2997-Sobolev Spaces - With Applications To Elliptic Partial Differential Equations (2nd)

Grundlehren der
mathematischen Wissenschaften
A Series of Comprehensive Studies in Mathematics
Series editors
M. Berger P. de la Harpe F. Hirzebruch
N.J. Hitchin L. Hrmander A. Kupiainen
G. Lebeau F.-H. Lin B.C. Ng
M. Ratner D. Serre Ya.G. Sinai
N.J.A. Sloane A.M. Vershik M. Waldschmidt
Editor-in-Chief
A. Chenciner J. Coates S.R.S. Varadhan
342
For further volumes:

http://www.springer.com/series/138
Vladimir Mazya
Sobolev Spaces
with Applications to Elliptic Partial
Differential Equations
2nd, revised and augmented Edition
Professor Vladimir Mazya

Department of Mathematics Sciences
University of Liverpool
Liverpool L69 7ZL,
UK
and
Department of Mathematics
Linkping University
Linkping 581 83,
Sweden
vlmaz@mai.liu.se
The 1st edition, published in 1985 in English under Vladimir G. Mazja in the Springer Series of Soviet
Mathematics was translated from Russian by Tatyana O. Shaposhnikova
ISSN 0072-7830
ISBN 978-3-642-15563-5
e-ISBN 978-3-642-15564-2
DOI 10.1007/978-3-642-15564-2
Springer Heidelberg Dordrecht London New York
Library of Congress Control Number: 2011921122
Mathematics Subject Classification: 46E35, 42B37, 26D10
c Springer-Verlag Berlin Heidelberg 1985, 2011

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations are
liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws
and regulations and therefore free for general use.
Cover design: VTEX, Vilnius
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
To Tatyana
Preface
Sobolev spaces, i.e., the classes of functions with derivatives in Lp , occupy

an outstanding place in analysis. During the last half-century a substantial
contribution to the study of these spaces has been made; so now solutions to
many important problems connected with them are known.
In the present monograph we consider various aspects of theory of Sobolev
spaces in particular, the so-called embedding theorems. Such theorems, originally established by S.L. Sobolev in the 1930s, proved to be a useful tool in
functional analysis and in the theory of linear and nonlinear partial dierential
equations.
A part of this book rst appeared in German as three booklets of TeubnerTexte f
ur Mathematik [552, 555]. In the Springer volume of Sobolev Spaces
[556] published in 1985, the material was expanded and revised.
As the years passed the area became immensely vast and underwent important changes, so the main contents of the 1985 volume had the potential
for further development, as shown by numerous references. Therefore, and
since the volume became a bibliographical rarity, Springer-Verlag oered me
the opportunity to prepare the second, updated edition of [556].
As in [556], the selection of topics was mainly inuenced by my involvement
in their study, so a considerable part of the text is a report on my work in the
eld. In comparison with [556], the present text is enhanced by more recent
results. New comments and the signicantly augmented list of references are
intended to create a broader and modern view of the area. The book diers
considerably from the monographs of other authors dealing with spaces of
dierentiable functions that were published in the last 50 years.
Each of the 18 chapters of the book is divided into sections and most of
the sections consist of subsections. The sections and subsections are numbered
by two and three numbers, respectively (3.1 is Sect. 1 in Chap. 3, 1.4.3 is
Subsect. 3 in Sect. 4 in Chap. 1). Inside subsections we use an independent
numbering of theorems, lemmas, propositions, corollaries, remarks, and so
on. If a subsection contains only one theorem or lemma then this theorem
or lemma has no number. In references to the material from another section
vii
viii
Preface
or subsection we rst indicate the number of this section or subsection. For

example, Theorem 1.2.1/1 means Theorem 1 in Subsect. 1.2.1, (2.6.6) denotes
formula (6) in Sect. 2.6.
The reader can obtain a general idea of the contents of the book from
the Introduction. Most of the references to the literature are collected in the
Comments. The list of notation is given at the end of the book.
The volume is addressed to students and researchers working in functional
analysis and in the theory of partial dierential operators. Prerequisites for
reading this book are undergraduate courses in these subjects.
Acknowledgments
My cordial thanks are to S. Bobkov, Yu.D. Burago, A. Cianchi, E. Milman,
S. Poborchi, P. Shvartsman, and I. Verbitsky who read parts of the book and
supplied me with their comments.
I wish to express my deep gratitude to M. Nieves for great help in the
technical preparation of the text.
The dedication of this book to its translator and my wife T.O. Shaposhnikova is a weak expression of my gratitude for her innite patience, useful
advice, and constant assistance.
LiverpoolLink
oping
January 2010
Vladimir Mazya
Contents
Basic Properties of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . .

1.1 The Spaces Llp (), Vpl () and Wpl () . . . . . . . . . . . . . . . . . . . . . .
1.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.2 Local Properties of Elements in the Space Llp () . . . . . .
1.1.3 Absolute Continuity of Functions in L1p () . . . . . . . . . . .
1.1.4 Spaces Wpl () and Vpl () . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.5 Approximation of Functions in Sobolev Spaces by
Smooth Functions in . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.6 Approximation of Functions in Sobolev Spaces by
...............................
Functions in C ()
1.1.7 Transformation of Coordinates in Norms of Sobolev
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.8 Domains Starshaped with Respect to a Ball . . . . . . . . . .
1.1.9 Domains of the Class C 0,1 and Domains Having the
Cone Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.10 Sobolev Integral Representation . . . . . . . . . . . . . . . . . . . . .
1.1.11 Generalized Poincare Inequality . . . . . . . . . . . . . . . . . . . . .
1.1.12 Completeness of Wpl () and Vpl () . . . . . . . . . . . . . . . . . .
1.1.13 The Space
Llp () and Its Completeness . . . . . . . . . . . . . .
1.1.14 Estimate of Intermediate Derivative and Spaces
l () and
W
Llp () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
p
1.1.15 Duals of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.16 Equivalent Norms in Wpl () . . . . . . . . . . . . . . . . . . . . . . . .
1.1.17 Extension of Functions in Vpl () onto Rn . . . . . . . . . . . . .
1.1.18 Removable Sets for Sobolev Functions . . . . . . . . . . . . . . . .
1.1.19 Comments to Sect. 1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Facts from Set Theory and Function Theory . . . . . . . . . . . . . . . .
1.2.1 Two Theorems on Coverings . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2 Theorem on Level Sets of a Smooth Function . . . . . . . . .
1
1
1
2
4
7
9
10
12
14
15
16
20
22
22
23
24
26
26
28
29
32
32
35
ix
Contents
1.3
1.4
1.5
1.6
1.7
1.2.3 Representation of the Lebesgue Integral as a Riemann

