Beruflich Dokumente
Kultur Dokumente
University of Ostrava
Faculty of Electrical Engineering and Computer Science
Department of Applied Mathematics
Author:
VSBTechnical
University of Ostrava
Referees:
VSBTechnical
University of Ostrava
Faculty of Electrical Engineering and Computer Science
Department of Applied Mathematics
September 2003
arka
Dedicated to S
iv
Abstract
This thesis treats with theoretical and computational aspects of threedimensional optimal shape
design problems that are governed by linear magnetostatics. The aim is to present a complete
process of mathematical modelling in a wellbalanced way. We stepbystep visit the world
of physics, functional analysis, computational mathematics, and we end up with reallife applications. Nevertheless, the main emphasis is put on an efficient implementation of numerical methods
for shape optimization which exploits an effective evaluation of gradients by the adjoint method
and a just recently introduced multilevel optimization approach. We also emphasize numerical
experiments with reallife problems of complex threedimensional geometries.
We begin from a description of the electromagnetic phenomena by Maxwells equations and
we derive their threedimensional (3d) and twodimensional (2d) magnetostatic cases with the
linear constitutive relation between the magnetic flux density and the magnetic strength density.
Then we start to develop a general theory that covers both 2d and 3d optimal interfaceshape
design problems that are constrained by a secondorder linear elliptic boundary vectorvalue problem (BVP). First we pose a weak formulation of the BVP with the homogeneous Dirichlet boundary condition. Whenever the kernel of the BVP operator is not trivial, we employ a regularization
technique such that the regularized solutions converge to the true one. The continuous weak formulation of the abstract BVP is discretized by the firstorder finite element method on triangles
and tetrahedra, respectively. We set an abstract continuous shape optimization problem, the state
problem of which involves one or more BVPs such that they only differ in the righthand sides,
i.e., different current excitations in case of magnetostatics. The design boundary is an interface
between two materials, rather than a part of the computational domain boundary, as it is usual in
optimal shape design for mechanics. We prove the existence of an optimal shape by checking the
continuity of the cost functional and the compactness of the set of admissible shapes. Then we discretize the continuous optimization problem by the finite element method and prove the existence
of the approximate solutions. The main theoretical result of this thesis is a proof of the convergence of the approximate optimized solutions to an optimal solution of the continuous problem,
where we also involve an inner approximation of the original computational domain with a Lipschitz boundary by a polyhedral (in the 3d case) or polygonal (in the 2d case) domain. Throughout
the abstract theory we introduce many assumptions that are checked for concrete applications
afterwards. These assumptions show the scope of the theory.
Concerning the computational aspects in optimization, we use the sequential quadratic programming method with a successive approximation of the Hessian. To justify the use, we verify
the smoothness of both the discretized cost and constraint functionals. Then we focus on the
calculation of gradients by means of the adjoint method and we derive an efficient algorithm for
that, including its Matlab implementation enclosed on the CD. We introduce a new multilevel
optimization approach as a possible adaptive optimization method.
Finally, we end up with physics again. We present two reallife applications with rather comv
vi
ABSTRACT
plex 3d geometries. After some motivation, we describe the optimization problem in terms of
verifying the theoretical assumptions, and we give numerical results. We present the speedup
of the adjoint method comparing to the numerical differentiation, and of the multilevel approach
comparing to the classical optimization. One optimized design was manufactured, we are provided
with measurements and, at the end, we discuss real improvements of the cost functional.
Acknowledgement
First of all I want to express my deep gratitude to my supervisor Prof. RNDr. Zdenek Dostal, CSc.,
the head of the Department of Applied Mathematics at the Technical University of Ostrava (TUO),
who introduced me to the magic of numerical mathematics and its applications in electromagnetism. I thank him for his kindness and patience during more than six years of my undergraduate
and doctoral studies for when he has been supervising me.
Special thanks are devoted to my very good friend, teacher, and colleague Doc. RNDr. Jir
Bouchala, Ph.D., who is an associated professor at the Department of Applied Mathematics, TUO.
Jirka convinced me that the mathematics is beautiful itself, even without applications. I thank to all
the other friends, colleagues, and teachers of mine from the Department of Applied Mathematics,
TUO, especially to Mgr. Vt Vondrak, Ph.D. and Mgr. Bohumil Krajc, Ph.D.
I am much obliged to o.Univ.Prof. Dipl.-Ing. Dr. Ulrich Langer, who is the head of the Institute
for Numerical Mathematics and the cospeaker of the Special Research Initiative SFB F013 Numerical and Symbolic Scientific Computing both associated to the Johannes Kepler University
Linz in Austria. Professor Langer offered me a Ph.D. position within the research project SFB
F013, where I spent one year and where I have been currently working again. Here I have made
a big progress in my research work and I have learned about efficient methods for solving large
discretized direct simulation problems based on the finite element and/or boundary element discretizations. I am also much indebted to my current or former colleagues, especially, I would like
to mention Dr. Dipl.-Ing. Joachim Schoberl, Dr. Dipl.-Ing. Michael Kuhn, MSc., and Dr. Dipl.Ing. Wolfram Muhlhuber.
I further thank to Prof. Ing. Jaromr Pistora, CSc., who is with the Institute of Physics, TUO,
and is the head of the Department of Education at the Faculty of Electrical Engineering. Professor
Pistora asked me to cooperate on development of a new generation of electromagnets used in the
research on magnetooptic effects. I also thank for the fruitful cooperation to the other colleagues
from the Institute of Physics, namely to Dr. Mgr. Kamil Postava, Dr. RNDr. Dalibor Ciprian,
Dr. Ing. Michal Lesna k, Ing. Martin Foldyna, and Ing. Igor Kopriva.
I very appreciate the possibility to consult my work with Prof. RNDr. Jaroslav Haslinger, DrSc.,
who is a world leading expert in optimal shape design. I thank very much to Prof. RNDr. Michal
Krz ek, DrSc. from the Mathematical Institute of the Czech Academy of Sciences that I could learn
a lot during several consultations in his office. I very much admire the enthusiasm of Prof. Krz ek
that he has devoted to math.
arka,
At the end let me express my deep grateness and love to my parents and to my girlfriend S
with whom my life is complete, no matter what the career is about.
This work has been supported by the Austrian Science Fund FWF within the SFB Numerical and
Symbolic Computing under the grant SFB F013, by the Czech Ministry of Education under the
research project CEZ: J17/98:272400019, and by the Grant Agency of the Czech Republic under
the grant 105/99/1698.
vii
viii
ACKNOWLEDGEMENT
Notation
N
R
i
C
C1 , . . . , C16
nonnegative integers
real numbers
imaginary unit
complex plane
fixed constants
Abbreviations
1d
2d
3d
PDE
BVP
FEM
BFGS
SQP
AD
onedimensional
twodimensional
threedimensional
partial differential equation
boundary (vector)value problem
finite element method
update formula for the Hessian matrix named after Broyden,
Fletcher, Goldfarb, and Shanno
sequential quadratic programming
automatic differentiation
Chapter 2
B
H
J
u
2d
J
u
p. 10
p. 10
p. 10
p. 10
p. 10
p. 10
p. 11
p. 11
p. 11
p. 14
p. 14
p. 15
p. 15
p. 15
p. 15
Chapter 3
k kU
V /U
Ker(L)
U0
h, i
(, )
ix
x
U
H = U U
Rn
AT
det(A)
e
A
A1
m
n
C()
C k ()
C ()
supp v
C0 ()
C 0,1 ()
L
div
grad
curl
nu
B
B
Lp ()
L ()
meas()
a.e.
D u
H k ()
(, )k,
k kk,
| |k,
n
H k ()
(, )n,k,
H01 ()
NOTATION
orthogonal complement to the space U
orthogonal decomposition of the Hilbert space H
Euclidean space consisting of ndimensional real vectors
transposed matrix
determinant of the matrix A
adjoint matrix
inverse matrix
dimension of the computational domain , m {2, 3}
domain, i.e., open, bounded, and connected subset of R m
closure of the domain
boundary of the domain
unit outer normal vector to the boundary
space of functions continuous over
space of functions which are continuous up to their kth partial
derivatives over , k N
space of infinitely differentiable functions over
support of the function v
space of infinitely differentiable functions with a compact support
in
space of Lipschitz continuous functions over
set of all the domains with Lipschitz continuous boundaries
divergence operator
gradient operator
curl operator
cross product, tangential component of the function u along the
boundary
linear vector firstorder differential operator, B : C 1 () 1 7
C() 2 , 1 , 2 N
adjoint
operator
related
1 to B by Greens theorem, B :
2
1
C ()
7 C()
trace operator related to B by Greens theorem, : C() 1 7
[C()]2
Lebesgue space of measurable functions defined over for which
the Lebesgue integral of their pth power is finite, p [1, )
Lebesgue space of measurable essentially bounded functions over
p. 16
p. 17
p. 17
p. 18
p. 18
p. 18
p. 18
p. 19
p. 19
p. 19
p. 19
p. 19
p. 20
p. 20
p. 21
p. 21
p. 21
p. 21
p. 21
p. 22
p. 22
p. 22
p. 23
p. 23
p. 23
p. 23
p. 24
p. 24
p. 24
p. 24
p. 25
p. 25
p. 25
p. 25
p. 25
p. 25
p. 25
p. 26
xi
H 1/2 ()
H 1/2 ()
H(B; )
(, )B,
k kB,
| |B,
H0 (B; )
Ker(B; )
H0, (B; )
(S)
D
f
a(, )
f ()
(W )
u
a (, )
(W )
u
p. 26
p. 26
p. 30
p. 30
p. 31
p. 31
p. 31
p. 31
p. 31
p. 32
p. 32
p. 32
p. 33
p. 33
p. 33
p. 33
p. 34
p. 34
p. 34
p. 34
Chapter 4
h
Vh
n
Dh
ah (, )
fh
f h ()
(Wh )
uh
An
fn
u n
h
n h
K i , K ei
Th
xh
n xh
p. 39
p. 39
p. 39
p. 39
p. 39
p. 39
p. 39
p. 40
p. 40
p. 42
p. 42
p. 42
p. 42
p. 42
p. 42
p. 43
p. 43
p. 43
xii
xhi
ei
P ei
ne
jei
ei
Eh
ih
h
G ei
i,j
ej i
hj
Ph
I0h
h
H0 B; h
(W ())
u ()
D ei
f ei
(Wh (h ))
uh (h )
Eih
aei (, )
f e ()
x ei
xej i
H ei
r
cr
x
cr
x
i
Rei , Rei
S ei , S ei
S eBi , SeBi
i
B n,e
B n
h ei
h
NOTATION
coordinates of the ith discretization node
the ith finite element
finite element space of the ith element
number of local degrees of freedom
the jth local degree of freedom of the ith element
set of all the degrees of freedom of the ith element
set of all the finite elements
the ith global degree of freedom
set of all the n global degrees of freedom
mapping from local to global degrees of freedom for the ith element, G ei : {1, . . . , ne } 7 {1, . . . , n}
Kroneckers symbol
the jth local shape (base) function of the ith element
the jth global shape (base) function
global finite element space
set of indices of those global degrees of freedom that determine
the trace along h
the finite element space V h
the problem (W ) for varying computational domain
solution to (W ())
the coefficient matrix Dh restricted to the ith element
the righthand side vector f h restricted to the ith element
finite element discretization of the problem (W (h ))
solution to (Wh (h ))
set of the elements neighbouring with the ith element
contribution to the bilinear form a h (, ) from the ith element
contribution to the linear functional f h () from the ith element
block vector of all the corners of the ith element
coordinates of the jth corner of the ith element
mapping from the element nodal indices to the global nodal indices, Hei : {1, . . . , m + 1} 7 {1, . . . , nxh }
reference element
block vector of all the reference element nodes
the ith corner of the reference element
linear mapping, the related matrix, from the reference element to
the ith element, Rei : K r 7 K ei
linear mapping, the related matrix, of finite element functions defined over K r to the ones defined over K ei , S ei : Pr 7 Pei
linear mapping, the related matrix, of the finite element functions
defined over
K r to the ones
K ei under the operator
2defined
over
2
ei 2
2
r
e
i
B, S B : L (K )
7 L (K )
elementwise constant vector of the operator B applied to the solution uh
i
block vector of the elementwise constant vectors B n,e
p. 43
p. 43
p. 43
p. 43
p. 43
p. 43
p. 43
p. 43
p. 43
p. 43
p. 44
p. 44
p. 44
p. 44
p. 45
p. 45
p. 45
p. 45
p. 45
p. 45
p. 45
p. 45
p. 46
p. 46
p. 46
p. 46
p. 46
p. 46
p. 47
p. 47
p. 47
p. 47
p. 48
p. 48
p. 50
p. 50
p. 50
p. 50
xiii
Xh
X0 B; ; h
h %
ei
h
h
h0
4h h
extension
extends functions by zero, X h :
linear
operator
that
2
h
2
h
L ( ) 7 L () , , N
h
space of functions extended from H0 B; h by Xh1
approximation of by polygonal domains h from inner
characteristic function of
interpolation
associated to the ith element, ei :
1 operator
C (K ei ) 7 Pei
i
h
1
7 Ph
global interpolation operator, h : C (h )
i 1
h
h
7 H0 B; h
global interpolation operator, h0 : C (h )
p. 51
p. 51
p. 51
p. 52
p. 52
p. 52
p. 52
p. 53
cost functional, I : U L2 () 2 v 7 R
cost functional, J : U 7 R
continuous setting of the shape optimization problem
solution to (P )
parameterized cost functional, Je : 7 R
continuous setting of the shape optimization problem solved for
design parameters
solution to (Pe )
the regularized bilinear form a (, ) controlled by the shape
the multistate problem (W v ()) regularized by the regularization
parameter
regularized cost functional
regularized setting of the shape optimization problem
solution to (P )
regularized and parameterized cost functional, Je : 7 R
p. 68
p. 68
p. 68
p. 68
p. 68
p. 68
p. 68
p. 68
p. 68
p. 69
p. 69
p. 69
p. 69
p. 69
p. 69
p. 69
p, 69
p. 70
p. 70
p. 71
p. 71
p. 72
p. 72
p. 72
p. 72
Chapter 5
l , u
U
n
n
F
0 (), 1 ()
graph()
D
D0 , D 1
a (, )
nv
v
fv
f v ()
(W v ())
uv ()
I
J
(P )
Je
(Pe)
p
a, (, )
(Wv ())
J
(P )
Je
p. 72
p. 72
p. 72
p. 73
p. 73
p. 73
p. 74
xiv
(Pe )
p
nh
ih
Th
P 1 (Th )
xhh ,j
NOTATION
regularized setting of the shape optimization problem solved for
design parameters
solution to (Pe )
number of elements in the discretization of
the ith element in the discretization of
discretization of
space of continuous functions that are linear over ih
the jth corner of the ith element in the discretization of
p. 74
p. 74
p. 74
p. 74
p. 74
p. 74
p. 74
nxh
xh,j
Uh
h
h
h0 (h ), h1 (h )
T h (h )
(Wv,h (h ))
h
uv,h
( )
ah,h (, )
Dhh
f v,h ()
Jh
(Ph )
h
Jeh
(Peh )
p. 74
p. 74
p. 74
p. 74
p. 75
p. 75
p. 75
p. 76
number of constraints
constraint function, : Rn 7 Rn
number of shape nodes, nh := nxh
designtoshape mapping, h : Rn 7 Rnh
shapetomesh mapping, xh : Rnh 7 Rmnxh
block vector of all the initial grid nodes, x h0 Rmnxh
block vector of all the grid displacements between the current and
initial grid, 4xh Rmnxh
matrix that identically maps the shape displacements h onto the
corresponding grid nodal coordinates x h , Mh : Rnh 7 Rmnxh
stiffness matrix for the initial grid x h0 of the auxiliary elasticity
(shapetomesh) problem
righthand side vector involving the inhomogeneous Dirichlet
condition h of the auxiliary elasticity (shapetomesh) problem
p. 84
p. 84
p. 84
p. 84
p. 84
p. 84
p. 84
p. 76
p. 76
p. 76
p. 76
p. 80
p. 80
p. 80
p. 81
p. 81
Chapter 6
n
nh
h
xh
xh0
4xh
Mh
K h (xh0 )
b h (h )
p. 84
p. 84
p. 84
xv
f v,n (xh )
h
u v,n
(x )
i
u v,n,e
(xh )
v,n,ei h
B
(x )
h
B v,n
(x )
h
I
(QP)
(QP 1 (p0 ))
Hess
Grad
(LS(p0 , sQP ))
(QP 2 (p0 , d))
Hk
G
righthand side vector that arises from the discretized linear functional f v,h () for the vth state
h
solution vector (corresponds to u v,h
( )) of the arising vth linear system
h
the solution vector u v,n
(x ) restricted to the ith element
elementwise constant vector of the operator B applied to the soh
lution uv,h
(x )
i
block vector of the elementwise constant vectors B v,n,e
(xh )
nh
h
revisited discretized cost functional, I : R
Rmnxh
2 n h n v
[R
] 7 R
quadratic programming problem
quadratic programming subproblem for the line search approach
Hessian, matrix of all the second partial derivatives of a scalar
function
gradient of a vector function whose columns are gradients of the
particular components of the function
line search problem
quadratic programming subproblem for the trust region method
the kth successive approximation of the Hessian, k N
matrix involving the sensitivity of the multistate system upon the
grid displacements
p. 85
p. 85
p. 85
p. 85
p. 85
p. 86
p. 90
p. 90
p. 90
p. 90
p. 90
p. 91
p. 92
p. 96
Chapter 7
MC
ORing
m
yoke
westp , . . .
westc , . . .
x,i,j
pi,j
in
I
nI
Sc
Jv
B avg,v
nvm
p. 105
p. 105
p. 106
p. 108
p. 108
p. 108
p. 111
p. 111
p. 111
p. 112
p. 112
p. 112
p. 112
p. 114
p. 114
p. 114
p. 114
p. 114
xvi
avg,v
Bmin
h
v,h
B avg,v,n
h
x,i
pi
NOTATION
minimal required magnitude of the magnetic field of the vth state
problem
discretization of , h : Rmnh 7 R
discretization of v , v,h : Rmnh 7 R
discretization of B avg,v
design interface
the ith node in the regular discretization of the 1d shape domain
p. 114
p. 116
p. 116
p. 116
p. 117
p. 118
p. 118
Contents
Abstract
Acknowledgement
vii
Notation
ix
Contents
xvii
Introduction
1.1 General aspects of optimization . . . . . . . . . . . . . . . . .
1.1.1 Optimization problems: Classification and connections
1.1.2 Optimization methods . . . . . . . . . . . . . . . . .
1.1.3 Iterative methods for linear systems of equations . . .
1.1.4 Commercial versus academic software tools . . . . . .
1.2 Optimal shape design . . . . . . . . . . . . . . . . . . . . . .
1.3 Computational electromagnetism . . . . . . . . . . . . . . . .
1.4 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
1
2
2
3
3
4
5
6
6
9
9
10
11
13
13
13
14
15
17
19
20
21
24
24
25
26
28
29
xvii
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
xviii
3.4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
30
32
34
36
37
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
39
39
39
42
42
43
45
46
46
48
49
50
51
52
52
54
57
58
62
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
67
67
68
68
69
71
72
74
74
75
80
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
83
83
84
84
84
85
86
87
89
90
CONTENTS
6.3
6.4
7
xix
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 91
. 92
. 93
. 94
. 96
. 98
. 98
. 100
. 101
.
.
.
.
.
.
.
.
.
.
.
.
105
105
108
108
109
112
113
115
117
119
120
121
124
.
.
.
.
.
.
.
.
.
.
.
.
127
Bibliography
129
Curriculum vitae
143
xx
CONTENTS
Chapter 1
Introduction
Nowadays dynamic progress in computer technology has made powerful computers to become
cheap. This has been influencing the development of numerical methods. Many both commercial
and academic simulation software tools are available for a large variety of problems. Computer
simulations replaced prototyping. A usual picture is that developers in a company are modelling
a new product on a computer, doing some calculations, and thinking what parameters and how to
shift to achieve better properties of the product. Still increasing standard of technologies brings
together experts from different areas. Developers work is now much more interdisciplinary. It
involves
experts in the area of main interest, e.g., engineers, physicists, medics, economists, etc.,
theoretical mathematicians who introduce correct theories that can be used for mathematical
modelling,
numerical analysts who design efficient numerical methods and analyze their properties,
e.g., speedup, convergence rate, etc.,
and computer scientists who effectively implement the methods on a proper platform.
The people who are experienced in more areas are especially welcome to coordinate the design
process.
As far as the direct simulation is fast enough, it is straightforward to automatize also the
synthesis (design) process. To this end, a developer has to exactly formulate
the objective criterion saying what design is better,
the design parameters that can be changed including their possible limit values,
and some additional constraining criteria that the product must satisfy.
The objective criterion (the optimization goal) might be the minimal weight, the maximal output
power, the minimal cost, the minimal loss, etc. The design parameters are for example size of
the product, microstructure of the used material, or shape of the product. We might additionally
require that the product must not exceed a given volume, weight, or that it must be robust, e.g.,
stiff enough. Once we know these exactly, we have formulated an optimization problem that can
be solved automatically.
1
CHAPTER 1. INTRODUCTION
S [123], M ARROCCO
for example, D I BARBA ET AL . [18], B RANDST ATTER
ET AL . [30], L UK A
AND P IRONNEAU [132], TAKAHASHI [206].
It turns out that there is much in common in topology and shape optimization. Recently there
have appeared several papers in this context, like C EA ET AL . [40], R IETZ AND P ETERSSON [171],
TANG AND C HANG [207].
CHAPTER 1. INTRODUCTION
years ago, the development of conjugate gradients methods has been running. The conjugate
gradients methods are looking for minimum of the quadratic functional in the directions that are
conjugated by the energy scalar product related to the system matrix. The research was initiated
by H ESTENSEN AND S TIEFEL [90] and from then an extensive literature to this topic has been
written, see Z OUTENDIJK [221], G OLUB AND VAN L OAN [69], A XELSSON [13], S AAD [179].
Nowadays, the key point is the construction of proper preconditioners. The ones that turned
out to be the best are based on multigrid techniques. They construct a hierarchy of finite element
discretizations such that they first minimize the low frequencies (eigenvalues) of the residual error on a coarse discretization, which is very fast, and then the higher frequencies on a finer one,
which is again fast, as the low frequencies are not any longer present. The hierarchy can be constructed either with respect to the computational grid (geometrical multigrids) or with respect to
the structure of the system matrix (algebraic multigrids). Various topics on multigrid techniques
are presented in H ACKBUSCH [78], B RAMBLE [28], B RAMBLE , PASCIAK , AND X U [29], H IPTMAIR [91], J UNG AND L ANGER [102], R EITZINGER [169], H AASE AND L ANGER [74], H AASE
ET AL . [75, 76]. Applications in electromagnetism can be found in S CHINNERL ET AL . [183, 184].
A software package based on algebraic multigrids was done by R EITZINGER [168].
for the analysis or optimization are presented in K UHN , L ANGER , AND S CH OBERL
[117], S ILVA
AND B ITTERCOUNT [191], R ASMUSSEN ET AL . [166], PARKINSON AND BALLING [153]. An
example of more educational software system is in T SCHERNIAK AND S IGMUND [210]. A typical
commercial software directed to the academy is M ATLAB [208].
Until recently, one could hardly work in both the industry and academy, as their objectives
were rather different. The nowadays trend seems to be towards interdiciplinary work. Industrial
partners are invited to talk at scientific symposia, and many companies invest to further education
of their staff. The gap between the industry and academy gets smaller, thus, the difference between the commercial and scientific software does so. The commercial software should take more
into account the latest research progress and the scientists developing research software packages should put more effort into the documentation, userinterface, and better coordination of the
development. The more communication between the industry and academy there is, the more
improvement can be done.
