(Series in Analysis 3) S. Kesavan-Symmetrization & Applications-World Scientific (2006)

\
Symmetrization &
Applications
\t
(acav
Symmetnzation & Applications
SERIES IN ANALYSIS
Series Editor: Professor Roderick Wong
City University of Hong Kong, Hong Kong, China
Published
Vol. 1 Wavelet Analysis

edited by Ding-Xuan Zhou
Vol. 2 Differential Equations and Asymptotic Theory in

Mathematical Physics
edited by Hua Chen and Roderick S.C. Wong
Vol. 3 Symmetrization and Applications

by S Kesavan
Symmettizatiort & Applications
S Kesavan
The Institute of Mathematical Sciences, India
l | p World Scientific
NEW JERSEY • LONDON • SINGAPORE • BEIJING • SHANGHAI • HONGKONG • TAIPEI • CHENNAI
Published by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
SYMMETRIZATION AND APPLICATIONS

Series in Analysis — Vol. 3
Copyright © 2006 by World Scientific Publishing Co. Pte. Ltd.
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN 981-256-733-X
Printed in Singapore by World Scientific Printers (S) Pte Ltd

Preface
The study of isoperimetric inequalities goes back to antiquity. The solution

of Dido's problem states that, of all plane domains of given perimeter, the
circular disc, and the disc alone, maximizes the enclosed area. Equivalently,
of all domains of given area, the disc minimizes the perimeter. This is
expressed via the classical isoperimetric inequality
L2 > 4TTA,
where L stands for the perimeter and A for the enclosed area of a domain
in the plane. Equality is attained only for the disc. In three dimensional
space, if S is the surface area of a body and V its volume, then
S 3 > 367ry2
and equality is attained only for the sphere. Thus, of all bodies of given
volume, the sphere, and the sphere alone, has least surface area. (Why are
soap bubbles spherical in shape? The soap bubble will attain a position of
equilibrium when the potential energy due to surface tension is minimal.
This energy is proportional to the surface area of the bubble. Hence, for a
given volume of air blown to form a bubble, it will take the spherical shape
which has the least surface area.) These two inequalities can be generalized
to all space dimensions.
In a broader sense, an isoperimetric problem tries to optimize a given
domain dependent functional keeping some geometric parameter of the do-
main (like its volume) fixed. The study of such problems started with the
conjecture of Saint Venant in 1856 regarding the optimal shape of the cross
section of a prism in order to maximize its torsional rigidity. This was
finally settled by Polya in 1948. In 1894, Lord Rayleigh, in his treatise
on the theory of sound, made conjectures regarding the vibrations of cer-
V
VI Symmetrization and Applications
tain elastic bodies. He conjectured that, of all fixed membranes of given

area, the circular membrane has the lowest fundamental frequency of vi-
bration. This was proved by Faber and Krahn towards the end of the first
quarter of the twentieth century. He also conjectured that of all vibrating
thin clamped plates of given area, the circular plate has the lowest fun-
damental frequency. This conjecture resisted solution for nearly a century
and was finally settled by Nadirashvili in 1992. While the Faber - Krahn
result can be extended to all dimensions, the conjecture regarding plates
remains unsolved in dimensions greater than three even today. Poincare
conjectured that of all bodies with given volume, the ball alone has the
least electrostatic capacity and this was proved by Szego in 1930.
The study of isoperimetric inequalities involves a fascinating interplay of
analysis, geometry and the theory of partial differential equations. Several
conjectures have been made and while many have been resolved, a large
number still remain open.
One of the first comprehensive treatises on this subject is the book by
Polya and Szego (1951). It has since been complemented by several review
articles and also by books such as those by Bandle (1980) and Mossino
(1984). Since then, several new results have been proved, and a few conjec-
tures have been resolved, especially those concerning eigenvalues of elliptic
partial differential operators.
As seen from the examples cited above, Nature often seems to choose
the perfect symmetric form, viz. spherical symmetry, when optimizing var-
ious characteristics of bodies. One of the principal tools in the study of
isoperimetric problems, especially when spherical symmetry is involved, is
Schwarz symmetrization, which is also known as the spherically symmetric
and decreasing rearrangement of functions. The aim of this book is to give
an introduction to the theory of Schwarz symmetrization and study some
of its applications. Other equally important types of symmetrization (for
example, Steiner symmetrization) are not treated here.
The first chapter introduces the notion of Schwarz symmetrization and
proves several of its properties. The principal result of the second chapter
is the famous inequality of Polya and Szego regarding Dirichlet integrals.
Since this proof depends on the classical isoperimetric inequality and the co-
area formula, these are also studied. The relationship between the classical
isoperimetric inequality and Sobolev's inequality is also discussed.
Usually, in their full generality, the proofs of the isoperimetric inequality,
the co-area formula and the Polya - Szego theorem involve heavy doses of
geometric measure theory. Here, a simple and elementary proof of the
Preface vn
isoperimetric inequality (that I learnt from a lecture by X. Cabre) has been

presented in the case of smooth domains. Simple versions of the co-area
formula that are enough to prove the Polya - Szego theorem are proved using
techniques mainly drawn from the study of partial differential equations.
The third chapter centers around Talenti's theorem which compares the
solution of a second order elliptic boundary value problem with that of a
'symmetrized problem'. Several applications of this result are studied.
The fourth chapter looks at various isoperimetric inequalities involving
eigenvalues of elliptic operators. The fifth (and last) chapter deals with
isoperimetric inequalities involving positive solutions of some non-linear
problems and uses them to obtain the radial symmetry of such solutions
when the domain is a ball.
Effort has been made to keep the exposition as simple and as self-
contained as possible. Occasionally, in order that we do not get mired
in technical details and thus lose the main thread of the argument, some
proofs, which are long and which involve completely different techniques,
have been omitted and the 'interested reader' is given appropriate refer-
ences. But such instances are few and far between. A knowledge of the
existence theory of weak solutions of elliptic partial differential equations
in Sobolev spaces is, however, assumed. Apart from this and a general
mathematical maturity at the graduate level, there are no other prerequi-
sites. The text is peppered with several exercises.
While the bibliography is fairly extensive, it is, as is to be expected,
far from exhaustive. Further references to a vast and rich literature can be
found in the works cited here. At the end of each of the last three chapters,
bibliographic comments indicate results and directions not treated in the
text.
I have taught most of the material covered in this book, as a short in-
troductory course, to graduate students at the Universita degli Studi di
Roma, La Sapienza, Rome, Italy. I wish to take this opportunity to thank
the university for its hospitality. Portions of this book were also taught at
instructional conferences in India organized by the Aligarh Muslim Univer-
sity, Aligarh, and the Tata Institute of Fundamental Research (TIFR) at
its Bangalore Centre (Indian Institute of Science Campus). I thank the or-
ganizers of these conferences for giving me the opportunity to deliver those
lectures.
I would like to thank the Institute of Mathematical Sciences, Chen-
nai for the excellent facilities it provided which greatly eased the task of
bringing out this volume. In particular, I wish to record my appreciation
viii Symmetrization and Applications
of the cooperation extended by the members of our library staff (and, of

course, their counterparts in the libraries of the TIFR Centre, Bangalore
and the TIFR, Mumbai) in tracing and procuring copies of research papers
in very old journals. I also thank Dr. M. Rajesh, who read portions of the
manuscript and made several helpful suggestions.
I learnt much about symmetrization from the works of, and through
discussions with, several experts. It is a pleasure to express my thanks to
Professor G. Talenti (especially for inviting me to the conference at Cortona,
which turned out to be a valuable experience, and for all that he has done for
me), to Professors G. TYombetti and V. Ferone (and all their colleagues) for
very enjoyable and useful visits to Naples, to Professor Jacqueline Mossino
for her warm hospitality at Orsay and to Professor Filomena Pacella for
giving me the opportunity to teach this course in Rome and for fruitful
collaboration (some of our joint work appears in this volume).
I am grateful to my family for the constant support given to me. I am
also very grateful to Professor P. G. Ciarlet who has been a constant source
of inspiration and encouragement. I also wish to thank Professor R. Wong,
the editor of the 'Series in Analysis', and the staff of World Scientific for
their cooperation in bringing out this volume.
Finally, I would like to dedicate this book to the memory of Professor
Jacques-Louis Lions, who was one of my teachers and to whom I owe more
than I can possibly express.
S. Kesavan
Notations
• MN denotes the N ~ dimensional Euclidean space.
• If x,y £ RN, then x.y is the usual scalar product in R ^ , i.e. if
x = (xi,...,xN), and y = (i/i,...,yjv), then x.y = J2iLixiVi-
N
• Ifxe R , then M = ( E ^ ! Afn-
N
• If E C R , then \E\ denotes its (N - dimensional) Lebesgue mea-
sure.
• If E C RN, then £* denotes the ball with centre at the origin such
that \E*\ = \E\.
• If E C R N , then |-E|jv_i denotes a suitable (iV — 1) - dimensional
measure (cf. Section 2.1).
• £ ( # ; r) denotes the open ball in MN with centre at x and radius r.
Br will also be used to denote the ball with centre at the origin
and radius r.
• LOW denotes the volume of the unit ball in MN (cf. Section 1.3).
• If Q C R ^ is a bounded domain, then d£l denotes its boundary.
Let u : Q C RN -> K.
• u + = max{u, 0} denotes the positive part of u.

u" — — min{u,0} denotes the negative part of u.
• Vu denotes the gradient of u.
• Au denotes the Laplacian of u, i.e.,
\—\ a u . ,„ .
Au dlv v
= Eas = (^
• u# (resp. u#) denotes the unidimensional decreasing (resp. in-
creasing) rearrangement of u (cf. Section 1.1 (resp. Section 1.4)).
• u* (resp. u*) denotes the spherically symmetric and decreasing
(resp. increasing) rearrangement of u (cf. Section 1.3 (resp. Section
1.4)).
Function Spaces
Let Vt C RN be an open set.
• C(fl) denotes the space of continuous functions on U.
ix
x Symmetrization and Applications
• Ck(Q.) denotes the space of functions which are k times continuously

differentiable in Q and whose derivatives upto this order possess
continuous extensions to Q.
• V(Q) denotes the space of C°° functions with compact support in
n.
Let 1 < p < oo.
• The norm in the usual Lebesgue space L P (H) is denoted by ||-||p,n-
• If m is a positive integer, then Wm'p(Cl) is the usual Sobolev space
of order m of functions in LP(U) all of whose (distribution) deriva-
tives upto order m are also in LP(U).
• The norm in this space is denoted by ||.|| m , p ,n-
• The closure of £>(H) in this space is denoted by W™,P(Q).
• When p = 2, the above spaces are denoted by Hm(Q) and H™^)
respectively.
Contents
Preface v
Notations ix
1. Symmetrization 1
1.1 The Decreasing Rearrangement 1
1.2 Some Rearrangement Inequalities 8
1.3 Schwarz Symmetrization 13
1.4 Variations on the Theme 15
2. Some Classical Inequalities 19

2.1 The Isoperimetric Inequality 19
2.2 The Co-area Formula 28
2.3 The Polya - Szego Theorem 35
2.4 Sobolev's Inequality 43
3. Comparison Theorems 47
3.1 Talenti's Theorem 47
3.2 The Equality Case 53
3.3 Sobolev Imbeddings 63
3.4 The Obstacle Problem 66
3.5 Electrostatic Capacity 70
3.6 The Saint Venant Problem 73
3.7 Comments 80
xi
xii Symmetrization and Applications
4. Eigenvalue Problems 83
4.1 The Faber - Krahn Inequality 83
4.2 The Szego - Weinberger Inequality 89
4.3 Chiti's Theorem 95
4.4 The Payne - Polya - Weinberger Conjecture 98
4.5 Rayleigh's Conjecture for Clamped Plates 105
4.6 The Buckling Problem 116
4.7 Comments 122
5. Nonlinear Problems 125

5.1 Payne - Rayner Type Inequalities 125
5.2 A System of Semilinear Equations 132
5.3 Comments 139
Bibliography 141
Index 147
Chapter 1
Symmetrization
1.1 The Decreasing Rearrangement
Schwarz symmetrization is a particular kind of rearrangement of functions

defined on a domain Q C M.N. Given a real valued function on such a
domain, we construct an associated function, on the ball centered at the
origin and of the same measure as fi, assuming the same range of values
and having special properties. In particular, we wish that this new function
be radial and radially decreasing. In order to define this, we first construct
the unidimensional decreasing rearrangement of the given function, which
we now proceed to do.
First of all, we need some notation. Given a (Lebesgue) measurable
subset E C M.N, we denote its N - dimensional (Lebesgue) measure by \E\.
Let fi C R ^ be a bounded measurable set. Let u : Q, —• R be a
measurable function. For t 6 R, the level set {u > t} is defined as
{u > t} = {x £ £1 | u(x) > £}.
The sets {u < t}, {u > £}, {u = t} and so on are defined by analogy. Then
the distribution function of u is given by
This function is a monotonically decreasing function of t and for t >

ess. sup(u), we have fiu(t) = 0, while for t < ess. inf('u), we have
(iu(t) = \Q\. Thus the range of fxu is the interval [0, |H|].
Definition 1.1.1 Let Q C R ^ be bounded and let u : Cl —» R be a mea-

surable function. Then the (unidimensional) decreasing rearrange-
1
2 Symmetrization and Applications
m e n t of u, denoted u # , is defined on [0, |fi|] by
u # (0) = ess. sup(u) 1

(1.1.1)
u*(s) = inf{« | /i u (t) < s}, s > 0. J
R e m a r k 1.1.1 Essentially, vft is just the inverse function of the distribu-

tion function fiu of u. However, since fiu(t) is just monotonically decreasing,
it can have jump discontinuities. If t is a point of discontinuity, then the
above definition fixes the value of u * in the interval [fiu(t+), fiu(t—)] as tM
R e m a r k 1.1.2 In classical texts (cf. [Hardy, Littlewood, and Polya (1952)]

or [Polya and Szego (1951)]), the distribution function is usually defined
as fj,u(t) = |{|«| > t}\. Consequently, the definition of the rearrangement
would correspond to, in our notation, that of \u\#. Throughout this text,
we will, however, deal with the 'rearrangement with sign' i.e. our definition
of the rearrangement will be based on the distribution function which takes
the sign of the function into account. •
E x a m p l e 1.1.1 [Kawohl (1986)]
0 0.5
Let H = ( - 2 , 2) C K. Define u : ft -> R as follows.
2 + y, - 2 < y < - l
l,-l<y<0
u(y)
1 + y, 0 < y < 0 . 5
2 - 2/, 0.5 < y < 2.
Symmetrization
Then, it is easy to check that
4 - 2t, 0 < t < 1

Hu(t) = { 3 - 2 t , 1 < t< 1.5
0, i > 1.5
and that
(3-s)/2, 0 < s < l
u#(s) 1, 1 < s < 2
(4-s)/2, 2 < s <4.
1.5 ,#
1.5
The following properties of the decreasing rearrangement are immediate

from its definition.
Proposition 1.1.1 Let u : Q —> M. where Q C>N is bounded. Then u&
is a non-increasing and left-continuous function.
Proof: (i) Let si < s2. Then \{u > t}\ < s\ implies that \{u > t}\ < s2.
Thus,
{t \ i*u{t) < si} C {t \fiu(t)<s2}.
Thus, by definition, it follows that u^(s\) > u*(s2).

(ii) Let s G (0, \Q\). By definition of u#, given e > 0, there exists a t
such that u#(s) < t < u#(s) + e and fiu(t) < s. Choose h > 0 such that
/iu(£) < s — h < s. Then, for all 0 < h' < /i, we have (iu(t) < s — h' < s
and so u#(s) < u&(s ~~ hf) < t < u*(s) + e. This proves that vft is
left-continuous. •
Proposition 1.1.2 The mapping u \—• u# is non-decreasing, i.e. if u <

v, where u and v are real valued functions on fi, then u^ < v&.
Proof: Since {u > t} C {v > t}, we have that
{t\ \{v>t}\<s} C {* | \{u>t}\<s}
and the result follows from the definition. •
Definition 1.1.2 Two real valued functions (with possibly different do-
mains of definition) are said to be equimeasurable if they have the same
distribution function. Equimeasurable functions are said to be rearrange-
ments of each other.
We now show that u# is indeed a rearrangement of u in this sense.
Proposition 1.1.3 The functions u : Q —> R and u* : [0, |fi|] —> R are
equimeasurable, i.e., for all t,
\{u>t}\ = \{u*>t}\. (1.1.2)
Proof: If vft(s) > t, then, by definition, it follows that \{u > t}\ > s.
Thus,
{5 | u*{s) >t) C {s\ \{u > t}\ > s}.
Since u# is non-increasing, we have
\{u#>t}\ - sup{s | u#(s) >t} < |{u>t}|. (1.1.3)
On the other hand, let |{u* > t}| — s. By the left-continuity and the
non-increasing nature of u # , it follows that u#(s) = t. Then, by definition,
\{u>t}\ <s. Thus,
|{«>«}| < |{ti # >OI- (1-1-4)
Applying (1.1.3) and (1.1.4) for t + h instead of t, we get
\{u#>t + h}\ < \{u>t + h}\ < | { w # > t + ft}|.
Passing to the limit as h j 0, we get
\{u*>t}\ ^ l{«>*}l < K« # >*}l

which proves (1.1.2). •
Symmetrization 5
Corollary 1.1.1 With the preceding notations, we have
\{u>t}\ = \{u*>t}\.
\{u>t}\ = \{u*>t}\.
|{«<t}| = |{«#<*}|-
Proof: The first relation of (1.1.5) has already been proved. The rest fol-
low easily by complementation and suitable limiting arguments. •
Remark 1.1.3 The above proposition, in the light of Definition 1.1.2, ex-
plains why u# is called a rearrangement of u. The decreasing rearrangement
is just one example of a wide variey of such constructions. The reader is
referred to the book of [Kawohl (1985)] for examples of different kinds of
rearrangements. •
Corollary 1.1.2 If u > 0, and if u e Lp(Cl) for 1 < p < oo; then
u* GLP((0,|fi|)) and
IMIp.n - llu#llp,(o,|n|)-
(Here the norms are the corresponding Lv - norms.)
Proof: If p — oo, then the result is contained in the definition of the

rearrangement. Let 1 < p < oo. By equimeasurability, both u and u# have
the same distribution function p, and as both functions are non-negative, it
follows (by a simple application of Fubini's theorem) that
/•CO
i«ii;,n = p / *p-vm* lu#llP,(o,!n|).

Jo
This result is also true without the non-negativity condition. In fact, as

a consequence of the equimeasurability, we have the following general and
powerful result.
Theorem 1.1.1 Let u : 0, —> R be measurable. Let F : R —• R be a

non-negative Borel measurable function. Then
r /-mi
/ F(u(x))dx = / F{u*(s))ds. (1.1.6)
Jn Jo
Proof: Let E = [£,oo] and set F(£) = XE(0> where XE is the indicator
(or characteristic) function of E. Then
/ F(u{x))dx = \{u>t}\ = \{u#>t}\ = / F{u#(s))ds.

Jn Jo
Similarly, the result holds for F = XE where E is any interval and hence if
E is any open set and, again, if E is any Borel set, by standard arguments.
Hence the result is true for any non-negative simple function F. If F is
any non-negative Borel function, it can be expressed as the limit of an
increasing sequence {Fn} of non-negative simple functions. Thus, for each
n we have
, ,|fi|
/ Fn{u(x))dx = / Fn{u#{s))ds. (1.1.7)
Jn Jo
We can pass to the limit as n —+ oo to get (1.1.6) using the monotone
convergence theorem. •
Corollary 1.1.3 Let F : R —+ R be a Borel function and let u : H —> R
be such that F(u) e L 1 ^ ) . Then F ( u # ) £ ^ ( ( 0 , |fi|)) and (1.1.6) is still
valid.
Proof: We write F = F+ — F~ and both F+ and F~ are non-negative
Borel functions and so (1.1.6) holds for each of them in place of F . If
F(u) G L1(Q)i then both jnF+(u(x))dx and J n F~(u(x))dx are finite and
we can subtract the relation for F~ from that of F+ to get (1.1.6). •
Corollary 1.1.4 Let u G Lp{Cl) forl<p<oo. Then u* G L p ((0, |fi|))
and the corresponding Lp norms are equal.
Proof: Take F(t) = \t\p in the preceding theorem. •
Remark 1.1.4 The result remains true for p = co since u*(0) —

ess. sup(u) and u # ( | n | ) = ess. inf(u). •
Remark 1.1.5 Since the proofs of Theorem 1.1.1. and its consequences
depended only on the equimeasurability, these results also hold for other
types of rearrangements. •
We now prove another important property of the (decreasing) rearrange-

ment which is a consequence of Theorem 1.1.1.
Lemma 1.1.1 Let u : [0,1] —• R be non-increasing. Then u = vft a.e.
Symmetrization 7
Proof: If t < u(s), then, since u is non-increasing, \{u > t}\ > s. Hence,
u*(s) > t, by definition. This implies that
u*{s) > u(s) (1.1.8)
for all s £ [0,/]. Now, let s be a point of continuity for u. Since u is

non-increasing,
\{u>u(s-h)}\ < s- h < s
for h > 0. Thus, by definition, u#(s) < u(s — h). Hence, as h J, 0, we get,
using the continuity of u at s, that
u*{s) < u(s). (1.1.9)
Thus, by (1.1.8) and (1.1.9), we get that u#(s) = u(s) at all points of
continuity of u. The result stands proved since u, being monotonic, admits
at most only a countable number of discontinuities. •
Proposition 1.1.4 Let tft : R —> R be a non-decreasing function. Let
u : Q, —* R, where Q, C M.N is bounded. Then
iP(u*) = (ip(u)}#, a.e. (1.1.10)
Proof: Step 1. If u, w : [0,1] —> R are equimeasurable and non-increasing,

then v — w a.e. For, as they are non-increasing, by the preceding lemma,
v = v# and w = w# a.e. As they are equimeasurable, by definition,
t># = w # .
Step 2. The result will be proved if we can show that 4>(u#) and (V"(u))*
are both equimeasurable and non-increasing on [0, |fi|]. That they are non-
increasing follows from the definition of the rearrangement and the fact
that ij) is non-decreasing. Now,
\{i>{u#) > t}\ = J^] X{ip(u#)>t}{s)ds = SaX{i>{u)>t}{x)dx
= \Mu)>t}\ = |{W«))#>t}|.
The equality of the two integrals above is a consequence of Theorem 1.1.1
since the function s •-» X{^>t}(5) ls a non-negative Borel function. The
last equality is a consequence of the equimeasurability of a function and its
rearrangement. •
Corollary 1.1.5 For u : Q -> R, we have ( u + ) # = ( u # ) + .
Exercise 1.1.1 Show by means of an example that, in general, given two

functions v,w : Q —+ R, (v + w)# and v# 4- w# are not equal. However, if
c G l , show that
(v -f c)* = w* + c.
Exercise 1.1.2 Let Q, be the annulus
{x G R 2 | 0 < # i < |z| < i?o}.
Let u(x) = (R% - |x| 2 )/2. Show that
1.2 Some R e a r r a n g e m e n t Inequalities
We begin by proving the Lp continuity of the rearrangement map.
P r o p o s i t i o n 1.2.1 Let p = 1 oroo, Then for f, g E Lp(fi.),
ll/ # -9 # ll P ,(0,|n|) < I I / - S I U - (1.2.1)
Proof: Let p = oo. Then for almost all # E fi, we have
| / ( x ) - p ( x ) | < ||/-9||oo,n.
Thus,
/ ( * ) - 1 1 / - s | | o o , n < 9(x) < /(x) + | | / - 9 | | o o , n .

By the monotonicity of the rearrangement map (cf. Proposition 1.1.2) and
by Exercise 1.1.1, we deduce that
/ * ( » ) - 11/ -3lloc,n < 9*(s) < /*(«) +||/-a||oo,n
which immediately gives (1.2.1) for p = oo.
Let p = 1. Set h = max{/,g}. Then as / < h and g < h, we have that

/ # < / i # andg* <h#. Now,
| / # - 5 # | < \f#-h#\ + \h#-g#\ = 2ft#-/#-*#

Symmetrization 9
Thus,
C \f*(s) - 9*{s)\ds < jM(2h#(8) - f#(s) - g#(s))ds
= Jn(2h(x)-f(x)-g(x))dx
= fn\f{x) - 9{x)\dx
which proves the result for p — 1. •
Theorem 1.2.1 Let 1 u in L P (H). Since ft is bounded, it follows that
un —» u in L 1 (fi) as well and so u# —> u * in L 1 ((0, |H|)). Hence, for a
subsequence, ujjf —> u# a.e. Further,
HunJlp,(o,|n|) = I K J L n -> Nlp,n = llw#llP,(o,|n|)-

Hence it follows that u*fc —> u# in L p ((0, |ft|)) as well. In fact, as the limit
is independent of the subsequence, it follows that the entire sequence {ujf}
converges to vft in Lp((0, \Cl\)) and this completes the proof. •
We saw in the previous section that the rearrangement preserves the

integral of a function over the domain (take F(t) — t in Theorem 1.1.1).
We now consider integrals over proper subsets.
Proposition 1.2.2 Let Ct C MN be bounded and let u : Cl —> R be an
integrable function. Let E C fi be a measurable subset. Then
r r\E\
I u(x)dx t}\ < s, then
\{v>t}\ = \{u>t}C\E\ < s.
Thus
{ t | \{u>t}\<s} C {t\,\{v>t}\<s}
and so v#(s) < u#(s). Thus

f f\E\ f\E\
\ u(x)dx = / v*{s)ds < / u#{s)ds, (1.2.3)
JE JO JO
which proves (1.2.2). If equality holds in (1.2.2), then we have equality
throughout in (1.2.3) and this is possible if, and only if, v# = u# a.e. in E
and the result is proved. •
Lemma 1.2.1 Let u : fi —> M. and let t G R. Define
Et = {x e O I u(x) > t}
Ft = {x e n J u(x) < t) = n\Et.
Define b : R x U -> R by
Xs t (z) */ t > 0
!
^ ' ^ -XF t (x) i / t < 0 .
TTien
i-oo
b(t,x)dt. (1.2.4)
/ -OO
Proof: If u(z) > 0,

+oo pu{x)
/
b(t,x)dt — / dt = u(x).
If u(a;) < 0, then -oo ^0
+oo /*0
di = u(x).
/ (x)
b{t,x)dt = -
-oo «/u(:
Lemma 1.2.2 Let / , g : fi —> R wt/i # integrable over fl. Let a < f t} J
Proof: Assume that a > 0 (the other case can be similarly treated). Setting
Et = {/ > £}, we have, by the preceding lemma,
/(a) = / XEt(x)dt
Jo
Symmetrization 11
Thus, by Fubini's theorem,

•6
/ f(x)g(x)dx = / g(x) / XEt(x)dtdx "= g{x)\Et{x)dxdt

Jo. Ja Jo Jo Jo,
which gives
/ f{x)g{x)dx = 1 1 g(x)dxdt-\- I I g{x)dxdt

Jo JO JO Ja JEt
from which (1.2.5) follows immediately. •
Exercise 1.2.1 In the above proposition, if b G R and if —oo < a < f <b,
show that
/ f(x)g(x)dx = b g(x)dx - / I / g(x)dzIx I dt.

Jo. Jo Ja \J{f<t}
We are now in a position to prove an important rearrangement inequal-
ity.
Theorem 1.2.2 (Hardy - Littlewood) Let f G LP(Q.) and g G L*(fi)
where (l/p) + (l/<?) = 1, 1 < p, Q < oo. Then
/ f(x)g(x)dx < / f#(s)g#(s)ds. (1.2.6)

Jo Jo
Proof: Assume first that / G L°°'(fi) n Lp (H). Let a and b be real numbers
such that a < f < b. Then, by the preceding lemma,
fn f(x)g(x)dx = afn g(x)dx + f* J{f>t} g(x)dxdt
a
= /J"' 9*(s)ds + $ba J{f>t} g(x)dxdt
<aCg*{s)ds + !hJ]o{S>^9*{s)dsdt
= aC9ns)ds + Jbaf}i{f*>t»g#(s)dsdt
= f^f*(s)9*(s)d8
where the last equality comes from applying (1.2.5) to / # and g#.
If 1 < p < oo, the general case can be completed by a density argument
since we know that the mapping u t—* u# is continuous from Lp(il) into
L*((o,|n|)). •
Exercise 1.2.2 Let f,g£ L2{9). Using (1.2.6), show that
||/#-3#lk(Q,|n|) < ||/-s||2,n-

Exercise 1.2.3 Let ip : R —> R be non-decreasing. Let / e LP(Q) and
</»(s) e Li{n), where (1/p) + (l/q) = 1, 1 < p,<? < oo. Show that
[ f(x)ij(9(x))dx < [ f#(s)i>(g#(s))ds.

JQ JO
Exercise 1.2.4 (a) Let u : Q, —• R. Show that {x{u>t})^ — X{-u#>t}-

(b) Show that, if u, v : Q -> R, then
/ u(a:)X{i;<t}(^)^ > / u # (s)x{ i; #<i}(s)rfs.

Jn " Jo
We saw that the rearrangement mapping is non-expansive between the
relevant Lp spaces when p = 1,2 or oo. In fact, it is true for all 1 o
and
J
-w-\ o if t>o
so that J — J+ + J-. Both J + and J_ are convex and differentiate
functions. Thus,
J+(/(x)-g(x)) = / J'+(f(x)-t)dt = / j;(/(x)-t)x{s<t}(x)*.

Jg(x) J—oo
Hence, after an application of Fubini's theorem, we get
[ J+{f(x)-g(x))dx = [ [ J'+(f(x)-t)X{9<t}(x)dxdt (1.2.8)

JO. J-oc JQ
Symmetrization 13
and, similarly
•\n\ ,.+«> r\n\
/ J+(f*(s)-g*(8))ds = / j;(/*(a)-t)X{9#<t}(«)A«it.
JOo J-oo JO
(1.2.9)
Now, since J+ is convex, we have that J^_ is non-decreasing. Thus, by
Proposition 1.1.4 and Exercise 1.1.1, we get
(J'+(f(x)-t))#(s) = J'+(f*(s)-t).
Hence, by virtue of Exercise 1.2.4, we deduce that
\n\
J,+ (f(x)~t)X{9<t}(x)dx
/ J+{f(x)-t)Xig<t}(x)dx > I/ J'+(f#(s)-t)X(9#<t}(s)ds.
Jn
'Q. Jo
J0
Thus, using (1.2.8) and (1.2.9), we deduce that

r r\n\
/ J+(f(x)-g(x))dx > / J+(f#(s)-g#(s))ds.
JQ JO
We can prove, in the same way, the same inequality with J_ in place of J +
Adding, we get (1.2.7). •
1.3 Schwarz Symmetrization
Henceforth, given a measurable subset E C R ^ of finite measure, we will

denote by E*, the open ball centered at the origin and having the same
measure as E, i.e. \E*\ = \E\. Given a vector x € MN, we denote its
Euclidean norm by |a;|. Finally, we will denote by U>N, the volume of the
unit ball in R N . Notice that
K
7T 2
UJN
r(f + i;
where T(s) is the usual gamma function.
Definition 1.3.1 Let ft, C MN be a bounded domain. Let u : SI —> R

be a measurable function. Then, its Schwarz symmetrization or, the
spherically symmetric and decreasing rearrangement is the function
u* : Q* -> R defined by
u*(x) = U#(LJN\X\N), xttt*.