Integral along a Halfaxis . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.2.4 Formula for the Integral of Modulus of the Gradient . . . 38
1.2.5 Comments to Sect. 1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Some Inequalities for Functions of One Variable . . . . . . . . . . . . . 40
1.3.1 Basic Facts on Hardy-type Inequalities . . . . . . . . . . . . . . . 40
1.3.2 Two-weight Extensions of Hardys Type Inequality in
the Case p q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.3.3 Two-Weight Extensions of Hardys Inequality in the
Case p > q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.3.4 Hardy-Type Inequalities with Indenite Weights . . . . . . 51
1.3.5 Three Inequalities for Functions on (0, ) . . . . . . . . . . . . 57
1.3.6 Estimates for Dierentiable Nonnegative Functions of
One Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
1.3.7 Comments to Sect. 1.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Embedding Theorems of Sobolev Type . . . . . . . . . . . . . . . . . . . . . 63
1.4.1 D.R. Adams Theorem on Riesz Potentials . . . . . . . . . . . . 64
1.4.2 Estimate for the Norm in Lq (Rn , ) by the Integral of
the Modulus of the Gradient . . . . . . . . . . . . . . . . . . . . . . . . 67
1.4.3 Estimate for the Norm in Lq (Rn , ) by the Integral of
the Modulus of the lth Order Gradient . . . . . . . . . . . . . . . 70
1.4.4 Corollaries of Previous Results . . . . . . . . . . . . . . . . . . . . . . 72
1.4.5 Generalized Sobolev Theorem . . . . . . . . . . . . . . . . . . . . . . . 73
1.4.6 Compactness Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
1.4.7 Multiplicative Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . 79
1.4.8 Comments to Sect. 1.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
More on Extension of Functions in Sobolev Spaces . . . . . . . . . . . 87
1.5.1 Survey of Results and Examples of Domains . . . . . . . . . . 87
1.5.2 Domains in EVp1 which Are Not Quasidisks . . . . . . . . . . . 91
1.5.3 Extension with Zero Boundary Data . . . . . . . . . . . . . . . . . 94
1.5.4 Comments to Sect. 1.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Inequalities for Functions with Zero Incomplete Cauchy Data . 99
1.6.1 Integral Representation for Functions of One
Independent Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
1.6.2 Integral Representation for Functions of Several
Variables with Zero Incomplete Cauchy Data . . . . . . . . . 100
1.6.3 Embedding Theorems for Functions with Zero
Incomplete Cauchy Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
1.6.4 Necessity of the Condition l 2k . . . . . . . . . . . . . . . . . . . . 105
Density of Bounded Functions in Sobolev Spaces . . . . . . . . . . . . 107
1.7.1 Lemma on Approximation of Functions in L1p () . . . . . . 107
1.7.2 Functions with Bounded Gradients Are Not Always
Dense in L1p () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
1.7.3 A Planar Bounded Domain for Which L21 () L ()
Is Not Dense in L21 () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Contents
xi
1.7.4 Density of Bounded Functions in L2p () for

Paraboloids in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
1.7.5 Comments to Sect. 1.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
1.8 Maximal Algebra in Wpl () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
1.8.1 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
1.8.2 The Space W22 () L () Is Not Always a Banach
Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
1.8.3 Comments to Sect. 1.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2
Inequalities for Functions Vanishing at the Boundary . . . . . . 123

2.1 Conditions for Validity of Integral Inequalities
(the Case p = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
2.1.1 Criterion Formulated in Terms of Arbitrary Admissible
Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
2.1.2 Criterion Formulated in Terms of Balls for = Rn . . . . 126
2.1.3 Inequality Involving the Norms in Lq (, ) and
Lr (, ) (Case p = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
2.1.4 Case q (0, 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
2.1.5 Inequality (2.1.10) Containing Particular Measures . . . . 132
2.1.6 Power Weight Norm of the Gradient on the
Right-Hand Side . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
2.1.7 Inequalities of HardySobolev Type as Corollaries of
Theorem 2.1.6/1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
2.1.8 Comments to Sect. 2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
2.2 (p, )-Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
2.2.1 Denition and Properties of the (p, )-Capacity . . . . . . . 141
2.2.2 Expression for the (p, )-Capacity Containing an
Integral over Level Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . 144
2.2.3 Lower Estimates for the (p, )-Capacity . . . . . . . . . . . . . . 146
2.2.4 p-Capacity of a Ball . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
2.2.5 (p, )-Capacity for p = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
2.2.6 The Measure mn1 and 2-Capacity . . . . . . . . . . . . . . . . . . 149
2.2.7 Comments to Sect. 2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
(the Case p > 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
2.3.1 The (p, )-Capacitary Inequality . . . . . . . . . . . . . . . . . . . . 152
2.3.2 Capacity Minimizing Function and Its Applications . . . . 156
2.3.3 Estimate for a Norm in a BirnbaumOrlicz Space . . . . . 157
2.3.4 Sobolev Type Inequality as Corollary of Theorem 2.3.3 . 160
2.3.5 Best Constant in the Sobolev Inequality (p > 1) . . . . . . . 160
2.3.6 Multiplicative Inequality (the Case p 1) . . . . . . . . . . . . 162
2.3.7 Estimate for the Norm in Lq (, ) with q < p (First
Necessary and Sucient Condition) . . . . . . . . . . . . . . . . . . 165
2.3.8 Estimate for the Norm in Lq (, ) with q < p (Second
Necessary and Sucient Condition) . . . . . . . . . . . . . . . . . . 167
xii
Contents
2.4
2.5
2.6
2.7
2.8
2.3.9 Inequality with the Norms in Lq (, ) and Lr (, )