CHAPTER 1. INTRODUCTION
and prove the corresponding convergence result. Finally, we discretize the state problem and,
consequently, the optimization problem by means of the finite element method. We prove the
convergence of the discretized optimized shapes to a continuous optimal one. The convergence
theory uses very standard tools of the functional and finite element analysis and it was inspired by
[85]. Nevertheless, one of its assumptions,
the monograph of H ASLINGER AND N EITTAANM AKI
namely the continuity of the mapping between the shape nodes and the remainding grid nodes,
is difficult to assure in practice. For nonacademic problems with complex geometries and for
fine discretizations one can hardly find such a continuous shapetomesh mapping, as for large
changes of the design shape some disturbed or even flipped elements can appear and the geometry
has to be remeshed. This brings the discontinuity of the cost functional into the business. In
Conclusion, possible outcomes are discussed.
In Chapter 6, we revisit our abstract optimization problem from the computational point of
view. We analyze the structure of the cost functional and present it as a compound mapping,
consisting of several smooth submappings. Therefore, we can prove the smoothness of the cost
functional, which justify us to use algorithms of the Newton type afterwards. We briefly mention
all the ingrediences of the sequential quadratic programming method. Then, we derive an efficient
method for the firstorder sensitivity analysis, including its implementation in Matlab, which is
enclosed on the CD. This is actually the heart of the whole thesis. Finally, we introduce a multilevel optimization algorithm, which is welldesigned to be adaptive with respect to aposteriori
the error analysis of the underlying finite element discretization of the state problem. We refer
to the very recent papers by S CHLEUPEN , M AUTE , AND R AMM [185], R AMM , M AUTE , AND
S CHWARZ [164], in which the finite element adaptivity is already used for calculating error of the
approximation of the cost functional.
At the beginning of Chapter 7, we present an application, which has arisen from the research on
magnetooptic effects. Our aim is to find optimal shapes of two electromagnets in order to minimize
inhomogeneities of the magnetic field in a certain area. The electromagnets have rather complex
3dimensional geometries. We formulate the cost functional, the set of admissible shapes, and the
state problem such that, simultaneously, we verify the related theoretical assumptions. We further
pose the corresponding reduced 2dimensional settings of the problems. Then, both 2d and 3d
numerical results are given. We discuss the speedup of the used adjoint method comparing to the
numerical differentiation, and the speedup of the multilevel approach with respect to the standard
approach. Finally, an optimized shape was manufactured and we are provided with physical measurements of the magnetic fields for both the original and optimized electromagnets. We present
the improvements of the magnetic field in terms of the cost functional.
In Conclusion, we summarize the results of this thesis and give directions of the further research.
CHAPTER 1. INTRODUCTION
Chapter 2
Mathematical modelling in
magnetostatics
In this chapter, we will start from Maxwells equations in a general timedependent 3dimensional
(3d) setting, we will pass through the timeharmonic case, and we will end up with the 3d magnetostatic boundary value problem. Neglecting the magnetic phenomena in a given direction, we
will arrive at the 2dimensional (2d) magnetostatic boundary value problem. Throughout this
chapter, we will formally describe the physical phenomena, rather than introduce all the necessary
assumptions on the smoothness of the domain or the differentiability of the physical quantities.
Mathematically correct settings will be introduced in Chapter 3.
For the theory of electromagnetism we refer to F EYNAM , L EIGHTON , AND S ANDS [60],
H AUS AND M ELCHER [87], S OLYMAR [197], and S TRATTON [201]. The monographs focused
[115], B OSSAVIT [26],
more on numerical modelling are given by K R I Z EK AND N EITTAANM AKI
VAN R IENEN [211], I DA AND BASTOS [99], KOST [112], M AYERGOYZ [133], S TEELE [199].
curl(H) = J + E +
B
curl(E) =
,
(2.1)
t
div(D) =
div(B) = 0
D = E and B = H,
(2.2)
where E denotes the electric field (electric intensity), D is the electric flux density, > 0 is
the permittivity, J is the external electric current density, > 0 is the electric conductivity,
0 is the charge density, H is the magnetic field, B is the magnetic flux density, > 0 is the
permeability, t 0 is the time and the differential operators are defined as follows:
div(v) :=
v2
v3
v1
+
+
,
x1 x2 x3
9
10
v2 v1
v3 v2
v1
v3
x2 x3 x3 x1 x1 x2
where E := E(x) and H := H(x) are complexvalued vector functions, i denotes the imaginary
unit, and Re(v) := (Re(v1 ), Re(v2 ), Re(v3 )) is the componentwise real part of the vector v :=
(v1 , v2 , v3 ). Moreover, we assume the charge density and Maxwells current to be zeros
= 0 and
D
= 0.
t
We also assume that the constitutive relations (2.2) are timeindependent, realvalued and linear
:= (x) and := (x),
rather than := (x, E) and := (x, H) in the nonlinear case. Finally, we assume the external
current density J and the conductivity to be timeindependent and realvalued
J := J(x) and := (x).
Now, Maxwells equations (2.1) can be rewritten as follows:
curl(H) = J + E
curl(E) = iB
div(D) = 0
div(B) = 0
in ,
(2.3)
where, according to (2.2), D = E and B = H. We prescribe that the electric field vanishes on
the boundary
n E = 0 on ,
(2.4)
where n denotes the outer unit normal to and where, given vectors u := (u 1 , u2 , u3 ) and
v := (v1 , v2 , v3 ),
u v := (u2 v3 u3 v2 , u3 v1 u1 v3 , u1 v2 u2 v1 )
is the vector cross product.
11
curl(E) = icurl(u)
in ,
div(iu) = 0 in
and last Maxwells equation is automatically fulfilled, since the vector identity div(curl(u)) = 0
holds. We consider the timeindependent case of (2.3). Taking := 0 and neglecting the electric
field, we arrive at the following magnetostatic boundary value problem
1
curl
curl(u) = J in
.
(2.5)
n u = 0 on
Let us assume that the magnetic field given by (2.5) does not significantly depend on the x 3
coordinate. This is often the case when J(x) = (0,0, J(x 1 , x2 )) and (x) = (x1 , x2 ) in a
large enough neighbourhood of the zeroplane Z := x R3 | x3 = 0 . We are interested in an
approximate solution of (2.5) in this neighbourhood. So, let us assume that
J(x) := (0, 0, J(x1 , x2 )) , (x) := (x1 , x2 ), and u(x) := (0, 0, u(x1 , x2 )) .
Using the latter, the problem (2.5) reduces to the following
1
div
grad(u) = J in 2d
u = 0 on 2d
where
2d := x0 = (x1 , x2 ) R2 | (x1 , x2 , 0)
(2.6)
represents a cross section of in the sense 2d {0} = Z, and where the differential operator
grad is defined as follows:
u u
grad(u) :=
,
,
x1 x2
where u is a scalar function. It is easy to see that the magnetic flux density is then given by
u
u
,
,0 ,
B=
x2 x1
where u solves (2.6).
12
Chapter 3
138], H IPTMAIR [93], N EITTAANM AKI AND S ARANEN [146, 147], S TEELE [199], S ILVESTER
AND F ERRARI [192]. The monograph by G IRAULT AND R AVIART [67] and the paper by H IPTMAIR [92] inspired us to build an abstract theoretical framework for the weak formulations and
consequent finite element discretizations of the linear elliptic boundary vectorvalue problems.
u + v = v + u,
(u + v) = u + v,
( + )u = u + u,
z V : u + z = v,
(u) = ()u,
1u = u.
Among others, the axioms imply that the following hold
0 V u V : 0 + u = u,
u V (u) V : u + (u) = 0.
13
14
u, v V.
The subset U V is called a subspace of V if it is also a linear vector space with respect to the
operations . and +.
Let V be a linear vector space. The mapping k k V : V 7 R is called a norm if for any
u, v V and any R the relations
ku + vkV kukV + kvkV ,
kukV = ||kukV ,
kukV 6= 0 if v 6= 0
(3.1)
hold. The space V equipped with a norm is called a normed linear vector space.
Let V be a normed linear vector space. The sequence {u n }
n=1 V is said to be convergent
if there exists u V such that
kun ukV 0, as n .
We denote it by un u in V .
Let M V be a subset of a normed linear vector space V . The subset M is said to be closed
if for any convergent sequence {un }
n=1 M the following is true
un u in V u M.
The subset M is said to be dense in V if the condition
u V {un }
n=1 M : un u in V
is satisfied. We denote it by
V = M in the norm k kV .
Let V be a normed linear vector space and U V be a subspace. The space
V /U := {[u] V | u V and v U : u + v [u]}
is called a quotient space. The space V /U equipped with the norm
k[u]kV /U := inf ku + vkV
vU
forms a normed linear vector space. Moreover, if U is a closed subspace, then the infimum is
realized on U and it becomes the minimum.
L(u) = L(u)
15
The set
Ker(L) := {u U | L(u) = 0}
(3.2)
is called the kernel of the operator L and it is a closed subspace of U . The linear operator L 1 :
V 7 U is called the inverse to L if
u U v V : L(u) = v L1 (v) = u.
The mapping f : U 7 R is called a functional. The space of continuous linear functionals
that are defined on a normed linear vector space U is called a dual space and it is denoted by U 0 .
The mapping h, i : U 0 U 7 R defined by
hf, ui := f (u),
f U 0 , u U,
is a norm. The space U 0 equipped with k kU 0 and with the following operations
hf + g, ui := hf, ui + hg, ui,
hf, ui := hf, ui
f, g U 0 , R, u U,
p
(u, u)H
is called the induced norm. Moreover, if the space H with the scalar product and the induced norm
is a Banach space, then it is called a Hilbert space. The following CauchySchwarz inequality
holds:
|(u, v)H | kukH kvkH
(3.3)
for any u, v H. Let U H be a closed subspace of H, then it is a Hilbert space, too.
16
Theorem 3.1. (Riesz theorem) Let H be a Hilbert space. Then for any f H 0 there exists exactly
one element u H such that
v H : (v, u)H = f (v).
(3.4)
Moreover,
kukH = kf kH 0 .
Proof. See O DEN
AND
(3.5)
Let H be a Hilbert space. The mapping a(, ) : H H 7 R is called a bilinear form if for
any fixed u H both the mappings a(, u) and a(u, ) are linear functionals. The bilinear form is
said to be continuous on H if there exists a positive constant C 1 such that
u, v H : |a(u, v)| C1 kukH kvkH .
The bilinear form is called Helliptic if there exists a positive constant C 2 such that
v H : |a(v, v)| C2 kvk2H
(3.6)
Lemma 3.1. (LaxMilgram lemma) Let H be a Hilbert space and let a(, ) be a continuous
bilinear form on H which is Helliptic with the constant C 2 . Then for any f H 0 there exists
exactly one element u V such that
v H : a(v, u) = f (v).
Moreover,
kukH
(3.7)
1
kf kH 0 .
C2
AND
1
a(v, v) f (v),
2
v H.
[115, p. 14].
N EITTAANM AKI
The normed linear vector spaces U and V are said to be isomorphically isometric if there exists
a onetoone linear operator L : U 7 V such that
u U : kL(u)kV = kukU .
The operator L is called an isomorphism.
Let H be a Hilbert space and U H be a closed subspace. The space U defined by
U := {u H | v U : (u, v)H = 0}
(3.8)
17
The vectors ei are the base vectors and the real numbers i are the coordinates of the vector u
in the base E. If the base consists of only a finite number of base vectors, we say that H is
finitedimensional, otherwise, H is infinitedimensional.
u v :=
n
X
ui vi ,
i=1
where u := (u1 , . . . , un ) Rn , v := (v1 , . . . , vn ) Rn stand for column vectors. Then the set
{e1 , . . . , en } forms the Euclidean base, where all the entries of the Euclidean base vector e i Rn
are zeros except for the ith entry which is one.
Let A := (A1 , . . . , Am ) : Rn 7 Rm be a linear vector operator acting between two Euclidean
spaces. Then we can represent A by the following matrix
A := .
..
.. , where ai,j := Ai (ej ) for i = 1, . . . , m, j = 1, . . . , n.
.
.
.
.
.
.
.
am,1 am,2 . . . am,n
We will also denote the matrix by A := (a i,j ) (ai,j )i,j Rmn . From the linearity of A it
follows that for a vector u := (u1 , u2 , . . . , un ) Rn
Pn
a1,j uj
Pj=1
n
j=1 a2,j uj
A(u) = A u, where A u :=
.
..
.
Pn
j=1 am,j uj
(3.9)
18
By the matrix AT := aTi,j Rnm we denote the transpose matrix to the matrix A, where
aTi,j := aj,i
for i = 1, . . . , n, j = 1, . . . , m.
u Rn ,
m
X
bi,k ak,j
for i = 1, . . . , p, j = 1, . . . , n.
k=1
The linear algebra provides a powerful tool for solving linear operator equations. Given a
linear mapping A : Rn 7 Rm and a vector f := (f1 , . . . , fm ) Rm , the linear operator
equation
A(u) = f ,
(3.10)
solved for u Rn , can be equivalently written as a system of linear algebraic equations, the
matrix form of which is
A u = f,
(3.11)
where the matrix A := (ai,j ) Rmn represents the linear operator A. Moreover, if m = n
and if there exists the inverse operator A1 : Rn 7 Rn , then the solution to the linear operator
equation (3.10) is represented by
u = A1 (f ).
The latter can be again written in terms of matrices. To this end, we introduce a multilinear form
det(A), called the determinant of the matrix A, which is recursively defined by
(
a1,1
,n = 1
det(A) := Pn
,
(3.12)
j+1
det(A1,j ) , n 2
j=1 (1)
where the matrix Ai,j R(n1)(n1) is made from the matrix A Rnn by excluding its ith
e := (af
row and jth column. Further, to the matrix A we associate the adjoint matrix A
i,j )
nn
R
by
i+j
af
det(Aj,i ) .
(3.13)
i,j := (1)
1
e
A
det(A)
A A1 = A1 A = I,
(3.14)
19
1
e f.
A
det(A)
(3.15)
Note that (3.15) is extremely inappropriate for practical calculations of u, since the computation
e is very timeconsuming. We will rather use (3.15) for analysis only.
of both det(A) and A
Given a nonsingular matrix A Rnn , the transposition and the inversion are mutually commutative and we abbreviate them as follows:
T
1
AT := A1 = AT
.
x3
x3
A
PSfrag replacements
e3
e2
A(e3 )
x2
e1
x1
x2
A(e2 )
x1
A(e1 )
20
Assumption 3.1. In all what follows we will assume that denotes a nonempty, bounded, open,
and connected subset of Rm , where m {2, 3}.
Nevertheless, all the results up to Chapter 5 are valid for any m N.
(3.16)
AND
Let k N {0}. The symbol C k () denotes the space of continuous realvalued functions
that are differentiable up to the order k. In particular, we denote the space of continuous functions
by
C() := C 0 (),
which, being equipped with the norm
kukC() := max |u(x)|,
x
x : u(x) .
n
Then v u C k () , where for x the function v u is defined by
(v u) (x) := v(u(x)).
21
i = 1, . . . , n, j = 1, . . . , l,
(3.17)
k=1
C k ()
k=1
where
|| u(x)
, || N
m
1
.
D u(x) := x1 ...xm
u(x)
, = (0, . . . , 0)
The following classical theorems of real analysis are due to RUDIN [178].
Theorem 3.3. (Taylors theorem) Let be an open subset of R m , m N, and let u C k .
Let further x := (x1 , . . . , xm ) and let z Rm be such that
t R : 0 t 1 x + tz .
22
Then
u(x + z) =
m
k1
Y
X
1 X
D u(x)
(xj )j + r(z),
i!
i=1
j=1
0||i
,i N
,i = 0
denotes the factorial of i N {0}, and where the remainder function r : 7 R satisfies
r(z)
= 0.
z0 |z|k1
lim
AND
m
Corollary 3.1. Let the assumptions on and u hold and let v C 1 () . Then the following
is satisfied
Z
Z
Z
(un) v ds ,
u div(v) dx =
grad(u) v dx +
where the differential operators grad and div are respectively defined as follows:
u
u
grad(u) :=
, u C 1 (),
,...,
x1
xm
div(v) :=
m
X
vi
,
xi
i=1
m
v := (v1 , . . . , vm ) C 1 () .
holds. Summing up the latter for the index i = 1, . . . , m, we get the assertion.
3
Corollary 3.2. Let R3 , L, and let u, v C 1 () . Then the following is satisfied
Z
Z
Z
(n u) v ds,
(3.18)
u curl(v) dx =
curl(u) v dx
3
where for u := (u1 , u2 , u3 ) C 1 () the differential operators curl is defined by
u3 u2 u1 u3 u2 u1
,
curl(u) :=
x2 x3 x3 x1 x1 x2
(3.19)
23
(3.20)
1 X
m
X
(k,l) uk
, bi
k=1 l=1
xl
(3.21)
(k,l)
where bi
R for i = 1, . . . , 2 , and where u := (u1 , . . . , u1 ) C 1 () 1 . We define the
adjoint operator B : C 1 () 2 7 C() 1 to the operator B by
B (v) :=
B1 (v), . . . , B1 (v)
, where
Bk (v)
:=
2 X
m
X
(k,l) vi
bi
i=1 l=1
xl
for k = 1, . . . , 2 ,
(3.22)
and where v := (v1 , . . . , v2 ) C 1 () 2 .
Moreover, we define the trace operator : C() 1 7 [C()]2 associated to B by
(u) := (1 (u), . . . , 2 (u)) , where i (u) :=
1 X
m
X
(k,l)
bi
k=1 l=1
uk | nl
for i = 1, . . . , 2 ,
(3.23)
Proof. Let u C 1 ()
and v C 1 () 2 be arbitrary. Using the previous definition, we
write down the lefthand side of (3.24)
Z
1
Z
Z
uk
vi
u B (v) dx =
B(u) v dx +
uk
vi dx +
dx =
xl
xl
1 X
2 X
m
X
(k,l)
bi
24
3
Theorem 3.5. (Stokes theorem) Let R 3 be an open set and let u := (u1 , u2 , u3 ) C 1 () ,
where is a 2dimensional surface with the boundary being a piecewise C 1 curve. Then the
following is satisfied
Z
Z
curl(u) n dx =
u n ds,
kukLp () :=
Z
|u| dx
1/p
u Lp ().
1 1
+ =1
p q
and let u Lp (), v Lq (), then the following Holder inequality holds
Z
Z
1/q
1/p Z
q
p
uv dx =
.
|v| dx
|u| dx
(3.25)
L () := u : R ess sup |u(x)| < +
x
u L ().
We say that the set is measurable in the Lebesgue sense if the following Lebesgue integral
exists
Z
dx,
meas() :=
25
1
Theorem 3.6. (Lebesgue dominated convergence theorem) Let {u n }
n=1 L () be a sequence
of functions measurable in the Lebesgues sense. Let u n u almost everywhere in , where
u : R is a function. If there exists a function v L 1 () such that |un | v almost
everywhere in for all n N, then u L 1 () and
Z
Z
u dx = lim
un dx.
n
AND
M AL Y [129, p. 26].
||
v C0 () :
zv dx = (1)
u D v dx.
(3.26)
We can easily see that for any u C k (), k N {0}, for a multiindex such that || k,
and for z := D u C(), which is the th classical derivative of u, the relation (3.26) holds
in virtue of Theorem 3.4. Therefore, we can extend the symbol D u and we denote the th
generalized derivative still by D u := z.
Now, for k N {0} we define the Sobolev spaces as follows:
H k () := {u L2 () | : || k D u L2 ()}.
The latter, equipped with the scalar product
X Z
(u, v)k, :=
D uD v dx,
||k
u, v H k (),
forms a Hilbert space with the following induced norm and seminorm
v
uX Z
q
u
|D u|2 dx,
kukk, := (u, u)k, , |u|k, := t
||=k
u H k (),
n
respectively. The Sobolev spaces of vector functions H k () , n N, equipped with the scalar
product
n
h
in
X
(u, v)n,k, :=
(ui , vi )k, , u, v H k () ,
i=1
26
AND
The function (u) is called the trace of u. The trace theorem enables us to define the space
H01 () := u H 1 () | (u) = 0 .
The spaces C
and [C0 ()]n , respectively, are dense in H 1 () and H01 () ,
i.e.,
1
n
n
n
H () = C
and H01 () = [C0 ()]n in the norm k kn,1, .
(3.27)
The next theorem extends Theorem 3.4.
Theorem 3.8. (Greens theorem) Let R m , L, and let u, v H 1 (). Then the relation
Z
Z
Z
v
u
u
(u)(v)ni ds for i = 1, . . . , m
v dx +
dx =
xi
xi
holds, where n := (n1 , . . . , nm ) denotes the outer unit normal to .
Proof. See N E C AS [141, p. 29].
Note that, avoiding some additional effort, yet we have not defined either the boundary integral
or the space L2 (), for which we refer to K UFNER , J OHN , AND F U C I K [116] or N E C AS [141].
The last theorem, of which we will make use later when analyzing the ellipticity of differential
z v dx,
u div(v) dx =
v [C0 ()] :
and we denote the generalized gradient by grad(u) := z. In particular, from (3.26) it is clear that
u
u
1
,...,
,
u H () : grad(u) =
x1
xm
27
where the partial derivatives are the generalized ones. We define the space
m
H(grad; ) := u L2 () | z L2 () : z = grad(u) .
uv dx +
u, v H(grad; ),
forms a Hilbert space. We introduce the following induced norm and seminorm
kukgrad, :=
(u, u)grad, ,
|u|grad, :=
sZ
kgrad(u)k2 dx,
u H(grad; ),
respectively. Clearly
kuk2grad, = kuk20, + |u|2grad, ,
u H(grad; ),
holds. From the definition of H 1 () it is obvious that H(grad; ) = H 1 (), (u, v)grad, =
(u, v)1, , kukgrad, = kuk1, , and |u|grad, = |u|1, . Therefore, Theorem 3.7 holds and we can
define the space
H0 (grad; ) := {u H(grad; ) | (u) = 0} ,
where the trace operator
(u) := (u)n,
where n := (n1 , . . . , nm ) denotes the outer unit normal to and is due to Theorem 3.8.
Obviously, H0 (grad; ) = H01 () holds and H0 (grad; )/Ker(grad; ) is equal to H 01 (),
since Ker(grad; ) = {0}, where due to (3.2)
Ker(grad; ) := {u H0 (grad; ) | grad(u) = 0} .
(3.28)
m
Theorem 3.10. (Greens
1 theorem
m in H(grad)) Let R , L, and let u H(grad; ),
v := (v1 , . . . , vm ) H () . Then the relation
grad(u) v dx +
u div(v) dx =
(u) v dx
m
holds, where the differential operator div is extended onto H 1 () as follows:
div(v) :=
m
X
vi
,
xi
i=1
m
v := (v1 , . . . , vm ) H 1 () .
Proof. We use Theorem 3.8 and similar arguments as in the proof of Corollary 3.1.
Finally, in Theorem 3.9 we replace the symbols kuk1, and |u|1, by the symbols kukgrad,
and |u|grad, , respectively, and the theorem holds with the same constant C 3 .
28
v [C0 ()] :
u curl(v) dx =
z v dx,
u, v H(curl; ),
u H(curl; ),
AND
u H(curl; ).
R AVIART [67, p. 34].
Theorem 3.11. (Trace theorem in H(curl)) Let R 3 , L. Then there exists exactly one
3
linear continuous operator : H(curl; ) 7 H 1/2 () such that
3
u C () : (u) = n u| ,
AND
R AVIART [67, Corollary 2.9], the space Ker(curl; ) is equal to the space
n
o
3
H0,0 (curl; ) := u L2 () p H01 () : u = grad(p)
(3.29)
AND
and, by H IPTMAIR [91, p. 9495], the quotient space H0 (curl; )/Ker(curl; ) is isomorphically isometric to
Z
1
u grad(v) dx = 0 ,
(3.30)
H0, (curl; ) := u H0 (curl; ) v H0 () :
29
i.e.,
The spaces C
3
and H0 (curl; ) = [C0 ()]3 in the norm k kcurl, .
