Topics in Mathematical Modeling

topics-in-mathematical-modeling 2008/12/5 8:30 page i #1
Jind rich Ne cas Center for Mathematical Modeling

Lecture notes
Volume 4
Topics in
mathematical
modeling
Volume edited by M. Benes and E. Feireisl
topics-in-mathematical-modeling 2008/12/5 8:30 page ii #2
topics-in-mathematical-modeling 2008/12/5 8:30 page iii #3
Lecture notes
Volume 4
Editorial board
Michal Benes
Pavel Dr abek
Eduard Feireisl
Miloslav Feistauer
Josef M alek
Jan Mal y
S arka Ne casov a
Jir Neustupa
Antonn Novotn y
Kumbakonam R. Rajagopal
Hans-Georg Roos
Tom as Roub cek
Daniel

Sev covi c
Vladimr

Sver ak
Managing editors
Petr Kaplick y
Vt Prusa
topics-in-mathematical-modeling 2008/12/5 8:30 page iv #4
topics-in-mathematical-modeling 2008/12/5 8:30 page v #5
Jindrich Necas Center for Mathematical Modeling
Lecture notes
Topics in mathematical modeling
Masato Kimura
Faculty of Mathematics
Kyushu University
Hakozaki 6-10-1, Higashi-ku, Fukuoka 812-8581
Japan
Philippe Laurencot
Institut de Mathematiques de Toulouse, CNRS UMR 5219
Universite de Toulouse
F31062 Toulouse cedex 9
France
Shigetoshi Yazaki
Faculty of Engineering
University of Miyazaki
1-1 Gakuen Kibanadai Nishi, Miyazaki 889-2192
Japan
Volume edited by M. Benes and E. Feireisl
topics-in-mathematical-modeling 2008/12/5 8:30 page vi #6
2000 Mathematics Subject Classication. 35-06, 49Q10, 53C44, 35B40, 34A34
Key words and phrases. mathematical modeling, shape derivative, moving
interfaces, diusive HamiltonJacobi eqautions, crystalline curvature ow
equations
Abstract. The volume provides a record of lectures given by the visitors of the Jindrich Necas
Center for Mathematical Modeling in academic year 2007/2008.
All rights reserved, no part of this publication may be reproduced or transmitted in any form
or by any means, electronic, mechanical, photocopying or otherwise, without the prior written
permission of the publisher.
c Jindrich Necas Center for Mathematical Modeling, 2008
c MATFYZPRESS Publishing House of the Faculty of Mathematics and Physics
Charles University in Prague, 2008
ISBN 978-80-7378-060-9
topics-in-mathematical-modeling 2008/12/5 8:30 page vii #7
Preface
This volume of the Lecture Notes contains texts prepared by Masato Kimura,
Philippe Lauren cot and Shigetoshi Yazaki. They were long term visiting scientists
at the Necas Center for Mathematical Modeling in the years 2007 and 2008, and
their topics are related each to other.
The Center creates a fruitful and productive environment for its members and
for its visitors. The visitors are invited to prepare mini-courses focused on their
research eld. The texts covering the contents of mini-courses is published in the
Lecture Notes of the Necas Center for Mathematical Modeling.
First part of the volume of the Lecture Notes covers the lecture series of Masato
Kimura on dynamics of hyperplanes in R
n
. The text discusses aspects of formu-
lation for the problems with moving interfaces including the shape derivatives of
energy functionals. The author then derives the motion laws known in this domain
by means of dierential and variational calculus on hypersurfaces. Finally, the
gradient-ow approach is used to put known examples of the hyperplane dynamics
into a general framework.
Second part of the volume contains the lecture notes for the course of Philippe
Lauren cot. This course was devoted to the aspects of diusive HamiltonJacobi
equations which play a key role in several recently studied domains of application
including the dynamics of hyperplanes, phase transitions and computer image pro-
cessing. The text presents results related to the long-term behavior of the solution
of diusive HamiltonJacobi equations as well as results for the solution extinction
in nite time.
Third part of the volume is devoted to the lecture notes for the course of
Shigetoshi Yazaki, which analyzed the motion of closed planar curves by crystalline
curvature with preservation of area. The text explains key issues related to the given
domain and provides insight into the physical application. The author discusses the
extent of numerical solution of the given problem as well.
The presented three lecture notes are suitable for the students wishing to learn
recent advances in the given eld of applied mathematics as well as they can serve as
reviews for general mathematical audience. They also can be seen as vivid examples
of a cooperation established within the Necas Center for Mathematical Modeling.
October 2008 M. Benes
E. Feireisl
vii
topics-in-mathematical-modeling 2008/12/5 8:30 page viii #8
topics-in-mathematical-modeling 2008/12/5 8:30 page ix #9
Contents
Preface vii
Part 1. Shape derivative of minimum potential energy: abstract
theory and applications
Masato Kimura 1
Preface 5
Chapter 1. Shape derivative of minimum potential energy: abstract theory
and applications 7
1. Introduction 7
2. Multilinear map and Frechet derivative 10
3. Minimization problems 12
4. Banach contraction mapping theorem 15
5. Implicit function theorem 17
6. Parameter variation formulas 19
7. Lipschitz deformation of domains 23
8. Potential energy in deformed domains 28
9. Applications 30
Bibliography 37
Part 2. Geometry of hypersurfaces and moving hypersurfaces in R
m
for the study of moving boundary problems
Masato Kimura 39
Preface 43
Chapter 1. Preliminaries 45
1. Notation 45
2. Plane curves 46
3. Parametric representation 47
4. Graph representation and principal curvatures 50
Chapter 2. Dierential calculus on hypersurfaces 53
1. Dierential operators on 53
2. Weingarten map and principal curvatures 54
Chapter 3. Signed distance function 59
ix
topics-in-mathematical-modeling 2008/12/5 8:30 page x #10
x CONTENTS
1. Signed distance function in general 59
2. Signed distance function for hypersurface 60
Chapter 4. Curvilinear coordinates 65
1. Dierential and integral formulas in ^
() 65
Chapter 5. Moving hypersurfaces 69
1. Normal time derivatives 69
2. Signed distance function for moving hypersurface 72
3. Time derivatives of geometric quantities 73
Chapter 6. Variational formulas 75
1. Transport identities 75
2. Transport identities for curvatures 78
Chapter 7. Gradient structure and moving boundary problems 81
1. General gradient ow of hypersurfaces 81
2. Prescribed normal velocity motion 82
3. Mean curvature ow 82
4. Anisotropic mean curvature ow 83
5. Gaussian curvature ow 83
6. Willmore ow 84
7. Volume preserving mean curvature ow 84
8. Surface diusion ow 85
9. HeleShaw moving boundary problem 86
Bibliography 89
Appendix A. 91
1. Adjugate matrix 91
2. Jacobis formula 92
Part 3. Large time behaviour for diusive HamiltonJacobi
equations
Philippe Laurencot 95
Chapter 1. Introduction 99
Chapter 2. Well-posedness and smoothing eects 103
1. Gradient estimates 105
2. Time derivative estimates 108
3. Hessian estimates 111
4. Existence 114
5. Uniqueness 116
Bibliographical notes 117
Chapter 3. Extinction in nite time 119
1. An integral condition for extinction 121
2. A pointwise condition for extinction 123
3. Non-extinction 124
topics-in-mathematical-modeling 2008/12/5 8:30 page xi #11
CONTENTS xi
4. A lower bound near the extinction time 126
Chapter 4. Temporal decay estimates for integrable initial data: = 1 129
1. Decay rates 131
2. Limit values of |u|
1
134
3. Improved decay rates: q (1, q
) 137
Chapter 5. Temporal growth estimates for integrable initial data: = 1 141
1
and |u|
142
2. Growth rates 149
Chapter 6. Convergence to self-similarity 151
1. The diusion-dominated case: = 1 151
2. The reaction-dominated case: = 1 154
Bibliography 159
Appendix A. Self-similar large time behaviour 163
1. Convergence to self-similarity for the heat equation 163
2. Convergence to self-similarity for HamiltonJacobi equations 164
Part 4. An area-preserving crystalline curvature ow equation
Shigetoshi Yazaki 169
Preface 173
Chapter 1. An area-preserving crystalline curvature ow equation:
introduction to mathematical aspects, numerical computations,
and a modeling perspective 175
1. Introduction 175
2. Plane curve 175
3. Moving plane curve 178
4. Curvature ow equations 179
5. Anisotropy 181
6. The Frank diagram and the Wul shape 183
7. Crystalline energy 187
8. Crystalline curvature ow equations 189
9. An area-preserving motion by crystalline curvature 195
10. Scenario of the proof of Theorem 1.50 196
11. Numerical scheme 198
12. Towards modeling the formation of negative ice crystals or vapor
gures produced by freezing of internal melt gures 202
Bibliography 207
Index 211
topics-in-mathematical-modeling 2008/12/5 8:30 page xii #12
xii CONTENTS
Appendix A. 213
1. Strange examples 213
2. A non-concave curve 213
3. Anisotropic inequalityproof of Lemma 1.53 213
topics-in-mathematical-modeling 2008/12/5 8:30 page 1 #13
Part 1
Shape derivative of minimum
potential energy: abstract theory
and applications
Masato Kimura
2000 Mathematics Subject Classication. 49-01, 49Q10
Key words and phrases. shape derivative, fracture mechanics, lipschitz
deformation on domains
Abstract. The text provides an elementary introduction to mathematical
foundation of shape derivative of potential energy and provides several ap-
plications of the theory to some elliptic problems. Among many important
applications the text focus on the energy release rate in fracture mechanics.
Contents
Preface 5
Chapter 1. Shape derivative of minimum potential energy: abstract theory
and applications 7
1. Introduction 7
2. Multilinear map and Frechet derivative 10
3. Minimization problems 12
4. Banach contraction mapping theorem 15
5. Implicit function theorem 17
6. Parameter variation formulas 19
7. Lipschitz deformation of domains 23
8. Potential energy in deformed domains 28
9. Applications 30
Bibliography 37
3
Preface
The principal aim of this short lecture note is to present a mathematical foun-
dation of shape derivatives of potential energies with applications to some elliptic
problems. Among many important applications, in particular, we will focus on the
energy release rate in fracture mechanics. I hope that this work will contribute to
further mathematical understanding of many applications of the shape derivative
including fracture mechanics.
This note was originally prepared for a series of lectures at Graduate School
of Mathematics, Kyushu University in 2005-2006. I also had occasions to give a
lecture at Department of Mathematics, Faculty of Nuclear Sciences and Physical
Engineering, Czech Technical University in Prague in March 2006, and an intensive
lecture at Graduate School of Informatics, Kyoto University in July 2006. I stayed
at Czech Technical University in Prague for one year from April 2007 as a visiting
researcher of the Necas Center for Mathematical Modeling LC06052 nanced by
MSMT. During the stay, I again had a chance to present a part of this work in
the workshop Matematika na Vysok ych

Skolach, Variacn Principy v Matematice
a Fyzice held at Herbertov in September 2007.
Based on the above lectures, this lecture note was completed during my stay
in Prague with a support of NCMM. I would like to express my sincere thanks
to all participants in the above lectures and to the organizers who gave me such
opportunities, in particular, Professor Yuusuke Iso and Dr. Masayoshi Kubo of
Kyoto University. I also would like to be grateful to Professor Michal Benes and
his colleagues in the Mathematical Modelling Group of Faculty of Nuclear Sciences
and Physical Engineering, Czech Technical University in Prague, for their kind
hospitalities and the stimulating research environment during my stay in Prague.
The mathematical results in this work have been obtained in a joint work
[13, 14] with Dr. Isao Wakano of Kyoto University. I wish to acknowledge his
cooperation in this research. Professor Kohji Ohtsuka of Hiroshima Kokusai Gakuin
University introduced me to his preceding work in this research area with fruitful
discussions. I would like to appreciate his kind encouragement during the research.
Masato Kimura
Prague, February 2008
5
CHAPTER 1
Shape derivative of minimum potential energy:
abstract theory and applications
1. Introduction
The shape derivative (or shape variation) is one of the fundamental techniques
in the theories of shape optimization, shape sensitivity analysis, and many free
boundary problems. The importance of these research elds has been greatly in-
creasing for the last few decades, because of their many applications to numerical
simulations of industrial problems. The expansion of such applications of the shape
derivative has been also supplying interesting and challenging problems to mathe-
maticians.
We present a mathematical foundation of shape derivatives of potential ener-
gies related to some elliptic problems. A central application which we keep in mind
is the energy release rate in a cracked domain, which plays an important role in
fracture mechanics. The application of the shape derivative to the energy release
rate was originally studied in a series of works by Ohtsuka [16, 17, 18], and math-
ematical justication of the existence and derivation of some formulas in a general
geometric situation was carried out there. To the works by Ohtsuka, we will give an
alternative aspect with a new mathematical framework, which enables us to have
rened results.
We treat a class of shape derivatives, which includes the energy release rate,
as a variational problem with a parameter in Banach spaces. A new parameter
variational principle and the classical implicit function theorem in Banach spaces
will play key roles in our abstract theory. Similar mathematical approaches to
the shape derivative for various problems are found in Sokolowski and Zolesio [22]
and [15] etc.
We briey give an explanation of the energy release rate in fracture mechanics.
Scientic investigation to understand crack evolution process in elastic body was
originated by Grith [9] and has been studied from various viewpoints in engineer-
ing, physics and mathematics since then. Griths idea in the fracture mechanics is
even now the fundamental theory in modeling and analysis of the crack behaviour.
We here make reference to only very few extended studies from mathematical point
of view, Cherepanov [4], Rice [20], Ohtsuka [16], [17], [18], Ohtsuka-Khludnev [19],
and Francfort-Marigo [6]. For more complete list of crack problems and fracture
mechanics, please see the references in the above papers.
The energy release rate G is a central concept in Griths theory and its various
extended theories such as [6]. According to such theories, we treat crack evolutions
in brittle materials with linear elasticity under a quasi-static situation, in which
7
8 1. SHAPE DERIVATIVE OF MINIMUM POTENTIAL ENERGY
applied boundary loading is supposed to change slowly and any inertial eect can
be ignored. The elastic energy at a xed moment is supposed to be given by
minimization of an elastostatic potential energy. According to the Griths theory,
the surface energy which is required in the crack evolution is supplied by relaxation
of the potential energy along crack growth.
Roughly speaking, the energy release rate is dened as follows. Let
be
a bounded domain in R
n
(n 2), which corresponds to the un-cracked material
under consideration. We assume that a crack exists in
, where is a closure
of an n 1 dimensional hypersurface. The cracked elastic body is represented by
. We consider a virtual crack extension (t) with parameter t [0, T), where
= (0) (t
1
) (t
2
) (0
t
1

t
2
< T).
Under the quasi-static assumption, the elastic potential energy E(t) in (t) :=
(t) is given by
E(t) := min
v
_
\(t)
W(x, v(x), v(x))dx, (1.1)
where v is a possible displacement eld in
(t) with a given boundary condition

and W(x, v(x), v(x)) represents a potential energy density including a body force.
A given displacement eld is imposed on a part of
, and no boundary condition

is imposed for the admissible displacement elds on the other part of (t) including
both sides of (t). In other words, the normal stress free condition is imposed for
the minimizer on (t) implicitly.
The energy release rate Gat t = 0 along the virtual crack extension (t)
0t<T
is given by
G := lim
t+0
E(0) E(t)
[(t) [
. (1.2)
Since E(t) E(0), G 0 follows if the limit exists. The Griths criterion for
the brittle crack extension is given by G G
c
, where G
c
is an energy required to
create a new crack per unit (n 1)-dimensional volume (i.e., length in 2d and area
in 3d) and it is a constant depending on the material property and the position.
Cherepanov [4] and Rice [20] studied the so-called J-integral for straight crack
in two dimensional linear elasticity, which is a path-independent integral expressions
of the energy release rate. Since these works, theoretical and practical studies of
crack evolutions have been much developed by means of such useful mathematical
expression of G in two dimensional case.
While most of these mathematically rigorous results have been restricted to two
dimensional linear elasticity (and often only for straight cracks), Ohtsuka [16][18]
and Ohtsuka-Khludnev [19] developed a mathematical formulation of the energy
release rate for general curved cracks in multi-dimensional linear or semi-linear
elliptic systems. They proved existence of the energy release rate, and obtained its
expression by a domain integral and by generalized J-integral.
Based on the idea in [17], we shall give a new mathematical framework for
shape derivative of potential energy including the energy release rate. Adopting
domain perturbation of Lipschitz class, the shape derivative is treated in an
abstract parameter variation problem in Banach spaces, where is considered as a
parameter belonging to Lipschitz class. Instead of estimating the limit (1.2) directly
as in [17] and [19], we treat it by means of the Frechet derivative.
1. INTRODUCTION 9
In our approach, the shape derivative of minimum potential energy is derived
as a Frechet derivative in a Banach space within an abstract parameter variation
formulas and it is given in a domain integral expression. The key tools in the
abstract parameter variation setting are the implicit function theorem and the
LaxMilgram theorem.
The organization of this lecture note is as follows. In Sections 25, some fun-
damental results from nonlinear functional analysis are systematically described,
particularly on the Frechet derivative and the implicit function theorem. Their
proofs are also given in most cases and some relatively simple proofs are left to
the readers. One of the primary sources of these sections is the introduction of the
book [11]. For more details of the nonlinear functional analysis, see [3] etc.
Based on the tools prepared in the previous sections, several abstract param-
eter variation formulas are established in Section 6. A framework of Lipschitz de-
formation of domains, which includes crack extensions, is established in Section 7,
Minimization problems with a general potential energy in deformed domains are
studied in Section 8 as applications of the abstract parameter variation formulas
of Section 6. Quadratic energy functionals corresponding to second order linear
elliptic equations are treated in Section 9.
We remark that the theorems obtained in Sections 89 include the results
in [17] and [19] under a weaker assumption for regularity of domain perturba-
tion. In [19], they assumed that the domain perturbation t (t) belongs to
C
2
([0, T], W
2,
(R
n
)
n
) and derived the domain integral expression of G, whereas
we prove it in Theorem 1.56 under a weaker assumption C
1
([0, T], W
1,
(R
n
)
n
).
In this note, we use the following notation. For an integer n N, R
n
denotes
the n-dimensional Euclidean space over R. Each point x R
n
is expressed by a
column vector x = (x
1
, , x
n
)
T
, where
T
stands for the transpose of the vector or
matrix.
For a real-valued function f = f(x), its gradient is given by a column vector
f(x) :=
_
f
x
1
(x), ,
f
x
n
(x)
_
T
=
_
_
_
f
x1
(x)
.
.
.
f
xn
(x)
_
_
_,
and its Hessian matrix is denoted by
2
f :=
T
(f) = (
T
f) =
_
_
_
_
2
f
x
2
1

2
f
x1xn
.
.
.
.
.
.
2
f
xnx1

2
f
x
2
n
_
_
_
_
R
nn
,
where we note that
2
does not mean the Laplacian operator in this lecture note.
We denote the Laplacian of f by
f := divf = tr
2
f =
n
i=1
2
f
x
2
i
.
2. Multilinear map and Frechet derivative
Our mathematical treatment for the shape derivative of the potential energy
is based on innite dimensional dierential calculus, i.e. the Frechet derivative in
Banach spaces. Several fundamental facts on multilinear maps and the Frechet de-
rivative in Banach spaces are collected in this section. For the readers convenience,
their denitions and basic theorems are stated with some exercises.
Definition 1.1 (Multilinear map). We suppose that X
k
(k = 1, , n, n N),
X and Y are real Banach spaces.
(1) A map g is called a multilinear map from
n
k=1
X
k
to Y , if g :
n
k=1
X
k

Y and the map [x
k
g(x
1
, , x
n
)] is linear from X
k
to Y for each
k = 1, , n. In the case that X
1
= = X
n
= X, g is called a n-linear
map from X to Y .
1
(2) A multilinear map g is called bounded, if there exists M > 0 such that
|g(x
1
, , x
n
)|
Y
M|x
1
|
X1
|x
n
|
Xn
_
(x
1
, , x
n
)
n
k=1
X
k
_
.
The set of bounded multilinear maps g :
n
k=1
X
k
Y is denoted by
B(X
1
, , X
n
; Y ). It will be shown (see Exercise 1.2) that
B(X
1
, , X
n
; Y ) is a Banach space with the norm
|g|
B(X1, ,Xn;Y )
:= sup
x
k
X
k
|g(x
1
, , x
n
)|
Y
|x
1
|
X1
|x
n
|
Xn
(g B(X
1
, , X
n
; Y )).
The set of bounded n-linear maps from X to Y also becomes a Banach
space and it is denoted by B
n
(X, Y ).
(3) An n-linear map g from X to Y is called symmetric, if g(x
1
, , x
n
) is
invariant under all permutations of x
1
, , x
n
. We write B
sym
n
(X, Y ) :=
g B
n
(X, Y ); g : symmetric.
Exercise 1.2. Prove that B(X
1
, , X
n
; Y ) is a Banach space by using the
natural identication B(X
1
, , X
n
; Y )

= B(X
1
, B(X
2
, , X
n
; Y )).
Similar to the case of linear operators (1-linear maps), the boundedness of an
n-linear map is equivalent to its continuity as shown in the next proposition.
Proposition 1.3. Let X
k
(k = 1, , n, n N), X and Y be Banach spaces.
Then the following three conditions are equivalent.
(1) g is a continuous multilinear map from
n
k=1
X
k
to Y .
(2) g is a multilinear map from
n
k=1
X
k
to Y , and is continuous at the origin
of
n
k=1
X
k
.
(3) g B(X
1
, , X
n
; Y ).
Proof. 1.2. This is trivial.
2.3. If not, for m N, there exist x
m
k
X
k
such that
|g(x
m
1
, , x
m
n
)|
Y
m
n
|x
m
1
|
X1
|x
m
n
|
Xn
. Let us dene
m
k
:= (m|x
m
k
|
X
k
)
1
x
m
k
.
Then |g(
m
1
, ,
m
n
)|
Y
1. Since
m
k
0 in X
k
as m , taking the limit as
1
Multilinear and n-linear are usually called, linear (n = 1), bilinear (n = 2), trilinear
(n = 3), etc.
2. MULTILINEAR MAP AND FR
ECHET DERIVATIVE 11
m , it follows that |g(0, , 0)|
Y
1 from the continuity of g at the origin.
This contradicts the fact g(0, , 0) = 0.
3.1. This is shown by a recursive argument with respect to n = 1, 2, .
Some basic properties and typical examples of the n-linear map will be found
in the following exercises.
Exercise 1.4. If n 2, prove that B
sym
n
(X, Y ) is a closed subspace of B
n
(X, Y ).
Exercise 1.5. Let us consider a bounded bilinear map f
0
B(X, Y ; Z) and
bounded multilinear maps g
1
B(X
1
, , X
n
; X) and g
2
B(Y
1
, , Y
m
; Y ), and
we dene
f(x
1
, , x
n
, y
1
, , y
m
) := f
0
(g
1
(x
1
, , x
n
), g
2
(y
1
, , y
m
)).
Then prove that f B(X
1
, , X
n
, Y
1
, , Y
m
; Z) and
|f|
B(X1, ,Xn,Y1, ,Ym;Z)
|f
0
|
B(X,Y ;Z)
|g
1
|
B(X1, ,XnY )
|g
2
|
B(Y1, ,Ym;Y )
.
Exercise 1.6. Let be a bounded domain in R
n
and let p, q, r [1, ]. For
u L
p
() and v L
q
(), we dene g(u, v)(x) := u(x)v(x) (x ). Prove that
g B(L
p
(), L
q
(); L
r
()) for r (p, q), where
(p, q) =
_
_
pq
p +q
(p, q [1, )),
min p, q (p = or q = ).
n
and let k, m N 0. For
(u
1
, , u
k
) C
m
()
k
, we dene
g(u
1
, , u
k
)(x) := u
1
(x) u
k
(x) (x ).
Prove that g B
sym
k
(C
m
(), C
m
()).
Let us dene the Frechet derivative in Banach spaces together with the Gateaux
derivative.
Definition 1.8 (Frechet derivative). Let X, Y be real Banach spaces, and let
U be an open set in X. A mapping f : U Y is Frechet dierentiable at x
0
U
if there is A B(X, Y ) such that
|f(x
0
+h) f(x
0
) Ah|
Y
= o(|h|
X
) as |h|
X
0.
In this case, we write A = f
(x
0
) or f
x
(x
0
), and f
(x
0
) is called the Frechet de-
rivative of f at x
0
. If f is Frechet dierentiable at any x U and the mapping
U x f
(x) B(X, Y ) is continuous, we say f C

1
(U, Y ).
If f
C
1
(U, B(X, Y )) then f
(x) = (f
(x) B(X, B(X, Y )) = B

2
(X, Y ).
If f
C
0
(U, B
2
(X, Y )), we say f C
2
(U, Y ). In the same manner, we can
dene C
n
(U, Y ) if the nth Frechet derivative f
(n)
C
0
(U, B
n
(X, Y )), and dene
C
(U, Y ) :=
nN
C
n
(U, Y ). In case that Y = R, we simply write C
n
(U) =
C
n
(U, R).
Definition 1.9 (G ateaux derivative). Under the condition of Denition 1.8, for
x
0
U and for (h
1
, h
2
, , h
n
) X
n
(n N), if there is O an open neighbourhood
of the origin of R
n
and [(t
1
, , t
n
) f(x
0
+t
1
h
1
+ +t
n
h
n
)] C
n
(O, Y ), then
we dene nth G ateaux derivative at x
0
as
d
n
f(x
0
, h
1
, h
2
, , h
n
) :=

n
t
1
t
n
f(x
0
+t
1
h
1
+ +t
n
h
n
)
t1==tn=0
.
It is not dicult to show the Gateaux dierentiability from the Frechet dier-
entiability.
Exercise 1.10. For f C
n
(U, Y ) (n 2), prove that there exists any nth
G ateaux derivative d
n
f(x, h
1
, h
2
, h
n
) for all x U and it is given by
d
n
f(x, h
1
, h
2
, , h
n
) = f
(n)
(x)[h
1
, , h
n
], (1.3)
and that f
(n)
(x) B
sym
n
(X, Y ) for x U.
On the other hand, a converse proposition is also valid. A proof of the next
theorem is found in 2.1E of [3], for example.
Theorem 1.11. Let X, Y be real Banach spaces and U be an open set in X.
For n N, a mapping f : U Y belongs to C
n
(U, Y ), if and only if the nth G ateaux
derivative d
n
f(x, h
1
, h
2
, , h
n
) exists for all x U and (h
1
, , h
n
) X
n
and
[(h
1
, , h
n
) d
n
f(x, h
1
, h
2
, , h
n
)] B
n
(X, Y ) (
x U),
and d
n
f C
0
(U, B
n
(X, Y )). Furthermore, (1.3) holds in this case.
We conclude this section with some examples of Frechet derivatives in exercises
below. These examples will help the readers to understand the notion of higher
Frechet and Gateaux derivatives.
Exercise 1.12. Let X, Y be real Banach spaces. For g B
n
(X, Y ) (n N),
we dene f(x) := g(x, , x) Y (x X). Prove that f C
(X, Y ) and f
(k)
= 0
for k n + 1. Moreover, in the case that g B
sym
n
(X, Y ), nd the kth derivative
f
(k)
B
k
(X, Y ) for k = 1, , n.
n
and let C
k
(R) (k
N 0). For u C
m
() (m N 0), we dene f(u)(x) := (u(x)) (x ).
Prove that f C
km
(C
m
(), C
m
()) if k m.
Exercise 1.14. Let X be a real innite dimensional Hilbert space and let
e
k
k=1
X be an orthonormal system. We dene
J(u) :=
k=1
a
2
k
(k
2
a
2
k
), a
k
:= (u, e
k
)
X
(u U := w X; |w|
X
< 1).
Prove that J C
(U) and calculate its Frechet derivatives.

3. Minimization problems
This section is devoted to abstract minimization problems in Banach spaces in
connection with the LaxMilgram theorem. We start our discussion with denitions
of global and local minimizers.
Definition 1.15. Let (S, d) be a metric space and J be a real valued functional
dened on S.
3. MINIMIZATION PROBLEMS 13
(1) u
0
S is called a global minimizer of J in S, if J(u
0
) J(u) for all u S.
(2) u
0
S is called a local minimizer of J, if there exists an open neighbour-
hood O S of u
0
such that u
0
is a global minimizer of J in O.
(3) A local minimizer u
0
of J is called isolated, if there exists > 0 and
J(u
0
) < J(u) for all u S with 0 < d(u, u
0
) < .
Fundamental variational principles in Banach spaces are stated as follows. Its
proof will be left for the readers exercise.
Proposition 1.16. Let X be a Banach space and U be an open subset of X.
We consider a functional J : U R.
(1) If J C
1
(U) and u
0
U is a local minimizer of J, we have J
(u
0
) = 0
X
.
(2) If J C
2
(U) and u
0
U is a local minimizer of J, we have J
(u
0
)[w, w]
0 for all w X.
(3) Suppose that J C
2
(U) and J
(u)[w, w] 0 for all w X and u U.

If u
0
U with J
(u
0
) = 0 X
and U is star-shaped
2
with respect to u
0
then u
0
is a global minimizer of J. Moreover if J
(u
0
)[w, w] = 0 implies
w = 0, then u
0
is a unique global minimizer.
(4) Let J C
1
(U) and u
0
U with J
(u
0
) = 0 X
. If J
(u
0
) B
2
(X, R)
exists and if there exists > 0 and J
(u
0
)[w, w] |w|
2
X
for all w X,
then u
0
is an isolated local minimizer of J.
Exercise 1.17. Prove Proposition 1.16.
Contrary to analogy with nite dimensional cases, the notion of the Gateaux
derivative is not sucient to describe the variational principles in innite dimen-
sional spaces like Proposition 1.16. Validity of the Frechet derivative is apparent
also from the next example.
Exercise 1.18. Let X be a Banach space and U be an open subset of X. We
assume that a functional J : U R and u
0
U satisfy the condition;
For each v X, t = 0 is a local minimizer of f(t) := J(u
0
+tv) in R.
(1) Prove that J
(u
0
) = 0 X
if J C
1
(U).
(2) Prove that J
(u
0
)[w, w] 0 for all w X if J C
2
(U).
(3) Is u
0
a local minimizer of J? (Hint: Study the functional in Exercise 1.14
with u
0
= 0.)
A unique existence of a global minimizer of a convex functional is obtained if
we assume the coercivity condition (1.4) below.
Theorem 1.19. Let X be a real reexive Banach space and let V be a closed
subspace of X. Assume that a real valued functional J C
2
(X) satises the coer-
civity condition:
J
(w)[v, v] |v|
2
X
(
v V,

w X), (1.4)
where is a positive constant. Then the functional J admits a unique global min-
imizer over V (g) := v X; v g V for arbitrary g X. Furthermore, u is
2
U is called star-shaped with respect to u
0
, if (1 s)u
0
+su U for all u U and s [0, 1].
the minimizer of J in V (g) if and only if u V (g) satises
J
(u)[v] = 0 (
v V ). (1.5)
Proof. For arbitrary v, w V (g), we dene f(t) := J((1 t)w + tv),
(0 t 1). Then we have
f
(t) = J
((1 t)w +tv)[v w],

f
(t) = J
((1 t)w +tv)[v w, v w] |v w|

2
X
,
and
J(v) J(w) =
_
1
0
_
f
(0) +
_
s
0
f
(t)dt
_
ds J
(w)[v w] +

2
|v w|
2
X
. (1.6)
Putting w = g, we have
J(v) J(g) |J
(g)|
X
|v g|
X
+

2
|v g|
2
X
.
Solving this quadratic inequality, we have
J(v) J(g)
|J
(g)|
2
X
2
, (1.7)
|v|
X
|g|
X
+
1
_
2|J
(g)|
X
+
_
2([J(g)[ +[J(v)[)
_
. (1.8)
From (1.7), the functional J is bounded from below. Let u
n
V (g) be a
minimizing sequence which attains the inmum of J in V (g). From (1.8), u
n
is bounded in X and there exists a subsequence u

n
which converges weakly in
X.
3
The subsequence is again denoted by u
n
and the weak limit is denoted by u.
Since V (g) is weakly closed
4
, u V (g) follows.
From (1.6) with v = u
n
and w = u,
J(u
n
) J(u) +J
(u)[u
n
u],
follows, and taking limit as n ,
inf
V (g)
J = lim
n
J(u
n
) J(u) +J
(w)[u u] = J(u).
Thus, u is a global minimizer of J in V (g). Since (1.5) holds, from (1.6), we have
J(v) J(u)

2
|v u|
2
X
(
v V (g)), (1.9)
and the uniqueness of the minimizer follows. Conversely, if (1.5) holds, from (1.9),
it follows that u is the minimizer of J.
Theorem 1.19 is related to the well-known LaxMilgram theorem (see [21], for
example).
3
A bounded closed convex subset of a reexive Banach space is sequentially compact with
respect to the weak topology. See [24] V.2 Theorem 1.
4
A convex subset of a Banach space is strongly closed if and only if it is weakly closed. See
Mazurs theorem [24] V.1 Theorem 2.
4. BANACH CONTRACTION MAPPING THEOREM 15
Theorem 1.20 (LaxMilgram theorem). For a reexive Banach space X, let
A B(X, X
) = B(X, B(X, R))

= B
2
(X, R) be selfadjoint, i.e., A
= A. We
assume the coercivity condition:
> 0 s.t.
X
Aw, w)
X
|w|
2
X
(
w X).
Then A becomes a linear topological isomorphism from X to X
, i.e., A is a bijective
bounded linear operator from X to X
and A
1
B(X
, X).
Exercise 1.21. In case that X is a real Hilbert space, prove the LaxMilgram
theorem, applying Theorem 1.19 to the functional J(u) :=
1
2 X
Au, u)
X
X
f, u)
X
for arbitrary xed f X
.
Theorem 1.19 and the LaxMilgram theorem have many important applications
in elliptic partial dierential equations. This is an example of a mixed boundary
value problem.
n
with a Lipschitz boundary
. We suppose that
D
is a relatively open subportion of , and
N
:=
D
with
D
,= . For xed f L
2
() and g
N
L
2
(
N
), we consider
J(v, a) :=
_
_
1
2
a(x)[v(x)[
2
f(x)v(x)
_
dx
_
N
a(x)g
N
(x)v(x)d
(v H
1
(), a C
0
()).
For g H
1
(), we dene closed ane subspaces of H
1
()
V (g) := v H
1
(); v = g on
D
, V := V (0).
(1) Prove that J C
(H
1
() C
0
()).
(2) We dene L B(V, V
) by
V
Lv, z)
V
:=
2
X
J(w, a)[v, z] (v, z V ). Show
that L does not depend on w H
1
() and write down the operator L as
a dierential operator formally.
(3) We x a C
0
() with a(x) > 0 (x ) and x g H
1
(). Applying
Theorem 1.19 to J(, a), prove the unique existence of a global minimizer
of J(, a) over V (g), and obtain the Euler equation (1.5) for it.
(4) Under the assumption of sucient regularity for the minimizer and the
other functions, show that the Euler equation (1.5) gives the following
boundary value problem of an elliptic partial dierential equation
5
:
_
_
div (a(x)u) = f in
u = g on
D
u
= g
N
on
N
.
(1.10)
4. Banach contraction mapping theorem
We discuss the continuity of xed points of contraction mappings with respect
to a parameter in an abstract metric space. The results in this section will be used
for a proof of the implicit function theorem in Section 5.
5
The unique minimizer of J is called a weak (or H
1
) solution of the mixed boundary value
problem (1.10).
Theorem 1.23 (Banach contraction mapping theorem). Let (S, d) be a com-
plete metric space. We suppose that T : S S be a contraction, i.e. there is
(0, 1) such that
d(T(x), T(y)) d(x, y) (
x, y S).
Then there uniquely exists a S such that T(a) = a. Moreover we have
d(a, T
n
(b))
n
d(a, b) (
b S, n N). (1.11)
Proof. For arbitrary b S and m, n N (n < m), since
d(T
m
(b), T
n
(b))
m1
k=n
d(T
k+1
(b), T
k
(b))
m1
k=n
k
d(T(b), b)

n
1
d(T(b), b),
T
n
(b)
n
is a Cauchy sequence in S. The limit of T
n
(b) is denoted by a S. Then
a = T(a) follows from
d(a, T(a)) d(a, T
n
(b)) +d(T
n
(b), T(a))
d(a, T
n
(b)) +d(T
n1
(b), a) 0 as n .
The inequality (1.11) is shown as
d(a, T
n
(b)) = d(T
n
(a), T
n
(b))
n
d(a, b) (
b S, n N),
and the uniqueness of the xed point a immediately follows from it.
We consider a family of uniform contraction mappings. We nd that, if it is
continuous with respect to a parameter, the continuity of the xed points follows.
Theorem 1.24. Let (S, d) be a complete metric space and let (, l) be a metric
space. We suppose that T : S S be a uniform contraction, i.e. there is
(0, 1) such that
d(T(x, ), T(y, )) d(x, y) (
x, y S, ). (1.12)
Then the unique xed point g() S, T(g(), ) = g() exists for all . If
T(x, ) C
0
(, S) for all x S, then g C
0
(, S).
Proof. We x
0
. For arbitrary , we have
d(g(), g(
0
)) = d(T(g(), ), T(g(
0
),
0
))
d(T(g(), ), T(g(
0
), )) +d(T(g(
0
), ), T(g(
0
),
0
)).
Applying (1.12) to the rst term, we have
d(g(), g(
0
))
1
1
d(T(g(
0
), ), T(g(
0
),
0
)), (1.13)
and g() converges to g(
0
)) as
0
from the continuity of T.
5. IMPLICIT FUNCTION THEOREM 17
5. Implicit function theorem
In this section, we prove the implicit function theorem in Banach spaces. We
rst present some results on the dierentiability of inverse operators and xed points
of contraction mappings.
Theorem 1.25. Let X, Y , Z be real Banach spaces, and let U be an open set
in X and A C
k
(U, B(Y, Z)) for some k N 0. We suppose that, for x U,
there exists (x) B(Z, Y ) such that (x) = [A(x)]
1
. Then C
k
(U, B(Z, Y ))
follows, and if k 1 we have
(x) = (x)(A
(x))(x) B(Z, Y ) (
x U,

X).
Proof. First, we assume k = 0. For arbitrary x
0
, x U, we dene Q :=
|(x) (x
0
)|
B(Z,Y )
and Q
0
:= |(x
0
)|
B(Z,Y )
. Since |(x)|
B(Z,Y )
Q+Q
0
, for
arbitrary z Z, we have
|((x) (x
0
))z|
Y
= |(x)(A(x
0
) A(x))(x
0
)z|
Y
(Q +Q
0
)|A(x
0
) A(x)|
B(Y,Z)
Q
0
|z|
Z
,
Hence, Q (Q+Q
0
)Q
0
|A(x
0
) A(x)|
B(Y,Z)
follows. Let x
0
U be xed and let
|x x
0
|
X
be suciently small so that |A(x) A(x
0
)|
B(Y,Z)
(2Q
0
)
1
. Then we
have
1
2
Q Q
2
0
|A(x) A(x
0
)|
B(Y,Z)
0 as x x
0
in X.
Hence, is continuous at x
0
.
Next, we assume k = 1. We x x
0
U and dene R(x) := A(x) A(x
0
)
A
(x
0
)[x x
0
] for x U. Since R(x) = o(|x x
0
|
X
) as x x
0
, we have
(x) (x
0
) + (x)(A
(x
0
)[x x
0
])(x
0
)
= ((x) (x
0
))(A
(x
0
)[x x
0
])(x
0
) (x)R(x)(x
0
)
= o(|x x
0
|
X
) as x x
0
.
It follows that there exists
(x
0
) B(X, B(Z, Y )), and furthermore we have
(x
0
) = (x
0
)(A
(x
0
))(x
0
) for X. Since
C
0
(U, B(X, B(Z, Y ))),
we have proved C
1
(U, B(Z, Y )).
If k = 2, since C
1
and A
C
1
, we have
C
1
and C
2
follows. For
k = 3, 4, , the assertion follows recursively.
Theorem 1.26 ([11]). Let X, Y be real Banach spaces, and let U, V be open
sets in X and Y , respectively. We assume that T C
k
(U V , V ) and V y
T(x, y) V is a contraction on V uniformly in x U, i.e., there exists (0, 1)
such that
|T(x, y
1
) T(x, y
2
)|
Y
|y
1
y
2
|
Y
(
x U,

y
1
, y
2
V ).
For x U, let f(x) V be the unique xed point of T(x, ) in V , i.e. T(x, f(x)) =
f(x). Then f C
k
(U, Y ) follows.
Proof. If k = 0, the assertion follows from Theorem 1.12. In the case k = 1,
we dene K(x) := T
y
(x, f(x)) B(Y ). Since |K(x)|
B(Y )
< 1, there exists
(I K(x))
1
=
m=0
K(x)
m
B(Y ), |(I K(x))
1
|
B(Y )

1
1
.
For xed x
0
U, we dene (x) := f(x) f(x
0
) (x U) and
A
0
:= (I K(x
0
))
1
T
x
(x
0
, f(x
0
)) B(X, Y ),
R(x) := (I K(x
0
))(x) T
x
(x
0
, f(x
0
))[x x
0
] (x U).
Then we have
f(x) f(x
0
) = (x) = A
0
(x x
0
) + (I K(x
0
))
1
R(x) (x U). (1.14)
Since
(x) = T(x, f(x
0
) +(x)) T(x
0
, f(x
0
))
= T
x
(x
0
, f(x
0
))[x x
0
] +T
y
(x
0
, f(x
0
))[(x)] +R(x),
it follows that R(x) = o(|x x
0
|
X
+ |(x)|
Y
) as |x x
0
|
X
0. From (1.14),
we also have (x) = O(|x x
0
|
X
) as |x x
0
|
X
0. From these estimates, there
exists the Frechet derivative of f at x
0
and f
(x
0
) = A
0
. Hence, we have proved
f
(x) = (I K(x))
1
T
x
(x, f(x)) B(X, Y ) (x U).
From Theorem 1.25, (I K())
1
C
0
(U, B(Y )) holds, and we have f C
1
(U, Y ).
For the case k 2, we can prove f C
k
(U, Y ) by a recursive argument for
k = 1, 2, .
Corollary 1.27. Under the condition of Theorem 1.26, we have
|f(x) f(x
0
)|
Y

1
1
|T(x, f(x
0
)) f(x
0
)|
Y
(
x
0
, x U
0
).
Proof. This immediately follows from the inequality (1.13).
The implicit function theorem in Banach spaces is stated as follows.
Theorem 1.28 (Implicit function theorem). Let X, Y , Z be real Banach spaces
and U, V be open sets in X and Y , respectively. We suppose that F : U V Z
and (x
0
, y
0
) U V satisfy the conditions;
(1) F(x
0
, y
0
) = 0.
(2) F C
0
(U V, Z).
(3) F(x, ) C
1
(V, Z) for x U and F
y
is continuous at (x, y) = (x
0
, y
0
).
(4) (F
y
(x
0
, y
0
))
1
B(Z, Y ).
Then there exist a convex open neighbourhood of (x
0
, y
0
), U
0
V
0
U V and
f C
0
(U
0
, V
0
), such that, for (x, y) U
0
V
0
, F(x, y) = 0 if and only if y = f(x).
Moreover, if F C
k
(U V, Z) (k N), then f C
k
(U
0
, V
0
).
Proof. Let us dene
T(x, y) := y F
y
(x
0
, y
0
)
1
F(x, y) Y (x, y) U V.
Then F(x, y) = 0 is equivalent to y = T(x, y).
6. PARAMETER VARIATION FORMULAS 19
For > 0, we dene U() := x U; |x x
0
|
X
< and V () := y
V ; |y y
0
|
Y
< . We x
0
> 0 such that y Y ; |y y
0
|
Y

0
V . Since
F
y
is continuous at (x
0
, y
0
), for arbitrary > 0, there exist (0,
0
] such that
|F
y
(x, y) F
y
(x
0
, y
0
)|
B(Y,Z)
< (
(x, y) U() V ()).

For x U() and y
1
, y
2
V (), from the equality
T(x, y
1
)T(x, y
2
) = F
y
(x
0
, y
0
)
1
_
1
0
F
y
(x
0
, y
0
)F
y
(x, y
1
+s(y
2
y
1
))[y
1
y
2
]ds,
we have
|T(x, y
1
) T(x, y
2
)|
Y
L|y
1
y
2
|
Y
,
where L := |F
y
(x
0
, y
0
)
1
|
B(Z,Y )
.
Since F C
0
(U V, Z), for arbitrary
0
> 0, there exists
0
(0, ] such that,
for x U(
0
), we have |F(x, y
0
)|
Z
<
0
. For (x, y) U(
0
) V (), we have
|T(x, y)y
0
|
Y
= |T(x, y)T(x, y
0
)F
y
(x
0
, y
0
)
1
F(x, y
0
)|
Y
< L|yy
0
|
Y
+L
0
.
Hence, let := (2L)
1
and we choose (0,
0
] as above and x it, and, for
0
:= (4L)
1
, let us choose
0
(0, ] as above. Then, for U
0
:= U(
0
) and
V
0
:= V (/2), we have T(x, y) V
0
for (x, y) U
0
V
0
and
|T(x, y
1
) T(x, y
2
)|
Y

1
2
|y
1
y
2
|
Y
(
x U
0
,

y
1
, y
2
V
0
).
Since T C
0
(U
0
V
0
, V
0
), from Theorem 1.26, there exists f C
0
(U
0
, V
0
) and the
assertion follows.
In the case that F C
k
(U V, Z) (k N), T C
k
(U
0
V
0
, V
0
) holds and
f C
k
(U
0
, V
0
) also follows from Theorem 1.26.
Corollary 1.29. Under the condition of the proof of Theorem 1.28, we have
|f(x) f(x
0
)|
Y
2|F
y
(x
0
, y
0
)
1
|
B(Z,Y )
|F(x, y
0
)|
Z
(
x U
0
).
Proof. This immediately follows from Corollary 1.27.
6. Parameter variation formulas
We establish a theory of parameter variation of abstract potential energy in
Banach spaces in this section. The implicit function theorem and the Lax-Milgram
theorem play essential roles in the arguments. The obtained results will provide
a fundamental framework for our analysis of shape derivative of elliptic potential
energies.
As in the next theorem, we consider a family of minimizers u() of a functional
J(u, ) in a Banach space X, where is a parameter in another Banach space M.
Our aim is to study the Frechet derivatives of u() and J
() = min
u
J(u, ) with
respect to the parameter M. A simple but basic observation in our analysis is
the relation (1.15) below.
Theorem 1.30. Let X and M be real Banach spaces. For |
0
X and an open
subset O
0
M, we consider a real valued functional J : |
0
O
0
R and a map
u : O
0
|
0
. We dene J
() := J(u(), ) for O
0
. We suppose the following
conditions.
(1) J C
0
(|
0
O
0
), J(w, ) C
1
(O
0
) for w |
0
, and
M
J C
0
(|
0

O
0
, M
).
(2) u C
0
(O
0
, X) and u() is a global minimizer of J(, ) in |
0
for each
O
0
.
Then we have J
C
1
(O
0
) and
J
() =
M
J(u(), ) ( O
0
). (1.15)
Remark 1.31. If |
0
is an open set and J C
1
(|
0
, O
0
) and u C
1
(O
0
, X),
the formula (1.15) is easily obtained as follows;
J
() = D
[J(u(), )] =
X
J(u(), )[u
()] +
M
J(u(), ) =
M
J(u(), ),
where the D
denotes the Frechet dierential operator with respect to M and

the last equality follows from
X
J(u(), ) = 0 X
.
Proof of Theorem 1.30. We x
0
O
0
and we dene u
0
:= u(
0
) and
r() := J
() J
(
0
)
M
J(u
0
,
0
)[
0
] ( O
0
).
Since u() is a global minimizer and u C
0
(O
0
, X), if is close to
0
, we have
r() J(u
0
, ) J(u
0
,
0
)
M
J(u
0
,
0
)[
0
] = o(|
0
|
M
),
and
r() J(u(), ) J(u(),
0
)
M
J(u
0
,
0
)[
0
]
=
_
1
0
M
J(u(),
0
+s(
0
))
M
J(u
0
,
0
)[
0
]ds
= o(|
0
|
M
).
It follows that r() = o(|
0
|
M
), and we have (1.15) and J
C
0
(O
0
, M
).
Theorem 1.30 is extended to the following C
k
-version.
Corollary 1.32. Under the condition of Theorem 1.30, we assume that |
0
is open. Let k N. If
M
J C
k
(|
0
O
0
, M
) and u C
k
(O
0
, X), then J

C
k+1
(O
0
).
Proof. This immediately follows from the formula (1.15).
In the above theorem, we have assumed the existence and the regularity of the
minimizer u(), whereas we can derive them from the LaxMilgram theorem if we
assume the coercivity of the functional J.
Theorem 1.33. Let X and M be real Banach spaces and | and O be open
subsets of X and M, respectively. We consider a real valued functional J : |O
R and x
0
O. We assume
(1) J(, ) C
2
(|) for O and
X
J C
0
(| O, X
).
(2) u
0
| satises
X
J(u
0
,
0
) = 0.
(3)
2
X
J is continuous at (w, ) = (u
0
,
0
).
(4) There exists > 0 such that
2
X
J(u
0
,
0
)[w, w] |w|
2
X
for w X.
Then there exist a convex open neighbourhood of (u
0
,
0
), |
0
O
0
| O and
u C
0
(O
0
, |
0
), such that, for O
0
, the following three conditions are equivalent.
6. PARAMETER VARIATION FORMULAS 21
(a) w |
0
is a local minimizer of J(, ).
(b) w |
0
satises
X
J(w, ) = 0.
(c) w = u().
In this case, u() is a global minimizer of J(, ) on |
0
.
Proof. We dene a map F :=
X
J from | O to X
and apply Theo-

rem 1.28 at (w, ) = (u
0
,
0
). From the LaxMilgram theorem (Theorem 1.20),
X
F(u
0
,
0
) =
2
X
J(u
0
,
0
) becomes a linear topological isomorphism from X to
X
. Then, from the implicit function theorem (Theorem 1.28), there exist a convex
open neighbourhood of (u
0
,
0
), |
0
O
0
| O and u C
0
(O
0
, U
0
), such that,
for O
0
, w |
0
satises
X
J(w, ) = 0 if and only if w = u().
From the continuity of
2
X
J at (u
0
,
0
), without loss of generality, (after re-
placing |
0
and O
0
with smaller ones if we need) we can assume that
2
X
J(v, )[w, w]

2
|w|
2
X
(
w X,

(v, ) |
0
O
0
). (1.16)
For O
0
, if w |
0
is a local minimizer of J(, ) in |
0
, the
X
J(w, ) = 0
follows. Conversely, if w |
0
satises
X
J(w, ) = 0, w is a local minimizer in |
0
from the condition (1.16). It also follows from (1.16) that u() is a global minimizer
of J(, ) on |
0
(Proposition 1.16 (3)).
Higher regularity of u() also follows from the implicit function theorem.
Theorem 1.34. Under the condition of Theorem 1.33, we additionally assume
that
X
J C
k
(| O, X
) for some k N. Then u C

k
(O
0
, |
0
).
Proof. The assertion follows from the implicit function theorem (Theorem1.28).
Under the condition of Theorem 1.33, we dene

J
() := J(u(), ) ( O
0
).
As a consequence of Theorem 1.34, a sucient condition for J
C
1
(O
0
) is J
C
1
(|O) and
X
J C
1
(|O, X
). Due to Theorem 1.30, however, the condition
X
J C
1
(| O, X
) is not necessary as shown in the next theorem.

that J C
k
(| O) for some k N, then J
C
k
(O
0
) and it satises (1.15).
Proof. From Theorem 1.30, J
C
1
(O
0
) and (1.15) immediately follows.
Since u C
k1
(O
0
, X) follows from Theorem 1.34, J
C
k
(O
0
) is obtained from
the formula (1.15).
In Theorem 1.35 with k = 1, we have derived J
C
1
(O
0
) without assuming
any dierentiability of u() more than u C
0
(O
0
, X) (this is from Theorem 1.33).
Actually, we have a H older regularity of u.
Proposition 1.36. Under the condition of Theorem 1.33, we additionally as-
sume that J C
1
(| O), then we have
|u() u
0
|
X
= o
_
|
0
|
1/2
M
_
as |
0
|
M
0.
Proof. Let
0
> 0 with v X; |v u
0
|
X

0
|
0
. For h X with
|h|
X
= 1, O
0
and (0,
0
], we have
J(u
0
+h, ) = J(u
0
, ) +
X
J(u
0
, )[h] +
2
_
1
0
(1 s)
2
X
J(u
0
+sh, )[h, h]ds.
X
J(u
0
, )[h] =
X
J(u
0
, )[h]
X
J(u
0
,
0
)[h]
=
_
1
(J(u
0
+h, ) J(u
0
, ))
_
1
0
(1 s)
2
X
J(u
0
+sh, )[h, h]ds
_
1
(J(u
0
+h,
0
) J(u
0
,
0
))
_
1
0
(1 s)
2
X
J(u
0
+sh,
0
)[h, h]ds
_
=
1
_
1
0
M
J(u
0
+h,
0
+t(
0
))
M
J(u
0
,
0
+t(
0
))[
0
]dt
_
1
0
(1 s)
2
X
J(u
0
+sh, )
2
X
J(u
0
+sh,
0
)[h, h]ds
Choosing := |
0
|
1/2
M
, we have
|
X
J(u
0
, )|
X
2() ( O
0
, |
0
|
M

2
0
),
where, for r > 0,
S(r) :=
_
(w, ) X M; |w u
0
|
X
r, |
0
|
M
r
2
_
,
(r) := sup
(w,)S(r)
|
M
J(w, )
M
J(u
0
,
0
)|
M
+ sup
(w,)S(r)
|
2
X
J(w, )
2
X
J(u
0
,
0
)|
B2(X,R)
.
We remark that (r) 0 as r +0. Hence, from Corollary 1.29, we have
|u() u
0
|
X
2
1
|
X
J(u
0
, )|
X
4
1
() = o().
Under the conditions of Theorem 1.33,
2
X
J(u(), ) can be regarded as a linear
topological isomorphism from X to X
from the LaxMilgram theorem. Therefore,

we can dene () B(X
, X) which satises
2
X
J(u(), )[()h, w] = h[w] (
w X,

h X
).
Theorem 1.37. Under the condition of Theorem 1.34 with k = 1,
u
() = ()h
0
() ( O
0
), (1.17)
holds, where h
0
() :=
M
X
J(u(), ) B(M, X
).
Proof. Dierentiating
X
J(u(), ) = 0 X
by , we have
2
X
J(u(), )[u
()] +
M
X
J(u(), ) = 0 B(M, X
).
This is equivalent to (1.17) from the LaxMilgram theorem.
The second Frechet derivative of J
is given by the next formula.

