o
o
o
o
o
o
o
o
o
o o o o
θ = 1,
num
= 0.05 θ = 1,
num
= 0.025
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o
o o
o
o
o
o
o
o
o
o
o
o
o
o o o o o o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o o o o
o o
o
o
o
o
o
o
o
o
o
o o o o o
θ = 0.5,
num
= 0.0 θ = 0.5,
num
= 0.0
0
0.5
1
1.5
2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o
o
o
o
o
o
o
o
o
o
o
o o o o o o o o o o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o o
o
o
o
o
o
o
o
o
o
o
o
o
o o o o o o o
θ = 0,
num
= −0.05 θ = 0,
num
= −0.025
Figure 2.11: Numerical solution of the convection diﬀusion equation
In all ﬁgures the numerical solution is given for t = 1; The true sulution is
given by the solid line; the dashed line represents the solution of the modiﬁed
equation. The dotted line is the numerical approximation; obtained with
central space discretisation: w = 0.5, ∆x = 0.1.
Version: September 13, 2004 41
0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o
o
o
o
o
o
o
o o o o o o o o o o o o o o
0.2
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
o
o
o
o
o
o
o
o
o o o o o o o o o o o o o
∆t = 0.1,
num
= −0.05 ∆t = 0.05,
num
= −0.025
Figure 2.12: Numerical solution of the convection diﬀusion equation
In these ﬁgures the numerical solution is given for t = 0.5; The true sulution is
given by the solid line; the dashed line represents the solution of the modiﬁed
equation. The dotted line is the numerical approximation; obtained with
central space discretisation (w = 0.5), and forward Euler timediscretisation
(θ = 0), ∆x = 0.1.
Figure 2.13: Nonconservative interpolation
42 P.W. Hemker
2.5 Criteria for good discretisation methods
In this section we want to summarise a set of rules and give criteria that can
help us to ﬁnd good discretisation methods. The last three rules are taken from
[19].
Rule 1. If the number of nodal points (coeﬃcients for the representation)
increases we may expect (and we should require) that the discrete solutions con
verge to a true solution of the continuous problem.
This property of a discretisation is usually mathematically formulated by an
asymptotic statement: if the meshwidth h of the discretisation is suﬃciently
small, the discretisation error (in the solution) (in some norm) should vanish
(at some rate) as h tends to zero.
Rule 2 Often we wish more than only an asymptotic validity. We wish
that a reasonably accurate solution is already obtained for a reasonably ﬁne (=
reasonably coarse) grid. We do not want to construct an extremely ﬁne grid
before we can be sure about the value of the approximation.
This requirement is more diﬃcult to formulate mathematically. Often it
is expressed that we want to obtain a “physically realistic” solution. I.e. one
requires that the discrete solution already for a coarse grid shows the same
“global character” as the continuous solution.
For example, if it is known that the continuous solution is positive or mono
tone, one often requires that also the discrete solution is positive or monotone.
E.g. (i) if concentrations of a chemical are computed, one doesn’t like to ﬁnd
negative concentrations by approximation errors; (ii) for the stationary solu
tion of a heatconduction problem without sources or sinks, one requires that
the temperature as computed does not show local extrema; (iii) if the continu
ous equation describes a conservation law, one requires that also the computed
quantity is not created or does not disappear in the computation.
For the class of (hyperbolic) problems in conservation form it is relatively
simple to satisfy the requirement that the discrete equations satisfy the conser
vation law, not only globally, for all Ω, but also locally, for each cell Ω
e
of the
discretisation of Ω.
A class of (elliptic) problems for which the requirement of “physical rele
vance” is usually not too hard to satisfy, are those problems that can be formu
lated as a minimisation problem. For such symmetric elliptic problems a convex
functional exists that is minimised by the solution of the PDE.
Often the functional has a physical meaning (e.g. the energy left in the
system) and, therefore, the minimising element in a subspace satisﬁes automat
ically the requirement of physical relevance. Some discretisation methods (ﬁnite
element methods) ﬁnd this minimising approximate solution.
Rule 3: Obeying the conservation law. A conservation law is most
reliably discretised if the discrete system satisﬁes the law on each subelement
of the discretisation separately. In order to satisfy the conservation law locally
in a box method, the ﬂux over boxinterfaces should be computed consistently.
I.e. the computed ﬂux J(u) n over the box interface should be the same when
it represents the ﬂow out of the one cell or when it represents the ﬂow into the
Version: September 13, 2004 43
other. The explanation is simple: the amount of the conserved quantity that
leaves the one cell should reappear (exactly) in the other cell.
Example 2.5.1
If we use a threepoint, parabolic approximation to compute the ﬂux over the
right boundary of cell Ω
i
(ﬁg.2.13) then the requirement of conservation is not
fulﬁlled. If we take the twopoint linear approximation the conservation require
ment is satisﬁed.
Example 2.5.2
See Section 2.3.1. In equation (2.12) we have (correctly) approximated J(u) by
J(u) = k(x
i+
1
2
)
u
i+1
−u
i
x
i+1
−x
i
.
If in equation (2.12) we had approximated k(x) by k(x
i
), then the discretisation
would not have been conservative.
Rule 4: Trivial (constant) solutions should satisfy the discrete so
lution.
If a diﬀerential operator only contains derivatives of a dependent variable,
then with an approximate solution u also u + constant should be a solution of
the discrete equations. I.e. the constant function should be in the kernel of the
discrete operator. (The rowsum of the linearisation of the discrete operator
should vanish.)
This rule can be seen as an instance for the more general principle: Seeking
a good discretisation for a complex (intricate) operator, it makes sense to ﬁnd a
discretisation that at least satisﬁes basic requirements for most simpliﬁcations
of the problem. (A good method for a problem is also a good method for any
simpliﬁcation of that problem.)
Rule 5: Guaranteed positivity of the solution. In order to guarantee
the positivity of the solution (in the absence of sources and sinks) for the lin
earisation of the discrete operator, the diagonal and the oﬀdiagonal elements
of the Jacobian matrix should have diﬀerent signs.
Explanation: If we write the linearised discrete equation for one space di
mension as
a
i
u
i
= a
i−1
u
i−1
+ a
i+1
u
i+1
+ s
i
(2.41)
the coeﬃcients a
i
, a
i−1
and a
i+1
should have the same sign to guarantee that
for all possible u
i−1
and u
i+1
. The same argument is simple to generalise for
two or three dimensions.
For discretisations with a higher order of accuracy (than one) it is diﬃcult
(often impossible) to satisfy this requirement. For higher order discretisations
more nonzero diagonals appear in the linearised operator.
The ﬁfth rule allows an extension for the case that a source term is present.
First, if the source term is neglected the discretisation should satisfy rule 5.
Further, (ii) if there is a source term, it should not disturb the positivity (mono
tonicity).
44 P.W. Hemker
Let s(u) the source term in the equation Lu = s(u) allow the linearisation
s(u) ≈ s
C
+ s
L
+ u, then the discretisation of Lu = s(u) leads to (see sect
(2.3.2))
(a
i
−s
L
i
)u
i
= a
i−1
u
i−1
+ a
i+1
u
i+1
+ s
C
i
,
then s
L
i
and a
i
should have opposite signs.
Chapter 3
Finite Element Methods
3.1 Introduction
This chapter is devoted entirely to the subject of Finite Elements. First we
give a number of examples that also will be treated later in applications of the
method. Then we give the theoretical foundation of the method. Subsequently
we discuss how the method can be used in practice and ﬁnally we show how
accurate the method is.
Excellent reference books on the subject are, e.g., [2] and [3].
3.2 Examples of boundary value problems
The ﬁnite element method is best applied to (partial) diﬀerential equations in
variational form. This is best illustrated by means of elliptic equations, which
are always of even order. In this section we discuss a number of typical examples
we will frequently use. Although we will restrict ourselves in these lectures to the
examples, in general we can apply the FEM also to systems of PDEs, nonlinear
equations or problems that involve diﬀerentiation in time.
3.2.1 Onedimensional second order
The simplest problem we meet is the homogeneous Dirichlet twopoint boundary
value problem (TPBVP) on an interval Ω = [a, b] ⊂ R:
−(a
2
u
x
)
x
+ a
1
u
x
+ a
0
u = f,
u(a) = u(b) = 0.
For smooth functions a
2
, a
1
, a
0
and f, and a
2
> 0, it is known that this problem
has a unique solution.
45
46 P.W. Hemker
3.2.2 Twodimensional second order
An example of a twodimensional, secondorder, elliptic equation is the steady
state convectiondiﬀusion equation
∇ (D∇ψ) −∇ (ψv) = f on Ω ⊂ R
2
, (3.1)
where ψ(x, y) is the unknown function (for instance the temperature or the
density of some matter), v(x, y) is a (given) velocity ﬁeld, D(x, y) a diﬀusion
tensor, and f(x, y) is the source function. Special examples are the diﬀusion
equation (v = 0), the potential equation (v = 0, D is is the identity matrix)
and the Laplace equation (v = 0, D is is the identity matrix, f = 0). Because
of the special diﬃculties in the numerical treatment of the convection diﬀusion
equation, we will discuss it in more detail in a later chapter.
We write the linear twodimensional secondorder equation in its most gen
eral form:
a
11
u
xx
+ 2a
12
u
xy
+ a
22
u
yy
+ a
1
u
x
+ a
2
u
y
+ a
0
u = f, (3.2)
on a domain Ω ⊂ R
2
, where the coeﬃcients a
ij
, a
i
and f are functions of x and
y.
We assume the equation to be elliptic on Ω, which means the coeﬃcients of
the principle part of (3.2) satisfy
a
11
ξ
2
1
+ 2a
12
ξ
1
ξ
2
+ a
22
ξ
2
2
= 0, ∀ξ = (ξ
1
, ξ
2
) = (0, 0) ∈ Ω. (3.3)
It is easy to see that (3.3) is equivalent with a
2
12
< a
22
a
11
. The sign of a
11
completely determines the sign of the left hand side of (3.3).
Deﬁne the coeﬃcient matrix by
A =
a
11
a
12
a
12
a
22
,
then (3.3) is equivalent with
ξ
T
A ξ > 0, ∀ξ = 0,
i.e. the matrix A is positive deﬁnite.
The elliptic equation (3.2) is deﬁned on an open connected domain Ω ⊂ R
2
,
and to obtain a unique solution we have to provide boundary conditions. It
appears that we need one condition along all the boundary Γ := ∂Ω. This
boundary condition can be of the form
u = g on Γ, (3.4)
or, more generally,
∂u
∂n
+ b
1
∂u
∂s
+ b
0
u = h on Γ, (3.5)
where n is the outward unit normal on Γ and s is the vector tangent to the
boundary. Boundary conditions of type (3.4) are called Dirichlet conditions,
boundary conditions of the form (3.5) mixed conditions. The special case when
b
0
= b
1
= 0 is called Neumann condition.
Version: September 13, 2004 47
Theorem 3.2.1 If the coeﬃcients in equation (3.2) are taken constant, the
principle part
a
11
u
xx
+ 2a
12
u
xy
+ a
22
u
yy
can be reduced to the form
∆u
by a linear coordinate transformation.
Proof: First we notice that the principle part of (3.2) can be written as
(∇, A∇u), where we use the notation ∇ = (∂/∂x, ∂/∂y)
T
. We also deﬁne
∇
= (∂/∂ξ, ∂/∂η)
T
. The transformation we need is of the form:
ξ
η
= T
x
y
.
A direct consequence is (ξ, η) = (x, y) T
T
. Also easy to check are
∂/∂ξ
∂/∂η
(ξ, η) = I,
and a similar expression for the other coordinates. Assuming the existence of a
W so that ∇ = W∇
we see
T
T
=
∂/∂x
∂/∂y
(x, y) T
T
=
∂/∂x
∂/∂y
(ξ, η)
= W
∂/∂ξ
∂/∂η
(ξ, η) = W.
Hence ∇ = T
T
∇
, and we write the principle part as:
(∇, A∇u) = (T
T
∇
, AT
T
∇
u)
= (∇
, TAT
T
∇
u),
by the deﬁnition of adjoint. Because A is positive deﬁnite we can choose T
such that TAT
T
= D, with D = diag¦α
11
, α
22
¦, where α
11
, α
22
are positive.
Because both α
11
and α
22
are positive we can reduce the principle part of the
elliptic equation by a simple scaling to
∆u.
This completes the proof.
A transformation of equation (3.2) along the lines of the above theorem reduces
the equation to
−∆u + β
1
u
x
+ β
2
u
y
+ β
0
u = f. (3.6)
By an additional transformation of the dependent variable u we can remove the
ﬁrst order terms from (3.6). A substitution
u(x, y) = v(x, y) exp¦
1
2
(β
1
x + β
2
y)¦
48 P.W. Hemker
reduces (3.6) to
−∆v + [β
0
+
1
4
(β
2
1
+ β
2
2
)]v = f(x, y) exp¦−
1
2
(β
1
x + β
2
y)¦.
Note that the right hand side of the equation is weighted by an exponential
factor. In case of variable coeﬃcients β
j
, if the ﬁeld
β is rotation free (.e., if
∇
β = 0, so that there exists a B(x, y) such that
β = ∇B ≡ grad B), we use
the transformation
u(x, y) = v(x, y) exp¦
1
2
β dx¦. (3.7)
Inserting this in (3.6) gives
−∆v + [β
0
+
1
4
(β
2
1
+ β
2
2
) −
1
2
((β
1
)
x
+ (β
2
)
y
)]v = f exp ¦−
1
2
β dx¦.
For constant β’s this is the same as the previous reduced equation.
These computations show that any uniqueness/existence result we obtain
for −∆u + au = f will also be valid for the general case. To understand the
existence and uniqueness of such problem, in Section 3.3 we will discuss parts
of the theory on elliptic partial diﬀerential equations.
3.2.3 Fourth order problems
The results on existence and uniqueness of elliptic problems are also applicable
to fourth order elliptic problems like the biharmonic problem
∆
2
u = f in Ω,
u = 0 on Γ,
∂u
∂n
= 0 on Γ.
To this form we may reduce the Stokes problem in ﬂuid dynamics and it also
models the displacement of a thin elastic plate, clamped at its boundary, under
a transversal load.
We may consider the onedimensional homogeneous biharmonic problem on
[a, b] ⊂ R, e.g.,
a
2
u
xxxx
−a
1
u
xx
+ a
0
u = f,
u(a) = u(b) = u
(a) = u
(b) = 0.
3.3 Abstract Elliptic Theory
A powerful tool to prove the existence and uniqueness of solutions of PDEs of
elliptic type is the (generalised) Theorem of LaxMilgram 3.3.18, which can be
seen as a variant of the Riesz Representation Theorem 3.3.13. To formulate
this theory properly we ﬁrst introduce some basic notions. After this we state
the theorem and give a number of applications.
Version: September 13, 2004 49
3.3.1 Introduction to Functional Analysis
Vectorspaces
A linear space or vector space X is a nonempty set of elements, provided with
an addition and a scalar multiplication over a scalar ﬁeld K. These operations
satisfy (for all x, y ∈ X and scalars α, β ∈ K) the following properties
1. x + y = y + x,
2. x + (y + z) = (x + y) + z,
3. ∃! 0 ∈ X, such that 0 + x = x + 0 = x,
4. ∀x ∈ X, ∃(−x) ∈ X such that x + (−x) = 0,
5. α(x + y) = αx + αy,
6. (α + β)x = αx + βx,
7. (αβ)x = α(βx),
8. 1 x = x. where 1 ∈ K.
The usual ﬁelds are the ﬁelds of the real (K = R) or the complex (K = C)
numbers.
A normed linear space is a linear space X provided with a norm  , i.e.
a mapping X → R, that satisﬁes for all x ∈ X and every scalar α ∈ K the
properties
1. x ≥ 0,
2. x = 0 ⇐⇒ x = 0,
3. x + y ≤ x +y,
4. αx = [α[ x.
Property 3 is called the triangle inequality. A seminorm is a similar mapping
satisfying 1, 3 and 4, but not necessarily 2. Two norms . and .
are called
equivalent if there exist c, C > 0, such that for all u ∈ X
cu ≤ u
≤ Cu.
An inner product space is a linear space X provided with a mapping (, ) :
X X → K (the inner product or scalar product), such that for all x, y ∈ X
and scalars α, β ∈ K
1. (αx + βy, z) = α(x, z) + β(y, z);
2. (x, y) = (y, x);
3. (x, x) ≥ 0, and (x, x) = 0 ⇐⇒ x = 0.
50 P.W. Hemker
Here z denotes the complex conjugate of z ∈ C.
We say that two elements x and y of an inner product space X are perpen
dicular to each other (notation x ⊥ y) if (x, y) = 0.
Every inner product space is a normed linear space because we can deﬁne a
norm:
x :=
(x, x), ∀x ∈ X.
Quite useful is the CauchySchwarz inequality:
[(x, y)[ ≤ xy. (3.8)
Equality holds in (3.8) if and only if x and y are linearly dependent.
With the norm we have introduced a metric or distance ρ(x, y) = x − y.
So we can discuss the convergence of Cauchy sequences in X and the closure X.
A Banach space is a normed linear space which is closed in the given norm. This
means that every Cauchy sequence converges in the norm given. Such spaces
for which X = X are called complete. A Hilbert space is a inner product space
which is closed in the norm induced by the inner product. Note that diﬀerent
norms can cause diﬀerent closures and that a space can be a Banach space when
equipped with one norm but will not be a Banach space in another norm. A
subset M ⊂ X is said to be dense in X if M = X. The set X is separable
if it contains a countable dense subset, or equivalently if it contains a dense
sequence.
We conclude this section with some elementary results on Banach and Hilbert
spaces.
Lemma 3.3.1 A closed linear subspace Y of a Banach space X is also a Banach
space.
Lemma 3.3.2 Let V be a closed linear subspace of a Hilbert space H and
V = H. Then there is a space V
⊥
(the orthogonal complement) with V ⊥ V
⊥
such that each element e ∈ H can be uniquely written as e = u+v, where u ∈ V
and v ∈ V
⊥
Corollary 3.3.3
Let V be a closed linear subspace of a Hilbert space H (V = H), then there
exists an element e = 0 in H such that for all v ∈ V :
(v, e) = 0.
Corollary 3.3.4
Let V be a closed linear subspace of a Hilbert space H and let u ∈ H. If
P : H → V is the orthogonal projection on V , then Pu ⊥ u −Pu. Further
(Pv, w) = (v, Pw)
for all v, w ∈ H, P
2
= P and P = 1.
Version: September 13, 2004 51
Function spaces
Functions on Ω (mappings Ω →R or Ω →C) can be considered as elements of
linear spaces. Such linear spaces are called function spaces. Here and in what
follows we denote by Ω an open set Ω ⊂ R
d
. The boundary of Ω is denoted by
Γ = ∂Ω. This boundary is supposed to be Lipschitz continuous
1
.
Before we give examples of function spaces we ﬁrst introduce the multi
integer or multiindex notation for higher derivatives of functions in more vari
ables. Let Ω ⊂ R
d
then we will denote the ith partial derivative D
i
u of the
function u by:
D
i
u = (
∂
∂x
i
)u, i = 1, 2, . . . d.
Further we introduce a multiindex α = (α
1
, α
2
, . . . α
d
) ∈ N
d
and we write
[α[ = α
1
+ α
2
+ . . . + α
d
.
Then the αth derivative of u is
D
α
u = D
α
1
1
D
α
2
2
D
α
d
d
u = (
∂
∂x
1
)
α
1
(
∂
∂x
2
)
α
2
(
∂
∂x
d
)
α
d
u. (3.9)
We will now give a number of examples of function spaces as well as norms
and inner products we can deﬁne on them.
(
0
(Ω) is the space of all continuous functions deﬁned on Ω and
(
m
(Ω) = ¦u ∈ (
0
(Ω); D
α
u ∈ (
0
(Ω), ∀ [α[ ≤ m¦.
As the elements of these spaces are not necessarily bounded functions we also
introduce the space
(
m
(Ω) = ¦u ∈ (
m
(Ω); D
α
u are bounded and uniformly continuous on Ω,
∀ 0 ≤ [α[ ≤ m¦.
This is a Banach space with norm
u
C
m
(Ω)
= max
α≤m
sup
x∈Ω
[D
α
u(x)[, (3.10)
in which α is a multiinteger. We will also use the associated seminorms:
[u[
C
m
(Ω)
= max
α=m
sup
x∈Ω
[D
α
u(x)[. (3.11)
The space (
∞
0
(Ω) is the subset of all functions in C
∞
(Ω) with compact support:
the function values and the derivatives vanish on the boundary. Analogously
we deﬁne the subsets (
m
0
(Ω) of (
m
(Ω).
1
A boundary Γ of a subset Ω of a ddimensional space is Lipschitz continuous if it is locally
a suﬃciently smooth (d − 1)dimensional manifold. In eﬀect we can deﬁne an outward unit
normal n a.e. on Γ. See [27].
52 P.W. Hemker
Example 3.3.5
Consider the function f : R
d
→R deﬁned by
f(x) =
exp(
1
x
2
−1
) [x[ < 1,
0 [x[ ≥ 1.
This function is arbitrarily often diﬀerentiable and its support is the unit ball
T
d
= ¦x ∈ R
d
[ [x[ ≤ 1¦. Then f ∈ (
∞
0
(R
d
).
Further we deﬁne the Lebesguespaces of integrable functions [30]. First, L
p
(Ω),
1 ≤ p < ∞, is the set of functions u on Ω, such that [u[
p
is Lebesgue integrable.
On L
p
(Ω) we deﬁne the (obvious) norm:
u
L
p
(Ω)
:=
Ω
[u(x)[
p
dx
1
p
. (3.12)
As a particular case , L
2
(Ω) is a Hilbert space with inner product:
(u, v)
L
2
(Ω)
:=
Ω
u(x)v(x)dx. (3.13)
By L
p
(Ω) with p = ∞ we denote the space L
∞
(Ω) of essentially bounded
functions, with norm
f
L
∞
(Ω)
:= ess sup
x∈Ω
[f[ = inf¦λ ∈ R [ [f(x)[ ≤ λ a.e.¦. (3.14)
Lemma 3.3.6 (
∞
0
(Ω) ⊂ L
2
(Ω) and (
∞
0
(Ω) is densely embedded in L
2
(Ω).
Proof: See [1].
Remark:
As (
∞
0
(Ω) is densely embedded in L
2
(Ω) we can say that L
2
(Ω) is the completion
of (
∞
0
(Ω) in the norm
u
L
2
(Ω)
=
Ω
[u(x)[
2
dx
1/2
.
It is important to note the following inequality due to H¨ older.
Lemma 3.3.7
Let p, q be two real numbers such that q ≥ 1, p ≥ 1,
1
q
+
1
p
= 1 and f ∈ L
p
(Ω),
g ∈ L
q
(Ω). Then fg ∈ L
1
(Ω) and
fg
L
1
(Ω)
≤ f
L
p
(Ω)
g
L
q
(Ω)
. (3.15)
Proof: [30, page 33].
The H¨ older inequality (3.15) is a generalisation of the CauchySchwarz inequal
ity (3.8).
Version: September 13, 2004 53
Linear operators and functionals
Let X and Y be normed linear spaces over a ﬁeld K. A linear operator T :
X → Y is a mapping such that T(u + v) = T(u) + T(v) and T(λu) = λT(u)
for all u, v ∈ X and scalar λ ∈ K. A conjugatelinear operator S : X → Y is a
mapping such that S(u + v) = S(u) + S(v) and S(λu) =
¯
λS(u).
The norm of a linear operator T is deﬁned by
T = T
Y ←X
:= sup
x∈X, x=0
Tx
Y
x
X
. (3.16)
A linear operator is bounded iﬀ T < ∞.
Lemma 3.3.8 A linear operator T is continuous iﬀ it is bounded.
Proof:
Assume T is bounded: T < ∞. Let ε > 0, then choose δ <
ε
T
. So for all x, y
such that x−y < δ we have (because T is linear): Tx−Ty = T(x−y) ≤
Tx −y < T
ε
T
= ε. So T is continuous.
Assume T is continuous, so in 0 we have: ∀ε > 0 ∃δ > 0 : ∀x : x < δ ⇒
Tx < ε. Take an y and consider y
∗
=
δ
2
y
y
. We have y
∗
 = δ/2 < δ, so
Ty
∗
 =
δ
2
Ty
y
< ε, so
Ty
y
<
2ε
δ
, or T ≤
2ε
δ
< ∞, which means that T is
bounded.
Deﬁnition 3.3.9 A bilinear operator F : X
1
X
2
→ Y is an operator that
is linear in each of its arguments. A sesquilinear operator is a mapping F :
X X → Y that is linear in its ﬁrst argument and conjugatelinear in its
second.
A bilinear operator F : X
1
X
2
→ Y is called bounded with bound F if
F := sup
u,v=0
F(u, v)
Y
u
X
1
v
X
2
< ∞.
A bilinear operator F : X X →C is called symmetric if F(u, v) = F(v, u).
Remark:
An inner product is an example of a sesquilinear operator. In fact every strictly
positive, symmetric, sesquilinear operator can be considered as an inner prod
uct.
A special class of operators are functionals, i.e. operators that map elements of
a Banach space to the ﬁeld, K, of real or complex numbers. So we obtain the
following deﬁnitions.
Deﬁnition 3.3.10
(1) A linear functional T is a linear operator T : X → K.
54 P.W. Hemker
(2) A bilinear functional F is a bilinear operator F : X
1
X
2
→ K.
(3) A sesquilinear functional F is a sesquilinear operator F : X X → K.
Examples: Functionals
• Let X = ((Ω) and f ∈ X, then for every x
0
∈ Ω is T : f → T(f) = f(x
0
)
a bounded linear functional.
• Let X = L
p
(Ω), f ∈ X and w ∈ (
∞
0
(Ω) a given function. Then
f → T(f) =
Ω
f(x)w(x)dx (3.17)
is a bounded linear functional. The function w is called the weighting
function.
• Equation (3.17) also deﬁnes a bounded linear functional if w is an element
in L
q
(Ω) under condition that p and q are conjugate exponents, that is
p, q ≥ 1 and
1
p
+
1
q
= 1.
This follows from H¨ olders inequality (3.15).
Dual spaces
Deﬁnition 3.3.11 The space of all bounded linear functionals X → K is called
the dual space X
of X. It is equipped with the norm
x

X
:= sup
x∈X, x=0
[x
(x)[
x
X
(3.18)
We sometimes write < x
, x > instead of x
(x) and we will call < ., . > the
duality paring between X and X
.
We want to characterise the dual of some spaces. First the following elemen
tary theorem.
Theorem 3.3.12 The dual space X
of a normed linear space X is a Banach
space.
Proof: Let x
n
be a Cauchysequence in X
. This means that
∀ > 0 ∃N ≥ 0 ∀n, m ≥ N : x
n
−x
m

X
< ,
or
∀ > 0 ∃N ≥ 0 ∀n, m ≥ N : ∀x ∈ X [x
n
(x) −x
m
(x)[ < .
Now R and C are complete, so for every x the sequence x
n
(x) has a limit, say
x
(x). So,
∀ > 0 ∃N ≥ 0 ∀n ≥ N : ∀x ∈ X [x
n
(x) −x
(x)[ < .
Version: September 13, 2004 55
Now deﬁne the function x
as the function that maps every x to the number
x
(x) then we see
∀ > 0 ∃N ≥ 0 ∀n ≥ N : x
n
−x

X
< ,
or rather that the Cauchysequence x
n
has limit x
.
To characterise the dual of a Hilbert space we have the following important
theorem. It says that the dual space can be identiﬁed with the space itself.
Theorem 3.3.13 (The Riesz Representation Theorem)
Let u : X → K be a bounded linear functional on a Hilbert space X, then there
exists one and only one y
u
∈ X such that
2
u(x) = (x, y
u
)
X
∀x ∈ X. (3.19)
Proof: Consider the null space N(u) = ¦x[u(x) = 0¦. Due to continuity of
u and completeness of X, this is a closed linear subspace of X. Let P be the
projection of X to N(u). Assuming u = 0 (or we could take y
u
= 0) there is
a y
0
such that u(y
0
) = 0. Deﬁne y
1
= y
0
− Py
0
, then y
1
⊥ N(u). Further
u(y
1
) = u(y
0
) − u(Py
0
) = u(y
0
) = 0. Normalisation yields y
2
=
1
u(y
1
)
y
1
, so
that u(y
2
) =
1
u(y
1
)
u(y
1
) = 1 and still y
2
⊥ N(u). Now deﬁne y
u
:=
y
2
y
2
2
.
Take x ∈ X arbitrarily, then z := x − u(x)y
2
∈ N(u), because u(z) =
u(x) −u(x)u(y
2
) = 0. So
(z, y
2
) = 0
⇐⇒ (x −u(x)y
2
, y
2
) = 0
⇐⇒ (x, y
2
) = u(x)(y
2
, y
2
)
⇐⇒ u(x) =
(x, y
2
)
(y
2
, y
2
)
= (x,
y
2
y
2

2
) = (x, y
u
).
Uniqueness is clear, since (x, z) = 0 for all x implies z = 0.
Corollary 3.3.14
The norms of u ∈ X
and z ∈ X with u(x) = (x, z)
X
are identical:
u
X
= sup
x∈X
[u(x)[
x
X
= sup
x∈X
[(x, z)[
X
x
X
= z
X
,
so the mapping R : X
→ X of u to its representation Ru = z is a linear bijective
isomorphism.
2
The converse is trivial: let u
y
(x) = (x, y), then u
y
∈ X
.
56 P.W. Hemker
The dual of a L
p
space is a L
q
space under some conditions, as is shown by the
following theorem.
Theorem 3.3.15 (Riesz for L
p
(Ω))
L
q
(Ω) is isometric isomorph to the dual space of L
p
(Ω) if p, q ≥ 1 and
1
p
+
1
q
= 1.
If p = 1 then q = ∞, but L
∞
can not be associated with the dual of L
1
.
Proof: See [1, page 40,41]
With the notion of the dual in mind we can give a weaker deﬁnition for conver
gence. For contrast we also state the usual deﬁnition.
Deﬁnition 3.3.16 A sequence ¦u
n
¦
n
in X is called (strongly) convergent to
u ∈ X if
lim
n→∞
u
n
−u
X
= 0.
A sequence ¦u
n
¦
n
in X is called weakly convergent to u ∈ X if
lim
n→∞
ϕ(u
n
) = ϕ(u) ∀ϕ ∈ X
.
As this deﬁnes the weakest topology which renders all elements of the dual
continuous, we call it the weak topology.
Strong convergence implies weak convergence, but the converse is not true
(cf. [22]).
Remark:
If X is a Hilbert space we have by the Riesz Representation Theorem 3.3.13
that weak convergence is equivalent to
lim
n→∞
(u
n
, v)
X
= (u, v)
X
, ∀v ∈ X.
3.3.2 The Generalised LaxMilgram Theorem
Deﬁnition 3.3.17
(1) A symmetric bilinear (sesquilinear) functional F : X X → K is called
coercive (or strictly positive) if there exists a γ > 0 such that for all x ∈ X
[F(x, x)[ ≥ γx
2
. (3.20)
(2) A nonsymmetric bilinear functional F : X
1
X
2
→ K is called subcoercive
if there exists a γ > 0 such that:
∀x ∈ X
1
∃z ∈ X
2
, z = 0 [F(x, z)[ ≥ γx
X
1
z
X
2
(3.21)
and
∀z ∈ X
2
, z = 0 ∃x ∈ X
1
[F(x, z)[ > 0. (3.22)
Version: September 13, 2004 57
It is easy to see that a coercive functional also satisﬁes the properties (3.21) and
(3.22) of a subcoercive functional.
Condition (3.22) is equivalent to ∀z ∈ X
2
, z = 0 F(, z) ≡ 0. Condition
(3.21) implies that ∃γ > 0 such that for all x ∈ X
1
:
sup
z∈X
2
[F(x, z)[
z
X
2
≥ γx
X
1
. (3.23)
This means that we can ﬁnd a largest possible γ as:
γ := inf
x∈X
1
sup
z∈X
2
[F(x, z)[
x z
. (3.24)
We now come to the central theorem of this chapter: the LaxMilgram Theorem.
It discusses, in an abstract setting, the existence as well as the uniqueness and
boundedness of solutions to elliptic PDEs. We will state it in a more generalised
version than usually, namely for the general, asymmetric instead of only the
symmetric case. The treatment is such that we can easily generalise it for
discrete equations later.
Theorem 3.3.18 (Generalised LaxMilgram)
Let X
1
and X
2
be two Hilbert spaces and let F : X
1
X
2
→ K be a bounded,
subcoercive, bilinear functional. Let f ∈ X
2
, then there exists one and only
one u
0
∈ X
1
such that ∀v ∈ X
2
,
F(u
0
, v) = f(v). (3.25)
Further the problem is stable: small changes in the data f cause only small
changes in the solution u
0
:
u
0

X
1
≤
1
γ
f
X
2
.
Proof: We give the proof in six parts.
(1) As F is bounded we have a bounded linear functional F(u, ) : X
2
→ K for
each u ∈ X
1
. Now, according to the Riesz Theorem for every u ∈ X
1
a unique
element z ∈ X
2
exists such that F(u, ) = (, z)
X
2
. This deﬁnes an operator
R : X
1
→ X
2
such that
F(u, ) = (, Ru)
X
2
. (3.26)
(2) The operator R is linear, due to the bilinearity of F.
(3) To prove that R is a surjection we ﬁrst prove that R(X
1
) is a closed subspace
of X
2
. We ﬁrst note that:
Ru
X
2
= sup
v∈X
2
[(Ru, v)[
v
X
2
= sup
v∈X
2
[F(u, v)[
v
X
2
.
58 P.W. Hemker
(a) An easy consequence is that R is bounded, and so continuous:
R
X
2
←X
1
= sup
u∈X
1
Ru
X
2
u
X
1
= sup
u∈X
1
1
u
sup
v∈X
2
[F(u, v)[
v
X
2
= F.
(b) F is subcoercive, so we get using (3.21) or (3.23):
Ru = sup
v
[F(u, v)[
v
≥ γu. (3.27)
(c) Now we show that R(X
1
) is closed or that every accumulation
point of R(X
1
) is an element of it. So let v be an accumulation
point of R(X
1
), then it is the limit point of some Cauchysequence
¦Ru
n
¦. Due to (3.27) we know Ru
m
−Ru
n

X
2
≥ γu
m
− u
n
, so
¦u
n
¦ is a Cauchy sequence in X
1
. Because X
1
is a Banach space
∃!u ∈ X
1
, u = limu
n
and so ∃!Ru ∈ X
2
as limit of ¦Ru
n
¦. Clearly
v = Ru, by continuity of R, so v ∈ R(X
1
). Hence it follows that
R(X
1
) = R(X
1
) ⊂ X
2
.
(4) We prove that the range of R actually is all of X
2
. To do so we ﬁrst assume
that R(X
1
) = X
2
. This means there exists a v
0
= 0 ∈ X
2
such that v
0
⊥ R(X
1
),
or (v
0
, Ru)
X
2
= 0 for all u ∈ X
1
(by Corollary 3.3.3). By (3.26) F(, v
0
) ≡ 0,
which contradicts the subcoercivity of F. This implies R(X
1
) = X
2
. Now we
have proved that R is surjective. So we obtain ∀v ∈ X
2
∃u ∈ X
1
Ru = v
which means ∃R
−1
: X
2
→ X
1
.
(5) In fact R
−1
is the solution operator for the equation (3.25). This is seen as
follows. If we take a righthandside of our equation f ∈ X
2
, then according to
the Riesz theorem ∃!v
0
(v, v
0
) = f(v), ∀v ∈ X
2
. If we deﬁne u
0
= R
−1
v
0
then
f() = (, v
0
) = (, Ru
0
) = F(u
0
, ),
so u
0
is the solution of equation (3.25).
(6) Finally we prove the stability of the problem, using (3.27): γu
0
 ≤ Ru
0
 =
v
0
 = f, or:
u
0
 ≤
1
γ
f.
Uniqueness is a direct consequence of this stability.
3.3.3 Distributions and Sobolev Spaces
Solutions to (weak formulations of) elliptic PDEs are often found among special
classes of functions: the so called Sobolev spaces. In this section we will discuss
these spaces and some of the properties of their elements. Because it involves
weak formulations of PDEs we ﬁrst treat the subject of distributions.
Version: September 13, 2004 59
Distributions
We deﬁne the function space T(Ω) as the set of functions (
∞
0
(Ω) supplied with
the following topology. We say that a sequence ϕ
n
∈ T(Ω) converges to ϕ ∈
T(Ω), notation
lim
n→∞
ϕ
n
= ϕ,
if and only if there exists a compact subset K ⊂ Ω such that for every function
ϕ
n
we have supp(ϕ
n
) ⊂ K and such that for all [α[ ≥ 0
D
α
ϕ
n
→ D
α
ϕ.
So T(Ω) is the same set of functions as (
∞
0
(Ω), but it has more structure.
Deﬁnition 3.3.19 A distribution (or generalised function) on Ω is a bounded
linear functional on T(Ω). The space of distributions is denoted by T
(Ω), being
the dual of T(Ω).
Example 3.3.20 (Generalised functions)
A function f ∈ L
p
(Ω) deﬁnes
3
a distribution T
f
: T(Ω) →R by:
T
f
(ϕ) =
Ω
f(x)ϕ(x)dx, ∀ϕ ∈ T(Ω).
In this way we can identify each L
p
function with a distribution from T
(Ω).
Sometimes we write T
f
(ϕ) =< f, ϕ >, in which < , > is the duality paring
between T(Ω) and T
(Ω).
Example 3.3.21
Let x
0
∈ Ω and deﬁne T
x
0
: T(Ω) →R as
T
x
0
(ϕ) = ϕ(x
0
).
This generalised function is called the Dirac delta function associated with point
x
0
and it is also called δ
x
0
. We write < δ
x
0
, ϕ > or
Ω
δ
x
0
(x)ϕ(x)dx.
However, the latter is rather formal because it has nothing to do with an actual
integration.
3
In fact we can extend the class of such functions to the locally integrable functons
L
1
loc
(Ω) =
¸
f  f∈L
1
(K) ∀compactK ⊂ Ω
¸
.
60 P.W. Hemker
The topology on T
(Ω) is deﬁned as follows. With T
n
, T ∈ T
(Ω) we say that
lim
n→∞
T
n
= T
when:
lim
n→∞
T
n
(ϕ) = T(ϕ), ∀ϕ ∈ T(Ω).
Because of the analogy with weak convergence we will call this the weak topology.
Deﬁnition 3.3.22 The distributional or generalised derivative D
α
T of a distri
bution T is deﬁned by
D
α
T(ϕ) = (−1)
α
T(D
α
ϕ), ∀ϕ ∈ T(Ω).
If we have a diﬀerentiable f, say f ∈ (
1
(Ω), then we may consider the dis
tribution T
f
associated with f, and the distribution T
Df
associated with the
derivative Df of f. Now we see, for any ϕ ∈ T(Ω)
T
Df
(ϕ) =
Ω
Df(x)ϕ(x)dx = −
Ω
f(x)Dϕ(x)dx = −T
f
(Dϕ) = DT
f
(ϕ).
This means that, for diﬀerentiable functions, the generalised derivative of the
(generalised) function is the generalised function of the derivative. However, for
all (generalised) functions all generalised derivatives (always) exist.
Remark:
The mapping D
α
: T
(Ω) → T
(Ω) is continuous because if
lim
n→∞
T
n
= T in T
(Ω),
then ∀ϕ ∈ T(Ω):
D
α
T
n
(ϕ) = (−1)
α
T
n
(D
α
ϕ) → (−1)
α
T(D
α
ϕ) = D
α
T(ϕ).
Example 3.3.23 (Heaviside and deltafunction)
We consider the Heaviside function H(x), which yields 0 if x < 0 and 1 if x ≥ 0.
Of course, this is not an L
2
function, (if Ω = R) but it is a distribution:
T
H
(ϕ) =
R
H(x)ϕ(x) dx =
∞
0
ϕ(x) dx,
which is clearly bounded as ϕ ∈ (
∞
0
(Ω). We compute the derivative DT
H
:
DT
H
(ϕ) = −T
H
(Dϕ) = −
R
H(x)Dϕ(x)dx
= −
∞
0
Dϕ(x)dx = ϕ(0) = T
δ
0
(ϕ).
This means DT
H
= T
δ
0
, i.e. the generalised derivative of the Heavisidefunction
is the Dirac deltafunction.
Version: September 13, 2004 61
Sobolev spaces
In the theory of elliptic PDEs, functions that, together with their (generalised)
derivatives, are elements of L
2
(Ω), play an important role. Therefore, we intro
duce the so called Sobolev spaces.
Deﬁnition 3.3.24 The Sobolev space W
n,p
(Ω), p ≥ 1, n = 0, 1, . . . , is the
space of all L
p
(Ω)functions u for which all distributional derivatives D
α
u, [α[ ≤
n are also elements of L
p
(Ω):
W
n,p
(Ω) = ¦u ∈ L
p
(Ω)[D
α
u ∈ L
p
(Ω), ∀ [α[ ≤ n¦.
This is a Banach space if we provide it with the norm:
u
n,p,Ω
:=
¸
α≤n
D
α
u
p
L
p
(Ω)
1
p
. (3.28)
In case of p = 2 we denote the space W
n,2
(Ω) as H
n
(Ω). Supplied with the
inner product
(u, v)
W
n,2
(Ω)
=
¸
α≤n
(D
α
u, D
α
v)
L
2
(Ω)
, (3.29)
these Sobolev spaces are Hilbert spaces. We write u
n,Ω
:= u
n,2,Ω
, or even
u
n
:= u
n,2,Ω
.
Another way of saying that a functional u is an element of W
n,p
(Ω) is that they
are L
p
functions u
α
, 0 ≤ [α[ ≤ n such that
Ω
u
α
(x) ϕ(x) dx = (−1)
α
Ω
u(x) D
α
ϕ(x) dx,
for all ϕ ∈ T(Ω).
Example 3.3.25
The deﬁnition is not trivial. For instance the Heaviside function on an interval
Ω = [a, b], −∞ < a < 0 < b < ∞ shows that u ∈ L
p
(Ω) does not imply
Du ∈ L
p
(Ω).
Example 3.3.26
Consider the basic hat function on the interval Ω = [0, 1]:
u(x) =
2x for x ∈ [0,
1
2
)
2 −2x for x ∈ [
1
2
, 1].
¸
.
This function is in L
2
(Ω). To look if it is also in H
1
(Ω) we must ﬁnd an L
2
(Ω)
function du that is the derivative of u in distributional sense:
Ω
du(x) ϕ(x) dx =
−
Ω
u(x) Dϕ(x) dx, for all ϕ ∈ T(Ω). Now take the function du such that
du =
∂u
∂x
, except for x = 0,
1
2
, 1 where we deﬁne it (arbitrarily) zero. This
62 P.W. Hemker
function du is a simple function, so it is in L
2
, so it remains to check if this is
the generalised derivative of u.
Ω
du(x) ϕ(x) dx =
1
2
0
du(x) ϕ(x) dx +
1
1
2
du(x) ϕ(x) dx
=
1
2
0
∂u
∂x
ϕ(x) dx +
1
1
2
∂u
∂x
ϕ(x) dx
=
lim
x↑
1
2
u(x) ϕ(x) −lim
x↓0
u(x)ϕ(x)
¸
−
1
2
0
u(x)
∂ϕ
∂x
(x) dx
+
lim
x↑1
u(x)ϕ(x) − lim
x↓
1
2
u(x) ϕ(x)
¸
−
1
1
2
u(x)
∂ϕ
∂x
(x) dx
= −
1
2
0
u(x)
∂ϕ
∂x
(x) dx −
1
1
2
u(x)
∂ϕ
∂x
(x) dx
= −
Ω
u(x) Dϕ(x) dx,
where we have used the fact that
lim
x↑
1
2
u(x) = lim
x↓
1
2
u(x)
and that ϕ ∈ T([0, 1]), so ϕ(0) = ϕ(1) = 0.
The proof shows that it is essential that u is continuous in the point
1
2
, where
its (generalised) derivative is discontinuous. This raises the question under
which conditions elements of Sobolev spaces are continuous or even continuously
diﬀerentiable.
4
This is the content of the next theorem.
Theorem 3.3.27 (Sobolev’s Embedding Theorem)
Let Ω ⊂ R
d
be a bounded open region and let n > k +d/p, then every element
from W
n,p
(Ω) is k times continuously diﬀerentiable. Or:
n > k + d/p ⇒ W
n,p
(Ω) ⊂ (
k
(Ω). (3.30)
Proof: [30, page 174]
Example 3.3.28
Note that the dimension d plays a role. So for d = 1 the functions of H
1
(Ω)
are continuous, but for d = 2 they do not have to be and indeed it should be
4
An element u of a Sobolev space is continuous, bounded or whatever, if there is an
equivalent function u
such that this function has the property.
Version: September 13, 2004 63
noted that the Sobolev’s inequality (3.30) is sharp: there are functions in H
1
(Ω)
(d = 2) that are not continuous. An example of this is
f(x, y) = log [ log(x
2
+ y
2
)[ on B1
2
(0, 0).
It is not continuous in (x, y) = (0, 0), but it is not diﬃcult to prove that f
L
2 ,