H(curl; ) = C
(3.31)
3
3
holds, where h(u), vi denotes the duality pairing between H 1/2 () and H 1/2 () .
AND
The last theorem is a Friedrichslike inequality and it will be useful for analyzing the ellipticity of differential operators defined on H(curl; ).
Theorem 3.13. (Friedrichs inequality in H(curl)) Let R 3 , L. Then there exists a
positive constant C4 C4 () such that
u H0, (curl; ) : kukcurl, C4 |u|curl, .
Proof. See H IPTMAIR [91, p. 96].
v C0 () :
u grad(v) dx =
zv dx,
u, v H(div; ),
u H(div; ),
AND
u H(div; ).
R AVIART [67, p. 2728].
30
Theorem 3.14. (Trace theorem in H(div)) Let R 3 , L. Then there exists exactly one
linear continuous operator : H(div; ) 7 H 1/2 () such that
3
u C () : (u) = n u| ,
AND
i.e.,
The spaces C
3
and H0 (div; ) = [C0 ()]3 in the norm k kdiv, .
H(div; ) = C
holds, where h(u), vi denotes the duality pairing between (u) H 1/2 () and v
H 1/2 ().
Proof. See G IRAULT
AND
1
operator B onto a subspace of L () . The function z
L2 () 2 is said to be the gen
(3.33)
31
which can be shown to be a scalar product on H(B; ). The space H(B; ), together with this
scalar product, forms a Hilbert space. The induced norm and seminorm are as follows:
sZ
q
kB(u)k2 dx, u H(B; ),
kukB, := (u, u)B, , |u|B, :=
u H(B; ).
(3.34)
(3.36)
(3.37)
holds.
Proof. Here, we use exactly the same technique as presented in H IPTMAIR [91, p. 9495].
Let us recall the norm in the quotient space H 0 (B; )/Ker(B; )
k[v]kH0 (B;)/Ker(B;) :=
min
wKer(B;)
kv + wkB, ,
v H0 (B; ).
We look for a subspace of H0 (B; ) that consists of the minimizers v + w(v) determined as
follows:
n
o
kv + w(v)k 2B, = kvk2B, +
min
2 (v, w(v)) 1 ,0, + kw(v)k21 ,0, .
w(v)Ker(B;)
Ker(B; )
32
and, since (, )1 ,0, is a scalar product on Ker(B; ), by Theorem 3.1 w(v) is unique. Therefore, the minimizer u := v + w(v) H0 (B; ) is uniquely characterized by
Ker(B; ) : (u, )1 ,0, = (u, )B, = 0.
(3.38)
The space H0, (B; ), see (3.36), which consists of such minimizers is a closed subspace of
H0 (B; ) and, due to (3.38) and (3.8),
H0, (B; ) = Ker(B; ) ,
which completes the proof.
Further, we assume that the following Greens formula holds.
Assumption 3.4. Let Rm , L, and let u H(B; ), v H 1 () 2 . We assume that
the relation
Z
Z
B(u) v dx +
u B (v) dx = h(u), vi
2
2
holds, in which the duality pairing between H 1/2 ()
and H 1/2 ()
is denoted by
h(u), vi .
Finally, we will need the ellipticity. To this end, we assume that the following Friedrichslike
inequality holds.
Assumption 3.5. Let Rm , L. We assume that there exists a positive constant C 5
C5 () such that
u H0, (B; ) : kukB, C5 |u|B, .
At the end, we summarize how the abstract operators B, B , and read in the spaces introduced above.
m
N
N
3
1
1
m
3
2
m
1
3
H(B; )
H(grad; )
H(div; )
H(curl; )
B
grad
div
curl
B
div
grad
curl
(v)
vn
nv
nv
33
where B, B , and are defined by Definition 3.2, and where D C 1 () 2 2 , 2 N, is a
uniformly positive definite realvalued matrix, i.e., there exists a constant C 6 > 0 such that
x v R2 : v (D(x) v) C6 kvk2 ,
(3.39)
1
2
1
and where f C() . The function u C () is called the classical solution to (S).
Now, we introduce a weak setting of (S), which will enable us to weaken the assumptions on
the differentiability of the data in (S) and to deal with problems of more practical purposes. Let
us take into account the extensions of definitions of B, B , and from Section 3.3.6, as well as
Assumptions 3.23.5 and Lemma 3.5 introduced there. We define the continuous bilinear form
a : H(B; ) H(B; ) 7 R and the continuous linear functional f : H(B; ) 7 R by
Z
B(v) (D B(u)) dx, u, v H(B; ),
(3.40)
a(v, u) :=
f (v) :=
f v dx,
v H(B; ),
(3.41)
respectively, where f L2 () 1 , D := (di,j ) is a matrix the entries of which di,j L (),
i, j = 1, . . . , 2 , and the condition (3.39) holds almost everywhere (a.e.) in . A weak formulation
of the problem (S) reads as follows:
)
Find u H0 (B; ):
.
(3.42)
a(v, u) = f (v) v H0 (B; )
Just by applying Corollary 3.3, we can see that the classical solution u C 2 () 1 of the
problem (S) is also a solution to (3.42). However, the problem (3.42) admits more general and
physically still reasonable data.
We can observe that if u H0 (B; ) is a solution to (3.42), then for any p Ker(B; )
the function u + p is a solution, too. This indicates a Neumannlike problem. Therefore, we
restrict our consideration onto the quotient space H 0 (B; )/Ker(B; ), which is by Lemma 3.5
isomorphically isometric to H0, (B; )), being a subspace of H0 (B; ). In this case, we have to
introduce a compatibility condition on the righthand side f
Z
p Ker(B; ) :
f p dx = 0.
(3.43)
v H0, (B; )
(W )
where f satisfies (3.43). It can be easily verified that a solution u to (W ) also solves the problem
(3.42). On the other hand, the next theorem shows that (W ) has a unique solution, unlike the
problem (3.42).
Theorem 3.16. There exists exactly one solution u H 0, (B; ) to the problem (W ). Moreover,
there exists a positive constant C7 such that
kukB, C7 kf k1 ,0, .
(3.44)
34
Proof. We will check the assumptions of Lemma 3.1 and the assertion then follows.
Since H0, (B; ) is a closed subspace of H0 (B; ), it is also a Hilbert space equipped with
the scalar product (3.33). The form a is clearly bilinear. Concerning the matrix D, we denote
d := max ess sup |di,j (x)|
(3.45)
i,j
where u, v H(B; ), and where we used that D [L ()]2 2 , (3.3), and (3.34), respectively. The H0, (B; )ellipticity of a follows from
Z
Z
C6
a(v, v) =
B(v) (D B(v)) dx C6
kB(v)k2 dx 2 kvk2B, , v H0, (B; ),
C5
(3.46)
where (3.39) and Assumption 3.5 were used, respectively. Finally, f is obviously a linear functional on H(B; ) and it is continuous thereon, too, since
Z
|f (v)| = f v dx kf k1 ,0, kvk1 ,0, kf k1 ,0, kvkB, , v H(B; ),
where we used (3.3) and (3.34), respectively. The assertion now follows from Lemma 3.1, where
the H0, (B; )ellipticity constant is
C6
C7 := 2 .
C5
(W )
35
Proof. The proof is fairly the same as the one of Theorem 3.16. The continuity of a is proven as
follows:
Z
Z
|a (v, u)| d B(v) B(u) dx + v u dx max {d, } kvkB, kukB, , (3.48)
where u, v H(B; ) and where d is given by (3.45). The H0 (B; )ellipticity of a follows
from
Z
Z
kB(v)k2 dx +
a (v, v) C6
|v|2 dx min{C6 , }kvk2B, , v H(B; ),
(3.49)
(3.50)
where u H0 (B; ) are the solutions to (W ) and u H0, (B; ) is the solution to (W ).
Proof. Let > 0 be arbitrary. Using (3.39) and the definitions of (W ) and (W ), we have
Z
2
2
kB(u ) B(u)k2 dx
kB(u ) B(u)k2 ,0, = |u u|B, =
Z
1
1
a(u u, u u)
B(u u) (D B(u u)) dx =
C6
C6
1
1
a (u u, u u)) =
(f (u u) a (u u, u)) =
C6
C6
Z
1
f (u u) a(u u, u) (u u) u dx
=
C6
(3.51)
Using the orthogonal decomposition (3.37), there exists u, H0, (B; ) and there exists
u,0 Ker(B; ) such that
u = u, + u,0 .
Using the latter, the condition (3.43), (3.38) and (3.25), the estimate (3.51) reads
kB(u ) B(u)k22 ,0, = kB(u, ) B(u)k22 ,0,
Z
1
ku, ukB, kuk1 ,0, .
(u, u) u dx
C6
C6
C5
|u, u|B, kuk1 ,0, .
C6
36
B := curl,
and
(v) := n v| ,
where n := (n1 , n2 , n3 ) denotes the outer unit normal to . Due to (2.5), we determine the
symbols
1
D := , f := J,
3
where L (), > 0 a.e. in , J L2 () . The condition (3.39) is now equivalent to
1 > 0 : (x) 1 a.e. in
(3.52)
1
.
1
Since Ker(curl; ) is equal to the space H 0,0 (curl; ) defined by (3.29) the condition (3.43)
reads as follows:
Z
1
J grad(p) dx = 0.
p H0 () :
Finally, we specify the terms in (W ). As we have seen in Section 3.3.4, the quotient space
H0 (curl; )/Ker(curl; ) is isomorphically isometric to the space H 0, (curl; ), which was
defined by (3.30). The bilinear form (3.40) and the linear form (3.41) are, respectively, determined
by
Z
1
a(v, u) :=
curl(v)
curl(u) dx, u, v H(curl; ),
and by
f (v) :=
J v dx,
v H(curl; ).
We have specified all the assumptions on the abstract weak formulation (W ) introduced in Section 3.4. Therefore, Theorem 3.16 holds with the H0, (curl; )ellipticity constant
C7 :=
1
,
0 C42
where C4 is given by Theorem 3.13. In case of the formulation (W ), we only recall the regularized
bilinear form
Z
Z
1
v u dx, u, v H(curl; ),
curl(u) dx +
curl(v)
a (v, u) :=
37
B := div,
and
(v) := v| n,
where n := (n1 , n2 ) denotes the outer unit normal to . Due to (2.6), we determine the symbols
D :=
1
,
f := J,
(3.53)
where L (), > 0 a.e. in , J L2 (). We again replace the condition (3.39) by (3.52).
Since Ker(grad; ) defined by (3.28) is equal to the zero space {0}, the condition (3.43) always
holds and does not need to be introduced in this case. Finally, we specify the terms in (W ). As
we have seen in Section 3.3.3, the quotient space H0 (grad; )/{0} is equal to the space H 01 ().
The bilinear form (3.40) and the linear functional (3.41) are, respectively, determined by
Z
1
grad(u) dx, u, v H 1 (),
a(v, u) :=
grad(v)
and by
f (v) :=
Jv dx,
v H 1 ().
1
,
0 C32
where C3 is given by Theorem 3.9. In this case, we do not need to introduce the regularized
problem (W ), since all the spaces are equal
H0, (grad; ) = H0 (grad; ) = H01 ().
38
Chapter 4
f (v) :=
f h v dx,
39
v H(B; ),
(4.2)
40
where f h L2 ()
1
is such that
h > 0 p Ker(B; ) :
2
i,j=1
f h p dx = 0,
h > 0 : ess sup dhi,j (x) ess sup |di,j (x)| ,
i, j = 1, . . . , 2 .
(4.3)
(4.4)
Find uh Vh :
,
ah vh , uh = f h vh
vh Vh
(Wh )
Theorem 4.1. For each > 0 and h > 0 there exists a unique solution u h Vh to the problem
(Wh ). Moreover, there exists a positive constant C 8 () such that
h > 0 : kuh kB, C8 ()
f h
.
1 ,0,
Proof. The proof is fairly the same as the one of Theorem 3.17.
Let > 0 and h > 0 be arbitrary. By definition, V h is a closed subspace of H0 (B; ),
therefore, it is also a Hilbert space. The form a h is obviously bilinear and f h is a linear functional
on Vh . We have the continuity of ah
h
(v,
w)
max{dh , } kvkB, kwkB, , v, w H(B; ),
a
where
h
d := max ess sup di,j (x) .
h
i,j
Vh ellipticity
of
ah
v H(B; ),
(4.5)
41
The following lemma, cf. B RAESS [27], C IARLET [45], or S TRANG AND F IX [200], says
that we can study the approximation properties of u h via the approximation of H0 (B; ) by its
subspaces Vh and via the approximation properties of the forms a h and f h .
Lemma 4.1. (1st Strangs lemma) Let > 0 and h > 0 be a regularization and discretization
parameter, respectively. Then there exists a positive constant C 9 (), independent of h, such that
h
h
h
v V :
u u
C9 ()
u vh
+
B,
B,
a v h , u h v h a h v h , u h v h
+
+
kuh vh kB,
)
f uh vh f h uh vh
+
,
kuh vh kB,
where u H0 (B; ) is a solution to (W ) and uh Vh are solutions to (Wh ).
Proof. Let h > 0 and vh Vh be arbitrary. By the triangle inequality,
u vh
+
uh vh
u uh
B,
B,
B,
(4.6)
(4.7)
Now, we use the Vh ellipticity of a , the bilinearity of a , ah , and the definitions of problems (W ), (Wh ), respectively, and we get
2
ah wh , wh = a wh , u vh a wh , u vh + ah wh , wh =
C8 ()
wh
B,
= a wh , u vh + a wh , vh ah wh , vh +
+ ah wh , uh a wh , u
=
= a wh , u vh + a wh , vh ah wh , vh +
+ f h wh f wh .
Dividing the latter by kwh kB, and using (3.48) yield
C8 ()
wh
B,
max{d, }
u vh
a vh , wh ah vh , wh
+
+
kwh kB,
B,
f wh f h wh
+
.
kwh kB,
(4.8)
Combining (4.5), (4.6), (4.7), and (4.8), the assertion is proved, where the constant is as follows:
max{d, }
1
C9 () := max 1 +
,
.
min{C6 , } min{C6 , }
42
(4.9)
where the matrix An Rnn and the righthand side vector f n Rn are as follows:
n
n
An := ah (vi , vj )
, f n := f h (vi )
,
i,j=1
i=1
(4.10)
respectively, and the solution vector u n := un,1 , . . . , un,n Rn corresponds to the approximate
solution uh in the following way
n
X
uh =
un,i vi .
i=1
The finite element method is a special case of the Galerkin method. The base {v 1 , . . . , vn }
of the space Vh is chosen such that the matrix An is sparse. In this case the system (4.9) can be
solved much faster and the matrix An takes less computer memory. The finite element method is
determined as follows:
The domain Rm is decomposed into smaller convex subdomains, e.g., line segments
for m = 1, triangles for m = 2, or tetrahedra for m = 3.
The base {v1 , . . . , vn } is chosen as simple functions, e.g., polynomials. The space V h is
called the space of finite elements.
The basis functions v1 , . . . , vn have small supports, which make the matrix A n to be sparse.
Ki ,
i=1
Ki 6= for each i = 1, . . . , nh ,
Ki 6= Kj Ki Kj = for each i, j = 1, . . . , nh ,
any face of any Ki is either a subset of the boundary h or a face of another element Kj ,
where i, j = 1, . . . , nh ,
each Ki has exactly m + 1 faces.
The last assumption means that in the cases of m = 1, m = 2, and m = 3 we deal with line
segments, triangles, and tetrahedra, respectively. This assumption will provide us to introduce a
reference element.
43
The set
T h := {Ki | i = 1, . . . , nh }
is called a discretization of h . In Fig. 4.1 we can see some discretizations, which were generated
where xhi :=
xhi,1 , . . . , xhi,m
discretization T h , where nxh N stands for the number of the discretization nodes.
if
jh |([C (K e )]1 )0 = ie ,
i = 1, . . . , ne ,
j = 1, . . . , n.
44
e
e
e
e
Since dim(P
e ) = nee , ne eN, and 1 , . . . , ne are linearly independent, then there
exists a basis 1 , . . . , ne P such that
(
1 ,i = j
e e
.
i j = i,j , i, j = 1, . . . , ne , where i,j :=
0 , i 6= j
These base functions are called shape functions. In the same virtue, we introduce global shape
functions h1 , . . . , hn : h 7 R1 such that
(4.13)
ih hj = i,j , i, j = 1, . . . , n.
e Eh,
i = 1, . . . , ne .
The global shape functions form a basis for the following space
(
)
n
X
Ph := vh =
vi hi v1 , . . . , vn R .
(4.14)
i=1
Hence, the space Ph consists of such functions that are elementwise in P e , i.e.,
vh Ph e E h : vh |K e Pe
We need Ph to be a subspace of H B; h . This property is called the conformity of the finite
elements. The following lemma gives a sufficient condition on the conformity.
Lemma 4.2. Let vh Ph . Then vh H B; h if for any two adjacent elements ei , ej E h ,
i 6= j, with a common face fi,j := Ki Kj , the trace is continuous over fi,j , i.e.,
Ki vh |Ki |fi,j = Kj vh |Kj |fi,j ,
(4.15)
where is given by (3.23). Note that the minus sign appeared, since the outer unit normal vectors
on fi,j satisfy nKi = nKj .
Proof. Let vh Ph and let (4.15) holds. Clearly vh L2 h 1 . We will prove that vh
H B; h by means of (3.32). We take zh L2 h 2 such that
zh |Ki := B vh |K i , i = 1, . . . , nh .
Let wh C0 h 2 be arbitrary. Then, Corollary 3.3 and (4.15) yield
Z
X Z
h
h
vh B wh dx =
v B w dx =
h
X Z
=
Ki T
=
=
Ki T h
h
Ki
z w dx +
zh wh dx +
zh wh dx,
Ki
X Z
fi,j
Ki
fi,j
Ki v
w ds
Ki vh wh ds +
Z
fi,j
Kj vh wh ds
45
The next assumption ensures the conformity of the finite elements, i.e., P h H B; h . We
refer to H IPTMAIR [92] for a unified way of the design of conforming finite elements.
Assumption 4.1. Let e E h be an element and let f K e denote a face. We assume that
the degrees of freedom connected to the face f are exactly the ones which determine the trace
K e (ve ) |f , where ve Pe .
Further, we introduce the set of indices of those degrees of freedom which determine the trace
h . Due to (4.14), we can write an arbitrary v h Ph as
vh =
n
X
vih hi ,
i=1
where vih R for i = 1, . . . , n and where hi denote the global shape functions. Then,
n
X
vih h hi .
h v h =
i=1
h
Then, the finite element space V h H0 B; h H0 B; h is defined by
h
o
n
H0 B; h := vh Ph i I0h : ih vh = 0 .
(4.16)
(4.17)
v H0 (B; )
(W ())
As long as L, the existence of the unique solution u () to (W ()) is given by Theorem 3.17.
Further, let h > 0 be a discretization parameter and let h be a nonempty polyhedral
subdomain. Then, h L. Let T h be a discretization of h and let E h be the corresponding set
of finite elements. Concerning the bilinear form a h and the linear functional f h given by (4.1) and
(4.2), respectively, we assume that for each e E h there exist a constant matrix De R2 2
and vector f e R1 such that
x K e : Dh (x) = De and f h (x) = f e .
The Galerkin approximation of the problem (W (h )) reads as follows:
h
Find uh h H0 B; h :
h ,
h
h
h
h
h
h
h
h
a v , u
=f v
v H0 B;
(4.18)
(Wh (h ))
h
where ah , f h , and H0 B; h are respectively
defined by (4.1), (4.2), and by (4.17). The exis
tence and uniqueness of the solution u h h to (Wh ) follows from Theorem 4.1.
46
where
n
o
Eih := e E h | j {1, . . . , ne } : G e (j) = i ,
i = 1, . . . , n,
(4.20)
is the set of the elements neighbouring with e i . Since ah is a bilinear form and f h is a linear
functional, we assemble the matrix A n and the righthand side vector f n , see (4.10), elementwise.
Due to (4.19), each element contributes only by its n e global degrees of freedom, i.e.,
(An )i,j
eEih Ejh
ne
X
ae ( ek , el ) ,
k,l=1
(f )i =
ne
X X
eEih
f e ( ek ) ,
i, j = 1, . . . , n,
where the local contributions to the matrix and to the righthand side vector are
Z
Z
e e e
e
e
e
a ( k , l ) :=
B( k ) (D B( l )) dx +
ek el dx, k, l = 1, . . . , ne ,
Ke
(4.21)
k=1
(4.22)
Ke
( ek )
:=
Ke
f e ek dx,
k = 1, . . . , ne ,
(4.23)
(4.24)
i=1
where u n := un,1 , . . . , un,n Rn denotes the solution to the linear system (4.9).
if
xei = xhj ,
i = 1, . . . , m + 1,
j = 1, . . . , nxh .
(4.26)
47
We further introduce a reference element r := K r , Pr , r such that the polyhedral domain K r
is determined by the following block column vector consisting of the reference corners
i
h
m(m+1)
r
cr , . . . , x[
cr := x
,
x
1
m+1 R
r
e
r
r ,...,x
r
cr := x
d
d
where x
i
i,1
i,m K , i = 1, . . . , m + 1, and where dim(P ) = dim(P ) = ne .
To each element e E h , we associate a onetoone linear mapping R e : K r 7 K e defined by
b + re ,
x := Re (b
x) := Re x
b Kr,
x K e, x
(4.27)
where Re Rmm is a nonsingular matrix and re Rm is a vector both of which are uniquely
determined by xe as follows:
cr + re = xei ,
Re x
i
i = 1, . . . , m + 1.
(4.28)
Obviously, both Re and re are continuously differentiable with respect to each coordinate of the
corners xe .
PSfrag replacements
x
c2
x2
Re
Ke
Kr
0
x
c1
x1
PSfrag replacements
Figure 4.2: A transformation between the reference and an element domain
b
v
Se
1 v
b
0
Kr
1
x
c1
S e (b
v)
x
c2
x2
Ke
x1
Figure 4.3: A transformation between the reference and an element shape function
i.e.,
r
Further, let us denote by br1 , . . . , c
ne the shape functions acting on the reference element r,
ir brj = i,j ,
i, j = 1, . . . , ne .
48
Assumption 4.2. We assume that there exist nonsingular matrices S e R1 1 and SeB
R2 2 , both of which are continuously differentiable with respect to the corners x e , such that
Se bri (b
x) = ei (x) ,
and
i = 1, . . . , ne ,
x) = Bx ( ei (x)) ,
SeB Bxb bri (b
i = 1, . . . , ne ,
where x := Re (b
x), and where Bx and Bxb , respectively, stand for the differential operator B,
defined by (3.21), with respect to the global coordinates x and withrespect to
co the reference
b. We define transformations S e : Pr 7 Pe and S eB : L2 (K r ) 2 7 L2 (K e ) 2
ordinates x
by
b(b
S e (b
v (b
x)) := Se v
x) and S eB (Bxb (b
v (b
x))) := SeB Bxb (b
v(b
x)) ,
b (b
where v
x) Pr .
The linear transformations S eB and S e are associated to the differential operator B and to the
identity operator, respectively. In general, the theory of differential forms, cf. H IPTMAIR [91],
can be used in order to derive a canonical transformation, see H IPTMAIR [92], which is related to
some differential operator and to some degrees of freedom.
Z
Kr
S crk Se brl |det(Re )| db
x,
(4.29)
( ek )
Kr
f e Se c
rk |det(Re )| db
x.