(the Case p 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
2.3.10 Estimate with a Charge on the Left-Hand Side . . . . . . 173
2.3.11 Multiplicative Inequality with the Norms in Lq (, )
and Lr (, ) (Case p 1) . . . . . . . . . . . . . . . . . . . . . . . . . . 174
2.3.12 On Nash and Moser Multiplicative Inequalities . . . . . . . . 176
2.3.13 Comments to Sect. 2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Continuity and Compactness of Embedding Operators of
1 () into BirnbaumOrlicz Spaces . . . . . . . . . . . . . 179

L1p () and W
p
2.4.1 Conditions for Boundedness of Embedding Operators . . 180
2.4.2 Criteria for Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
2.4.3 Comments to Sect. 2.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Structure of the Negative Spectrum of the Multidimensional
Schrodinger Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
2.5.1 Preliminaries and Notation . . . . . . . . . . . . . . . . . . . . . . . . . 188
2.5.2 Positivity of the Form S1 [u, u] . . . . . . . . . . . . . . . . . . . . . . 189
2.5.3 Semiboundedness of the Schr
odinger Operator . . . . . . . . 190
2.5.4 Discreteness of the Negative Spectrum . . . . . . . . . . . . . . . 193
2.5.5 Discreteness of the Negative Spectrum of the Operator
Sh for all h . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
2.5.6 Finiteness of the Negative Spectrum . . . . . . . . . . . . . . . . . 197
2.5.7 Inniteness and Finiteness of the Negative Spectrum
of the Operator Sh for all h . . . . . . . . . . . . . . . . . . . . . . . . . 199
2.5.8 Proofs of Lemmas 2.5.1/1 and 2.5.1/2 . . . . . . . . . . . . . . . . 200
2.5.9 Comments to Sect. 2.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Properties of Sobolev Spaces Generated by Quadratic Forms
with Variable Coecients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
2.6.1 Degenerate Quadratic Form . . . . . . . . . . . . . . . . . . . . . . . . 205
2.6.2 Completion in the Metric of a Generalized Dirichlet
Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
2.6.3 Comments to Sect. 2.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
Dilation Invariant Sharp Hardys Inequalities . . . . . . . . . . . . . . . 213
2.7.1 Hardys Inequality with Sharp Sobolev Remainder
Term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
2.7.2 Two-Weight Hardys Inequalities . . . . . . . . . . . . . . . . . . . . 214
2.7.3 Comments to Sect. 2.7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Sharp HardyLeray Inequality for Axisymmetric
Divergence-Free Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
2.8.1 Statement of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
2.8.2 Proof of Theorem 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
2.8.3 Proof of Theorem 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
2.8.4 Comments to Sect. 2.8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
Contents
xiii
Conductor and Capacitary Inequalities with Applications

to Sobolev-Type Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
3.2 Comparison of Inequalities (3.1.4) and (3.1.5) . . . . . . . . . . . . . . . 233
3.3 Conductor Inequality (3.1.1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
3.4 Applications of the Conductor Inequality (3.1.1) . . . . . . . . . . . . . 236
3.5 p-Capacity Depending on and Its Applications to a
Conductor Inequality and Inequality (3.4.1) . . . . . . . . . . . . . . . . . 241
3.6 Compactness and Essential Norm . . . . . . . . . . . . . . . . . . . . . . . . . . 243
3.7 Inequality (3.1.10) with Integer l 2 . . . . . . . . . . . . . . . . . . . . . . . 245
3.8 Two-Weight Inequalities Involving Fractional Sobolev Norms . . 249
3.9 Comments to Chap. 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Generalizations for Functions on Manifolds and

Topological Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
4.2 Integral Inequalities for Functions on Riemannian Manifolds . . 257
4.3 The First DirichletLaplace Eigenvalue and Isoperimetric
Constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
4.4 Conductor Inequalities for a Dirichlet-Type Integral with a
Locality Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
4.5 Conductor Inequality for a Dirichlet-Type Integral without
Locality Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
4.6 Sharp Capacitary Inequalities and Their Applications . . . . . . . . 273
4.7 Capacitary Improvement of the FaberKrahn Inequality . . . . . . 278
4.8 Two-Weight Sobolev Inequality with Sharp Constant . . . . . . . . 282
4.9 Comments to Chap. 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
Integrability of Functions in the Space L11 () . . . . . . . . . . . . . . 287

5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
5.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
1
5.1.2 Lemmas on Approximation of Functions in Wp,r
()
1
and Lp () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
5.2 Classes of Sets J , H and the Embedding L11 () Lq () . . 290
5.2.1 Classes J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
5.2.2 Technical Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
5.2.3 Embedding L11 () Lq () . . . . . . . . . . . . . . . . . . . . . . . . . 295
5.2.4 Area Minimizing Function M and Embedding of
L11 () into Lq () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
5.2.5 Example of a Domain in J1 . . . . . . . . . . . . . . . . . . . . . . . . 299
5.3 Subareal Mappings and the Classes J and H . . . . . . . . . . . . 300
5.3.1 Subareal Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
5.3.2 Estimate for the Function in Terms of Subareal
Mappings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
5.3.3 Estimates for the Function for Special Domains . . . . . 303
xiv
Contents
5.4 Two-Sided Estimates for the Function for the Domain in

Nikod
yms Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
5.5 Compactness of the Embedding L11 () Lq () (q 1) . . . . . . 311
5.5.1 Class J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
5.5.2 Compactness Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
1
(, ) Lq () . . . . . . . . . . . . . . . . . . . . . . . . . 314
5.6 Embedding W1,r
5.6.1 Class K, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
5.6.2 Examples of Sets in K, . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
5.6.3 Continuity of the Embedding Operator
1
(, ) Lq () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
W1,r
5.7 Comments to Chap. 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
6
Integrability of Functions in the Space L1p () . . . . . . . . . . . . . . 323