Observe that, if R is the radius of Q*, then
Ja, u*(x)dx = / n . u#(uN\x\N)dx = f*u#(ijNrN)NuNrN-1dr
= fWu#(a)ds =fWu#(s)ds.
We can translate all the results obtained in the preceding sections for
the decreasing unidimensional rearrangement to get the corresponding re-
sults for the Schwarz symmetrization. In particular, we easily deduce the
following:
u* is radially symmetric and decreasing.

u, u& and u* are all equimeasurable.
If F : R —* R is a Borel measurable function such that either F > 0
or F{u) eLl{n), then
/ F{u*(x))dx = / F(u (x))dx. (1.3.1)
In particular, u and u* have the same LP norms and JL u(x)dx =

J^, u*(x)dx when u is integrable over il.
• If ip : M —> R is a non-decreasing function, then
• The mapping u —i » u* is a non-expansive mapping from Lp(fl) into

LP(n*) for 1 <p< oo.
• If £ C O is a measurable subset, then,
r r\E\ r
I u{x)dx < I u#(s)ds = \ u*(x)dx. (1.3.2)
JE JO JE*
Equality occurs if, and only if, {U\E)* — U*\E*-

• (Hardy - Littlewood) If / € L p (fi) and g 6 L«(fi) where (1/p) +
(l/g) = l,then,
f f(x)g(x)dx < J f*(s)g*{s)ds = / r(x)g*(x)dx.

Jn Jo Jn*
(1.3.3)
We conclude this section by computing the Schwarz symmetrization of

the function in Example 1.1.1.
Symmetrization 15
Example 1.3.1 Let ft = (-2,2) c E and let u : ft - be as in Example

1.1.1. Then ft* = (-2,2) as well and u* is given by
x, 0 < a; < 0.5

•x) = u*(x) — 1, 0.5 < x < 1
x, 1 < x < 2.
1.5
\
0.5 1
1.4 Variations on the Theme
In this section, we briefly describe some variants of the notions discussed

previously. Just as we defined the unidimensional decreasing rearrange-
ment, we can define the increasing rearrangement as well.
Definition 1.4.1 Let ft C be bounded and let u : ft be a

measurable function. The (unidimensional) increasing rearrangement

of u, denoted u#, is defined on [0, \Cl\] by
u#(|fi|) - e s s . sup(u)
V }
u # ( s ) = inf{t | \{u < t } | > s}, se [0, |H|). "'
As before, u# is essentially the inverse of the function m(t) = \{u < t}\.
It is non-decreasing, and equimeasurable with u. If u < v, where u and v
are measurable functions defined on Q, then u# #.
Exercise 1.4.1 Let u : Q —» R be measurable. Show that
u*{s) = Mini-a).
Proposition 1.4.1 Let u : 0, —> R be a measurable function. Then,
u* = - ( - u ) # a.e.. (1.4.2)
Proof: Notice that

\{u#>t}\= \{u>t}\ =\{-u<-t}\
= |{(-«)#<-*}| = |{-(-u)#>t}|.
Since u * and —(—u)# are both non-increasing and equimeasurable, they

are equal a.e. •
Corollary 1.4.1 (Hardy - Littlewood) Let f e LP(Q) and g £ LQ(Q)

where (1/p) + (l/q) = 1. Then
•\n\
/ /(x)5(x)di > / f#(s)g#(s)d8. (1.4.3)
Proof: We have
- f f(x)g(x)dx < f (-f)*(s)g#(s)ds = - [ U(s)g*(s)ds

Jo. Jn Jn.
and the result follows immediately. •
Definition 1.4.2 Let u : Q, —> R be a measurable function where Q C R ^

is bounded. Then the spherically symmetric and increasing rear-
rangement of u is defined oni]* by
u.(x) = « # K | x r ) . (1.4.4)
Symmetrization 17
Exercise 1.4.2 Compute u# and u* when u is as in the Example 1.1.1.
Exercise 1.4.3 Let / : fi —* R be a measurable function where Q C B>N is

bounded. Show that
(/+)#(*)(/~)#(s) = 0 for all 0 < s < |fi|.
The next variant we would like to discuss is the definition of the radially
symmetric and decreasing rearrangement of a function defined on all of RN.
Definition 1.4.3 Let / : R ^ —• R be a Borel measurable function. It

is said to vanish at infinity if, for every t > 0, the sets {|/| > t} are of
finite measure.
Definition 1.4.4 Let / : R w - > E b e a Borel measurable function van-

ishing at infinity. Then its spherically symmetric and decreasing re-
arrangement , /*, is defined on R ^ by
/•OO
/'W = / X{\f\>ty{x)dt. (1.4.5)

Jo
Since the integrand in (1.4.5) is in terms of characteristic functions of
balls, clearly, / * is radially symmetric. Also it is obvious that it is radially
decreasing, i.e. if x,y G R ^ are such that \x\ < \y\, then f*(x) > f*(y)-
Assume that f*(x) > t. Let, if possible, x f {\f\ > t}*. Then, for
all t' > t, x £ {\f\ > t'Y since these are smaller concentric balls. Thus,
it follows from (1.4.5) that f*(x) < t, which is a contradiction. Thus,
xe{|/|>t}'.
Conversely, if x G {|/| > t}*, then, clearly, t is not the supremum of | / |
and so we can find a t' > t such that x G {|/| > £'}* which is a smaller
concentric ball. Now, for all t" < t', it is evident that x G {|/| > t"}*.
Thus, it follows from (1.4.5) that f*(x) >t'>t. Thus, we have shown that
{/*><} = {\f\>ty.
Since the sets {/* > t) are all open, it follows that /* is measurable and
in fact, lower-semicontinuous. It also follows that /* and | / | are equimea-
surable and the results of the preceding sections follow as a consequence.
In particular, the Lp norms of / and /* are the same. Notice that we are
essentially symmetrizing | / | rather than / (cf. Remark 1.1.2).
It can be shown that the map / i-» /* is non-expansive from LP(RN) into
itself and that the Hardy - Littlewood inequality holds. Finally, we state
an important property of this rearrangement in R . We refer the reader

to the book of [Lieb and Loss (1997)] for the proof and for generalizations.
T h e o r e m 1.4.1 (Riesz' inequality) Let f,g and h be non-negative Borel
measurable functions on M.N vanishing at infinity. Define
I{f,9,h) = f(x)g{x-y)h(y)dxdy.
Then
I(f,9,h) < I(f%9\h*). (1.4.6)

Chapter 2
Some Classical Inequalities
2.1 The Isoperimetric Inequality
In this chapter, we will prove two important and very classical inequalities
- the isoperimetric inequality and the Polya - Szego inequality. The former
is one of the oldest results in the calculus of variations, solving a problem
that goes back to antiquity. The latter describes an important property
of rearrangements; in particular, it describes how rearrangements affect
certain Sobolev norms under certain conditions. The proof of the Polya -
Szego inequality uses the isoperimetric inequality and another important
result from geometric measure theory called the co-area formula. We will
prove particular cases of the co-area formula in the next section. In the
last section, we discuss the Sobolev inequality and its connection to the
isoperimetric inequality.
In this section, we will prove the isoperimetric inequality. This inequal-
ity, in R 2 , solves an ancient problem in the calculus of variations, known as
Dido's problem: of all simple closed plane curves of fixed length, find that
which encloses the maximum area. (Equivalent ly, of all simply connected
plane domains of fixed area, find that which has the least perimeter).
The solution to the above problem is that the circular disc , and the
disc alone, maximizes the enclosed area, given the perimeter, or minimizes
the perimeter, given the area.
There are two aspects to the proof of this statement. First, the existence
of an optimal domain must be proved and then it must be shown to be the
disc. An ingenious geometric proof was given by Steiner (cf. [Courant and
Robbins (1996)]). An analytic way of approaching this problem is via the
following inequality: given a plane domain with perimeter L and area At
19
we have
L2 > 4TTA (2.1.1)
Equality is attained only for the disc. This proves that, given the perime-
ter L, the maximum possible area that can be enclosed is L2/4ir which is
achieved by the disc, and the disc alone. This completely solves Dido's
problem.
R e m a r k 2.1.1 The inequality (2.1.1) is called the (classical) isoperimet-

ric inequality. The prefix 'iso' means same in Greek. Today an isoperi-
metric inequality refers to an inequality obtained when we optimize some
domain dependent functional keeping some geometric parameter (like the
measure, for instance) of the domain fixed. •
Several easy proofs of (2.1.1) are available. We refer the reader to the
survey article of [Osserman (1978)] for a nice proof. The inequality can
also easily be proved using Fourier series. We will defer a proof of this
inequality to later in this section, when we will give a simple proof (at least
for smooth domains) in all space dimensions.
The isoperimetric inequality in RN, i.e. the generalization of (2.1.1),
holds in the greatest possible generality, i.e. however irregular be the
boundary of the domain, but we need suitable definitions to even state
it. In the plane, if the boundary of a domain is not a rectifiable curve, we
can set its perimeter L — oo and (2.1.1) holds trivially. If the boundary
is rectifiable, then the length is the limit of the perimeters of approximat-
ing polygons and we can deduce (2.1.1) in the most general case from the
special case of polygonal domains. On the other hand, in higher dimen-
sions, there are several notions of 'surface area', each one being suited for
a specific purpose, although they all agree for smooth surfaces. In the cal-
culus of variations, one would like to assume the least possible regularity
and then deduce the smoothness of the domain as a consequence of some
extremal property. As [Osserman (1978)] puts it, 'it is only by examining
the consequences of a definition of surface area that one can decide on its
appropriateness or correctness. The validity of the isoperimetric inequality
is, in fact, one criterion that has been used.'
We briefly describe below some notions of surface area used in the lit-
erature. We assume, henceforth, that N >3.
Some Classical Inequalities 21
The Minkowski Content (cf. [Federer (1969)], [Osserman (1978)])
Let Ec1&N. For r > 0, define
Er = {xeRN | p(x,E) <r}
where p(x, E) — mi{\x — y\ \ y e E} is the distance of x from E. Let k be

an integer such that 1 < k < N — 1. Set
M fc (£) - liminf ^ .. (2.1.2)

*v ; r - o uN-krN~k K }
Mk(E) is called the k - dimensional Minkowski content of E. (If E is a

k - dimensional manifold, then the above definition is based on the idea
that, for small r, the volume of Er is roughly the k - dimensional measure
of E times the 'cross - sectional area' (given by the volume of the (N — k)
- dimensional ball of radius r, lying in the (N — k) - dimensional space
orthogonal to the k - dimensional tangent space at each point of E).
Example 2.1.1 Let E = S^~x be the sphere of radius p centered at the

origin in M^. For r > 0,
Er = B(0;p + r ) \ B ( 0 ; p - r )
where J5(0; r) denotes the open ball centered at the origin and of radius r.
Hence
\Er\ = uN{p+r)N-LON(p~r)N = UN(2Np

N 1
-r + 2(^pN-3r3 +
Thus,
MN-i(S?-L) - liminf ^ - = liminf ^ ^ = NuNp

H
r—*0 oJir r—*0 2r
(since uj\ = 2), which indeed the surface area of the sphere of radius p
Hausdorff Measures (cf. [Evans and Gariepy (1992)])
Let E C M.N and let 0 < s < oo and 0 < 6 < oo. Define
Hl(E) = infj5>(s) diam(Cj) | E c U°°=1Cj, diam(C,-) < 5 \
(2.1.3)
where diam(C) denotes the diameter of a subset C and
Notice that LJ(N) = u>#, the volume of the unit ball in UN (cf. Section
1.3). Further, define
H3{E) = limHs6(E) = supHi(E). (2.1.4)

*-*° <5>0
It can be shown that Hs defines a regular (outer) measure on Borel sets
and is called the s- dimensional Hausdorff measure of E. It can be shown
that H° is the counting measure and that on RN, HN is the Lebesgue
measure. Further, Hs = 0 on RN for all s > TV, It can also be shown that,
if E C RN and if 0 < s < t < oo, H3{E) < oo implies that H*(E) = 0
and that, if ^(E) > 0, then HS(E) = oo. In view of this, the Hausdorff
dimension of a subset E C R is defined as
inf G
^dim(^) = {* [°>°°) I WW = °>- (2-1-5)
Notice that the Haudsorff dimension of E is at most TV and if it is equal
to s, then HÊ) = 0 for all i > s and HÊ) = oo for all t < s. The
Hausdorff dimension may be any real number in the closed interval [0, oo].
If E is a smooth (TV - 1)- dimensional manifold, then HN~1(E) agrees with
the usual (TV — 1) - dimensional surface area.
The de Giorgi Perimeter (cf. [de Giorgi (1954)], [Mossino (1984)])
Let O C RN be an open set and let E C ft be a measurable set. The

de Giorgi perimeter of E with respect to Q, denoted Pî(E), is defined as
the total variation of the characteristic function \E of E. In mathematical
terms,
Pn(E) = Sup{^j§fZ : SefZW,*?^}

(2.1.6)
where
/ N
| * | | 2 = max £>i(s)|a , * = (*I,...,¥>N).
A=\
One could also interpret PQ(E) as the surface area of that part of the bound-
ary of E contained in H where a normal can be unambiguously defined. If
H were smooth, then, by the divergence theorem,
/ div($)dx = / $.vda
JQ JdQ
where v is the unit outer normal on d£l and thus P R w(0) would be the
'usual' surface area of d£l.
We also have the following property of the de Giorgi perimeter:
P<i(E) = Pn{Q\E).
For,
/ div($)dx = / div($)dx - / div($)dx = - I div($)dx
JQ\E J to JE JE
if $ e {V{Q))N.
In each of the above figures, if Q is the outer domain containing £ , then

PQ(E) will be the length of the curve C.
Finally, if E C U C ft', we have
PQ(E) < Pci'(E)
as is immediate to verify.
We are now in a position to state and prove the (classical) isoperimentric
inequality in all dimensions N > 2. Let Q C RN be a bounded domain. Let
us denote by |dS"2|;v-i, a suitable (2V — 1) - dimensional surface measure of
the boundary dfl. Then the classical isoperimetric inequality in l w reads
|0fi|jv-i > i V ^ i n i 1 " ^ (2.1.7)

where LO^ is the volume of the unit ball in MN (cf. Section 1.3). Equality
occurs only for the ball.
Usually, the proof in higher dimensions relies on the Brunn - Minkowski
inequality which reads as follows: if A and B are subsets of HN and
A + B = {x + y\xeA, y€B},
then
\A + B\& > |i4|* + | B | *

which is a result from geometric measure theory (cf. [Federer (1969)]). The
equality case is even more difficult to prove (cf. [Brothers (1987)]).
We will now present a fairly elementary proof of (2.1.7) in the case of
smooth domains. The author learnt of this proof from a lecture given by X.
Cabre (see also [Kesavan (2002)]). It uses simple ideas from calculus and
partial differential equations (Neumann problem for the Laplace operator)
and is based on Alexandrov's method of moving planes.
Definition 2.1.1 Let ft C RN be a bounded domain and let u G Cl(fl).
The lower contact set of u is the set defined by
S(u) = {xeft\u{y)>u(x) + Vu(x).(y~x) for all y G ft}. (2.1.8)

In other words, the lower contact set is the set of all points in ft such
that the (graph of) u lies entirely above the tangent plane at the point
(x, u{x)). If, in addition, we also have that u G C2(fl), then it is easy to see
that, if x G S(u), then, the Hessian D2{u){x) of second partial derivatives
is a (symmetric) positive semi-definite matrix, i.e. for all £ G RN, we have
-P2 (#)£•£ > 0- For, if t is small enough such that x + t£ G ft, then
t2
u(x + t£) = u{x) + tVu(x)4+ — D2u(x)£.£ + o(t2)
and since x G S(u), it follows that
t2D2u(x)£4 + o{t2) > 0

and the conclusion follows on dividing by t2 and letting t —> 0.
Let us consider a plane with normal m lying below the graph of u and
moving parallel to itself. It will ultimately make contact with the graph of
u and let us assume that the first point of contact, (xo, U(XQ)) is such that
XQ G ft. Then, for all x G ft, we have
u{x) > U(XQ) + m.(x — XQ)

i.e. the function g(x) — u(x) — U(XQ) ~ m.(x — XQ) is non-negative and
attains its minimum, viz. zero, at XQ G 0. Thus, Vg(xo) = 0 which means
that m = VU(XQ). Thus, the plane, in this position, is a tangent to u and,
as u lies entirely above it, it follows that xo G S(u) and s o m G Vu(5(u)).
Let us now consider the following Neumann problem for the Laplacian.
Au = c in SI I .
where c is a constant. By the usual compatibility condition for the existence

of a solution, it follows that
c = J-J-L. (2.1.10)
Of course, the solution is unique up to an additive constant and we can

consider any such solution, which will be in C2(Q) if O is sufficiently smooth.
Lemma 2.1.1 (Cabre) Let Q, C MN be a smooth bounded domain and let
u be a solution of (2.1.9). If B(0; 1) is the open unit ball in R ^ (centered
at the origin), then,
B(0;1) C Vu(S(u)). (2.1.11)
Proof: Let us consider a plane with normal m lying below the graph of
u and moving parallel to itself, assume that its first point of contact with
u is (XQ,U(XQ)) where XQ G dQ. Thus, if g were denned as before, it is a
non-negative function attaining its minimum on the boundary of H, viz. at
XQ. Thus, if v is the unit outer normal on dfi, we have that |j(a:o) < 0.
Hence,
d u
^ i \
rn.v > —(XQ) = 1i
oi/
by virtue of (2.1.9). In particular, it follows that \m\ > 1. Consequently,
a plane with normal m, lying below u and moving parallel to itself must
necessarily first touch u at an 'interior point' if \m\ < 1 and if the point
of contact is (cco)U(xo) with XQ G S~2, we saw that m — VU(XQ) and that
XQ G S'(U). This proves the lemma. •
We can now easily prove the isoperimetric inequality (2.1.7) for smooth
domains.
Theorem 2.1.1 Let Q C RN, N > 2, be a sufficiently smooth domain
and dQ, its boundary. Then (2.1.7) holds.
P r o o f : In view of (2.1.11), we have
u>N = | S ( 0 ; 1 ) | < | V u ( 5 ( u ) ) | = f
JVu(S(u
By the change of variable formula, we get
/ dx < J \det(D2(u{x))\dx.
JVu(S(u)) JS{u)
We do not have equality since Vi/ may not be a diffeomorphism. Now, if x G
S(u), we saw t h a t D2u{x) is a symmetric and positive semi-definite matrix.
T h u s , its eigenvalues are all non-negative and hence so is its determinant.
Further, by the classical inequality between the arithmetic and geometric
means (AM-GM inequality), we also get t h a t
2
, ,~2 / n ftr(D u(x))
det(D2u(x)) < ^ V
N
K )}
where tv(A) stands for t h e trace of a matrix A. B u t tv(D2u{x)) — Au(x) =

c where c is given by (2.1.10). Thus,
"N - L {-Rm-) dx
- L \TNW) dx =
NWT^
which immediately yields (2.1.7). •
If fi is a ball of radius R, we have \8Q\N-I = NLONRN~1 while |fi| —

N
LUNR . Thus, equality is attained in (2.1.7) for a ball and so, of all domains
of fixed volume, the ball has the least surface area.
T h e o r e m 2.1.2 Let Q C M.N be as in the preceding theorem. If equality

is attained in (2.1.7), then Q, (assumed to be smooth) must be a ball.
Tracing back the proof of (2.1.7), we observe the following:
uN - | V u ( 5 ( u ) ) | - fs, sdetD2u(x)dx =
fea is a
= / f l M (^ )^ = / n ( ^ r^
Since the integrand in the last two integrals is a constant, it follows t h a t
| f i \ 5 ( u ) | — 0 and so S(u) is dense in H. B u t by the definition of S(u) (cf.
(2.1.8)), and t h e fact t h a t u £ C 2 (H), it follows t h a t S(u) = Q.
Next, we also deduce that det{D2u(x)) = (tr(D 2 u(x))/N) N and so

we have equality in the AM-GM inequality applied to the eigenvalues of
D2u(x), Thus, all the eigenvalues are equal and thus, since D2u(x) is also
symmetric, it follows that D2u(x) = X(x)I, where I is the identity matrix.
Also, X(x) = tr{D2u(x))/N = \d£l\#-i/N]£l\ = X which is independent of
xeil. Thus,
D2u(x) = XI, for all x € fi. (2.1.12)
Finally, |Vit(5(u))| = |B(0;1)| and so B(0; 1) is dense in Vu(S(u)).

Hence, for all x G Q, we deduce that |Vu(x)| < 1. But Vu.v = | j = 1 on
c?fi. Thus, if r is a tangent vector on the boundary of fi, we have Vu.r — 0.
Thus, u is constant on d£l,
We can therefore assume that the solution u that we are considering
satisfies the following:
u= 0 on dU \ (2.1.13)
^ = 1 on dQ. )
By the maximum principle, it follows that u < 0 in Q and so u must attain
a minimum — M at a point zo G Q, where M > 0. Let B be the largest
possible open ball contained in 0, with centre at xo. Then, dBndfl / 0.
Since B is convex, by Taylor's theorem, we can write (for x G B)
u(x) = u(x0) + VU(X0).{X~XQ) + -D2u(£,){x - X0).(X - XO)
where £ is a point in the line segment connecting x and XQ. Using (2.1.12)
and the fact that U(XQ) = — M and that VU{XQ) = 0, we get
u(x) = -M+-\x~xQ\2
for all x G B and hence for all x G B.

If x G & B n<9Q, then
0 = - M + - | x - x0|2. (2.1.14)
But, since u(x), for x G B, depends only on |a; —xo|, it follows that u(x) = 0
for all x G 3B. Now, since u < 0 in £7, we deduce then that dB = dfi and
so H = B, the ball centered at x$ and of radius y ' x " - This completes the
proof. •
2.2 The Co-area Formula
In this section, we will prove some special cases of the co-area formula,
which is one of the fundamental results of geometric measure theory (cf.
[Federer (1969)] or [Evans and Gariepy (1992)]). For our purposes, these
special cases will suffice. The general result states that if u : MN —• R is
Lipschitz continuous and if g : M.N —» R is integrable, where N > 2, then,
/ g(x)\Vu{x)\dx = i If g(x)d<j(x) dt (2.2.1)

JRN J-OO \J{u=t}
where da{x) stands for integration with respect to an appropriate (N — 1)

- dimensional measure on the level set {u = t}.
Our first result will deal with the special case when g = 1.
Theorem 2.2.1 (Fleming - Rischell) Let Q C MN be an open set and let
l i G ^ f f i ) . Then,
/ \Vu\dx = / Pn({u>t})dt. (2.2.2)

JQ J-OO
Proof: Let <S> £ (V(Q))N. Then, by Lemma 1.2.1,
/ u div($)dx = / ( / 6(t,a;)div($)dx) dt
JQ J-OO \JQ J
after an application of Fubini's theorem. Using the definition of &(.,.), we
get
r fQ r r+oo r
/ u d\v(§)dx = - / / d\v($)dxdt + / / d\v($)dxdt.
JQ J-OO JFt Jo JEt
But, since $ e (V(Q))N, we have JF div($)dx — - JE div($)dx and so
/ u div($)dx ( = div($)dx) dt
JQ \Jst J-OOJ
and the result follows on taking the supremum over all such $ ^ 0. •
Remark 2.2.1 If u : Q. —> R is an integrable function and if the integral on
the right-hand side of (2.2.2) is finite, then, for any $ £ (D(£l))N, we have
/ u d\\($)dx < C||*||.

JQ
It then follows that u e W x , 1 (n) and then JQ \Vu\dx is given by (2.2.2).

Corollary 2.2.1 IfuE W 1,1 (fi) is such that u>to a.e., then
/ \Vu\dx = / PQ{{u>t})dt. (2.2.3)

12 J to
Proof: If t 0 < 0, then b(t,x) = 0 for t < t 0 and if t 0 > 0, then
,, , Jlfor0<t<t0
b{t x) =
' \ofort<0
for all x G f l . Thus,
r+oo
u(x) = / b(t,x)dt if t 0 < 0
Jtn
and
u(x) = £ 0 + / b(t,x)dt if t 0 > 0

./to
and the result now follows as in the proof of the previous theorem. •
Corollary 2.2.2 Let u G Wl^{Q). Then
Pn({u>t}) = - j ( j \Vu\dx\. (2.2.4)
Proof: Set v = (u - t)+ + t. Then v > t and so, by (2.2.3), we have

r r+oo
/ \Vv\dx = / Pn({v>r})dr. (2.2.5)
Jn Jt
Let r > t. If v > r, then, a fortiori, v > t and so (u —1) + > 0 which implies
that u > t as well. Then, (u — t ) + = u — t and s o v ^ u which means that
u > T as well. The converse is immediate. Thus, for all r > t, we have
{v > r } = {u > r } .
On the set {it < t}, we have that v = t and so, Vv = 0 a.e. on this set.
Thus,
/ \Vv\dx = / \Vv\dx + / \Wv\dx = / |Vu|da\

Jfl J{u<t} J{u>t} J{u>t}
Substituting this in (2.2.5) and differentiating both sides with respect to t,
we get (2.2.4). •
Corollary 2.2.3 Let ft be a bounded domain and let u G WQ' (O) be such
that u > 0 a.e. inU. Then, for t > 0,
-T- I / |V«|da:| > NLO»\{U > t } | 1 - ^ - (2.2.6)

dt
\J{u>t} J
Proof: We can extend u by zero outside Q and then the resulting function,
u, will be an element of W1^1(RN). If t > 0, then,
{ x e R ^ I u(x) >t} = {xen\ u(x) > t}.

Now,
/ \Vu\dx = f Pa({u>t})dt
JQ JO
and
/ |Vu|dx = / PRN({u>t})dt = / PRN({u>t})dt

JRN JO JO
But the integrals on the left-hand sides of the above relations are equal.
Further,
Pn({«>*}) <iW{«>*})
and so we have
P^{{u>t}) = Pn({u>t})
for almost every t > 0 and hence for all t > 0 by a density argument using
the definition of the perimeter. The result now follows from the classical
isoperimetric inequality (2.1.7). •
Lemma 2.2.1 Let 1 < p, q < c© be such that (l/p) + (l/<?) = 1 and let
f G LP(Q) and g G Lq(Q, where Q C RN is an open set. Set
F{t) = [ g(f-t)dx.
Then
F'(t) = - / gdx. (2.2.7)

Proof: Define h — (/ - £) + + t, for t e R . Then h > t and so by Lemma

1.2.2,
-t-oo
/ ghdx = t I gdx 4- / / #d:c dr.
in in Jt \J{h>T} J
On the other hand, by the definition of h, we also have
/ ghdx = t / gdx-\- I g(f -t)+dx = t / gdx + / g(f~t)dx.

in in in in J{f>t}
Comparing the two relations, we get
/ g(f~t)dx= [ If gdx) dr. (2.2.8)

i{/>i} it \i{/i>r} /
Now, arguing exactly as in the proof of Corollary 2.2.2, we easily see that,
for all r > t,
{h>r} = {/>r}.
Using this in (2.2.8) and differentiating both sides of that relation with
respect to t, we get (2.2.7). •
We now prove another particular case of the co-area formula.
Theorem 2.2.2 Let Q C RN be an open set and let u £ £>(fi). Ifu>0,
then, for 1 < p < oof we have
[ \Vu\pdx = [ \[ iVu^da) dt (2.2.9)

in Jo \J{u=t} J
where M = max a . G û(x). Further, if u* : ft* —» R is its Schwarz sym-
metrization, then we also have
[ \Vu*\pdx < [ if \Vu*\p-lda)dt (2.2.10)

in- Jo \J{u*=t} J
Proof: Step 1. Since u is smooth, by Sard's thoerem, for almost every t in
the range of u, we have that |Vu| ^ 0 on the level set {« = £}. Thus, {u = t}
will be an (N — 1) - dimensional surface and, further {u = t} = d{u > t).
Also, \{u* = t}\ = |{w = t}| = 0.
Step 2. Let 2 0. Set v = (u-1)+ € WQ'P{Q). Using this test function in (2.2.11),
and observing that v is supported on the set {u > t} and that on this set
v = u — t, we get
/ \Vu\vdx = [ f{u-t)dx.
J{u>t} J{u>t}
Thus, differentiating with respect to t, and using Lemma 2.2.1, we get
- ^ I / \Vu\*dx) = f fdx. (2.2.12)

at \J{u>t} J J{u>t)
Integrating this relation over the range of u, i.e. on [0, M], we get
f \Vu\pdx = [ (f fdx)dt. (2.2.13)

JO. J0 \J{u>t} J
Now, for almost every t in the interval [0, Af], the properties mentioned
earlier a t the beginning of this proof hold. For such a t, by the definition
of / a n d by Green's theorem, we get,
f fdx = - [ \Vu\p-2Vu.vda = [ |Vu| p-1 d<7

J{u>t} J{u=t} J{u=t)
since, on the set {u = t}, the tangential derivatives of u vanish and, since
u > t inside this surface,we have — VUM = |Vu|. Substituting this in
(2.2.13), we get (2.2.9).
Step 3. Let 1 0. Define
fe - -dwfdVup+e^Vu
so that, for all v £ Wo' p (ft), we have
f(\Vu\2 + e)E?1Vu.Vvdx = [ fevdx. (2.2.14)

Jn Jn
Once again, we set v = (u — t)+ for t > 0 and go through the steps as in
the previous case to get
/"(iVuf + e ^ l V u f d x = f ( f {\Vu\2 + e)^\Vu\da]dt

Jn Jo \J{u=t} J
(2.2.15)
We now pass to the limit as £ —> 0 on both sides of (2.2.15). On the left-hand
side, the integrand coverges pointwise to |Vu| p ; further, since 1 < p < 2,
we have
1-y
12 \ 2
2 2
(|w| + e )^|v«| = (|^"' +e ) ' iv«|' < |vu|"
which is integrable. Thus, by the dominated convergence theorem, we get
lim / ( i V u f + e ^ l V i i f d x = f \Vu\pdx (2.2.16)

e >0
- 7n Ja
On the right-hand side, we once again observe that
(IVuf + e J ^ I V u l < \Vu\v~1

which is integrable on the set {u = t}. So, once again, by the dominated
convergence theorem,
/ (|Vu| 2 + e ) ^ | V u | d a ^ f \Vu\p-xda
J{u=t} J{u=t}
as e —• 0. Further,
/ (IVup + e J ^ I V u l d f f ^ / |Vu| p_1 dc7

J{u=t) J{u=t}
and
p l
/ f \Vu\'" _~1 dadt < oo
"""
JO J{u=t}
since u € T>(Q,). Thus another application of the dominated convergence

theorem yields
lim / ( / (IVw^ + e J ^ l V u l d c r J dt = f ( f IVU^"1^ ) dt
(2.2.17)
and the result follows from the relations (2.2.15) - (2.2.17).
Step 4. Finally, we consider the rearrangement u* of u. Let R be the

radius of Q*. Let us write (by abuse of notation) u*(x) = u*(\x\) and thus
consider it as a function of a single variable, for convenience. Since it is
monotonically decreasing, it is differentiable almost everywhere and using
polar coordinates, we get
/ n , \Vu*\?dx = / * \u*'(r)\pNu>NrN-1dr
= Jo* \u*' {r^N^r"'1^™*' (r))dr
</o M i^, { u , = ( t r l iK-û-i^
= /oM(/{«-t}iv«-rl^)rf«
using the change of variable t — u*(r), observing that — u* is monotonic
increasing (which is why we only have an inequality in the above set of
relations) and using the fact that the gradient of u* is constant on each of
the sets {u* — t}. •
Theorem 2.2.3 Let u G V(Q) be such that u > 0. Let n denote the
distribution function of u. Then, for almost every t in the range of u, we
have
-V(t) = f # T = / # r . (2.2.18)
where u* is the Schwarz symmetrization of u.
Proof: Step 1. As in the proof of the previous theorem, for almost every
i, we have the properties stated in Step 1 of the earlier proof.
Step 2. Let e > 0. Define
/
= "div {^WTl) •
Multiplying by (u — t)+ and integrating by parts, we get, as before,
/ J^S-dx = f f{u-t)dz.
2
J{u>t} | V u | + £ J{u>t}
Differentiating with respect to t, we deduce that
d_ f V 12
\Vu\
dt f l <
dx = fdxf
J{u>t}\Vu\* + e J{u>t}J
Step 3. Let t be such that |Vu| ^ 0 on the set {u = t}. For sufficiently small
h > 0, the same holds for the set {t — h T} ) Jt-h \J{u=r} |Vu|2+£ a

)
using the definition of / . Once again, applying the dominated convergence

theorem, we can pass to the limit as e —* 0 to get
d<T
fx(t - h) - fi{t) = J dx = f [
J{t-h<u<t} Jt-h J{U=T}
and the first relation in (2.2.18) follows by dividing by h and taking the
limit as h —> 0.
Step 4. Let r(t) be the radius of the ball {u* > t}. Then /i(t) = toN(r(t))N
and so, fi'(t) = N(jN(r(t))N~lr'(t). Notice that fj,(t) and r(t) are mono-
tonically decreasing functions and hence are differentiable for almost every
t. But u*(r(t)) = t for almost every t (where we now write, again by abuse
of notation, u*(x) = u*(|x|). Thus, we get, by implicit differentiation,
since u* is decreasing and the second relation in (2.2.18) follows. •
2.3 The Polya - Szego Theorem
We will now prove one of the most important properties of Schwarz sym-
metrization.
Theorem 2.3.1 (Polya - Szego) Let 1 0. Then
[ \Vu*\pdx < f \Vu\pdx. (2.3.1)

Jn* Jo.
In particular, u* G WQ'P(Q*).
Proof: Step 1. The case p = 1.