7. LIPSCHITZ DEFORMATION OF DOMAINS 23
that J C
2
(| O) then J
C
2
(O
0
) and it satises
J
()[
1
,
2
] =
2
M
J(u(), )[
1
,
2
]
X
()h
0
()[
1
], h
0
()[
2
])
X
( O
0
,
1
,
2
M).
Proof. Dierentiating the formula (1.15) by and substituting (1.17), we
obtain the formula.
7. Lipschitz deformation of domains
A systematic investigation of domain deformation with Lipschitz domain map-
pings is carried out in this section. The domain mapping method is one of the
important techniques of the shape sensitivity analysis and optimal shape design
theory.
We consider a domain deformation with Lipschitz transform : (),
where is a bounded domain in R
n
(n N) and is R
n
-valued W
1,
function. It
is known that a function in W
1,
is Lipschitz in the following sense. For a function
u : R
k
, we dene
[u[
Lip,
:= sup
x,y,x=y
[u(x) u(y)[
[x y[
.
If [u[
Lip,
< , u is called uniformly Lipschitz continuous on .
6
Proposition 1.39. Let be a domain in R
n
. For u W
1,
(), there is
u C
0
() such that u(x) = u(x) for a.e. x , in other words, we can regard
W
1,
() C
0
(). If is convex, W
1,
() = C
0,1
() as a subset of C
0
().
Moreover, we have
|u|
L
()
= [u[
Lip,
(u W
1,
() C
0
()).
Exercise 1.40. Prove Proposition 1.39. (Hint: W
1,
C
0,1
is shown by
using the Friedrichs mollier. The converse is obtained by direct calculation of the
distributional derivative as a limit of a nite dierence.)
In the following arguments, we x a bounded convex domain
0
R
n
and we
identify W
1,
(
0
, R
n
) with C
0,1
(
0
, R
n
). We denote by
0
the identity map on
R
n
, i.e.,
0
(x) = x (x R
n
).
Proposition 1.41. Suppose that W
1,
(
0
, R
n
) satises [
0
[
Lip,0
< 1.
Then is a bi-Lipschitz transform from
0
to (
0
), i.e. is bijective from
0
onto an open set and and
1
are both uniformly Lipschitz continuous.
Proof. Let :=
0
and := [[
Lip,0
(0, 1). First, we show that (
0
)
is open. We x arbitrary y
0
(
0
) with y
0
= (x
0
), x
0

0
. Let > 0 such
that B
(x
0
)
0
, where B
(x
0
) := x R
n
; [x x
0
[ < . For y B
(1)
(y
0
),
we show that y (
0
). It is easily checked that T() := y () is a uniform
contraction on B
(x
0
). From the contraction mapping theorem (Theorem 1.23),
6
For a Lipschitz domain R
n
and (0, 1], we dene C
0,
() := {u C
0
()
L
();

K > 0, |u(x) u(y)| K|x y|
x, y )}. u C
0,
() is called an uniform
-H older function if (0, 1), and u C
0,1
() is called a uniform Lipschitz function.
there is a xed point x = T(x) = y(x) in B
(x
0
), that is y = (x). Hence, (
0
)
is an open set. Since [(x
1
) (x
2
)[ [x
1
x
2
[ [(x
1
) (x
2
)[ (1 )[x
1
x
2
[
for x
1
, x
2

0
, it follows that is injective and
1
satises uniform Lipschitz
condition on (
0
).
For an arbitrary open set with
0
, we dene
O() :=
_
W
1,
(
0
, R
n
); [
0
[
Lip,0
< 1, ()
0
_
. (1.18)
We note that O() is an open subset of W
1,
(
0
, R
n
). For O(), we denote
by () the deformed domain, i.e., () := (), hereafter.
A virtual crack extension in a cracked elastic body for the energy release rate
(1.2) is treated as follows. Let
be an un-cracked elastic body in R

n
with

0
,
and let
be an initial crack. We assume that is a closure of an (n 1)-

dimensional smooth hypersurface. Then, we consider the energy release rate of a
cracked domain :=
along a virtual crack extension (t) parametrized by

0 t < T, where (t) is closed and satises
= (0) (t
1
) (t
2
) (0
t
1

t
2
< T).
We suppose that the virtual crack extension is expressed by a parametrized
domain mappings (t) O() with (0) =
0
, (t)() = (t), ((t)) =
(t),
and (t)(x) = x in a neighbourhood of
. Without loss of generality, we choose

a parameter t by t = H
n1
((t) ), where H
n1
denotes the (n 1)-dimensional
Hausdor measure.
If the elastic potential energy in the deformed domain () is denoted by
E
(), the energy release rate G along (t)

0t<T
at t = 0 is expressed as
G = E
(
0
)[ (0)],
where E
is the Frechet derivative of E
with respect to and is the t derivative

of . Namely, the calculation and the mathematical justication of the energy
release rate G are reduced to those of the Frechet derivative of E
.
Example 1.42. Let
be a bounded domain in R
2
. We dene
(t) :=
_
(x
1
, 0)
T
R
2
; c
0
x
1
c
1
+t
_
(0 t T),
and assume that (T)
. We consider a virtual crack extension (t)

0t<T
in a cracked domain :=
(0). We suppose that

0
be a bounded convex
domain with

0
and that a cut o function q W
1,
(
0
) with supp(q)
x
; x
1
> c
0
and q(x) = 1 at x = (c
1
, 0)
T
. Then we dene
(t)(x) := x +t q(x)
_
1
0
_
(x
0
, t 0).
If [(t)
0
[
Lip,0
= t [q[
Lip,0
< 1, then (t) is a bi-Lipschitz transform from
onto itself and satises (t)((0)) = (t) and ((t)) =
(t).
Example 1.43. Let =
(0, 1) be a bounded cylindrical domain in R

n
,
where
is a bounded Lipschitz domain in R

n1
. For xed h W
1,
(
), we
consider a domain perturbation of :
(t) := (x
, x
n
)
R; t h(x
) < x
n
< 1 ,
for small [t[ << 1. We x a bounded convex domain
with
. Then,
this Lipschitz boundary deformation is expressed by (t) O() as follows. Let
q W
1,
(
0
) be a cut o function with supp(q) (x
, x
n
)
0
; x
n
< 1 and
q(x
, 0) = 1 for x
. Then we dene
(t)(x) :=
_
x
x
n
+tq(x)h(x
)
_
(x = (x
, x
n
)
T

0
, t 0).
If [(t)
0
[
Lip,0
= t [qh[
Lip,0
< 1, then (t) is a bi-Lipschitz transform from
onto itself and satises ((t)) = (t).

For each deformation O(), we dene a pushforward operator
which
transforms a function v on to a function
v := v
1
on (), if satises
Proposition 1.41. We dene
T
(x) :=
_
j
x
i
(x)
_
i,j
R
nn
, (x
0
),
A() :=
_
T
_
1
L
(
0
, R
nn
), () := det
T
L
(
0
, R).
These Jacobi matrices and Jacobian appear in the pullback of dierentiation and
integration on () to . For a function v on , we have
[(
v)] = A()v a.e in (v W

1,1
()),
_
()
(
v)(y)dy =
_
v(x)()(x)dx (v L
1
()).
These equalities are well known in the case C
1
. However, for W
1,
= C
0,1
,
these are not so trivial. See [5] and [25] etc. for details.
Exercise 1.44. Under the condition of Proposition 1.41, prove that
ess- inf
0
() > 0.
Frechet derivatives of () and A() with respect to are obtained as follows.
Theorem 1.45. Let be an open subset of
0
.
(1) It holds that C
(W
1,
(, R
n
), L
()), and the (n + 1)-th Frechet

derivative of vanishes, i.e.,
(n+1)
= 0 B
n+1
(W
1,
(, R
n
), L
()).
In particular, we have
(
0
)[] = div for W
1,
(, R
n
).
(2) We dene O
0
() := W
1,
(, R
n
); ess- inf
() > 0, which is an
open subset of W
1,
(, R
n
). Then A C
(O
0
(), L
(, R
nn
)) holds
and, in particular, we have the formula:
A
(
0
)[] =
T
for W
1,
(, R
n
).
Proof. Since the determinant is a polynomial of degree n, the proposition
C
(W
1,
(, R
n
), L
()) is clear. For xed W

1,
(, R
n
), we dene
m
ij
(t) :=
ij
+t
j
x
i
L
() (i, j = 1, , n, t R),
where
ij
is Kroneckers delta. Then we have
(
0
)[] =
d
dt
t=0
(
0
+t)
=
d
dt
t=0
det (m
ij
(t)) =
d
dt
t=0
_
Sn
sgn()m
1(1)
(t) m
n(n)
(t)
_
=
d
dt
t=0
(m
11
(t) m
nn
(t)) =
n
i=1
m
11
(0) m
ii
(0) m
nn
(0) = div.
Let the (i, j) component of A() be denoted by a
ij
() L
(). Then we
have a
ij
() =
ij
()/(), where
ij
() is the (i, j) cofactor of
T
, which is
a polynomial of

k
x
l
of degree n 1. Since ess- inf
() > 0 for O
0
(),
a
ij
C
(O
0
(), L
()) follows. For xed W

1,
(, R
n
), dierentiating the
identity
A(
0
+t)(I +t
T
) = I (I: identity matrix of degree n),
by t R at t = 0, we have
A
(
0
)[] +A(
0
)
T
= O.
Since A(
0
) = I, we have A
(
0
)[] =
T
.
We conclude this section with three basic propositions for the pushforward
operator
.
Proposition 1.46. Under the condition of Proposition 1.41, for p [1, ],
is a linear topological isomorphism from L

p
() onto L
p
(()), and a linear
topological isomorphism from W
1,p
() onto W
1,p
(()). More precisely, we have
m()|v|
L
p
()
|
v|
L
p
(())
M()
n
|v|
L
p
()
(v L
p
(), 1 p ),
m()
M()
|v|
L
p
()
|(
v)|
L
p
(())

M()
n1
m()
|v|
L
p
()
(v W
1,p
(), 1 p ),
where
m() := min
_
1, ess- inf
()
_
,
M() := max
_
1, n max
1i,jn
_
_
_
_
j
x
i
_
_
_
_
L
()
_
.
Proof. Since we have
m() ()(x) M()
n
a.e. x ,
the estimates for L
p
norms are obtained. We also have
[v(x)[ M()[(
v)((x))[ a.e. x ,
from the equality v = (
T
)(
v).
7
Note that
ij
is given by a cofactor of
T
and is estimated as [
ij
[ M()
n1
a.e. in . Hence, we have
[(
v)((x))[
M()
n1
m()
[v(x)[ a.e. x .
From these estimates, we obtain the L
p
estimates for the gradients.
Proposition 1.47. Let k N 0, l 0, 1, and p [1, ]. We suppose
that f W
k+l,p
(
0
) if p [1, ), and f C
k+l
(
0
) if p = . Then the mapping
[ f ] belongs to C
k
(O(), W
l,p
()).
Proof. We x a domain
1
with a smooth boundary such that
1

1

0
. Let
be a Friedrichs mollier and dene f
:=

f, where

f is
the zero extension of f to R
n
. Then f
(R
n
) and f
tends to f strongly in
W
k+l,p
(
1
) as +0.
First, we prove the theorem for k = 0. We x
1
O(), and suppose that
O() satises |
1
|
W
1,
(0)
for suciently small xed > 0. Without
loss of generality, we assume that (
1
)
1
and ()
1
. We have
|f
1
f |
W
l,p
()
|f
1
f

1
|
W
l,p
()
+|f

1
f
|
W
l,p
()
+|f
f |
W
l,p
()
.
The last term (and the rst term as its special case) is estimated as
|f
f |
W
l,p
()
C(, p)|f
f|
W
l,p
(())
C(, p)|f
f|
W
l,p
(1)
,
where C(, p) is a positive constant depending only on and p. For the second
term, if p [1, ), we have the estimates
|f

1
f
|
p
L
p
()
=
_
[f
(
1
(x)) f
((x))[
p
dx
_
|f
|
L
(0)
|
1
|
L
(0)
_
p
dx = [[|f
|
p
L
(0)
|
1
|
p
L
(0)
,
and
|(f

1
f
)|
p
L
p
()
=
_
[
T
1
(x)]f
(
1
(x)) [
T
(x)]f
((x))
p
dx
[
T
1
(x)] (f
(
1
(x)) f
((x)))
_
[
T
1
(x)] [
T
(x)]
_
f
((x))
_
p
dx
C(, , p)|
1
|
p
W
1,
(0)
,
where C(, , p) is a positive constant depending only on , and p. For the case
p = , similarly we have
|f

1
f
|
W
1,
()
C
(, )|
1
|
W
1,
(0)
.
7
For x R
n
and n n matrix A = (a
ij
), |Ax| (
P
i,j
a
2
ij
)
1/2
|x| holds, where |x| =
(
P
n
i=1
x
2
i
)
1/2
.
From these estimates, it follows that the mapping [ f ] is continuous at
=
1
. Since
1
is arbitrary in O(), it belongs to C
0
(O(), W
l,p
()).
Let us proceed to the case k = 1. For an arbitrary W
1,
(
0
), we consider
the Gateaux derivative
d
dt
[
t=0
f ( +t) in W
l,p
(). We dene
F(t) := f ( +t) W
l,p
(), F
(t) := f
( +t) W
l,p
().
Then it is easy to see that there exists a > 0 such that
lim
+0
|F
F|
C
1
([a,a],W
l,p
())
= 0,
and that
d
dt
f (+t) = lim
+0
d
dt
f
(+t) = lim
+0
(f
(+t)) = (f (+t)) .
Hence [ f ] is Gateaux dierentiable and the derivative is given by (f )
W
l,p
(). Since [ f ] C
0
(O(), W
l,p
()), in other words, we have
[ f ] C
0
(O(), B(W
1,
(
0
), W
l,p
())). From Theorem 1.11, it follows
that [ f ] belongs to C
1
(O(), W
l,p
()).
For the case k 2, we can prove the assertion in the same way.
Proposition 1.48. We assume W
1,
(
0
, R
n
) with supp() .
(1) If [[
Lip,0
< 1, then =
0
+ is a bi-Lipschitz transform from onto
itself.
(2) For t R with [t[
Lip,0
< 1, we dene a bi-Lipschitz transform (t) =
0
+ t from to itself. Let l 0, 1 and p [1, ]. Suppose that
f W
l,p
() if p [1, ), and f C
l
() W
l,
() if p = . Then
(t)
f f strongly in W
l,p
() as t 0.
Proof. From Proposition 1.41, the claim 1 is clear. For the claim 2, let us x
t
0
> 0 with [t
0
[
Lip,0
< 1. Then, from Proposition 1.46, there exist C > 0 such
that the following inequalities hold for [t[ t
0
,
|(t)
f f|
W
l,p
()
= |(t)
(f f (t))|
W
l,p
()
C|f f (t)|
W
l,p
()
.
Since [ f ] C
0
(O(), W
l,p
()) from Proposition 1.47, we obtain
|f f (t)|
W
l,p
()
= |f
0
f (t)|
W
l,p
()
0,
as t 0.
8. Potential energy in deformed domains
Let
0
be a xed bounded convex open set of R
n
(n 2). We consider an
open set whose closure is contained in
0
. We suppose that v H
1
() be a
scalar valued function which describes a physical state in R
n
. For such v(x),
we introduce the following energy functional:
E(v, ) :=
_
W(x, v(x), v(x))dx,

where
W(, , ) R for (, , )
0
R R
n
,
8. POTENTIAL ENERGY IN DEFORMED DOMAINS 29
is a given energy density function. We assume some suitable regularity condi-
tions and boundedness of its derivatives in the following argument. For sim-
plicity, the partial derivatives of W with respect to , and will be denoted
by
W = (
W
1
, ,
W
n
)
T
, W
=
W
, and
W = (
W
1
, ,
W
n
)
T
, respec-
tively. Moreover, for v H
1
(), we often write W(v(x)) = W(x, v(x), v(x)),
W(v(x)) =
W(x, v(x), v(x)), etc.

We consider the following minimization problem.
Problem 1.49. Let V be a closed subspace of H
1
() with H
1
0
() V H
1
(),
and let V (g) := v H
1
(); v g V for g H
1
(). For given g H
1
(), nd
a local minimizer u of E(, ) in V (g), i.e. u V (g) and there exists > 0 such
that
E(u, ) E(w, ) (
w V (g) with |w u|
H
1
()
< ). (1.19)
If u is a local minimizer, under suitable regularity conditions for W, formally
we obtain the variation formulas;
_
(u(x))v(x) +
W(u(x)) v(x) dx = 0 (
v V ), (1.20)
div [
W(u(x))] +W
(u(x)) = 0 in
For xed and V H
1
() as Problem 1.49, we consider a family of min-
imization problems which is parametrized by O(), where O() is dened
by (1.18).
We dene
V (, g) :=
(V (g)) = v H
1
(());
1
(v) g V
( O(), g H
1
()).
Problem 1.50. For given O() and g H
1
(), nd a local minimizer
u() of E(, ()) in V (, g), i.e. u() V (, g) and there exists > 0 such that
E(u(), ()) E(w, ()) (
w V (, g) with |w u()|
H
1
(())
< ).
(1.21)
In many mathematical models of actual (quasi-)static systems, the potential
energy in a deformed domain should be a local minimum. So, we dene
E
() := E(u(), ()), (1.22)

for local minimizers u().
We note that v V is equivalent to
(v + g) V (, g). Using the formulas

in 7, we have
E(
(v + g), ()) = c(v, ) (v V ),

where
c(v, ) :=

c(v +g, ) (v H
1
()),
c(v, ) :=
_
W ((x), v(x), [A()(x)]v(x)) ()(x)dx (v H

1
()).
We dene v() :=
1
(u()) g V . Then u() is a local minimizer of E(, ())

in V (, g), if and only if v() is a local minimizer of c(, ) in V . The following
theorem is a direct consequence of Theorem 1.30.
Theorem 1.51. Suppose that

c C
1
(H
1
() O()). Let |
0
V and let O
0
be an open subset of O() with
0
O
0
. If v C
0
(O
0
, V ) and v() is a global
minimizer of c(, ) in |
0
, then we have E
C
1
(O
0
) and
E
() = c
(v(), ),
where E
is dened by (1.22) with u() =
(v() +g).
We remark that
E
(
0
) = c
(v(
0
),
0
) =

c
(u(
0
),
0
).
Under the suitable regularity conditions for W(, , ), we have the following for-
mula. For W
1,
(
0
, R
n
),
(u,
0
)[]
=
d
dt
_
W (x +t(x), u(x), [A(

0
+t)(x)]u(x)) (
0
+t)(x)dx
t=0
=
_
W(u) (
W(u))
T
(
T
)u +W(u)div
_
dx. (1.23)
By means of the above domain mappings, we obtain another type of variation
formula for Problem 1.49, so called interior variation which is dierent from (1.20).
Theorem 1.52. Suppose that

c C
1
(H
1
()O()). If u is a local minimizer
of E(, ) in V (g) as in Problem 1.49, then we have
(u,
0
)[] = 0 ( W
1,
(), supp() ).
Proof. For W
1,
() with supp() , we dene (t) =
0
t for t 0.
From Proposition 1.46 and 1.48, if [t[
Lip,0
< 1, the corresponding pushforward
operator (t)
is a linear topological isomorphism from H

1
() onto itself, and
lim
t0
|(t)
u u|
H
1
()
= 0.
Since
E((t)
u, ) E(u, ) = E((0)
u, ),
for 0 t << 1, we obtain
(u,
0
)[] =
d
dt
c(u, (t))
t=0
=
d
dt
E((t)
u, )
t=0
0.
Hence, we have

c
(u,
0
)[] = 0.
9. Applications
Based on the discussions in previous sections, we study the shape derivative of
quadratic elliptic potential energies which correspond to scalar valued linear elliptic
problems of second order. However, our framework is applicable even to vector
valued elliptic systems such as linear elasticity problems and to some semilinear
problems as shown in [17].
9. APPLICATIONS 31
Example 1.53. We consider the following potential energy.
W(, , ) =
1
2
_
T
B() +b()
2
_
f(), (1.24)
E(v, ) =
_
_
1
2
_
(v)
T
B(x)v +b(x)v
2
_
f(x)v
_
dx,
where k N0 and B() is an nn symmetric matrix (which is denoted by R
nn
sym
)
and it satises B = (b
ij
) C
k
(
0
, R
nn
sym
) and

0
> 0 such that
T
B()
0
[[
2
(

0
,

R
n
), and
b C
k
(
0
, R), f W
k,2
(
0
, R), (1.25)
with the condition b() 0 for
0
. We remark that
W(, , ) =
1
2
_
_
n
i,j=1
b
ij
()
i
j
+b()
2
_
_
f(),
W
(, , ) = b() f(),
W(, , ) = B().
We suppose that
D
is a nonempty Lipschitz portion of and that a bounded
trace operator
0
: H
1
() L
2
(
D
) is dened and
V := v H
1
();
0
(v) = 0 ,= 0, (1.26)
holds.
8
Then the minimization problem 1.21 corresponds to the following linear
elliptic boundary value problem.
_
_
div(B(x)u) +b(x)u = f(x) in (),
u = g on
D
,
(Bu) = 0 on
D
,
where is a unit normal vector on .
From the Poincare inequality and the LaxMilgram theorem, there exists a
unique global minimizer u() V (g, ) for O().
The Poisson equation is a special case of Example 1.53.
Example 1.54. We consider the following potential energy which corresponds
to the Poisson equation.
W(, , ) =
1
2
[[
2
f(),
E(v, ) =
_
_
1
2
[v[
2
f(x)v
_
dx,
where k N 0 and f W
k,2
(
0
, R). This is a special case of Example 1.53.
We remark that
W(, , ) = f(), W
(, , ) = f(),
W(, , ) = .
8
Under the condition (1.26), from the well known Poincare inequality, the coercivity of this
functional follows. We remark that, if b() > 0 for
0
, we do not need the condition (1.26).
Under the same boundary condition, the minimization problem 1.21 corresponds
to the following boundary value problem of the Poisson equation:
_
_
u = f(x) in (),
u = g on
D
,
u
= 0 on
D
.
Lemma 1.55. For the potential energy of Example 1.53,

c C
k
(H
1
()O())
and (1.23) holds.
Proof. In the case of Example 1.53,

c becomes
c(v, )
=
_
_
1
2
(A()v)
T
B((x)) (A()v) +
1
2
b((x))v
2
f((x))v
_
()dx
=
_
_
1
2
1
(v, )(x) +
1
2
2
(v, )(x)
3
(v, )(x)
_
()(x)dx
where
1
(v, ) := (A()v)
T
(B) (A()v) ,
2
(v, ) := (b)v
2
,
3
(v, ) :=
(f )v. Under the assumptions, from Theorem 1.45 and Proposition 1.47, we have
[(v, ) A()v] C
(H
1
() O(), L
2
()
n
),
[ B ] C
k
(O(), L
(, R
nn
)),
[(w, B) w
T
Bw] C
(L
2
()
n
L
(, R
nn
), L
1
()).
From these regularities, it follows that
1
C
k
(H
1
() O(), L
1
()). Similarly,
from the following regularities
[ b ] C
k
(O(), L
()),
[(b, v) bv
2
] C
(L
() H
1
(), L
1
()),
[ f ] C
k
(O(), L
2
())
[(f, v) fv] C
(L
2
() H
1
(), L
1
()),
2
and
3
also belong to C
k
(H
1
() O(), L
1
()). Since C
(O(), L
())
from Theorem 1.45, we conclude that

c C
k
(H
1
() O()).
We obtain the following theorem from Lemma 1.55 and Theorem 1.35.
Theorem 1.56. For the potential energy of Example 1.53, E
C
k
(O())
holds, and if k 1 we have
E
(
0
)[] =
_
W(u) (
W(u))
T
(
T
)u +W(u)div
_
dx
( W
1,
(
0
, R
n
)),
9. APPLICATIONS 33
where u is the global minimizer of E(, ) in V (g). In particular, in the case of
Example 1.54, we have
E
(
0
)[] =
_
_
(f )u (
T
u)(
T
)(u) +
_
1
2
[u[
2
fu
_
div
_
dx
( W
1,
(
0
, R
n
)).
We consider the case of Example 1.53. The global minimizer u belongs to
H
1
(), but not to H
2
() in general. The regularity loss is caused by geometrical
singularities such as a crack tip and a corner of the domain boundary, or by an
incompatible point of a mixed boundary condition.
Theorem 1.57. We consider Example 1.53 with k = 1. Let ( be a nonempty
open set in R
n
with a Lipschitz boundary. If there exists a sequence of subdomains
l
in which the GaussGreen formula holds and
1

2
with
_
l=1
l
= ,
and if the global minimizer u belongs to H
2
(
l
() for each l, then we have
E
(
0
)[] = lim
l
_
l
W(u) (
W(u) ) (u ) dH
n1
x
, (1.27)
for W
1,
(
0
, R
n
) with supp() (.
Proof. We dene

l
:=
l
(. Under the conditions, we have
div [
W(u)] +W
(u) = 0 in L
2
(
l
),
which is equivalent to
div(Bu) +bu = f in L
2
(
l
).
From the following equalities:
9
_
e
l
W(u(x))div(x)dx =
_
l
W(u(x))(x) dH
n1
x

_
e
x
[W(u(x))] (x)dx,
x
[W(u(x))] =
W(u(x)) +W
(u(x))u(x) + [
2
u(x)]
W(u(x)) in

l
,
_
e
l
W
(u(x))u(x) (x)dx =
_
e
l
div [
W(u)] u(x) (x)dx

=
_
l
(
W(u) ) (u(x) (x)) dH

n1
x

_
e
W(u)(u(x) (x)) dx
=
_
l
(
W(u) ) (u(x) (x)) dH

n1
x
W(u)
_
(
2
u(x))(x) + (
T
)u
_
dx,
9
In this note, u(x) denotes the gradient of u as a column vector and
2
u(x) denotes the
Hessian matrix.
_
e
x
[W(u(x))] (x)dx
=
_
l
_
W(u(x)) + [
2
u(x)]
W(u(x))
_
(x)dx
=
_
l
(
W(u) ) (u(x) (x)) dH

n1
x
W(u)
_
(
2
u(x))(x) + (
T
)u
_
dx
=
_
l
(
W(u) ) (u(x) (x)) dH

n1
x
+
_
l
_
W(u) dx
T
W(u)(
T
)u
_
dx,
we obtain
E
(
0
)[] =
_
W(u) (
W(u))
T
(
T
)u +W(u)div
_
dx
=
_
\
l
_
W(u) (
W(u))
T
(
T
)u +W(u)div
_
dx
+
_
l
_
W(u) (
W(u))
T
(
T
)u
_
dx
+
_
l
W(u) dH
n1
x

_
x
[W(u)] dx
=
_
\
l
_
W(u) (
W(u))
T
(
T
)u +W(u)div
_
dx
+
_
l
W(u) (
W(u) ) (u ) dH
n1
x
,
where denotes the outward unit normal of
l
. The rst term tends to 0 if l .
Hence, we have the formula (1.27).
Corollary 1.58. In particular, for the case of Example 1.54, we have
E
(
0
)[] = lim
l
_
l
__
1
2
[u[
2
fu
_
(u ) (u )
_
dH
n1
x
. (1.28)
In the absence of singularity, Corollary 1.58 has the following form.
Theorem 1.59. Let be a bounded domain with C
2
-boundary. Under the
condition of Example 1.54, we assume that
D
= with g 0. Then the following
formula holds.
E
(
0
)[] =
1
2
_
[u[
2
dH
n1
x
.
Proof. From the regularity theorem of elliptic boundary value problems, we
have u H
2
(). So, we can choose
l
= . We remark that
u = (u )( ), [u [ = [u[ on ,
9. APPLICATIONS 35
since u and its tangential derivatives vanish on the boundary. Hence we have
E
(
0
)[] =
_
__
1
2
[u[
2
fu
_
(u ) (u )
_
dH
n1
x
=
_
__
1
2
[u[
2
_

_
[u[
2
_

_
dH
n1
x
=
1
2
_
[u[
2
dH
n1
x
.
This theorem can be generalized as follows.

Theorem 1.60. Under the condition of Example 1.54, we assume that
W
1,
(
0
, R
n
) satises
supp() supp( g [
D
) = , supp()
D

D
= ,
( : an open set of R
n
s.t. ( supp() and ( is C
2
-class.
Then the following formula holds.
E
(
0
)[] =
1
2
_
D
[u[
2
dH
n1
x

_
\D
f u dH
n1
x
.
Bibliography
[1] R. A. Adams, Sobolev Spaces. Pure and Applied Mathematics, Vol.65, Aca-
demic Press, New York-London (1975).
[2] G. I. Barenblatt, The mathematical theory of equilibrium of cracks in brittle
fracture. Advances in Appl. Mech., Vol.7 (1962) 55-129.
[3] M. S. Berger, Nonlinearity and Functional Analysis. Lectures on nonlinear
problems in mathematical analysis, Pure and Applied Mathematics. Academic
Press, New York-London (1977).
[4] G. P. Cherepanov, On Crack propagation in continuous media. Prikl. Math.
Mekh. Vol.31, No.3 (1967), 476-493.
[5] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Func-
tions. Studies in Advanced Mathematics, CRC Press, Boca Raton, FL (1992).
[6] G. A. Francfort and J.-J. Marigo, Revisiting brittle fracture as an energy min-
imization problem. J. Mech. Phys. Solids, Vol.46 (1998), 1319-1342.
[7] D. Gilbarg and N. S. Trudinger, Elliptic Partial Dierential Equations of Sec-
ond Order. Second edition, Grundlehren der Mathematischen Wissenschaften,
224. Springer-Verlag, Berlin (1983).
[8] V. Girault and P.-A. Raviart, Finite Element Methods for Navier-Stokes Equa-
tions. Theory and Algorithms. Springer Series in Computational Mathematics,
5. Springer-Verlag, Berlin, (1986).
[9] A. A. Grith, The phenomenon of rupture and ow in solids. Phil. Trans.
Royal Soc. London, A221 (1920), 163-198.
[10] P. Grisvard, Elliptic Problems in Nonsmooth Domains. Monographs and Stud-
ies in Mathematics, 24, Pitman (Advanced Publishing Program), Boston, MA,
(1985).
[11] D. Henry, Geometric Theory of Semilinear Parabolic Equations. Lecture Notes
in Mathematics, vol. 840, Springer Verlag (1981).
[12] G. R. Irwin, Fracture dynamics. Fracturing of Metals, American Society of
Metals, (1948) 147-166.
[13] M. Kimura and I. Wakano, New mathematical approach to the energy release
rate in crack extension. Trans. Japan Soc. Indust. Appl. Math. Vol.16, No.3
(2006), 345-358. (in Japanese)
[14] M. Kimura and I. Wakano, Shape derivative of potential energy and energy
release rate in fracture mechanics. (in preparation)
[15] A. Novruzi and M. Pierre, Structure of shape derivatives. J. Evol. Equ. Vol.2,
No.3 (2002), 365-382.
[16] K. Ohtsuka, Generalized J-integral and three-dimensional fracture mechanics
I. Hiroshima Math. J., Vol.11 (1981), 21-52.
37
38 Bibliography
[17] K. Ohtsuka, Generalized J-integral and its applications I Basic theory .
Japan J. Appl. Math. Vol.2, No.2 (1985), 329-350.
[18] K. Ohtsuka, Generalized J-integral and three-dimensional fracture mechanics
II. Hiroshima Math. J., Vol.16 (1986), 327-352.
[19] K. Ohtsuka and A. Khludnev, Generalized J-integral method for sensitivity
analysis of static shape design. Control & Cybernetics, Vol.29 (2000), 513-533.
[20] J. R. Rice, A path-independent integral and the approximate analysis of strain
concentration by notches and cracks. J. Appl. Mech., Vol.35 (1968), 379-386.
[21] M. Schechter, Principles of Functional Analysis. second edition, Graduate
Studies in Mathematics, Vol.36, American Mathematical Society, Providence,
RI, (2002).
[22] J. Sokolowski and J.-P. Zolesio, Introduction to Shape Optimization. Shape
Sensitivity Analysis. Springer Series in Computational Mathematics, Vol.16,
Springer-Verlag, Berlin, (1992).
[23] I. Wakano, Analysis for stress intensity factors in two dimensional elasticity.
Proc. Japan Acad., Vol.73, Ser.A, No.5 (1997), 86-88.
[24] K. Yosida, Functional Analysis. sixth edition, Springer-Verlag, Berlin-New
York (1980).
[25] W. P. Ziemer, Weakly Dierentiable Functions, Sobolev Spaces and Functions
of Bounded Variation. Graduate Texts in Mathematics, 120, Springer-Verlag,
New York (1989).
Part 2
Geometry of hypersurfaces and
moving hypersurfaces in R
m
for the
study of moving boundary
problems
Masato Kimura
2000 Mathematics Subject Classication. 14Q10, 35R35, 53C44
Key words and phrases. hypersurfaces, moving hypersurfaces, gradient ows
Abstract. The text provides a record of an intensive lecture course on the
moving boundary problems. Keeping the applications to moving boundary
problems in R
2
or R
3
in mind, the text systematically constructs the math-
ematical foundations of hypersurfaces and moving hypersurfaces in R
m
for
m 2. Main basic concepts treated in the text are dierential and inte-
gral formulas for hypersurfaces, geometric quantities and the signed distance
functions for moving hypersurfaces, the variational formulas and the transport
identities, and the gradient structure of moving boundary problems.
Contents
Preface 43
Chapter 1. Preliminaries 45
1. Notation 45
2. Plane curves 46
3. Parametric representation 47
4. Graph representation and principal curvatures 50
Chapter 2. Dierential calculus on hypersurfaces 53
1. Dierential operators on 53
2. Weingarten map and principal curvatures 54
Chapter 3. Signed distance function 59
1. Signed distance function in general 59
2. Signed distance function for hypersurface 60
Chapter 4. Curvilinear coordinates 65
() 65
Chapter 5. Moving hypersurfaces 69
1. Normal time derivatives 69
2. Signed distance function for moving hypersurface 72
3. Time derivatives of geometric quantities 73
Chapter 6. Variational formulas 75
1. Transport identities 75
2. Transport identities for curvatures 78
Chapter 7. Gradient structure and moving boundary problems 81
1. General gradient ow of hypersurfaces 81
2. Prescribed normal velocity motion 82
3. Mean curvature ow 82
4. Anisotropic mean curvature ow 83
5. Gaussian curvature ow 83
6. Willmore ow 84
7. Volume preserving mean curvature ow 84
8. Surface diusion ow 85
9. HeleShaw moving boundary problem 86
Bibliography 89
41
42 CONTENTS
Appendix A. 91
1. Adjugate matrix 91
2. Jacobis formula 92
Preface
From the end of March 2007, I had an occasion to spend one year at De-
partment of Mathematics, Faculty of Nuclear Sciences and Physical Engineering,
Czech Technical University in Prague as a visiting researcher of the Necas Center
for Mathematical Modeling. One of the common research subjects in the Mathe-
matical Modelling Group of Professor Michal Benes in Czech Technical University,
where I visited, is the moving boundary problems related to physics, chemistry,
biology and engineering. This is a note for my intensive lectures on the moving
boundary problems which I gave at Czech Technical University in November 2007.
The principal object of my intensive lectures was to construct a basic mathe-
matical framework for the moving boundary problems and to derive several useful
formulas without assuming any knowledge on the dierential geometry. Most of
the topics in this note are basically included in the course of the classical dier-
ential geometry. However, the standard description of the dierential geometry is
not always suitable for the applied mathematicians who intend to study moving
boundary problems.
Keeping the applications to moving boundary problems in R
2
or R
3
in mind,
we systematically construct the mathematical foundations of hypersurfaces and
moving hypersurfaces in R
m
for m 2. Main basic concepts treated in this note
are dierential and integral formulas for hypersurfaces, geometric quantities and the
signed distance functions for moving hypersurfaces, the variational formulas and the
transport identities, and the gradient structure of moving boundary problems. By
the word geometric, we just mean that the quantity or the operator does not
depend on the choice of the Cartesian coordinate in R
m
.
On the other hand, we had to omit many important topics, such as the con-
struction of the partition of unity, the hight function and the domain mapping
method, the Sobolev spaces on hypersurfaces, the well-posedness of some moving
boundary problems, etc. We are not able to touch on these issues except for some
examples and exercises without proofs.
Finally, I would like to thank Professor Michal Benes and his colleagues for giv-
ing me this opportunity and for attending the lectures. Especially, I am grateful to
Mr. Tom as Oberhuber who carefully read the rst manuscript and gave me several
useful comments on this work.
Masato Kimura
Fukuoka, July 2008
43
CHAPTER 1
Preliminaries
In this chapter, we collect our notation in this lecture note and several basic
facts on the geometries of hypersurfaces in the standard style of geometrical treat-
ment. We start from the well-known Frenets formula for plane curves. In general
dimension, we introduce the standard local parametric representation of a hyper-
surface and dene the surface integral on the hypersurface. In the last section, we
give a denition of the principal curvatures based on the graph representation of
the hypersurface in general dimension.
1. Notation
We use the following notation throughout this note. For an integer m N
with m 2, R
m
denotes the m-dimensional Euclidean space over R. Each point
x R
m
is expressed by a column vector x = (x
1
, , x
m
)
T
, where
T
stands for the
transpose of the vector or matrix.
For two column vectors a = (a
1
, , a
m
)
T
R
m
and b = (b
1
, , b
m
)
T
R
m
,
their inner product is denoted by
a b := a
T
b =
m
i=1
a
i
b
i
,
and [x[ :=

x x. For two square matrices A = (a
ij
) R
mm
and B = (b
ij
)
R
mm
, their componentwise inner product is also denoted by
A: B := tr
_
A
T
B
_
=
m
i=1
m
j=1
a
ij
b
ij
,
where tr stands for the trace of a square matrix.
For a real-valued function f = f(x), its gradient is given by a column vector
f(x) =
x
f(x) :=
_
f
x
1
(x), ,
f
x
m
(x)
_
T
=
_
_
_
f
x1
(x)
.
.
.
f
xm
(x)
_
_
_,
and its transpose is denoted by
T
T
f(x) =
_
f
x
1
(x), ,
f
x
m
(x)
_
.
For a (column) vector-valued function h(x) = (h
1
(x), , h
m
(x))
T
R
m
, its trans-
pose is denoted by h
T
(x) = h(x)
T
= (h
1
(x), , h
m
(x)). The above gradient
45
46 1. PRELIMINARIES
operator acts on h as follows.
h
T
(x) := (h
1
(x), , h
m
(x)) =
_
_
_
h1
x1
(x)
hm
x1
(x)
.
.
.
.
.
.
h1
xm
(x)
hm
xm
(x)
_
_
_ R
mm
.
T
h(x) := (h
T
(x))
T
=
_
_
_
T
h
1
(x)
.
.
.
T
h
m
(x)
_
_
_ =
_
_
_
h1
x1
(x)
h1
xm
(x)
.
.
.
.
.
.
hm
x1
(x)
hm
xm
(x)
_
_
_ R
mm
.
We note that
T
does not mean the divergence operator in this paper. The diver-
gence operator is denoted by
div
x
h(x) = divh(x) :=
m
i=1
h
i
x
i
(x) = tr(h
T
) = tr(
T
h).
This div acts on a column vector.
For a real-valued function f = f(x), its Hessian matrix is denoted by
2
f :=
T
(f) = (
T
f) =
_
_
_
_
2
f
x
2
1

2
f
x1xm
.
.
.
.
.
.
2
f
xmx1

2
f
x
2
m
_
_
_
_
R
mm
,
where we note that
2
does not mean the Laplacian operator in this paper. We
denote the Laplacian of f by
f := divf = tr
2
f =
m
i=1
2
f
x
2
i
.
2. Plane curves
We start from the plane curve. Let be a C
2
-class curve in R
2
parametrized
by a length parameter s 1, where 1 is an interval:
= (s) = (
1
(s),
2
(s))
T
R
2
; s 1, C
2
(1, R
2
), [
(s)[ = 1.
The tangential and normal unit vectors are given by
(s) :=
(s), (s) := (
2
(s),
1
(s))
T
.
Exercise 1.1. Prove that
(s) | (s) and
(s) | (s).
The signed curvature (s) is dened by the formulas
(s) = (s)(s),
(s) = (s)(s).
These are known as the Frenet formula for plane curves. We draw the readers
attention to the above sign convention of and , which we use throughout this
note.
3. PARAMETRIC REPRESENTATION 47
Example 1.2. We consider a circle of radius r > 0: = x R
2
; [x[ = r.
Then we have the following relations:
(s) =
_
r cos
_
s
r
_
, r sin
_
s
r
__
T
(0 s < 2r),
(s) =
_
sin
_
s
r
_
, cos
_
s
r
__
T
, (s) =
_
cos
_
s
r
_
, sin
_
s
r
__
T
(s) =
_
1
r
cos
_
s
r
_
,
1
r
sin
_
s
r
_
_
T
=
1
r
(s), (s) =
1
r
.
Exercise 1.3. For given by a graph = u() ( 1
1
R) in - plane with
u C
2
(1
1
) and
ds
d
> 0, prove that
=
u
()
(1 +[u
()[
2
)
3
2
.
Exercise 1.4. For = () = (
1
(),
2
())
T
R
2
; 1
2
R with
C
2
(1
2
) and
ds
d
> 0, prove that
=

1
()
2
()
1
()
2
()
[
()[
3
.
3. Parametric representation
Let m N and m 2. We consider (m1)-dimensional hypersurfaces embed-
ded in R
m
.
Definition 1.5. Let k N. A subset R
m
is called a C
k
-class hypersurface
in R
m
, if, for each x , there exist,
_
_
_
O: a nonempty bounded domain in R
m
,
| : a nonempty bounded domain in R
m1
,
C
k
(|, R
m
)
(1.1)
such that x O and they satisfy the conditions:
: | O is bijective, and rank(
T
()) = m1 (
|). (1.2)
The triplet (O, |, ) gives a kind of local embedding C
k
-coordinate of the hyper-
surface . We dene
C
k
() := (O, |, ); (O, |, ) satises (1.1) and (1.2) .
If a subset (O
, |
)
M
C
k
() satises
M
O
, then we call
(O
, |
)
M
a local embedding C
k
-coordinate system of . Moreover, if
is compact, we can choose its nite subset (O
j
, |
j
,
j
)
j=1,2, ,J
such that
it is a local C
k
-embedding coordinate system of , too.
Exercise 1.6. Let k N. For a C
k
-class hypersurface in R
m
, suppose that
(O, |, ) and (

O,

|, ) are two triplets belonging to C
k
() with := O

O ,= .
Then, prove that
1
C
k
(
1
(),
1
()).
To represent a hypersurface, there are mainly four ways;
(i) parametric representation by local coordinate (Denition 1.5),
48 1. PRELIMINARIES
(ii) local graph representation (next section),
(iii) level set approach,
(iv) domain mapping method.
Although the domain mapping method is important and widely used, we omit
it in this note except for Section 1.
The level set approach enables us to treat a hypersurface globally in a xed
orthogonal coordinate system. It is useful especially to represent moving hypersur-
faces.
Exercise 1.7. Let O be an open set in R
m
, and let w C
k
(O). Then, for a
xed c R, we dene a (c-)level set of w by
:= x O; w(x) = c.
Prove that is a C
k
m
provided ,= and w(x) ,= 0 for
all x . (Conversely, any C
k
-class hypersurface can be represented as a level set
of a C
k
-class function. An example of such functions is given by the signed distance
function (Section 2).)
Definition 1.8. Let k N and let l be an integer with 0 l k. For a
C
k
m
and a function f dened on , f is called of C
l
-class
if f C
l
(|) for any (O, |, ) C
k
() . The set of all C
l
-class functions on
is denoted by C
l
().
Exercise 1.9. Let k N and let l be an integer with 0 l k. Let
(O
, |
)
M
be a local embedding C
k
-coordinate system of . Suppose that
a function f dened on satises f
C
l
(|
) for all M. Then, prove

that f C
l
().
Let (O, |, ) C
k
(). Then, for each |,

1
(), ,

m1
() are m 1
independent R
m
vectors, which are tangent to at x = () . We denote by
T
x
() the tangent space at x = () :
T
x
() := span
_
1
(), ,

m1
()
_
.
This is an m1 dimensional subspace of R
m
.
Exercise 1.10. Prove that T
x
() does not depend on the choice of local C
k
-
embedding coordinate of which covers x.
We dene the following functions of |:
G() :=
_
T
()
_ _
()
_
R
(m1)(m1)
sym
,
i.e., G() = (g
ij
()), g
ij
=

j
,
g() := det G().

i
() := (
1
(), ,
i1
(),
i+1
(), ,
m
())
T
R
m1
,
() := (
1
(), ,
m
())
T
R
m
,
i
() := (1)
m+i
det
_

i
()
_
.
3. PARAMETRIC REPRESENTATION 49
Theorem 1.11. Under the above conditions, for |, we have ()
T
()
()
and
det
_
(), ()
_
= g() = [()[
2
> 0.
Proof. Since rank(
T
()) = m 1, at least one of

i
is not zero, i.e.,
() ,= 0. From the equality
j
=
m
i=1
j
=
m
i=1
(1)
m+i
det
_

i
_
j
= det
_
(),

j
_
= 0,
for each j = 1, , m1, we obtain () T
()
()
.
We dene J() :=
_
(), ()
_
. Then we have
det J =
m
i=1
(1)
m+i
det
_

i
_
i
=
m
i=1
[
i
[
2
= [[
2
,
and
(det J)
2
= det(J
T
J) = det
__

T
_
_
,
_
_
= det
_
G 0
0
T
[[
2
_
= (det G)[[
2
.
Hence, we obtain the assertions.
Corollary 1.12. We can dene a unit normal vector eld C
k1
(
O, R
m
) by
(x) :=
()
[()[
(x = () O),
and we have
_
g() = det
_
(), (())
_
.
We remark that the tangent space is expressed as
T
x
() = y R
m
; y (x) = 0.
Definition 1.13. A C
k
m
is called oriented, if there
exists a (single-valued) unit normal vector eld on , i.e., C
k1
(, R
m
) with
(x) T
x
()
and [(x)[ = 1 (x ).
Under the conditions of Exercise 1.7, the unit normal vector eld is expressed
as
(x) =
w(x)
[w(x)[
or
w(x)
[w(x)[
(x ),
and becomes an oriented hypersurface (see Lemma 2.1).
We dene the surface integral on as follows. Let k N and let be a C
k
-class
hypersurface in R
m
. For a function f dened on , we consider the surface integral:
_
f dH
m1
=
_
f(x) dH
m1
x
,
50 1. PRELIMINARIES
where H
m1
x
is the m1 dimensional Hausdor measure with respect to x (see [5],
[20] etc.), provided f is H
m1
-integrable. In particular, if supp(f) ( O) and
f L
1
(|) for (O, |, ) C
k
(), we have
_
f(x) dH
m1
x
=
_
U
f(())
_
g() d.
4. Graph representation and principal curvatures
Let be a C
k
m
(k N). We x an arbitrary point
on and assume that it is the origin 0 R
m
without loss of generality. We
choose a suitable orthogonal coordinate (
1
,
2
, ,
m1
, )
T
of R
m
such that the
tangent space and the unit normal vector at 0 are given by
T
0
() =
_
(, 0)
T
R
m
; = (
1
, ,
m1
)
T
R
m1
_
,
(0) = (0, , 0, 1)
T
R
m
.
In a neighborhood of the origin, from the implicit function theorem, is locally
expressed by a graph = u() ( |), where | is a bounded domain of R
m1
with 0 |, and
u C
k
(|),
u(0) = 0 R
m1
.
The unit normal vector of in the neighborhood of the origin is given by
(x) =
1
_
1 +[
u()[
2
_

u()
1
_ _
x =
_

u()
_

_
. (1.3)
Proposition 1.14. Let be a bounded C
k
-class hypersurface (k 1). Suppose
that is contained in a C
k
-class hypersurface

. Then, for f C
l
() (0 l k),
there exists an open set O R
m
with O and

f C
l
(O) such that

f[
= f.
Proof. First, we assume that
supp(f)
:= (, u())
T
R
m
; | (1.4)
in the above local graph representation. Then we can construct

f by

f(, ) :=
f(, u()) in a neighborhood of
. For general f, applying a partition of unity, we

decompose f =
N
j=1
f
j
, where each f
j
satises the condition (1.4), and we dene
f :=
N
j=1
f
j
.
We assume that k 2 hereafter, and denote by
2
u the Hessian matrix of u()
u() =
_

2
u
j
()
_
i,j=1, ,m1
R
(m1)(m1)
sym
.
Definition 1.15 (principal curvatures and directions). Let the eigenvalues
and the eigenvectors of the symmetric matrix
2
u(0) be denoted by
i
R and
e
i
R
m1
(i = 1, , m1) with
u(0)e
i
=
i
e
i
, e
i
e
j
=
ij
.
Then
i
(i = 1, , m 1) are called principal curvatures of at 0 , and
e
i
:= (e
i
, 0)
T
T
0
() R
m
is called a principal direction with respect to
i
.
The sum of the principal curvatures is denoted by :=
m1
i=1

i
, and call mean
4. GRAPH REPRESENTATION AND PRINCIPAL CURVATURES 51
curvature in stead of /(m1) in this lecture note. The Gauss-Kronecker curvature
is also denoted by
g
:=
m1
i=1

i
.
The meaning of the principal curvatures is claried in the next proposition.
Proposition 1.16. Under the conditions of Denition 1.15, for = (
, 0)
T
T
0
() with
R
m1
and [[ = [
[ = 1, we dene a plane curve in - plane

with a coordinate (, )
T
R
2
by
:=
_
(, u(
))
T
R
2
; 1
_
,
where 1 is an open interval with 0 1. Let the curvature of
at (, ) = (0, 0)
be denoted by
. Then, it is given as
= (
)
T
_
u(0)
_

=
m1
i=1
i
[ e
i
[
2
.
In particular,
ei
=
i
holds.
Proof. We obtain the assertions from the equality:
=
d
2
d
2
u(
=0
= (
)
T
_
u(0)
_

The following lemma will be used in proving Theorem 2.10.

