Anda di halaman 1dari 53

Numerical Methods for PDEs

(PDE2)
Jaroslav Vondrejc1
1
Technische Universitat Braunschweig, Carl-Friedrich-Gau-Fakultat, Institute of
Scientific Computing, Muhlenpfordstrasse 23, 38106 Braunschweig, Germany

August 16, 2017

Contents

I Theory to PDEs 4

1 Introduction and repetition 4


1.1 Literature: . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Vector spaces, norms, scalar product, and some inequalities . . . . . . . . . . . . 4

2 Introduction to Lebesgue spaces 6


2.1 Lebesgue spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Some inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Completeness in the sense of metric spaces . . . . . . . . . . . . . . . . . . . . . . 8

3 Sobolev spaces 10
3.1 Prerequisites for Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.1 Continuous functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.2 Weak derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1.3 Norms on finite dimensional spaces . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Definition of Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Sobolev inequalities & embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.4 Trace operator values on boundary . . . . . . . . . . . . . . . . . . . . . . . . 15
3.5 Singularities of Sobolev functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Properties of Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

4 Hilbert and Banach spaces 18


4.1 Linear operators, linear functionals, and dual space . . . . . . . . . . . . . . . . . 19

5 Existence of solutions to elliptic PDEs 20


5.1 Abstract formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.1.1 Minimization problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Existence of solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.1 Existence from minimisation problem . . . . . . . . . . . . . . . . . . . . 21
5.2.2 Lax-Milgram and some properties . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Applications to PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.4 Derivation of convection-diffusion equation . . . . . . . . . . . . . . . . . . . . . . 23

1
5.4.1 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.4.2 Heat equation homogeneous Dirichlet boundary conditions . . . . . . . . 24
5.4.3 Helmholtz equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.4.4 Helmholtz equation with Neumann boundary conditions . . . . . . . . . . 26
5.4.5 Neumann problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

II Discretisation of variational problems 27

6 Abstract approaches 27
6.1 Discretisation approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

7 Finite element method in general setting 29


7.1 Basic aspects of FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7.2 Simplicial finite elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7.3 Rectangular finite elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
7.4 Convergence of finite element method . . . . . . . . . . . . . . . . . . . . . . . . 30
7.5 Influence of numerical integration . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

III Saddle point problems 32

8 Introduction 32
8.1 Motivation example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
8.1.1 Minimisation problem without a constraint . . . . . . . . . . . . . . . . . 32
8.1.2 Minimisation problem with a constraint . . . . . . . . . . . . . . . . . . . 32
8.1.3 Minimisation with a constraint that do not work . . . . . . . . . . . . . . 33
8.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

9 Abstract formulations 34
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
9.2 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

10 Discrete (matrix-vector) formulations 36


10.1 Linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
10.1.1 Direct solvers and Krylov subspace methods . . . . . . . . . . . . . . . . . 37
10.1.2 Conjugate gradients on a subspace V . . . . . . . . . . . . . . . . . . . . . 38
10.2 Eigenvalues of saddle-point linear system . . . . . . . . . . . . . . . . . . . . . . 38

11 Mixed-formulations and its numerical approximation 40


11.1 Applications to PDEs: Poisson equation with general boundary conditions . . . . 40
11.1.1 Connection of mixed-formulations by Green formula . . . . . . . . . . . . 41
11.1.2 Ellipticity and inf-sup condition in continuous setting . . . . . . . . . . . 42
11.2 Galerkin approximation to saddle-point problems . . . . . . . . . . . . . . . . . . 42
11.2.1 Abstract setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
11.3 Poisson equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
11.3.1 Primal-mixed formulation: . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
11.3.2 Dual-mixed formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

12 Duality with Fenchel conjugate 43


12.1 Min-max problems Poisson equation . . . . . . . . . . . . . . . . . . . . . . . . . 43
12.2 Perturbation duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2
IV Appendix 46

A Introduction to Lebesgue integration 46


A.1 Lebesgue measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
A.2 Abstract measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
A.3 Measurable and continuous functions . . . . . . . . . . . . . . . . . . . . . . . . . 49
A.4 Lebesgue integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
A.4.1 Alternative integrals to the Lebesgue one . . . . . . . . . . . . . . . . . . 51
A.4.2 Properties of Lebesgue integration . . . . . . . . . . . . . . . . . . . . . . 52

3
Part I
Theory to PDEs
1 Introduction and repetition
1.1 Literature:
[Braess, 2007]: Chapter II, 1-3: Sobolev spaces and PDEs;

[Ern and Guermond, 2004]: Appendix A and B: overview of Banach and Hilbert spaces
and functional analysis

[Langtangen and Logg, 2017]: Tutorial to FEniCS.

1.2 Vector spaces, norms, scalar product, and some inequalities


Definition 1.1 (Vector space). Space V over a field of real numbers R is a vector space if there
are two operations

addition of two vectors: u + v V for all u, v V

multiplication with scalar: au V for all a R and u V

(explanation: it means that the space is also closed under those two operations) and following
conditions are satisfied for all u, v V and a, b R:

u+v =v+u

V contains a zero vector 0 V such that u + 0 = u

there exist u V such that u + (u) = 0

1u = u

a(bu) = (ab)u

a(u + v) = au + av and (a + b)u = au + bu

Examples 1.2. The vector space has to be closed for addition and scalar multiplication.
For example the space Rd is closed for those two operations for d N.

Let a R2 , (e.g. a = (2, 1)), then the set {x R2 : a x = 0} is also closed under those
two operations for scalar product .

The set V = {x R2 : kxk2 1} is not a vector space, because it is not closed under
addition nor scalar multiplication, e.g. u = (1, 0) V because kuk2 = 1 but u + u =
(2, 0)
/ V because ku + uk = 2.

Examples 1.3. R, Rn , Lp for p 1, H 1 , Rnn are all vector spaces

Definition 1.4 (Norm). For vector space V , norm is a map k k : V R satisfying for v V
and a R:

(i) kavk = |a|kvk

(ii) subaditivity ku + vk kuk + kvk

4
(iii) kvk = 0 imply v = 0 (zero vector)

Remark 1.5 (Minkowski functional geometrical interpretation of norm). Let K balanced


(aK K for all |a| 1) and convex set, then Minkowski functional: pK (x) = inf{r > 0 : x
rK}, defines a norm. This is a geometrical interpretation of a norm.

Examples 1.6. Let p 1:

for Rn , we have vector norms


(
maxi |xi | for p =
kxkp = P (1.1) {?}
( i |xi |p )1/p otherwise

for Rnn , we have matrix norms

kAxkp
kAkp = sup , (1.2) {?}
xRn ,x6=0 kxkp

which is a norm of linear operator, when we consider the matrix A as a linear operator on
Rn

Definition 1.7 (Scalar product). For vector space V , scalar product is a symmetric positive
definite bilinear map (, ) : V V R, i.e. for u, v, w V and a, b R

(u, v) = (v, u) (symmetric)

(au + bv, w) = a(u, w) + b(v, w) (linear in first argument)

(u, u) 0 with equality if u = 0 (positive-definite)

5
2 Introduction to Lebesgue spaces
2.1 Lebesgue spaces
Definition 2.1 (Lp norm). Let p 1 and f : R be a measurable function. Then
(
ess supx |f (x)| := inf{a : f (x) a for almost all x } for p =
kf kLp () := R 1 (2.1) eq:Lp_norm
( |f (x)|p dx) p otherwise

We usually shorten the notation so we write equivalently kf kLp () = kf kLp = kf kp .

Remark 2.2 (Lp spaces). Let (, , ) be a measure space. Then we define

Lp := {f : R : f is measurable and kf kp < }. (2.2) {?}

However, the function k k : R is not a norm on this space because kf k = 0 do not imply
that f = 0. It only implies that f = 0 almost everywhere. Therefore, we identify all functions
that equal almost everywhere; let

[f ] := {g Lp : f = g almost everywhere}. (2.3) {?}

be a class (set) of those functions. In short words, the functions are equal almost everywhere if
their integrals are the same for different integration domains. For more details, see Appendix A.
Definition 2.3 (Lp space). We define Lp space as a set of function classes equal almost every-
where

Lp = {[f ] : f Lp }. (2.4) {?}

Lemma 2.4 (norm on Lp ). Each map in (2.1) is a norm on corresponding Lp space.


Proof. For a proof, we have to show that mapping k kp satisfy norm properties, i.e.
1 1 1 1
(i) kavk = |a|kvk, i.e. ( |av|p ) p = ( |a|p |v|p ) p = (|a|p |v|p ) p = |a|( |v|p ) p
R R R R

(ii) subaditivity is exactly Minkowski inequality stated in Lemma 2.9.


(iii) kvk = 0 imply v = 0 (zero vector)

Examples 2.5 (One-dimensional integrals). Decide whether the following functions f belong to
Lp (I) for
f (x) = x2 1 on I = (0, 1)
1
f (x) = x on I = (0, 1)
1
f (x) = x on I = (1, )
f (x) = x for R and I = (0, 1) or on I = (1, )
(
log x for = 1
Z
f (x) = x+1
+1 otherwise
R1 R
if = 1 then 0 |f | = and 1 |f | =
R1 R
if > 1 then 0 |f | < and 1 |f | =
R1 R
if < 1 then 0 |f | = and 1 |f | <

6
2.2 Some inequalities
hdef:convexi
Definition 2.6. A function f : R R is convex if f (tx + (1 t)y) tf (x) + (1 t)f (y) for
t [0, 1] and x, y R.
1
Lemma 2.7 (Young inequality). Let p, q (1, ) such that p + 1q = 1 and a, b are nonnegative
numbers. Then
ap aq
ab + (2.5) ?eq:Young_ineq?
p q
Proof. We use the fact that x 7 exp(x) is a convex function. Then
1 1 1 1 ap aq
ab = exp[log a + log b] = exp[ log ap + log bq ] exp[log ap ] + exp[log bq ] = + .
p q p q p q
where the inequality is from Def. 2.6 with x = log ap , y = log bq . and t = p1 .

Lemma 2.8 (Holder inequality). Let f, g : R be measurable functions and p, q 1 be


conjugate values, i.e. p1 + 1q = 1. Then

kf gk1 kf kp kgkq (2.6) eq:Holder_ineq

Proof. The proof is a quite straightforward for p = 1 and q = , when we take essential
supremum. For other cases, we can take u = kffkp and v = kgk g
q
, which has a unit norm
kukp = kvkq = 1. Using the Young inequality, we calculate
Z Z Z
1 p 1
kuvk1 = |uv| |u| + |v|q = 1.
p q
kf gk1
Since kuvk1 = kf kp kgkq 1, we have the required inequality.
hlem:Minkowskii Lemma 2.9 (Minkowski inequality). Let f, g : R be measurable functions and 1 p .
Then

kf + gkp kf kp + kgkp (2.7) ?eq:Minkowski_i

Proof. For the case when p = , the proof comes from the triangle inequality and the properties
of supremum. Otherwise, we can calculate
Z Z Z Z
p p p1 p1
kf + gkp = |f + g| |f + g| (|f | + |g|) = |f + g| |f | + |f + g|p1 |g|

Z  p1 "Z  1 Z 1 #
p p p
|f + g|p |f |p + |g|p

= kf + gkp1
p (kf kp + kgkp )
p
where we have used Holder inequality (2.6) with coefficients p1 and p.

Lemma 2.10. Let be bounded and f Lp () for some p 1. Then f Lq () for any q
such that 1 q p.
Proof. We will proof it by showing that kf kqLq is finite. We can calculate it with the help of
p p
Holder inequality with constants pq , q (it satisfy the assumption of Holder space pq q
p + p = 1)
Z Z
q q q
kf kLq = |f | = 1|f |q k1k pq
p kf k p ,
L
L

where k1k p is finite for finite .


L pq

7
Remark 2.11. The previous lemma fails when the domain is unbounded, e.g. = R. Then
the norm also controls the singularities in infinity. However, some interpolation between Lp
spaces can be established. See following lemma.

Lemma 2.12 (Interpolation between Lp spaces). Assume 1 p r q such that

1 (1 )
= + (2.8) {?}
r s t
and f Lp Lq . Then also f Lr and it holds
T

kf kLr kf kLp kf kLq (2.9) {?}

Proof. [Evans, 2000, Appendix B.2]

2.3 Completeness in the sense of metric spaces


Definition 2.13 (Completeness in the sense of metric spaces). Let V be a vector space with
norm k k. Then it is a complete space if for each Cauchy sequence {fi }iN in V (i.e. for each
> 0, there is N N such that for all m, n > N the following holds kfm fn k < ), there exist
f V such that

fi f in V,

which means limi kfi f k = 0.

Lemma 2.14. The Lp space is a complete space.

Proof. See [Rudin, 1986].