Since u > 0 in H and u = 0 on dft, we have seen that PQ({U > t}) =
Pu*({u > t}) for t > 0 (cf. Corollary 2.2.3). Further, by the classical
isoperimetric inequality, we also have
PRs({u>t}) >PR*({u*>t}).
Thus, from the Fleming - Rischell theorem (Theorem 2.2.1) and Remark
2.2.1, it follows that
/ \Vu\dx = / PRN({u>t})dt
Jn Jo
/•OO
•oo /•*
> / PRN({U* >t})dt = / \Vu*\dx

Jo Jo.
Step 2. Let 1 0. Let M =
maxxeû(x). By virtue of Theorem 2.2.2, it suffices to show that, for
almost every t G (0, M),
/ \Vu\p-xda > f \Vu*\p~lda. (2.3.2)

J{u=t} J{u*=t}
Since u is smooth, we can assume, by Sard's theorem, that |Vu| does not
vanish on the set {u = t] for almost every t G (0, M). Thus the properties
enunciated in Step 1 of the proof of Theorem 2.2.1 are valid. Let us define
a measure v on {u = t] by dv = da/\Vu\. By Holder's inequality, we have
/ \Vu\du < [ [ \Vu\pdv) \ [ dv)

J{u=t} \J{u^t} J \J{u=t} J
Thus, taking into account the definition of v we get
I {««*}
\Vu\p-ld<7 >
" ( / { ^ } ^ P
" 1
U{u=t}^)P-1
But by the classical isoperimetric inequality, we have,
/ da > I da.
J{u=t) J{u*=t}
Combining this with (2.2.18), we get
/{u=t> i ^ r 1 ^ > T';w'P ) P = -"'WIV«I<..=.> \p
Thus, we have proved (2.3.1) when u G V(Q) and u > 0.
Step 3. Let 1 0. Then,
there exists a sequence {un} such that un G T>(Q), un > 0 and u n —> ii in
W 0 llP (fi). By Step 2,
/ |V<|p<to < / \Vun\pdx.

JQ* JQ
Hence the sequence {u*n} is bounded in W0 ,P(Q,*) and as 1 u* in
LP(Q*). Thus, we deduce that u* G WQ'P(Q*) and that u$ ->• u* weakly in
that space. By the weak lower semi-continuity of the norm, we get
f \Vu*\pdx < liminf f \Vvl\pdx

J a- «-°° JQ-
< liminf
n
f \Vun\pdx = / \Vu\pdx.
^ ° ° JQ JQ
This completes the proof. •
Remark 2.3.1 If we already knew that u* G W0,P(Q*), then we can also

prove (2.3.1) as follows. Let u be a smooth and non-negative function with
maximum M. Set tp(t) — Jtu>ty \Vu\pdx for t > 0. Then
/ \Vu\pdx = - / ^'(t)dt.
Now, by Holder's inequality,
f \Vu\dx < ( / \Vu\pdx) ( [ dx

J{t-h<u<t) \J{t-h<u<t} J \J{t-h<u<t}
which, on letting h —> 0, gives,
~~ ( \Vu\dx < ( V W ) * ( V W ) ^
dt
J{u>t}
where fi is the distribution function of u. Combining this with the result
of Corollary 2.2.3, we get
-*/(*) > (N4r(»(t)r-%/(-»'(t)r-\

Thus,
Jn^dx > {N4r J"(^r*- (-3-3)

On the other hand, using (2.2.10) and (2.2.18), and using the fact that we
have equality in the isoperimetric inequality for the sphere {u* — £}, we
get
j^ \vu*\*dx < j™ ivu^^r 1 !^* = t}\N^dt

fM \{u* = t}\pN ,
dt (234)
=i hm^ --
Thus (2.3.1) follows for smooth and non-negative u from (2.3.3) and (2.3.4).
Remark 2.3.2 When p = 2, we can prove (2.3.1) by an entirely different

method using neither the isoperimetric inequality nor the co-area formula.
This proof is due to [Lieb (1977)].
By extending u by zero outside its domain, we can consider it as a func-
tion in /f 1 (R j/v ). If u € H 1 (R JV ), then we can define its Fourier transform
u. For t > 0, we set
Then, by the properties of the Fourier transform, we get

1
f t 2 IT, \ \j f (l-exp(-47r 2 [e| 2 t),^, 2 ^
- / (u*-(Kt*u)u)dx = / |u(OI *
where Kt * u denotes the convolution product of Kt and u. Now, the

integrand on the right-hand side converges pointwise, as t —» 0, to
4^\2\u(0\2-
Further, since 1 — exp(—at) < at, we have that the integrand is bounded,
for all t, by the same function. This function is integrable for u G HX{R.N)
and, in fact,
/ \vu\2dx = [ 47r2|£|2|e(£)|2d£.
Thus, by the dominated convergence theorem, we get
/ \Vu\2dx = limi / (u2 - (Kt * u)u)dx (2.3.5)
for u G /f 1 (R j V ). Further, since (1 — exp(—at))/t —» a as t —> 0, we have

that (1 - exp(-at))/t > a/2 for sufficiently small t. Thus, if the limit on
the right-hand side of (2.3.5) exists, it follows that u G Hl(JlN) and the
semi-norm is given by this relation.
Now, we know that the Schwarz symmetrization preserves the L2 - norm.
Further, by Riesz' inequality (cf. Theorem 1.4.1) applied to the functions
i/, Kt and u, we get
j R N JRN exp (" | J 4 7 y | j u(y)u{x) dydx <

^ IR^ JRNexp(^^^ju*(y)u*(x)dydx.
It now follows, from (2.3.5) and its corresponding version for u*, that u* €
H1^**) and that
/ \Vu*\2dx < [ \Vu\2dx.

Jn* Jn
R e m a r k 2.3.3 Let u G V{Q) be such that u > 0. If F : R - • R is any

convex and non-decreasing function, then one can show that
/ F(|Vw*|)dr < / F{\Vu \)da

J{u*=t} J{u=t}
for almost every t. We need to use Jensen's inequality in place of Holder's
inequality in the proof. If, in addition, G : R —> R is a non-negative and
bounded Borel function, we can also show that
[ G(u*)F(\Vu*\)dx 0, then for any 0 < a < |H|, we have
that u# e W 1 , p ((a, |H|)). In particular, v# will be absolutely continuous
on any such interval (a, |fi|), a > 0. Thus, all the discontinuities of u # are
concentrated at zero. To see this, we have that u* € WQ'P(Q*) and so, if R
is the radius of S7*,
/ \Vu*\pdx = [ NwNrN-1(\u#'{u>NrN)\NwNrN-l)pdr.
Jn* Jo
Making a change of variable s = uj^rN', we get
r i /"|n|
/ \Vu*\pdx = {NUJ$)P / \u# (s)\psp-%ds
JQ' JO
and so the integral on the right-hand side of the above relation is finite.
Thus u*'(s)s1-^ €L*((0,|n|)). F o r a > 0 , we know that s1 N is bounded
away from zero. Thus we deduce that w* G Lp((a, |H|)) and the result fol-
lows. •
Remark 2.3.5 Let 1 < p < oo. We saw, in Section 1.4, that we can define
the symmetric decreasing rearrangement for certain functions defined on
RN. In particular, it is defined for functions in W1,P(RN) and the Polya
- Szego theorem is still true (in this case we do not need to specify that
the function is non-negative, since the rearrangement essentially deals with
the absolute value of the function concerned). While the rearrangement
is continuous from LP(WN) into itself, the same is not true for W 1,P (R JV )
when N > 2. [Almgren and Lieb (1989)] showed that the continuity of the
rearrangement mapping is restricted to a class of functions which possess a
property that they call co-area regularity and that both co-area regular and
irregular functions are dense in Whp(RN). However, [Coron (1984)] has
shown that the rearrangement mapping is indeed continuous from W1,P(M)
into itself. •
Let us now look at the case when p = oo. The result can be stated as
follows.
T h e o r e m 2.3.2 Let u>0 vanish on dQ, and be Lipschitz continuous with

Lipschitz constant L. Then u* is also Lipschitz continuous with Lipschitz
constant less than, or equal to, L.
P r o o f : If u is as in the statement of t h e theorem, then it is differentiable

almost everywhere (Rademacher's theorem; cf. [Evans and Gariepy (1992)])
and further |Vu| < L almost everywhere. Thus u is non-negative and is in
all t h e spaces W 0 1 , p (n). T h e n we saw in t h e preceding remark t h a t u * is
continuous in the interval (0, |fi|). Let s € (0, |fi|) and set t = u#($). Let
u # ( s + k) — t — h. If fi is the distribution function of u, then
/ \Vu\dx < L(n(t-h)-n(t)).

J{t-h<u<t]
On the other hand, using Theorem 2.2.1, t h e fact t h a t , for t > 0, we have
Pn({u > t}) = P^N({U > £}), and the classical isoperimetric inequality, we
get
I{t-h<u<t} \Vu\dx = St-h P

MU > r
))dr
>^H-h{^r))l'Ur
>Nu^h{fi{t))1-^.
Thus
1+N
u*(s) — u#(s -f k) < —ks
Nu3
and we get
0 < -**L < _ ^ 5 - i +

ds
Nu£
T h u s , if \x\ > \y\, we have 0 < u*(x) < u*(y) and

f"N\x\N #
dd u #
u*(y)-u*(x) = / —T-da < L(\x\-\y\) < L|x-y|.
>N\y\r
This completes the proof.
We conclude this section by considering the case when equality is at-

tained in t h e Polya - Szego inequality (2.3.1). Obviously, if fi is a ball and
if u is radially symmetric and decreasing (after a translation of the origin
to the centre of this ball), i.e. u = u*, we do have equality in (2.3.1). A

natural question is to ask if the converse is true, i.e. if equality in (2.3.1)
implies that H is a ball and if, after a translation of the origin, u = u*.
Unfortunately this is not true.
Example 2.3.1 Consider ft = (-2,2) C R and let u be as in Example

1.1.1. Its Schwarz symmetrization was computed in Example 1.3.1. It is
immediate to check that in this case
J \u'{x)\2dx = f \u*'\2dx = 3
but uû*.
We state below a theorem of [Brothers and Ziemer (1988)] which gives

a necessary condition for equality in (2.3.1) to imply that u = u*.
Theorem 2.3.3 Let 0 < u < M, and let u e WQ'P(Q) for 1 < p < oo.
Assume that
\{x e 0* | 0 < u*{x) < M, Vu*(x) = 0}| = 0. (2.3.6)
Then, if we have equality in (2.3.1), we also have that Q = ft* (up to a

translation of the origin) and u = u*.
The proof of this result is quite long and uses a lot of machinery from
geometric measure theory.
Notice that in Example 2.3.1, we have that u* = 0 on the set
(—1,-1/2) U (1/2,1) which has unit measure. On this interval u* = 1
which lies strictly between the maximum and minimum values of u. Thus
the condition (2.3.6) is not met.
Remark 2.3.6 By virtue of the co-area formula (cf. (2.2.18)), we can show
that the condition (2.3.6) is equivalent to the absolute continuity of the
distribution function of u*. The corresponding condition for u also implies
the absolute continuity of the distribution function. However, if u were not
radially symmetric, it is not known whether the absolute continuity of the
distribution function implies that (2.3.6) (with u replacing u*) is valid. •
2.4 Sobolev's Inequality
Let 1 < p < oo and let Q, C RN be an open set. Then, we have vari-
ous imbedding theorems for the Sobolev space Whp{Q) (cf. for instance,
[Adams (1975)] or [Kesavan (1989)]). The main result states that if p < TV,
then WltP(RN) C LP*{RN) where
1
= --L (2.4.1)
K }
p* p N
The above inclusion is continuous and we have the following inequality,
known as Sobolev's inequality: there exists a constant C > 0 such that, for
all u e W ^ R * ) ,
||Vu|| PiK * > C||U|| P . ( R AT. (2.4.2)
Since the extension by zero outside a domain 0, provides a continuous

inclusion W01,p(ft) <-> Whp(RN), we still have the inequality (2.4.2) valid for
functions in W0'p(fl) (where the norms are replaced by the corresponding
norms over Q).
A simple dimensional analysis tells us how the exponent p* occurs in a
natural fashion. Let A > 0. If u G W^P{RN), define
u\(x) = u(Xx).
Then, a routine computation yields
"„ Vtt ^' B " = X-^f-T )JRfc (2.4.3)
Thus, if an inequality of the type (2.4.2) holds for all functions in W 1 , p (R A r ),

then, necessarily, the exponent of A in the relation (2.4.3) must vanish and
so we get the value of p* given by (2.4.1).
The best constant in (2.4.2) is given by
5= inf Jj^fc (2.4.4)

This constant is called the Sobolev constant and has several interesting
properties. It also plays a crucial role in the solution of several semilinear
elliptic partial differential equations where the right-hand side has a 'critical
growth' (cf. [Brezis and Nirenberg (1983)]). It can be seen easily, using the
same scaling argument used earlier, that the same constant works for all
domains Q C RN.
It is of interest to calculate this constant exactly and also see if it can

be attained for some function u £ W1,P(RN). This was done by [Aubin
(1976)] and [Talenti (1976a)]. We sketch briefly the arguments leading to
this, using symmetrization results.
First of all, since u and \u\ have the same Lp and Sobolev norms, it
is enough to look for non-negative functions satisfying (2.4.4). Now, if
u e W1,P(RN) is non-negative, it vanishes at infinity. So its Schwarz sym-
metrization can be defined (cf. Section 1.4). The Polya - Szego theorem
still holds (we can approximate by non-negative functions from V(RN)).
Then, since all Lp norms are preserved by symmetrization, it follows that
J(u) > J{u*) where J is given by
m = f^.
Hence, it now suffices to look for u which is non-negative, radial, radially de-
creasing and vanishing at infinity when seeking to minimize J. Converting
the corresponding integrals in the norms defining J using polar coordinates,
we now have essentially a minimization problem involving functions of only
one variable r — \x\, [Talenti (1976a)] showed that the minimum is attained
only for functions of the form
P
u(x) = ( W & | X | F M (2.4.5)
where a and b are positive constants. Further,
s _ ^N, ^__j ( r(f + i)r(A0 j
when 1 < p < oo and S — Nu^ when p = 1.

An immediate consequence is that for any domain fi strictly contained
in MN, the infimum in (2.4.4) is never attained by a function in WQ,P(Q).
For, if it were attained, then the extension by zero is also a minimizer for
all of R ^ ( since the minimum S is independent of the domain) but this is
impossible by Talenti's result since the mmimizers in MN never vanish (cf.
(2.4.5)).
We have a very interesting connection between the Sobolev inequality
for p = I and the classical isoperimetric inequality. In fact each implies
the other. Assume that the Sobolev inequality holds for all functions u G
Wo'p{n) for a bounded domain H c RN. In this case p* = N/(N-l). Let
e > 0. Define
1, if p(x>dQ) > e
^ ^ {p(x,dQ)/e, if p(x,dQ)<e
where p(x, dft) is the distance of a point x from dQ. Then (p£ G W 0 ' (S7)
N
and J"fi (pe~l dx —> J^ da; = |Q| as e —> 0. Further, if
£>£ - { x e f i | p(a;,5n) < e},

we have
, , / 1/e in £>£
|v
^ = \ o in n\z>e.
Thus,
/ |V^|rfa; = ^ £!
./n £
The limit of the right-hand side is, in fact, the Minkowski content
M;v_i(<9ft). Thus, passing to the limit in the Sobolev inequality for <pe
as e —> 0, we get
|0fi|jv_i >Nu^\n\'-^
which is exactly the isoperimetric inequality.

Conversely, we prove the Sobolev inequality for p = 1 assuming the
isoperimetric inequality. Let u e WQ' (fi). We can assume, without loss of
generality that u > 0, since, otherwise we can always work with |u| instead
of u. On one hand, combining the Fleming - Rischell theorem (cf. Theorem
2.2.1) and the isoperimetric inequality, we get
/ \Vu\dx > NLJ^ [ (i{t)l-TTdt (2.4.6)

Jo, Jo
where p(t) is the distribution function of u. On the other hand (cf. the
proof of Corollary 1.1.2),
/•OO
N
- / t^p,(t)dt. (2.4.7)
1
Jo
Now, since p,(t) is monotonically decreasing, we have
N
I / P(T) dr) > t*-i^(i)£
for t > 0. Thus,
1 . . , . N-l ff , ^ N~\ \
t
N-l d
/ n{T)~R~dr\
N dt
Hence
N f°° i / f°° N-i \*£T
iV J-^ t—^(t)dt < (j( p{T)-*-dr) . (2
Combining (2.4.6), (2.4.7) and (2.4.8), we get
/ \Vu\dx > N^\\u\\^_

N 1 a
Jo,
which is the Sobolev inequality when p = 1.
Chapter 3
Comparison Theorems
3.1 Talenti's Theorem
Symmetrization is a useful tool in proving theorems which compare so-

lutions of different boundary value problems and get estimates on their
various norms. The main result of this section, in its present form, is one
due to Talenti, though earlier versions have been obtained by researchers
like Bandle and Weinberger.
Let fi C RN be a bounded domain. Let A = (a^(:r)) be an N x N
matrix whose coefficients are functions defined on fi. Let 0 < a < ft. We
set M(a, {3, Q) to be the collection of all such matrices such that
a\tf<AU<m\2, a. e. (x)
for all £ G RN. Associated to such a matrix is a uniformly elliptic second
order differential operator defined by
C(u) = -div{AVu).
We then consider the elliptic boundary value problem: given / G L2(Q),

find u such that
c( /infi 1
i^z-^>
0 on dQ. J (3.1.D
If we define the bilinear form
a(u,v) = / AVu.Vvdx, for n,v G H^(Q) (3.1.2)
then the weak formulation of the above problem is to find u G HQ{Q) such
47
that
a(u,v) = J fvdx for all v G flj(fi). (3.1.3)
By virtue of the Lax - Milgram lemma, the problem (3.1.3) admits a unique
solution u e HQ((1) (see, for instance, [Kesavan (1989)]) called t h e weak
solution of (3.1.2).
T h e o r e m 3.1.1 [Talenti (1916b)] Let H C RN be a bounded domain and

let 0 < a < (3. Let A e M(a,0,n) and let f £ L 2 ( H ) . Let u e H&{n) be
the (weak) solution of the problem (3.1.3). Let f* € L2(Q*) be the Schwarz
symmetrization of f. Consider the problem
-aAv = f*inn* |
Ifu>0inQ, then u* < v a.e. in W, where u* is the Schwarz symmetriza-

tion ofu.
P r o o f : Step 1. W i t h o u t loss of generality, let us assume t h a t a — 1. Let

t > 0. Set v = (u — t)+. T h e n v G HQ(Q) a n d we can use it in the weak
formulation (3.1.3). We then get (cf. the proof of Theorem 2.2.2), as usual,
-r, I Y] a i : j ^ ^ d x = / fdx. (3.1.5)

11
J{u>t} ^ X
° 3 dxi J{u>t}
Step 2. By the ellipticity of the differential operator, we have
/ \Vu\2dx < V" a,ij-—-^—dx.

dx
J{t<u<t+h} J{t<u<t+h} rj^l j oxi
Observe t h a t Lu>t\ \Vu\2dx is a decreasing function of t. Thus, dividing

both sides of the above relation by h and letting h —> 0, we get, by virtue
of (3.1.5),
0 < -•£ / \Vu\2dx < f fdx. (3.1.6)

dt
J{u>t} J{u>t}
Once again, by the Cauchy - Schwarz inequality,
(\ [ \Vu\dx) < \ \ ( \Vu\2dx) ( i / dx)

n n
\n> J{t<u<t+h} I \ J{t<u<t+h] J \ J{t<u<t+h} J
Comparison Theorems 49
Thus, letting h —* 0 and using (3.1.6), we get
(~ [ \Vu\dx) < -n'(t) [ fdx (3.1.7)

dt
V J{u>t} J J{u>t}
where fi(t) is the distribution function of u. Let us set
F(t) = / f#(s)ds.
Jo
Recall that (cf. Proposition 1.2.2)
r rv(t)
/ fdx
fdx < J f#{s)ds = F(ti{t)).
Jfu>t\
{u>t} JO
In view of (3.1.6), we then deduce that F(fi(t)) is non-negative for all t > 0.
We combine the above inequalities with the result of Corollary 2.2.3 (which
involves the co-area formula and the isoperimetric inequality and the fact
that u > 0) to get
(JVwjfoVM 2 -* < -/i'(t)F(/i(t)).

Equivalently, we write
1 < (AT<)-2(/,(t))*-2F(Mi))(V(t)). (3.1.8)
We integrate this inequality from t' to t and make a change of variable on

the right-hand side using f = //(£). Observing that — fi is monotonically
increasing, we get
t-t' < (iV<)"2 / " " ' S*-2F{i)dt. (3.1.9)
As remarked earlier, the integrand on the right-hand side is non-negative.
Step 3. Set t' = 0 and t = u # ( s ) - 77 where 77 > 0 and 5 e (0, |fi|). Then
/i(t') = |H| and
H(t) = | { u > u # ( s ) - 7 ? } | > s
by the definition of the unidimensional rearrangement. Since we remarked

already that the integrand in (3.1.9) is non-negative, we then get
u*{s)-v < (Nu»)-2 J £»-2F({i)dZ.

Since rj > 0 was arbitrary, we deduce that

•l«l
2
u*(s) < (Ni4 ) " / f*-aF«)de. (3.1.10)
Step 4. Since ft* is a ball and / * is radially symmetric, it follows from the
symmetry properties of the Laplacian that v is also radially symmetric. In
fact, if we set (by abuse of notation) v(x) = v(\x\), then v(r) satisfies the
following two point boundary value problem:
~v"(r)-^v'(r) = r,0<r<RA
[6A }
v>(o) =V(R)= o }
where R is the radius of fi*. Integrating this equation once and using the
definition of /* in terms of the unidimensional decreasing rearrangement
(cf. Section 1.3), we get
N
•r"- V(r) = f aN~1f#(u;NaN)da = (NUN)-1 P"" f#(s)di
Jo Jo
Setting G(r) to be the last term in the above relation and integrating again,
we get v(r) = f r1~NG(r)dT, which, on changing the variable to £ =
UNTN , yields
v(r) = ( i V < ) - 2 / Z»-2[ / f#(s)ds df. (3.1.12)

Ju>NrN \Jo J
Since the integrand is non-negative, it is now clear that v is radially de-
creasing and so we can write
v*(s) = (iV<4)" 2 / £*-2F(0# (3.1.13)

JS
In view of (3.1.10) and (3.1.13), it follows that u#{s) < v#(s) and the proof
is complete. •
R e m a r k 3.1.1 As a by-product of this proof, we have seen that if t > 0,

then
/ fdx > 0
J{u>t}
where u is the solution of (3.1.3) (cf. (3.1.6)).
Remark 3.1.2 The above proof involved essentially the co-area formula
and the classical isoperimetric inequality to get the crucial inequality
(3.1.8). It is also possible to prove this result just using the Polya - Szego
inequality for the case p = 2, which, in turn, can be proved without using
the co-area formula or the isoperimetric inequality (cf. Remark 2.3.2). For
details, see [Lions (1981a)]. •
Proposition 3.1.1 With the hypotheses and notations of the preceding

theorem, we have
IMIp.fi < \\v\\P,n* for all l < p < oo (3.1.14)
and
l|Vw||,,n < ||Vv||„, n . for l<q< 2. (3.1.15)
Proof: The relation (3.1.14) is an immediate consequence of the fact that

0 < u* < v in £}* and the fact that the Schwarz symmetrization preserves
all LP norms.
Let 1 < q < 2. Let t > 0. By Holder's inequality,
f \Vu\gdx t} lv«N* < (-£ f{u>t} \vufdx)§ (-M'(t))V
< (F(M0))*(V(*))^-
Let M < oo denote the maximum value of u. Integrating the above relation
between 0 and M, and using (3.1.8), we get
J\Vu\"dx < J [(iV^)-2(^(t))*-2(F(M(t)))2]2 {-n'{t))dt

'Q JO
and so, by the change of variable £ = fi(t), we get

• |fi[
/ \Vu\qdx < / (N^)- 1
^"^^) df. (3.1.16)
Ja Jo •- -1
On the other hand, from the explicit expression for v (cf. (3.1.12)) we get,
using polar coordinates,
/ \Vv\qdx = f {(NtoN)-1r1-NF{u>NrN)]q NuNrN-ldr

Jn* Jo
which reduces to the right-hand side of (3.1.16) thus proving (3.1.15). •
Exercise 3.1.1 With the notations and hypotheses of Theorem 3.1.1, prove
that
liv«||!,n. - ||v«u!,n > ||v«||| in -||Vu*||?, n . + ||V(«-u')||?, n ..

Hence deduce (3.1.15) for the case q = 2.
Exercise 3.1.2 Let u e H^{fl) be the solution of (3.1.2) and / e L 2 (fi).

We do not assume that u is non-negative. Let v £ HQ(Q*) be the solution
of the problem
- a A v = (|/|)* in ft*
v= 0 on dn'
Then show that, for all 1 2. The following
example shows that the result is false when q = oo and, therefore, for large
enough q as well.
E x a m p l e 3.1.1 Let SI C R 2 be the region enclosed by the ellipse whose

equation is
x2 y2
Then |fi| = nab and |ft*| — TTR2 where R — yab. Consider the problems:
- A iM = 1 in U \
u = 0 on Oil )
and
•Av = 1 in fi*
vÔondtt*.
Then
.2 „2
l-£-fcr , , , a&-*2—2
= a n d l , ( a : , y )
^ ' ^ 2(^T6%
Then, it is easy to see that
._ . -v/a6 . ._ . a26
max v v = —-— and max Vu > ——-=•
2
2 a + ¥
(the latter inequality being obtained by just evaluating |Vu| at the point
(0, b)). Let a = Xb where A > 1. Then, if
4A3 > (A2 + l ) 2
(which holds, for instance, if A = 2), we have
a2b \fah
2 2 >
a +b 2
and so, for all such ellipses, we get HVuHo^n > ||Vv||oo,n*-
3.2 The Equality Case
In the previous section, we obtained estimates for various norms of u, so-

lution of problem (3.1.3) in terms of those of v, solution of (3.1.4), when
u > 0. In this section we examine the case when equality occurs in any of
these estimates.
Throughout this section, we assume, without loss of generality, that the
ellipticity constant a = 1 and that u and v are the solutions of problems
(3.1.3) and (3.1.4) respectively. We also assume that / e L2(Q) is non-zero
and non-negative, so that, by the strong maximum principle, u > 0 in S7.
We denote by F the indefinite integral of / # as before.
Proposition 3.2.1 Let / > 0. Equality occurs in any one of the inequal-
ities of (3.1.14) or (3.1.15) if, and only if, u* — v (a.e.). Thus, either we
have equality in all these inequalities or all of these inequalities are strict.
Proof: Step 1. Let 1 < p < oo. If we have ||u|| p ,n = |M|p,n+, then, as
u and u* have the same Lp norms, and since 0 < u* < v, it follows that
u* — v (a.e.). The converse is trivially true in this case.
Step 2. let p = oo. Let us denote the distribution functions of u and v by

fj, and f respectively. Then,
/i(t) = |{U" > t}| < |{l>>t}| = l/(i)

since u* < v. Now, from the explicit formula for v (cf. (3.1.12)) we get
Differentiating with respect to t, we get
1 = (JVc4)- 2 K*))*- 2 F(Kt))(-i/(i)). (3-2.1)

Comparing this with (3.1.8), we deduce that
where H(X) = ( W w ^ A 1 - * ) - 2 ^ ) . If H is a primitive of H, then the

above relation can be rewritten as
£<«*» > o
where C(t) = 5(i/(t)) - if (/*(*)).
Since 1/(0) = /i(0) - |fi|, we have C(0) = 0. If M = H ^ n = Hoo,n«,
then (,{M) = 0 as well. But we have just seen that f is non-decreasing and
so C = 0 on the interval [0, M\. But H is strictly increasing (unless F = 0,
i e . / = 0, in which case u = 0 and u* = v = 0). Hence £ = 0 implies
that fi(t) = v(t) for all t G [0, M]. Thus u* and v are equimeasurable and
non-increasing and so u* = i; (a.e.). Again, the converse is trivially true.
Step 3. Let 1 < q < 2. Assume that ||Vu|| 9l n = ||Vu||ff)n+. Retracing the
proof of Proposition 3.1.1, we easily deduce that, for almost all t,
1 = (N4)-2(^(t))i-2F(n(t))(~^(t)).
But then this shows that C,'(t) = 0 and so, as in the previous step, we
conclude that u* = v (a.e.).
Conversely, if u* = v, then, by the Polya- Szego theorem and by Propo-
sition 3.1.1, we get
||Vu||9,n* - l|Vu*|| 9 , n . < \\Vu\\q,n < ||Vu||,,n
and so we have equality throughout and this completes the proof. •

Proposition 3.2.2 Let / > 0. If u* = v, then Q is a ball.

Proof: Let /i and v denote the distribution functions of u and v respec-
tively. If u* = v, then \i = v and so (cf. (3.2.1)) we get
i = (Jv w *)- 2 Mt))#- a F(,i(i))(VW)

for £ > 0. Going back to the proof of Theorem 3.1.1, we see that this
implies that we have equality in the result of Corollary 2.2.3, which in turn
implies that the set {u > t} achieves equality in the classical isoperimetric
inequality. Thus, for t > 0, the set {u > t} is a ball. Now Q = {u > 0}
is then the increasing union of such balls. In particular this implies that
Q is convex. Further, if we choose a sequence {tn} decreasing to zero such
that each set {u > tn] is a ball of radius Rn and centre at xn G fi, then
Rn increases to R and |H| — lim n _ f00 LONRU ~ ^NRN and, by extracting a
subsequence, if necessary, xn —> x G Ct. If y G fi, then y G {u > tn} for n
sufficiently large. We then have
\y-%\ < \y-xn\ + \xn -x\ < R+\xn-x\.