(4.30)
Now we employ the Gaussian quadrature method, cf. R ALSTON [163], C IARLET AND L I [46]. Having a sufficient number of Gaussian integration points, we can calculate the integrals exactly. Then, the matrix and the righthand side vector in (4.10) are evaluated elementwise,
where the contributions of the elements are
ONS
Ae :=
nG
X
i=1
nG
T
T
X
e
e c
G e c
e
e c
G
G
G +
wiG B eB xc
w
B
x
,
B
xG
x
i
B
i
i
i
i
i=1
f :=
nG
X
i=1
T
G
g e,
wiG B e xc
i
d
G, . . . , x
G K r are the Gaussian integration points, w G , . . . , w G R are the Gaussian
where xc
g
nG
1
1
integration weights, and where
i
i
h
h
r
r
x) , . . . , c
x) ,
x) , B e (b
x) := Se br1 (b
x) , . . . , Bxb c
B eB (b
x) := SeB Bxb br1 (b
ne (b
ne (b
49
I e := |det(Re )| Ine ,
g e := |det(Re )| f e ,
where Ine Rne ne is the unit matrix. Note that we will employ the lowestorder, i.e., linear,
finite elements. In this case, since both D h and f h are elementwise constant, then all the integrands
are linear over the element, thus, we will employ only one Gaussian point being the mass point of
K r , and the corresponding weight. These are respectively as follows:
1
, for m = 1
2
G
1
1
x1 :=
, for m = 2 , w1G := 1.
3, 3
1 1 1
, for m = 3
6, 6, 6
The name BDB integrators is due to the structure of the contributions to the bilinear form. The
differential operator is involved in the matrix B e while the matrix D e or I e provide the material
properties and geometrical parameters of the element domain K e .
for u n := un,1 , . . . , un,n Rn . In fact, we rather look for B uh (x) , which is elementwise
constant, since we have employed the lowest, i.e., the firstorder finite elements only. Therefore,
50
we can describe B uh (x) by the block column vector
h
n,en i
1
B n := B n,e
, . . . , B h R 2 n h
such that
i
B uh (x) |Ki = B n,e
,
where
i
B n,e
:=
ne
X
j=1
un,G ei(j) SeBi Bxb brj (b
x ) R 2
for i = 1, . . . , nh .
This procedure which assembles the vector B n is called postprocessing and it is depicted in Algorithm 2.
Algorithm 2 Finite element method: postprocessing
Given u n
B n := 0
for i := 1, . . . , nh do
i
B n,e
:= 0
for j := 1, . . . , ne do
i
i
x)
Evaluate B n,e
:= B n,e
+ un,G ei(j) SeBi Bxb brj (b
end for
k := 2 (i 1)
for j := 1, . . . , 2 do
i
[B n ]k+j := [B n,e
]j
end for
end for
(4.31)
Convention 4.1. In what follows, we will assume that there exists h > 0, being, e.g., the minimum
diameter of a sphere (or circle) containing , such that any considered discretization parameter
h fulfills
h h.
(4.32)
The aim of this section
is to prove a convergence, in some sense, of the approximate finite
element solutions uh h to the true solution u ().
51
h
2
h
polyhedral subdomain. We define a linear extension operator X : L
7 L () by
Xh (v(x))
:=
v(x)
0
, x h
,
, x \ h
h i
.
v(x) L2 h
(4.33)
h
Proof. Let vh H0 B; h be arbitrary. We denote bh := Xh2 B vh . Clearly, bh
1 h 2
2
2
= Xh1 vh B () dx.
The latter implies that bh = B Xh1 vh and Xh1 vh H(B; ). Further, let H 1 () 2 ,
then |h H 1 h 2 . By (4.33), by Assumption 3.4, and since h vh = 0, we get
Z
Z
E
D
h
h
h
h
Xh1 vh B () dx =
dx +
B X 1 v
=
,
X 1 v
Z
Z
=
B vh |h dx +
vh B (|h ) dx =
h
h
E
D
h
= h v , |h
= 0,
h
2
stands for the trace operator along . Therefore,
where : H(B; ) 7 H 1/2 ()
h
h
h
h
H0 (B; ). Since H0 B;
X 1 v
is a finitedimensional Hilbert space (of the dih
h
h
mension less than n) and X1 : H0 B;
7 H0 (B; ) is a linear mapping, then the set
h
X0 B; ; h , defined by (4.34), is obviously a closed subspace of H 0 (B; ), hence, again a
finitedimensional Hilbert space.
m
Let
L be a domain and let h > 0 be a discretization parameter. We say that the
h R ,
class h>0 , h approximates from the inner if the following is satisfied
xh h x :
xh x
h.
(4.35)
52
ie ( e (v)) = ie (v), i = 1, . . . , ne ,
holds for any v C K e 1 . Further, we introduce a global interpolation operator h :
i1
i1
h
h
7 Ph such that for any v C h
C h
ih h (v) = ih (v),
i = 1, . . . , n,
(4.37)
h (v) |K e := e (v|K e ) ,
K e T h,
h0 (v)
:=
ih (v)
0
(4.38)
i1
7
h
, i 6 I0h
,
, i I0h
We suppose that T h are regular discretizations in the sense of the following three assumptions.
Assumption 4.5. We assume that there exists a positive constant C 11 such that
h > 0 K e T h : C10 (K e ) C11 .
53
Assumption 4.6. We assume that for each v [C 0 ()]1 there exist positive constants C12
C12 (v) and C13 C13 (v) such that
h > 0 K e T h : K e h h je v|K e kSe k C12 and je v|K e kSeB k C13 ,
(4.39)
Assumption 4.7. We assume that for each v [C 0 ()]1 there exists a positive constant C14
C14 (v) such that
h > 0 K e T h x Re (b
x) K e :
n
e
X
Bx e v| e (x)
=
x)
C14 ,
ie v|K e SeB Bxb bri (b
K
i=1
Let us note that Assumption 4.5 is replaced, in case of m = 2, by the minimum angle con
dition, see Z L AMAL
[219, p. 397], or by the maximum angle condition, see K R I Z EK AND N EIT [115, p. 67], and, in case of m = 3, by either the minimum or maximum angle conTAANM AKI
[115,
ditions between the edges as well as between the faces, see K R I Z EK AND N EITTAANM AKI
p. 83]. For the used kind of elements we will show that Assumptions 4.6 and 4.7 follow from the
angle conditions.
Lemma 4.4. Let Rm , L be a domain and let h h>0 be a class of its nonempty polyhedral subdomains such that h % . Let further v [C0 ()]1 . Then, under Assumption 4.6,
the following convergence holds
X
h
h
h
h
0, as h 0+ .
i (v|h ) i
v|h 0 v|h
h =
B,
iI h
B,h
Proof. The proof bases on Theorem 3.6. Let us write the square of the norm
2
X
h
h
(v|
)
h
i
i
iI h
0
B,h
2
Z
X
h
h
=
i (v|h ) i (x)
dx+
iI h
0
2
Z
X
h
h
dx.
+
)B
(x)
(v|
x
h
i
i
iI h
(4.40)
Due to (4.19), both hi and Bx ( hi ) have small supports. We take an arbitrary x . Since
h % and K e T h : he h, then due to (4.35) there exists h0 := miny kx yk/2 such
that
h h0 : x h \ h h
54
(4.41)
(v|
x
)
)S
(b
x
)
)
(x)
(v|
=
,
e
h
i
j
j
i
j
K
j=1,...,ne x
bK r
iI h
j=1
0
ne
X
X h
je (v|K e )SeB Bxb brj (b
i (v|h )Bx hi (x)
x)
=
iI h
j=1
0
x)
,
ne C13 max max
Bxb brj (b
j=1,...,ne x
bK r
therefore, the integrands themselves are also bounded. Having the boundeness and by (4.41)
having also the convergence of the integrands to zero in , as h 0 + , we now apply Theorem 3.6
to both the integrals in (4.40), which yields
2
2
Z
X
Z
X
h
h
h
h
i (v|h )Bx i (x)
i (v|h ) i (x)
dx 0 and
dx 0, as h 0+ .
iI h
iI h
1 ,0,
0, as h 0+ ,
where f h (x) := f (x) in \ h . Then for each > 0 the following convergence holds
u () in H0 (B; ), as h 0+ ,
Xh1 uh h
(4.43)
where Xh1 : L2 h 1 7 L2 () 1 is the linear extension operator defined by (4.33),
h
uh h H0 B; h is the solution to (Wh (h )), and u() H0 (B; ) is the solution
to (W ()).
55
(4.44)
h
From Lemma 4.3 we know that the set X0 B; ; h is a closed subspace of H0 (B; ). Therefore, the function uh is the Galerkin approximation to the solution u of (W ()) in the space
h
h
X0 B; ; h and we can employ Lemma 4.1, which for any vh X0 B; ; h yields
(
a vh , uh vh ah vh , uh vh
h
h
+
C9 ()
u v
+
u u
kuh vh kB,
B,
B,
(4.45)
)
f uh vh f h uh vh
.
+
kuh vh kB,
[115, Th. 4.16], originally
Now the idea of the proof is like in K R I Z EK AND N EITTAANM AKI
f [C0 ()]1
from D OKTOR [56]. Let > 0 be arbitrary. By Assumption 3.3, there exists u
such that
f kB,
ku () u
.
(4.46)
6C9 ()
In the estimate (4.45) we choose
f |h .
vh := Xh1 h0 u
We estimate the first term on the righthand side of (4.45). By the triangle inequality (3.1),
by Lemma 4.3, and by (4.46), we get
h
h
h
f
f
f
f
u
v
=
u
u
+
u
ku
u
k
+
u
B,
B,
B,
B,
(4.47)
f vh
.
+
u
6C9 ()
B,
The second term on the righthand side of (4.47) reads
Z
2
2
h
f |h
h =
f h0 u
f v
\h (x) kf
u (x)k2 + kB (f
u (x))k2 dx +
u
=
u
B,
B,
Z
2
f |h
f h0 u
u (x)k2 + kB (f
u (x))k2 dx +
u
,
= (1 h (x)) kf
h
B,
(4.48)
where k k denotes the Euclidean norm. Since h % , then (4.35) holds and we get
(1 h (x)) kf
u (x)k2 + kB(f
u (x))k2 0 a.e. in , as h 0+ .
h > 0 : (1 h (x)) kf
u (x)k2 + kB(f
u (x))k2 kf
u (x)k2 + kB(f
u (x))k2 a.e. in .
56
Using the triangle inequality (3.1), the second term on the righthand side of (4.48) is estimated
as follows:
f |h
h
f h u
f |h + h u
f |h h0 u
f |h
h =
u
f h0 u
u
B,
B,
h
h
h
f u
f |h
h +
u
f |h 0 u
f |h
h .
u
B,
B,
(4.49)
The definition (4.38), Assumption 4.4, and Assumption 4.5, respectively, yield
2
f h u
f |h
u
B,h
K e T h
kf
u e (f
u |K e )k2B,K e
X
K e T h
C10 (K e )he kf
u k1 ,2,K e
2
2 2
C11
h kf
u k21 ,2, .
By Lemma 4.4, the second term on the righthand side of (4.49) tends toward zero. Therefore, the
righthand side of (4.48) tends to zero, as h 0+ , i.e., there exists h1 > 0 such that
h h1 :
u vh
.
(4.50)
3C9 ()
B,
Further, we estimate the second term on the righthand side of (4.45). The nominator reads as
follows:
a vh , uh vh ah vh , uh vh =
Z
T
h
h
h
DD
B vh
dx
= B u v
sZ
2
h
2
u v h
max dhi,j (x) di,j (x) kB(vh (x))k dx,
B,
i,j
where we
used the CauchySchwarz inequality (3.3) in L2 () 2 . After dividing the latter by
uh vh
, we get
B,
sZ
2
a v h , u h v h a h v h , u h v h
h
2
(x)
d
(x)
d
max
kB(vh (x))k dx.
i,j
i,j
i,j
kuh vh kB,
(4.51)
Now, we use Theorem 3.6 to show that the integral on the righthand side of (4.51) vanishes, as
h 0+ . First, we prove the boundeness of the integrand. We take an arbitrary x . If x 6 h ,
then vh (x) = 0 and the integrand vanishes. If x K e for some K e T h , K e h h , then
by (4.3) and Assumption 4.6 the square of the integrand reads
max dhi,j (x) di,j (x)
B vh (x)
=
i,j
X
f |K e SeB Bxb brj (b
je u
= max dhi,j (x) di,j (x)
x)
i,j
j:G e (j)6I h
0
x)
.
2 max kdi,j kL () ne C13 (f
u ) max max
Bxb brj (b
i,j
j=1,...,ne x
bK r
57
where we used (4.3) and Assumption 4.7. Therefore, the integrand on the righthand side of (4.51)
is bounded by a constant independent of h, and due to the assumption (4.42) it converges to zero
almost everywhere in . Then, by Theorem 3.6
Z
2
2
max dhi,j di,j
B vh
dx 0, as h 0+ .
i,j
.
h h2 :
h
h
ku v kB,
3C9 ()
(4.52)
Finally, we estimate the third term on the righthand side of (4.45). The nominator reads as
follows:
Z
h
h
h
h
h
h
h
h
f f u v dx
f u v f u v =
h
h
(4.53)
v
f f h
u
1 ,0,
1 ,0,
h
,
f f h
u v h
1 ,0,
B,
where we used the CauchySchwarz inequality in L2 () . Dividing (4.53) by
uh vh
B,
and using the assumption (4.43), it follows that there exists h3 > 0 such that
f uh vh f h uh vh
h h3 :
.
(4.54)
h
h
ku v kB,
3C9 ()
1
At the end, combining (4.45), (4.50), (4.52), and (4.54), and recalling the notation (4.44), we have
proven the statement, i.e., for any > 0 there exists h 0 := min{h1 , h2 , h3 } such that
.
h h0 :
u () Xh1 uh h
B,
58
e := {1e , 2e , 3e } ,
where ie : C K e 7 R is such that for v C K e
ie (v) := v(xei ) ,
i = 1, 2, 3,
(4.55)
where xe1 := xe1,1 , xe1,2 , xe2 := xe2,1 , xe2,2 , xe3 := xe3,1 , xe3,2 are the corners of K e .
We concern the space H 1 (K e ) with the trace operator K e (v) := v|K e , v P e . From (4.55)
it is easy to see that the following couples of degrees of freedom ( 1e , 2e ), (2e , 3e ), and (3e , 1e )
for any v P e determine the traces v|he3 , v|he2 , and v|he1 along the edges he3 := (xe1 , xe2 ), he1 :=
(xe2 , xe3 ), and he2 := (xe3 , xe1 ), respectively, see also Fig. 4.4. Therefore, Assumption 4.1 is fulfilled
and we say that the linear Lagrange elements are H 1 (K e )conforming.
According to (4.27), (4.28), and Fig. 4.2, we specify the transformation from the reference
element to the element e by
e
R :=
r :=
xe1,1
xe1,2
(4.56)
cr := (1, 0),
x
2
cr := (0, 1).
x
3
(4.57)
Se := 1.
(4.58)
(4.59)
x) := x
c1 ,
b2r (b
x) := x
c2 ,
b3r (b
b := (c
where x
x1 , x
c2 ) K r
(4.60)
and where K r is the triangle in Fig. 4.2 the corners of which are given by (4.57).
Now we will state the element approximation property such that both Assumption
4.4 and
e
e
h
1
Assumption 4.5 will be fulfilled. Suppose that we have a discretization T := K , . . . , K nh
59
Definition 4.1. A family F := T h | h > 0 of triangulations is said to satisfy the minimum
angle condition if there exists a constant 0 such that for any T h F and any K e T h we have
0 < 0 K e <
,
2
(4.61)
h
h
where : C h 7 H 1 h is defined by (4.38), using the degrees of freedom (4.55).
The next two lemmas fulfill Assumptions 4.6 and 4.7, respectively.
Lemma 4.6. Let v C0 (). Then there exist positive constants C 12 C12 (v) and C13
C13 (v) such that for any discretization parameter h > 0 satisfying (4.32), for any subdomain
h satisfying Assumption 4.3, and for any discretization T h which satisfies the minimum
angle condition (4.61) the following holds
K e T h : K e h h je v| e kSe k C12 and je v| e
Se
C13 ,
K
grad
Proof. Let v C0 () be an arbitrary function, h > 0 be a discretization parameter satisfying (4.32), h be a polygonal subdomain satisfying Assumption 4.3, and T h be a discretization which satisfies the minimum angle condition (4.61).
Let K e h be arbitrary and let x K e . Since Se = 1, the first estimate is as follows:
e
j v| e kSe k = v xej max |v(z)| , j = 1, 2, 3,
K
z
Let now K e T h be such that K e h h and let x K e . Then, by Assumption 4.3, there
exist y and xh h such that
kx yk =
x xh + xh y
x xh
+
xh y
2he ,
where he is by definition the maximum side of K e . Since y , then v(y) = 0. Now, we use
Theorem 3.3
v(x) = v(y) + grad(v(z)) (x z) = grad(v(z)) (x z),
(4.62)
PSfrag replacements
60
xe3
he2
2e
xe1
he1
3e
e
1e
2e
1e = K e
xe2
he3 = he
Figure 4.4: A Lagrange triangle K e
(4.63)
To prove the second estimate, we exploit the structure of the matrix (R e )T . Using (3.9)
and (3.14), we get
e
xj+1,i xe1,i
e T
1
T
e
e
= max
= max (R )
(R )
= max (R )
i,j det (Re )
i,j
i,j
i,j
i,j
(4.64)
he
.
2 meas (K e )
From Fig. 4.4, it is clear that
he e
.
2
Using (4.63), (4.64), Fig. 4.4, and the minimum angle condition (4.61), the second estimate reads
as follows:
e
e T
2he
j v| e kSeB k = v xej
max
kgrad(v(z))k
)
(R
K
z
e
2 (1e + 2e )
max kgrad(v(z))k
z
e
4 e
1
max kgrad(v(z))k e2 = 4 max kgrad(v(z))k
z
z
tan (K e )
1
,
4 max kgrad(v(z))k
z
tan (0 )
meas (K e ) =
4 maxz kgrad(v(z))k
.
tan (0 )
Lemma 4.7. Let v C0 (). Then there exists a positive constant C 14 C14 (v) such that for
any discretization parameter h > 0, for any subdomain h , and for any discretization T h
which satisfies the minimum angle condition (4.61), the following holds
K e T h x Re (b
x) K e :
3
X
gradx e v| e (x)
=
x)
C14 ,
ie v|K e Segrad gradxb bir (b
K
i=1
61
gradxb b2r (b
x) = (1, 0),
gradxb b3r (b
x) = (0, 1),
b := (c
where x
x1 , x
c2 ) K r . Now, using the latter and the definition of R e , we exploit the structure
of the matrix Segrad (Re )T . It holds that
3
X
x)
ie v|K e Segrad gradxb bir (b
i=1
e
2
x3,2 xe1,2 (v (xe2 ) v (xe1 )) + xe2,2 xe1,2 (v (xe1 ) v (xe3 ))
,
where we also used (3.14). Since he is the maximum side, it follows that
e
xi,j xe1,j he ,
i = 1, 2, 3,
j = 1, 2.
i = 2, 3.
Finally, like at the end of the previous proof, from Fig. 4.4 it is clear that
|det (Re )| = 2 meas (K e ) = he e
and, due to the minimum angle condition (4.61) and Fig. 4.4, the estimate (4.65) is as follows:
3
X
2
e
e
r
b
x)
e e 2he max kgrad(v(z))k he
i v|K e Sgrad gradxb i (b
h
z
i=1
e + e
2 e
2 2 max kgrad(v(z))k 1 e 2 2 2 max kgrad(v(z))k e2 =
z
z
1
1
4 2 max kgrad(v(z))k
,
= 4 2 max kgrad(v(z))k
z
z
tan (K e )
tan (0 )
where K e is the minimum angle of the triangle K e . Hence,
1
.
tan (0 )
62
Kr
cbr6
1
x
c1
x
c3
1
0
x3
Re
cbr3
cbr1
cbr
cbr5
cbr2
Ke
ce3 ce5
ce6
ce2
ce1
c2
1 x
ce4
x2
x1
4.4.2 Linear N
ed
elec elements on tetrahedra
Here, we state a type of finite elements, which is frequently used for the approximation of the space
H(curl; ), where R3 , L. These elements can be used for solving the 3dimensional
linear magnetostatic problem, which was introduced in Section 3.4.3. The elements are defined
over tetrahedra and the degrees of freedom are calculated as integrals along the edges. The elements were first introduced by N E D E LEC [142] and, since then, they have become a standard.
The linear Nedelec element is a triple E := K e , Pe , e , where K e R3 is a tetrahedral
domain,
Pe := p(x) := ae x + be ae , be R3 , x := (x1 , x2 , x3 ) K e ,
e := {1e , 2e , 3e , 4e , 5e , 6e } ,
3
3
7 R is such that for v C K e
where ie : C K e
Z
e
v tei ds, i = 1, . . . , 6,
i (v) :=
cei
where cei stand for the oriented edges, see Fig. 4.5, and tei are the related unit tangential vectors.
Now, we concern the space H(curl; K e ) and the corresponding trace operator K e (v) :=
e
n v on K e , where v Pe and ne denotes the unit outer normal vector to K e . By N E LEC [142, Theorem 1], Assumption 4.1 is fulfilled, thus, the linear Nedelec finite elements are
DE
H(curl; K e )conforming.
The transformation Re in Fig. 4.5 is determined by
e
e
x1,1
x2,1 xe1,1 xe3,1 xe1,1 xe4,1 xe1,1
(4.66)
Re := xe2,2 xe1,2 xe3,2 xe1,2 xe4,2 xe1,2 , re := xe1,2 ,
e
e
e
e
e
e
e
x1,3
x2,3 x1,3 x3,3 x1,3 x4,3 x1,3
where xei := xei,1 , xei,2 , xei,3 , i=1,. . . ,4, are the corners of the tetrahedron K e , which correspond
to the following corners of K r
cr := (0, 0, 0),
x
1
cr := (1, 0, 0),
x
2
cr := (0, 1, 0),
x
3
cr := (0, 0, 1).
x
4
(4.67)
63
(4.68)
AND
T HOMAS [167,
(4.69)
b (b
b + re . The reference shape functions read as follows:
where v(x) := Se v
x) and x := Re x
0
1
1
0
b + 0 ,
b + 1 ,
br1 (b
x) := 1 x
br2 (b
x) := 0 x
1
0
1
0
1
0
0
0
b + 0 ,
b + 0 ,
br3 (b
x) := 1 x
br4 (b
x) := 0 x
(4.70)
0
1
1
0
1
0
0
0
b + 0 ,
b + 0 ,
br5 (b
x) := 0 x
br6 (b
x) := 1 x
0
0
0
0
b := (c
where x
x1 , x
c2 , x
c3 ) K r and K r is the reference tetrahedron, see Fig. 4.5, the corners of
which are given by (4.67).
Now, we will state the element approximation property such that both Assumption
4.4 and As
e
sumption 4.5 will be fulfilled. Suppose that we have a decomposition T h := K e1 , . . . , K nh .
The following definition and lemma are due to N E D E LEC [142, p. 327].
Definition 4.2. A family F := T h | h > 0 of decompositions into tetrahedra is said to be
regular if there exists a constant C 15 > 0 such that for any T h F and any K e T h we have
he
C15 ,
e
(4.71)
where j = 1, . . . , 6.
64
where he stands for the maximum edge size. Since S e := (Re )T , then, using (3.9) and (3.14), it
follows that
1
1
(he )2
fe
fe
.
(4.73)
kSe k =
R
=
R
|det (Re )|
6 meas (K e )
3 meas (K e )
Since e denotes the radius of the largest sphere inscribed in K e , from the regularity condition (4.71) it is obvious that
e 3
h
4
4
e 3
e
.
(4.74)
meas (K ) ( )
3
3
C15
Putting the latter into (4.73) and combining that with (4.72), the first estimate reads as follows:
3
e
j v| e kSe k max kv(z)k (C15 ) ,
K
z
4
hence,
(C15 )3
.