Examples 2.15.

spaces R, N are complete with norm corresponding to the norm given by the absolute value,
i.e. k k := | |.

space Q of rational numbers is not complete space; the proof is by counter example. The
number is irrational. Let set the sequence {xi }
i=1 using elements being the partial decimal
expressions of , i.e.

x1 = 3.1, x2 = 3.14, x3 = 3.141, x4 = 3.1415,

and so on. The the sequence converges to and all xi are rational.

Definition 2.16 (Closure). Let V be a vector space with norm k k. Then we denote by V or
cl V the closure of the set, i.e. the set consisting of all elements of V and also all limiting points
(the points that arise as a limits of all convergent subsequences, convergent with respect to k k).
kk
If we can consider more norms, we write V

Examples 2.17.

Since the Lp space is complete we have Lp = Lp

(0, 1) = [0, 1]
||
R=Q

8
the space of continuous functions C(I) on closed interval I is not complete if we consider
Lp norm with 1 p < . Consider e.g. a function on [0, 1]
(
1 nx for x = [0, n1 ]
fn (x) = ,
0 otherwise

which converges to
(
1 for x = 0
f (x) = .
0 otherwise

9
3 Sobolev spaces
Sobolev spaces are fundamental for the analysis of PDEs because naturally the norm of solution
kukLp and the norm of its gradient kukLp naturally arises. We will establish the relations
between different norms. This is important not only for theoretical analysis of PDEs but also
for numerics, where we can have norms derived e.g. from Lp -norms or Sobolev norms. On finite
dimensional spaces occurring in numerics, all norms are equivalent but there can be a huge
constant of equivalence, which blow up for infinite dimensional spaces spaces of functions.
The Sobolev spaces are also very important for the convergence of FEM method; particularly,
the smoother the solution (measuring in Sobolev norms), the higher convergence rate.

3.1 Prerequisites for Sobolev spaces


3.1.1 Continuous functions
These functions are introduces here because they will be useful for Sobolev spaces; it turns out
that some Sobolev spaces are continuous, see embedding theorem.

Definition 3.1 (Continuous function). Let f : R be a function on a measure space (, , ).


Then f is continuous if for every open interval I, there is an open set A such that f (A) I.

Definition 3.2 (Lipschitz functions). Let f : R be a function satisfying

|f (x) f (y)| C|x y| (3.1) {?}

for all x, y . This function is called Lipschitz continuous.

Definition 3.3 (Holder spaces). Let 0 < 1 and f : R be a function satisfying

|f (x) f (y)| C|x y| (3.2) {?}

for all x, y . This function is called Holder continuous.

Lemma 3.4. This space is a Banach space (complete vector space with norm), where the norm
is defined as

|f (x) f (y)|
kf kC 0, = sup
x,y,x6=y |x y|

Definition 3.5 (Infinitely differentiable functions). Let C () denote the space of infinitely
differentiable functions f : R. Then C0 () is its subset containing functions with compact
support, i.e.

supp f := {x : f (x) 6= 0}. (3.3) {?}

3.1.2 Weak derivatives


u(x)
Definition 3.6. Let u L2 (), then v = xi is a (weak) derivative of u if
Z Z
(x)
v(x)(x)dx = u dx (3.4) eq:weak_derivat
xi
for all test functions C0 , infinitely differentiable functions with zero values on boundary.

Remark 3.7. If a function has a derivative, then classical derivative coincide with a weak
derivative concept; the formula (3.4) is a Greens formula for integration by parts.

10
Example 3.8 (Weak derivative). Show that a derivative of f (x) := |x| on I = [1, 1] is g such
that g(x) = 1 for x 0 and g(x) = 1 for x < 0:
Z 1 Z 0 Z 1
0 0
|x| (x)dx = x (x)dx x0 (x)dx
1 1 0
Z 0 Z 1 Z 1
= 1(x)dx + 1(x)dx = g(x)(x)dx.
1 0 1

Notice that the formula hold independently on the value of g at x = 0! The function g is in
L2 (I).

Example 3.9 (Weak derivative that is a measure). Show that a derivative of g = f 0 on [1, 1]
from previous example has a derivative g 0 = 20 , which equals to Dirac measure concentrated at
zero:
Z 1 Z 0 Z 1
0 0
g (x)dx = 1 (x)dx + 10 (x)dx
1 1 0
= (x)|01 (x)|01 = [(0) (1)] + [(1) (0)] = 2(0),

because (1) = (1) = 0. Now, the derivative of g, is some object g 0 that has to satisfy
Z Z
0
g (x)(x) dx = 2(0) = 20 (x)(x) dx, (3.5) eq:dirac_int
I I

which is exactly the double of Dirac delta measure g 0 = 20 . It can be imagined that it is
something like
(
for x = 0
0 (x) = ,
0 otherwise

but it has only a good meaning when it is evaluated under integral with some test function as in
(3.5).

Remark 3.10 (Multi-index notation). Let : P R with Rd is sufficiently smooth and


1 d
= (1 , 2 , . . . , d ) is multiindex of order || = i i , than D = x 1 . . . .
x d 1 d

3.1.3 Norms on finite dimensional spaces


Definition 3.11 (Equivalent norms). Two norms k k and ||| ||| on a vector space V are
equivalent if there are two positive constants , > 0 such that

kxk |||x||| kxk (3.6) {?}

hold for all x V .

Lemma 3.12 (Norms on finite dimensional spaces). Let there are two norms k k and ||| ||| on
finite dimensional space V . Then those two norms are equivalent.

Example 3.13. Show that k k1 , k k2 , and k k are equivalent norms on Rn .


X
kvk1 = |vi | (3.7) {?}
i

11

Proof. First we prove kvk1 nkvk2
n n
X X
kvk1 = |vi | = 1 |vi | = 1 v k1k2 kvk2 = nkvk2 (3.8) {?}
i=1 i=1

where 1 = (1, . . . , 1) Rn . Then we prove kvk2 nkvk
!1 !1 1
2 2


X X 2
2 2 2
kvk2 = |vi | max |vi | = n max |vi | = n max |vi | = nkvk . (3.9) {?}
i i i
i=1 i=1

Finally, we prove kvk kvk1


n
X
max |vi | = |vj | |vj | + |vi | = kvk1 (3.10) {?}
i
i=1,i6=j

where j is an index where the max norm attain a maximum.

3.2 Definition of Sobolev spaces


For more information see e.g. [Braess, 2007, Chapter II, section 1] or [Evans, 2000, section 5].
Notation: We denote with L2 () space of scalar valued functions f : R. For vector valued
functions f : RN such that Rd is a part (or the whole) d-dimensional space and N N.
We say that f L2 (; RN ) if fi L2 () for all i. This space can be equipped with a norm
Z
kf k = kf (x)kRN dx. (3.11) {?}

Definition 3.14 (Sobolev norms). Let p 1 and D f : R be a measurable function for


any || k. Then
P
||k ess supx |D f (x)|p for p =
kf kW k,p () := P 1 (3.12) ?eq:Sobolev_nor
f (x)|p dx p
R

||k |D otherwise

We usually shorten the notation so we write equivalently kf kL2 () = kf kL2 = kf k2 .

Remark 3.15. We will show soon that above definition is really a norm on a suitable space of
functions.

Remark 3.16. In previous semester, we have introduced the Sobolev spaces in the following
sense
Sobolev_space2i
W 1,p = {v Lp () | v Lp (; Rd )} (3.13a) {?}
W01,p = {v W 1,p () | v(x) = 0 for x } (3.13b) {?}

where we always shorten

H 1 = W 1,2 , H01 = W01,2 .

But what is a meaning of function values on boundary ? It will be explained in section about
trace operator.

12
Definition 3.17 (Sobolev spaces). Let be a subset of Rd with piece-wise Lipschitz boundary.
Then we define for p [1, )
:Sobolev_spacei
kkW 1,p kkW 1,p
W 1,p = C , W01,p = C0 . (3.14a) {?}

Particularly, we have H 1 = W 1,2 and H01 = W01,2 spaces.


Remark 3.18. From the previous definition of Sobolev spaces, it is clear that for each Sobolev
function f and for each n N, there exists a smooth continuous function fn C such that
1
kf fn kH 1
.
n
Remark 3.19 (Equivalence of definitions). The definitions in (3.14) are equivalent with those
in (3.13). It means that for each measurable function f such that kf kH 1 < , there exists
fn C such that kf fn kH 1 n1 .

3.3 Sobolev inequalities & embeddings


There is a question if a function f belongs to W 1,p (), does it also belongs to other spaces such
as Lq ()? The answer depends on p and dimension d. So we distinguish three cases
1 p < d, p = d, d < p .
Lemma 3.20 (Embeddings). Let f W 1,p () for bounded . Then
dp
(i) If 1 p < d, then f Lq () for all q such that 1 q p = dp and

kf kLq () Ckf kW 1,p () , (3.15) eq:ineq_embed


where C depends only on p, d and .
(ii) If p = d, then f Lq () for 1 q < .
d
(iii) If p > d, then f C 0, () for = 1 p (0, 1] and

kf kC 0, () Ckf kW 1,p () . (3.16) {?}


dp
Remark 3.21 (Note on Sobolev conjugate). First, we say that value p = dp is a Sobolev
conjugate for p such that 1 p < d; this value occur in embedding theorems. This Sobolev
conjugate can also be deduced by heuristic analysis (not rigorous). Assume that the inequality
kf kLq () Ckf kW 1,p () , (3.17) {?}
hold for some sufficiently continuous function u C0 on = Rn . We define the rescaled
function
u (x) = u(x) for > 0. (3.18) {?}
Now, we will investigated if the inequality holds if we change a parameter . We calculate
1 p
ku kqLq = kukqLq , ku kpLp = kukpLp . (3.19) {?}
n d
If we put it into inequality, we get
1
kukLq C kukLp (3.20) {?}
d/q d/p
1 dp + dq
kukLq C kukLp (3.21) {?}
dp
But then 1 dp + d
q = 0 giving q = dp , because otherwise we can send to 0 or to get a
contradiction.

13
Remark 3.22 (Embeddings inequality). The inequalities in previous lemma do not only say

that if f W 1,p () than f is in some other space such as Lp but it also says following: if we
have a Cauchy sequence {fi }iN in W 1,p (), than this sequence is Cauchy also in other spaces,
e.g.
kfm fn kLq () Ckfm fn kW 1,p () . (3.22) {?}
for some m, n > N N. Since those spaces are complete, we have
kfm f kLq () Ckfm f kW 1,p () 0 for m . (3.23) {?}
with a limit function f ; the convergence in one space means a convergence in another space,
which is generally not true in infinite dimensional spaces.
re-fried-ineqi?
Lemma 3.23 (Poincare-Friedrichs inequality). Assume that Rd is contained in d-dimensional
cube of side s. Then
kukL2 skukL2 (3.24) eq:poincare-fri
for all u H01 ()
Proof. See [Braess, 2007, Chapter II, 1.5].

This proof holds only for functions with zero values on boundary, or alternatively with zero
mean, see next Remark.
Remark 3.24 (Poincare-Wirtinger inequality). The variant of Poincare inequality is following
inequality
ku u kL2 CkukL2 (3.25) eq:Poincare-Wir
1
R
for u = u denoting the mean over .
Lemma 3.25 (Equivalent norms on H01 ). On H01 () with bounded in s-sized cube, the mapping
u 7 kukL2 (;Rd ) is a norm and it is equivalent norm to k kH01 .
Proof. We will use Poincare-Friedrichs inequality to get estimates
kuk2H 1 = kuk2L + kuk2L s2 kuk2L + kuk2L = (s2 + 1)kuk2L (s2 + 1)kuk2H 1 ,
which reveals
1
kukH 1 kukH 1 kukH 1 .
(s2 + 1)1/2
This relation is exactly required for the equivalent norms. So, in order to show that u 7
kukL2 (;Rd ) is a norm, we have to show homogeneity, triangle inequality, and separation of
points. The homogeneity comes from linearity of gradient and linearity of integration, i.e.
Z 1/2 Z 1/2
2
kaukL2 = kaukL2 = kau(x)k2 dx = a2 ku(x)k22 dx

Z 1/2
2
=a ku(x)k2 dx = akukL2

Triangle inequality comes from Holder inequality. For separation of points, we have to show that
kuk2L = 0 imply u = 0 a.e. But if kuk2L = 0, then from the Poincare-Friedrichs inequality,
we have also k0kL2 = 0, which is a norm and it implies u = 0 a.e.

Remark 3.26. Let I = [a, b] then H 1 (I) C(I).


Outline of proof. See [Braess, 2007, Chapter II, 1.8].

14
3.4 Trace operator values on boundary
Assume that our computational domain Rd has a boundary . For a continuous functions
u C() we can easily consider its values on boundary. The problem arises for Sobolev functions
u W 1,p (), which are generally not continuous. Since has d-dimensional Lebesgue measure
zero, it is not clear what the function value on boundary means. This is solved out with trace
operator.