6.1 Conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
6.1.1 Equivalence of Certain Denitions of Conductivity . . . . . 324
6.1.2 Some Properties of Conductivity . . . . . . . . . . . . . . . . . . . . 326
6.1.3 Dirichlet Principle with Prescribed Level Surfaces and
Its Corollaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
6.2 Multiplicative Inequality for Functions Which Vanish on a
Subset of . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
6.3 Classes of Sets Ip, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
6.3.1 Denition and Simple Properties of Ip, . . . . . . . . . . . . . 331
6.3.2 Identity of the Classes I1, and J . . . . . . . . . . . . . . . . . 333
6.3.3 Necessary and Sucient Condition for the Validity of
1
a Multiplicative Inequality for Functions in Wp,s
() . . . 334
1
6.3.4 Criterion for the Embedding Wp,s () Lq (),
p q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
6.3.5 Function M,p and the Relationship of the Classes
Ip, and J . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
6.3.6 Estimates for the Conductivity Minimizing Function
M,p for Certain Domains . . . . . . . . . . . . . . . . . . . . . . . . . . 338
1
() Lq () for q < p . . . . . . . . . . . . . . . . . . . 341
6.4 Embedding Wp,s
6.4.1 Estimate for the Norm in Lq () with q < p for
Functions which Vanish on a Subset of . . . . . . . . . . . . . 341
1
6.4.2 Class Hp, and the Embedding Wp,s
() Lq () for
0 < q < p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342

6.4.3 Embedding L1p () Lq () for a Domain with Finite
Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
6.4.4 Sucient Condition for Belonging to Hp, . . . . . . . . . . . . 345
6.4.5 Necessary Conditions for Belonging to the Classes
Ip, and Hp, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
6.4.6 Examples of Domains in Hp, . . . . . . . . . . . . . . . . . . . . . . 347
6.4.7 Other Descriptions of the Classes Ip, and Hp, . . . . . . 348
6.4.8 Integral Inequalities for Domains with Power Cusps . . . . 350
Contents
xv
6.5 More on the Nikod

ym Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
6.6 Some Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
1
6.7 Inclusion Wp,r
() Lq () (r > q) for Domains with Innite
Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
6.7.1 Classes J and Ip, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364

1
() Lq () (r > q) . . . . . . . . . . . . . . . . 367
6.7.2 Embedding Wp,r
6.7.3 Example of a Domain in the Class Ip, . . . . . . . . . . . . . . 368

(0)
6.8
6.9
6.10
6.11
6.7.4 Space L1p () and Its Embedding into Lq () . . . . . . . . . . 370

6.7.5 Poincare-Type Inequality for Domains with Innite
Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Compactness of the Embedding L1p () Lq () . . . . . . . . . . . . . 374
6.8.1 Class Ip, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
6.8.2 Compactness Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
6.8.3 Sucient Conditions for Compactness of the
Embedding L1p () Lq () . . . . . . . . . . . . . . . . . . . . . . . . 376
6.8.4 Compactness Theorem for an Arbitrary Domain with
Finite Volume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
6.8.5 Examples of Domains in the class Ip, . . . . . . . . . . . . . . . 378
Embedding Llp () Lq () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Applications to the Neumann Problem for Strongly Elliptic
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
6.10.1 Second-Order Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381
6.10.2 Neumann Problem for Operators of Arbitrary Order . . . 382
6.10.3 Neumann Problem for a Special Domain . . . . . . . . . . . . . 385
6.10.4 Counterexample to Inequality (6.10.7) . . . . . . . . . . . . . . . 389
Inequalities Containing Integrals over the Boundary . . . . . . . . . 390
1
(, ) Lq () . . . . . . . . . . . . . . . . . . . 390
(n1)
(n1)
and J
. . . . . . . . . . . . . . . . . . . . . . . . 393
Classes Ip,
(n1)
(n1)
Examples of Domains in Ip,
and J
. . . . . . . . . 394
Estimates for the Norm in Lq () . . . . . . . . . . . . . . . . . . . 395
(n1)
Class Ip,
and Compactness Theorems . . . . . . . . . . . . 397
Criteria of Solvability of Boundary Value Problems for
Second-Order Elliptic Equations . . . . . . . . . . . . . . . . . . . . . 399
6.12 Comments to Chap. 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
6.11.2
6.11.3
6.11.4
6.11.5
6.11.6
Continuity and Boundedness of Functions in Sobolev

Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
7.1 The Embedding Wp1 () C() L () . . . . . . . . . . . . . . . . . . 406
7.1.1 Criteria for Continuity of Embedding Operators of
Wp1 () and L1p () into C() L () . . . . . . . . . . . . . . . 406
7.1.2 Sucient Condition in Terms of the Isoperimetric
Function for the Embedding Wp1 () C() L () . 409
xvi
Contents
7.2
7.3
7.4
7.5
7.6
7.7
8
7.1.3 Isoperimetric Function and a BrezisGallouet

Wainger-Type Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . 410
Multiplicative Estimate for Modulus of a Function in Wp1 () . 412
7.2.1 Conditions for Validity of a Multiplicative Inequality . . . 412
7.2.2 Multiplicative Inequality in the Limit Case
r = (p n)/n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414
Continuity Modulus of Functions in L1p () . . . . . . . . . . . . . . . . . . 416
Boundedness of Functions with Derivatives in Birnbaum
Orlicz Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
Compactness of the Embedding Wp1 () C() L () . . . . 422
7.5.1 Compactness Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
7.5.2 Sucient Condition for the Compactness in Terms of
the Isoperimetric Function . . . . . . . . . . . . . . . . . . . . . . . . . . 423
7.5.3 Domain for Which the Embedding Operator of Wp1 ()
into C() L () is Bounded but not Compact . . . . . 424
Generalizations to Sobolev Spaces of an Arbitrary Integer
Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
7.6.1 The (p, l)-Conductivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426
7.6.2 Embedding Llp () C() L () . . . . . . . . . . . . . . . . . 427
7.6.3 Embedding Vpl () C() L () . . . . . . . . . . . . . . . . . 428
7.6.4 Compactness of the Embedding
Llp () C() L () . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
7.6.5 Sucient Conditions for the Continuity and the
Compactness of the Embedding
Llp () C() L () . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
k (),
7.6.6 Embedding Operators for the Space Wpl () W
p
l > 2k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432
Comments to Chap. 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 434
Localization Moduli of Sobolev Embeddings for General

Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435
8.1 Localization Moduli and Their Properties . . . . . . . . . . . . . . . . . . 437
8.2 Counterexample for the Case p = q . . . . . . . . . . . . . . . . . . . . . . . . 442
8.3 Critical Sobolev Exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
8.4 Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
8.5 Measures of Noncompactness for Power Cusp-Shaped
Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
8.6 Finiteness of the Negative Spectrum of a Schrodinger
Operator on -Cusp Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
8.7 Relations of Measures of Noncompactness with Local
Isoconductivity and Isoperimetric Constants . . . . . . . . . . . . . . . . 456
8.8 Comments to Chap. 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Contents
xvii
Space of Functions of Bounded Variation . . . . . . . . . . . . . . . . . . 459

9.1 Properties of the Set Perimeter and Functions in BV () . . . . . 459
9.1.1 Denitions of the Space BV () and of the Relative
Perimeter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
9.1.2 Approximation of Functions in BV () . . . . . . . . . . . . . . . 460
9.1.3 Approximation of Sets with Finite Perimeter . . . . . . . . . . 463
9.1.4 Compactness of the Family of Sets with Uniformly
Bounded Relative Perimeters . . . . . . . . . . . . . . . . . . . . . . . 464
9.1.5 Isoperimetric Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
9.1.6 Integral Formula for the Norm in BV () . . . . . . . . . . . . . 465
9.1.7 Embedding BV () Lq () . . . . . . . . . . . . . . . . . . . . . . . . 466
9.2 GaussGreen Formula for Lipschitz Functions . . . . . . . . . . . . . . . 467
9.2.1 Normal in the Sense of Federer and Reduced
Boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
9.2.2 GaussGreen Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
9.2.3 Several Auxiliary Assertions . . . . . . . . . . . . . . . . . . . . . . . . 468
9.2.4 Study of the Set N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
9.2.5 Relations Between varE and s on E . . . . . . . . . . . . . . 473
9.3 Extension of Functions in BV () onto Rn . . . . . . . . . . . . . . . . . . 477
9.3.1 Proof of Necessity of (9.3.2) . . . . . . . . . . . . . . . . . . . . . . . . 478
9.3.2 Three Lemmas on PC (E ) . . . . . . . . . . . . . . . . . . . . . . . . . 478
9.3.3 Proof of Suciency of (9.3.2) . . . . . . . . . . . . . . . . . . . . . . . 480
9.3.4 Equivalent Statement of Theorem 9.3 . . . . . . . . . . . . . . . . 482
9.3.5 One More Extension Theorem . . . . . . . . . . . . . . . . . . . . . . 483
9.4 Exact Constants for Certain Convex Domains . . . . . . . . . . . . . . . 484
9.4.1 Lemmas on Approximations by Polyhedra . . . . . . . . . . . . 484
9.4.2 Property of PC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
9.4.3 Expression for the Set Function (E ) for a Convex
Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 486
9.4.4 The Function || for a Convex Domain . . . . . . . . . . . . . . 487
9.5 Rough Trace of Functions in BV () and Certain Integral
Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
9.5.1 Denition of the Rough Trace and Its Properties . . . . . . 489
9.5.2 Integrability of the Rough Trace . . . . . . . . . . . . . . . . . . . . . 492
9.5.3 Exact Constants in Certain Integral Estimates for the
Rough Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
9.5.4 More on Integrability of the Rough Trace . . . . . . . . . . . . . 495
9.5.5 Extension of a Function in BV () to C by
a Constant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496
9.5.6 Multiplicative Estimates for the Rough Trace . . . . . . . . . 497
9.5.7 Estimate for the Norm in Ln/(n1) () of a Function
in BV () with Integrable Rough Trace . . . . . . . . . . . . . . 499
9.6 Traces of Functions in BV () on the Boundary and
GaussGreen Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
9.6.1 Denition of the Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
xviii
Contents
9.6.2 Coincidence of the Trace and the Rough Trace . . . . . . . . 501

9.6.3 Trace of the Characteristic Function . . . . . . . . . . . . . . . . . 504
9.6.4 Integrability of the Trace of a Function in BV () . . . . . 504
9.6.5 GaussGreen Formula for Functions in BV () . . . . . . . . 505
9.7 Comments to Chap. 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
10 Certain Function Spaces, Capacities, and Potentials . . . . . . . 511
10.1 Spaces of Functions Dierentiable of Arbitrary Positive Order . 512
10.1.1 Spaces wpl , Wpl , blp , Bpl for l > 0 . . . . . . . . . . . . . . . . . . . . . 512
10.1.2 Riesz and Bessel Potential Spaces . . . . . . . . . . . . . . . . . . . 516
10.1.3 Other Properties of the Introduced Function Spaces . . . 519
10.2 Bourgain, Brezis, and Mironescu Theorem Concerning
Limiting Embeddings of Fractional Sobolev Spaces . . . . . . . . . . 521
10.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
10.2.2 Hardy-Type Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522
10.2.3 Sobolev Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
ps (Rn ) as s 0 . . . . . . . . . 528
10.2.4 Asymptotics of the Norm in W
10.3 On the Brezis and Mironescu Conjecture Concerning a
GagliardoNirenberg Inequality for Fractional Sobolev Norms . 530
10.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 530
10.3.2 Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
10.4 Some Facts from Nonlinear Potential Theory . . . . . . . . . . . . . . . . 536
10.4.1 Capacity cap(e, Spl ) and Its Properties . . . . . . . . . . . . . . . 536
10.4.2 Nonlinear Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
10.4.3 Metric Properties of Capacity . . . . . . . . . . . . . . . . . . . . . . . 541
10.4.4 Rened Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 544
10.5 Comments to Chap. 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
11 Capacitary and Trace Inequalities for Functions in Rn
with Derivatives of an Arbitrary Order . . . . . . . . . . . . . . . . . . . . 549
11.1 Description of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
11.2 Capacitary Inequality of an Arbitrary Order . . . . . . . . . . . . . . . . 552
11.2.1 A Proof Based on the Smooth Truncation of
a Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 552
11.2.2 A Proof Based on the Maximum Principle for
Nonlinear Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554
11.3 Conditions for the Validity of Embedding Theorems in Terms
of Isocapacitary Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 556
11.4 Counterexample to the Capacitary Inequality for the Norm
in L22 () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 558
11.5 Ball and Pointwise Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564
11.6 Conditions for Embedding into Lq () for p > q > 0 . . . . . . . . . . 570
11.6.1 Criterion in Terms of the Capacity Minimizing
Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
11.6.2 Two Simple Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
Contents
xix
11.7 Cartan-Type Theorem and Estimates for Capacities . . . . . . . . . 575