4
Similarly as in the proof of Lemma 4.6, let K e T h be such that K e h h . Then there
exists xh h and, by Assumption 4.3, there exists y such that
kx yk =
x xh + xh y
x xh
+
xh y
2he ,
where he is by definition the maximum side of K e . Since y , then v(y) = 0. Now we use
Theorem 3.3
vi (x) = vi (y) + grad(vi (z)) (x z) = grad(vi (z)) (x z)
for i = 1, 2, 3
and
kv(x)k max max kgrad(vi (z))k 2he
i{1,2,3} z
for i = 1, 2, 3,
1
kRe k
|det (Re )|
maxi,j xei+1,j xe1,j
6 meas (K e )
he
(C15 )3
,
6 meas (K e )
8 (he )2
(4.75)
65
where we used (4.74). Combining the latter with (4.75), the second estimate is as follows:
3
e
j v| e kSe k max max kgrad(vi (z))k 2 (he )2 (C15 )
curl
K
i{1,2,3} z
8 (he )2
1
max max kgrad(vi (z))k (C15 )3 ,
4 i{1,2,3} z
hence,
C13 :=
(4.76)
1
max max kgrad(vi (z))k (C15 )3 .
4 i{1,2,3} z
Lemma 4.10. Let v [C0 ()]3 . Then there exists a positive constant C 14 C14 (v) such that
for any discretization parameter h > 0, for any subdomain h , and for any discretization
T h which satisfy the regularity condition (4.71), the following holds
K e T h x Re (b
x) K e :
6
X
e
e
e
r
curlx v| e
=
x)
C14 .
i v|K e Scurl curlxb bi (b
K
(4.77)
i=1
Proof. The proof is similar to that of Lemma 4.7. Let v [C0 ()]1 be an arbitrary function,
h > 0 be a discretization parameter, h be a polygonal subdomain, T h be a discretization of
h which satisfies the regularity condition (4.71), and let K e T h be an element domain. The
rotations of the reference shape functions, see (4.70), are constant over K r
curlxb br1 (b
x) = (0, 2, 2),
curlxb br2 (b
x) = (2, 0, 2),
curlxb br3 (b
x) = (2, 2, 0),
curlxb br4 (b
x) = (0, 0, 2),
curlxb br5 (b
x) = (2, 0, 0),
curlxb br6 (b
x) = (0, 2, 0),
b := (c
where x
x1 , x
c2 , x
c3 ) K r . Let us simplify the rest of the proof by the following notation
ie := ie v|K e
for i = 1, 2, . . . , 6.
Now, we exploit the structure of the matrix S ecurl . It holds that
e
6
X
e
1
e
br (b
v|
=
R
curl
x
)
=
e
b
x
i
i
K
det(Re )
i=1
e
x2,1 xe1,1 (2e 3e + 5e ) +
2
xe2,2 xe1,2 (2e 3e + 5e ) +
=
6 meas(K e )
xe2,3 xe1,3 (2e 3e + 5e ) +
Let f2e , f3e , and f4e stand for the faces that are respectively opposite to the nodes x e2 , xe3 , and xe4 .
The following oriented closed curves
(xe1 , xe4 , xe3 , xe1 ) ,
and
66
see also Fig. 4.5, are respectively the positively oriented boundaries of the faces f 2e , f3e , and f4e
with the outer unit normal vectors n e2 , ne3 , and ne4 . Now, using Theorem 3.5, we arrive at
e
R
v|
curl
(x)
n2 (x) dS
e
e
x
K
f2
R
1
hence
3
curlx e v| e (x)
3 maxx kcurlx (v(x))k (C15 ) ,
K
8
C14 :=
Chapter 5
{nk }
k=1 {n }n=1 and U such that nk in V , as k .
The next fundamental theorem of functional analysis examine the existence of a solution to
the problem (P ).
67
68
Theorem 5.1. Let U be a compact subset of the normed linear space V and let J : U 7 R be a
continuous functional. Then there exists a solution to the problem (P ).
Proof. See H ASLINGER
AND
[85, p. 67].
N EITTAANM AKI
(5.1)
We further employ the box constraints, i.e., there exist l , u R such that
x : l (x) u .
(5.2)
(5.3)
(5.4)
Proof. Let {n }
n=1 U be an arbitrary sequence of shapes. By (5.2) the sequence is uniformly bounded and by (5.1) it is equicontinuous. Then by Theorem 3.2 there exist a subsequence
{nk }
k=1 {n }n=1 and C() such that
nk in , as k .
It is easy to see that satisfies both (5.1) and (5.2), which completes the proof.
In Chapter 7, we will deal with an application where we will be at the end looking for smooth
shapes, e.g., Bezier curves or patches, cf. FARIN [59], rather than for continuous ones. To this
such that graph() 0 () 1 (), meas (0 ()) > 0, and meas (1 ()) > 0,
(5.6)
69
PSfrag replacements
0 ()
1 ()
x2
x1
From the positive definiteness of D0 and D1 , the relation (3.39) follows. The bilinear form
(3.40) now reads
Z
Z
a (v, u) :=
B(v) (D0 B(u)) dx +
B(v) (D1 B(u)) dx, u, v H(B; ),
0 ()
1 ()
(5.8)
where both the operator B and the space H 0 (B; ) were described in Section 3.3.6.
Concerning the linear functional (3.41), we distinguish several righthand sides f , e.g., several
current excitations in case of magnetostatics. The linear functional (3.41) reads as follows:
Z
v
f (v) :=
f v v dx, v H(B; ), for v = 1, 2, . . . , nv ,
(5.9)
where nv N is a number of the considered righthand sides f v L2 () 1 , 1 N, such that
they fulfill
Z
p Ker(B; ) :
f v p dx = 0
for each v = 1, 2, . . . , nv ,
70
Assumption 5.2. We assume that for each v = 1, . . . , n v the righthand sides f v are independent
from U.
Now for each v = 1, 2, . . . , nv the state problem (W ) is rewritten as follows:
Find uv () H0, (B; ):
a (v, uv ()) = f v (v)
v H0, (B; )
(W v ())
v = 1, 2, . . . , nv .
Proof. Taking an arbitrary shape U and any v = 1, 2, . . . , n v , the proof is the same as the
one of Theorem 3.16, where the symbols a, f , D, and f are replaced by a , f v , D , and f v ,
respectively.
(5.10)
kun uk2B,
(5.11)
(5.12)
Now, we write down the righthand side of (5.11) and using (3.45) and the CauchySchwarz
71
inequality in L2 () 2 we get
Z
|a (u, un u) an (u, un u)| =
+
0 ()
1 ()
0 (n )
1 (n )
B(u) (D1 B(un u)) dx
Z
0 () 0 (n ) B(u) (D0 B(un u)) dx +
Z
+
1 () 1 (n ) B(u) (D1 B(un u)) dx
d
0 () 0 (n ) B(u)
0, , +
2
+
1 () 1 (n ) B(u)
0, , kB(un u)k0,2 , .
2
(5.13)
A1 (n) :=
1 () 1 (n ) B(u)
0,
2 ,
(5.14)
(5.15)
and, since B(u) [L2 ()]2 , the functions on the lefthand side of (5.15) are in L1 () and each
bounded by kB(u)k2 L1 () from above. Now, using Theorem 3.6, we arrive at
A0 (n) 0 and A1 (n) 0, as n .
(5.16)
Combining (5.10), (5.11), (5.14), and (5.16), we have proven the statement
uv (n ) uv () in H(B; ), as n .
72
The continuous optimization problem then, in accordance with Section 5.1, reads as follows:
)
Find U:
.
(P )
J ( ) J () U
Theorem 5.2. There exists U that is a solution to (P ).
Proof. By Lemma 5.1,
subset of the normed linear space C(). Using the con U is acompact
n
tinuity of I on U L2 () 2 v and Lemma 5.3, the continuity of J on U follows. Now,
Theorem 5.1 completes the proof.
Moreover, we use (5.5) to define the cost functional Je : 7 R
Je(p) := J (F (p)),
p .
Then, by the compactness of , by the continuity of F on , and by the same arguments as in the
proof of Theorem 5.2, there exists a solution p to the optimization problem
)
Find p :
.
(Pe )
Je(p ) Je(p) p
In this section, we will show a convergence of solutions of optimization problems whose state
problems are regularized, as described in Section 3.4.1, to a continuous solution .
Let > 0 be a regularization parameter. Due to (3.47) and (5.8), we introduce the regularized
bilinear form controlled by the shape U
Z
a, (v, u) := a (v, u) +
v u dx, u, v H(B; ).
(Wv ())
Lemma 5.4. Let > 0, U. Then for each v = 1, 2, . . . , n v there exists a unique solution
uv () H0 (B; ) to the problem (Wv ()). Moreover, there exists a positive constant C 8 ()
such that
kuv ()kB, C8 ()kf v k1 ,0, for each v = 1, 2, . . . , nv .
Proof. Taking > 0, an arbitrary shape U, and any v = 1, 2, . . . , n v , the proof is the same as
the one of Theorem 3.17, while the symbols a , f , D, and f are replaced by a, , f v , D , and f v ,
respectively.
Lemma 5.5. Let > 0. Then for each v = 1, 2, . . . , n v the mapping uv : U 7 H0 (B; ) is
continuous on U.
Proof. Taking any > 0, the proof is the same as the one of Lemma 5.3, where all the proper
symbols are subscribed with .
73
(5.17)
where uv () H0 (B; ) are the solutions to (Wv ()) and uv () H0, (B; ) is the solution
to (W v ()).
Proof. Taking an arbitrary shape U and any v = 1, 2, . . . , n v , the proof is the same as the
one of Theorem 3.18, where we replace the symbols related to the problems (W ) and (W ) by the
corresponding symbols related to (W v ()) and (Wv ()), respectively.
Now, we return to the shape optimization problem. We introduce the regularized cost functional by
J () := I , B u1 () , B u2 () , . . . , B(un v ()) , U.
The regularized shape optimization problem then reads as follows:
)
Find U:
.
J ( ) J () U
(P )
(5.19)
Using Lemma 5.6 and the continuity of I, the righthand side of (5.19) converges as follows:
Jnk () J (), as k .
Using (5.18), Lemma 5.5, Lemma 5.6, and the continuity of I, the lefthand side of (5.19) also
converges
Jnk n J ( ), as k .
k
J ( ) J ().
74
p .
(Pe )
In the same fashion as in the case of (P ), we can derive the existence theory as well as the
approximation property of the problem ( Pe ).
In this section, we introduce a setting of the shape optimization problem (P ) discretized by the
finite element method. We will prove a convergence of the approximate solutions to the true one.
Let > 0 be a regularization parameter and let h > 0 be a discretization parameter, see (4.31).
With any h > 0 we associate a nonempty polyhedral computational subdomain h such
that (4.35) is satisfied.
xh
for i = 1, . . . , nxh .
(5.21)
h
equipped with the uniform
h convergence (5.4). Obviously for each h > 0: U U and, by the
h
1
definition of P T , U is finite dimensional.
75
We introduce an interpolation operator h : U 7 P 1 Th such that for each xhh ,j , a corner
of ih , it holds that
h
i
h
h
h
U : () xh ,j = xh ,j ,
i
i = 1, . . . , nh , j = 1, . . . , m.
(5.22)
AND
G LOWINSKI [19].
Further, let h > 0 be given. Since for any U: (5.1) implies (5.20), and (5.2) implies
(5.21), then also U implies h () U h . Moreover, as F : 7 U, then for any p it
follows that h (F (p)) U h . Therefore, we use also for the discretized setting.
Finally, like in the continuous case we assume that a discretized shape h controls the decomposition of h into the subdomains h0 (h ) and h1 (h ) as follows:
h = h0 (h ) h1 (h ), h0 (h ) h1 (h ) =
PSfrag replacements
such that graph(h ) h0 (h ) h1 (h ), meas h0 (h ) > 0, and meas h1 (h ) > 0,
(5.23)
x3
PSfrag replacements
h0 (h )
h1 (h )
h0 (h )
h1 (h )
x1
x2
7h
1h
2h
8h
h h
10
11
9h
3h
4h
5h
6h
h
12
x2
x1
Figure 5.2: Decomposition of h
Ki (h ) T h (h ) : Ki (h ) h0 (h ) or Ki (h ) h1 (h ).
(5.24)
76
o
n
By E h := e1 , . . . , enh we denote the corresponding set of finite elements. For any h > 0 and
h
h
h
v,h
:
Find u H0 B;
v,h h
h , v = 1, . . . , nv , (W ( ))
ah,h vh , uv,h
h
= f v,h vh
vh H0 B; h
h
where thefinite element space H0 B; h is defined by (4.17), where further for each v h , wh
H B; h we define
Z
Z
h
h
h
h
h
h
a,h v , w :=
B v D h B w
dx +
vh wh dx,
(5.25)
h
D0
h
Dh (x) := D1
, x h0 h
, x h1 h ,
, x \ h
and where for each vh H B; h we set
Z
v,h
h
f
v :=
f v,h vh dx,
h
(5.26)
v = 1, . . . , nv ,
in which, due to (4.18), f v,h L2 h 1 are elementwise constant and such that
v,h
0, as h 0+ , v = 1, . . . , nv ,
f f v
1 ,0,
(5.27)
Assumption 5.4. We assume that for each v = 1, . . . , n v the righthand side f v,h is independent
of the shape h U h .
The following lemma assures that for any > 0, h > 0, and v = 1, . . . , n v fixed the mapping
h
: U h 7 H0 B; h is well defined.
uv,h
Lemma 5.9. For each > 0, h > 0, h U h , and v = 1, . . . , nv there exists a unique solution
h
uv,h
h H0 B; h to the problem (Wv,h (h )).
Proof. Since h is a polyhedron, then h L and the statement follows by the same arguments
as in the proof of Theorem 3.17.
77
Lemma 5.10. Let > 0, h > 0. Then for each v = 1, 2, . . . , n v the mapping uv,h
: U h 7
h
H0 B; h is continuous on U h .
Proof. We take an arbitrary > 0, h > 0, and v = 1, . . . , n v . Note that we cannot use the
same technique as in the proof of Lemma 5.3, since the settings (Wv,h (h )) differ from h U h .
Therefore, the estimate (5.11) cannot be established. We will rather exploit the algebraic structure
of the mapping uv,h
.
In the similar manner as in (4.24), the solution to (Wv,h (h )) reads as follows:
uv,h
n
X
i=1
h
h
uv,n
hi xh h ,
,i x
(5.28)
where xh h denotes a vector of global coordinates of all element domains corners, which are
by Assumption 5.3 continuously dependent on the shape h U h , where further hi xh denotes
the global shape
functions, and where we use the same notation for both the functions u v,n
h
v,n
h
h
and u x
v,n
h
h
Rn ,
,
.
.
.
,
u
xh h
:= uv,n
x
u v,n
xh h
h u v,n
,n
,1
which is the solution to the linear system (4.9). In this case, (4.9) reads as follows:
An xh u v,n
xh = f v,n xh .
(5.29)
Now, let us take a look into the assembling of the matrix and the righthand side vector
in (5.29). Due to (4.21), the element contributions to them are
An
v,n
i,j
n
X
n
X X
v,e
(5.30)
( ek (xe )) ,
eEih k=1
for i, j = 1, . . . , n, where Eih denotes the set of elements neighbouring with e i , see (4.20), and
xe is the vector of coordinates of the element domain corners, see (4.25). Using the map from
the reference element r, the element contributions to the bilinear form and linear functional, see
also (4.29) and (4.30), respectively are
ae,xe ( ek (xe ) , el (xe )) =
Z
=
rk Dexe SeB (xe ) Bxb brl
|det(Re (xe ))| db
x+
SeB (xe ) Bxb c
Kr
f v,e ( ek (xe )) =
Kr
Z
Kr
x,
Se (xe ) c
rk Se (xe ) brl |det(Re (xe ))| db
x,
f e Se (xe ) c
rk |det(Re (xe ))| db
(5.31)
78
The expressions (5.28)(5.31) specify the function uv,h
h . Now, we will prove its continu
ity. Let h U h be an arbitrary discretized admissible shape and let hp p=1 U h be such a
sequence that
hp h in U h , as p .
Let us denote for each element e E h , where E h stands for the set of finite elements,
xep := xe hp and xe := xe h .
xep xe in Rm(m+1) , as p .
Again by Assumption 5.3, for each e E h and each p N the element domains K e xep as
well as K e (xe ) still form a triangle (in case of m = 2) ora tetrahedron (in case of m = 3), and it
follows from the definition (4.28) that the matrices Re xep , Re (xe ) are nonsingular, and therefore
det Re xep > 0, |det(Re (xe ))| > 0.
(5.32)
(5.33)
b are
Now, the only symbols
in the integrals
(5.31) that depend on the integration variable x
r
r
r
r
c
b
c
b
k (b
x), l (b
x), Bxb k (b
x) , and Bxb l (b
x) . However, they are each independent from the vector
xe . We can expand the matrix multiplications in the integrands, which leads to the following finite
linear combinations of integrals
ae,xe ( ek (xe ) , el (xe ))
f v,e ( ek (xe )) =
N
X
i=1
M
X
j=1
ce,i (xe )
e
dv,e
j (x )
Kr
Kr
Fi (b
x) db
x,
Gj (b
x) db
x.
and
ae,xep ek xep , el xep ae,xe ( ek (xe ) , el (xe )) in R, as p
f v,e ek xep
f v,e ( ek (xe )) in R, as p
for k, l = 1, . . . , ne . By Assumption 5.3 the topology of the mesh T h h does not change with
any h U h , hence, the sets Eih and Ejh in (5.30) remain unchanged. It follows that
An xh hp
An xh h
, as p ,
i,j
i,j
79
in R, as p ,
f v,n xh h
f v,n xh hp
i
u v,n
u
x
xh h
p
(5.34)
Using the map from the reference element r, the global shape function reads as follows:
X X
X X
hi xh =
Se (xe ) brj ,
ej (xe ) =
eEih j:G e (j)=i
(5.35)
Combining (5.28), (5.34), and (5.35), we have completed the proof, i.e.,
v,h
h
h in U h , as p .
uv,h
uv () in H0 (B; ), as n ,
where uv,hn hn is the solution to (Wv,hn (hn )) and uv () is the solution to (Wv ()).
Proof. It is enough to prove that the assumption (4.42) is fulfilled and the rest is then fairly the
same as the proof of Theorem 4.2, while (4.43) is replaced by the assumption (5.27).
Given an arbitrary x \ (0 () 1 ()), it follows from (5.6) that either x 0 ()
or x 1 (). Thus, by (5.7) either
D (x) = D0 or D (x) = D1 ,
respectively. Having hn % , hn , as n , and due
n0 (x) N
to (5.23), there exists
such that for each n N, n n0 (x) either x h0 n hn or x h1 n hn , respectively.
Therefore, either
Dhn (x) = D0 or Dhn (x) = D1 ,
respectively. Thus, we have verified the assumption (4.42), i.e., for any i, j = 1, . . . , 2 :
hn
dhn ,i,j (x) d,i,j (x) 0 a.e. in , as n ,
where Dhn (x) := dhnhn ,i,j (x)
i,j
80
,
.
.
.
,
B
X
h
,
h U h , (5.36)
h
where Xh1 : H0 B; h 7 H0 (B; ) is due to (4.33) and Lemma 4.3. The relevant setting of
the shape optimization problem reads as follows:
Find h U h :
.
(Ph )
h
h
h
h
h
h
U
J J
Theorem 5.5. Let > 0 and h > 0. Then there exists h U h that is a solution to (Ph ).
Proof. Taking any > 0 and h > 0, the proof is fairly the same as the one of Theorem 5.2, where
we use the symbol Jh instead of J , and Lemma 5.10 instead of Lemma 5.3.
Theorem 5.6. Let > 0 be a fixed regularization parameter. Let {h n }
n=1 R be a sequence
n=1
hn k
in U, as k ,
hn
solution to (Phn ).
U hn , a o
Proof. By Theorem 5.5, for each > 0 and hn > 0 there
n exists
o n
hn k
By Lemma 5.1, there exist a subsequence of shapes
h n
and a shape
k=1
n=1
U such that
hn
k in U, as k .
(5.37)
h
Let U be an arbitrary shape. By Lemma 5.8, there exists a sequence nk k=1 , hnk U hnk
such that
hnk in U, as k .
(5.38)
hn k
(5.39)
Using (5.37), (5.38), Lemma 5.11, and the continuity of I, both the left and righthand side
of (5.39) converge
h
hn
n
J k k J ( ) and J hnk hnk J () , as k .
J ( ) J ().
81
Find ph :
.
Jeh ph Jeh (p) p
(Peh )
Since is a compact set and h F : 7 U h is a continuous mapping, we can state and prove the
existence theorem for (Peh ) similarly to Theorem 5.5. We can also state the convergence theorem,
the proof of which is even simpler than the one of Theorem 5.6, as the set of admissible design
parameters is not changed by discretization.
Remark 5.1. In cases of complex geometries, as those in Chapter 7, Assumption 5.3 is a serious
bottleneck of this discretization approach. For small discretization parameters and large changes
in the design we cannot guarantee that the perturbed elements still satisfy some regularity condition. They might be even flipped. In this case, we have to remesh the geometry and solve the
optimization problem again, but now on a grid of different topology. Then certainly the cost functional is not continuous any more and the just introduced convergence theory cannot be applied.
Nevertheless, in literature this approach is still the most frequently used one as far as a finite element discretization is concerned. In practice, after we get an optimized shape we should compare
the value of a very fine discretized cost functional for the optimized design with that value for the
initial one. If we can see a progress then the optimization surely did a good job. Some solutions to
this obstacle are discussed in Conclusion.
82
Chapter 6
AND
AND O LHOFF [82], B ROCKMAN [34], G RIEWANK [70], M AKINEN
[131], N EITTAANM AKI
S ALMENJOKI [145].
Find ph :
,
(Peh )
Jeh ph Jeh (p) p
where Jeh : 7 R denotes the discretized and regularized cost functional and R n ,
n N, is the set of admissible design parameters.
83
84
hi (p) := h F (p) [F (p)] xh,i
for i = 1, . . . , nh ,
(6.1)
where F : 7 U is due to (5.5) and h : U 7 P 1 Th is defined by (5.22).
coordinates onto the corresponding grid nodal coordinates. Both K h and b h arise from the finite
element discretization of the auxiliary linear elasticity problem. For finite elements in elasticity, we
refer to Z IENKIEWICZ [217]. The matrix Mh might also involve some symmetry assumptions on
the geometry, as we will state later in Chapter 7. Solving the equation (6.2) takes approximately
the same computational effort as solving one state problem. Nevertheless, the mapping is very
general, which fits to our intent in developing a robust and efficient numerical method for shape
optimization.
85
where both the system matrix and the right-hand side vectors are assembled by means of Algorithm 1
ne
X X
An xh
:=
ae,xe ( ek (xe ) , el (xe )) ,
i,j
f v,n xh
:=
ne
X X
(6.4)
f v,e ( ek (xe ))
eEih k=1
for i, j = 1, . . . , n, where Eih denotes the set of elements neighbouring with e i , see (4.20), and
xe Rm(m+1) is the vector of coordinates of the element domain corners, see (4.25), which is
also included in xh by means of the mapping H e , see (4.26). Components of the solution to (6.3)
are denoted by
v,n
n
v,n
v = 1, . . . , nv .
u v,n
:= u,1 , . . . , u,n R ,
Using the map from the reference element r, the element contributions to the bilinear form and
linear functional, respectively, see also (4.29) and (4.30), are
Z
Kr
v,e
Se (xe ) c
rk (b
x) Se (xe ) brl (b
x) |det(Re (xe ))| db
x,
( ek (xe ))
:=
Kr
x) |det(Re (xe ))| db
x,
f v,e Se (xe ) c
rk (b
(6.5)
x ;x =
n
X
i=1
h
uv,n
hi xh ; x ,
,i x
v = 1, . . . , nv , xh Rmnxh , x h ,
(6.6)
is the solution to the state problem (W v,h (h )), where hi xh ; x denote the global shape functions. Moreover, for e E h we introduce the element solution vector by
v,n,e
v,n,e
u v,n,e
:=
u
,
.