∂f
∂x

L
2 and 
∂f
∂y

L
2 do exist.
Example 3.3.29
Remember that, according to Lemma 3.3.6, we have that (
∞
0
(Ω) is dense in
L
2
(Ω). Unfortunately, for a true subset Ω ⊂ R
d
, the space (
∞
0
(Ω) is not dense
in W
n,p
(Ω). For instance, consider H
1
(Ω). Assuming that (
∞
0
(Ω) is not dense
we compute the orthogonal complement:
0 = (u, ϕ)
H
1 = (u, ϕ)
L
2 + (Du, Dϕ)
L
2 = (u, ϕ)
L
2 −(∆u, ϕ)
L
2 = (u −∆u, ϕ)
L
2.
So the orthogonal complement of (
∞
0
(Ω) consists of those u that satisfy u−∆u =
0 distributionally. A non trivial example is the function e
(ξ,x)
, with [ξ[ = 1. This
function is in H
1
(Ω), but not in H
1
0
(Ω) (in terms of the trace operator in the
next section: the trace on Γ is not zero). Luckily however, (
∞
(Ω) is dense
in W
n,p
(Ω) under the Sobolev norm. This implies that we might introduce
W
n,p
(Ω) as the closure of (
∞
(Ω) in the norm  
n,p,Ω
.
We now give the following deﬁnition.
Deﬁnition 3.3.30 W
n,p
0
(Ω) is the closure of (
∞
0
(Ω) in the norm  
n,p,Ω
. In
analogy to W
n,2
(Ω) = H
n
(Ω) we write also W
n,2
0
(Ω) = H
n
0
(Ω).
Corollary 3.3.31
T(Ω) is dense in H
n
0
(Ω).
From Lemma 3.3.6 we see that H
0
0
(Ω) = H
0
(Ω). For n ≥ 1 however H
n
0
(Ω) =
H
n
(Ω), as is shown by Example 3.3.29.
We are interested in the dual spaces of the Sobolev spaces. First a deﬁnition.
Theorem 3.3.32 Let p = ∞ and
1
p
+
1
q
= 1, then
(W
p,m
0
(Ω))
∼
= W
q,−m
(Ω).
Proof: [1, page 48]
For the Sobolev spaces with p = q = 2 we write H
−m
= (H
m
0
)
, m > 0. In
particular we use H
−1
(Ω) := W
2,−1
(Ω), the dual of H
1
0
(Ω).
64 P.W. Hemker
Trace operator
For a function f ∈ (
∞
(Ω) we deﬁne the trace γ
0
f as γ
0
f = f[
Γ
i.e. the restriction
of f to Γ, the boundary of the domain Ω.
5
If Ω is a bounded area and has a Lipschitz continuous boundary then one
can show that the linear mapping
γ
0
: (
∞
(Ω) → L
2
(Γ)
is bounded in the sense that
γ
0
u
L
2
(Γ)
≤ Cu
H
1
(Ω)
.
So one can extend the operator γ
0
to a bounded linear operator
γ
0
: H
1
(Ω) → L
2
(Γ).
H
1
0
(Ω) is exactly the kernel of γ
0
:
H
1
0
(Ω) = ¦u ∈ H
1
(Ω) : γ
0
u = 0¦.
We also deﬁne the trace operator γ
1
:
γ
1
u :=
∂u
∂n
=
d
¸
i=1
γ
0
∂u
∂x
i
n
i
.
3.3.4 The Poincar´eFriedrichs Inequality.
To prove the coercivity of a bilinear functional one can sometimes use the fol
lowing lemma.
Lemma 3.3.33 (Poincar´e’s Lemma)
If Ω ⊂ R
d
is a bounded and open set then for all v ∈ H
1
0
(Ω)
Ω
[v[
2
dx ≤ C(Ω)
Ω
[∇v[dx. (3.31)
Proof: It is suﬃcient to prove the statement for functions in T(Ω) as T(Ω) is
dense in H
1
0
(Ω).
We enclose Ω in a rectangle S:
Ω ⊂ S = [a
1
, b
1
] [a
d
, b
d
].
We extend u to S by setting u(x) = 0 whenever x ∈ S ` Ω.
As u ∈ T(Ω) we have u ∈ T(S) and for every i = 1, . . . , d:
u(x) =
x
i
a
i
∂u
∂x
i
(x
1
, . . . , x
i−1
, t, x
i+1
, . . . , x
d
), dt
5
For more on trace operators we refer to [1, page 113].
Version: September 13, 2004 65
so we get by the CauchySchwarzLemma (3.8):
[u(x)[
2
≤
x
i
a
i
1
2
dt
x
i
a
i
[
∂u
∂x
i
(. . . , t, . . .)[
2
dt
≤ (b
i
−a
i
)
b
i
a
i
[
∂u
∂x
i
(. . . , t, . . .)[
2
dt
Now integrate over S:
b
1
a
1
b
d
a
d
[u(x)[
2
dx ≤ (b
i
−a
i
)
b
1
a
1
b
d
a
d
b
i
a
i
[
∂u
∂x
i
(. . . , t, . . .)[
2
dxdt
= (b
i
−a
i
)
2
b
1
a
1
b
d
a
d
[
∂u
∂x
i
(. . . , t, . . .)[
2
dx.
So we get for every i = 1, 2, . . . , n,
Ω
[u(x)[
2
dx ≤ (b
i
−a
i
)
2
Ω
[∇
i
u[
2
dx
and so
Ω
[u(x)[
2
dx ≤ C(Ω)
Ω
[∇u[
2
dx.
Corollary 3.3.34
When the set Ω is bounded, the seminorm [ [
1,Ω
is a norm over the space
H
1
0
(Ω) equivalent to the norm  
1,Ω
.
The inequality (3.31) is also known as the Poincar´eFriedrichs inequality.
3.3.5 Variational formulations for diﬀerential equations
A Dirichlet problem
Let Ω ⊂ R
d
be an open bounded area with Lipschitz continuous boundary Γ.
Let u ∈ (
2
(Ω) be a solution of the Poisson problem with homogeneous Dirichlet
boundary conditions
−∆u = f in Ω, (3.32)
u = 0 on Γ, (3.33)
then, by partial integration, for every ϕ ∈ (
∞
(Ω)
Ω
ϕf dx = −
Ω
ϕ∆udx = −
Ω
ϕ
¸
i
∇
i
∇
i
udx (3.34)
=
¸
i
Ω
∇
i
ϕ∇
i
udx −
Γ
ϕ(∇u n) ds, (3.35)
66 P.W. Hemker
where n is the outward unit normal. For all ϕ ∈ T(Ω) we get
Ω
ϕfdx =
¸
i
Ω
∇
i
ϕ∇
i
u dx.
This leads to the following formulation of a problem similar to (3.32): ﬁnd a u
with u = 0 on Γ such that
Ω
∇
i
ϕ∇
i
u dx =
Ω
ϕfdx, ∀ϕ ∈ (
∞
0
(Ω). (3.36)
We see that a solution of (3.32) is also a solution of (3.36).
Equation (3.36) is of the following form: given the bilinear form B : SV →
R deﬁned by
B(u, ϕ) =
¸
i
Ω
∇
i
ϕ∇
i
u dx
and the linear form l : V →R, deﬁned by
l(ϕ) =
Ω
ϕf dx,
ﬁnd a u ∈ S such that
B(u, ϕ) = l(ϕ) for all ϕ ∈ V. (3.37)
In our case (3.36) we see that B : H
1
0
(Ω) H
1
0
(Ω) →R is a symmetrical bilinear
form deﬁned on H
1
0
(Ω), and from Poincar´e’s Lemma 3.3.33 we get:
B(u, u) =
Ω
[∇u[
2
dx ≥ γ u
1,Ω
,
which tells us B : H
1
0
(Ω) H
1
0
(Ω) → R is coercive. Now it is an immediate
consequence of the generalised theorem of LaxMilgram 3.3.18 that for every
f ∈ H
−1
(Ω) there exists a unique solution u ∈ H
1
0
(Ω) to (3.36) for which
u
H
1
0
(Ω)
≤
1
γ
f
H
−1
(Ω)
.
If (3.32) has a solution, we can ﬁnd this solution by solving (3.36).
It is clear that there will be cases where (3.36) has a solution but (3.32)
hasn’t: for instance when u ∈ H
1
0
(Ω) but u ∈ (
2
0
(Ω). Obviously we have
generalised the meaning of (3.32). Hence we will call the solution u ∈ H
1
0
(Ω)
the weak solution.
More general problems
Let u ∈ (
2
(Ω) be a solution to
−∆u + cu = f in Ω, (3.38)
u
n
+ βu = h on Γ, (3.39)
Version: September 13, 2004 67
then for every ϕ ∈ (
∞
(Ω):
Ω
ϕfdx = −
Ω
∆u ϕ dx +
Ω
cϕ u dx
=
Ω
∇
i
ϕ∇
i
u dx +
Ω
cϕ u dx −
Γ
ϕ∇
i
u n
i
dx
=
Ω
[∇
i
ϕ∇
i
u + c ϕu ] dx −
Γ
ϕ(h −βu) ds
=
Ω
[∇
i
ϕ∇
i
u + c ϕu ] dx +
Γ
βϕu ds −
Γ
ϕh ds.
This leads to the problem: ﬁnd u ∈ S such that
B(u, ϕ) = l(ϕ), ∀ϕ ∈ V,
where
B(u, ϕ) =
Ω
∇
i
ϕ∇
i
u dx +
Ω
c ϕ u dx +
Γ
β ϕ u ds
and
l(ϕ) =
Ω
ϕ f dx +
Γ
ϕ h ds.
The bilinear operator B : H
1
(Ω) H
1
(Ω) →R is bounded and symmetric, and
it is coercive if c ≥ 0, β ≥ 0, except when c = 0 and β = 0:
[B(ϕ, ϕ)[ =
Ω
[∇ϕ[
2
dx + c
Ω
[ϕ[
2
dx + β
Γ
[ϕ[
2
ds
=
Ω
[∇ϕ[
2
dx + c
Ω
[ϕ[
2
dx + β
Γ
[ϕ[
2
ds
≥ C
¸
[ϕ[
2
+[∇ϕ[
2
¸
= Cϕ
1,Ω
, ∀ϕ ∈ H
1
(Ω),
for some C > 0, and also
[B(ϕ, ϕ)[ ≤ C
¸
[ϕ[
2
+[∇ϕ[
2
¸
= Cϕ
1,Ω
, ∀ϕ ∈ H
1
(Ω),
for some C > 0. From the LaxMilgram Theorem we conclude that there exists
one and only one solution in H
1
(Ω). So, if the solution to (3.38) with boundary
condition (3.39) exists, we can ﬁnd it by solving the weak problem.
Symmetric problems and minimisation of quadratic func
tionals
In the equations (3.32) and (3.38), where B(u, v) is a symmetric, coercive bilin
ear form, the functional J(ϕ) = B(ϕ, ϕ) can be considered as a quadratic form.
B(ϕ, ϕ) ≥ 0 and B(ϕ, ϕ) = 0 if and only if ϕ = 0. For problem (3.32) this
functional is deﬁned on X = H
1
0
(Ω), for problem (3.38) on X = H
1
(Ω).
68 P.W. Hemker
Theorem 3.3.35 Under the circumstances that B(ϕ, ϕ) is a quadratic form,
we can also characterise the solution of the problem: ﬁnd u ∈ X such that
B(u, ϕ) = l(ϕ) for all ϕ ∈ X
as: ﬁnd u ∈ X which minimises the quadratic functional
J(u) =
1
2
B(u, u) −l(u).
Proof:
J(u + ϕ) =
1
2
B(u + ϕ, u + ϕ) −l(u + ϕ)
=
1
2
B(u, u) −l(u) + B(u, ϕ) −l(ϕ) +
1
2
B(ϕ, ϕ).
For a u that satisﬁes B(u, ϕ) = l(ϕ) we see that
J(u + ϕ) = J(u) +
1
2
B(ϕ, ϕ) ≥ J(u).
Obviously, there is one and only one u ∈ X that minimises J().
Asymmetric problems
A quite diﬀerent situation exists for problems with ﬁrst derivatives and/or more
general boundary values:
−∆u + b
j
∇
j
u + cu = f on Ω,
u
n
+ αu
s
+ βu = h on Γ.
(3.40)
Then for ϕ ∈ (
∞
(Ω)
Ω
ϕfdx =
Ω
[∇
i
ϕ∇
i
u + ϕ b
j
∇
j
u + ϕ c u] dx −
Γ
ϕ(h −βu −αu
s
)ds.
This again can be written in the form
B(u, ϕ) = l(ϕ), ∀ϕ ∈ V,
but in this case the form B(u, ϕ) is not symmetric, due to the ﬁrst order deriva
tive. To prove that the conditions of LaxMilgram are satisﬁed is a bit more
hard to do. As we have seen in Theorem 3.2.1 we can get rid of the ﬁrst order
terms by a transformation of the variables. This helps us to prove existence
and uniqueness (as in Section 3.3.5), but for numerical purposes the transfor
mation is of no use. We will see an example of this in the chapter concerning
the convectiondiﬀusion equation.
Version: September 13, 2004 69
3.4 The technique of the ﬁnite element method
In this section we treat the practical aspects of the ﬁnite element method. We
give a simple error estimate and we discuss the choice of basic functions (in
one or two dimensions) and the construction of the discrete equations for an
example problem. We will also discuss the use of isoparametric elements for
the case of curved boundaries. The techniques explained here for the one and
twodimensional case can readily be generalised to three space dimensions.
3.4.1 The principles
The basic principle of the ﬁnite element method is the following. It is a Galerkin
method applied to the variational problem: ﬁnd u ∈ S such that
B(u, ϕ) = l(ϕ), ∀ ϕ ∈ V, (3.41)
in which S and V are (inﬁnitely dimensional) vector spaces. Usually we have a
conforming ﬁnite element discretisation for which the discrete system of equa
tions read: ﬁnd u
h
∈ S
h
such that
B(u
h
, ϕ
h
) = l(ϕ
h
), ∀ ϕ
h
∈ V
h
, (3.42)
where S
h
and V
h
are ﬁnite dimensional subspaces of S and V , with dim(S
h
) =
dim(V
h
). We also may replace B(, ) with an approximation B
h
(, ) and l()
with an approximation l
h
(). The consequences of this are discussed in Section
3.5.
If B(, ) is a symmetric and coercive bilinear form and if S = V then we
can show that this type of discretisation gives an approximate solution if S
h
=
V
h
⊂ S = V . The solution to (3.41) is the minimising function of a quadratic
functional J(u) over S. The discrete function u
h
minimises the same functional
over the subspace. This implies that an optimal solution is found in the norm
induced by the problem (the energy norm). We describe this in the following
theorem.
Theorem 3.4.1 Let (, )
B
= B(, ) be the inner product induced by the bi
linear form B and let e
h
= u
h
− u be the error in the approximation, then, in
the norm .
B
induced by B, the solution of (3.42) is in S
h
= V
h
the optimal
approximation of u, the solution to (3.41):
e
h

B
≤ inf
v
h
∈S
h
u −v
h

B
.
Proof: First,
e
h

2
B
= B(e
h
, e
h
) = B(e
h
, u
h
−u).
Now because B(u, v
h
) = f(v
h
) for all v
h
∈ V , and B(u
h
, v
h
) = f(v
h
) for all
v
h
∈ V
h
, we have B(u
h
− u, v
h
) = B(e
h
, v
h
) = 0 for all v
h
∈ V
h
and because
u
h
∈ S
h
= V
h
we get B(e
h
, u
h
) = 0. So
B(e
h
, u
h
−u) = B(e
h
, v
h
−u) ≤ e
h

B
u −v
h

B
∀v
h
∈ V
h
.
70 P.W. Hemker
This gives
e
h

B
≤ inf
v
h
∈V
h
u −v
h

B
.
We want to investigate the approximation in other norms (L
p
and Sobolev
norms) and for more general problems. To this purpose we will ﬁrst consider
some speciﬁc approximation spaces V
h
and in the next chapter discuss in more
detail the error estimates that are valid when we use these approximations.
3.4.2 Piecewise Lagrange Interpolation in one dimension
We divide an interval Ω = [a, b] ⊂ R in N (not necessarily equal) subintervals
e
j
= [x
j−1
, x
j
] j = 1, . . . , N.
We denote this partition by
Ω
N
: a = x
0
< x
1
< . . . < x
N
= b.
Further we deﬁne h
j
= x
j
−x
j−1
and h = max¦h
j
¦, the meshwidth of the mesh
Ω
N
.
Deﬁnition 3.4.2
(1) Let Ω
be a (connected) subinterval of Ω, then P
k
(Ω
) is the set of all
polynomials of degree less or equal k on Ω
.
(2) We deﬁne P
k
(Ω
N
) as the space of continuous, piecewise polynomial functions
on Ω:
P
k
(Ω
N
) = ¦f[f ∈ ((Ω); f[
e
i
∈ P
k
(e
i
), i = 1, . . . , N¦.
A polynomial in P
k
(e
i
) needs k +1 (independent) values in order to be deﬁned
correctly and uniquely.
Lemma 3.4.3 P
k
(Ω
N
) ⊂ H
1
(Ω). Also P
k
(Ω) ⊂ P
k
(Ω
N
), for each partition
Ω
N
.
Proof: We prove the ﬁrst, the second is obvious.
Take a polynomial p ∈ P
k
(Ω
N
), then p ∈ L
2
(Ω). Deﬁne the distribution
dp
dx
by dp(x) =
∂p
∂x
(x), x = x
i
and
dp
dx
(x
i
) = 0, then we can prove in a way similar
to Example 3.3.26 (and using the continuity of p) that dp is the distributional
derivative of p. Also dp ∈ L
2
(Ω), so p ∈ H
1
(Ω).
Now we want to determine the dimension and a basis for the space P
k
(Ω
N
) for
k = 1, 2, 3. For this purpose we ﬁrst deﬁne on the subinterval e
i
a new variable
t = (x − x
j−1
)/h
j
, so that t = 0 corresponds with x = x
j−1
and t = 1 with
x = x
j
.
The case k = 1.
Version: September 13, 2004 71
An element from P
1
(Ω
N
) is a piecewise linear function. Such a function is
determined by its values at the points x
i
. So there are N + 1 basis functions.
A natural choice is the set ¦ψ
j
¦
N
j=1
where
ψ
j
(x
i
) = δ
ij
i, j = 0, . . . , N. (3.43)
These are the so called standard hat functions. In ﬁgure 3.1 the function ψ
i
is
given.
a b x
i
Figure 3.1: Hat function ψ
j
.
It turns out that a basis function ψ
i
vanishes everywhere on Ω except on the two
subintervals to which x
i
belongs. For each subinterval e
j
there are two nonzero
basisfunctions:
ψ
j−1
(x) = 1 −(x −x
j−1
)/h
j
= (1 −t),
ψ
j
(x) = (x −x
j−1
)/h
j
= t.
The case k = 2.
An element from P
2
(Ω
N
) is a piecewise quadratic function. The functions are
not completely determined by their values on the points in the partition Ω
N
.
We need one additional value in each interval. We take in each e
j
an interior
point, for instance x
j−
1
2
= (x
j−1
+x
j
)/2, the midpoint of the interval. We now
choose the basis functions ¦ψ
i
¦ ⊂ P
2
(Ω
N
) as
ψ
i
(x
j
) = δ
i,j
i, j = 0, 1/2, 1, 3/2, . . . , (2N −1)/2, N.
The number of basis functions is 2N+1. Again ψ
i
= 0 on at most two segments
e
i
, namely those to which x
i
belongs. On each segment e
j
there are at most
three basis functions unequal to zero
ψ
j−1
(x) = (1 −t)(1 −2t),
ψ
j−
1
2
(x) = 4t(1 −t),
ψ
j
(x) = −t(1 −2t).
In ﬁgure 3.2 the functions ψ
j
and ψ
j−
1
2
are shown. All basisfunctions ψ
i
,
i integer, are similar to ψ
j
(with the exception for the two functions at the
72 P.W. Hemker
e
j
x x x x
x x x
j−1 j+1
0 1 2 N
j
Figure 3.2: Second order Lagrange.
boundary). Note that the support of ψ
j−
1
2
is completely contained in e
j
. All
functions ψ
i−
1
2
, i integer, are translated versions of ψ
j−
1
2
.
The case k = 3.
This can be treated completely analogous to the case k = 2. We get the space
P
3
(Ω
N
) of piecewise cubic functions. Now we have two extra degrees of freedom
for each interval e
j
. We can choose x
j+k/3−1
= x
j−1
+
k
3
h
j
for k = 1, 2 and the
basis ¦ψ
i
¦ is deﬁned by
ψ
i
(x
j
) = δ
i,j
i, j = 0, 1/3, 2/3, 1, 4/3, . . . , N.
On each interval e
j
we now have four nonzero functions:
ψ
j−1
(x) = −
1
2
(3t −1)(3t −2)(t −1),
ψ
j−
2
3
(x) =
9
2
(3t −2)t(t −1),
ψ
j−
1
2
(x) = −
9
2
(3t −1)t(t −1),
ψ
j
(x) =
1
2
(3t −1)(3t −2)t.
These functions are shown in Figure 3.3. Again the basis functions correspond
ing to the interior points have their support in a single interval e
j
. Obviously,
the dimension of the approximation space P
3
(Ω
N
) is 3N + 1.
Figure 3.3: Third order Lagrange.
Version: September 13, 2004 73
Remark:
For the same spaces P
k
(Ω
N
), k = 1, 2, 3, we can also choose diﬀerent basis
functions. One possibility is to use a hierarchical basis. This means that we
start with the basis for the lower dimensional approximating space (smaller k
or smaller N) and that we extend this basis to a basis of the current approxi
mating space. For instance, we make such a basis for k = 2 by starting with the
N + 1 basis functions already used as basis for the case k = 1. To this set we
add one quadratic basisfunction, for each interval e
j
. We can take for instance
ψ
j−1/2
. Similarly the basis for k = 3 can be formed by adding one cubic basis
function per interval to the (hierarchical) basis that was constructed for k = 2.
3.4.3 The construction of the discrete equations
As an example of a ﬁnite element discretisation we discuss the discretisation
of a twopoint boundaryvalue problem (TPBVP). We consider the following
problem
−(a
2
u
x
)
x
+ a
1
u
x
+ a
0
u = s in Ω,
u = 0 on Γ,
where Ω = [a, b], Γ = ∂Ω and a
2
, a
1
, a
0
and s are functions on Ω. To apply the
ﬁnite element method we ﬁrst have to formulate the problem in its the weak
form, so we seek an approximation for the function u ∈ H
1
0
(Ω) that satisﬁes
(a
2
u
x
, v
x
) + (a
1
u
x
, v) + (a
0
u, v) = (s, v), ∀ v ∈ H
1
0
(Ω). (3.44)
For our ﬁnite element discrete approximation we choose
S
h
= ¦u ∈ P
k
(Ω
N
) [ u(a) = u(b) = 0¦ ⊂ H
1
0
(Ω),
and we take our weighting functions to be the same space, so V
h
= S
h
.
With P
k
(Ω
N
) = Span ¦ψ
i
¦ we can write u
h
(x) =
¸
j
c
j
ψ
j
(x). This yields
the discrete system of kN + 1 equations
¸
j
c
j
Ω
a
2
ψ
j
ψ
i
+ a
1
ψ
j
ψ
i
+ a
0
ψ
j
ψ
i
dx =
Ω
sψ
i
dx,
one equation for every weighting function ψ
i
. Here the prime denotes diﬀerentia
tion with respect to x. The integration interval can be split into the subintervals
e
j
. We get
¸
j
c
j
¦
¸
k
e
k
a
2
ψ
j
ψ
i
+ a
1
ψ
j
ψ
i
+ a
0
ψ
j
ψ
i
dx¦ =
¸
k
e
k
sψ
i
dx,
74 P.W. Hemker
where we notice that the contributions from most of the subintervals vanish
Thus the discrete equations are a linear system
¸
j
m
ij
c
j
= s
i
, (3.45)
in which
m
ij
=
¸
k
e
k
a
2
ψ
j
ψ
i
+ a
1
ψ
j
ψ
i
+ ψ
j
ψ
i
dx,
s
i
=
¸
k
e
k
sψ
i
dx.
From equation (3.45) the coeﬃcients c
j
are to be solved, to obtain the discrete
solution
u
h
(x) =
¸
j
c
j
ψ
j
(x).
The matrix (m
ij
) is called the stiﬀness matrix and the vector (s
i
) the load
vector. (m
ij
) is composed of contributions from the separate intervals e
k
. These
contributions vanish if e
k
is not a subset of the support of both ψ
i
and ψ
j
. The
contribution of e
k
to (m
ij
) or (s
i
) is called the elementary stiﬀness matrix or
the elementary load vector corresponding to this interval.
In practice the linear system is built up by scanning every element in the
partition and adding the elementary stiﬀness matrix and elementary loadvector
to the full stiﬀness matrix or loadvector. This means that the matrix and vector
elements are not computed one by one. This technique is practical not only for
onedimensional, but also for two and threedimensional problems. It is called
the assembly of the stiﬀness matrix and load vector.
Now we show explicitly how such a matrix/vector is drawn up for a two
point boundaryvalue problem. For simplicity we take the coeﬃcients a
2
, a
1
and a
0
constant. Deﬁne
m
2,k,i,j
=
e
k
ψ
j
ψ
i
dx,
m
1,k,i,j
=
e
k
ψ
j
ψ
i
dx,
m
0,k,i,j
=
e
k
ψ
j
ψ
i
dx,
then the element matrix entries are
m
ij
=
¸
k
m
k,i,j
,
with
m
k,i,j
= a
2
m
2,k,i,j
+ a
1
m
1,k,i,j
+ a
0
m
0,k,i,j
.
Version: September 13, 2004 75
Example 3.4.4
If we take piecewise linear basisfunctions as in (3.43), for which ψ
k−1
= 1 − t,
ψ
k
= t, ψ
k−1
= −
1
h
k
and ψ
k
=
1
h
k
, we get
m
k,i,j
=
a
2
h
k
1 −1
−1 1
+ a
1
−
1
2
1
2
−
1
2
1
2
+ a
0
h
k
1
3
1
6
1
6
1
3
. (3.46)
Note that for convenience we have given only the essential part of the matrix:
the elements (k − 1, k − 1), (k, k − 1), (k − 1, k) and (k, k); all other elements
vanish.
6
The elementary stiﬀness matrices for higher order piecewise polynomials can
be computed in a similar way.
Use of quadrature rules
The computation of the matrix entries is more laborious when the coeﬃcients
a
2
, a
1
and a
0
are not (piecewise) constant. Only in special cases we can compute
the integrals m
k,i,j
exactly. In most cases it is necessary to compute the integral
using a quadrature rule:
b
a
f(x) dx ≈
l
¸
m=0
w
m
f(x
m
),
where w
m
is the weight for quadrature node x
m
. Replacing the integral by
a summation causes a new error, so that the accuracy of the quadrature may
inﬂuence the accuracy of the discretisation. We usually want to use quadrature
rules that preserve the order of accuracy of the discretisation method. The
requirements for such a quadrature rule are discussed in section 3.5.
An interesting quadrature in combination with Lagrangian interpolation is
the method that uses the same nodal points for the integration. Assume that
the approximation uses polynomials of degree l (so V
h
= P
l
(Ω
N
)) and take y
k,m
,
m = 0, . . . , l as the nodes in e
k
:
x
k−1
= y
k,0
< y
k,1
< < y
k,l
= x
k
.
Or rather,
y
k,m
:= x
k+
m
l
−1
,
so that
ψ
i
(y
k,m
) = δ
i,k+
m
l
−1
.
The elementary matrix entry becomes (k = 0, , N and i = n
i
+
m
i
l
− 1,
j = n
j
+
m
j
l
−1 where n
i
, n
j
= 0, , N, m
i
, m
j
= 0, , l)
m
k,i,j
≈
l
¸
m=0
w
m
¸
a
2
(y
k,m
)ψ
j
(y
k,m
)ψ
i
(y
k,m
)
6
Be aware of the boundaries however, see a next section.
76 P.W. Hemker
+ a
1
(y
k,m
)ψ
j
(y
k,m
)ψ
i
(y
k,m
) + a
0
(y
m
)ψ
j
(y
k,m
)ψ
i
(y
k,m
)
¸
=
l
¸
m=0
w
m
¸
a
2
(y
k,m
)ψ
j
(y
k,m
)ψ
i
(y
k,m
) (3.47)
+ a
1
(y
k,m
)ψ
j
(y
k,m
)δ
i,k+m/l−1
+ a
0
(y
k,m
)δ
j,k+m/l−1
δ
i,k+m/l−1
¸
=
l
¸
m=0
w
m
¸
a
2
(y
k,m
)ψ
j
(y
k,m
)ψ
i
(y
k,m
)
¸
+a
1
(y
k,m
i
)ψ
j
(y
n
i
,m
i
) w
m
i
+ a
0
(x
i
)δ
i,j
w
m
i
.
Example 3.4.5
In case of piecewise linear approximation (l = 1) and the trapezoidal rule
(w
0
= w
1
=
1
2
), we get
m
k,i,j
≈
a
2
(x
k−1
) + a
2
(x
k
)
2h
k
1 −1
−1 1
(3.48)
+
1
2
−a
1
(x
k−1
) a
1
(x
k−1
)
−a
1
(x
k
) a
1
(x
k
)
+
h
k
2
a
0
(x
k−1
) 0
0 a
0
(x
k
)
.
We observe that the zero order term contribution is a diagonal matrix, which
is diﬀerent from formula (3.46) for constant coeﬃcients. Using the quadrature
also for the elementary load vector, we obtain
s
k,i
≈
h
k
2
s(x
k−1
)
s(x
k
)
.
Remark:
It is known that for an accurate quadrature we better take the nodes within
each element [x
j−1
, x
j
] not equidistant. To obtain the most accurate quadrature
(and assuming that the endpoints of the interval are be nodal points) the nodes
should be placed as for Lobatto quadrature. These points, then, may determine
the nodal points for the Lagrange interpolation.
Example 3.4.6
As a last example, we give the construction of the element matrix based on
piecewise quadratic functions and the Simpson quadrature (l = 2, w
0
= w
2
=
1
6
,
w
1
=
2
3
). For simplicity we choose a
2
(x) piecewise constant on the partitioning
Ω
N
. We obtain
m
k,i,j
≈
a
2,k
6h
k
¸
14 −16 2
−16 32 −16
2 −16 14
¸
+
1
6
¸
−3a
1
(x
k−1
) 4a
1
(x
k−1
) −a
1
(x
k−1
)
−4a
1
(x
k−
1
2
) 0 4a
1
(x
k−
1
2
)
a
1
(x
k
) −4a
1
(x
k
) 3a
1
(x
k
)
¸
(3.49)
Version: September 13, 2004 77
+
h
k
6
¸
a
0
(x
k−1
) 0 0
0 4a
0
(x
k−
1
2
) 0
0 0 a
0
(x
k
)
¸
.
For a
1
and a
0
also piecewise constant this formula simpliﬁes to
m
k,i,j
≈
a
2,k
6h
k
¸
14 −16 2
−16 32 −16
2 −16 14
¸
+
a
1,k
6
¸
−3 4 −1
−4 0 4
1 −4 3
¸
+
+
a
0,k
h
k
6
¸
1 0 0
0 4 0
0 0 1
¸
. (3.50)
Again the contribution of the zeroorder term is a diagonal matrix. The ele
mentary loadvector follows after direct computation, using the same Simpson
quadrature
s
k,i
≈
h
6
¸
s(x
k−1
)
4s(x
k−
1
2
)
s(x
k
)
¸
. (3.51)
For our twopoint boundaryvalue problem and piecewise second and higher
order functions the shape of the complete matrix and righthand vector is given
in Figure 3.4.
Figure 3.4: Matrix and righthandside constructed by means of C
0
piecewise
cubic polynomials.
Static condensation, lumping
The solution of the general matrix system (3.45) is not a topic we will discuss
here. See for instance [6]. We however make some remarks.
78 P.W. Hemker
First notice, that due to the relatively small support of the basis functions,
we have a linear system that is sparse: only a fairly limited amount of the matrix
elements is nonzero.
The use of Lagrange interpolation has a nice advantage. We can reduce the
set of equations in ﬁgure 3.4 to a tridiagonal form by socalled static conden
sation. With this we mean the elimination of the unknowns belonging to the
internal nodes x
i
, for which the basis functions ψ
i
have a support that is com
pletely contained in a single element e
k
. The elimination of these unknowns is
a strictly local process that can be done elementwise: it causes no changes in
the contributions of other elements to the stiﬀness matrix and the load vector.
The same technique can be applied for 2 or 3dimensional problems, for each
basis function ψ
i
(x) that has its support completely in a single element e
k
.
Let a linear system be partitioned as
A
11
A
12
A
21
A
22
x
1
x
2
=
b
1
b
2
and let the second set of variables and equations be available for elimination,
then after elimination we get the system
(A
11
−A
12
A
−1
22
A
21
)x
1
= b
1
−A
12
A
−1
22
b
2
,
so there remains just a number of additional contributions in the matrix and
righthandside.
When all interior variables are eliminated in this way we get the tridiagonal
system of equations shown in ﬁgure 3.5.
x x
x x x
0
1
2
3
4
Figure 3.5: Matrix and righthand side after lumping.
We have seen that the zero order (not diﬀerentiated) terms consist only of
diagonal elements if we use a quadrature. The technique in which these terms
are placed on the main diagonal is called lumping
7
. This technique is interesting
7
Originally the oﬀdiagonal elements where added to the main diagonal without any ex
planation.
Version: September 13, 2004 79
for two reasons:
1. The part a
0
(x)u(x) = s(x) of the equation is discretised pointwise. This
can be preferable when a
0
(x) or s(x) is strongly varying or dominating.
2. For an initialboundaryvalue problem (IBVP), when there is time depen
dence in the problem, we can use lumping together with the method of
semidiscretisation to reduce the IBVP to an explicit set of ordinary dif
ferential equations. As an example we may consider a parabolic equation
of the form ϕ
t
= Lϕ, where L is a diﬀerential operator in the space vari
ables (for instance Lϕ = pϕ
xx
+qϕ
x
+rϕ+s), then a semidiscretisation
with S
h
= V
h
= Span ¦ϕ
i
¦ yields the following set of equations: ﬁnd
ϕ
h
(x, t) =
¸
j
c
j
(t)ϕ
j
(x) such that
(
∂
∂t
ϕ
h
, ϕ
i
) = (Lϕ
h
, ϕ
i
), ∀ϕ
i
.
So we have to compute the set ¦c
j
(t)¦ for which
¸
j
dc
j
(t)
dt
(ϕ
j
, ϕ
i
) = (Lϕ
h
(t, ), ϕ
i
), ∀ϕ
i
.
The matrix (ϕ
j
, ϕ
i
) is called the mass matrix. By using the technique
“lumping” we can recast it to a diagonal matrix (ϕ
j
, ϕ
i
) = w
i
δ
ij
so that we
obtain an explicit set of ordinary diﬀerential equations for the coeﬃcients
¦c
i
(t)¦:
dc
i
(t)
dt
=
1
w
i
(Lϕ
h
(t, ), ϕ
i
), ∀ i.
Treatment of boundary conditions
Until now we have limited ourselves to problems with homogeneous Dirichlet
boundary conditions. In this section we treat also inhomogeneous and more
general boundary conditions.
First we handle inhomogeneous Dirichlet boundary conditions. Consider
Ω = [a, b] and the equation
Lu := −(a
2
u
x
)
x
+ a
1
u
x
+ a
0
u = s on Ω, (3.52)
u = g on Γ. (3.53)
The weak form of this problem can be written as: ﬁnd a u ∈ H
1
(Ω) such that
u = g on Γ
and
B(u, v) := (a
2
u
x
, v
x
) + (a
1
u
x
, v) + (a
0
u, v) = (s, v), ∀v ∈ H
1
0
(Ω).
80 P.W. Hemker
For theoretical purposes, this problem is easily reduced to the problem with
homogeneous boundary conditions by choosing a function g ∈ H
1
(Ω) (the ex
tension of the g deﬁned on Γ) that satisﬁes the boundary conditions
8
. Deﬁning
the function w := u −g we have w ∈ H
1
0
(Ω) and w satisﬁes the equation
Lw = s −Lg,
or,
B(w, v) = (s, v) −B(g, v).
Thus the problem is reduced to the problem with homogeneous boundary con
ditions. This gives existence and uniqueness.
In practice it is also easy to implement nonhomogeneous Dirichlet boundary
conditions. Because the value of the function u at the boundary points is given
we have to compute less unknowns. This means that there are no discrete
equations needed for the boundary points.
We can handle Dirichlet boundary conditions in two ways. To show this, we
ﬁrst assume the equations are constructed elementwise as shown in the previous
section (ﬁgure 3.4). The two approaches are as follows.
I) Replace the equations for the boundary points by equations that force the
boundary conditions. For example, in the stiﬀness matrix put 1 on the main
diagonal and 0 at the other entries of this row, and put the boundary value in
the righthandside. This forces the boundary value to be part of the solution.
In this way we get a linear system of equations of the original form, of which
the size corresponds with the total number of nodes.
II) In the other approach (see Figure 3.6) we eliminate the known “unknowns”
on the boundary. The set of equations is reduced, but more important is that
the righthandside has to be adjusted at a number of places. Now the size of the
system corresponds with the interior nodes. In practice, this is less convenient,
in particular if we have a combination of Dirichlet and Neumann conditions.
In this case, however, a symmetric operator B(u, v) guarantees a symmetric
stiﬀness matrix.
If we consider more general boundary conditions than Dirichlet boundary
conditions, we see that the whole procedure can be carried out in the same
manner as before. In the discrete operator and righthandside additional terms
will appear due to the boundary condition. See for instance the B(u, ϕ) and
the l(ϕ) of section 3.3.5. Every boundary integral gives a contribution to the
elementary matrix/righthandside at each boundary point. In practice we ﬁrst
add all contributions from the interior elements to the stiﬀnessmatrix and later
the contributions from the boundary segments are added. Also these boundary
integrals can be computed by quadrature.
8
That we can actually ﬁnd such a function puts an extra condition on the boundary
condition and on the smoothness of the boundary Γ. With γ
0
the trace operator we have the
requirement that
γ
0
u = γ
0
g.
Version: September 13, 2004 81
A 0 0 0
a
a
a
1
2
3
u
0
−a u
−a u
−a u
1 0
2 0
3 0
u
3
−a u
−a u
−a u
1 3
2 3
3 3
...
...
...
...
...
...
...
Figure 3.6: Treatment of boundary conditions.
3.4.4 Other piecewise polynomial approximations in one
dimension
In every ﬁnite element method we use an approximation of the solution of the
form
u
h
(x) =
¸
j
c
j
ϕ
i
(x),
where the use of continuous, piecewise polynomial basis functions ϕ
i
(x) with a
small support is preferred. Not only Lagrange interpolation but also Hermite
interpolation can be used. Here the functions ϕ
i
(x) are not only determined by
their function values at the nodes, but also by the values of the derivatives. The
approximating basis functions thus may have larger smoothness, which gives us
the opportunity to solve problems which contain higher order derivatives. We
may ﬁnd solutions in H
k
0
(Ω) for k > 1. For instance it allows the solution of
the homogeneous Dirichlet problem for the biharmonic operator ∆
2
:
∆
2
u = f in Ω,
u = ∂
n
u = 0 on Γ.
A polynomial of degree k needs k +1 (independent) values on an interval to be
deﬁned correctly.
Deﬁnition 3.4.7 Let f be a function in (
m−1
([a, b]), then the Hermite inter
polate of f on a partition Ω
N
is the function p ∈ P
2m−1
(Ω
N
) ∩ (
m−1
([a, b]) for
which, for all x
i
∈ Ω
N
and k = 0, 1, . . . , m−1,
D
k
p(x
i
) = D
k
f(x
i
).
Note that the Hermitian interpolation of a polynomial q of degree 2m−1 is the
polynomial itself. Moreover it is not just an element of (
m−1
([a, b]), but even
from (
2m−1
([a, b]).
82 P.W. Hemker
A usual Hermitebasis for a ﬁnite element method in one space dimension is
that of the cubic Hermite splines (m=2): on every interval [x
i
, x
i+1
] the P
3
(∆)
function is determined by four parameters f(x
i
), f
(x
i
), f(x
i+1
) and f
(x
i+1
).
For every element e
j
= [x
j
, x
j+1
] there are four nonzero basisfunctions: ϕ
0,j
,
ϕ
1,j
, ϕ
0,j+1
and ϕ
1,j+1
such that D
k
ϕ
l,j
(x
i
) = δ
kl
δ
ij
. Let t = (x −x
j
)/(x
j+1
−
x
j
) then the form of these functions is on element e
j
ϕ
0,j
(t) = (t −1)
2
(2t + 1),
ϕ
0,j+1
(t) = t
2
(3 −2t),
ϕ
1,j
(t) = t(t −1)
2
(x
j+1
−x
j
),
ϕ
1,j+1
(t) = t
2
(t −1)(x
j+1
−x
j
).
Extending these functions to the entire interval [a, b] such that D
k
ϕ
l,j
(x
i
) =
δ
kl
δ
ij
we get functions ϕ
j,k
(x) ∈ H
2
([a, b]), because the derivatives are continu
ous. Except for the functions that are associated with the boundary points, all
functions ϕ
l,j
have a support that extends over two subintervals e
k
.
The smoothness of the approximating function u
h
(x) implies that it can be
used for solving fourth order elliptic problems like
au
xxxx
−bu
xx
+ cu = f, (3.54)
u(a) = u
(a) = u(b) = u
(b) = 0,
of which the corresponding variational formulation is as follows: ﬁnd u ∈ H
2
0
(Ω)
such that
Ω
a u
xx
ϕ
xx
+ b u
x
ϕ
x
+ cu ϕ dx =
f ϕ dx,
for all ϕ ∈ H
2
0
(Ω).
This problem can not be solved by Lagrangian interpolation because the
space spanned by the Lagrange interpolation polynomials ¦ϕ
i
¦ is not in H
2
(Ω).
The discrete set of equations for the two point boundary value problem
discretised with Hermite interpolation polynomials has a blocktridiagonal form
(see Figure 3.7).
Exercise 3.4.8
Consider the fourth order problem (Ω = [0, 1])
u
xxxx
−2u
xx
+ u = 1,
u(0) = u(1) = u
x
(0) = u
x
(1) = 0.
Prove that the analytic solution is
u(x) = 1 + e
x
(c
1
+ c
2
x) + e
−x
(c
3
+ c
4
x),
c
1
= −
(2e −1)(e
2
−2e −1)
e
4
−6e
2
+ 1
,
c
2
=
(e −1)(e
2
−2e −1)
e
4
−6e
2
+ 1
,
c
3
= −1 −c
1
,
c
4
= c
3
−c
1
−c
2
.
Version: September 13, 2004 83
X X X
X X X X X
X
X X X X X X
X X X X X X
X X X X X X
X X X X X X
X X X X X X
X X X X X X
X X X X X X
X X X X X X
X X X X
X X X X
Figure 3.7: Blocktridiagonal matrix.
Make a uniform discretisation with h = 1/N, compute and solve the linear
system. This gives discrete approximations for u(x
i
) and u
(x
i
). Compute
max
i
[u(x
i
) −u
h
(x
i
)[
and
max
i
[u
(x
i
) −u
h
(x
i
)[,
for N = 10, 20, 40. Compare the order of convergence you observe with the
order you expected.
Other piecewise polynomials that might be used are for instance the splines.
These are P
m
(∆)∩(
m−1
([a, b])functions. Such functions have a single degree of
freedom for each interval and they have a maximum smoothness for a polynomial
of degree m. However, in general, the support of these basic spline stretches over
many adjacent elements. Application for a ﬁnite element method is therefore
unusual, because it means that many nonzero elements will appear in each row
of the stiﬀness matrix. In other words: the use of splines causes a nonsparse
system of linear equations (although in the onedimensional case it will be a
band system).
3.4.5 Approximation in two dimensions
In the introduction to discretisation methods we have seen some possible repre
sentations of a twodimensional function. Not all of them are useful for the solu
tion of second order elliptic problems by a (conforming) ﬁnite element method.
The most important which remain are shown in Figure 3.8. A dot indicates that
at this point we take the function value, a circle that we also take the value of
the derivative, two circles the value of two derivatives, etc.
84 P.W. Hemker
Figure 3.8: Some representations of a twodimensional function.
Among these are the piecewise Lagrange interpolation polynomials on a trian
gulation. Notice that the basisfunctions deﬁned by ϕ
i
∈ P
k
(Ω
N
) (functions
with [α[ ≤ k) and ϕ
i
(x
j
) = δ
ij
are indeed continuous.
Example 3.4.9
On the standard triangle shown in ﬁgure 3.9 we ﬁnd basis functions for a P
2
(∆)
Lagrange approximation by
Φ
1
(x, y) = 2 (x + y −1) (x + y −
1
2
),
Φ
2
(x, y) = 2x (x −
1
2
),
Φ
3
(x, y) = 2y (y −
1
2
),
Φ
4
(x, y) = 4x y,
Φ
5
(x, y) = −4y (x + y −1),
Φ
6
(x, y) = −4x (x + y −1).
That the dimension of P
2
(∆) is indeed 6 can be easily seen because any element
is of the form
p(x, y) =
¸
α≤2
a
α
x
α
1
y
α
2
.
This is only possible for the set of values
α = (α
1
, α
2
) ∈ ¦(0, 0), (1, 0), (0, 1), (2, 0), (0, 2), (1, 1)¦.
To express the symmetry in this formulas more clearly, we may use barycentric
coordinates (r, s, t) deﬁned by r = x, s = y and t = 1 −x −y.
9
Then we get
Φ
1
(r, s, t) = t (2t −1),
Φ
2
(r, s, t) = r (2r −1),
Φ
3
(r, s, t) = s (2s −1), (3.55)
Φ
4
(r, s, t) = 4 r s,
Φ
5
(r, s, t) = 4 s t,
Φ
6
(r, s, t) = 4 r t.
9
So: t = 0 ⇒ x + y = 1, r = 0 ⇒ x = 0 and s = 0 ⇒ y = 0. Which give the three
boundaries of the triangle.
Version: September 13, 2004 85
(0,0) (1,0)
(0,1)
x
y
Figure 3.9: Lagrange for a triangle.
These functions are also used for isoparametrical elements, which we shall dis
cuss in the following section.
The number of degrees of freedom for a general kth order polynomial in 2 or 3
dimensions is exactly the number of points that can be distributed nicely over
a triangle or tetrahedron. If, on a rectangle or a parallelepiped, a number of
nodes is distributed in a regular way we will not be able to construct exactly
all kth order polynomials. For a parallelepiped one usually takes polynomials
of the form P
k
(x) P
k
(y) (so the tensor product of the two spaces) and for the
nodes one takes the Cartesian product of the onedimensional partitions.
Example 3.4.10
We consider the bilinear interpolation on a rectangle. A bilinear basis function
is of the form
a + bx + cy + dxy,
or,
(a
1
+ a
2
x)(b
1
+ b
2
y).
The ﬁrst form shows clearly that such a function has four degrees of freedom on
each rectangle and the second form shows that the space of bilinear functions
is the product of the two spaces of linear functions.
Having four degrees of freedom, such a bilinear function is completely deter
mined by the function values at the vertices of the rectangle.
Note that in two dimensions, on triangles, the piecewise Hermite interpolation
polynomials are continuous, but are not in (
1
(Ω). If we want to ﬁnd a piecewise
polynomial in (
1
(Ω), then we have to use a relatively intricate construction. One
of the more simple ones is a P
5
(Ω)function which has 21 degrees of freedom for
each triangle (the Argyris triangle, cf [4, p. 71], see Figure 3.10).
Note that the function u along a boundary is a 5th order function determined
by its function value and ﬁrst and second order derivatives at the vertices and
the value of
∂u
∂n
(n is the outward unit normal) at the edges midpoints.
The construction of the discrete equations and the treatment of the boundary
86 P.W. Hemker
u
u
n
u
u
n
u
u u
x y
u u u
xx xy yy
Figure 3.10: The Argyris triangle.
conditions is in principle similar to the onedimensional case: the matrix and
the righthandside will be build up elementwise. To handle the boundary con
ditions we have to compute boundary integrals. When basis functions φ
j
(x)
have a support that is completely contained in a single element, then we can
eliminate the corresponding variables by static condensation. When computing
the elementary matrices we again use a quadrature and also the combination
of the elementary basisfunctions for Lagrange interpolation and an associated
quadrature can lead to lumping.
3.4.6 Isoparametric elements
By triangulation of the domain, the solution of a problem on polygonal areas
can be performed in a relatively simple way. However when there is a curved
boundary we can approximate it by means of isoparametric elements. To deﬁne
these elements we ﬁrst introduce the following
Deﬁnition 3.4.11 Let v = ¦z
1
, . . . , z
N
k
¦ be a set of N
k
points in R
2
. Then v
is called a kunisolvent set if for every sequence of real numbers (α
1
, . . . , α
N
k
)
there is precisely one polynomial of degree ≤ k such that
P
k
(z
i
) = α
i
, i = 1, . . . , N
k
.
Examples of this are: for k = 1: the vertices of a triangle; k = 2: the vertices
and the midpoints of the edges. Examples for k > 2 are easily found.
An isoparametric element of degree k = 2 is deﬁned as follows. Let v =
¦z
1
, z
2
, . . . , z
6
¦ be a 2unisolvent set in R
2
. The matching isoparametric ele
ment is
I
∆
= ¦(x, y)[(x, y)
T
=
6
¸
i=1
z
i
Φ
i
(r, s, t); 0 ≤ r, s, t ≤ 1; r + s + t = 1¦,
where r, s and t are the barycentric coordinates as introduced in Section 3.4.5
and Φ
i
the functions deﬁned in (3.55). On the isoparametric element we now
have a parametrisation in r, s, t, with r + s + t = 1, and basis functions can be
Version: September 13, 2004 87
deﬁned on this element like it was done in case of triangular elements
u
h
(x)[
I
∆
=
6
¸
i=1
a
i
Φ
i
(r, s, t).
We see that with kth order isoparametric elements the boundary is approxi
mated by piecewise kth order polynomials.
The dependent and independent variables (u and (x, y) respectively) are
both parametrised by the same type of functions. This is the reason to call
these elements isoparametric elements.
3.5 Error estimates for the ﬁnite element method
In this section we derive error estimates for the ﬁnite element method. First
we will give the discrete version of the Generalised LaxMilgram theorem. It
gives the uniqueness of the solution to the discrete equation and it gives a ﬁrst
estimate of the error.
This theorem is in fact only applicable when we use ﬁnite dimensional sub
spaces of our original Hilbert spaces. We have a more general case when we
don’t have such subspaces or when the operators in the variational equation are
replaced by approximations (for instance by quadrature). We give also error
estimates for this case.
The error estimates depend on how good we can interpolate elements of
Banach spaces in subspaces of these Banach spaces, so we have to discuss the
interpolation theory in Banach spaces, preceded by a necessary discussion of
the formalism of the ﬁnite element method. This will give us estimates in the
Sobolev norms .
m,q,Ω
. We will also give an estimate in the L
2
norm, but for
this we need additional requirements on the problem we consider.
We also discuss pointwise convergence and superconvergence, i.e the phe
nomenon that at some particular points the accuracy of the approximation is
better than the global error estimates indicate. Finally we study the inﬂuence
of the use of particular quadrature rules on Banach space interpolation and on
superconvergence.
3.5.1 The discrete version of the Generalised LaxMilgram
In Section 3.3.2 we saw that under certain conditions a variational problem
has a solution, which is bounded by the righthandside of the equation. This
Generalised LaxMilgram Theorem 3.3.18 can be applied not only to continuous
problems, but also to the associated discrete problems, with trial functions in
S
h
⊂ S and test functions in V
h
⊂ V . Taking these discrete spaces as subspaces
of the continuous spaces we obtain a so called conforming ﬁnite element methods.
In the following theorem we show that the discrete solution obtained in this
way is “quasioptimal”. This means that, apart from a constant factor, it is (in
the Snorm) not worse than the best approximation of the solution u ∈ S in the
discrete space S
h
.
88 P.W. Hemker
Theorem 3.5.1 (Discrete Generalised LaxMilgram)
Let the requirements of the Generalised LaxMilgram Theorem 3.3.18 be sat
isﬁed. So, let X
1
and X
2
be two Hilbert spaces and let F : X
1
X
2
→ R be
a bilinear functional that is bounded and subcoercive. Further, let X
1
h
⊂ X
1
and X
2
h
⊂ X
2
be linear subspaces, such that
∀x
h
∈ X
1
h
∃z
h
∈ X
2
h
, z
h
= 0 [F(x
h
, z
h
)[ ≥ γ
h
x
h
 z
h
,
and
∀z
h
∈ X
2
h
, z
h
= 0 ∃x
h
∈ X
1
h
[F(x
h
, z
h
)[ > 0.
Finally, let f ∈ (X
2
)
, and u
0
∈ X
1
be the solution of
F(u
0
, v) = f(v) for all v ∈ X
2
.
Then there exists a unique u
h
∈ X
1
h
, the solution of the discrete problem
F(u
h
, v
h
) = f(v
h
) for all v
h
∈ X
2
h
,
and the following error estimate holds
u
h
−u
0