Lemma 1.17. Under the conditions of Denition 1.15, let y C
1
(1, ) with
y(0) = 0 and y
(0) = e
i
, where 1 is an open interval with 0 1. Then, we have
d
ds
(y(s))
s=0
=
i
e
i
.
Proof. We denote by
m
(x) the m-th component of (x). At x = (, u())
T
,
we have
(x) =
m
(x)
_

u()
1
_
.
We note that
d
ds
m
(y(s))
s=0
= (0)
d
ds
(y(s))
s=0
=
1
2
d
ds
[(y(s))[
2
s=0
= 0.
We dene (s) R
m1
for small s such that y(s) = ((s), u((s)))
T
. Since
d
ds
(0) =
e
i
, we obtain
d
ds
(y(s))
s=0
=
d
ds
_
m
(y(s))
_

u((s))
1
__
s=0
=
m
(0)
_
_

d
ds
u((s))
s=0
0
_
_
=
_

2
u(0)e
i
0
_
=
i
e
i
.

CHAPTER 2
Dierential calculus on hypersurfaces
In this chapter, we introduce some dierential operators on hypersurfaces with
their basic formulas and dene the Weingarten map, which will play an essential
role in our study of principal curvatures on the hypersurfaces.
Throughout this chapter, we suppose that is an oriented C
k
-class hypersur-
face in R
m
with k 1 and m 2, and that O is an open set of R
m
with O.
1. Dierential operators on
We derive several reduction formulas from the dierential operators in R
m
to
the ones on hypersurfaces in this section.
Lemma 2.1. If f C
1
(O) and f[
= 0, then f(x) T
x
()
.
Proof. We assume that is locally parametrized as x = () R
m
by R
m1
. Since T
x
() =
1
(), ,

m1
()) for x = (), the assertion
follows from
0 =

i
f(()) = (f(x))
T

i
() (i = 1, , m1).
Definition 2.2 (Gradient on ). For f C

1
(), we dene
f(x) :=
x
f(x) (x ),
where
x
:= (I (x)(x)
T
) is the orthogonal projection from R
m
to T
x
(),
and

f C
1
(O) is arbitrary C
1
-extension of f to an open neighborhood of with
f[
= f. From Lemma 2.1,
f does not depend on the choice of

f.
Definition 2.3 (Divergence on ). For h C
1
(, R
m
), we dene
div
h := tr
h
T
.
Definition 2.4 (LaplaceBeltrami operator). Let k 2. For f C
2
(), we
dene
f := div
f.
This
is called the LaplaceBeltrami operator on .

Dierential rules for these dierential operators on are collected in the fol-
lowing three propositions.
Proposition 2.5. We suppose that f C
k
() and h C
k
(, R
m
).
(i)
f C
k1
(, R
m
).
53
54 2. DIFFERENTIAL CALCULUS ON HYPERSURFACES
(ii) div
h C
k1
().
(iii)
f C
k2
(, R
m
) (if k 2).
Proposition 2.6 (Product rules). Suppose that f, g C
1
() and that h
C
1
(, R
m
). Then we have the following formulas on .
(i)
(fg) = (
f)g + (
g)f.
(ii) div
(fh) = f div
h +
f h.
(iii)
(fg) = (
f) g + 2(
f) (
g) +f (
g)
(if k 2 and f, g C
2
()).
Proposition 2.7 (Chain rules). Suppose that f C
1
() and g C
1
(R), and
that y C
1
(R, R
m
) with y(s) for s R. Then we have the following formulas
on .
(i)
d
ds
(f y(s)) =
T
f(y(s)) y
(s) (s R).
(ii)
(g f)(x) = (
f(x)) (g
f(x))) (x ).
Exercise 2.8. Prove the above three propositions.
The reduction formula for div
is given as follows.
Proposition 2.9. For h C
1
(O, R
m
), we have
div
h = divh
h
on .
Proof. From the denition of div
, we obtain
div
h = tr
_
(I
T
)h
T
_
= divh tr
_
h
T
_
= divh
h
,
on .
2. Weingarten map and principal curvatures
We assume that k 2 in this section.
Theorem 2.10. We dene W C
k2
(, R
mm
) by
W(x) :=
T
(x) x .
Then, W(x) is symmetric and
_
_
_
W(x)e
i
=
i
e
i
(i = 1, , m1)
W(x)(x) = 0
holds, where
i
and e
i
are the principal curvatures and the corresponding principal
directions at x with e
i
e
j
=
ij
.
Proof. We x x . The equality W(x)(x) = 0 is clear from the denition
of W. For each principal direction e
i
, there exists y C
1
(1, ) such that y(0) = x
and y
(0) = e
i
, where 1 is an open interval with 0 1. Then, from Lemma 1.17,
we have
i
e
i
=
d
ds
[(y(s))][
s=0
=
_
(y(0))
_
y
(0) = W(x)e
i
. (2.1)
2. WEINGARTEN MAP AND PRINCIPAL CURVATURES 55
Since the eigenvectors of matrix W(x) consists of the orthonormal basis e
1
, ,
e
m1
, (x), it follows that the matrix W(x) is symmetric.
Definition 2.11 (Weingarten map). We consider the linear mapping corre-
sponding to the matrix W(x)
W(x) : R
m
T
x
() R
m
p W(x)p =
m1
i=1
i
(p e
i
)e
i
which does not depend on the choice of the orthogonal coordinate system. We
call it the (extended) Weingarten map at x . If we restrict the map W(x) to
the tangent space T
x
(), this is usually called the Weingarten map or the second
fundamental tensor.
The following corollary is clear from the fact that
1
(x), ,
m1
(x), 0 are
the eigenvalues of W(x).
Corollary 2.12.
(x) = tr W(x) = div
(x) (x ),
g
(x) = det (W(x) +(x)(x)
T
) (x ).
The following proposition is useful to calculate the mean curvature.
Proposition 2.13. If an extension of satises
C
1
(O, R
m
), [ (x)[ = 1 (x O), [
= ,
then = div holds on .
Proof. We x x and = (x). For R, if [[ << 1, then
[ (x )[
2
= 1.
Dierentiating this equality by at = 0, we obtain
0 =
d
d
[ (x )[
2
=0
= 2 (x)
T
_
T
(x)()
_
= 2
T

(x).
Hence, from Proposition 2.9, we have
div (x) = div
(x) +

(x) = div
(x) = (x).
Example 2.14. Under the conditions of Exercise 1.7, if k 2 and is in the

direction from the domain w > c to w < c, then the mean curvature of the level
set is given as
(x) = div
_
w(x)
[w(x)[
_
(x = x O; w(x) = c).
56 2. DIFFERENTIAL CALCULUS ON HYPERSURFACES
Exercise 2.15. If is given by a graph = u() with u C
2
(R
m1
) and mth
component of is positive, i.e.,
m
> 0, then prove that
(x) = div
u()
_
1 +[
u()[
2
_
for x =
_

u()
_
.
Proposition 2.16. For f C
2
(O), we have
f = f +
f

2
f
2
on .
Proof. At a point x , we have
f = div(f) = div
(f) +
T
(
2
f)
= div
f +
f
_
+

2
f
2
=
f + (div
)
f
+
T
_
f
_
+

2
f
2
=
f
f
+

2
f
2
.
The next theorem is an extension of the Frenets formula for plane curves to
R
m
.
Theorem 2.17. Let be the identity map on , i.e., (x) := x for x .
Then we have the following formulas:
T
(x) = I (x)
T
(x) =
x
() (x ),
(x) = (x)(x) (x ).
Proof. We extend to whole R
m
as (x) = (
1
(x), ,
m
(x))
T
:= x. Then
we have
T
= (I
T
)
T
= (I
T
) = I (
1
, ,
m
),
T
= (div
(
1
), , div
(
m
))
= (div
)
T
T
(
T
) = (div
)
T
=
T
.
One of the most important tools in our analysis is the following GaussGreen
formula on .
Theorem 2.18 (GaussGreen formula on ). Let h C
1
(, R
m
) and f
C
1
(). We assume that supp(hf) is compact in .
1
Then we have
_
fdH
m1
=
_
(div
h + h)fdH
m1
.
1
If is compact, this condition is always satised.
2. WEINGARTEN MAP AND PRINCIPAL CURVATURES 57
A proof of this theorem will be given in Section 1.
Corollary 2.19. For f C
1
() and g C
2
() with compact supp(fg), we
have
_
g dH
m1
=
_
g dH
m1
.
CHAPTER 3
Signed distance function
We study the signed distance function for a hypersurface in this chapter. In
the rst section, we show the Lipschitz property of the signed distance function for
a general closed set. In the second section, we give various dierential formulas for
smooth hypersurfaces. The signed distance function has many good properties and
it is a useful mathematical tool in mathematical and numerical analysis for free
boundary problems (see [12, 14, 15, 16] etc.).
1. Signed distance function in general
We begin from the situation without any regularity assumption on .
We suppose that there are open subsets
+
and
+
of R
m
with
+

= ,
and we dene a closed set := R
m
(
+

). Then we dene
d(x) =
_
_
_
dist(x, ) (x
+
),
0 (x ),
dist(x, ) (x
),
where
dist(x, ) := inf
y
[x y[.
This d(x) is called the signed distance function for .
Exercise 3.1. Prove that, for a closed set R
m
and x R
m
, there exists
x such that
[x x[ = min
y
[x y[ = dist(x, ).
Theorem 3.2. The signed distance function is Lipschitz continuous with Lip-
schitz constant 1, i.e.,
[d(x) d(y)[ [x y[ (x, y R
m
).
Proof. We consider the following two cases depending on the value of d(x)d(y).
If d(x)d(y) 0, without loss of generality, we can assume that x and y are both
in
+
. In case that they are both in
, we can prove the assertion in the

same way. Let y satisfy d(y) = [y y[. Since d(x) = dist(x, ), we have
d(x) d(y) = d(x) [y y[ [x y[ [y y[ [x y[.
The other inequality d(y) d(x) [x y[ is obtained in the same way. Hence we
obtain [d(x) d(y)[ [x y[.
59
60 3. SIGNED DISTANCE FUNCTION
If d(x)d(y) < 0, without loss of generality, we can assume that x
+
and
y
. Since
+
and
are open sets, there exists (0, 1) such that z :=

(1 )x +y . Then we have
[d(x) d(y)[ = d(x) d(y) = dist(x, ) + dist(y, ) [x z[ +[z y[ = [x y[.
Corollary 3.3. The signed distance function d is dierentiable a.e. in R

m
,
and its gradient d coincides with the distribution sense gradient of d. Moreover,
d W
1,
(R
m
) and it satises
[d(x)[ 1 a.e. in R
m
.
Proof. This corollary follows from the Rademachers Theorem (see [5] etc.).
Let

d be an another signed distance function for a closed set

= R
m
(
).
We assume that and

are both compact (bounded and closed). The Hausdor
distance between two compact sets and

is dened by
dist
H
(,

) := max
_
max
x
dist(x,

), max
x
dist( x, )
_
.
Theorem 3.4. Under the assumptions, we have dist
H
(,

) |d

d|
L
(R
m
)
.
Proof. For x , since d(x) = 0, we obtain
dist(x,

) = [
d(x)[ = [
d(x) d(x)[ |d

d|
L
(R
m
)
,
and
max
x
dist(x,

) |d

d|
L
(R
m
)
.
In the same way, we obtain the other inequality.
2. Signed distance function for hypersurface
In this section, we assume that is an oriented m1 dimensional hypersurface
embedded in R
m
(m 2) of C
k
-class (k N, k 2). Then, C
k1
(, R
m
)
follows. We dene
X(y, r) := y r(y) R
m
(y , r R), (3.1)
^
(
0
) := X(y, r); y
0
, [r[ < (
0
, > 0).
^
(
0
) := X(y, r); y
0
, 0 < r < (
0
, > 0).
From Theorem 2.10 and 2.17, we have
X(y, r) = I (y)
T
(y) +rW(y) (y , r R). (3.2)
Lemma 3.5. We x a point x
and denote by
i
(i = 1, , m 1) the
principal curvatures of at x
. We dene
+
:= max
1im1
i
,
:= min
1im1
i
,
2. SIGNED DISTANCE FUNCTION FOR HYPERSURFACE 61
:=
_
_
_
+
if
+
> 0
if
+
0,
+
:=
_
_
_
if
< 0
+ if
0.
For any
<
<
+
<
+
, there exists a bounded domain O
R
m
with
x
such that X is C
k1
-dieomorphism from (O
) [
,
+
] to its image
in R
m
.
Proof. For r [
,
+
], 1 +r
i
> 0 (i = 1, , m1) holds. It follows that
rank(I +rW(x
)) = m1 and I +rW(x
) is bijective from T
x
() onto itself.
We consider a local embedding coordinate (O, |, ) C
k
() with x
O,
and we dene
:=
1
(x
) | R
m1
. Let us dene a new coordinate
= (, r)
T
= (
1
, ,
m1
, r)
T
| R R
m
. We dene
(, r) := X((), r) ((, r) | R). (3.3)
Then its Jacobian matrix is given by
) =
_
,

r
_
=
_
X(y, r)
T
(), (y)
_
,
where y = () for |. From (3.2), we obtain
X(y, r)
T
() =
_
I (y)
T
(y) +rW(y)
_
()
= (I +rW(y))
T
(),
and
) =
_
(I +rW(y))
T
(), (y)
_
. (3.4)
Since

i
(
)
i=1, ,m1
is a basis of T
x
(), if r [
,
+
], so is the set of
column vectors of (I +rW(x
))
T
), too. Hence, from
,
m
) =
_
(I +
m
W(x
))
T
), (x
)
_
,
it follows that det
T
,
m
) ,= 0.
We choose a bounded domain |
| such that
and that becomes a

C
k1
-dieomorphism from|
,
+
] onto its image. There also exists a bounded
domain O
R
m
such that O
= (|
). Then, it satises the assertion.

Proposition 3.6. We suppose that
0
is a bounded hypersurface and it satises
0
. Then there exists
0
> 0 such that X is C
k1
-dieomorphism from
0
[
0
,
0
] onto ^
0
(
0
).
In the following argument, for simple description, we suppose the following
condition:
> 0 s.t. X is a C
k1
-dieomorphism from (, ) onto ^
(). (3.5)
Then, we can dene
((x), d(x)) := X
1
(x) (x ^
()). (3.6)
We remark that (x) stands for the perpendicular foot on from x ^
().
The function d(x) is none other than the signed distance function for .
62 3. SIGNED DISTANCE FUNCTION
For simplicity, we often write the perpendicular foot from x ^
() as x :=
(x) . Moreover, for a function f dened on , we dene an extension of f by
f(x) := f( x) (x ^
()).
We remark that
x = x d(x)( x) (x ^
()). (3.7)
We have the following theorem.
Theorem 3.8. Under the above conditions, C
k1
(^
(), R
m
) and d
C
k
(^
()) hold. Furthermore they satisfy the following formulas for x ^
(),
d(x) = ( x), (x) = x d(x)d(x),
2
d(x) =
_
I +d(x)

W
_
1
W, d(x) =
m1
i=1
i
( x)
1 +d(x)
i
( x)
,
T
(x) =
_
I +d(x)

W
_
1
x
=
x
_
I +d(x)

W
_
1
, (3.8)
where x := (x) and

W := W( x).
Proof. From the inverse function theorem, it follows that and d belong to
C
k1
-class.
We dierentiate the equation:
r = d(X(y, r)), (3.9)
by y , then we have
0
T
=
T
d(X(y, r))
T
X(y, r). (3.10)

The principal direction at x corresponding to
i
( x) is denoted by e
i
for i =
1, , m 1. Substituting y = x and r = d(x) to (3.10) and multiplying it to e
i
,
we obtain
0 =
T
d(x)
T
X( x, d(x))e
i
=
T
d(x)(I ( x)
T
( x) +d(x)

W)e
i
=
T
d(x)(1 +d(x)
i
( x))e
i
.
Since 1 +d(x)
i
( x) > 0, we have d(x) e
i
= 0 for i = 1, , m1. On the other
hand, dierentiating (3.9) by r and substituting y = x and r = d(x), we obtain
1 =
T
d(X(y, r))
X
r
(y, r)[
y= x,r=d(x)
=
T
d(x)( x).
Thus we obtain d(x) = ( x) = ((x)). Since and both belong to C
k1
-
class, d C
k1
(^
(), R
m
) and hence d C
k
(^
()) follows. The equation

(x) = x d(x)d(x) also follows from (3.7).
Applying the gradient operator
T
to (3.7), we have
I =
T
(x) ( x)
T
d(x) d(x)
T
( x)
T
(x)
= (I +d(x)

W)
T
(x) +( x)
T
( x).
Since
det(I +d(x)

W) =
m1
i=1
(1 +d(x)
i
( x)) > 0,
we obtain
T
(x) =
_
I +d(x)

W
_
1
x
.
2. SIGNED DISTANCE FUNCTION FOR HYPERSURFACE 63
We remark that (I + d(x)

W)
1
and
x
are commutative because
x

W =

W =
W
x
. Hence nally, we arrive at
2
d(x) =
T
(d(x)) =
T
(((x))) =
T
( x)
T
(x)
=

W
_
I +d(x)

W
_
1
x
=
_
I +d(x)

W
_
1
W
x
=
_
I +d(x)

W
_
1
W.
The Laplacian d(x) = tr
2
d(x) is easily calculated from this formula.
Theorem 3.9. If f C
k1
() then

f C
k1
(^
()) and
f(x) = (I +d(x)

W(x))
1
f( x) (x ^
()),
in particular,
f(x) =
f(x) (x ). (3.11)
Moreover, if k 3, then the equation
2

f(x) =
(
T
f)(x) + (x)
T
f(x) W(x) (x ),
holds. In particular,

f(x) =
f(x) (x ).
Proof. Since

f = f ,

f C
k1
(^
()) is clear. From Theorem 3.8, we

have
f(x) =
T
(x)
f( x)
= (I +d(x)

W(x))
1
f( x)
= (I +d(x)

W(x))
1
f( x).
The latter part is shown as follows. To the equation
f(x) =
_
(I +d(x)

W(x))
f(x)
_
T
=
T
f(x)(I +d(x)

W(x)) (x ^
()),
we multiply at x , then, from (3.11), we obtain
(
T
f)(x) = (
T
f)(x)
= (
T
f(x) +d(x)
T
f(x)

W(x))
=
2

f(x) +d(x)
_
f(x)

W(x)
_
+d(x)
_
f(x)

W(x))
_
=
2

f(x) (x)
T
f(x)W(x),
and

f(x) = tr
2

f(x) = tr
_
(
T
f)(x) +(x) (W(x)
f(x))
T
_
= div
f(x) +(x) (W(x)
f(x)) =
f(x).

CHAPTER 4
Curvilinear coordinates
In this chapter, we study various dierential and integral formulas in a curvi-
linear coordinate of ^
(). A proof of the GaussGreens theorem on is also

given.
()
Let be a compact C
k
m
with k 2. We suppose the
condition (3.5) in this section.
For r (, ), we dene
r
:= x ^
(); d(x) = r.
Then, we have
^
() =
_
|r|<
r
.
By d
r
,
r
,
r
and W
r
, we denote the signed distance function, the unit normal
vector, the mean curvature and the Weingarten map, respectively, for
r
. Then we
have the following proposition.
Proposition 4.1. Under the above conditions, for r (, ),
r
is a C
k
-class
hypersurface and
d
r
(x) = d(x) r (x ^
())
r
(x) = ( x) (x
r
)
r
(x) =
m1
i=1
i
( x)
1 +r
i
( x)
(x
r
)
W
r
(x) = (I +r

W)
1

W (x
r
)
Proof. The equality d
r
(x) = d(x) r is clear, and the other assertions follow
from Theorem 3.8.
Theorem 4.2. For f C
0
(
r
), we have
_
r
f(x) dH
m1
x
=
_
f(X(y, r)) det(I +rW(y)) dH

m1
y
. (4.1)
Proof. We x a local coordinate (O, |, ) C
k
(). First, we assume that
supp(f X(, r)) ( O). (4.2)
65
66 4. CURVILINEAR COORDINATES
Then, for dened by (3.3), we have
(, r) = (I +rW(y))
T
().
Similar to Section 3, we dene
G
r
() :=
_
T
(, r)
_ _
(, r)
_
=
_
T
()
_
(I +rW(y))
2
_
()
_
,
g
r
() := det G
r
() = (det (I +rW(y)))
2
det G(),
_
g
r
() = det (I +rW(y))
_
g().
Hence, the surface integral on
r
is given by (4.1) under the condition (4.2). For
general f, applying a partition of unity, we decompose f =
N
j=1
f
j
, where each f
j
satises the condition (4.2).
Theorem 4.3. For x ^
(), we dene
y := (x) , r := d(x) (, ).
Then, we have
dx = det(I +rW(y))dH
m1
y
dr in ^
().
i.e.
_
N
()
f(x)dx =
_

f(X(y, r)) det(I +rW(y)) dH

m1
y
dr,
for f C
0
(^
()). In other words, we have

_
N
()
f(x)dx =
_

_
r
f(x) dH
m1
x
dr .
Proof. From (3.4) and Corollary 1.12, the Jacobian becomes
det
T
(, r) = det (I +rW(y))
_
g().
This gives the formulas of the theorem.
From this theorem, we immediately have the following corollary.
Corollary 4.4. For f C
0
(^
()),
lim
r0
1
2r
_
N
r
()
f(x) dx =
_
f(x) dH
m1
x
,
holds.
We shall prove the GaussGreen theorem on by using Corollary 4.4.
Proof of Theorem 2.18. We decompose the vector eld h = h
0
+ h
1
as
follows.
h
0
(x) :=
x
h(x), h
1
(x) := h(x) h
0
(x) (x ).
1. DIFFERENTIAL AND INTEGRAL FORMULAS IN N
() 67
We have
_
f dH
m1
=
_
h
0

f dH
m1
= lim
r0
1
2r
_
N
r
()
h
0

f dx
= lim
r0
1
2r
_
N
r
()
(div h
0
)

f dx
=
_
(div h
0
)

f dH
m1
=
_
(div
h
0
) f dH
m1
The term div
h
0
is computed as follows:
div
h
0
= div
_
(I
T
)h
_
= div
h (div
)
T
h
T
(
T
h)
= div
h
T
h.
Substituting this to the above equation, we obtain the GaussGreens theorem
on .
CHAPTER 5
Moving hypersurfaces
We extend our formulation to moving hypersurfaces, i.e., time dependent hy-
persurfaces. In the rst section, we introduce a new geometric quantity the normal
velocity and a new derivative the normal time derivative. We give several formu-
las for the signed distance function in Section 2, and for the normal derivatives of
several geometric quantities in Section 3.
1. Normal time derivatives
Let us consider a time dependent hypersurface (t) (t 1), where 1 is an
interval. We assume that each (t) is a nonempty oriented (m 1)-dimensional
hypersurface in R
m
. We dene
/ =
_
tI
((t) t) R
m+1
, (5.1)
and we call / a moving hypersurface in R
m
. The geometric quantities of (t) such
as , and W etc. are denoted by (x, t), (x, t) and W(x, t) etc. for x (t).
In this note, we assume that the moving hypersurface / has the following
regularity.
Definition 5.1 (C
2,1
-class moving hypersurface). An oriented moving hyper-
surface / of the form (5.1) is called of C
2,1
-class, if and only if / is a C
1
-class
m-dimensional hypersurface in R
m
R and C
1
(/, R
m
).
Proposition 5.2. Let /be an oriented moving hypersurface of the form (5.1).
Then, / belongs to C
2,1
-class if and only if / is locally represented by a graph
= u(, t) ( | R
m1
, 1
0
1) in a suitable Cartesian coordinate (, ) of
R
m
with
u C
1
(| 1
0
),
u C
1
(| 1
0
, R
m1
). (5.2)
More precisely, for any (x, t) /, there exist a suitable Cartesian coordinate (, )
of R
m
, and a bounded domain | R
m1
, an open subinterval 1
0
1, a function
u satisfying (5.2), and a bounded domain Q R
m
R, such that (x, t)

O and
x (t); (x, t) Q = (, u(, t)); | (t 1
0
).
In the following arguments, we always assume that / is of C
2,1
class. We call
y a C
1
-trajectory on / if y C
1
(1
0
, R
m
) with y(t) (t) for t 1
0
, where 1
0
is
a subinterval of 1.
Definition 5.4 (Normal velocity). We dene a normal velocity at (x, t) /
by v(x, t) := y
(t) (x, t), where y is a C

1
-trajectory on / through (x, t).
69
70 5. MOVING HYPERSURFACES
Theorem 5.5. The normal velocity v C
0
(/) is well-dened. Moreover, it
satises v(, t) C
1
((t)) and
v C
0
(/, R
m
).
Proof. Let y(t) be a C
1
-trajectory on /. Without loss of generality, we
suppose that y(t) / Q for t 1
0
under the condition of Proposition 5.2.
Then, there exists C
1
(1
0
, |) such that y(t) = ((t), u((t), t))
T
for t 1
0
.
From (1.3) and the equality
y
(t) =
_

(t)
u((t), t)
(t) +u
t
((t), t)
_
, (5.3)
we obtain
v(y(t), t) = (x, t) y
(t) =
u
t
((t), t)
_
1 +[
u((t), t)[
2
.
For xed t 1
0
and x = (, )
T
= y(t) (t), since (t) = |, it follows that
v(x, t) =
u
t
(, t)
_
1 +[
u(, t)[
2
,
and the right hand side does not depend on the choice of the C
1
-trajectory. More-
over, the assertions on the regularity follow from this expression.
Proposition 5.6. Under the condition of Proposition 5.2, we have
v(x, t) =
u
t
(, t)
_
1 +[
u(, t)[
2
(x = (, u(, t))
T
(t)), (5.4)
for | and t 1
0
.
Definition 5.7. A C
1
-trajectory y(t) (t 1
0
) on / is called a normal tra-
jectory on /, if and only if y
(t) T
y(t)
((t))
for t 1
0
.
We remark that y
(t) = v(y(t), t)(y(t), t) holds for any normal trajectory

y(t).
Proposition 5.8. For (x
0
, t
0
) /, there uniquely exists a normal trajectory
on / through (x
0
, t
0
).
Proof. Under the conditions of the proof of Theorem 5.5, y(t) is a normal
trajectory on / through (x
0
, t
0
) if and only if
_
_
_
y
(t) = v(y(t), t) (y(t), t)

y(t
0
) = x
0
.
(5.5)
Substituting (1.3), (5.3) and (5.4), we obtain
_
(t) =
u
t
((t), t)
1 +[
u((t), t)[
2

u((t), t),
(t
0
) =
0
,
(5.6)
where x
0
= (
0
, u(
0
, t
0
)). The system of ordinary dierential equations (5.6) is
equivalent to (5.5) and permits a unique solution since the right hand side of the
rst equation of (5.6) satises the local Lipschitz condition with respect to the
-variable.
1. NORMAL TIME DERIVATIVES 71
Definition 5.9 (Normal time derivative). Let f C
1
(/, R
k
). For (x
0
, t
0
)
/ and the normal trajectory y(t) on / through (x
0
, t
0
), we dene
D
t
f(x
0
, t
0
) :=
d
dt
[f(y(t), t)]
t=t0
,
and D
t
f is called the normal time derivative of f on /.
We denote the identity map on / by (x, t) := x ((x, t) /)). Then we
have
D
t
= v on /.
We also remark that even if a function f = f(x) does not depend on the time
variable t, the normal time derivative of f may not be zero. In general, for f(x, t) =
f(x), we have
D
t
f(x, t) =
d
dt
f(y(t)) = v(x, t) (
T
f(x)) (x, t) (x = y(t) (t)).
Proposition 5.10. Let Q be an open neighbourhood of / in R
m+1
. For
f C
1
(Q),
D
t
f(x, t) = f
t
(x, t) +v(x, t)
f
(x, t) ((x, t) /),

holds, where
f
denotes the partial derivative with respect to x in the direction

(x, t).
Proof. Let y(t) be a normal trajectory on /. Then, at x = y(t) (t), we
have
D
t
f(x, t) =
d
dt
f(y(t), t) =
T
f(x, t) y
(t) +f
t
(x, t)
= v(x, t)
T
f(x, t) (x, t) +f
t
(x, t).
Proposition 5.11. For f C

1
(/) and a C
1
-trajectory y(t) on /, we obtain
d
dt
f(y(t), t) = D
t
f(x, t) +
T
(t)
f(x, t) y
(t) (x = y(t) (t)).

Proof. Let

f C
1
(Q) be an extension of f. Then, for x = y(t) (t), we
obtain
d
dt
f(y(t), t) =
d
dt
f(y(t), t)
=
T
f(x, t) y
(t) +

f
t
(x, t)
=
_
(t)
f(x, t) +(x, t)
(x, t)
_
T
y
(t) +

f
t
(x, t)
=
T
(t)
f(x, t) y
(t) +
_
(x, t)
T
(x, t) y
(t) +

f
t
(x, t)
_
=
T
(t)
f(x, t) y
(t) +D
t
f(x, t).
2. Signed distance function for moving hypersurface

Let /be a C
2,1
-class moving hypersurface of the form (5.1). For simplicity, we
assume that there exists > 0 and (t) satises the condition (3.5) for each t 1.
The signed distance function and the perpendicular foot and other quantities for
(t) are denoted by d(x, t) and (x, t), etc. We dene
^
(/) := (x, t) R
m
R; x ^
((t)), t 1.
Lemma 5.12. Under the above condition, C
1
(^
(/), R
m
) and d C
1
(^
(/))
hold.
Exercise 5.13. Using the implicit function theorem, prove Lemma 5.12.
Theorem 5.14. Under the above condition, d C
1
(^
(/), R
m
) and
C
1
(^
(/), R
m
) hold, and, for (x, t) ^
(/), we have
d
t
(x, t) = v( x, t),
d
t
(x, t) = (I +d(x, t)W( x, t))
1
(t)
v( x, t),
t
(x, t) = v( x, t)( x, t) d(x, t)(I +d(x, t)W( x, t))
1
(t)
v( x, t),
where x = (x, t) (t). Furthermore, if W C
1
(/, R
mm
), then
2
d
C
1
(^
(/), R
mm
) holds and, for (x, t) /, we obtain
2
d
t
(x, t) =
(t)
(
T
(t)
v)(x, t) +(x, t)
T
(t)
v(x, t) W(x, t),
d
t
(x, t) =
(t)
v(x, t).
Proof. Since
d(x, t) = ((x, t), t) ((x, t) ^
(/)),
from Lemma 5.12, we obtain d C
1
(^
(/), R
m
).
Dierentiating the equality
x = (x, t) +d(x, t)d(x, t) ((x, t) ^
(/)),
by t, then we obtain
0 =
t
(x, t) +d
t
(x, t)d(x, t) +d(x, t)d
t
(x, t). (5.7)
Taking an inner product with d(x, t), we get
d
t
(x, t) = d
t
(x, t)[d(x, t)[
2
= d(x, t) (
t
(x, t) +d(x, t)d
t
(x, t))
= ((x, t), t)
t
(x, t) d(x, t)d(x, t) d
t
(x, t)
= v((x, t), t) d(x, t)
1
2
t
[d(x, t)[
2
= v((x, t), t).
Hence, d
t
= v holds and the expression of d
t
follows from Theorem 3.9. The time
derivative of is obtained from (5.7).
In the case W C
1
(/, R
mm
), from the expression
2
d(x, t) =
_
I +d(x, t)

W
_
1
W ((x, t) ^
(/)),
3. TIME DERIVATIVES OF GEOMETRIC QUANTITIES 73
where

W = W((x, t), t),
2
d C
1
(^
(/), R
mm
) holds. The last two equalities
also follow from d
t
= v and Theorem 3.9.
3. Time derivatives of geometric quantities
The normal time derivatives of the geometric quantities of a moving hypersur-
face / are given as follows.
Theorem 5.15. If / belongs to C
2,1
-class,
D
t
=
(t)
v,
holds on /. Furthermore, if W C
1
(/, R
mm
), then the following equalities
hold on /:
D
t
W = vW
2
+
(t)
(
T
(t)
v) +
_
T
(t)
v
_
W,
D
t
= v
m1
i=1
2
i
+
(t)
v,
D
t
g
= v
g
+ adj(W +
T
) :
(t)
(
T
(t)
v).
Proof. Let y(t) be a normal trajectory on /. Then, at x = y(t) (t), we
have
D
t
(x, t) =
d
dt
((y(t), t)) =
d
dt
(d(y(t), t))
=
2
d(x, t)y
(t) d
t
(x, t)
= W(x, t)(x, t)v(x, t) v(x, t)
= v(x, t) =
(t)
v(x, t).
The second equality is similarly calculated as follows.
D
t
W(x, t) =
d
dt
_
2
d(y(t), t)
_
= v(x, t)

2
d(x, t) +
2
d
t
(x, t).
From the equality
2
d(x, t) =
d
dr
_
2
d(x +r(x, t), t)
r=0
=
d
dr
_
(I rW(x, t))
1
W(x, t)
r=0
= W(x, t)
2
,
and Theorem 5.14, we obtain the expression of D
t
W. The normal time derivative
of the mean curvature is obtained from
D
t
= D
t
(trW) = tr(D
t
W).
For the Gauss-Kronecker curvature
g
, from the Jacobis formula, we have
D
t
g
= D
t
det(W +
T
) = tr
_
AD
t
(W +
T
)
_
,
where A = adj(W +
T
) denotes the adjugate matrix of W +
T
(see Section 1).
From Proposition A.1, it satises
AW =
g
(I
T
), A =
g
, Ae
i
=
_
_
j=i
j
_
_
e
i
(i = 1, , m1).
Since
D
t
(W +
T
)
= D
t
W + (D
t
)
T
+(D
t
)
T
= vW
2
+
(t)
(
T
(t)
v) +
_
T
(t)
v
_
W (
(t)
v)
T

T
(t)
v
= vW
2
+
(t)
(
T
(t)
v) (
(t)
v)
T
+
_
T
(t)
v
_
(W I),
we obtain
D
t
g
= tr
_
v
g
W +A
(t)
(
T
(t)
v) A(
(t)
v)
T
+
g
T
(t)
v
_
(W I)
_
= v
g
+ tr
_
A
(t)
(
T
(t)
v)
_
_
A
(t)
v
_
+
g

_
(W I)
(t)
v
_
= v
g
+ tr
_
A
(t)
(
T
(t)
v)
_
= v
g
+A :
(t)
(
T
(t)
v) .

CHAPTER 6
Variational formulas
One of the most important applications of our formulation for moving hypersur-
faces is to derive various variational formulas with respect to the shape deformation
of hypersurfaces. In this chapter, we prove some basic transport identities, i.e., vari-
ations of volume or surface integrals in the rst section. A transport identity of the
symmetric polynomials of the principal curvatures is derived in the second section.
This is an interesting extension of the Gauss-Bonnet theorem for two dimensional
surfaces to multidimensional cases.
1. Transport identities
We suppose that / is an oriented C
2,1
-class moving hypersurface of the form
(5.1) and that (t) is compact for each t 1.
Theorem 6.1. If f C
1
(/), then
d
dt
_
(t)
f(x, t) dH
m1
=
_
(t)
(D
t
f fv) dH
m1
.
d
dt
[(t)[ =
_
(t)
v dH
m1
. (6.1)
Proof. Under the condition of Proposition 5.2, we dene
(, t) := (, u(, t))
T
R
m
( |, t 1
0
).
We rst suppose the condition:
supp(f(, t)) (|, t) (t 1
0
).
Then, from Corollary 1.12, we have
_
(t)
f(x, t) dH
m1
=
_
U
f((, t), t)
_
g(, t)d,
where
A(, t) :=
_
(, t), ((, t), t)

_
,
g(, t) := (det A(, t))
2
.
Without loss of generality, we assume that
_
g(, t) = det A(, t) > 0.
From Proposition A.3, we have
t
_
g(, t) =

t
det A(, t) = tr
_
A(, t)
1
A
t
(, t)
_ _
g(, t).
75
76 6. VARIATIONAL FORMULAS
The last term is computed as follows. Since x (t) and | are corresponding
to each other by the map x = (, t), we can write it conversely as = (x, t),
where
x = ((x, t), t) (x (t)).
It follows that, for x = (, t),
A(, t)
1
=
_

T
(t)
(x, t)
T
(x, t)
_
,
and
tr
_
A(, t)
1
A
t
(, t)
_
= tr
_
_
_
_
_
T
(t)
T
_
_
_

t
(, t),

t
[((, t), t)]
_
_
_
_
= tr
_
(
T
(t)
)(
T

t
(, t))
1
2
t
[[
2
_
= tr
_
(
T
(t)
)(
T

t
(, t))
_
= tr
_
(
(t)
T
)(
T
t
(, t))
_
= tr
T
t
((x, t), t)
_
= div
(t),x
(
t
((x, t), t)) .
Hence, applying Proposition 5.11 and the GaussGreen formula (Theorem 2.18),
we obtain
d
dt
_
(t)
f(x, t) dH
m1
=
d
dt
_
U
f((, t), t)
_
g(, t)d
=
_
U
_
t
(f((, t), t))
g +f

t
_
g(, t)
_
d
=
_
U
__
D
t
f + (
T
(t)
f)
t
_
g +f div
(t),x
(
t
((x, t), t))
g
_
d
=
_
(t)
_
D
t
f + (
T
(t)
f)
t
+f div
(t),x
(
t
((x, t), t))
_
dH
m1
x
=
_
(t)
(D
t
f fv) dH
m1
.
For more general f, we can apply the above result with a partition of unity
of /.
We additionally suppose that there exist a bounded open set
(t) and a
unbounded open set
+
(t) of R
m
satisfying the following conditions:
R
m
(t) =
(t)
+
(t),
(t)
+
(t) = , ^
((t))
(t),
1. TRANSPORT IDENTITIES 77
for suciently small > 0. We dene
Q :=
_
tI
(t) t.
Then / is represented by the domain mapping method as follows.
Proposition 6.2. Under the above conditions, for a xed t
0
1, there ex-
ist a bounded domain
0
R
m
with a smooth boundary
0
=
0
and
C
1
(1
0
, C
1
(
0
, R
m
)) with an open subinterval 1
0
1 including t
0
, such that
(t) = (
0
, t), (t) = (
0
, t) (t 1
0
),
and that (, t) is a C
1
-dieomorphism from
0
to
(t) for each t 1

0
.
We remark that (, t) represents the map from y
0
R
m
to x
(t)
R
m
as x = (y, t), and that the condition C
1
(1, C
1
(
0
, R
m
)) stands for that
(y, t),
y
T
(y, t),

t
(y, t),
y
T
t
(y, t) =

t
T
(y, t),
are all continuous with respect to (y, t)
0
1
0
.
Theorem 6.4. Under the condition of Proposition 6.2, if u C
1
(Q), then we
have
d
dt
_
(t)
u(x, t) dx =
_
(t)
u
t
(x, t) dx
_
(t)
u(x, t)v(x, t) dH
m1
.
d
dt
[
(t)[ =
_
(t)
v dH
m1
.
Proof. Without loss of generality, we assume that
det
y
T
(y, t) > 0 (y
0
).
We dene (, t) := (, t)
1
C
1
(
(t),
0
). From the Jacobis formula (Propo-
sition A.3), we have
t
(det
T
y
(y, t)) = (det
T
y
) tr
_
(
x
T
)(
y
T
t
)
= (det
T
y
) tr
x
_
T
t
((x, t), t)
= (det
T
y
(y, t)) div
x
[
t
((x, t), t)] .
Hence, we obtain
d
dt
_
(t)
u(x, t) dx
=
d
dt
_
0
u((y, t), t) det
y
T
(y, t) dy
=
_
0
_
(
T
t
(y, t)
x
u(x, t) +u
t
(x, t))(det
y
T
(y, t))
+u(x, t) (det
T
y
(y, t)) div
x
[
t
((x, t), t)]
_
dy
=
_
(t)
_
T
t
(y, t)
x
u(x, t) +u
t
(x, t) +u(x, t) div
x
[
t
((x, t), t)]
_
dx
=
_
(t)
u
t
(x, t) dx
_
(t)
u(x, t)
t
((x, t), t) (x, t) dH
m1
x
.
Since v(x, t) =
t
((x, t), t) (x, t), the transport identity follows.
2. Transport identities for curvatures
We again suppose that / is an oriented C
2,1
-class moving hypersurface of the
form (5.1) and that (t) is compact for each t 1.
Theorem 6.5 (Transport identity for
g
).
d
dt
_
(t)
g
dH
m1
= 0. (6.2)
This theorem indicates that
_
g
dH
m1
is a topological invariant. In R
2
, it
is well known that
_
dH
m1
= #(connected components of ) 2.
In the three dimensional case, Theorem 6.5 follows from the famous Gauss-Bonnet
theorem for two dimensional manifolds.
If (t) is a solution of the Gaussian curvature ow v =
g
(Section 5), then we
have
d
dt
[
(t)[ =
_
(t)
vdH
m1
=
_
(t)
g
dH
m1
,
where the last integral does not depend on t. This indicates that the solution of
the Gaussian curvature ow possesses the constant volume speed property.
A direct approach to prove Theorem 6.5 is as follows. From Theorem 6.1 and
Theorem 5.15, we have
d
dt
_
(t)
g
dH
m1
=
_
(t)
(D
t
g
v
g
) dH
m1
=
_
(t)
_
adj(W +
T
) :
(t)
(
T
(t)
v)
_
dH
m1
.
2. TRANSPORT IDENTITIES FOR CURVATURES 79
Using the integration by parts on (t), it is possible to show that the last integral
vanishes, but it is lengthy and complicated. So, we choose another way to prove
Theorem 6.5 in a more general form.
For r R, we dene
A
r
(x, t) := I +rW(x, t) ((x, t) /).
Then we have
det A
r
=
m1
i=1
(1 +r
i
). (6.3)
Lemma 6.6.
d
dt
[
r
(t)[ =
_
(t)
_

r
det A
r
(x, t)
_
v(x, t) dH
m1
x
.
Proof. For y
r
(t), the perpendicular foot on (t) from y is denoted by
x = (y, t) (t). From Proposition 4.1 and (6.3), we have
v
r
(y, t) =

t
d
r
(y, t) =

t
(d(y, t) r) = v(x, t).
and
r
(y, t) =
m1
i=1
i
1 +r
i
=
m1
i=1
r
log(1 +r
i
)
=

r
log det A
r
=
1
det A
r
r
det A
r
,
where
i
=
i
(x, t) and A
r
= A
r
(x, t).
d
dt
[
r
(t)[ = =
_
r(t)
r
v
r
dH
m1
=
_
(t)
1
det A
r
_

r
det A
r
_
v det A
r
dH
m1
=
_
(t)
_

r
det A
r
_
v dH
m1
.
Let S
l
= S
l
(
1
, ,
m1
) be the lth elementary symmetric polynomial of
1
, ,
m1
, i.e.,
S
0
= 1, S
1
=
m1
i=1
i
= , S
2
=
i<j
j
, , S
m1
=
m1
i=1
i
=
g
.
Then we have
det A
r
=
m1
i=1
(1 +r
i
) =
m
l=0
S
l
r
l
,
where S
m
:= 0.
Theorem 6.7.
d
dt
_
(t)
S
l
dH
m1
= (l + 1)
_
(t)
S
l+1
v dH
m1
(l = 0, , m1).
Proof. Since
[
r
(t)[ =
_
(t)
det A
r
dH
m1
=
m1
l=0
r
l
_
(t)
S
l
dH
m1
,
we have
d
dt
[
r
(t)[ =
m1
l=0
r
l
d
dt
_
(t)
S
l
dH
m1
. (6.4)
On the other hand, from Lemma 6.6 and the equality
r
det A
r
=
m1
l=0
(l + 1)S
l+1
r
l
,
we also obtain
d
dt
[
r
(t)[ =
m1
l=0
r
l
_
(l + 1)
_
(t)
S
l+1
v dH
m1
_
. (6.5)
Comparing (6.4) and (6.5), we obtain the result.
Theorem 6.7 includes the two important formulas (6.1) with l = 0 and (6.2)
with l = m1. This theorem enables us to comprehend them in a unied approach.
CHAPTER 7
Gradient structure and moving boundary
problems
Many geometric moving boundary problems have their own gradient ow struc-
tures. A gradient structure is dened by an energy functional and an inner product
for each hypersurface. We consider several examples of moving boundary problems
and give formal gradient structures for them. In this chapter, we suppose / is a
suciently smooth moving hypersurface with compact (t).
1. General gradient ow of hypersurfaces
We illustrate a general framework of moving boundary problems with a formal
gradient structure. Let ( be a set of hypersurfaces in R
m
with some regularity
conditions and with a constraint R() = 0. We suppose that (t) ( for t 1.
Then formally we can write
d
dt
R((t)) = R
((t)), v(, t)) ,

where the right hand side is linear with respect to the normal velocity v. This leads
to the constraint for the normal velocity:
v T
(t)
(() := v ; R
((t)), v) = 0.
An typical example of the constraint is the volume preserving condition:
[
[ = a
0
, R() := [
[ a
0
= 0, (7.1)
where a
0
> 0 is a xed constant. In this case, we have
d
dt
R((t)) =
_
(t)
v dH
m1
,
v T
(() :=
_
v ;
_
v dH
m1
= 0
_
. (7.2)
We consider an energy:
E : ( E() R.
Then the rst variation of the energy
d
dt
E((t)) becomes a linear functional of
v T
(().
To consider a gradient ow of E(), we need to identify the rst variation of the
energy with a linear functional on T
((). We introduce an inner product f, g)
81
82 7. GRADIENT STRUCTURE AND MOVING BOUNDARY PROBLEMS
dened on a linear space T
((), by which we identify the variation of energy with

a element of T
((), namely,
d
dt
E((t)) = E((t)), v)
(t)
.
The expression of the rst variation E() depends on the choice of the inner
product. The most standard choice is the L
2
-inner product:
f, g)
=
_
fg dH
m1
.
The gradient ow of E() with respect to , )
is formally given as
v = E((t)).
An important property of the gradient ow is the energy decreasing property:
d
dt
E((t)) = E((t)), v)
(t)
= v, v)
(t)
0.
We will show various examples of moving boundary problems with the gradient
structure in the following sections.
2. Prescribed normal velocity motion
Condition: f C
(R
m
) is given.
Energy: E() :=
_
f(x)dH
m
Variation of energy:
d
dt
E((t)) =
_
(t)
f(x)v dH
m1
Inner product: f, g)
:=
_
fg dH
m1
Gradient ow: v = f(x)
We remark that, if f 1, then E() = [
[ and this is the constant speed motion

v = 1, which is deeply related to the eikonal equation (see [18] etc.).
3. Mean curvature ow
The mean curvature ow is one of the most important moving boundary prob-
lems and is related to many applications. We refer [6] and [18] and reference therein.
Energy: E() :=
_
dH
m1
d
dt
E((t)) =
_
(t)
v dH
m1
:=
_
fg dH
m1
Gradient ow: v =
5. GAUSSIAN CURVATURE FLOW 83
4. Anisotropic mean curvature ow
We also consider an anisotropic surface energy
E() :=
_
((x)) dH
m1
,
where C
2
(S
m1
, R
+
) is a given anisotropic energy density, where S
m1
:=
x R
m
; [x[ = 1. We extend to the whole space as
(x) := (x/[x[) (x R
m
, x ,= 0).
The variation of the energy is given as
d
dt
E((t)) =
_
(t)
_
2
() : W +()
_
v dH
m1
. (7.3)
Then we have the gradient ow as follows.
:=
_
fg dH
m1
Gradient ow: v =
2
() : W +()
In the case m = 2, we can write () := () with ( = (cos , sin)
T
). Then the
gradient ow becomes
v = (
() +()) . (7.4)
Exercise 7.1. Prove (7.3) and (7.4).
For more details about the anisotropic mean curvature ow, see [6] and the
references therein.
5. Gaussian curvature ow
Let (t) be convex (i.e.
(t) is a convex domain). In the case m = 3, the

moving boundary problem
v =
g
,
is called the Gaussian curvature ow and is well studied (see [6], [19] and refer-
ences therein). In particular, it possesses the constant volume speed property (see
Section 2).
We consider a more general form in R
m
using the result of Theorem 6.7.
Energy: E() :=
1
l + 1
_
S
l
(
1
, ,
m1
) dH
m1
d
dt
E((t)) =
_
(t)
S
l+1
(
1
, ,
m1
) v dH
m1
:=
_
fg dH
m1
Gradient ow: v = S
l+1
(
1
, ,
m1
)
This is an extension of the mean curvature ow (l = 0) and the Gaussian curvature
ow (m = 3 and l = 2).
6. Willmore ow
The following gradient ow is called the Willmore ow (see [1], [17] and the
references therein).
Energy: E() :=
1
2
_
2
dH
m1
d
dt
E((t)) =
_
(t)
_
+
m1
i=1
2
i

1
2
3
_
v dH
m1
:=
_
fg dH
m1
Gradient ow: v =

m1
i=1
2
i
+
1
2
3
Gradient ow (m=3): v =

1
2
3
+ 2
g
7. Volume preserving mean curvature ow
We suppose the volume preserving condition (7.1) and the constraint on v
(7.2) with the same energy and inner product as Section 3. Since we impose the
constraint, the inner product is
f, g)
=
_
fg dH
m1
for f, g T
(().
The variation of the energy is identied with an element of T
(() as follows. We
dene
) :=
1
[[
_
dH
m1
.
Then we have
d
dt
E((t)) =
_
(t)
v dH
m1
=
_
(t)
() )v dH
m1
= ) , v(, t))
(t)
.
Hence, we obtain
E() = ) .
The gradient ow under the volume preserving condition becomes as follows.
Gradient ow: v = )
This is called the volume preserving mean curvature ow (area preserving mean
curvature ow, if m = 2), which keeps the volume enclosed by (t) and decreases
the surface area. See [11] etc., for details.
8. SURFACE DIFFUSION FLOW 85
8. Surface diusion ow
We again suppose the volume preserving condition (7.1) and the constraint on
v (7.2) with the same energy as Section 3. But we consider another inner product
on T
(().
We consider a Sobolev space on :
H
1
() := f : R; f L
2
(), [
f[ L
2
(),
with the norm
|f|
H
1
()
:=
__
([f[
2
+[
f[
2
) dH
m1
_1
2
.
Its dual space is denoted by H
1
(). For f L
2
(), it is identied with an element
of H
1
() by
H
1
()
f, g )
H
1
()
=
_
fg dH
m1
(
g H
1
()).
We dene the following Sobolev spaces with constraint.
H
1
() := f H
1
();
_
f dH
m1
= 0,
H
1
() := f H
1
();
H
1
()
1, f )
H
1
()
= 0.
Then it follows that the LaplaceBeltrami operator can be extended to a topo-
logical linear isomorphism from

H
1
() onto

H
1
(). We denote it by

B(

H
1
(),

H
1
()). Our inner product is dened as
f, g)
:=
H
1
()
f, g
_
H
1
()
(f, g T
(()

H
1
()) .
This can be regarded as an inner product of

H
1
().
Then we can identify the variation of the energy in the following sense. Since
_
fdH
m1
= 0, we formally have
f) = f f) (f H
1
()),
and
d
dt
E((t)) =
_
(t)
( )) v dH
m1
=
_
(t)
) v dH
m1
=
, v(, t))
(t)
.
Hence, we obtain
E() =
The gradient ow with respect to the inner product becomes as follows.