11.8 Embedding Theorems for the Space Spl (Conditions in Terms
of Balls, p > 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
11.9 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
11.9.1 Compactness Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
11.9.2 Equivalence of Continuity and Compactness of the
Embedding Hpl Lq () for p > q . . . . . . . . . . . . . . . . . . 583
11.9.3 Applications to the Theory of Elliptic Operators . . . . 586
11.9.4 Criteria for the RellichKato Inequality . . . . . . . . . . . . 586
11.10 Embedding Theorems for p = 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 588
11.10.1 Integrability with Respect to a Measure . . . . . . . . . . . . 588
11.10.2 Criterion for an Upper Estimate of a Dierence
Seminorm (the Case p = 1) . . . . . . . . . . . . . . . . . . . . . . . 590
11.10.3 Embedding into a Riesz Potential Space . . . . . . . . . . . . 596
11.11 Criteria for an Upper Estimate of a Dierence Seminorm
(the Case p > 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
11.11.1 Case q > p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 597
11.11.2 Capacitary Sucient Condition in the Case q = p . . . 603
11.12 Comments to Chap. 11 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
12 Pointwise Interpolation Inequalities for Derivatives and
Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611
12.1 Pointwise Interpolation Inequalities for Riesz and Bessel
Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
12.1.1 Estimate for the Maximal Operator of a Convolution 612
12.1.2 Pointwise Interpolation Inequality for Riesz
Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613
12.1.3 Estimates for |Jw | . . . . . . . . . . . . . . . . . . . . . . . . . . . . 614
12.1.4 Estimates for |Jw ( )| . . . . . . . . . . . . . . . . . . . . . . . 619
12.1.5 Pointwise Interpolation Inequality for Bessel
Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 620
12.1.6 Pointwise Estimates Involving Mk u and l u . . . . . . 622
12.1.7 Application: Weighted Norm Interpolation
Inequalities for Potentials . . . . . . . . . . . . . . . . . . . . . . . . 623
12.2 Sharp Pointwise Inequalities for u . . . . . . . . . . . . . . . . . . . . . . 624
12.2.1 The Case of Nonnegative Functions . . . . . . . . . . . . . . . . 624
12.2.2 Functions with Unrestricted Sign. Main Result . . . . . . 624
12.2.3 Proof of Inequality (12.2.6) . . . . . . . . . . . . . . . . . . . . . . . 626
12.2.4 Proof of Sharpness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
12.2.5 Particular Case (r) = r , > 0 . . . . . . . . . . . . . . . . . . 634
12.2.6 One-Dimensional Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 636
12.3 Pointwise Interpolation Inequalities Involving Fractional
Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
12.3.1 Inequalities with Fractional Derivatives on the
Right-Hand Sides . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 638
xx
Contents
12.4
12.5
12.3.2 Inequality with a Fractional Derivative Operator on

the Left-Hand Side . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
12.3.3 Application: Weighted GagliardoNirenberg-Type
Inequalities for Derivatives . . . . . . . . . . . . . . . . . . . . . . . 643
Application of (12.3.11) to Composition Operator in
Fractional Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
12.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
12.4.2 Proof of Inequality (12.4.1) . . . . . . . . . . . . . . . . . . . . . . . 645
12.4.3 Continuity of the Map (12.4.2) . . . . . . . . . . . . . . . . . . . . 648
Comments to Chap. 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 653
13 A Variant of Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657

13.1 Capacity Cap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657
13.1.1 Simple Properties of Cap(e,
Llp ()) . . . . . . . . . . . . . . . . 657
13.1.2 Capacity of a Continuum . . . . . . . . . . . . . . . . . . . . . . . . . 660
13.1.3 Capacity of a Bounded Cylinder . . . . . . . . . . . . . . . . . . 662
13.1.4 Sets of Zero Capacity Cap(, Wpl ) . . . . . . . . . . . . . . . . . . 663
13.2 On (p, l)-Polar Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 663
13.3 Equivalence of Two Capacities . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
13.4 Removable Singularities of l-Harmonic Functions in Lm
2 . . . . . 666
13.5 Comments to Chap. 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 668
14 Integral Inequality for Functions on a Cube . . . . . . . . . . . . . . . 669
14.1 Connection Between the Best Constant and Capacity
(Case k = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
14.1.1 Denition of a (p, l)-Negligible Set . . . . . . . . . . . . . . . . . 670
14.1.2 Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 670
14.1.3 Variant of Theorem 14.1.2 and Its Corollaries . . . . . . . 673
14.2 Connection Between Best Constant and the (p, l)-Inner
Diameter (Case k = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
14.2.1 Set Function lp,q (G) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675
14.2.2 Denition of the (p, l)-Inner Diameter . . . . . . . . . . . . . . 676
14.2.3 Estimates for the Best Constant in (14.1.3) by the
(p, l)-Inner Diameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 676
14.3 Estimates for the Best Constant C in the General Case . . . . . 679
14.3.1 Necessary and Sucient Condition for Validity of
the Basic Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 679
14.3.2 Polynomial Capacities of Function Classes . . . . . . . . . . 680
14.3.3 Estimates for the Best Constant C in the Basic
Inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 681
Llp (Q2d )) . . . . . . . . . 684
14.3.4 Class C0 (e) and Capacity Capk (e,
14.3.5 Lower Bound for Capk . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
14.3.6 Estimates for the Best Constant in the Case of
Small (p, l)-Inner Diameter . . . . . . . . . . . . . . . . . . . . . . . 687
Contents
xxi
14.3.7 A Logarithmic Sobolev Inequality with Application

to the Uniqueness Theorem for Analytic Functions
in the Class L1p (U ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 689
14.4
Comments to Chap. 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691
15 Embedding of the Space