Thus, it follows that \y — x\< R and so SI is a convex set contained in the
ball centered at x and whose volume is equal to that of this ball. Hence £~2
must be the ball centered at x and of radius R. I
Thus, by translating the origin we can assume that u* = v implies that

Q = fi*. We now ask the question as to whether it also implies that f = f*
and u = u*. We answer this question in the affirmative. We give below
the proof given in [Kesavan (1991)]. Alternate proofs can also be found in
[Kesavan (1988)] and in [Alvino, Lions and Trombetti (1986)].
We will henceforth assume that fl = U* and that u € HQ{Q) is the
solution of (3.1.3). Let v G HQ(Q.) be the solution of (3.1.4). We assume
that / > 0. We need the following two lemmas to prove the result mentioned
above. The first lemma deals with the particular case when / = 1.
Lemma 3.2.1 Assume that f = 1. Ifv = u*, then u — u*. Let g — g* >
0 belong to L2(Q). If ip G HQ(£1) is the solution of the dual problem
<X,<P) = [ 9Xdx for all X € H£(Q) (3.2.2)
and if (p G HQ(Q) satisfies —A(p = g, then
 0), the variational formula-
tion (3.1.3), the preservation of the integral under symmetrization and the
variational formulation of (3.1.4). Thus, we have equality throughout and
so
l|Vti*|||, n = ||V«||l i n = f udx.
On the other hand, by the variational formulation of the symmetrized prob-

lem (3.1.4),
/ Vu*.Vudx = / udx.
Jo. JuQ
These two relations immediately imply that u — u*. Now, by Talenti's
theorem, we know that <p* < (p. Further,
fn gudx = a(u, <p) — JQ (pdx — JQ <p*dx < fn (pdx
= fnVu*.Vipdx — J^gu*dx — fngudx
since u = u*. So again we have equality throughout and we deduce that

ip* = {p. Thus, once again,
l|Vv?*||2,fi < l|Vv3|ll,n < a(v,v) = fn9<pdx < JQ9*V*dx
= fng?dx = ||Vvlli, n = l|V^||!,n.

Thus,
IIVvllLn = \\V<p\\in = [ 9<fdx.

JQ
On the other hand, we also have
/ V<p*.V<pdx = / gipdx
Jn JQ
by the variational formulation of the problem for ft = (p*. Again, these two
relations imply that <p = <p*. •
R e m a r k 3 . 2 . 1 T h e dual problem (3.2.2) is the weak formulation of the

problem:
-div(AtV<p) = g in H
tp = 0 on dQ
where A1 is the transpose of the matrix A. T h e above lemma states t h a t if

equality occurs in Talenti's theorem when / = 1 then it also occurs for all
dual problems with t h e right-hand side radial and decreasing. •
T h e next lemma is useful in obtaining identities which cannot normally

be obtained by using test functions in t h e various variational formulations.
L e m m a 3.2.2 Let f e L 2 (H) be non-negative. Let (p G C(Cl) be radially

symmetric and strictly decreasing. If
/ f<pdx = / f*(pdx,
Jn Jn
then for any other function ijj sharing the same properties with <p, we have
f wdx = I r tftdx.
Jo, JU
P r o o f : Notice t h a t , by hypotheses, every ball centered at the origin and

contained in the ball fi is a level set of tp and of ip and vice versa. Now,
(cf. L e m m a 1.2.2)
/ f<pdx = ipm{n J fdx-r I ( / fdx) dt (3.2.3)

JO. JQ Jifmin \J{v»t} )
and
f f rVmax / r \
/ f*ipdx = ipmin f*dx+ / f*dx\dt. (3.2.4)
JCI JO, Jpmin \A^»t} /
T h e first terms in the right-hand sides of (3.2.3) and (3.2.4) are equal. Since
{ip > t} is a ball centered at the origin, we also have (cf. (1.3.2))
f fdx < f f*dx.

J{ip>t} J{t}
Hence the hypothesis that the left-hand sides of (3.2.3) and (3.2.4) are equal
implies that, for almost every t,
[ fdx = f f*dx. (3.2.5)
By our earlier observation on the level sets of ip and ip, it follows that (3.2.5)
still holds if tp is replaced by ip. Hence, using the formulae corresponding
to (3.2.3) and (3.2.4) with 0. Then, if u* = v, it follows that
u — u* and that f = f*.
Step 1. First of all,
HV«1li, n < | | V u | | ! ( n < a ( W , u ) - fnfudx < fnf*u*dx
= jnf*vdx = \\Vv\\la = ||Vu*||| i n .

Thus,
l | V u l ! i n = ||V«||l ( n - a(u,u) = f fudx = [ f*u*dx. (3.2.6)

Jo. Jo
Step 2. Let z G i/g(Q) be the solution of the dual problem with right-hand
side = 1:
a(x»*) = / Xdx for all X e i?o( n )

Jo
and let w £ flo(fi) be the solution of the corresponding symmetrized prob-
lem —Aw — 1 in fi. By Talenti's theorem, it follows that z* < w. Again,
Jnfzdx = a(w, z) = Jn udx — fû*dx — f^'S7w.Vu*dx\
\ (3.2.7)
= / n /*«Kfc > fnfz'dx > Jnfzdx )
and so equality holds throughout. In particular,
/ f*wdx = I f*z*dx.
Jo, Jo
Now, since / ^ 0, the above equality holds in the ball {/* > 0} which is of
positive radius. In particular, z* = w in this ball. But since both z* and w
are radially decreasing (for w{x) = (R2 — |a;|2)/2iV) where R is the radius
of fl), it follows that z*(0) = iu(0), i.e. the L°° norms of 2* and w are
equal. But then, by Proposition 3.2.1, it follows that z* = w in Q. Thus,

we are in the situation of Lemma 3.2.1 (applied to the dual problem with
right-hand side = 1). Hence, we deduce that z ~ z* = w. It then follows
from equality holding throughout in (3.2.7) that
/ fwdx = f f*wdx. (3.2.8)

Jn Jn
Step 3. Now both the functions iv and u* = v are strictly decreasing radial
functions (cf. (3.1.12)) and so we can apply Lemma 3.2.2. Thus, we deduce
from (3.2.8) that
a(u,u*) =
J fu*dx = f f*u*dx. (3.2.9)
Ju Jn
We again appeal to Lemma 3.2.1 and apply it to the dual problem. Since
we already know that —Aw* = /*, u* must also solve the dual of the dual
with right-hand side /*, i.e.
a(u*,X) = /Vxdxforallxetfo1^). (3.2.10)

Jn
In particular,
a(u*,u*) - / f*u*dx (3.2.11)

Jn
and
( Vu*.Vudx = J f*udx = a{u*,u). (3.2.12)

Jn Jn
Now,
0 < ||V(«-u*)||!in < a(u-u\u-u*).
Expanding this, and using (3.2.6), (3.2.9), (3.2.11) and (3.2.12), we get
0 < 2 ( / \Vu\2dx - J Vu.Vu*dx) < ( f \Vu\2dx - f Vu.Vu*dx

\Jn Jn J \Jn Jn
and so
/' \Vu*\2dx = f \Vu\2dx = [ Vu.Vu*dx

Jn Jn Jn
which yields u = u*.
Step 4. Finally, Since u solves the original problem (3.1.3) with right-hand
side / and u* solves the same problem, but with right-hand side / * (cf.
3.2.10), we deduce from the above t h a t / = / * . •
R e m a r k 3 . 2 . 2 T h e above proof just involves repeated use of t h e Polya-

Szego and Hardy-Lit tie wood inequalities. Another proof found in [Kesavan
(1991)] shows t h a t t h e maximum of u is attained at a unique interior point
and t h a t in each ball {u > t}, the m a x i m u m occurs at the centre. Thus,
all the level sets {u > t} are concentric and thus u is radial and decreasing
and so u — u*. To do this, L e m m a 3.2.2 is used in a different way. [Alvino,
Lions and Trombetti (1986)] also give a proof of the fact t h a t the level sets
are concentric, using other arguments. Of course, one can also verify fairly
easily t h a t t h e condition of Brothers and Ziemer (cf. (2.3.6)) is satisfied and
so the fact t h a t ||Vu*||2,n — ||Vu||2,n will immediately imply t h a t u = u*.
B u t the argument t h a t we have presented here and the ones cited above
are self-contained and fairly elementary. However, later on, when studying
the case of equality in nonlinear problems, we will, indeed, appeal to t h e
result of Brothers and Ziemer. •
One of t h e immediate consequences of t h e above theorem is t h e following

characterization of radially decreasing functions.
Corollary 3.2.1 Let fl = ft*, let w € H^(ft) satisfy - A w = 1 in ft. Let

f G L2(ft) be non-negative. Then, the following are equivalent:
(ii) Jn fwdx = Ja f*wdx. _

(iiijf^fipdx — f~f*(pdx for any ip &" C(ft) which is radial and strictly
decreasing.
P r o o f : T h a t (i) =!> (ii) is obvious. T h e fact t h a t (ii) <=*• (iii) is the statement
of L e m m a 3.2.2. We now show t h a t (ii) => (i). let u,v G H Q ( ^ ) such t h a t
-Au = f and —Av — f* in ft. Then, by Talenti's theorem, u* < v. By
(ii) we get
Jnfwdx = jnVu.Vwdx = jnudx = JQu*dx < }Qvdx
— fn Viv.Vvdx = fn f*wdx — /n fwdx
and so it follows t h a t u* = v and by t h e preceding theorem, t h a t u = u*

and / — / * , which proves (i). •
We now look at the matrix A associated with a problem for which the
equality case occurs. Consider the following hypothesis:
(H) Let A be a matrix such that there exists / > 0 in L2(Q) with the prop-
erty that if u € HQ(Q) is the solution of (3.2.4), then u* satisfies -Au* = /*
in H*.
If (H) holds, then we just saw that Q. — Q*, u = u* and that f = f*.
We also saw that if —Aw = 1, then w also satisfies
a(xM = /xforallxetfo1^).
Proposition 3.2.3 Assume that (H) is valid- If A is symmetric, then,

for almost every x £ Vl, we have
N
2â>ij{x)xj = Xi. (3.2.13)
3=1
Proof: Since w(x) — (R2 — \x\2)/2N , where R is the radius of fi, satisfies
the dual problem with right-hand side = 1 as well as —Aw = 1, we get that
a(w,w) = I \Vw\
This is the same as

N
I z~-, aij{x)XjXidx ~ I y ^ x{dx.
By the ellipticity condition, we then deduce that for almost every a;6fi,
N N
y ^ aij(x)xjXi = ^^x2. (3.2.14)
i,j = l z=l
Now, by the ellipticity condition, A(x) is a symmetric matrix whose first

eigenvalue is greater than, or equal to, 1 and the above condition then
shows that the first eigenvalue is indeed unity with x as eigenvector. This
completes the proof. •
If A is not symmetric, we do not have a condition like (3.2.13) as the

following example shows.
Example 3.2.1 Let N = 2. Define
a-ij(x) = 8ij-(i-j)(xl + xl), l<ij<2.
Then if C{u) = -div(v4Vu), we have that w(x) = (R2 - \x\2)/4 satisfies
C(w) = -Aw = 1
in the ball centered at the origin and of radius R. However condition

(3.2.13) is not satisfied.
Example 3.2.2 Even if A is symmetric, the differential operator C need

not even resemble the Laplacian. Again, let N = 2. Let £(x\1X2) be a
non-negative, smooth and bounded function. Define
MX\ = l + îî'^) ~x\X2^{xi,x2)

l-x1x2£(x1,x2) l + x2£(x1,x2).
If w is as above and C the corresponding differential operator, then, in the

ball centered at the origin and of radius R, we have
C(w) = -Aw = 1.
If an = 1 for 1 < i < N, then we can say more about A.
Proposition 3.2.4 Assume that an = 1 for all 1 < i < N. Assume that
(H) holds. Then
(i) ai:i = -aji fori j= j .
In particular, if A is also symmetric, then C = —A.
(it) If (f is any radial function, then C(<p) = —A(p.
Proof: Since an = 1 for all 1 < i < N, the ellipticity condition yields
for all £ £ MN. If { e ^ } ^ is the canonical basis for RN, we choose £ = e^+e^
and £ — e^ — e^ successively in the above inequality to get (i).
By virtue of condition (H) and Lemma 3.2.1, we know that if — Aw = 1
in Q (where w is given by w(x) — (R2 — \x\2)/2N), then w also solves
-div(i4*Vw) = linfi.
Using this information together with (i), we get
=
Yl'fofaiW**} -^faT.&iifcM = °- (3.2.15)
We also get, again by (i),
2_\.aij{x)xix3 — 0- (3.2.16)
i^3
Let <p = <p(r) be a radial function (where r = \x\). Then
Now,
E w sir (««(*)&) = £** si- ( « « ( ^ ^ )
where ^ ( r ) = <p'(r)/r. Thanks to (3.2.15) and (3.2.16), both the terms on

the right-hand side of the last relation vanish and this proves (ii). •
3.3 Sobolev Imbeddings
In Section 2.4, we saw how we could use symmetrization to evaluate the

best Sobolev constant and also saw that the classical Sobolev inequality
for the case p — 1 was equivalent to the classical isoperimetric inequality.
In this section, we apply Talenti's result to get estimates on the constants
involved in certain other Sobolev imbeddings. We first start by estimating
various norms of solutions of elliptic boundary value problems.
Proposition 3.3.1 Let Q, C RN be a bounded domain. Let f e L2(S1).

Let u G HQ(U) be the (weak) solution of the following problem:
-Au = f in n |
[6 6 l)
u = o on on.) --
l
Let2<p<oo. Assume that f € LP(ft). Set 0 = $ - . Then,
(i) if (3>0, we have u € L°°(ft) and
\\u\Ua < (^r20~l\nf\\f\\p>Ci; (3.3.2)

(ii) if ^ > —p > 0, we have u G Lg(Q,), and
I f1 X
2 + (3.3.3)
||u|kn < ( A ^ r | / T W * / (1-**)*<** Ip.n
Uo
when /? < 0 and
||u||,, n < (Ar«;^)- a |n|i Q f ( - l o g * ) * * ) ' l l / I U (3.3.4)
w/ten p = 0.
Proof: Without loss of generality, we can assume that / > 0 (cf. Exercise
3.1.2). Then, as we saw in Theorem 3.1.1,
u*(s) < (Nv$)-* J™ £*-*(£ f#{r,)dn\dt.
Case 1. Let p = oo. Then,
|u#( a )| < ( J V w * r 2 l l / l l o o , n / ' % * - 1 d e
from which it follows that
\u#(s)\ < (iV<)"2 ( - 0 (|fi|* - , * ) 11/m.n
which yields (on setting s = 0 in the above inequality) the estimate (3.3.2)
when p = oo.
Case 2. If p < oo, then, by Holder's inequality,
/ ftirtdq < ||/ # || P l (o,|n|)€ 1 -* = 11/llp.nC1-*-

Jo
Thus,
\u*(s)\ < {N4r2\\f\\p.n rt*-*-1** (3-3.5)

Js
which then yields (3.3.2) when /? > 0.
Case 3. Let 0 < 0. Let tf"1 > - 0 . If /? / 0, then (3.3.5) gives
|u#(s)| < (N4)-2\\f\\p,n\&\-l{sV-\nf)

and so
IMIS,n = H«#C(o,|n|) < (N4r29\P\-9\\f\\l,n Tit0 -Vfyds

J0
which yields (3.3.3) after a change of variable t — s/\£}\ in the integral

above.
Case 4. Finally, if /? = 0, then (3.3.5) gives
\u*(s)\ < (JVw*)- 2 ||/|| P i n(log(|n|)-log S )
and hence
ll«li;,n < (Arw*)-a«||/n;-n jT'"' ( - log ( ^ ) y d*
from which we easily deduce (3.3.4). •
Let u e W2'p{n) n W 0 1,p (n). If we set / = - A u , then u is the solution

of (3.3.1) with / e ^P(fi). Further,
11/Hp.n = ||Ati|| Pl n < HtiHa^n.
The following result is, therefore, a direct consequence of the preceding

proposition.
T h e o r e m 3.3.1 Let £1 C RN be a bounded domain and let 2 0, we have
^ ( f l ) n i v 0 l , p ( f i ) c L°°(fi)
and, /or all u £ W2'P{Q) n W01,p(ft), we Jiaue
IMIoo.fi < C||u||2,Pjn
where
C < (NUJ")-2^-1^0;
(iij if {3 < 0 and q~l > ~{3, then
w2'p(n)nw^'p(Q) cLg{n)
and, for all u e W2*(Q) n W^V{U), we have
HU.n < C||«||2>P,n

where
c < |(;v4)-2i/?rW+*l/oCi-^)'*!* >//ô

" | (tfu/*)- a |fi|i ( ^ ( - l o g i ) * * ) * '/ 0 = 0.
3.4 The Obstacle Problem
In this section we will apply the techniques used to prove Talenti's result
to study a simple example of a variational inequality.
Let fi C RN be a bounded domain. Let A G M(l, /3, Q) be a matrix as
in Section 3.1. Let a(.,.) be the associated bilinear form (cf. (3.1.2)). Let
K = {w€H£(U) \W>0 in fi}.
Given / € L2(Q), we consider the following variational inequality: find

u € K such that, for all w G K,
a(u, w — u) > / f(w ~ u)dx. (3.4.1)

Jn
This problem has a unique solution (cf., for instance, [Kesavan (1989)])
and, if A is symmetric, then u is the minimizer, over K, of the energy given
by
J(w) — -a(w,w) — I fwdx.
This is a free boundary problem. The coincidence set is the set given
by
fio = {x e H | u(x) = 0}.
One can, in M2, imagine a membrane stretched over the region Q with an
obstacle along the plane and the membrane being fixed along the boundary
d£l. The energy J is the strain energy and the equilibrium position of the
membrane occurs when this energy is minimized over all possible configura-
tions. It can be shown that on the set H+ = fi\fio> the function u satisfies
the differential equation C(u) = / , where C is the associated second order
elliptic differential operator (cf. Section 3.1).
As in the case of Talenti's theorem, we can compare the solution of this

problem with that of a 'symmetrized problem' and obtain useful informa-
tion on the size of the coincidence set.
As before, given f E L 2 (fi), we define
F(0 - / f*(v)dv.
Jo
Lemma 3.4.1 Let u e K be the solution of (3.4-1)- Then, F > 0 on the
interval [0, \{u > 0}|]. Farther, for all 0 < t' <t < ess. sup.(u), we have
i H{«>*'}l ,
t~t' < (NLJ{[)-2 ^-2F(£)<e (3.4.2)
J\{u>t}\
Proof: Let t > 0. Set w = (u - t)+. Then
a(u, w) — j fwdx = I (C(u) — f)(u — t)dx — 0

Jn Ju>t
since u satisfies C(u) = f on the set {u > t} C {u > 0}. Further,
N
a(u,{u-t)+) = / âtj-^-^dx.
ux Udj
J{u>t} i j= 1 3 -<-
Thus, we get
/ yzai3^—^~dx =
i f{u-t)dx.
Hu>t} ^ dx3 dxi J{u>t}
Notice that, by the ellipticity condition, the integrand on the left is non-
negative and so the integral is a decreasing function of t. Thus, we get
N
0
^ -if E ^7r-7rdx = f fd* < ^(MW)
where we have set // to be the distribution function of u. This immediately
implies that F > 0 in the interval [0, \{u > 0}|]. The inequality (3.4,2) now
follows exactly as in Step 2 of the proof of Theorem 3.1.1. •
Corollary 3.4.1 Let u e K he the solution of (3.4-1)'• Then
2 1 0 2
u#(s) < /(A^)- /. ^ ^*- ^ ^ if Se[0,|{U>0}|]
\o if se[|{u>o}|,|n|].
(3.4.3)
and
— - 7 - W < {Nu>%)-*s*-'£F{s), for s 6 ( 0 , | { u > 0 } | ) . (3.4.4)
Proof: Let 0 < s < s' < \{u > 0}|. Then u*{s') < u*(s) since u* is de-
creasing and further, by definition of the rearrangement, |{u # > u#{s')}\ <
s'. If e > 0, then, as u# is decreasing, we have
|{t/ # > u # ( s ) - e } | > | { u # > u # ( s ) } | > s.
Set t' - u#{s') and t = u#(s) - e. Then, by (3.4.2), we get
i r3'
2
U#(s) - £ u#{s') < (Nug)- / f*-2F(0df
since the integrand is non-negative in the interval concerned. Letting e —> 0,

we get
u#(s)-u#(s') < (Nvji)-2 f f*"2FK)df (3.4.5)

is
If we now set s' = \{u > 0}|, then u # ( s ' ) = 0 and we get (3.4.3). The
estimate (3.4.4) follows directly from (3.4.5). •
Let us now consider the symmetrized problem: find v e K* C H^(Q*),

such that for all w e K*,
f Vv.V(w-v)dx > f f*(w-v)dx (3.4.6)

Jo,' Jn*
where
K* = {weH^n*) | w>0 in £T}.
This problem has a unique solution v E K* which minimizes the corre-

sponding energy
wdx
over K*. Further, i; satisfies the equation —Av = f* over the set {v > 0}.
By the Polya - Szego and the Hardy - Littlewood inequalities, it is clear
that J*(w) > J*(w*) and so, by the uniqueness of the minimizer of J*,
it follows that v = v*. Thus, the set {v > 0} is a ball with centre at the
origin. It is then immediate to see that
v#(s) = {{N4)~2 fl{V>°}] t*-'F(t)d(, s€[0,\{v> 0}|] (3 4 7)

u K
\o, 8€[|{i;>o}|,|n|]. --'
Theorem 3.4.1 Let u and v be the solutions of (3.4-1) and (3.4-6) re-
spectively. Then
'0 if f <0 a.e.,

dx and
|{u>0}| < |{^>0}| W */ Inf ^°> / ^ °
so the unique solution of F(s) = 0
in ( | { / > 0 } U ^ | ) 5 otherwise.
Further, u* < v.
Proof: (i) If / < 0 a.e., then it is immediate to check that u — 0 and v = 0
are the solutions to problems (3.4.1) and (3.4.6) respectively.
(ii) Let / n fdx > 0. Then F(0) - 0 and F(\Q\) = f^1 f*{s)ds = JQ fdx >
0. Further, since / * is decreasing, it follows that F is concave and thus
F > 0 in the interval [0, |H|]. It then follows that z G HQ(Q*) such that
—Az = f* is non-negative and is given by the right-hand side of (3.1.12).
It is now obvious that z also solves (3.4.6) and so v = z and it follows that
{«>o}| = |n|.
(iii) Let us now assume that Jafdx < 0 and that | { / > 0}| > 0. Since
F is concave and F(0) — 0 while F(|fi|) < 0, it follows that there exists
a unique s 0 G (|{/ > 0}|, |fi|) such that F(s0) = 0. Notice that F > 0 in
(0, so) and that F < 0 in ($o, |^|)- Now, v is strictly radially decreasing in
the set where {v > 0} and
^ ( S ) = -(iv4)" 2 s *- 2 F( s )
in that set. Hence it follows that \{v > 0}\ — so where the derivative
vanishes for the first time. We already saw that F > 0 on the interval
[0, \{u > 0}|] and so we have
| { u > 0 } | < so = \{v>0}\.
That u* < v^ is now immediate from Corollary 3.4.1. Hence u* < v. •

Proposition 3.4.1 Let u and v be the solutions of (3.4-1) and (3.4.6)

respectively. Then
J{u) > J*(v).
Proof: By the ellipticity condition, we have
J(u) > \2 ( \Vu\2dx- f fudx.

Jn Ja
Notice that u > 0 and so by the Polya - Szego and the Hardy - Littlewood
inequalities we get
J(u) > J*{u*).
But u* E K* and so by the variational characterization of v, we have
J*(u*) > J*(v)
which completes the proof. •
3.5 Electrostatic Capacity
The isoperimetric inequality for the problem of electrostatic capacity is one

of the early successes of the theory of symmetrization and is due to [Szego
(1930)].
Let Qo be a bounded domain in M3 and let T C ô be a subdomain.
The region ft = £IQ\T represents an electrical condenser. We denote the
boundary of UQ by To and the boundary of T by Ti- The electrostatic
potential is the function u E Hl(fl) which satisfies the following:
Au = 0 in Q
u = Q on r 0 } (3.5.i;
u = 1 on IY
The electrostatic capacity of H is given by
C{Q) = 1-- / [ ^da (3.5.2)

47T ./ r , dv
where v is the unit outward normal on the boundary of tt.
Multiplying the differential equation in (3.5.1) bu u and integrating by

parts using Green's theorem, we easily see that
C(fi) = - ^ / \Vu\2dx.
4TT Jn
By the maximum principle, u attains its extreme values on dQ, and not
in Q. Thus, 0 0 on T\.
Proposition 3.5.1 (Dirichlet Principle) The electrostatic potential u is
the minimizer of the functional
J{w) = / \Vwf, dx
JQ
JQ'
over the set
K = {we Hl(Q) \ w = 0 on TQ and w = 1 on Ti}.
Proof: Let w G K. Then, there exists ip e HQ(Q,) such that w = u + (p.

Then,
/ Vu.Vt/;cfa; = - / Au.(pdx = 0.
7n Vo
Thus,
J \Vw\2dx = f {\Vu\2 + \Vip\2)dx > f \Vu\2dx

JQ JQ JQ
which proves the claim. •
The Dirichlet principle for the capacity immediately yields an isoperi-

metric inequality for the same.
Theorem 3.5.1 (Szego) Let QQ and T* be balls centered at the origin
such that | n j | - |fi 0 | and \T*\ = \T\. Set S = fi5\T*. Then,
C(to) > C(fi). (3.5.3)
Proof: Let u be the solution of (3.5.1) and denote by u its extension to T

by unity. Then u > 0 and it belongs to HQ(QQ). Now, by the Polya - Szego
inequality,
4nC{n) = f \Vu\2dx > [ \Vu*\2dx = [\Vu*\2dx.

JQQ JQ* JQ
But on H, u* is non-negative; further, u* = 0 on OCIQ and u* = 1 on dT*

Thus, by the Dirichlet principle applied to fi, we deduce that
L' \Vu*\2dx
n
> 47rC(n)
R e m a r k 3.5.1 We can formulate the problem discussed above in all di-

mensions N. The least capacity, given the volumes of Qo and T, will always
be for the spherical annulus bounded by dT* and <9fi£. •
We can explicitly calculate C(Q). Let the outer and inner radii of the
annular domain H be Ro and R\ respectively. Thus 0 < R\ < RQ. AS ob-
served in the proof of the preceding theorem, where we extended functions
to the inner region by unity, the functional to be minimized will be
/ \Vw\2dx
over the set of non-negative functions vanishing on OQQ and = 1 on T*.

Further, the electrostatic potential, v, satisfies 0 < v < 1 in the annulus.
Consequently v* will also be equal to unity in T* and will vanish on 3Qg.
Since Schwarz symmetrization decreases this integral, it follows that
/ \Vv\2dx > [ \Vv*\2dx.
By the uniqueness of the electrostatic potential, it then follows that v = v*.

Thus, writing (as usual, by abuse of notation) v = v(\x\), the function v(r)
satisfies
u " M + ^v'{r) = 0, Ri < r < Ro

v(R{] = 1 ; v{Ro) = 0. }
This gives
„2-N R2-JV
v(r)
logr-log-Rp pr _ 9
logflx-logfio' JV _ Z
When N — 3, using the above formula and the definiton (3.5.2) of the
capacity, we get that

R0R1
C(Q)
Ro-Ri
3.6 The Saint Venant Problem
One of the important conjectures in mathematical physics solved by Polya

was that of Saint Venant regarding the torsional rigidity of an elastic cable.
Consider a cylindrical cable in M3 with uniform cross section Ho C K 2 .
Assume that fii c Ho is a subdomain. Let \Q\\ = a. Then, the warping
function is denned as the unique solution of the following problem:
-Art = 2 in Q = QoVh
u —0 on <9Qo
u = c, an unknown constant, on dQ\
• U %** = 2a-
The torsional rigidity of the cable is defined as the quantity
z
S(Sl) - / / \Vu\ dx. (3.6.1)
Jn
The conjecture of Saint Venant states that of all cross sections (UQ)
of given area and of all holes (fix) of given area, the torsional rigidity is
maximal when QQ and Hi are concentric circles. The first proof of this
conjecture is due to [Polya (1948)]. The result was extended to an arbi-
trary number of holes by [Polya and Weinstein (1950)] and [Payne (1962)]
proved an important inequality for the torsinal rigidity. We will give below,
following the treatment of [Mossino (1984)], a proof of Payne's inequality
and the theorem of Polya and Weinstein.
Remark 3.6.1 In the absence of any hole, i.e. if fii = 0, the result is an
immediate consequence of Talenti's theorem. For, u will satisfy —Au — 2
in O and will vanish on <9H. If v satisfied the same equation on fi*, then,
by Talenti's theorem, we have
[ \Vu\2dx < f \Vv\2dx
(cf. Proposition 3.1.1) which is precisely the conjecture of Saint Venant in

this situation. •
Let QQ C R 2 and let ft; C fto be pairwise disjoint subdomains for

1 < i < m. Let I \ = dft^, 0 < i < m. Let |ftj| = a* for 1 < i < m. Finally,
set ft = f2o\ U ^ ! ft*. Let v be the unit outward normal on the boundary
of ft. The warping function on ft is defined as the solution of the following
problem:
-Au = 2 in ft
u —0 on To
u — Ci > 0, an unknown constant, (3.6.2)
on Ti, 1 < i < m
Jr Wda = 2a^ l<i<m.
Proposition 3.6.1 Assume that the Ti, 0 < i < m are all sufficiently
smooth. Then, the problem (3.6.2) admits a unique solution.
Proof: Step 1. (Uniqueness) If u\ and u-i are two possible solutions to the
above problem, then, let w — u\ ~ u-i. Then, — Aw = 0. Multiplying this
by w and integrating by parts over ft using Green's theorem, and using the
other conditions defining the warping function in (3.6.2), we get
(du1_du1)
/ -Aw.wdx — / \Vw\2dx ~ y^(ci,j - c2,i) /
Ja Jn ~r^ Jvr . \ dv dv
The integrals over I \ all vanish by the last condition in (3.6.2). Thus, by
Poincare's inequality (w = 0 on To), it follows that w — 0 which proves the
uniqueness of the solution.
Step 2. Let c — (ci, ...,c m ) G R m be given. Let v = u(c) be the unique

solution of the problem:
-Av = 0 in ft
v — 0 on TQ
v — Ci on Ti, 1 Rm defined by A(c) = d is thus linear. If A(c) = 0, then
0 = / -Av.vdx = / iVvfidx-y^Ci / -^-da.

Jn Ja £f M dv
By hypothesis, the integrals on I \ all vanish and so we get that J*n |Vv\ 2 dx =
0 and again, by Poincare's inequality, it follows that v = 0. Thus, c = 0
and so A is one-one. Therefore, it is onto as well. It is now evident that
the affine linear map T : Rm -> M™ defined by
T(c) = A(c) + /9
where /3 = (/?i, ...,/? m ) is also onto. Thus, if we choose di = 2a* — ft, then,
if A(c) = d, we get u = v(c) + w satisfies
-Au =2 in 0
u =0 on To
u — Cj, on r\, 1 < i < m
%** = 2a,i, 1< i < m.
Step 3. We now show that the Ci are all strictly positive. If I \ , 1 < i < m
are all sufficiently smooth, then u G H2(Q) c C(fl), by the Sobolev imbed-
ding theorem, since Q C R 2 . Thus u will be continuous. Let xo G H be
such that u attains its minimum there. If XQ G fi, then VU(XQ) = 0 and
Au > 0. But Au = —2 < 0. Thus w attains its minimum on the boundary.
If XQ G Ti for some 1 0. Thus XQ G TO and so u > 0 and u > 0 on fi\r0. This
proves that Ci > 0 for 1 < i < m. •
Remark 3.6.2 The smoothness of the boundary was used only in Step 3,
to prove the positivity of the unknown constants. Otherwise the existence
of u vanishing on To and constant on each r \ , 1 < i < m is alsways true.
Even the smoothness required is minimal. It is enough if the boundary is
Lipschitz continuous. •
Remark 3.6.3 We have, in fact, proved that we can assign arbitrary val-
ues to the integrals of the normal derivatives over I \ , 1 < i < m as is
evident from Step 2 of the above proof. The special values 2a\ = 2|fij| not
only ensure that the c^ are all positive, but also give us a neat variational
characterization for the problem (3.6.2). Set
W — {v E #0(^0) I v = fy, a constant, in Qi, 1 < i < m}.
Then, it is a simple exercise to check that the variational formulation of

(3.6.2) is to find u e W such that, for all veW,
/ Vu.Vvdx = 2 / vdx.
JQQ Jn
If u is the warping function, we define the torsional rigidity of Q, as

before, by
S{Q) = / \Vu\2dx.
Jo.
Example 3.6.1 (The case of the circular annulus) Let m = 1. Let
n = {x e M2 1 0 < Ri < \x\ < R0}.