X
1 ≤
¸
1 +
F
γ
h
inf
x
h
∈X
1
h
u
0
−x
h

X
1. (3.56)
Proof: Let R : X
1
→ X
2
be the Riesz mapping as used in the proof of the
Generalised LaxMilgram Theorem 3.3.18
(Ru, v)
X
2 = F(u, v), for all u ∈ X
1
, v ∈ X
2
,
and let S : X
1
h
→ X
2
h
be the analogue mapping for X
1
h
and X
2
h
, i.e.,
(Su
h
, v
h
)
X
2 = F(u
h
, v
h
), for all u
h
∈ X
1
h
, v
h
∈ X
2
h
.
Then R = F and S
−1
 ≤ γ
−1
h
.
Let P
1
h
: X
1
→ X
1
h
and P
2
h
: X
2
→ X
2
h
be orthogonal projections
X
1
R
−→ X
2
P
1
h
↓ ↓ P
2
h
.
X
1
h
S
−→ X
2
h
i) Consider point (5) in the GLM theorem. With v
0
∈ X
2
such that f(v) =
(v
0
, v)
X
2 for all v ∈ X
2
(Riesz) and v
h
∈ X
2
h
such that f(v) = (v
h
, v)
X
2 for all
v ∈ X
2
h
(Riesz) we have (v
0
− v
h
, P
2
h
v) = 0 for all v ∈ X
2
; so Corollary 3.3.4
gives that P
2
h
(v
0
−v
h
) = 0. It follows that P
2
h
v
0
= P
2
h
v
h
= v
h
, so v
h
= P
2
h
v
0
.
From Lax Milgram we know that u
0
= R
−1
v
0
and u
h
= S
−1
v
h
, so that
u
h
= S
−1
P
2
h
Ru
0
.
ii) For every x
h
∈ X
1
h
we have Sx
1
h
= P
2
h
Rx
1
h
because, for all v
h
∈ X
2
h
,
(P
2
h
Rx
1
h
, v
h
) = (Rx
1
h
, P
2
h
v
h
) = (Rx
1
h
, v
h
)
= F(x
1
h
, v
h
) = (Sx
1
h
, v
h
).
Version: September 13, 2004 89
We want to know how well the discrete solution approximates the solution of
the continuous problem, so we are interested in u
h
−u
0
. First we look at
u
h
−P
1
h
u
0
 = S
−1
P
2
h
Ru
0
−P
1
h
u
0

≤ S
−1
 P
2
h
Ru
0
−SP
1
h
u
0

= S
−1
 P
2
h
Ru
0
−P
2
h
RP
1
h
u
0

≤ S
−1
 P
2
h
 R u
0
−P
1
h
u
0

(∗)
= S
−1
 R u
0
−P
1
h
u
0
.
In (∗) we used that P
2
h
 = 1 by Corollary 3.3.4. Hence
u
h
−u
0
 ≤ u
h
−P
1
h
u
0
 +P
1
h
u
0
−u
0

≤ ¦1 +S
−1
R¦ u
0
−P
1
h
u
0
.
From this easily follows
u
h
−u
0

X
1 ≤ ¦1 +F/γ
h
¦ inf
x
h
∈X
1
h
u
0
−x
h

X
1.
Corollary 3.5.2
Given a problem that satisﬁes the conditions of the Generalised Lax Milgram
Theorem, then for a conforming ﬁnite element method a suﬃcient condition for
convergence is the existence of a family of subspaces (X
1
h
) of X
1
such that, for
each u ∈ X
1
,
lim
h→0
inf
x
h
∈X
1
h
u −x
h
 = 0.
So the problem we were facing is reduced to an approximation problem: evaluate
the distance d(u, X
1
h
) = inf
x
h
∈X
1
h
u − x
h
 between a function u ∈ X
1
and a
subspace X
1
h
⊂ X
1
. Thus, if we can prove that d(u, X
1
h
) = O(h
k
), for some
k > 0, we have as an immediate consequence that
u −u
h
 = O(h
k
)
We then will call k the order of convergence.
The norm in which the error is measured is more or less induced by the
problem: the conditions on coercivity and continuity are stated in the X
1
norm.
For an elliptic problem of order 2m this will usually be the H
m
(Ω)norm. This
will be discussed later.
Measuring the error in other norms (and especially the L
2
norm) requires a
further analysis of the problem.
90 P.W. Hemker
3.5.2 More general error estimate
In this section we study an error estimate for more general problems. We can
approximate both the bilinear form B and the linear form f by discrete versions
B
h
and f
h
. These discrete operators will work on some spaces V
h
and S
h
that
are now not necessarily contained in V and S. This means we have to expect a
supplementary consistency error in our estimates. The estimate thus obtained
will also be valid for nonconforming ﬁnite element methods.
We consider a solution u
∗
∈ S ⊂ S
∗
(S
∗
a normed linear space) of the
problem: ﬁnd u ∈ S such that
B(u, v) = f(v), ∀v ∈ V. (3.57)
Let B
h
: S
∗
V
∗
→R and f
h
: V
∗
→R. We also consider the discrete problem
associated with (3.57): ﬁnd u
h
∈ S
h
such that
B
h
(u
h
, v
h
) = f
h
(v
h
), ∀v
h
∈ V
h
, (3.58)
in which S
h
⊂ S
∗
and V
h
⊂ V
∗
, but not necessarily S
h
⊂ S, V
h
⊂ V .
Note that we have assumed no relation between B and B
h
or f and f
h
.
What can we say about u
h
−u
∗

S
∗?
Theorem 3.5.3 Assume there exists a u
h
∈ S
h
that satisﬁes equation (3.58).
Let B
h
: S
∗
V
∗
→R be bounded with B = M
h
, so that
∃M
h
> 0 ∀s ∈ S
∗
, v ∈ V
∗
[B
h
(s, v)[ ≤ M
h
s
S
∗v
V
∗.
Let B
h
: S
h
V
h
→R be coercive:
∃α
h
> 0 ∀s
h
∈ S
h
∃v
h
∈ V
h
α
h
s
h

S
∗v
h

V
∗ ≤ [B
h
(s
h
, v
h
)[.
Then the following estimate is valid
u
h
−u
∗

S
∗ ≤
¸
1 +
M
h
α
h
inf
s
h
∈S
h
u
∗
−s
h

S
∗
+
1
α
h
sup
v
h
∈V
h
[B
h
(u
∗
, v
h
) −f
h
(v
h
)[
v
h

V
∗
. (3.59)
Proof: In the proof norms are either .
S
∗ or .
V
∗.
Because B
h
is coercive we know that for each s
h
∈ S
h
there exists a w
h
∈ V
h
such that
α
h
u
h
−s
h
 w
h
 ≤ [B
h
(u
h
−s
h
, w
h
)[
= [B
h
(u
∗
−s
h
, w
h
) −B
h
(u
∗
, w
h
) + B
h
(u
h
, w
h
)[
≤ [B
h
(u
∗
−s
h
, w
h
)[ +[B
h
(u
h
, w
h
) −B
h
(u
∗
, w
h
)[
≤ M
h
u
∗
−s
h
 w
h
 +[B
h
(u
∗
, w
h
) −f
h
(w
h
)[,
Version: September 13, 2004 91
because u
h
is a solution to (3.58). It follows that for arbitrary s
h
∈ S
h
u
h
−s
h
 ≤
M
h
α
h
u
∗
−s
h
 +
1
α
h
[B
h
(u
∗
, w
h
) −f
h
(w
h
)[
w
h

.
Now choose an arbitrary s
h
∈ S
h
then
u
h
−u
∗
 ≤ u
h
−s
h
 +s
h
−u
∗

≤
¸
1 +
M
h
α
h
u
∗
−s
h
 +
1
α
h
[B
h
(u
∗
, w
h
) −f
h
(w
h
)[
w
h

.
From this the theorem follows.
We see that the error is determined by
1. the conditioning of the discrete problem: M
h
/α
h
;
2. the interpolation error: inf
s
h
u
∗
−s
h

S
∗;
3. the consistency error: [B
h
(u
∗
, v
h
) −f
h
(v
h
)[.
Remarks:
• In the proof of the theorem we have not used the fact that u
∗
is a solution
to (3.57). However we can expect the consistency error to be rather big if
we choose a u
∗
that is not a solution.
• The conditioning of the problem can also be formulated as: there exist
α
h
> 0 and M
h
> 0 so that for all s
h
∈ S
h
α
h
s
h

S
h
≤ sup
v
h
∈V
h
[B
h
(s
h
, v
h
)[
v
h

V
h
≤ M
h
s
h

S
h
.
• If the discrete spaces are embedded in continuous spaces , i.e. if S
h
⊂ S
and V
h
⊂ V , and if B
h
is the restriction of a B : S V → R and f
h
the
restriction of a f : V →R to these subspaces, i.e. if
B
h
(u
h
, v
h
) = B(u
h
, v
h
) and f
h
(v
h
) = f(v
h
), ∀u
h
∈ S
h
, ∀v
h
∈ V
h
,
then
B
h
(u
∗
, w
h
) −f
h
(w
h
) = B(u
∗
, w
h
) −f(w
h
) = 0, ∀w
h
∈ V
h
.
It follows that in this case the estimate is reduced to the result from
Theorem 3.5.1. A diﬀerence is the fact that for the present estimate we
have assumed the existence of the solution u
∗
.
It turns out from theorem 3.5.3 that, apart from studying interpolation theory
for Banach spaces, we also have to study the approximation of bilinear and linear
operators. In this we will restrict ourselves to errors caused by quadrature. See
Section 3.5.6.
92 P.W. Hemker
3.5.3 The formalisation of the ﬁnite element method
We do not want to describe the theory of interpolation by piecewise polynomials
in all detail (for instance: we will prove few theorems) but we will show some of
the main results in the next section. To gain some insight in the derivation of
these results, we show in this section a formal way to introduce ﬁnite elements.
Let Ω ⊂ R
d
be an open bounded domain with a Lipschitz continuous bound
ary ∂Ω. A partition ∆
h
of Ω is deﬁned as
∆
h
= ¦ Ω
e
[
¸
e
Ω
e
= Ω, Ω
e
i
∩ Ω
e
j
= ∅ if i = j,
Ω
e
= ∅, ∂Ω
e
Lipschitz continuous¦.
Let u ∈ (
m
(Ω) with m ≥ 0. For a local ﬁnite element approximation of u on
Ω
e
∈ ∆
h
we choose a set of functions P
e
= ¦φ
e
j
¦
j=1,···,N
e
and we approximate u
on Ω
e
by
u
h
(x) =
N
e
¸
j=1
a
e
j
φ
e
j
(x) for x ∈ Ω
e
.
As a rule we choose ¦φ
e
j
¦ such that
1. there exists a set of nodal points of Ω
e
: ¦b
e
j
[ b
e
j
∈ Ω
e
, i = 1, , N
e
¦ and
2. there exists a set of derivatives D
α
i
, where α
i
are multiintegers, such that
D
α
i
φ
e
j
(b
e
i
) = δ
ij
. (3.60)
Usually we choose ¦φ
e
j
¦ such that it contains all polynomials of degree k
P
k
(Ω
e
) ⊂ Span(φ
e
j
).
We formalise the above construction.
Deﬁnition 3.5.4 Let P be a set of realvalued functions deﬁned over a domain
Ω. Let L be a ﬁnite set of linearly independent linear functionals l
i
, i = 1, , N
over P. Then L is Punisolvent if for every set of scalars α
i
, i = 1, , N, there
is a unique p ∈ P such that
l
i
(p) = α
i
.
Corollary 3.5.5
Let L be a Punisolvent set then there exists a unique set of piecewise polyno
mials ¦p
j
, j = 1, , N, p
j
∈ P¦ such that,
l
i
(p
j
) = δ
ij
. (3.61)
Example 3.5.6
Take Ω = [0, 1] and make a partition 0 = x
1
< x
2
< < x
N+1
= 1. Let
Ω
i
= [x
i
, x
i+1
], P
Ω
i
= P
1
(Ω
i
) and L
Ω
i
= ¦l
i
, l
i+1
¦, with l
i
(f) = f(x
i
) and
l
i+1
(f) = f(x
i+1
). Then ¦L
Ω
i
¦ is a ¦P
Ω
i
¦unisolvent set: take any two numbers
α and β then there is one and only one polynomial p ∈ P
Ω
i
such that
p(x
i
) = α and p(x
i+1
) = β.
Version: September 13, 2004 93
Deﬁnition 3.5.7 A ﬁnite element is a triple ¦Ω
e
, P
e
, L
e
¦ in which
P
e
= ¦φ
e
j
∈ (
∞
(Ω
e
), j = 1, , N
e
¦
are the basis functions of the ﬁnite element and the P
e
unisolvent set
L
e
= ¦l
e
i
∈
(
m
(Ω
e
)
, i = 1, , N
e
¦
gives the degrees of freedoms of the ﬁnite element.
Deﬁnition 3.5.8 A global ﬁnite element approximation of a function u on Ω is
deﬁned by
u
h
(x) =
N
¸
j=1
a
j
φ
j
(x)
for x ∈ Ω, where we choose the set P = ¦φ
j
¦ such that
1. there exists a partition ∆
h
= ¦Ω
e
¦
e
of Ω;
2. for every Ω
e
∈ ∆
h
there is a ﬁnite element ¦Ω
e
, P
e
, L
e
¦;
3. by L = ∪
e
L
e
we denote the set of degrees of freedom of the global ap
proximation. The number of (diﬀerent) elements in L is N =dim(L), so
N ≤
¸
e
N
e
, and N is the total number of degrees of freedom;
4. by P we denote the set of basis functions of the global approximation; P
is the set of N functions given by
P = ¦φ
j
[ ∃e ÷ , φ
j
[
Ω
e
∈ P
e
; l
k
(φ
m
) = δ
km
¦ .
The Punisolvent set L of degrees of freedom implies a natural Pinterpolation
of a (
m
(Ω) function
Π : (
m
(Ω) → Span(P). (3.62)
This Π is deﬁned as
v → Πv =
¸
j
l
j
(v)φ
j
. (3.63)
Π is a projection because l
i
(Πv) = l
i
(v) for all l
i
∈ L and any v ∈ (
m
(Ω).
Usually we take the same type of L
e
and P
e
on every Ω
e
in ∆. So we take
a set of ﬁnite elements ¦Ω
e
, P
e
, L
e
¦, that are all is equivalent to a reference
or master element (
ˆ
Ω,
ˆ
P,
ˆ
L). Then all ﬁnite elements are constructed from
(
ˆ
Ω,
ˆ
P,
ˆ
L) by an aﬃne or isoparametric transformation. Error estimates for the
master element are then easily transferred to every element.
94 P.W. Hemker
3.5.4 Interpolation theory and applications
Interpolation theory in Sobolev spaces
In Section 3.5.3 we saw the construction of an interpolation operator Π :
(
m
(Ω) → Span(P). It is clear that a function on an area Ω is approximated
better if more degrees of freedom are available. This can be done either by
making the elements Ω
e
smaller or by taking a higher order of approximation.
In this section we choose for the ﬁrst option and study the quality of the
approximation if we take smaller and smaller elements Ω
e
. In fact, we consider
a sequence ¦∆
h
¦
h→0
of partitions, in which h denotes the largest diameter of
the elements Ω
e
∈ ∆
h
. Further we assume that the reﬁnement of the whole area
will take place in a regular way so that also the smallest of the inscribed circles
of the elements in the partition will be O(h).
On each partition ∆
h
we assume a global ﬁnite element approximation and
we denote by S
h
= Span(P) the space of approximating functions. In this
way we also introduce a sequence of projections ¦Π
h
¦
h→0
with Π
h
: (
m
(Ω) →
S
h
. With an appropriate choice of the functionals in L (imposed by Sobolev’s
Lemma), the operator Π
h
: W
l,p
(Ω) → S
h
can be deﬁned
Π
h
u =
¸
j
l
j
(u)φ
j
.
We are interested in the behaviour of the error of approximation u −Π
h
u for
h → 0. A result is given in theorem 3.5.18 below. The proof of this theorem
and a number of preceding lemmas will not be given explicitly. They can be
found in [4], or its more recent version [18].
The plan of our treatment is as follows. First we deﬁne on each element
in the partition a projection operator which shall be used for the approxima
tion. Assuming that all elements are similar to one another (and thus there
will be a unique reference element) we can ﬁnd estimates for these projection
operators on each element. From this we can ﬁnd error estimates for the global
approximation. We apply this to analyse a secondorder problem in detail.
Deﬁnition 3.5.9
Let Ω
e
be a bounded polygon then deﬁne
h
e
= diam(Ω
e
),
ρ
e
= sup¦diam(S); S a ball contained in Ω
e
¦,
meas(Ω
e
) =
Ω
e
dx.
Deﬁnition 3.5.10 Two open subsets Ω and
ˆ
Ω of R
n
are aﬃneequivalent if
there is an invertible aﬃne mapping
F : ˆ x ∈ R
n
→ F(ˆ x) = Bˆ x + b ∈ R
n
such that Ω = F(
ˆ
Ω).
Version: September 13, 2004 95
If we denote the outer and the inner diameter by respectively h and ρ, then we
ﬁnd
B ≤ h
Ω
e
/ρ
ˆ
Ω
and B
−1
 ≤ h
ˆ
Ω
/ρ
Ω
e
.
Usually we have the following correspondences between points and functions
ˆ x ∈
ˆ
Ω → x = F(ˆ x) ∈ Ω, (3.64)
(ˆ v :
ˆ
Ω →R) → (v = ˆ v ◦ F
−1
: Ω →R). (3.65)
So we have
ˆ v(ˆ x) = v(x).
Analogously we deﬁne the aﬃneequivalence of two ﬁnite elements and all ﬁnite
elements in a partitioning are aﬃne equivalent to a master element (
ˆ
Ω,
ˆ
P,
ˆ
L) if
for each Ω
e
∈ ∆
h
there exists an aﬃne mapping F
e
such that for this element
• x = F
e
(ˆ x),
• ψ
e
j
(x) =
ˆ
ψ
j
(ˆ x),
• l
e
i
(ψ
e
j
(x)) =
ˆ
l
i
(
ˆ
ψ
j
(ˆ x)).
Before we give the basic error theorems we recall the deﬁnitions of the norms
and seminorms with which we work.
[v[
m,p,Ω
e
:=
¸
¸
α=m
Ω
e
[D
α
v[
p
dx
¸
1
p
,
v
m,p,Ω
e
:=
¸
¸
l≤m
[v[
p
l,p,Ω
e
¸
1
p
=
¸
¸
α≤m
Ω
e
[D
α
v[
p
dx
¸
1
p
,
[v[
m,∞,Ω
e
:= max
α=m
¦ess.sup
x
[D
α
v(x)[¦,
v
m,∞,Ω
e
:= max
α≤m
¦ess.sup
x
[D
α
v(x)[¦.
Lemma 3.5.11 Assume the integers k ≥ 0 and m ≥ 0 and the numbers p, q ∈
[1, ∞] are such that the Sobolev spaces W
k+1,p
(
ˆ
Ω) and W
m,q
(
ˆ
Ω) satisfy the
inclusion
W
k+1,p
(
ˆ
Ω) → W
m,q
(
ˆ
Ω).
Further let
ˆ
Π ∈ L(W
k+1,p
(
ˆ
Ω); W
m,q
(
ˆ
Ω)) be a polynomial preserving mapping:
∀ˆ p ∈ P
k
(
ˆ
Ω),
ˆ
Πˆ p = ˆ p.
For every open set Ω
e
that is aﬃneequivalent to
ˆ
Ω deﬁne the mapping Π
e
by
(Π
e
v)ˆ=
ˆ
Πˆ v,
96 P.W. Hemker
with ˆ v ∈ W
k+1,p
(
ˆ
Ω) and v ∈ W
k+1,p
(Ω
e
) corresponding as in (3.65). Then
there exists a constant C = C(
ˆ
Π,
ˆ
Ω) such that for all aﬃneequivalent sets Ω
e
and all v ∈ W
k+1,p
(Ω
e
)
[v −Π
e
v[
m,q,Ω
e
≤ C(meas(Ω
e
))
1/q−1/p
h
k+1
e
ρ
m
e
[v[
k+1,p,Ω
e
,
with h
e
, ρ
e
and meas(Ω
e
) as in deﬁnition 3.5.9.
We specialise this immediately to aﬃneequivalent ﬁnite element families. We
formulate a number of requirements on the spaces we use:
1. we have a set of degrees of freedom
ˆ
L which uses derivatives up to certain
order s. By Sobolev’s Lemma, this sets requirements on the Sobolev
spaces. This gives condition (3.66),
2. the projections will be measured in a weaker norm  
m,q,Ω
e
: condition
(3.67),
3. the approximating functions will at least have to include all polynomials
up to certain order: condition (3.68).
Lemma 3.5.12 Let (
ˆ
Ω,
ˆ
P,
ˆ
L) be a master element for which s is the largest
derivative in
ˆ
L. If for some k ≥ 0, m ≥ 0 and p, q ∈ [1, ∞]
W
k+1,p
(
ˆ
Ω) → (
s
(
ˆ
Ω), (3.66)
W
k+1,p
(
ˆ
Ω) → W
m,q
(
ˆ
Ω), (3.67)
P
k
(
ˆ
Ω) ⊂
ˆ
P ⊂ W
m,q
(
ˆ
Ω), (3.68)
then there is a C = C(
ˆ
Ω,
ˆ
P,
ˆ
L) such that for aﬃneequivalent ﬁnite elements
(Ω
e
, P
e
, L
e
) and all v ∈ W
k+1,p
(Ω
e
)
[v −Π
e
v[
m,q,Ω
e
≤ C(meas(Ω
e
))
1/q−1/p
h
k+1
e
ρ
m
e
[v[
k+1,p,Ω
e
.
Obviously, the term
h
k+1
e
ρ
m
e
appearing in this lemma is undesirable, so we would
like to dispose of the ρ
e
. This can be done provided the elements do not become
”ﬂat” in the limit: for h
e
→ 0 we want that ρ
e
= O(h
e
).
Deﬁnition 3.5.13 A sequence of partitionings ¦∆
h
¦ of ﬁnite elements ¦(Ω
e
, P
e
, L
e
)¦
h
10
is regular if the following two conditions are satisﬁed:
1. There exists a constant σ such that
∀e,
h
e
ρ
e
≤ σ.
10
Note that we use e as a kind of parameter of the family.
Version: September 13, 2004 97
2. The diameter h = max
Ω
e
∈∆
h
h
e
tends to zero.
Remark:
The regularity of the elements can also be stated in a more geometrical sense:
let θ
Ω
denote the smallest angle in an element Ω, then there must be a θ
0
> 0
such that
∀Ω ∈ T
h
, θ
Ω
≥ θ
0
> 0.
For regular families we can convert the error estimate 3.5.12 in estimates for
the associated norms.
Lemma 3.5.14 Let there be given a regular aﬃne family of ﬁnite elements
(Ω
e
, P
e
, L
e
) whose reference element (
ˆ
Ω,
ˆ
P,
ˆ
L) satisﬁes conditions (3.66), (3.67)
and (3.68). Then there exists a constant C = C(
ˆ
Ω,
ˆ
P,
ˆ
L) such that for all
members Ω of the family and all v ∈ W
k+1,p
(Ω),
v −Π
e
v
m,q,Ω
e
≤ C(meas(Ω
e
))
1/q−1/p
h
k+1−m
e
[v[
k+1,p,Ω
e
.
Thus, we can give an error estimate on each element in a partition ∆
h
of our
domain Ω
e
. Assuming the family is aﬃneequivalent we would like to give
an error estimate for the global ﬁnite element approximation. The assumptions
above have to be applied in an even more strict sense: on each level of reﬁnement.
Assumptions:
(H1) We consider a regular family of ﬁnite elements T
h
. That is, there exists a
constant σ > 0 such that
∀ Ω
e
∈ ∪
h
T
h
,
h
e
ρ
e
≤ σ.
and
h = max
Ω
e
∈F
h
h
e
→ 0 .
(H2) All the ﬁnite elements (Ω
e
, P
e
, L
e
), Ω
e
∈ ∪
h
T
h
are aﬃneequivalent to
a single reference element (
ˆ
Ω,
ˆ
P,
ˆ
L).
(H3) All the elements are of class (
0
.
To summarise the result we deﬁne a norm and a seminorm for piecewise
smooth functions on a given partitioning ∆
h
.
Deﬁnition 3.5.15 For functions that are (together with their derivatives up to
order m) integrable on the elements Ω
e
∈ ∆
h
, we introduce the following norm
and seminorm if p < ∞
v
m,p,∆
h
:=
¸
Ω
e
∈∆
h
v
p
m,p,Ω
e
¸1
p
,
98 P.W. Hemker
[v[
m,p,∆
h
:=
¸
Ω
e
∈∆
h
[v[
p
m,p,Ω
e
¸1
p
.
In case p = 2 we drop the p in the subscripts. We denote the space of functions
with ﬁnite norm .
m,∆
h
by H
m
∆
h
.
If p = ∞ we deﬁne
v
m,∞,∆
h
:= max
Ω
e
∈∆
h
v
m,∞,Ω
e
,
[v[
m,∞,∆
h
:= max
Ω
e
∈∆
h
[v[
m,∞,Ω
e
.
Example 3.5.16
The norm  
m,p,∆
h
on the partitioning ∆
h
of Ω is not the same as the norm
 
m,p,Ω
on Ω itself. This is seen as follows. Take Ω = [a, b] and consider the
piecewise linear function in ﬁgure 3.11. This function is an element of H
1
(Ω),
Figure 3.11: Piecewise H
2
function.
but it is not in H
2
(Ω). If we take however a partition ∆
h
such that the elements
Ω
e
correspond with the smooth parts of the function, it is piecewise H
2
, so in
H
2
∆
h
(Ω).
A number of approximation results is summarised in the following lemma.
Lemma 3.5.17 Let T
h
be a regular family of ﬁnite elements satisfying (H1),
(H2) and (H3) that are all aﬃne equivalent with the master element (
ˆ
Ω,
ˆ
L,
ˆ
P).
Let the requirements (3.66), (3.67) and (3.68) be satisﬁed for every element.
Consider the global interpolation operator
Π
h
: W
k+1,p
(Ω) → S
h
⊂ W
m,p
(Ω)
deﬁned by
Π
h
[
Ω
e
= Π
e
then there is a constant C = C(
ˆ
Ω,
ˆ
L,
ˆ
P) > 0, such that,
[v −Π
e
v[
m,p,Ω
e
≤ C h
k+1−m
e
[v[
k+1,p,Ω
e
,
and
[v −Π
h
v[
m,p,∆
h
≤ C h
k+1−m
[v[
k+1,p,∆
h
.
Version: September 13, 2004 99
From this easily follows
v −Π
e
v
m,p,Ω
e
≤ C h
k+1−m
e
[v[
k+1,p,Ω
e
, (3.69)
and
v −Π
h
v
m,p,Ω
≤ C h
k+1−m
[v[
k+1,p,Ω
. (3.70)
Remark:
If the functions v and Π
h
v are not suﬃciently smooth estimate (3.70) is not
valid, but we still have
v −Π
h
v
m,p,∆
h
≤ C h
k+1−m
[v[
k+1,p,∆
h
.
As a direct consequence of this lemma we can state the basic interpolation
theorem that can be applied to elliptic problems. We thus take p = q = 2.
Theorem 3.5.18 Under conditions as in lemma 3.5.17, for a regular family of
ﬁnite element approximations, the following error estimate is valid
u −Π
h
u
m,Ω
≤ C h
k+1−m
[u[
k+1,∆
h
.
Here C is independent of h, 0 ≤ m ≤ k + 1, u ∈ H
k+1
(Ω) and Π
h
is the
interpolation operator that leaves invariant all polynomials of degree k or less.
Deﬁnition 3.5.19 The spaces S
h
, based on a regular family T
h
of ﬁnite el
ements, such that the requirements of lemma 3.5.12 are satisﬁed for p =
2, i.e.they are kth degree piecewise polynomials in H
m
(Ω), are denoted by
S
k,m
h
(Ω), 0 ≤ m ≤ k + 1.
The spaces S
k,m
h
(Ω) contain all kth order piecewise polynomials and are subsets
of H
m
(Ω). By the preceding theorem we know that for a sequence ¦S
k,m
h
(Ω)¦
h→0
,
h ≤ h
0
, u ∈ H
r
(Ω), r ≥ 0 and 0 ≤ s ≤ min(r, m) there exists a u
h
∈ S
k,m
h
(Ω)
and a constant C ≥ 0, independent of h, such that
u −u
h

s
≤ C h
σ
u
r
,
where σ = min(k + 1 −s, r −s).
Corollary 3.5.20
Under the same conditions as in theorem 3.5.18, let u be the solution to a
second order elliptic problem and let u
h
be the associated discrete solution.
Then
u −u
h

1,Ω
= O(h
k
).
Suﬃcient for convergence for a second order problem is thus k = 1, or approxi
mation by piecewise linear functions.
Proof: Because we consider a second order problem we have m = 1 and the
theorem follows by combining theorem 3.5.1 and theorem 3.5.18.
100 P.W. Hemker
Inverse inequalities
We now treat a property of piecewise polynomial spaces that has no immediate
use at this point. However it is an easy consequence of the foregoing and we
will use it in section 3.5.6.
Deﬁnition 3.5.21 (H4) A family T
h
of triangulations satisﬁes an inverse as
sumption if there exists a ν such that
∀ T ∈ ∪
h
T
h
,
h
h
T
≤ ν.
In fact an inverse assumption implies that the elements can not diﬀer too much
in size, which in turn implies that functions in the approximation space have
derivatives that will not ”blow up”, unless the function itself ”blows up”.
It is clear that this assumption is by no means restrictive in practice: the
’usual’ ﬁnite element spaces satisfy an inverse assumption.
For such families of spaces we establish the following equivalence between
seminorms. Note that σ appears in the regularity assumption (H1).
Theorem 3.5.22 Let there be given a regular family of triangulations T
h
in
R
d
, satisfying an inverse assumption. Let 0 ≤ m ≤ k, p, q ∈ [1, ∞], and
ˆ
P ⊂ W
k,p
(
ˆ
Ω) ∩ W
m,q
(
ˆ
Ω),
then ∃C = C(ν, σ, k, m, p, q) such that for all v
h
∈ V
h
[v
h
[
k,p,∆
h
≤
C
meas(Ω
h
)
max{0,1/q−1/p}
h
k−m
[v
h
[
m,q,∆
h
, (3.71)
with the usual adaptation if p or q is ∞.
Proof: See [4, p. 141].
Example 3.5.23
Assume V
h
∈ W
l,r
for some (l, r), then the above inequalities can be given for
the seminorms [.[
l,r,Ω
. For instance suppose assumption (H3) is valid and that
ˆ
P ⊂ H
1
(Ω), then
[v
h
[
0,∞,Ω
≤
C
h
n/2
[v
h
[
0,Ω
,
and
[v
h
[
1,Ω
≤
C
h
[v
h
[
0,Ω
. (3.72)
If
ˆ
P ⊂ W
1,∞
(Ω), then
[v
h
[
1,∞,Ω
≤
C
h
[v
h
[
0,∞,Ω
.
Version: September 13, 2004 101
Example 3.5.24
We can also easily give inequalities for the norms: for instance, from (3.72) we
ﬁnd
v
h

1,Ω
≤
C
h
[v
h
[
0,Ω
. (3.73)
This can be applied for the following case. Take Ω = [a, b]. Consider S
1,1
h
(Ω),
the space of piecewise linear functions on a regular partition ∆
h
, then (3.73) is
valid for each v
h
∈ S
1,1
h
(Ω).
Motivated by this we come to the following deﬁnition, suited for the function
spaces S
k,m
h
(Ω).
Deﬁnition 3.5.25 A family of piecewise polynomial spaces ¦S
k,m
h
(Ω)¦
h→0
has
the inverse property if
∃C : ∀s ≤ m, h
m
u
h

m
≤ C h
s
u
h

s
, ∀u
h
∈ S
k,m
h
(Ω).
Error estimate in the L
2
norm
In Section 3.5.4 we saw that under conditions we can assure that u−u
h

1,Ω
=
O(h
k
). This means that at least u −u
h

0,Ω
= O(h
k
). In this section we show
that, under mild additional assumptions, u −u
h

0,Ω
= O(h
k+1
).
Theorem 3.5.26 (AubinNitsche Lemma)
Let Lu = f be a problem of order 2m and let V
h
⊂ V = H
m
(Ω). The continuous
problem is to ﬁnd u such that B(u, v) = f(v) for all v ∈ V . The discrete problem
is: ﬁnd u
h
such that B(u
h
, v
h
) = f(v
h
) for all v
h
∈ V
h
. Assume that for all
g ∈ H
−s
there is a solution to L
T
z = g (the adjoint problem), such that
z
2m−s
≤ C g
−s
, s ∈ [0, m] ,
(the regularity of the adjoint problem). Then
u −u
h

s,Ω
≤ C h
m−s
u −u
h

m,Ω
. (3.74)
Proof: Deﬁne the error e = u
h
−u. For any 0 ≤ s ≤ m:
e
s
= sup
g∈H
−s
[g(e)[
g
−s
,
We ﬁrst evaluate [g(e)[. Let z be the solution to the adjoint problem B(w, z) =
g(w), for all w ∈ H
s
(Ω). In particular B(e, z) = g(e), or, as B(e, v
h
) = 0 for all
v
h
∈ V
h
,
B(e, z −v
h
) = g(e).
So
[g(e)[ ≤ B e
m
inf
v
h
∈V
h
z −v
h

m
.
102 P.W. Hemker
Now, provided r ≤ k + 1 −m,
[g(e)[ ≤ B e
m
C h
r
z
m+r
,
due to theorem 3.5.18. If we put r = m−s and provided 2m−s ≤ k + 1,
[g(e)[ ≤ B e
m
C h
m−s
z
2m−s
≤ B e
m
C h
m−s
g
−s
,
because of the regularity of the adjoint problem. So, if we compute the error in
the snorm, we get
e
s
= sup
g∈H
−s
[g(e)[
g
−s
≤ C h
m−s
e
m
.
Corollary 3.5.27
For the error in L
2
norm for a second order problem we get
u −u
h

0,Ω
≤ C h
k+1
[u[
k+1,∆
h
,
provided k ≥ 1.
Remarks:
• The regularity of the adjoint problem is not very restrictive. For example
in case of a symmetric problem it is a trivial matter.
3.5.5 Pointwise error estimate and superconvergence
In this section we restrict ourselves to a second order linear ordinary diﬀerential
equation on Ω = [a, b] ⊂ R
Ly ≡ −
d
dx
(a
2
(x)
d
dx
y) + a
1
(x)
d
dx
y + a
0
(x)y = s(x), (3.75)
with a
2
(x) = 0 and homogeneous Dirichlet boundary conditions
y(a) = y(b) = 0.
We will use the Greens function for equation (3.75): i.e. the function G(x; ξ)
such that
11
y(x) = −
b
a
G(x; ξ) s(ξ) dξ, (3.76)
11
The function G is the resolvent kernel of the diﬀerential equation.
Version: September 13, 2004 103
for any function s(x). A priori, it is not clear that such a function G(x; ξ) can
exist, but we will construct it. Therefore, let ϕ
1
and ϕ
2
be two independent
solutions to the adjoint (transposed) equation
L
T
y ≡ −
d
dx
(a
2
(x)
d
dx
y) −
d
dx
(a
1
(x)y) + a
0
(x)y = 0,
with boundary conditions ϕ
1
(a) = 0, ϕ
1
(a) = 1, ϕ
2
(b) = 0 and ϕ
2
(b) = 1.
Now G(x; ξ) is constructed as
G(x; ξ) =
ϕ
1
(x)ϕ
2
(ξ)/z(x) if x < ξ
ϕ
1
(ξ)ϕ
2
(x)/z(x) if ξ < x,
(3.77)
where
12
z(x) := a
2
(x) (ϕ
1
(x)ϕ
2
(x) −ϕ
1
(x)ϕ
2
(x)) .
Remark:
It is easily seen that this function z(x) satisﬁes the equation a
2
z
+a
1
z = 0. So
z has a unique sign. If z ≡ 0 then ϕ
1
and ϕ
2
are linearly dependent. (In that
particular case the homogeneous problem has a nontrivial solution and G(x, ξ)
does not exist.)
By substitution of (3.76) in (3.75), it is not hard to verify that G(x; ξ) is the
Green’s function indeed. Other properties of this Greens function G(x; ξ) are:
1. G(x; ) ∈ H
1
0
(Ω) ∩ C
2
( (a, x) ∪ (x, b) ),
2. L
T
G(x; ) ≡ 0 on (a, x) ∪ (x, b),
3.
jump
ξ=x
∂
∂ξ
G(x; ξ) := lim
h→0
∂G(x; ξ + h)
∂ξ
− lim
h→0
∂G(x; ξ −h)
∂ξ
=
1
a
2
(x)
.
If x
i
is a nodal point in a partition ∆
h
of Ω, then G(x
i
, ) can be approxi
mated well in S
k,1
h
(Ω). Under this condition we see that
inf
v
h
∈S
k,1
h
G(x
i
; ) −v
h

m,∆
h
≤ C h
k+1−m
G(x
i
; )
k+1,∆
h
.
So, for m = 1
inf
v
h
∈S
k,1
h
(Ω)
G(x; ) −v
h

1
≤
C h
k
[G[
k+1,∆
h
if x = x
i
,
C h
0
[G[
1
if x = x
i
.
(3.78)
The last estimate is true because in this case G(x; ) is not piecewise H
k+1
on
∆
h
, but it still is continuous, so in H
1
.
12
The determinant ϕ
1
ϕ
2
−ϕ
1
ϕ
2
=
ϕ
1
ϕ
1
ϕ
2
ϕ
2
is called the Wronskian.
104 P.W. Hemker
Theorem 3.5.28 If equation (3.75) is discretised by a ﬁnite element method
with S
h
= V
h
= S
k,1
h
(Ω), then we have the pointwise error estimate
[y(x) −y
h
(x)[ ≤
C h
2k
if x = x
i
∈ ∆
h
,
C h
k
if x ∈ ∆
h
.
(3.79)
Proof: y ∈ H
1
0
(Ω) is the solution to B(y, v) = f(v), ∀v ∈ H
1
0
(Ω). Similarly
y
h
∈ S
k,1
h
(Ω) is the solution to B(y
h
, v
h
) = f(v
h
), ∀v
h
∈ S
k,0
h
(Ω), so for e :=
y −y
h
we have B(e, v
h
) = 0 ∀v
h
∈ S
k,1
h
(Ω). Now
y(x) = −
b
a
G(x; ξ) f(ξ) dξ = −
G(x; ξ)Ly(ξ)dξ = −B(y, G(x; )),
and analogously for y
h
, so
e(x) = −B(e, G(x; ))
= −B(e, G(x; ) −v
h
), ∀v
h
∈ S
k,1
h
(Ω).
This gives
[e(x)[ ≤ B e
1
inf
v
h
G(x; ) −v
h

1
.
Because B is bounded, e
1
= O(h
k
) and (3.78) the theorem follows.
The phenomenon that the order of accuracy of an approximation at particular
points is better than the estimate in a global norm, is called superconvergence.
Remark:
If this same problem is tackled with functions in S
3,2
(Ω), there will be no super
convergence. Then estimate (3.78) fails to be better at the points of ∆
h
.
3.5.6 The inﬂuence of the quadrature rule
In the previous sections we have given estimates for the error y −y
h
, that we
obtain when we solve the discrete equation instead of the continuous equation.
As we have seen in section 3.4.3 it may be so that we have to compute the
integrals in the discrete equations using a quadrature rule. So we not only have
the error y−y
h
 but also an additional error y
h
− ˜ y
h
 (which can be associated
with the consistency error of theorem 3.5.3). This could disturb all the error
estimates we have given so far. In order to take this into account we require that
the additional error should be at most of the same order as the discretisation
error. This will lead to some requirement for accuracy of the quadrature rule.
In this section we will formulate these requirements for two cases: ﬁrst we
show under which conditions the global error estimate is preserved (as in the
orem 3.5.18) and further we discuss the how we can preserve the pointwise
superconvergence of theorem 3.5.28.
Version: September 13, 2004 105
Deﬁnition 3.5.29 A quadrature rule of the form
Ω
f dx ≈
¸
i
w
i
f(x
i
)
is called to be of degree t if it integrates any polynomial f of degree t exactly.
Theorem 3.5.30 Given a repeated quadrature rule of degree t, applied to a
partition ∆
h
, then there exists, for f ∈ ((Ω) ∩ W
t+1,p
(Ω
e
), ∀Ω
e
, a piecewise
polynomial Π
h
f of degree t such that it interpolates f on the nodes, i.e.
f(x
i
) = Π
h
f(x
i
)
and such that
f −Π
h
f
m,p,∆
h
≤ C h
t+1−m
[f[
t+1,p,∆
h
.
Proof: Take a (t + 1)points interpolating polynomial of degree t on every
Ω
e
: each of this gives the estimate of lemma 3.5.17. Provided the set of nodes
contains the quadrature points the estimate is valid.
Corollary 3.5.31
When a quadrature rule of degree t is used the quadrature error is O(h
t+1
):
Ω
e
[f −Π
h
f[ dx ≤ C h
t+1
¸
α=t+1
Ω
e
[D
α
f(x)[ dx. (3.80)
Example 3.5.32
A t +1point NewtonCoates formula has degree t. An npoints Gauss quadra
ture has degree 2n −1; an npoint Lobatto formula has degree 2n −3.
In the following we will restrict ourselves to second order elliptic problems, and
we use a quadrature of degree t, we will assume that all the coeﬃcients in the
partial diﬀerential equation are (piecewise) C
t+1
(Ω)functions.
In order to study the inﬂuence of the quadrature on the accuracy of the FEM
we have to distinguish three diﬀerent problems:
1. the continuous problem
ﬁnd y ∈ S such that B(y, v) = f(v), for all v ∈ V ;
2. the discrete problem
ﬁnd y
h
∈ S
h
such that B(y
h
, v
h
) = f(v
h
) for all v
h
∈ V
h
;
3. the discrete problem with quadrature
ﬁnd ˜ y
h
∈ S
h
such that B
h
(˜ y
h
, v
h
) = f
h
(v
h
), for all v
h
∈ V
h
.
106 P.W. Hemker
Here f
h
is a linear and B
h
is a bilinear functional in which the integrals (see
Section 3.4.3) are replaced by a repeated quadrature rule of degree t on the
partition ∆
h
.
To compute the additional error we ﬁrst have to make an estimate of the
amount by which f
h
deviates from f, and B
h
deviates from B.
Lemma 3.5.33 When using a quadrature rule of degree t we have
[f(v
h
) −f
h
(v
h
)[ ≤ C h
t+1
v
h

k,∆
h
(3.81)
and
[B(˜ y
h
, v
h
) −B
h
(˜ y
h
, v
h
)[ ≤ C h
t+1
˜ y
h

k,∆
h
v
h

k,∆
h
. (3.82)
Proof: We only give the proof for f
h
and f, the other goes analogously. Using
the operator Π
h
, as introduced in Theorem 3.5.30, we have
E := [f(v
h
) −f
h
(v
h
)[ =
¸
e
[
Ω
e
((fv) −Π
h
(fv)) dx[.
Using (3.80) we have
E ≤ C h
t+1
¸
e
Ω
e
[D
t+1
(fv)[ dx,
so, by Leibniz’ rule and the fact that v
h
∈ V
h
,
E ≤ C h
t+1
k
¸
l=0
¸
e
Ω
e
[(D
t+1−l
f)(D
l
v
h
)[ dx ≤ C h
t+1
v
h

k,∆
h
.
The following theorem states the conditions under which the accuracy as given
in theorem 3.5.18 is preserved.
Theorem 3.5.34 Let S = V = H
1
(Ω), ∆
h
a regular partition of Ω and assume
S
h
and V
h
satisfy the inverse property. Further, let B : S
h
V
h
→R be bounded
and coercive and let f
h
and B
h
be the approximations of f and B obtained by
quadrature t ≥ 2k −2. Then
y − ˜ y
h