Llp () into Other Function
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
15.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693
15.2 Embedding
Llp () D () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694
15.2.1 Auxiliary Assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694
15.2.2 Case = Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696
15.2.3 Case n = p l, p > 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
15.2.4 Case n < p l and Noninteger n/p . . . . . . . . . . . . . . . . . . 697
15.2.5 Case n < pl, 1 < p < , and Integer n/p . . . . . . . . . . . 698
15.3 Embedding
Llp () Lq (, loc) . . . . . . . . . . . . . . . . . . . . . . . . . . 701
15.4 Embedding
Llp () Lq () (the Case p q) . . . . . . . . . . . . . . . 703
15.4.1 A Condition in Terms of the (p, l)-Inner Diameter . . . 703
15.4.2 A Condition in Terms of Capacity . . . . . . . . . . . . . . . . . 704
15.5 Embedding
Llp () Lq () (the Case p > q 1) . . . . . . . . . . . 707
15.5.1 Denitions and Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . 707
15.5.2 Basic Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 710
15.5.3 Embedding
Llp () Lq () for an Innite Funnel . 712
15.6 Compactness of the Embedding
Llp () Lq () . . . . . . . . . . . . 714
15.6.1 Case p q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714
15.6.2 Case p > q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
15.7 Application to the Dirichlet Problem for a Strongly Elliptic
Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716
15.7.1 Dirichlet Problem with Nonhomogeneous Boundary
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717
15.7.2 Dirichlet Problem with Homogeneous Boundary
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 718
15.7.3 Discreteness of the Spectrum of the Dirichlet
Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
15.7.4 Dirichlet Problem for a Nonselfadjoint Operator . . . . . 719
15.8 Applications to the Theory of Quasilinear Elliptic Equations . 721
15.8.1 Solvability of the Dirichlet Problem for Quasilinear
Equations in Unbounded Domains . . . . . . . . . . . . . . . . . 721
15.8.2 A Weighted Multiplicative Inequality . . . . . . . . . . . . . . 725
15.8.3 Uniqueness of a Solution to the Dirichlet Problem
with an Exceptional Set for Equations of Arbitrary
Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 727
15.8.4 Uniqueness of a Solution to the Neumann Problem
for Quasilinear Second-Order Equation . . . . . . . . . . . . . 730
15.9 Comments to Chap. 15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733
xxii
Contents
16 Embedding
Llp (, ) Wrm () . . . . . . . . . . . . . . . . . . . . . . . . . . . . 737
16.1 Auxiliary Assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 737
16.2 Continuity of the Embedding Operator
Llp (, ) Wrm () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739

16.3 Compactness of the Embedding Operator
Llp (, ) Wrm () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742

16.3.1 Essential Norm of the Embedding Operator . . . . . . . . 742
16.3.2 Criteria for Compactness . . . . . . . . . . . . . . . . . . . . . . . . . 744
16.4 Closability of Embedding Operators . . . . . . . . . . . . . . . . . . . . . . 746
16.5 Application: Positive Deniteness and Discreteness of the
Spectrum of a Strongly Elliptic Operator . . . . . . . . . . . . . . . . . . 749
16.6 Comments to Chap. 16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 751
17 Approximation in Weighted Sobolev Spaces . . . . . . . . . . . . . . . 755
17.1 Main Results and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 755
17.2 Capacities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
17.3 Applications of Lemma 17.2/3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 761
17.4 Proof of Theorem 17.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765
17.5 Comments to Chap. 17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 768
18 Spectrum of the Schr
odinger Operator and the Dirichlet
Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 769
18.1 Main Results on the Schrodinger Operator . . . . . . . . . . . . . . . . 770
18.2 Discreteness of Spectrum: Necessity . . . . . . . . . . . . . . . . . . . . . . 773
18.3 Discreteness of Spectrum: Suciency . . . . . . . . . . . . . . . . . . . . . 781
18.4 A Suciency Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783
18.5 Positivity of HV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 787
18.6 Structure of the Essential Spectrum of HV . . . . . . . . . . . . . . . . . 787
18.7 Two-Sided Estimates of the First Eigenvalue of the Dirichlet
Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
18.7.1 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
18.7.2 Lower Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 790
18.7.3 Upper Bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795
18.7.4 Comments to Chap. 18 . . . . . . . . . . . . . . . . . . . . . . . . . . . 800
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 803
List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 849
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 859
Introduction
In [711713] Sobolev proved general integral inequalities for dierentiable

functions of several variables and applied them to a number of problems of
mathematical physics. Sobolev considered the Banach space Wpl () of functions in Lp (), p 1, with generalized derivatives of order l integrable with
power p. In particular, using these theorems on the potential-type integrals as
well as an integral representation of functions, Sobolev established the embedding of Wpl () into Lq () or C() under certain conditions on the exponents
p, l, and q.1
Later the Sobolev theorems were generalized and rened in various ways
(Kondrashov, Ilin, Gagliardo, Nirenberg, et al.). In these studies the domains
of functions possess the so-called cone property (each point of a domain is
the vertex of a spherical cone with xed height and angle which is situated
inside the domain). Simple examples show that this condition is precise, e.g.,
if the boundary contains an outward cusp then a function in Wp1 () is not,
in general, summable with power pn/(n p), n > p, contrary to the Sobolev
inequality. On the other hand, looking at Fig. 1, the reader can easily see that
the cone property is unnecessary for the embedding Wp1 () L2p/(2p) (),
2 > p. Indeed, by unifying with its mirror image, we obtain a new domain
with the cone property for which the above embedding holds by the Sobolev
theorem. Consequently, the same is valid for the initial domain although it
does not possess the cone property.
Now we note that, even before the Sobolev results, it was known that certain integral inequalities hold under fairly weak requirements on the domain.
For instance, the Friedrichs inequality ([292], 1927)

u dx K
2
(grad u) dx +

2
u ds
A sketch of a fairly rich prehistory of Sobolev spaces can be found in Naumann