Then, a\ — TTR2.
Consider u(x) = (RQ ~ \x\2)/2. Then it is immediate to check that

—Au — 2 in fi and that u — 0 on To = {x \ \x\ = RQ}. Further, u is a
constant (being a radial function) on Y\ = {x \ \x\ — R\}. Now, if r = |a;|,
du ,
_ = - t i (r) = r
on Fi (where we have written, by abuse of notation, u — u(r)) and so
/ JLfo = 2-KR\ = 2a 2 .
Thus u is the warping function and the torsional rigidity in this case is
given by
S(n) = 2TT / r(u'(r))2dr = -(R40-Rt).
Let m(t) = \{u < t}\ for t > 0. Now,
{u < t} = {x \ R2 - 2t < |x| 2 < R20}.
Thus m{t) = 2irt

Lemma 3.6.1 Let u be the warping function, i.e. the solution of (3.6.2)
and let m(t) = \{u < t}\. Then,
2irt < m{t). (3.6.3)
Equality is attained for the circular annulus.

Proof: Without loss of generality, we can assume that the subdomains f2;
are numbered such that
0 = c0 < c\ < c2 < ... < c m < c m + i = u m a x
where umax stands for the maximum value of u.
Step 1. Let 0 < t < umax. set w - t - (u - t)~ £ H1^). Thus,
ft, if it > t
w =<
(u, if u < t.
On one hand, we have
/ -Au.wdx - 2 / wdx = 2t\Q\ + 2 / (u - t)dx.

in Jn J{u<t}
On the other hand, by Green's theorem,
Ja -Au.wdx = fn VuVwdx - /an ^wda
= I{u<t} \û\2dx ~ 2 E C l <t a%Ci - 2t £ C i > t a{.

Thus,
/ \Vu\2dx = 2 ]T ai°i + 2t XI ai + 2i l n l + 2 / (u ~ *)da;.

Hu<t) Ci<t c .>( J{u<t}
Differentiating with respect to t, we get (for almost every t)
— I \Vu\2dx = 2 V ^ + 2|Q| -2m(t). (3.6.4)

dt
•>{«<*} t>t
Step 2. As in the proof of Talenti's theorem, a combination of the Fleming
- Rischell theorem and the Cauchy - Schwarz inequality gives
(PQ({u<t}f = \AJu<t \Vu?dx\ < m'(t)(~J \Vu\2dxY

(3.6.5)
Step 3. Now
(Pa{{u<t}))2 = PR2({u>t}UUCi>tni)
by the properties of the de Giorgi perimeter, since u > 0 in fi and vanishes

on its external boundary. Hence, by the classical isoperimetric inequality,
we have
(Pn({u<t}))2 > 4nA
where
A = \{u>t}\+^2a,i = \n\-m(t)+Ylai-
ci>t a>t
Thus, in view of (3.6.4) and (3.6.5), we get
4TT ( \Q\ - m(t) + J2 a<) ^ 2m 'W I lnl ~ m ( f ) + J2 ai

I'
\ ci>t J \ a>t J
Since the term in parantheses on both sides is strictly positive, we get
m'(t) > 27r for almost every t which yields (3.6.3) on integrating between
0 and t (since m(0) = 0). •
Lemma 3.6.2 Let u be the solution of (3.6.2). Then,

m
|Q|2
S(ft) < ^ + 2 ^ ^ ^ (3.6.6)
In particular, we have Payne's inequality:

m
|Q|2 2u
s{n) < v~+ mBxy;oi. 1=1
(3.6.7)
Proof: We have, from (3.6.2),
2 / udx = / -An.urfx = / |Vu| 2 dx — / —da

Jn JQ Jn Jan ®v
which gives
.|fi|
S(fi) 2 / udx + 2 >_JaiCi = 2 u#(s)ds + 2yâjCi
where u# is the uindimensional increasing rearrangement of u (cf. Section

1.4). Now, set t = u#(s) in the preceding lemma. Thus,
2iru#(s) < m(u#{s)) = \{u a<iCi
7o r^
which gives (3.6.6). Since, for all 1 < i < m, we have Q < ii m a x , the
inequality (3.6.7) follows immediately. •
Theorem 3.6.1 (Polya - Weinstein) Let u be the solution of (3.6.2). Let

So be the torsional rigidity of the circular annulus having the same measure
as H and having the hole of area Yl^Li ai- Then,
S(fi) < ^ +2 w f > <ff + ^f;ai = So. (3.6.8)
Thus, of all multiply connected domains of fixed area with an arbitrary

number of holes of fixed total area, the circular annulus has the maximum
torsional rigidity.
Proof: The first inequality above is Payne's inequality proved in the pre-
ceding lemma. Now, u m a x — u#(|fi|), by definition and we saw, in the proof
of the preceding lemma, that
2TTU#(S) < s.
Thus, u#(|fi|) < |0|/27r and the second inequality in (3.6.8) follows. Fi-
nally, it remains for us to show that So is given by the quantity mentioned
in (3.6.8). Indeed, if RQ and R\ are, respectively, the outer and inner radii
of the annulus, then J2iLi ai = nî anc * 1^1 = ^(ô ~~ -^l)- Thus, as seen
in Example 3.6.1,
So = i(*8 - .Rf) = f ( ^ - i ? ? ) ( ^ + i??)
ini)W+i om|.
M^M iSli.
Remark 3.6.4 We saw that u#(s) < s/2ir. Now, we know that (cf. Exer-
cise 1.4.1) u*(s) = u # (|fi| - 5). Thus,
u#{s) < (\U\~S)/2TZ.
If v is the warping function for the annulus, then (cf. Example 3.6.1),
v(x) = (R% - \x\2)/2 and so (cf. Exercise 1.1.2)
Thus, u#{s) <v#(s). M
3.7 Comments
In this chapter, we have given a sampling of results to show how sym-

metrization could be used to produce comparison results for solutions of
partial differential equations and to get estimates for these solutions.
As mentioned earlier, the first results in this direction were those of
[Bandle (1975)] and [Weinberger (1962)] and the result of [Talenti (1976b)]
that we presented here was a refinement of these. The method of proof is
quite robust and has been used in a variety of situations. The basic idea,
always, is to get a differential inequality for the distribution function of the
solution, which will reduce to an equality for the symmetrized problem, and
then deduce the comparison result. There is an enormous amount of litera-
ture on this topic. The most general results, using Schwarz symmetrization,
are given in [Alvino, Lions and Trombetti (1990)]. See also [Talenti (1985)].
Comparison results using Steiner symmetrization can be found in [Alvino,
Diaz, Lions and Trombetti (1996)]. While most of these results are valid for
second order elliptic problems with Dirichlet boundary conditions (because
of the implicit involvement of the maximum principle), there have also been
attempts to tackle other types of boundary conditions (cf. [Ferone (1986)]
for results on a Neumann problem and [Brandolini, Posteraro and Volpicelh
(2003)] for an elliptic problem with a mixed boundary condition) as well as
fourth order operators (cf. [Ferone and Kawohl (2003)]).
The above mentioned method for obtaining comparison results also
works for parabolic problems. The reader is referred to the works of [Bandle
(1976a)] and of [Mossino and Rakotoson (1986)].
We mentioned earlier, in Section 3.1 (cf. Remark 3.1.2) that P. L. Lions

gave an alternative proof of Talenti's theorem without using the isoperi-
metric inequality and based on Lieb's proof of the Polya - Szego inequality
for the case p — 2 (cf. Remark 2.3.2). This proof relies on the fact that
the fundamental solution of the heat equation in R ^ is spherically symmet-
ric, positive and decreasing, i.e. it is equal to its Schwarz symmetrization.
This idea has been devoloped to give an alternative approach to the study
of camparison results for solutions of elliptic and parabolic equations by
[Alvino, Lions and Trombetti (1991)].
In Section 3.4, we presented a very elementary example of a variational
inequality, viz. the obstacle problem. Schwarz symmetrization has been
successfully used to obtain comparison results for various variational in-
equalities. We cite, in particular, the works of [Bandle and Mossino (1984)],
[Maderna and Salsa (1984)], [Alvino, Matarasso and Trombetti (1992)],
[Posteraro and Volpicelli (1993)] and [Posteraro (1995)] and the references
contained therein.
The proofs of the isoperimetric inequalities for the electrostatic capac-
ity and the torsional rigidity, thereby settling long standing conjectures of
mathematical physics, were amongst the earliest successes of the method of
symmetrization. See [Mossino (1984)] for generalizations of the results on
capacity, applied to condensers in series and in parallel. See also the work
of [Ferone (1988)].
The equality case was independently investigated by [Alvino, Lions and
Trombetti (1986)] and by [Kesavan (1988)], [Kesavan (1991)].
The bibliography cited above is, obviously, far from being exhaustive.
The interested reader will find further references quoted in these works.
Chapter 4
Eigenvalue Problems
4.1 The Faber - Krahn Inequality
We now look at some examples of eigenvalue problems for elliptic partial dif-
ferential operators and some isoperimetric inequalities which are associated
to them. The first of these is the famous Faber-Krahn inequality which, in
fact, was first conjectured by [Lord Rayleigh (1894)] in his treatise on the
theory of sound in 1894, but was proved independently by [Faber (1923)]
and [Krahn (1924)] towards the end of the first quarter of the twentieth
century.
We return to the notations of Section 3.1. Let CI C R ^ be a bounded
domain. Let A E M(a,{3, fi) and let
C(u) = - d i v ( ^ V u )
be the corresponding second order elliptic differential operator in divergence

form. Let a(.,.) be the associated bilinear form defined on HQ(Q) X HQ(Q)
(cf.(3.1.2)). Given / € L2(ft), let u e H&(tt) be the unique solution of the
problem:
a(u,v) — / fvdx, for all v e HQ(Q).

JQ
Let us set u = Qf. Then Q can be thought of as a linear mapping of L2(U)

into itself, since HQ(U) <—> L2(S7). Since, Q is bounded, the above inclusion
is compact, by the Rellich - Kondrosov theorem (cf. [Kesavan (1989)]), and
G is thus a compact linear operator on L2{U).
Let us now assume that the matrix A is, in addition, symmetric, i.e.
a
ij — aji f° r a n 1 ^ i>3 ^ N. Then, for f,g G L2(Q), we have, setting
83
u = Qf and w = Qg,
(Gf,9) = (fl»w) = a ( w , « ) = a(u,w) = (/,w) = (f,Gg)
where (.,.) denotes t h e inner-product in L 2 ( Q ) . Thus, Q is a compact

self-adjoint operator on L 2 (fi) a n d hence its spectrum coinsists of 0 a n d
a sequence of eigenvalues (i£ [ 0. There exists a n orthonormal basis for
L2(fl) consisting of eigenvectors of Q.
Let us now consider the eigenvalue problem for t h e operator C We look
for (X,u) e R x (H 0 1 (fi)\{0}) such t h a t
a(u,v) = A(«,v) for all v £ H£(Q). (4.1.1)
This is equivalent t o solving
u = Q(\u).
Thus, from our preceding observations, we have a sequence \^ = (fj,^)~l of

eigenvalues a n d associated eigenfunctions {</?£} (forming a n orthonormal
basis for L2(Q,)) such t h a t
a(rf,v) = A j J ( ^ , « ) for all v E H*(Q) (4.1.2)
and, further,
0 < Af < \£ < \$ < ... < A^ < . . . - > o o . (4.1.3)
T h a n k s t o t h e strong m a x i m u m principle, t h e first eigenvalue Af, also

called the principal eigenvalue, is simple (hence t h e strict inequality follow-
ing it in (4.1.3)) and, further, t h e eigenfunction does n o t vanish inside fi.
Thus, we will henceforth assume t h a t tpf > 0 in fl. Notice, further, t h a t
all t h e eigenfunctions ip£ are, in fact, in HQ(£1).
T h e eigenvalues {A^} can be given a variational characterization using
t h e Rayleigh quotient which is given by
RA(u) = ^ ^ , u^Q, ueH^Q). (4.1.4)
Let Vn be t h e subspace of HQ(Q) spanned by { v i S y ^ , — j ^ n } - Then,
A£ = RA(<p£)= m a x RA{v)
= min RA(V)
v±Vn~i
= m i n m a x RA (V)
dim W=n v£W
Eigenvalue Problems 85
where t h e minimum in the last relation is taken over all n - dimensional

subspaces of HQ(Q).
In particular, we have
Af - min RAM. (4.1.5)
Further, if w G HQ(£1) is such t h a t RA{W) = Af, then w will be an eigen-

function corresponding to Af.
For a proof of all the above mentioned results, see, for instance, Kesa-
van [Kesavan (1989)]. T h u s , to get an upper bound for Af, it suffices to
take any function v € H^ft) and compute RA(V). On the other hand, it is
quite difficult to get lower bounds for \f.
Towards t h e end of the nineteenth century, [Lord Rayleigh (1894)] con-
jectured that, when C — —A, the Laplace operator, t h e disc has the least
principal eigenvalue amongst all plane domains of equal area. If we denote
t h e first eigenvalue of t h e Laplace operator as Ai(fi), then, in our notation,
this reads as
Ai(fi) > Ai(fi*) (4.1.6)
This was independently proved by [Faber (1923)] and [Krahn (1924)] nearly
a quarter of a century later and (4.1.6) is now known as the Faber-Krahn
inequality.
Using the Polya- Szego theorem (cf. Theorem 2.3.1), we can prove a
more general result in all space dimensions.
T h e o r e m 4.1.1 Let A e M{a,(3,Q) where H C RN is a bounded domain.

Let {A^(fi)} denote the sequence of eigenvalues of the operator C. Let
{Xk(Q)} denote the eigenvalues of the Laplace operator. Then
Af(fi) > aAi(fi) > a A i ( £ T ) . (4.1.7)
P r o o f : Let </?f > 0 be the first eigenfunction of C. Then, by t h e Rayleigh

quotient characterization of eigenvalues, we have
1 {)
~ ~mh - KII2,« •
But
L\^t?dx Jn \Vv\*dx _
— . „ > mm —-—= = Ai(iZ;,
which proves the first inequality in (4.1.7). Now, let (pi > 0 be the first
eigenfunction of the Laplace operator in ft. Since (pi > 0, we can apply
the Polya - Szego theorem, and the fact that the Schwarz symmetrization
preserves the L2 - norm, to get
JnlVvil 2 ^ > Jn. IVpII'ds
Now ifl e HQ(Q*) and so, again appealing to the variational characteriza-
tion of the principal eigenvalue, we easily deduce that
Ai(fi) > Ai(fi*)

and the proof is complete. •
When ft is a ball, we can explicitly solve for the first eigenvalue and
eigenfunction of the Laplacian in terms of Bessel functions. Let ft — H* =
BR, the ball with centre at the origin and of radius R. Since symmetrization
decreases the integral J*fl \Vipi\2dx and keeps the L2(ft) - norm unaltered,
it follows that the Rayleigh quotient of (pi is greater than that of y?J. But
Since X$ft) is the minimum of the Rayleigh quotient, it follows that (p\
achieves this minimum, and hence that it is an eigenfunction as well. Now,
the simplicity of the principal eigenvalue implies that (pi = <p\. Thus, the
first positive eigenfunction of the Laplacian, with homogeneous Dirichlet
boundary conditions, is a radial function, which is radially decreasing, when
ft is a ball. Thus we can set <pi{x) = u(\x$, where u now solves the ordinary
differential equation (got from the Laplacian, in polar coordinates, applied
to a radial function)
N - 1
u"{r) + u'(r) + Xu(r) = 0.
A straight forward computation shows that the solution can be written as
u(r) = cr1'^JKî(VXr)
where c is an arbitrary constant and Jp(p) is the Bessel function (of the
first kind) of order p satisfying the differential equation
The boundary condition u(R) = 0 determines A. If J P J I denotes the first

positive zero of the function Jp then, clearly, we must have yf\R = JK-I I,
Ai(Bfl) = -£=H. (4.1.8)
T h e eigenfunction is given by
Vi(s) = I s f - f j ^ p t ^ M ) . (4.1.9)
We now examine the case when equality occurs in the Faber-Krahn

inequality. Faber and K r a h n proved t h a t equality occurs in the case of
plane domains only when H is a disc of the same area. In general, in the
literature (cf., for instance, [Payne (1967)]), while explicit reference is m a d e
to t h e case of plane domains, no mention is made of t h e equality case for
higher dimensions. To the best of the our knowledge, t h e only discussion
of this is to be found in [Kawohl (1985)] and [Kesavan (1988)]; the former
uses ideas from t h e theory of Steiner symmetrization while the latter uses
Schwarz symmetrization and the equality case of Talenti's theorem (cf.
Section 3.2). We give this proof now.
T h e o r e m 4.1.2 Let Q C E N be a bounded domain such that
Ai(ft) - A^fl*).
Then, CI is a ball
P r o o f : Let (pi > 0 be t h e first normalized eigenfunction for t h e Laplace

operator. T h u s ,
—A(fi — A i ^ i in fi
tpi = 0 on dQ,
Let w
where Ai = Ai(Q) - Ai(fi*) and HîH^n = 1- G # o ( ^ * ) be the
solution of
—Aw — Xi(pl in Q*
w = 0 on dQ*.
Then, by Talenti's theorem (cf. Theorem 3.1.1) <p* < w. Thus,
—Aiu < Xiw.
Multiplying this inequality by w and integrating over Q*, we get
/ \Vw\2dx < Ai / w2dx.

Since Ai is also the principal eigenvalue of the Laplace operator on ft*,

which is the minimum of the Rayleigh quotient, we deduce that
JQ.w2dx
Since, Ai, the minimum of the Rayleigh quotient, is achieved at w, it fol-
lows, as already noted earlier, that w is an eigenfunction of the Laplacian,
corresponding to Ai, i.e. —Aw = Xiw. Thus, w = tp\ and so we are in
the equality case of Talenti's theorem and hence, by Proposition 3.2.2, it
follows that ft is a ball. •
Two domains fti and SI2 in ^tN a r e said to be isospectral if Afc(fti) =

Afc(^2) for all positive integers k. A celebrated question of [Kac (1966)]
was whether isospectral plane domains are isometric, i.e. congruent to each
other. Of course, this question could be posed in all dimensions and also in
the context of manifolds. While most cases were settled in the negative, the
case of plane domains remained open for several decades and was finally
settled by [Gordon, Webb and Volpert (1992)], again in the negative.
Nevertheless, the spectrum of the Laplacian contains several pieces of
information about the geometry of the domain. For instance, from the Weyl
asymptotic formula (cf. [Protter (1987)])
it follows that two isospectral domains have the same volume. In the case of
plane domains, a further refinement of this result due to Pleijel (cf. [Protter
(1987)]) states that
yVAfc(n)t - lal L l
£r[ 47rt 4 V2TT£
where L is the perimeter of the domain. Thus, if two plane domains are
isospectral, they have the same area as well as the same perimeter. If one
of them is a disc, then, L 2 — 4-KA and so, from the isoperimetric theorem,
the same being true for the other, it follows that the other domain is also
a disc.
We can now prove this in all dimensions. If two domains in *RN are
isospectral, they have the same volume. If one of them is a ball, we can,
without loss of generality, call the domains ft and ft*. Since they are
isospectral, in particular, Ai(fi) = Ai(H*) and so, by Theorem 4.1.2, it fol-

lows that Q, is a ball.
Exercise 4.1.1 Let Q C RN be a bounded domain and let P : O —> R be

a strictly positive continuous function. Consider the eigenvalue problem:
—Aw = XPw in Q
w — 0 on dQ.
Show that this problem admits an increasing sequence {\n,p(Q)} of positive

eigenvalues which tends to infinity and that the first eigenvalue Ai ( p(fi)
admits an eigenfunction of constant sign. Show that
J
Ai lP Q) = min "' ' •
O#VG^(0) J^Pv^dx
With usual notations, if P* denotes the Schwarz symmetrization of P , show

that
Ai,p(fi) > AI,P-(Q*).
4.2 The Szego - Weinberger Inequality
In the previous section, we discussed isoperimetric inequalities for the prin-

cipal eigenvalue under homogeneous Dirichlet boundary conditions. We
now consider Neumann boundary conditions. Let ft C R N be a bounded
domain. Consider the eigenvalue problem
-Au = »u in n 1
K }
fs = o on m\
where v is the unit outward normal on dft. As in the previous case, we

have an increasing sequence of eigenvalues
0 — Ho < Pi(^) < M2(^) ••• < A*n(^) < ... —> oo
with an associated family of eigenfunctions (which now belong to H1(Q))

forming an orthonormal basis for L2(fl). The eigenfunction corresponding
to fj,0 = 0 is the constant function. The principal (Neumann) eigenvalue is,
therefore, pi (SI) which can be given the following variational characteriza-
tion (via the Rayleigh quotient):
Ml fi = min J n f ' ' • (4.2.2)

In vdx=0
In the case of a ball BR C M^ (JV > 2) of radius # > 0, we can explicitly

write down the eigenvalues in terms of Bessel functions, by the method
of separation of variables (cf., for instance, [Courant and Hilbert (1953)]).
Using polar coordinates, we can write u(x) — w(\x\)v(u), where u G SN~l,
the unit sphere in RN. Then, it can be seen that w satisfies the equation
7V_ i TV - 1
w"{r) H w'(r) — w(r) + fMl{BR)w(r) =0, for 0 < r < R
(4.2.3)
and the conditions w(0) = w'(R) = 0. As in the case of the Dirichlet
problem, we can easily verify that the solution to this equation is given by
l
w (r) = cr%~ Jx(^(B^r) (4.2.4)
where c is an arbitrary constant and Jp denotes, as usual, the Bessel func-

tion of the first kind of order p. Once again, the eigenvalue is determined
from the boundary condition, which, in this case, reads as w'(R) = 0. Thus,
for instance, if R = 1, we get
where /xi = ni(Bi). In particular, if N = 2, /ij is the square of the first

positive zero of J[.
We will now express fii (BR) in terms of integrals of w and its derivative
which will be useful in the sequel. We rewrite (4.2.3) as
W(B«Mr) = _ - ^ A ( r " - V ( r ) ) + ^««;(r).
Multiplying this on bothsides by w and integrating over BR, we get
HI{BR) / w2{\x\)dx = NuN / - —(r*-V(r))u;(r)dr

J B pi " 0
N 1
f ~ n (X2J
Evaluating the integral of the first term on the right by parts and using the
fact that w(Q) = w'(R) = 0, we get
Nu>N f ~-^{rN~1w,(r))w(r)dr = NujN [ rN~lw'(rfdr

Jo dr J0
I. w'(\x\)2dx.
Thus,
^BR) ~ jBRM\*\)>d* • (4 2 5)
' '
Just as in the case of the Dirichlet eigenvalue problem, where the first
eigenfunction could be chosen to be of constant sign, here we can choose
w such that w'(r) > 0 in (0,i?). Thus, w is a non-negative and increasing
function.
The isoperimetric inequality for the first (Neumann) eigenvalue reads
as
MiW < M ^ * ) - (4-2.6)
This was proved by [Szego (1954)] for simply connected plane domains and
in full generality by [Weinberger (1956)]. In order to prove this inequality,
we need two technical lemmas which we now establish.
Lemma 4.2.1 LetfiC R ^ be a bounded domain, let g : [0, oo) —* [0, oo)
be a continuous function. Define
Pi(x) = g{\x\)^- for l<i<N. (4.2.7)

\x\
Then, we can choose the origin in such a way that, for all 1 < % < N, we
have
f P,{x )dx = 0. (4.2.8)

Ju
Proof: Choose a ball B centred at the origin such that fi C B. Let y G B.
Define F : B - • RN by
F^/rflx-,,,)^ dx.
Taking the inner-product of F(y) with y, we get

{x
F(y)-y = f 9(\x-y\) f~lypdx.
Jn \x-y\
If \y\ — R, the radius of B, then, for all X E U ,
x-y-\y\2 < \x\.\y\-\y\2 < o

and so F(y).y < 0 for all y on the boundary of B. It then follows from the
Brouwer fixed point theorem that (cf. [Kesavan (2004)] or [Lions (1969)])
there exits y0 e B such that F(y0) — 0. We shift the origin to y0 and
(4.2.8) follows. •
The next lemma is the result of a simple application of the Schwarz

symmetrization.
Lemma 4.2.2 Let Q C M.N be a bounded domain. Let g : R —> R be a
non-increasing function. Then
f g(\x\)dx < f g(\x\)dx. (4.2.9)

Jn Jn*
Proof: Let g(x) = g(\x\) and let h = ^|n- Then
/ g(\x\)dx — I hdx — / h*dx.

Jn Jn Jn*
The proof will be complete if we show that, for all x 6 SI*, we have h*(x) <
g{\x\). Indeed, h*(x) and g(\x\) are both radial functions which are radially
decreasing. Hence it is sufficient to show that their distribution functions
satisfy the same inequality. Now, given any t G R,
\{h'>t}\ = \{h>t}\ = | { x e n | 5 ( | x | ) > t } | < \{X e RN | g(\x\) > t}\

Since g is non-increasing, the set appearing in the last term of the above
chain of relations is a ball with centre at the origin. Since {h* > t} is also
a ball centred at the origin and is contained in S~T, we get
\{h*>t}\ < |$rn{xeR"|9(|:r|)>t}| = \{xen*\g(\x\)>t}\

which establishes the claim. •
Corollary 4.2.1 In the above lemma, if g is non-decreasing, then the
inequality in (4-2.9) is reversed.
Theorem 4.2.1 Let Cl c R ^ be a bounded domain. Then (4-2.6) holds.
Proof: Step 1. Define
w(r) for 0 < r < R

w(R) for r>R
where R is the radius of ft* and w is the function satisfying (4.2.3) - (4.2.5).
Then g is a non-decreasing and non-negative function. Setting
Pi(x) = 9 ( M ) g , l<i<N
we can assume that (by changing the origin, if necessary) JQ P{(x)dx — 0

for all 1 < i < N (cf. Lemma 4.2.1). Then, by virtue of the variational
characterization of the principal eigenvalue (cf. (4.2.2)), we have
Mi (ft) / Pfdx < J \VPi\2dx, l<i<N.
Summing over alii, for 1 < i < N, we get
Mn)<(/ o £lv^)/(/ n E Ptdx

A straightforward computation yields
(JZWPiAw = 9'{rf +—N 3-1— g(r)

, , 2 def
=
D,
B(r)
v
where r = |x| and
i=\ )
Thus,
fQB(\x\)dx
Mi (ft) <
Jng(\x\)2dx
Step 2. Now, for 0 < r < R,
B'(r) = -2 Hi{BR)g{r)g'{r) + — - ( r f f ' ( r ) - p ( r ) ) :
by virtue of (4.2.3). Notice that BR is none other than ft*. Since 5 is

non-decreasing it follows that B'(r) < 0 for 0 < r < R and hence that B is
non-increasing for all r. Now, by Lemma 4.2.2 and its corollary, we get
/ B(\x\)dx < f B(|x|)dx
/ g(\x\)2dx> f g(\x\)2dx.
Thus, it follows that
J n . B{\x\)dx
l(n)
" * fn.9(\*\ydx = "l(n }
by virtue of (4.2.5). This completes the proof. •
N
Corollary 4.2.2 Let ft C R he a bounded domain and let Ai(fi) and
fii (Q) be, respectively, the principal Dirichlet and Neumann eigenvalues of
the Laplacian. Then
/ii(H) < Ai(fi). (4.2.10)

Proof: In view of (4.2.6) and the Faber - Krahn inequality (4.1.6), it
suffices to prove (4.2.10) when U = Q*. Let R be the radius of Q*.
Now, the first (Dirichlet) eigenfunction ip\ associated to Ai(fi*) satisfies
the differential equation
<p'{(r) + ^^V'lir) + \i(n*)<Pi(r) = 0
where, by abuse of notation, we have set fi(x) ~ y>i(|x|). Differentiating

this relation and setting v = (p[, we get
TV - 1 TV — 1
v"{r) + V(r) — v(r) + A i ( f i > ( r ) = 0. (4.2.11)
Now, v(Q) = <p[{0) = 0 and v(R) < 0 (since <pi(r) is decreasing). On

the other hand, from the differential equation for <p\f we get that v'(R) =
<p"(R) > 0. Thus, starting from (4.2.11) and proceeding exactly as in the
derivation of (4.2.5) from (4.2.3), and taking into account that v(R)v'(R) <
0, v(0) = 0, we get
t ,n.^ JBR {AW2 + ^v(\x\?) dx

Al(fi >
> fBRv(\x\)>dx •
On the other hand, since v = <p[ < 0 we can again prove, as in Lemma 4.2.1,
that Pi{x) — v(\x\)xi/\x\ satisfies (4.2.8) for 1 < i < N and then proceed as
in Step 1 of the proof of Theorem 4.2.1 to show that the right-hand side of
the preceding inequality is greater than or equal to /xi(fi*). This completes

the proof. •
4.3 Chiti's T h e o r e m
In this section, we present a comparison theorem (in the spirit of Talenti's

theorem) for the first eigenfunction of the Laplacian (with homogeneous
Dirichlet boundary conditions), due to [Chiti (1982a)].
Let O C R N (N > 2) be a bounded domain and let Ai(fi) be the
principal eigenvalue of the (Dirichlet) Laplacian. Let 5 be the ball with
centre at the origin such that Ai(5) = Ai(ft). By the Faber - Krahn
inequality (4.1.6), it then follows that Ai (Q*) < Ai (S), with strict inequality
if H is not a ball. Thus, by the variational characterization of the first
eigenvalue, it follows that 5 c f l * . In fact, the radius of S, denoted 7 - 1 , is
given by
-• - m
where B\ is the unit ball in RN, in view of the formula (4.1.8) for the first
eigenvalue in a ball. The eigenfunction associated to this eigenvalue is given
by
z(x) = c | i | 1 - * J f l _ 1 ( v ' M n ) | 2 ; | ) , (4.3.2)
where c is a normalizing constant.

Lemma 4.3.1 Choose c in (4-3.2) such that
2(0) = ¥>*(0) = max.<pi(x)

x£Sl
where ip\ is the eigenfunction associated to Ai(fi) in Q. Then z < ip\ in S.

Proof: The function z satisfies
- A z = XiZ in 5 1
z = Q on OS J
where, we have set Ai = Ai (Q). We can now proceed exactly as in the proof
of Talenti's theorem to obtain
- £ ( * # W ) < \I(NLO%)~2S%~2 J* z#{v)dri (4.3.3)

for 0 < s < \S\ (start with the relation corresponding to (3.1.9) and proceed
as in the proof of Corollary 3.4.1). We also note that z — z*. Similarly,
-^(vfto) < Xi(Nu,*)-2a*-* j\t{v)dn (4.3.4)
for 0 < s < |fl| = |n*|. Since | 5 | < |fi*|, we have tpf(\S\) > 0 while
z # ( | S | ) = 0 . Now, by choice, z#(0) = z*{0) = v?I(°) = ¥>* (<>)• If 2 = 2* £
y>! in St we can find a f c > l such that, for s G [0, \S\], we have
k(ff(s) > z#(s).
Let us choose the smallest such k such that the above inequality holds.
Now, there exists SQ G (0, \S\) such that
k<pf(s0) = z#(s0).
Define v^ by
#/• \ (ktpfis),
f k(pf(s), for 0 < s < s 0
for so < s < | 5 |
Then, u* is monotonically decreasing and u*(|5|) — 0. Further, by virtue

of (4.3.3) and (4.3.4), we easily see that
—^v#(s) < A i ( A ^ ^ ) - 2 s ^ - 2 / v#(rj)dn.

«s Jo
If we set v(x) = v#{uN\x\N), then, v G H&(S). Then,
/ 5 \Vv\2dx =,/ 0 7 " 1 I v ' M l ^ w j v r ^ - ^ r
= (JVw Af ) 3 / 0 7 " 1 r 2JV - 2 |v#'(u;jvr JV )| 2 r JV - 1 dr
= (^<) 2 /d 5| |^ # '( 5 )l 2s2 "^^

<Ai/0|S|(-^#'W)/o^#(^^
= Ai/ ( J 5 | V #(f?)/f l (-i;# , W)(fed»?