Lemma 3.27 (Trace theorem). Let be a bounded domain with sufficiently nice boundary
(e.g. C 1 boundary or piece-wise smooth with cone property). Then there exists a bounded linear
mapping T : W 1,p () Lp () for 1 p < such that
\
T u = u| for all u W 1,p () C() (3.26) {?}
kT ukL2 () CkukW 1,p for all u W 1,p () (3.27) eq:ineq_trace

and the constant C depends only on p and .

Definition 3.28. The operator T from previous definition is called trace operator.

Remark 3.29 (Trace as limit of continuous functions). Using the denseness of continuous
functions in Sobolev spaces, for each f H 1 (), there exists sequence of continuous functions
{fn }nN such that fn f in H 1 , i.e. kf fn kH 1 0. Each continuous function has a good
meaning of values on boundary. Therefore, the trace at the boundary can be understood as a
limit of continuous functions at boundary

T f = u| = lim fn | .
n

However, it is necessary to show that the limit is independent of the choice of a sequence; there
are more sequences converging to f .

3.5 Singularities of Sobolev functions


Example 3.30 (L2 functions). Let f (x) : Rd R such that f (x) = kxk2 and integrate over
= {x Rd : kxk2 < 1}.

For d = 1,
(
2
1 2+1 |1 < for > 21
Z
kf k2L2 = 2 |x|2 dx = 2+1 x 0
0 for = 12

Use polar coordinates for d = 2, (r, t) = [r cos t, r sin t] = x, which is a map from (0, 1)
(0, 2) with determinant of Jacobian det = r
Z 1 Z 2 Z 1 (
< for > 1.
Z
2 2 2 2+1
kf kL2 = kxk2 dx = r r dt dr = 2 r dr .
0 0 0 = otherwise

Use polar coordinates for d = 3, (r, t, s) = [r cos t sin s, r sin t sin s, r cos s] = x, which is a
map from (0, 1) (0, 2) (0, ) with determinant of Jacobian det = r2 cos s
(
< for > 23 .
Z Z 1 Z 2 Z Z 1
kf k2L2 = kxk2
2 dx = r 2 2
r sin s dt dr ds = 4 r 2+2
dr .
0 0 0 0 = otherwise

15
for general dimension d, it leads to
(
for > d2 .
Z 1Z 1
<
Z Z
kf k2L2 = kxk2
2 dx = r 2
dS dr = C 2 d1
r r dr .
0 Br 0 = otherwise

where dS denote integration over a ball of radius r, i.e. Br = {x Rd : kxk2 = r}. The
constant C corresponds to the surface area of unit ball.
Example 3.31 (Sobolev
Pd functions). Let f : R be a function on Rd defined again as
f (x) = kxk2 = ( i=1 xi )/2 . The function is measurable so we will show that it belongs to
2

Sobolev space H 1 if the square of Sobolev norm kf k2H 1 = kf k2L2 + kf k2L2 is finite; the crucial
term is kf k2L2 because this term has stronger singularity. The gradient and Euclidean norm is
expressed as
d
X /21
f (x) = /2 x2i 2x, kf (x)k2 = kxk2
2 kxk2 = kxk1
2 .
i=1

So, we can calculate norm of gradient


Z Z
2(1) d
kf k2L2 = kf (x)k22 dx = 2 kxk2 dx < for > 1 . (3.28) {?}
2
To conclude, the singularities for Lp or Sobolev spaces W 1,p are stronger for higher dimensions.
Particularly, this signifies for more complicated analysis of PDEs in three- than two-dimensional
setting.

3.6 Properties of Sobolev spaces


Lemma 3.32. Sobolev spaces H 1 () and H01 () are a Hilbert spaces with inner product:
Z Z
(u, v)H 1 = u(x)v(x)dx + u(x) v(x)dx (3.29) {?}

and induced norm


1
kukH 1 = ((u, u)H 1 ) 2 (3.30) {?}

Proof that the space is complete for addition of two vectors. Let f, g H 1 we will show that
kf + gkH 1 < . It can be shown using Mikowski inequality

kf + gkH 1 = kf + gkL2 + k(f + g)kL2 = kf kL2 + kgkL2 + kf kL2 + kgkL2 = kf kH 1 + kgkH 1

Proof that norm is really a norm. I will show only that H 1 -norm, to show that (, )H 1 is an
inner product it follows similarly. Hence
 1/2
kukH 1 () = kuk2L2 () + kuk2L2 (;Rd ) ,

is a norm for u : R where Rd . It is necessary to confirm following requirements:


homogeneity: for c R it holds kcukH 1 () = |c| kukH 1 () because
 1/2  1/2
kcukH 1 () = kcuk2L2 () + kcuk2L2 (;Rd ) = |c|2 kuk2L2 () + |c|2 kuk2L2 (;Rd )
 1/2
= |c| kuk2L2 () + kuk2L2 (;Rd ) = |c| kukH 1 () .

16
triangle inequality: ku + vkH 1 () kukH 1 () + kvkH 1 () is based on triangle inequality
(Minkowski inequality) of L2 () norm, i.e.
 1/2
ku + vkH 1 () = ku + vk2L2 () + ku + vk2L2 (;Rd )
 1/2
(kukL2 () + kvkL2 () )2 + (kukL2 (;Rd ) + kvkL2 (;Rd ) )2 .

Then, the triangle inequality of Euclidean norm, i.e.

for a, b R2 kak2 = (a21 + a22 )1/2 ka + bk2 kak2 + kbk2 ,

is used for a = (kukL2 , kukL2 ) and b = (kvkL2 , kvkL2 ):


 1/2
ku + vkH 1 () (kukL2 () + kvkL2 () )2 + (kukL2 (;Rd ) + kvkL2 (;Rd ) )2
 1/2  1/2
kuk2L2 () + kuk2L2 (;Rd ) + kvk2L2 () + kvk2L2 (;Rd )
= kukH 1 () + kvkH 1 () .

separates points: kukH 1 () = 0 implies u(x) = 0; it follows the same arguments as for L2
space.

17
4 Hilbert and Banach spaces
Definition 4.1 (Hilbert space). Let V be a vector space over R with scalar product. Then V is
Hilbert space if it is complete space.
Lemma 4.2 (Cauchy-Schwartz p inequality). The following hold (u, v) kukkvk for norm in-
duced by scalar product kuk = (u, u).
Proof. For v = 0, the proof is trivial, so we can assume v 6= 0. Then

u v 2 kuk2 2(u, v) kvk2 2(u, v)


0k k = + =2 , (4.1) {?}
kuk kvk kuk2 kukkvk kvk2 kukkvk
(u,v)
which gives kukkvk 1.

p The triangle inequality ku + vk kuk + kvk holds for norm


Lemma 4.3 (Triangle inequality).
induced by scalar product kuk = (u, u).
Proof. Using Schwartz inequality, we prove
ku + vk2 = (u + v, u + v) = kuk2 + 2(u, v) + kvk2 kuk2 + 2kukkvk + kvk2 = (kuk + kvk)2 .

Lemma 4.4. A norm k k R is a continuous function.


Proof. From triangle inequality, we have
ku + vk kvk kuk. (4.2) {?}
Now, we use substitution w = u + v to get
kwk kvk kw vk. (4.3) {?}
Since the same hold if we interchange the variables, i.e. kvk kwk kv wk = kw vk, the
inequality also holds with absolute value on left-hand side

kwk kvk kw vk. (4.4) {?}
This proves the continuity of the norm, i.e. for all > 0, such that

kw vk < kwk kvk , (4.5) {?}
which is shown for := . The small changes of the arguments result in small changes in
response.
p
Example 4.5. Show that norm induced by a scalar product, i.e. kuk = (u, u) is really a norm
(show that it satisfies the axioms of norm).
Examples 4.6.
spaces R, R are Hilbert spaces
space C is also a Hilbert space but with scalar product that is sesquilinear
L2 () with scalar product (u, v) = uv is a Hilbert space
R

Definition 4.7 (Banach space). Let V be a vector space over R with norm. Then V is Hilbert
space if it is complete space.
Examples 4.8.
Lp () for p 1 with norm (2.1) is a Banach space.

18
4.1 Linear operators, linear functionals, and dual space
Definition 4.9. Let V, W be two vector spaces over R. Then a map L : V W is linear if

L(au + bv) = aL(u) + bL(v) (4.6) {?}

holds for all a, b R and u, v V .

Definition 4.10 (Bounded linear operators, norm). Let V, W be two Banach spaces. Then
linear mapping L : V W is bounded linear operator if there exists C > 0 such that

kL(u)kW CkukV for all u V. (4.7) {?}

For simplicity, we also write Lu instead of L(u). The norm of the operator is defined as

kLk = sup kLukW . (4.8) {?}


uV,kukV =1

Lemma 4.11. A linear operator is bounded if and only if is continuous.

Proof. Assume that an operator L is bounded. Then the continuity follows from inequality

kL(u) L(ui )kW = kL(u ui )kW Cku ui kV 0 for i . (4.9) {?}

Conversely, let an operator is continuous. Then, from definition of continuity, for each > 0
there exists such that following hold for all khk

kL(h)kW = kL(h) L(0)kW < . (4.10) {?}

So from linearity and homogeneity of norm, we have the estimate for arbitrary v V

kvkV v kvkV v
kL(v)k = k L( )k = kL( )k kvkV . (4.11) {?}
kvkV kvkV

Definition 4.12 (Linear functionals, dual space). For W = R from previous definition, we call
the map L continuous linear functional. Then the dual space V of V is a space of all bounded
linear functionals, i.e.

V = {f : V R; kf k < }.

We then also equivalently write





L(u) = L, u = L, u V V

Definition 4.13 (Separable Hilbert space). Hilbert space H is separable if there exists countable
basis.

19
5 Existence of solutions to elliptic PDEs
In this section, we will focus on existence results of PDEs, which will be treated within the
following abstract weak formulation.

5.1 Abstract formulation


ef:abstract_WFi
Definition 5.1 (Abstract weak formulation). Let V be a Banach space (complete vector space
with norm kk), a : V V R be a bilinear form that is uniformly elliptic with positive constant
> 0, i.e.

kuk2 a(u, u) u V (5.1) {?}

and bounded (continuous) with constant <

a(u, v) kukkvk u, v V. (5.2) {?}

And let F : V R, i.e. F V , be a continuous linear functional

F (v) kF k kvk v V. (5.3) {?}

Then the following problem is called abstract weak formulation: Find u V satisfying

a(u, v) = F (v) for all v V. (5.4) {?}

5.1.1 Minimization problem


According to [Braess, 2007, Theorem 2.2]:
equiv_weak_mini Lemma 5.2 (Equivalence to minimization problem). Assume the assumption of abstract weak
formulation and additionally symmetricity of bilinear form, i.e.

a(u, v) = a(v, u) u V. (5.5) {?}

Then the following linear functional J : V R defined as


1
J(v) = a(v, v) F (v) (5.6) {?}
2
attains an minimum over V at u V , i.e. writting it as minimization problem

u = arg min J(v), (5.7) {?}


vV

if and only if the following weak formulation holds

a(u, v) = F (v) v V. (5.8) eq:weak_abstrac

Proof. For u, v V and t R, we have : R R such that


1
(t) = J(u + tv) = a(u + tv, u + tv) F (u + tv)
2
t2
= J(u) + t[a(u, v) F (v)] + J(v, v)
2
If u V satisfies (5.8), then for t = 1
1
J(u + v) = J(u) + J(v, v) > J(u) v V, v 6= 0.
2

20
Hence, u is a unique minimizer.
Conversely, if J has a minimum at u V , then the derivative of must vanish for t = 0 for
arbitrary v V , i.e.
0 (0) = a(u, v) l(v) = 0.
Since v V is arbitrary, we have (5.8).

Lemma 5.3. Functional J from previous lemma is continuous.