Gradient ow: v =
This is called the surface diusion ow, which keeps the volume enclosed by (t)
and decreases the surface area. See [2] etc., for details.
9. HeleShaw moving boundary problem
HeleShaw ow is two-dimensional slow ow of viscous uid between two par-
allel horizontal plates with a thin gap. It was studied by H. S. HeleShaw [9]
experimentally at rst, and some mathematical models have been studied in sev-
eral papers [3], [4], [10], [13], etc.
A typical mathematical model for the moving boundary problem of the Hele
Shaw ow with the surface tension eect is as follows.
_
_
p(x, t) = 0 (x
(t)),
p(x, t) = (x (t)),
v(x, t) = (x, t) p(x, t) (x (t)),
with a given initial boundary (0), where p(x, t) is a pressure eld of the uid in
(t).
This HeleShaw moving boundary problem also has a gradient structure. Let
H
1
2
() be the trace space on from H
1
(
). Then it is well known that, for

q H
1
2
(), there exists u = u(q) H
1
(
) with
u = 0 in
, u[
= q.
We dene the Dirichlet-to-Neumann map:
q :=
u(q)
B(H
1
2
(), H
1
2
()),
where H
1
2
() := (H
1
2
())
. We dene quotient spaces
H
1
2
() :=
_
f H
1
2
();
_
f dH
m1
= 0
_
,
1
2
() :=
_
f H
1
2
();
H
1
2 ()
1, f )
H
1
2 ()
= 0
_
.
Then

:=
H
1
2 ()
belongs to B(

H
1
2
(),

H
1
2
()) and is bijective. We dene
U(f) := u(
1
f). Their inner products are dened by

(f, g)
H
1
2 ()
:=
_
u(f) u(g)dx,
(f, g)
1
2 ()
:=
_
U(f) U(g)dx,
We remark that
(f, g)
H
1
2 ()
=
_
f) g dH
m1
(if
f L
2
()),
(f, g)
1
2 ()
=
_
f) g dH
m1
(if g L
2
()).
We again suppose the volume preserving condition (7.1) and the constraint on
v (7.2) with the same energy as Section 3. We adopt the inner product of

H
1
2
():
f, g)
:= (f, g)
1
2 ()
.
9. HELESHAW MOVING BOUNDARY PROBLEM 87
Then we can identify the variation of the energy in the following sense. Since
_
f) dH
m1
= 0, we formally have
f) = f f) (f H
1
2
()),
and
d
dt
E((t)) =
_
(t)
( )) v dH
m1
=
_
(t)
) v dH
m1
=
, v(, t))
(t)
.
Hence, we obtain
E() =
=
u()
.
The gradient ow with respect to the inner product becomes as follows.
Gradient ow: v =
u()
This is nothing but the HeleShaw moving boundary problem. It also keeps the
volume enclosed by (t) and decreases the surface area. Physically, these properties
correspond to the volume conservation law and the surface tension eect. See the
above references for details.
Bibliography
[1] K. Deckelnick and G. Dziuk, Error estimates for the Willmore ow of graphs,
Interfaces and Free Boundaries Vol.8 (2006), pp.21-46.
[2] C. M. Elliott and H. Garcke, Existence results for diusive surface motion
laws, Adv. Math. Sci. Appl. Vol.7 (1997), pp.467-490.
[3] J. Escher and G. Simonett; Classical solutions of multidimensional Hele-Shaw
models, SIAM J. Math. Anal. Vol.28 (1997), pp.1028-1047.
[4] J. Escher and G. Simonett; Classical solutions for Hele-Shaw models with
surface tension, Adv. Di. Eq. Vol.2 (1997), pp.619-642.
[5] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Func-
tions, Studies in Advanced Mathematics, CRC Press, Boca Raton, FL (1992).
[6] Y. Giga, Surface Evolution Equations, A Level Set Approach, Monographs in
Mathematics Vol.99, Birkhauser Verlag, Basel (2006).
[7] D. Gilberg and N. S. Trudinger, Elliptic Partial Dierential Equations of Sec-
ond Order, second edition, Springer-Verlag (1983).
[8] M. E. Gurtin, Thermomechanics of Evolving Phase Boundaries in the Plane,
Oxford Mathematical Monographs, The Clarendon Press, Oxford (1993).
[9] H. S. Hele-Shaw, The ow of water, Nature Vol.58 (1898), pp.34-36.
[10] S. D. Howison; Fingering in Hele-Shaw cells, J. Fluid Mech. Vol.167 (1986),
pp.439-453.
[11] G. Huisken, The volume preserving mean curvature ow, J. Reine Angew.
Math. Vol.382 (1987), pp.35-48.
[12] M. Kimura, Numerical analysis of moving boundary problems using the bound-
ary tracking method, Japan J. Indust. Appl. Math. Vol.14 (1997), pp.373-398.
[13] M. Kimura, Time local existence of a moving boundary of the Hele-Shaw ow
with suction, Euro. J. Appl. Math. Vol.10 (2000), pp.581-605.
[14] M. Kimura, Shape derivative and linearization of moving boundary prob-
lems, GAKUTO International Series, Mathematical Sciences and Applications
Vol.13 (2000), pp.167-179.
[15] M. Kimura, Numerical analysis of moving boundary problems, Sugaku Expo-
sitions, Vol.15 (2002), pp.71-88.
[16] M. Kimura and H. Notsu, A level set method using the signed distance func-
tion, Japan J. Indust. Appl. Math. Vol.19 (2002), pp.415-446.
[17] T. Oberhuber, Finite dierence scheme for the Willmore ow of graphs, Ky-
bernetika Vol.43, No. 6 (2007), pp.855-867.
[18] J. A. Sethian, Level Set Methods and Fast Marching Methods, Evolving In-
terfaces in Computational Geometry, Fluid Mechanics, Computer Vision, and
Materials Science, second edition, Cambridge Monographs on Applied and
89
90 Bibliography
Computational Mathematics, Vol.3, Cambridge University Press, Cambridge
(1999).
[19] T. K. Ushijima and H. Yagisita, Approximation of the Gauss curvature ow by
a three-dimensional crystalline motion, Proceedings of Czech-Japanese Semi-
nar in Applied Mathematics 2005, COE Lect. Note Vol.3, Kyushu Univ. (2006),
pp.139-145.
[20] W. P. Ziemer, Weakly Dierentiable Functions, Sobolev Spaces and Functions
of Bounded Variation, Graduate Texts in Mathematics, Vol.120, Springer-
Verlag, New York (1989).
APPENDIX A
1. Adjugate matrix
For A R
mm
, we denote by adjA R
mm
the transpose of the cofactor
matrix of A. It is called the adjugate of A. It is also called adjoint, but the
adjoint matrix of A often means the conjugate transpose A
:= A
T
. To avoid the
ambiguity, we adopt the word adjugate for adjA. It is well known that
A(adjA) = (adjA) A = (det A) I,
and, for invertible A, the inverse matrix is given by
A
1
=
1
det A
adjA.
We remark the next proposition.
Proposition A.1. Let A be diagonalizable. Namely, there exists an invertible
matrix P such that
A = P
1
_
_
_
1
O
.
.
.
O
m
_
_
_P,
where
1
, ,
m
are the eigenvalues of A. Then, the adjugate of A is represented
in the form:
adjA = P
1
_
_
_
1
O
.
.
.
O
m
_
_
_P,
where
i
:=
1jm, j=i
j
(i = 1, , m).
Proof. The assertion is clear if det A =
1

m
,= 0. For general A, since
det(A +I) = (
1
+) (
m
+) ,= 0 for 0 < << 1, we have
adj(A +I) = P
1
_
_
_
1
() O
.
.
.
O
m
()
_
_
_P,
where
i
() :=
1jm, j=i
(
j
+) (i = 1, , m).
Taking the limit +0, we obtain the assertion.
91
92 A
2. Jacobis formula
Lemma A.2. Let A = (a
1
, , a
m
), B = (b
1
, , b
m
) R
mm
. Then we
have
tr (adjA) B =
m
k=1
det (a
1
, , a
k1
, b
k
, a
k+1
, , a
m
) .
Proof. It is sucient to prove the assertion of the lemma in the case detA ,= 0.
We remark that adjA = (detA)A
1
. The jth column vector of the identity matrix
I R
mm
is denoted by
j
= (
1j
, ,
mj
)
T
R
m
. We have
m
k=1
det (a
1
, , a
k1
, b
k
, a
k+1
, , a
m
)
=
m
k=1
(det A) det
_
A
1
(a
1
, , a
k1
, b
k
, a
k+1
, , a
m
)
_
= (det A)
m
k=1
det
_
1
, ,
k1
, A
1
b
k
,
k+1
, ,
m
_
= (det A)
m
k=1
(the kth component of A
1
b
k
)
= (det A) tr (A
1
B).
For general A, we can derive the formula in the similar way to the proof of Propo-
sition A.1.
Proposition A.3 (Jacobis formula). We suppose that A C
1
(1, R
mm
).
Then we have
d
dt
detA(t) = tr (adjA(t)) A
(t) (t 1).
In particular, if detA(t) ,= 0 then
d
dt
detA(t) = detA(t) tr
_
A(t)
1
A
(t)
_
(t 1).
2. JACOBIS FORMULA 93
Proof. Let a
j
(t) = (a
1j
(t), , a
mj
(t))
T
R
m
be the jth column vector of
the matrix A(t) R
mm
. The t derivative is denoted by

=
d
dt
. We have
d
dt
detA(t)
=
d
dt
Sm
sgn() a
(1)1
a
(m)m
=
Sm
sgn()
_
m
k=1
a
(1)1
a
(k1)k1
a
(k)k
a
(k+1)k+1
a
(m)m
_
=
m
k=1
_

Sm
sgn() a
(1)1
a
(k1)k1
a
(k)k
a
(k+1)k+1
a
(m)m
_
=
m
k=1
det (a
1
, , a
k1
, a
k
, a
k+1
, , a
m
)
= tr (adjA(t)) A
(t) .

Part 3
Large time behaviour for diusive
HamiltonJacobi equations
Philippe Laurencot
2000 Mathematics Subject Classication. Primary 35B40, 35B33, 35K15, 35K55,
49L25
Key words and phrases. diusive HamiltonJacobi equation, large time behaviour,
gradient estimates, extinction, convergence to self-similarity
Abstract. The large time behaviour of non-negative and integrable solutions
to the semilinear diusive HamiltonJacobi equation tu u |u|
q
= 0
in (0, ) R
N
is surveyed in these lecture notes, the HamiltonJacobi term
acting either as an absorption or a source term. Temporal decay estimates are
established for the solution, its gradient and time derivative as well as one-sided
estimates for its Hessian matrix. Next, according to the values of the param-
eters q > 0 and N 1, the time evolution is shown to be either dominated
by the diusion term, the reaction term, or governed by a balance between
both terms. The occurrence of nite time extinction is also investigated for
q (0, 1).
Acknowledgement. These notes grew out from lectures which I gave at Fu-
dan University, Shangai, in November 2006 and at the Necas Center for Math-
ematical Modeling, Praha, in April 2008. I thank Professors Songmu Zheng
and Eduard Feireisl for their kind invitations as well as both institutions for
their hospitality and support. I also thank Sad Benachour and the referee for
useful comments on the manuscript.
Contents
Chapter 1. Introduction 99
Chapter 2. Well-posedness and smoothing eects 103
1. Gradient estimates 105
1.1. Gradient estimates: = 1 and q > 1 106
1.2. Gradient estimates: q > 1 107
1.3. Gradient estimates: q = 1 107
1.4. Gradient estimates: q (0, 1) 108
2. Time derivative estimates 108
3. Hessian estimates 111
4. Existence 114
5. Uniqueness 116
Chapter 3. Extinction in nite time 119
1. An integral condition for extinction 121
2. A pointwise condition for extinction 123
3. Non-extinction 124
4. A lower bound near the extinction time 126
Chapter 4. Temporal decay estimates for integrable initial data: = 1 129
1. Decay rates 131
1.1. Decay rates: q = N/(N + 1) 132
1.2. Decay rates: q = 1 133
1
134
) 137
Chapter 5. Temporal growth estimates for integrable initial data: = 1 141
1
and |u|
142
1.1. The case q 2 142
1.2. The case q (q
, 2) 144
1.3. The case q [1, q
] 147
2. Growth rates 149
Chapter 6. Convergence to self-similarity 151
1. The diusion-dominated case: = 1 151
97
98 CONTENTS
2. The reaction-dominated case: = 1 154
Bibliography 159
Appendix A. Self-similar large time behaviour 163
1. Convergence to self-similarity for the heat equation 163
2. Convergence to self-similarity for HamiltonJacobi equations 164
CHAPTER 1
Introduction
The aim of these lectures is to present several qualitative properties of non-
negative solutions to the Cauchy problem for a semilinear parabolic equation in-
volving a nonlinearity depending on the gradient of the solution which might be
called a diusive (or viscous) HamiltonJacobi equation and reads
t
u u + [u[
q
= 0 , (t, x) Q
:= (0, ) R
N
, (1.1)
u(0) = u
0
, x R
N
, (1.2)
where u = u(t, x) is a function of the time t 0 and the space variable x R
N
,
N 1, and the parameters and q are such that
q (0, ) and 1, 1 . (1.3)
The equation (1.1) features two mechanisms acting on the space variable, the
linear diusion u and the nonlinearity [u[
q
. Taken apart, these two mecha-
nisms have dierent properties and there is thus a competition between them in
(1.1). The main issue to be addressed in these notes is whether one of these two
terms dominates the dynamics at large times. To start the discussion on this mat-
ter, it is obvious that the sign of is of utmost importance: indeed, if = 1, the
nonlinear term [u[
q
is non-negative and thus acts as an absorption term which
enhances the eect of the linear diusion. While, if = 1, [u[
q
is non-positive
and is thus a source term which opposes to the diusive eect. Observe also that,
in contrast to the semilinear parabolic equation
t
v v v
p
= 0 , (t, x) Q
,
which has been thoroughly studied, the nonlinearity in (1.1) is only eective in the
spatial regions where u is steep and not at. In particular, constants are solutions
to (1.1).
In order to determine which of these mechanisms governs the large time dy-
namics, let us rst describe briey some salient features of their own and restrict
ourselves to non-negative integrable and bounded continuous initial data: we thus
assume that
u
0
L
1
(R
N
) B((R
N
) , u
0
0, u
0
, 0 . (1.4)
On the one hand, it is well-known that the linear heat equation
t
c c = 0 , (t, x) Q
, (1.5)
99
100 1. INTRODUCTION
with initial condition c(0) = u
0
has a unique smooth solution c : t e
t
u
0
which
is given by
c(t, x) :=
_
e
t
u
0
_
(x) = (g(t) u
0
) (x) =
1
t
N/2
_
R
N
G
_
x y
t
1/2
_
u
0
(y) dy (1.6)
with the notation
g(t, x) :=
1
t
N/2
G
_
x
t
1/2
_
and G(x) :=
1
(4)
N/2
exp
_
[x[
2
4
_
(1.7)
for (t, x) Q
. An important property enjoyed by non-negative and integrable

solutions to (1.5) is the invariance through time evolution of the L
1
-norm, that is
_
_
e
t
u
0
_
_
1
= |u
0
|
1
for every t 0 . (1.8)
On the other hand, the HamiltonJacobi equation
t
h + [h[
q
= 0 , (t, x) Q
, (1.9)
does not have smooth solutions and might even have several weak solutions, a
drawback which is also met in the study of scalar conservation laws. Still, there is a
selection principle which allows to single out one solution among the weak solutions,
the so-called viscosity solution. We refer the reader to [4, 6, 25, 27] for the precise
denition and several properties of viscosity solutions. In particular, (1.9) has a
unique viscosity solution with initial condition h(0) = u
0
. Moreover, if q 1, it is
given by the Hopf-Lax-Oleinik representation formula
h(t, x) := inf
yR
N
_
u
0
(y) +
q 1
q
q/(q1)
[x y[
q/(q1)
t
1/(q1)
_
(1.10)
if q > 1 and
h(t, x) := inf
{|yx|t}
u
0
(y) if q = 1 , (1.11)
for (t, x) Q
. According to (1.10) and (1.11) the HamiltonJacobi equation (1.9)

enjoys the following invariance property
inf
xR
N
h(t, x) = inf
xR
N
u
0
(x) for every t 0 . (1.12)
A natural guess is then that
|u(t)|
1
and inf
xR
N
u(t, x)
are two relevant quantities for the large time dynamics.
Let us rst look at the case = 1. In that case, u(t) turns out to be integrable
for each t 0 so that
inf
xR
N
u(t, x) = inf
xR
N
u
0
(x) = 0 ,
and this quantity does not seem to provide any useful information on the large time
behaviour. Concerning the L
1
-norm of u it formally follows from (1.1) that
|u(t)|
1
= |u
0
|
1
_
t
0
_
R
N
[u(s, x)[
q
dxds , t 0 , (1.13)
1. INTRODUCTION 101
from which we readily deduce that t |u(t)|
1
is non-increasing and
I
1
() := lim
t
|u(t)|
1
= |u
0
|
1

_

0
_
R
N
[u(s, x)[
q
dxds 0 .
The question is then whether I
1
() > 0 or I
1
() = 0. In the latter case it means
that the damping of the nonlinearity is suciently strong so as to drive the L
1
-
norm of u(t) to zero as t . The nonlinear term is thus expected to play a
non-negligible role for large times.
In the same vein, u is a subsolution to the linear heat equation and we infer
from the comparison principle that
|u(t)|

_
_
e
t
u
0
_
_
C t
N/2
, t (0, ) .
As the nonlinear term [u[
q
acts as an absorption term in that case it enhances
the dissipation due to the diusion and the main issue is to gure out whether this
additional dissipative mechanism speeds up the convergence to zero of the L
-norm
of u(t) as t .
When = 1, the nonlinear term [u[
q
is a source term and thus slows down
or even impedes the dissipation of the diusion. Indeed, it formally follows from
(1.1) by integration that
|u(t)|
1
= |u
0
|
1
+
_
t
0
_
R
N
[u(s, x)[
q
dxds , t 0 .
Consequently t |u(t)|
1
and
t inf
xR
N
u(t, x) = |u(t)|
are non-decreasing. Setting

I
1
() := lim
t
|u(t)|
1
[|u
0
|
1
, ] and I
() := lim
t
|u(t)|
[0, |u
0
|
] ,
the questions here are whether I
1
() is nite or not and whether I
() is positive
or zero. While we expect the diusion to be dominant for large times if I
1
() <
and the nonlinearity to be dominant for large times if I
() > 0, we shall see

below that the situation is more complicated than for = 1.
We nally collect some elementary properties of (1.1).
If u is a non-negative solution to (1.1), (1.2) then
u(t, x) :=
1
2
|u
0
|
+ u(t, x) , (t, x) [0, ) R

N
, (1.14)
is a non-negative solution to
t
u u +[ u[
q
= 0 , (t, x) Q
, (1.15)
u(0) = (1 )
|u
0
|
2
+ u
0
, x R
N
. (1.16)
Of course, u = u if = 1.
The equation (1.1) is an autonomous equation, that is, if u is a solution
to (1.1), (1.2), and t
0
> 0, then u(t, x) := u(t + t
0
, x) solves (1.1) with
initial condition u(t
0
).
102 1. INTRODUCTION
If u is a solution to (1.1), (1.2) and > 0, then the function u
dened
by
u
(t, x) :=
2q
u
_
2(q1)
t,
q1
x
_
, (t, x) [0, ) R
N
is also a solution to (1.1) with initial condition u
(0).
Finally there are two cases for which (1.1) can be transformed to a linear
equation:
if q = 2, the celebrated HopfCole transformation reduces (1.1) to the
linear heat equation. Indeed, introducing v = e
u
, we have v(t) =
e
t
e
u0
for t 0.
If q = 1, N = 1 and u
0
is non-increasing on (0, ) and non-decreasing
on (, 0), then so is u(t) for each t 0 and u actually solves the linear
equation
t
u
2
x
u sign(x)
x
u = 0 , (t, x) [0, ) R.
Notations. In the following, C and C
i
, i 1, denote positive constants that only
depend on N and q and may vary from place to place. The dependence of these
constants upon additional parameters is indicated explicitly.
For p [1, ] and w L
p
(R
N
), |w|
p
denotes the norm of w in L
p
(R
N
).
We also dene the space B((R
N
) of bounded and continuous functions in R
N
by
B((R
N
) := ((R
N
) L
(R
N
) while B|((R
N
) denotes the space of bounded and
uniformly continuous functions in R
N
.
Given a real number r R, we dene its positive part r
+
by r
+
:= max r, 0.
CHAPTER 2
Well-posedness and smoothing eects
We begin with the well-posedness of (1.1), (1.2) in B((R
N
).
Theorem 2.1. [32] Let u
0
be a non-negative function in B((R
N
) and
1, 1. Then the initial-value problem (1.1), (1.2) has a unique non-negative
classical solution
u B(([0, ) R
N
) C
1,2
(Q
) .
In addition, t |u(t)|
is a non-increasing function of time and

0 u(t, x) |u
0
|
, (t, x) [0, ) R
N
. (2.1)
We also have
u(t) = e
t
u
0
_
t
0
e
(ts)
[u(s)[
q
ds , t 0 . (2.2)
Recall that t e
t
u
0
denotes the solution to the linear heat equation (1.5) with
initial condition u
0
and is given by (1.6), (1.7). Furthermore, if u
0
W
1,
(R
N
),
then
|u(t)|
|u
0
|
, t 0 . (2.3)
While the uniqueness stated in Theorem 2.1 is a consequence of the comparison
principle [32] (see Section 5 below), the proof of the existence of a solution to (1.1),
(1.2) requires more ingredients among which the following gradient estimates are
crucial.
Theorem 2.2. [11, 32] Let u
0
N
),
1, 1, and denote by u the corresponding classical solution to (1.1), (1.2).
Then u(t) L
(R
N
) and
|u(t)|
C
1
|u(s)|
(t s)
1/2
if q > 0 , (2.4)
|u(t)|
C
2
|u(s)|
1/q
(t s)
1/q
if q > 0 , q ,= 1 , (2.5)
for t > s 0.
Furthermore, if = 1 and q > 1, then
_
_
_u
(q1)/q
(t)
_
_
_
C
3
|u(s)|
(q1)/q
(t s)
1/2
, (2.6)
_
_
_u
(q1)/q
(t)
_
_
_
C
4
t
1/q
(2.7)
for every t > s 0.
103
104 2. WELL-POSEDNESS AND SMOOTHING EFFECTS
As a consequence of Theorem 2.2 there is a smoothing eect from L
(R
N
) to
W
1,
(R
N
) for solutions to (1.1), (1.2). But Theorem 2.2 also provides quantitative
information about the temporal decay of |u(t)|
which will be used in the study

of the large time behaviour. In fact, non-negative solutions to the linear heat
equation (1.5) also satisfy (2.4) and (2.6) while (2.5) and (2.7) are also enjoyed
by non-negative (viscosity) solutions to (1.9). Thus, in that particular case, non-
negative solutions to (1.1), (1.2) inherit the smoothing eects from both terms
acting on the space variable.
The proof of the gradient estimates (2.4)(2.7) relies on a modied Bernstein
technique [21]: introducing a new unknown function v dened by v := f
1
(u) for
some strictly monotone function f to be determined, one looks for time-dependent
supersolutions to the nonlinear parabolic equation satised by w := [v[
2
and then
apply the comparison principle.
Smoothing eects for the time derivative are also available for both (1.5) and
(1.9) and it turns out that this property is still shared by (1.1).
0
N
), 1, 1,
and denote by u the corresponding classical solution to (1.1), (1.2). Then, for
i 1, 2, x R
N
and t > 0, we have
C
6

0
(t)
t
u(t, x) C
5

i
(t) , (t, x) Q
, (2.8)
with
0
(t) :=
_
_
_
|u
0
|
t
1
if q ,= 1 ,
|u
0
|
_
t
1
+t
1/2
_
if q = 1 ,
1
(t) :=
_
_
_
|u
0
|
1/(2q1)
t
(q+1)/(2q1)
if q 2 ,
|u
0
|
t
1
if q (0, 2) ,
2
(t) :=
_
_
|u
0
|
t
1
if q 2 ,
|u
0
|
1/(2q1)
t
(q+1)/(2q1)
if q (1, 2) ,
|u
0
|
t
3/2
(1 + log (1 +t)) if q = 1 ,
|u
0
|
(2q)/q
t
2/q
if q (0, 1) .
The proof of Theorem 2.3 also relies on the comparison principle applied to the
equation satised by (
t
u A)/(u) for some suitable choices of 1, 1,
A R and (
2
(R
N
; (0, )).
The situation is slightly dierent for the Hessian matrix: indeed, while it also
enjoys a smoothing eect for (1.5), this is no longer true for (1.9). It is then not clear
whether a smoothing eect is available for (1.1). Still, non-negative solutions to
(1.9) become instantaneously semiconcave if = 1 [27, Chapter 3.3] and semiconvex
if = 1 and non-negative solutions to (1.1) also enjoy this property if q (1, 2).
1. GRADIENT ESTIMATES 105
0
N
), 1, 1,
and denote by u the corresponding classical solution to (1.1), (1.2). If q (1, 2],
then the Hessian matrix D
2
u = (
i
j
u)
1i,jN
satises
D
2
u(t, x) C
7
|u
0
|
(2q)/q
t
2/q
I
N
(2.9)
for (t, x) Q
, that is,
i,j=1
(
i
j
u) (t, x)
i

j
C
7
|u
0
|
(2q)/q
[[
2
t
2/q
for each R
N
. Here I
N
denotes the N N identity matrix.
Furthermore, if u
0
W
2,
(R
N
),
D
2
u(t, x) |D
2
u
0
|
I
N
. (2.10)
The proof of Theorem 2.4 still relies on the comparison principle. For q = 2,
the estimate (2.9) follows from the analysis of Hamilton [34] (since, if f is a non-
negative solution to the linear heat equation
t
f = f, the function log f solves
(1.1) with q = 2). It is also established in [41, Lemma 5.1], still for q = 2.
The estimate (2.9) may also be seen as an extension to a multidimensional
setting of a weak form of the Oleinik gradient estimate for scalar conservation laws.
Indeed, if N = 1 and U =
x
u, then U is a solution to
t
U
2
x
U +
x
([U[
q
) = 0
in (0, ) R. The estimate (2.9) then reads

x
U C |u
0
|
(2q)/q
t
2/q
for t > 0, respectively, and we thus recover the results of [29, 37] in that case.
When q = 1 and = 1 it is rather log u that becomes instantaneously semi-
convex, a property reminiscent from the Aronson-Benilan one-sided inequality for
the Laplacian of log
_
e
t
u
0
_
[3].
Proposition 2.5. Let u
0
N
) and denote by
u the corresponding classical solution to (1.1), (1.2). If = 1 and q = 1, then
the matrix D
2
log u(t, x) +(1/2t) I
N
is a non-negative matrix for each (t, x) Q
,
that is, for each R
N
,
N
i,j=1
(
i
j
log u) (t, x)
i

j
+
[[
2
2t
0 . (2.11)
The remainder of this chapter is devoted to the proof of the above results: since
the nonlinearity and the initial condition need not have the required regularity to
apply the comparison principle, we give a formal proof of Theorems 2.2, 2.3, 2.4 and
Proposition 2.5 in Sections 1, 2 and 3. We sketch rigorous proofs in Section 4 (using
approximation arguments) and complete the proof of Theorem 2.1 in Section 5.
1. Gradient estimates
The proof of the gradient estimates listed in Theorem 2.2 relies on a modica-
tion of the Bernstein technique [21]. The cornerstone of such a technique is to nd
a strictly monotone function f such that the equation satised by w := [v[
2
with
v := f
1
(u) has a supersolution which is a sole function of time.
Let us thus consider a strictly monotone function f and put
v := f
1
(u) and w := [v[
2
. (2.12)
It follows from (1.1) that v solves
t
v v
_
f
_
(v) [v[
2
+
1
f
(v)
F
_
(f
(v))
2
[v[
2
_
= 0 ,
with F(r) := r
q/2
for r 0. Next
t
w = 2
N
i=1
i
v
i
v + 2
_
f
_
(v) v w
+ 2
_
f
(v) w
2
+ 2
_
f
2
_
(v) F
_
(f
(v))
2
w
_
w
2 f
(v) F
_
(f
(v))
2
w
_
v w 4 f
(v) F
_
(f
(v))
2
w
_
w
2
.
Using the inequality
w 2
N
i=1
i
v
i
v ,
we obtain
Lw 2
_
f
(v) w
2
+ 2
_
f
2
_
(v)
_
(f
(v))
2
w
_
w 0 , (2.13)
where
Lz :=
t
z z + 2
_
f
(v) F
_
(f
(v))
2
w
_
_
f
_
(v)
_
v z , (2.14)
and
(r) := 2 r F
(r) F(r) = (q 1) r
q/2
, r [0, ) . (2.15)
At this point, assuming that u
0
W
1,
(R
N
) and choosing f(r) = r for r 0,
we nd that L(|u
0
|
) = 0 which, together with (2.13) and the comparison

principle, gives (2.3).
After this preparation we are in a position to establish (2.4)(2.7) by a suitable
choice of the function f.
1.1. Gradient estimates: = 1 and q > 1. In that case, we choose
f(r) = r
q/(q1)
for r 0 and notice that
2
_
f
(v) =
2
q 1
1
v
2
and
_
f
2
_
(v) =
1
qf(v)
. (2.16)
On the one hand, since f and are non-negative and
2
_
f
(v)
2
q 1
|u
0
|
2(q1)/q
by (2.1) it follows from (2.13) that

Lw +
2
q 1
|u
0
|
2(q1)/q
w
2
0 . (2.17)
1. GRADIENT ESTIMATES 107
The function
Y
1
(t) :=
(q 1) |u
0
|
2(q1)/q
2t
, t (0, ) ,
is clearly a supersolution to (2.17) and we deduce from the comparison principle
that w(t, x) Y
1
(t) for (t, x) Q
. We have thus proved (2.6) for s = 0 and use

the autonomous property of (1.1) to obtain (2.6) for t > s 0.
We next turn to the proof of (2.7). For that purpose we infer from (2.13), (2.15)
and (2.16) that
Lw + 2
_
q
q 1
_
q1
w
(q+2)/2
0 . (2.18)
The function
Y
2
(t) :=
_
q 1
q
_
2
(q 1)
2/q
t
2/q
, t (0, ) ,
is a supersolution to (2.18) and (2.7) follows by the comparison principle.
1.2. Gradient estimates: q > 1. We put
u :=
1
2
|u
0
|
+ u
and notice that u solves (1.15), (1.16) and satises 0 u(t, x) |u
0
|
for all
(t, x) Q
. Since
u(t, x) =
q
q 1
u
1/q
u
(q1)/q
(t, x)
and u enjoys the properties (2.6) and (2.7) as shown in the previous section 1.1, we
infer from (2.6) and (2.7) that
|u(t)|

q
q 1
|u
0
|
1/q
_
_
_ u
(q1)/q
(t)
_
_
_
_
q C
3
q 1
|u
0
|
t
1/2
,
q C
4
q 1
|u
0
|
1/q
t
1/q
,
hence (2.4) and (2.5).
1.3. Gradient estimates: q = 1. In that case, the function dened in
(2.15) is equal to zero and we choose f(r) = |u
0
|
r
2
for r
_
0, |u
0
|
1/2
_
. Then
2
_
f
(v) =
2
v
2

2
|u
0
|
and (2.13) reads

Lw +
2
|u
0
|
w
2
0 . (2.19)
The function t |u
0
|
/(2t) is clearly a supersolution to (2.19) and the compar-

ison principle ensures that [v(t, x)[ |u
0
|
1/2
(2t)
1/2
for (t, x) Q
. Finally,
[u(t, x)[ 2 v(t, x) [v(t, x)[
2 (|u
0
|
u(t, x))
1/2
|u
0
|
1/2
t
1/2
2 |u
0
|
t
1/2
,
hence (2.4) for q = 1.
1.4. Gradient estimates: q (0, 1). We choose
f(r) =
1 +
2
|u
0
|
r
2
for r
_
0,
2 |u
0
|
1/2
1 +
_
,
so that
2
_
f
(v) =
2
v
2

2
|u
0
|
,
2
_
f
2
_
(v)
_
(f
(v))
2
w
_
w = 2
q
(1 q) v
q2
w
(q+2)/2
2
q
(1 q) |u
0
|
(q2)/2
w
(q+2)/2
.
We then deduce from (2.13) that
Lw +
2
|u
0
|
w
2
0 , (2.20)
Lw + 2
q
(1 q) |u
0
|
(q2)/2
w
(q+2)/2
0 . (2.21)
On the one hand we proceed as in the previous section to show that (2.4) follows
from (2.20). On the other hand, the function
Y
3
(t) :=
_
2
1q
|u
0
|
(2q)/2
q(1 q)
_
2/q
t
2/q
, t > 0 ,
is a supersolution to (2.21). By the comparison principle, we have w(t, x) Y
3
(t)
for (t, x) Q
from which (2.5) readily follows.

2. Time derivative estimates
For R
N
, we put F() := [[
q
and consider a positive function (
2
(R
N
)
and two real numbers 1, 1 and A R. Introducing
w :=
1
(u)
(
t
u A) , (2.22)
we infer from (1.1) that
t
w =
()(u)
(u)
[u F(u)] w
+
1
(u)
[(u) w (F)(u) (u) w] .
Observing that
()(u) F(u) = (u) (F)(u) ,
we obtain
t
w =
w
(u)
_
_
(u)
i,j
(
i
)(u)
2
j
i
u
_
_
+b w + w,
where
b := 2
(u)
(u)
(F)(u) .
2. TIME DERIVATIVE ESTIMATES 109
Since
(u) =
i,j
(
i
)(u)
2
j
i
u +
i,j,k
(
i
k
)(u)
j
k
u
j
i
u ,
we end up with
t
w b w w =
w
(u)
i,j,k
(
i
k
)()(u)
j
k
u
j
i
u , (2.23)
and we are left to nd appropriate functions for which the quadratic form on the
right-hand side of (2.23) will be negative.
Let (
2
(R) be a convex and even function such that (0) = 0 and
(s) 0
for s 0. For (0, ) and R
N
, [[ |u
0
|
, we put
() := +
N
4
_
_
N
1/2
|u
0
|
i=1
(
i
)
_
, (2.24)
and note that () and (
i
k
)() = 0 if i ,= k and (
2
i
)() = N
(
i
)/4.
Then (2.23) reads
t
w b w w +
N
4(u)
i,j
(
i
u) (
i
j
u)
2
w = 0 . (2.25)
Before going further, let us recall that, by the Cauchy-Schwarz inequality,
N
i,j
(
i
j
u)
2
N
i
(
2
i
u)
2
(u)
2
. (2.26)
(a) (s) = s
2
/, > 0 . Then
(s) = 2/ and
(u)
2
= ((u) w + A+ F(u))
2
.
We then infer from (2.25) that
Lw :=
t
w b w w +
(u)
2
w
3
+
A+ F(u)
w
2
+
1
2(u)
_
_
N
i,j
(
i
j
u)
2
(u)
2
+ (A + F(u))
2
_
_
w = 0 .
We rst take = , = and A = F(|u
0
|
) = |u
0
|
q
. Then (u)
and it follows from (2.3) and (2.26) that L
_
t
1/2
_
0. The comparison principle
then entails that w(t, x) t
1/2
for (t, x) Q
, hence

t
u(t, x) |u
0
|
q
(u)
t
1/2
|u
0
|
q
_
+
N
2
|u
0
|
2
4
_
t
1/2
.
Choosing = |u
0
|
, we conclude that

t
u(t, x) |u
0
|
q
C |u
0
|
t
1/2
for (t, x) Q
. The equation (1.1) being autonomous, we infer from the previous

inequality that

t
u(t, x) |u(t/2)|
q
2 C |u(t/2)|
t
1/2
,
and we use (2.4) and (2.5) to obtain that

t
u(t, x) C |u
0
|
t
1
, (t, x) Q
, (2.27)
if q ,= 1 and

t
u(t, x) C |u
0
|
_
t
1/2
+t
1
_
, (t, x) Q
, (2.28)
if q = 1.
We next take = , = and A = 0. Then (u) and it follows from
(2.3) and (2.26) that L
_
t
1/2
_
0. The comparison principle then entails that
w(t, x) t
1/2
for (t, x) Q
, hence

t
u(t, x)
(u)
t
1/2

_
+
N
2
|u
0
|
2
4
_
t
1/2
.
Choosing = |u
0
|
, we conclude that

t
u(t, x) C |u
0
|
t
1/2
for (t, x) Q
and we proceed as before (with the help of (2.4) and (2.5)) to

conclude that

t
u(t, x) C |u
0
|
t
1
, (t, x) Q
. (2.29)
(b)
(s) = 1/(A+[s[
q
) . In that case,
is non-increasing. Consequently,
(
i
u)
([u[) and
(u)
2
= ((u) w +A + F(u))
2
= 4 (u) (A + F(u)) w + ((u) w A F(u))
2
.
We then infer from (2.25) that
Lw :=
t
w b w w +
([u[) (A + F(u)) w
2
+

([u[)
4(u)
_
_
N
i,j
(
i
u)
([u[)
(
i
j
u)
2
(u)
2
_
_
w
+

([u[)
4(u)
[(u) w A F(u)]
2
w = 0 .
Taking = yields that
([u[) (A+ F(u)) = 1 and thus L

_
t
1
_
0. We
then deduce from the comparison principle that

t
u(t, x) A +
_
+
N
4

_
N
1/2
|u
0
|
_
_
t
1
for (t, x) Q
. Letting 0 in the above inequality, we end up with

t
u(t, x) A +
N
4

_
N
1/2
|u
0
|
_
t
1
, (t, x) Q
. (2.30)
If q (0, 1), we may take A = 0 to obtain that (s) = s
2q
/((2 q)(1 q))
and infer from (2.30) that

t
u(t, x) C |u
0
|
2q
t
1
, (t, x) Q
.
3. HESSIAN ESTIMATES 111
Using once more the fact that (1.1) is an autonomous equation and (2.5) we nally
obtain

t
u(t, x) C |u
0
|
(2q)/q
t
2/q
, (t, x) Q
. (2.31)
If q = 1 we have
(s) =
_
s
0
s z
A +z
dz s log
_
1 +
s
A
_
, s 0 ,
which, together with (2.30), yields

t
u(t, x) A +C |u
0
|
log
_
1 +
|u
0
|
A
_
t
1
, (t, x) Q
.
The choice A = |u
0
|
log (2 +t) t
1
then gives

t
u(t, x) C |u
0
|
1 + log (1 +t)
t
, (t, x) Q
.
Owing to (2.4), we may proceed as before to conclude that

t
u(t, x) C |u
0
|
1 + log (1 +t)
t
3/2
, (t, x) Q
. (2.32)
Finally, if q > 1, we have
(s) =
_
s
0
s z
A+z
q
dz A
(q1)/q
s
_

0
dz
1 +z
q
, s 0 .
Choosing A = |u
0
|
q/(2q1)
t
q/(2q1)
we infer from (2.30) that

t
u(t, x) C |u
0
|
q/(2q1)
t
q/(2q1)
, (t, x) Q
.
Recalling (2.5) we may proceed as before to conclude that

t
u(t, x) C |u
0
|
1/(2q1)
t
(q+1)/(2q1)
, (t, x) Q
, (2.33)
and thus complete the proof of Theorem 2.3.
3. Hessian estimates
Let us rst handle the case q (1, 2].
Proof of Theorem 2.4. For 1 i, j N, we put w
ij
:=
i
j
u. It follows
from (1.1) that
t
w
ij
w
ij
= q
i
_
[u[
q2
_
N
k=1
k
u w
jk
__
= q [u[
q2
N
k=1
w
ik
w
jk
q [u[
q2
N
k=1
k
u
i
w
jk
q (q 2) [u[
q4
_
N
k=1
k
u w
ik
_ _
N
k=1
k
u w
jk
_
. (2.34)
Consider now R
N
0 and set
w :=
N
i=1
N
j=1
w
ij

i

j
.
Multiplying (2.34) by
i

j
and summing up the resulting identities yield
t
w w = q [u[
q2
N
k=1
_
N
i=1
w
ik

i
_
2
q [u[
q2
u w
q (q 2) [u[
q4
_
_
N
i=1
N
j=1
j
u w
ij

i
_
_
2
. (2.35)
Thanks to the following inequalities
[u[
q4
_
_
N
i=1
N
j=1
j
u w
ij

i
_
_
2
[u[
q4
N
j=1
[
j
u[
2
N
j=1
_
N
i=1
w
ij

i
_
2
[u[
q2
N
k=1
_
N
i=1
w
ik

i
_
2
,
and
w
2
[[
2
N
k=1
_
N
i=1
w
ik

i
_
2
,
and since q 2, the right-hand side of (2.35) can be bounded from above. We thus
obtain
t
w w q (q 1) [u[
q2
N
k=1
_
N
i=1
w
ik

i
_
2
q [u[
q2
u w
q [u[
q2
u w
q (q 1) [u[
q2
[[
2
w
2
.
Consequently,
Lw 0 in R
N
(0, ) , (2.36)
where the parabolic dierential operator L is given by
Lz := z
t
z +q [u[
q2
u z +
q (q 1) [u[
q2
[[
2
z
2
.
On the one hand, since q (1, 2] and [u(t, x)[ |u
0
|
, it is straightforward
to check that
W
1
(t) :=
_
1
|w(0)|
+
q (q 1) t
[[
2
|u
0
|
2q
_
1
, t > 0 ,
satises LW
1
0 with W
1
(0) w(0, x) for all x R
N
. The comparison principle
then entails that w(t, x) W
1
(t) for (t, x) Q
, from which we conclude that

w(t, x) |w(0)|
|D
2
u
0
|
[[
2
,
whence (2.10). Observe that we also obtain that
w(t, x)
[[
2
|u
0
|
2q
q (q 1) t
for (t, x) Q
. (2.37)
3. HESSIAN ESTIMATES 113
On the other hand, we infer from (2.5) that
W
2
(t) :=
2 C
2q
2
[[
2
q
2
(q 1)
|u
0
|
(2q)/q
t
2/q
, t > 0 ,
satises LW
2
0 with W
2
(0) = w(0, x) for all x R
N
. We then use again
the comparison principle as above and obtain (2.9).
For further use, we report the following weaker version of Theorem 2.4.
Corollary 2.6. Under the assumptions of Theorem 2.4 there is C
9
such that
u(t, x)
C
9
|u
0
|
(2q)/q
t
2/q
, (2.38)
for (t, x) (0, ) R
N
.
In addition, if u
0
W
2,
(R
N
),
sup
xR
N
u(t, x) sup
xR
N
u
0
(x) , t 0 . (2.39)
Proof. Consider i 1, . . . , N and dene
i
= (
i
j
) R
N
by
i
i
= 1 and
i
j
= 0 if j ,= i. We take =
i
in (2.36) and obtain that L(
2
i
u) 0, that is,
t
_

2
i
u
_
_

2
i
u
_
+q [u[
q2
u
_

2
i
u
_
+q (q1) [u[
q2
_

2
i
u
_
2
0
in Q
. Summing the above inequality over i 1, . . . , N and recalling that

[u[
2
N
N
i=1
_
2
i
u
_
2
,
we end up with
( u)
t
( u) +q [u[
q2
u ( u) +
q (q 1) [u[
q2
N
[ u[
2
0
in Q
. We next proceed as in the proof of Theorem 2.4 to complete the proof of

Corollary 2.6.
We next establish the log-semiconvexity of u in the particular case = 1 and

q = 1.
Proof of Proposition 2.5. We follow the approach of [34] and introduce
v := log u. We deduce from (1.1) and the positivity of u [32, Corollary 4.2] that v
solves
t
v v = [v[ +[v[
2
in Q
. (2.40)
We next put w
ij
:=
i
j
v for 1 i, j N, P := 1/[v[ and deduce from (2.40)
that w
ij
solves
t
w
ij
= w
ij
+ 2
N
k=1
w
ik
w
jk
+ 2 v w
ij
+P
N
k=1
w
ik
w
jk
+ P v w
ij
P
3
_
N
k=1
w
ik

k
v
_ _
N
l=1
w
jl

l
v
_
Consider now R
N
0 and put
w :=
N
i,j=1
w
ij

i

j
.
Then
t
w = w + 2
N
k=1
_
N
i=1
w
ik

i
_
2
+ 2 v w +P
N
k=1
_
N
i=1
w
ik

i
_
2
+ P v w P
3
_
_
N
i,k=1
w
ik

i

k
v
_
_
2
.
By the Cauchy-Schwarz inequality, we have
_
_
N
i,k=1
w
ik

i

k
v
_
_
2
[v[
2
N
k=1
_
N
i=1
w
ik

i
_
2
.
Therefore, we may estimate from below the last term of the right-hand side of the
equation satised by w and obtain
t
w w + (2 +P) v w + 2
N
k=1
_
N
i=1
w
ik

i
_
2
.
Using once more the Cauchy-Schwarz inequality, we realize that
w
2
[[
2
N
k=1
_
N
i=1
w
ik

i
_
2
,
whence
t
w w + (2 +P) v w +
2 w
2
[[
2
.
The comparison principle then ensures that
w(t, x)
[[
2
2t
, (t, x) Q
,
which gives the result.
4. Existence
Two dierent approaches have been used to investigate the well-posedness of
(1.1), (1.2), one relying on the comparison principle and the previous gradient esti-
mates [11, 32] and the other one on the integral formulation (2.2) and the smooth-
ing properties of the heat semigroup (e
t
)
t0
[20]. Both methods actually allow to
handle dierent classes of initial data but in dierent ranges of the parameter q.
We sketch the rst approach which is actually a compactness method and is thus
based on a sequence of approximations to (1.1), (1.2). In view of the transformation
(1.14), we only need to consider the case = 1. We thus assume that = 1 and
4. EXISTENCE 115
consider a non-negative function u
0
B((R
N
). There is a sequence of functions
(u
0,k
)
k1
such that, for each integer k 1, u
0,k
B(
(R
N
),
0 u
0,k
(x) u
0,k+1
(x) u
0
(x) , x R
N
, (2.41)
and (u
0,k
) converges uniformly towards u
0
on compact subsets of R
N
. In addition,
if u
0
W
1,
(R
N
) we may assume that
|u
0,k
|