1,Ω
= O(h
k
).
Proof: Let y
h
be the solution of the discrete problem and ˜ y
h
the solution to
the discrete equation with quadrature. We seek a upper bound of y
h
− ˜ y
h
 by
considering
y
h
− ˜ y
h

1
≤
1
γ
h
sup
v
h
∈V
h
[B(y
h
− ˜ y
h
, v
h
)[
v
h

1
,
Version: September 13, 2004 107
in which γ
h
is the (discrete) coercivity constant of B.
So compute
[B(y
h
− ˜ y
h
, v
h
)[ ≤ [B(y
h
, v
h
) −B
h
(˜ y
h
, v
h
)[ +[B
h
(˜ y
h
, v
h
) −B(˜ y
h
, v
h
)[
i
= [f(v
h
) −f
h
(v
h
)[ +[B
h
(˜ y
h
, v
h
) −B(˜ y
h
, v
h
)[
ii
≤ C h
t+1
v
h

k,∆
h
+ C h
t+1
˜ y
h

k,∆
h
v
h

k,∆
h
iii
≤ C h
t+2−k
v
h

1
+
C h
t+2−k
v
h

1
¦y
h

k,∆
h
+y
h
− ˜ y
h

k,∆
h
¦
iiii
≤ C h
t+2−k
v
h

1
¦1 +y
h

k,∆
h
¦ +
C h
t+3−2k
v
h

1
˜ y
h
−y
h

1,∆
h
.
We used (i) that y
h
is a solution of the discrete problem and ˜ y
h
is the solution
to the discrete problem with quadrature, (ii) the estimates of Lemma 3.5.33
for the quadrature, (iii) the fact that V
h
has the inverse property and (iiii) that
S
h
has the inverse property.
Thus,
y
h
− ˜ y
h

1
¦1 −C h
t+3−2k
¦ ≤ C h
t+2−k
¦1 +y
h

k,∆
h
¦.
Provided h is small enough (C h
t+3−2k
< 1), we have
y
h
− ˜ y
h

1
≤ C h
t+2−k
¦1 +y
h

k,∆
h
¦. (3.83)
To preserve the accuracy t should satisfy both t +2 −k ≥ k and t +3 −2k > 0,
both of which are satisﬁed if t ≥ 2k −2.
For the superconvergence as treated in Section 3.5.5 we now answer the same
question: what should be the degree of the quadrature to preserve the pointwise
superconvergence accuracy?
Theorem 3.5.35 Let the two point boundary value problem (3.75) be discre
tised by a ﬁnite element method with S
h
= V
h
= S
k,1
h
(Ω), and let the quadrature
rule be of degree t, with t ≥ 2k −1, then
[y(x) − ˜ y
h
(x)[ =
O(h
2k
), x ∈ ∆
h
,
O(h
k
), x ∈ ∆
h
.
Proof: We introduce G
i
= G(x
i
, ), x
i
a nodal point, and consider
[y
h
(x
i
) − ˜ y
h
(x
i
)[ = [B(y
h
− ˜ y
h
, G
i
)[.
First
B(y
h
− ˜ y
h
, G
i
) = B(y
h
− ˜ y
h
, G
i
−v
h
) + B(y
h
− ˜ y
h
, v
h
)
108 P.W. Hemker
= B(y
h
− ˜ y
h
, G
i
−v
h
) + B(y
h
, v
h
) −B
h
(˜ y
h
, v
h
)
+B
h
(˜ y
h
, v
h
) −B(˜ y
h
, v
h
)
= B(y
h
− ˜ y
h
, G
i
−v
h
) + f(v
h
) −f
h
(v
h
)
+B
h
(˜ y
h
, v
h
) −B(˜ y
h
, v
h
),
so, using the continuity of B
[B(y
h
− ˜ y
h
, G
i
)[ ≤ M y
h
− ˜ y
h