Proof. For a sequence {vi }i=1 converging to v, we will show that J(vi ) converges to J(v). Hence
1 1
|J(vi ) J(v)| = | a(vi , vi ) F (vi ) a(v, v) + F (v)| (5.9) {?}
2 2
1 1
|F (v) F (vi )| + | a(vi , vi ) a(vi , v) a(v, v)| (5.10) {?}
2 2
1 1
= |F (v vi )| + | a(vi , vi v) + a(vi v, v)| (5.11) {?}
2 2

Ckv vi kV + kvi kV kvi vkV + kvi vkV kvkV . (5.12) {?}
2 2
Since kv vi kV converges to zero and kvkV and kvi kV are uniformly bounded, we have
lim |J(vi ) J(v)| = 0.
i

5.2 Existence of solution


5.2.1 Existence from minimisation problem
Theorem 5.4 (Existence theorem from minimization problem). Assume the assumptions from
Lemma 5.2. Let V be a closed convex set in V . Then there exists unique solution of the
minimization problem
u = arg min J(v). (5.13) {?}
vV

Proof. First, we have to show that the functional J is bounded from below
1 kf k2 kf k2
J(v) kvk2V kf k kvkV = (kvkV kf kV )2 .
2 2 2 2
Then, we denote M = inf vV J(v), and let {vi }iN be an infimizing sequence (limi J(vi ) =
M ). Then we calculate
kvn vm k2V a(vn vm , vn vm ),
= 2a(vn , vn ) + 2a(vm , vm ) a(vn + vm , vn + vm ),
vn + vm vn + vm vn + vm
= 2a(vn , vn ) 4F (vn ) + 2a(vm , vm ) 4F (vm ) 4a( , ) + 8F ( ),
2 2 2
vn + vm
= 4J(vn ) + 4J(vm ) 8J( ),
2
4J(vn ) + 4J(vm ) 8M ;
the first inequality comes from the ellipticity assumption, the last inequality comes from the
fact that V is a convex, which implies vn +v
2
m
V and thus M J( vn +v
2
m
). Since {vi }iN is an
infimizing sequence, we have
kvn vm k2V 0 for m, n ,

21
which means that {vi }iN is a Cauchy sequence in V , and its limit exists
u = lim ui
i

since the space is complete. Since V is closed then also u V . The continuity of J implies that
J(u) = lim J(vi ) = inf J(v).
i vV

We now show the uniqueness of the solution. Assume that there are two solutions u1 , u2 V .
Then u1 , u2 , u1 , u2 , . . . is also an infimizing sequence, which has to be Cauchy sequence implying
u1 = u2 .

5.2.2 Lax-Milgram and some properties


Lemma 5.5 (Lax-Milgram). There is a unique solution of abstract weak formulation in Def. 5.1,
(no symmetry of the bilinear form is needed).
ori_continuousi Remark 5.6 (A priori estimates). Assuming abstract weak formulation, we have
kuk2V a(u, u) = F (u) kF k kukV , (5.14) {?}
which implies that
kF k
kukV . (5.15) {?}

kF k
It means that the solution has to be contained in a ball of radius , i.e.
kF k
u {v V : kvk } (5.16) {?}

Proof of Lax-Milgram lemma: See [Evans, 2000, section 6.2] or [Braess, 2007, Chapter III, sec-
tion 3].

5.3 Applications to PDEs


Here, we apply the abstract theory to concrete PDEs. In the later, we assume that our compu-
tational domain is a nice (reasonable) open subset of Rd in dimension d. All the PDEs, will
be cast into abstract variational formulation Weak formulation: find u V satisfying
a(u, v) = F (v) v V.
It means that for each PDE, we have to determine bilinear form a, linear functional F , and trial
space V , which fits into Lax-Milgram lemma. Moreover, the source terms f , material coefficients
A have to be specified. Particularly, the following notation will be used
f L2 () source term
c L (, Rdd ) scalar material coefficients that are uniformly positive and bounded,
i.e. there exists 0 < < < such that
c(x)

A L (, Rdd ) tensor valued material coefficients that are uniformly positive and
bounded, i.e. there exists 0 < < < such that
kuk22 A(x)u u, A(x)u u kuk2 kvk2

holds for all u Rd .

22
5.4 Derivation of convection-diffusion equation
we want to derive a convection-diffusion equation:

u(x, t) D(x)u(x, t) + [v(x, t)u(x, t)] = f (x, t), (5.17) eq:conv-diff_eq

where

u(, t) : R is a concentration density (concentration per volume) at time t;

D(, t) : R is a given diffusion parameter, time independent;

v(, t) : Rd is a given velocity field at time t

f (, t) : Rd is a given source at time t

We derive it from continuity equation. We start from total concentration at some specimen
Rd , i.e. Z
Q= u(x, t) dx.

Then we derive the heat equation from postulate: change of total concentration in time
equals to a concentration flux through boundary and a source. It can be written as
Z Z
Q = n q(x, t) + f (x, t) dx (5.18) {?}

where q : R3 is a concentration flux density. Then we use Fourier law q(x) = D(x)u(x, t)+
v(x, t)u(x, t), which is a material
R law obtained Rfrom e.g. from observation or experiments, and
divergence (Gauss) theorem: n f (x) ds = f (x) dx to derive
Z Z
u(x, t) D(x)u(x, t) + [v(x, t)u(x, t)]dx = f (x, t)dx. (5.19) {?}

Since Q is defined on domain constant in time, we moved derivative into the integral. Then we
use following lemma, from which we can deduce the equation (5.17).

Lemma 5.7. Let be arbitrary set and the following hold


Z
f (x) dx = 0. (5.20) {?}

Then f (x) = 0 almost everywhere.


R R
Note: Compare divergence theorem to Stokes theorem: tf (x) ds = m(f (x)) dx.

5.4.1 Boundary conditions


initial boundary conditions for u(x, t) prescribes u(x, 0) for x
Dirichlet boundary conditions u(x, t) = uD for D
Neumann boundary conditions n (x)u(x, t) = qN for x D

23
5.4.2 Heat equation homogeneous Dirichlet boundary conditions
Strong formulation: find u C 2 () C 0 () such that
T

A(x)u(x) = f (x) for x


u(x) = 0 for x
Weak form:
V = H01 ()
kuk = kukH 1
Z
a(u, v) = A(x)u(x) v(x) dx
Z

F (v) = f (x) v(x) dx



Lemma 5.8. There exists a unique solution to above heat equation.
Proof. We will proof the existence of unique solution with Lax-Milgram lemma; we have to
satisfy its assumptions. Linearity of bilinear of a and linearity of F comes from the properties
of integrals. Ellipticity follows from
Z
a(u, u) = A(x)u(x) u(x) dx
Z

u(x) u(x) dx = kuk2L2 = kuk2L2 + kuk2L2
2 2
2 2 2
kukL2 + kukL2 min{ , }kukH 1
2 2s 2 2s
where the last inequality comes from Poincare-Friedrichs inequality (3.24). Boundedness of
bilinear form comes from Holder inequality and properties of A
Z Z
a(u, v) = A(x)u(x) v(x) dx u(x) v(x) dx = ku vkL1 (;Rd ) ,

kukL2 (;Rd ) kvkL2 (;Rd ) kukH 1 () kvkH 1 () .
Boundedness of F comes from Holder inequality
Z
F (v) = f (x) v(x) dx = kf vkL1 kf kL2 kvkL2 kf kL2 kvkH 1 .

Lemma 5.9. The solution continuously depends on load. It means that for two loads f, f L2
and corresponding solutions
a(u, v) = F (v) := (f, v)L2 , (5.21) {?}
a(u, v) = F (v) := (f, v)L2 , (5.22) {?}
we have
1
ku ukH kf fkL2 (5.23) {?}

Proof.
ku uk2H a(u u, u u) = a(u, u u) a(u, u u) = (f, u u)L2 (f, u u)L2
= (f f, u u)L2 kf fkL2 ku ukL2 kf fkL2 ku ukH

24
5.4.3 Helmholtz equation
Strong formulation: find u C 2 () C 0 () such that
T

u(x) + b(x)u(x) = f (x) for x


u(x) = 0 for x

Weak form:

V = H01 ()
kuk = kukH 1
Z Z
a(u, v) = u(x) v(x) dx + b(x)u(x)v(x) dx
Z

F (v) = f (x) v(x) dx


Lemma 5.10. There exists a unique solution to above heat equation for b L () and b(x)
b0 > 0.

Proof. We will proof the existence of unique solution with Lax-Milgram lemma; we have to satisfy
its assumptions. Linearity of bilinear form a and linearity of F comes from the properties of
integrals. Ellipticity follows from
Z Z
a(u, u) = u(x) u(x) dx + b(x)u(x)u(x) dx
Z Z

u(x) u(x) dx + b0 u(x)u(x) dx min{1, b0 }kuk2H 1 .


Boundedness of bilinear form comes from Holder inequality


Z Z
a(u, v) = u(x) v(x) dx + b(x)u(x)v(x) dx
Z Z
u(x) v(x) dx + kbkL u(x)v(x) dx

kukL2 kvkL2 + kbk L kukL2 kvkL2 max{, kbkL }(kukL2 kvkL2 + kukL2 kvkL2 )
= max{, kbkL }(kukL2 , kukL2 ) (kvkL2 , kvkL2 )
max{, kbkL }k(kukL2 , kukL2 )k2 k(kvkL2 , kvkL2 )k2
max{, kbkL }kukH 1 kvkH 1 .

where we have used that the Sobolev norm can be written as

kuk2H 1 = kuk2L2 + kuk2L2 = k(kukL2 , kukL2 )k22 .

Boundedness of F comes from Holder inequality (the same like previous proof for heat
equation)
Z
F (v) = f (x) v(x) dx = kf vkL1 kf kL2 kvkL2 kf kL2 kvkH 1 .

25
5.4.4 Helmholtz equation with Neumann boundary conditions
Strong formulation: find u C 2 () C 0 () such that
T

A(x)u(x) + u(x) = f (x) for x


n A(x)u(x) = uN (x) for x
Weak form for A(x) = 1:
V = H 1 ()
kuk = kukH 1
Z Z
a(u, v) = u(x) v(x) dx + u(x) v(x) dx
Z Z

F (v) = f (x) v(x) dx + uN (x)v(x) ds x



Here ds x means a measure on boundary .
Lemma 5.11. There exists a unique solution to above heat equation for uN L2 ().
Proof. The proof is similar like for the previous Helmholtz equation with Dirichlet boundary
conditions. Here we have to proof boundedness (continuity) of F . Hence
F (v) = kf vkL1 () + kuN vkL1 ()
kf kL2 () kvkL2 () + kuN kL2 () kvkL2 ()
Ckf kL2 () kvkH 1 () + CT kuN kL2 () kvkH 1 ()
= (Ckf kL2 () + CT kuN kL2 () )kvkH 1 ()
where the constant C comes from inequality (3.15) and CT from inequality of trace operator
(3.27).

5.4.5 Neumann problem


Strong formulation: find u C 2 () C 0 () such that
T

u(x) = f (x) for x


n A(x)u(x) = uN (x) for x
Weak form:
Z
1
V = {v H () : v = 0}

kuk = kukH 1
Z
a(u, v) = u(x) v(x) dx
Z Z
F (v) = f (x) v(x) dx + uN (x)v(x) ds x

Here ds x means a measure on boundary .
Lemma 5.12. There exists a unique solution to above heat equation for uN L2 () if
Z Z
f+ uN = 0. (5.24) {?}

Proof. The proof is similar like for previous examples. The only difference is for ellipticity
condition, where a Poincare-Wirtinger (3.25) inequality has to be used.

26
Part II
Discretisation of variational problems
6 Abstract approaches
6.1 Discretisation approaches
Assume the abstract variational formulation from Definition 5.1. Since the corresponding
trial/test space V is infinite dimensional, we have to approximate this problem.
Here, we will mainly talk about conforming methods. It is based on approximation of the
infinite dimensional space V with a discrete one Vh that is a subspace

Vh V ; (6.1) {?}

otherwise we talk about non-conforming approximation (e.g. discontinuous Galerkin method).


For example, H01 (I) Sobolev space on unit interval is approximated with space of trigonometric
polynomials Vh = span{k |k (x) = sin(2ikx), k Z, h1 < k < h1 }, or finite element space
(space of consisting of continuous functions that are piece-wise polynomials).

Definition 6.1 (Galerkin method or also Bubnov-Galerkin method). Find uh Vh V such


that

a(uh , v) = F (v) v Vh . (6.2) {?}

Definition 6.2 (Ritz method). Find uh Vh V such that


1
uh = arg min a(vh , vh ) F (vh ) (6.3) {?}
vh Vh |2 {z }
J(vh )

Remark 6.3 (Ritz-Galerkin method). Since the Ritz method corresponds to Galerkin method
when the bilinear form is symmetric, we often talk about Ritz-Galerkin method.

Although, we focus on Ritz-Galerkin method, there are also other approaches, e.g. when the
trial and test space are different.

Definition 6.4 (Petrov-Galerkin method). Find uh Vh V such that

a(uh , v) = F (v) v Wh V. (6.4) {?}

Remark 6.5. Using a priori estimates, based on ellipticity and continuity (boundedness) of
linear functional, we can show that the approximate solutions of Galerkin approximations are
bounded
kF k
kuh k . (6.5) {?}

See Remark 5.6 for a case in a continuous setting.

Lemma 6.6 (Ceas lemma). The solution of Galerkin approximation satisfies

C
ku uh kV inf ku vh kV , (6.6) {?}
vh Vh
with a constant ellipticity constant and continutity constant C of bilinear form.