_
1 +
K
1
k
_
|u
0
|
, (2.42)
for some constant K
1
> 0 depending only on the approximation process. Next,
since [[
q
, R
N
, is not regular enough for small values of q, we set
F
(r) :=
_
2
+r
_
q/2
q
, r 0 , (2.43)
for (0, 1). Then the Cauchy problem
t
u
k,
u
k,
+F
_
[u
k,
[
2
_
= 0 , (t, x) Q
, (2.44)
u
k,
(0) = u
0,k
+ , x R
N
, (2.45)
has a unique classical solution u
k,
(
(3+)/2,3+
([0, ) R
N
) for some (0, 1)
[38]. Observing that and |u
0
|
+ are solutions to (2.44) with u

k,
(0, x)
|u
0
|
+ , the comparison principle warrants that

u
k,
(t, x) |u
0
|
+ , (t, x) [0, ) R
N
.
Moreover, arguing as in Section 1, we obtain similar gradient estimates for u
k,
which does not depend on k 1 and have a mild dependence on (0, 1). More
precisely, introducing v
k,
:= f
1
(u
k,
) and w
k,
:= [v
k,
[
2
for some strictly
monotone function f to be specied, we proceed along the lines of Section 1 to
establish that:
if q > 1 (with f(r) = r
q/(q1)
),
u
(q1)/q
k,
(t, x)

_
q 1
2
_
1/2
(|u
0
|
+)
(q1)/q
t
1/2
, t > 0 ,
if q > 2 (with f(r) = r
q/(q1)
),
u
(q1)/q
k,
(t, x)

_
2
qK
2
()
+
K
3
()
K
2
()

(q1)/2
_
1/q
t
1/q
, t
_
0,
(1q)/2
_
,
for every (0, q 1) with
K
2
() :=
2(q 1 )
q
_
q
q 1
_
q
and K
3
() :=
4
q
_
q 2
_
(q2)/2
,
if q (1, 2] (with f(r) = r
q/(q1)
),
u
(q1)/q
k,
(t, x)

q 1
q
_
1
q 1
+
(q1)/2
_
1/q
t
1/q
, t
_
0,
(1q)/2
_
,
if q (0, 1] (with f(r) = |u
0
|
+ +
q/2
r
2
),
[u
k,
(t, x)[
_
2 + 2
q/4
_
1/2
_
|u
0
|
+
q/2
_
t
1/2
, t
_
0,
q/4
_
,
if q (0, 1) (with f(r) = |u
0
|
+ +
q/2
r
2
),
[u
k,
(t, x)[
_
2 +q
q/4
2
q
q(1 q)
_1/q
_
|u
0
|
+ +
q/2
_
1/q
t
1/q
, t
_
0,
q/4
_
.
In addition, we have
|u
k,
(t)|
|u
0,k
|
, t 0 .
Combining these estimates with classical parabolic regularity results we rst
let 0 and then k to obtain the existence part of Theorem 2.1, together
with Theorem 2.2. The proofs of Theorems 2.3, 2.4 and Proposition 2.5 are done
in a similar way.
5. Uniqueness
So far, given u
0
B((R
N
), we have constructed a solution u to (1.1), (1.2) en-
joying the properties stated in Theorems 2.2, 2.3, 2.4 and Proposition 2.5. To prove
the uniqueness assertion of Theorem 2.1, we follow the proof of [32, Theorem 4]
and show that any other solution to (1.1), (1.2) in the sense of Theorem 2.1 coin-
cides with the solution u already obtained in the previous section. More precisely,
assume that
v B(([0, ) R
N
) (
1,2
(Q
)
is such that

t
v v + [v[
q
0 , (t, x) Q
, (2.46)
and
v(0, x) u
0
(x) , x R
N
, (2.47)
for some 1, +1.
Then we claim that
v(t, x) u(t, x) , (t, x) [0, ) R
N
. (2.48)
Indeed, x T > 0, (0, 1) and put := min 1/q, 1 and
A
:=
_
_
2q
_
N T
(q1)/q
+q
_
C
2
|u
0
|
1/q
+ T
1/q
_
q1
_
if q > 1 ,
2 (
q
+ N ) if q (0, 1] .
Introducing
z(t, x) := (u v)(t, x) A
_
1 +[x[
2
_
1/2
, (t, x) [0, T) R
N
,
we infer from the boundedness of u and v and (2.47) that there is R
> 0 such that

z(t, x) < 0 for t [0, T] and [x[ R
and also for (t, x) 0 R

N
. Conse-
quently, if z is positive somewhere in [0, T] R
N
it must have a positive maximum
at some point (t
0
, x
0
) (0, T] [x[ R
. This implies that z(t

0
, x
0
) = 0,
z(t
0
, x
0
) 0 and
t
z(t
0
, x
0
) 0 so that
t
z(t
0
, x
0
) z(t
0
, x
0
) 0 . (2.49)
Now, by (1.1) and (2.46) we have
t
z(t
0
, x
0
) z(t
0
, x
0
)
[v(t
0
, x
0
)[
q
[u(t
0
, x
0
)[
q
t
1
0
+N
BIBLIOGRAPHICAL NOTES 117
and we use the relation z(t
0
, x
0
) = 0 to compute v(t
0
, x
0
) and obtain
t
z(t
0
, x
0
) z(t
0
, x
0
)
u(t
0
, x
0
)
x
0
(1 +[x
0
[
2
)
1/2
q
[u(t
0
, x
0
)[
q
t
1
0
+N .
Either q > 1 and it follows from the mean value theorem, (2.5) and the choice of
A
that
t
z(t
0
, x
0
) z(t
0
, x
0
) q
_
[u(t
0
, x
0
)[ +
[x
0
[
(1 +[x
0
[
2
)
1/2
_
q1
[x
0
[
(1 +[x
0
[
2
)
1/2
q
t
(q1)/q
0
+N
q
_
C
2
|u
0
|
1/q
t
1/q
0
+
_
q1
q
t
(q1)/q
0
+N

A
2q
t
1
0
< 0 ,
which contradicts (2.49). Or q (0, 1] and we deduce from the H older continuity
of r r
q
and the choice of A
that
t
z(t
0
, x
0
) z(t
0
, x
0
)
_
[x
0
[
(1 +[x
0
[
2
)
1/2
_
q
A
+N
A
2
< 0 ,
again contradicting (2.49). Therefore, z cannot take positive values in [0, T] R
N
and thus
u(t, x) v(t, x) + A
+
_
1 +[x[
2
_
1/2
, (t, x) [0, T) R
N
.
Since A
0 as 0 we may pass to the limit as 0 in the previous inequality

and conclude that (2.48) holds true.
The uniqueness statement of Theorem 2.1 then readily follows.
Bibliographical notes
The gradient estimates (2.5) and (2.7) (with = 1) are also true for non-
negative viscosity solutions to the non-diusive Hamiton-Jacobi equation (1.9): this
fact is proved in [42] by a dierent method, still using the comparison principle.
No such gradient estimate seems to be available for q = 1 as a consequence of the
lack of strict convexity (or concavity) of [u[.
An alternative approach to study the well-posedness of (1.1), (1.2) in Lebesgue
spaces L
p
(R
N
), p 1, is employed in [20]. It relies on the formulation (2.2) of
(1.1), (1.2) which is the starting point of a xed point procedure. This approach is
restricted to q [1, 2) but provides the existence and uniqueness of a weak solution
(which is classical for positive times) for u
0
L
r
(R
N
) whenever r 1 and
r >
N(q 1)
2 q
or r =
N(q 1)
2 q
> 1 .
When q = 1, previous existence and uniqueness results were obtained in [15, 16,
17, 18].
CHAPTER 3
Extinction in nite time
Let us rst recall that, if X is a vector space, a solution f : [0, ) X to
an evolution problem enjoys the extinction in nite time property if there exists
T
> 0 such that

_
_
_
f(t) ,= 0 for t [0, T
) ,
f(t) 0 for t T
.
A typical example of an evolution problem which exhibits the previous property
is the ordinary dierential equation f
(t) + f(t)
= 0 with f(0) = f
0
> 0 and
(0, 1). Then, for t 0,
f(t) = (1 )
1/(1)
(T
t)
1/(1)
+
with T
:=
f
1
0
1
.
Extinction in nite time also shows up for non-negative solutions to the fast diusion
equation
t
v = v
m
in R
N
if m (0, (N2)
+
/N) and to the p-Laplacian equation
t
v = div
_
[v[
p2
v
_
in R
N
if p (0, 2N/(N +1)) (see, e.g., [22, 35]). Noticing
that the exponent below which extinction in nite time takes place is smaller than
one in the previous examples, it is rather natural to wonder whether extinction
phenomena could also occur for non-negative solutions to (1.1), (1.2) when q
(0, 1). It is however obvious that extinction in nite time cannot take place if
= 1: indeed, in that case, we have u(t) e
t
u
0
for each t 0 by the comparison
principle, which readily implies that u(t, x) > 0 for (t, x) Q
if u
0
, 0.
Thus, throughout this chapter we assume that
= 1 and q (0, 1) . (3.1)
There are two dierent conditions on the initial condition u
0
which warrant that
the corresponding classical solution u to (1.1), (1.2) vanishes after a nite time: the
rst one requires some integrability at innity for u
0
while the second involves a
pointwise bound.
Theorem 3.1. Consider a non-negative function u
0
B((R
N
) and denote by
u the corresponding classical solution to (1.1), (1.2). If u
0
L
1
(R
N
; [x[
m
dx) for
some m 0 and q (0, (m+N)/(m+N + 1)), then there exists T
[0, ) such
that
u(t, x) = 0 for (t, x) [T
, ) R
N
.
0
N
) satisfying
limsup
|x|
[x[
q/(1q)
u
0
(x) < . (3.2)
119
120 3. EXTINCTION IN FINITE TIME
Denoting by u the corresponding classical solution to (1.1), (1.2), there exists
T
[0, ) such that

u(t, x) = 0 for (t, x) [T
, ) R
N
.
Let us rst point out that the sets of initial data involved in Theorems 3.1
and 3.2 have a non-empty intersection but are dierent. Indeed, both results apply
for compactly supported initial data but, if m 0 and q (0, (m+N)/(m+N+1)),
a function u
0
satisfying (3.2) need not be in L
1
(R
N
; [x[
m
dx). Next, as we shall see
below, the proofs of Theorems 3.1 and 3.2 rely on dierent arguments : indeed the
latter is achieved by constructing suitable supersolutions while the former follows
from a dierential inequality involving the L
-norm of u.
The next issue to be considered is the optimality of the algebraic growth con-
dition (3.2). In that direction, we report the following result:
Theorem 3.3. [14] Consider u
0
B((R
N
) satisfying
u
0
(x) > 0, x R
N
, (3.3)
lim
|x|
[x[
q/(1q)
u
0
(x) = . (3.4)
Denoting by u the corresponding classical solution to (1.1), (1.2), we have u(t, x) >
0 for every (t, x) Q
.
It turns out that Theorem 3.3 also guarantees the optimality of the exponent
(m + N)/(m + N + 1) in Theorem 3.1: indeed, if m 0 and q ((m + N)/(m +
N +1), 1), there is at least a non-negative function u
0
L
1
(R
N
; [x[
m
dx) such that
the corresponding classical solution u to (1.1), (1.2) satises u(t, x) > 0 for every
(t, x) Q
(see Corollary 3.10 below).

Remark 3.4. An explicit computation shows that, if T > 0 and K [0, ),
the function z dened by
z(t, x) := (T t)
1/(1q)
+
_
K
(1q)/q
+
(1 q)
(q1)/q
q
[x[
_q/(1q)
for (t, x) Q
, is a solution to the HamiltonJacobi equation

t
z + [z[
q
= 0 in
Q
. Observe that z(t, x) > 0 for every (t, x) [0, T) R

N
and z(T) 0. The
supersolutions and subsolutions we will construct in the proofs of Theorems 3.2
and 3.3 are somehow related to this function.
We nally establish a lower bound for the L
-norm of u near the extinction

time.
Proposition 3.5. Let u
0
N
) with compact
support and u denote the corresponding classical solution to (1.1), (1.2). There is
a positive real number depending only on N, q and [supp u
0
[ such that
|u(t)|
(T
t)
1/(1q)
, t [0, T
] , (3.5)
where T
(0, ) is the extinction time of u. Here, [supp u

0
[ denotes the (Lebesgue)
measure of supp u
0
.
1. AN INTEGRAL CONDITION FOR EXTINCTION 121
1. An integral condition for extinction
Consider m 0 and a non-negative function u
0
L
1
(R
N
; [x[
m
dx) B((R
N
).
We denote by u the corresponding classical solution to (1.1), (1.2) and rst establish
that u(t) still belongs to L
1
(R
N
; [x[
m
dx) for t > 0.
Lemma 3.6. There exists K
1
> 0 depending only on m, N, q and u
0
such that
_
R
N
[x[
m
u(t, x) dx K
1
, t 0 . (3.6)
Proof. For x R
N
and > 0, we put b
(x) :=
_
1 + [x[
2
_
1/2
. We rst
multiply (1.1) by b
1
(x)
m
and integrate over R
N
: owing to the non-negativity of u
and = 1, we obtain
d
dt
_
R
N
b
1
(x)
m
u(t, x) dx
m
_
R
N
b
1
(x)
m2
x u(t, x) dx
_
R
N
_
Nm b
1
(x)
m2
+m(m2) [x[
2
b
1
(x)
m4
_
u(t, x) dx
C
_
R
N
b
1
(x)
m
u(t, x) dx,
whence
_
R
N
b
1
(x)
m
u(t, x) dx C
_
R
N
b
1
(x)
m
u
0
(x) dx, t [0, 1] . (3.7)
Next, for t 1, we multiply (1.1) by b
(x)
m
and integrate over R
N
to obtain
d
dt
_
R
N
b
(x)
m
u(t, x) dx +
_
R
N
b
(x)
m
[u(t, x)[
q
dx
m
_
R
N
b
(x)
m2
[x[ [u(t, x)[ dx
m
1/2
|u(t)|
1q
_
R
N
b
(x)
m
[u(t, x)[
q
dx.
Since |u(t)|
C
2
|u
0
|
1/q
for t 1 by (2.5), we end up with

d
dt
_
R
N
b
(x)
m
u(t, x) dx
+
_
1 m C
1q
2
|u
0
|
(1q)/q

1/2
_
_
R
N
b
(x)
m
[u(t, x)[
q
dx 0
for t 1. Choosing = 1/
_
m
2
C
2(1q)
2
|u
0
|
2(1q)/q
_
, we conclude that
_
R
N
b
(x)
m
u(t, x) dx
_
R
N
b
(x)
m
u(1, x) dx, t 1 . (3.8)
Lemma 3.6 now readily follows from (3.7) and (3.8) as b
(x)
m
[x[
m
.

Another useful tool for the proof of Theorem 3.1 is the following interpolation
inequality:
Lemma 3.7. If w L
1
(R
N
; [x[
m
dx) L
(R
N
), then w L
1
(R
N
) and there
is a constant K
2
> 0 depending only on N such that
|w|
1
K
2
|w|
m/(m+N)
__
R
N
[x[
m
w(x) dx
_
N/(m+N)
.
Proof. Consider w L
1
(R
N
; [x[
m
dx) L
(R
N
) and R > 0. Then
|w|
1
=
_
{|x|R}
[w(x)[ dx +
_
{|x|>R}
[w(x)[ dx
C R
N
|w|
+
1
R
m
_
R
N
[x[
m
w(x) dx.
Choosing
R =
__
R
N
[x[
m
w(x) dx
_
1/(m+N)
|w|
1/(m+N)
yields Lemma 3.7.
Proof of Theorem 3.1. Owing to (2.5), (3.6), Lemma 3.7 and the Gagliardo
Nirenberg inequality
|w|
C |w|
N/(N+1)
|w|
1/(N+1)
1
for w L
1
(R
N
) W
1,
(R
N
) , (3.9)
we have for t > s 0
|u(t)|
q
C |u(t)|
qN/(N+1)
|u(t)|
q/(N+1)
1
C (t s)
N/(N+1)
|u(s)|
N/(N+1)
|u(t)|
qm/((N+1)(N+m))
__
R
N
[x[
m
u(t, x) dx
_
qN/((N+1)(N+m))
C (t s)
N/(N+1)
|u(s)|
N/(N+1)
|u(t)|
qm/((N+1)(N+m))
,
whence
|u(t)|
q
C (t s)
(m+N)/(m+N+1)
|u(s)|
(m+N)/(m+N+1)
.
We multiply the above inequality by 1/t and integrate with respect to t over (s, )
to deduce that
Z(s) :=
_

s
|u(t)|
q
t
dt C s
(m+N)/(m+N+1)
|u(s)|
(m+N)/(m+N+1)
.
Since dZ/ds(s) = |u(s)|
q
/s, the above inequality also reads

s
1q
dZ
ds
(s) +C Z(s)
q(m+N+1)/(m+N)
0 , s 0 .
Introducing

Z(s) := Z
_
s
1/q
_
for s 0, we readily obtain
d
Z
ds
(s) +C

Z(s)
q(m+N+1)/(m+N)
0 , s 0 ,
2. A POINTWISE CONDITION FOR EXTINCTION 123
whence

Z(s) = 0 for s suciently large since q < (m+N)/(m+N+1). Consequently
u(t) = 0 for t large enough.
2. A pointwise condition for extinction

We introduce the following positive real numbers
=
2 q
2(1 q)
and = 1 =
q
2(1 q)
. (3.10)
A classical tool to establish extinction in nite time is to construct supersolutions
to (1.1) enjoying such a property and in fact to look for self-similar supersolutions.
Lemma 3.8. Consider T (0, ) and B 1 such that
B
q/2
+ + 4 ( + 1)
min , (2)
q
. (3.11)
The function W
B,T
dened by
W
B,T
(t, x) = (T t)
f
B
_
[x[ (T t)
1/2
_
, (t, x) [0, T] R
N
,
f
B
(y) =
_
A +B y
2
_
, y R, with A =

B,
is a supersolution to (1.1) in [0, T) R
N
.
Proof. We denote by L the parabolic operator dened by
Lw =
t
w w +[w[
q
, (3.12)
and put
y
0
=
_

B
_
1/2
.
Let (t, x) [0, T) R
N
and compute LW
B,T
(t, x). Putting y = [x[ (T t)
1/2
we
obtain
LW
B,T
(t, x) = (T t)
1
_
A+B y
2
_
1
H(y),
where
H(y) := 2 N B A + (2 B)
q
y
q
_
A +B y
2
_
(2q)/2
( +) B y
2
4 ( + 1)
B
2
y
2
A+B y
2
B + (2 )
q
B
(2+q)/2
y
2
( +) B y
2
4 ( + 1)
B
2
y
2
A +B y
2
.
Now, on the one hand, we have for y [0, y
0
]
H(y) B ( +) B y
2
0
4 ( + 1)
B
2
y
2
0
A +B y
2
0
B ( +)

4 ( + 1)
B
1 + B
.
As < and B B
q/2
1 we further obtain, thanks to (3.11),
H(y) B ( +) 4 ( + 1) 0.
On the other hand, there holds for y y
0
H(y) (2 )
q
B
(2+q)/2
y
2
( +) B y
2
4 ( + 1)
B
2
y
2
A +B y
2
_
(2 )
q
B
q/2
( +) 4 ( + 1)
B
A +B y
2
0
_
B y
2
_
(2 )
q
B
q/2
( +) 4 ( + 1)
B
1 + B
_
B y
2
,
hence H(y) 0 by (3.11). The proof of Lemma 3.8 is thus complete.
Proof of Theorem 3.2. Owing to (3.2) and the boundedness of u

0
there is
a constant C
0
> 0 such that
u
0
(x) C
0
(1 +[x[)
q/(1q)
, x R
N
.
Since (1 +[x[)
2
1 +[x[
2
for x R
N
we further obtain that u
0
satises
u
0
(x) C
0
_
1 +[x[
2
_
, x R
N
,
the real number being dened in (3.10). We next x B 1 such that (3.11) holds
true and consider T > 0 such that
T 1 and T C
1/
0
B
/
. (3.13)
With this choice of the parameters T and B the function W
B,T
dened in Lemma 3.8
is a supersolution to (1.1). In addition, > and we infer from (3.13) that, for
x R
N
,
W
B,T
(0, x) = T
+
[x[
2
T
_
C
0
_
1 +[x[
2
_
u
0
(x) .
The comparison principle [32, Theorem 4] then ensures that u(t, x) W
B,T
(t, x) for
(t, x) [0, T]R
N
. Therefore, since u is non-negative and W
B,T
vanishes identically
at time T, we conclude that u(T) 0, whence u(t, x) = 0 for (t, x) [T, ) R
N
.
3. Non-extinction
The proof of Theorem 3.3 also relies on a comparison argument, but now with
suitable subsolutions.
Lemma 3.9. Let T, a and b be three positive real numbers and dene
w
a,b,T
(t, x) = (T t)
1/(1q)
_
a +b [x[
2
_
, (t, x) [0, T] R
N
,
the parameter being dened in (3.10). If
:= b
q/2
q
q
(1 q)
1q
1 < 0 and b
(1 q) [[
N q (N + 2 q)
a
T
, (3.14)
then the function w
a,b,T
is a subsolution to (1.1) on [0, T] R
N
.
3. NON-EXTINCTION 125
Proof. The parabolic operator L being dened by (3.12), we compute Lw
a,b,T
.
Let (t, x) [0, T] R
N
and put y = a +b [x[
2
. Since b[x[
2
y, we have
Lw
a,b,T
(t, x)
(T t)
q/(1q)
1 q
y
2q/(22q)
F(y),
where
F(y) := y
2
+b q (T t)
_
a (2 q)
1 q
+
(N 2) q (N 1)
1 q
y
_
.
Now we have
F(y) y
2
+b q (T t)
_
a (2 q)
1 q
+
N
1 q
y
_
y
2
+
b N q T
1 q
y +
a b q (2 q) T
1 q
=: G(y).
As G
= 2 < 0 by (3.14) and y a, we infer from (3.14) that

G
(y) G
(a) = 2 a +
b N q T
1 q
0 .
Consequently, G is a non-increasing function on [a, ) and (3.14) ensures that
G(y) G(a) = a
_
a +
b q T
1 q
(N + 2 q)
_
0
for y [a, ). Therefore Lw
a,b,T
(t, x) 0 and the proof of Lemma 3.9 is complete.
Proof of Theorem 3.3. We x b (0, 1) such that b

q/2
q
q
(1 q)
1q
1/2
and T > 0. On the one hand it follows from (3.4) that there is R
T
> 0 such that
u
0
(x)
_
b T
2/q
[x[
2
_
q/(22q)
, [x[ R
T
. (3.15)
On the other hand the continuity of u
0
and (3.3) ensure that
m
T
:= min
{|x|RT }
u
0
(x) > 0 .
We then dene a(T) by
a(T) = T
2/q
m
(22q)/q
T
+
N q (N + 2 q)
(1 q) [[
b T, (3.16)
where is dened in (3.14) and notice that
_
a(T) T
2/q
_
q/(22q)
m
T
. (3.17)
Owing to (3.16) and Lemma 3.9 the function w
a(T),b,T
is a subsolution to (1.1) on
[0, T] R
N
and we infer from (3.15) and (3.17) that w
a(T),b,T
(0, x) u
0
(x) for
x R
N
. The comparison principle [32, Theorem 4] then entails that
w
a(T),b,T
(t, x) u(t, x), (t, x) [0, T] R
N
.
Therefore u(T/2, x) > 0 for each x R
N
and, as T is arbitrary in (0, ), Theo-
rem 3.3 follows.

Corollary 3.10. Consider m 0 and q ((m + N)/(m + N + 1), 1). Then
there is at least a non-negative function u
0
B((R
N
) L
1
(R
N
; [x[
m
dx) such
that the corresponding classical solution u to (1.1), (1.2) satises u(t, x) > 0 for
(t, x) [0, ) R
N
.
In other words, the assertion of Theorem 3.1 is false for q ((m + N)/(m +
N + 1), 1).
Proof. The condition on q implies that m + N < q/(1 q). We may then
choose (0, ) such that
m+N <
q
1 q
.
We next put
u
0
(x) = (1 +[x[)
q/(1q)
, x R
N
.
Thanks to the choice of , u
0
satises the assumptions of Theorem 3.3 and u
0
belongs to L
1
(R
N
; [x[
m
dx), which yields the expected result.
4. A lower bound near the extinction time
Since u
0
is compactly supported, it clearly satises (3.2) and Theorem 3.2
implies that T
< . Introducing the positivity set

T(t) :=
_
x R
N
: u(t, x) > 0
_
of u at time t 0, we proceed as in [31, Theorem 9] to prove that
T(t)
_
x R
N
: d(x, T(0)) <
_
|u
0
|
A
0
_
(1q)/(2q)
_
(3.18)
where A
0
:= (1 q)
(2q)/(1q)
(N(1 q) + q)
1/(1q)
/(2 q). Indeed, con-
sider x
0
R
N
such that d(x
0
, T(0))
(2q)/(1q)
|u
0
|
/A
0
. Introducing S(x) :=
A
0
[x x
0
[
(2q)/(1q)
for x R
N
, we have u
0
(x) = 0 S(x) if x , T(0) and
u
0
(x) |u
0
|
A
0
d(x
0
, T(0))
(2q)/(1q)
A
0
[x x
0
[
(2q)/(1q)
= S(x)
if x T(0), the last inequality following from the choice of x
0
. Therefore,
u
0
(x) S(x) for x R
N
and S is actually a stationary solution to (1.1). The
comparison principle then entails that u(t, x) S(x) for (t, x) [0, ) R
N
. In
particular, u(t, x
0
) S(x
0
) = 0 for t 0 which, together with the non-negativity
of u, implies that u(t, x
0
) = 0 for t 0 and completes the proof of (3.18).
Consider next t [0, T
) and s [0, t). On the one hand, the gradient estimate

(2.5), (3.18) and the GagliardoNirenberg inequality (3.9) give
|u(t)|
C |u(t)|
N/(N+1)
|u(t)|
1/(N+1)
1
C |u(s)|
N/q(N+1)
(t s)
N/q(N+1)
|u(t)|
1/(N+1)
[T(t)[
1/(N+1)
C([T(0)[) |u(s)|
N/q(N+1)
(t s)
N/q(N+1)
|u(t)|
1/(N+1)
.
As t < T
, we have |u(t)|
> 0 and we deduce from the previous inequality that

(t s) |u(t)|
q
C([T(0)[) |u(s)|
, 0 s < t < T
. (3.19)
Now, for T (0, T
), we put
m(T) := inf
s[0,T)
_
|u(s)|
(T s)
1/(1q)
_
> 0 .
We infer from (3.19) that, for s [0, T) and t (s, T),
(t s)(T t)
q/(1q)
(T s)
1/(1q)
_
|u(t)|
(T t)
1/(1q)
_
q
=
t s
(T s)
1/(1q)
|u(t)|
q
C([T(0)[)
|u(s)|
(T s)
1/(1q)
.
Choosing t = (1 q)T +qs (s, T) in the previous inequality leads us to
C([T(0)[) m(T)
q
|u(s)|
(T s)
1/(1q)
, s [0, T) .
Therefore, C([T(0)[) m(T)
q
m(T) and the positivity of m(T) allows us to
conclude that m(T) C([T(0)[) > 0. We have thus proved that |u(s)|

C([T(0)[) (T s)
1/(1q)
for s [0, T) and T (0, T
) with a positive constant

C([T(0)[) which does not depend on T. We then let T T
in the previous
inequality to complete the proof of Proposition 3.5.
Theorem 3.1 improves [13, Theorem 1] which only deals with the case m = 0.
The proof given here is also simpler. The extinction in nite time of non-negative
solutions to (1.1), (1.2) with a compactly supported initial condition u
0
is also
established in [31, Corollary 9.1] by a comparison argument but with a dierent
supersolution (travelling wave). Proposition 3.5 seems to be new and is a rst step
towards a better understanding of the behaviour near the extinction time.
CHAPTER 4
Temporal decay estimates for integrable initial
data: = 1
Throughout this section we assume that
= 1 and u
0
is a non-negative function in L
1
(R
N
) B((R
N
) (4.1)
and denote by u the corresponding classical solution to (1.1), (1.2). We then in-
vestigate the time behaviour of the L
p
-norms of u for p = 1 and p = . As a rst
step, we check that u(t) actually remains in L
1
(R
N
) for t > 0.
Lemma 4.1. If q > 0, then u (([0, ), L
1
(R
N
)) and t |u(t)|
1
is a
non-increasing function of time with
I
1
() := lim
t
|u(t)|
1
[0, |u
0
|
1
] . (4.2)
In addition, [u[
q
L
1
(Q
) and
_

0
_
R
N
[u(t, x)[
q
dxdt |u
0
|
1
. (4.3)
We now turn to the time evolution of the L
p
-norms of u for p = 1 and p = .
The main issue here is to gure out whether the additional dissipative mechanism
[u[
q
speeds up the convergence to zero in L
p
(R
N
) for p = . For that purpose,
we introduce the following positive real numbers:
q
:=
N + 2
N + 1
, (4.4)
_
_
a :=
(N + 1)(q
q)
2(q 1)
for q (1, q
)
b :=
1
q(N + 1) N
for q >
N
N + 1
,
(4.5)
and rst study the behaviour of the L
-norm of u.
Proposition 4.2. [13]
(a): If q = N/(N + 1), then there is a constant C
10
such that
|u(t)|
C
10
|u
0
|
exp
_
C
10
|u
0
|
1/(N+1)
1
t
_
, t (0, ) . (4.6)
(b): If q (N/(N + 1), q
], then there is a constant C

10
such that
|u(t)|
C
10
|u
0
|
qb
1
t
Nb
, t (0, ) . (4.7)
129
130 4. TEMPORAL DECAY ESTIMATES FOR INTEGRABLE INITIAL DATA: = 1
(c): If q q
, then there is a constant C

10
such that
|u(t)|
C
10
|u
0
|
1
t
N/2
, t (0, ) . (4.8)
Observe that, for q [N/(N + 1), q
) the temporal decay estimates (4.6) and

(4.7) are faster for large times than the one which follows from the heat equation by
comparison, thus showing the inuence of the absorption term when q < q
. The
decay to zero is even faster for q (0, N/(N + 1)) as u vanishes identically after a
nite time by Theorem 3.1 (with m = 0). For q > q
, the absorption term does not

speed up the convergence of |u(t)|
to zero.
We next investigate the behaviour of the L
1
-norm of u.
Proposition 4.3. [11, 14, 19] We have
I
1
() > 0 q > q
, (4.9)
where the critical exponent q
is given by (4.4).
Recalling that the L
1
-norm of non-negative solutions to the linear heat equation
remains constant (and positive if u
0
, 0) throughout time evolution, a consequence
of Proposition 4.3 is that the absorption is suciently strong for q (0, q
] so as
to drive the L
1
-norm of u(t) to zero as t .
When q (1, q
), it is actually possible to obtain more precise information on

the convergence to zero of |u(t)|
1
.
Proposition 4.4. [2] If q (1, q
), there is C
11
such that
|u(t)|
1
C
11
_
_
|x|t
1/2
u
0
(x) dx +t
a
_
, t 0 . (4.10)
As a consequence of Proposition 4.4 we realize that, if the initial condition u
0
is such that
sup
R>0
_
R
2a
_
{|x|R}
u
0
(x) dx
_
< , (4.11)
then |u(t)|
1
C t
a
for t > 0. In addition, since (1.1) is autonomous, we infer
from Proposition 4.2 (b) that
|u(t)|
2
Nb
C
10
_
_
_
_
u
_
t
2
__
_
_
_
qb
1
t
Nb
C t
(2a+N)/2
for t > 0. It is however not possible for a solution to (1.1), (1.2) to decay to zero
in L
(R
N
) at a faster algebraic rate as the following result shows:
Proposition 4.5. [13, 19] Assume that
(a): either q (1, q
] and there is > a such that

|u(t)|
L
C t
(N/2)
for t 1,
(b): or q = q
and there is > N + 1 such that

|u(t)|
L
C t
N/2
(log t)
for t 1,
then u 0.
1. DECAY RATES 131
Combining the outcome of Proposition 4.4 and Proposition 4.5 we conclude
that, if q (1, q
), there are initial data for which |u(t)|

1
decays as t
a
but
|u(t)|
1
cannot decay at a faster algebraic rate. This is in sharp contrast with the
case q = 1 for which exponential decay rates are possible [15, 16]. Nevertheless,
|u(t)|
1
also cannot decay arbitrarily fast in that case since
|u(t)|
1
C t
3/2
e
t/4
, t 1 ,
if N = 1 and
|u(t)|
1
e
(Ct
1/3
+t)/4
, t 1 ,
if N 2.
1. Decay rates
We rst state an easy consequence of the comparison principle.
Lemma 4.6. There is a constant C such that, for q (0, ) and t > 0,
|u(t)|
C |u
0
|
1
t
N/2
, (4.12)
|u(t)|
C |u
0
|
1
t
(N+1)/2
. (4.13)
Proof. Since = 1, u is a subsolution to the linear heat equation and the
comparison principle ensures that u(t, x)
_
e
t
u
0
_
(x) for (t, x) [0, ) R
N
.
Since u
0
is non-negative, we infer from the temporal decay estimates for integrable
solutions to the linear heat equation that
|u(t)|

_
_
e
t
u
0
_
_
C |u
0
|
1
t
N/2
for t > 0. It next follows from (2.4) with s = t/2 and the previous inequality that
|u(t)|
C
1
_
_
_
_
u
_
t
2
__
_
_
_
_
2
t
_
1/2
C |u
0
|
1
t
(N+1)/2
for t > 0.
We next use the gradient estimate (2.5) to obtain another decay rate.
Lemma 4.7. There is a constant C such that, for q (N/(N + 1), ), q ,= 1,
and t > 0,
|u(t)|
C |u
0
|
qb
1
t
Nb
, (4.14)
|u(t)|
C |u
0
|
b
1
t
(N+1)b
. (4.15)
Proof. For t > 0, we infer from (2.5) (with s = t/2), Lemma 4.1 and the
GagliardoNirenberg inequality (3.9) that
|u(t)|
C |u(t)|
1/(N+1)
1
|u(t)|
N/(N+1)
C |u
0
|
1/(N+1)
1
_
_
_
_
u
_
t
2
__
_
_
_
N/q(N+1)
_
2
t
_
N/q(N+1)
.
Multiplying both sides of the above inequality by t
Nb
leads to
t
Nb
|u(t)|
C |u
0
|
1/(N+1)
1
_
_
t
2
_
Nb
_
_
_
_
u
_
t
2
__
_
_
_
_
N/q(N+1)
for t > 0. Introducing
(T) := sup
t[0,T]
_
t
Nb
|u(t)|
_
for T > 0, it follows from the previous inequality that, for t [0, T],
t
Nb
|u(t)|
C |u
0
|
1/(N+1)
1
(T)
N/q(N+1)
.
Consequently,
(T) C |u
0
|
1/(N+1)
1
(T)
N/q(N+1)
,
(T) C |u
0
|
qb
1
,
and thus
|u(t)|
C |u
0
|
qb
1
t
Nb
for t [0, T]. Since T is arbitrary, we have proved (4.14). Using again (2.5) (with
s = t/2) and (4.14), we further obtain that
|u(t)|
C
2
_
_
_
_
u
_
t
2
__
_
_
_
1/q
_
2
t
_
1/q
C |u
0
|
b
1
t
(N+1)b
for t > 0.
The temporal decay rates stated in (b) and (c) of Proposition 4.2 for q
(N/(N + 1), ), q ,= 1, follow from Lemma 4.6 for q q
and Lemma 4.7 for

q < q
. We next turn to the remaining cases q = N/(N + 1) and q = 1.

1.1. Decay rates: q = N/(N +1). We proceed as in the proof of Lemma 4.7
and infer from (2.5), Lemma 4.1 and the GagliardoNirenberg inequality (3.9) that,
for t > s > 0,
|u(t)|
C |u(t)|
1/(N+1)
1
|u(t)|
N/(N+1)
C |u
0
|
1/(N+1)
1
|u(s)|
N/q(N+1)
(t s)
N/q(N+1)
C |u
0
|
1/(N+1)
1
|u(s)|
(t s)
1
.
Let B be a positive real number to be specied later and assume that t > B.
Choosing s = t B > 0 and multiplying both sides of the above inequality by e
t/B
lead to
e
t/B
|u(t)|

C e
B
|u
0
|
1/(N+1)
1
_
e
(tB)/B
|u(t B)|
_
for t > B. Introducing
(T) := sup
t[0,T]
_
e
t/B
|u(t)|
_
for T > B, it follows from the previous inequality that, for t (B, T],
e
t/B
|u(t)|

C e
B
|u
0
|
1/(N+1)
1
(T) ,
while, for t [0, B],
e
t/B
|u(t)|
e |u
0
|

1. DECAY RATES 133
by (2.1). Therefore,
(T)
C e
B
|u
0
|
1/(N+1)
1
(T) +e |u
0
|
.
Choosing B = 2 C e |u
0
|
1/(N+1)
1
we end up with
(T) 2 e |u
0
|
.
As T is arbitrary we conclude that
|u(t)|
2 e |u
0
|
exp
_
t
2 C e |u
0
|
1/(N+1)
1
_
for t 0 and the proof of (4.6) is complete.
1.2. Decay rates: q = 1. We give two proofs of Proposition 4.2 for q = 1, the
rst one relying on a Moser technique [13] and the second one on the L
1
-euclidean
logarithmic Sobolev inequality.
First proof of Proposition 4.2: q = 1. We employ a Moser technique as
in [26, Section 4]. Consider r 1, s
2
(0, ) and s
1
[0, s
2
). It follows from (1.1)
after multiplication by r u
r1
and integration over (s
1
, s
2
) R
N
that
|u(s
2
)|
r
r
+
_
s2
s1
_
R
N
[(u
r
) (s, x)[ dxds |u(s
1
)|
r
r
.
We next use the Sobolev inequality to obtain that
_
s2
s1
|u
r
(s)|
1
ds C |u(s
1
)|
r
r
, (4.16)
where 1
= if N = 1 and 1
= N/(N 1) otherwise.
Fix t (0, ). If N = 1 we choose r = 1, s
1
= 0 and s
2
= t in (4.16) and
use the monotonicity of s |u(s)|
to obtain (4.7) for q = 1. Assume now that

N 2. As u is non-negative s |u(s)|
m
is a non-increasing function for each
m [1, ] and we infer from (4.16) that
|u(s
2
)|
rN/(N1)
C
1/r
(s
2
s
1
)
1/r
|u(s
1
)|
r
, 0 s
1
< s
2
. (4.17)
Introducing for k 0 the sequences r
k
= (N/(N 1))
k
and t
k
= t(12
(k+1)
), we
proceed as in [26, Section 4] and write (4.17) with r = r
k
, s
1
= t
k
and s
2
= t
k+1
.
Arguing by induction we eventually arrive at the following inequality:
|u(t)|
r
k
|u(t
k
)|
r
k
C
k
2
k
t
k
|u(t/2)|
1
(4.18)
for k 1 with
k
:=
k1
n=0
_
N 1
N
_
n
and
k
:=
k1
n=0
(n + 2)
_
N 1
N
_
n
, k 0.
We may now let k in (4.18) to obtain (4.7) for q = 1.
We now give an alternative proof of Proposition 4.2 for q = 1 employing the

L
1
-euclidean logarithmic Sobolev inequality [8, Theorem 2] which we recall now.
Theorem 4.8. For each w W
1,1
(R
N
) and (0, ), we have
_
R
N
[w(x)[ log
_
[w(x)[
|w|
1
_
dx +N log
_
e
NL
1
_
|w|
1
|w|
1
(4.19)
with L
1
:= ((N + 2)/2)
1/N
/
_
N
1/2
_
.
Second proof of Proposition 4.2: q = 1. Fix t (0, ) and let
(
1
([0, t)) be an increasing function such that (0) = 1 and (s) as s t.
Introducing F(s) := |u(s)|
(s)
for s [0, t), we infer from (1.1) that
F
(s)
F(s)
=

(s)
(s)
2
_
R
N
u(s, x)
(s)
F(s)
(s)
log
_
u(s, x)
(s)
F(s)
(s)
_
dx
+
_
R
N
u(s, x)
(s)1
F(s)
(s)
(u(s, x) [u(s, x)[) dx
(s)
(s)
2
_
R
N
u(s, x)
(s)
F(s)
(s)
log
_
u(s, x)
(s)
F(s)
(s)
_
dx
1
(s)
_
R
N
u
(s)
(s, x)
F(s)
(s)
dx.
It then follows from (4.19) with w = u
(s)
and = (s)/
(s) that
F
(s)
F(s)
N

(s)
(s)
2
log
_
NL
1
e
(s)
(s)
_
, s [0, t) .
With the choice (s) = t/(t s), we obtain
F
(s)
F(s)

N
t
log
_
e
NL
1
(t s)
_
, s [0, t) ,
which yields after integration
log F(s) log F(0)
N
t
_
log
_
e
NL
1
_
s + (t s) (t s) log (t s)
_
+ N (1 log t)
for s [0, t). We then let s t in the previous inequality and end up with
log |u(t)|
log |u
0
|
1
+N log (NL
1
) N log t ,
that is, (4.7) for q = 1 with C
10
= (NL
1
)
N
.

1
Lemma 4.9. [19, Lemma 3.2] If q > 1 and u
0
, 0 then |u(t)|
1
> 0 for every
t 0.
Proof. Since u is a classical solution to
t
u u + A u = 0 in Q
with
A := [u[
q2
u ((Q
; R
N
), the assertion of Lemma 4.9 readily follows from
the strong maximum principle.

2. LIMIT VALUES OF u1 135
Proof of Proposition 4.3: q > q
. Combining (2.6) (with the choice s =

t/2) and (4.12) yields
_
_
_u
(q1)/q
(t)
_
_
_
L
C |u
0
|
(q1)/q
1
t
(q(N+1)N)/(2q)
, t > 0 . (4.20)
Consider now s (0, ) and t (s, ). Since u = (q/(q1)) u
1/q
u
(q1)/q
,
it follows from (1.1) and (4.20) that
|u(t)|
1
= |u(s)|
1
_
q
q 1
_
q
_
t
s
_
R
N
u
(q1)/q
(, x)
q
u(, x) dxd
|u(s)|
1
C
_
t
s
(q(N+1)N)/2
|u()|
1
d .
Owing to the monotonicity of |u()|
1
, we further obtain
|u(t)|
1
|u(s)|
1
_
1 C
_
t
s
(q(N+1)N)/2
d
_
.
Since q > q
the right-hand side of the above inequality has a nite limit as t .

We may then let t and use (4.2) to obtain
I
1
() |u(s)|
L
1
_
1 C s
(N+1)(qq)/2
_
, s > 0 .
Consequently, for s large enough, we have I
1
() > |u(s)|
1
/2 which is positive by
Lemma 4.9.
Proof of Proposition 4.3: q [1, q
]. It rst follows from (4.3) that

(t) :=
_

t
|u(s)|
q
q
ds
t
0 . (4.21)
We next consider a (
-smooth function in R
N
such that 0 1 and
(x) = 0 if [x[ 1 and (x) = 1 if [x[ 2 .
For R > 0 and x R
N
we put
R
(x) = (x/R). We multiply (1.1) by
R
(x) and
integrate over (t
1
, t
2
) R
N
to obtain
_
R
N
u(t
2
, x)
R
(x) dx
_
R
N
u(t
1
, x)
R
(x) dx
1
R
_
t2
t1
_
x
R
_
u(s, x) dxds .
Using the H older inequality and the properties of we deduce that
_
{|x|2R}
u(t
2
, x) dx
_
{|x|R}
u(t
1
, x) dx+C R
(NqNq)/q
(t
1
)
1/q
(t
2
t
1
)
(q1)/q
.
Combining the above inequality with (4.7) and Lemma 4.1 yields
|u(t
2
)|
1
=
_
{|x|2R}
u(t
2
, x) dx +
_
{|x|2R}
u(t
2
, x) dx
C R
N
|u(t
2
)|
+
_
{|x|R}
u(t
1
, x) dx
+ C R
(NqNq)/q
(t
1
)
1/q
(t
2
t
1
)
(q1)/q
_
{|x|R}
u(t
1
, x) dx +C R
N
(t
2
t
1
)
Nb
+ C R
(NqNq)/q
(t
1
)
1/q
(t
2
t
1
)
(q1)/q
.
Choosing
R = R(t
1
, t
2
) := (t
1
)
1/(q+N)
(t
2
t
1
)
(qNb+q1)/(q+N)
we are led to
|u(t
2
)|
1

_
{|x|R(t1,t2)}
u(t
1
, x) dx
+ C (t
1
)
N/(q+N)
(t
2
t
1
)
qNb(N+1)(qq)/(q+N)
.
Since q
q 1 > N/(N + 1) (so that b > 0) we may let t

2
in the previous
inequality to conclude that I
1
() 0 if q [1, q
) and I
1
() C (t
1
)
N/(q+N)
if
q = q
. We have used here that R(t

1
, t
2
) as t
2
and that u(t
1
) L
1
(R
N
).
Owing to the non-negativity of I
1
(), we readily obtain that I
1
() = 0 if q [1, q
).
When q = q
, we let t
1
and use (4.21) to conclude that I
1
() = 0 also in
that case.
Proof of Proposition 4.3: q [N/(N + 1), 1). As in the previous proof we

consider a (
N
such that 0 1 and
(x) = 0 if [x[ 1 and (x) = 1 if [x[ 2 .
For R > 0 and x R
N
we put
R
(x) = (x/R). We multiply (1.1) by
R
(x) and
integrate over (1, t) R
N
to obtain
_
{|x|2R}
u(t, x) dx
_
{|x|R}
u(1, x) dx
+
||
R
_
t
1
|u(s)|
1q
_
R
N
[u(s)[
q
dxds .
3. IMPROVED DECAY RATES: q (1, q) 137
We then infer from (2.5) and (4.3) that
_
{|x|2R}
u(t, x) dx
_
{|x|R}
u(1, x) dx
+
C
R
_

1
s
(1q)/q
_
R
N
[u(s)[
q
dxds
_
{|x|R}
u(1, x) dx +
C
R
_

1
_
R
N
[u(s)[
q
dxds
_
{|x|R}
u(1, x) dx +
C
R
.
Therefore
|u(t)|
1
C R
N
_
_
e
t
u
0
_
_
+
_
{|x|R}
u(1, x) dx +
C
R
,
and the choice R = R(t) :=
_
_
e
t
u
0
_
_
1/(N+1)
yields
|u(t)|
1

_
{|x|R(t)}
u(1, x) dx +
_
_
e
t
u
0
_
_
1/(N+1)
.
Letting t gives the result since
_
_
e
t
u
0
_
_
0 as t .
Proof of Proposition 4.3: q (0, N/(N + 1)). As u(t) vanishes identically

after a nite time by Theorem 3.1 (with m = 0), we clearly have I
1
() = 0.
)
Proof of Proposition 4.4. We x T > 0. For R > 0 and t [0, T], we
have
|u(t)|
1
=
_
{|x|2R}
u(t, x) dx +
_
{|x|>2R}
u(t, x) dx.
On the one hand, it follows from (4.7) that
_
{|x|2R}
u(t, x) dx C R
N
|u(t)|
C R
N
_
_
_
_
u
_
t
2
__
_
_
_
qb
1
t
Nb
. (4.22)
On the other hand, consider a (
N
such that 0 1
and
(x) = 0 if [x[ 1 and (x) = 1 if [x[ 2 .
For x R
N
we put
R
(x) = (x/R). It follows from (1.1) and the H older and
Young inequalities that
d
dt
_
R
N
R
(x)
q/(q1)
u(t, x) dx +
_
R
N
R
(x)
q/(q1)
[u(t, x)[
q
dx
=
q
q 1
_
R
N
R
(x)
1/(q1)
R
(x) u(t, x) dx
C R
((N1)qN)/q
__
R
N
R
(x)
q/(q1)
[u(t, x)[
q
dx
_
1/q
1
2
_
R
N
R
(x)
q/(q1)
[u(t, x)[
q
dx +C R
((N1)qN)/(q1)
,
whence, after integration with respect to time,
_
R
N
R
(x)
q/(q1)
u(t, x) dx
_
R
N
R
(x)
q/(q1)
u
0
(x) dx +C t R
((N1)qN)/(q1)
.
Using the properties of and since t [0, T], we end up with
_
{|x|>2R}
u(t, x) dx
_
{|x|R}
u
0
(x) dx +C T R
((N1)qN)/(q1)
. (4.23)
Collecting (4.22) and (4.23), we obtain
|u(t)|
1

_
{|x|R}
u
0
(x) dx +C T R
((N1)qN)/(q1)
(4.24)
+ C R
N
_
_
_
_
u
_
t
2
__
_
_
_
qb
1
t
Nb
.
Now, we put
z(t) := t
1/(q1)
R
1/b(q1)
|u(t)|
1
, t [0, T]
Y (T, R) := T
1/(q1)
R
1/b(q1)
_
_
{|x|R}
u
0
(x) dx +C T R
((N1)qN)/(q1)
_
.
Multiplying both sides of (4.24) by t
1/(q1)
R
1/b(q1)
we obtain for t [0, T]
z(t) Y (T, R)
_
t
T
_
1/(q1)
+C
_
t
b
R
_
(1Nb(q1)qb)/b(q1)
_
z
_
t
2
__
qb
Y (T, R) +C
_
z
_
t
2
__
qb
as Nb(q 1) 1 +qb = 0 by the denition (4.5) of b. Consequently,
sup
[0,T]
z Y (T, R) +C
_
sup
[0,T]
z
_
qb
.
Since q > 1, we have qb < 1 and infer from the Young inequality that
sup
[0,T]
z Y (T, R) +C +
1
2
sup
[0,T]
z
sup
[0,T]
z 2 Y (T, R) +C .
In particular, z(T) 2 Y (T, R) + C, that is,
|u(T)|
1
2
_
{|x|R}
u
0
(x) dx+C T R
((N1)qN)/(q1)
+C R
1/((q1)b)
T
1/(q1)
.
The choice R = T
1/2
then yields (4.10).
We nally show that, when q (1, q
], the L
-norm of u cannot decay to zero

at an arbitrary fast rate.
Proof of Proposition 4.5. It follows from (1.1) that, for t s 2,
|u(s)|
1
|u(t)|
1
C
_
t
s
_
_
_
_
u
(q1)/q
_
()
_
_
_
q
|u()|
1
d .
We infer from (2.6) (with s = /2) and the time monotonicity of the L
1
-norm of u
that
|u(s)|
1
|u(t)|
1
C |u(s)|
1
_
t
s/2
|u()|
q1

q/2
d .
But it is easy to check that the assumptions (a) or (b) imply that |u()|
q1

q/2
belongs to L
1
(1, ). Consequently we may nd s
0
large enough such that
|u(s
0
)|
1
|u(t)|
1

1
2
|u(s
0
)|
1
for t s
0
.
We then pass to the limit as t and deduce from (4.9) that |u(s
0
)|
1

2 I
1
() = 0, whence u(s
0
) 0. Applying Lemma 4.9 we conclude that u 0.
For q (1, 2), an alternative proof of Lemma 4.9 relying on L
p
-estimates may
be found in [2, Lemma 4.1] (and applies also if the Laplacian is replaced by a
degenerate diusion). Proposition 4.2 is proved in [13] but the argument given
here is simpler. The alternative proof of Proposition 4.2 for q = 1 built upon the
L
1
-euclidean logarithmic Sobolev inequality is also new. Proposition 4.3 is proved
in [15, 16] when q = 1 and in [1] when q (1, q
) and q = 2. The general case q > 1

is performed in [11, 19] and is extended to q (0, 1] in [13, 14]. Next, if q (1, q
)
and u
0
fulls (4.11), we have |u(t)|
1
C t
a
for t > 0 as already mentioned. This
property was previously established in [12] under the stronger condition
R(u
0
) := inf
_
R > 0 such that sup
|x|R
_
[x[
(2q)/(q1)
u
0
(x)
_

q
_
< ,
the proof combining (2.7) and the comparison with the supersolution
x
q
[x[
(2q)/(q1)
with
q
:=
_
(q 1)(2(1 +a))
1/(q1)
_
/(2 q) .
Proposition 4.4 is proved in [2] with a slightly dierent proof. It unfortunately
does not cover the critical exponent q = q
. In that case, it has been shown in

[30] that, for initial data decaying suciently rapidly as [x[ , then |u(t)|
1

C (log t)
(N+1)
and |u(t)|
C (log t)
(N+1)
t
N/2
for t 1, the proof relying
on completely dierent arguments (invariant manifold). Note that, according to
Proposition 4.4, t
N/2
|u(t)|
cannot decay at a faster logarithmic rate. Finally,

Proposition 4.5 is actually a slight improvement of [19, Corollary 3.5].
CHAPTER 5
Temporal growth estimates for integrable initial
data: = 1
Throughout this section we assume that
= 1 and u
0
is a non-negative function in L
1
(R
N
) W
1,
(R
N
) (5.1)
and denote by u the corresponding classical solution to (1.1), (1.2). We then in-
vestigate the time behaviour of the L
p
-norms of u for p = 1 and p = . As a rst
step, we specify conditions ensuring that u(t) remains in L
1
(R
N
) for t > 0.
Lemma 5.1. If q 1, then u (([0, ), L
1
(R
N
)) and t |u(t)|
1
is a
non-decreasing function of time with
I
1
() := lim
t
|u(t)|
1
[|u
0
|
1
, ] . (5.2)
In addition,
|u(t)|
1
= |u
0
|
1
+
_
t
0
|u(s)|
q
q
ds , t 0 . (5.3)
Remark 5.2. It is not known whether Lemma 5.1 remains valid for q (0, 1)
when = 1.
When = 1, the nonlinear term [u[
q
is a source term and thus slows down
or even impedes the dissipation of the diusion. In that case, the counterpart of
Proposition 4.3 (though less complete and more complicated) reads:
Proposition 5.3. [9, 31, 39] We have
_
_
I
1
() [|u
0
|
1
, ) and I
() = 0 if q 2 ,
I
1
() [|u
0
|
1
, ] and I
() [0, |u
0
|
] if q (q
, 2) ,
I
1
() = and I
() (0, |u
0
|
] if q [1, q
] ,
(5.4)
where q
is dened in (4.4) and

I
() := lim
t
|u(t)|
[0, |u
0
|
] . (5.5)
In addition, if I
() > 0 then I
1
() = and
|u(t)|
1
C t
N/q
, t 0 . (5.6)
Observe that, if q [1, q
], the strength of the source term leads to the un-

boundedness of the L
1
-norm of u(t) as t . The behaviour of |u(t)|
1
is much
more complex when q ranges in (q
, 2) since it is unbounded for large times only for

141
142 5. TEMPORAL GROWTH ESTIMATES FOR INTEGRABLE INITIAL DATA: = 1
some initial data. Let us also emphasize at this point that the converse of the last
statement of Proposition 5.3 is false, that is, there is at least one initial condition
u
0
such that I
1
() = and I
() = 0, see Proposition 5.6 below.