Thus
IsWvfte < Al
Is v2dx
and, as Ai is also the minimum of the Rayleigh quotient on 5 , it follows
that this minimum is achieved for v and so v is an eigenfunction associated
to Ai on S. Then v has to be a multiple of z and so, from the definition
of t;, it follows that k<pf(s) — z#(s) for 0 < s < so as well and so, since
(p\ and z agree at the origin, k = 1 and ipf (s) > z#(s) for all 0 < s < \S\
which proves the lemma. •
We now prove a result due to [Chiti (1982a)] which will be useful in the
sequel.
Theorem 4.3.1 Let c be chosen in (4-3.2) such that
/ tp\dx = / <pfdx = / z2dx. (4.3.5)

Jn Ja* JS
Then, there exists r\ G (0,7 _ 1 ) such that
<pl{x) < z(x), for 0 < | x | < n 1

z{x) < tpl{x), for n < |x| < 7 " 1 . J { }
Proof: Suppose z(0) < <£*(0). Then, for some fe > 1, we have kz(0) —
(^J(O). By Lemma 4.3.1, it then follows that kz#(s) < <pf(s) for all s G
[0,|S|]. But then, (4.3.5) cannot hold. Thus, it follows that 2(0) > y?;(0).
If 2(0) = <pl(0), and (4.3.5) holds, the preceding lemma again implies
that \S\ = |H*| and, in fact z = <p\. Thus the theorem is trivially true.
So let us now assume that z(0) > <pl(0). Thus, \S\ < |H*|. Therefore,
z*(|5|) = 0 and <pf (\S\) > 0. Now, z = z* and so z& is continuous and
(pf is, in fact, absolutely continuous on any interval (a, |H|), for a > 0 (cf.
Remark 2.3.4). So, for a > 0, sufficiently small, z*{a) > (pf(0) > (pf{a).
Consequently, there exists Si € (0, \S\) such that z*(si) = <pf(s\). Choose
si to be the largest such number with the additional property that <fif{t) <
z#(t) for all t € [0, si]. Thus, in an interval immediately to the right of s\,
we have <pf(s) > z#{s). We now claim that this holds for all s 6 (si, \S\].
If not, there exists «2 £ (s\, \S\) such that (pffa) = z#{s2) and ipf (s) >
z#(s) for all s G (si,S2)* Again, these follow from the continuity of the
unidimensional rearrangements. We now define
W#(S) - / * # ( S ) ' f°r SG[°'Sl]U[S2,|5|]

\<pf{s), for s e [ s i , 5 2 ] .
Then, w# is monotonically decreasing and w(x) = W&(LJN\X\N) is such

that w € HQ(S). AS in the proof of the preceding lemma, we can show that
ds Jo
Again, from this it will follow that the Rayleigh quotient (over S) of w is
equal to Ai and hence that w is an eigenfunction for X\, Consequently,
w# _ 2 # a n ( j s o (pY(s) = z#(s) in [si, s2] contradicting the maximality of
si. This completes the proof. •
4.4 The Payne - Polya - Weinberger Conjecture
Let H C Mw be a bounded domain and consider the eigenvalue problem for

the Laplacian with homogeneous Dirichlet boundary conditions:
—Au — Xu in n 1
(
u = 0 on da J '
In 1955-56, Payne, Polya and Weinberger considered bounds for the eigen-
values and showed that
W <,
when N — 2 and conjectured that that the right-hand side could be replaced
by
Ai(fi*)
This result was extended to all dimensions by Thompson (1969) who showed
that
A2(fi) < 1 + 4
Ai(Q) ~ N
and, again, it was conjectured that the right-hand side could be replaced
by
2
A2(n*) = Jj,i
2
i \
Ai(fi*) \U
where, as usual, jPtk denotes the fc-th positive zero of the Bessel function
Jp (of the first kind) of order p.
This problem, for TV = 2, was studied by Brands who obtained the value
2.686 for the ratio of the first two eigenvalues, by de Vries who obtained
2.658 and finally by Chiti who got the estimate 2.586.
The conjecture was finally settled positively by [Ashbaugh and Benguria
(1992a)] in 1992 and we sketch their proof below. Their paper and the
survey article of (Ashbaugh (1999)1 gi y e complete references to the previous
works cited above.
As in the preceding sections, we will denote the eigenfunction (which is
positive in ft) associated to Ai(fi) by tp\.
Lemma 4.4.1 Let P ^ 0 be a sufficiently smooth function such that
Pipi G #o(fi). Assume further that fQ Pip\dx = 0. Then
Proof: Since Py?i € #o(fi) and is assumed to be orthogonal to <fii, we

have
A2(n) (4 4 3)
- fnP'tfdx • -'
Now, as V ( i V i ) — ViVP + PVffiii we have
| V ( / V i ) | 2 - vflVPI 2 + P 2 | V ^ i | 2 + 2<plPVP.V<pl
= </>2|VP|2 + P2\V<pi\2 + îV(P 2 ).Vv?i.
Further,
Ja <pMP2).V<pidx = - / n nP2îdx - / n P a |Vy>i\ 2 dx
= \i{n)fat*<pldx-fap*\v<p1\*dz.
Thus,
f \V(P<pi)\2dx = [ \VP\2tpldx + \i(Q) [ P2(p2dx.

Jn Jn Jii
The inequality (4.4.2) now follows on substituting this in (4.4.3). •
The idea is to use the above inequality for N trial functions of the form
Pi(x) = S ( M ) g , l<i<N (4.4.4)
for a suitable function g. This is reminiscent of the method used in the

proof of the Szego - Weinberger inequality Indeed, the first necessity is
then to ensure that the functions P%<p\ are orthogonal to the eigenfunction
(pi. This is done in the following lemma, exactly as in Lemma 4.2.1.
Lemma 4.4.2 Let g : [0, oo) —> [0, oo) be a sufficiently smooth function.
Then, for a suitable choice of the origin, we have JQ Piipfdx = 0 for all
1 < i < N, where Pi is of the form in (4.4-4)-
Proof: Let B b e a ball of radius R, with centre at the origin, such that
f l c B . For y E B, define
F(V) = [ g(\x-y\)^_Mx-y)2dx.
JQ \X - y\
Then, it is immediate to verify that, for \y\ = R, we have F(y).y < 0 and
so, by Brouwer's theorem, there exists yo 6 B such that F(yo) = 0. We
shift the origin to yo- This proves the lemma. •
We now set
<x = 3%-\,i = \Ai(î)
where B\ is the unit ball in RN. Notice that 7 is precisely the quantity
defined in the previous section (cf. (4.3.1)) and is the reciprocal of the
radius of the ball, S, whose first (Dirichlet) eigenvalue is also Ai(H). We
also define
W(t) = J J f(WA'f-!(«*)> for o < t 1.
Then u> is a C1 function. Finally, we set
g(r) = w{~fr). (4.4.5)

Lemma 4.4.3 (Gap Inequality) If g is defined as in (4-4-5), we have
^ ' ^ < - ^ J ng (N)V l( ^ J • (4A6)
Proof; If we define P e , for 1 < i < N as in (4.4.4), then, thanks to Lemma

4.4.2, we may assume that the functions Pnp\ are orthogonal to (p\ in L2(U).
Thus, by Lemma 4.4.1, it follows that
(A 2 (n) - A:(H)) / P?y\dx < f \VPi\2ip\dx.

Jn Jn
Summing over i, for 1 < i < JV, we get
L(ZliP?)tfdx
But by definition of the Pi, we get
(£P2)(Z) = g(\x\)2
i=\
and
|a;|
i=i
as already observed in the proof of Theorem 4.2.1. Thus, we get (4.4.6). •
Notice that from (4.4.5), we get

(SV)) 2 + ^ S ( r ) 2
= l2B(ir)
where
B(t) = («,'(*))2 + ^ - i ^ ( t ) a .
The choice of the function w has been made so that, as in the case of
the proof of the Szego - Weinberger inequality, the gap inequality (4.4.6)
above becomes an equality for the unit ball. We now prove this fact.
Lemma 4.4.4 Let B\ be the unit ball in RN. Then,
L B(T)JN Aar)2rdr
J0 w(r)jijN__1{ar)'irar
Proof: Step 1. In the case of the unit ball, we have already seen that
Xi(Bi) = a2 and that
tpi{x) = \x\l-^J^_x{a\x\).
The eigenfunction corresponding to A2(#i) = 01 is no longer radial; by

separation of variables, it has a radial part v(|x|) given by
v(r) = r 1 2
JN (fir)
which satisfies the differential equation
v"(r) + — — v'ir) ~ v(r) + (32v(r) = 0

r TZ
and the conditions v(0) = v(l) = 0. Proceeding exactly as in the proof of
relation (4.2.5), we can again show that
2
2
S* lV$v'{\x$)
JBl U IJ
+ ^v{\x\)J 2]dx
U
0 = \2(Bi) = 2~T " • (4.4.8)
Step 2. In the case when fi = B\, we have 5 = i?i as well and so 7 = 1.

With our preceding notations, we have, for 0 < r < 1,
•J$(/fr) v(r)
g(r) = w{r) =
^-i(ar) î(r)
where, by abuse of notation, we have set (p$x) = <^?i(|x|).
Step 3. We saw, in the proof of Lemma 4.4.1, that for suitable functions P ,
j \V{P^)\2dx = [ \VP\2tpldx + Ai(fi) / P2<p2dx.

Jo, Jo, Jo,
Applying this to Q = B\ and to each of the functions Pi{x) = #(|x|) A- and
then summing over i for 1 P«|a)(*) = B(\x$.
1=1
In exactly the same way, since (Pi<pi)(x) = v(\x\)^r, it follows that

N
(£MPapi)\2)(x) = (v>(\x\))2 + ^v(\x\)2.
1=1
Thus, as (g<pi)(x) = v(\x\), we conclude from (4.4.8) and (4.4.9) that
(32 [ v{\x\)2dx = / Bîp^xfdx + a2 f v(\x\)2dx.

JBI JBI JBX
The relation (4.4.7) now follows easily from this on passing to polar coor-
dinates. •
Lemma 4.4.5 With the preceding definitions of w and B, we have that
w is an increasing function and that B is a decreasing function.
The proof of this lemma is the crux of the argument of [Ashbaugh and
Benguria (1992a)] and [Ashbaugh and Benguria (1992b)]); it is rather long
and technical and is proved using the properties of Bessel functions. The
interested reader is referred to the papers cited above.
Let us now go back to the notations of the preceding section and set
5 to be the ball centred at the origin and with radius 7 _ 1 so that its
first (Dirichlet) eigenvalue is exactly Ai(fi). Let z be the corresponding
eigenfunction as in Theorem 4.3.1. We then have the following result.
Lemma 4.4.6 Let f : R —• R be an increasing function. Then
f {{\x\)v$x)2dx > I f{\x$z{xfdx. (4.4.10)

Ja* Js
The inequality is reversed if f is decreasing.
Proof: Using polar coordinates, we get
J f(\x\)z{x)2dx - f }{\x\)<el{xfdx =
Js Jn*
NLON I'* f{r)((z(r)2 -ipXrftr"-^
Jo
n~x
+NuN / f{r)(z{r)2-ip$r)2)rN-xdr
Jr\
.2„7V-1
I /(r)tf(t
where r\ is the number occuring in Chiti's theorem (cf. Theorem 4.3.1)

and R is the radius of ft*. (Here, we have, by abuse of notation, written
z(x) — z(\x$ and <fi(x) = î(|z|), as they are both radial functions.) In
the last integral, we have used the fact that for r > 7 - 1 , we have z(r) = 0.
Since / is increasing, f(r) < f(ri) for 0 < r < n and f(r) > f(ri) for
n <r < R. Now the term z(r)2 - <pl(r)2 is non-negative in the first range
and non-positive in the range n < r < 7 - 1 , by Chiti's theorem (Theorem
4.3.1). Thus we immediately deduce that
[ f(\x\)z(x)2dx- f f(\x\)<pl(x)2dx <

Js Jn*
NtvNf(r1)^$z{r)2-<p*1{r)2)rN-1dr
+ / (z(r)2-<pl(r)2)rN-ldr
Jrl
dr
= }{n)\jsz2dx- J y2dx = 0
by the choice of the normalization of z in Theorem 4.3.1. •
We are now in a position to give the proof of the Payne - Polya -

Weinberger conjecture as done by [Ashbaugh and Benguria (1992a)].
Theorem 4-4.1 Let Q C MN be a bounded domain. Then,
xm < w) £
Ai(fi) " Ai(n*) a2' l ;
Proof: By the gap inequality (4.4.6), we get
Now, B(7r) is a decreasing function as observed in Lemma 4.4.5. Then, as

seen in the proof of Lemma 4.2.2, we have
(B(7W)|n)'(.) < B(7(.))

infi*. Thus,
[ B(i\x$<pi(x)2dx < I B(i\x\)<p\{x)2dx < [ B(-y\x\)z(x)2dx.

Jn Jn* JS
The last inequality above follows from Lemma 4.4.6, since, again, B(-yr) is
decreasing. Similarly w(-yr) is increasing. If w* is the increasing rearrange-
ment of w on H* (cf. Section 1.4) we have (cf. Corollary 1.4.1)
/.(7i*i)V(*)^> / «tea*.
Jn Jn*
Again, as w(-yr) is increasing, we have that wm > w in H* and, by yet
another application of Lemma 4.4.6, we get
/ wt<pl2dx > / ^(7|x|) 2 î(x) 2 dx > / w^xDzix^dx.

Jn* Jn* Js
Using these estimates, we get

2
2JsB(y\x\)z(x) dx
A 2 (fi)-Ai(fl) <Y $sw(i\x\) z(x)2dx

2
_ Ai(Q) Jo ^ ( r ) J j ^ - i ( t t r ) 2 r d r
a
J0 w(r)2JN_1(ar)2rdr
on using polar coordinates. But, by Lemma 4.4.4, we know that the ratio
of the integrals in the last term is just (32 — a2 and the result now follows
immediately. •
4.5 Rayleigh's Conjecture for Clamped Plates
The Dirichlet eigenvalues for the Laplacian in the plane give the frequencies
of vibration of a membrane occupying the region of a domain and which is
fixed along its boundary. The Neumann eigenvalues give the frequencies of
vibration of the free membrane. We now look at the biharmonic operator
A 2 . Let fi C M2 be a bounded domain. If this represents the middle surface
of a thin plate which is clamped along the boundary, then the frequencies
of vibration of the plate are given by the eigenvalues of the problem:
A2u = Au in Q 1 , .
(4 5 X)
u = ^ = o on a n / ' -'
The same problem can, of course, be posed in all space dimensions.

Henceforth, we will assume that 0 C MN is a bounded domain and consider
the problem (4.5.1). As in the case of the Laplacian, we can show that there
exists an increasing sequence {An(Q)} of positive eigenvalues, tending to
infinity, and an associated orthonormal basis of eigenfunctions (which now
belong to H,g(fi)).
The principal difference from the case of the Laplacian is that it is no
longer true that the principal frequency Aj(fi) is a simple eigenvalue. The
corresponding eigenfunction need not be of constant sign either.
However, in the case of the ball, the principal eigenvalue is simple and
the eigenfunction is of constant sign and, in addition, is radial. If we set
and if Jp and Ip are, respectively, the Bessel and modified Bessel functions,
then the principal eigenvalue of a ball of radius R is given by
*•<*>-(*)'-« (i3&)*
where kv is the first positive zero of the equation
Mt)Iv+i(t) + Jv+i(t)Iv{t) = 0. (4.5.2)

For N = 2, [Lord Rayleigh (1894)] conjectured that of all plates of equal
area, the circular plate has the least principal frequency (or fundamental
tone). Of course, this conjecture can be stated in all dimensions and can
be written as
Aj(ft) > Ai(JT). (4.5.3)
To this day, this conjecture has not been fully settled. Szego proved
it under the assumption that the first eigenfunction could be taken to be
positive in H. But, as already mentioned, this is not true in general. In fact,
there is no simple criterion to determine whether a given domain admits
a first eigenfunction which is positive. The result is false even for simple
domains like the square. The next important attempt at its solution was
due to [Talenti (1981)] who proved the weaker inequality
Ai(fi) >cjvAi(fi*) (4.5.4)
where 0 < c^ < 1 is a computable constant depending only on the dimen-

sion N.
The inequality (4.5.3) was finally proved for all plane domains by [Nadi-
rashvili (1995)] by modifying Talenti's approach. This was further refined
by [Ashbaugh and Benguria (1995)] and they also established the inequal-
ity (4.5.3) for domains in IR3. The conjecture remains open in dimensions
N > 4. Pushing the argument of Ashbaugh and Benguria further, [Ash-
baugh and Laugesen (1996)] proved an inequality of type (4.5.4) but with
an improved constant CN which, in fact, tends to unity as N —• oo.
We will now sketch the proofs of these results below.
The starting point of all investigations is the Rayleigh quotient charac-
terization of the first eigenvalue given by the following:
Jn
Ax H = min ; ' . 4.5.5
Theorem 4.5.1 (Szego) Let the first eigenjunction u\ be positive in Q.

Then, (4.5.3) holds.
Proof: Set / = — Au\. Consider the symmetrized problem:
- A t ; = /* in Q*
v = 0 on dQ\
Since m G Ho(fi), an elementary application of Green's formula yields
/ fdx = 0.
'n
Jn
Then, again by Green's formula,
/ ^-da = [ f*dx = 0
Jan' ov yn„
and, as v is radial, | ^ is a constant on dft* and thus vanishes. Consequently,
v e H$(n*).
Now, since u\ > 0 in Q, by Talenti's theorem (cf. Proposition 3.1.1),
we have
/ u\dx < / v2dx.

Jn Jn*
Thus, by virtue of (4.5.5), we get
A m\ - hf2dx > in- f*2dx > A m M

Since we cannot assume, in general, that the first eigenfunction u\ is

positive, we will henceforth set
H+ = {uf > 0} and fi_ = {uj~ > 0}.
By classical regularity results, we have |fi| — |fi+| + |fi_|. Let us denote
fi+ - B a and ST_ = Bfcl
the balls centered at the origin and of radii a and 6 respectively. If R is the
radius of fi*, we thus have
aN+bN = _RW. (4.5.6)

We now set / — Au and define, for 0 < s < |S~21,
G( 8 ) = ( / + ) # ( « ) - ( / - ) # ( | n | - s ) , \ .
F(s) = -G(\n\-s). j â-'>
Both i*1 and G are decreasing functions. Notice that (/~)#(|S7| — s) =

(/")#(«) (cf. Exercise 1.4.1). Now, it is immediate to see that
f\n\ ,|nj
/ F{s)ds = / G(s)ds - 0
Jo Jo
for,
/•M f\n\ r
/ G(s)<fo - / ((/+)#-(/")#)ds - (f+-r)dx = 0
Jo Jo Jn
since f+ — f~ = f has zero mean. We now define
«#(*) = (No;*)-2 JIB'1 £*-2 j * F f t ) * ^ 1

[ (4.5.8)
2 2
w#(s) = ( i v 4 ) ~ j f ' l ^~ /o G(i,)«iude J
and set v(x) = ^ ( ( ^ v l ^ l ^ ) and WJ(X) = ^ ( u ^ a ; ^ ) . (Since v and w are

radial functions, we will, by abuse of notation, write v(x) = v(\x\) and
w(x) — w(\x\) at times, to simplify the exposition.) The following lemma
lists the important properties of v and w.
L e m m a 4.5.1 Let v and w be defined by (4-5.8). Then
v > 0 in Ba and w > 0 in Bf>, (4.5.9)

0 on dBa and w ~ 0 on dBb, (4.5.10)
> - w<°> - > - > - »• <"•»>

aN~^(a) = bN~^(b) (4.5.12)
dr dr
and
-Av = F* in Ba and ~ Aw = G* in Bh (4.5.13)
where F*{x) = F(LJN\X\N) and G*{x) = G(coN\x\N).

Proof: Since F and G are decreasing, it follows that the indefinite integrals
Jo ^(ri)^,ri a n a - Jo ^C7?)^7? a r e c o n c a v e functions and, since they vanish for
s — 0 and s — |H|, it follows that they are both non-negative. Consequently
v# and «;* are also non-negative and so (4.5.9) follows.
By definition, it is clear that (4.5.10) is true. Now
_ ( r ) = -{Nu;N)-1rl-N J F(V)dV
and a similar relation holds for w with F replaced by G. The relations

(4.5.11) follow easily from this. The relations (4.5.13) are also valid since we
have already seen that the explicit solutions of these differential equations
are indeed given by v and w (cf. Step 4 of the proof of Theorem 3.1.1).
Finally, we prove (4.5.12). In view of (4.5.6), it is easy to see that
/ Avdx = - J F*dx = - f G*dx = [ Awdx.

JBa JBa Bb h
JB JB
JEb
Indeed, since Jl Gds = 0,
r r\Ba\ r\n\
/ F*dx = - / Fds = Gds
JBa JO J\EBb\
= - Gds = - / G*dx.
Jo By
JB_
Thus, by Green's formula, we have
f ^V ri — f ^W A
JdBa dv ° JdBb dv
which implies (4.5.12) since v and w are radial and so their normal deriva-
tives on the boundaries of the respective balls are constants. •
Proposition 4.5.1 [Talenti (1981)] With the above notations, we have
(uj)* <v in W (45 14)

(ui)* < w in £1* lit] -
Further,
J {Auifdx = J (Av)2dx+ [ (Aw)2dx (4.5.15)

JQ Jm JQ-
and
f {Amfdx - f {Avfdx = [ (Awfdx. (4.5.16)

Ju Jn* Jn*
Proof: Let t > 0. Retracing the proof of Theorem 3.1.1, we get
(JVw*) a /i(*) 2 -* < -At) I (-f)dx

J{ui>t]
where (j,(t) = \{ui > t}\. Now,
hu^-f)dx = ku^-ndx+j{ui>t} f-dx

<-C)U+)#(s)ds + f^t\f-)#(s)ds
= - /oM(t)(/+)#(|fi| - *)ds + C\r)*(s)ds

= r ( t ) F(S)ds.
Thus, we get
1 < (A^rVW*-2/ F(s)ds(-n'(t))

Jo
which yields
1 f^V 2 ft
t < (Nu£)-2 ^~2 F{ri)d^
Jn(t) Jo
from which we conclude that (cf. the proof of Theorem 3.1.1) (w + ) # (s) <
v#(s) since (i(0) = \Ba\. In an identical manner, we can show that
N \2
(Nu$yv(ty-" < •v\t) \ fdx < -v'(t) I G(s)ds
J{u<-t} J0
where u(t) = \{u < —t}\. This will then give (u )*(s) < w(s). This proves
(4.5.14).
Now,
•|n| •|0|
/ (Aufdx = / (/+) 2 <fa+ / {f-fdx = / (f+)*2ds+ / (D*2ds.
Jn Jn JQ JO JO
(4.5.17)
On the other hand, since ( / + ) * ( / ~ ) # = 0 (cf. Exercise 1.4.3), we have
jf* 1 G{sfds = f™{{f+)#(s))2ds + / d B f c l ( ( / - ) # ( | n i " *))2ds

(4.5.18)
+ 2 2
jr\(f )#(s)) ds+fZ((nns)) ds \Q\
\Ba\
Similarly
/ ••IPSaall f\a\ f\Da\

2 + 2
/ F(s) ds = / ((f )*(s)) ds + ((r)#(s))2ds. (4.5.19)
JO y|Bb| JO
The relation (4.5.15) follows from relations (4.5.17) - (4.5.19).

Finally, since
,|n| ,|n|
/ ((f-)*(\n\~s))2ds = / ((/-)#(*))*&,
JO JO
it is easy to see that

•mi
/•mi /-mi /• 2
/ G(sfds = / ((/+)#(S)2 + ( / - ) # ( s ) 2 ) ^ - / f dx.
Jo
'0 Jo Jn
This proves one part of (4.5.16). The other part follows in the same way.
Let 0 < t < 1. We set
v is radial
2
Kt = lvzH (Bi v(x) = 0 for \x\ = t, when t ^ 0
Vv(x) = 0 for |x| = 1.
We now define, for 0 < t < 1,
p(t LN ) = max
o?v€Kt JBi(Av)2dx'
T h e o r e m 4.5.2 [Talenti (1981)] Let QcUN be a bounded domain. With

the above notations, we have
Ai(Q*) 1
A , ' < -TTT m a x ( p ( t ) + p ( l - t ) ) . (4.5.20)
Ai(fi) p ( l ) 0<t<l V V > r\ u \ )
P r o o f : Let, as usual, u\ be an eigenfunction associated to Ai(fi). Then,
in view of (4.5.14) and (4.5.16), we have
A i ( n ) - 1 = ([(Aurfdx) (f (u+)*2dx+ f (ui)*2dx)
fB v2dx JB w2dx
< . ,\ ,o, +
fBR(Av)*dx fBR(Aw)*dx-
By t h e definition of p, we then have, by a simple rescaling,
aN\ fbN
Ai(H)-1 < R4 Pl +p
w) [w
Further, since it is known t h a t t h e principal eigenfunction in a ball is radial,
Ai(fi*) = R~4/p(l) (cf. (4.5.5)). T h e result now follows by observing t h a t
a and b satisfy (4.5.6). •
R e m a r k 4 . 5 . 1 Using the properties of the Bessel and t h e modified Bessel

functions, [Talenti (1981)] showed t h a t p is an increasing function and so
we get the inequality
Ai(n) > Âi(n*)
in all space dimensions. However, a more careful study of the maximum

occuring in the right-hand side of (4.5.20) can be carried out and in fact
we get the existence of a constant c^ such t h a t
Ai(fi) > CJVAI(Q*).
[Talenti (1981)] showed t h a t c 2 ra 0.97768, c 3 « 0.73910, c 4 %

0.65242, c 5 % 0.60925 and c 6 » 0.58394. We see t h a t cN decreases
to 0.5 as N increases. However, [Ashbaugh and Laugesen (1996)] have im-

proved this to a constant which tends to unity as n —> oo. •
In order to improve on the results of Talenti and, in fact, to prove

Rayleigh's conjecture in dimensions N = 2 and TV = 3, we present another
approach based on the relation (4.5.15). This leads to a new variational pin-
ciple, first proposed by [Nadirashvili (1995)] and later refined by [Ashbaugh
and Benguria (1995)].
From (4.5.14), it follows t h a t
ju\dx< f V*dx+[ ^
and so, in view of (4.5.15), we get
fQ. (Av)2dx 4- j o . {Aw)2dx
We use (4.5.21) to formulate a very useful variational principle. Let a, b

and R satisfy (4.5.6). Define
(4 5 22)
'<"*> = JByd* + JBb^ • --
and set
J{a,b) = inf €{(p,?p)
where the infimum is taken over all pairs of radial functions ((p,ip) ^ (0,0)
with (p e H2{Ba) n H&{Ba) and ip e H2{Bb) D H&(Bb) such t h a t
dl d
.N •l Pf~\ = b"-^(b).
a"-^(a) uN-\ ^ (4.5.23)
ar ar
Proposition 4.5.2 There exists a pair (tpo,ipo) such that S((po,ipo) —
J(a,b). Further,
A2ip0 - (i<po in B,
2,0
A2tp ,. in
= iiipQ ,„ „l}
Bb (4-5.24)
together with the condition (in addition to (4-5.23), (po ~ 0 on dBa and
ipQ = Q on dBb, built into the set of admissible functions for the minimiza-
tion problem)
Ay? 0 (a) + AipQ(b) = 0. (4.5.25)
Finally, we also have
fi = J(a,b).
P r o o f : We can look for (ip, ip) satisfying the required conditions as well as
/ <p2dx + / ip2dx = 1
JBa JBb
when looking for a minimizer of £. Since, J ( a , b) < oo, if {((pn, ipn)} is such
a minimizing sequence, it follows from t h e fact t h a t z •—> || A U / | | 2 , B is a norm
on H2(B) n HQ(B) for any bounded domain B equivalent to the Sobolev
norm ||2||2,2,B, t h a t {<pn} and {•ipn} are bounded in H2(Ba) and H2(Bb)
respectively. Hence, for a subsequence, tpn —>• <p0 in H2(Ba) D H^Ba) and
fpn —*• ipo in i / 2 ( S b ) n / f o ( B ( , ) weakly. Since the norm is weakly sequentially
lower semi-continuous, it follows t h a t €(<po,tpo) — J(a,b). This establishes
the existence of a minimizer for £.
Now, (fo,ijJo) satisfies the Euler equations which can be written as
/ ( A V o -ti<po)<pdx+ / (A2V>o " vipo)ipdx

JBa JBb
+ I A<p0^da+ J AiPo^-da = 0 (4.5.26)
JdBa du JdBb dv
for all admissible pairs ((p,ip). Here fi is a Lagrange multiplier. Taking,
successively, (Ba) and ip = 0 and then ip = 0 and ip G T>(Bf,), we get
(4.5.24). Using this in (4.5.26) and t h e fact t h a t t h e pair ((p}ip) satisfies
(4.5.23) and t h e fact t h a t all these functions are radial, we get (4.5.25).
By applying Green's formula to the numerator of £(<po,ipo) using the
relations (4.5.24) and the various boundary conditions, we easily deduce
t h a t /J, = J(a,b). This completes the proof. •
It now follows from (4.5.21) t h a t Ai(fi) > J(a,b). If we show t h a t

J(a,b) > J(0,R) for all 0 < a m i > : ( A t 2 ' X = Ai(n-)

w
Jn* w dx
which will prove the conjecture.
At this point, the symmetrization arguments cease and it now becomes

an independent minimization problem treated, using different approaches,
by [Nadirashvili (1995)] for N = 2 and by [Ashbaugh and Benguria (1995)]
for N = 2,3. We sketch the latter method. The proofs are involved and
technical, and are based on a difficult and intricate analysis involving the
Bessel functions.
From (4.5.24), it follows that we can write
<Po{r) = [AJu(kr) + BIv(kr)] r~v

Mr) = \CJu(kr) + DIv{kr)] r~v
where we set v = -y — 1 and fi = k4, The number k (and hence \x = J(a, b))
is now determined from the boundary conditions stated in the preceding
proposition. Thus, we get four homogeneous linear equations for the un-
knowns A,B,C and D. A nontriviai solution exists if the determinant of
the corresponding matrix vanishes. This relation takes the form
hu{k) = Mka) + fv(kb) = 0 (4.5.27)
where
Jv+i{x) I„+i{x)
f„(x) = x
JJx) + IJx)
Let us assume, without loss of generality, that aN + bN = 1. Let ku{a)

denote the first positive zero of hu and let kv denote the first positive zero
of fu. In fact, kv = Ai(i?i), which can be proved via the identity
Jv{x)lv+\(x) + Jv+\(x)Iu(x) = xl~N Ju{x)Iy(x).
Thus, in order to show that J (a, b) > J(0,1), we need to show that ^ ( o ) >
&i,(0) = ku. This is the main body of the work done by [Ashbaugh and
Benguria (1995)]. In particular, a necessary condition for the above to hold
turns out to be
where, as usual, j V t \ denotes the first positive zero of Jv. This inequal-
ity holds only for N = 2 and N — 3. Hence their proof works only for
these dimensions. [Ashbaugh and Laugesen (1996)] show that, for all space
dimensions iV,
kv{a) > ku(2N) = 2&ju>i

for 0 < a < 2 ^ . Thus,
Ai(0) > 2*j}tl = CivAi(fi*)
where
which tends to unity as N —» oo, thus improving on Talenti's inequality.
4.6 The Buckling Problem
Let fl c I 2 be a bounded domain in the horizontal plane occupied by

the middle surface of a thin plate. Let us assume that the plate is clamped
along the boundary. If this plate is subjected to a uniform compressive load
acting laterally on the boundary, then, upto some point, the plate will just
be compressed in the plane itself. Beyond a certain critical load, the plate
will then buckle out of the plane. In the language of bifurcation theory (cf.,
for instance, [Kesavan (2004)]), if the strength of the force is parametrized
by a parameter C? then for all values of £, we have the trivial solution,
viz. the zero vertical displacement. When £ crosses a critical value, non-
trivial solutions appear and we are at a bifurcation point. This situation is
modelled by the von Karman equations, which is a nonlinear problem, and
the critical values when bifurcation takes place are the eigenvalues of the
linearized problem. This eigenvalue problem is given by
A2u = -CAu in n
du = o
|H on on.r (4 6 i:
"-
Once again, the above problem admits an increasing sequence {Cu(^)}
of positive eigenvalues tending to infinity and an associated orthonormal
basis of eigenfunctions (which belong to HQ(U)). The first eigenvalue fi(fi)
is called the buckling load.
Inspired by Rayleigh's conjecture for vibrations of plates, Polya and
Szego conjectured that the circular plate has the minimal buckling load
amongst all plates of equal area. The problem (4.6.1) and the conjecture
of Polya and Szego can be readily extended to all dimensions. Thus, we
assume, henceforth, that Q, C RN is a bounded domain and the conjecture
can now be written as
Ci(n) > Ci(fl'). (4.6.2)

Szego proved this conjecture when the first eigenfunction is of constant

sign in Q. Again, this is an unreasonable assumption since, even for fairly
simple domains, the first eigenfunction changes sign and the first eigenvalue
is not simple. Nevertheless, these properties are indeed true for a ball.
This conjecture remains open even now. Unlike the case of plate vibra-
tions, it has not been proved even for low dimensions. The best result, as
of now, is that of [Ashbaugh and Laugesen (1996)] who show that there
exists a constant d^ > 0 which tends to unity as N —> oo, such that
C i W > dN<;(n*). (4.6.3)
We present these arguments below.