27
Proof. The solution u V and approximate solution uh Vh satisfy

a(u, v) = F (v) v V,
a(uh , v) = F (v) v Vh .

Their substraction reveals Galerkin orthogonality

a(u uh , v) = 0 v Vh ,

with respect to subspace Vh V . Therefore for arbitrary vh Vh and a special choice of


v = vh uh Vh , we have a variant of it, particularly a(u uh , vh uh ) = 0. This can be used
to prove the estimate using the V-ellipticity

ku uh k2V a(u uh , u uh ) = a(u uh , u vh ) + a(u uh , vh uh )


Cku uh kV ku vh kV .

The inequality
C
ku uh kV ku vh kV (6.7) {?}

holds for all vh Vh and therefore it has to hold also with infimum.

Remark 6.7 (Ceas lemma in symmetric case energetic-like estimate). When the bilinear
form a is symmetric, it is also a scalar product, which induces an energetic norm
p
kvkA = a(v, v).

Therefore, for v = u uh , we have

ku uh k2A = a(u uh , u uh ) ku uh kA ku vh kA

where we have used Galerkin orthogonality and Cauchy-Schwartz inequality. Therefore, in en-
ergetic norm, the estimate is without a constant, i.e.

ku uh kA inf ku vh kA ,
vh Vh

and FEM thus provides the best approximation in an energetic norm.


This result can be also used to improve the estimate in V norm, i.e.

ku uh k2V a(u uh , u uh ) = ku uh k2A ku vh k2A Cku vh k2V .

It does not provide the better estimate


r
C
ku uh kV inf ku vh kV .
vh Vh
but it also says that energetic norm is equivalent to the norm V , typically some Sobolev norm.

Remark 6.8. Ceas lemma is a base for a proof of convergence of approximate solutions to the
continuous one. For details, see a corresponding section below.

28
7 Finite element method in general setting
Here, we will focus on some elements of a conforming method, i.e. when the approximation
space Vh is contained in the trial space V , written as Vh V .
For additional informationf, I refer to lecture; see also attached lecture notes written by Dr.
Noemi Friedman or see a literature for a detailed overview.

7.1 Basic aspects of FEM


For a set of finite element T , the following has to be satisfied

for each T T , the set T is closed and interior is nonempty and connected

the boundary of each element is Lipschitz continuous

the intersection of two elements is either empty set or a common vertex, edge, or facet.
For conforming method

Now we have a look on the fact why we need continuous basis function (continuous over
elements) for approximation of Sobolev functions.

Lemma 7.1. Let Vh be composed of polynomials over element, which are continuous on facets
(edges in 2D or faces in 3D), i.e. Vh C 0 (). Then

Vh H 1 () (7.1) {?}

Proof. Since any function v Vh is continuous, it is also in L2 . In order to show that this
function is in H 1 , we have to find its weak derivative and show that it is integrable. Notice that
this function generally does not have normal derivative on facets. I.e. we have to find wi (as a
weak partial derivative w.r.t. xi ) such that
Z Z
(x)
wi (x)(x)dx = v(x) dx. (7.2) {?}
xi

We can calculate using Greens theorem


Z XZ
(x) (x)
v(x) dx = v(x)|T dx (7.3) {?}
xi T xi
T T
X Z v(x)|T XZ
= (x)dx + vi (x)|T (x)dx (7.4) {?}
xi
T T T T T T
| {z }
=0
Z
= wi (x)(x)dx (7.5) {?}

where the sum over all boundary integrals cancels because of continuity over elements and zero
values of at . The weak derivative is just a function with derivatives defined in interior
points of the elements.

Remark 7.2. For conforming approximation of H 1 space, the approximation space has to con-
sists of continuous functions. Similarly approximation of H 2 has to consists of functions with
continuous derivatives over elements. On other side L2 can be approximated with discontinous
Galerkin space (no continuity over elements is required).

29
7.2 Simplicial finite elements
This FEM space is based on complete polynomials:

X
Pk = p : Rd R | p(x) = c x1 i . . . xd d (7.6) {?}
P
; i i <k

The dimension of such space is


  Qd
d+k (k + i)
dim Pk = = i=1 (7.7) {?}
k k!
The simplicial finite element is based on simplex element T (triangle, tetrahedron, etc.),
which can be obtained as a convex combination of 3, 4, etc. points. The FEM space approxi-
mating H 1 can be expressed as

Vh = {v C 0 (Rd ) | v|T Pk } H 1 . (7.8) {?}

In Fenics, this space is called Lagrange, CG, or Continuous Galerkin.


The basis of such functions can be expressed in barycentric coordinates, for details see e.g.
[Ciarlet, 1991]. The numerical integration is not so trivial as for rectangular finite elements and
can be found e.g. in [Wandzurat and Xiao, 2003].
Isoparametric mappings:

7.3 Rectangular finite elements


This space is based on tensor product space.

X
Qk = p : Rd R | p(x) = c x1 i . . . xd d (7.9) {?}

;i <k

The dimension of such space is

dim Qk = (k + 1)d . (7.10) {?}

Those polynomials are uniquely determined on a set


i1 id
Mk = {x = ( , . . . , ) | ij {0, 1, . . . , k} for j = 1, . . . , d} (7.11) {?}
k k
Isoparametric mappings:

7.4 Convergence of finite element method


The convergence of the finite element method depends on the regularity (smoothness) of the
solution. The more smooth the solution is, the higher convergence rate can be expected.
Generally, we only have a convergence that can be arbitrarily slow. It is possible to prove
that

ku uh kH 1 0 for h 0 (7.12) {?}

where h is a characteristic size of some regular mesh (some non-degenerate mesh see literature
for details, e.g. [Gockenbach, 2006]). By hT , we denote the diameter of the element

hT = {kx yk2 | x, y T }. (7.13) {?}

30
The the characteristic size of the mesh can be defined as the maximum of hT over all elements,
i.e.

h = max{hT | T T }, (7.14) {?}

where T is a set of all elements.


If the solution of the PDE is smoother than the convergence can be with some rate.

Theorem 7.3. Let an elliptic PDEs, defined for a trial space H 1 , has a solution from H s for
s > 1. Then the following estimates with a suitable constants can be derived

ku uh kL2 Chs |u|H s (7.15) {?}


s1
ku uh kH 1 Ch |u|H s (7.16) {?}

when we use a polynomial approximation of order at least s. The value s is called an order of
convergence; for s = 1 it is linear convergence, for s = 2 quadratic convergence, etc. It means
that for quadratic convergence we need at least quadratic basis functions.

7.5 Influence of numerical integration


In the FEM and particularly in elliptic problems, we have to evaluate bilinear form a(u, v) and
a linear functional F (v) for a trial and test functions u and v, respectively. The evaluation
consists of evaluating integrals and is typically provided element-wise
Z XZ
(x)dx = (x)dx (7.17) {?}
T T T

along with the numerical integration at elements


Z n
X
(x)dx wi (xi ). (7.18) {?}
T i=1

This integration (typically over reference elements) with integration points xi and corresponding
weights wi introduces another numerical error (first error is introduced by Galerkin approxima-
tion).
Therefore, the bilinear forms and linear functional are approximated by

a(u, v) ah (u, v), F (v) Fh (v), (7.19) {?}

where u, v are trial and test functions from approximation (FEM) spaces.
The error in Galerkin approximation is controlled by Ceas lemma. In the case of numerical
integration, we have its generalisation: first Strang lemma
" #
a(vh , wh ) ah (vh , wh ) F (wh ) Fh (wh )
ku uh k inf ku vh k + sup + sup (7.20) {?}
vh Vh wh Vh kwh k wh Vh kwh k

Observe that for an exact integration, the first Strangs lemma reduces to the Ceas lemma. The
additional terms are called inconsistency error.
There is also a second Strangs lemma, which is useful for non-conforming methods (the
approximation space is not contained in the trial space Vh * V ) such as discontinuous Galerkin
method (when the finite element basis functions are not continuous over elements they are not
in a Sobolev space).

31
Part III
Saddle point problems
This topic is covered by the following book [Braess, 2007], particularly by chapter III, paragraph
4 Saddle point problems and paragraph 5 Mixed methods for the Poisson equation.

8 Introduction
8.1 Motivation example
vation-examplei
Let assume the minimisation of the following function J : Rd R defined as
1
J(v) = (v12 + v22 ) v1 v2 2v1 v2 = Av v b v (8.1) {?}
2
for
 
2 1
A= , b = (2, 1). (8.2) {?}
1 2

8.1.1 Minimisation problem without a constraint


Minimisation problem without a constraint

u = arg min f (v) (8.3) {?}


vRd

is equivalent to weak formulation Au v = b v for all v Rd , which is also equivalent to the


solution of linear system for u R2

Au = b, u = A1 b = (0.83, 0.6) (8.4) {?}

8.1.2 Minimisation problem with a constraint


Assume the same problem as in previous section, but now with a constraint x1 + 3x2 = 2 which
can be written as

Bu = g for B = (1, 3) and g = 2 (8.5) {?}

or in a weak form

Bu w = g w w R. (8.6) {?}

Minimisation problem with a constraint is equivalent to minimisation

u = arg min f (v) (8.7) {?}


vVg

over a space
Vg = {v R2 | Bv = g}.
However, this is not a vector space!
The solution can be searched using Lagrange multipliers

(u, ) = arg min arg max L(v, ), (8.8) {?}


vR2 R

L(v, ) = J(v) + (Bv g) . (8.9) {?}

32
Similarly this min-max problem is equivalent to the linear system:

A BT
    
u b
= (8.10) {?}
B 0 g

with constraint solution x = (0.35714286, 0.28571429) and Lagrange multiplier.

8.1.3 Minimisation with a constraint that do not work


Here, we show that not all constraint optimisation problems are well-defined. Assume that we
have two constraints

u1 + 3u2 = 2, 2u1 + 6u2 = 4, (8.11) {?}

where the first constraint is the same as in initial constraint problem and second constraint is
just its multiple. The minimisation problem

u = arg min J(v) (8.12) {?}


vVg

is well-defined because the minimisation space Vg remains the same. However, the min-max
problem with Lagrange multipliers

min max J(v) + (Bv g) 1 + (2Bv 2g) 2 , (8.13) {?}


vR2 R2

do not have unique Lagrange multipliers because any combination = 1 + 22 with from
previous example is a possible solution. This is reflected in corresponding linear system

A B T 2B T

x f
B 0 0 1 = g , (8.14) {?}
2B 0 0 2 2g

where the matrix is singular (see that the last two equations are linearly dependent).
Alternatively, the corresponding constraint

u1 + 3u2 = 2, 2u1 + 6u2 = 2, (8.15) {?}

is satisfied by no vectors and thus there is no solution.


The characterisation of well-posed saddle point problems is needed and will be investigated
in the following sections.

8.2 Applications
optimisation

pure Neumann problem

Stokes problem

dual problems

a posteriori error estimates

and many others

33
9 Abstract formulations
9.1 Introduction
In the motivation section 8.1, we have seen that some constraint minimisation problems leads
to linear system (saddle-point problems) that are not-defined. The proper treatment will be
developed here in abstract setting that also covers the PDEs with constraint.

9.2 Model problem


Two Hilbert spaces X and M , two bilinear forms a : X X R and b : X M R, and two
linear functionals f : X R and g : M R. We typically write that those linear functionals
are from the corresponding dual space f X 0 and g M 0 and write





f (x) = f, x X 0 X = f, x , g(x) = g, x M 0 M = g, x . (9.1) {?}
We consider a minimisation of a quadratic functional over X
1
J(v) = a(v, v) f (v) (9.2) {?}
2
subject to constraint on X
b(u, w) = g(v) w M. (9.3) {?}
Since the constraint minimisation corresponds to minimisation over some subspaces, we
define the following spaces
Vg = {v X | b(v, ) = g() M }, (9.4) eq:Vspace
V = {v X | b(v, ) = 0 M }, (9.5) {?}
which will be found useful later on.
Problem 9.1 (Minimisation with a constraint). The minimisation problem can be written as
minimisation over a constraint space:
min J(v) = J(u) (9.6) {?}
vVg

where u is called a minimiser. However, it is usually not easy to find a basis for V to allow
direct numerical approximation; the space Vg is not a vector space. Therefore, we can write it
as
min J(v) + Vg (v) = J(u) + Vg (u) = J(u) (9.7) {?}
vX

where Vg : X R is an indicator function of a set Vg , and is defined as


(
0 for u Vg
Vg = . (9.8) {?}
otherwise
Therefore this indicator function enforces the constraint and minimisation space can be enlarged
to the whole X.
Problem 9.2 (Lagrange multipliers). Still, the last formulation is not suitable for approximation
because numerically it is difficult to deal with indicator function. Another possibility is to come
to the Lagrange multipliers or min-max problem, i.e.
min max J(v) + b(v, ) g() . (9.9) {?}
vX M | {z }
L(v,)

34
hprob:saddlei Lemma 9.3 (Saddle-point problem). From the minmax problem with Lagrangian, stationary
equation (optimality condition) is expressed as: Find (u, ) such that
:saddle_systemi
a(u, v) + b(v, ) = f (v) v X (9.10a) {?}
b(u, ) = g() M (9.10b) {?}

Proof. Let define a function of two variables : R2 R defined as

(r, s) = L(u + rv, + s)

The gradient of thus defines the (Gateaux) directional derivative of Lagrangian L at point
(u, ) in the direction of (v, ).
We can calculate split the Lagrangian

r2
L(u + rv, + s) = L(u, ) + r[a(u, v) + b(v, ) f (v)] + a(v, v) + s[b(u, ) g()],
2
which is used to calculate partial derivatives; one w.r.t variable r is then
r2
L(u + rv, + s) L(u, + s) r[a(u, v) + b(v, ) f (v)] + 2 a(v, v)
r (r = 0, s) = lim = lim
r0 r r0 r
= a(u, v) + b(v, ) f (v).