Additional properties are gathered in the next proposition.
Proposition 5.4. [31]
(a): If q (q
, 2) and I
() = 0, then
|u(t)|
C t
(2q)/(2(q1))
, t > 0 . (5.7)
(b): If q (1, 2) and I
() > 0, then
0 t
(2q)/q
(|u(t)|
()) C |u
0
|
(2q)/q
, t > 0 . (5.8)
(c): If q = 1, then
0
t
1/2
log t
(|u(t)|
()) C |u
0
|
, t 2 . (5.9)
1
and |u|
This section is devoted to the proof of Proposition 5.3 which requires to handle
separately the cases q 2, q (q
, 2) and q [1, q
].
We rst check the last statement of Proposition 5.3. We thus assume that
I
() > 0 and x t > 0. For k 1, let x

k
R
N
(depending possibly on t) be
such that |u(t)|
1/k u(t, x
k
). For R > 0, it follows from (2.5) (with s = 0)
and the time monotonicity of |u|
that
|u(t)|
1

_
{|xx
k
|R}
u(t, x) dx
_
{|xx
k
|R}
(u(t, x
k
) [x x
k
[ |u(t)|
) dx
C
_
R
N
N
_
|u(t)|
1
k
_
C
1
R
N+1
N + 1
|u
0
|
1/q
t
1/q
_
C R
N
_
I
()
1
k
C
1
R t
1/q
_
.
Letting k and choosing R = R(t) :=
_
I
()t
1/q
_
/(2C
1
) gives the result.
1.1. The case q 2. We put
A := |u
0
|
|u
0
|
q2
and B := |u
0
|
q2
, (5.10)
and consider t > 0 and an integer p 1. We multiply (1.1) by u
p
and integrate
over (0, t) R
N
to obtain
p
_
t
0
_
R
N
u
p1
[u[
2
dxds =
1
p + 1
_
R
N
_
u
p+1
0
u
p+1
(t)
_
dx
+
_
t
0
_
R
N
u
p
[u[
q
dxds ,
1. LIMIT VALUES OF u1 AND u 143
whence, thanks to (2.3),
_
t
0
_
R
N
u
p1
[u[
2
dxds
|u
0
|
1
|u
0
|
p
p(p + 1)
+
B
p
_
t
0
_
R
N
u
p
[u[
2
dxds . (5.11)
We next claim that
_
t
0
_
R
N
[u[
q
dxds |u
0
|
1
k
i=1
A
i
(i + 1)!
+
B
k+1
k!
_
t
0
_
R
N
u
k
[u[
2
dxds (5.12)
for every integer k 1. Indeed, it follows from (2.3) and (5.11) with p = 1 that
_
t
0
_
R
N
[u[
q
dxds B
_
t
0
_
R
N
[u[
2
dxds
A
2
|u
0
|
1
+B
2
_
t
0
_
R
N
u [u[
2
dxds
whence (5.12) for k = 1. We next proceed by induction and assume that (5.12)
holds true for some k 1. Inserting (5.11) with p = k + 1 in (5.12) yields
_
t
0
_
R
N
[u[
q
dxds |u
0
|
1
k
i=1
A
i
(i + 1)!
+
B
k+1
k!
|u
0
|
1
|u
0
|
k+1
(k + 1)(k + 2)
+
B
k+2
(k + 1)!
_
t
0
_
R
N
u
k+1
[u[
2
dxds
|u
0
|
1
k+1
i=1
A
i
(i + 1)!
+
B
k+2
(k + 1)!
_
t
0
_
R
N
u
k+1
[u[
2
dxds ,
whence (5.12) for k + 1 and the proof of (5.12) is complete.
Finally, let p
0
1 be the smallest integer satisfying p
0
> A. Using (5.11) with
p = p
0
+ 1 and (2.3), we obtain that
(p
0
+ 1)
_
t
0
_
R
N
u
p0
[u[
2
dxds
|u
0
|
1
|u
0
|
p0+1
p
0
+ 2
+A
_
t
0
_
R
N
u
p0
[u[
2
dxds ,
from which we deduce, since p
0
+ 1 A 1,
_
t
0
_
R
N
u
p0
[u[
2
dxds
|u
0
|
1
|u
0
|
p0+1
p
0
+ 2
.
We insert the above estimate in (5.12) with k = p
0
and end up with
_
t
0
_
R
N
[u[
q
dxds |u
0
|
1
p0
i=1
A
i
(i + 1)!
+
B
p0+1
p
0
!
|u
0
|
1
|u
0
|
p0+1
p
0
+ 2
|u
0
|
1
p0+1
i=1
A
i
i!
.
Consequently, [u[
q
L
1
(Q
) and we may integrate (1.1) over Q
to conclude
that
I
1
() |u
0
|
1
p0+1
i=0
A
i
i!
< ,
and the proof of (5.4) for q 2 is complete.
1.2. The case q (q
, 2). In that case, the statement of Proposition 5.3 is an

obvious consequence of Theorem 2.1 and Lemma 5.1.
Nevertheless, some additional information on I
1
() and I
() may be ob-
tained when q (q
, 2). The rst result in that direction is the following:

Proposition 5.5. Assume that q (q
, 2). There is = (N, q) > 0 such that

I
1
() < (and thus I
() = 0 by Proposition 5.3) whenever

|u
0
|
1
|u
0
|
(N+1)(qq)
. (5.13)
Proof. Owing to (5.3), it is clear that I
1
() < whenever t |u(t)|
q
q

L
1
(0, ). Such a property is certainly true if we know that
t |u(t)|
k
k
L
1
(0, ) for some k [1, q] (5.14)
since |u|
q
q
|u
0
|
qk
|u|
k
k
by (2.3). We proceed in two steps. Under a small-
ness assumption on u
0
, we obtain a rst temporal decay estimate for |u(t)|
k
. In a
second step, we improve this estimate so as to obtain the required time integrability
of t |u(t)|
k
k
.
Step 1. Fix k (q
, q] (close to q
) such that
_
_
q
< k < min

_
N
N 1
, 2
_
, p :=
N(k 1)
2 k
(1, k) and
1
2(k 1)

N
2k
>
1
q
.
(5.15)
Such a choice is always possible since the function k N(k 1)/(2 k) is equal
to 1 for k = q
and the function k (1/(2(k 1))) (N/(2k)) is equal to 1/q
for k = q
. Putting A(s, x) := [u(s, x)[

qk
, the integral identity (2.2) reads
u(t) = e
t
u
0
+
_
t
0
e
(ts)
A(s)[u(s)[
k
ds , t 0 . (5.16)
Recall that |A(s)|
:= |u
0
|
qk
in view of (2.3).
We next introduce
:=
N
2
_
1
p

1
k
_
+
1
2
=
1
2(k 1)

N
2k
>
1
2
and
M(t) := sup
s(0,t)
s
|u(s)|
k
for t > 0. Using a xed point argument in a suitable space and the uniqueness
assertion of Theorem 2.1, we proceed as in [20, 39] to establish that M(t) is nite
for each t > 0 and converges to zero as t 0. It follows from (5.16) and the decay
properties of the heat semigroup that
|u(t)|
k
C t
|u
0
|
p
+C
_
t
0
(t s)
(k+N(k1))/(2k)
A
_
_
[u(s)[
k
_
_
1
ds
C t
|u
0
|
p
+C A
M(t)
k
_
t
0
(t s)
(k+N(k1))/(2k)
s
k
ds .
Consequently,
t
|u(t)|
k
C |u
0
|
p
+C A
M(t)
k
_
1
0
(1 s)
(k+N(k1))/(2k)
s
k
ds ,
and the last integral in the right-hand side of the previous inequality is nite thanks
to the choice of k (q
, N/(N 1)). Thus

M(t) K
1
_
|u
0
|
p
+A
M(t)
k
_
for t > 0 . (5.17)
Now, assume that
|u
0
|
k1
p
A
2
(k+1)
K
k
1
. (5.18)
Since M(t) 0 as t 0 and M is non-decreasing, there is a maximal t
0
(0, ]
such that M(t) 2K
1
|u
0
|
p
for t [0, t
0
). Assuming for contradiction that t
0
< ,
we have M(t
0
) = 2K
1
|u
0
|
p
while (5.17) implies that
M(t
0
) K
1
_
|u
0
|
p
+A
(2K
1
)
k
|u
0
|
k
p
_
3K
1
2
|u
0
|
p
,
hence a contradiction. Consequently, t
0
= and we deduce that
M(t) 2K
1
|u
0
|
p
for all t > 0 (5.19)
whenever u
0
fulls (5.18). Observe that, since k < 1, the estimate (5.19) does not
guarantee (5.14). A better estimate will be obtained in the next step.
Step 2. We combine (2.5) and (5.19) to obtain
|u(t)|
q
q
|u(t)|
qk
|u(t)|
k
k
C |u
0
|
(qk)/q
t
1k((1/q))
for t 1. Recalling that > 1/q by (5.15), we conclude that t |u(t)|
q
q
be-
longs to L
1
(0, ), whence I
1
() < . Finally, owing to the GagliardoNirenberg
inequality
|w|
p
C |w|
(N(p1))/(p(N+1))
|w|
(p+N)/(p(N+1))
1
, w L
1
(R
N
) W
1,
(R
N
) ,
we realize that
|u
0
|
k1
p
A
= |u
0
|
k1
p
|u
0
|
qk
C |u
0
|
1/(N+1)
1
|u
0
|
qq
,
so that choosing suciently small in (5.13) implies (5.18).
Before identifying a class of initial data for which I
() > 0 (and thus

I
1
() = by Proposition 5.3) we report the following result.
Proposition 5.6. [20] Assume that q (q
, 2). There exists u

0
L
1
(R
N
)
B(
2
(R
N
) such that I
1
() = and I
() = 0. More precisely, this particular

solution to (1.1), (1.2) enjoys the following properties: for t 0,
|u(t)|
1
= (1+t)
((N+1)(qq))/(2(q1))
|u
0
|
1
and |u(t)|
=
|u
0
|
(1 +t)
(2q)/(2(q1))
.
Proposition 5.6 is established in [20, Theorem 3.5] by constructing a self-similar
solution to (1.1) of the form
(t, x) t
(2q)/(2(q1))
W
s
_
[x[t
1/2
_
.
The prole W
s
then satises a nonlinear ordinary dierential equation, which turns
out to have an integrable and smooth solution W
s
if q (q
, 2). Taking u
0
= W
s
gives the result.
We end up this section with the following result:
Proposition 5.7. [9] There is a constant K = K(q) > 0 such that, if u
0

W
2,
(R
N
) satises
|u
0
|
inf
yR
N
u
0
(2q)/q
> K , (5.20)
then I
() > 0.
Proof. Since u is a classical solution to (1.1), (1.2) with = 1, it follows
from (1.1) that
u(t, x) u
0
(x) +
_
t
0
u(s, x) ds u
0
(x) +
_
t
0
inf
yR
N
u(s, y) ds
for every x R
N
and t 0. Therefore,
|u(t)|
|u
0
|
+
_
t
0
inf
yR
N
u(s, y) ds ,
and we infer from (2.38) and (2.39) that
|u(t)|
|u
0
|
+T inf
yR
N
u
0
C |u
0
|
(2q)/q
_
t
T
s
2/q
ds
for T > 0 and t > T. Since q < 2, we may let t in the above inequality
and obtain with the choice T = |u
0
|
(2q)/2
inf
yR
N u
0
q/2
that there is a
constant K depending only on q such that
I
() |u
0
|
K
q/2
inf
yR
N
u
0
(2q)/2
|u
0
|
(2q)/2
. (5.21)
Therefore, if |u
0
|
> K
inf
yR
N u
0
(2q)/q
, we readily conclude from (5.21)
that I
() > 0, whence Proposition 5.7.
Unfortunately, the conditions (5.13) and (5.20) do not involve the same quan-
tities. Still, we can prove that if u
0
fulls
|u
0
|
|D
2
u
0
|
(2q)/q
>
(which clearly implies (5.20) since q < 2), the quantity |u
0
|
1
|u
0
|
(N+1)(qq)
cannot be small. Indeed, there is a constant C depending only on q and N such

that
_
|u
0
|
|D
2
u
0
|
(2q)/q
_
q(N+1)/2
C|u
0
|
1
|u
0
|
(N+1)(qq)
. (5.22)
To prove (5.22), we put B = |u
0
|
|D
2
u
0
|
(2q)/q
and note that the Gagliardo

Nirenberg inequalities
|u
0
|
C |u
0
|
N/(N+1)
|u
0
|
1/(N+1)
1
,
|u
0
|
C |D
2
u
0
|
(N+1)/(N+2)
|u
0
|
1/(N+2)
1
,
imply that
|u
0
|
(2q)(N+2)
C |D
2
u
0
|
(2q)(N+1)
|u
0
|
2q
1
= C B
q(N+1)
|u
0
|
q(N+1)
|u
0
|
2q
1
C B
q(N+1)
|u
0
|
qN
|u
0
|
2
1
,
whence the above claim.
1.3. The case q [1, q
]. We rst show that I

1
() = .
Lemma 5.8. Assume that q [1, q
]. Then I
1
() = .
Proof. For T > 0 and t > T, we infer from (1.1) that
|u(t)|
1
= |u(T)|
1
+
_
t
T
|u(s)|
q
q
ds
_
t
T
|u(s)|
q
q
ds . (5.23)
Case 1: N 2. Since 1 q q
2, it follows from (5.23), the pointwise

inequality u(s, x)
_
e
s
u
0
_
(x) and the Sobolev inequality that
|u(t)|
1
C
_
t
T
|u(s)|
q
Q
ds C
_
t
T
_
_
e
s
u
0
_
_
q
Q
ds
with Q := Nq/(N q). Since
_
_
e
s
u
0
_
_
Q
|u
0
|
1
|g(s)|
Q
_
_
e
s
u
0
|u
0
|
1
g(s)
_
_
Q
= s
N(Q1)/(2Q)
|u
0
|
1
|G|
Q
_
_
e
s
u
0
|u
0
|
1
g(s)
_
_
Q
|u
0
|
1
|G|
Q
2
s
(Nq(N+1))/(2q)
for s T and T large enough by Proposition A.1 (the functions g and G being
dened in (1.7)), we conclude that
|u(t)|
1
C
_
t
((N+1)(qq))/2
T
((N+1)(qq))/2
_
if q [1, q
)
and
|u(t)|
1
C (log t log T) if q = q
,
which gives the expected result for N 2.
Case 2: N = 1. It follows from (1.1), the time monotonicity of |u|
1
and the
GagliardoNirenberg inequality
|w|
2q1
C |
x
w|
q
q
|w|
q1
1
, w L
1
(R) W
1,q
(R) ,
that, for t > T > 0,
|u(t)|
1
C
_
t
T
|u(s)|
2q1
|u(s)|
(q1)
1
ds C |u(t)|
(q1)
1
_
t
T
|u(s)|
2q1
ds .
Consequently,
|u(t)|
q
1
C
_
t
T
|u(s)|
2q1
ds ,
and we proceed as in the previous case to complete the proof.

Having excluded the niteness of I
1
(), we turn to I
() and establish the

following lemma:
Lemma 5.9. [31] Assume that q (1, 2) and I
() = 0. Then |u(t)|

C t
(2q)/(2(q1))
for all t 0.
Proof. infer from (1.1) that, for each T > t,
u(T, x) = u(t, x) +
_
T
t
t
u(s, x) ds u(t, x) +
_
T
t
u(s, x) ds .
Since (1.1) is autonomous, we infer from (2.9) that
u(s, x) N C
7
_
_
_
_
u
_
t
2
__
_
_
_
(2q)/q
_
s
t
2
_
2/q
for s (t, T) ,
so that
u(T, x) u(t, x) N C
7
_
_
_
_
u
_
t
2
__
_
_
_
(2q)/q
_
T
t
_
s
t
2
_
2/q
ds
u(t, x) C
_
_
_
_
u
_
t
2
__
_
_
_
(2q)/q
t
(2q)/q
.
Letting T we obtain
u(t, x) C
_
_
_
_
u
_
t
2
__
_
_
_
(2q)/q
t
(2q)/q
+I
()
= C
_
_
_
_
u
_
t
2
__
_
_
_
(2q)/q
t
(2q)/q
,
whence
t
(2q)/(2q2)
|u(t)|
C
_
_
t
2
_
(2q)/(2q2)
_
_
_
_
u
_
t
2
__
_
_
_
_
(2q)/q
.
Introducing
A(t) := sup
s[0,t]
_
s
(2q)/(2q2)
|u(s)|
_
,
we deduce from the previous inequality that A(t) C A(t)
(2q)/q
, hence A(t)
C
q/(2q2)
. This bound being valid for each t 0, the proof of Lemma 5.9 is
complete.
Proof of Proposition 5.3: q [1, q
]. Since u(t) e
(tT)
u(T) for T 0
and t > T by the comparison principle, we have
(t T)
N/2
|u(t)|
|u(T)|
1
|G|
(t T)
N/2
_
_
_e
(tT)
u(T) |u(T)|
1
g(t T)
_
_
_
,
and we infer from Proposition A.1 that
liminf
t
_
t
N/2
|u(t)|
_
liminf
t
_
(t T)
N/2
|u(t)|
_
|u(T)|
1
|G|
(5.24)
2. GROWTH RATES 149
for every T > 0, the functions g and G being still dened by (1.7).
We rst consider the case q > 1 and assume for contradiction that I
() = 0.
Then Lemma 5.9 entails that
t
N/2
|u(t)|
C t
((N+1)(qq))/(2(q1))
,
from which we deduce that
limsup
t
_
t
N/2
|u(t)|
_
C (5.25)
as q [1, q
]. Combining (5.24) and (5.25) yields that |u(T)|

1
|G|
C for every
T > 0 and thus contradicts Lemma 5.8. Therefore I
() > 0.
Next, if q = 1, we observe that (2.3) ensures that
t
u u
[u[
q
|u
0
|
q1
in Q
. Consequently, u := u/|u
0
|
satises
t
u u [ u[
q
in Q
. We then
infer from the comparison principle that u(t, x) v(t, x) for (t, x) [0, ) R
N
,
where v denotes the unique classical solution to
t
v v = [v[
q
in Q
with
initial condition v(0) = u
0
/|u
0
|
. Therefore,
I
() = |u
0
|
lim
t
| u(t)|
|u
0
|
lim
t
|v(t)|
> 0 ,
which completes the proof of Proposition 5.3.
2. Growth rates
First, the assertion (a) of Proposition 5.4 is nothing but Lemma 5.9. We then
turn to the proof of the assertions (b) and (c).
Proof of Proposition 5.4 (b). For (t, x) Q
, we infer from (1.1) and

(2.38) that, for each T > t,
u(T, x) = u(t, x) +
_
T
t
t
u(s, x) ds u(t, x) +
_
T
t
u(s, x) ds
u(t, x) C
9
|u
0
|
(2q)/q
_
T
t
s
2/q
ds
u(t, x) C |u
0
|
(2q)/q
t
(2q)/q
,
|u(T)|
|u(t)|
C |u
0
|
(2q)/q
t
(2q)/q
.
Letting T gives (5.8).

Proof of Proposition 5.4 (c). We proceed as in the previous proof but use
(2.8) instead of (2.38) to obtain
u(T, x) = u(t, x) +
_
T
t
t
u(s, x) ds u(t, x) C
5
_
T
t
2
(s) ds
u(t, x) C |u
0
|
_
T
t
s
3/2
(1 + log (1 +s)) ds
u(t, x) C |u
0
|
t
1/2
(2 + log (1 +t)) ,
|u(T)|
|u(t)|
C |u
0
|
t
1/2
(2 + log (1 +t)) .
Letting T gives (5.9).
Lemma 5.8 was rst proved in [39]. The proof given here is slightly dierent
and is inspired by the proof of [31, Theorem 3]. The fact that I
() > 0 for
q [1, q
] was rst establish in [9] for suciently large initial data (fullling a
condition similar to (5.20)). The proof for general initial data is due to Gilding
[31, Theorem 3]. When q (1, 2) and I
() > 0 the growth rate (5.6) seems to

be optimal, see Section 2. When q = 1 the growth rate (5.6) is also optimal in the
sense that there is a large class of initial data (including compactly supported ones)
such that |u(t)|
1
C t
N
for t 1 [40]. Concerning self-similar solutions to (1.1)
for = 1, their existence is investigated in [20, 33].
CHAPTER 6
Convergence to self-similarity
1. The diusion-dominated case: = 1
In this section, we shall prove that, when = 1, the diusion governs the large
time dynamics as soon as I
1
() is non-zero. More precisely, we have the following
result which is similar to the one known for the linear heat equation (1.5) and
recalled in Section 1 below.
Theorem 6.1. Assume that = 1 and q > q
. Let u
0
L
1
(R
N
) W
1,
(R
N
)
be a non-negative function and denote by u the corresponding classical solution to
(1.1), (1.2). Then
I
1
() := lim
t
|u(t)|
1
(0, ) , (6.1)
and, for every p [1, ],
lim
t
t
N(p1)/2p
|u(t) I
1
() g(t)|
p
= 0 , (6.2)
lim
t
t
N(p1)/2p+(1/2)
|u(t) I
1
() g(t)|
p
= 0 , (6.3)
where the functions g and G are dened in (1.7).
The proof of Theorem 6.1 relies on the following properties of the inhomoge-
neous heat equation.
Lemma 6.2. Assume that v is the solution to the Cauchy problem for the linear
inhomogeneous heat equation
t
v v = f , (t, x) Q
,
v(0) = v
0
, x R
N
,
with v
0
L
1
(R
N
) and f L
1
(Q
). Then
lim
t
|v(t) K
g(t)|
1
= 0 , (6.4)
where
K
:= lim
t
_
R
N
v(t, x) dx =
_
R
N
v
0
(x) dx +
_

0
_
R
N
f(t, x) dx dt .
Assume further that there is p [1, ] such that f(t) L
p
(R
N
) for every t > 0
and
lim
t
t
1+(N/2)(11/p)
|f(t)|
p
= 0 . (6.5)
Then
lim
t
t
(N/2)(11/p)
|v(t) K
g(t)|
p
= 0 , (6.6)
151
152 6. CONVERGENCE TO SELF-SIMILARITY
and
lim
t
t
(N/2)(11/p)+1/2
|v(t) K
g(t)|
p
= 0 . (6.7)
Proof. We rst observe that the assumptions on v
0
and f warrant that K
is nite and put

K(T) :=
_
R
N
v(T, x) dx =
_
R
N
v
0
(x) dx +
_
T
0
_
R
N
f(t, x) dx dt for T > 0 .
Moreover, by the Duhamel formula,
|v(t)|
1
|g(t) v
0
|
1
+
_
t
0
|g(t ) f()|
1
d |v
0
|
1
+
_

0
|f()|
1
d
for t 0 and thus v L
(0, ; L
1
(R
N
)).
We now prove (6.4). Let T > 0 and t (T, ). By the Duhamel formula,
v(t) = g(t T) v(T) +
_
t
T
g(t ) f() d ,
so that, by the Young inequality,
|v(t) g(t T) v(T)|
1

_
t
T
|g(t )|
1
|f()|
1
d
_

T
|f()|
1
d .
Then
|v(t) K
g(t)|
1
|v(t) g(t T) v(T)|
1
+|(g(t T) g(t)) v(T)|
1
+ |g(t) v(T) K(T) g(t)|
1
+[K(T) K
[ |g(t)|
1
2
_

T
|f()|
1
d + sup
0
|v()|
1
|(g(t T) g(t))|
1
+ |g(t) v(T) K(T) g(t)|
1
.
Owing to Proposition A.1 and the properties of g, we may let t in the above
inequality and deduce that
limsup
t
|v(t) K
g(t)|
1
2
_

T
|f()|
1
d .
We next use the time integrability of f to pass to the limit as T and complete
the proof of (6.4).
We next assume (6.5) and prove (6.7). Let T > 0 and t (T, ). By the
Duhamel formula,
v(t) = g(t T) v(T) +
_
t
T
g(t ) f() d .
1. THE DIFFUSION-DOMINATED CASE: = 1 153
It follows from the Young inequality that
t
(N/2)(11/p)+1/2
|v(t) g(t T) v(T)|
p
C t
(N/2)(11/p)+1/2
_
(T+t)/2
T
(t )
(N/2)(11/p)1/2
|f()|
1
d
+ C t
(N/2)(11/p)+1/2
_
t
(T+t)/2
(t )
1/2
|f()|
p
d
C
_
t
t T
_
(N/2)(11/p)+1/2
_

T
|f()|
1
d
+ C sup
T
_
(N/2)(11/p)+1
|f()|
p
_
_
t
(T+t)/2
(t )
1/2
1/2
d
C
_
t
t T
_
(N/2)(11/p)+1/2
_

T
|f()|
1
d
+ C sup
T
_
(N/2)(11/p)+1
|f()|
p
_
.
It also follows from Proposition A.1 and the properties of g that
lim
t
t
(N/2)(11/p)+1/2
|g(t T) v(T) K(T) g(t T)|
p
= 0 ,
and
lim
t
t
(N/2)(11/p)+1/2
|g(t T) g(t)|
p
= 0
for every p [1, ]. Since
|v(t) K
g(t)|
p
|v(t) g(t T) v(T)|
p
+ |g(t T) v(T) K(T) g(t T)|
p
+ [K(T) K
[ |g(t T)|
p
+ [K
[ |g(t T) g(t)|
p
,
the previous relations imply that
limsup
t
t
(N/2)(11/p)+1/2
|v(t) K
g(t)|
p
C
__

T
|f()|
1
d + sup
T
_
(N/2)(11/p)+1
|f()|
p
_
_
.
The above inequality being valid for any T > 0, we may let T and conclude
that (6.7) holds true. The assertion (6.6) then follows from (6.4) and (6.7) by the
GagliardoNirenberg inequality.
Proof of Theorem 6.1. Consider p [1, ] and t > 0. Since = 1, u is a

subsolution to the linear heat equation and therefore satises
|u(t)|
p
|e
t
u
0
|
p
C t
(N/2)(11/p)
|u
0
|
1
as a consequence of the comparison principle. Combining this estimate with (2.6)
(with s = t/2) and (4.12) gives
|[u(t)[
q
|
p
=
_
q
q 1
_
q _
_
_u(t)
u
(q1)/q
(t)
q
_
_
_
p
_
q
q 1
_
q _
_
_u
(q1)/q
(t)
_
_
_
q
|u(t)|
p
C |u
0
|
1
_
_
_
_
u
_
t
2
__
_
_
_
q1
t
q/2
t
N(p1)/2p
C |u
0
|
q
1
t
(q+N(q1))/2
t
N(p1)/2p
,
whence
t
(N/2)(11/p)+1
|[u(t)[
q
|
p
C |u
0
|
q
1
t
(N+1)(qq)/2
t
0
because q > q
. Theorem 6.1 then readily follows from Lemma 6.2 with f(t, x) =
[u(t, x)[
q
.
Remark 6.3. The convergence (6.2) in Theorem 6.1 is also true when = 1
and either q 2 or q (q
, 2) and u
0
is suitably small. Indeed, in both cases, I
1
()
is positive and nite by Proposition 5.3 and Proposition 5.5, respectively. The proof
again relies on Lemma 6.2 but establishing that f = [u[
q
fulls the condition (6.5)
for all p [1, ] is more delicate [9].
2. The reaction-dominated case: = 1
The aim of this section is to show that, when = 1 and I
() > 0, the
large time behaviour of the solution u to (1.1), (1.2) is dominated by the nonlinear
HamiltonJacobi term in the sense that u behaves as a self-similar solution to (1.9)
for large times (recall that I
() is the limit of |u(t)|
as t , see (5.5)).
Such a result however holds for initial data decaying to zero as [x[ and we
dene the space (
0
(R
N
) of such functions by
(
0
(R
N
) :=
_
z B((R
N
) : lim
R
sup
{|x|R}
[z(x)[ = 0
_
.
Theorem 6.4. Assume that = 1 and q (1, 2). Let u
0
(
0
(R
N
)
W
1,
(R
N
) be a non-negative function and denote by u the corresponding classical
solution to (1.1), (1.2). Assume further that
I
() := lim
t
|u(t)|
> 0 . (6.8)
Then
lim
t
|u(t) h
(t)|
= 0 , (6.9)
where h
is given by
h
(t, x) := H
_
x
t
1/q
_
and H
(x) :=
_
I
()
q
[x[
q/(q1)
_
+
for (t, x) Q
and
q
:= (q 1) q
q/(q1)
.
2. THE REACTION-DOMINATED CASE: = 1 155
A similar result is valid for viscosity solutions to (1.9), see Section 2 below.
Thus Theorem 6.4 clearly asserts the sole domination of the HamiltonJacobi term
for large times. In fact, h
is the unique viscosity solution in B|((Q
) to (1.9)
with the bounded and upper semicontinuous initial condition h
(0, x) = 0 if x ,= 0
and h
(0, 0) = I
().
We next point out that there is no loss of generality in assuming that q < 2 in
Theorem 6.4. Indeed, according to Proposition 5.3, we always have I
() = 0 if
q 2.
Proof of Theorem 6.4. We change the variables and the unknown function
so that the convergence (6.9) reduces to a convergence to a steady state. More
precisely, we introduce the self-similar (or scaling) variables
=
1
q
log (1 +t) , y =
x
(1 +t)
1/q
,
and the new unknown function v dened by
u(t, x) = v
_
log (1 +t)
q
,
x
(1 +t)
1/q
_
, (t, x) [0, ) R
N
.
Then v(, y) = u (e
q
1, ye
) and it follows from (1.1), (1.2) that v solves
v y v q [v[
q
q e
(2q)
v = 0 , (, y) Q
, (6.10)
v(0) = u
0
, y R
N
. (6.11)
We also infer from (2.1), (2.3), (2.5) and (2.9) that there is a positive constant K
1
depending only on N, q and u
0
such that
|v()|
+|v()|
K
1
, 0 , (6.12)
v(, y) K
1
, (, y) [1, ) R
N
. (6.13)
At this point, since q < 2, it is clear that the diusion in (6.10) becomes
weaker and weaker as time increases to innity and the behaviour of v for large
times is expected to look like that of the solution to the HamiltonJacobi equation
z y z q [z[
q
= 0. To investigate the large time behaviour of rst-
order HamiltonJacobi equations, a useful technique has been developed in [43, 45]
which relies on the relaxed half-limits method introduced in [7]. More precisely, for
(, y) Q
we dene the relaxed half-limits v
and v
by
v
(y) := liminf
(,z,s)(,y,)
v( +s, z) and v
(y) := limsup
(,z,s)(,y,)
v( +s, z) (6.14)
and rst note that the right-hand sides of the above denitions indeed do not depend
on > 0. In addition,
0 v
(x) v
(x) I
() for x R
N
(6.15)
by (5.5) and (6.14), while (6.12) and the Rademacher theorem clearly ensure that
v
and v
both belong to W
1,
(R
N
). Finally, by [6, Theor`eme 4.1] applied to
equation (6.10), v
and v
are viscosity subsolution and supersolution, respectively,

to the HamiltonJacobi equation
H(y, z) := y z q [z[
q
= 0 in R
N
. (6.16)
The next step is to show that v
and v
actually coincide. However, the equation

(6.16) has several solutions as y
_
c
q
[y[
q/(q1)
_
+
solves (6.16) for any c > 0.
The information obtained so far on v
and v
are thus not sucient. The next two

results provide the lacking information.
Lemma 6.5. Given (0, 1), there is R
> 1/ such that

v(, y) for 0 and y R
N
B(0, R
) . (6.17)
Proof. We rst construct a supersolution to (6.10) in (0, ) R
N
B(0, R)
for R large enough. To this end, consider R
_
4 +q 2
q1
|u
0
|
q1
_
1/q
and dene
R
(y) = |u
0
|
R
2
[y[
2
for y R
N
. Let L be the parabolic operator dened by
Lw(, y) :=
w(, y) y w(, y) q [w(, y)[

q
q e
(2q)
w(, y)
for (, y) Q
(so that Lv = 0). Then, if y R

N
B(0, R), we have
L
R
(y) = 2 |u
0
|
R
2
[y[
2
q 2
q
|u
0
|
q
R
2q
[y[
3q
2 (4 N) |u
0
|
R
2
[y[
4
e
(2q)
= 2 |u
0
|
R
2
[y[
2
_
1 +
N 4
[y[
2
e
(2q)
q
(2 |u
0
|
)
q1
[y[
q
_
R
2
[y[
2
_
q1
_
2 |u
0
|
R
2
[y[
2
_
1
4
[y[
2
e
(2q)
q
(2 |u
0
|
)
q1
R
q
_
2 |u
0
|
R
2
[y[
2
_
1
_
4 +q (2 |u
0
|
)
q1
_
R
q
_
0 .
Consequently, for R
_
4 +q 2
q1
|u
0
|
q1
_
1/q
,
R
is a supersolution to (6.10) in
(0, ) R
N
B(0, R).
Consider next (0, 1). Since u
0
(
0
(R
N
), there is r

_
4 +q 2
q1
|u
0
|
q1
_
1/q
,
r
1/ such that u
0
(y) /2 if [y[ r
. Then, if y R
N
B(0, r
), we have
u
0
(y)/2 0
r
(y) while v(, y)/2 |u
0
|
=
r
(y) if 0 and [y[ = r
.
Since
r
is a supersolution to (6.10) as previously established and v /2 is a so-
lution to (6.10), the comparison principle entails that v(, y) /2
r
(y) for
0 and y R
N
B(0, r
). Since
r
(y) 0 as [y[ , there is R
such
that
r
(y) /2 for [y[ R
which completes the proof of (6.17).
It actually follows from Lemma 6.5 that v() belongs to (

0
(R
N
) for every 0
in a way which is uniform with respect to 0.
Lemma 6.6. For y R
N
, we have
H
(y) v
(y) v
(y) . (6.18)
Proof. For 0, y R
N
,
0
R and R
N
, we set H(, y,
0
, ) :=
0
y q [[
q
+ q K
1
e
(2q)
, the constant K
1
being dened in (6.12) and
(6.13). Since v(, y) K
1
for (, y) (1, ) R
N
by (6.13), we deduce
2. THE REACTION-DOMINATED CASE: = 1 157
from (6.10) that H(, y,
v(, y), v(, y)) = q e

(2q)
(v(, y) + K
1
) 0 for
(, y) (1, ) R
N
. Consequently,
v is a supersolution to H(, y,
z, z) = 0 in (1, ) R
N
. (6.19)
Next x
0
> 1 and denote by w the (viscosity) solution to
w y w q [w[
q
= 0 , (, y) (
0
, ) R
N
,
w(
0
) = v(
0
) , y R
N
.
On the one hand, the function w dened by
w(, y) := w(, y) q K
1
_

0
e
(2q)s
ds , (, y) (
0
, ) R
N
,
is the (viscosity) solution to H(, y,
w, w) = 0 in (
0
, ) R
N
with initial
condition w(
0
) = v(
0
). Recalling (6.19), we infer from the comparison principle
that
w(, y) v(, y) for (, y) (
0
, ) R
N
. (6.20)
On the other hand, it follows from Corollary A.4 that
lim
sup
yR
N
w()
_
|v(
0
)|

q
[y[
q/(q1)
_
+
= 0 .
We may then pass to the limit as in (6.20) to conclude that
_
|v(
0
)|

q
[y[
q/(q1)
_
+
q K
1
_

0
e
(2q)s
ds v
(y) v
(y)
for y R
N
. Letting
0
in the above inequality gives (6.18).
We are now in a position to complete the proof of Theorem 6.4. Fix (0, 1).
On the one hand, it readily follows from Lemma 6.5 that v
(y) for [y[ R

1/. On the other hand, H
(0) = I
() and the continuity of H
ensures that
there is r
(0, ) such that H
(y) I
() for y B(0, r
). Recalling (6.15),
we conclude that
_
_
_
v
(y) 0 H
(y) if [y[ = R
,
v
(y) I
() H
(y) if [y[ = r
.
(6.21)
Now, introducing (y) =
q
[y[
q/(q1)
/2 and recalling that H is dened in (6.16),
we have
H(y, (y)) =
q
q
2(q 1)
[y[
q/(q1)
_
1
1
2
q1
_
> 0 if r
< [y[ < R
. (6.22)
Summarizing, putting
:=
_
y R
N
: r
< [y[ < R
_
, we have shown that H
and v
are supersolution and subsolution to (6.16) in
, respectively, with
v
on
by (6.21). Owing to (6.22) and the concavity of H with

respect to its second variable, we may apply [36, Theorem 1] to conclude that
v
in
. Letting 0, we end up with v
in R
N
. Recalling
(6.18), we have thus established that v
= v
= H
in R
N
. In addition, owing to
[4, Lemma V.1.9] or [6, Lemme 4.1], the equality v
= v
and (6.14) provide the

convergence of (v())
>0
towards H
uniformly on every compact subset of R

N
as
. Combining this last property with Lemma 6.5 implies that
lim
|v() H
= 0 . (6.23)
Theorem 6.4 then readily follows after writing the convergence (6.23) in the original
variables (t, x) and noticing that |h
(1 +t) h
(t)|
0 as t .
Theorem 6.1 was rst proved for p = 1 in [23] when = 1 and [39] when
= 1. The extension to the other L
p
-norms and the gradient was done in [9].
Theorem 6.4 is proved in [9, Theorem 2.6] with a dierent method which uses rather
the Hopf-Lax-Oleinik representation formula than the relaxed half-limits technique.
When = 1 and q (1, q
), the large time behaviour of non-negative solutions

to (1.1), (1.2) is also self-similar (at least for initial data vanishing suciently
rapidly for large values of x) but neither the diusion nor the absorption dominate
for large times. It is actually a self-similar solution W to (1.1) which characterises
the large time behaviour of non-negative solutions u to (1.1), (1.2) with an initial
condition satisfying (1.4) and the decay condition
ess lim
|x|
[x[
(2q)/(q1)
u
0
(x) = 0 .
More precisely, it has been shown in [9] that
lim
t
_
t
(N+1)(qq)/(2(q1))
|u(t) W(t)|
1
+t
(2q)/(2(q1))
|u(t) W(t)|
_
= 0
and W satises W(t, x) = t
(2q)/(2(q1))
W
_
1, xt
1/2
_
for (t, x) Q
. In fact,
W is a very singular solution to (1.1) in the sense of [24], that is, W is a classical
solution to (1.1) in Q
, W(t) L
1
(R
N
) for each t > 0,
lim
t0
_
B(0,r)
W(t, x) dx = and lim
t0
_
{|x|r}
W(t, x) dx = 0
for every r > 0. The existence and uniqueness of such a solution have been studied
in [10, 12, 28, 44].
Finally, convergence to self-similarity also takes place for the critical exponent
q = q
when = 1, the scaling prole being the heat kernel g as for q > q
(see
Theorem 6.1) but with an additional logarithmic scaling factor. More precisely, it
is shown in [30] that, if q = q
and u
0
satises (1.4) together with u
0
L
2
(R
N
; (1+
[x[
2m
) dx) for some m > N/2, the corresponding solution u to (1.1), (1.2) satises
lim
t
t
(N(p1))/2p
(log t)
N+1
_
_
_
_
u(t)
M
(log t)
N+1
g(t)
_
_
_
_
p
= 0
with M
:= (N + 1)
N+1
|G|
(N+2)
q
(recall that g and G are dened in (1.7)).
Bibliography
[1] Laurent Amour and Matania Ben-Artzi, Global existence and decay for
viscous HamiltonJacobi equations, Nonlinear Anal. 31 (1998), 621628.
[2] Daniele Andreucci, Anatoli F. Tedeev and Maura Ughi, The Cauchy
problem for degenerate parabolic equations with source and damping,
Ukrainian Math. Bull. 1 (2004), 123.
[3] Donald G. Aronson and Philippe Benilan, Regularite des solutions de
lequation des milieux poreux dans R
N
, C. R. Acad. Sci. Paris Ser. I
Math. 288 (1979), 103105.
[4] Martino Bardi and Italo Capuzzo-Dolcetta, Optimal Control and Vis-
cosity Solutions of HamiltonJacobi-Bellman Equations, Systems Con-
trol Found. Appl., Birkhauser, Boston, 1997.
[5] Guy Barles, Asymptotic behavior of viscosity solutions of rst order
HamiltonJacobi equations, Ricerche Mat. 34 (1985), 227260.
[6] Guy Barles, Solutions de Viscosite des Equations dHamiltonJacobi,
Mathematiques & Applications 17, Springer-Verlag, Berlin, 1994.
[7] Guy Barles and Benot Perthame, Exit time problems in optimal control
and vanishing viscosity method, SIAM J. Control Optim. 26 (1988),
11331148.
[8] William Beckner, Geometric asymptotics and the logarithmic Sobolev
inequality, Forum Math. 11 (1999), 105137.
[9] Sad Benachour, Grzegorz Karch and Philippe Lauren cot, Asymptotic
proles of solutions to viscous HamiltonJacobi equations, J. Math.
Pures Appl. (9) 83 (2004), 12751308.
[10] Sad Benachour, Herbert Koch and Philippe Lauren cot, Very singular
solutions to a nonlinear parabolic equation with absorption. II Unique-
ness, Proc. Roy. Soc. Edinburgh Sect. A 134 (2004), 3954.
[11] Sad Benachour and Philippe Lauren cot, Global solutions to viscous
HamiltonJacobi equations with irregular initial data, Comm. Partial
Dierential Equations 24 (1999), 19992021.
[12] Sad Benachour and Philippe Lauren cot, Very singular solutions to a
nonlinear parabolic equation with absorption. I. Existence, Proc. Roy.
Soc. Edinburgh Sect. A 131 (2001), 2744.
[13] Sad Benachour, Philippe Lauren cot and Didier Schmitt, Extinction and
decay estimates for viscous HamiltonJacobi equations in R
N
, Proc.
Amer. Math. Soc. 130 (2002), 11031111.
159
160 Bibliography
[14] Sad Benachour, Philippe Lauren cot, Didier Schmitt and Philippe Sou-
plet, Extinction and non-extinction for viscous HamiltonJacobi equa-
tions in R
N
, Asymptot. Anal. 31 (2002), 229246.
[15] Sad Benachour, Bernard Roynette and Pierre Vallois, Solutions fonda-
mentales de u
t
1
2
u
xx
= [u
x
[, Asterisque 236 (1996), 4171.
[16] Sad Benachour, Bernard Roynette and Pierre Vallois, Asymptotic esti-
mates of solutions of u
t
1
2
u = [u[ in R
+
R
d
, d 2, J. Funct.
Anal. 144 (1997), 301324.
[17] Matania Ben-Artzi, Global existence and decay for a nonlinear parabolic
equation, Nonlinear Anal. 19 (1992), 763768.
[18] Matania Ben-Artzi, Jonathan Goodman and Arnon Levy, Remarks on
a nonlinear parabolic equation, Trans. Amer. Math. Soc. 352 (1999),
731751.
[19] Matania Ben-Artzi and Herbert Koch, Decay of mass for a semilin-
ear parabolic equation, Comm. Partial Dierential Equations 24 (1999),
869881.
[20] Matania Ben-Artzi, Philippe Souplet and Fred B. Weissler, The lo-
cal theory for viscous HamiltonJacobi equations in Lebesgue spaces, J.
Math. Pures Appl. (9) 81 (2002), 343378.
[21] Philippe Benilan, Evolution equations and accretive operators, Lecture
notes taken by S. Lenhardt, Univ. of Kentucky, Spring 1981.
[22] Philippe Benilan and Michael G. Crandall, The continuous dependence
on of solutions of u
t
(u) = 0, Indiana Univ. Math. J. 30 (1981),
161177.
[23] Piotr Biler, Mohammed Guedda and Grzegorz Karch, Asymptotic prop-
erties of solutions of the viscous HamiltonJacobi equation, J. Evolution
Equations 4 (2004), 7597.
[24] Ham Brezis, Lambertus A. Peletier and David Terman, A very singu-
lar solution of the heat equation with absorption, Arch. Rational Mech.
Anal. 95 (1986), 185209.
[25] Michael G. Crandall, Hitoshi Ishii and Pierre-Louis Lions, Users guide
to viscosity solutions of second order partial dierential equations, Bull.
Amer. Math. Soc. (N.S.) 27 (1992), 167.
[26] Lawrence C. Evans, Regularity properties for the heat equation subject
to nonlinear boundary constraints, Nonlinear Anal. 1 (1977), 593602.
[27] Lawrence C. Evans, Partial Dierential Equations, Grad. Stud. Math.
19, Amer. Math. Soc., Providence, RI, 1998.
[28] Zhong Bo Fang and Minkyu Kwak, Complete classication of shape
functions of self-similar solutions, J. Math. Anal. Appl. 330 (2007),
14471464.
[29] Eduard Feireisl and Philippe Lauren cot, The L
1
-stability of constant
states of degenerate convection-diusion equations, Asymptot. Anal. 19
(1999), 267288.
[30] Thierry Gallay and Philippe Lauren cot, Asymptotic behavior for a vis-
cous HamiltonJacobi equation with critical exponent, Indiana Univ.
Math. J. 56 (2007), 459479.
Bibliography 161
[31] Brian H. Gilding, The Cauchy problem for u
t
= u +[u[
q
, large-time
behaviour, J. Math. Pures Appl. (9) 84 (2005), 753785.
[32] Brian H. Gilding, Mohammed Guedda and Robert Kersner, The Cauchy
problem for u
t
= u+[u[
q
, J. Math. Anal. Appl. 284 (2003), 733755.
[33] Mohammed Guedda and Robert Kersner, Self-similar solutions to the
generalized deterministic KPZ equation, NoDEA Nonlinear Dierential
Equations Appl. 10 (2003), 113.
[34] Richard S. Hamilton, A matrix Harnack estimate for the heat equation,
Comm. Anal. Geom. 1 (1993), 113126.
[35] Miguel A. Herrero and Juan Luis V azquez, Asymptotic behaviour of
the solutions of a strongly nonlinear parabolic problem, Ann. Fac. Sci.
Toulouse Math. (5) 3 (1981), 113127.
[36] Hitoshi Ishii, A simple, direct proof of uniqueness for solutions of the
HamiltonJacobi equations of eikonal type, Proc. Amer. Math. Soc. 100
(1987), 247251.
[37] Yong Jung Kim, An Oleinik type estimate for a convection-diusion
equation and convergence to N-waves, J. Dierential Equations 199
(2004), 269289.
[38] Olga A. Ladyzenskaja, Vsevolod A. Solonnikov and Nina N. Uralceva,
Linear and Quasilinear Equations of Parabolic Type, Transl. Math.
Monogr. 23, Amer. Math. Soc., Providence, RI, 1988.
[39] Philippe Lauren cot and Philippe Souplet, On the growth of mass for a
viscous HamiltonJacobi equation, J. Anal. Math. 89 (2003), 367383.
[40] Philippe Lauren cot and Philippe Souplet, Optimal growth rates for a
viscous HamiltonJacobi equation, J. Evolution Equations 5 (2005),
123135.
[41] Chi-Tien Lin and Eitan Tadmor, L
1
-stability and error estimates for
approximate HamiltonJacobi solutions, Numer. Math. 87 (2001), 701
735.
[42] Pierre-Louis Lions, Regularizing eects for rst-order HamiltonJacobi
equations, Appl. Anal. 20 (1985), 283307.
[43] Gawtum Namah and Jean-Michel Roquejore, Remarks on the long
time behaviour of the solutions of HamiltonJacobi equations, Comm.
Partial Dierential Equations 24 (1999), 883893.
[44] Yuanwei Qi and Mingxin Wang, The self-similar proles of generalized
KPZ equation, Pacic J. Math. 201 (2001), 223240.
[45] Jean-Michel Roquejore, Convergence to steady states or periodic solu-
tions in a class of HamiltonJacobi equations, J. Math. Pures Appl. (9)
80 (2001), 85104.
APPENDIX A
Self-similar large time behaviour
1. Convergence to self-similarity for the heat equation
We recall the following well-known result.
Proposition A.1. Consider u
0
L
1
(R
N
) and put
M
0
:=
_
R
N
u
0
(x) dx.
Then, for p [1, ], we have
lim
t
t
N(p1)/(2p)
_
_
e
t
u
0
M
0
g(t)
_
_
p
= 0 , (A.1)
lim
t
t
N(p1)/(2p)+(1/2)
_
_
e
t
u
0
M
0
g(t)
_
_
p
= 0 , (A.2)
where the functions g and G are given by (1.7).
Proof. For (t, x) Q
, we put
c(t, x) :=
_
e
t
u
0
_
(x) =
1
t
N/2
_
R
N
G
_
x y
t
1/2
_
u
0
(y) dy
and z(t, x) := c(t, x) M
0
g(t, x). An alternative formula for z reads
z(t, x) =
1
t
N/2
_
R
N
_
G
_
x y
t
1/2
_
G
_
x
t
1/2
_
_
u
0
(y) dy .
On the one hand,
|z(t)|
1

_
R
N
[u
0
(y)[
_
R
N
G(x) G
_
x yt
1/2
_
dxdy .
Noting that
_
R
N
G(x) G
_
x yt
1/2
_
dx 2 |G|
1
<
and
lim
t
_
R
N
G(x) G
_
x yt
1/2
_
dx = 0
for each y R
N
, we infer from the Lebesgue dominated convergence theorem that
lim
t
|z(t)|
1
= 0 . (A.3)
On the other hand, since c(t) = e
(t/2)
c(t/2) and g(t) = e
(t/2)
g(t/2), we have
[z(t, x)[
1
(2t)
N/2
_
R
N
z
_
t
2
, y
_
exp
_
[x y[
2
2t
_
dy
1
(2t)
N/2
_
_
_
_
z
_
t
2
__
_
_
_
1
,
163
164 A. SELF-SIMILAR LARGE TIME BEHAVIOUR
hence
t
N/2
|z(t)|

1
(2)
N/2
_
_
_
_
z
_
t
2
__
_
_
_
1
. (A.4)
Combining (A.3) and (A.4) yields (A.1) for p = . The result for p (1, ) then
follows by interpolation with (A.3).
Similarly,
z(t, x) =
2
(N+1)/2
t
(N+1)/2
_
R
N
G
_
2
1/2
(x y)
t
1/2
_
z
_
t
2
, y
_
dy
and we infer from the Young inequality that
|z(t)|
p

2
1/2
t
1/2
|G|
1
_
_
_
_
z
_
t
2
__
_
_
_
p
for p [1, ]. We now use (A.1) to estimate the right-hand side of the previous
inequality and obtain (A.2).
2. Convergence to self-similarity for HamiltonJacobi equations

This section is devoted to the large time behaviour of non-negative (viscosity)
solutions to the HamiltonJacobi equation (1.9). We begin with the absorption
case = 1 and establish the following result for compactly supported initial data.
Theorem A.2. Assume that q > 1 and = 1. Consider a non-negative
function u
0
((R
N
) with compact support and denote by h the corresponding
viscosity solution to the HamiltonJacobi equation (1.9). Then
lim
t
_
_
_t
1/(q1)
h(t) h
_
_
_
= 0 (A.5)
with
h
(x) =
q
d
_
x, R
N
T
0
_
q/(q1)
, x R
N
,
q
:= (q 1) q
q/(q1)
and T
0
:=
_
x R
N
: u
0
(x) > 0
_
.
Proof. We rst recall that h is given by the HopfLaxOleinik formula (1.10)
h(t, x) := inf
yR
N
_
u
0
(y) +
q
[x y[
q/(q1)
t
1/(q1)
_
for (t, x) Q
. (A.6)
We next introduce the positivity set T(t) of h at time t > 0 and claim that
T(t) :=
_
x R
N
: h(t, x) > 0
_
= T
0
. (A.7)
Indeed, if t > 0 and x , T
0
, we have u
0
(x) = 0 and the choice y = x in the
right-hand side of (A.6) ensures that 0 h(t, x) u
0
(x) = 0, whence x , T(t).
Next, if t > 0 and x T
0
, we assume for contradiction that h(t, x) = 0. Since
u
0
(y) +
q
[x y[
q/(q1)
t
1/(q1)
as [y[ , the inmum in (A.6) is
attained and there is y R
N
such that
0 = h(t, x) = u
0
(y) +
q
[x y[
q/(q1)
t
1/(q1)
.
This readily implies that y = x and u
0
(y) = u
0
(x) = 0, and a contradiction.
Therefore, x T(t) and the proof of (A.7) is complete.
2. CONVERGENCE TO SELF-SIMILARITY FOR HAMILTONJACOBI EQUATIONS 165
We next introduce (t, x) := t
1/(q1)
h(t, x) so that
(t, x) := inf
yR
N
_
t
1/(q1)
u
0
(y) +
q
[x y[
q/(q1)
_
(A.8)
for (t, x) [0, ) R
N
. We rst note that
0 (t
1
, x) (t
2
, x)
q
d
_
x, R
N
T
0
_
q/(q1)
= h
(x) (A.9)
for x R
N
and 0 < t
1
< t
2
, the last inequality being deduced from (A.8) by taking
the inmum only on R
N
T
0
where u
0
vanishes. We next consider x T
0
and

_
0, d
_
x, R
N
T
0
_
/2
_
. Setting
m
:= inf
_
u
0
(y) : d
_
y, R
N
T
0
_

_
> 0 , t
:=
_
q
m
_
q1
d
_
x, R
N
T
0
_
q
and choosing t t
, we have the following alternative for y R

N
: either d
_
y, R
N
T
0
_
and
t
1/(q1)
u
0
(y) +
q
[x y[
q/(q1)
m
t
1/(q1)

q
d
_
x, R
N
T
0
_
q/(q1)
or d
_
y, R
N
T
0
_
< and
t
1/(q1)
u
0
(y) +
q
[x y[
q/(q1)

q
_
d
_
x, R
N
T
0
_
d
_
y, R
N
T
0
__
q/(q1)

q
_
d
_
x, R
N
T
0
_
_
q/(q1)
.
Consequently,
(t, x)
q
_
d
_
x, R
N
T
0
_
_
q/(q1)
for t [t
, ) R
N
. (A.10)
The convergence (A.5) then readily follows from (A.9) and (A.10) after passing to
the limit 0.
We next turn to the case = 1 and prove the following result.