1
G
i
−v
h

1
+[f(v
h
) −f
h
(v
h
)[
+[B
h
(˜ y
h
, v
h
) −B(˜ y
h
, v
h
)[
≤ M y
h
− ˜ y
h

1
G
i
−v
h

1
+ C h
t+1
v
h

k,∆
h
+C h
t+1
˜ y
h

k,∆
h
v
h

k,∆
h
(3.83)
≤ M C h
t+2−k
¦1 +y
h

k,∆
h
¦ G
i
−v
h

1
+C h
t+1
v
h

k,∆
h
¦1 +y
h

k,∆
h
¦
≤ M C h
t+2−k
¦1 +y
h

k,∆
h
¦ G
i
−v
h

1
+C h
t+1
G
i
−v
h

k,∆
h
¦1 +y
h

k,∆
h
¦
+C h
t+1
G
i

k,∆
h
¦1 +y
h

k,∆
h
¦
(3.78)
≤ C h
t+1
G
i

k,∆
h
+ C h
t+1
G
i

k,∆
h
≤ C h
t+1
G
i

k,∆
h
.
According to theorem 3.5.28 y(x) − y
h
(x) = O(h
2k
) in nodal points x = x
i
∈
∆
h
, so we conclude that by the condition t +1 ≥ 2k also y(x) − ˜ y
h
(x) = O(h
2k
)
in nodal points.
Example 3.5.36
A Gaussian quadrature on k points or a Lobatto quadrature on k +1 points are
suﬃciently accurate to preserve superconvergence.
Acknowledgement
In 1992 H. Eleveld prepared an earlier version of these notes on ﬁnite element
methods, based on lectures given at the University of Amsterdam. Final cor
rections 2004.
Bibliography
[1] R.A. Adams. Sobolev Spaces, volume vol 65 of Pure and applied mathemat
ics series. Academic Press, New York, 1975.
[2] D. Braess. Finite Elements. Cambridge University Press, 1997.
[3] S. C. Brenner and L. Ridgeway Scott. The Mathematical Theory of Finite
Element Methods. Springer Verlag, 1994.
[4] P. G. Ciarlet. The ﬁnite element method for elliptic problems, volume vol 4
of Studies in mathematics and its application. NorthHolland, Amsterdam,
1978.
[5] G.E. Forsythe and W.R. Wasow. FiniteDiﬀerence Methods for Partial
Diﬀerential Equations. Wiley, New York, 1960.
[6] G. H. Golub and Ch. F. van Loan. Matrix Computations. Johns Hopkins
University Press, Baltimore, London, 1989. 2nd edition.
[7] W. Hackbusch. Theorie und Numerik elliptischer Diﬀerentialgleichungen.
B. G. Teubner, Stuttgart, 1986.
[8] A. Harten, P. D. Lax, and B. Van Leer. On upstream diﬀerencing and
Godunovtype schemes for hyperbolic conservation laws. SIAM Review,
25:35–61, 1983.
[9] P.W. Hemker and B. Koren. Defect correction and nonlinear multigrid for
the steady Euler equations. Technical Report NMN8801, CWI, 1988.
[10] P.W. Hemker and B. Koren. Eﬃcient multidimensional upwinding for the
steady Euler equations. Technical Report NMR9107, CWI, 1991.
[11] C. Hirsch. Numerical Computation of Internal and External Flows, Volume
1: Fundamentals of Numerical Discretisation. J. Wiley, 1988.
[12] C. Hirsch. Numerical Computation of Internal and External Flows, Volume
2: Computational Methods ﬁr Inviscid and Viscous Flow. J. Wiley, 1990.
[13] R. Jeltsch. Properties of discrete hyperbolic conservation laws. Technical
Report 8854, Inst. f. Geometrie u. Praktische Mathematik, RWTH, Aken,
1988.
109
110 P.W. Hemker
[14] C. Johnson, A. Szepessy, and P. Hansbo. Defect correction and nonlinear
multigrid for the steady Euler equations. Technical Report 21, Chalmers
Univ. Techn., 1987.
[15] B. Koren. Multigrid and Defect Correction for the Steady NavierStokes
equations, Applications to aerodynamics. Number 74 in CWI Tract series.
CWI, Amsterdam, 1991.
[16] P. D. Lax. Hyperbolic Systems of Conservation Laws and the Mathematical
Theory of Shock Waves, volume 11 of Regional Conference Series in Applied
Mathematics. SIAM Publication, Philadelphia, 1973.
[17] R.J. LeVeque. Numerical Methods for Conservation Laws. Lectures in
Mathematics. Birkh¨ auser, Basel, 1990.
[18] J. L. Lions (eds.) P. G. Ciarlet. Handbook of Numerical Analysis, Vol II:
Finite Element Methods, part I.
[19] S. V. Patankar. Numerical Heat Transfer and Fluid Flow. Hemisphere
McGrawHill, 1980.
[20] D. W. Peaceman. Fundamentals of Numerical Reservoir Simulation. Else
vier Science Publ., 1977.
[21] R. D. Richtmyer and K. W. Morton. Diﬀerence Methods for Initial Value
Problems. Interscience Publ., 1967.
[22] W. Rudin. Functional Analysis. McGrawHill, New Delhi, etc., 1973.
[23] J. Smoller. Shock Waves and Reaction Diﬀusion Equations, volume 258 of
Grundlehren der mathematische Wissenschaften. Springer Verlag, 1983.
[24] S.P. Spekreijse. Multigrid Solution of the Steady Euler Equations, volume 46
of CWI Tracts. CWI, 1988.
[25] J. C. Strikwerda. Finite Diﬀerence Schemes and Partial Diﬀerential Equa
tions. Wadswordth and Brooks, Paciﬁc Grove, Calif., 1989.
[26] R. Struijs, H. Deconinck, and G. Coussement. Multidimensional upwind
schemes for the compressible ﬂow equations. Technical report, Von Karman
Institute, 1989.
[27] P. A. Raviart. V. Girault. Finite Elements Methods for Navier Stokes
Equations, volume vol. 5 of Springer series in Computational Mathematics.
SpringerVerlag, Berlin, et.c, 1986.
[28] B. Van Leer. On numerical dispersion by upwind diﬀerencing. Technical
Report 8551, Delft Univ. Techn., 1985.
[29] B. Van Leer. Flux vector splitting for the 1990’s. Technical report, Univ.
Michigan, 1990.
[30] K. Yosida. Functional Analysis. Springer Verlag, Berlin, etc., 1974.
Index
adjoint problem, 101
aﬃneequivalent, 94
ampliﬁcation factor, 34
asymmetric problem, 68
AubinNitsche Lemma, 101
backward Euler method, 33
Banach space, 50
barycentric coordinates, 84
bilinear functional, 54
bilinear operator, 53
block centered, 21
boundary conditions, 46
bounded operator, 53
box method, 28
BubnovGalerkin, 28
Burgers’ equation, 12, 13
Cauchy Riemann, 22
CauchySchwarz inequality, 50
cell centered, 21
cell vertex, 21
central approximation, 32
characteristic polynomial, 5
characteristics, 7
coercive, 56
collocation method, 27
compact support, 51
compatibility condition, 24
complete, 50
conforming method, 69
conjugate exponents, 54
conjugatelinear operator, 53
conservation law, 6
consistency error, 91
convection equation, 31
convective ﬂux, 2
convergence
strong, 56
weak, 56
CranckNicolson method, 33
curved boundary, 86
degree, 105
degrees of freedoms, 93
delta function, 59
dense, 50
diﬀusion equation, 3, 29, 46
diﬀusive ﬂux, 2
Dirac delta function, 59
Dirichlet boundary conditions, 46
Discrete LaxMilgram Theorem, 88
discrete solution, 87
distance, 50
distribution, 59
distributional derivative, 60
divergence, 2
downwind approximation, 32
dual space, 54
elementary load vector, 74
elementary stiﬀness matrix, 74
elliptic, 5, 6
elliptic equation, 46
energy norm, 69
entropy condition, 15, 16
equivalent diﬀerential equation, 37
Euler equations, 6
ﬁnite diﬀerence approximation, 19
ﬁnite element, 93
ﬁnite element approximation, 21
ﬁnite volume approximation, 21
ﬁnite volume method, 28
111
112 P.W. Hemker
forward Euler method, 33
function spaces, 51
functionals, 53
Galerkin method, 28
Gauss’ theorem, 2
generalised derivative, 60
generalised function, 59
global discretisation error, 36
global ﬁnite element approximation,
93
gradient, 2
Greens function, 102
Heaviside function, 60
Hermite interpolate, 81
hierarchical basis, 73
Hilbert space, 50
hyperbolic, 5, 7
IBVP, 79
initialboundaryvalue problem, 79
inner product, 49
inner product space, 49
interpolation error, 91
inverse property, 101
inviscid Burgers’ equation, 8
isoparametric elements, 86
Lagrange interpolation, 70
Laplace equation, 46
Laplace’s equation, 3
LaxMilgram Theorem, generalised,
57
linear functional, 53
linear operator, 53
linear space, 49
linearisation, 30
Lipschitz continuous, 51
load vector, 74
local discretisation error, 36
lumping, 78
mass matrix, 79
master element, 93
mesh, 19, 70
meshwidth, 70
metric, 50
mixed boundary conditions, 46
modiﬁed equation, 37
multiindex, 51
multiinteger, 51
Neumann boundary conditions, 46
nodal points, 92
nonunique solution, 14
nonconforming method, 90
norm, 49
normed linear space, 49
numerical diﬀusion, 37
Oleinik, 15, 16
order of the diﬀerential equation, 1
ordinary diﬀerential equation, 1
parabolic, 5
partial diﬀerential equation, 2
partitioning, 19
PDE, 1
PetrovGalerkin, 28
piecewise Lagrange interpolation, 70
piecewise polynomials, 21
Poincar´e’s Lemma, 64
Poisson equation, 3
positive
strictly, 56
positive deﬁnite, 46
potential equation, 46
principle part, 3, 5, 46
projection Π
h
, 94, 98, 105
quadrature rule, 75
quasilinear, 3
RankineHugoniot, 12
reference element, 93
regular family, 97
regularity, 101
scalar ﬁeld, 49
scalar product, 49
semidiscretisation, 31, 79
seminorm, 49
separable, 50
Version: September 13, 2004 113
sesquilinear functional, 54
sesquilinear operator, 53
Sobolev space, 61
Sobolev’s Embeddings Theorem, 62
spectral method, 21
splines, 83
staggered grid, 22
standard hat functions, 71
static condensation, 78
steady equations, 4
stiﬀness matrix, 74
Stokes equations, 24
strictly positive, 56
subcoercive, 56
superconvergence, 104
symmetric operator, 53
symmetric problem, 67
topology, 60
trace, 64
triangle inequality, 49
unisolvent set, 86
upwind approximation, 32
variational method, 26
vector space, 49
vertex centered, 21
wave equation, 3
weak solution, 11, 66
weak topology, 60
weakly convergent, 56
weighted residual method, 27
weighting function, 54
Wronskian, 103
ii
P.W. Hemker
Contents
1 Introduction to PDEs 1.1 Classiﬁcation . . . . . . . . . . . . . . . . . . . . . 1.1.1 Ordinary and partial diﬀerential equations . 1.1.2 Elliptic, parabolic or hyperbolic equations . 1.1.3 Conservation laws . . . . . . . . . . . . . . 1.2 Hyperbolic equations . . . . . . . . . . . . . . . . . 1.2.1 Characteristics . . . . . . . . . . . . . . . . 1.2.2 Discontinuous solutions . . . . . . . . . . . 1 1 1 4 6 7 8 10 17 17 19 19 21 21 21 22 24 24 26 27 27 28 28 29 29 30 31 31 31 32 35 42
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
2 Discretisation Principles 2.1 Discrete representation of the solution . . . . . . . . . . . . 2.1.1 Discretisation of the domain, solution and equation . 2.1.2 Finite diﬀerence approximation . . . . . . . . . . . . 2.1.3 Finite element approximation . . . . . . . . . . . . . 2.1.4 Finite volume approximation . . . . . . . . . . . . . 2.1.5 Spectral methods . . . . . . . . . . . . . . . . . . . . 2.1.6 Staggered grid . . . . . . . . . . . . . . . . . . . . . 2.2 Techniques for discretisation of PDEs . . . . . . . . . . . . 2.2.1 Finite diﬀerence methods . . . . . . . . . . . . . . . 2.2.2 Variational method . . . . . . . . . . . . . . . . . . . 2.2.3 Weighted residual methods . . . . . . . . . . . . . . 2.2.4 Collocation methods . . . . . . . . . . . . . . . . . . 2.2.5 Galerkin methods . . . . . . . . . . . . . . . . . . . 2.2.6 Box methods = Finite Volume methods . . . . . . . 2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 The diﬀusion equation . . . . . . . . . . . . . . . . . 2.3.2 The source term and solution of the system . . . . . 2.3.3 The convection equation . . . . . . . . . . . . . . . . 2.4 Techniques for timediscretisation . . . . . . . . . . . . . . . 2.4.1 Timespace discretisation or semidiscretisation . . . 2.4.2 FDM for a linear hyperbolic problem . . . . . . . . . 2.4.3 The equivalent diﬀerential equation . . . . . . . . . 2.5 Criteria for good discretisation methods . . . . . . . . . . . iii
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . 90 3. 86 3. . . .5 Error estimates for the ﬁnite element method . 45 3. . . . . . . . . . . .2 More general error estimate . . . .1 The discrete version of the Generalised LaxMilgram . . . . . . . . .3. . . .3 Distributions and Sobolev Spaces . . . . . . . . . .4. . .5. . . . . . . . 87 3.1 Onedimensional second order . . . . . . . . . . .4 The technique of the ﬁnite element method .5 Variational formulations for diﬀerential equations . . . .5 Pointwise error estimate and superconvergence . . . . . . . . . .2 Piecewise Lagrange Interpolation in one dimension . . . . . .2. .6 The inﬂuence of the quadrature rule . 102 3. .2 Twodimensional second order . . . . . . . . . Hemker 3 Finite Element Methods 45 3. . . . . . . 56 3. . . . . . . . . . . . . . . . . 49 3. . 64 e 3. . . . . . . . . . 83 3. . .1 Introduction to Functional Analysis . .5. . . . . . . 70 3. . . . . . . . . .3 Fourth order problems . . . . . . . . . . . .4. .5. . .6 Isoparametric elements . . . . . . . 46 3. . . . 104 References Index 109 111 . . . . . . . . .3. . . 87 3. . . . . . . . . . . . . . .W. . . . . . . . .2 Examples of boundary value problems . . . . . . . . . . .2. . . 58 3. . . . . . . . . . . . 81 3. . . . 94 3. . . .4. .4 The Poincar´Friedrichs Inequality. . . . . . . . . . 69 3. . .4 Interpolation theory and applications . . . . . . . .4. 48 3.3 The formalisation of the ﬁnite element method . . . . .iv P. 45 3.3 The construction of the discrete equations . .1 Introduction . . . . 69 3. . .1 The principles . . .5. . . . . . .5 Approximation in two dimensions . . . . . . . . . . . .3.5.4 Other piecewise polynomial approximations in one dimension . . . 92 3. . . . . . .3. . . . . . . . . . . .5. . . . . . . .4. . . . 48 3. . . .3. . . . . . . . . . . 73 3. . . . . .2 The Generalised LaxMilgram Theorem . .4. . . 65 3. .3 Abstract Elliptic Theory . . 45 3.2. . . . . . . .
usually one or more additional (initial) conditions are needed to determine the unknown function uniquely. usually with a large number of unknowns. The second step is the solution of the large algebraic system of equations. etc. In the later part. as e. 1. and some elementary properties of the PDEs are explained. First. and to determine the solutions with some accuracy.1 Ordinary and partial diﬀerential equations An ordinary diﬀerential equation (ODE) implicitly describes a function depending on a single variable and the ODE expresses a relation between the solution and one or more of its derivatives. electromagnetics. Beside the ODE. In Chapter 3 we study the Finite Element Method for the discretisation of elliptic equations. The order of the diﬀerential equation is the order of the highest derivative in the equation.Chapter 1 Introduction to PDEs 1. The numerical techniques to compute (approximations to) the solution are often split in two parts. in order to determine the unknowns that describe (approximately) the solution of the PDE. To get insight into the quantitative behaviour of the solutions of the PDEs. This transformation is called the discretisation process.. in particular.1 Classiﬁcation In these lectures we study numerical methods for partial diﬀerential equations (PDEs). for diﬀerent kinds of PDEs. the PDE (the solution of which is a continuous function or set of functions) is transformed into a set of linear or nonlinear algebraic equations. we will concentrate on so called elliptic equations. These equations are used in many areas of application. in most cases analytic methods fail and we have to rely on numerical methods. 1 .1. mechanics. In Chapter 2 we will give a survey of the methods that are used for discretisation. In the present chapter diﬀerent kinds of diﬀerential equations are classiﬁed.g.
z) reads ∂ (ρu) + div (ρvu) + div Jd = s. .2) In this equation we distinguish between the diﬀusive ﬂux. x. ∂t (1.1. This relation describes a partial diﬀerential equation (PDE). ∂t (1. y. An example of a second order ODE is u(x) = −c2 u (x). z). y. y. again additional relations are required: initial and/or boundary conditions. y. x. and the complete equation describing the behaviour of u(t.2 Let u(t. y. x. Hemker In many interesting cases a function depends on more independent variables and a relation is given for the function and its partial derivatives. 2 The gradient gradient of a scalar function φ(x. z) denote the mass fraction of a chemical species. m. in a medium with density ρ(t.1) where s(t. If diﬀusion also plays a role. u (0) = 1. Jd .W. x. To determine the unknown function. u(0) = 1. y. y. where d is the diﬀusion coeﬃcient. In the presence of a velocity ﬁeld v(t. x. z) the conservation of the species m is expressed by ∂ (ρu) + div (ρuv) = s. In many practical problems the independent variables represent the time and space coordinates (t. divergence of the vector ﬁeld v is deﬁned by div v = important relation for the divergence is Gauss’ theorem div u dΩ = Ω 3 Γ 1 The v= ∂vx ∂x + ∂vy ∂y + ∂vz ∂z . P. Example 1. u(0) = 0. z) is the vector ﬁeld deﬁned by grad φ = φ= ∂φ/∂x ∂φ/∂y ∂φ/∂z . where Ω ⊂ R is a bounded volume with surface Γ and n is the outward pointing normal on Γ. and the convective ﬂux.2 Example 1.1 An example of a ﬁrst order ODE is u(x) = u (x). x. then the diﬀusive ﬂux Jd is described by Fick’s law of diﬀusion 2 Jd = −d grad u. z) is a possible source 1 . An u · n dΓ. Jc = ρuv. z).1.
2004 3 The principle part of a diﬀerential equation is the part (the set of terms) that contain the highest derivatives.5) where a is a positive real number.1) is (ρu)t + div (ρuv) and the principle part of (1.g. All the above mentioned examples are diﬀerential equations that are linear in the dependent variable u. the principle part of (1. the less computational (1. equations in which the highest derivatives appear linearly.3) where f (x. E. x. and often we select values of the independent variables at which the values of u will be calculated. (1. y) is given at the boundary of Ω. (x.e. An important subclass of the nonlinear diﬀerential equations are the quasilinear equations. PDEs can have quite diverse properties. where φ and ψ are arbitrary (twice diﬀerentiable) functions. without a velocity ﬁeld or a source term. The fewer the number of degrees of freedom in the approximation of u(t.3) is called Laplace’s equation.6) .g. the equation (1. • The diﬀusion equation ut = d uxx . Although most theory available is for such linear partial diﬀerential equations. t.3 • With Ω ⊂ R2 an open domain. Notice that the equation (in its threedimensional form) is obtained from (1.1. y. • A diﬀerent kind of PDE is the wave equation utt = a2 uxx . z). the dependent variable u in the PDE is a function of time. Most typical properties are related with the shape of the principle part of the equation. z). i. In general. y) is a prescribed function. Example 1. The solution of these equations can be determined if the value of u(x. y. In the special case of f ≡ 0 equation (1. and the three space coordinates. In numerical methods we have to choose how to represent the function u. A few of such typical forms are given in the next example. y) = φ(at − x) + ψ(at + x). (1.1) or (1. many important equations are nonlinear. the Poisson equation is uxx + uyy = f (x.4) describes the onedimensional unsteady case of simple diﬀusion.2) is (ρu)t − div (d grad u). (1.Version: September 13. y).2) in the case of a timeindependent solution with v = 0 and s = 0. E. It is simple to see that a solution of this equation is given by u(x.2) would be nonlinear if the velocity ﬁeld v was dependent on u or the source s was a nonlinear function of u.
no such direction is found in (1. The notion of direction “forward in time” shows also that in timedependent problems all or part of the sideconditions will be given as “initial conditions”. To understand the behaviour of the solution it is important to notice that in some equations we can distinguish a direction of “ﬂow of information”. y)u = g(x.g. x. y)η + f (x. that determine the solution (i. y).3). we consider the general linear second order diﬀerential equation in two space dimensions: Lu := a(x. If a solution is sought that is not timedependent the solution is called steady and the timevariable can be neglected.to disregard all or most information that was obtained about the solution for t < t1 (or t < t1 − τ . z). y)ξ 2 + 2 b(x.4 P. in equation (1. y)ux + 2 e(x. if such a direction can be identiﬁed.2 Elliptic.6) of equation (1. The knowledge also may be used to reduce the computer storage and/or the computer time needed to ﬁnd the numerical solution. y. The corresponding equations are the steady equations. E.1) the ﬂow of information follows the vector ﬁeld v.2) or (1.1). y)ξ + 2 e(x. the state of the physical system) at a later stage.e. in the positive xdirection) with speed a. Having computed the solution for t ≤ t1 .W.e. t2 ] is determined. Also. For the solution (1. contains all information to determine the solution for t > t1 . (1. in a timedependent problem as (1. (1. This character is mainly determined by the highest derivatives that appear in the equation. whereas the information in the part ψ ﬂows backward as time proceeds. y)uxy + c(x. with t0 < t1 < t2 .4) we can always distinguish a unique direction “forward in time”. y)η 2 + + 2 d(x. For timedependent PDEs such as (1. the behaviour of the solution can be better understood. t2 ]. parabolic or hyperbolic equations In the previous section we noticed that diﬀerent PDEs may show a diﬀerent character. y). 1. However. the solution at time t = t1 .7) L is a linear diﬀerential operator of order two. for which the solution has to be computed over a long timeinterval [t0 . y)uy + f (x. The motivation for the discussion of “direction of ﬂow of information” is that.e. y)ξη + c(x.8) . by a judicious choice of the coordinate system) the number of space dimensions can often be reduced to two or one. y)uyy + 2 d(x. not all problems require all independent variables. it may be eﬃcient from the point of view of computer resources. Fortunately. y)uxx + 2 b(x. Hemker eﬀort is needed to determine an approximation of u(t.1. for some small τ > 0) before the solution on (t1 . By symmetry considerations (i.5) we see that the information contained in the part φ of the solution ﬂows forward (i. To characterise some typical classes of PDEs. With this diﬀerential operator we associate the characteristic polynomial Pxy = a(x.
j=1 aij (x) ξi ξj . . x) = 1 1 1 f (x + at) + f (x − at) + 2 2 2a x+at g(ξ) dξ. the general second order equation reads n n Lu := i. it is hyperbolic at (x. x−at (1.. the diﬀusion equation and the wave equation are elliptic. ξ2 . (1. y) if a(x. ut (0. y)c(x.j=1 aij (x)uxi xj + i=1 ai (x)uxi + a(x)u. parabolic or hyperbolic (without reference to a speciﬁc point) if it satisﬁes the corresponding (in)equalities for all (x.4). It is clear that in more dimensions not all linear second order PDEs can be classiﬁed as elliptic. . The diﬀerential equation is called elliptic.. the characteristic polynomial in ξ1 . (1. with initial condition u(0..9) Now.. 2004 5 The second order diﬀerential equation (1. y) if a(x. The characterisation of the diﬀerent PDEs is important for distinguishing the kind of additional conditions that is needed to specify a (unique) solution for the equation and for understanding the “ﬂow of information” that may appear in its solution. xn ). y) > 0. x) = g(x).. y)c(x. x) = √ 2 πdt +∞ −∞ f (ξ) e−(ξ−x) 2 /(4dt) dξ.Version: September 13. where Ω is the domain on which the diﬀerential equation is deﬁned. with x∈R. x + at]. y) < 0. . ξn is n Px (ξ1 . The parabolic equation For the linear diﬀusion equation (1. y)∈Ω. ξn ) = i. y)c(x.7) is called elliptic at (x.10) The operator is elliptic if the matrix (aij (x)) is positive or negative deﬁnite. y) − b2 (x.. x) = f (x). y) if a(x.. We show this for the three linear examples given earlier in this section.. for the principle part of the equation. x) depends only on the initial solution on the interval [x − at. parabolic or hyperbolic. this terminology is related with the geometric interpretation of the characteristic polynomial. y)− b2 (x. Thus. x2 . and it is parabolic at (x. The hyperbolic equation For the wave equation (1. the general solution for t ≥ 0 reads u(t.12) . y) − b2 (x. For an equation of more variables (x1 . it is parabolic if one eigenvalue vanishes and the other n − 1 have the same sign.. parabolic and hyperbolic. y) = 0.5) with initial conditions u(0. (1.11) This means that the solution at (t. respectively. it is hyperbolic if (aij (x)) has one negative (positive) eigenvalue and n − 1 positive (negative) eigenvalues. x) = f (x). the solution reads 1 u(t. We see that the Poisson equation.
v).14) where v(t. a physical system can often be described with u = u(v). v) + div J(t. ˜ − 2Rr cos(θ − φ) + r 2 (1. v)) = s(t. written in polar coordinates (r. we consider a system of equations. (1.6 P. J = J(v). v(t.14) over Ω. we see that the increase of the amount of u inside the volume Ω should be caused either by inﬂow over its boundary or by a source described in s(t.13) This shows that the behaviour of the solution inside a circle. v) is the ﬂux of the state variables3 .14) is called a system of conservation laws because the equations describe the conservation of the quantities u.15) ∂t ∂ ∂p (ρvi ) + div (ρvi v) = − . The Euler equations of gas dynamics describe the ﬂow of an inviscid nonheatconducting compressible ﬂuid (a gas). r = R. Then. momentum and energy. φ) = ˜ 1 2π 2π 0 R2 R2 − r 2 u(R. This is seen by considering an arbitrary volume Ω and by integration of (1. ∂t (1. x) speciﬁc energy (temperature) and p(t. is completely determined by all the values of the solution at the boundary of that circle. φ). u(t. Generally. x. s(t. x) may be caused by a coordinate transformation. these conservation laws read respectively ∂ (ρ) + div (ρv) = 0.3 Conservation laws A slight extension of equation (1. the dependence on (t. With ρ(t. 1. is given by u(r. x) over the interval −∞ < x < +∞. x) velocity. x.W. x. Equation(1. 2.1. x. the function of an elliptic second order equation is determined by its values at the boundary of its domain of deﬁnition. now each component of J is a vector. x. Example 1. In natural coordinates. v) is a possible source term. x. x. Using Gauss’ theorem. x) the presure of the gas. y) = u(r. v) is a state vector and J(t. e(t.4 The Euler equations. x) is the solution sought.1. x) density. it is known from the theory of harmonic functions that the solution u(x. s = s(v). . φ) in the interior ˜ of a circle with radius R. ∂t ∂xi 3 Because i = 1. They represent the conservation of mass. 3.1) is the system of equations ∂ u(t. θ) dθ. v). Hemker This implies that the solution at x∈R for any t > 0 depends on all of the solution u(0. The elliptic equation For Laplace’s equation.
For a perfect gas we have the state equation 1 p = (γ − 1)ρ(e − v2 ). Along the characteristic the solution is now determined by a simple ODE. 1.18) The solution of the system reads v(t.2 Hyperbolic equations The second order equation (1. ∂t and p = p(ρ. If we make the change of dependent variables v = u + p and w = u − p.17) ∂p ∂u ∂t = −a ∂x .5) allows decomposition into a system of two ﬁrst order equations. v. along which the solution is constant. we consider the inhomogeneous equation ∂u − v grad u = s(u). is ∂u − v · grad u = 0. 3. ∂w ∂w ∂t − a ∂x = 0.1 that the solution of this PDE is uniquely determined as soon as one value of the solution is given for each characteristic. e. These variables play the role of the vector v in (1. this equation (1. These lines are called the characteristics of the diﬀerential equation (or of the diﬀerential operator). x) = w(x + at).Version: September 13. Notice that they are diﬀerent from the conserved quantities u.16) With suitable boundary conditions the variables ρ.20) We assume that the vector ﬁeld v(x) is suﬃciently smooth so that it determines a complete family of characteristics in Ω. the ﬁrst order operators are ∂ ∂ ∂ ∂ found to be ∂t + a ∂x and ∂t − a ∂x .18) the property of having a characteristic direction. A typical ﬁrst order “oneway” wave equation in Rd . (1. (1. v. d = 1. ∂t (1.17) uncouples and transforms into ∂v ∂v ∂t + a ∂x = 0. v). The decomposition ﬁrst can be written4 as a system ∂p ∂u ∂t = −a ∂x . These equations show that the quantities u + p (respectively u − p) are constant along the line x − at = constant (x + at = constant). ∂t (1. 2004 ∂ (ρe) + div (ρev) = −div (pv).16). x) = v(x − at). e and p can be solved from (1.19) It shares with (1. . w(t.15) and (1. 2 7 (1.2.5). To illustrate further the importance of the characteristics.14). Therefore also the ﬁrst order equations are called hyperbolic PDE. Similar to the factorisation of the characteristic polynomial P (τ. 4 The minus sign is by convention only. Then we will show in Section 1. 2. ξ) = τ 2 − a2 ξ 2 = (τ + aξ)(τ − aξ) of (1.
23) Further.8 P. y. if A is a point (x. Q.or more.22) i. y. u). and P.1 Inviscid Burgers’ equation.2. Hemker 1. We write it down for dimension d = 3. y. Let u(x.e. If the solution is prescribed on a suitable (d − 1)dimensional manifold (suitable means that the manifold intersects all characteristics once) then the solution is apparently uniquely determined. It follows that the characteristic direction can be computed from P Q R S = = = · dx dy dz du (1. z) be the solution of (1. −1) ⊥ (dx. R. u) is tangential to the solution in each point of the solution. y. if a value of (x. −1) ⊥ (P. r = uz . y. and r. We consider the equation P p + Qq + Rr = S . for the one.W. u) is given in one point of a characteristic. z)space are characteristics (characteristic direction lines). z. y. and A(s) is an (arbitrary) curve in a solution surface.21) where p = ux . Apparently. Or. q. Example 1.dimensional case the treatment is completely analogous. u).2.21). S are functions of (x. u) on a solution.e. by deﬁnition of p. y. z. The projections of these characteristic solutions on the (x. z. u). For a solution along a characteristic solution line (parametrised by s) we have (P. q. then the characteristic and the solution on it can be found by the solution of the simple ODE (1.1 Characteristics We consider a general ﬁrst order (quasi)linear DE in more dimensions. (1. R. q = uy . dy. q. (1. y. I. the two vectors are linearly dependent. Q. z.21) is satisﬁed as well as du = pdx + qdy + rdz. Q. z. then at A we ﬁnd (p. R. (p. dz. z) are the independent variables and u is the dependent variable. S). y. q. r. inviscid Burgers’ equation A wellknown model problem is the inviscid Burgers’ equation 1 ut + ( u2 ) x = 0 . R. z. ˙ ˙ ˙ ˙ with ˙ = ∂/∂s. (x. S) (x. then. (p. the vector ﬁeld (P. two.23). S)(x. −1) ⊥ (x. du). Q. it follows that. 2 . Streamlines in this vector ﬁeld are characteristic solutions. r. ˙ ˙ ˙ ˙ (1. r.
dt dx du The characteristic satisﬁes dx/dt = u and du/dt = 0.2.g. ∂xj (1. m ∂ui = bk . we introduce a new set of independent variables (φ1 . Under what conditions is a solution determined in a layer in the neighbourhood of C? We easily extend the question to the case of systems of equations. φ2 . m . Let a 1 ux + b 1 uy + c 1 uz + d 1 v x + e 1 v y + f 1 v z + · · · = g 1 . x2 .2)). 2. x2 . (1. we introduce local coordinates around the surface Φ.Version: September 13. or. a 2 ux + b 2 uy + c 2 uz + d 2 v x + e 2 v y + f 2 v z + · · · = g 2 . with d=3. · · · . φ1 = 0. in short. 2004 9 Figure 1. We may study the above arguments more generally in ddimensional space.e. (The surface Φ is our (d − 1)dimensional manifold. (I.) Now we have ∂ui = ∂xj ∂ui ∂φl · ∂φl ∂xj l=1. e. For simplicity we take d = 3. k = 1. akij j = 1.3 . x3 ) = 0. Let a solution be prescribed along a (d − 1)dimensional manifold C. 2. If the initial solution is increasing in some area.) Moreover. · · · . This will be considered further in the next section. φ3 ) such that the surface is determined by φ1 = constant. x3 ) such that a surface Φ is determined by φ(x1 .25) We introduce a function φ(x1 . 3 i = 1. the characteristics will intersect and no smooth solution will further be possible.1: The generation of a shock for Burgers’ equation (See also example (1. and the characteristics are straight lines.24) ···.2. so that the solution is constant along the characteristics. We notice that 1 u 0 = = .
Also. then we can also ﬁnd the normal derivatives (and hence the analytic solution in a layer around Φ). 3 j = 1. on the contrary. then in general. with a∈R. m .e.2. x) is easily understood for t > 0. (iii) ∂ui /∂φl . If the solution u(x. ∂φl ∂xj (1.3 If. m i=1. 2.···. x) = u0 (x − at). 3 i = 1. x∈R.2. x) = 0. · · · . y.31) where a(u) is a continuous function depending on the solution.30) (1. with a(x) a continuous function.27) This is a linear m × msystem for the computation of ∂ui /∂φ1 . in this section we ﬁrst consider the onedimensional case. 3 i = 1. on the surface φ1 = 0. x) + a(u) ux (t. x) = u0 (x) . and the initial condition u(0. x) + a ux (t. det = 0. We have seen that the solution of ut (t. provided that ∂φ1 det akij = 0. j = 1. (1. The reason is that characteristic lines for the equation (1.25) becomes akij l. The problem gets considerably more interesting if the problem is quasilinear. For the computation we need (i) the coefﬁcients akij . i. z) of the system of PDEs is known in the surface Φ.2 Discontinuous solutions For simplicity. m P. k = 1. x) = 0. 2. l = 2. Together with the initial condition (1.2.m j=1.no smooth solution exists in the neighbourhood of Φ.29) is u(t. if we consider ut (t.26) and we get akij ∂φ1 ∂ui = bk − ∂xj ∂φ1 akij l = 2.e.30). (1. 2. even if u0 (x) is a smooth function. ∂φl ∂xj (1.31) may intersect. the normal component of ui on the surface Φ. · · · .28) ∂xj j=1. Hemker ∂ui ∂φl = bk . a . 2. As we have seen. the behaviour of u(t. 3. it now turns out that not in all cases a smooth solution exists for all t > 0.W. (ii) the righthandsides bk .10 Equation (1. 1. (1. i.3 ∂ui ∂φl . · · · .
where we obtain weak solutions of (1.33) ∞ for all φ∈C0 (R ×R ). As the functions u (t. (1. on the half space t ≥ 0. At points where characteristics intersect a shockline is formed and a discontinuity appears. and to consider the integral ∞ 0 Rd φ(t. that have the characteristic function on [t0 .2) and it means that we add a small diﬀusion term to the original equation without diﬀusion. because div(−vu) = −u j (vj u) = −u j vj − vj j u = −vgrad u. We see that here the requirement of a diﬀerentiable u has disappeared ∞ from the formulation of the equation. [t0 . an arbitrary smooth function with compact support. x) = lim →0 u (t. x) for > 0. x) + a(u) ux (t. The class of possible solutions is extended from the class of diﬀerentiable functions to the class of integrable functions. Instead of (1. x) N (u)dx dt = 0. (for a given set of initial data). and different characteristics can carry diﬀerent values.e.e. In our case. 2004 11 classical solution of (1. then that solution is called the physically relevant weak solution of (1. we present the second approach. The integral form reads ut + div J(u) dx dt = 0.t1 ] 5 The (1. with N (u) = ut + div J(u). after partial integration this comes down to ∞ − 0 + d Rd uφt + J(u) · grad φ dx dt = − Rd φ(0.1) to (1.31). 6 Notice that (1.33) and (1.32) This corresponds with the transition from (1. (1. x) exists. and a (nonsmooth) limit function u0 (t.34) is found by integration over an area R = [t0 .35) d Ω general principle of weak solution for a diﬀerential equation N (u) = 0 on R is + d ∞ actually to take an arbitrary φ∈C0 (R × R ).34) with J(u) = −vu.36). The ﬁrst approach is to generalise the concept of a solution. x) u0 (x) dx (1. The characteristic line ends. (We see this form later in (1.) However. There are essentially two ways out of this diﬃculty. Under reasonable conditions for the initial function u0 (x) the parabolic equation (1. it can be shown that (1. i. and the equation is considered in a weak sense.31) should be constant on a characteristic line.32) always has a unique smooth solution u (t.36) are equivalent. t1 ] × Ω as its limit. The integral or weak form5 of the conservation law in divergence form6 ut + div J(u) = 0. and it remains to characterise the “physically relevant” weak solutions. i. By taking a sequence of functions φ∈C0 (R × R ). t1 ] × Ω. Before we consider this question. But this cannot be described by the diﬀerential equation. 0 < 1.Version: September 13.31). it is written in an integral form. x) converge to such a weak solution. x) = uxx . with div v = 0.31) we now consider the slightly perturbed equation ut (t. Now we return to the ﬁrst approach. the weak solutions turn out to be nonunique.19) is written as (1. + d .
Because such solutions are not allowed. For d = 2 we immediately see (cf. Apparently two types of discontinuities may appear in ﬁrst order hyperbolic equations: linear and nonlinear ones.W.35) is a generalisation of (1.40) 2 and it follows that the discontinuity moves with a speed s = (uleft + uright )/2.34).2 Burgers’ equation. (1. after a suﬃciently long period this would lead to a multiple valued function (if that would be possible). The ﬁrst type is present in the initial condition and is carried along a characteristic. is neglected.2. Hence. Hemker Ω Ω [t0 .37) dt D where [u] denotes the jump of u over the discontinuity. a discontinuity appears of which the speed s is determined by u2 J(u) = . (1. we see7 s= J⊥ (uR ) − J⊥ (uL ) .38) This expression for the shock speed is called the RankineHugoniot relation. Figure ??? ) that at a discontinuity D we have dx [u] = [J⊥ (u)] . Denoting the propagation speed of the discontinuity by s. In one space dimension it reads ut + uux = uxx .t1 ] ∂Ω n · J(u) dΓ dt.e. x) is discontinuous. (1.s. The second type may appear even if the initial solution is smooth.36) The integral form holds also if there is a (d − 1)dimensional manifold on which u(t. (1. in a positive solution the top of a wave moves faster than the bottom. . i. Thus. It is convected with a speed v that does not essentially depend on the solution.12 which can also be written as u(t1 ) dΩ − u(t0 ) dΩ = − P. The appearance of a nonlinear discontinuity is often shown for Burgers’ equation. (1. if we take = 0.39) If the r. Example 1.h. and the speed essentially depends the solution. uR − u L (1. 7 In more dimensions we get J⊥ (uR ) − J⊥ (uL ) = s(uR − uL ) where J⊥ denotes the ﬂux vector component in the direction perpendicular to the discontinuity.
A nontrivial steady solution does not appear for u−∞ < 0 < u∞ . x > 0. We can use the relation (1.42) Notice that the initial condition is already a weak solution of the reduced Burgers’ equation ((1. so the RankineHugoniot condition is satisﬁed.2. . after a ﬁrst dissipative phase. and its shape is given by u(t. now it appears that these solutions are not always uniquely determined.39) with = 0).h. otherwise.38) to solve explicitly some initial value problems that are not classically solvable. With u(0.2) has two shocks. then the r. propagating with speeds ±1/2 t. In that case. We take again the inviscid Burgers’ equation as an example. √ The right shock has left. 2004 t 13 x Figure 1.41) u(∞.2: A curved shock solution for Burgers’ equation If we do allow > 0.3 A discontinuous solution. the nonlinear convection will level out all diﬀerences in u. (1.s. u∞ = −q < 0.and right states uL = 1/ t and uR = 0. The dissipated shock moves at a speed a = (uL + uR )/2. 8 For some initial conditions the shape of the dissipated shock is easy to determine for Burgers’ equation. x) = the steady solution of (1. Example 1. − t < x < t. However.Version: September 13. The other shock behaves similarly. x < 0. (1. the following wave is also a solution to the inviscid Burgers’ equation: √ √ x/t. x) = uL + u R uL − u R − tanh 2 2 (uL + uR )(x − at) 2 .39) is u−∞ = q > 0. x) = −q tanh qx 2 . will always have a signiﬁcant value in those regions the solution is steep and the dissipation will prevent the discontinuity to appear8 . x) = (1. u(t.43) 0. √ This solution (see Figure 1. As an example (from [17]).
x < 1 + (t − 1)/2.47) 1. at/2 < x < (1 + a)t/2. or x > 1 + (t − 1)/2. . 1.3: A continuous solution developing a discontinuity u=0 t u=0 u=1 x u=0 t u = x/t u=1 x Figure 1. x < at/2.44) Example 1. x > 0. (1. u(t. 1) 0.2. Hemker u=1 u=0 t=1 u=1 x=t u=0 x u = (1−x)/(1−t) Figure 1.5 A nonunique solution. 0. 1 − x. x) = 0. x) = (1. We take Burgers’ equation with initial function x < 0. (1. (1 + a)t/2 < x. t ≤ x ≤ 1.W. u(t.4: A discontinuous and a continuous solution solution Example 1.14 t x = 1 + (t−1)/2 P. (1 − x)/(1 − t). (1. x > 1. 0 ≤ x ≤ 1.4 A unique solution. 1. or 1.45) 0. u0 (x) = x > 1.4)) are a family of discontinuous functions for a∈(0. a. We take Burgers’ equation with initial function u0 (x) = The solution we ﬁnd reads (see Figure (1.3)) x < t.46) The solutions we ﬁnd are (see Figure (1. x < 0.2.
We again take Burgers’ equation. The interested reader is referred to Chapter 16 of [23]. x) = lim →0 u (t. 2004 t 2x=−(1+a)t u = −a u=a 2x=(1+a)t 15 u = −1 u=1 x Figure 1. t > 0. Such conditions are generally called entropy conditions. x) = 1. −a.51) where E is independent9 of x. +a. min{f (u)} and max{f (u)}.5)). 0.48) Example 1. This condition implies that if we ﬁx t > 0. 2x < −(1 + a)t. however. x) = −1. with some a∈(0.49) Figure (1. and we let x go from −∞ to +∞. with u the solution of ut + f (u)x = uxx . (1 + a)t < 2x. 1). 1.Version: September 13. . then we can only jump down (in one direction) over the discontinuity. x < 0. we formulate a condition that guarantees unicity of the discontinuous solution. x > t. and h. t. We may extend the previous example to obtain a more complex nonunique solution. t) E ≤ . x + h) − u(x.2. now with initial function u0 (x) = The solution we ﬁnd reads (see ua (t. x/t. (1. (1. −1.51). x). u(t. −(1 + a)t < 2x < 0. 0 < 2x < (1 + a)t. (1. h > 0. We do not go into details. u(t. First. one can show that there exists a unique solution that satisﬁes the “entropy” condition [Oleinik] (1. 0 < x < t. 9 E depends on u 0 ∞ . In the case of a single equation ut + f (u)x = 0 with f > 0. h t (1. +1. x > 0.50) To pick out the “physically relevant solution” we have to ﬁnd the solution u(t. Later we shall see a relationship between these entropy conditions and the physical relevance of a solution.6 Another nonunique solution.5: Another discontinuous solution and a continuous function x < 0.
Combining both equations. This shock speed can move in the positive or in the negative direction.53) The entropy condition implies that a shock must satisfy f (ρL ) > f (ρR ). The inequality (1. We had assumed f > 0 (i. With 0 < ρL < ρR < ρm a shock wave travels with speed (1. u(ρ) = um (1 − ρ/ρm ). t) car speed. which implies ρL < ρR . Because cars don’t disappear we have ρt + (ρu)x = 0 . uL − u R for some ζ ∈[uR . (1. Condition (1.52) means that the characteristic lines on both sides of the discontinuity run into the discontinuity. 0) = ρL . With 0 < ρR < ρL < ρm we rather have a “rarefaction wave’. Let u be given as u = u(ρ). and the shock speed for a jump in the car density from ρL to ρR : s= f (ρL ) − f (ρR ) = um (1 − (ρL + ρR )/ρm ) .g. e. uL ].51) says that. x < 0. Hemker This entropy condition of Oleinik guarantees uniqueness of the weak solution.e.52) has been extended by Lax [16].) Let 0 ≤ ρ(x. x > 0. t) ≤ ρm denote car density on a road.7 Traﬃc ﬂow (shocks and rarefaction waves).W.2. We easily derive the characteristic speed f (ρ) = um (1 − 2ρ/ρm ) . ρL − ρ R (1.53). From the RankineHugoniot condition we have s= f (uL ) − f (uR ) = f (ζ) .52) Sometimes this inequality is also called the entropy condition. f is convex) and hence we ﬁnally ﬁnd that Oleiniks entropy condition implies f (uL ) > s > f (uR ) . we have uL > uR . ρR . with f (ρ) = ρum (1 − ρ/ρm ). .51) of Oleinik has not been extended to systems of conservation laws. However (1. at discontinuities. several situations can be distinguished. with initial data ρ(x. Example 1. (This example is taken from [17].16 P. Now. and u(x. The entropy condition (1. gives ρt + f (ρ)x = 0 .
One may obtain these through measurements. and in general. In practice it is clear that there are many restrictions in gathering reliable data from reallife measurements. Numerical mathematics takes the mathematical equations as a starting point and strives for an eﬃcient and accurate approximation of the the quantitative data. Because computational capacity is always ﬁnite. technical or other reallife systems. In the numerical modelling of these systems. the solution has to be approximated by a function that can be represented by a ﬁnite set of numbers.e.it cannot be described by a ﬁnite number of real numbers. the solution of a PDE is a steady or timedependent (vector) function in d space variables (d = 1. 3). 2. Therefore numerical methods are used to get insight in their quantitative behaviour.1.Chapter 2 Discretisation Principles In many cases it is important to obtain quantitative data about physical. and let the solution of the equation be a simple continuous function u(x).1 Let d = 1. most of these equations do not allow an explicit solution in closed form. x∈[a. limitations of accuracy are caused by possible inaccuracies in the mathematical model. b] = Ω. A few possible ways to represent this function are: 17 . we can also try to derive quantitative data from these models. and although much analytical theory is available for PDEs.1 Discrete representation of the solution In many problems from practice. 2. However. because the model is often a simpliﬁcation of reality. Thus the solution is an element of an inﬁnitelydimensional space. Example 2. and also because there is always a limitation of computing resources. if we have suﬃciently accurate mathematical models for a system. the problem has to be discretised: i. Models for technical problems often take the form of a PDE or a system of PDEs.
which forms the basis for the approximating function space. (ii) goniometric functions {φi (x)} = {1. sin x. iii) a table of function values and values of derivatives of that function (either equidistant or nonequidistant) ui.g.W. e. N. Possible choices for {φi } are. xi = a + i(b − a)/N .a denser set of function values or more coeﬃcients to compute. or (iii) piecewise polynomials on a partition of the interval [a. k = 0.g. · · · . In all these cases the solution at the intermediate values of x can be approximated by interpolation. 1.1) approximates the function that is to be represented. This can be done e.g. N. 2. Another way to approximate the function is by choosing a priory. selecting a basis in this space. we need in general. b]. cos x. · · · . ii) a table of nonequidistant function values ui ≈ u(xi ). 1.a ﬁnitedimensional function space. N } so that the function N φ(x) = i=1 ai φi (x) (2. 1. i = 0. N }. and we compute coeﬃcients {ai i = 1. sin(2x). cos(2x). (i) the polynomials φi (x) = xi .1: Piecewise polynomial approximations i) a table of equidistant function values ui ≈ u(xi ). · · · . · · · . a spline or a piecewise polynomial as the interpolating function. If . 1.18 P. · · · . Hemker Piecewise linear Piecewise polynomial Figure 2. 2. piecewise linear or piecewise parabolic functions. Then we have to select a family of functions {φi (x)i = 1. To obtain a more accurate approximation of the function u(x).k = d dx k u(xi ). by a choosing a polynomial. N. i = 0. 2. and determining the coeﬃcients of the approximation with respect to that basis. · · ·}. e. a ≤ x0 ≤ x1 ≤ · · · ≤ · · · ≤ xN ≤ b. i = 0.
Before we describe the techniques how to discretise the equations.2 or 2. in the discretisation process) it is convenient to distinguish a number of stages: i) the discretisation of Ω. where δi.Version: September 13.1.1. (i. a (regular or irregular) mesh of points. the domain of deﬁnition of the PDE. or a partitioning of the domain in triangles or quadrilaterals (See Figure 2. We notice that the classiﬁcation of approximation methods does not yield disjoint classes of methods. This grid is almost always arranged in (uniform or nonuniform) rectangular manner. the domain Ω is represented by a ﬁnite subset of points {xi } = Ωh ⊂ Ω.2. the same method can be considered as a pointwise approximation and as a functional approximation as well. or a partition in subintervals. The discrete equations are obtained by replacing the diﬀerentials in the PDE by ﬁnite diﬀerences in the discrete equations. This parameter is called the meshsize. . These points are the so called “nodal points” of the grid.j denotes the Kronecker delta.1) the coeﬃcients may coincide with function values or values of derivatives. To ﬁnd the discrete representation of the solution. the selection of the representation for the solution). ii) the discretisation of u. the solution (i.1 Discretisation of the domain. 2. Example (iii) can also be seen as an interpolation method for a set of function values.2 Finite diﬀerence approximation For the ﬁnite diﬀerence method (FDM). the largest distance between two neighbouring points. If there is a set of nodal point {xi } and a set of basis function {φi }. one usually selects in the domain where the PDE is deﬁned. and iii) the discretisation of the equation. we ﬁrst describe diﬀerent methods for the discretisation of the domain and of the solution. In the ﬁnite diﬀerence method (FDM) the function u is represented by a set of function values ui that approximate u(xi ). In the representation (2. For a smaller h the approximation gets (in general) (i) more accurate.e. or the size of the largest interval in the partition is usually denoted by a parameter h.e. 2004 19 a partition of nodal points is made. solution and equation In the transformation of the PDE into a ﬁnite set of equations. see Figure 2. and (ii) more laborious (timeconsuming) (”more expensive”) to compute.3). given at the subinterval endpoints.j . 2. such that they satisfy the relation φi (xj ) = δi.
Hemker Uniform regular grid Non−uniform regular grid Figure 2.3: Partition of twodimensional domains .W.20 P.2: Sets of nodal points in rectangular regions Figure 2.
5) that are located around vertices of a grid (solid lines). If the discretisation is constructed on cells (dashed lined in Figure 2. and ∪Ωi = Ω. If there is a single coeﬃcient ai associated with each volume (so that ai may represent u(x) for x∈Ωi ). so that Ωi ∩ Ωj = ∅.1. then the mesh is called vertex centered.2) where the (possibly discontinuous) functions φi (x) are deﬁned only on a single volume.3 Finite element approximation In the ﬁnite element method (FEM) the domain Ω is partioned in a a ﬁnite set of elements {Ωi }. (2.5 Spectral methods uh = ai φi (x) The function u is again approximated by . Further they are constructed so that their support extends only over a small number of elements. the function u is approximated by uh = aj φj (x) . then the grid is called a block centered or cell centered. 2.Version: September 13. and ∪Ωi = Ω. or on the boundaries of the volumes. 2004 21 Piecewise constant Piecewise linear Piecewise linear Piecewise quadratic Hermite piecewise cubic Figure 2. Usually the functions φi are continuous polynomials of a low degree. for i = j. where φi are functions that are polynomials on each Ωi (i. then the method is called cell vertex (see ﬁgure 2. for i = j. Usually one takes for Ωi triangles or quadrangles. piecewise polynomials). Then the function is approximated by uh = ai φi (x) .e. If a (single) coeﬃcient is associated with each vertex. Now in A cell centered (or block centered) ﬁnite volume method (FVM). 2.5).1.1.4: Examples of nodal points in a triangular domain 2.4 Finite volume approximation In this case the domain Ω is also partitioned in a ﬁnite set of volumes (mostly triangles or quadrangles) so that Ωi ∩ Ωj = ∅.