Analogically, one can calculate partial derivative w.r.t. variable s:

r (r, s = 0) = b(u, ) g()

The derivatives have to be equal to zero for all possible directions (v, ), which gives the required
saddle-point system (9.10).

Remark 9.4. From the min-max problem, we can see that we have saddle-point at (u, )

L(u, ) L(u, ) L(v, ) for all , v.

Theorem 9.5 (Existence theorem). Assume that we have a saddle-point problem 11.2 and
spaces (9.4) with following assumption:

V-ellipticity of bilinear form a for positive constant > 0

a(v, v) kvkX v V (9.11) {?}

inf-sup (Babuska-Brezzi, BB, or LBB) condition for positive constant > 0

b(v, )
inf sup (9.12) {?}
M vX kvkX kkM

Then for each functional f and g, there exist unique solution.

Remark 9.6 (Comments on inf-sup condition). The condition is equivalent to

b(v, )
sup kkM M (9.13) {?}
vX kvkX

35
10 Discrete (matrix-vector) formulations
Spaces:

X = Rn , M = Rm , for n > m. (10.1) {?}

Bilinear form and linear functionals

a(u, v) = Au v b(u, v) = Bu w (10.2) {?}


f (v) = f v f (v) = f v (10.3) {?}

holding for all vectors u, v X and w M , where matrices A Rnn and B Rmn and
vectors f X and g M .
Saddle point problem: find (u, ) X M such that

A BT
    
u f
= (10.4) {?}
B 0 g
or

a(u, v) + b(v, ) = f (v) v X (10.5a) {?}


b(u, ) = g() M. (10.5b) {?}

The following two conditions are crucial (along with continuity of the forms) for the existence
and uniqueness of the solution. Those conditions will be examined in general setting in next
section.
Remark 10.1 (Ellipticity condition). What does it mean ellipticity condition?

a(u, u) kukV (10.6) {?}


a(u, u) = Au u kukRn (10.7) {?}

It means that the symmetric part of the matrix A is positive definite, its eigenvalues are positive,
recall spectral decomposition.
Remark 10.2 (Inf-sup condition). What does it mean inf-sup condition?
b(u, )
sup kkV (10.8) {?}
uX kukX
b(u, ) Bu
sup = sup kk (10.9) {?}
uX kukX uX kukX

the condition is not satisfied if Bu = 0, which means that there is a vector orthogonal to
range B. Therefore B has to be a full rank matrix.
Remark 10.3. The spaces from (9.4) can be characterised as

Vg = {v X | b(v, ) = g() M } = {v X | Bv = g}, (10.10) {?}


V = {v X | b(v, ) = 0 M } = {v X | Bv = 0} = ker B, (10.11) {?}
T
V = {v X | (v, w)X = 0 w X} = range B (10.12) {?}

Proof of existence for this special case. First, we find a vector x0 Vg , which satisfy a con-
straint, e.g.

x0 = B T (BB T )1 g Vg , (10.13) eq:mat_x0


T T 1
Bx0 = BB (BB ) g = g; (10.14) {?}

36
the inverse exists because BB T is positive definite

BB T v v = B T v B T v > 0 (10.15) {?}

for all v M , since B has a full rank.


Now, we can decompose x = x0 + x, which leads to an adjusted linear system

A BT
    
x f Ax0
= , (10.16) {?}
B 0 0

where now x V . Then, if we multiply the first equation with v V and use B T v = Bv = 0,
we get a problem for x V only (without )

Ax v = (f Ax0 ) v v V. (10.17) {?}

It has a unique solution since A is positive definite on V


1
x = arg min Ay y (f Ax0 ) y. (10.18) {?}
yV 2

Then we need to find . From a first equation

B T = (f Ax0 Ax), (10.19) {?}

we can find solution as

= (BB T )1 B(f Ax0 Ax) (10.20) {?}

Proof of solution uniqueness. It is based on standard arguments. Assume that we have two
solutions (x1 , 1 ) and (x2 , 2 )

A BT A BT
         
x1 f x2 f
= = (10.21) {?}
B 0 1 g B 0 2 g

Then we show that both solutions have to coincide. After substraction of both equations we
have
A BT
    
x1 x2 0
= (10.22) {?}
B 0 1 2 0

From second equation, we have x1 x2 V . From first equation, after multiplication with
(x1 x2 ), we obtain

A(x1 x2 ) (x1 x2 ) = 0 (10.23) {?}

which implies x1 = x2 from positivity of A. Then from first equation, we have B T (1 2 ) = 0


which implies 1 = 2 as B is of full rank (thanks to inf-sup condition).

10.1 Linear systems


10.1.1 Direct solvers and Krylov subspace methods
BiCG Biconjugate gradient method

BiCGstab stabilized Biconjugate gradient method

37
10.1.2 Conjugate gradients on a subspace V
In previous section, we have seen that it is possible to find as solution as x = x0 + x where x0
is obtained from (10.13) and x can be found from a minimisation problem
1
x = arg min Ay y (f Ax0 ) y. (10.24) eq:min_over_V
yV 2

We can use the orthogonal projection on a space V . First, we can show that the following
operator/matrix P : X X

P = B T (BB T )1 B (10.25) {?}

is projection, indeed

P 2 = B T (BB T )1 BB T (BB T )1 B = B T (BB T )1 B = P.

It is also orthogonal projection because

P T = [B T (BB T )1 B]T = B T (BB T )1 B = P.

Then we show that P is projection on V = range B T = {B T X | M }. We show that


P v = v for all v V and P w = 0 for all w V

P B T = B T (BB T )1 BB T = B T
P w = B T (BB T )1 Bw = 0

Then Q = (I P ) is also an orthogonal projection, now on V . The minimisation over a subspace


V from (10.24) is equivalent to weak form

Ax v = (f Ax0 ) v v V.

From which we can deduce the linear system

Ax Qv = (f Ax0 ) Qv v V,
Ax Qv = (f Ax0 ) Qv v X,
QAx v = Q(f Ax0 ) v v X,
QAx = Q(f Ax0 ).

Now the problem states: find x V satisfying

QAx = Q(f Ax0 ).

This can be solved by conjugate gradients for initial approximation x(0) V . Indeed the Krylov
subspace for residuum r(0) V

Kk (r0 ) ={r(0) , QAr(0) , . . . , (QA)k1 r(0) } V

10.2 Eigenvalues of saddle-point linear system


Let we have a saddle-point linear system

A BT
 

B C

38
with symmetric and positive definite matrix A Rnn , symmetric and positive semidefinite
matrix C Rmm , and a full rank matrix B Rmn . Then the system is similar to
 
A 0
0 (C + BA1 B T )
as
A BT In A1 B T
     
In 0 A 0
=
B C BA1 Im 0 (C + BA1 B T ) 0 Im
1 T
  
A 0 In A B
= 1 T
B (C + BA B ) 0 Im
.
We note that the matrices V, W are similar if there exists invertible matrix P such that

V = P 1 W P.

Similar matrices U, V have the same eigenvalues and the eigenvector u, w are connect as v = P w
because

V v = v
P 1 V v = P 1 v
P 1 V P w = P 1 P w
W w = w.

Therefore, the matrix


 
A 0
(10.26) {?}
B (C + BA1 B T )

has n positive eigenvalues because A is positive definite and m negative eigenvalues because the
matrix (C + BA1 B T ) is negative definite, i.e. for all Rm

(C + BA1 B T ), Rm = C, Rm + BA1 B T , Rm
  
(10.27) {?}
= C, Rm + A1 B T , B T Rn > 0
 
(10.28) {?}
| {z } | {z }
0 >0

because of positive semidefiniteness of C and positive definiteness of A and thus also of A1 .

39
11 Mixed-formulations and its numerical approximation
11.1 Applications to PDEs: Poisson equation with general boundary condi-
tions
Set of differential equations on = (0, 1) (0, 1) R2
= f, = Ae, e = u (11.1) {?}
with Dirichlet boundary conditions (for simplicity considered homogeneous)
u(x) = uD (x) for x D , (11.2) {?}
n (x) = N (x) for x N (11.3) {?}
T S
where N D = and N D = .
Weak formulation is based on a bilinear form a : V V R and linear functional F : V V
for some Hilbert space V , stating find u V such that
a(u, v) = F (v) v V (11.4) {?}

Primal formulation:
In order to satisfy boundary conditions, we have to introduce two spaces
VD = {v H 1 () : u|D = uD }, (11.5) {?}
1
V = {v H () : u|D = 0}. (11.6) {?}
The boundary conditions have to be some that there exists at least one function satisfying this
boundary condition u0 Vd . Then we search a solution u = u0 + u, where u V and satisfy
Z Z Z
a(u, v) = Au v = F (v) = fv + N v a(u0 , v) (11.7) {?}
N

which is also equivalent to


1 1
u = arg min a(v, v) F (v) = a(u, u) F (u) (11.8) {?}
vV 2 2

Primal-mixed formulation:

Strong form:
A1 u = 0, (11.9) {?}
div = f (11.10) {?}
The primal-mixed formulation is obtained by a multiplication with test function and by a
modification of a second equation with the help of Greens theorem.
Spaces and forms
X = L2 (; Rd ), M = {v H 1 () : u|D = 0}. (11.11) {?}
a(, ) = (A1 , )L2 b( , u) = ( , u)L2 (11.12) {?}
Similarly to primal problem, we have to decompose u = u0 + u with
u0 MD = {v H 1 () : u|D = uD }.
The weak formulation than states: Find in (, u) X M
(A1 , )L2 ( , u)L2 = ( , u0 )L2 X (11.13) {?}
(, v)L2 = (f, v)L2 (N , v)L2 (N ) v M (11.14) {?}

40
Dual-mixed formulation:
Strong form:

A1 u = 0, (11.15) {?}
div = f (11.16) {?}

The dual-mixed formulation is obtained by a multiplication with test function and by a


modification of the first equation with the help of Greens theorem.
Spaces:

X = Hdiv () = {v L2 (; Rd ) | v L2 () and n v|N = 0}, M = L2 () (11.17) {?}


XN = Hdiv () = {v L2 (; Rd ) | v L2 () and n v|N = N }, (11.18) {?}

Because of boundary conditions, we decompose the flux = 0 + , where X and 0 XN .


Find in (, u) X M

(A1 , )L2 + ( , u)L2 = n , uD L2 () (A1 0 , )L2 X




(11.19) {?}
( , v)L2 = (f, v)L2 ( 0 , v)L2 v M (11.20) {?}

11.1.1 Connection of mixed-formulations by Green formula


For sufficiently smooth functions, we can use Green formula
Z Z Z Z Z Z
(u) = ( )u + u = n u = n u + n u (11.21) {?}
N D

The essential boundary conditions have to be incorporated into the finite dimensional space.
Dirichlet boundary conditions are essential in primal-mixed formulation, while Neumann bound-
ary conditions are essential in dual-mixed formulation.

Remark 11.1. In the Green formulation, it is necessary to be mathematically precise with


boundary terms. For a Sobolev function u H 1 () we know that its value on boundary (given
by trace operator) is a L2 -function, i.e. u| = uD L2 (). However, the trace operator is not
onto L2 (), which means that if there is function uD L2 () there do not necessarily exists
u H 1 () such that u| = u0 . The functions uD has to be from better space H 1/2 () which is
something between L2 and H 1 . Similarly the div operator has values on boundary in H 1/2 ()
which is a space of linear functionals to H 1/2 (). Therefore, we can find in a literature that
Z


n u = n , u H 1/2 ()H 1/2 () , (11.22) {?}
N

which is a correct expression of a boundary term.