Theorem A.3. Assume that q > 1 and = 1. Consider a non-negative
function u
0
(
0
(R
N
) and denote by h the corresponding viscosity solution to the
HamiltonJacobi equation (1.9). Then
lim
t
sup
xR
N
h(t, x) H
_
x
t
1/q
_
= 0 (A.11)
with
H
(x) =
_
|u
0
|

q
[x[
q/(q1)
_
+
, x R
N
.
Proof. As before, h is given by the HopfLaxOleinik formula (1.10)
h(t, x) := sup
yR
N
_
u
0
(y)
q
[x y[
q/(q1)
t
1/(q1)
_
for (t, x) Q
. (A.12)
We rst show that
|h(t)|
= |u
0
|
, t 0 . (A.13)
Indeed, recall that h(t, x) u
0
(x) for (t, x) [0, ) R
N
which ensures that
|h(t)|
|u
0
|
while the reverse inequality readily follows from the representa-

tion formula (A.12).
Next, it is rather easy to deduce from (A.12) that h(t, x) |u
0
|
as t
for each x R
N
. The goal is now to study more precisely this convergence. To this
end we introduce the self-similar (or scaling) variables
s =
1
q
log (1 +t) , =
x
(1 +t)
1/q
,
and the new unknown function dened by
h(t, x) =
_
log (1 +t)
q
,
x
(1 +t)
1/q
_
, (t, x) [0, ) R
N
.
Then (s, ) = h(e
qs
1, e
s
) and it follows from (A.12) that
(s, ) = sup
yR
N
_
u
0
(y)
q
y e
s
q/(q1)
_
1 e
qs
_
1/(q1)
_
for (s, ) [0, ) R
N
. Since (s, ) u
0
(e
s
) 0, we have in fact
(s, ) = sup
yR
N
_
_
u
0
(y)
q
y e
s
q/(q1)
_
1 e
qs
_
1/(q1)
_
+
_
for (s, ) [0, ) R
N
. Let us point out here that, as a consequence of (1.9), is
actually the viscosity solution to
s
= +q [[
q
, (s, ) Q
, (A.14)
(0) = u
0
, R
N
. (A.15)
Consider now (0, 1). As u
0
(
0
(R
N
), there is R
> 0 such that

u
0
(y) for [y[ R
. (A.16)
On the one hand, if (s, ) [log R
, ) R
N
and y R
N
, we have either [y[ R
and
_
u
0
(y)
q
[ y e
s
[
q/(q1)
(1 e
qs
)
1/(q1)
_
+
_
u
0
(y)
q
[[
q/(q1)
_
+
_
u
0
(y)
q
[ y e
s
[
q/(q1)
(1 e
qs
)
1/(q1)
_
+
+
_
u
0
(y)
q
[[
q/(q1)
_
+
2
2. CONVERGENCE TO SELF-SIMILARITY FOR HAMILTONJACOBI EQUATIONS 167
by (A.16) or y B(0, R
) and
_
u
0
(y)
q
[ y e
s
[
q/(q1)
(1 e
qs
)
1/(q1)
_
+
_
u
0
(y)
q
[[
q/(q1)
_
+
y e
s
q/(q1)
_
_
1 e
qs
_
1/(q1)
1
_
+
q
y e
s
q/(q1)
[[
q/(q1)

q
_
[[ +R
e
s
_
q/(q1)
_
_
1 e
qs
_
1/(q1)
1
_
+
q
q
q 1
_
[[ +[y[ e
s
_
1/(q1)
[y[ e
s

q
(1 +[[)
1/(q1)
_
q
q 1
+ 1 +[[
_
_
_
1 e
qs
_
1/(q1)
1 +R
e
s
_
as s log R
. Combining the above two estimates gives
(s, ) sup
yR
N
_
_
u
0
(y)
q
[[
q/(q1)
_
+
_
C ([[ + 1)
q/(q1)
_
_
1 e
qs
_
1/(q1)
1 +R
e
s
_
+ 2 ,
whence
[(s, ) H
()[ (A.17)
C ([[ + 1)
q/(q1)
_
_
1 e
qs
_
1/(q1)
1 +R
e
s
_
+ 2
for (s, ) [log R
, ) R
N
. On the other hand, if s log(R
), [[ :=
1 + (|u
0
|
/
q
)
(q1)/q
and y R
N
, we have either [ y e
s
[ 1 and
u
0
(y)
q
y e
s
q/(q1)
_
1 e
qs
_
1/(q1)
_
1 e
qs
_
1/(q1)
_
|u
0
|
_
1 e
qs
_
1/(q1)
y e
s
q/(q1)
_
_
1 e
qs
_
1/(q1)
_
|u
0
|

q
( 1)
q/(q1)
_
0 ,
or [ y e
s
[ < 1 and
[y[ [ e
s
[ [y e
s
[ e
s
( 1) e
s
= e
s
R
,
so that
u
0
(y)
q
y e
s
q/(q1)
_
1 e
qs
_
1/(q1)

by (A.16). Therefore,
(s, ) for (s, ) [log R
, ) R
N
B(0, ) . (A.18)
It then follows easily from (A.17) and (A.18) that |(s) H
0 as
s . Returning to the original variables (t, x), we conclude that
lim
t
sup
xR
N
h(t, x) H
_
x
(1 +t)
1/q
_
= 0 .
Since
lim
t
sup
xR
N
_
x
(1 +t)
1/q
_
H
_
x
t
1/q
_
= 0 ,
the convergence (A.11) follows.
We have actually shown the following result.
Corollary A.4. Let q > 1 and consider a non-negative function u
0
(
0
(R
N
).
If denotes the corresponding viscosity solution to (A.14), (A.15), then
lim
s
|(s) H
= 0 ,
the function H
being dened in Theorem A.3.

Remark A.5. Related results may be found in [5, Section II] for the Hamilton
Jacobi equation
t
w + H(w) = 0 in Q
, the function H being non-negative and

continuous with suitable properties.
Part 4
An area-preserving crystalline
curvature ow equation
Shigetoshi Yazaki
2000 Mathematics Subject Classication. 34A34, 34K25, 39A12, 52A10, 53A04,
74N05, 34K12, 34K28, 65L12, 65M06, 65M12, 65D99, 41A25, 82D25
Key words and phrases. evolving interfaces, polygonal curves, curvature,
curve-shortening, crystalline curvature, anisotropy, interfacial energy, the Frank
diagram, the Wul shape, admissibility, area-preserving, single ice crystal,
internal melting, Tyndall gure, negative crystal, vapor gure, crystal growth
Abstract. In this note, we will show how to derive crystalline curvature ow
equation and its area-preserving version. This ow is regarded as a gradient
ow of total interfacial energy with the singular anisotropy, then the usual
dierential calculation is not allowed in the classical sense. Because of this, a
dierence calculation is used instead of dierential calculation and polygonal
curves are used instead of smooth curves. This consideration will be focused
in the former and main part of this note, which includes the following three
topics:
plane curves, geometric quantities on them, and evolution of them;
anisotropy, the Frank diagram, and the Wul shape, etc.; and
crystalline curvature ow equation and its area-preserving version.
In the latter part, we will see the following two applications:
a numerical scheme of the area-preserving crystalline curvature ow; and
a modeling perspective on producing negative ice crystals.
Contents
Preface 173
Chapter 1. An area-preserving crystalline curvature ow equation:
introduction to mathematical aspects, numerical computations,
and a modeling perspective 175
1. Introduction 175
2. Plane curve 175
3. Moving plane curve 178
4. Curvature ow equations 179
5. Anisotropy 181
6. The Frank diagram and the Wul shape 183
7. Crystalline energy 187
8. Crystalline curvature ow equations 189
9. An area-preserving motion by crystalline curvature 195
10. Scenario of the proof of Theorem 1.50 196
11. Numerical scheme 198
gures produced by freezing of internal melt gures 202
Bibliography 207
Index 211
Appendix A. 213
1. Strange examples 213
2. A non-concave curve 213
3. Anisotropic inequalityproof of Lemma 1.53 213
171
Preface
The so-called crystalline curvature ow equation appeared at the end of 1980s
in the eld of mathematical sciences. This equation is regarded as a weak form of
an equation of interfacial motion with a singular interfacial energy. The equation
is described as a system of ODEs and governs motion of polygonal curves in two
dimensional case. If the number of vertices of curves is large, the motion can
approximate the one described by PDE with a smooth interfacial energy. On the
other hand, if the number is small, the motion is sometimes strikingly dierent from
the one in the smooth case. This phenomena suggest that research of this eld will
continue to a wide range of applications.
The lecture note will target beginners of the research eld. Therefore, only two
dimensional closed curves will be focused and I will put emphasis on a fundamental
part of mathematical introduction to crystalline curvature ow equations. I hope all
the readers will follow several calculations and challenge to solve given exercises. In
addition, we will touch on topics of a numerical scheme and a modeling of negative
crystal, and they may give directions of applications.
The note is based on the lectures which were held at Czech Technical University
in Prague, September 4 and 6, 2006, and December 4, 6, 11 and 13, 2007, while the
author was visiting Czech Republic within the period from August to September,
2006, and from March, 2007 to March, 2008; this visit was sponsored by Faculty of
Nuclear Sciences and Physical Engineering, Czech Technical University in Prague,
within the Jindrich Necas Center for Mathematical Modeling (Project of the Czech
Ministry of Education, Youth and Sports LC 06052). The author would like to
thank Professor Michal Benes for giving me this opportunity not only to have the
series of lectures but also to visit Czech Republic as a young researcher.
Shigetoshi Yazaki
Miyazaki, June 2008
173
CHAPTER 1
An area-preserving crystalline curvature ow
equation: introduction to mathematical aspects,
numerical computations, and a modeling
perspective
1. Introduction
Motion by curvature is generally referred as a motion of curves in the plane or
surfaces in space which change its shape in time and depend on its bend, especially
on its curvature. They are also called curvature ows, since one tracks ows of the
family of curves and surfaces parametrized in time. The curvature ow equation
is a general term which describes such ows, and has been investigated by many
scientists and mathematicians since the 1950s. At the end of 1980s, J. E. Tay-
lor, and S. Angenent and M. E. Gurtin focused on motion of polygonal curves by
crystalline curvature in the plane, and since then crystalline curvature ow equa-
tion has been studied under various kinds of evolution law by several authors. We
refer the reader to the pioneer works Taylor [Tay91a, Tay93] and Angenent and
Gurtin [AG89], and the surveys by Taylor, Cahn and Handwerker [TCH92] and the
books by Gurtin [Gur93] for a geometric and physical background. Also one can
nd essentially the same method of crystalline as a numerical scheme for curvature
ow equation in Roberts [Rob93]. We refer the reader Almgren and Taylor [AT95]
for detailed history. Besides this crystalline strategy, other strategies by subdier-
ential and level-set method have been extensively studied. See Giga [Gig00] and
references therein.
This note will focus mainly on motion of smooth and piecewise linear curves
in the plane, and touch on a numerical scheme of crystalline curvature ow equa-
tion which is a kind of so-called direct approach. On motion of surfaces, indirect
approach (ex. level set methods), other curvature ows, and physical background,
the reader is referred to the books by Gurtin [Gur93] and Sethian [Set96], and
the surveys by [TCH92] and Giga [Gig95, Gig98, Gig00], and references therein,
respectively. In the last section, a modeling perspective will be reported.
2. Plane curve
In this section we will explain plane curve and several geometric quantities
on it.
A plane curve is dened as the following map x:
= x() R
2
; Q,
175
176 1. AN AREA-PRESERVING CRYSTALLINE CURVATURE FLOW EQUATION
where Q R is an interval
1
, and the map x : Q R
2
is continuous and
surjective
2
. The plane curve x is said to be closed if Q = [a, b] and x(a) = x(b).
If x is not closed, then it is called open. The plane curve x is said to be simple if
the map
3
x is injective
4
. A simple and closed plane curve is called Jordan curve.
If x is a Jordan curve, then is also called Jordan curve.
Remark 1.1 (Jordan curve theorem). If is a Jordan curve, then R
2
has
two components
5
, where is the boundary of each.
In what follows, curves mean plane curves, and we will omit the term plane.
A curve x is said to be C
k
-curve (k = 1, 2, . . .) if the map x is C
k
-class function.
We put
g() = [
x()[.
Here and hereafter, we denote
F = F/. A curve x is said to be immersed if

x is C
1
-curve and g() ,= 0 holds for all .
For closed curves, we take = R/Z with parameter (modulo 1), and use the
interval = [0, 1]. Then a closed plane curve x can be described as
= x() R
2
; .
Remark 1.2. If x is a closed immersed curve, then
s() =
_

0
g() d
is increasing C
1
-function for , since g is continuous function and
s() =
g() > 0 holds. Therefore, by virtue of an implicit function theorem, is a func-
tion of s, and the closed immersed curve x can be parametrized by the arc-length
parameter s:
= x(s) R
2
; s [0, L], L =
_
ds =
_
g() d,
where L is the total length of .
A curve x is said to be embedded if x is immersed and simple. Hence, if has
self-intersection or self-tangency, then x is not embedded curve. Note that all the
above terminologies are dened for the map x, not for the set . This distinction
is important. See section 1.
Let x be an immersed curve. The unit tangent vector is dened as t =
x/[
x[ =
s
x, and the unit normal vector n is dened such as det(n, t) = 1,
or t = n
, where (a, b)
= (b, a). Note that the derivative with respect to

the arc-length
s
along the curve is uniquely dened, while the arc-length s is
unique only up to a constant. The positive direction of x is the direction which
is increasing. If the curve is a Jordan curve, we take the positive direction as
1
Non-empty and connected subset in R.
2
Onto map, i.e., for any x there exists a Q for which x() = x.
3
If x is closed, we take Q = [a, b).
4
One-to-one map, i.e., for any
1
,
2
Q if
1
=
2
, then x(
1
) = x(
2
) holds.
5
So-called inside D and outside E. Both D and E are connected, respectively, but its
direct sum D E is not connected.
2. PLANE CURVE 177
anticlockwise, i.e., the direction which one can always see enclosed region on the left
side when traveling along the curve in the direction of increasing. See Figure 1.
Let x be an immersed C
2
-curve. The curvature K of x is dened as
s
t = Kn and
s
n = Kt, (1.1)
and it is often called the curvature in the direction of n. The signed convention
of the curvature K is that K = 1 if and only if x is the unit circle. See Figure 1.
K > 0
K < 0
n
t
positive direction
Figure 1. The normal vector n, the tangent vector t, and the
curvature K on a closed embedded curve.
Remark 1.3. Equations (1.1) are called FrenetSerret formulae for plane
curves. From n n = 1, we obtain
s
(n n) = 2(
s
n) n = 0. Here and hereafter,
a b is the Euclidean inner product between a and b. Hence there exists a function
c(s) such as
s
n = ct. Similarly, from t t = 1 and t n = 0, we obtain
s
(t t) =
2(
s
t) t = 0 and
s
(t n) = (
s
t) n + t (
s
n) = (
s
t) n + c = 0. Therefore
s
t = cn holds. We dened c as the curvature K.
Let be a Jordan curve and let be a region enclosed by
6
. The Jordan curve
is said to be convex or oval if for any
1
,
2
(
1
,=
2
) the line segment
7
S = [x(
1
), x(
2
)] is not contained outside of , i.e., S holds. In particular, if
S = x(
1
), x(
2
) holds for any
1
,
2
, then is called strictly convex.
The curve is called concave or nonconvex, if it is not convex.
Lemma 1.4. A closed embedded C
2
-curve is convex if and only if K() 0
holds for . Equivalently, the curve is concave if and only if K takes negative
value at some point.
Lemma 1.5. A closed embedded C
2
-curve is strictly convex, if K() > 0 holds
for .
Exercise 1.6. Prove the above two lemmas.
Remark 1.7. Convex curves admit including a line segment (K = 0 on some
portion of ), while strictly convex curve does not admit it. This does not mean
that the converse of Lemma 1.5 is true. See section 2.
6
is called inside of , and outside of is R
2
\.
7
In detail, S = {(1 )x(
1
) +x(
2
); [0, 1]}.
3. Moving plane curve
We consider a moving curve with the time t in some interval, for instance
t [0, T) with some T > 0. Let a smooth-closed-immersed moving curve be
(t) = x(, t) R
2
; .
By V we denote the growth speed at each point of (t) in the direction of n:
V =
t
x n.
It is well known that information of V determines the shape of (t), while the
tangential component v does not eect on the shape [EG87]:
t
x = V n +vt.
Here and hereafter
t
F means
t
F(, t), and
s
F is given in terms of :
s
F = g
1
F, g =
_
(
x)
2
+ (
y)
2
= [
x[, x(, t) = (x(, t), y(, t)).

That is, the arc-length element is ds = g d.
In what follows, we will see that fundamental properties of evolution of geo-
metric quantities.
Since
t
and
can commute:
t
t
, using the FrenetSerret formulae
we obtain:
Lemma 1.8.
t
g = (KV +
s
v)g.
This lemma and t = g
1
x yield:
Lemma 1.9.
t
t = (
s
V Kv)n and
t
n = (
s
V Kv)t.
Let = (, t) be the normal angle such as n = (cos , sin ). Then we have:
Lemma 1.10.
s
= K and
t
=
s
V +Kv.
By using the above lemmas, we have the evolution of K:
Lemma 1.11.
t
K = (
2
s
V +K
2
V ) +v
s
K.
Since s depends on t,
t
and
s
can not commute:
Lemma 1.12.
t
s
F =
s
t
F (KV +
s
v)
s
F.
We denote by L(t) the total length of (t): L =
_
ds =
_
g d. The evolution
of L(t) is given by:
Lemma 1.13.
t
L =
_
KV ds.
We denote by (t) the enclosed region of (t) and by A(t) the area of (t).
Using the relation
A =
1
2
_
div
_
x
y
_
dxdy =
1
2
_
x nds =
1
2
_
(x n)g d,
we have the evolution of A(t):
Lemma 1.14.
t
A =
_
V ds.
Exercise 1.15. Show above lemmas.
4. CURVATURE FLOW EQUATIONS 179
4. Curvature ow equations
To catch mathematical characteristics of curvature ows, we shall formulate
the typical problem. We mainly focus on smooth, embedded and closed curve, and
sometimes mention the case where curves are immersed.
Classical curvature ow. The classical and the typical equation which mo-
tivates investigation of curvature ow equations is
V = K. (1.2)
This is called the classical curvature ow equation. Figure 2 indicates evolution
of curve by (1.2). Convex part (K > 0) of nonconvex curve (t) moves towards the
direction of n (inward), and concave part (K < 0) moves towards the direction
of n (outward). Any embedded curve becomes convex in nite time [Gra87], and
any convex curve shrinks to a single point in nite time and its asymptotic shape
is a circle [GH86]. In the case where the solution curve is immersed, for example,
Figure 2. Evolution of a simple closed curve by (1.2) (from left to right).
the curve has a single small loop as in Figure 3, we can observe that the single
loop shrinks and the curve may have cusp in nite time. A detailed analysis of the
singularity near the extinction time was done by Angenent and Velazquez [AV95].
Figure 3. Evolution of a immersed closed curve by (1.2) (from
left to right). The small loop disappears in nite time.
Example of graph. In the case where (t) is described by a graph y = u(x, t),
we have the partial dierential equation
t
u =

2
x
u
1+(xu)
2
equivalent to (1.2) with
V = (0,
t
u)
T
n, n =
(xu,1)
T
1+(xu)
2
and K =

2
x
u
(1+(xu)
2
)
3/2
, where (a, b)
T
=
_
a
b
_
.
Figure 4 indicates its numerical example.
Gradient ow. As Figure 2 and Figure 4 suggested, the circumference L(t)
of the solution curve (t) is decreasing in time. This is the reason why (1.2) is
often called the curve-shortening equation. Actually, (1.2) is characterized by
the gradient ow of L as follows. By Lemma 1.13, the rate of change of the total
length L = L((t)) of (t) is
t
L((t)) =
_
KV ds.
Figure 4. Evolution of a graph y = u(x, t) by the curvature ow
equation
t
u =

2
x
u
1+(xu)
2
(0 < x < 1) (from left to right). The
boundary condition is u(0) = u(1) = 0, and the initial curve is
given by u(x, 0) = 0.1 sin(x) 0.3 sin(2x) + 0.2 sin(5x).
Let the inner product on be
F, G)
=
_
FGds.
Then we have
t
L((t)) = K, V )
. If the left-hand-side is regarded as a linear

functional of V in a suitable space with the metric , )
, then the gradient ow is

given by V = K, and
t
L((t)) = K, K)
0. Hence, (1.2) is the equation

which requires motion of curve in the most decreasing direction of L in the sense
of the above metric.
Exercise 1.16. Show
t
A(t) = 2.
Area-preserving curvature ow. Besides the curve-shortening ow, there
has also been interest in area-preserving ows. The typical ow is the gradient ow
of L along curves which enclose a xed area as follows:
V = K K, (1.3)
where K is the average of K: K =
_
K ds
__
ds = 2/L if is simple and closed,

K = 2/L if is immersed with the rotation number being = 2, 3, . . . We have
(1.3) as the gradient ow of L in the metric , )
under the constraint for V :

1, V )
= 0.
Figure 5. Evolution of a convex curve by (1.3) (from left to right).
Exercise 1.17. Show that
t
L(t) 0 and
t
A(t) = 0.
Gage [Gag86] proved that any convex curve converges to a circle as time tends
to innity (see Figure 5), and conjectured that a nonconvex curve may intersect;
this conjecture was proved rigorously by Mayer and Simonett [MS00].
5. ANISOTROPY 181
Figure 6. Evolution of a immersed closed curve by (1.3) (from
left to right). The small loop disappears in nite time (at least
numerically).
Open problems. The following two problems are still open: One is the prob-
lem whether the small loop may shrink or not. Figure 6 is a numerical exam-
ple, and it suggests that extinction occurs in nite time. The other one is the
problem whether any embedded curve becomes eventually convex or not, even if
self-intersection occurred. In the curve-shortening case, these problems have been
already analyzed, as it was mentioned above.
5. Anisotropy
In the eld of material sciences and crystallography, we need to explain the
anisotropyphenomenon of interface motion which depends on the normal direc-
tion, i.e., model equations have to contain eects of anisotropy of materials or
crystals.
Note. There are two kinds of anisotropy: one is kinetic anisotropy which ap-
pears in the problem of growth form, and the other is equilibrium anisotropy which
appears in the problem of equilibrium form.
Interfacial energy. For example, if the crystal growth of snow akes is re-
garded as motion of closed plane curve, then some points on the interface are
easy to growth in the case where their normal directions are in the six special direc-
tions. To explain these phenomena, it is convenient to dene an interfacial energy
on the curve which has line density (n) > 0. Integration of over is the total
interfacial energy:
E
=
_
(n) ds.
In the case where 1, E
is nothing but the total length L, and its gradient ow

was (1.2). For a general , what is the gradient ow of E
?
Extension of . The function (n) can be extended to the function x R
2
by putting
(x) =
_
[x[
_
x
|x|
_
, x ,= 0,
0, x = 0.
This extension is called the extension of positively homogeneous of degree 1 (in short,
PHD1) , since
(x) = (x)
holds for 0 and x R
2
. We will use the same notation for the extended
function. Hereafter we assume that is positive for x ,= 0 and positively homoge-
neous of degree 1 function. Following Kobayashi and Giga [KG01], we call such
a PPHD1 function.
Remark 1.18. The isotropic case (n) 1 holds if and only if (x) = [x[.
We assume that C
2
(R
2
0). Then the following properties hold:
(x) = (x) ( is homogeneous of degree 0),
(x) = x (x) (In particular, (n) = n (n) holds),
Hess (x) =
1
Hess (x) (Hess is homogeneous of degree 1),

where Hess =
_
11

12
21

22
_
is the Hessian of (
ij
=
2
/x
i
x
j
for i, j 1, 2
and x = (x
1
, x
2
)).
Exercise 1.19. Show these properties. (The answer is in [KG01, Appendix 1].)
Weighted curvature ow. The gradient ow of the total interfacial energy
E
((t)) =
_
(n) ds =
_
(n)g d
is given by
V =
(n),
(n) = (Hess (n)n
) n
K. (1.4)
This
(n) is called weighted curvature or anisotropic curvature , and (1.4) is

called the weighted curvature ow equation or anisotropic curvature ow
equation .
Exercise 1.20. Show
t
E
((t)) =
_
(n)V ds.
Exercise 1.21. Try to calculate the gradient ow of F
=
_
(x) ds.
Exercise 1.22. Try to calculate the gradient ow of G
=
_
(n)
ds,
= 1, 2.
Weighted curvature. Meaning of the weighted curvature
(n) will be clearer

as follows: Let be the exterior normal angle such as n = n() = (cos , sin ) and
t = t() = (sin , cos ). Put () = (n()). Then we obtain
(n()) = ( () +
2
())K.
Exercise 1.23. Prove this equation.
Example of . Let us construct an example of characteristic . For a natural
number M, we extend the line segment y = x (x [0, 2/M]) to 2/M-periodic
function p(x) such as
p(x) =
2
M
_
arctan
_
tan
_
M
2
x

2
__
+

2
_
,
and by (x) we denote a concave and symmetric function with respect to the central
line x = /M dened on [0, 2/M]. For example,
(x) =
sin x + sin(2/M x)
sin(2/M)
.
The following function
p
is a 2/M-periodic function with the number of peak
being M on [0, 2/M] (Figure 8 (left)).
p
() = (p()), R.
6. THE FRANK DIAGRAM AND THE WULFF SHAPE 183
Note that
p
is not dierentiable at 2k/M (k Z). Then by using
(x) =
_
x
2
+
2
, we construct a smooth function
(Figure 8 (left)):
() =
(
p
() 1) + 1, R.
As tends to innity,
(x) converges to [x[, and then
converges to
p
formally.
In general, if the curve (t) is strictly convex, then the solution K of (1.4)
satises the partial dierential equation
t
K = K
2
(
2
(V ) +(V )). See the book

of Gurtin [Gur93] or Exercise 1.24 below. Hence if +
2
> 0 holds, then the

PDE is quasi-linear strictly parabolic type and its Cauchy problem is solvable. We
note that
satises
+
2
> 0.
Exercise 1.24. Check the following story: Using the height function h = x n,
the curve = x is described as x = hn + (
h)t, if is strictly convex. Hence
h = K
1
h holds, and then we have
t
(K
1
) =
2
V +V by V =
t
h.
Figure 7 indicates evolution of solution curves of (1.4) by using the interfacial
energy density =
in the case where M = 6 and = 15. The asymptotic shape

seems to be a round hexagon which is strikingly dierent from the one in Figure 2.
Figure 7. Evolution of a simple closed curve by (1.4) (from left
to right). The initial curve is the same as in Figure 2.
6. The Frank diagram and the Wul shape
We will consider
p
and
by graphical constructions.
The Frank diagram. Figure 8 (left) indicates the graph of =
p
() and
=
p
(), respectively; and Figure 8 (middle) indicates the polar coordinate of
each :
(
=
_
(n())n(); S
1
_
.
Incidentally, to observe the characteristic of , the following Frank diagram
(Figure 8 (right)) is more useful than a graph of and (
:
T
=
_
n()
(n())
; S
1
_
=
_
x R
2
; (x) = 1
_
.
From this denition T
= (
1/
holds.
Exercise 1.25. Show the following three proposition are equivalent, if is a
PPHD1 function. (The answer is in [KG01, Appendix 2].)
(i) We denote the set enclosed by T
by

T
=
_
x R
2
; (x) 1
_
. Then
is convex.
(ii) The function satises ((1 )x + y) (1 )(x) + (y) for
x, y R
2
and [0, 1]. (That is, is convex.)
Figure 8. In each gure, the solid line is
p
(), and the dotted
line is
() ( = 15), respectively. M = 6 in any case. (Left):

graphs of
p
() and
(). The horizontal axis is 0 2, and

the vertical axis is magnied for easy to observe. At each point
= 2k/M (k Z), the function
p
() (resp.
()) achieves the

minimum value 1 (resp. 1 +1/). (Middle): (
p
and (
. (Right):
the Frank diagram T
p
and T
. T
p
is a regular M-polygon.
(iii) The function satises (x + y) (x) + (y) for x, y R
2
. (That is,
is subadditive.)
The interfacial energy can be classied by the Frank diagram T
. Since the
sign of the curvature of the curve T
agrees with the sign of +

2
, then as
mentioned above, the initial value problem of (1.4) can be caught in the framework
of parabolic partial dierential equations. The function =
is such an example.
However, in the case where =
p
, its Frank diagram T
p
is a regular M-polygon,
and on the each edge
p
+
2

p
= 0, and
p
may not be dierentiable at each
vertex. To treat such interfacial energy, crystalline curvature ow in the title will
enter the stage. It will be mentioned later. Before it, let us dene another shape
of characterizing .
The Wul shape. The following shape is called the Wul shape:
S
1
_
x R
2
; x n() (n())
_
.
Figure 9 indicates the Wul shape of
p
and
, respectively, in the case where

M = 4 and M = 6.
The Wul shape is a convex set by means of constructions. In particular, if
is smooth and T
is strictly convex, then
is also a strictly convex set with

a smooth boundary. In this case, the distance from the origin (which is inside of
) to L
equals to (n()) = (), where L
is the tangent line which is passing

through the point on
with its outward normal vector being n(). From this

fact, we note that the curvature of
is given by ( +
2
)
1
.
Exercise 1.26. Show that the curvature of
is ( +
2
)
1
if
is a
strictly convex set with a smooth boundary.
In the physical context,
describes the equilibrium of crystal, i.e., in the

plane case,
is the answer of the following Wul problemon the equilibrium

shape of crystals (by Gibbs (1878), Curie (1885), and Wul (1901)):
6. THE FRANK DIAGRAM AND THE WULFF SHAPE 185
Figure 9. Comparison of
p
() and
() ( = 15) by the Wul

shape. (Left, middle left): The Wul shapes
p
and
with
M = 4. (Middle right, right): The Wul shapes
p
and
with M = 6. Two
p
s are regular M-polygons, and two
s
are round regular M-polygons, respectively.
What is the shape which has the least total interfacial energy
E
of the curve for the xed enclosed area?

This suggests that the asymptotic shape of a solution curve of the gradient ow
of E
(1.4) is
in the case =
(Figure 7). However, in the mathematical

context, there are not so many known results.
Open problems. Convexied phenomena or formation of convexity is one of
the interesting problems. In the isotropic case 1, Grayson [Gra87] proved
that convexity is formed (see Figure 2). However in the anisotropic case, even if
is smooth, the formation of convexity has not been completed, except in the
symmetric case ( +) = () by Chou and Zhu [CZ99].
Incidentally, the Frank diagram T
has been considered after the Wul shape
has appeared (Frank (1963) and Meijering (1963)). See the book on crystal
growth in detail. We also refer the reader to Kobayashi and Giga [KG01] for a lot
of examples of (
, T
and
, and for discussion of anisotropy and the eect of

curvature.
Numerical examples. Now, going back the story, we observe the behavior of
a solution curve y = u(x, t) by (1.4) with =
( = 1000) from a view point of

an approximation of =
p
.
Since the total interfacial energy is given by
E
=
_
1
0
W(
x
u) dx, W() = (())
_
1 +
2
, () = arctan
1
,
we obtain the partial dierential equation
t
u =
_
1 + (
x
u)
2
x
(
W(
x
u)) =
+
2

1 + (
x
u)
2
2
x
u,
which is equivalent to (1.4).
Figure 10 indicates a numerical example of the case where M = 4 and = 1000.
As time goes by, one can observe that line segment parts satisfying
x
u = 0 expands,
and the other parts (which are put between two parallel line segments satisfying
x
u = 0) do not move, and will disappear in nite time. Actually, since
+
2

takes a huge value only at = k/2 (k Z) and almost zero at other points, the
movement of points satisfying
x
u 0 can be observed and other points do not
seem to move.
equation
t
u =
_
1 + (
x
u)
2
x
(
W(
x
u)) (0 < x < 1) (from left
to right). Here W() =
(())
_
1 +
2
( = 1000, M = 4),
() = /2+arctan. The boundary conditions are u(0) = u(1) =
0, and the initial curve is the same as the one in Figure 4.
Figure 11 indicates a numerical example of the case where M = 6 and = 1000.
In this case,
is almost a regular hexagon (it almost agrees with

p
in Figure
9 (middle right), and in the set of normal angles of
one can observe the

movement of points on the graph satisfying that their normal angles almost agree
with = k/3 (k = 1, 2), and other points do not seem to move. Since the end
points are xed, Figure 11 (far right) is the nal shape.
equation
t
u =
_
1 + (
x
u)
2
x
(
W(
x
u)) (0 < x < 1) (from left
to right). Here W() = f
(())
_
1 +
2
( = 1000, M = 6),
() = /2+arctan. The boundary conditions are u(0) = u(1) =
0, and the initial curve is the same as the one in Figure 4.
When tends to innity, how does the equation
t
u =
_
1 + (
x
u)
2
x
(
W(
x
u))
change? In the case where M = 4,
p
() = [ cos [ + [ sin [ holds, and then we
have W() = [[ + 1. From this we obtain formally
2
W() = 2(), i.e., the

partial dierential equation becomes
t
u = 2
_
1 + (
x
u)
2
(
x
u)
2
x
u, where is
Dirac delta function. As Figure 10 suggested, the right-hand-side is considerable
as a nonlocal quantity [Gig97].
By the observation of Figure 10 and Figure 11, it seems that there is no problem
if curves are restricted to the following special class of polygonal curves: the set of
normal vectors of the curves agrees with the one of the Wul shape.
7. CRYSTALLINE ENERGY 187
7. Crystalline energy
If the Frank diagram T
is a convex polygon, is called crystalline en-

ergy. When T
is a J-sided convex polygon, there exists a set of angles

i
[
1
<
2
< <
J
<
1
+ 2 such that the position vectors of vertices are labeled
n(
i
)/(n(
i
)) in an anticlockwise order (
J+1
=
1
,
0
=
J
) and
i+1

i

(0, ) holds for i = 1, 2, . . . , J:
T
=
J
_
i=1
_

i
(
i
)
,

i+1
(
i+1
)
_
.
Here and hereafter, we denote
i
= n(
i
) (i). See Figure 12 (left). In this case,
the Wul shape is also a J-sided convex polygon with the outward normal vector
of the i-th edge being
i
:
=
J
i=1
_
x R
2
; x
i
(
i
)
_
.
See Figure 12 (right). We dene a set of normal vectors of
by ^
=
1
,
2
, . . . ,
J
.
5

6
Figure 12. Example of the Frank diagram (left) and the Wul
shape (right) in the case where is a crystalline energy and J = 6.
This Frank diagram is the same regular hexagon T
p
as in Figure
8 (right), and this Wul shape is the same regular hexagon
p
as in Figure 9 (middle right).
Remark 1.27. A point on the (i 1)-th edge of J-sided polygon T
can be
described as
()
= (1 )

i1
(
i1
)
+

i
(
i
)
for some (0, 1), where = n() with some (
i1
,
i
); or one can describe
without using :
=
(1 )
i1
+
i
[(1 )
i1
+
i
[
, =
(
i1
)
(1 )(
i
) +(
i1
)
.
Exercise 1.28. Show this expression of .
Let L
i
be the straight line passing through (
i
)
i
in the direction
i
for all
i, and let y
i
be the intersection point of L
i1
and L
i
:
y
i
= (
i
)
i
l
i
, l
=
(
i1
) (
i1

i
)(
i
)
i1

i
.
It is easy to check that the following claim.
Proposition 1.29. y
i
= () holds for each i.
Proposition 1.29 asserts that the straight line passing through () in the
direction
passes through y
i
.
We have
l
=
1 a
, a
= (
i1

i
)
(
i
)
(
i1
)
, b
i1

i
(
i1
)
.
In the same way, we have
y
i+1
= (
i
)
i
+l
+
i
, l
+
=
(
i+1
) (
i+1

i
)(
i
)
i+1

i
=
1 a
+
b
+
,
where
a
+
= (
i+1

i
)
(
i
)
(
i+1
)
, b
+
=

i+1

i
(
i+1
)
.
Since the angle between
i1
and
i
satises
i1

i
(0, ), b
> 0 holds, and

also b
+
> 0 holds by the similar reason.
Proposition 1.31. (y
i+1
y
i
)
i
= l
+l
+
> 0 holds.
Since T
is a J-sided convex polygon, there exists (0, 1) such that for each i
= (1 )

i1
(
i1
)
+

i+1
(
i+1
)
holds, and that

i
<
1
(
i
)
,
i
= 0
hold. From the inequality and equation, and from the facts b
> 0 and b
+
> 0, we
have
=
b
+b
+
, a
(a
a
+
) =
a
b
+
+a
+
b
+b
+
< 1.
Hence it holds that
l
+l
+
=
1
b
b
+
(b
+b
+
(a
b
+
+a
+
b
)) > 0.
Therefore
is J-sided convex polygon with the i-th vertex being y

i
and the the
length of the i-th edge being l
+l
+
> 0 for each i.

Exercise 1.32. Show that if T
is the regular J-sided polygon,
is also the
regular J-sided polygon.
Remark 1.33. There exist two crystalline energies
1
and
2
such that
1
agrees with
2
except the position of the center. See Figure 13.
Remark 1.34. As we saw the above, if the Frank diagram is a J-sided convex
polygon, then the Wul shape is also the J-sided convex polygon. But the converse
is not true. See Figure 14.
Remark 1.35. Remark 1.33 and Remark 1.34 suggest that
does not have

all the information of . Actually,

=
holds, where is convexication of .

See [KG01].
8. CRYSTALLINE CURVATURE FLOW EQUATIONS 189
T
=
1
:
=
2
:
Figure 13. Example of the same shape of the Wul shape with
dierent :
1
,=
2
.
8. Crystalline curvature ow equations
We will dene crystalline curvature ow equations.
Polygonal curves. Let T be a simple closed N-sided polygonal curve in the
plane R
2
, and label the position vector of vertices p
i
(i = 1, 2, . . . , N) in an anti-
clockwise order:
T =
N
_
i=1
o
i
,
where o
i
= [p
i
, p
i+1
] is the i-th edge (p
N+1
= p
1
, p
0
= p
N
). The length of o
i
is d
i
= [p
i+1
p
i
[, and then the i-th unit tangent vector is t
i
= (p
i+1
p
i
)/d
i
and the i-th unit outward normal vector is n
i
= t
i
. We dene a set of normal
vectors of T by ^ = n
1
, n
2
, . . . , n
N
. Let
i
be the exterior normal angle of
o
i
. Then n
i
= n(
i
) and t
i
= t(
i
) hold. We dene the i-th hight function
h
i
= p
i
n
i
= p
i+1
n
i
. See Figure 15. By using N-tuple h = (h
1
, h
2
, . . . , h
N
), d
i
is described as d
i
= D
i
(h), where
D
i
(h) =

i1,i
_
1 (n
i1
n
i
)
2
(h
i1
(n
i1
n
i
)h
i
)
+

i,i+1
_
1 (n
i
n
i+1
)
2
(h
i+1
(n
i
n
i+1
)h
i
),
T
=
1
:
=
2
:
Figure 14. The Wul shapes are both regular hexagons, while
the corresponding Frank diagrams are not polygons with dierent
:
1
,=
2
.
where
i,j
= sgn(det(n
i
, n
j
)) for i = 1, 2, . . . , N (h
N+1
= h
1
, h
0
= h
N
). Since
n
i
n
j
= cos(
i

j
), we have another expression:
D
i
(h) = (cot
i
+ cot
i+1
)h
i
+h
i1
cosec
i
+h
i+1
cosec
i+1
, (1.5)
where
i
=
i

i1
for i = 1, 2, . . . , N. Note that 0 < [
i
[ < holds for all i.
Furthermore, the i-th vertex p
i
(i = 1, 2, . . . , N) is described as follows:
p
i
= h
i
n
i
+
h
i1
(n
i1
n
i
)h
i
n
i1
t
i
t
i
. (1.6)
Exercise 1.36. From the relation n
T
i1
p
i
= h
i1
and n
T
i
p
i
= h
i
, we obtain
p
i
=
_
n
T
i1
n
T
i
_
1
_
h
i1
h
i
_
.
From this, show (1.6). This idea can be found in [BKY08].
Exercise 1.37. From (1.6) and the relation t
i
n
i+1
= sin
i+1
, show (1.5).
Exercise 1.38. From (1.6), check that d
i
= [p
i+1
p
i
[ and d
i
= (p
i+1
p
i
) t
i
hold.
Exercise 1.39. Fix any a R
2
. By k
i
= a n
i
we denote distance from the
origin to a in the direction n
i
, and by k = (k
1
, k
2
, . . . , k
N
) we denote their N-tuple
(k
N+1
= k
1
, k
0
= k
N
). Show that D
i
(k) = 0 holds for i = 1, 2, . . . , N. From this
pi+1
Si
pi
i j
nj
tj
hj
dj
Figure 15. Example of polygonal curves (N = 56).
fact, D
i
(h +k) = D
i
(h) +D
i
(k) = d
i
holds for all i. This means that the shape of
curve does not change for a shift of the origin.
Admissible curves. Following [HGGD05], we call curve T essentially ad-
missible if and only if the consecutive outward unit normal vectors n
i
, n
i+1
^
(n
N+1
= n
1
, n
0
= n
N
) satisfy
n
i+
=
(1 )n
i
+n
i+1
[(1 )n
i
+n
i+1
[
, ^
for (0, 1) and i = 1, 2, . . . , N. See Figure 16 (middle).