6 Staggered grid For systems of equations all methods mentioned above are easy to generalise. All functions can simply be replaced by vector functions.g. . Example 2. (We see that the diﬀerent types of methods are not always clearly to distinguish. with “staggered grids”. Hemker Cell centered Cell vertex Vertex centered Figure 2. This is the case e.3(ii)) are constructed at the cell vertices.1. The discrete equations corresponding with equation (2. (2. the values of v for the midpoints of the horizontal edges.3) Ω is a rectangular domain. we write the CauchyRiemann equations as L We see that det(L) = − ∂ ∂x u v = ∂ ∂x ∂ ∂y ∂ ∂y ∂ − ∂x 2 u v 2 = f1 f2 .2 We consider the Cauchy Riemann equations (i) (ii) ux + vy = f1 uy − vx = f2 on Ω ⊂ R2 .3(i)) are related with the midpoints of the cells.4) − ∂ ∂y = −∆ . The discrete equations corresponding with equation (2. To understand the character of these equations. Approximations are computed for u and v. The values of u are calculated for the midpoints of the vertical edges. We explain this by two examples.1. (2.22 P. it is also possible to approximate the diﬀerent (dependent) variables by diﬀerent representations.W.5: Examples of nodal points in a quadrangular domain where the functions φi (x) are deﬁned on all of Ω (they have a support that extends over more than only a small part of Ω) then the methods are called global or spectral methods.) 2. One way is to approximate the diﬀerent components of the solution on diﬀerent point sets. However. and Ω is partitioned in a number of equal subrectangles.
it is easily seen that the boundary function g(x.6: A staggered grid for the CauchyRiemann equations v u p v p v p v u v p v p v p v u v p v p v p v u v p v p v p v u v p v p v p v u v p v p v p v u u v p v p v p v u v p v p v p v u v p v p v p v u v p v p v p v u v p v p v p v u v p v p v p v u u u u u u u u u u u u u u u u u u u u u u u u u u u u u Figure 2. y) should satisfy a condition.7: A staggered grid for Stokes equations This shows that the system is elliptic and second order. and 2 × 5 for equation (ii)). and the number of equations is 28 (3 × 6 for equation (i). From ux +vy = div v = f1 . Now the number of unknowns is 27 (viz.n dΓ = ∂Ω g(x.Version: September 13. the system seems to be overdetermined. or ∆v = −curl f .n = g(x.ny = v. and 2 × 6 for v). y) . 3 × 5 for u.nx + v. three dimensions in two dimensions curl v = ∂v2 ∂x1 − curl v = ∂v1 . However. it needs one boundary condition along all the boundary. it follows that ux + vy dΩ = Ω 1 In div v dΩ = Ω ∂Ω v. In our discrete system. It is also directly seen by elimination that 1 ∆u = div f . this condition ﬁxes the values of u and v on the boundary. y) dΓ . ∂x2 ∂v3 ∂x2 ∂v1 ∂x3 ∂v2 ∂x1 − − − ∂v2 ∂x3 ∂v3 ∂x1 ∂v1 ∂x2 . The system being second order elliptic. 2004 v u 1 v 1 v 1 v 23 v u 2 u 2 u 1 v 1 v 1 v u 2 u 2 u v 1 v 1 v 1 v u 2 u 2 u v 1 v 1 v 1 v u 2 u 2 u v 1 v 1 v 1 v u 2 u 2 u v 1 v 1 v 1 v u u u u u variables v u v u v u v u v u v u v u v u v u v v u v u v u v v u v u v u v v u v u v u 1 v u v u 1 2 1 v u 1 2 1 2 1 2 equations 1 2 1 2 1 1 2 1 2 1 1 2 1 2 1 1 1 1 Figure 2. We take u. . Hence.
and gk denotes the value of g(x. −1 1 2.1 Techniques for discretisation of PDEs Finite diﬀerence methods For this method. η∆u − px = 0. Example 2. y) in the kth boundary edge element.24 i. Hemker g(x. The discrete equation is constructed by replacing diﬀerential operators by diﬀerence operators. y) in the ith cell in the interior of Ω.5) i. (2.3 Stationary Stokes equations Analogous to the staggered grid for the CauchyRiemann equations.6) Here f1. y) dΓ . ux + vy = 0.1. for which we need two conditions over all of the boundary. one (linear) dependence relation exists between the 28 equations and we get a solvable system of equations. If the boundary values satisfy condition (2. we have already seen examples. that in totalh k gk = h i uR.6). This is a 4th order elliptic operator. L v = 0 η∆ 2 0 p 0 p 1 2 The character of the system of equations is determined by the determinant of the principle part of the diﬀerential operator detL = η∆ · 2 1 + η∆ · 2 2 + 0 = η∆2 .W.i denotes the value of f1 (x. we can construct one for the stationary Stokes equations. the compatibility condition f1 (x.i − h uL. The .i − h vB. η∆v − py = 0.i = h2 i i f1. 0 u η∆ 0 u 1 v = 0 .i (2.e.i . the operator 1 by −1 1 and 2 by . This can be written as a system with linear diﬀerential operator L.2 2.i + h i i vT. The operator ∆ is represented by the ﬁnite diﬀerence stencil 1 1 1 −4 1 . gk = h k∈boundary i∈interior f1. y) Ω = Ω ∂Ω P.e.2. Similarly it follows from adding all discrete equations (uR − uL )/h + (uT − uB )/h = f1 for each cell.
As this property is shared by most discretisations of PDEs. by which the solution can often be found in a relatively eﬃcient2 manner. In the neighbourhood of xij . most entries in the matrix vanish. the coeﬃcients a are expressed in these function values uij . 2 Compared with a full system of equations of the same dimensions. in the interior points of the domain the same diﬀerence approximation can be used. In this way a “sparse system” of algebraic equations appears. Whereas the systems that appear as a result of discretisations can become very large. in the PDE at this point the diﬀerentials are replaced by diﬀerence approximations.Version: September 13. With the values uij in the nodal points xij around x. thus. 2004 25 Figure 2. That does not mean that the complete function u is considered as a polynomial. Diﬀerent diﬀerence approximations appear near the boundary of the domain. for each unknown uij an equation is set up at the point xij . Now. In this way. they mostly have this nice property. or in parts of the domain where the nodal points are not regularly placed. By the diﬀerence approximations the value of the unknown function at a particular point is only coupled with the values at neighbouring points. the approximation of u(x) is. considered as a piecewise polynomial (of a low degree). The (linear or nonlinear) algebraic system of equations that results from the discretisation process has an important property. These coeﬃcients apq are associated with the derivatives of the approximation in x. The corresponding matrix (of the linearisation) only contains a few nonzero elements in each row. The interpretation of the polynomial representation depends on the nodal point where the interpretation is made.8: Diﬀerence approximation and polynomial interpretation construction of the diﬀerence operators is as follows: in the neighbourhood of a point x∈Ω the approximation of u(x) is replaced by a truncated Taylor series: u(x) = a0 + a10 x + a01 y + a11 x2 + · · · . For a regular grid. special methods are sought (and found) for the eﬃcient solution of such sparse systems of equations. .
j h 2 + ui. y) = 0 on Γ. y) ∂y 2 + 2f (x. y) . This minimisation problem also leads to a linear system. This gives another possibility to construct the discrete equations.2). we can use the variational method also if we seek an approximation in the form uh (x. y) dΩ attains its minimum. and square meshes with a meshsize h) we can write: ﬁnd uh = {uij } such that uij = 0 on Γ. On a regular square grid (i. This equation can be written as the minimisation problem: ﬁnd a continuous function u such that u = 0 on Γ. We show this by means of the Poisson equation ∆u = f. so that with φi (x.2. Ω a rectangle. .e. on Γ . now in the unknown coeﬃcients (ai ). This system also yields a diﬀerence equation for each internal point of the grid. Now we compute the coeﬃcients ai such that ∂ ai φi (x. can be written as a minimisation problem. Hemker 2. y) i ai φi (x. The minimisation of this quadratic functional in the unknowns {uij } leads to linear system in {uij }.j − ui. We take the same example (2.j h 2 + 2fij ui. u = 0. y) ∂x 2 + i Ω i ∂ ai φi (x.2. The variational method for coeﬃcients For problems that can be written as minimisation problems. the symmetric elliptic equations.j+1 − ui.2 Variational method Variational method for diﬀerence methods An important class of PDEs.) x y Now we can set up a discrete analogue for this minimisation problem [5]. and J(u) = Ω ∂u ∂x 2 + ∂u ∂y 2 + 2f u dΩ exists and is minimal. The same technique is applicable on irregular grids.W.26 P.j is minimal. (Exists: u2 + u2 and f u should be integrable functions. on Ω. y) = i ai φi (x. and Jh (uh ) = ij ui+1.
We will treat the most important. for f ∈V. Of course. In this way. This means that the discrete equations are N (uh )(xi ) = s(xi ). Here lxi is deﬁned by lxi (f ) = f (xi ). let S be a set of functions in which the solution u is sought: N : S → V . · · · . in an even more general notation. (2. 2. then for an arbitrary linear functional lw ∈V the equation3 lw (N (u) − s) = 0 is satisﬁed. there are many diﬀerent applications of the weighted residual principle. . Now. NΓ (u) = sΓ . on Γ. 3 The dual space of a Banach space V (denoted by V ) is the family of all bounded linear functionals deﬁned on V . the discretisation yields an n × nsystem of equations.8) on Ω. N } . for any function w we have wΩ (NΩ (u) − sΩ ) dΩ + wΓ (NΓ (u) − sΓ ) dΓ = 0.2. N (u)∈V . In a weighted residual method an ndimensional subspace Sh ⊂ S is selected D and an ndimensional subspace Vh ⊂ V . (2. i = 1. · · · . then. furnished with the obvious addition and scalar multiplication. where n = dim(Sh ). This means that the approximate solution uh should satisfy the original PDE precisely in n given points. The discretisation now reads: ﬁnd uh ∈Sh such that D lw (N (uh ) − s) = 0 for all lw ∈Vh .3 Weighted residual methods A quite general discretisation technique is based on the following consideration. 2004 27 2. let the equation be N (u) = s and let s. for all lw ∈V . for i = 1. N.7) Ω Γ Or.2.Version: September 13. with V a Banach space. then the PDE can be formulated as: ﬁnd u∈S such that lw (N (u) − s) = 0. Let the PDE be given by NΩ (u) = sΩ .4 Collocation methods Collocation methods are weighted residual methods with the functionals lw in D the space Vh that is spanned by {lxi  xi ∈Ω.
The weight functions are now the characteristic functions on these Ωi . and ∪Ωi = Ω.5 Galerkin methods When the functionals l are deﬁned by integrals with integrable weight functions. (2. for i = 1. Then the discrete equation gets the form ∂Ωi J(uh ) · n dΓi = Ωi s(uh ) dΩi . then the discretisation satisﬁes a discrete conservation law. The volumes Ωi are called “boxes”. “cells” or “control volumes”. if i = j. Then the discrete equations read w(x) f (x) dΩ . · · · .6 Box methods = Finite Volume methods Box methods are a special kind of Galerkin methods. Almost all usual Finite Elements methods are of the (Bubnov) Galerkin type.28 P.N I. Typically lw ∈Vh is deﬁned by lw (f ) = Ω D The space Vh is now deﬁned by n linearly independent functions wi (x). 2. such that Vh := span (wi ) = Sh then the methods are called BubnovGalerkin methods.2.···. wi (x) = 1. wi (x) = 0.e. if x∈Ωi . Here the domain Ω is partitioned in a number of volumes {Ωi }i=1. Ω φj N ( ai φi ) dΩ = Ω φj s dΩ. N. and the discrete equations read Ωi N ai φi dΩi = Ωi s dΩi . Hemker 2. D we obtain “Galerkin”methods. When applied to symmetric positive deﬁnite elliptic PDEs.2. if x∈Ωi .W. This means that for an arbitrary union of cells G = ∪Ωi hold: div J(uh ) dΩ = G G s(uh ) dΩ . these methods can also be regarded as variational methods.9) If the computation of J(uh ) at an interface between two cells is made consistently for both neighbouring cells. on Ω. Ωi ∩ Ωj = ∅. An obvious choice is the wi = φi . If we take Vh := span (wi ) = Sh . If we choose {wi (x)}i=1. .N . the method is called a PetrovGalerkin method.2. Box methods are particularly popular for the discretisation of equations in divergence form div J(u) = s(u).···.
.g. in this section we discretise a steady onedimensional diﬀusion equation. e.3 2.s.) The discrete approximation is represented by the function values at the nodal points ui ≈ u(xi ). u) the heat generation per unit length.9: Nodal points and boxes on the real line 2. and approximations ux (xe ) and ux (xw ). but with (ii) we can. on Ω ⊂ R. Starting from ui ≈ u(xi ) we can approximate u. We divide Ω in (unequal) boxes Ωi and we choose for each Ωi a nodal point xi (see Figure 2. 2004 x w x e 29 x i−1 x i x i+1 Figure 2. u = 0.10) together with a boundary condition (e.12) where si is the mean value of s(x. u) over the volume Ωi . and s(x. u)dx = − xw (kux )x dx = − k(x)ux (x)xe . We may choose the boxwalls halfway the nodal points. (2.1 Examples The diﬀusion equation To show the principle of discretisation. by i) taking a constant value ui on Ωi . or we may put the nodal points in the midpoint of the boxes. by some source.Version: September 13.11). This may be seen as the description of heatconduction through a bar. ii) taking a piecewise linear approximation between the nodal points. with a sourceterm s and a variable diﬀusion coeﬃcient k.3. − d dx k(x) d u dx = s(x.9). k(x) the heat diﬀusion coeﬃcient. Then we obtain k(xe ) ui − ui−1 ui+1 − ui − k(xw ) =− xi+1 − xi xi − xi−1 xe xw s(x. Integration of the equation over Ωi yields xe xw xe s(x. xw (2. With (i) we cannot approximate the r.11) To set up the discrete equations we need an approximation of u(x) over Ωi . uh ) dx = −si (uh ) (xe − xw ) .g.h. (Such choice is not relevant at the moment. on ∂Ω). u). where u(x) denotes the temperature. (2. of (2.
12). if s = s(u) is a nonlinear function of u. The processes are identical if we take sL = 0 for all i. Hence the linear system is tridiagonal.W.2 The source term and solution of the system If s is independent of u. i. whereas in (ii) linear dependence of s on u is taken into account.g. and (ii) the equation has to be solved iteratively. The ﬁnal solution of the two processes (i) and (ii) is the same if si (uh ) = sC (uh ) + sL (uh ) ui . This makes that the linear system to solve is changed: k(xe ) xi+1 −xi ui+1 k(xe ) k(xw ) xi+1 −xi + xi −xi−1 k(x + xi −xw ) ui−1 = −(xe i−1 − − xw )sC (uh ) i L − (xe − xw )si (uh ) ui + (n) (n) (2. In this case we have to imation of s as si = si (uh ) + sL (uh )(ui i adapt (2. On the other hand. that have to be determined for each nodal i i point.12) slightly.12) we have used the simplest possible representation for uh that still allows an approximation of (kux ). The (n) source term si can either be approximated by a constant si := si (uh ). or by linearisation.e. then determine sC (uh ). where (n) the previous approximation uh is used directly to obtain a (hopefully better) (n+1) approximation uh by (2. In the construction of (2. (i. Hemker 2. The diﬀerence between (i) and (ii) is that in (i) si depends completely on the previous approximation. i i L An advantage of the alternative (ii) is that. taking a linear approx(n) (n) (n+1) (n) C − ui . or if s depends linearly on u. in the assumptions for the approximation of u) when the diﬀerent parts in the diﬀerential operator are discretised. the process i i i reads: ﬁrst make an initial estimate for uh . the values in the neighbouring cells. then (2.13). This band structure of the matrix makes that the linear system can be solved very eﬃciently.30 P. i i and solve the linear system (2.3. we could take si = s(ui ) piecewise constant. Here.e.12) or (2. then (i) si (uh ) should be approximated. . then process (ii) may converge in one step. sC and sL are constants. This process is repeated until convergence is reached.13) ui is only coupled with ui−1 nd ui+1 . Apparently. and if we know sL (u) = ds/du. if s is a linear i function of u. E. it is not necessary to be consistent in the choice for uh .13) L where si and sC are such that sL (uh )ui + sC (uh ) = si (uh ). However. by a judicious choice of si we may expect that the iterative process (ii) converges faster than process (i). We see that by (2. whereas a piecewise linear approximation was used to determine ux . and sL (uh ).12) is a linear system of equations that can be solved immediately (as soon as the values of the endpoints u0 and un are given) if we take si (uh ) = s(ui ). Other representations are possible as well.
The discrete solution is determined in a number of subsequent timesteps.1) and. and also the values at the odd nodes. again. x).16) ui+1 is coupled with ui−1 only. This is also called instability: the discrete solution is not bounded by the rhs of the discrete equation. this leads to v (ui+1 − ui−1 ) = si (uh ) (xe − xw ) . in a 2dimensional . u = 0 at the left boundary of the interval. The approximations of u at the even nodes are coupled to each other.4. Notice that the technique for the discretisation of the equations in the interior domain cannot be used at the outﬂow boundary. x)dx = xw xw x (vu)x dx = v(x)u(x)xe .3. analogous to (2.16) For a constant v(x) = v. ∂x Now we consider the problem (2. 2 2 (2.Version: September 13.3 The convection equation ∂ (v(x)u(x)) = s(u.17) Here we see that an important problem appears. Because the linear system is singular there is a family of possible solutions.14) where v(x) is a given function.1 Techniques for timediscretisation Timespace discretisation or semidiscretisation We have seen that (because the timedirection is unique) it makes sense to treat the timedependence of a problem diﬀerent from the spacedependence. we see that in equation (2.11) gives xe xe s(u. There a diﬀerent discretisation method should be used. Thus. As a boundary condition we use e. Usually one uses the same timesteps for all nodal points in a grid. a piecewise linear approximation for uh .3. 2 (2. In fact it is this particular discretisation at the outﬂow boundary that determines the behaviour of the discrete solution at the odd points. w (2. In the simplest case of a constant v and s. but there is no coupling between the odd and the even nodes.g. If the problem is deﬁned for 0 ≤ i ≤ N .4 2.15) We take the same grid as in (2. because no value un is available. The homogeneous equation allows nontrivial solutions. 2004 31 2. 2. We consider this completely intolerable. This yields the discrete equation v(xe ) ui + ui−1 ui+1 + ui − v(xw ) = si (uh ) (xe − xw ) . Integration. then there is no boundary condition to determine the values at the odd nodes. and the boundary condition is given at the inﬂow boundary x0 (i = 0).
18) over Ωi yields ∂ ∂t x=xi+ 1 u dx + f (u(x)) Ωi 2 = 0. with ui the mean value of u over the cell Ωi . Another possibility for discretisation of timedependent PDEs is to defer the timediscretisation and to apply semidiscretisation: the solution of the equations is represented by {uij (t)}. Hemker ﬁnite diﬀerence discretisation. The value of the solution at the point xij (or another coeﬃcient used for the spacediscretisation) is considered as a function in time. We do this by means of the simple convection equation ∂u ∂ + f (u) = 0. one possible choice is fxi+ 1 = 2 1 (f (ui ) + f (ui+1 )).2 FDM for a linear hyperbolic problem In this section we will show and analyse a number of characteristic problems that are encountered in the treatment of timedependent problems. Then we arrive at the equation (xi+ 1 − xi− 1 ) 2 2 ∂ ui (t) + fxi+ 1 − fxi− 1 = 0 . and not the past on the future) the values {un } are computed on basis of the ij k known values {uij }. where ij (i. Then. the solutions are represented by {un }. Halfway the 1 nodal points we situate the cell interfaces of the cell Ωi at xi± 2 . as e. fxi± 1 denotes an approximation for f (u(xi± 1 )).g. (2. As in section 2 2. 2 . downwind and central ﬂux computations. To construct a system of equations for the variables {ui } we still have to decide how we relate f (u(xi± 1 )) to the values of {ui }. k = n − 1. 2. the semidiscretised PDE leads to a very large system of ODEs for the unknown functions {uij (t)}. Integration of (2. By the special character of the timevariable (the future depends on the past.3. ∀i. · · ·.32 P. 2 2 For the space discretisation we use upwind. ∂t ∂x (2.3. in d uij (t) = Fij ({ukl (t)}kl ). n − 2. j) denotes the nodal point xij and n the timecoordinate tn . dt For the solution of these equations any suitable technique for the solution of ODEs can be used.18) As a grid we simply take a set of equidistant nodal point {xi }.4.W.20) Here. ∂t 2 2 (2. j.19) x=xi− 1 2 We represent the discrete approximation by {ui }.
An alternative possibility is the backward Euler method: 1 uh (tn+1 ) − uh (tn ) = gh (uh (tn+1 )). fxi+ 1 = f (ui ) or fxi+ 1 = f (ui+1 ). We write fi+ 1 = wf (ui ) + (1 − w)f (ui+1 ).24) In this case the new values uh (tn+1 ) are expressed explicitly in the old values uh (tn ). 2 2 33 respectively the backward and the forward discretisation4 .g.22) ∂ uh (t) = gh (uh (t)). 2.23) simply by timestepping with steps ∆t. ∂t (2.25) These two methods are “implicit” timediscretisation methods. ∆t (2.g. ∆t The fully discretised system reads: un+1 − un = i i 4 With (2. Other choices are e.Version: September 13. the approximation fx = f (ui ) is called upwind approximation. (2. This can be done e.27) ∆t ∆x θ ∆t ∆x (1 n+1 n+1 {wfi−1 + (1 − 2w)fin+1 − (1 − w)fi+1 }+ n n n − θ) {wfi−1 + (1 − 2w)fi − (1 − w)fi+1 }. by the forward Euler method: 1 uh (tn+1 ) − uh (tn ) = gh (uh (tn )). ∂t ∆x ∆x ∆x 1 where ∆x = xi+ 2 − xi− 1 . .23) We realise the time discretisation of (2. For convenience we will combine the three choices in one formula. We summarise the three possibilities. The other approximation is called downstream approximation. in one parametrised formula: 1 uh (tn+1 ) − uh (tn ) = (1 − θ)gh (uh (tn )) + θgh (uh (tn+1 )) . · · · . 2 Now we have completed our semidiscretisation: w 1 − 2w 1−w ∂ ui (t) = f (ui−1 ) + f (ui ) − f (ui+1 ). After some analysis we make a choice for w. 2004 the central approximation. with a parameter w. ∆t (2. again. ∆t or also the CranckNicolson method 1 uh (tn+1 ) − uh (tn ) = ( gh (uh (tn )) + gh (uh (tn+1 )) )/2. Here we have to solve a system of equations to compute the new values uh (tn+1 ).21) (2. 3. i+ 1 2 because the ﬂow of information is from left to right and for the approximation of fx the upstream value ui is used.26) (2. We write this equation brieﬂy as 2 (2.28) i+ 1 2 ∂f /∂u > 0. i = 1.
This can also be written as γ= = 1 + λ(1 − θ){(1 − 2w) + we−jωh − (1 − w)e+jωh } 1 − λθ{(1 − 2w) + we−jωh − (1 − w)e+jωh } (2. In order to study the behaviour of the solution of the discrete system. We set du = f constant. we consider the behaviour in time of the approximate solution in the domain Ω = (−∞. The equation now is ut + f ux = 0. x). it doesn’t grow or shrink. it is simple to solve it also analytically. +∞). i Using equation (2. In time the solution doesn’t really change.28) becomes un+1 − un = λθ {wun+1 + (1 − 2w)un+1 − (1 − w)un+1 }+ i i i−1 i i+1 λ(1 − θ) {wun + (1 − 2w)un − (1 − w)un } . and we combine a few parameters in a single: λ = f ∆t/∆x. we can see which properties of the true solution are inherited by the numerical counterparts. it only moves with the velocity f .30) i where γ is a complex number that satisﬁes γ − 1 = λθ {wγe−jωh + (1 − 2w)γ − (1 − w)γe+jωh }+ λ(1 − θ) {we−jωh + (1 − 2w) − (1 − w)e+jωh } .29) we see immediately that the solution at a later stage is given by un = γ n ejωhi . We see that γ2 = {1 + λ(1 − θ)(1 − 2w)(1 − cos(ωh))}2 + λ2 (1 − θ)2 sin2 (ωh) · {1 − λθ(1 − 2w)(1 − cos(ωh))}2 + λ2 θ 2 sin2 (ωh) (2.W. (We refer to a textbook in Fourier analysis to understand the relevance of this choice. Hemker k where fj represents f (uk ).31) 1 + λ(1 − θ){(1 − 2w)(1 − cos(ωh)) − j sin(ωh)} . for an initial function u(t0 . that was introduced in the space discretisation. We use the notation j := √ −1 in order not to interfere with the index i previously used. is u0 = ejωhi . With the diﬀerent possible choices for the numerical method. Further we can see what is the eﬀect of the parameter λ. . Because it is diﬃcult to analyse nonlinear equations we simplify the df problem by linearising it. and of θ that was introduced in the time discretisation. i−1 i i+1 (2.34 P. We will use this knowledge about the analytic solution to check if our numerical techniques yield reliable results. and the solution is ψ(x − f t) for an arbitrary initial solution ψ(x) = u(0.) The discrete initial function. The solutions i increases for t → ∞ if γ > 1. Now equation (2. j In general. then.32) (2. to compute un+1 from un . 1 − λθ{(1 − 2w)(1 − cos(ωh)) − j sin(ωh)} γ is the ampliﬁcation factor for the timeintegration: un  = γn .33) Now we want to study the eﬀect of the parameter w. (2. we have to solve a system of nonlinear i i equations. x) = ejωx . and it decreases for γ < 1.29) Having simpliﬁed the equation so far.
4. 1 This implies instability for θ ∈[0. 2004 • If w = 1 2 35 (i. the amount to which the true solution at the nodal points. The downwind discretisation is unstable for all reasonable cases.e. the timestep) small enough. with f < 0 it is the other way. – f < 0 ⇔ λ < 0 ⇔ stability for: λ(1 − 2θ) ≥ 1. If θ < 1 stability may depend on λ. γ < 1 ⇔ (1 − θ)2 ≤ θ 2 ⇔ θ ≥ • If w = 1 (i. Depending on the sign of f we ﬁnd: – f > 0 ⇔ λ > 0 ⇔ stability for: λ(1 − 2θ) ≤ 1. 2 2 1 stability is obtained for λ ≤ 1−2θ and we obtain stability only for λ (i.e.e. only for λ large(!) enough. i. 1 ].33) reduces to γ2 = 1 + λ2 (1 − θ)2 sin2 (ωh) .e. we obtain stability for λ small enough.does not satisfy the discrete equation 5 ) 5 For the diﬀerential operator L : X → Y and its discrete operator L : X → Y . 2 ] we get 1 stability for λ ≤ 1−2θ . if θ ∈[0.e. • If w = 0 (i. This means unconditional instability (forward Euler). if θ > 2 then get stability for 2 1 λ ≥ 2θ−1 . i. We recognise a symmetry If f > 0 then w = 1 corresponds with upwind discretisation and w = 0 with downwind. 1 + λ2 θ 2 sin2 (ωh) (2. for “forward” discretisation in space) stability ⇔ γ ≤ 1 ⇔ λ(λ(1 − 2θ) + 1) ≤ 0 – f > 0 ⇔ λ > 0 ⇔ stability for λ(1 − 2θ) ≤ 1. 1 ]. and we ﬁnd 1 1 – f < 0 ⇔ λ < 0. This implies instability for θ ∈[0. If θ > 1 then stability is guar2 2 1 anteed for λ ≥ 2θ−1 and we obtain stability only for λ large (!) enough. 2.3 The equivalent diﬀerential equation In this section we study the local discretisation error (i. for central discretisation in space) equation(2. iii) θ = 1 ⇔ γ < 1.Version: September 13. this implies stability for θ ≥ 2 . For 2 1 θ < 2 it is stable if the timestep is small enough. 2 ii) θ = 0 ⇔ γ > 1. This implies stability for θ ≥ 1 . This means unconditional stability (backward Euler). for “backward” discretisation in space) stability ⇔ γ ≤ 1 ⇔ λ(λ(1 − 2θ) − 1) ≤ 0.e. We notice that the upwind discretisation is unconditionally stable when θ ≥ 1 .e. and the h h h restriction operators Rh : X → Xh and Rh : Y → Yh .34) 1 · 2 Depending on θ we can draw the following conclusion: i) θ = 1 ⇔ γ = 1. we remember the deﬁnitions for global . This implies neutral stability (CrankNicolson).
37) Substitution of this in (2. ∂x2 i ∂u ∂2u ∂u n = ∂x − θ ∆t ∂x∂t + O(∆t2 ). and t = tn + θ∆t. 1 + 2 ∆x2 + O(∆x3 ). . We consider again the linear problem ∂ ∂ u+f u = 0.36 for (2. We obtain (i) (ii) (iii) (iv) (v) (vi) (vii) (viii) (ix) (x) un+1 = u + (1 − θ)∆t i ui−1 = ui − ∆x ui+1 = ui + ∆x 1 2 (ui+1 ∂u ∂x i ∂u ∂x i un = u − θ∆t i ∂u ∂t + 1 θ 2 ∆t2 2 ∂u ∂t ∂2u ∂t2 + 1 ∆x2 2 + 1 (1 − θ)2 ∆t2 2 ∂2u ∂x2 ∂2u ∂x2 i i + O(∆t3 ). ∂x i ∂u n+1 ∂2u = ∂u + (1 − θ) ∆t ∂x∂t + O(∆t2 ). ∂x i ∂x n n+1 = ∂u + O(∆t2 ).W. We denote u := u(xi . u(xi ) − ui  = (Rh u)i − ui . t ). ∂2u ∂t2 + O(∆t3 ).36) In order to compute the local truncation error. and for the local discretisation error (Lh Rh u − Rh Lu)i . ∂x2 i i − ui−1 ) = +∆x (2. we expand the function u(x.29) also as 1 1 un+1 − un = −λθ{ 2 (un+1 − un+1 ) + ( 2 − w)(un+1 − 2un+1 + un+1 )}+ i i+1 i−1 i+1 i i−1 i 1 1 n n −λ(1 − θ){ 2 (ui+1 − ui−1 ) + ( 2 − w)(un − 2un + un )}. t) in a Taylor series around the point x = xi . ∂t ∂x P. + O(∆x3 ). i+1 i i−1 (2. Hemker (2.35) and its discretisation equation (2.36) yields for the lefthandside n un+1 − ui i ∂u ∂t 1 + 2 (1 − 2θ)∆t2 ∂2u ∂t2 = ∆t + O(∆t3 ) = and for the righthandside −λθ ∆x = −λ ∆x ∂u n+1 ∂x i 1 + ( 2 − w)∆x2 ∂2u ∂x2 n+1 i ∂2u ∂x2 −λ(1 − θ) ∆x discretisation error ∂u n ∂x i ∂u ∂x + O(∆t2 ) + ( 1 − w)∆x 2 1 + ( 2 − w)∆x2 n i 2 + O(∆x3 ) + O(∆x3 ) ∂2u ∂x2 + O(∆t2 ) + O(∆x3 ) . ∂u 3 ∂x i + O(∆x ).28). 2 (ui+1 − 2ui + ui−1 ) = ∆x2 ∂ u + O(∆x3 ). We can write (2.28). (1 − θ) ∂u i + θ ∂u i ∂x ∂x ∂x n n+1 2 ∂2u ∂2u (1 − θ) ∂x2 + θ ∂x2 = ∂ u + O(∆t2 ).
and the discrete solutions are much more smeared than the true solution of the diﬀerential equation. 1 + O(∆t2 ) + ( − w)∆x 2 ∂2u ∂t2 1 + f (w − )∆x 2 1 = (θ − )∆t 2 ∂2u ∂x2 1 This means that the method is ﬁrst order consistent in time and space. however.38) 1 1 f ∆x (θ − )λ + (w − ) uxx = 2 2 . 2004 37 Making use of the parameter λ = f ∆t/∆x we get the expression for the truncation error (in equation (2.2 corresponds exactly with a positive numerical diﬀusion! The numerical diﬀusion.e in the numerical solutions maxima or minima appear that do not exist in the true solution. can cause problems in a convection diﬀusion problem if the diﬀusion coeﬃcient in the equation is (much) smaller than the numerical diﬀusion. If we make computations with one of the above schemes.39) So.35) we ﬁnd that utt = (f )2 uxx . If θ = 2 1 it is second order in time.40) is also called the numerical diﬀusion of the diﬀerence Remark: Notice that the stability of the diﬀerence scheme as studied in Section 2. (2. i. num (2. and if w = 2 it is second order in space. we may see the diﬀerence scheme as a second order discretisation of a modiﬁed equation (or the equivalent diﬀerential equation) 1 1 ut = f ∆x (θ − )λ + (w − ) uxx − f ux . we notice: (i) the numerical diﬀusion can have a signiﬁcant eﬀect. .Version: September 13. (iii) unstable scheme are useless.4.36) τ (u) = rhs − lhs) τ (u) =− ∆t −f ∂u ∂t ∂u ∂x 1 − ( − θ)∆t 2 ∂2u ∂t2 + O(∆t2 ) ∂2u ∂x2 + O(∆x2 ) + O(∆x2 ) + O(∆t2 ) . (ii) the second order scheme (without numerical diﬀusion) shows an “overshoot”. so that the principal term in the truncation error can be written as f 1 1 (θ − )f ∆t + (w − )∆x uxx = 2 2 num uxx (2. The numerical diﬀusion may then overrule the ‘physical’ diﬀusion. Diﬀerentiating the PDE (2. 2 2 For this reason scheme.
and we solve the problem for diﬀerent values of w and θ.38 Example 2.05 (λ = 0.01 + num )uxx − ux . as well as the solution of the modiﬁed equation ut = (0. We discretise the term uxx by uxx ≈ 1 1 n+1 (ui+1 − 2un+1 + un+1 ) + (un − 2un + un ) .01 and either ∆t = 0.1 We solve the equation ut = 0.1 (λ = 1.10 and 2. We take a ﬁxed ∆x = 0.4. and boundary condition u = 1 at x = 0 (at the “inﬂow” boundary). Hemker on the domain t ≥ 0. with initial condition u = 0 at t = 0. . x ≥ 0.W.11 we show the results for t = 1 (and in Figure 2.0). i i−1 i i−1 2 2 i+1 The other terms are discretised by (2.29).12 for t = 0.5) or ∆t = 0. For comparison the true solution is also shown.01 uxx − ux P. In the Figures 2.5).
2 0.05 1o o o o o o o o o o 1o o o o o o o o o o 0.025 Figure 2. num = 0. .10: Numerical solution of the convection diﬀusion equation In all ﬁgures the numerical solution is given for t = 1.6 o o 0.4 0. num = 0.6 0.2 0.4 1. ∆x = 0.4 0.0. The true sulution is given by the solid line.4 1.8 o o 0.4 0.1 1.8 o 0.2 0.2 0.4 1.6 0.8 0.2 1. 2004 39 ∆t = 0.5. num = 0. 1. The dotted line is the numerical approximation.8 2 θ = 0.2 1. obtained with upwind space discretisation: w = 1.4 o 0.6 o o o 0.8 2 θ = 1.8 o o 0.4 1.6 o o 0.2 0 0.2 0.2 0 0. the dashed line represents the solution of the modiﬁed equation.2 0 0.2 0 0.8 1 1.6 0.2 o o o o o 0 o 0 o o o 0.6 o o 0.Version: September 13.1.8 0.2 num = 0.2 1.6 1.8 2 0.4 o o o o o o o o 0.2 o o o 0 o o o o o o o 0 o o o o o 0.8 2 0.6 o o o o o o o o o o o o o 0.2 o o 0.6 1.05 1.6 0.6 1.4 0.2 0.4 0.2 0 0.6 0.4 o 0.2 ∆t = 0.8 2 θ = 0.6 1.2 1.2 0.8 1 1.05 1o o o o 1o o o o o o o o o o o o 0.4 o o 0.4 0.2 1.4 1.6 0.8 1 1.8 1 1.2 0.8 o o 0.6 0.8 1 1.2 θ = 1.6 1.075 1o o o o o o o o 1o o o o o o o o 0.2 θ = 0. num = 0.2 o o o o o o o 0 0 0.2 0 0.8 2 0.2 1.5. 1.4 0.6 1.2 1.8 1 1.4 1.4 0.1 1.0 θ = 0.2 num = 0.
2 1.2 ∆t = 0.2 0.40 P.4 0. 1.8 2 0.6 0. the dashed line represents the solution of the modiﬁed equation.4 o o 0.2 o o o 0 o o o o o 0 o o o o o o 0.4 1.8 o o o 0.2 0 0.4 0.6 1.2 0 0.4 0.8 2 θ = 0.2 1.4 o o 0.4 0.8 2 θ = 0.8 2 0.4 1.2 1.6 0.2 o o o o o o o o 0.6 1o 0.4 o o 0.2 0 0.8 1 1.6 0.2 o o o o 0 0 o o o o 0.11: Numerical solution of the convection diﬀusion equation In all ﬁgures the numerical solution is given for t = 1.4 1. obtained with central space discretisation: w = 0.8 0.2 o o o o o o o o o o o o o o o 0 o o o o o o o o 0 0 0.0 2 1o o o o o o 1.8 2 0.6 0.2 0.6 1.2 θ = 0.4 1.5.8 1 1.2 o o o o 0.2 0.5.6 0.8 1 1.4 0.05 1. num = −0.8 1 1.8 1 1. o o o num = 0.8 o 0.2 0 0.05 1o o o o o o o 1o o o o o o o o o o 0.8 1 1.6 0.5.6 o o 0.6 1.1 1.6 1.8 0. num = −0. Hemker ∆t = 0.2 0.025 1o o o o 1o o o o o o o o 0.4 0. .6 1.2 1.4 1.2 0.8 2 θ = 1.2 o o o o num = 0.5 o o o 0.2 0. o o num = 0.05 θ = 0.2 1.4 0. o o num = 0.4 1.2 1.8 o o 0. The dotted line is the numerical approximation. ∆x = 0. The true sulution is given by the solid line.6 1.6 o o 0.025 Figure 2.W.6 0.2 1.1.5 o 0.2 θ = 1.2 0 0.0 1.6 0.4 o o 0.
8 o 0.2 0.05 ∆t = 0.13: Nonconservative interpolation . The dotted line is the numerical approximation.5).2 1.6 1.4 0.4 1.8 2 ∆t = 0.8 0.1.2 0 0 o o o o o o o o o o o o o 0.4 0. num = −0.4 1. obtained with central space discretisation (w = 0.5.12: Numerical solution of the convection diﬀusion equation In these ﬁgures the numerical solution is given for t = 0. ∆x = 0.6 o o o 0. The true sulution is given by the solid line.Version: September 13. and forward Euler timediscretisation (θ = 0). 2004 41 o 1. Figure 2.05.025 Figure 2.1.6 0.8 2 0. num = −0.2 1o 0.6 o 0.2 1.2 o o o 1. the dashed line represents the solution of the modiﬁed equation.2 0.4 1o 1.6 1.2 0.2 o o o o o o o o o o o o o o o o o o 0 0.8 1 1.4 0.6 0.4 0.8 1 1.2 0 0.
42 P. Often the functional has a physical meaning (e. if it is known that the continuous solution is positive or monotone. the energy left in the system) and. We wish that a reasonably accurate solution is already obtained for a reasonably ﬁne (= reasonably coarse) grid. Hemker 2. I.shows the same “global character” as the continuous solution. the ﬂux over boxinterfaces should be computed consistently. I. In order to satisfy the conservation law locally in a box method. If the number of nodal points (coeﬃcients for the representation) increases we may expect (and we should require) that the discrete solutions converge to a true solution of the continuous problem.g. The last three rules are taken from [19]. E. not only globally. the computed ﬂux J(u) · n over the box interface should be the same when it represents the ﬂow out of the one cell or when it represents the ﬂow into the .W. one doesn’t like to ﬁnd negative concentrations by approximation errors. Often it is expressed that we want to obtain a “physically realistic” solution. A conservation law is most reliably discretised if the discrete system satisﬁes the law on each subelement of the discretisation separately. Rule 2 Often we wish more than only an asymptotic validity. This requirement is more diﬃcult to formulate mathematically. for all Ω. A class of (elliptic) problems for which the requirement of “physical relevance” is usually not too hard to satisfy. For such symmetric elliptic problems a convex functional exists that is minimised by the solution of the PDE. Rule 1. are those problems that can be formulated as a minimisation problem. the minimising element in a subspace satisﬁes automatically the requirement of physical relevance. but also locally. For the class of (hyperbolic) problems in conservation form it is relatively simple to satisfy the requirement that the discrete equations satisfy the conservation law. Rule 3: Obeying the conservation law. one requires that also the computed quantity is not created or does not disappear in the computation. one requires that the discrete solution already for a coarse grid. For example. therefore.e. (iii) if the continuous equation describes a conservation law. one requires that the temperature as computed does not show local extrema. (i) if concentrations of a chemical are computed.e. the discretisation error (in the solution) (in some norm) should vanish (at some rate) as h tends to zero. one often requires that also the discrete solution is positive or monotone. This property of a discretisation is usually mathematically formulated by an asymptotic statement: if the meshwidth h of the discretisation is suﬃciently small. for each cell Ωe of the discretisation of Ω.g. We do not want to construct an extremely ﬁne grid before we can be sure about the value of the approximation. (ii) for the stationary solution of a heatconduction problem without sources or sinks. Some discretisation methods (ﬁnite element methods) ﬁnd this minimising approximate solution.5 Criteria for good discretisation methods In this section we want to summarise a set of rules and give criteria that can help us to ﬁnd good discretisation methods.
I. 2004 43 other. The same argument is simple to generalise for two or three dimensions.5. If a diﬀerential operator only contains derivatives of a dependent variable.1. .2. First.1 If we use a threepoint. In order to guarantee the positivity of the solution (in the absence of sources and sinks) for the linearisation of the discrete operator. Further. Explanation: If we write the linearised discrete equation for one space dimension as ai ui = ai−1 ui−1 + ai+1 ui+1 + si (2.satisﬁes basic requirements for most simpliﬁcations of the problem. (The rowsum of the linearisation of the discrete operator should vanish. the constant function should be in the kernel of the discrete operator. Rule 4: Trivial (constant) solutions should satisfy the discrete solution.) Rule 5: Guaranteed positivity of the solution.12) we had approximated k(x) by k(xi ). xi+1 − xi If in equation (2. Example 2. If we take the twopoint linear approximation the conservation requirement is satisﬁed. In equation (2.5. (ii) if there is a source term. then with an approximate solution u also u + constant should be a solution of the discrete equations.12) we have (correctly) approximated J(u) by J(u) = k(xi+ 1 ) 2 ui+1 − ui . parabolic approximation to compute the ﬂux over the right boundary of cell Ωi (ﬁg. The explanation is simple: the amount of the conserved quantity that leaves the one cell should reappear (exactly) in the other cell. (A good method for a problem is also a good method for any simpliﬁcation of that problem.) This rule can be seen as an instance for the more general principle: Seeking a good discretisation for a complex (intricate) operator. The ﬁfth rule allows an extension for the case that a source term is present. it should not disturb the positivity (monotonicity).3.13) then the requirement of conservation is not fulﬁlled. it makes sense to ﬁnd a discretisation that at least.41) the coeﬃcients ai . Example 2. then the discretisation would not have been conservative. if the source term is neglected the discretisation should satisfy rule 5.2 See Section 2. For discretisations with a higher order of accuracy (than one) it is diﬃcult (often impossible) to satisfy this requirement. the diagonal and the oﬀdiagonal elements of the Jacobian matrix should have diﬀerent signs. For higher order discretisations more nonzero diagonals appear in the linearised operator. ai−1 and ai+1 should have the same sign to guarantee that for all possible ui−1 and ui+1 .e.Version: September 13.
2)) (ai − sL )ui = ai−1 ui−1 + ai+1 ui+1 + sC . Hemker Let s(u) the source term in the equation Lu = s(u) allow the linearisation s(u) ≈ sC + sL + u.44 P. i . then the discretisation of Lu = s(u) leads to (see sect (2.3.W. i i then sL and ai should have opposite signs.
45 ..1 Introduction This chapter is devoted entirely to the subject of Finite Elements. Then we give the theoretical foundation of the method. This is best illustrated by means of elliptic equations. Excellent reference books on the subject are. 3.Chapter 3 Finite Element Methods 3. First we give a number of examples that also will be treated later in applications of the method. nonlinear equations or problems that involve diﬀerentiation in time. 3. Although we will restrict ourselves in these lectures to the examples. in general we can apply the FEM also to systems of PDEs. e.g. a0 and f .2 Examples of boundary value problems The ﬁnite element method is best applied to (partial) diﬀerential equations in variational form. b] ⊂ R: −(a2 ux )x + a1 ux + a0 u = f. it is known that this problem has a unique solution. In this section we discuss a number of typical examples we will frequently use. u(a) = u(b) = 0. Subsequently we discuss how the method can be used in practice and ﬁnally we show how accurate the method is. For smooth functions a2 .1 Onedimensional second order The simplest problem we meet is the homogeneous Dirichlet twopoint boundaryvalue problem (TPBVP) on an interval Ω = [a. a1 . [2] and [3]. and a2 > 0.2. which are always of even order.
2) is deﬁned on an open connected domain Ω ⊂ R2 .46 P. we will discuss it in more detail in a later chapter. ∀ξ = (ξ1 . The special case when b0 = b1 = 0 is called Neumann condition.3) is equivalent with completely determines the sign of the left hand side of (3. D is is the identity matrix) and the Laplace equation (v = 0. This boundary condition can be of the form u = g on Γ. Deﬁne the coeﬃcient matrix by A= then (3. The sign of a11 It is easy to see that (3. .W. ∀ξ = 0.3) < a22 a11 .1) where ψ(x.4) a11 a12 a12 a22 . (3. f = 0). ai and f are functions of x and y. The elliptic equation (3. and f (x.e. boundary conditions of the form (3.5) mixed conditions. It appears that we need one condition along all the boundary Γ := ∂Ω. more generally. or. the matrix A is positive deﬁnite. where the coeﬃcients aij . a2 12 ∂u ∂u + b1 + b0 u = h on Γ. v(x. Special examples are the diﬀusion equation (v = 0). Because of the special diﬃculties in the numerical treatment of the convection diﬀusion equation. y) is the source function.3) is equivalent with ξ T A ξ > 0. secondorder. and to obtain a unique solution we have to provide boundary conditions. ξ2 ) = (0. We assume the equation to be elliptic on Ω.2 Twodimensional second order An example of a twodimensional.2. Hemker 3. elliptic equation is the steadystate convectiondiﬀusion equation · (D ψ) − · (ψv) = f on Ω ⊂ R2 . (3. D(x. We write the linear twodimensional secondorder equation in its most general form: a11 uxx + 2a12 uxy + a22 uyy + a1 ux + a2 uy + a0 u = f. (3. 0) ∈ Ω. the potential equation (v = 0. D is is the identity matrix.2) satisfy 2 2 a11 ξ1 + 2a12 ξ1 ξ2 + a22 ξ2 = 0. y) a diﬀusion tensor.2) on a domain Ω ⊂ R2 . which means the coeﬃcients of the principle part of (3. (3. y) is the unknown function (for instance the temperature or the density of some matter). i.4) are called Dirichlet conditions. y) is a (given) velocity ﬁeld.5) ∂n ∂s where n is the outward unit normal on Γ and s is the vector tangent to the boundary.3). (3. Boundary conditions of type (3.
η) = W. y) = v(x. η) = I. AT T u) = ( . η) . the principle part a11 uxx + 2a12 uxy + a22 uyy can be reduced to the form ∆u by a linear coordinate transformation. A u). Also easy to check are ∂/∂ξ ∂/∂η (ξ. and a similar expression for the other coordinates. We also deﬁne T = (∂/∂ξ. T AT T u).2) can be written as ( .1 If the coeﬃcients in equation (3. ∂/∂η) . A direct consequence is (ξ. A substitution 1 u(x. A transformation of equation (3. Because both α11 and α22 are positive we can reduce the principle part of the elliptic equation by a simple scaling to ∆u. Assuming the existence of a W so that = W we see TT = = W Hence = TT ∂/∂x ∂/∂y ∂/∂ξ ∂/∂η (x. (3.Version: September 13. ∂/∂y)T . Proof: First we notice that the principle part of (3.2. y) T T . y) T T = (ξ. ∂/∂x ∂/∂y (ξ. α22 are positive. by the deﬁnition of adjoint. where α11 .6).6) By an additional transformation of the dependent variable u we can remove the ﬁrst order terms from (3. y) · exp{ (β1 x + β2 y)} 2 .2) along the lines of the above theorem reduces the equation to −∆u + β1 ux + β2 uy + β0 u = f. α22 }. 2004 47 Theorem 3. and we write the principle part as: ( . Because A is positive deﬁnite we can choose T such that T AT T = D. where we use the notation = (∂/∂x.2) are taken constant. A u) = (T T . The transformation we need is of the form: ξ η =T x y . with D = diag{α11 . This completes the proof. η) = (x.
In case of variable coeﬃcients βj . We may consider the onedimensional homogeneous biharmonic problem on [a.13.. 3.7) For constant β’s this is the same as the previous reduced equation. 3. To understand the existence and uniqueness of such problem. After this we state the theorem and give a number of applications.. b] ⊂ R. Hemker 1 1 2 2 −∆v + [β0 + (β1 + β2 )]v = f (x. These computations show that any uniqueness/existence result we obtain for −∆u + au = f will also be valid for the general case. so that there exists a B(x. if × β = 0. y) exp{− (β1 x + β2 y)}. y) such that β = B ≡ grad B). To formulate this theory properly we ﬁrst introduce some basic notions. 1 2 β · dx}.2. e. in Section 3.3 we will discuss parts of the theory on elliptic partial diﬀerential equations.3 Fourth order problems The results on existence and uniqueness of elliptic problems are also applicable to fourth order elliptic problems like the biharmonic problem ∆2 u = f in Ω. if the ﬁeld β is rotation free (. y) = v(x. clamped at its boundary.g. we use the transformation u(x.18.3.W.6) gives 1 2 1 1 2 −∆v + [β0 + (β1 + β2 ) − ((β1 )x + (β2 )y )]v = f · exp {− 4 2 2 β · dx}. ∂u = 0 on Γ.48 reduces (3. u = 0 on Γ. u(a) = u(b) = u (a) = u (b) = 0. under a transversal load. a2 uxxxx − a1 uxx + a0 u = f.3. y) · exp{ Inserting this in (3. ∂n To this form we may reduce the Stokes problem in ﬂuid dynamics and it also models the displacement of a thin elastic plate.e.3 Abstract Elliptic Theory A powerful tool to prove the existence and uniqueness of solutions of PDEs of elliptic type is the (generalised) Theorem of LaxMilgram 3.6) to P. which can be seen as a variant of the Riesz Representation Theorem 3. (3. 4 2 Note that the right hand side of the equation is weighted by an exponential factor. .
x ≥ 0. i. 8. ∃! 0 ∈ X. 4. 1 · x = x. x).e. that satisﬁes for all x ∈ X and every scalar α ∈ K the properties 1. z) + β(y. 3 and 4. (x. These operations satisfy (for all x. 7. y ∈ X and scalars α. 2. Property 3 is called the triangle inequality. a mapping X → R. x) ≥ 0. and (x. x + y ≤ x + y . x) = 0 ⇐⇒ x = 0. x + (y + z) = (x + y) + z. β ∈ K 1. where 1 ∈ K.Version: September 13. provided with an addition and a scalar multiplication over a scalar ﬁeld K. α(x + y) = αx + αy. 3. x = 0 ⇐⇒ x = 0. 4. 2. 3. An inner product space is a linear space X provided with a mapping (·. and . such that for all u ∈ X c u ≤ u ≤C u . 3. such that 0 + x = x + 0 = x.3. β ∈ K) the following properties 1. ∃(−x) ∈ X such that x + (−x) = 0. ∀x ∈ X. (αβ)x = α(βx). y ∈ X and scalars α. (αx + βy. z) = α(x. αx = α x . A normed linear space is a linear space X provided with a norm · . 2. (x. C > 0. but not necessarily 2. such that for all x. 5. The usual ﬁelds are the ﬁelds of the real (K = R) or the complex (K = C) numbers. x + y = y + x.1 Introduction to Functional Analysis Vectorspaces A linear space or vector space X is a nonempty set of elements. are called equivalent if there exist c. Two norms . 2004 49 3. (α + β)x = αx + βx. y) = (y. A seminorm is a similar mapping satisfying 1. ·) : X × X → K (the inner product or scalar product). . z). 6.
then P u ⊥ u − P u. .3. Then there is a space V ⊥ (the orthogonal complement) with V ⊥ V ⊥ such that each element e ∈ H can be uniquely written as e = u+v. ∀x ∈ X.3. Further (P v.1 A closed linear subspace Y of a Banach space X is also a Banach space.8) if and only if x and y are linearly dependent. We say that two elements x and y of an inner product space X are perpendicular to each other (notation x ⊥ y) if (x.3 Let V be a closed linear subspace of a Hilbert space H (V = H). y) = 0. Note that diﬀerent norms can cause diﬀerent closures and that a space can be a Banach space when equipped with one norm but will not be a Banach space in another norm.W.8) Equality holds in (3. Quite useful is the CauchySchwarz inequality: (x. Lemma 3. If P : H → V is the orthogonal projection on V . Such spaces for which X = X are called complete.2 Let V be a closed linear subspace of a Hilbert space H and V = H. A Banach space is a normed linear space which is closed in the given norm. The set X is separable if it contains a countable dense subset. then there exists an element e = 0 in H such that for all v ∈ V : (v. (3. where u ∈ V and v ∈ V ⊥ Corollary 3.3. w ∈ H. This means that every Cauchy sequence converges in the norm given. So we can discuss the convergence of Cauchy sequences in X and the closure X. Hemker Here z denotes the complex conjugate of z ∈ C. Corollary 3. With the norm we have introduced a metric or distance ρ(x. A Hilbert space is a inner product space which is closed in the norm induced by the inner product.50 P. w) = (v.4 Let V be a closed linear subspace of a Hilbert space H and let u ∈ H. or equivalently if it contains a dense sequence. Lemma 3. y) = x − y . e) = 0. x). P w) for all v. We conclude this section with some elementary results on Banach and Hilbert spaces. P 2 = P and P = 1. y) ≤ x y .3. A subset M ⊂ X is said to be dense in X if M = X. Every inner product space is a normed linear space because we can deﬁne a norm: x := (x.
C 0 (Ω) is the space of all continuous functions deﬁned on Ω and C m (Ω) = {u ∈ C 0 (Ω). Let Ω ⊂ Rd then we will denote the ith partial derivative Di u of the function u by: ∂ Di u = ( )u. See [27].9) We will now give a number of examples of function spaces as well as norms and inner products we can deﬁne on them. 2. . 1 A boundary Γ of a subset Ω of a ddimensional space is Lipschitz continuous if it is locally a suﬃciently smooth (d − 1)dimensional manifold. ∀ 0 ≤ α ≤ m}. + αd . . Before we give examples of function spaces we ﬁrst introduce the multiinteger or multiindex notation for higher derivatives of functions in more variables.10) in which α is a multiinteger. In eﬀect we can deﬁne an outward unit normal n a. As the elements of these spaces are not necessarily bounded functions we also introduce the space C m (Ω) = {u ∈ C m (Ω). . ∀ α ≤ m}. α≤m x∈Ω (3. Then the αth derivative of u is α α α D α u = D 1 1 D2 2 · · · D d d u = ( ∂ α1 ∂ α2 ∂ αd ) ( ) ···( ) u. d. We will also use the associated seminorms: uC m (Ω) = max sup D α u(x). ∂x1 ∂x2 ∂xd (3. 2004 51 Function spaces Functions on Ω (mappings Ω → R or Ω → C) can be considered as elements of linear spaces. This boundary is supposed to be Lipschitz continuous 1 . i = 1. . . . The boundary of Ω is denoted by Γ = ∂Ω.11) ∞ The space C0 (Ω) is the subset of all functions in C ∞ (Ω) with compact support: the function values and the derivatives vanish on the boundary. . α=m x∈Ω (3.Version: September 13. ∂xi Further we introduce a multiindex α = (α1 . . on Γ. αd ) ∈ Nd and we write α = α1 + α2 + . α2 . Such linear spaces are called function spaces. This is a Banach space with norm u C m (Ω) = max sup D α u(x). . Analogously m we deﬁne the subsets C0 (Ω) of C m (Ω).e. D α u are bounded and uniformly continuous on Ω. D α u ∈ C 0 (Ω). Here and in what follows we denote by Ω an open set Ω ⊂ Rd .