.
.
,
u
Rne , where uv,n,e
:= uv,n
,ne
,1
,i
,G e(i) for i = 1, . . . , ne .
As we look for B uv,h
rather than for uv,h
1
xh
R2 nh (6.7)
xh := B xh , u v,n
xh , . . . , B
xh := B v,n,e
B v,n
86
and where xh Rmnxh contains all the grid nodes, xe Rmne contains the grid nodes related to
the element e E h , where further
h
ei
v,n,ei
h
v,h
h
B xh , u v,n
x
|
:=
B
x
,
u
x
:=
B
u
x
;
x
| Ki =
Ki
x
=
=
ne
X
j=1
ne
X
j=1
ei ei
ei
(x
;
x)
=
uv,n
(x
)
B
x
j
,G ei(j)
ei ei
ei
i
br (b
x
)
uv,n,e
(x
)
S
(x
)
B
b
x
j
,j
B
(6.8)
for i = 1, . . . , nh ,
where x := Rei b
x+rei Ki , and where ei E h is the element related
to Ki . Recall
that since we
v,h h
employ the lowest, i.e., firstorder finite elements, the function B x u (x ; x) is elementwise
constant.
xh
where h := h (p) is a vector of shape control coordinates, where for v = 1, . . . , n v B v,n
n
n
xh [R 2 h ] v 7 R is the revised
R2 nh is given by (6.7) and (6.8), and where I h : Rnh Rmn
cost functional which is for p and for x h := xh h (p) defined by
:=
xh , . . . , B n v ,n xh
I h h(p), xh , B 1,n
:= I h (F (p)) , Xh1 B u1,h
, . . . , Xh1 B un v ,h
,
7
in which h : U 7 U h is defined by (5.22), F : 7 U is due to (5.5), Xh1 : H0 B; h
v,h
H0 (B; ) is due to (4.33) and Lemma 4.3, and where u is the solution (6.6).
The complete evaluation of the cost functional proceeds as follows:
v,n
v,n
K h 4xh =b h(h )
B (xh ,u v,n
)
h F
An
FEM
u =f
p h xh An , f v,n
u v,n
nv ,n
B (xh ,u v,n
I h(h ,xh ,B 1,n
)
,...,B
)
Jeh (p).
B v,n
(6.9)
The cost functional Jeh is compounded of the following submappings:
87
FEM which assembles the system matrix A n and the righthand side vectors f 1,n , . . . ,
f nv ,n by means of the finite element method, as described in Algorithm 1,
v,n
An u v,n
that solve the nv linear systems of algebraic equations,
=f
B v,n
which is a blockcolumn vector whose individual vectors represent the elementwise
Assumption 6.2. We assume that for each h > 0 such that h h the following hold:
x : [F ()] (x) C 2 () ,
K h R(mnxh )(mnxh ) is nonsingular,
mn h
x
b h h C 2 (Rnh )
,
mm
e E h : Re (xe ) C 2 Rm(m+1)
,
2 2
e E h : SeB (xe ) C 2 Rm(m+1)
,
1 1
e E h : Se (xe ) C 2 Rm(m+1)
, and
nv ,n
C 2 (Rnh (R2 nh )nv ) .
I h h , B 1,n
, . . . , B
Lemma 6.1. Under Assumptions 5.1, 5.3, 5.4, and 6.2, for any h > 0 such that h h it holds
that
Jeh C 2 ().
Proof. We will stepbystep use Assumption 6.2 and apply Lemma 3.4 to prove the smoothness
of the individual submappings.
Let h > 0 be given such that h h. By Assumption 6.2, for each i = 1, . . . , nxh we have
[F ()] xh,i C 2 (), therefore,
h C 2 ()
n
(6.10)
6 0 and
Again by Assumption 6.2 and by Lemma 3.3, det K h =
h
Kh
i1
:=
1
gh ,
K
h
det K
gh denotes the adjoint matrix, which was defined by (3.13). Then, due to the latter
holds, where K
and by Assumption 6.2, we get
h
i1
mn h
x
b h h C 2 (Rnh )
xh h = K h
.
(6.11)
88
Let us further prove the smoothness of the solutions u v,n
xh to the discretized multistate
problem (6.3). Let > 0 and v = 1, . . . , nv be arbitrary. By Assumption 6.2, for each e E h we
get
h
h
imm
h
i2 2
i1 1
Re C 2 Rm(m+1)
, SeB C 2 Rm(m+1)
, Se C 2 Rm(m+1)
.
Then, also due to the definition (3.12), det(Re ) C 2 Rm(m+1) . From Assumption 5.3 it follows
that the element K e must not flip, so the determinant does not change its sign, i.e.,
|det(Re )| C 2 Rm(m+1) .
Now let us look at the element contributions (6.5). Having Assumptions 5.1 and 5.4, only the refb. After some matrixvector
erence shape functions c
rk and brl depend on the integration variable x
multiplications we get the following structure of the element bilinear form and linear functional,
respectively,
Z
N
X
e
e
e
e
e e
e
c,i (x )
a,xe ( k (x ) , l (x )) =
Fi (b
x) db
x,
f
v,e
( ek (xe ))
i=1
M
X
i=1
Kr
e
dv,e
i (x )
Kr
Gi (b
x) db
x,
m(m+1) , since they arise as sumations and multiplications
e
2
where both ce,i (xe ) , dv,e
i (x ) C R
x) and
of the entries of SeB , De , Se , and f v,e multiplied then by |det(Re )|, and where both Fi (b
Gi (b
x) are common for all e E h . Thus, it follows that both the element bilinear form and element
linear functional are smooth, i.e.,
h
ine ne
h
ine
ae,xe ( ek (xe ) , el (xe )) C 2 Rm(m+1)
, f v,e ( ek (xe )) C 2 Rm(m+1)
. (6.12)
Now we employ Assumption 5.3, which assures that the topology of the discretization T
Hence, neither Eih nor Ejh in (6.3) depends on xh . From (6.12) it follows that
n
nn
and f v,n xh C 2 (Rmnxh ) ,
An xh C 2 (Rmnxh )
a consequence of which is
is fixed.
det An xh
C 2 (Rmnxh ) .
Lemma 5.9 provides us the existence of the solution u v,n
xh to (6.3). Hence, there exist the
1
. Then, by Lemma 3.3, det An xh 6= 0 and
inverse matrix An xh
h i1
1
fn xh C 2 (Rmnxh ) nn
A
An xh
:=
n h
det(A (x ))
fn xh denotes the adjoint matrix, which was defined by (3.13). From (3.15) we
holds, where A
get
h i1
n
f v,n xh C 2 (Rmnxh )
for v = 1, . . . , nv .
(6.13)
u v,n xh = An xh
89
Convention 6.1. Just for the purposes of this chapter let us skip in our notation the discretization
parameter h, the superscript n in (6.3), and the regularization parameter that will be each
fixed for the moment. If not stated otherwise, all the symbols in the sequel will be considered
as discretized ones, even if they were previously reserved for the continuous setting. Hence, we
consider the following discrete optimization problem with inequality constraints
Je(p)
Find p := arg min
n
pR
,
(P)
subject to (p) 0
where Je : Rn 7 R and (p) := (1 (p), . . . , n (p)) : Rn 7 Rn . The problem (P) is
governed by the following multistate problem
A(x) u v (x) = f v (x)
(P v (x))
for v = 1, . . . , nv .
90
subject to L(p) 0
where Q : Rn 7 R denotes a quadratic function and L : R n 7 Rn denotes a linear vector
function.
Basically, there are two approaches to the approximation of the problem (P) by a subproblem
(QP). In both of them the input p0 Rn denotes initial design parameters. The first approach
is called a line search approach where we look for an optimal Newton direction s QP being the
solution to the following subproblem
i o
nh
e, p0 (s)
Find sQP := arg min
J
Q
sRn
,
(QP 1 (p0 ))
subject to [L (, p0 )] (s) 0
where s := p p0 stands for a directional vector from the initial design p 0 to the current one
p, Q Je, p0 stands for quadratic Taylors expansion, see Theorem 3.3, of the function Je at the
point p0 while skipping the constant term Je(p0 )
h
i
1
(6.15)
Q Je, p0 (s) := grad Je(p0 ) s + s Hess Je(p0 ) s , s Rn ,
2
in which Hess Je(p0 ) Rn n denotes the Hessian matrix whose entries are as follows:
h
i
Hess Je(p0 )
i,j
:=
2 Je(p0 )
,
pi pj
i, j = 1, . . . , n ,
and where L Je, p0 denotes linear Taylors expansion, see Theorem 3.3, of the vector function
at the point p0
[L(, p0 )] (s) := (p0 ) + Grad((p0 )) s,
s R n ,
(6.16)
subject to p0 + sQP 0
and
pQP := p0 + QP sQP
91
is the solution.
The second approach is called a trust region method. It supposes that the quadratic subproblem
approximates the original problem well, but just in a given neighbourhood of p 0 . Hence, given an
initial point p0 and a trust region diameter d > 0, we solve the following quadratic subproblem
nh
i
o
e, p0 (p p0 )
Find pQP := arg min
J
Q
pRn
,
(QP 2 (p0 , d))
subject to [L (, p0 )] (p p0 ) 0
kp p0 k
2
where Q and L are respectively given by (6.15) and (6.16).
pQP := p0 + QP sQP
pk+1 := pQP
k := k + 1
end while
p := pk
pk+1 := pQP
Update dk
dk+1
k := k + 1
end while
p := pk
Let us note that there are many aspects to deal with, as to find a proper convergence criterion
or to modify the quadratic subproblem when it does not admit a solution, which is the case if
the Hessian matrix or its certain invariant is not positive definite. Here, we want to mention the
92
BFGS modification, see F LETCHER [62], named after its authors Broyden, Fletcher, Goldfarb, and
Shanno. It is originally based on the idea of DAVIDON [52, 53]. The method was a revolutionary
improvement of the SQP algorithm. At each iteration, it requires to evaluate only the gradient of
the objective and constraint functions, while the Hessian matrix is iteratively built up by measuring
changes in the gradients. For k 0, k N, the BFGS formula is the following, cf. N OCEDAL
AND W RIGHT [148, p. 25],
Hk+1 := Hk
(6.17)
where Hk and Hk+1 are two successive approximations of the Hessian matrices Hess Je(pk )
and Hess Je(pk+1 ) , respectively, and where
sk := pk+1 pk
and
yk := grad Je(pk+1 ) grad Je(pk ) .
The SQP method with the BFGS update is classified as a quasiNewton method.
,
xi
2
where > 0. The approximation error decreases with 2 until a computer roundoff error becomes
significant. Therefore, we have to choose such that neither the approximation nor roundoff error
is large. Another possibility is using a numerical differentiation formula of a higher mth order,
m N, for which the approximation error decreases with m+1 . However, evaluating the gradient
approximation needs 2mn evaluations of f , which is in the case of shape optimization very time
consuming. Hence, we have to balance between the time issue and the precision. The advantages
of the method are robustness and easy implementation.
The automatic differentiation, see G RIEWANK [70], differentiate the function f symbolically.
The input of the method is a routine that evaluates f (x) and the output is again a routine which
now evaluates grad(f (x)). An implementation of the method is very difficult, since it involves a
syntax recognizing, and it relies on the programming language that f is coded in. Nowadays, there
are free software packages available. The method is robust and precise up to the computer round
off error, but it is too much both time and memory consuming in case of shape optimization, since
the routine for solving the linear system which arises from the finite element discretization is also
differentiated symbolically.
93
Here, we will focus on the semianalytical methods, cf. H AUG , C HOI , AND KOMKOV [86],
that bases on the algebraic approach to sensitivity analysis, cf. H ASLINGER AND N EITTAANM A KI [85]. The methods respect the structure of the shape optimization problem, in which solution
to a linear system of algebraic equations is involved. The cost functional is a compound map and
its gradient is then a product of the gradients of the individual submappings. Most difficult to
evaluate is differentiation of the solution to the linear system with respect to nodal coordinates
of the discretization grid. This is performed by solution to other linear systems with the original
but transposed system matrix and with new righthand side vectors. The method is precise up
to the numerical error of the linear system solver. The computational time roughly corresponds
to the computation of the function f . The method is not robust, as it covers just the shape optimization problems, nevertheless, some other classes of optimization problems, e.g., the topology
optimization, have a similar structure. Thus, an extension of the method is straightforward. The
semianalytical methods might also be combined with both the numerical and automatic differentiation.
ne
n
X
X
Bjv,e (xe , u v,e ) uv,e
xk,l () o (p)
p (x)
+
,
v,e
up
xk,l
o pi
p=1
o=1
where
(p) := (1 (p), . . . , n (p)) Rn denotes control coordinates of the shape and for
o = 1, . . . , n , i = 1, . . . , n it holds that
[F (p)](x,o )
o (p)
=
,
pi
pi
94
x() := [x1 (), . . . , xnx ()] Rmnx denotes a block column vector consisting of all the
grid nodes, where for o = 1, . . . , n
[M]1,o
4x()
4x()
b()
x()
..
:=
+
=
,
, where K (x0 )
.
o
o
o
o
[M]nx ,o
xe () := xe1 (), . . . , xem+1 () Rm(m+1) denotes a block column vector consisting of
the corners of the element domain K e , e E,
xel () := xel,1 (), . . . , xel,m () Rm denotes coordinates of the lth corner of the
element domain K e , e E,
u v (x) := (uv1 (x), . . . , uvn (x)) Rn denotes the solution to vth state problem, i.e., to the
system of linear algebraic equations (P v (x)),
v,e
ne denotes the solution of the problem (P v (x))
u v,e (x) := (uv,e
1 (x), . . . , une (x)) R
associated to an element e E in such a way that
v
uv,e
j (x) = uG e(j) (x),
B v (x) := [B v,e1 (x), . . . , B v,en (x)] R2 n denotes a block column vector resulting
after the application of the operator B to the solution of the problem (P v (x)),
2 denotes the value of B(uv (x)) over the element
B v,e (x) := (B1v,e (x), . . . , Bv,e
2 (x)) R
domain K e , e E, such that
ne
X
v,e
e
v,e
e
e
br (b
B (x) := B (x , u ) :=
uv,e
x
)
,
(x)
S
(x
)
B
b
x
j
B
j
j=1
and
Je(p) := I , x, B 1 , . . . , B nv R denotes the value of the cost functional.
The main computational effort in the formula (6.19) is connected with the bracket term, which is
the sensitivity of B(x, u v (x)) with respect to the grid nodal coordinates x, i.e., with the derivatives
Bjv,e (xe , u v,e (x))
xk,l
:=
ne
X
Bjv,e (xe , u v,e ) uv,e
p (x)
uv,e
p
p=1
xk,l
if z {1, . . . , m + 1} : H e (z) 6= k,
#
"
ne
X
Bjv,e (xe , u v,e )
SeB (xe )
v,e
r
x)
, if He (z) = k,
Bxb bp (b
:=
up (x)
xk,l
xez,l
xk,l
:= 0,
p=1
i
(6.20)
95
In (6.20) it remains to express the derivative u v (x)/xk,l . To this goal, let us differentiate the
state equation (P v (x)), where v = 1, . . . , nv , with respect to the lth coordinate of a node x k ,
where k = 1, . . . , nx . We arrive at the following linear system of equations
A(x)
u v (x)
f v (x) A(x)
=
u v (x)
xk,l
xk,l
xk,l
A(x)
.
..
..
:=
.
xk,l
An,1 (x)
...
xk,l
A1,n (x)
xk,l
..
.
An,n (x)
xk,l
u v (x)
:=
xk,l
uv1 (x)
xk,l
..
.
f v (x)
:=
xk,l
(6.21)
f1v (x)
xk,l
..
.
fnv (x)
xk,l
uvn (x)
xk,l
v
(fi (x))ni=1 ,
where
aexe eo (xe ) , ep (xe )
= 0, if z {1, . . . , m + 1} : H e (z) 6= k,
xk,l
aexe eo (xe ) , ep (xe )
=
xk,l
!
Z
SeB (xe )
e
e
e
r
b
br
S
(x
)
B
=
|det(Re (xe ))| db
x+
b
b p
x
x
o
B
xez,l
Kr
!!
Z
e (xe )
S
e
e
e
r
r
B
Bxb bp
+
SB (x ) Bxb bo D
|det(Re (xe ))| db
x+
xez,l
r
K
Z
|det(Re (xe ))|
db
x+
+
SeB (xe ) Bxb bro De SeB (xe ) Bxb brp
xez,l
r
K
!
Z
Se (xe ) br
e e
br |det(Re (xe ))| db
+
S
(x
)
x+
o
p
xez,l
Kr
!
Z
e (xe )
S
e e
r
r
x+
bp |det(Re (xe ))| db
+
S (x ) bo
xez,l
Kr
Z
|det(Re (xe ))|
db
x, if He (z) = k.
+
Se (xe ) bro Se (xe ) brp
xez,l
Kr
(6.22)
96
Note that none of A(x)/xk,l is evaluated itself. In Section 6.3.4 we will rather assemble the
vector
[G(x, u v (x))]T v Rmnx ,
where
A(x)
A(x)
v
v
u (x), . . . ,
u (x) Rn(mnx ) ,
G(x, u (x)) :=
x1,1
xnx ,m+1
(6.23)
and where v Rn .
n 1
n 1
nv
h
X
Gradx (B(x, u v )) +
+ Grad(x()) gradx I , x, B 1 , . . . , B nv +
{z
} |
{z
}
|
{z
} v=1 |
n mnx
mnx 1
(mnx )(2 n )
io
+ G(x, u ) A(x)1 Gradu v (B(x, u v )) gradB v I , x, B 1 , . . . , B nv
,
| {z } | {z } |
{z
} |
{z
}
v T
(mnx )n
nn
n(2 n )
(2 n )1
(6.24)
in which matrix (or vector) size is written under the brackets, and where the gradients are
grad Je(p) :=
Je(p)
p1
..
.
Je(p)
pn
, grad I , x, B 1 , . . . , B nv :=
1 (p)
p1
..
.
1 (p)
pn
.
.
..
..
:=
x1,m ()
xnx ,1 ()
x1,1 ()
.
.
.
.
.
.
n
n
n
I (,x,B 1 ,...,B nv )
1
..
.
I (,x,B 1 ,...,B nv )
n
...
..
.
...
n (p)
p1
..
.
n (p)
pn
...
..
.
xnx ,1 ()
1
...
xnx ,1 ()
n
gradx1 I , x, B 1 , . . . , B nv
..
gradx I , x, B 1 , . . . , B nv :=
,
.
1
nv
gradxnx I , x, B , . . . , B
,
97
in which for k = q, . . . , nx
gradxk I , x, B 1 , . . . , B nv
where further
:=
I (,x,B 1 ,...,B nv )
xk,1
..
.
I (,x,B 1 ,...,B nv )
xk,m
..
..
Gradx (B(x, u v )) :=
.
.
e
v,e
1
1
Gradxnx (B(x , u
)) . . .
in which for k = 1, . . . , nx
where further
..
,
.
e n
v,en
Gradxnx (B(x , u
))
z {1, . . . , m + 1} : H e (z) 6= k,
B2(xe ,u v,e )
B1(xe ,u v,e )
.
.
.
e
e
xz,1
xz,1
..
..
..
, if He (z) = k,
.
.
.
e
v,e
B2(x ,u )
B1(xe ,u v,e )
...
xe
xe
if
z,m
z,m
gradB v
gradB v,e1 I , x, B 1 , . . . , B nv
..
I , x, B 1 , . . . , B nv :=
,
.
1
nv
v,e
gradB n I , x, B , . . . , B
in which for e E
gradB v,e I , x, B 1 , . . . , B nv
:=
I (,x,B 1 ,...,B nv )
B1v,e
..
.
I (,x,B 1 ,...,B nv )
Bv,e
2
Now, all the art is how to evaluate the expression (6.24) efficiently. Basically, there are two
possibilities. Either we proceed from left to right, then it is called the direct method, or the other
way round, which is called the adjoint method. The main computational effort is in calculating
the state sensitivity. In case of the direct method, we would solve n v n systems consisting of n
linear equations, while, in case of the adjoint method, we have to solve just n v systems of n linear
equations. This is why we prefer the latter. Let us note that if the constraint function were state
dependent, the adjoint method would arrive at solving n v (1 + n ) systems of n linear equations.
98
Assemble 2 := G(x, u v )T
:= + 1 + 2
end for
:= + I x
Assemble := Grad(x())
:= + I
Assemble
:= Grad((p))
e
grad J (p) :=
Only the gradients of I have to be provided by the user. All the other parts are more or less
independent. The particular assembling procedures are depicted in Algorithms 79. On the CD
there are enclosed the corresponding M ATLAB [208] routines used for optimal shape design in
2dimensional magnetostatics.
Algorithm 6 Adjoint method: the shape part (Assemble )
Given p, , and v
:= 0
for i := 1, . . . , n do
for j := 1, . . . , n do
i := i + [F (p)](x,j ) pi vj
end for
end for
to L UK A S , M UHLHUBER
, AND K UHN [125].
99
for o := 1, . . . , n do
o := b() o
for j := 1, . . . , nx do
o := o + Mj,o j
end for
end for
Algorithm 8 Adjoint method: FEM preprocessor part (Assemble 2 )
Given x, u v , and
2 := 0
for i := 1, . . . , n do
for z := 1, . . . , m + 1 do
for l := 1, . . . , m do
k := m (Hei (z) 1) + l
for o, p := 1, . . . , ne do
if o = p or G ei(o) 6 I0h and G ei(p) 6I0h then
.
xez,li uvG ei(p)
The library supports routines for evaluating the cost and constraint functions and their gradients with respect to the shape design parameters. The library can be used with any gradient or
Newtonlike optimization algorithm, as SQP with BFGS (BFGSSQP), see (6.17). The library
uses 3 external modules: a mesh generator, a finite element method (FEM) module, and a solver of
linear algebraic systems of equations. The mesh generator runs just once at the very beginning and
it discretizes the domain for the initial design p 0 . It provides initial grid nodes x0 := x ((p0 ))
and the discretization T . Having some grid nodes x, the FEM preprocessor assembles the matrix A(x) and the righthand side vector f v (x) for each state v = 1, . . . , nv . Then, the solver
of linear systems provides the solution u v to the FEM postprocessor that assembles the solution
B v . The FEM module is moreover supposed to assemble the corresponding gradientvector multiplications in Algorithm 5, namely 1 , 2 , and . The efficiency of the library strongly depends
on the linear system solver. We have used the software tools developed by K UHN , L ANGER ,
AND S CH OBERL
[117] at the University Linz in Austria, where the conjugate gradient method,
cf. G OLUB AND VAN L OAN [69], with a multigrid preconditioning, cf. H ACKBUSCH [78], is
involved.
Now, let us explain how the optimization proceeds in terms of Fig. 6.1. Given an initial vector
p0 of design parameters and a discretization parameter h, the BFGSSQP algorithm starts its run
while at the same time the mesh generator provides the initial grid x 0 and the grid topological
information T . Then, a quadratic programming subproblem is going
solved, see also Algo to be
e
e
rithms 3 and 4, which requires evaluation of J (p0 ), (p0 ), grad J (p0 ) , and Grad((p0 )).