Exercise 1.40. If T is an essentially admissible curve, then ^ ^
holds.
But the converse is not true. Make a counterexample.
Remark 1.41. In the case where T is a convex polygon, T is an essentially
admissible if and only if ^ ^
holds.
Exercise 1.42. Check this remark.
We call curve T admissible if and only if T is an essentially admissible curve
and ^ = ^
(especially, ^ ^
) holds. In other words, T is admissible if and only

if ^ = ^
holds and any adjacent two normal vectors in the set ^
, for example
i
and
i+1
are also adjacent in the set ^, i.e.,
i
,
i+1
= n
j
, n
j+1
^ holds
for some i and j (
J+1
=
1
, n
N+1
= n
1
). See Figure 16 (left).
Gradient ow. We consider a moving essentially admissible and an N-sided
curve T(t) with the time t in some interval and with the N-tuple of hight func-
tions h(t) = (h
1
(t), h
2
(t), . . . , h
N
(t)). Then we have d
i
(t) = D
i
(h(t)). The total
n
1
(a) Admissible polygonal curve
(N = 10).
n
1
n
3
3
n
9
5
n
12
6
(b) Essentially admissible polygo-
nal curve (N = 13). Note that
N {n
3
, n
9
, n
12
} = N holds.
n
1
2
n
3
(
3
)
2
n
6
(
3
)
4
n
9
5
(
6
)
1
(c) This polygonal curve (N = 9) is
not admissible, nor essentially admis-
sible, since N N (
6
N) holds,
and in addition there exist ,
(0, 1) such that n

3+
= n
6+
=
3
, n
9+
=
6
N holds.
Figure 16. Three typical examples of polygonal curves for the
Wul shape in Figure 12 (right).
crystalline energy on T(t) is dened as E
(T(t)) =
N
i=1
(n
i
)d
i
(t), and the rate
of change of it is given as
t
E
(T(t)) =
N
i=1
(n
i
)
t
D
i
(h) =
N
i=1
(n
i
)D
i
(V ),
where V = (V
1
, V
2
, . . . , V
N
) and V
i
=
t
h
i
(t) is the velocity on o
i
in the direction of
n
i
for i = 1, 2, . . . , N. Here we have used the relation
t
d
i
(t) = D
i
(V ) from (1.5),
which is equivalent to the time derivative of (1.6):
t
p
i
= V
i
n
i
+v
i
t
i
, v
i
=
V
i1
(n
i1
n
i
)V
i
n
i1
t
i
, i = 1, 2, . . . , N. (1.7)
We denote = ((n
1
), (n
2
), . . . , (n
N
)), and then we have
N
i=1
(n
i
)D
i
(V ) =
N
i=1
V
i
D
i
() =
N
i=1
D
i
()
d
i
V
i
d
i
.
Hence we have
t
E
(T(t)) =
N
i=1
(n
i
)V
i
d
i
,
(n
i
) =
D
i
()
d
i
, i = 1, 2, . . . , N.
Here
(n
i
) is called crystalline curvature on the i-th edge o
i
, which is the
crystalline version of weighted curvature derived from
t
E
((t)) =
_
(n)V ds.
For two N-tuples F = (F
1
, F
2
, . . . , F
N
), G = (G
1
, G
2
, . . . , G
N
) R
N
, let us dene
the inner product on T as follows:
F, G)
P
=
N
i=1
F
i
G
i
d
i
.
Therefore by analogue of gradient ow of E
((t)), we have the gradient ow of

E
(T(t)) such as
V
i
=
(n
i
), i = 1, 2, . . . , N.
This is called crystalline curvature ow equation.
The numerator of crystalline curvature
(n
i
) is described as
D
i
() =
i
l
(n
i
), (1.8)
where
i
= (
i1,i
+
i,i+1
)/2 takes +1 (resp. 1) if T is convex (resp. concave)
around o
i
in the direction of n
i
, otherwise
i
= 0; and l
(n) is the length of the

j-th edge of
if n =
j
for some j, otherwise l
(n) = 0.
Remark 1.43. Equation (1.8) can be derived as follows. We dene
l
, ) =
(n(
)) (n() n(
))(n())
n(
) n()
=
(n(
)) (n()) cos(
)
sin(
)
,
and l
+
(,
) = l
, ). Then we have
l
(
j
) = l
(
j1
,
j
) +l
+
(
j
,
j+1
), D
i
() = l
(
i1
,
i
) +l
+
(
i
,
i+1
).
By geometric consideration, we have
l
(,
j
) =
_
l
(
j1
,
j
), [
j1
,
j
),
l
+
(
j
, ) = l
+
(
j
,
j+1
), (
j
,
j+1
],
and
l
+
(
j
, ) =
_
l
+
(
j
,
j+1
), (
j
,
j+1
],
l
(,
j
) = l
(
j1
,
j
), [
j1
,
j
).
The two cases are possible. The rst case is n
i
^
, i.e., there exists j such that

n
i
=
j
and
i
=
j
; and we have four subcases as follows:
(i)
i1
[
j1
,
j
) and
i+1
(
j
,
j+1
],
(ii)
i1
(
j
,
j+1
] and
i+1
[
j1
,
j
),
(iii)
i1
,
i+1
(
j
,
j+1
], and
(iv)
i1
,
i+1
[
j1
,
j
).
Hence, D
i
() is equal to
(i) l
(
j
),
(ii) l
(
j
),
(iii) 0, and
(iv) 0, respectively.
The second case is n
i
, ^
, i.e., there exists j such that

i
(
j1
,
j
); and we
have four subcases similarly as above. In this case, we regard that there is a zero-
length edge of
between the (j 1)-th and the j-th edges, and for all subcases
D
i
() = 0 holds. Thus we obtain (1.8).
Exercise 1.44. Follow this remark.
a
1
b
5

6
n
12
(a) a is the length of edge
l(n
1
) > 0 and b is the
length of degenerate edge
l(n
12
) = 0.
d
1
n
1
d
12
n
12
(b) d
1
(resp. d
12
) is the length of
the edge whose normal vector is n
1
(resp. n
12
). Hence the crystalline
curvature is (n
1
) = l(n
1
)/d
1
(resp. (n
12
) = 0).
Figure 17. The Wul shape (left) is the same hexagon as in Fig-
ure 12 (right). Essentially admissible polygonal curve (right) is the
same curve as in Figure 16(b).
Remark 1.45. If T is an admissible and convex polygon, then n
i
=
i
and
i
= 1 for all i = 1, 2, . . . , N = J; and moreover, if T =
, then the crystalline

curvature is 1.
Note. To handle the gradient ow of E
(1.4) in the case where is crys-

talline, we have restricted smooth curve to admissible curve and have introduced
the crystalline curvature dened on each edge. This strategy was proposed by Tay-
lor [Tay90, Tay91a, Tay91b, Tay93] and independently by Angenent and Gurtin
[AG89]. Also one can nd essentially the same method as a numerical scheme for
9. AN AREA-PRESERVING MOTION BY CRYSTALLINE CURVATURE 195
curvature ow equation in Roberts [Rob93]. We refer the reader Almgren and Tay-
lor [AT95] for detailed history. In the physical context, the region enclosed by T
represents the crystal. See also Taylor, Cahn and Handwerker [TCH92] and Gurtin
[Gur93] for physical background. Besides this crystalline strategy, other strategies
by subdierential and level-set method have been extensively studied. See Giga
[Gig97, Gig98, Gig00] and references therein.
9. An area-preserving motion by crystalline curvature
The enclosed area A(t) of T(t) is given by
A =
1
2
N
i=1
h
i
d
i
,
and its rate of change is
t
A(t) =
1
2
N
i=1
V
i
d
i
+
1
2
N
i=1
h
i
D
i
(V ) =
N
i=1
V
i
d
i
.
Then the gradient ow of E
along T which encloses a xed area is

V
i
=
(n
i
), i = 1, 2, . . . , N, (1.9)
where
N
i=1

(n
i
)d
i
N
k=1
d
k
=
N
i=1
D
i
()
L
=
N
i=1
i
l
(n
i
)
L
is the average of the crystalline curvature. The gradient ow (1.9) is in the metric
, )
P
under the constraint for V : 1, V )
P
= 0, 1 = (1, 1, . . . , 1).
Exercise 1.46. Check the following two basic properties:
t
E
(T(t)) 0 and
t
A(t) = 0.
Problem 1. For a given N-sided essentially admissible closed curve T
0
, nd a
family of N-sided essentially admissible curves
0t<T
T(t) satisfying
_
V
i
(t) =
(n
i
), 0 t < T, i = 1, 2, . . . , N,
T(0) = T
0
.
Remark 1.47. Problem 1 is equivalent to
t
d
i
(t) = D
i
(V ), or equivalent to
(1.7); and since it is described as the N-system of ordinary dierential equations,
the maximal existence time is positive: T > 0.
Remark 1.48. Problem 1 is closely related to general area-preserving motion
of polygonal curves. See [BKY07, BKY08].
Known results. What might happen to T(t) as t tends to T ? For this
question, we have the following three results. The rst result is the case where
motion is isotropic and polygon is admissible.
Theorem 1.49. Let the interfacial energy be isotropic 1. Assume the initial
polygon T
0
is an N-sided admissible convex polygon. Then a solution admissible
polygon T(t) exists globally in time keeping the area enclosed by the polygon constant
A, and T(t) converges to the shape of the boundary of the Wul shape
in the
Hausdor metric as t tends to innity, where
=
_
2A/
N
k=1
l
1
(n
k
) is constant.
In particular, if T
0
is centrally symmetric with respect to the origin, then we have
an exponential rate of convergence.
This theorem is proved by Yazaki [Yaz02] by using the isoperimetric inequality
and the theory of dynamical systems. We note that
is the circumscribed
polygon of a circle with radius
, and then this result is a semidiscrete version of

Gage [Gag86].
The second result is the case where motion is anisotropic and polygon is ad-
missible.
Theorem 1.50. Let the crystalline energy be > 0. Assume the initial polygon
T
0
is an N-sided admissible convex polygon. Then a solution admissible polygon
T(t) exists globally in time keeping the area enclosed by the polygon constant A,
and T(t) converges to the shape of the boundary of the Wul shape
in the
Hausdor metric as t tends to innity, where
(n
i
) = (n
i
)/w
, w
=
_
[
[/A
for i = 1, 2, . . . , N and [
[ =
N
k=1
(n
k
)l
(n
k
)/2 is enclosed area of
.
This theorem is proved in Yazaki [Yaz04, Part I] by using the anisoperimet-
ric inequality or Br unn and Minkowskis inequality and the theory of dynamical
systems. For readers convenience, the proof will be shown in the next section.
The last result is the case where motion is anisotropic and polygon is essentially
admissible.
Theorem 1.51. Let the crystalline energy be > 0. Assume the initial polygon
T
0
is an N-sided essentially admissible convex polygon. If the maximal existence
time of a solution essentially admissible polygon T(t) is nite T < , then there
exists the i-th edge o
i
such that lim
tT
d
i
(t) = 0 and l
(n
i
) = 0 hold. That is,
the normal vector of vanishing edge does not belong to ^
, and inf
0<t<T
d
k
(t) > 0
holds for all n
k
^
.
This theorem is proved in [Yaz07b].
Open problems. For any essentially admissible convex polygon T
0
, is T a
nite value? This is still open. If the answer of this question is yes, then we
have the nite time sequence T
1
< T
2
< < T
M
such that T(T
i
) is essentially
admissible for i = 1, 2, . . . , M 1 and T(T
M
) is admissible. In the general case
where V
i
= F(n
i
,
(n
i
)) for all i under certain conditions of F, the answer of the
above question is yes. See [Yaz07a]. However, F does not include
.
Open problems in the case where P
0
is nonconvex. Even if T
0
is ad-
missible, we have no theoretical results at this stage. See section 11 in detail.
10. Scenario of the proof of Theorem 1.50
Put the crystalline curvature such as
w
i
(t) =
(n
i
) =
l
(n
i
)
d
i
, i = 1, 2, . . . , N.
Then we have
t
w
i
(t) = l
1
i
w
i
(t)
2
D
i
(V ), D
i
(V ) = D
i
(1)
D
i
(w),
10. SCENARIO OF THE PROOF OF THEOREM 1.50 197
where l
i
= l
(n
i
), 1 = (1, 1, . . . , 1) and w = (w
1
, w
2
, . . . , w
N
). Note that
D
i
(1) = tan

i
2
+ tan

i+1
2
holds for i = 1, 2, . . . , N. Hence from the average
N
k=1
l
k
/L and the length
L =
N
j=1
l
j
w
j
(t)
1
, we can restate Problem 1 as the following Problem 2.
Problem 2. Find a function w(t) = (w
1
(t), w
2
(t), . . . , w
N
(t)) (C[0, T)
C
1
(0, T))
N
satisfying
t
w
i
(t) = l
1
i
D
i
(w(t))w
i
(t)
2
l
1
i
D
i
(1)
N
k=1
l
k
N
j=1
l
j
w
j
(t)
1
w
i
(t)
2
,
i = 1, 2, . . . , N, 0 < t < T
and
w
i
(0) = w
0
i
, i = 1, 2, . . . , N,
w
N+1
(t) = w
1
(t), w
0
(t) = w
N
(t), 0 < t < T,
where w
0
i
is the i-th initial crystalline curvature of T
0
.
Problem 1 and Problem 2 are equivalent except the indeniteness of position
of a solution polygon. See [Yaz02, Remark 2.1]. Since Problem 2 is the initial
value problem of ordinary dierential equations, there exists a unique time local
solution. Moreover, by using a similar argument as in Taylor [Tay93, Proposition
3.1], Ishii and Soner [IS99, Lemma 3.4], Yazaki [Yaz02, Lemma 3.1], we obtain the
time global solvability.
Lemma 1.52 (time global existence). A solution w of Problem 2 and a solution
polygon of Problem 1 exist globally in time, i.e., a solution polygon does not develop
singularities in a nite time.
In the following, we will show three lemmas, which play an important role in a
scenario of the proof of Theorem 1.50.
Let the anisoperimetric ratio be
(T) =
E
(T)
2
4[
[A
for an N-sided admissible convex polygon T associated with
. Here A (resp.
[
[) is the enclosed area of T (resp. the area of
). The rst key lemma is the

anisotropic version of the isoperimetric inequality.
Lemma 1.53 (anisoperimetric inequality). For a polygon T associated with
,
the anisoperimetric inequality
(T) 1
holds. The equality
(T) 1 holds if and only if w

i
=
(n
i
) const. for all i,
i.e., (the enclosed region of T) satises = k
for some constant k > 0.

We will prove this lemma in Appendix 3 using the mixed area and the Br unn
and Minkowskis inequality.
From this lemma and lemma 1.52, for a solution polygon T(t),
(T(t)) =
E
(T(t))
2
4[
[A
1
holds for t 0. Moreover, we have the following second key lemma.
Lemma 1.54. lim
t
E
(T(t)) = 2
_
[
[A and lim
t
(T(t)) = 1 hold.
Proof of this lemma closely follows [Yaz02, Lemma 5.6].
From this lemma, if a solution polygon T(t) is N-sided polygon at the time
innity, then (t) approaches to k
(k > 0) as t tends to innity. There-

fore, to complete the proof of Theorem 1.50, it is required to show the estimate
inf
0<t<
min
1iN
d
i
(t) > 0. As a matter of fact, the following the third key
lemma holds. This lemma is a strong assertion compared with the above estimate.
Lemma 1.55. Let w
be the equilibrium point of the rst evolution equation in

Problem 2. Then w
=
_
[
[/A holds. Moreover, the equilibrium point w
is
asymptotically stable and
lim
t
w
i
(t) = w
holds for i = 1, 2, . . . , N.
One can prove this lemma by the general theory of dynamical systems or the
Lyapunov theorem (see [Yaz02, Lemma 5.8]).
Proof of Theorem 1.50. From Lemma 1.55, we have
lim
t
d
i
(t) =
l
i
w
, i = 1, 2, . . . , N.
From this limit and the theory of the generalized eigenvalue space, there exists a
vector c(t) R
2
such that h
i
(t) c(t) n
i
converges to
(n
i
) = (n
i
)/w
for
all i as t tends to innity. Hence for any > 0 there exists t
> 0 such that

T(t)
(1+)
(1)
holds for t t
. Then the assertion holds.

11. Numerical scheme
The aims are to construct a numerical scheme which enjoys two basic properties
t
E
0 and
t
A = 0 (see Exercise 1.46), and to investigate what might happen to
solution T(t) of the evolution equation (1.9) as t tends to T . We discretize the
system of ordinary equations
t
d
i
(t) = D
i
(V ) or (1.7) with the initial essentially
admissible closed curve T
0
= T
0
at the time t
0
= 0. Let m = 0, 1, 2, . . . be a step
number. By a
m
, we denote the approximation of a(t
m
) at the time t
m
=
m1
i=0

i
(the time step
m
will be dened in the following procedure).
The following procedure is an extension of [UY04].
Procedure 1. Fix parameters [0, 1] and , (0, 1). For a given essen-
tially admissible N-sided closed curve T
m
=
N
i=1
[p
m
i
, p
m
i+1
], we dene T
m+1
=
N
i=1
[p
m+1
i
, p
m+1
i+1
] as follows:
(i) the i-th length: d
m
i
= [p
m
i+1
p
m
i
[ (i);
(ii) the m-th variable time step:
m
= (d
m
min
)
2
/, where
= (1) min, 1, = 2[l
(n)[
max
(2/[ sin[
min
+[ tan(/2)[
max
);
11. NUMERICAL SCHEME 199
(iii) the i-th length d
m+1
i
:
(D
d)
m
i
= (cot
i
+ cot
i+1
)V
m+
i
+ V
m+
i1
cosec
i
+ V
m+
i+1
cosec
i+1
(i),
V
m+
i
=
P
N
j=1
jl(nj)
P
N
k=1
d
m+
k
il(ni)
d
m+
i
, d
m+
i
= (1 )d
m
i
+d
m+1
i
;
(iv) the i-th hight h
m+1
i
: (D
h)
m
i
= V
m+
i
(i);
(v) the i-th vertex: p
m+1
i
= h
m+1
i
n
i
+
h
m+1
i1
(ni1ni)h
m+1
i
ni1ti
t
i
(i);
(vi) the new time: t
m+1
= t
m
+
m
.
Here we have used the notation: a
min
= min
1iN
a
i
, [a[
min
= min
1iN
[a
i
[,
[a[
max
= max
1iN
[a
i
[, (D
a)
m
i
= (a
m+1
i
a
m
i
)/
m
, and the periodicity F
N+1
= F
1
,
F
0
= F
N
.
Two basic properties. One is that the total energy E
m
is decreasing in steps:
(D
)
m
0 for any [0, 1]. The other is that the enclosed area A
m
of T
m
is
preserved: (D
A)
m
= 0 if = 1/2.
Iteration. In (3), if (0, 1], we solve the following iteration starting from
z
0
i
= d
m
i
:
z
k+1
i
z
0
i
m
= (cot
i
+ cot
i+1
)
V
m+
i
+

V
m+
i1
cosec
i
+

V
m+
i+1
cosec
i+1
,
V
m+
i
=
N
j=1
j
l
(n
j
)
N
k=1
d
m+
k

i
l
(n
i
)
d
m+
i
,

d
m+
i
= (1 )z
0
i
+z
k
i
, k = 0, 1, . . .
Convergence lim
k
z
k
i
= d
m+1
i
and positivity d
m+1
i
(1 )d
m
i
> 0 hold for all i
([UY04]).
The maximal existence time. Since the above positivity d
m
i
> 0 holds, we
can keep iterating Procedure 1 in nitely many steps (even if T
m
self-intersects at a
step m, we can continue). Then the maximal existence time is t
= lim
m
m
k=0
k
.
Here we have two questions: one is whether t
is nite or not, and the other is what

might happen to T
m
as m tends to innity. It is known that t
= holds if W
is an N-sided regular polygon and T

0
is an admissible convex polygon ([UY04]).
Extension. At the maximal existence time t
, it is possible that at least one

edge, for instance the i-th edge, may disappear. If
i
= 0 or l
(n
i
) = 0, then T
is still essentially admissible. Hence we can continue Procedure 1 starting from the
initial curve T
. In practice, if some edges are small enough, we eliminate them

articially as the following procedure.
Procedure 2. Put a positive parameter 1. For every step m, do the
followings:
(i) dene D = mind
m
i
;
i
,= 0, l
(n
i
) > 0 (this is well-dened);
(ii) nd the value k such that d
m
k
= mind
m
i
;
i
= 0 (if it exists);
nd the value j such that d
m
j
= mind
m
i
;
i
,= 0, l
(n
i
) = 0 (if it exists);
(iii) if k or j exists, check the followings:
(a) if d
m
k
/D < and if d
m
k
d
m
j
or the value j does not exist, then
eliminate the k-th edge (see Figure 18 (left));
(b) if d
m
j
/D < and if d
m
j
< d
m
k
or the value k does not exist, then
eliminate the j-th edge (see Figure 18 (right));
(c) otherwise, exit from Procedure 2;
(iv) in (3), if (a) occurred, then do renumbering within the new number N :=
N 2;
else if (b) occurred, then do renumbering within the new number N :=
N 1.
Numerical computation will be continued repeating Procedure 1 and 2.
k
(a) In the type of (3) (a), the
number of particles decreases
from N to N 2.
j
(b) In the type of (3) (b), the
number of particles decreases
from N to N 1.
Figure 18. Two types of elimination of some edges in Procedure 2 (3).
Numerical simulations. In the following 8 gures on each line, from left to
right, they indicate
, T
0
, T
m1
, . . . , T
m6
(0 < m
1
< < m
6
).
The case where P
0
is convex and admissible. Figure 19 and Figure 20
indicate that asymptotic behavior of solution polygon which does not break the
results of Theorem 1.49 and Theorem 1.50. See also Gage [Gag86] for the smooth
case. On the convergence between (t
m
) and T(t
m
), see Ushijima and Yazaki
[UY04].
Figure 19. A numerical example of Theorem 1.49: Convergence
of a hexagon to the regular hexagon
.
Figure 20. A numerical example of Theorem 1.50: Even when
the initial polygon is quite dierent shape from
, it converges
to
.
The case where P
0
is convex and essentially admissible. Known result is
only Theorem 1.51. Figure 21 suggests possibility of positive answer of the question
in open problem just after Theorem 1.51.
11. NUMERICAL SCHEME 201
Figure 21. A numerical example of Theorem 1.51: An essentially
admissible convex polygon converges to
in a nite time.
Figure 22. A numerical example of convexied phenomena.
The case where P
0
is nonconvex and admissible. Figure 22 suggests that
convexied phenomena holds.
An example of self-intersection. In smooth case, a self-intersection is con-
jectured in Gage [Gag86], and is proved in Mayer and Simonett [MS00]. Figure 23
suggests that self-intersection is possible to occur.
Figure 23. A numerical example of self-intersection. The motion
is delicate. See the symbolic motion in Figure 24.
Open problems. At this stage, we have the following three open problems:
(i) Does T(t) become convex in nite time?
(ii) Will the admissibility be preserved?
(iii) Does T(t) self-intersect?
We have information related to these three questions. (1) In the case where
V
i
= (n
i
)[
(n
i
)[
1
(n
i
), there exist (0, 1), and T
0
such that non-
convex solution curve T(t) shrinks homothetically, i.e., there exists a nonconvex
self-similar solution polygonal curve. See Ishiwata, Ushijima, Yagisita and Yazaki
[IUYY04]. (2) In the case where V
i
= a(n
i
)[
(n
i
)[
1
(n
i
), for any 1, a()
and T
0
if is symmetric, then the solution curve keeps admissibility. See Giga and
Giga [GGig00]. However, there exist (0, 1), a() and T
0
such that the admissi-
bility collapses in nite time, i.e., we have the examples that admissible nonconvex
polygonal curve becomes nonadmissible in nite time. See Hirota, Ishiwata and
Yazaki [HIY06, HIY07]. (3) Figure 23 suggests the possibility of self-intersection.
Remark 1.56. Thanks to the tangential velocity v
i
in (1.7), numerical compu-
tation of crystalline curvature ow equations is quite stable. The v
i
corresponds to
p
1
p
2
p
3
p
4
p
5
p
6
p
7
p
8
p
9
p
10
p
11
p
12
Figure 24. Symbolic motion of a numerical example in Figure 23.

The initial 12-sided nonconvex polygonal curve T
0
in Figure 23 is
(far left). Let be the distance between the 6-th and the 12-th
edges (left), which is small enough. Then disappears immedi-
ately, and moreover, becomes negative (middle left), and becomes
positive again (middle right). After that, the vertical two edges,
the thick line segments in (right), disappear. Finally, the horizon-
tal two edges, the thick line segments in (far right), disappear, and
becomes rectangle such as Figure 23 (far right).
v =
s
V/K (this is equivalent to
t
= 0 by Lemma 1.10.) in the continuous case
(see [Yaz07c]), and it is utilized as redistribution of grid points on evolving curve.
See [
SY07,

SY, BKP
STY07].
gures produced by freezing of internal melt gures
This section is based on the work with Tetsuya Ishiwata [IY07, IY08].
When a block of ice is exposed to solar beams or other radiation, internal
melting of ice occurs. That is, internal melting starts from some interior points
of ice without melting the exterior portions, and each water region forms a ower
of six petals, which is called Tyndall gure (see Figure 25). The gure is lled
Figure 25. Tyndall gures seen from the direction of 45
to c-axis
[Nak56, No. 17].
with water except for a vapor bubble (black spot in Figure 25). This phenomenon
was rst observed by Tyndall (1858). When Tyndall gure is refrozen, the vapor
bubble remains in the ice as a hexagonal disk (see Figure 26). This hexagonal
disk is lled with water vapor saturated at that temperature and surrounded by
12. FORMATION OF NEGATIVE ICE CRYSTALS OR VAPOR FIGURES 203
Figure 26. Natural negative crystals in an ice single crystal
[Nak56, No. 1].
ice. McConnel found these disks in the ice of Davos lake [McC89]. Nakaya called
this hexagonal disk vapor gure and investigated its properties precisely [Nak56].
Adams and Lewis (1934) called it negative crystal. (Although Nakaya said this
term does not seem adequate with a certain reason, hereafter we use the term
negative crystal for avoiding confusion.)
Negative crystal is useful to determine the structure and orientation of ice
or solids. Because, within a single ice crystal, all negative crystals are similarly
oriented, that is, corresponding edges of hexagon are parallel each other (see Fig-
ure 27). Furukawa and Kohata made hexagonal prisms experimentally in a single
Figure 27. A cluster of minute negative crystals [Nak56, No. 53].
ice crystal, and investigated the habit change of negative crystals with respect to
the temperatures and the evaporation mechanisms of ice surfaces [FK93].
To the best of authors knowledge, after the Furukawa and Kohatas experimen-
tal research, there have been no published results on negative crystals, and there
are no dynamical model equations describing the process of formation of negative
crystals. In the present talk, we will focus on the process of formation of negative
crystals after Tyndall gures are refrozen, and try to propose a model equation of
interfacial motion which tracks the deformation of negative crystals in time.
Formation of negative crystals. Figure 28 indicates aftereect of freezing
of Tyndall gures from the initial stage of refrozen process to the nal stage of the
formation of negative crystals. The aim of this talk is to propose a model equation,
Figure 28. From left to right, upper to lower: (a) Start of freez-
ing, t = 0min. (b) Freezing proceeds, t = 3min. (c) Freezing pro-
ceeds further, t = 11min. (d) The bubble is separated, t = 17min.
(e) The separated liquid lm migrates, t = 28min. (f) After freez-
ing, cloudy layers and a vapor gure are left, t = 1hr 21min.
[Nak56, No. 52a52f].
revealed the process in Figure 28 from (e) to (f). This process may be described as
the following:
Negative crystal changes the shape from oval to hexagon.
Thus, our model will be assumed that
(i) water vapor region is simply connected and bounded region in the plane
R
2
(we denote it by );
(ii) is surrounded by a single ice crystal (i.e., ice region is R
2
);
12. FORMATION OF NEGATIVE ICE CRYSTALS OR VAPOR FIGURES 205
(iii) moving boundary is interface of the water vapor region and the single
ice crystal; and
(iv) c-axis (main axis) of the single ice crystal is perpendicular to the plane.
The evolution law of moving interface is similar to the growth of snow crystal,
since deformation of negative crystal is regarded as crystal growth in the air. As a
model of snow crystal growth, we refer the Yokoyama-Kuroda model [YK90], which
is based on the diusion process and the surface kinetic process by Burton-Cabrera-
Frank (BCF) theory [BCF51]. Meanwhile, we assume the existence of interfacial
energy (density) on the boundary . The equilibrium shape of negative crystal
is a regular hexagon, and if the region is very close to a regular hexagon, then
the evolution process may be described as a gradient ow of total interfacial energy
subject to a xed enclosed area. Therefore, we can divide the process of formation
of negative crystals into two stages as follows:
(i) in the former stage by the diusion and the surface kinetic, oval changes
to a hexagon; and
(ii) in the latter stage by a gradient ow of interfacial energy, the hexagon
converges to a regular hexagon.
A simple modeling of these stages is proposed as an area-preserving crystalline
curvature ow for negative polygonal curves. See [IY07, IY08].
Bibliography
[AT95] F. Almgren and J. E. Taylor, Flat ow is motion by crystalline curvature
for curves with crystalline energies, J. Di. Geom. 42 (1995) 122.
[AG89] S. Angenent and M. E. Gurtin, Multiphase thermomechanics with interfa-
cial structure, 2. Evolution of an isothermal interface, Arch. Rational Mech.
Anal. 108 (1989) 323391.
[AV95] S. Angenent and J. J. L Velazquez, Asymptotic shape of cusp singularities
in curve shortening, Duke Math. J. 77 (1995) 71110.
[BKP
STY07] M. Benes, M. Kimura, P. Paus, D.

Sevcovic, T. Tsujikawa and
S. Yazaki, Application of a curvature adjusted method in image segmen-
tation, submitted to Bulletin of Inst. of Mathematics, Academia Sinica at
Taipei (arXiv:0712.2334v1).
[BKY07] M. Benes, M. Kimura and S. Yazaki, Analytical and numerical aspects on
motion of polygonal curves with constant area speed, Proceedings of Slovak-
Austrian Mathematical Congress (Podbanske, 2007), within MAGIA 2007,
Dept. of Mathematics and Descriptive Geometry, Faculty of Civil Engineer-
ing, Slovak University of Technology, ISBN 978-80-227-2796-9 (2007) 127141.
(arXiv:0802.2204v3 is a new version including a minor change.)
[BKY08] M. Benes, M. Kimura and S. Yazaki, Second order numerical scheme
for motion of polygonal curves with constant area speed, preprint
(arXiv:0805.4469v1).
[BCF51] W. K. Burton, N. Cabrera and F. C. Frank, The growth of crystals and
the equilibrium structure of their surfaces, Philos. Trans. R. Soc. Landon,
Ser. A 243 (1951) 299358.
[CZ99] K.-S. Chou and X.-P. Zhu, A convexity theorem for a class of anisotropic
ows of plane curves, Indiana Univ. Math. J. 48 (1999) 139154.
[EG87] C. L. Epstein and M. Gage, The curve shortening ow, Wave motion: the-
ory, modelling, and computation (Berkeley, Calif., 1986), Math. Sci. Res. Inst.
Publ., 7, Springer, New York (1987) 1559.
[FK93] Y. Furukawa and S. Kohata, Temperature dependence of the growth form
of negative crystal in an ice single crystal and evaporation kinetics for its
surfaces, J. Crystal Growth 129 (1993) 571581.
[Gag86] M. Gage, On an area-preserving evolution equations for plane curves, Con-
temporary Math. 51 (1986) 5162.
[GH86] M. Gage and R. S. Hamilton, The heat equation shrinking convex plane
curves, J. Di. Geom. 23 (1986) 6996.
[Gig95] Y. Giga, A level set method for surface evolution equations, S ugaku 47
(1995) 321340; English transl., Sugaku Expositions 10 (1997) 217241.
207
208 Bibliography
[Gig97] Y. Giga, Interesting diusion equations, S ugaku 49 (1997) 193196.
(Japanese)
[Gig98] Y. Giga, Interface dynamics eects of curvature, Hokkaido University
Technical Report Series 56 (1998). (Japanese)
[Gig00] Y. Giga, Anisotropic curvature eects in interface dynamics, S ugaku 52
(2000) 113127; English transl., Sugaku Expositions 16 (2003) 135152.
[GGig00] M.-H. Giga and Y. Giga, Crystalline and level set ow convergence of
a crystalline algorithm for a general anisotropic curvature ow in the plane,
Free boundary problems: theory and applications, I (Chiba, 1999), GAKUTO
Internat. Ser. Math. Sci. Appl., Gakkotosho, Tokyo 13 (2000) 6479.
[Gra87] M. A. Grayson, The heat equation shrinks emmbedded plane curves to
round points, J. Di. Geom. 26 (1987) 285314.
[Gur93] M. E. Gurtin, Thermomechanics of evolving phase boundaries in the plane,
Oxford, Clarendon Press (1993).
[HIY06] C. Hirota, T. Ishiwata and S. Yazaki, Note on the asymptotic behavior of
solutions to an anisotropic crystalline curvature ow, Recent Advances on El-
liptic and Parabolic Issues (Editors: Michel Chipot and Hirokazu Ninomiya),
Proceedings of the 2004 Swiss-Japan Seminar, World Scientic (2006) 129
143.
[HIY07] C. Hirota, T. Ishiwata and S. Yazaki, Numerical study and examples on
singularities of solutions to anisotropic crystalline curvature ows of noncon-
vex polygonal curves, Proceedings of MSJ-IRI 2005 Asymptotic Analysis
and Singularities (Sendai, 2005), Advanced Studies in Pure Mathematics
(ASPM) 47 (2007) 543564.
[HGGD05] H. Hontani, M.-H. Giga, Y. Giga and K. Deguchi, Expanding selfsimilar
solutions of a crystalline ow with applications to contour gure analysis,
Discrete Applied Mathematics 147 (2005) 265285.
[IS99] K. Ishii and H. M. Soner , Regularity and convergence of crystalline motion,
SIAM J. Math. Anal. 30 (1999) 1937.
[IUYY04] T. Ishiwata, T. K. Ushijima, H. Yagisita and S. Yazaki, Two examples of
nonconvex self-similar solution curves for a crystalline curvature ow, Proc.
Japan Academy 80, Ser. A, No. 8 (2004) 151154.
[IY07] T. Ishiwata and S. Yazaki, Towards modelling the formation of negative
ice crystals or vapor gures produced by freezing of internal melt gures.
Symposium on Mathematical Models of Phenomena and Evolution Equations
(Kyoto 2006). RIMS Koky uroku 1542 (2007) 111.
[IY08] T. Ishiwata and S. Yazaki, Interface motion of a negative crystal and its
analysis. Nonlinear Evolution Equations and Mathematical Modeling (Kyoto
2007). RIMS Koky uroku 1588 (2008) 2329.
[KG01] R. Kobayashi and Y. Giga, On anisotropy and curvature eects for growing
crystals, Japan J. Indust. Appl. Math. 18 (2001) 207230.
[MS00] U. F. Mayer and G. Simonett, Self-intersections for the surface diusion and
the volume-preserving mean curvature ow, Dierential Integral Equations 13
(2000) 11891199.
[McC89] J. C. McConnel, The crystallization of lake ice, Nature 39 (1889) 367.
[Nak56] U. Nayaka, Properties of single crystals of ice, revealed by internal melting,
SIPRE (Snow, Ice and Permafrost Reseach Establishment) Research Paper
Bibliography 209
13 (1956).
[Rob93] S. Roberts, A line element algorithm for curve ow problems in the plane,
CMA Research Report 58 (1989); J. Austral. Math. Soc. Ser. B 35 (1993)
244261.
[Sch93] R. Schneider, Convex bodies: the Brunn-Minkowski theory, Cambridge
University Press (1993).
[Set96] J. A. Sethian, Level set methods: evolving interfaces in geometry, uid
mechanics, computer vision, and materials science, Cambridge Monographs
on Applied and Computational Mathematics 3, Cambridge University Press,
Cambridge (1996).
[
SY07] D.

Sevcovic and S. Yazaki, On a motion of plane curves with a curvature
adjusted tangential velocity, submitted to Proceedings of Equadi 2007 Con-
ference (arXiv:0711.2568v1).
[
SY] D.

Sevcovic and S. Yazaki, Evolution of plane curves with curvature adjusted
tangential redistribution velocity, in preparation.
[Tay90] J. E. Taylor, Crystals, in equilibrium and otherwise, videotape of 1989
AMS-MAA lecture, Selected Lectures in Math., Amer. Math. Soc. (1990).
[Tay91a] J. E. Taylor, Constructions and conjectures in crystalline nondierential
geometry, Proceedings of the Conference on Dierential Geometry, Rio de
Janeiro, Pitman Monographs Surveys Pure Appl. Math. 52 (1991) 321336,
Pitman London.
[Tay91b] J. E. Taylor, Motion by crystalline curvature, Computing Optimal Ge-
ometries, Selected Lectures in Math., Amer. Math. Soc. (1991) 6365 plus
video.
[Tay93] J. E. Taylor, Motion of curves by crystalline curvature, including triple
junctions and boundary points, Di. Geom.: partial di. eqs. on manifolds
(Los Angeles, CA, 1990), Proc. Sympos. Pure Math., 54 (1993), Part I, 417
438, AMS, Providencd, RI.
[TCH92] J. E. Taylor, J. Cahn and C. Handwerker, Geometric models of crystal
growth, Acta Metall., 40 (1992) 14431474.
[Tot72] L. F. Toth, Lagerungen in der Ebene auf der Kugel und im Raum, Springer-
Verlag, Heidelberg (1972); Japanese transl., Misuzu Shobo (1983).
[UY04] T. K. Ushijima and S. Yazaki, Convergence of a crystalline approxima-
tion for an area-preserving motion, Journal of Computational and Applied
Mathematics 166 (2004) 427452.
[Yaz02] S. Yazaki, On an area-preserving crystalline motion, Calc. Var. 14 (2002)
85105.
[Yaz04] S. Yazaki, On an anisotropic area-preserving crystalline motion and motion
of nonadmissible polygons by crystalline curvature, RIMS Koky uroku 1356
(2004) 4458.
[Yaz07a] S. Yazaki, Motion of nonadmissible convex polygons by crystalline curva-
ture, Publications of Research Institute for Mathematical Sciences 43 (2007),
155170.
[Yaz07b] S. Yazaki, Asymptotic behavior of solutions to an area-preserving motion
by crystalline curvature, Kybernetika 43 (2007) 903912.
210 Bibliography
[Yaz07c] S. Yazaki, On the tangential velocity arising in a crystalline approximation
of evolving plane curves, Kybernetika 43 (2007) 913918.
[YK90] E. Yokoyama and T. Kuroda, Pattern formation in growth of snow crystals
occurring in the surface kinetic process and the diusion process, Physical
Review A 41(4) (1990) 20382049.
Index
anisotropy, 49
crystal
negative, 71, 72
crystals
ice, 70
curvature, 45
anisotropic, 50
crystalline, 61
weighted, 50
curve
C
k
, 44
admissible, 59
closed, 44
concave, 45
convex, 45
embedded, 44
essentially admissible, 59
immersed, 44
Jordan, 44
noncovex, 45
open, 44
oval, 45
plane, 44
polygonal, 57
simple, 44
strictly convex, 45
diagram
Frank, 51
energy
crystalline, 55
interfacial, 49
equation
anisotropic curvature ow, 50
classical curvature ow, 47
crystalline curvature ow, 57, 61
curve shortening, 47
weighted curvature ow, 50
extension
positively homogeneous of degree 1, 49
gure
Tyndall, 70
vapor, 70
ow
area preserving, 48
classical curvature, 47
gradient, 47, 59
weighted curvature, 50
formulae
FrenetSerret, 45
inequality
anisoperimetric, 65
motion
area preserving, 63
PHD1, 49
PPHD1, 49
shape
Wul, 52
vector
positive direction of, 45
unit tangent, 44
211
APPENDIX A
1. Strange examples
The following set describes the unit circle:
= x() = (cos f(), sin f()); [0, 1],
if we take the range of f is greater than or equals 2. However, the following circles
are strange in the sense of contrary to intuition.
Non-closed circle: If f() = (2 +1), then x is not closed, since x(0) ,=
x(1).
Non-simple circle: If f() = (2 1)(4(2 1)
2
1)/3, then the map
f : [0, 1] [0, 2] is surjective but not injective. Hence x is not simple.
Non-immersed circle: If f() = ((21)
3
+1), then the map f : [0, 1]
[0, 2] is bijective. Hence x is simple and C
1
-closed curve but not im-
mersed, since g() = [
f()[ = 0 holds at = 1/2.

Non-embedded circle: If f() = 2n (n = 2, 3, . . .), then x is the n-
multiply covered unit circle. Since the map x : is not injective, x
is immersed but not embedded.
2. A non-concave curve
We need some attention concerning the converse of Lemma 1.5. Strictly convex
curves admit nite number of zeros of K (such points are considered as degenerate
line segment). Figure 1 indicates such an example. Thus we can say that
A closed embedded C
2
-curve is strictly convex, if and only if
K > 0 holds in except nite number of zero points.
3. Anisotropic inequalityproof of Lemma 1.53
The result of Lemma 1.53 follows from a classical convex geometry by using a
concept of mixed area and the Br unn and Minkowskis inequality.
Let T
0
and T
1
be N-sided admissible convex polygons associated with
.
For i = 0, 1, by h
(i)
j
and d
(i)
j
, we denote the j-th hight function and the length
of the j-th edge, respectively. Note that d
(i)
j
= D
j
(h
(i)
) holds for i = 1, 2, . . . , N,
h
(i)
= (h
(i)
1
, h
(i)
2
, . . . , h
(i)
N
). Let the enclosed region of T
i
be
i
=
N
j=1
x R
2
; x
n
j
h
(i)
j
for i = 0, 1. Dene the linear interpolant of
0
and
1
by
=
(1 )
0
+
1
= (1 )x + y; x
0
, y
1
, and T
= (1 )T
0
+ T
1
for 0 1. Then the j-th height function and the length of the j-th edge of T
213
214 A
1+/2
2+/2
3+/2
4+/2
5+/2
6+/2
3/2
5/2
0 0.19 0.29 0.43 0.68 0.78 0.90 1 0.5
/2
=1 (=0)
=0.190811
=0.294048
=0.431905
=0.675084
=0.779306
=0.902935
=0.5
Figure 1. Graph of versus (left), and the C
2
-class strictly
convex curve = x(); [0, 1] (right). Here () =
() = (2 cos(2))/2 if [0, 0.5), () = ( 0.5) +
if [0.5, 1], and x() =
_
0
(cos (), sin ()) d. In the
right gure, the set ; () is indicated, where =
i + /2 (i = 1, 2, . . . , 6), 3/2, 5/2. Note that (0.5) = 3/2
and (1) = 5/2. Since g() 1, the curvature is given by
K() =
=
2
[ sin(2)[. Hence K() is positive except two
points = 0.5 and = 1 ( = 0).
are given as h
()
j
= (1 )h
(0)
j
+h
(1)
j
and d
()
j
= (1 )d
(0)
j
+d
(1)
j
, respectively.
Let A(
) be the area of
. Then we have
A(
) =
1
2
N
j=1
d
(s)
j
h
(s)
j
= (1)
2
A(
0
)+
2
A(
1
)+
(1 )
2
N
j=1
_
d
(1)
j
h
(0)
j
+d
(0)
j
h
(1)
j
_
.
By summation by parts:
N
j=1
d
(1)
j
h
(0)
j
=
N
j=1
D
j
(h
(1)
)h
(0)
j
=
N
j=1
D
j
(h
(0)
)h
(1)
j
=
N
j=1
d
(0)
j
h
(1)
j
, (A.1)
we have
A(
) =
1
2
N
j=1
d
()
j
h
()
j
= (1 )
2
A(
0
) +
2
A(
1
) + 2(1 )
1
2
N
j=1
d
(0)
j
h
(1)
j
.
The coecient of 2(1 ) is called the mixed area of
0
and
1
, and denoted
by
A(
0
,
1
) =
1
2
N
j=1
d
(0)
j
h
(1)
j
.
Note that A(
0
,
1
) = A(
1
,
0
) holds by (A.1).
H. Br unn and H. Minkowski proved the following inequality:
_
A(
) (1 )
_
A(
0
) +
_
A(
1
), 0 1. (A.2)
3. ANISOTROPIC INEQUALITYPROOF OF LEMMA 1.53 215
Equality holds if and only if
0
= k
1
(k > 0). The Br unn and Minkowskis
inequality (A.2) is equivalent to the following inequality:
A(
0
,
1
)
_
A(
0
)A(
1
). (A.3)
From this inequality, we obtain the anisoperimetric inequality as follows. Let
=
0
be the enclosed region of a polygon T and let the enclosed area be A() = A.
For the crystalline energy > 0, let
1
be the Wul shape
. Then the area of
is A(
) = [
[ =
N
k=1
(n
k
)l
(n
k
)/2, and the mixed area is A(,
) =
N
j=1
(n
j
)d
j
/2 = E
(T)/2, which is a half of the total interfacial energy on T.

Hence by (A.3), E
(T)/2
_
A[
[, namely,
(T) =
E
(T)
2
4[
[A
1. (A.4)
The equality
(T) 1 holds if and only if = k
for some constant k > 0, i.e.,
(n
j
) const. for j = 1, 2, . . . , N.
In particular, if 1, then E
= L and [
[ =
N
j=1
l
(n
j
)/2, so we have
1 =
L
2
2
N
j=1
l
1
(n
j
)A
1, (A.5)
which is the isoperimetric inequality of polygons. Note that L. Fejes Toth writes
in his book [Tot72] such that this inequality was done by S. Lhuilier. We refer
R. Schneider [Sch93] for general theory of convex bodies. See also [Yaz02] for the
proof of (A.5) by using a solution of crystalline curvature ow equation.
The isoperimetric inequality (A.5) represents the variational problem: what is
the shape which has the least total length of a polygon for the xed enclosed area?
The answer (it corresponds to the case where 1 = 1) is the boundary of the Wul
shape
, which is the circumscribed polygon of the circle with radius
=
_
2A/
N
i=1
l
1
(n
i
) (cf. Theorem 1.49). Similarly, the anisoperimetric inequality
(A.4) represents the variational problem: what is the shape which has the least
total interfacial energy of a polygon for the xed enclosed area? The answer (it
corresponds to the case where
= 1) is the boundary of the Wul shape
(cf. Theorem 1.50).

(Eds.) M. Benes and E. Feireisl
Topics in mathematical modeling
Published by
MATFYZPRESS
Publishing House of the Faculty of Mathematics and Physics
Charles University, Prague
Sokolovsk a 83, CZ 186 75 Praha 8
as the 254. publication
The volume was typeset by the authors using L
A
T
E
X
Printed by
Reproduction center UK MFF
Sokolovsk a 83, CZ 186 75 Praha 8
First edition
Praha 2008
ISBN 978-80-7378-060-9
Jind rich Ne cas
Jindrich Necas was born in Prague on December 14
th
, 1929. He studied mathe-
matics at the Faculty of Natural Sciences at the Charles University from 1948 to
1952. After a brief stint as a member of the Faculty of Civil Engineering at the
Czech Technical University, he joined the Czechoslovak Academy of Sciences where
he served as the Head of the Department of Partial Dierential Equations. He held
joint appointments at the Czechoslovak Academy of Sciences and the Charles Uni-
versity from 1967 and became a full time member of the Faculty of Mathematics
and Physics at the Charles University in 1977. He spent the rest of his life there, a
signicant portion of it as the Head of the Department of Mathematical Analysis
and the Department of Mathematical Modeling.
His initial interest in continuum mechanics led naturally to his abiding passion
to various aspects of the applications of mathematics. He can be rightfully consid-
ered as the father of modern methods in partial dierential equations in the Czech
Republic, both through his contributions and through those of his numerous stu-
dents. He has made signicant contributions to both linear and non-linear theories
of partial dierential equations. That which immediately strikes a person conver-
sant with his contributions is their breadth without the depth being compromised
in the least bit. He made seminal contributions to the study of Rellich identities and
inequalities, proved an innite dimensional version of Sards Theorem for analytic
functionals, established important results of the type of Fredholm alternative, and
most importantly established a signicant body of work concerning the regularity
of partial dierential equations that had a bearing on both elliptic and parabolic
equations. At the same time, Necas also made important contributions to rigorous
studies in mechanics. Notice must be made of his work, with his collaborators, on
the linearized elastic and inelastic response of solids, the challenging eld of contact
mechanics, a variety of aspects of the Navier-Stokes theory that includes regularity
issues as well as important results concerning transonic ows, and nally non-linear
uid theories that include uids with shear-rate dependent viscosities, multi-polar
uids, and nally incompressible uids with pressure dependent viscosities.
Necas was a prolic writer. He authored or co-authored eight books. Special
mention must be made of his book Les methodes directes en theorie des equations
elliptiques which has already had tremendous impact on the progress of the subject
and will have a lasting inuence in the eld. He has written a hundred and forty
seven papers in archival journals as well as numerous papers in the proceedings of
conferences all of which have had a signicant impact in various areas of applications
of mathematics and mechanics.
Jindrich Necas passed away on December 5
th
, 2002. However, the legacy that
Necas has left behind will be cherished by generations of mathematicians in the
Czech Republic in particular, and the world of mathematical analysts in general.
The Necas Center for Mathematical Modeling is a collaborative eort between the
Faculty of Mathematics and Physics of the Charles University, the Institute of
Mathematics of the Academy of Sciences of the Czech Republic and the Faculty of
Nuclear Sciences and Physical Engineering of the Czech Technical University.
The goal of the Center is to provide a place for interaction between mathemati-
cians, physicists, and engineers with a view towards achieving a better understand-
ing of, and to develop a better mathematical representation of the world that we
live in. The Center provides a forum for experts from dierent parts of the world
to interact and exchange ideas with Czech scientists.
The main focus of the Center is in the following areas, though not restricted
only to them: non-linear theoretical, numerical and computer analysis of problems
in the physics of continua; thermodynamics of compressible and incompressible
uids and solids; the mathematics of interacting continua; analysis of the equations
governing biochemical reactions; modeling of the non-linear response of materials.
The Jindrich Necas Center conducts workshops, house post-doctoral scholars
for periods up to one year and senior scientists for durations up to one term. The
Center is expected to become world renowned in its intended eld of interest.
ISBN 978-80-7378-060-9

Topics in Mathematical Modeling

Caricato da

Informazioni sul documento

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Topics in Mathematical Modeling

Caricato da

Copyright:

Formati disponibili

topics-in-mathematical-modeling 2008/12/5 8:30 page i #1

Jind rich Ne cas Center for Mathematical Modeling

(t) with a given boundary condition

, and no boundary condition

(x) B(X, Y ) is continuous, we say f C

(x) B(X, B(X, Y )) = B

(U) and calculate its Frechet derivatives.

(u)[w, w] 0 for all w X and u U.

((1 t)w +tv)[v w],

((1 t)w +tv)[v w, v w] |v w|

is bounded in X and there exists a subsequence u

) = B(X, B(X, R))

(x, y) U() V ()).

denotes the Frechet dierential operator with respect to M and

and apply Theo-

) for some k N. Then u C

Under the condition of Theorem 1.33, we dene

). Due to Theorem 1.30, however, the condition

) is not necessary as shown in the next theorem.

from the LaxMilgram theorem. Therefore,

is given by the next formula.

be an un-cracked elastic body in R

be an initial crack. We assume that is a closure of an (n 1)-

along a virtual crack extension (t) parametrized by

. Without loss of generality, we choose

(), the energy release rate G along (t)

is the Frechet derivative of E

with respect to and is the t derivative

. We consider a virtual crack extension (t)

(0). We suppose that

onto itself and satises (t)((0)) = (t) and ((t)) =

(0, 1) be a bounded cylindrical domain in R

is a bounded Lipschitz domain in R

onto itself and satises ((t)) = (t).

v)] = A()v a.e in (v W

()), and the (n + 1)-th Frechet

()) is clear. For xed W

()) follows. For xed W

is a linear topological isomorphism from L

be a Friedrichs mollier and dene f

W(x, v(x), v(x))dx,

W(x, v(x), v(x)), etc.

() := E(u(), ()), (1.22)

(v + g) V (, g). Using the formulas

(v + g), ()) = c(v, ) (v V ),

W ((x), v(x), [A()(x)]v(x)) ()(x)dx (v H

(u()) g V . Then u() is a local minimizer of E(, ())

is dened by (1.22) with u() =

W (x +t(x), u(x), [A(

is a linear topological isomorphism from H

W(u)] u(x) (x)dx

W(u) ) (u(x) (x)) dH

W(u) ) (u(x) (x)) dH

W(u) ) (u(x) (x)) dH

W(u) ) (u(x) (x)) dH

This theorem can be generalized as follows.

(s) | (s) and

) for all M. Then, prove

()) = m 1, at least one of

. For general f, applying a partition of unity, we

u the Hessian matrix of u()

[ = 1, we dene a plane curve in - plane

The following lemma will be used in proving Theorem 2.10.

topics-in-mathematical-modeling 2008/12/5 8:30 page 52 #64