As usual, the eigenvalues admit a Rayleigh quotient characterization.
In particular,
Cl(n) = min k^L. (4.6.4)
Exercise 4.6.1 Show that
Ci(fi) > Ai(n)*
and that
Ai(n) > Ai(fi)Ci(fi)
where Ai(O) is the first eigenvalue of the vibrating plate problem and Ai(fi)
is the first (Dirichlet) eigenvalue of the Laplacian.
The proof of Szego's result follows the same lines as the one in the case
of vibrating plates (cf. Theorem 4.5.1).
Theorem 4.6.1 (Szego) Let us assume that the first eigenvalue Ci(^)
admits an eigenfunction which is positive in H. Then (4-6.2) holds.
Proof: Let w > 0 be such an eigenfunction. Set / — — Aw. Consider the

symmetrized problem
-Av = f* in IT
v = 0 on dfl.
Since w G HQ(U), we have that JQ fdx = 0 and so it follows easily that

| g = 0 on dQ* (cf. the proof of Theorem 4.5.1). Thus, v e H§(Q*). Now,
by Talenti's theorem (cf. Proposition 3.1.1), since w > 0 in f2, we have
/ \Vw\2dx < [ \Vv\2dx.

Jn Jn*
Thus,
This completes t h e proof. •
As already mentioned, in general, the hypothesis of the above theorem

is not verified for most domains and, in fact, no general criterion is known
to determine whether this is true for a given domain. We now present,
following [Ashbaugh and Laugesen (1996)], a proof of the weaker inequality
(4.6.3). We begin with Payne's inequality connecting fi(fi) and t h e second
(Dirichlet) eigenvalue of the Laplacian.
T h e o r e m 4.6.2 [Payne (1955)] Let A2(H) denote the second eigenvalue

of the Laplacian (with homogeneous Dirichlet boundary conditions). Then,
Ci(fi) > A 2 (fi). (4.6.5)
Proof: Let ipi G # o ( ^ ) be the eigenfunction associated to Ai(ft), the

first eigenvalue of the Laplacian and let w\ G HQ(Q) be an eigenfunction
associated to Ci(^)- Then,
M n ) = m i n kp^L ( 4.6.6)
v±<f! Jn v*ax
where the minimum is taken over all non-zero functions in HQ(Q) orthog-
onal (in L2(Q)) to v?i. For 1 < i < N, define
ipi = aiWi -f ——.

dxi
T h e n ipi G HQ(£1). We choose a^ such t h a t ipi is orthogonal (in L2(Q)) to

ipi. It is clear t h a t ipi =£ 0 since w\ G HQ(Q) while its first derivatives are
only in HQ(Q). Now, using the boundary conditions on w\ and integration
by parts, it is immediate to verify t h a t
f \ViPi\2dx = a2 J \Vwx\2dx+ [ dx
Jn Jn Jn dxi)
and that
Using ipi as a test function in (4.6.6) for each 1 
(Awi)2dx.
SJC ' ( S ) * - J£
dx =
Also,
/ \Vwi\2dx J (-Awi)widx < ( J (Awi)2dx) ( f w\ dx

Jn \JQ J Via
by the Cauchy - Schwarz inequality. Finally, since w\ is the an eigenfunction
associated to Ci(^)> w e have
f(AWl)2dx = diU) f \Vwtf 'dx
and so
/ i v ^ i 2 < Ci(n) f w\ dx.

Using these observations, we get
&?=!<$) fa\Vvi\2<te + fn(*v>i)2dx

A2(fi) < < Ci(fi)
(Zli°Z)JnV2d* + fn\Vwi\2dx
Lemma 4.6.1 Let ft C MN be a bounded domain. With the usual nota-
tions,
A 2 (n) > 2*jltl (4.6.7)

\n\
where v = y — 1.
Proof: Let (p2 G Hl{£l) be an eigenfunction associated to A2(f2). Since if 2

has to change sign in Q, we have two non-empty open sets
Q+ = {<fi2 > 0} and Q- = {(fi2 < 0}.
In each component of these open sets, ^2 will be of constant sign, will

vanish on its boundary and will verify — Ay>2 — ^2(^)^2- Thus, in each
such component, we have that \i($£) is the first eigenvalue of the (Dirichlet)
Laplacian. If C C fi± is one such component, then since the first eigenvalue
is monotonically decreasing with the domain (as is easy to see from the
variational characterization), we have A2(fi) = Ai(C) > Ai(H±). Thus,
2
UN
A2(fi) > Ai(Ot) > A i ( n i ) - jl%l
|0±l
(We cannot say that A2(fi) = Ai(H±) since these subdomains may not be
connected.) Thus,
UJW
2A2(fi) >jlilU
.m+l* in. I* * ^ 2 , r I ffl
since |n+| + |fi_| < |fi| and the function t •—• £~"tf is monotonic increasing
for t > 0 and is convex.
In case equality holds, then retracing the proof, we see that both Q+ and
fi_ will be balls (by the case of equality in the Faber - Krahn inequality)
and both of equal volume (equal to |fi|/2). In this case, Q will no longer
be connected. Thus, for bounded domains, we do have a strict inequality
and this completes the proof. •
Combining (4.6.5) and (4.6.7), we immediately get the following result.
Corollary 4.6.1 (Bramble - Payne Inequality) Let N = 2 and let fi C R 2

be a bounded domain. Then
Ci(n) > 27rj» 1 |fi|- 1 .
Lemma 4.6.2 Let v = y — 1. Let n C M.N be a bounded domain. Then

2_
Ci(n*) = J2+1.1 ( j g ) " • (4-6.8)

Proof: Without loss of generality, we can assume that |Q| = UJ^, i.e. 0*
is the unit ball in WLN. Define
w{x) = JvUv+iMW1* -Jv(jv+i,i)-
Then w is smooth and vanishes on dQ* = {\x\ — 1}. Further, from the
properties of Bessel functions, we also have that
d fJu(t)\
SJv(t J„+i(t)
dt \ tu t"
Thus f^f = o on dn* as well. Thus, w € flg(fi*). Now, since w is radial,
we will set, by abuse of notaion, w(x) = w(\x\). Then,
_ d2w N ~ 1 dw
dr2 r dr
After a straight forward computation, it also follows, from Bessel's equation,
that
&w + $+IA(W + Jvtiv+i,i)) = °-
Hence,
AWjJ+i.iAti; = 0
and so
j2+i,i > Ci(n*)-

On the other hand, by Payne's inequality (4.6.5)
Ci(n') > A2(n*) = # + u .

Thus, 0(fi*) = JH-I,I f° r t n e unit
ball an
d (4.6.8) is proven.
Combining (4.6.5), (4.6.7) and (4.6.8), we get (4.6.3) with
dN = 2 * ^
72
We then have
dN = 1 „4-_log4+0(iv_§)
which tends to unity as N —» oo.

4.7 Comments
In this section, we survey some related results and also indicate several
fascinating open problems relating to the eigenvalues of the Laplacian and
the biharmonic operators.
The Faber - Krahn inequality for the principal eigenvalue of the Lapla-
cian opens up an entire line of investigation of isoperimetric inequalities for
eigenvalues. In particular, one can pose the question for all the eigenvalues.
For each positive integer k, is there an optimal domain which minimizes
Afc, the k-th eigenvalue of the Dirichlet Laplacian, amongst all domains of
given volume? To start with, the second eigenvalue is minimized, amongst
all open sets of given volume, by the union of two disjoint balls each with
half the given volume (cf. Lemma 4.6.1). This, however, is a set which is
not connected. Unfortunately, the minimum is not attained over connected
sets. One can consider two disjoint balls connected by a thin tube and thin
out the tube so that, in the limit, we obtain two disjoint balls. Thus we can
produce a sequence of connected open sets so that their second eigenvalue
tends to the minimum value, which, however, is attained for a disconnected
set. One can try asking the question for convex domains. The existence
of an optimal convex domain which minimizes the second eigenvalue has
been established by [Cox and Ross (1995)]. It was widely believed that the
optimal convex domain, which minimizes the second eigenvalue amongst all
convex domains of given area in the plane, is the 'stadium' i.e. the closed
covex hull of two equal circles which touch each other. However, [Henrot
and Oudet (2001)] have recently shown that this is not the case.
It has been shown by [Bucur and Henrot (2000)] that there exists an
optimal set (connected, in case of dimensions 2 and 3) which minimizes the
third eigenvalue. Though one does not know yet what this domain looks
like, it is conjectured that in the case of dimensions 2 and 3, it is the ball and
that it is the union of three equal disjoint balls for higher dimensions. The
problem remains open for all higher eigenvalues as well as for all eigenvalues
with other boundary conditions (like the Neumann problem).
Another interesting problem is the optimization of the principal eigen-
value in domains with an obstacle. Let fi C RN be a bounded domain
and let w C fi be a subdomain. We consider the eigenvalue problem
(with Dirichlet boundary conditions) for the Laplacian in the domain U\uJ.
The conjecture is that the first eigenvalue is minimal when the obstacle u>
touches the outer boundary Oil (i.e. the infimum of the first eigenvalue,
amongst all possible positions of the obstacle, is reached as the obstacle
approaches the boundary) and that the maximum occurs when the obsta-
cle is close to the 'centre' of ft. For results in this direction, see [Harrell,
Kroger and Kurata (2001)], though much remains to be done. In the spe-
cial case, when ft and u are both balls, [Kesvan (2003)] has shown that the
maximum indeed occurs when the balls are concentric (see also Ramm and
Shivakumar at www.math.ksu.edu/ramm/r.html, publication 383).
There are several results and open questions regarding optimization of
various eigenvalues of the Laplacian under diverse constraints. For a nice
survey of these, see the article of [Henrot (2003)].
Coming to the ratios of eigenvalues, [Payne, Polya and Weinberger
(1956)] made several conjectures in their original paper. Of these, only
two remain open. They are to show that
A2(ft) + A3(ft) < A2(ft*) + A3(ft*)
Ai(fi) " Ax(ft*)
for D e l 2 (and the corresponding result for (A2(ft) + ... +A;v+i(ft))/Ai(ft)
when ft C M N ) and to show that
A m +i(ft) < A m +i(fi*)

Am (ft) - Am(ft*)
w
for ft C R for all m > 1. Of course, we presented the proof of Ashbaugh
and Benguria for the case m — 1 in Section 4.4. [Ashbaugh and Benguria
(1993)] have also shown that
A4(ft) A2(ft*)
A2(ft) ~ Ai(ft*)
and the cases m = 2 and 3 follow from this stronger inequality. All higher
cases remain open as yet.
As far as isoperimetric inequalities for eigenvalues of the biharmonic
operator (A 2 ) are concerned, almost all the corresponding problems are
open. One of the most recent results concerning the buckling problem
(cf. Section 4.6) is that of [Ashbaugh and Bucur (2003)] where they prove
the existence of an optimal domain for the buckling load. Willms and
Weinberger have indicated how to prove the conjecture for the buckling
load (cf. (4.6.2)) if there exists an optimal domain which is regular (C2).
The above result is a positive step in that direction. The proof involves
variational and domain differentiation techniques.
Finally, there are several results and open problems regarding universal
inequalities for eigenvalues, i.e. inequalities between various eigenvalues
that are valid for all domains. The interested reader is referred to the
survey article of [Ashbaugh (1999)].
Chapter 5
Nonlinear Problems
5.1 Payne - Rayner Type Inequalities
In this chapter, we will look at some isoperimetric inequalities associated

to solutions of some nonlinear problems and apply them to get symmetry
results for solutions in a ball.
Let n C M^ be a bounded domain. Then, as a consequence of the
Holder inequality, we have the continuous inclusion
Thus, for all u G L2(Q,), we have ||U||I,Q < C||u|l2,n- If u = yi > 0, the first
eigenfunction of the Dirichlet Laplacian (cf. Section 4.1), and if N = 2,
[Payne and Rayner (1972)] showed that
This inequality is known as the Payne-Rayner inequality or as a reverse

Holder inequality. This inequality has been extended in several directions.
It was generalized to higher dimensions by [Payne and Rayner (1973)] them-
selves but their inequality involves the first eigenvalue of an auxiliary prob-
lem and, for this reason, their generalization is not entirely satisfactory. It
was shown by [Kohler-Jobin (1977)] that
<pidx < " !N_2

Jd
where, as in the previous chapter, we have set v = y — 1. A generalization

of the Payne-Rayner inequality for the first eigenfunction of second order
elliptic differential operators was also obtained by [Chiti (1982b)].
125
A more interesting generalization of this result is for solutions of semi-

linear equations of the form
-Au = f(u) in O \
u>0 in 0 > (5.1.2)
u = 0 on dCl J
where / : K —» M is a positive continuous function. In this case, [Payne,

Sperb and Stakgold (1977)] showed t h a t , when N = 2,
8TT / F(u)rfx < ( / /(u)da: ) (5.1.3)

Jo. \Jn )
where F is the primitive of / such t h a t F(Q) = 0. Notice t h a t , if we set
/(£) = Ai(Q)t, we recover the original Payne - Rayner inequality (5.1.1).
In this section, we will prove a further generalization of this inequality
to all dimensions and for the p-Laplacian. This result is due to [Mossino
(1983)]. We will also discuss the equality case and use it to derive a sym-
metry result for t h e solution of a class of nonlinear equations following the
work of [Kesavan and Pacella (1994)].
Let H C WLN be a bounded domain. Let / : M. —> R be a continuous
function. Let 1 < p < oo. Consider the following boundary value problem:
u = 0 on m. }
} (5 1 4)
' -
T h e weak formulation of this problem reads as follows: Find u G WQ,P(Q)

such t h a t , for all v G W 0 1,p (ft),
/ |Vu]p~2Vu.V^:r = f f(u)vdx. (5.1.5)
We assume, henceforth, t h a t the following hypotheses are satisfied by the

nonlinearity / :
(HI) There exist constants A > 0 and B > 0 such t h a t f(t) < A\t\s + B,
where
s e if p< N
N —v
and s > 0 otherwise.
(H2) For t > 0, we have f(t) > 0.

Nonlinear Problems 127
T h e hypothesis (HI) guarantees that a weak solution of (5.1.4) in

W 0 (fi) is, in fact, in Cha{Q) for a suitable a G (0,1) (cf. [Guedda,
1,p
and Veron (1989)]).

We can now state the generalised Payne - Rayner type inequality for
positive solutions of (5.1.4). In all t h a t follows, we assume t h a t 1 < p < co
and t h a t q is its conjugate exponent, i.e.
v~x + q~l = i.
It will be immediate to see t h a t if p = q = 2 and if N = 2, the inequality

(5.1.6) below is precisely t h a t of Payne, Sperb and Stakgold (5.1.3).
T h e o r e m 5.1.1 [Mossmo (1983)1 Assume that the hypotheses (Hi) and

(H2) are satisfied. Let F be the primitive of f such that F(0) — 0. Ifu > 0
is a solution of (5.1-4), then
i \ i /*,f3' ,i i x [ r
i2\i Jr)(N^N)q S9("^)F(u#(s))ds < / f{u)<k
(5.1.6)
P r o o f : Let t > 0. We set //(t) = \{u > t}\. Since (u\{u>ty)# = ti#|{u#>(},
we have
*(*) =f / /(«)<& = / " f(u*)ds.

J{u>t} JO
Step 1. Since u e W 0 1 , p (fi), if , for some t such t h a t 0 < t < M where M is

t h e m a x i m u m of u on U (recall t h a t u is, in fact, Holder continuous, thanks
to ( H i ) ) , the set {u — t) has positive measure, then, by Stampacchia's
theorem (cf., for instance, [Kesavan (1989)]), Vu = 0 almost everywhere on
this set, and hence, on a set of positive measure as well. Thus, f(u) = 0 on
this set as well, contradicting (H2). Thus, \{u = t}\ = 0 for all 0 < t < M
and so ji(t) is continuous on (0, M ) . Consequently, we have u * ( / i ( t ) ) — t
on this interval and so,
*'(*) = / ( « # ( / i ( t ) ) / i ' ( t ) = f(t)n'(t). (5.1.7)
Step 2. We now use t h e variational formulation of (5.1.4), viz. (5.1.5). We

set v = (u — t ) + , as in the proof of Theorem 3.1.1, for 0 < t < M to get
~ f \Vu\vdx = f f(u)dx = *(t).

dt
J{u>t} J{u>t}
Again, an application of Holder's inequality yields (as in Theorem 3.1.1),
~ [ \Vu\dx) < -/i'(t)*(t)5

at
J{u>t} J
Then, by a combination of the co-area formula and the isoperimetric theo-
rem (cf. Corollary 2.2.3), we get
( J V a ; | ) V W ' - * < (Px»({u>t}))q < -//(«)*(*)*• (5.1.8)
Multiplying this inequality throughout by / ( t ) , which is strictly positive,

and using (5.1.7), we get
( M 4 ) W * / W < -*(i)2*'(i). (5.1.9)
Step 3. We now integrate (5.1.9) on both sides with respect to t between 0

and M. The right-hand side gives
f0M-Ht)iv(t)dt = -±f0M£mr)dt
< i*(o)9
since 3>(M) = 0 (the inequality appears in the above computations since

we only know that $ is monotonically increasing). On the other hand, we
have
/ 0 M / ( « ) / * ( * ) ' - * * = /o M f(t) (Q ~ # ) / 0 " W J-*-1**
= l(l-jf) J0M f(t) C s"-^X{u#>t)(S)dsdt
= i (i - jr) C /W C s«i-^X{u*>t}(s)dsdt
=«(i4)/rl/;,l"/«''('"i)^
Combining these two results, we get (5.1.6) and the proof is complete. •
We now treat the equality case.

Theorem 5.1.2 [Kesavan and Pacella (1994)1 Equality is achieved in

(5.1.6) if, and only if, Q is a ball and (up to a translation of the origin)
u = u*.
Proof: Step 1. Retracing the proof of the preceding theorem, we see that,
if equality is achieved in (5.1.6), then it also the case with (5.1.9) for almost
every t such that 0 < t < M. Since $'(t) = /(£)/*'(t) and since f(t) > 0,
we also have equality throughout in (5.1.8). Thus, for almost every t, the
level sets {u > t} obey equality in the classical isoperimetric inequality and
hence are balls. In particular, as in the proof of Proposition 3.2.2, we can
deduce that all level sets {u > t} and S7 are balls. So, by translating the
origin, if necessary, we may assume, henceforth, that n = fi*.
Step 2. Let 0 < t < M. Then, the set of all points where Vu(x) = 0 and
u(x) = t is empty. To see this, we first observe that if >£ > 0 in a bounded
domain G, and if the set
C = {xeG\ *{x) = 0}
has empty interior, then, by a result analogous to the Hopf boundary lemma
(cf. [Guedda, and Veron (1989)]), §^ < 0 at all points of dG, where
w G Cl{G) is a positive solution of the problem:
-div(|Vw| p - 2 Vw) = * in G
(5.1.10)
w — 0 on dG.
Now, setting G = fit — {u > t}, and w — u — t, we see that w satisfies

(5.1.10) for *(x) = f(u{x)) = f(w(x) + t) > 0. So, if G = fit, the set
C is empty and thus, | ^ < 0 on <9fif, or, in other words, Vu / 0 on
dft,t — {u = t}. (In fact, if there exists x G {u = t} but not on 5fit.,
then, for every e > 0, we have x G fit-e. By the convexity of level sets (all
level sets are balls), the entire line segment joining x to any point of dCtt
would then have the same property and so u would be constant on that line
segment (= t). This would contradict the fact that u G C1,a(Q) and that
^ < 0 on d£lf This proves our claim. In particular, the measure of the
set of all points x G fi such that Vu(x) ~ 0 and 0 < u(x) < M is zero.
Step 3. As a result of Step 2, it follows that (cf. Remark 2.3.6) ji(t) is an

absolutely continuous function. Now, from the fact that we have equality
in (5.1.9), we get
i = -(^*)-VWMW*"**W (5.1.11)
We can integrate this to get
/ /(u#( 5 ))<fa dg, (5.1.12)

Jo
using the change of variable £ — //(£).
Step 4. Consider the problem:

-div(|V^| p - 2 V^) - f(u*) in n - f2*
(5.1.13)
v= 0 on W .
As in Step 4 of the proof of Theorem 3.1.1, it is easy to show that this
problem has a radial, and radially decreasing solution given by
v(x) = (Nu^)~q / £*-# / f{u#(s))ds #

Ju>N\x\N JO
and its distribution function also satisfies (5.1.12). Since u* and v are both
radially decreasing and equimeasurable, it follows that v = u*.
Now, the function t *-> f(t)t is a Borel function and so
f \Vu\pdx = f f{u)udx = f f(u*)u*dx = / \Vu*\pdx

JQ JQ JQ* JQ*
since u* solves (5.1.13). Since u* = v is now given by the formula above, a
straight forward computation reveals that, thanks to the hypothesis (H2),
Vu*(x) ^ 0 when 0 < u*(x) < M. Thus, by the theorem of Brothers and
Ziemer (cf. Theorem 2.3.3) we have that u = u*. •
We conclude this section with an application of the above result.

Theorem 5.1.3 [Kesavan and Pacella (1994)] Assume that (HI) and
(H2) are satisfied and that p — N. Then, every positive solution of (5.1-4)
when Q — BR, the ball of radius R, is radial and radially decreasing.
Proof: For a solution u G WQ'P(Q) of (5.1.4), we have the following Po-
hozaev type identity (cf. [Guedda, and Veron (1989)])
du
[ F(u)dx+(l ) / }{u)udx = - J (XM) da (5.1.14)
JQ \ P J JQ Q JdQ dv
where v is the unit outward normal on the boundary. If H = BR, then

v = x/R. By Green's formula, we get
du
[ f(u)dx = f \Vur2^da = f da.
av
Jn Jdn Jdsi
Applying Holder's inequality to the last term, we get
du
f f(u)dx
)dx << (NuNRN-l)i(f
(NLJN
Thus, Pohozaev's identity (5.1.14) now gives

1<?
N f F(u)dx+(l- —) f f{u)udx > -
R )dx
q(NujNRN-l)p JBR
JBR \ P J J BR
(5.1.15)
since x.v — R on t h e boundary of BR. Now setting p — N, we get q —
p/(p - 1) = N/(N - 1) so t h a t (5.1.15) becomes
N2
/ f(u)dx (NuN)"~ w1-1 / F(u)dx. (5.1.16)
JBR N -1 JBR
On the other hand, (5.1.6) yields,

N
JV-1
N2
[ /(«)dx N
> -(NUJN)^
1
J F(u)dx.
JBR
(5.1.17)
J BR
Thus, from the above two inequalities, we see t h a t equality is attained in
(5.1.6) and so, by the preceding theorem, we have t h a t u — u*t which com-
pletes the proof. •
R e m a r k 5 . 1 . 1 W h e n p = 2, the p - Laplacian reduces to the Laplace oper-

ator. T h e result of Theorem 5.1.3 was first given by [Lions (1981b)] in this
case and, for quite some time, several unsuccesful a t t e m p t s were made to
produce such a proof for all dimensions for semilinear equations involving
t h e Laplacian. T h e real generalization of Lions' result is as above, for the
iV-Laplacian in TV dimensions. •
R e m a r k 5 . 1 . 2 It was shown by [Gidas, Ni and Nirenberg (1979)] that, in

t h e case of the Laplacian, the symmetry result is true for all nonlinearities
(albeit with some smoothness assumptions). However, we need the hypoth-
esis (H2) t h a t f(t) > 0 for t > 0. This is not unreasonable considering the
fact t h a t [Kichenassamy and Smoller (1990)] have constructed an exam-

ple of a function / with zero mean (hence of changing sign) such t h a t the
problem (5.1.4) admits a non-negative and non-radial solution in t h e ball
for p ^ 2. This difference in the behaviour of the p-Laplacian when p / 2
is because in those cases, in general, we do not have t h e strong maximum
principle due to the degenerate n a t u r e of the operator. For more comments
on symmetry of solutions, see Section 5.3. •
5.2 A S y s t e m of S e m i l i n e a r E q u a t i o n s
In this section, we will establish an isoperimetric inequality of Payne -

Rayner type for a system of semilinear equations. We will also show t h a t
equality occurs only in the spherically symmetric case. As in the previous
section, with t h e help of a Pohozaev type identity, we will show t h a t the
solutions in a ball are spherically symmetric. However, in this case, we will
need to restrict our attention to the case when t h e space dimension N = 2.
Let fi C MN be a bounded domain. Let fi : R —> K, 1 < i < m
be a family of continuous functions which are positive and monotonically
increasing. Let A — {a%j) be a real symmetric matrix of order m with
non-negative constant coefficients. Let u = (ui) € (HQ(Q.) r\C(Q))m be a
(weak) solution vector to the problem:
-Aui = fi(Yl™=i aijuj) in

^
\ \<i<m. (5.2.1)
U{ = 0 on dQ
Since the fi are all positive, it follows from the maximum principle t h a t
all t h e Ui are positive. To exclude the case where an equation decouples
itself from the rest, we assume t h a t in each row of the matrix A, there is
at least one off-diagonal coefficient which is non-zero.
We now define w G ( ^ ( f i ) ) ™ by
m
Wi = y j a j j U j , 1 < i < m. (5.2.2)
j=i
Since the coefficients of A are all non-negative, it follows t h a t t h e Wi are

positive as well. Let Fi denote the primitive of fi such t h a t Fi (0) — 0 for
1 < i < m. We then define
mi(r) = f fi(w*)dx and M{(r) - f F{w*)dx (5.2.3)

for 0 < r < R and 1 < i < m, where i? is the radius of fi*.
Lemma 5.2.1 For all I < i < m and for almost every 0 < r < R, we
have
m
N x
-NuNr - w*\r) < Ôijmj-(r). (5.2.4)
Proof: Since w* (r) is a monotonically decreasing function (as we have done

several times before, we write, by abuse of notation, w*(x) = w*(\x\)), its
derivative exists almost everywhere. If this derivative vanishes for some
r, then (5.2.4) is trivially true. If the derivative is non-zero, then, clearly,
Br = {w* > c] for some c > 0. Thus, it is enough to prove (5.2.4) when
Br is a level set of w*.
It follows immediately from the definition that w satisfies the system of
equations:
i 1 0 be less than the maximum value of w^ and set S \ c = {wi > c}.
Let rc be the radius of H* c . Then,
-NujNr^wf(rc) = - j ^ °£d* = J ^ \Vw*\da
- /n«,c Aw*dx = HT=i /oi. c a

ijf3(wj)dx
<J2T=iavIniJj(wj)dx-
The inequality in the second line above comes from the classical isoperi-
metric inequality, as in the proof of the Polya - Szego theorem (cf. Step
2 of the proof of Theorem 2.3.1). The last inequality above uses the fact
that, when fj is monotonically increasing, then, (fj(u)j))* = fj{w*) and
the property (1.3.2).
The result now follows for any Br which is of the form fi*c and the
proof is complete. •
Theorem 5.2.1 [Kesavan and Pacella (1999)] Let Q, C RN be a bounded
domain. Let fi : ]R —> R, 1 < i < m be a family of continuous, positive and
monotonically increasing functions. Let Fi be the primitive of fi such that
Fi{0) = 0 for 1 4N(N-l)uN^2 \x\N-2Fi(w*(x))dx (5.2.6)
where, for 1 < i < m,
fhi =
I fi(u)i)dx. (5.2.7)
Jn
Further, equality is achieved in this inequality if, and only if, Q is a ball
and all the Wi are radially symmetric and radially decreasing.
Proof: Let Mi and rrii be given by (5.2.3) for 1 N(N -l)uNrN-2Ft(wUr)) - ft(w;(r))J2alJmJ(r).
j=i
Multiplying throughout by Nuj^r1^-1 and using the expressions for the

first derivatives of Mi and m^, we get
771
N 2
NuN^^M'^r) > N(N -l)uNr - Ml{r) -Y,aijmj(r)m,i(r).
3=1
We sum the above inequalities over all 1 < i < m and integrate over the
interval [0, R] where R is the radius of fi*. On one hand, since M[(R) =
NtJNRN~lFi(0) = 0, we get, on integrating the left-hand side by parts,
pR t-R
\ NuNrN-xM'l{r)dr = -N(N-l)uN rN~2Ml(r)dr
Jo Jo
which, except for a change of sign, is exactly the first integral on the right-
hand side. On the other hand, since a^ = a^ we also have
J2 oymj(r)mJ(r) - -— J^ a^mii^m^r)
Since t h e rrii are all monotonic increasing, we thus deduce t h a t
rJV-2F<«(r))NwjvrJV-1dr.
i,j = l i=l ^°
Now
m^tf) = / fi{w*)dx — I fi(wi)dx = m*.
T h e integral on t h e right-hand side of t h e preceding inequality is precisely

the expression, on passing to polar coordinates, for t h e integral on t h e
right-hand side of (5.2.6) and so t h e proof t h a t inequality is complete.
If equality is achieved in (5.2.6), then, retracing t h e argument above,
we see t h a t for almost every 0 < r < R, we have equality in (5.2.4). Hence,
for almost every value c > 0,
/ \Vw*\da = [ \Vwi\da.
Integrating this over t h e range of values c of Wi, we get, using t h e co-area

formula,
/ \Vw*\2dx = / \Vwi\2dx
Jn* Jo.
for each 1 < i < m. Since t h e right-hand side of (5.2.4) is positive, it
follows t h a t , in the case of equality in (5.2.6), t h a t w* (r) < 0 for almost
every r. Thus by t h e theorem of Brothers and Ziemer (cf. Theorem 2.3.3),
we deduce t h a t (up t o a translation of t h e origin) Q, — Q* and that, for all
1 < % < m , W{ — w*. This completes t h e proof. •
Corollary 5.2.1 With the hypotheses and notations as above, if N — 2,

then
m m .
Y j aijfhiThj > 871"^ / Fi(vji)dx. (5.2.8)
P r o o f : W h e n TV = 2, t h e right-hand side of (5.2.6) reduces to

m ,, m „
8TTV / Fi{w*)dx - 8TTV / F^w^dx
i=i Jn* i=1 Jn
using the properties of the Schwarz symmetrization and so we get (5.2.8).
R e m a r k 5.2.1 Let B ~ (bij(x)) be a matrix of order N such that, for all

£ G R ^ and for almost all x G H, we have
K|2 < B(x)U < m2- (5.2.9)
Let u = (ui) G (HQ($l)nC(£l))m be a positive solution vector to the system:
mil
-div(BVm) = fi(ZT-i «««*) \,l<i<m.
Ui = 0 on dQ, J '
Let Wi be defined by (5.2.2). If c > 0 is a value of Wi, then,
-BVwtM = -Vwt.B^ = -^(BI/.I/) > - ^

of ov
where v is the unit outward normal on dH^c, thanks to the condition (5.2.9).
Now the proof of Lemma 5.2.1 will go through mutatis mutandis and so The-
orem 5.2.1 is still valid. Observe, further, since the isoperimetric inequality
(5.2.6) only involves the Wi, it is still true if w = (wi) satisfies
-div(B,Vt«() = £ ™ , o y / i t o ) i n n \ i < j < m

Wi = 0 on dCl J ' — —
where all the Z?j, 1 < i < m, satisfy (5.2.9). •
We now establish a Pohozaev type identity for solutions of the system

(5.2.1).
T h e o r e m 5.2.2 Let u be a solution vector of the system (5.2.1). Let w

be given by (5.2.2). Then
Y ] (2-N) f Vui.Vwidx + 2N [ Fi(wi)dx (5.2.10)

iî L Jn Vn
Proof: We start with the identity
div((z.Vuj)Vui) = Vui.(l + (x.V))Vuj + (a;.Vuj)A^

Integrating this by parts over Q, we get
f V7 dui dUi
•da
j X.VUj
x.VUj-^-
'an
= / Vui.Vujdx+ / Y^ -rr^Xki:—^—dx + i (x.VuAAuidx.
Jn îfjiidxi dx
k9xi Jn
Integrating the middle term on the right-hand side twice by parts and
collecting the terms, we get
du' du'
/ (X.VUJ)-^ + (x.Vui)—j- - (x.u^Vui.Vuj) da
J an
= (2-7V) / Vui.Vujdx+ j [(x.Vuj)&Ui + (x.Vui)&Uj]dx.
As already remarked earlier, since the Ui and u>i vanish on <9f2, we have
Vui = jfif-v and Vwi = ^ u on d$l. In view of this, the left-hand side of
the above relation can be simplified as
/ (x.i/)———-da.
Jdn du du
Multiplying the resulting identity by a%j and summing over i and j , we get,
using the symmetry of A,
„ m
dui duj
/ fa-") V] ai du du
N
f f ft
(2-TV) / Vui.Vwidx + 2 \ Y^ XkÂuidx
in ,/n~ dxk
The second term on the right-hand side can be computed, by integration

by parts, as follows:
m r N p. m ~ N n
^fc^—lAuidx - - 2 V / y^xk—-f(u>i)dx
1=1 Jil fc=l
i=1-•«*=! oxk e y e r dxp.k
m
n» N
xk {Fi{wi))dx
W^ ^
um „
Fi(wi)dx.
There are no boundary integrals since Wi vanishes on dQ, and Ft(0) = 0.