100
101
p
p0
Je(p), (p)
p
grad Je(p)
BFGSSQP
Grad((p))
p
Computation
of Je,
Bv
T
Computation
of grad(Je), Grad()
T Bv
Bv
I , I x , I B v
((p0 ))
x0
User interface
Solver
of linear systems
Mesh
generator
A, f v
uv
T
FEM
pre/post
processor
Adjoint method
1, 2,
IB v ,
x0 , T Shapetomesh
mapping
(p0 )
T
p
Designtoshape
mapping
p0
S CH OBERL
[117], S CHINNERL , L ANGER , AND L ERCH [183], or S CHINNERL ET AL . [184]. We
want to establish a hierarchy of discretizations to our continuous shape optimization problem ( Pe)
such that the optimized design achieved at a coarse level is used as the initial design at a next finer
discretization level. The first results can be found in L UK A S [123] and in L UK A S [128].
In this section, we employ the full notation with both the regularization parameter and the
discretization parameter h
Find ph :
,
(Peh )
h
h
h
e
e
J p J (p) p
102
where Jeh : 7 R denotes the discretized and regularized cost functional and
:= {p Rn | (p) 0 }
h
Solve (Pe ) with the initial design p0
ph
1 and h1 , and the optimization algorithm proceeds ending up with a coarse optimized design p h11 .
Then, we refine both the regularization and the discretization parameters and run the optimization
algorithm with smaller values of 2 and h2 while using the design ph11 as the initial one at this
second level. We end up with a finer optimized design p h22 , and so further. The approach is
described in Algorithm 11.
Algorithm 11 Hierarchical optimization approach
Given 1 , h1 , and p1,0
l := 1
while l > 1 and a terminate criterion is not satisfied do
Discretize (Pe) at the level l
(Pehll )
pl+1,0 := phll
l := l + 1
end while
hl1
is the optimized design
pl1
The hierarchical approach in shape optimization has turned out to be much more effective than
the classical one whenever the coarse optimized design p h00 approximates the true one rather well.
The crucial part of the algorithm is the refinement step. The updated values of l+1 and hl+1 must
not be too smaller than those of l and hl , since SQP would take many iterations. On the other
hand, if the refinement is rather coarse, i.e., the values of l+1 and hl+1 are comparable to l and
hl , there is hardly any progress in the SQP algorithm and the hierarchical approach takes many
iterations. In Section 7.4 we provide some numerical experiments without any use of multigrid
yet. The idea, which we want to investigate in the future, is that the refinement strategy should
benefit from the aposteriori finite element error analysis and from multigrid techniques. The first
papers in this context have appeared just recently, see R AMM , M AUTE , AND S CHWARZ [164],
and S CHLEUPEN , M AUTE , AND R AMM [185]. In the paper by S CHERZER [182] a multilevel
approach is used for solving nonlinear illposed problems.
103
Another improvement can be done, when applying the multilevel approach on the level of
mathematical modelling. It means that in those application where we can reduce the problem complexity by neglecting a dimension or some physical phenomena we can first solve the discretized
reduced problem and then use the result as the initial design for the more complex problem. A
typical example might be solving a shape optimization problem governed first by 2d linear magnetostatic state problem, then, prolong the optimized design dimensionally and use it as the initial
design for shape optimization governed by 3d linear magnetostatics, and finally use the resulting
shape as the initial design for shape optimization governed by 3d nonlinear magnetostatic state
problem. In Section 7.4, we will give a numerical test of the 2d/3d dimensional step.
104
Chapter 7
in L UK A S , M UHLHUBER
, AND K UHN [125].
also note that the electromagnets have been developed at the Institute of Physics, V SBTechnical
University of Ostrava, Czech Republic in the research group of Prof. Jaromr Pistora. Some
instances have been already delivered to the following laboratories:
105
106
pole
ferromagnetic
yoke
pole head
0.1
x2 [m]
0.05
coil
0.05
0.1
0.15
0.2
0.2
magnetization
area m
magnetization
planes
0.1
0
x1 [m]
0.1
0.2
107
the pole heads. In this area the magnetic field is homogeneous enough with respect to the normal
vector of some polarization plane, see Fig. 7.2. We pass an optical (light) beam of a given polarPSfrag replacements
ization vector to the sample. There it reflects and components of the reflected polarization vector
are measured in terms of the Kerr rotation and ellipticity, respectively. Briefly saying, we measure
the polarization state of the reflected beam. The Kerr rotation means the difference between the
angle of the main ellipticity axis of the reflected beam from that one before the reflection. Typical
measured data is depicted in Fig. 7.3, which was measured by Ing. Igor Kopriva at the Institute of
Physics, VSBTechnical
University of Ostrava, see also KOP R IVA ET AL . [111].
0.15
0.1
0.05
0
*
+
0.05
0.1
0.15
0.2
20
40
60
80
100
120
sample orientation in magnetic field [degree]
140
108
magnetic field are minimized, but the field itself is still strong enough.
d3
2
2
2
,
x := (x1 , x2 , x3 ) R (x1 ) + (x2 ) < (r) and |x3 | <
2
3
d1 := d2 := 0.4 [m],
d3 := 0.02 [m],
r := 0.2 [m].
Figure 7.4: Geometrical models of the Maltese Cross and ORing electromagnets
parts are the ferromagnetic yoke and the poles. The blue parts are the coils. In Fig. 7.5 and in
Fig. 7.6 we can see dimensions in milimeters for geometrical models of the Maltese Cross and
of the ORing electromagnet, respectively. The symbol yoke stands for the domain occupied by
the ferromagnetic yoke, the symbols westp , northwestp , northp , northeastp , eastp , southeastp ,
southp , and southwestp denote the domains occupied by the particular poles, and the symbols
westc , northwestc , northc , northeastc , eastc , southeastc , southc , and southwestc denote the
domains that are occupied by the corresponding coils. In Fig. 7.7 we can see the west pole of
109
x2
45
x2
PSfrag replacements
45
northp
yoke
50
westc
160
eastp
65
eastc
x1
southc
x3
50
45
westp
northc
30
45
yoke
45
southp
22.5
25
45
x2
northp
northc
yoke
50
northeastc
eastc
southwestc
southeastc
southwestp
160
eastp
x1
southeastp
southp
20
40
45
yoke
southc
x3
50
northwestc
westc
westp
northeastp
45
northwestp
40
x2
20
40
3
.
8
The box constraints (5.2) are chosen such that the shape of the west pole head must not be either
higher than the bottom of the north coil or penetrate with the neighbouring pole head. Therefore,
we choose
l := 0.012 [m], u := 0.05 [m]
for the Maltese Cross and
l := 0.028 [m],
u := 0.05 [m]
for the ORing. Then, the set U of admissible shapes is given by (5.3) and Lemma 5.1 holds.
Since from the practical point of view we cannot manufacture any shape, we will restrict
ourselves to those that are described by a Bezier patch of a fixed number of design parameters
n := n,1 n,2 ,
and we choose
n,1 := 4,
n,2 := 3.
111
x2
pole head
x2
23
20
x1
20
westp
41.5
westc
60
PSfrag replacements
30
135
x3
20
41.5
electric currents
60
The mapping F : 7 U, see also (5.5), is the following (tensor product) Bezier mapping that
involves the symmetry
(x1 , x3 ) := [F (x1 , x3 )] (p) :=
n,1 n,2
XX
2n,1 1 2x1 + dpole,1
2n,1 1 2x1 + dpole,1
:=
pi,j i
+ i
2dpole,1
2dpole,1
i=1 j=1
2n,2 1 2x3 + dpole,3
2n,2 1 2x3 + dpole,3
j
+ j
, (x1 , x3 ) , (7.1)
2dpole,3
2dpole,3
where for n N, i N, i n, and t R such that 0 t 1
in (t) :=
(n 1)!
ti1 (1 t)ni ,
(i 1)! (n i)!
(7.2)
112
15
x3 [mm]
x3 [mm]
15
10
5
0
5
Sfrag replacements
10
5
0
5
10
PSfrag replacements
15
30
25
10
15
30
15
15
25
10
20
10
15
20
10
x2 [mm]
10
5
15
x2 [mm]
x1 [mm]
5
10
15
15
x1 [mm]
Figure 7.8: Bezier design parameters and the corresponding shape of the north pole head
1
,
0
D1 :=
1
,
1
where 0 := 4107 [H.m1 ] and 1 := 51000 are the permeabilities of the air and the ferromagnetic parts, respectively.
Further, we consider
nv := 2 and nv := 3
variations of the current excitations in case of the Maltese Cross and the ORing electromagnet,
respectively. In both the cases the righthand side f v (x) is calculated from the electric current I,
from the number of turns nI
I := 5 [A] or I := 1.41 [A],
nI := 600,
respectively, for the Maltese Cross or the ORing, and from the crosssection area through the
coils, see also Figs. 7.57.7,
Sc := 0.03 0.01 = 3 104 [m2 ]
or
Sc := 0.023
0.02 + 0.01075
= 4.6125 104 [m2 ].
2
The current densities f v Jv satisfy (5.2.2), i.e., they are divergencefree. The absolute value of
Jv (x) is nonzero only in the subdomains westc , . . . , southwestc where the direct electric currents
are located
nI I
|Jv (x)| =
.
Sc
These subdomains are independent of the shape , therefore, Assumption 5.2 is satisfied.
Now, we will describe the directions of J v (x) for both the electromagnets and each variation
v of the current excitation. During the description we will be referring to Fig. 7.7 and to Figs. 7.9
7.13. We say that two coils are pumped (excited) in the same sense if there is a magnetic circuit
113
that goes through both of them. Otherwise, the coils are excited in the opposite sense. In Fig. 7.9
the north and the south coils are excited in the same sense, the other two coils are switched off. In
Fig. 7.10 all the coils are pumped by currents such that the west and the north one are in the same
sense, the south and east one as well, but the west and the south one are in the opposite sense, as
well as the north and the east one are. In Fig. 7.11 the north and the south coils are pumped in the
same sense, the others are switched off. In Fig. 7.12 the situation is similar to Fig. 7.10 while the
northwest, northeast, southeast, and southwest coils are switched off. Finally, in Fig. 7.13 the
following 4 couples of coils are excited in the same sense: the southwest and south, the west and
southeast, the northwest and east, and the north and northeast coil.
Figure 7.9: Magnetic flux lines for the vertical current excitation (v := 1) for the Maltese Cross
(7.3)
114
Figure 7.10: Magnetic flux lines for the diagonal current excitation (v := 2) for the Maltese Cross
where
Bv (x) := curlx ([uv ()] (x)) ,
1
(B (x)) :=
meas(m )
avg,v 2
Bmin
m
2
avg,v
v (Bv (x)) := max 0, Bmin
B avg,v (Bv (x))
,
:= 106 ,
(7.4)
(7.5)
where uv () stands for the solution to (W v ()) and where the following is the average magnetic
flux density
Z
1
avg,v
v
|Bv (x) nvm | dx.
(7.6)
B
(B (x)) :=
meas(m )
m
nvm
:=
nvm ,
(0, 1, 0)
(1/ 2, 1/ 2, 0)
,v = 1
,v = 2
or
0)
(0, 1,
v
nm := (1/ 2, 1/ 2, 0)
(cos(/8), sin(/8), 0)
,v = 1
,v = 2
,v = 3
in case of the Maltese Cross or the ORing, respectively. The minimal average magnetic flux
densities are
avg,1
avg,2
Bmin
:= 0.1 [T], Bmin
:= 0.15 [T]
for both the geometries and, additionally, in case of ORings superdiagonal excitation it is
avg,3
Bmin
:= 0.3 [T].
115
Figure 7.11: Magnetic flux lines for the vertical current excitation (v := 1) for the ORing
116
Figure 7.12: Magnetic flux lines for the diagonal current excitation (v := 2) for the ORing
Finally, the discretized (and regularized) cost functional is given by (5.36) and by the relations
(7.3)(7.6) which arrive at the following expressions
I
nv ,n
B 1,n
, . . . , B
nv h
i
1 X
v,h
:=
h (B v,n
(B v,n
)+
) ,
nv
(7.7)
v=1
where B v,n
is the elementwise constant magnetic field given by (6.7)(6.8) and where
h (B v,n
) :=
1
meas(m )
avg,v 2
Bmin
eE h :K e m
|B v,n,e
B avg,v,n (B v,n,e
) nvm |2 meas (K e ) ,
(7.8)
2
avg,v
v,h (B v,n ) := max 0, Bmin
B avg,v,n (B v,n
,
)
B avg,v,n (B v,n
) :=
meas(m )
eE h :K e
|B v,n,e
nvm | meas (K e ) .
(7.9)
(7.10)
In order to justify using a Newtonlike optimization algorithm, we still have to satisfy Assumptions 6.1 and 6.2. Concerning Assumption 6.1, components of the constraint functional
: Rn 7 Rn , where n := 2n = 2n,1 n,2 , are as follows:
(
l pi,j
, i n,1 , j n,2
k (p) :=
for k = in,2 + j = 1, . . . , n
pin,1 ,jn,2 u , i > n,1 , j > n,2
Hence, Assumption 6.1 is obviously satisfied. Now, we shall verify Assumption 6.2. The smoothness of the designtoshape mapping F with respect to p is easy to see from (7.1). Concerning
the smoothness of the shapetomesh mapping x h , which is given by (6.2), it is well known that
117
Figure 7.13: Magnetic flux lines for the superdiagonal current excitation (v := 3) for the ORing
the stiffness matrix K h (x0 ) is nonsingular as far as we consider a Dirichlet boundary condition
at a certain part of either the boundary h or an interface h0 (h ) h1 (h ). Since b h h
involves the following nonhomogeneous Dirichlet design interface boundary condition
4xh = Mh h on h ,
where h denotes the design interface, then, x h , see (6.2), is smooth with respect to h . Further, due to (4.66), (4.68), and (4.69) we can see that each of Re (xe ), Se (xe ), and Securl (xe ),
respectively, is smooth as far as K e does not flip. The last item of Assumption 6.2 easily follows
from (7.7)(7.10).
:=
dpole,1
0,
2
The Lipschitz constant as well as the box constraints remain. The set of admissible shapes is given
by (5.3).
Concerning , we choose the following number of design parameters
n := n,1 := 4,
118
which is the number of control Bezier nodes. The domain is decomposed into (n 1) subintervals with the n nodes
x,i :=
(i 1)dpole,1
n 1
for i = 1, . . . , n .
n
X
i=1
pi
i2n 1
2x1 + dpole,1
2dpole,1
i2n 1
2x1 + dpole,1
2dpole,1
(7.11)
where x1 and in is given by (7.2). The mapping F is depicted in Fig. 7.14, where the red
line connects the design parameters and the blue line is the resulting shape.
50
45
x2 [mm]
40
35
30
25
PSfrag replacements
20
15
10
x1 [mm]
10
15
Figure 7.14: Bezier design parameters and the corresponding 2d shape of the north pole head
We concern the 2dimensional magnetostatics with the differential operator
B := grad.
The space H(grad; 2d ) is equivalent to the space H 1 (2d ), hence, Assumptions 3.23.5 are
satisfied by Theorems 3.73.9 and by (3.27). Further, we can proceed throughout the rest of Section 7.2.4. We only recall that the compatibility condition (5.2.2) is satisfied, as Ker(grad; 2d ) =
{0}.
As far as calculation of the 2d continuous cost functional is concerned, the magnetization area
is
m := [0.005, 0.005] [0.005, 0.005] [m]
119
and the expressions (7.3)(7.10) remain, where for v = 1, . . . , nv the vectors nvm are as follows:
nvm
:=
(0, 1)
(1/ 2, 1/ 2)
,v = 1
,v = 2
or
(0, 1)
v
nm := (1/ 2, 1/ 2)
(cos(/8), sin(/8))
,v = 1
,v = 2
,v = 3
in case of the Maltese Cross or the ORing electromagnet, respectively. The values of minimal
magnetic flux densities remain as well.
Now, we do not need to introduce any regularization of the state problem, as the bilinear form
is elliptic on the whole space H0 (grad; 2d ) H01 (2d ). Concerning the finite element discretization with a discretization parameter h > 0 such that h h, we approximate the domain
2d by a polygonal domain h2d , while for the Maltese Cross h2d := 2d and for the ORing it is
like in Figs. 7.117.13. Then, Assumption 4.3 holds. We use linear Lagrange elements that are described in Section 4.4.1. The discretization T h (h ) satisfies the minimum angle condition (4.61)
and, therefore, Assumptions 4.14.2 and Assumptions 4.44.7 are satisfied. The remaining specification of the 2d discretized shape optimization problem is as in Section 7.2.5. The only difference
is that the smoothness of Re (xe ), Se (xe ), and Segrad (xe ) is now due to (4.56), (4.58), and (4.59),
respectively.
extension package for shape optimization, see L UK A S , M UHLHUBER , AND K UHN [125], which
were all developed in the research project SFB F013 at the University Linz in Austria. All the
Figure 7.15: Optimized 2d and 3d pole heads of the Maltese Cross electromagnet
variables including the symmetry. Concerning discretization of the state problem, the discretiza-
120
tion parameters are
h := 0.05 [m],
h := 0.025 [m],
and
h := 0.0125 [m]
at the first (coarsest), second, and third (finest) level, respectively. There are 12272 degrees of
freedom at the last (3rd, finest) level. The optimization took 8 SQP iterations at the first (coarsest)
level, 35 ones at the second level, and 25 ones at the third (finest) level, which was all done in 1
hour and 59 minutes, see also Fig. 7.17. The cost functional decreased from 1.97 10 6 (1st level)
to 1.49 106 (3rd level). The 3d shape is determined by 12 design variables with the symmetry
involved. The state problem at the finest level is discretized by 29541 degrees of freedom. Within
the multilevel approach we made a step between the 2d and 3d model such that at the first level we
had a coarse discretization of the 2d problem, at the second level we had a coarse discretization
of the 3d problem, and at the last third level we had a fine discretization of the 3d problem. The
calculation proceeded in 6, 50, and 37 SQP iterations at the respective levels, i.e., 93 SQP iterations
in total, which took 29 hours and 46 minutes, see also Fig. 7.18. The cost functional decreased
from 2.57 106 (2nd level) to 7.32 107 (3rd level).
The 2d optimized pole head of the ORing electromagnet is depicted in Fig. 7.16 while the
initial shape was again a rectangle. The shape is described by 7 design variables including the
121
number of des.
variables
number of
unknowns
SQP
iters.
CPU
time
1386
56s
4705
47
27m 31s
4970
53
30m 12s
12272
72
1h 58m 55s
12324
125
6h 44m 0s
3d optimization is employed. This means that we have to prolong the 2d coarse optimized design
into the third dimension by constant. This is the step that goes through the thick line in Fig. 7.18.
Then, we proceed on, as we did in Fig. 7.17. We use the 3d coarse optimized design as the initial
guess at the third level (h := 0.025 [m]). From the last line in Fig. 7.18, we can see that the whole
calculation took almost 30 hours. We tried to compare this general multilevel approach with the
classical one, but the calculation took more than 4 days and several remeshings of the geometry
had to be done. Unfortunately, in 4 days we were still not able to achieve the optimal solution,
hence, the calculation was stopped.
provide a Matlab implementation, see L UK A S [119], of the method, which is enclosed on the CD.
Let us consider a 2d academic optimization problem governed by linear magnetostatics. Its
geometry is depicted in Fig. 7.19. Due to the symmetry and since we employ only one state
problem, the computational domainis the topleft quarter
2d := (0.2, 0) (0, 0.1) [m].
122
optimized
designs
number of des.
variables
number of
unknowns
SQP
iters.
CPU
time
2777
44s
12086
47
7h 27m
12
29541
93
29h 46m
meas(m ) kBreq k2
u := 0.01 [m].
123
x2
40
40
20
40
20
westc
40
40
eastc
20
PSfrag replacements
40
40
40
40
eastp
x1
20
westp
40
eastc
westc
x2 [m]
x2 [m]
0.1
0.05
PSfrag replacements
0.05
PSfrag replacements
0
0.2
0.1
x1 [m]
0
0.2
0.1
x1 [m]
Figure 7.20: Initial design and the magnetic field of the twocoils problem
improved from 0.0077 to 0.0042. In Table 7.1 there is a comparison of the SQP method using the
firstorder numerical differentiation with the SQP method using the adjoint method for calculating
gradients. Both the calculations took 4 SQP iterations. We can see that the SQP with the numerical
differentiation needed 21 evaluations of the state problem while only 5 were needed by the adjoint
method plus additional 4 evaluations of the adjoint state problem. In fact, the numerical differentiation took 5 evaluations of the cost functional plus additional 4 (number of design variables) times
4 (number of SQP iterations) evaluations, which give the total 21 evaluations. The cost functional
was evaluated in about 18 seconds while the adjoint state problem in about 5 seconds. Enclosed
there is a CD with the Matlab implementation, see also L UK A S [119], where you can run
> optimization(n); % numerical differentiation
> optimization(a); % adjoint method
to see this comparison.
124
0.1
x2 [m]
x2 [m]
0.1
0.05
PSfrag replacements
0.05
PSfrag replacements
0
0.2
0.1
0
0.2
x1 [m]
0.1
x1 [m]
Figure 7.21: Optimized design and the magnetic field of the twocoils problem
number of cost func. evals.
number of adjoint problem evals.
number of SQP iterations
CPU time
numerical differentiation
21
0
4
6min 26sec
adjoint method
5
4
4
1min 53sec
Figure 7.22: Initial and optimized 2d pole heads of the Maltese Cross electromagnet
In Fig. 7.23 there are distributions of the normal component of the magnetic flux density depicted. The blue solid line is the normal magnetic flux density along the magnetization plane for
the diagonal excitation, see also Fig. 7.10, of the initial design, see Fig. 7.22. The red solid line
is the normal magnetic flux density for the diagonal excitation of the optimized design. The blue
and red dashed lines, respectively, are the normal magnetic flux densities along the magnetiza-
125
tion plane for the vertical excitation, see also Fig. 7.9, of the initial and optimized designs. In
Fig. 7.23 we can see a significant improvement of the homogeneity of the magnetic field. The cost
functional calculated from the measured data shows that it decreases 4.5times. The cost functional calculated from the computer simulated magnetic field decreases only twice. The relative
differences between the measured and the calculated magnetic fields are about 30%, which might
be caused by saturation of the magnetic field in the corners. Employing a nonlinear governing
magnetostatic state problem should improve also the mismatch of the magnetic fields. Nevertheless, the significant improvement of the cost functional shows that the optimization works well, no
matter how big the nonlinearities in the direct magnetic field problem are.
1800
init. shape, vertical excit.
init. shape, diagonal excit.
opt. shape, vertical excit.
opt. shape, diagonal excit.
1600
1400
1200
1000
800
PSfrag replacements
600
8
distance [mm]
Figure 7.23: Magnetic field for the initial and optimized design of the MC electromagnet
126
Chapter 8
Conclusion
This thesis treated with the shape optimization in both two and threedimensional linear magnetostatics. The aim was to present a complete picture of the mathematical modelling process. We
dealt with both the theoretical and computational aspects and demonstrated them on an application
being of a practical purpose in the research on magnetooptic effects.
Let us summarize the main results obtained in the thesis.
In Section 3.4 we developed an abstract theory for weak formulations of linear elliptic
secondorder boundary vectorvalue problems (BVP).
In Theorem 4.2 we proved the convergence of the solution of the finite element approximation to our abstract BVP while also dealing with an inner approximation of the original
domain with the Lipschitz boundary by a sequence of domains with polyhedral (or polygonal) boundaries.
In Sections 4.4.1 and 4.4.2 we concretized the abstract framework for the linear Lagrange
and Nedelec elements on triangles or tetrahedra, respectively.
In Chapter 5 we introduced an abstract shape optimization problem and its finite element
approximation. We proved both the existence and convergence theorems while they rely on
Lemma 5.3 and Theorem 4.2, respectively.
In Section 6.3.4 there is the heart of the thesis. There we developed an efficient implementation of the adjoint method for the firstorder sensitivity analysis. We provided also a Matlab
implementation, which is enclosed on the CD.