15) ≤ f Lp (Ω) · g Lq (Ω) . q be two real numbers such that q ≥ 1.3.5 Consider the function f : Rd → R deﬁned by f (x) = P. The H¨lder inequality (3. g ∈ Lq (Ω). such that up is Lebesgue integrable.3.}. Then f ∈ C0 (Rd ).6 C0 (Ω) ⊂ L2 (Ω) and C0 (Ω) is densely embedded in L2 (Ω). Lp (Ω). (3.3. := Ω u(x) dx p 1 p .14) ∞ ∞ Lemma 3.8). (3. page 33]. Proof: See [1]. o Lemma 3. Remark: ∞ As C0 (Ω) is densely embedded in L2 (Ω) we can say that L2 (Ω) is the completion ∞ of C0 (Ω) in the norm 1/2 u L2 (Ω) = Ω u(x)2 dx . Further we deﬁne the Lebesguespaces of integrable functions [30]. It is important to note the following inequality due to H¨lder. L2 (Ω) is a Hilbert space with inner product: (u. with norm f L∞ (Ω) := ess sup f  = inf{λ ∈ R  f (x) ≤ λ a.7 Let p.W. First. L1 (Ω) 1 q + 1 p = 1 and f ∈ Lp (Ω). 0 x ≥ 1.13) By Lp (Ω) with p = ∞ we denote the space L∞ (Ω) of essentially bounded functions. is the set of functions u on Ω. (3.12) As a particular case . x∈Ω (3. v)L2 (Ω) := Ω u(x)v(x)dx. Then f g ∈ L1 (Ω) and fg Proof: [30.e.15) is a generalisation of the CauchySchwarz inequalo ity (3. . 1 ≤ p < ∞.52 Example 3. p ≥ 1. On Lp (Ω) we deﬁne the (obvious) norm: u Lp (Ω) This function is arbitrarily often diﬀerentiable and its support is the unit ball ∞ Td = {x ∈ Rd  x ≤ 1}. Hemker 1 exp( x2 −1 ) x < 1.
v) = F (v. In fact every strictly positive. x X (3. So for all x. Proof: ε Assume T is bounded: T < ∞. We have y ∗ = δ/2 < δ.10 (1) A linear functional T is a linear operator T : X → K. A bilinear operator F : X1 × X2 → Y is called bounded with bound F if F := sup u. operators that map elements of a Banach space to the ﬁeld. A special class of operators are functionals. .e. i. v ∈ X and scalar λ ∈ K.v=0 F (u. So we obtain the following deﬁnitions. then choose δ < T . of real or complex numbers. A sesquilinear operator is a mapping F : X × X → Y that is linear in its ﬁrst argument and conjugatelinear in its second.3.3.16) A linear operator is bounded iﬀ T < ∞. 2004 53 Linear operators and functionals Let X and Y be normed linear spaces over a ﬁeld K. Let ε > 0. u). Remark: An inner product is an example of a sesquilinear operator.9 A bilinear operator F : X1 × X2 → Y is an operator that is linear in each of its arguments. K.8 A linear operator T is continuous iﬀ it is bounded. so Ty y < 2ε δ . Ty y < ε.Version: September 13. which means that T is Deﬁnition 3. so in 0 we have: ∀ε > 0 ∃δ > 0 : ∀x : x < δ ⇒ y δ T x < ε. So T is continuous. or T ≤ 2ε δ < ∞. The norm of a linear operator T is deﬁned by T = T Y ←X := sup x∈X. A linear operator T : X → Y is a mapping such that T (u + v) = T (u) + T (v) and T (λu) = λT (u) for all u. v) u X1 v Y X2 < ∞. symmetric. Deﬁnition 3. so δ T y∗ = 2 · bounded. A conjugatelinear operator S : X → Y is a ¯ mapping such that S(u + v) = S(u) + S(v) and S(λu) = λS(u). Lemma 3. Take an y and consider y ∗ = 2 · y . sesquilinear operator can be considered as an inner product. A bilinear operator F : X × X → C is called symmetric if F (u. x=0 Tx Y .3. y such that x − y < δ we have (because T is linear): T x − T y = T (x − y) ≤ ε T x − y < T · T = ε. Assume T is continuous.
so for every x the sequence xn (x) has a limit. Theorem 3. (3) A sesquilinear functional F is a sesquilinear operator F : X × X → K.15). . . ∞ • Let X = Lp (Ω). q ≥ 1 and 1 1 + = 1. say x (x). Proof: Let xn be a Cauchysequence in X ..17) is a bounded linear functional. then for every x0 ∈ Ω is T : f → T (f ) = f (x0 ) a bounded linear functional. • Equation (3. Hemker (2) A bilinear functional F is a bilinear operator F : X1 × X2 → K. Then f → T (f ) = f (x)w(x)dx Ω (3. > the duality paring between X and X . First the following elementary theorem.W. that is p. It is equipped with the norm x X := sup x∈X.18) We sometimes write < x . x n − xm X < . m ≥ N : or ∀ > 0 ∃N ≥ 0 ∀n. Examples: Functionals • Let X = C(Ω) and f ∈ X.12 The dual space X of a normed linear space X is a Banach space.54 P.3. m ≥ N : ∀x ∈ X xn (x) − xm (x) < . This means that ∀ > 0 ∃N ≥ 0 ∀n. We want to characterise the dual of some spaces. So.11 The space of all bounded linear functionals X → K is called the dual space X of X. x (x) x=0 x X (3. o Dual spaces Deﬁnition 3.17) also deﬁnes a bounded linear functional if w is an element in Lq (Ω) under condition that p and q are conjugate exponents. The function w is called the weighting function. x > instead of x (x) and we will call < . Now R and C are complete. f ∈ X and w ∈ C0 (Ω) a given function. p q This follows from H¨lders inequality (3.3. ∀ > 0 ∃N ≥ 0 ∀n ≥ N : ∀x ∈ X xn (x) − x (x) < .
so the mapping R : X → X of u to its representation Ru = z is a linear bijective isomorphism. Deﬁne y1 = y0 − P y0 . 2004 55 Now deﬁne the function x as the function that maps every x to the number x (x) then we see ∀ > 0 ∃N ≥ 0 ∀n ≥ N : xn − x X < . or rather that the Cauchysequence xn has limit x . z)X u(x) = sup = z x X x X x∈X X.3. Corollary 3.13 (The Riesz Representation Theorem) Let u : X → K be a bounded linear functional on a Hilbert space X.3. It says that the dual space can be identiﬁed with the space itself. Further 1 u(y1 ) = u(y0 ) − u(P y0 ) = u(y0 ) = 0. then z := x − u(x)y2 ∈ N (u). Normalisation yields y2 = u(y1 ) y1 . z) = 0 for all x implies z = 0. then there exists one and only one yu ∈ X such that 2 u(x) = (x. yu )X ∀x ∈ X. y2 ) y2 2 Uniqueness is clear. ⇐⇒ ⇐⇒ (x. To characterise the dual of a Hilbert space we have the following important theorem. yu ).19) Proof: Consider the null space N (u) = {xu(x) = 0}. this is a closed linear subspace of X.14 The norms of u ∈ X and z ∈ X with u(x) = (x. because u(z) = u(x) − u(x)u(y2 ) = 0. y2 ) = u(x)(y2 . Let P be the projection of X to N (u).Version: September 13. Theorem 3. 2 The converse is trivial: let uy (x) = (x. y2 ) y2 u(x) = = (x. then y1 ⊥ N (u). y2 ) (x. ) = (x. y). so that u(y2 ) = 1 u(y1 ) Take x ∈ X arbitrarily. . (y2 . Assuming u = 0 (or we could take yu = 0) there is a y0 such that u(y0 ) = 0. So ⇐⇒ (z. y2 ) = 0 (x − u(x)y2 . since (x. then uy ∈ X . z)X are identical: u X = sup x∈X (x. Due to continuity of u and completeness of X. Now deﬁne yu := y2 y2 2 . y2 ) = 0 u(y1 ) = 1 and still y2 ⊥ N (u). (3.
3.3.16 A sequence {un }n in X is called (strongly) convergent to u ∈ X if lim un − u X = 0. Theorem 3. Hemker The dual of a Lp space is a Lq space under some conditions.3.20) (2) A nonsymmetric bilinear functional F : X1 × X2 → K is called subcoercive if there exists a γ > 0 such that: ∀x ∈ X1 ∃z ∈ X2 . Strong convergence implies weak convergence. z) ≥ γ x and ∀z ∈ X2 . Proof: See [1. Deﬁnition 3. x) ≥ γ x 2 . z) > 0.17 (1) A symmetric bilinear (sesquilinear) functional F : X × X → K is called coercive (or strictly positive) if there exists a γ > 0 such that for all x ∈ X F (x.3. Remark: If X is a Hilbert space we have by the Riesz Representation Theorem 3.2 The Generalised LaxMilgram Theorem Deﬁnition 3. q ≥ 1 and p + 1 = 1. but the converse is not true (cf. v)X = (u. [22]).3. z = 0 ∃x ∈ X1 F (x. but L can not be associated with the dual of L1 . (3. (3. For contrast we also state the usual deﬁnition.3. as is shown by the following theorem. we call it the weak topology. n→∞ A sequence {un }n in X is called weakly convergent to u ∈ X if n→∞ lim ϕ(un ) = ϕ(u) ∀ϕ ∈ X . v)X .21) . q ∞ If p = 1 then q = ∞. As this deﬁnes the weakest topology which renders all elements of the dual continuous.13 that weak convergence is equivalent to n→∞ lim (un .22) X1 z X2 (3.41] With the notion of the dual in mind we can give a weaker deﬁnition for convergence.W. ∀v ∈ X. z = 0 F (x.56 P. page 40.15 (Riesz for Lp (Ω)) 1 Lq (Ω) is isometric isomorph to the dual space of Lp (Ω) if p.
x z (3.18 (Generalised LaxMilgram) Let X1 and X2 be two Hilbert spaces and let F : X1 × X2 → K be a bounded. the existence as well as the uniqueness and boundedness of solutions to elliptic PDEs. F (u0 . subcoercive. It discusses. z) ≡ 0. z) . (3. z)X2 . z = 0 F (·. Proof: We give the proof in six parts. z) ≥γ x z X2 X1 . Ru)X2 . in an abstract setting. v X2 v X2 v∈X2 . namely for the general. according to the Riesz Theorem for every u ∈ X1 a unique element z ∈ X2 exists such that F (u.22) is equivalent to ∀z ∈ X2 . ·) : X2 → K for each u ∈ X1 . due to the bilinearity of F .25) Further the problem is stable: small changes in the data f cause only small changes in the solution u0 : u0 X1 ≤ 1 f γ X2 .23) This means that we can ﬁnd a largest possible γ as: γ := inf sup F (x. (1) As F is bounded we have a bounded linear functional F (u. Now.24) x∈X1 z∈X2 We now come to the central theorem of this chapter: the LaxMilgram Theorem.22) of a subcoercive functional.26) (2) The operator R is linear. Condition (3. v) (Ru. Theorem 3. The treatment is such that we can easily generalise it for discrete equations later. then there exists one and only one u0 ∈ X1 such that ∀v ∈ X2 . (3) To prove that R is a surjection we ﬁrst prove that R(X1 ) is a closed subspace of X2 . ·) = (·. v) = sup . (3. (3.21) implies that ∃γ > 0 such that for all x ∈ X1 : sup z∈X2 F (x. We ﬁrst note that: Ru X2 = sup v∈X2 F (u. This deﬁnes an operator R : X1 → X2 such that F (u. We will state it in a more generalised version than usually.Version: September 13. asymmetric instead of only the symmetric case. Let f ∈ X2 . 2004 57 It is easy to see that a coercive functional also satisﬁes the properties (3. ·) = (·. Condition (3. bilinear functional.21) and (3. v) = f (v).3.
23): Ru = sup v F (u. ∀v ∈ X2 . (6) Finally we prove the stability of the problem. This is seen as follows.27) (c) Now we show that R(X1 ) is closed or that every accumulation point of R(X1 ) is an element of it.3 Distributions and Sobolev Spaces Solutions to (weak formulations of) elliptic PDEs are often found among special classes of functions: the so called Sobolev spaces. Because it involves weak formulations of PDEs we ﬁrst treat the subject of distributions. v0 ) = (·. So we obtain ∀v ∈ X2 ∃u ∈ X1 Ru = v which means ∃R−1 : X2 → X1 .58 P.3). v0 ) ≡ 0.25). v) = F . By (3. v (3. then it is the limit point of some Cauchysequence {Run }. . v0 ) = f (v). so {un } is a Cauchy sequence in X1 .27) we know Rum − Run X2 ≥ γ um − un . (4) We prove that the range of R actually is all of X2 . So let v be an accumulation point of R(X1 ). v) ≥γ u .25). by continuity of R. Hence it follows that R(X1 ) = R(X1 ) ⊂ X2 . so u0 is the solution of equation (3. or: 1 f . Because X1 is a Banach space ∃!u ∈ X1 . Clearly v = Ru. To do so we ﬁrst assume that R(X1 ) = X2 . Ru0 ) = F (u0 .3. If we take a righthandside of our equation f ∈ X2 . which contradicts the subcoercivity of F .W. Ru)X2 = 0 for all u ∈ X1 (by Corollary 3. then according to the Riesz theorem ∃!v0 (v.3. (5) In fact R−1 is the solution operator for the equation (3.27): γ u0 ≤ Ru0 = v0 = f . In this section we will discuss these spaces and some of the properties of their elements. If we deﬁne u0 = R−1 v0 then f (·) = (·. and so continuous: R X2 ←X1 = sup u∈X1 Ru X2 1 = sup u X1 u∈X1 u sup v∈X2 F (u.21) or (3. Hemker (a) An easy consequence is that R is bounded. v X2 (b) F is subcoercive. 3. This means there exists a v0 = 0 ∈ X2 such that v0 ⊥ R(X1 ). ·). u = lim un and so ∃!Ru ∈ X2 as limit of {Run }. Due to (3. so we get using (3. using (3. so v ∈ R(X1 ). This implies R(X1 ) = X2 . Now we have proved that R is surjective.26) F (·. u0 ≤ γ Uniqueness is a direct consequence of this stability. or (v0 .
19 A distribution (or generalised function) on Ω is a bounded linear functional on D(Ω). ϕ >. We say that a sequence ϕn ∈ D(Ω) converges to ϕ ∈ D(Ω). In this way we can identify each Lp function with a distribution from D (Ω). 2004 59 Distributions ∞ We deﬁne the function space D(Ω) as the set of functions C0 (Ω) supplied with the following topology. Example 3. Ω However. 3 In fact we can extend the class of such functions to the locally integrable functons L1 (Ω) = loc f  f ∈L1 (K) ∀compactK ⊂ Ω . Sometimes we write Tf (ϕ) =< f. Deﬁnition 3.Version: September 13. · > is the duality paring between D(Ω) and D (Ω). . being the dual of D(Ω). We write < δx0 . the latter is rather formal because it has nothing to do with an actual integration.3. Example 3. n→∞ if and only if there exists a compact subset K ⊂ Ω such that for every function ϕn we have supp(ϕn ) ⊂ K and such that for all α ≥ 0 D α ϕn → D α ϕ.3. in which < ·.3. notation lim ϕn = ϕ. ∞ So D(Ω) is the same set of functions as C0 (Ω). ϕ > or δx0 (x)ϕ(x)dx.20 (Generalised functions) A function f ∈ Lp (Ω) deﬁnes3 a distribution Tf : D(Ω) → R by: Tf (ϕ) = Ω f (x)ϕ(x)dx. This generalised function is called the Dirac delta function associated with point x0 and it is also called δx0 . ∀ϕ ∈ D(Ω).21 Let x0 ∈ Ω and deﬁne Tx0 : D(Ω) → R as Tx0 (ϕ) = ϕ(x0 ). The space of distributions is denoted by D (Ω). but it has more structure.
this is not an L2 function. for any ϕ ∈ D(Ω) TDf (ϕ) = Ω Df (x)ϕ(x)dx = − Ω f (x)Dϕ(x)dx = −Tf (Dϕ) = DTf (ϕ). i. for all (generalised) functions all generalised derivatives (always) exist.W.22 The distributional or generalised derivative D α T of a distribution T is deﬁned by D α T (ϕ) = (−1)α T (D α ϕ). Of course. which yields 0 if x < 0 and 1 if x ≥ 0. 0 This means DTH = Tδ0 . ∞ which is clearly bounded as ϕ ∈ C0 (Ω). then ∀ϕ ∈ D(Ω): D α Tn (ϕ) = (−1)α Tn (D α ϕ) → (−1)α T (D α ϕ) = D α T (ϕ). the generalised derivative of the Heavisidefunction is the Dirac deltafunction. However. This means that. Example 3. the generalised derivative of the (generalised) function is the generalised function of the derivative. Deﬁnition 3. ∀ϕ ∈ D(Ω). then we may consider the distribution Tf associated with f .60 P. . (if Ω = R) but it is a distribution: ∞ TH (ϕ) = R H(x)ϕ(x) dx = 0 ϕ(x) dx.e. If we have a diﬀerentiable f . Remark: The mapping D α : D (Ω) → D (Ω) is continuous because if n→∞ lim Tn = T in D (Ω).3.3. We compute the derivative DTH : DTH (ϕ) = −TH (Dϕ) = − ∞ R H(x)Dϕ(x)dx = − Dϕ(x)dx = ϕ(0) = Tδ0 (ϕ). say f ∈ C 1 (Ω). Now we see. Because of the analogy with weak convergence we will call this the weak topology. for diﬀerentiable functions. Hemker The topology on D (Ω) is deﬁned as follows. ∀ϕ ∈ D(Ω). T ∈ D (Ω) we say that n→∞ lim Tn = T when: n→∞ lim Tn (ϕ) = T (ϕ). and the distribution TDf associated with the derivative Df of f . With Tn .23 (Heaviside and deltafunction) We consider the Heaviside function H(x).
24 The Sobolev space W n. 2 . 1 ) 2 . 1]. α ≤ n are also elements of Lp (Ω): W n. for all ϕ ∈ D(Ω).2 (Ω) = (D α u. ∀ α ≤ n}. This is a Banach space if we provide it with the norm: Dα u p Lp (Ω) 1 p u n.3. (3.p (Ω) is that they are Lp functions uα . is the space of all Lp (Ω)functions u for which all distributional derivatives D α u. are elements of L2 (Ω). . (3. 1.26 Consider the basic hat function on the interval Ω = [0.p (Ω). we introduce the so called Sobolev spaces. b].Ω .Ω .Ω := u n.2. p ≥ 1. play an important role. This function is in L2 (Ω). v)W n. Example 3. We write u u n := u n. Now take the function du such that 1 du = ∂u . except for x = 0. For instance the Heaviside function on an interval Ω = [a.28) α≤n In case of p = 2 we denote the space W n. 1 x ∈ [ 2 . . 0 ≤ α ≤ n such that uα (x) ϕ(x) dx = (−1)α Ω Ω u(x) D α ϕ(x) dx.3. . Deﬁnition 3.2. .2 (Ω) as H n (Ω). 1 where we deﬁne it (arbitrarily) zero. Therefore.p. for all ϕ ∈ D(Ω). n.Ω := . To look if it is also in H 1 (Ω) we must ﬁnd an L2 (Ω)function du that is the derivative of u in distributional sense: Ω du(x) ϕ(x) dx = − Ω u(x) Dϕ(x) dx. together with their (generalised) derivatives. Supplied with the inner product (u.3. functions that.p (Ω) = {u ∈ Lp (Ω)D α u ∈ Lp (Ω). 2004 61 Sobolev spaces In the theory of elliptic PDEs. Example 3.29) α≤n these Sobolev spaces are Hilbert spaces. or even Another way of saying that a functional u is an element of W n. D α v)L2 (Ω) . −∞ < a < 0 < b < ∞ shows that u ∈ Lp (Ω) does not imply Du ∈ Lp (Ω).Version: September 13. This ∂x . n = 0.25 The deﬁnition is not trivial. 1]: u(x) = 2x 2 − 2x for for x ∈ [0.
1 2 1 du(x) ϕ(x) dx = Ω 0 1 2 du(x) ϕ(x) dx + 1 2 du(x) ϕ(x) dx 1 = 0 ∂u ϕ(x) dx + ∂x 1 2 ∂u ϕ(x) dx ∂x 1 2 = + = = − − x↑ 2 lim u(x) ϕ(x) − lim u(x)ϕ(x) 1 x↓0 − − u(x) 0 1 ∂ϕ (x) dx ∂x ∂ϕ (x) dx ∂x lim u(x)ϕ(x) − lim u(x) ϕ(x) 1 x↑1 x↓ 2 1 2 u(x) 1 2 u(x) 0 ∂ϕ (x) dx − ∂x 1 u(x) 1 2 ∂ϕ (x) dx ∂x u(x) Dϕ(x) dx. 1]). where 2 its (generalised) derivative is discontinuous. but for d = 2 they do not have to be and indeed it should be 4 An element u of a Sobolev space is continuous. The proof shows that it is essential that u is continuous in the point 1 . Ω where we have used the fact that 1 x↑ 2 lim u(x) = lim u(x) 1 x↓ 2 and that ϕ ∈ D([0.p (Ω) ⊂ C k (Ω).30) Example 3. so it is in L2 . if there is an equivalent function u such that this function has the property. Hemker function du is a simple function.3.p (Ω) is k times continuously diﬀerentiable. So for d = 1 the functions of H 1 (Ω) are continuous. . This raises the question under which conditions elements of Sobolev spaces are continuous or even continuously diﬀerentiable.3. page 174] (3. Theorem 3. bounded or whatever. Or: n > k + d/p ⇒ W n.27 (Sobolev’s Embedding Theorem) Let Ω ⊂ Rd be a bounded open region and let n > k + d/p.62 P.4 This is the content of the next theorem. then every element from W n.W. so ϕ(0) = ϕ(1) = 0. so it remains to check if this is the generalised derivative of u. Proof: [30.28 Note that the dimension d plays a role.
Ω .m (W0 (Ω)) ∼ W q. for a true subset Ω ⊂ R .Ω .2 (Ω) = H n (Ω) we write also W0 (Ω) = H0 (Ω). In 1 particular we use H −1 (Ω) := W 2.p. Assuming that C0 (Ω) is not dense we compute the orthogonal complement: 0 = (u.p (Ω) under the Sobolev norm.3. For n ≥ 1 however H0 (Ω) = n H (Ω).6 we see that H0 (Ω) = H 0 (Ω).31 n D(Ω) is dense in H0 (Ω).3. as is shown by Example 3.29. 0 n From Lemma 3. L2 . n. A non trivial example is the function e(ξ. 2 It is not continuous in (x.32 Let p = ∞ and 1 p + 1 q = 1. C ∞ (Ω) is dense in W n.3.3. ϕ)L2 + (Du. = Proof: [1. but not in H0 (Ω) (in terms of the trace operator in the next section: the trace on Γ is not zero).30 W0 (Ω) is the closure of C0 (Ω) in the norm n.p (Ω) as the closure of C ∞ (Ω) in the norm · n. In Corollary 3. y) = log  log(x2 + y 2 ) on B 1 (0. ϕ)L2 = (u − ∆u. consider H 1 (Ω). page 48] m For the Sobolev spaces with p = q = 2 we write H −m = (H0 ) .p (Ω).2 n analogy to W n. Unfortunately. ϕ)H 1 = (u. 2004 63 noted that the Sobolev’s inequality (3. ∞ So the orthogonal complement of C0 (Ω) consists of those u that satisfy u−∆u = 0 distributionally. m > 0. This implies that we might introduce W n. ϕ)L2 . . y) = (0. This 1 function is in H 1 (Ω).6. 0).3. Dϕ)L2 = (u.Version: September 13.x) . ϕ)L2 − (∆u. we have that C0 (Ω) is dense in d 2 ∞ L (Ω). the space C0 (Ω) is not dense ∞ in W n. For instance.−m (Ω). Example 3.−1 (Ω). the dual of H0 (Ω). then p. Theorem 3. We are interested in the dual spaces of the Sobolev spaces. with ξ = 1. We now give the following deﬁnition. according to Lemma 3. Luckily however. An example of this is f (x. 0).p ∞ Deﬁnition 3.3.3. · n. but it is not diﬃcult to prove that f ∂f ∂f 2 2 ∂x L and ∂y L do exist.29 ∞ Remember that. First a deﬁnition.p.30) is sharp: there are functions in H 1 (Ω) (d = 2) that are not continuous.
64 P. . We also deﬁne the trace operator γ1 : γ1 u := ∂u = ∂n d γ0 i=1 ∂u ∂xi ni . xd ).3. Lemma 3. 1 H0 (Ω) is exactly the kernel of γ0 : 1 H0 (Ω) = {u ∈ H 1 (Ω) : γ0 u = 0}. . 5 If Ω is a bounded area and has a Lipschitz continuous boundary then one can show that the linear mapping γ0 : C ∞ (Ω) → L2 (Γ) is bounded in the sense that γ0 u L2 (Γ) ≤C u H 1 (Ω) . page 113]. the restriction of f to Γ. So one can extend the operator γ0 to a bounded linear operator γ0 : H 1 (Ω) → L2 (Γ). . t. . bd ]. .4 The Poincar´Friedrichs Inequality.3. . e To prove the coercivity of a bilinear functional one can sometimes use the following lemma. . .31) Proof: It is suﬃcient to prove the statement for functions in D(Ω) as D(Ω) is 1 dense in H0 (Ω). (3. Hemker Trace operator For a function f ∈ C ∞ (Ω) we deﬁne the trace γ0 f as γ0 f = f Γ i. d: xi u(x) = ai 5 For ∂u (x1 . the boundary of the domain Ω. 3. . xi−1 . We enclose Ω in a rectangle S: Ω ⊂ S = [a1 . . xi+1 . . dt ∂xi more on trace operators we refer to [1.W. b1 ] × · · · × [ad .e. .33 (Poincar´’s Lemma) e 1 If Ω ⊂ Rd is a bounded and open set then for all v ∈ H0 (Ω) Ω v2 dx ≤ C(Ω) Ω  vdx. As u ∈ D(Ω) we have u ∈ D(S) and for every i = 1. . We extend u to S by setting u(x) = 0 whenever x ∈ S \ Ω.
. t. . . .Version: September 13. ∂xi So we get for every i = 1.)2 dx. t. . . Let u ∈ C 2 (Ω) be a solution of the Poisson problem with homogeneous Dirichlet boundary conditions −∆u = f in Ω. t.35) − Γ ϕ( u · n) ds. . Corollary 3.5 Variational formulations for diﬀerential equations A Dirichlet problem Let Ω ⊂ Rd be an open bounded area with Lipschitz continuous boundary Γ. t. . . . . . . . . . n. by partial integration.31) is also known as the Poincar´Friedrichs inequality. . .Ω is a norm over the space 1 H0 (Ω) equivalent to the norm · 1. .3. The inequality (3. 2004 so we get by the CauchySchwarzLemma (3. 2. .Ω .)2 dt ∂xi ≤ (bi − ai ) · Now integrate over S: b1 a1 bd ∂u (.)2 dt ∂xi ··· ad u(x)2 dx b1 bd bi ai ≤ (bi − ai ) = (bi − ai )2 a1 ··· b1 ad bd  ∂u (.33) u = 0 on Γ. . . . .32) (3. . u(x)2 dx ≤ (bi − ai )2 u(x)2 dx ≤ C(Ω)  i u 2 dx Ω Ω and so Ω Ω  u2 dx. then.34 When the set Ω is bounded.3. . . .8): u(x)2 xi 65 ≤ ai 12 dt · xi ai bi ai   ∂u (. the seminorm  · 1.)2 dx dt ∂xi a1 ··· ad  ∂u (. . e 3. for every ϕ ∈ C ∞ (Ω) Ω (3. ϕf dx = − = i Ω ϕ∆u dx = − iϕ Ω i u dx ϕ Ω i i i u dx (3.34) (3.
Now it is an immediate consequence of the generalised theorem of LaxMilgram 3. u) = Ω  u2 dx ≥ γ u 1.33 we get: e B(u.36) for which u 1 H0 (Ω) ≤ 1 f γ H −1 (Ω) . ﬁnd a u ∈ S such that B(u.32).32) 1 2 hasn’t: for instance when u ∈ H0 (Ω) but u ∈ C0 (Ω).18 that for every 1 f ∈ H −1 (Ω) there exists a unique solution u ∈ H0 (Ω) to (3. More general problems Let u ∈ C 2 (Ω) be a solution to −∆u + cu = f in Ω. For all ϕ ∈ D(Ω) we get ϕf dx = Ω i Ω iϕ iu P. we can ﬁnd this solution by solving (3. Hemker dx.32) has a solution.W.32) is also a solution of (3. Hence we will call the solution u ∈ H0 (Ω) the weak solution.32): ﬁnd a u with u = 0 on Γ such that iϕ Ω iu dx = Ω ∞ ϕf dx.36) is of the following form: given the bilinear form B : S ×V → R deﬁned by B(u. and from Poincar´’s Lemma 3. (3. ϕ) = i Ω iϕ iu dx and the linear form l : V → R.Ω . It is clear that there will be cases where (3.39) . 1 1 which tells us B : H0 (Ω) × H0 (Ω) → R is coercive.36). deﬁned by l(ϕ) = Ω ϕf dx. This leads to the following formulation of a problem similar to (3. Obviously we have 1 generalised the meaning of (3. un + βu = h on Γ.36). ϕ) = l(ϕ) for all ϕ ∈ V.36) we see that B : H0 (Ω) × H0 (Ω) → R is a symmetrical bilinear 1 form deﬁned on H0 (Ω).36) has a solution but (3. Equation (3.66 where n is the outward unit normal. (3.3. (3. ∀ϕ ∈ C0 (Ω). If (3.38) (3.36) We see that a solution of (3.37) 1 1 In our case (3.3.
32) this 1 functional is deﬁned on X = H0 (Ω). B(ϕ. ϕ) ≥ 0 and B(ϕ. if the solution to (3. 1. Γ ϕ2 ds ≥ C ϕ2 +  ϕ2 = C ϕ for some C > 0.32) and (3. for some C > 0. For problem (3. Symmetric problems and minimisation of quadratic functionals In the equations (3. β ≥ 0. for problem (3. From the LaxMilgram Theorem we conclude that there exists one and only one solution in H 1 (Ω).38) on X = H 1 (Ω). and also B(ϕ. ϕ) = Ω iϕ iu dx + Ω c ϕ u dx + Γ β ϕ u ds and l(ϕ) = Ω ϕ f dx + Γ ϕ h ds.38).Ω . where B(u. ϕ) can be considered as a quadratic form.39) exists. ∀ϕ ∈ V. v) is a symmetric. the functional J(ϕ) = B(ϕ. ϕ) = l(ϕ). coercive bilinear form. ∀ϕ ∈ H 1 (Ω). . ϕ) ≤ C ϕ2 +  ϕ2 = C ϕ ∀ϕ ∈ H 1 (Ω). So. Γ = Γ This leads to the problem: ﬁnd u ∈ S such that B(u.Version: September 13. The bilinear operator B : H 1 (Ω) × H 1 (Ω) → R is bounded and symmetric. except when c = 0 and β = 0: B(ϕ. 2004 then for every ϕ ∈ C ∞ (Ω): ϕf dx = Ω 67 − Ω ∆u ϕ dx + Ω iϕ iu iu iu Ω cϕ u dx cϕ u dx − Γ = = Ω dx + Ω ϕ Γ iu ni dx [ [ Ω iϕ iϕ + c ϕu ] dx − + c ϕu ] dx + ϕ(h − βu) ds βϕu ds − ϕh ds. we can ﬁnd it by solving the weak problem.Ω . ϕ) = = Ω Ω  ϕ2 dx + c Ω ϕ2 dx + β Γ ϕ2 ds  ϕ2 dx + c Ω ϕ2 dx + β 1. and it is coercive if c ≥ 0.38) with boundary condition (3. where B(u. ϕ) = 0 if and only if ϕ = 0.
ϕ) is not symmetric. on Γ. This again can be written in the form B(u. but in this case the form B(u. due to the ﬁrst order derivative. (3. To prove that the conditions of LaxMilgram are satisﬁed is a bit more hard to do.68 P. 2 Obviously. u + ϕ) − l(u + ϕ) 2 1 1 B(u.40) [ iϕ iu + ϕ bj ju + ϕ c u] dx − Γ ϕ(h − βu − αus )ds. . u) − l(u) + B(u. u) − l(u). ϕ) = l(ϕ) for all ϕ ∈ X as: ﬁnd u ∈ X which minimises the quadratic functional J(u) = Proof: J(u + ϕ) = = 1 B(u + ϕ. ϕ).W. ∀ϕ ∈ V. ϕ) = l(ϕ) we see that 1 J(u + ϕ) = J(u) + B(ϕ. Asymmetric problems A quite diﬀerent situation exists for problems with ﬁrst derivatives and/or more general boundary values: −∆u + bj j u + cu = f un + αus + βu = h Then for ϕ ∈ C ∞ (Ω) ϕf dx = Ω Ω on Ω.3. As we have seen in Theorem 3.3. Hemker Theorem 3. ϕ) ≥ J(u).1 we can get rid of the ﬁrst order terms by a transformation of the variables. We will see an example of this in the chapter concerning the convectiondiﬀusion equation. ϕ) − l(ϕ) + B(ϕ. ϕ) is a quadratic form. but for numerical purposes the transformation is of no use. there is one and only one u ∈ X that minimises J(·).35 Under the circumstances that B(ϕ. 2 For a u that satisﬁes B(u. we can also characterise the solution of the problem: ﬁnd u ∈ X such that B(u.5). ϕ) = l(ϕ).2. This helps us to prove existence and uniqueness (as in Section 3. 2 2 1 B(u.
then. uh ) = 0. vh − u) ≤ eh B = B(eh . ·) is a symmetric and coercive bilinear form and if S = V then we can show that this type of discretisation gives an approximate solution if Sh = Vh ⊂ S = V .4 The technique of the ﬁnite element method In this section we treat the practical aspects of the ﬁnite element method. uh − u). ∀ ϕh ∈ Vh . with dim(Sh ) = dim(Vh ). in the norm .42) is in Sh = Vh the optimal approximation of u.41) is the minimising function of a quadratic functional J(u) over S. We also may replace B(·. ϕh ) = l(ϕh ). B induced by B. The solution to (3. (3. the solution to (3. eh ) = B(eh . uh − u) = B(eh . ·)B = B(·.42) where Sh and Vh are ﬁnite dimensional subspaces of S and V . ·) with an approximation Bh (·. It is a Galerkin method applied to the variational problem: ﬁnd u ∈ S such that B(u. (3. If B(·. Theorem 3. Now because B(u. We give a simple error estimate and we discuss the choice of basic functions (in one or two dimensions) and the construction of the discrete equations for an example problem. vh ) = f (vh ) for all vh ∈ V . . The techniques explained here for the one and twodimensional case can readily be generalised to three space dimensions. vh ) = f (vh ) for all vh ∈ Vh . the solution of (3. The discrete function uh minimises the same functional over the subspace. vh ) = 0 for all vh ∈ Vh and because uh ∈ Sh = Vh we get B(eh . we have B(uh − u. We describe this in the following theorem. 2004 69 3. ∀ ϕ ∈ V. 3.41): eh Proof: First. and B(uh . u − vh B ∀vh ∈ Vh .1 Let (·. Usually we have a conforming ﬁnite element discretisation for which the discrete system of equations read: ﬁnd uh ∈ Sh such that B(uh . vh ) = B(eh . ·) and l(·) with an approximation lh (·).4.1 The principles The basic principle of the ﬁnite element method is the following. The consequences of this are discussed in Section 3. So B(eh .5. ϕ) = l(ϕ).41) in which S and V are (inﬁnitely dimensional) vector spaces. This implies that an optimal solution is found in the norm induced by the problem (the energy norm). We will also discuss the use of isoparametric elements for the case of curved boundaries.4. eh 2 B B ≤ inf vh ∈Sh u − vh B.Version: September 13. ·) be the inner product induced by the bilinear form B and let eh = uh − u be the error in the approximation.
We want to investigate the approximation in other norms (Lp and Sobolev norms) and for more general problems. Also dp ∈ L2 (Ω). 3. for each partition ΩN .3 Pk (ΩN ) ⊂ H 1 (Ω).4. (2) We deﬁne Pk (ΩN ) as the space of continuous. . . dp Take a polynomial p ∈ Pk (ΩN ). f ei ∈ Pk (ei ). so that t = 0 corresponds with x = xj−1 and t = 1 with x = xj . Deﬁnition 3. To this purpose we will ﬁrst consider some speciﬁc approximation spaces Vh and in the next chapter discuss in more detail the error estimates that are valid when we use these approximations. 2. Lemma 3. Deﬁne the distribution dx ∂p dp by dp(x) = ∂x (x). xj ] j = 1. . The case k = 1. Also Pk (Ω) ⊂ Pk (ΩN ).70 This gives eh B P. Further we deﬁne hj = xj − xj−1 and h = max{hj }.4. 3. .3. Hemker ≤ inf vh ∈Vh u − vh B. so p ∈ H 1 (Ω). < xN = b. N }. . the second is obvious. piecewise polynomial functions on Ω: Pk (ΩN ) = {f f ∈ C(Ω).26 (and using the continuity of p) that dp is the distributional derivative of p. N.2 Piecewise Lagrange Interpolation in one dimension We divide an interval Ω = [a. A polynomial in Pk (ei ) needs k + 1 (independent) values in order to be deﬁned correctly and uniquely.W. then we can prove in a way similar to Example 3. then Pk (Ω ) is the set of all polynomials of degree less or equal k on Ω . i = 1. For this purpose we ﬁrst deﬁne on the subinterval ei a new variable t = (x − xj−1 )/hj . . the meshwidth of the mesh ΩN . Now we want to determine the dimension and a basis for the space Pk (ΩN ) for k = 1. . . .4. x = xi and dx (xi ) = 0. . then p ∈ L2 (Ω). . We denote this partition by ΩN : a = x0 < x1 < . b] ⊂ R in N (not necessarily equal) subintervals ej = [xj−1 . Proof: We prove the ﬁrst.2 (1) Let Ω be a (connected) subinterval of Ω.
It turns out that a basis function ψi vanishes everywhere on Ω except on the two subintervals to which xi belongs. 3/2.1: Hat function ψj . We now 2 choose the basis functions {ψi } ⊂ P2 (ΩN ) as ψi (xj ) = δi. The number of basis functions is 2N + 1.Version: September 13. 1/2. N. The functions are not completely determined by their values on the points in the partition ΩN . .j i. We need one additional value in each interval. For each subinterval ej there are two nonzero basisfunctions: ψj−1 (x) = 1 − (x − xj−1 )/hj = (1 − t). 1.43) These are the so called standard hat functions. . (3. . 2 i integer. (2N − 1)/2. 1 ψj− 2 (x) = ψj (x) = In ﬁgure 3. In ﬁgure 3. j = 0. . On each segment ej there are at most three basis functions unequal to zero ψj−1 (x) = (1 − t)(1 − 2t). j = 0. . −t(1 − 2t).2 the functions ψj and ψj− 1 are shown. The case k = 2. are similar to ψj (with the exception for the two functions at the . for instance xj− 1 = (xj−1 + xj )/2. Again ψi = 0 on at most two segments ei . 2004 71 An element from P1 (ΩN ) is a piecewise linear function.1 the function ψi is given. All basisfunctions ψi . N A natural choice is the set {ψj }j=1 where ψj (xi ) = δij i. . N. Such a function is determined by its values at the points xi . An element from P2 (ΩN ) is a piecewise quadratic function. We take in each ej an interior point. . namely those to which xi belongs. So there are N + 1 basis functions. . ψj (x) = (x − xj−1 )/hj = t. 4t(1 − t). the midpoint of the interval. a x i b Figure 3.
Hemker x x j−1 j x j+1 x 0 x 1 x 2 x e j N Figure 3. . are translated versions of ψj− 2 . 1. On each interval ej we now have four nonzero functions: ψj−1 (x) = 1 − 2 (3t − 1)(3t − 2)(t − 1). We get the space P3 (ΩN ) of piecewise cubic functions. N. i integer. . These functions are shown in Figure 3. . 1/3.72 P. Now we have two extra degrees of freedom for each interval ej .j i. 2 and the 3 basis {ψi } is deﬁned by ψi (xj ) = δi. 2 The case k = 3. 1 2 (3t ψj− 2 (x) = 3 1 ψj− 2 (x) = − 2)t(t − 1). We can choose xj+k/3−1 = xj−1 + k hj for k = 1.3. boundary). j = 0. This can be treated completely analogous to the case k = 2. Figure 3. Obviously. 4/3. ψj (x) = − 1)(3t − 2)t.2: Second order Lagrange.3: Third order Lagrange. . . Again the basis functions corresponding to the interior points have their support in a single interval ej .W. 9 2 (3t 9 − 2 (3t − 1)t(t − 1). 2/3. the dimension of the approximation space P3 (ΩN ) is 3N + 1. Note that the support of ψj− 1 is completely contained in ej . All 2 1 functions ψi− 1 .
To this set we add one quadratic basisfunction. One possibility is to use a hierarchical basis. We consider the following problem −(a2 ux )x + a1 ux + a0 u = s in Ω.4. (3. and we take our weighting functions to be the same space. To apply the ﬁnite element method we ﬁrst have to formulate the problem in its the weak 1 form. Similarly the basis for k = 3 can be formed by adding one cubic basis function per interval to the (hierarchical) basis that was constructed for k = 2. This means that we start with the basis for the lower dimensional approximating space (smaller k or smaller N ) and that we extend this basis to a basis of the current approximating space. u=0 on Γ. v). The integration interval can be split into the subintervals ej . for each interval ej . k = 1. ek j k ek . v) + (a0 u. We get cj { a2 ψj ψi + a1 ψj ψi + a0 ψj ψi dx} = k sψi dx. a1 . Ω one equation for every weighting function ψi .Version: September 13. so Vh = Sh . 2. With Pk (ΩN ) = Span {ψi } we can write uh (x) = j cj ψj (x). Γ = ∂Ω and a2 . ∀ v ∈ H0 (Ω). 3.3 The construction of the discrete equations As an example of a ﬁnite element discretisation we discuss the discretisation of a twopoint boundaryvalue problem (TPBVP). 2004 73 Remark: For the same spaces Pk (ΩN ). We can take for instance ψj−1/2 . For instance.44) For our ﬁnite element discrete approximation we choose 1 Sh = {u ∈ Pk (ΩN )  u(a) = u(b) = 0} ⊂ H0 (Ω). where Ω = [a. v) = (s. b]. vx ) + (a1 ux . a0 and s are functions on Ω. so we seek an approximation for the function u ∈ H0 (Ω) that satisﬁes 1 (a2 ux . 3. we make such a basis for k = 2 by starting with the N + 1 basis functions already used as basis for the case k = 1. Here the prime denotes diﬀerentiation with respect to x. we can also choose diﬀerent basis functions. This yields the discrete system of kN + 1 equations cj j Ω a2 ψj ψi + a1 ψj ψi + a0 ψj ψi dx = sψi dx.
k.i.k. with mk. but also for two and threedimensional problems.i. This technique is practical not only for onedimensional.i. These contributions vanish if ek is not a subset of the support of both ψi and ψj .i. = ek = ek mk. sψi dx.i. The contribution of ek to (mij ) or (si ) is called the elementary stiﬀness matrix or the elementary load vector corresponding to this interval. For simplicity we take the coeﬃcients a2 . This means that the matrix and vector elements are not computed one by one. a1 and a0 constant.k.k.W. .j m0.74 P.j + a0 m0.i.k.45) the coeﬃcients cj are to be solved.j = a2 m2.j + a1 m1.45) in which mij si = k ek a2 ψj ψi + a1 ψj ψi + ψj ψi dx. ψj ψi dx. = k ek From equation (3. Hemker where we notice that the contributions from most of the subintervals vanish Thus the discrete equations are a linear system mij cj = si . to obtain the discrete solution cj ψj (x). Deﬁne m2.j then the element matrix entries are mij = k = ek ψj ψi dx. In practice the linear system is built up by scanning every element in the partition and adding the elementary stiﬀness matrix and elementary loadvector to the full stiﬀness matrix or loadvector.j .j . It is called the assembly of the stiﬀness matrix and load vector.i. uh (x) = j The matrix (mij ) is called the stiﬀness matrix and the vector (si ) the load vector. ψj ψi dx.k. Now we show explicitly how such a matrix/vector is drawn up for a twopoint boundaryvalue problem.i. j (3. (mij ) is composed of contributions from the separate intervals ek .j m1.
m )ψi (yk. all other elements vanish. 2004 Example 3. · · · . see a next section.m . Or rather.4 75 If we take piecewise linear basisfunctions as in (3. .k+ m −1 . . yk.m ) m=0 aware of the boundaries however.i.m ) = δi. (k. l as the nodes in ek : xk−1 = yk.43). we get mk.4.l = xk . m = 0.i. ψk−1 = − hk and ψk = hk . . (k − 1.m := xk+ m −1 .j = a2 hk 1 −1 −1 1 + a1 1 −2 1 −2 1 2 1 2 + a 0 hk 1 3 1 6 1 6 1 3 .1 < · · · < yk. Replacing the integral by a summation causes a new error.i.Version: September 13. N and i = ni + m j = nj + l j − 1 where ni . k) and (k.m )ψj (yk.j exactly. In most cases it is necessary to compute the integral using a quadrature rule: b a l f (x) dx ≈ wm f (xm ). l The elementary matrix entry becomes (k = 0. k). l) l mi l − 1. for which ψk−1 = 1 − t. · · · . .46) Note that for convenience we have given only the essential part of the matrix: the elements (k − 1. k − 1). (3. so that the accuracy of the quadrature may inﬂuence the accuracy of the discretisation. a1 and a0 are not (piecewise) constant. Assume that the approximation uses polynomials of degree l (so Vh = Pl (ΩN )) and take yk. 1 1 ψk = t. Only in special cases we can compute the integrals mk.j 6 Be ≈ wm a2 (yk. Use of quadrature rules The computation of the matrix entries is more laborious when the coeﬃcients a2 . An interesting quadrature in combination with Lagrangian interpolation is the method that uses the same nodal points for the integration.5. · · · . The requirements for such a quadrature rule are discussed in section 3. mj = 0. l so that ψi (yk. mi . m=0 where wm is the weight for quadrature node xm . nj = 0. 6 The elementary stiﬀness matrices for higher order piecewise polynomials can be computed in a similar way. We usually want to use quadrature rules that preserve the order of accuracy of the discretisation method. N .0 < yk. k − 1). . mk.
m )ψj (yk. Using the quadrature also for the elementary load vector. w0 = w2 = 6 .m )ψi (yk.j ≈ a2 (xk−1 ) + a2 (xk ) 2hk + 1 2 1 −1 −1 1 + hk 2 a0 (xk−1 ) 0 0 a0 (xk ) (3. which is diﬀerent from formula (3. We obtain 14 −16 2 a2. Example 3.47) + a1 (yk.W.m )ψj (yk. We observe that the zero order term contribution is a diagonal matrix. xj ] not equidistant.4.48) . For simplicity we choose a2 (x) piecewise constant on the partitioning ΩN . Example 3. Hemker + a1 (yk.76 P.m ) + a0 (ym )ψj (yk.mi )ψj (yni . To obtain the most accurate quadrature (and assuming that the endpoints of the interval are be nodal points) the nodes should be placed as for Lobatto quadrature.m ) +a1 (yk.k −16 32 −16 mk.m )δj.5 In case of piecewise linear approximation (l = 1) and the trapezoidal rule (w0 = w1 = 1 ).k+m/l−1 δi.m )δi.k+m/l−1 l = m=0 wm a2 (yk.i.m )ψi (yk.i.m )ψi (yk.m ) l = m=0 wm a2 (yk.m )ψj (yk. may determine the nodal points for the Lagrange interpolation. we obtain sk. These points.i ≈ hk 2 s(xk−1 ) s(xk ) . −a1 (xk−1 ) a1 (xk−1 ) −a1 (xk ) a1 (xk ) Remark: It is known that for an accurate quadrature we better take the nodes within each element [xj−1 .6 As a last example.j wmi .j ≈ 6hk 2 −16 14 + −3a1 (xk−1 ) 4a1 (xk−1 ) −a1 (xk−1 ) 1 1 1 0 4a1 (xk− 2 ) −4a1 (xk− 2 ) 6 a1 (xk ) −4a1 (xk ) 3a1 (xk ) (3. 2 w1 = 3 ).4.mi ) wmi + a0 (xi )δi.46) for constant coeﬃcients.m )ψj (yk. we get 2 mk. we give the construction of the element matrix based on 1 piecewise quadratic functions and the Simpson quadrature (l = 2.m ) (3.m )ψi (yk.k+m/l−1 + a0 (yk. then.49) .
45) is not a topic we will discuss here. Static condensation. 2004 a (x ) 0 0 hk 0 k−1 . Again the contribution of the zeroorder term is a diagonal matrix.k mk. The elementary loadvector follows after direct computation.k k 0 4 0 . using the same Simpson quadrature s(xk−1 ) h 1 4s(xk− 2 ) .k −4 0 4 + 6 −16 14 1 −4 3 1 0 0 a h (3.i.i ≈ (3. sk.j ≈ 6hk −16 2 For our twopoint boundaryvalue problem and piecewise second and higherorder functions the shape of the complete matrix and righthand vector is given in Figure 3.51) 6 s(xk ) constant this formula simpliﬁes to −16 2 −3 4 −1 a 32 −16 + 1. 0 4a0 (xk− 1 ) 0 2 6 0 0 a0 (xk ) 77 + For a1 and a0 also piecewise 14 a2.4. See for instance [6]. We however make some remarks. lumping The solution of the general matrix system (3. 6 0 0 1 Figure 3.50) + 0.4: Matrix and righthandside constructed by means of C 0 piecewise cubic polynomials. .Version: September 13.
The use of Lagrange interpolation has a nice advantage. The elimination of these unknowns is a strictly local process that can be done elementwise: it causes no changes in the contributions of other elements to the stiﬀness matrix and the load vector.or 3dimensional problems. We can reduce the set of equations in ﬁgure 3. Let a linear system be partitioned as A11 A21 A12 A22 x1 x2 = b1 b2 and let the second set of variables and equations be available for elimination.4 to a tridiagonal form by socalled static condensation. Hemker First notice. With this we mean the elimination of the unknowns belonging to the internal nodes xi . x 1 x 3 x 0 x 2 x 4 Figure 3. The technique in which these terms are placed on the main diagonal is called lumping 7 . that due to the relatively small support of the basis functions. we have a linear system that is sparse: only a fairly limited amount of the matrix elements is nonzero. for each basis function ψi (x) that has its support completely in a single element ek .W. then after elimination we get the system (A11 − A12 A−1 A21 )x1 = b1 − A12 A−1 b2 . We have seen that the zero order (not diﬀerentiated) terms consist only of diagonal elements if we use a quadrature. for which the basis functions ψi have a support that is completely contained in a single element ek .5: Matrix and righthand side after lumping. .78 P. 22 22 so there remains just a number of additional contributions in the matrix and righthandside.5. When all interior variables are eliminated in this way we get the tridiagonal system of equations shown in ﬁgure 3. This technique is interesting 7 Originally the oﬀdiagonal elements where added to the main diagonal without any explanation. The same technique can be applied for 2.
∀ϕi . ϕi ). then a semidiscretisation with Sh = Vh = Span {ϕi } yields the following set of equations: ﬁnd ϕh (x.Version: September 13.52) (3. First we handle inhomogeneous Dirichlet boundary conditions. ·). By using the technique “lumping” we can recast it to a diagonal matrix (ϕj . v) + (a0 u. In this section we treat also inhomogeneous and more general boundary conditions. u = g on Γ. dt j The matrix (ϕj . As an example we may consider a parabolic equation of the form ϕt = Lϕ. v). ϕi ) = (Lϕh . ϕi ). ϕi ) = wi δij so that we obtain an explicit set of ordinary diﬀerential equations for the coeﬃcients {ci (t)}: dci (t) 1 = (Lϕh (t. v) := (a2 ux . ϕi ) is called the mass matrix. ϕi ). v) = (s. ·). ϕi ) = (Lϕh (t. For an initialboundaryvalue problem (IBVP). (3. where L is a diﬀerential operator in the space variables (for instance Lϕ = pϕxx + qϕx + rϕ + s). t) = j cj (t)ϕj (x) such that ( ∂ ϕh . The part a0 (x)u(x) = s(x) of the equation is discretised pointwise. ∂t So we have to compute the set {cj (t)} for which dcj (t) (ϕj . we can use lumping together with the method of semidiscretisation to reduce the IBVP to an explicit set of ordinary differential equations. vx ) + (a1 ux . when there is time dependence in the problem. ∀v ∈ H0 (Ω).53) The weak form of this problem can be written as: ﬁnd a u ∈ H 1 (Ω) such that u = g on Γ and 1 B(u. dt wi Treatment of boundary conditions Until now we have limited ourselves to problems with homogeneous Dirichlet boundary conditions. 2004 for two reasons: 79 1. b] and the equation Lu := −(a2 ux )x + a1 ux + a0 u = s on Ω. 2. ∀ϕi . . Consider Ω = [a. ∀ i. This can be preferable when a0 (x) or s(x) is strongly varying or dominating.
80
P.W. Hemker
For theoretical purposes, this problem is easily reduced to the problem with homogeneous boundary conditions by choosing a function g ∈ H 1 (Ω) (the extension of the g deﬁned on Γ) that satisﬁes the boundary conditions 8 . Deﬁning 1 the function w := u − g we have w ∈ H0 (Ω) and w satisﬁes the equation Lw = s − Lg, or, Thus the problem is reduced to the problem with homogeneous boundary conditions. This gives existence and uniqueness. In practice it is also easy to implement nonhomogeneous Dirichlet boundary conditions. Because the value of the function u at the boundary points is given we have to compute less unknowns. This means that there are no discrete equations needed for the boundary points. We can handle Dirichlet boundary conditions in two ways. To show this, we ﬁrst assume the equations are constructed elementwise as shown in the previous section (ﬁgure 3.4). The two approaches are as follows. I) Replace the equations for the boundary points by equations that force the boundary conditions. For example, in the stiﬀness matrix put 1 on the main diagonal and 0 at the other entries of this row, and put the boundary value in the righthandside. This forces the boundary value to be part of the solution. In this way we get a linear system of equations of the original form, of which the size corresponds with the total number of nodes. II) In the other approach (see Figure 3.6) we eliminate the known “unknowns” on the boundary. The set of equations is reduced, but more important is that the righthandside has to be adjusted at a number of places. Now the size of the system corresponds with the interior nodes. In practice, this is less convenient, in particular if we have a combination of Dirichlet and Neumann conditions. In this case, however, a symmetric operator B(u, v) guarantees a symmetric stiﬀness matrix. If we consider more general boundary conditions than Dirichlet boundary conditions, we see that the whole procedure can be carried out in the same manner as before. In the discrete operator and righthandside additional terms will appear due to the boundary condition. See for instance the B(u, ϕ) and the l(ϕ) of section 3.3.5. Every boundary integral gives a contribution to the elementary matrix/righthandside at each boundary point. In practice we ﬁrst add all contributions from the interior elements to the stiﬀnessmatrix and later the contributions from the boundary segments are added. Also these boundary integrals can be computed by quadrature.
8 That we can actually ﬁnd such a function puts an extra condition on the boundary condition and on the smoothness of the boundary Γ. With γ0 the trace operator we have the requirement that γ0 u = γ0 g.
B(w, v) = (s, v) − B(g, v).
Version: September 13, 2004
A 0 0 0 u 0
81
a a a 3 1 2
...−a u 1 0 ...−a u 2 0 ...−a u 3 0
...−a u 1 3 ...−a u 2 3 ...−a u 3 3 u ...
3
Figure 3.6: Treatment of boundary conditions.
3.4.4
Other piecewise polynomial approximations in one dimension
In every ﬁnite element method we use an approximation of the solution of the form cj ϕi (x), uh (x) =
j
where the use of continuous, piecewise polynomial basis functions ϕi (x) with a small support is preferred. Not only Lagrange interpolation but also Hermite interpolation can be used. Here the functions ϕi (x) are not only determined by their function values at the nodes, but also by the values of the derivatives. The approximating basis functions thus may have larger smoothness, which gives us the opportunity to solve problems which contain higher order derivatives. We k may ﬁnd solutions in H0 (Ω) for k > 1. For instance it allows the solution of the homogeneous Dirichlet problem for the biharmonic operator ∆2 : ∆2 u = f u = ∂n u = 0 in on Ω, Γ.
A polynomial of degree k needs k + 1 (independent) values on an interval to be deﬁned correctly. Deﬁnition 3.4.7 Let f be a function in C m−1 ([a, b]), then the Hermite interpolate of f on a partition ΩN is the function p ∈ P2m−1 (ΩN ) ∩ C m−1 ([a, b]) for which, for all xi ∈ ΩN and k = 0, 1, . . . , m − 1, D k p(xi ) = D k f (xi ). Note that the Hermitian interpolation of a polynomial q of degree 2m − 1 is the polynomial itself. Moreover it is not just an element of C m−1 ([a, b]), but even from C 2m−1 ([a, b]).
82
P.W. Hemker
A usual Hermitebasis for a ﬁnite element method in one space dimension is that of the cubic Hermite splines (m=2): on every interval [xi , xi+1 ] the P3 (∆)function is determined by four parameters f (xi ), f (xi ), f (xi+1 ) and f (xi+1 ). For every element ej = [xj , xj+1 ] there are four nonzero basisfunctions: ϕ0,j , ϕ1,j , ϕ0,j+1 and ϕ1,j+1 such that D k ϕl,j (xi ) = δkl δij . Let t = (x − xj )/(xj+1 − xj ) then the form of these functions is on element ej ϕ0,j (t) ϕ0,j+1 (t) ϕ1,j (t) ϕ1,j+1 (t) = = = = (t − 1)2 (2t + 1), t2 (3 − 2t), t(t − 1)2 (xj+1 − xj ), t2 (t − 1)(xj+1 − xj ).
Extending these functions to the entire interval [a, b] such that D k ϕl,j (xi ) = δkl δij we get functions ϕj,k (x) ∈ H 2 ([a, b]), because the derivatives are continuous. Except for the functions that are associated with the boundary points, all functions ϕl,j have a support that extends over two subintervals ek . The smoothness of the approximating function uh (x) implies that it can be used for solving fourth order elliptic problems like auxxxx − buxx + cu = f, u(a) = u (a) = u(b) = u (b) = 0,
2 of which the corresponding variational formulation is as follows: ﬁnd u ∈ H0 (Ω) such that a uxx ϕxx + b ux ϕx + cu ϕ dx = f ϕ dx, Ω 2 for all ϕ ∈ H0 (Ω). This problem can not be solved by Lagrangian interpolation because the space spanned by the Lagrange interpolation polynomials {ϕi } is not in H 2 (Ω). The discrete set of equations for the two point boundary value problem discretised with Hermite interpolation polynomials has a blocktridiagonal form (see Figure 3.7).
(3.54)
Exercise 3.4.8 Consider the fourth order problem (Ω = [0, 1]) uxxxx − 2uxx + u = 1, u(0) = u(1) = ux (0) = ux (1) = 0. Prove that the analytic solution is u(x) = 1 + ex (c1 + c2 x) + e−x (c3 + c4 x), (2e − 1)(e2 − 2e − 1) c1 = − , e4 − 6e2 + 1 (e − 1)(e2 − 2e − 1) c2 = , e4 − 6e2 + 1 c3 = −1 − c1 , c4 = c 3 − c 1 − c 2 .