The identity (5.2.10) now follows immediately. •
We now consider the case when Q is a ball in the plane M2.
Theorem 5.2.3 Let Q — BR C R 2 ; the ball with centre at the origin

and of radius R. Let fi : IR —*• IR be positive and monotonically increasing
continuous functions for 1 < i < m. Let A be a symmetric matrix of
order m with non-negative coefficients. Assume, further, that A is positive
definite. Let u — (ui) G (HQ(£1) D C(Q))m be a positive solution vector
to (5.2.1). Then m is radially symmetric and radially decreasing for each
1 < i < m.
Proof: Let w be given by (5.2.2). Observe that v — x/R, where u is the

unit outward normal on the boundary. Thus x.v — R for x e dQ. Then,
since TV = 2, the Pohozaev identity (5.2.10) reads as
R E ^ ^ = 4 E / ^\{wi)dx. (5.2.11)
Let us set
1 f dui
Since — Aui — fi(wi), it follows, from Green's formula, that
ffii = -2-KRJJLU l<i<m. (5.2.12;
Since A is positive definite, we have
P m Q o m
ai da 27rR
= iJ / 2^ i7hJ~fr7 ~ 1^ ^jtoto-
Thus, in view of (5.2.12), we have
f v~^ dui du _ _ _
-ft / > Oii— T^dcr > —- > aumim-i
Thus, in view of (5.2.11), we get
\_] aijihifrij < 871-yj / Fi(u)i)dx.
But this is exactly the inequality (5.2.8) but in the opposite direction. Thus
it follows that we have equality in (5.2.8) and so, by Theorem 5.3.1, all the
Wi and hence, all the Ui, are radial and radially decreasing. •
R e m a r k 5.2.2 The fact that the differential operator was linear and the
structure of the nonlinearities helped us to define the auxiliary functions
Wi from the solutions Ui and derive the Payne - Rayner type inequality. If
we tried to imitate this procedure for the p-Laplacian, we would encounter
a term of the form
m
Y2 aijmj (r)mi(r)
i,j=X
which cannot be explicitly integrated in closed form. The positive deflnite-

ness was needed to obtain a suitable lower bound for the boundary integral
in Pohozaev's identity, in order to deduce the symmetry result. •
A similar problem was studied by [Chanillo and Kiessling (1995)]. In

their case fi(t) = exp(t) for all 1 < i < m. They not only assume that the
matrix is symmetric and with non-negative coefficients, but also that the
row sums are equal to unity. We do not need this hypothesis here. On the
other hand, since their problem is posed on all of M2, they are able to do
away with the boundary term in Pohozaev's identity and so they do not
need the additional hypothesis of positive definiteness of the matrix A.
5.3 Comments
In the preceding sections, we proved Payne - Rayner type isoperimetric in-

equalities for positive solutions of certain nonlinear problems. Then, using
a Pohozaev type identity, we deduced the radial symmetry of positive so-
lutions in a ball. As already mentioned, the idea of such a proof goes back
to [Lions (1981b)] who considered the solution of a semilinear equation in-
volving the Laplacian in a ball in the plane. In that paper, it is implicitly
assumed that equality in the isoperimetric inequality occurs only in the
spherically symmetric situation. This was rigorously proved in the general
case of the p-Laplacian by [Kesavan and Pacella (1994)], who also found
the correct generalization of Lions' result to all dimensions.
Of course, as already observed, the symmetry result here depends heav-
ily on the fact that the nonlinearity is positive. Again, for the p-Laplacian,
this is quite reasonable. However, for the Laplacian, any C1 function /
will do. The study of symmetry properties of solutions based on sym-
metry properties of the domain first started with the work of [Gidas, Ni
and Nirenberg (1979)] (though the ideas were already present in earlier
works of Alexandrov and of Serrin). It was further refined and improved by
[Berestycki and Nirenberg (1991)]. These depend heavily on the availability
of maximum principles and the method is generally known as the method
of moving planes.
Strong maximum principles are generally not available for degenerate
operators like the p-Laplacian. While symmetry results using variants of
the moving plane method do exist (cf. (Grossi, Kesavan, Pacella and Ra-
maswamy (1998)], [Damascelli, Pacella and Ramaswamy (1999)] and [Dam-
ascelli and Pacella (2000)], for instance), mention must be made of the
work of [Brock (2000)], who uses symmetrization techniques to prove such
results. He has developed a procedure known as continuous Steiner sym-
metrization and has applied it to obtain symmetry results for solutions of
the p-Laplacian.
There are several other applications of symmetrization to nonlinear
problems. Of course, the study of free boundary value problems, like the
obstacle problem (cf. Section 3.4) can already be thought of as falling into
this category. We also refer the reader to the works of [Bandle and Sperb
(1983)], [Mossino (1979)] and [Mossino and Temam (1981)] for isoperimetric
inequalities for free boundary value problems occuring in models describ-
ing the physics of plasmas. See also [Bandle (1976b)] for an isoperimetric
inequality for solutions of a nonlinear eigenvalue problem. These are just a
few instances of applications of symmetrization and the literature is vast,
rich and varied.
Needless to say that there is a wealth of literature on isoperimetric
inequalities for linear and nonlinear problems using techniques other than
symmetrization or techniques depending on other types of symmetrization.
References to such works could be found in those cited in this text.
Bibliography
Adams, R. A. (1975) Sobolev Spaces, Academic Press.

Almgren, F. J. Jr. and Lieb, E. H. (1989) Symmetric decreasing rearrangement
is sometimes continuous, J. Amer. Math. Soc, 2, 4, pp. 683-773.
Alvino, A., Diaz, J. I., Lions, P. L. and Trombetti, G. (1996) Elliptic equations
and Steiner symmetrization, Comm. Pure Appl. Math., 49, pp. 217-236.
Alvino, A., Lions, P. L. and Trombetti, G. (1986) A remark on comparison
results via symmetrization, Proc. Roy. Soc. Edin., 102 A, pp. 37-48.
Alvino, A., Lions, P. L., and Trombetti, G. (1990) Comparison results for ellip-
tic and parabolic equations via Schwarz symmetrization, Ann. Inst. Henri
Poincare Anal. Non Lineaire, 7, 2, pp. 37-65.
Alvino, A., Lions, P. L. and Trombetti, G. (1991) Comparison results for elliptic
and parabolic equations via symmetrization: a new approach, Diff. Int.
Eqns., 4, 1, pp. 25-50.
Alvino, A., Matarasso, S. and Trombetti, G. (1992) Variational inequalities and
rearrangements, Rend. Acad. Naz. Lincei, Serie IX, 3, 4, pp. 271-285.
Ashbaugh, M. S. (1999) Isoperimetric and universal inequalities for eigenval-
ues, Spectral theory and Geometry (Edinburgh, 1998), London Math. Soc.
Lecture Note Ser., 273, pp. 95-139.
Ashbaugh, M. S. and Benguria, R. D. (1992a) A sharp bound for the ratio of
the first two eigenvalues of Dirichlet Laplacians and extensions, Ann. Math.,
135, pp. 601-628.
Ashbaugh, M. S. and Benguria, R. D. (1992b) A second proof of the Payne -
Polya - Weinberger conjecture, Comm. Math. Phys., 147, pp. 181-190.
Ashbaugh, M. S. and Benguria, R. D. (1993) Isoperimetric bounds for higher
eigenvalue ratios for the n-dimensional fixed membrane problem, Proc. Roy.
Soc. Edin., 123 A, pp. 977-985.
Ashbaugh, M. S. and Benguria, R. D. (1995) On Rayleigh's conjecture for the
clamped plate and its generalization to three dimensions, Duke Math. J.,
78, 1, pp. 1-17.
Ashbaugh, M. S. and Bucur, D. (2003) On the isoperimetric inequality for the
buckling of a clamped plate, Z. Angew. Math. Phys., 54, 5, pp. 756- 770.
Ashbaugh, M. S. and Laugesen, R. S. (1996) Fundamental tones and buck-
141
ling loads of clamped plates, Ann. Scuola Norm. Sup. Pisa, Serie IV, 23,
pp. 383-402.
Aubin, T. (1976) Problemes isoperimetriques et espaces de Sobolev, J. Diff.
Geom., 11, pp. 573-598.
Bandle, C. (1975) Bounds for solutions of boundary value problems, Math. Anal.
AppL, 54, pp. 707-716.
Bandle, C. (1976a) On symmetrizations in parabolic equations, J. Anal. Math.,
30, pp. 98-112.
Bandle, C. (1976b) Isoperimetric inequalities for a nonlinear eigenvalue problem,
Proc. Amer. Math. Soc, 56, pp. 243-246.
Bandle, C. (1980) Isoperimetric Inequalities and Applications, Pitman.
Bandle, C. and Mossino, J. (1984) Application du rearrangement a une
inequation variationnelle, Ann. Mat. Pura AppL, IV, 138, pp. 1-14.
Bandle, C. and Sperb, R. (1983) Qualitative behaviour and bounds in a nonlinear
plasma problem, SIAM J. Math. Anal, 14, 1, pp. 142-151.
Berestycki, H. and Nirenberg, L. (1991) On the method of moving planes and the
sliding method, Boll, da Soc. Brasiliera di Mat., Nova Ser., 22, pp. 1-37.
Brandolini, B., Posteraro, M. R. and Volpicelli, R. (2003) Comparison results for
a linear elliptic equation with mixed boundary conditions, Diff. Int. Eqns.,
16, 5, pp. 625-639.
Brezis, H. and Nirenberg, L. (1983) Positive solutions of nonlinear elliptic equa-
tions involving critical Sobolev exponents, Comm. Pure AppL Math., 36,
pp. 437-477.
Brock, F. (2000) Continuous rearrangement and symmetry of solutions of elliptic
problems, Proc. Indian Acad. Sc, Math. Sc, 110, 2, pp. 157-204.
Brothers, J. E. (1987) The isoperimetric theorem, Centre for Mathematical
Analysis Research Report 5, University of Canberra.
Brothers, J. E. and Ziemer, W. P. (1988) Minimal rearrangements of Sobolev
functions, J. Reine Angew. Math., 384, pp. 153-179.
Bucur, D. and Henrot, A. (2000) Minimization of the third eigenvalue of the
Dirichlet Laplacian, Proc. Roy. Soc. Lon., 456, pp. 985-996.
Chanillo, S. and Kiessling, M. K.-H. (1995) Conformally invariant systems of
nonlinear PDE of Liouville type, Geom. Fund. Anal., 5, 6, pp. 924-947.
Chiti, G. (1982a) An isoperimetric inequality for the eigenfunctions of linear
second order elliptic operators, Boll. Un. Mat. Ital. A(6), 1, pp. 145-151.
Chiti, G. (1982b) A reverse Holder inequality for the eigenfunctions of linear
second order elliptic equations, Z. A. M. P., 33, pp. 143-148.
Coron, J.-M. (1984) The continuity of rearrangement in W1,P(K), Ann. Scuola
Norm. Sup. Pisa, Serie IV, 11, pp. 57-85.
Courant, R. and Hilbert, D. (1953) M e t h o d s of Mathematical Physics,
Volume I, John Wiley and Sons.
Courant, R. and Robbins, H. (1996) W h a t is Mathematics?, Second Edition
(Revised by Ian Stewart), Oxford University Press.
Cox, S. J. and Ross, M. (1995) Extremal eigenvalue problems for starlike planar
domains, J. Diff. Eqns., 120, pp. 174-197.
Damascelli, L. and Pacella, F. (2000) Monotonicity and symmetry results for p -
Bibliography 143
Laplace equations and applications, Adv. Diff. Eqns., 5, 7-9, pp. 1179-1200.
Damascelli, L., Pacella, F. and Ramaswamy, M. (1999) Symmetry of ground
states of p- Laplace equations via the moving plane method, Arch. Rational
Mech. Anal., 148, 4, pp. 291-308.
de Giorgi, E. (1954) Su una teoria generale della misura (r — l)-dimensionale in
uno spazio ad r dimensioni, Ann. Mat. Pura Appl., 36, pp. 191-213.
Evans, L. C. and Gariepy, R. F. (1992) Measure Theory and Fine Proper-
ties of Functions, Studies in Advanced Mathematics, CRC Press.
Faber, G. (1923) Beweiss dass unter alien homogenen Membranen von gleicher
Flache und gleicher Spannung die kreisformgige den leifsten Grundton gibt,
Sitz. bayer Acad. Wiss., pp. 169-172.
Federer, H. (1969) Geometric Measure Theory, Springer - Verlag.
Ferone, V. (1986) Symmetrization in a Neumann problem, Matematiche (Cata-
nia), 41, pp. 67-78.
Ferone, V. (1988) Symmetrization results in electrostatic problems, Ricerche
Mat, 37, 2, pp. 359-370.
Ferone, V. and Kawohl, B. (2003) Rearrangements and fourth order equations,
Quart. Appl. Math., 6 1 , 2, pp. 337-343.
Gidas, B., Ni, W. and Nirenberg, L. (1979) Symmetry and related properties
via the maximum principle, Comm. Math. Phys., 68, pp. 209-243.
Gordon, C., Webb, D. L. and Wolpert, S. (1992) One cannot hear the shape of
a drum, Bulletin (New series) AMS, 27, 1, pp. 134-138.
Grossi, M., Kesavan, S., Pacella, F. and Ramaswamy, M. (1998) Symmetry of
positive solutions of some nonlinear equations, Topol. Meth. Nonlin. Anal.,
12, 1, pp. 47-59.
Guedda, M. and Veron, L. (1989) Quasilinear elliptic equation involving critical
Sobolev exponents, Nonlin. Anal, Theory, Meth. Appl., 13, 8, pp. 879-902.
Hardy, G. H., Littlewood, J. E. and Polya, G. (1952) Inequalities, Cambridge
University Press, Second Edition.
Harrell II, E. M., Kroger, P. and Kurata, K. (2001) On the placement of an
obstacle or a well so as to optimize the fundamental eigenvalue, SIAM J.
Math. Anal., 33, 1, pp. 240-259.
Henrot, A. (2003) Minimization problems for eigenvalues of the Laplacian, J.
Evol. Equ., 3, 3, pp. 443-461.
Henrot, A. and Oudet, E. (2001) Le stade ne minimise pas A2 parmi les ouverts
convexes du plan, C. R. Acad. 5c. Paris, t . 332, Serie 1, pp. 417-422.
Kac, M. Can one hear the shape of a drum?, Amer. Math. Monthly, 73, pp. 1-23.
Kawohl, B. (1985) Rearrangements and C o n v e x i t y of Level sets in P D E ,
Lecture Notes in Mathematics, 1150, Springer-Verlag.
Kawohl, B. (1986) On the isoperimetric nature of a rearrangement inequality
and its consequences for some variational problems, Arch. Rational Mech.
Anal, 94, pp. 227-243.
Kesavan, S. (1988) Some remarks on a result of Talenti, Ann. Scuola Norm.
Sup. Pisa, Serie IV, 15, pp. 453-465.
Kesavan, S. (1989) Topics in Functional Analysis and Applications, Wiley
- Eastern.
Kesavan, S. (1991) On a comparison theorem via symmetrization, Proc. Roy.

Soc. Edin., 119 A, pp. 159-167.
Kesavan, S. (2002) The isoperimetric inequality, Resonance, 7, 9, pp. 8-18.
Kesavan, S. (2003) On two functionals connected to the Laplacian in a class of
doubly connected domains, Proc. Roy. Soc. Edin., 133 A, pp. 617-624.
Kesavan, S. (2004) Nonlinear Functional Analysis - A First Course, Texts
and Readings in Mathematics (TRIM), 28, Hindustan Book Agency, New
Delhi.
Kesavan, S. and Pacella, F. (1994) Symmetry of positive solutions of a quasi-
linear elliptic equation via isoperimetric inequalities, Applicable Anal., 54,
pp. 27-37.
Kesavan, S. and Pacella, F. (1999) Symmetry of solutions of a system of semi-
linear elliptic equations, Advances in Math. Sc. Appi, 9, 1, pp. 361-369.
Kichenassamy, S. and SmoUer, J. (1990) On the existence of radial solutions of
quasi-linear elliptic equations, Nonlinearity, 3, pp. 677-694.
Kohler-Jobin, M. T. (1977) Sur la premiere fonction propre d'une membrane:
une extension a N dimensions de l'inegalite isoperimetrique de Payne-
Rayner, Z. A. M. P., 28, pp. 1137-1140.
Krahn, E. (1924) Uber eine von Rayleigh formulierte Minimaleigenschaftdes
Kreises, Math. Ann., 94, pp. 97-100.
Lieb, E. H. (1977) Existence and uniqueness of the minimizing solution of
Choquard's nonlinear equation, Stud. Appl. Math, 57, pp. 93-105.
Lieb, E. H. and Loss, M. (1997) Analysis, Graduate Studies in Mathematics,
14, American Mathematical Society.
Lions, J. L. (1969) Quelques M e t h o d e s de Resolution des Problemes
aux Limites Nonlineaires, Dunod Gauthier-Villars.
Lions, P. L. (1981a) Quelques remarques sur la symetrisation de Schwarz,
College de France Seminar, Vol. 1, Eds. H. Brezis and J. L. Lions, Pit-
man Research Notes in Mathematics, 53, pp. 308-319.
Lions, P. L. (1981b) Two geometrical properties of semilinear problems, Appli-
cable Anal., 12, pp. 267-272.
Lord Rayleigh (1894) T h e Theory of Sound, 2nd Edition, Macmillan.
Maderna, C. and Salsa, S. (1984) Some special properties of solutions to obstacle
problems, Rend. Sem. Mat. Univ. Padova, 71, pp. 121-129.
Mossino, J. (1979) Estimations a priori et rearrangement dans un probleme de
la physique des plasmas, C. R. Acad. Sc. Paris, 288, Serie A, pp. 263-266.
Mossino, J. (1983) A generalization of the Payne - Rayner isoperimetric inequal-
ity, Boll. Un. Mat. Ital. A(6), 2, 3, pp. 335-342.
Mossino, J. (1984) Inegalites Isoperimetriques et Applications e n
Physique, Hermann.
Mossino, J. and Rakotoson, J. M. (1986) Isoperimetric inequalities in parabolic
equations, Ann. Scuola Norm. Sup. Pisa, Serie IV, 13, 1, pp. 51-73.
Mossino, J. and Temam, R. (1981) Directional derivative of the increasing re-
arrangement mapping and application to a queer differential equation in
plasma physics, Duke Math. J., 48, pp. 475-495.
Nadirashvili, N. S. (1995) Rayleigh's conjecture on the principal frequency of
Bibliography 145
the clamped plate, Arch. Rational Mech. Anal., 129, pp. 1-10.
Osserman, R. (1978) The isoperimetric inequality, Bull. A. M. S., 84, 6,
pp. 1182-1238.
Payne, L. E. (1955) Inequalities for eigenvalues of membranes and plates, J.
Rational Mech. Anal., 4, pp. 517-529.
Payne, L. E. (1962) Some isoperimetric inequalities in the torsional problem for
multiply connected regions, Studies in Mathematical Analysis and Related
Topics, University of California Press, Stanford.
Payne, L. E. (1967) Isoperimetric inequalities and their applications, SIAM
Review, 9, 3, pp. 453-488.
Payne, L. E., Polya, G. and Weinberger, H. F. (1956) On the ratio of consecutive
eigenvalues, J. Math. Phys., 35, pp. 289-298.
Payne, L. E. and Rayner, M. E. (1972) An isoperimetric inequality for the first
eigenfunction in the fixed membrane problem, Z. A. M. P., 23, pp. 13-15.
Payne, L. E. and Rayner, M. E. (1973) Some isoperimetric norm bounds for
solutions of the Helmholtz equation, Z. A. M. P., 24, pp. 105-110.
Payne, L. E., Sperb, R. and Stakgold, I. (1977) On Hopf type maximum princi-
ples for convex domains, Nonlin. Anal., Theory, Meth. Appl., 1, 5, pp. 547-
559.
Polya, G. Torsional rigidity, principal frequency, electrostatic capacity and sym-
metrization, Quart. J. Appl. Math., 6, pp. 267-277.
Polya, G. and Szego, G. (1951) Isoperimetric Inequalities in Mathematical
Physics, Ann. Math. Studies, 27, Princeton.
Polya, G. and Weinstein, A. (1950) On the torsional rigidity of multiply con-
nected cross-sections, Ann. Math., 52, pp. 154-163.
Posteraro, M. R. (1995) Estimates for solutions of nonlinear variational inequal-
ities, Ann. Inst. Henri Poincare Anal. Non Lineaire, 12, 5, pp. 577-597.
Posteraro, M. R. and Volpicelli, R. (1993) A comparison result in a class of
variational inequalities, Rev. Mat. Univ. Complut. Madrid, 6, 2, pp. 295-
310.
Protter, M. H. (1987) 'Can one hear the shape of a drum?' Revisited, SIAM
Review, 29, 2, pp. 185-195.
Szego, G. (1930) Uber einige Extremalaufgaben der Potentialtheorie, Math. Z.,
3 1 , pp. 583-593.
Szego, G. (1954) Inequalities for certain eigenvalues of a membrane of given
area, J. Rational Mech. Anal., 3, pp. 343-356.
Talenti, G. (1976a) Best constants in Sobolev inequality, Ann. Mat. Pura Appl,
110, pp. 353-372.
Talenti, G. (1976b) Elliptic equations and rearrangements, Ann. Scuola Norm.
Pisa, Serie IV, 3, pp. 697-718.
Talenti, G. (1981) On the first eigenvalue of the clamped plate, Ann. Mat. Pura
Appl. (Ser. 4), 129, pp. 265-280.
Talenti, G. (1985) Linear elliptic p.d.e.'s: level sets, rearrangements and esti-
mates of solutions, Boll. Un. Mat. Ital. B(6), 4, pp. 917-949.
Weinberger, H. F. (1962) Symmetrization in uniformly elliptic problems, Studies
in Math. Anal., Stanford University Press.
Weinberger, H. F. (1956) An isoperimetric inequality for the n-dimensional free

membrane problem, J. Rational Mech. Anal., 5, pp. 633-636.
Index
Bessel functions, 86, 90, 103 Faber-Krahn inequality, 85

modified, 106 Fleming - Rischell theorem, 28
zeroes of, 86 fundamental tone, 106
Bramble - Payne inequality, 120
Brunn - Minkowski inequality, 24 gap inequality, 101
buckling load, 116
Hardy - Littlewood inequality, 11, 14,
Cabre's lemma, 25 16
Chiti's theorem, 97 Hausdorff dimension, 22
co-area formula, 28, 31 Hausdorff measures, 21
co-area regularity, 40
conjecture isoperimetric inequality, 20, 23
Polya - Szego, 116 isospectral domains, 88
PPW, 99
Rayleigh (membranes), 85 level set, 1
Rayleigh (plates), 106 lower contact set, 24
Saint Venant, 73
Minkowski content, 21
de Giorgi perimeter, 22 moving planes, method of, 24, 140
Dido's problem, 19
Dirichlet principle, 71 Polya - Szego theorem, 35
distribution function, 1 Payne's inequality, 78, 118
Payne-Rayner inequality, 125
electrostatic capacity, 70 Payne-Sperb-Stakgold inequality, 126
electrostatic potential, 70 Pohozaev's identity, 130, 136
equality case
Faber - Krahn inequality, 87 Rayleigh quotient, 84, 90, 107, 117
Payne - Rayner inequality, 128 rearrangement
Talenti's theorem, 53 spherically symmetric
isoperimetric inequality, 26 decreasing, 13, 17
Polya - Szego inequality, 41 increasing, 16
equimeasurable, 4 unidimensional
147
decreasing, 2
increasing, 16
Riesz' inequality, 18, 39
Schwarz symmetrization, 13
Sobolev constant, 43, 44
Sobolev's inequality, 43
Szego-Weinberger inequality, 91
Talenti's theorem, 48
torsional rigidity, 73, 76
vanishing at infinity, 17
warping function, 73, 74

BN981 256-733-X
www.worldscientitic.com

(Series in Analysis 3) S. Kesavan-Symmetrization &amp; Applications-World Scientific (2006)

Încărcat de

Informații document

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

(Series in Analysis 3) S. Kesavan-Symmetrization &amp; Applications-World Scientific (2006)

Încărcat de

Drepturi de autor:

Formate disponibile

\

Vol. 1 Wavelet Analysis

Vol. 2 Differential Equations and Asymptotic Theory in

Vol. 3 Symmetrization and Applications

British Library Cataloguing-in-Publication Data

SYMMETRIZATION AND APPLICATIONS

Printed in Singapore by World Scientific Printers (S) Pte Ltd

The study of isoperimetric inequalities goes back to antiquity. The solution

tain elastic bodies. He conjectured that, of all fixed membranes of given

isoperimetric inequality (that I learnt from a lecture by X. Cabre) has been

of the cooperation extended by the members of our library staff (and, of

• u + = max{u, 0} denotes the positive part of u.

Let Vt C RN be an open set.

• C(fl) denotes the space of continuous functions on U.

• Ck(Q.) denotes the space of functions which are k times continuously

2. Some Classical Inequalities 19

5. Nonlinear Problems 125

1.1 The Decreasing Rearrangement

Schwarz symmetrization is a particular kind of rearrangement of functions

{u > t} = {x £ £1 | u(x) > £}.

This function is a monotonically decreasing function of t and for t >

Definition 1.1.1 Let Q C R ^ be bounded and let u : Cl —» R be a mea-

m e n t of u, denoted u # , is defined on [0, |fi|] by

u # (0) = ess. sup(u) 1

R e m a r k 1.1.1 Essentially, vft is just the inverse function of the distribu-

R e m a r k 1.1.2 In classical texts (cf. [Hardy, Littlewood, and Polya (1952)]

E x a m p l e 1.1.1 [Kawohl (1986)]

Let H = ( - 2 , 2) C K. Define u : ft -> R as follows.

Then, it is easy to check that

4 - 2t, 0 < t < 1

The following properties of the decreasing rearrangement are immediate

{t \ i*u{t) < si} C {t \fiu(t)<s2}.

Thus, by definition, it follows that u^(s\) > u*(s2).

Proposition 1.1.2 The mapping u \—• u# is non-decreasing, i.e. if u <

Proof: Since {u > t} C {v > t}, we have that

{t\ \{v>t}\<s} C {* | \{u>t}\<s}

and the result follows from the definition. •

We now show that u# is indeed a rearrangement of u in this sense.

\{u>t}\ = \{u*>t}\. (1.1.2)

{5 | u*{s) >t) C {s\ \{u > t}\ > s}.

Since u# is non-increasing, we have

\{u#>t}\ - sup{s | u#(s) >t} < |{u>t}|. (1.1.3)

|{«>«}| < |{ti # >OI- (1-1-4)

Applying (1.1.3) and (1.1.4) for t + h instead of t, we get

\{u#>t + h}\ < \{u>t + h}\ < | { w # > t + ft}|.

Passing to the limit as h j 0, we get

\{u*>t}\ ^ l{«>*}l < K« # >*}l

Corollary 1.1.1 With the preceding notations, we have

(Here the norms are the corresponding Lv - norms.)

Proof: If p — oo, then the result is contained in the definition of the

i«ii;,n = p / *p-vm* lu#llP,(o,!n|).

This result is also true without the non-negativity condition. In fact, as

Theorem 1.1.1 Let u : 0, —> R be measurable. Let F : R —• R be a

/ F(u{x))dx = \{u>t}\ = \{u#>t}\ = / F{u#(s))ds.

Remark 1.1.4 The result remains true for p = co since u*(0) —

We now prove another important property of the (decreasing) rearrange-

u*{s) > u(s) (1.1.8)

for all s £ [0,/]. Now, let s be a point of continuity for u. Since u is

\{u>u(s-h)}\ < s- h < s

u*{s) < u(s). (1.1.9)

iP(u*) = (ip(u)}#, a.e. (1.1.10)

Proof: Step 1. If u, w : [0,1] —> R are equimeasurable and non-increasing,

\{i>{u#) > t}\ = J^] X{ip(u#)>t}{s)ds = SaX{i>{u)>t}{x)dx

Corollary 1.1.5 For u : Q -> R, we have ( u + ) # = ( u # ) + .

(Series in Analysis 3) S. Kesavan-Symmetrization & Applications-World Scientific (2006)

(Series in Analysis 3) S. Kesavan-Symmetrization & Applications-World Scientific (2006)

\{u>t}\ ^ l{«>}l < K« # >*}l

i«ii;,n = p / p-vm lu#llP,(o,!n|).

/ * ( » ) - 11/ -3lloc,n < 9(s) < /(«) +||/-a||oo,n

C \f(s) - 9{s)\ds < jM(2h#(8) - f#(s) - g#(s))ds

u(x) = U#(LJN\X\N), xttt.

Ja, u(x)dx = / n . u#(uN\x\N)dx = fu#(ijNrN)NuNrN-1dr

Equality occurs if, and only if, {U\E)* — U\E-

f f(x)g(x)dx < J f(s)g{s)ds = / r(x)g*(x)dx.

- f f(x)g(x)dx < f (-f)(s)g#(s)ds = - [ U(s)g(s)ds