In dual-mixed formulation, u is from better space:

In dual-mixed formulation, it seems that u L2 () only. However, from the first equation
tested with C (; Rd ) such that n = 0, we have

(A1 , )L2 = ( , u)L2 . (11.23) {?}

It means A1 is a weak gradient of u, i.e. u = A1 in the weak sense. Since A1


L2 (; Rd ) then also u L2 (; Rd ) and consequently u H 1 ().

41
11.1.2 Ellipticity and inf-sup condition in continuous setting
Primal-mixed formulation: We can derive the inf-sup inequality by choosing appropriate
element from the space X, i.e.

b( , ) ( , ) (, )
sup = sup = kkL2 CkkH 1 (11.24) {?}
X k kX X k kL2 kkL2

where the last inequality comes from Friedrichs inequality.

Dual-mixed formulation: We have to consider that the norm of Hdiv -space is

kukH 1 = kuk2L2 + k uk2L2 . (11.25) {?}


div

We show that ellipticity holds on a subspace V composed of functions satisfying constraint


= 0. Then we simply have

a( , ) = k k2L2 = k k2L2 + k k2L2 = k kH 1 , (11.26) {?}


div

which gives the ellipticity condition to be 1.


For inf-sup condition, it is necessary to derive alternative to Friedrichs inequality

kvkL2 Ck div vkL2 (11.27) {?}

Then, it is possible to derive

b( , v) (div , v)L2 (w, v)L2


sup sup sup (11.28) {?}
X k kHdiv X (1 + C)k div kL2 wL2 (1 + C)kwkL2
(v, v)L2 1
= kvkL2 . (11.29) {?}
(1 + C)kvkL2 1+C

11.2 Galerkin approximation to saddle-point problems


11.2.1 Abstract setting
hprob:saddlei
Problem 11.2 (Galerkin approximation to saddle-point problem). Let Xh and Mh be finite
element spaces. Then a Galerkin approximation to a saddle point problem state: Find (uh , h )
Xh Mh such that

a(uh , v) + b(v, h ) = f (v) v Xh


b(uh , w) = g(w) w Mh

The discrete finite element spaces Xh and Mh have to satisfy inf-sup condition. However,
sometimes it is difficult to prove.

11.3 Poisson equation


11.3.1 Primal-mixed formulation:

Xh = {v L2 (; Rd ) : vi |T Pk (T ) for all i = 1, . . . , d and T Th }, X = L2 (; Rd )


Mh = {v H01 () : v|T Pk (T )} M = H01 (; Rd ),

where Th is regular triangulation with triangles denoted by T . The inf-sup condition is satisfied
as in the continuous setting because Mh Xh .

42
11.3.2 Dual-mixed formulation

Xh = RTk X = Hdiv ()
= {v L2 (; Rd ) : v|T = P + px for all p, Pi Pk (T ), v n is continous on T \ and T Th }
Mh = {v L2 () : v|T Pk (T ) and T Th } M = L2 ()

The space RTk is composed of Raviart-Thomas elements.


In two-dimensional setting, Xh = RTk can be written as
 
Xh = RTk = {v L2 (; Rd ) : v = PP12 + p ( xx12 ) for all p, Pi Pk (T ) and all i = 1, . . . , d}

Showing that RTk space is really a subspace of Hdiv : We derive for P RTk
Z X Z X Z X Z Z
P = P = P + P n = P
T Th T T Th T T Th T
| {z }
=0 because of continuity

where P is defined element-wise.

Basis of Raviart-Thomas elements in two-dimensional setting: Let Pi for i = 1, . . . , 3


be vertices of triangle. Then the basis functions are expressed as
1
i (x) = (x Pi ) (11.30) {?}
2|T |
where = 1 because the orientation of edges has to be considered. These basis functions
(similarly to Lagrange basis functions) satisfy Dirac delta property
(
1 for i, j, k {1, 2, 3} and i 6= j 6= k,
Z
njk i (ejk ) = (11.31) {?}
Ejk 0 otherwise

where Ejk = {tPj + (t 1)Pk | t [0, 1]} is edge opposite to Pi , point ejk = 12 (Pj + Pk ) is the
middle point of Ejk , and njk is the outer normal of triangle T that is orthogonal to (Pj Pk ),
i.e. (Pj Pk ) njk = 0.

12 Duality with Fenchel conjugate


For detailed information, see books [Fortin and Brezzi, 1991] or more advanced [Ekeland and Temam, 1976].

12.1 Min-max problems Poisson equation


Here, for simplicity, only homogeneous boundary conditions are considered.

Primal-mixed formulation: We start with a primal formulation


1
inf (Av, v) (f, v)
vH01 () 2

and use a following trick (Fenchel-conjugate function)


1 1
(Av, v) = sup ( , v) (A1 , ) (12.1) {?}
2 L2 (;Rd ) 2

43
which has a stationary point given by condition

( , v) (A1 , ) = 0q (12.2) {?}

and thus = Av.


Therefore, after substitution into primal formulation, we get
1
inf sup (A1 , ) (f, v) + ( , v) (12.3) {?}
vH01 () L2 (;Rd ) 2

which has a primal-mixed formulation as stationary condition.

Dual-mixed formulation:
1 1
inf (A , ) for Zf = {v L2 (; Rd ) | v f = 0}
Zf 2
1 1
inf sup (A , ) ( f )
L L2 () |2
2
{z }
=L(,)

We can investigate a variation of L at a stationary point (, ), i.e.


1
L( + , + ) = L(, u) + (A1 , ) (, ) + (, f ) + (A1 , ), (12.4) {?}
2
which leads to saddle point problem for (, )

(A1 , ) (, ) = 0 (12.5) {?}


( , ) = (f, ) (12.6) {?}

From the first equation, after Green theorem, we can deduce that

= u.

12.2 Perturbation duality


Primal problem
1
inf J(u, u) = inf (u, 0) J(u, u) = (Au, u) (f, u) (12.7) {?}
uH01 uH01 2

where : H01 L2 is a perturbed problem, which we can define as

1
(u, p) = J(u, u + p) = (A(u + p), u + p) (f, u). (12.8) {?}
2
Dual problem

sup (0, p ) (12.9) {?}


p

where : H 1 L2 is a Fenchel conjugate function

(u , p ) = sup sup[ u , u + p , p (u, p)].





(12.10) {?}
u p

44
Using substitution (u + p) = q, we derive

(0, p ) = sup sup[ p , p (u, p)]




(12.11) {?}
u p
1
= sup sup[ p , p + (f, u) (A(u + p), u + p)]


(12.12) {?}
u p 2


1
= sup sup[ p , u p , q + (f, u) (Aq, q)] (12.13) {?}
u q 2
1
= sup[ p , u + (f, u)] + sup[ p , q (Aq, q)].

(12.14) {?}
u q 2

The second supremum has a stationary point for Aq = p , which result in


1 1
sup[ p , q (Aq, q)] = (A1 p , p ).

q 2 2

Therefore
(

for p + f = 0
(0, p ) = 1 1 (12.15) {?}
2 (A p , p ) otherwise

Duality: From

(0, p ) = sup[ p , p (u, p)]




(12.16) {?}
u,p

we have for every u and p = 0

(0, p ) p , 0 (u, 0)


(12.17) {?}

(0, p ) (u, 0) (12.18) {?}

sup (0, p ) inf (u, 0) (12.19) {?}
p u

because left- and right-hand side holds for any p and u. In many cases, it is possible to show
even equality. Indeed, we can continue with the duality to obtain

inf

(u , 0) sup (0, p ) inf (u, 0). (12.20) {?}
u p u

In our case of Poisson equation and in many others, we have (u , 0) = (u, 0), which gives
us equality in the primal-dual relation

sup (0, p ) = inf (u, 0) (12.21) {?}


p X uH01
1 1
sup (A1 p , p ) = inf (Au, u) (f, u) (12.22) {?}
p X 2 uH01 2

where X = {p | p = f }. If this is approximated with some conforming method such that


Xh X and Yh H01 , we get

sup (0, ph ) sup (0, p ) = inf (u, 0) inf (uh , 0) = (uh , 0). (12.23) {?}
ph Xh p X uH01 uh Yh

This is a basis for a guaranteed error estimates.

45
Part IV
Appendix
This part is here only for interested students; it will not be taught and thus examined.

A Introduction to Lebesgue integration


hsec:Leb_integi In
this section, we build a Lebesgue integration, which is a fundamental part for definition of
Sobolev spaces, and later on analysis of PDEs.

A.1 Lebesgue measures


In order to develop the Lebesgue integral, we have to measure the sizes of the sets. d-dimensional
Lebesgue measure corresponds to measuring lengths, areas, or volumes for one-, two-, or three-
dimensional objects, respectively.

Definition A.1 (Open and closed interval). Let a, b R such that a < b then we define an open
and a closed interval as

(a, b) := {x R : a < x < b} [a, b] := {x R : a x b} (A.1) {?}

Definition A.2 (Outer and inner Lebesgue measure on R). Let A R be an arbitrary set, then

X
[
A := inf{ (bi ai ) : (ai , bi ) A} (A.2) {?}
i=1 i=1

is called outer Lebesgue measure of a set A. Let I be bounded interval such that A I, then

A := I (I A) (A.3) {?}

is called an inner measure of A.

Remark A.3. The outer Lebesgue measure starts from measuring the sizes of intervals (both
open and close gives the same value). Then outer measure is defined as in infimum over the
sizes of open intervals that cover the set A. The inner measure can be defined as a supremum
of closed disjoint intervals that are contained in A, i.e.

X
[ \
A := sup{ (bi ai ) : [ai , bi ] A and [ai , bi ] [aj , bj ] = for i 6= j}. (A.4) {?}
i=1 i=1

For some sets the outer and inner measure coincide, we call this measurable sets, see next
definition. However, there are sets where outer and inner measure differs. This is the reason
why proper theory has to be developed.

Definition A.4 (Lebesgue measurable sets). We say that A R is Lebesgue measurable if

A = A. (A.5) {?}

Definition A.5 (Lebesgue measure on R). We say that := from previous definition is a
Lebesgue measure on measurable sets.

46
Definition A.6 (Lebesgue measure on Rd with dimension d from N). Let I = (a1 , b1 )(a2 , b2 )
(ad , bd ) be a parallelepiped (square in two-dimensional setting d = 2) and vol I = (a1 , b1 )
(a2 , b2 ) . . . (ad , bd ) its volume. Then the d-dimensional Lebesgue measure is defined as
X [
A := A = { vol Ik : Ik A, Ik is a parallelepiped for each k N}. (A.6) {?}
kN kN

In order to distingish the dimensionality of the measure, we will denote it in subscript as d .


Examples A.7.
(0, 2) = 2, [0, 2] = 2, (0, 2] = 2, (0, 1023 ) = 1023
{1} = 0, {1, 2, 3} = 0, {1} = 0, , N = 0, Z = 0,Q = 0,R = . For example, how
to show that N = 0? We show that the outer measure equals to zero. Hence

X
[
= inf{ (bi ai ) : (ai , bi ) N}
i=1 i=1

X [
= inf{ i
: (i i+1 , i + i+1 ) N for some > 0}
2 2 2
i=1 i=1

where each point i N is covered by open interval (i 2i+1 , i+ 2i+1 ) with size 2i
. Therefore

[
= inf{ : (i ,i + ) N for some > 0} = 0.
2i+1 2i+1
i=1

2 {(1, 2) (2, 4)} = 2, 2 {(1, 2) [2, 4]} = 2, 2 {(1, 2)} = 0, 3 {(1, 2) [2, 4]} = 0,
1 {(1, 2) [2, 4]} =
There exists non-measurable sets, see e.g. Hausdorff paradox and BanachTarski paradox.
Remark A.8 (Finite, countable, and uncountable sets). There are different notions how to
measure the size of the set. One option is the length of intervals, area of surfaces, or volume
of three-dimensional objects. Another notion is cardinality, which measure somehow a number
of elements of the set. We say that two sets have the same cardinality if there exists bijection
between them (one-to-one map which is onto). For examples set A = {2, 4, 6} is finite, it has
three elements so its cardinality is #A = 3. For infinite sets, we say that countable sets have the
same cardinality as natural numbers N, e.g. Z, Q, N2 ; other infinite sets are uncountable sets,
e.g. R, R2 , C.
Example A.9 (Countable sets). Show that #N = #Z. We will find an isomorphism between
them, i.e. map f : N Z defined as
(
i
for i|2 = 0
f (i) = 2 i1 (A.7) {?}
2 otherwise
where i|2 denotes the remainder after division by 2 (sometimes called modulo operation, e.g.
0|2 = 2|2 = 4|2 = 0, 1|2 = 3|2 = 1).
Example A.10 (Uncountable sets). Show that #(0, 1) = #(0, 1) (0, 1). There is again an
isomorphism defined as follows. Let a (0, 1), then each such number can be expressed as
decimal number 0, a1 a2 a3 . . . , i.e. each number is defined with a sequence {ai }iN . Then
f (a) = (0.a1 a3 a5 . . . , 0.a2 a4 a6 . . . ) (A.8) {?}
and again each tuple (0.a1 a3 a5 . . . , 0.a2 a4 a6 . . . ) represents one point in square (0, 1) (0, 1).
Remark A.11 (Cantor set). Cantor set is an uncountable set of Lebesgue measure zero.