Make a uniform discretisation with h = 1/N .4. The most important which remain are shown in Figure 3. Such functions have a single degree of freedom for each interval and they have a maximum smoothness for a polynomial of degree m. 3. . a circle that we also take the value of the derivative. b])functions.7: Blocktridiagonal matrix. In other words: the use of splines causes a nonsparse system of linear equations (although in the onedimensional case it will be a band system). 20. 2004 X X X X X X X X X X X X X X X X X X X X 83 X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X Figure 3. the support of these basic spline stretches over many adjacent elements.5 Approximation in two dimensions In the introduction to discretisation methods we have seen some possible representations of a twodimensional function. 40. etc. in general. These are Pm (∆)∩C m−1 ([a.8. A dot indicates that at this point we take the function value. compute and solve the linear system. because it means that many nonzero elements will appear in each row of the stiﬀness matrix. two circles the value of two derivatives. Not all of them are useful for the solution of second order elliptic problems by a (conforming) ﬁnite element method. This gives discrete approximations for u(xi ) and u (xi ).Version: September 13. However. Other piecewise polynomials that might be used are for instance the splines. Compute max u(xi ) − uh (xi ) i and max u (xi ) − uh (xi ). i for N = 10. Compare the order of convergence you observe with the order you expected. Application for a ﬁnite element method is therefore unusual.
r = 0 ⇒ x = 0 and s = 0 ⇒ y = 0. Φ5 (x.55) . t) = 4 r s. Among these are the piecewise Lagrange interpolation polynomials on a triangulation. Which give the three boundaries of the triangle.9 we ﬁnd basis functions for a P2 (∆) Lagrange approximation by Φ1 (x. (1. Φ6 (x. 1 2x (x − 2 ). Φ3 (r. 1)}. (0. t) deﬁned by r = x. Φ5 (r. y) = Φ3 (x. 9 Then we get Φ1 (r. 9 So: (3. t) = r (2r − 1). Φ2 (r. 2). 1 2y (y − 2 ). (1. t = 0 ⇒ x + y = 1. s. s. y) = 1 2 (x + y − 1) (x + y − 2 ). s = y and t = 1 − x − y. s. p(x. t) = s (2s − 1). s. Example 3. t) = 4 r t. (0. 0).9 On the standard triangle shown in ﬁgure 3.4. y) = −4y (x + y − 1). 0). Φ6 (r. 0). t) = 4 s t. y) = Φ4 (x. (2. Φ4 (r. s. 1). t) = t (2t − 1). we may use barycentric coordinates (r. y) = α≤2 This is only possible for the set of values α = (α1 . To express the symmetry in this formulas more clearly. Notice that the basisfunctions deﬁned by ϕi ∈ Pk (ΩN ) (functions with α ≤ k) and ϕi (xj ) = δij are indeed continuous. y) = −4x (x + y − 1). s.W. That the dimension of P2 (∆) is indeed 6 can be easily seen because any element is of the form a α xα 1 y α 2 . y) = Φ2 (x. s. α2 ) ∈ {(0. 4x y. Hemker Figure 3.84 P.8: Some representations of a twodimensional function.
If we want to ﬁnd a piecewise polynomial in C 1 (Ω).Version: September 13. but are not in C 1 (Ω). For a parallelepiped one usually takes polynomials of the form Pk (x) · Pk (y) (so the tensor product of the two spaces) and for the nodes one takes the Cartesian product of the onedimensional partitions. Having four degrees of freedom.0) Figure 3. If. a number of nodes is distributed in a regular way we will not be able to construct exactly all kth order polynomials. on triangles. 71].9: Lagrange for a triangle. Example 3. A bilinear basis function is of the form a + bx + cy + dxy. One of the more simple ones is a P5 (Ω)function which has 21 degrees of freedom for each triangle (the Argyris triangle.4. or. These functions are also used for isoparametrical elements. then we have to use a relatively intricate construction. on a rectangle or a parallelepiped.0) (1. p. The number of degrees of freedom for a general kth order polynomial in 2 or 3 dimensions is exactly the number of points that can be distributed nicely over a triangle or tetrahedron. Note that in two dimensions. such a bilinear function is completely determined by the function values at the vertices of the rectangle. The ﬁrst form shows clearly that such a function has four degrees of freedom on each rectangle and the second form shows that the space of bilinear functions is the product of the two spaces of linear functions. the piecewise Hermite interpolation polynomials are continuous.10).1) x (0. cf [4. 2004 y 85 (0.10 We consider the bilinear interpolation on a rectangle. which we shall discuss in the following section. see Figure 3. Note that the function u along a boundary is a 5th order function determined by its function value and ﬁrst and second order derivatives at the vertices and ∂u the value of ∂n (n is the outward unit normal) at the edges midpoints. The construction of the discrete equations and the treatment of the boundary . (a1 + a2 x)(b1 + b2 y).
y)(x. . the solution of a problem on polygonal areas can be performed in a relatively simple way. .5 and Φi the functions deﬁned in (3. and basis functions can be . On the isoparametric element we now have a parametrisation in r. .86 u u ux uy u xx u xy u yy u n P. Examples of this are: for k = 1: the vertices of a triangle. k = 2: the vertices and the midpoints of the edges. y)T = i=1 zi Φi (r. Then v is called a kunisolvent set if for every sequence of real numbers (α1 . then we can eliminate the corresponding variables by static condensation. z6 } be a 2unisolvent set in R2 . z2 . .55). t). However when there is a curved boundary we can approximate it by means of isoparametric elements. s. Examples for k > 2 are easily found.11 Let v = {z1 .4. Hemker u un Figure 3. s and t are the barycentric coordinates as introduced in Section 3. s. When basis functions φj (x) have a support that is completely contained in a single element.4. To handle the boundary conditions we have to compute boundary integrals. where r. An isoparametric element of degree k = 2 is deﬁned as follows.10: The Argyris triangle. s. To deﬁne these elements we ﬁrst introduce the following Deﬁnition 3. Nk . . . . . . zNk } be a set of Nk points in R2 . . i = 1. with r + s + t = 1. .6 Isoparametric elements By triangulation of the domain. Let v = {z1 . conditions is in principle similar to the onedimensional case: the matrix and the righthandside will be build up elementwise. r + s + t = 1}. t. . 0 ≤ r. When computing the elementary matrices we again use a quadrature and also the combination of the elementary basisfunctions for Lagrange interpolation and an associated quadrature can lead to lumping. 3. t ≤ 1. . . . αNk ) there is precisely one polynomial of degree ≤ k such that Pk (zi ) = αi .4.W. The matching isoparametric element is 6 I∆ = {(x. .
In the following theorem we show that the discrete solution obtained in this way is “quasioptimal”. This means that. but also to the associated discrete problems. which is bounded by the righthandside of the equation. This will give us estimates in the Sobolev norms .3. t). 3. y) respectively) are both parametrised by the same type of functions. preceded by a necessary discussion of the formalism of the ﬁnite element method. It gives the uniqueness of the solution to the discrete equation and it gives a ﬁrst estimate of the error.18 can be applied not only to continuous problems. We have a more general case when we don’t have such subspaces or when the operators in the variational equation are replaced by approximations (for instance by quadrature).5 Error estimates for the ﬁnite element method In this section we derive error estimates for the ﬁnite element method. We also discuss pointwise convergence and superconvergence. The dependent and independent variables (u and (x. This is the reason to call these elements isoparametric elements. m. We give also error estimates for this case.1 The discrete version of the Generalised LaxMilgram In Section 3. We see that with kth order isoparametric elements the boundary is approximated by piecewise kth order polynomials. but for this we need additional requirements on the problem we consider. 2004 deﬁned on this element like it was done in case of triangular elements 6 87 uh (x)I∆ = i=1 ai Φi (r.q.5. it is (in the Snorm) not worse than the best approximation of the solution u ∈ S in the discrete space Sh .e the phenomenon that at some particular points the accuracy of the approximation is better than the global error estimates indicate. Finally we study the inﬂuence of the use of particular quadrature rules on Banach space interpolation and on superconvergence. The error estimates depend on how good we can interpolate elements of Banach spaces in subspaces of these Banach spaces. This theorem is in fact only applicable when we use ﬁnite dimensional subspaces of our original Hilbert spaces.Ω . i. This Generalised LaxMilgram Theorem 3. with trial functions in Sh ⊂ S and test functions in Vh ⊂ V . We will also give an estimate in the L2 norm. . s. 3. so we have to discuss the interpolation theory in Banach spaces.2 we saw that under certain conditions.Version: September 13.3. First we will give the discrete version of the Generalised LaxMilgram theorem. apart from a constant factor. Taking these discrete spaces as subspaces of the continuous spaces we obtain a so called conforming ﬁnite element methods.a variational problem has a solution.
such that 1 2 ∀xh ∈ Xh ∃zh ∈ Xh . so that 2 uh = S −1 Ph Ru0 . and Finally. zh ) > 0. for all vh ∈ Xh .18 (Ru. 2 2 1 ii) For every xh ∈ Xh we have Sx1 = Ph Rx1 because.5. With v0 ∈ X 2 such that f (v) = 2 (v0 . 1 Then there exists a unique uh ∈ Xh . −1 Then R = F and S −1 ≤ γh . So. Ph v) = 0 for all v ∈ X 2 . v) = f (v) for all v ∈ X 2 . vh ∈ Xh . −1 −1 From Lax Milgram we know that u0 = R v0 and uh = S vh . Ph vh ) = (Rx1 .3. vh )X 2 = F (uh . vh ) = (Rx1 . i.3. for all uh ∈ Xh . F (u0 .1 (Discrete Generalised LaxMilgram) Let the requirements of the Generalised LaxMilgram Theorem 3. Hemker Theorem 3. vh ) = f (vh ) for all vh ∈ Xh .4 2 2 2 2 gives that Ph (v0 − vh ) = 0. 1 2 1 2 and let S : Xh → Xh be the analogue mapping for Xh and Xh . h h 2 2 (Ph Rx1 . v)X 2 = F (u. vh ). and the following error estimate holds uh − u0 X1 ≤ 1+ F γh 1 xh ∈Xh inf u0 − xh X1 . vh ) h h h 1 = F (xh . 1 1 1 2 2 Let Ph : X → Xh and Ph : X 2 → Xh be orthogonal projections 1 Ph 1 Xh X1 ↓ −→ −→ X 2 ↓ S 2 Xh R 2 Ph . so vh = Ph v0 . vh ). i) Consider point (5) in the GLM theorem. v)X 2 for all 2 2 v ∈ Xh (Riesz) we have (v0 − vh . v).88 P. zh ) ≥ γh xh zh . (3. 2 1 (Suh . vh ) = (Sx1 . let X 1 and X 2 be two Hilbert spaces and let F : X 1 × X 2 → R be 1 a bilinear functional that is bounded and subcoercive.18 be satisﬁed.e. zh = 0 ∃xh ∈ Xh F (xh . for all u ∈ X 1 . so Corollary 3. zh = 0 F (xh .W. v)X 2 for all v ∈ X 2 (Riesz) and vh ∈ Xh such that f (v) = (vh . Further. the solution of the discrete problem 2 F (uh .. It follows that Ph v0 = Ph vh = vh . h .56) Proof: Let R : X 1 → X 2 be the Riesz mapping as used in the proof of the Generalised LaxMilgram Theorem 3. and u0 ∈ X 1 be the solution of 2 1 ∀zh ∈ Xh .3. v ∈ X 2 . let Xh ⊂ X 1 2 2 and Xh ⊂ X be linear subspaces. let f ∈ (X 2 ) .
For an elliptic problem of order 2m this will usually be the H m (Ω)norm.3. Hence uh − u0 From this easily follows uh − u0 X1 1 1 ≤ u h − P h u0 + P h u0 − u 0 1 ≤ {1 + S −1 R } u0 − Ph u0 . Measuring the error in other norms (and especially the L2 norm) requires a further analysis of the problem. Xh ) = inf xh ∈Xh u − xh between a function u ∈ X 1 and a 1 1 1 subspace Xh ⊂ X . 2 In (∗) we used that Ph = 1 by Corollary 3. 1 h→0 xh ∈Xh lim inf u − xh = 0. we have as an immediate consequence that u − uh = O(hk ) We then will call k the order of convergence.4. for 1 each u ∈ X . First we look at 1 u h − P h u0 = ≤ = ≤ 2 1 S −1 Ph Ru0 − Ph u0 S −1 S −1 S −1 (∗) 2 1 Ph Ru0 − SPh u0 2 2 1 Ph Ru0 − Ph RPh u0 2 1 P h R u 0 − P h u0 = S −1 R 1 u 0 − P h u0 . ≤ {1 + F /γh } inf 1 xh ∈Xh u0 − xh X1 . 2004 89 We want to know how well the discrete solution approximates the solution of the continuous problem. if we can prove that d(u. then for a conforming ﬁnite element method a suﬃcient condition for 1 convergence is the existence of a family of subspaces (Xh ) of X 1 such that. Corollary 3. The norm in which the error is measured is more or less induced by the problem: the conditions on coercivity and continuity are stated in the X 1 norm.2 Given a problem that satisﬁes the conditions of the Generalised Lax Milgram Theorem. . so we are interested in uh − u0 . Xh ) = O(hk ). Thus.Version: September 13.5. This will be discussed later. So the problem we were facing is reduced to an approximation problem: evaluate 1 1 the distance d(u. for some k > 0.
v) ≤ Mh s Let Bh : Sh × Vh → R be coercive: ∃αh > 0 ∀sh ∈ Sh ∃vh ∈ Vh αh sh Then the following estimate is valid uh − u∗ S∗ S∗ S∗ v V ∗.58). ∀v ∈ V. vh ) − fh (vh ) sup . We consider a solution u∗ ∈ S ⊂ S ∗ (S ∗ a normed linear space) of the problem: ﬁnd u ∈ S such that B(u. V ∗ . wh ) Mh u∗ − sh wh + Bh (u∗ . Let Bh : S ∗ × V ∗ → R be bounded with B = Mh .59) Proof: In the proof norms are either . v) = f (v). vh ) = fh (vh ).5. wh ) + Bh (uh .57) Let Bh : S ∗ × V ∗ → R and fh : V ∗ → R.57): ﬁnd uh ∈ Sh such that Bh (uh . What can we say about uh − u∗ S ∗ ? Theorem 3. (3. Vh ⊂ V . We can approximate both the bilinear form B and the linear form f by discrete versions Bh and fh . wh ) − fh (wh ). αh vh ∈Vh vh V ∗ (3. so that ∃Mh > 0 ∀s ∈ S ∗ . Hemker 3. wh ) + Bh (uh . Because Bh is coercive we know that for each sh ∈ Sh there exists a wh ∈ Vh such that αh u h − s h wh ≤ = ≤ ≤ Bh (uh − sh . vh V∗ ≤ Bh (sh . ∀vh ∈ Vh . We also consider the discrete problem associated with (3. v ∈ V ∗ Bh (s. wh ) Bh (u∗ − sh .5. Note that we have assumed no relation between B and Bh or f and fh .90 P.2 More general error estimate In this section we study an error estimate for more general problems. wh ) − Bh (u∗ . . ≤ + 1+ Mh αh sh ∈Sh inf u∗ − s h S∗ 1 Bh (u∗ .58) in which Sh ⊂ S ∗ and Vh ⊂ V ∗ . wh ) Bh (u∗ − sh . This means we have to expect a supplementary consistency error in our estimates. S ∗ or . but not necessarily Sh ⊂ S. wh ) − Bh (u∗ .3 Assume there exists a uh ∈ Sh that satisﬁes equation (3. (3. vh ).W. These discrete operators will work on some spaces Vh and Sh that are now not necessarily contained in V and S. The estimate thus obtained will also be valid for nonconforming ﬁnite element methods.
6. we also have to study the approximation of bilinear and linear operators. 2. i. wh ) − f (wh ) = 0. vh ) = B(uh . the conditioning of the discrete problem: Mh /αh . ∀vh ∈ Vh . vh ) − fh (vh ). apart from studying interpolation theory for Banach spaces. • If the discrete spaces are embedded in continuous spaces .5. However we can expect the consistency error to be rather big if we choose a u∗ that is not a solution. i. A diﬀerence is the fact that for the present estimate we have assumed the existence of the solution u∗ . if Bh (uh . then Bh (u∗ .5. the interpolation error: inf sh u∗ − sh S∗ .e. wh ) − fh (wh ) = B(u∗ . Remarks: • In the proof of the theorem we have not used the fact that u∗ is a solution to (3.57). vh ) and fh (vh ) = f (vh ).Version: September 13. ∀wh ∈ Vh . the consistency error: Bh (u∗ . αh αh wh u h − s h + s h − u∗ 1 Bh (u∗ . if Sh ⊂ S and Vh ⊂ V . and if Bh is the restriction of a B : S × V → R and fh the restriction of a f : V → R to these subspaces. αh αh wh 91 Now choose an arbitrary sh ∈ Sh then uh − u∗ ≤ ≤ From this the theorem follows. wh ) − fh (wh ) Mh u∗ − s h + 1+ . vh ) ≤ Mh sh v h Vh Sh . 3. See Section 3.5. It follows that for arbitrary sh ∈ Sh uh − s h ≤ Mh ∗ 1 Bh (u∗ . It turns out from theorem 3. It follows that in this case the estimate is reduced to the result from Theorem 3.e. ∀uh ∈ Sh .58). In this we will restrict ourselves to errors caused by quadrature. 2004 because uh is a solution to (3.1.3 that. . • The conditioning of the problem can also be formulated as: there exist αh > 0 and Mh > 0 so that for all sh ∈ Sh αh s h Sh ≤ sup vh ∈Vh Bh (sh . We see that the error is determined by 1. wh ) − fh (wh ) u − sh + .
li+1 }. Ωei ∩ Ωej = ∅ if i = j.3 The formalisation of the ﬁnite element method We do not want to describe the theory of interpolation by piecewise polynomials in all detail (for instance: we will prove few theorems) but we will show some of the main results in the next section. · · · .···. uh (x) = j=1 ae φe (x) for x ∈ Ωe .92 P. Corollary 3. i = 1. i = 1. Let L be a ﬁnite set of linearly independent linear functionals li . 1] and make a partition 0 = x1 < x2 < · · · < xN +1 = 1. · · · . Let Ωi = [xi . Let Ω ⊂ Rd be an open bounded domain with a Lipschitz continuous boundary ∂Ω.5.W. Let u ∈ C m (Ω) with m ≥ 0. j i 2.5. Then L is P unisolvent if for every set of scalars αi .61) Example 3. N . Hemker 3. ∂Ωe Lipschitz continuous}. where αi are multiintegers. · · · . · · · . i = 1. .4 Let P be a set of realvalued functions deﬁned over a domain Ω. (3. j We formalise the above construction. Deﬁnition 3.5 Let L be a P unisolvent set then there exists a unique set of piecewise polynomials {pj . N over P . there exists a set of derivatives D αi . PΩi = P1 (Ωi ) and LΩi = {li . Then {LΩi } is a {PΩi }unisolvent set: take any two numbers α and β then there is one and only one polynomial p ∈ PΩi such that p(xi ) = α and p(xi+1 ) = β. we show in this section a formal way to introduce ﬁnite elements. For a local ﬁnite element approximation of u on Ωe ∈ ∆h we choose a set of functions Pe = {φe }j=1. pj ∈ P } such that. such that (3.Ne and we approximate u j on Ωe by Ne Ωe = ∅. j j As a rule we choose {φe } such that j 1. there is a unique p ∈ P such that li (p) = αi . j = 1.60) Usually we choose {φe } such that it contains all polynomials of degree k j Pk (Ωe ) ⊂ Span(φe ). xi+1 ]. A partition ∆h of Ω is deﬁned as ∆h = { Ω e  e Ωe = Ω. To gain some insight in the derivation of these results. there exists a set of nodal points of Ωe : {be  be ∈ Ωe . with li (f ) = f (xi ) and li+1 (f ) = f (xi+1 ).5. Ne } and j j D αi φe (be ) = δij .6 Take Ω = [0. N . li (pj ) = δij .5.
So we take a set of ﬁnite elements {Ωe . Le } in which e Pe = {φj ∈ C ∞ (Ωe ). i = 1. 3. Ne } gives the degrees of freedoms of the ﬁnite element.63) Π is a projection because li (Πv) = li (v) for all li ∈ L and any v ∈ C m (Ω). The number of (diﬀerent) elements in L is N =dim(L).8 A global ﬁnite element approximation of a function u on Ω is deﬁned by N uh (x) = j=1 aj φj (x) for x ∈ Ω. that are all is equivalent to a reference ˆ ˆ ˆ or master element (Ω. The P unisolvent set L of degrees of freedom implies a natural P interpolation of a C m (Ω). P .Version: September 13. there exists a partition ∆h = {Ωe }e of Ω.5. j = 1. Pe . Error estimates for the master element are then easily transferred to every element. v → Πv = lj (v)φj .62) This Π is deﬁned as (3. 2004 Deﬁnition 3. Le }. Le }. so N ≤ e Ne . L). L) by an aﬃne or isoparametric transformation. j (3. · · · . where we choose the set P = {φj } such that 1.5. 4. and N is the total number of degrees of freedom. Pe . φj Ωe ∈ Pe . by L = ∪e Le we denote the set of degrees of freedom of the global approximation.7 A ﬁnite element is a triple {Ωe . lk (φm ) = δkm } . for every Ωe ∈ ∆h there is a ﬁnite element {Ωe . 2. by P we denote the set of basis functions of the global approximation.function Π : C m (Ω) → Span(P ). P is the set of N functions given by P = {φj  ∃e . Then all ﬁnite elements are constructed from ˆ ˆ ˆ (Ω. Pe . Deﬁnition 3. Ne } 93 are the basis functions of the ﬁnite element and the Pe unisolvent set e Le = {li ∈ C m (Ωe ) . · · · . P . . Usually we take the same type of Le and Pe on every Ωe in ∆.
The proof of this theorem and a number of preceding lemmas will not be given explicitly. Hemker 3. We are interested in the behaviour of the error of approximation u − Πh u for h → 0. In fact. They can be found in [4]. meas(Ωe ) = Ωe dx.94 P.p (Ω) → Sh can be deﬁned Πh u = j lj (u)φj . In this way we also introduce a sequence of projections {Πh }h→0 with Πh : C m (Ω) → Sh . On each partition ∆h we assume a global ﬁnite element approximation and we denote by Sh = Span(P ) the space of approximating functions. We apply this to analyse a secondorder problem in detail. First we deﬁne on each element in the partition a projection operator which shall be used for the approximation. From this we can ﬁnd error estimates for the global approximation.18 below. This can be done either by making the elements Ωe smaller or by taking a higher order of approximation. the operator Πh : W l.5. In this section we choose for the ﬁrst option and study the quality of the approximation if we take smaller and smaller elements Ωe . . ˆ Deﬁnition 3.5. Assuming that all elements are similar to one another (and thus there will be a unique reference element) we can ﬁnd estimates for these projection operators on each element. in which h denotes the largest diameter of the elements Ωe ∈ ∆h .5. A result is given in theorem 3.3 we saw the construction of an interpolation operator Π : C m (Ω) → Span(P ). With an appropriate choice of the functionals in L (imposed by Sobolev’s Lemma).10 Two open subsets Ω and Ω of Rn are aﬃneequivalent if there is an invertible aﬃne mapping F : x ∈ Rn → F (ˆ) = B x + b ∈ Rn ˆ x ˆ ˆ such that Ω = F (Ω). S a ball contained in Ωe }. or its more recent version [18]. Further we assume that the reﬁnement of the whole area will take place in a regular way so that also the smallest of the inscribed circles of the elements in the partition will be O(h). we consider a sequence {∆h }h→0 of partitions. Deﬁnition 3. The plan of our treatment is as follows.5. ρe = sup{diam(S).9 Let Ωe be a bounded polygon then deﬁne he = diam(Ωe ). It is clear that a function on an area Ω is approximated better if more degrees of freedom are available.5.W.4 Interpolation theory and applications Interpolation theory in Sobolev spaces In Section 3.
.p (Ω) and W m.p.11 Assume the integers k ≥ 0 and m ≥ 0 and the numbers p. ˆx Analogously we deﬁne the aﬃneequivalence of two ﬁnite elements and all ﬁnite ˆ ˆ ˆ elements in a partitioning are aﬃne equivalent to a master element (Ω.q (Ω).supx D α v(x)}. ∞] are such that the Sobolev spaces W k+1.∞.5. (3.Ωe := α=m 1 p Ωe D α vp dx .Ωe := max {ess. then we ﬁnd B ≤ hΩe /ρΩ and B −1 ≤ hΩ /ρΩe .∞.65) Before we give the basic error theorems we recall the deﬁnitions of the norms and seminorms with which we work.Version: September 13. q ∈ ˆ ˆ [1. p ˆ For every open set Ωe that is aﬃneequivalent to Ω deﬁne the mapping Πe by ˆv (Πe v)ˆ= Πˆ.Ωe α≤m vp e = l. v m. ˆ ˆ ˆ Further let Π ∈ L(W k+1. L) if for each Ωe ∈ ∆h there exists an aﬃne mapping Fe such that for this element • x = Fe (ˆ). e e ˆ ˆ x • li (ψj (x)) = li (ψj (ˆ)). ˆ ˆ Usually we have the following correspondences between points and functions x ∈ Ω → x = F (ˆ) ∈ Ω.p (Ω).Ω α≤m Ωe D α vp dx .supx D α v(x)}. α=m m.Ωe := v l≤m vm. W m.p. Πˆ = p. 2004 95 If we denote the outer and the inner diameter by respectively h and ρ. v ˆ ˆ So we have v (ˆ) = v(x).q (Ω)) be a polynomial preserving mapping: ˆ ˆp ˆ ∀ˆ ∈ Pk (Ω). 1 p vm.p. P . Lemma 3. x e ˆ x • ψj (x) = ψj (ˆ).q (Ω) satisfy the inclusion ˆ ˆ W k+1. 1 p := max {ess.64) (3.p (Ω) → W m. ˆ ˆ x (ˆ : Ω → R) → (v = v ◦ F −1 : Ω → R).
W. Pe . 2.67). This can be done provided the elements do not become ”ﬂat” in the limit: for he → 0 we want that ρe = O(he ).9.Ωe . We specialise this immediately to aﬃneequivalent ﬁnite element families. ˆ ˆ W k+1.q. P .5. There exists a constant σ such that ∀e. ˆ ˆ ˆ Lemma 3. (3. the approximating functions will at least have to include all polynomials up to certain order: condition (3.68) ˆ ˆ ˆ then there is a C = C(Ω. Pe . ρe that we use e as a kind of parameter of the family. m.p (Ωe ) corresponding as in (3.5.67) (3.q. ∞] ˆ ˆ W k+1.p and all v ∈ W (Ωe ) v − Πe vm. By Sobolev’s Lemma.Ωe ≤ C(meas(Ωe ))1/q−1/p hk+1 hk+1 e vk+1.p (Ω) → W m. we have a set of degrees of freedom L which uses derivatives up to certain order s.66) (3.Ωe : hk+1 e vk+1. ρe and meas(Ωe ) as in deﬁnition 3. Le ) and all v ∈ W k+1. ρm e e Obviously. Deﬁnition 3. ρm e condition 3.96 P.5.p. We formulate a number of requirements on the spaces we use: ˆ 1. L) such that for aﬃneequivalent ﬁnite elements (Ωe .65).p (Ω) and v ∈ W k+1.12 Let (Ω. Ω) such that for all aﬃneequivalent sets Ωe k+1.p.66).68). · m.q (Ω). Hemker ˆ with v ∈ W k+1.q ˆ ˆ ˆ Pk (Ω) ⊂ P ⊂ W (Ω). the term ρm appearing in this lemma is undesirable.p (Ω) → C s (Ω).Ωe .13 A sequence of partitionings {∆h } of ﬁnite elements {(Ωe .q.p (Ωe ) v − Πe vm. This gives condition (3. m ≥ 0 and p. this sets requirements on the Sobolev spaces. Le )}h 10 is regular if the following two conditions are satisﬁed: 1. If for some k ≥ 0. 10 Note he ≤ σ. L) be a master element for which s is the largest ˆ derivative in L. the projections will be measured in a weaker norm (3. so we would e like to dispose of the ρe . q ∈ [1. Then ˆ ˆ ˆ there exists a constant C = C(Π. P . .Ωe ≤ C(meas(Ωe ))1/q−1/p with he .
5.q. The diameter h = maxΩe ∈∆h he tends to zero.Ωe . θΩ ≥ θ0 > 0. L) such that for all k+1. Lemma 3. That is. Assumptions: (H1) We consider a regular family of ﬁnite elements Fh . Thus. 2004 2.p. P . Pe .Version: September 13. Deﬁnition 3.∆h := Ωe ∈∆h v p m.5.67) ˆ ˆ ˆ and (3. For regular families we can convert the error estimate 3.15 For functions that are (together with their derivatives up to order m) integrable on the elements Ωe ∈ ∆h . Ωe ∈Fh he ≤ σ. ρe (H2) All the ﬁnite elements (Ωe . (3. 97 Remark: The regularity of the elements can also be stated in a more geometrical sense: let θΩ denote the smallest angle in an element Ω.66).Ωe . L). and h = max he → 0 . P . L) satisﬁes conditions (3. (H3) All the elements are of class C 0 . Pe . Le ) whose reference element (Ω. then there must be a θ0 > 0 such that ∀Ω ∈ Fh .5.p. Assuming the family is aﬃneequivalent we would like to give an error estimate for the global ﬁnite element approximation.p members Ω of the family and all v ∈ W (Ω). we introduce the following norm and seminorm if p < ∞ 1 p v m. we can give an error estimate on each element in a partition ∆h of our domain Ωe . The assumptions above have to be applied in an even more strict sense: on each level of reﬁnement.68). P . Then there exists a constant C = C(Ω.p. Ωe ∈ ∪h Fh are aﬃneequivalent to ˆ ˆ ˆ a single reference element (Ω. Le ). .14 Let there be given a regular aﬃne family of ﬁnite elements ˆ ˆ ˆ (Ωe .12 in estimates for the associated norms.Ωe k+1−m ≤ C(meas(Ωe ))1/q−1/p he vk+1. there exists a constant σ > 0 such that ∀ Ωe ∈ ∪ h Fh . v − Πe v m. To summarise the result we deﬁne a norm and a seminorm for piecewise smooth functions on a given partitioning ∆h .
A number of approximation results is summarised in the following lemma. Figure 3.p (Ω) → Sh ⊂ W m. In case p = 2 we drop the p in the subscripts.∆h .Ω on Ω itself. P ) > 0. Lemma 3.Ωe .p.p.5. but it is not in H 2 (Ω).98 1 p P. If we take however a partition ∆h such that the elements Ωe correspond with the smooth parts of the function. Let the requirements (3. such that.68) be satisﬁed for every element.∆h := max vm.16 The norm · m. Hemker vm. b] and consider the piecewise linear function in ﬁgure 3.Ωe . Consider the global interpolation operator Πh : W k+1. This is seen as follows.∞. P ).p.Ωe .∆h on the partitioning ∆h of Ω is not the same as the norm · m.p (Ω) deﬁned by ˆ ˆ ˆ then there is a constant C = C(Ω. (3.∞. This function is an element of H 1 (Ω).∆h ≤ C hk+1−m vk+1.∞. Ωe ∈∆h Example 3.17 Let Fh be a regular family of ﬁnite elements satisfying (H1). Π h Ω e = Π e and v − Πh vm.p. L.11.∞.11: Piecewise H 2 function.p. ˆ ˆ ˆ (H2) and (H3) that are all aﬃne equivalent with the master element (Ω. it is piecewise H 2 . vm.5.67) and (3.∆h by H∆h . Take Ω = [a.p.∆h := max Ωe ∈∆h v m. m.∆h := Ωe ∈∆h vp m.p. L.p. . k+1−m v − Πe vm. so in 2 H∆h (Ω).W. We denote the space of functions m with ﬁnite norm .Ωe .Ωe ≤ C he vk+1. If p = ∞ we deﬁne v m.66).
i. let u be the solution to a second order elliptic problem and let uh be the associated discrete solution.70) ≤ C hk+1−m vk+1.17.p. r − s).m The spaces Sh (Ω) contain all kth order piecewise polynomials and are subsets k. Theorem 3.p.12 are satisﬁed for p = 2.p.∆h .18 Under conditions as in lemma 3.∆h ≤ C hk+1−m vk+1. the following error estimate is valid u − Πh u m.Ω = O(hk ).5. r ≥ 0 and 0 ≤ s ≤ min(r.5.5.18. s ≤ C hσ u r .Ω ≤ C hk+1−m uk+1. Suﬃcient for convergence for a second order problem is thus k = 1.p. are denoted by k. such that the requirements of lemma 3.p. Proof: Because we consider a second order problem we have m = 1 and the theorem follows by combining theorem 3. u ∈ H k+1 (Ω) and Πh is the interpolation operator that leaves invariant all polynomials of degree k or less.18. m) there exists a uh ∈ Sh (Ω) and a constant C ≥ 0.e. but we still have v − Πh v m.m Sh (Ω). Deﬁnition 3.m h ≤ h0 . 2004 From this easily follows v − Πe v and v − Πh v m. Remark: If the functions v and Πh v are not suﬃciently smooth estimate (3. Here C is independent of h. or approximation by piecewise linear functions. .m m of H (Ω).they are kth degree piecewise polynomials in H m (Ω). 0 ≤ m ≤ k + 1. 99 (3. for a regular family of ﬁnite element approximations.∆h .5. u ∈ H r (Ω).5. 0 ≤ m ≤ k + 1.5. We thus take p = q = 2. By the preceding theorem we know that for a sequence {Sh (Ω)}h→0 .1 and theorem 3.5. As a direct consequence of this lemma we can state the basic interpolation theorem that can be applied to elliptic problems.Version: September 13.p.Ω m. Corollary 3.Ωe k+1−m ≤ C he vk+1. k.70) is not valid.19 The spaces Sh .5. k.Ω .69) (3. independent of h. Then u − uh 1. such that u − uh where σ = min(k + 1 − s.20 Under the same conditions as in theorem 3. based on a regular family Fh of ﬁnite elements.Ωe .
Ω . satisfying an inverse assumption. and ˆ ˆ ˆ P ⊂ W k.q.∞ (Ω). q) such that for all vh ∈ Vh vh k. then ∃C = C(ν.Ω ≤ n/2 vh 0.r for some (l.r. Example 3.71) with the usual adaptation if p or q is ∞. p.Ω .∞. Note that σ appears in the regularity assumption (H1). p. unless the function itself ”blows up”.Ω .Ω ≤ C vh 0.6.23 Assume Vh ∈ W l. then C vh 0. h and vh 1. Theorem 3. h (3. q ∈ [1. k. then vh 1.5. For such families of spaces we establish the following equivalence between seminorms.5.p. which in turn implies that functions in the approximation space have derivatives that will not ”blow up”. then the above inequalities can be given for the seminorms .∆h . It is clear that this assumption is by no means restrictive in practice: the ’usual’ ﬁnite element spaces satisfy an inverse assumption. m. (3.21 (H4) A family Th of triangulations satisﬁes an inverse assumption if there exists a ν such that ∀ T ∈ ∪ h Th .Ω . Deﬁnition 3.72) . p. Proof: See [4.5. For instance suppose assumption (H3) is valid and that ˆ P ⊂ H 1 (Ω). r).22 Let there be given a regular family of triangulations Th in Rd . Hemker Inverse inequalities We now treat a property of piecewise polynomial spaces that has no immediate use at this point. 141].W.l. σ.∞. Let 0 ≤ m ≤ k.∆h ≤ C meas(Ωh )max{0.100 P.∞. h C vh 0.1/q−1/p} hk−m vh m.p (Ω) ∩ W m.q (Ω).5. ∞]. hT In fact an inverse assumption implies that the elements can not diﬀer too much in size. However it is an easy consequence of the foregoing and we will use it in section 3. h ≤ ν.Ω ≤ ˆ If P ⊂ W 1.
Ω = O(hk ).26 (AubinNitsche Lemma) Let Lu = f be a problem of order 2m and let Vh ⊂ V = H m (Ω). suited for the function k. Let z be the solution to the adjoint problem B(w. the space of piecewise linear functions on a regular partition ∆h . vh ) = 0 for all vh ∈ Vh . Motivated by this we come to the following deﬁnition. under mild additional assumptions.73) is 1. Error estimate in the L2 norm In Section 3. z − vh ) = g(e). In this section we show that. or. For any 0 ≤ s ≤ m: e s = sup g∈H −s g(e) .74) Proof: Deﬁne the error e = uh − u.1 valid for each vh ∈ Sh (Ω). s ∈ [0. for all w ∈ H s (Ω).25 A family of piecewise polynomial spaces {Sh (Ω)}h→0 has the inverse property if ∃C : ∀s ≤ m.5. from (3. 2004 101 Example 3. k. (3. m] .4 we saw that under conditions we can assure that u − uh 1.72) we ﬁnd C (3. such that z 2m−s ≤C g −s .24 We can also easily give inequalities for the norms: for instance. B(e.Ω . This means that at least u − uh 0. g −s We ﬁrst evaluate g(e). Theorem 3.1 This can be applied for the following case.Ω = O(hk+1 ).Ω .m Deﬁnition 3.73) vh 1. Consider Sh (Ω). So g(e) ≤ B e m vh ∈Vh inf z − vh m. hm uh m k. h 1. The continuous problem is to ﬁnd u such that B(u. In particular B(e. .5. Take Ω = [a.Ω = O(hk ). Then u − uh s. Assume that for all g ∈ H −s there is a solution to LT z = g (the adjoint problem). vh ) = f (vh ) for all vh ∈ Vh .m ≤ C hs uh s . then (3.m spaces Sh (Ω).Version: September 13. b]. z) = g(w).Ω ≤ C hm−s u − uh m.5. ∀uh ∈ Sh (Ω).5. The discrete problem is: ﬁnd uh such that B(uh . v) = f (v) for all v ∈ V . z) = g(e).Ω ≤ vh 0. u − uh 0. (the regularity of the adjoint problem). as B(e.
if we compute the error in the snorm. ξ) s(ξ) dξ. the function G(x.76) function G is the resolvent kernel of the diﬀerential equation. Hemker C hr z m+r . 0. b] ⊂ R Ly ≡ − d d d (a2 (x) y) + a1 (x) y + a0 (x)y = s(x). We will use the Greens function for equation (3.75): i.Ω ≤ C hk+1 uk+1. Corollary 3. dx dx dx (3.5. For example in case of a symmetric problem it is a trivial matter. So. because of the regularity of the adjoint problem. a (3. If we put r = m − s and provided 2m − s ≤ k + 1. g(e) ≤ B e m P.W. due to theorem 3. g(e) ≤ ≤ B B e e m m C hm−s z C hm−s g 2m−s −s . we get e s = sup g∈H −s g(e) ≤ C hm−s e g −s m. provided r ≤ k + 1 − m.5 Pointwise error estimate and superconvergence In this section we restrict ourselves to a second order linear ordinary diﬀerential equation on Ω = [a.102 Now.5. 3.27 For the error in L2 norm for a second order problem we get u − uh provided k ≥ 1. .75) with a2 (x) = 0 and homogeneous Dirichlet boundary conditions y(a) = y(b) = 0.e.18. ξ) such that 11 b y(x) = − 11 The G(x.5.∆h . Remarks: • The regularity of the adjoint problem is not very restrictive.
ϕ1 (a) = 1. (3. x) ∪ (x. LT G(x. ϕ1 (x)ϕ2 (ξ)/z(x) if x < ξ ϕ1 (ξ)ϕ2 (x)/z(x) if ξ < x. let ϕ1 and ϕ2 be two independent solutions to the adjoint (transposed) equation LT y ≡ − d d d (a2 (x) y) − (a1 (x)y) + a0 (x)y = 0. h→0 h→0 ∂ξ ∂ξ ∂ξ a2 (x) If xi is a nodal point in a partition ∆h of Ω. ξ + h) ∂G(x. 3. x) ∪ (x. Remark: It is easily seen that this function z(x) satisﬁes the equation a2 z + a1 z = 0. Therefore. .77) 2. Under this condition we see that k. b). ·) − vh 1 k. 2004 103 for any function s(x). ξ) := lim − lim = . x = xi . ξ − h) 1 G(x. ξ) are: 1 1. it is not hard to verify that G(x. If z ≡ 0 then ϕ1 and ϕ2 are linearly dependent. ξ) is constructed as G(x. b) ). for m = 1 inf G(x. dx dx dx with boundary conditions ϕ1 (a) = 0.76) in (3. Other properties of this Greens function G(x.1 vh ∈Sh inf G(xi .∆h if if x = xi .∆h .) By substitution of (3.1 vh ∈Sh (Ω) ≤ C h0 G1 C hk Gk+1. ·) − vh m. So. ·) ∈ H0 (Ω) ∩ C 2 ( (a.∆h ≤ C hk+1−m G(xi . ·) is not piecewise H k+1 on ∆h . G(x. A priori.Version: September 13. but we will construct it. (In that particular case the homogeneous problem has a nontrivial solution and G(x. but it still is continuous. then G(xi . So z has a unique sign. so in H 1 .1 mated well in Sh (Ω). ϕ2 (b) = 0 and ϕ2 (b) = 1. ξ) is the Green’s function indeed. ·) k+1. it is not clear that such a function G(x.78) The last estimate is true because in this case G(x. ξ) can exist. ξ) does not exist. Now G(x. jumpξ=x ∂ ∂G(x. ·) can be approxik. ·) ≡ 0 on (a. (3.75). ξ) = where12 z(x) := a2 (x) (ϕ1 (x)ϕ2 (x) − ϕ1 (x)ϕ2 (x)) . 12 The determinant ϕ1 ϕ2 − ϕ1 ϕ2 = ϕ1 ϕ2 ϕ1 ϕ2 is called the Wronskian.
vh ) = f (vh ). so for e := k. e 1 k. ξ) f (ξ) dξ = − G(x.75) is discretised by a ﬁnite element method k. Similarly k.4.W.78) fails to be better at the points of ∆h .1 k.104 P. ξ)Ly(ξ)dξ = −B(y. e k 1 inf G(x. that we obtain when we solve the discrete equation instead of the continuous equation.5.0 yh ∈ Sh (Ω) is the solution to B(yh . there will be no superconvergence.3). G(x. As we have seen in section 3. Hemker Theorem 3. In this section we will formulate these requirements for two cases: ﬁrst we show under which conditions the global error estimate is preserved (as in theorem 3. ·)) This gives e(x) ≤ B Because B is bounded. v) = f (v). This could disturb all the error estimates we have given so far. ·)).28. so e(x) = −B(e. .5. Now b y(x) = − a G(x. ∀vh ∈ Sh (Ω). and analogously for yh . ·) − vh ).1 = −B(e.3 it may be so that we have to compute the integrals in the discrete equations using a quadrature rule. (3.5.78) the theorem follows. ∀v ∈ H0 (Ω). G(x.1 with Sh = Vh = Sh (Ω).1 y − yh we have B(e.2 (Ω).5. x ∈ ∆h . Then estimate (3. So we not only have the error y−yh but also an additional error yh − yh (which can be associated ˜ with the consistency error of theorem 3. In order to take this into account we require that the additional error should be at most of the same order as the discretisation error.5. 3.18) and further we discuss the how we can preserve the pointwise superconvergence of theorem 3. This will lead to some requirement for accuracy of the quadrature rule. G(x. ∀vh ∈ Sh (Ω). is called superconvergence. then we have the pointwise error estimate y(x) − yh (x) ≤ C h2k C hk if if x = x i ∈ ∆h . ·) − vh 1 . Remark: If this same problem is tackled with functions in S 3. vh = O(h ) and (3.79) 1 1 Proof: y ∈ H0 (Ω) is the solution to B(y. The phenomenon that the order of accuracy of an approximation at particular points is better than the estimate in a global norm. vh ) = 0 ∀vh ∈ Sh (Ω).28 If equation (3.6 The inﬂuence of the quadrature rule In the previous sections we have given estimates for the error y − yh .
32 A t + 1point NewtonCoates formula has degree t. Provided the set of nodes contains the quadrature points the estimate is valid. ˜ y . Proof: Take a (t + 1)points interpolating polynomial of degree t on every Ωe : each of this gives the estimate of lemma 3.Version: September 13. and we use a quadrature of degree t. vh ) = fh (vh ). 3. ∀Ωe . applied to a partition ∆h . v) = f (v).∆h .5. (3.p (Ωe ). the continuous problem ﬁnd y ∈ S such that B(y.5. for all vh ∈ Vh . we will assume that all the coeﬃcients in the partial diﬀerential equation are (piecewise) C t+1 (Ω)functions. 2004 Deﬁnition 3. f (xi ) = Πh f (xi ) and such that f − Πh f m. for f ∈ C(Ω) ∩ W t+1. for all v ∈ V .p. the discrete problem ﬁnd yh ∈ Sh such that B(yh .31 When a quadrature rule of degree t is used the quadrature error is O(ht+1 ): Ωe f − Πh f  dx ≤ C ht+1 α=t+1 Ωe D α f (x) dx.∆h ≤ C ht+1−m f t+1. then there exists.29 A quadrature rule of the form f dx ≈ wi f (xi ) i 105 Ω is called to be of degree t if it integrates any polynomial f of degree t exactly. In order to study the inﬂuence of the quadrature on the accuracy of the FEM we have to distinguish three diﬀerent problems: 1. the discrete problem with quadrature ﬁnd yh ∈ Sh such that Bh (˜h .e.p.30 Given a repeated quadrature rule of degree t. an npoint Lobatto formula has degree 2n − 3. An npoints Gauss quadrature has degree 2n − 1. vh ) = f (vh ) for all vh ∈ Vh . In the following we will restrict ourselves to second order elliptic problems.17. i. Theorem 3.5. a piecewise polynomial Πh f of degree t such that it interpolates f on the nodes. 2.5.5.80) Example 3. Corollary 3.
Then y − yh ˜ 1.5. Hemker Here fh is a linear and Bh is a bilinear functional in which the integrals (see Section 3.4. let B : Sh ×Vh → R be bounded and coercive and let fh and Bh be the approximations of f and B obtained by quadrature t ≥ 2k − 2. and Bh deviates from B. Using the operator Πh .5.80) we have E ≤ C ht+1 D t+1 (f v) dx.∆h k.3) are replaced by a repeated quadrature rule of degree t on the partition ∆h . the other goes analogously.∆h .∆h (3. Theorem 3.5. e Ωe e Ωe so.∆h .5.33 When using a quadrature rule of degree t we have f (vh ) − fh (vh ) ≤ C ht+1 vh and y y ˜ B(˜h . Further.30. vh ) ˜ yh − y h 1 ≤ ˜ sup . as introduced in Theorem 3. Lemma 3. The following theorem states the conditions under which the accuracy as given in theorem 3. To compute the additional error we ﬁrst have to make an estimate of the amount by which fh deviates from f . by Leibniz’ rule and the fact that vh ∈ Vh . we have E := f (vh ) − fh (vh ) = Using (3.W.Ω = O(hk ). Proof: Let yh be the solution of the discrete problem and yh the solution to ˜ the discrete equation with quadrature.81) vh k.18 is preserved.82) Proof: We only give the proof for fh and f .  ((f v) − Πh (f v)) dx.34 Let S = V = H 1 (Ω).106 P. γh vh ∈Vh vh 1 . vh ) − Bh (˜h . ∆h a regular partition of Ω and assume Sh and Vh satisfy the inverse property. k E ≤ C ht+1 l=0 e Ωe (D t+1−l f )(D l vh ) dx ≤ C ht+1 vh k. (3. vh ) ≤ C ht+1 yh k. We seek a upper bound of yh − yh by ˜ considering 1 B(yh − yh .
Thus. ·). xi a nodal point. ˜ ˜ First B(yh − yh .∆h } 1.∆h iii ≤ iiii C ht+2−k vh 1 + C ht+2−k vh 1 { yh C ht+2−k vh C ht+3−2k vh 1 1 k. Gi − vh ) + B(yh − yh . (ii) the estimates of Lemma 3. yh − yh 1 {1 − C ht+3−2k } ≤ C ht+2−k {1 + yh ˜ Provided h is small enough (C ht+3−2k < 1).35 Let the two point boundary value problem (3. vh ) − Bh (˜h . Proof: We introduce Gi = G(xi .∆h }. (3. (iii) the fact that Vh has the inverse property and (iiii) that Sh has the inverse property. x ∈ ∆h . 2004 in which γh is the (discrete) coercivity constant of B. and consider yh (xi ) − yh (xi ) = B(yh − yh .∆h + C ht+1 yh ˜ k.5 we now answer the same question: what should be the degree of the quadrature to preserve the pointwise superconvergence accuracy? Theorem 3.33 for the quadrature. vh ) y y y f (vh ) − fh (vh ) + Bh (˜h . O(hk ). we have yh − y h ˜ 1 k. So compute B(yh − yh . vh ) + Bh (˜h .∆h ˜ + yh − yh k. then y(x) − yh (x) = ˜ O(h2k ). vh ) − B(˜h .∆h }.∆h .5. x ∈ ∆h .1 tised by a ﬁnite element method with Sh = Vh = Sh (Ω). with t ≥ 2k − 1. Gi ) = B(yh − yh .∆h } ≤ {1 + yh y h − yh ˜ + We used (i) that yh is a solution of the discrete problem and yh is the solution ˜ to the discrete problem with quadrature. vh ) ˜ ≤ = ≤ ii i 107 B(yh .5. vh ) y y C ht+1 vh k. Gi ). k. ≤ C ht+2−k {1 + yh k.5.Version: September 13.75) be discrek. vh ) − B(˜h . both of which are satisﬁed if t ≥ 2k − 2. and let the quadrature rule be of degree t. vh ) ˜ ˜ ˜ .83) To preserve the accuracy t should satisfy both t + 2 − k ≥ k and t + 3 − 2k > 0.∆h vh k. For the superconvergence as treated in Section 3.
vh ) − B(˜h . Gi − vh ) + B(yh . vh ) ˜ y +Bh (˜h . .∆h ˜ k.36 A Gaussian quadrature on k points or a Lobatto quadrature on k + 1 points are suﬃciently accurate to preserve superconvergence.∆h {1 + yh k. y y so. Acknowledgement In 1992 H. so we conclude that by the condition t + 1 ≥ 2k also y(x) − yh (x) = O(h2k ) ˜ in nodal points. Final corrections 2004.∆h .108 P.∆h ≤ ≤ (3. Gi − vh ) + f (vh ) − fh (vh ) ˜ +Bh (˜h .5. Hemker = B(yh − yh .∆h {1 + yh k.5. Example 3. Gi ) ˜ ≤ ≤ (3.28 y(x) − yh (x) = O(h2k ) in nodal points x = xi ∈ ∆h . vh ) − Bh (˜h .83) M yh − yh 1 Gi − vh 1 + f (vh ) − fh (vh ) ˜ +Bh (˜h .∆h } Gi − vh 1 +C ht+1 Gi − vh k. vh ) − B(˜h . According to theorem 3.W. vh ).∆h {1 + yh k.∆h } C ht+1 Gi C ht+1 Gi k.∆h k. vh ) y y = B(yh − yh . Eleveld prepared an earlier version of these notes on ﬁnite element methods.∆h M C ht+2−k {1 + yh k.∆h } 1 ≤ ≤ + C ht+1 Gi k. based on lectures given at the University of Amsterdam.∆h } Gi − vh +C ht+1 vh k. using the continuity of B B(yh − yh .∆h } +C ht+1 Gi k. vh ) y y M yh − yh 1 Gi − vh 1 + C ht+1 vh ˜ +C ht+1 yh k.∆h vh k. vh ) − B(˜h .78) M C ht+2−k {1 + yh k.
RWTH. Koren. [9] P. J. Braess. Jeltsch. Academic Press. Numerical Computation of Internal and External Flows. Defect correction and nonlinear multigrid for the steady Euler equations. Hemker and B. [13] R. J. D. 25:35–61. Lax. SIAM Review. Numerical Computation of Internal and External Flows. Ciarlet. Sobolev Spaces. Volume 1: Fundamentals of Numerical Discretisation. van Loan. [5] G. Properties of discrete hyperbolic conservation laws. 1988. 1983. Cambridge University Press. [10] P. G. f. Hirsch. P. Hirsch.W.A. [6] G. H. CWI. Ridgeway Scott. Teubner. 109 . Volume 2: Computational Methods ﬁr Inviscid and Viscous Flow. Theorie und Numerik elliptischer Diﬀerentialgleichungen. Eﬃcient multidimensional upwinding for the steady Euler equations. 1991. FiniteDiﬀerence Methods for Partial Diﬀerential Equations. 1988. Wiley. Koren. 1975. New York. Baltimore. Wiley. G. 1997. 2nd edition. and B.Bibliography [1] R. Hemker and B.W. 1960. [3] S. [11] C. volume vol 65 of Pure and applied mathematics series. Harten. 1986. [4] P. 1988. The ﬁnite element method for elliptic problems. Technical Report NMN8801. Johns Hopkins University Press. Finite Elements. Forsythe and W. volume vol 4 of Studies in mathematics and its application. The Mathematical Theory of Finite Element Methods. Praktische Mathematik. Technical Report NMR9107. Van Leer. F. Brenner and L. Golub and Ch. B. Springer Verlag. CWI. 1978. Technical Report 8854. [8] A. C. New York. London. Wasow. [12] C. [2] D. Aken.E. Stuttgart. On upstream diﬀerencing and Godunovtype schemes for hyperbolic conservation laws. 1989. NorthHolland. Wiley.R. 1994. Hackbusch. Geometrie u. Adams. Amsterdam. Matrix Computations. [7] W. 1990. Inst.
110
P.W. Hemker
[14] C. Johnson, A. Szepessy, and P. Hansbo. Defect correction and nonlinear multigrid for the steady Euler equations. Technical Report 21, Chalmers Univ. Techn., 1987. [15] B. Koren. Multigrid and Defect Correction for the Steady NavierStokes equations, Applications to aerodynamics. Number 74 in CWI Tract series. CWI, Amsterdam, 1991. [16] P. D. Lax. Hyperbolic Systems of Conservation Laws and the Mathematical Theory of Shock Waves, volume 11 of Regional Conference Series in Applied Mathematics. SIAM Publication, Philadelphia, 1973. [17] R.J. LeVeque. Numerical Methods for Conservation Laws. Lectures in Mathematics. Birkh¨user, Basel, 1990. a [18] J. L. Lions (eds.) P. G. Ciarlet. Handbook of Numerical Analysis, Vol II: Finite Element Methods, part I. [19] S. V. Patankar. Numerical Heat Transfer and Fluid Flow. Hemisphere McGrawHill, 1980. [20] D. W. Peaceman. Fundamentals of Numerical Reservoir Simulation. Elsevier Science Publ., 1977. [21] R. D. Richtmyer and K. W. Morton. Diﬀerence Methods for Initial Value Problems. Interscience Publ., 1967. [22] W. Rudin. Functional Analysis. McGrawHill, New Delhi, etc., 1973. [23] J. Smoller. Shock Waves and Reaction Diﬀusion Equations, volume 258 of Grundlehren der mathematische Wissenschaften. Springer Verlag, 1983. [24] S.P. Spekreijse. Multigrid Solution of the Steady Euler Equations, volume 46 of CWI Tracts. CWI, 1988. [25] J. C. Strikwerda. Finite Diﬀerence Schemes and Partial Diﬀerential Equations. Wadswordth and Brooks, Paciﬁc Grove, Calif., 1989. [26] R. Struijs, H. Deconinck, and G. Coussement. Multidimensional upwind schemes for the compressible ﬂow equations. Technical report, Von Karman Institute, 1989. [27] P. A. Raviart. V. Girault. Finite Elements Methods for Navier Stokes Equations, volume vol. 5 of Springer series in Computational Mathematics. SpringerVerlag, Berlin, et.c, 1986. [28] B. Van Leer. On numerical dispersion by upwind diﬀerencing. Technical Report 8551, Delft Univ. Techn., 1985. [29] B. Van Leer. Flux vector splitting for the 1990’s. Technical report, Univ. Michigan, 1990. [30] K. Yosida. Functional Analysis. Springer Verlag, Berlin, etc., 1974.
Index
adjoint problem, 101 aﬃneequivalent, 94 ampliﬁcation factor, 34 asymmetric problem, 68 AubinNitsche Lemma, 101 backward Euler method, 33 Banach space, 50 barycentric coordinates, 84 bilinear functional, 54 bilinear operator, 53 block centered, 21 boundary conditions, 46 bounded operator, 53 box method, 28 BubnovGalerkin, 28 Burgers’ equation, 12, 13 Cauchy Riemann, 22 CauchySchwarz inequality, 50 cell centered, 21 cell vertex, 21 central approximation, 32 characteristic polynomial, 5 characteristics, 7 coercive, 56 collocation method, 27 compact support, 51 compatibility condition, 24 complete, 50 conforming method, 69 conjugate exponents, 54 conjugatelinear operator, 53 conservation law, 6 consistency error, 91 convection equation, 31 convective ﬂux, 2 111 convergence strong, 56 weak, 56 CranckNicolson method, 33 curved boundary, 86 degree, 105 degrees of freedoms, 93 delta function, 59 dense, 50 diﬀusion equation, 3, 29, 46 diﬀusive ﬂux, 2 Dirac delta function, 59 Dirichlet boundary conditions, 46 Discrete LaxMilgram Theorem, 88 discrete solution, 87 distance, 50 distribution, 59 distributional derivative, 60 divergence, 2 downwind approximation, 32 dual space, 54 elementary load vector, 74 elementary stiﬀness matrix, 74 elliptic, 5, 6 elliptic equation, 46 energy norm, 69 entropy condition, 15, 16 equivalent diﬀerential equation, 37 Euler equations, 6 ﬁnite ﬁnite ﬁnite ﬁnite ﬁnite diﬀerence approximation, 19 element, 93 element approximation, 21 volume approximation, 21 volume method, 28
112 forward Euler method, 33 function spaces, 51 functionals, 53 Galerkin method, 28 Gauss’ theorem, 2 generalised derivative, 60 generalised function, 59 global discretisation error, 36 global ﬁnite element approximation, 93 gradient, 2 Greens function, 102 Heaviside function, 60 Hermite interpolate, 81 hierarchical basis, 73 Hilbert space, 50 hyperbolic, 5, 7 IBVP, 79 initialboundaryvalue problem, 79 inner product, 49 inner product space, 49 interpolation error, 91 inverse property, 101 inviscid Burgers’ equation, 8 isoparametric elements, 86 Lagrange interpolation, 70 Laplace equation, 46 Laplace’s equation, 3 LaxMilgram Theorem, generalised, 57 linear functional, 53 linear operator, 53 linear space, 49 linearisation, 30 Lipschitz continuous, 51 load vector, 74 local discretisation error, 36 lumping, 78 mass matrix, 79 master element, 93 mesh, 19, 70 meshwidth, 70
P.W. Hemker metric, 50 mixed boundary conditions, 46 modiﬁed equation, 37 multiindex, 51 multiinteger, 51 Neumann boundary conditions, 46 nodal points, 92 nonunique solution, 14 nonconforming method, 90 norm, 49 normed linear space, 49 numerical diﬀusion, 37 Oleinik, 15, 16 order of the diﬀerential equation, 1 ordinary diﬀerential equation, 1 parabolic, 5 partial diﬀerential equation, 2 partitioning, 19 PDE, 1 PetrovGalerkin, 28 piecewise Lagrange interpolation, 70 piecewise polynomials, 21 Poincar´’s Lemma, 64 e Poisson equation, 3 positive strictly, 56 positive deﬁnite, 46 potential equation, 46 principle part, 3, 5, 46 projection Πh , 94, 98, 105 quadrature rule, 75 quasilinear, 3 RankineHugoniot, 12 reference element, 93 regular family, 97 regularity, 101 scalar ﬁeld, 49 scalar product, 49 semidiscretisation, 31, 79 seminorm, 49 separable, 50
67 topology. 62 spectral method. 4 stiﬀness matrix. 54 Wronskian. 66 weak topology. 56 weighted residual method. 21 splines. 86 upwind approximation. 21 wave equation. 103 113 . 24 strictly positive. 53 symmetric problem. 104 symmetric operator. 74 Stokes equations. 3 weak solution. 2004 sesquilinear functional. 64 triangle inequality.Version: September 13. 11. 56 superconvergence. 60 trace. 26 vector space. 22 standard hat functions. 56 subcoercive. 53 Sobolev space. 49 vertex centered. 78 steady equations. 49 unisolvent set. 83 staggered grid. 27 weighting function. 71 static condensation. 61 Sobolev’s Embeddings Theorem. 60 weakly convergent. 32 variational method. 54 sesquilinear operator.
This action might not be possible to undo. Are you sure you want to continue?