In Section 6.4 we introduced a multilevel optimization approach, which is a rather new
technique. It is the first step towards adaptive optimization algorithms, as they have been
recently presented in R AMM , M AUTE , AND S CHWARZ [164] and in S CHLEUPEN , M AUTE ,
AND R AMM [185]. The efficiency of our multilevel optimization approach was documented
on numerical tests in Section 7.4.1.
Finally, in Chapter 7 we presented a reallife application arising from the research on magnetooptic effects. We began with the physical description, went through the mathematical
settings, and ended up with the manufacture of the optimized design and with the discussion
of real improvements based on the physical measurements of the magnetic field.
127
CHAPTER 8. CONCLUSION
128
In Chapter 5, we met one serious obstacle, see Remark 5.1, that the standard approximation
theory does not completely cover problems of complex geometries. Namely, it is due to that we
can hardly find a continuous mapping between the shape design nodes and the remaining nodes
in the discretization grid. For fine discretizations and large changes in the design shape some
elements flip. One possible outcome is in the use of the multilevel optimization techniques where
on the fine grids the difference between the initial and optimized shapes is not that big. Another
outcome might be when using composite finite elements that were developed for the treatment with
complicated geometries in the papers by H ACKBUSCH AND S AUTER [79, 80]. It is connected to an
idea which was given to me in January 2002 by RNDr. Jan Chleboun, CSc. from the Mathematical
Institute of the Czech Academy of Sciences. The idea is to use a fixed regular grid independent
of the geometry and to resolve the fine details of the geometry within special elements that arise
by the intersection of the geometry and the regular grid. This will move all the programming
effort into the development of such special finite elements instead the shapetomesh mapping.
We can also avoid this problem by using a boundary element discretization. From its matter, this
is very suited for optimal shape design, as we need to handle only the boundary discretization.
Nevertheless, construction of efficient multigrid solvers as well as using the method for nonlinear
governing state problems are still topics of the current research.
Finally, let us draw the further directions of this research. They are mainly focused
on development and rigorous analysis of the adaptive multilevel techniques in the shape
optimization,
on synergies among the inverse and shape optimization problems, namely, on the regularization techniques and numerical methods, e.g., the homogenization or levelset methods,
on common aspects in the topology and shape optimization,
on development of a userfriendly and welldocumented scientific computing software tool
for structural (both shape and topology) optimization,
and on reallife applications in both electromagnetism and mechanics involving complex
geometries and nonlinearities of the state problem, provided correct mathematical settings,
i.e., the existence of a solution at least.
Bibliography
[1] Ansys, http://www.ansys.com.
[2] Fluent, http://www.fluent.com.
[3] J. C. Adam, A. Gourdin-Serveniere, J. C. Nedelec, and P. A. Raviart, Study of an implicit scheme for integrating Maxwells equations, Comput. Methods Appl. Mech. Eng.
22 (1980), 327346.
[4] R. A. Adams, Sobolev spaces, Academic Press, New York, 1975.
[5] G. Allaire, Homogenization and twoscale convergence, SIAM J. Math. Anal. 23 (1992),
14821518.
[6]
[7] G. Allaire, E. Bonnetier, G. A. Francfort, and F. Jouve, Shape optimization by the homogenization method, Numer. Math. 76 (1997), 2768.
[8] G. Allaire, F. Jouve, and A.-M. Toader, A levelset method for shape optimization, C. R.
Acad. Sci. Paris 334 (2002), 11251130.
[9] C. Amrouche, C. Bernardi, M. Dauge, and V. Girault, Vector potentials in three
dimensional nonsmooth domains, Math. Methods Appl. Sci. 21 (1998), 823864.
[10] T. Apel and J. Schoberl, Multigrid methods for anisotropic edge refinement, SIAM J. Numer. Anal. 40 (2002), no. 5, 19932006.
[11] J. Arora, Introduction to optimum design, McGraw-Hill, New York, 1989.
[12] J. P. Aubin, Approximation of elliptic boundaryvalue problems, Pure and Applied Mathematics, vol. 26, John Wiley & Sons, New York, 1972.
[13] O. Axelsson, Iterative solution methods, Cambridge University Press, 1994.
[14] I. Babuska and A. K. Aziz, Survey lectures on the mathematical foundations of the finite
element method, The Mathematical Foundations of the Finite Element Method with Applications to Partial Differential Equations, Academic Press, New York and London, 1972.
[15]
, On the angle condition in the finite element method, SIAM J. Numer. Anal. 13
(1976), 214226.
129
BIBLIOGRAPHY
130
, Optimization of structural topology, shape and material, Springer, Berlin, Heidelberg, 1995.
[22] M. P. Bendse and O. Sigmund, Topology optimization. Theory, methods and applications,
Springer, Berlin Heidelberg, 2003.
[23] P. T. Boggs and J. W. Tolle, Sequential quadratic programming, Acta Numer. 4 (1995),
151.
[24] A. D. Borner, Mathematische Modelle f u r Optimum Shape Design Probleme in der Magnetostatik, Ph.D. thesis, Universitat Hamburg, Germany, 1985.
[25] T. Borrvall, Computational topology optimization of elastic continua by design restrictions,
Ph.D. thesis, Linkoping University, Sweden, 2000.
[26] A. Bossavit, Computational electromagnetism. Variational formulations, complementarity,
edge elements, Orlando, FL: Academic Press, 1998.
[27] D. Braess, Finite elements, Cambridge, 2001.
[28] J. H. Bramble, Multigrid methods, Pitman Research Notes in Mathematics, vol. 294, John
Wiley & Sons, 1993.
[29] J. H. Bramble, J. E. Pasciak, and J. Xu, Parallel multilevel preconditioners, Math. Comput.
55 (1990), 19111922.
[30] R. B. Brandtstatter, W. Ring, Ch. Magele, and K. R. Richter, Shape design with great geometrical deformations using continuously moving finite element nodes, IEEE Transactions
on Magnetics 34 (1998), no. 5, 28772880.
[31] C. A. Brebbia (ed.), Topics in boundary element research, vol. 6 electromagnetic applications, Springer, Berlin, 1989.
[32] S. C. Brenner and L. R. Scott, The mathematical theory of finite element methods, Springer,
1994.
[33] F. Brezzi and M. Fortin, Mixed and hybrid finite element methods, Springer, New York,
1991.
BIBLIOGRAPHY
131
[34] R. A. Brockman, Geometric sensitivity analysis with isoparametric finite elements, Commun. Appl. Numer. Methods 3 (1987), 495499.
[35] D. Bucur and J.-P. Zolesio, Existence and stability of the optimum in shape optimization,
ZAMM 76 (1996), no. 2, 7780.
[36] M. Burger, A level set method for inverse problems, Inverse Probl. 17 (2001), no. 5, 1327
1355.
[37] M. Burger and W. Muhlhuber, Iterative regularization of parameter identification problems
by sequential quadratic programming methods, Inverse Probl. 18 (2002), no. 4, 943969.
[38]
BIBLIOGRAPHY
132
[51] R. Courant and D. Hilbert, Methods of mathematical physics, Interscience Publishers, New
York, 1962.
[52] W. C. Davidon, Variable metric method for minimization, Technical report ANL5990 (revised), Argonne National Laboratory, 1959.
[53]
[54] M. C. Delfour and J.-P. Zolesio, Shape and geometries: Analysis, differential calculus, and
optimization, Advances in Design and Control, vol. 4, SIAM, 2001.
[55] J. E. Dennis and R. B. Schnabel, Numerical methods for unconstrained optimization and
nonlinear equations, Prentice Hall, 1983.
[56] P. Doktor, On the density of smooth functions in certain subspaces of Sobolev spaces, Comment. Math. Univ. Carol. 14 (1973), 609622.
, An optimal positive definite update for sparse Hessian matrices, SIAM J. Optim.
5 (1995), 192218.
BIBLIOGRAPHY
133
[71] C. Grossmann and H.-G. Ross, Numerik partieller Differentialgleichungen, Teubner Studienbucher, Stuttgart, 1992.
[72] C. Grossmann and J. Terno, Numerik der Optimierung, 2nd ed., Teubner Studienbucher,
Stuttgart, 1997.
[73] G. Haase, M. Kuhn, U. Langer, S. Reitzinger, and J. Schoberl, Parallel Maxwell solvers,
Scientific computing in electrical engineering (Ursula van Rienen, Michael Gunther, and
Dirk Hecht, eds.), Lect. Notes Comput. Sci. Eng., vol. 18, 2001, pp. 7178.
[74] G. Haase and U. Langer, On the use of multigrid preconditioners in the domain decomposition method, Parallel algorithms for partial differential equations, Proc. 6th GAMM-Semin.,
Notes Numer. Fluid Mech., vol. 31, 1991, pp. 101110.
[75] G. Haase, U. Langer, A. Meyer, and S.V. Nepomnyaschikh, Hierarchical extension operators and local multigrid methods in domain decomposition preconditioners, East West J.
Numer. Math. 31 (1994), no. 45, 269272.
[76] G. Haase, U. Langer, S. Reitzinger, and J. Schoberl, Algebraic multigrid methods based on
element preconditioning, Int. J. Comput. Math. 78 (2001), no. 4, 575598.
[77] G. Haase and E. Lindner, Advanced solving techniques in optimization of machine components, Comput. Assist. Mech. Eng. Sci. 6 (1999), no. 3, 337343.
[78] W. Hackbusch, Multigrid methods and applications, Springer, Berlin, 1985.
[79] W. Hackbusch and S. A. Sauter, Composite finite elements for problems containing small
geometric details - part II: Implementation and numerical results, Comput. Vis. Sci. 1
(1997), 1525.
[80]
, Composite finite elements for the approximation of PDEs on domains with complicated microstructures, Numer. Math. 75 (1997), 447472.
[81] W. W. Hager, D. W. Hearn, and P. M. Pardalos (eds.), Large scale optimization, Kluwer
Academic Publishers, Dordrecht, 1994.
[82] J. S. Hansen, Z. S. Liu, and N. Olhoff, Shape sensitivity analysis using a fixed basis function
finite element approach, Struct. Multidisc. Optim. 21 (2001), 177196.
[83] J. Haslinger and R. A. E. Makinen, Introduction to shape optimization. Theory, approximation, and computation, Advances in Design and Control, vol. 7, SIAM, Philadelphia,
2003.
[84] J. Haslinger, M. Miettinen, and D. Panagiotopoulos, Finite element method for hemivariational inequalities. Theory, methods and applications, Nonconvex Optimization and its
Applications, vol. 35, Kluwer Academic Publishers, Dordrecht, 1999.
[85] J. Haslinger and P. Neittaanmaki, Finite element approximation for optimal shape, material
and topology design, 2nd ed., John Wiley & Sons Ltd., Chinchester, 1997.
[86] E. J. Haug, K. K. Choi, and V. Komkov, Sensitivity analysis in structural design, Mathematics in Science and Engineering, vol. 177, Academic Press, 1986.
BIBLIOGRAPHY
134
[87] H. A. Haus and J. R. Melcher, Electromagnetic fields and energy, Englewood Cliffs, Prentice Hall, 1989.
[88] M. R. Hestensen, Optimization theory, John Wiley & Sons, 1975.
, Conjugate direction methods in optimization, Springer, 1980.
[89]
[90] M. R. Hestensen and E. Stiefel, Methods of conjugate gradients for solving linear systems,
J. Res. Natl. Inst. Stand. Technol. 49 (1952), 409436.
[91] R. Hiptmair, Multilevel preconditioning for mixed problems in three dimensions, Ph.D. thesis, University of Augsburg, Germany, 1996.
[92]
[93]
339.
[94] R. Hiptmair and C. Schwab, Natural boundary element methods for the electric field integral equation on polyhedra, SIAM J. Numer. Anal. 40 (2002), no. 1, 6686.
[95] I. Hlavac ek, J. Haslinger, J. Necas, and J. Lovsek, Numerical solution of variational inequalities, Springer, Berlin, 1988.
[96] R. H. W. Hoppe, S. I. Petrova, and V. H. Schulz, Numerical analysis and applications
topology optimization of conductive media described by Maxwells equations, Lecture
Notes Comput. Sci. (1988), 414422.
[97]
, 3d structural optimization in electromagnetics, Proceedings of Domain Decomposition Methods and Applications (Lyon, Oct. 912, 2000) (M. Garbey et al., ed.), 2001.
[98] T. J. R. Hughes, The finite element method, PrenticeHall, Englewood Cliffs, New Jersey,
1987.
[99] N. Ida and J. P. A. Bastos, Electromagnetics and calculation of fields, Springer, 1997.
[100] J. B. Jacobsen, N. Olhoff, and E. Rnholt, Generalized shape optimization of three
dimensional structures using materials with optimum microstuctures, Mechanics of Materials 28 (1998), no. 14, 207225.
[101] C. Johnson, Numerical solution of partial differential equations by the finite element
method, Cambridge University Press, 1990.
[102] M. Jung and U. Langer, Applications of multilevel methods to practical problems, Surv.
Math. Ind. 1 (1991), no. 3, 217257.
[103]
[104] J. Kacur, J. Necas, J. Polak, and J. Soucek, Convergence of a method for solving the magnetostatic field in nonlinear media, Appl. Math. Praha 13 (1968), 456465.
[105] A. M. Kalamkarov and A. G. Kolpakov, Analysis, design and optimization of composite
structures, John Wiley & Sons, New York; London; Sydney, 1997.
BIBLIOGRAPHY
135
[121]
[122]
, Shape optimization of homogeneous electromagnets, Tech. report, Specialforschungsbereich SFB F013, University Linz, 2000.
BIBLIOGRAPHY
136
[123]
[124] D. Lukas, I. Kopriva, D. Ciprian, and J. Pistora, Shape optimization of homogeneous electromagnets and their application to measurements of magnetooptic effects, Records of COMPUMAG 2001, vol. 4, 2001, pp. 156157.
[125] D. Lukas, W. Muhlhuber, and M. Kuhn, An object-oriented library for the shape optimization problems governed by systems of linear elliptic partial differential equations, Transac
, A comparison of three mixed methods for the time dependent Maxwells equations,
SIAM J. Sci. Statist. Comput. 13 (1992), 10971122.
[138]
, A finite elemeny method for approximating the timeharmonic Maxwells equations, Numer. Math. 63 (1992), 243261.
[139] W. Muhlhuber, Efficient solvers for optimal design problems with PDE constraints, Ph.D.
thesis, Johannes Kepler University Linz, Austria, 2002.
BIBLIOGRAPHY
137
[140] F. Murat and J. Simon, Studies on optimal shape design problems, Lecture Notes in Computational Science, vol. 41, Springer, Berlin, 1976.
[141] J. Necas, Les methodes directes en theorie des e quations elliptiques, Academia, Prague,
1967.
[142] J. C. Nedelec, Mixed finite elements in R 3 , Numer. Math. 35 (1980), 315341.
[143]
[144] P. Neittaanmaki, M. Rudnicki, and A. Savini, Inverse problems and optimal design in electricity and magnetism, Oxford University Press, 1995.
[145] P. Neittaanmaki and K. Salmenjoki, Sensitivity analysis for optimal shape design problems,
Struct. Multidisc. Optim. 1 (1989), 241251.
[146] P. Neittaanmaki and J. Saranen, Finite element approximation of electromagnetic fields in
the three dimensional space, Numer. Funct. Anal. Optim. 2 (1981), 487506.
[147]
, Finite element approximations of vector fields given by curl and divergence, Math.
Methods Appl. Sci. 3 (1981), 328335.
, Asymptotic commutativity of differentiation and discretization in shape optimization, Math. and Comput. Modelling 29 (1999), 1937.
[157] J. Petersson, On continuity of the designtostate mappings for trusses with variable topology, Int. J. Eng. Sci. 39 (2001), no. 10, 11191141.
[158] J. Petersson and J. Haslinger, An approximation theory for optimum sheets in unilateral
contact, Q. Appl. Math. 56 (1998), no. 2, 309332.
[159] O. Pironneau, Optimal shape design for elliptic systems, Springer Series in Computational
Physics, Springer, New-York, 1984.
BIBLIOGRAPHY
138
, Algebraic multigrid methods for large scale finite element equations, Ph.D. thesis,
Johannes Kepler University Linz, Austria, 2001.
[176] W. Rudin, Real and complex analysis, McGraw Hill, New York, 1966.
[177]
[178]
BIBLIOGRAPHY
139
[179] Y. Saad, Iterative methods for sparse linear systems, PWS, 1996, 2nd edition available at
http://www.cs.umn.edu/saad/books.html.
[180] M. Save and W. Prager (eds.), Structural optimization, vol. 1, optimality criteria, Plenum
Press, New York, 1985.
[181] M. Save and W. Prager (eds.), Structural optimization, vol. 2, mathematical optimization,
Plenum Press, New York, 1990.
[182] O. Scherzer, An iterative multi level algorithm for solving nonlinear illposed problems,
Numer. Math. 80 (1998), 579600.
[183] M. Schinnerl, U. Langer, and R. Lerch, Multigrid simulation of electromagnetic actuators,
ZAMM, Z. Angew. Math. Mech. 81 (2001), no. 3, 729730.
[184] M. Schinnerl, J. Schoberl, M. Kaltenbacher, U. Langer, and R. Lerch, Multigrid methods for
the fast numerical simulation of coupled magnetomechanical systems, ZAMM Z. Angew.
Math. Mech. 80 (2000), no. 1, 117120.
[185] A. Schleupen, K. Maute, and E. Ramm, Adaptive FEprocedures in shape optimization,
Struct. Multidisc. Optim. 19 (2000), 282302.
[186] J. Schoberl, NETGEN - An advancing front 2D/3D-mesh generator based on abstract rules,
Comput. Vis. Sci. (1997), 4152.
[187] C. Schwab, p and hp finite element methods: Theory and applications to solid and fluid
mechanics, Oxford University Press, 1998.
[188] J. Sethian and A. Wiegmann, Structural boundary design via level set and immersed interface methods, J. Comp. Phys. 163 (2000), 489528.
[189] R. E. Showalter, Hilbert space methods for partial differential equations, Pitman, London,
San Francisco, 1977.
[190] O. Sigmund and J. Petersson, Numerical instabilities in topology optimization: A survey
on procedures dealing with chechkerboards, meshdependencies and local minima, Struct.
Multidisc. Optim. 16 (1998), no. 1, 6875.
[191] C. A. C. Silva and M. L. Bittencourt, An objectoriented structural optimization program,
Struct. Multidisc. Optim. 20 (2000), 154166.
[192] P. P. Silvester and R. L. Ferrari, Finite elements for electrical engineers, Cambridge University Press, 1983.
[193] J. Simon, Differentiation with respect to the domain in boundary value problems, Numer.
Funct. Anal. Optimization 2 (1980), 649687.
[194] S. L. Sobolev, Applications of functional analysis in mathematical physics, American Mathematical Society, Providence, 1963.
[195] J. Sokolowski and A. Zochowski, On the topological derivative in shape optimization,
SIAM J. Control Optimization 37 (1999), 12511272.
BIBLIOGRAPHY
140
[196] J. Sokolowski and J.-P. Zolesio, Introduction to shape optimization, Springer Series in Computational Mathematics, no. 16, Springer, Berlin, 1992.
[197] L. Solymar, Lectures on electromagnetic theory, Oxford University Press, 1987.
[198] W. Stadler, Nonexistence of solutions in optimal structural design, Optim. Control Appl.
Methods 7 (1986), 243258.
[199] C. W. Steele, Numerical computation of electric and magnetic fields, Van Nostrand Reynold
Co., New York, 1987.
[200] G. Strang and G. J. Fix, An analysis of the finite element method, Prentice-Hall, 1973.
[201] J. A. Stratton, Electromagnetic theory, Mc GrawHill Book Co., Inc., New York, 1941.
[202] K. Suzuki and N. Kikuchi, A homogenization method for shape and topology optimization,
Comput. Methods Appl. Mech. Eng. 93 (1991), no. 3, 291318.
[203] K. Svanberg, The method of moving asymptotes a new method for structural optimization,
Int. J. Numer. Methods Eng. 24 (1987), 359373.
[204]
[205] B. Szabo and I. Babuska, Finite element analysis, John Wiley & Sons, 1991.
[206] N. Takahashi, Optimization of die press model, Proceedings of the TEAM Workshop in the
Sixth Round (Okayama, Japan), March 1996.
[207] P.-S. Tang and K.-H. Chang, Integration of topology and shape optimization for design of
structural components, Struct. Multidisc. Optim. 22 (2001), 6582.
[208] The MathWorks, Inc., Matlab user manual, 1993.
[209] H. Thomas, L. Zhou, and U. Schramm, Issues of commercial optimization software development, Struct. Multidisc. Optim. 23 (2002), no. 2, 97110.
[210] D. Tscherniak and O. Sigmund, A webbased topology optimization program, Struct. Multidisc. Optim. 22 (2001), no. 3, 179187.
[211] U. van Rienen, Numerical methods in computational electrodynamics. linear systems in
practical applications, Lecture Notes in Computational Science and Engineering, vol. 12,
Springer, Berlin, 2001.
[212] K. Washizu, Variational methods in elasticity and plasticity, Pergamon, Oxford, Great
Britain, 1983.
[213] Y. M. Xie and G. P. Steven, Evolutionary structural optimization, Springer, Berlin, 1997.
[214] J. Yoo and N. Kikuchi, Topology optimization in magnetic fields using the homogenization
design method, Int. J. Numer. Meth. Eng. 48 (2000), 14631479.
[215] K. Yuge, N. Iwai, and N. Kikuchi, Optimization of 2d structures subjected to nonlinear
deformations using the homogenization method, Struct. Multidisc. Optim. 17 (1999), no. 4,
286298.
BIBLIOGRAPHY
141
[216] K. Yuge and N. Kikuchi, Optimization of a frame structure subjected to a plastic deformation, Struct. Multidisc. Optim. 10 (1995), no. 34, 197208.
[217] O. C. Zienkiewicz, The finite element method in engineering science, London etc.:
McGraw-Hill XIV, 1971.
[218] O. C. Zienkiewicz and R. L. Taylor, The finite element method (parts 13), Butterworth
Heinemann, Oxford, 2000.
[219] M. Zlamal, On the finite element method, Numer. Math. 12 (1968), 394409.
[220] J.-P. Zolesio, The material derivative (or speed) method for shape optimization, Optimization of Distributed Parameter Structures, Part II (E. J. Haug and J. Cea, eds.), NATO Adv.
Study Inst., 1981, pp. 10891151.
[221] G. Zoutendijk, Methods of feasible directions, Elsevier, Amsterdam, 1960.
[222] J. Zowe, M. Kocvara, and M. P. Bendse, Free material optimization via mathematical
programming, Math. Program. 79 (1997), no. 13, Ser. B, 445466.
[223] A. K. Zvedin and V. A. Kotov, Modern magnetooptics and magnetooptical materials, Institute of Physics Publishing Bristol and Philadelphia, 1997.
142
BIBLIOGRAPHY
Curriculum vitae
Personal data
Name:
Date and place of birth:
Education
1990 - 1994:
1994 - 1999:
since November 1999:
VSBTechnical
University of Ostrava, Czech Republic
Career history
August 1996:
AprilJuly 1999:
April 2000March 2001:
VSBTechnical
University of Ostrava, Czech Republic
research fellow at the SFB013 Numerical and Symbolic Scientific Computing, Project F1309 Multilevel Solvers for Large
Scale Discretized Optimization Problems, University Linz, Austria
Publications
M.Sc. thesis [127], 1 article to appear in a refereed journal [120], 2 articles in refereed conference
proceedings [123, 124], 4 refereed technical reports [121, 122, 125, 126], 5 contributed talks and
3 posters at international conferences, and 1 tutorial of a student conference project [119]
Research interests
Shape optimization, finite element method, mathematical modelling, scientific computing, magnetostatics
For more information see: http://lukas.am.vsb.cz
143