47
A.2 Abstract measures
In previous section, we have introduced Lebesgue measure, which measures the sizes of sets
(their lengths, areas, or volumes). There are some other types of measuring sets, which do not
exactly corresponds to our understanding of sizes, areas, or volumes, but those measuring has
still a good sense. Those abstract theory is very useful also in engineering, e.g. in probability
theory.
In this section and later on, we will talk about measure space, which is a triple (, , ),
where is our basic set (e.g. computational domain, the set over which we usually integrate),
is a -algebra (the set of subsets of , which we can measure), and [0, ] is a measure,
i.e. a map that takes a set A (which also means A ) and assigns to this set its measure
(A) [0, ], which can also attain infinity; measurable space is a pair (, ) of set and some
-algebra.

Definition A.12 (-algebra). Let be a set and 2 be a power set (set of all possible sets).
Then 2 (we take only some of all possible subsets) is a -algebra if

(i) is in

(ii) is closed under completion, i.e. if A then also \ A


S
(iii) is closed under countable unions, i.e. if Ai for i N then iN Ai

Remark A.13. It can be proved that -algebra is closed under countable intersections, it also
contains empty set, etc.

Definition A.14 (Borel -algebra). Borel -algebra is a -algebra generated by all open sets,
i.e. smallest -algebra that contains all open sets.

Examples A.15 (Examples of -algebras).

= {1, 2, 3}, = {, , {1}, {2}, {3}, {2, 3}, {1, 3}, {1, 2}} is the biggest -algebra, and
equals to the space of all subsets = 2

= {1, 2, 3}, = {, } is the smallest -algebra

= {1, 2, 3}, = {, , {1}, {2, 3}} is another -algebra

= (0, 1), = {, } is the smallest -algebra

= (0, 1), = {, , (0, 21 ), [ 12 , 1)} is another -algebra

= (0, 1), the set of all possible subsets 2 is not a -algebra because it contains unmea-
surable sets

Definition A.16 (Measure). Let be a collection of subsets of . Then a function :


[0, ] is a measure if

(i) is a -algebra.

(ii) = 0

(iii) -aditivity: for each sequence {Ai }iN of pairwise disjoint sets from
X
(iN Ai ) = Ai (A.9) {?}
iN

48
Remark A.17. The measure allows the values. When the measure of the whole set is finite
< , we talk about finite measure. Generalizations of here introduced measure include e.g.
sign measures (allowing for negative values of measure) or complex measures (with values in
complex plain).

Examples A.18 (Measures).

Dirac measure x A = 1 if x A and x A = 0 otherwise

discrete measure: measures on finite sets, i.e. let = {1, 2, 3, 4, 5}, then we define A =
#A (the number of elements=cardinality) for each A , or A = 2 ;

probability measure ifR = 1; for example let f = 1/2 on (0, 2) and f = 0 otherwise, then
, defined as A := A f d with Lebesgue measure , is a probability measure

A.3 Measurable and continuous functions


Measurable functions are all those that can be integrated. We cannot measure all sets because
there are unmeasurable sets, sets that do not fulfil natural requirements such as aditivity. As
it will be seen in next section about integration, we also have to confine functions that can be
integrated, they are called measurable functions.

Definition A.19 (Measurable function). Let f : R be a function on a measurable space


(, ). Then f is measurable if preimage of an open interval I

f 1 (I) := {x : f (x) I} (A.10) {?}

is contained in -algebra, i.e. f 1 (I) , for each open interval I.

Definition A.20 (Equality almost everywhere). Let f, g : R be two functions on measurable


space (, , ). Then we say that f = g -almost everywhere (or briefly: almost everywhere) if
there exist D such that ( \ D) = 0 and

f (x) = g(x) for x D. (A.11) {?}

We then also write f = g a.e. (or for almost all x or shortly for a.a. x ). In the similar
setting, we can talk about other relations such as f < g a.e.

Remark A.21 (Operations with measurable functions). One example of measurable function is
characteristic function of measurable set. But this set is also closed under standard operations,
also limit operations. It means that if f, g, fi are measurable for i N and a R then also
following functions are measurable

af , f + g, max{f, g}, min{f, g}, |f |, f g, f /g

sup fn , inf fn , lim inf fn , lim sup fn , lim fn

h g for continuous function h ( is a composition of two functions)

A.4 Lebesgue integration


In this section, we develop Lebesgue integration, which starts from non-negative simple func-
tions.

49
heq:simple_funi
Definition A.22 (Nonnegative simple function). Let s : R be a measurable function on a
measurable space (, ). Then we say that it is a non-negative simple function if it has finite
range (finite number of elements) on measurable sets, i.e. it can be expressed as
n
X
s= i Ai (A.12) {?}
i=1

where Ai are pairwise disjoint sets from , and i R are positive i 0 for all i, and A is a
characteristic function of set A, i.e.
(
1 for x A
A = (A.13) {?}
0 otherwise
Examples A.23 (Simple functions).
function f : R R, defined as f (x) = 1 for all x R, is a simple function because its
range is {1}, which is a finite set
function f : R R, defined as f (x) = x for all x R, is not a simple function because its
range is R, which is not a finite set
function f : R R, defined as f (x) = x for all x (0, 1] and f (x) = 0 otherwise, is not a
simple function because its range is [0, 1], which is not a finite set (is uncountable)
function f : R R, defined as f (x) = n1 for all x (n, n + 1], n N and f (x) = 0
otherwise, is not a simple function because its range is {1/n}nN = {1, 12 , 13 , . . . }, which is
not a finite set (is countable)
function f : R R, defined as f (x) = n1 for all x (n, n + 1], n N, such that
n 1000 and f (x) = 0 otherwise, is a simple function because its range is {1/n}1000
n=1 =
{1, 12 , 13 , . . . , 1/1000}, which is a finite set
some other function such as sin(x), exp(x), x2 , x3 are not simple function because their
range is not finite
Definition A.24 (Lebesgue integration of nonnegative simple functions). Let s : R be a
nonnegative simple function like in Def. A.22 and D be a measurable set. Then we define
integral as
Z Xn
s d = i (D Ai ). (A.14) {?}
D i=1

For values like 0 that can possibly occur in the sum, we consider 0 = 0
Definition A.25 (Lebesgue integration of positive functions). Let f : R be a nonnegative
measurable function on measurable space (, , ), i.e. f 0 a.e.. Then we define integral on
D as
Z Z
f d = sup{ s d : s is a nonnegative simple function such that 0 s f }. (A.15) {?}
D D
Definition A.26 (Lebesgue integration). Let f : R be a measurable function on measurable
space (, , ), i.e. f 0 a.e., and f + = max{f, 0} and f = max{f, 0} be its decomposition
into positive and negative parts. Then we define integral on D as
Z Z Z
f d = +
f d f d (A.16) {?}
D D D
when at least one of the integrals is finite.

50
Remark A.27 (Integrals with discrete measure).
assume that (, , ) is measurable space with = {1, 2, 3, 4, 5}, = 2 and discrete measure
A = #A (the number of elements=cardinality) for each A ; then functions u, v : R
can be represented as standard Euclidean vectors u, v R5 in dimension 5 with components
ui = u(i) and v i = v(i). These maps u, v are measurable and are even simple functions
5
X
u= u(i){i}
i=1

We can also consider integrals of them, for example


Z 5
X
u(i)v(i) d(i) = u(i)v(i) = u v,
i=1

which corresponds to scalar product.

A.4.1 Alternative integrals to the Lebesgue one


Remark A.28 (Newton integration). Let f : (a, b) R be a function and F its primitive
function (antiderivative), i.e. F 0 (x) = f (x) for all x (a, b). The Newton integral is a value
Z b
N f = lim F (x) lim F (x)
a xb xa+

Remark A.29 (Riemann integration). Let f : (a, b) R be a bounded function. If a = 0 <


1 < . . . < n = b be a sequence of points in the interval (a, b), we call it partition. The the
following two values are upper Riemann sum and lower Riemann sum
n
X
R f = inf{ sup f (x)(i i1 ) : for all partions of (a, b)},
i=1 x(i1 ,i )
Xn
R f = sup{ inf f (x)(i i1 ) : for all partions of (a, b)}.
x(i1 ,i )
i=1

A function is then Riemann integrable if R f = R f .

Remark A.30 (Difference between integrals). If the Newton, Riemann, and Lebesgue integrals
exists then they equal each other. Usually, the function is not Newton integrable if there is no
primitive function (antiderivative); then still it can have Riemann and Lebesgue integral. It can
also happen that a function is Newton or Riemann integrable but not Lebesgue integrable; this
happens if
Z Z Z
f d = +
f d f d = ,
D D D

i.e. when the Lebesgue integral is not absolutely integrable, i.e f (x) = sinx
x on (0, ) for
(0, 1).
There are also functions, which are Lebesgue integrable but not Riemann integrable, e.g.
Dirichlet function (the indicator function of all rational numbers, Q ). The advantage of
Lebesgue integrals lies in its generality (different measures can be considered), which is espe-
cially advantageous in higher dimensions. It also allows for easy limit theorems, discussed in
next section, which are necessary in analysis of PDEs.

51
A.4.2 Properties of Lebesgue integration
Lemma A.31 (Basic properties). Let a, b R and f, g : R be measurable.
R R R
(i) af + bg d = a f + b g d
R R
(ii) f = g almost everywhere then f = g
R R
(iii) f g almost everywhere then f g
R R
(iv) if f < then also |f | <
R R
(v) | f | |f |
R
Lemma A.32. Let f : R be a positive measurable function and f d = 0. Then f = 0
almost everywhere.

Proof by contradiction. Assume Rthat f 6= 0, then f C > 0 on some A with positive


measure. Therefore 0 < CA f , which is a contradiction.

Direct proof. Let An =R {x : n1 f (x)}.


S It means that 0 An nf , which we can
integrate 0 An n An f = 0. Then also n=1 An = {x : f (x) > 0} has a measure zero
thanks to -aditivity of measures.
R
Lemma A.33. Let f : R be a measurable function and E f d = 0 for any measurable set
E . Then f = 0 almost everywhere on .

Proof. Set E := {x : f (x) 0}, then E f d = E f + d = 0 and using the previous theorem
R R

we get f + = 0 a.e. The same can be proven for f .

The following theorems describe the convergence theorems, which are of particular in PDEs.

Lemma A.34 (Monotone convergence theorem). Let {fi }i N be a sequence of measurable


functions with 0 f1 f2 . . . , then
Z Z
lim fk = lim fk .
k k

Lemma A.35 (Dominated convergence theorem). Let R{fi }i N be a sequence ofR measurable
functions such that limk fk = f a.e. and fk g with |g| d < . Then also |f | d <
and
Z Z
lim fk d = f d
k

References
aess2007bookFEM [Braess, 2007] Braess, D. (2007). Finite elements: Theory, fast solvers, and applications in solid
mechanics. Cambridge University Press.

arlet1991HNAFEM [Ciarlet, 1991] Ciarlet, P. G. (1991). Handbook of Numerical Analysis: Finite Element Methods.
Elsevier Science Publishers B.V. (North-Holland).

eland1976convex [Ekeland and Temam, 1976] Ekeland, I. and Temam, R. (1976). Convex Analysis and Varia-
tional Problems. North-Holland, Amsterdam.

uermond2013book [Ern and Guermond, 2004] Ern, A. and Guermond, J.-L. (2004). Theory and practice of finite
elements. Springer Science & Business Media, New York.

52
Evans2000PDE [Evans, 2000] Evans, L. (2000). Partial Differential Equations, volume 19. American Mathe-
matical Society.

Fortin1991 [Fortin and Brezzi, 1991] Fortin, M. and Brezzi, F. (1991). Mixed and hybrid finite element
methods. Springer Science & Business Media.

Gockenbach2006 [Gockenbach, 2006] Gockenbach, M. (2006). Understanding and implementing the finite element
method.

Fenics2017 [Langtangen and Logg, 2017] Langtangen, H. P. and Logg, A. (2017). Solving PDEs in Python.
Springer International Publishing.

Rudin1986real [Rudin, 1986] Rudin, W. (1986). Real and complex analysis. McGraw-Hill, New York, third
edition.

Wandzurat2003 [Wandzurat and Xiao, 2003] Wandzurat, S. and Xiao, H. (2003). Symmetric quadrature rules
on a triangle. Comput. & Math. with Appl., 45(12):18291840.

53

Anda mungkin juga menyukai