[624].
xxiii
xxiv
Introduction
Fig. 1.
was established under the sole assumption that is a bounded domain for
which the GaussGreen formula holds. In 1933, Nikod
ym [637] gave an example of a domain such that the square integrability of the gradient does not
imply the square integrability of the function dened in . The monograph
of Courant and Hilbert [216], Chap. 7, contains sucient conditions for the
validity of the Poincare inequality

u2 dx K
(grad u)2 dx +
1
mn
2
u dx
(see [663, p. 76] and [664, pp. 98104]) and of the Rellich lemma [672] on the
compactness in L2 () of the set bounded in the metric

(grad u)2 + u2 dx.
The previous historical remarks naturally suggest the problem of describing the properties of domains that are equivalent to various properties of
embedding operators.
Starting to work on this problem in 1959 as a fourth-year undergraduate
student, I discovered that Sobolev-type theorems for functions with gradients
in Lp () are valid if and only if some isoperimetric and isocapacitary inequalities hold. Such necessary and sucient conditions appeared in the early 1960s
in my works [527529, 531, 533, 534]. For p = 1 these conditions coincide with
isoperimetric inequalities between the volume and the area of a part of the
boundary of an arbitrary subset of the domain.
For p > 1, geometric functionals such as volume and area prove to be
insucient for an adequate description of the properties of domains. Here
inequalities between the volume and the p-capacity or the p-conductivity arise.
Introduction
xxv
Similar ideas were applied to complete characterizations of weight functions and measures in the norms involved in embedding theorems. Moreover,
the method of proof of the criteria does not use specic properties of the Euclidean space. The arguments can be carried over to the case of Riemannian
manifolds and even abstract metric spaces. A considerable part of the present
book (Chaps. 29 and 11) is devoted to the development of this isoperimetric
and isocapacitary ideology.
However, this theory does not exhaust the material of the book even conceptually. Without aiming at completeness, I mention that other areas of the
study in the book are related to the following questions. How massive must a
subset e of a domain be in order that the inequality
uLq () Cl uLp ()
holds for all smooth functions vanishing on e? How does the class of domains
admissible for integral inequalities depend upon additional requirements imposed upon the behavior of functions at the boundary? What are the conditions on domains and measures involved in the norms ensuring density of
a space of dierentiable functions in another one? We shall study the criteria of compactness of Sobolev-type embedding operators. Sometimes the best
constants in functional inequalities will be discussed. The embedding and extension operators involving BirbaumOrlicz spaces, the space BV of functions
whose gradients are measures, and Besov and Bessel potential spaces of functions with fractional smoothness will also be dealt with.
The investigation of the above-mentioned and similar problems is not only
of interest in its own right. By virtue of well-known general considerations it
leads to conditions for the solvability of boundary value problems for elliptic
equations and to theorems on the structure of the spectrum of the corresponding operators. Such applications are also included.
I describe briey the contents of the book. More details can be found in
the Introductions to the chapters.
Chapter 1 gives prerequisites to the theory. Along with classical facts this
chapter contains certain new results. It addresses miscellaneous topics related
to the theory of Sobolev spaces. Some of this material is of independent interest and some (Sects. 1.11.3) will be used in the sequel. The core of the
chapter is a generalized version of Sobolev embedding theorems (Sect. 1.4).
We also deal with various extension and approximation theorems (Sects. 1.5
and 1.7), and with maximal algebras in Sobolev spaces (Sect. 1.8). Section 1.6
is devoted to inequalities for functions vanishing on the boundary along with
their derivatives up to some order.
The idea of the equivalence of isoperimetric and isocapacitary inequalities
on the one hand and embedding theorems on the other hand is crucial for
Chap. 2. Most of this chapter deals with the necessary and sucient conditions
for the validity of integral inequalities for gradients of functions that vanish
at the boundary. Of special importance for applications are multidimensional
inequalities of the HardySobolev type proved in Sect. 2.1. The basic results
xxvi
Introduction
of Chap. 2 are applied to the spectral theory of the Schrodinger operator in

Sect. 2.5.
Chapters 3 and 4 briey address the so-called conductor and capacitary
inequalities, which are stronger than inequalities of the Sobolev type and are
valid for functions dened on quite general topological spaces.
The space L1p () of functions with gradients in Lp () is studied in
Chaps. 58. Chapter 5 deals with the case p = 1. Here, the necessary and
sucient conditions for the validity of embedding theorems stated in terms of
the classes J characterized by isoperimetric inequalities are found. We also
check whether some concrete domains belong to these classes. In Chaps. 6
and 7 we extend the presentation to the case p > 1. Here the criteria are
formulated in terms of the p-conductivity. In Chap. 6 we discuss theorems
on embeddings into Lq () and Lq (). Chapter 7 concerns embeddings into
L () C(). In particular, we present the necessary and sucient conditions for the validity of the previously mentioned Friedrichs and Poincare
inequalities and of the Rellich compactness lemma. In Chap. 9 we study the
essential norm and other noncompactness characteristics of the embedding
operator L1p () Lq ().
Throughout the book and especially in Chaps. 58 we include numerous
examples of domains that illustrate possible pathologies of embedding operators. For instance, in Sect. 1.1 we show that the square integrability of second
derivatives and of the function do not imply the square integrability of the
rst derivatives. In Sect. 7.5 we consider the domain for which the embedding
operator of Wp1 () into L () C() is continuous without being compact. This is impossible for domains with good boundaries. The results of
Chaps. 57 show that not only the classes of domains determine the parameters p, q, and so on in embedding theorems, but that a feedback takes place.
The criteria for the validity of integral inequalities are applied in Chap. 6 to
the theory of elliptic boundary value problems. The exhaustive results on embedding operators can be restated as necessary and sucient conditions for
the unique solvability and for the discreteness of the spectrum of boundary
value problems, in particular, of the Neumann problem.
Chapter 9, written together with Yu.D. Burago, is devoted to the study
of the space BV () consisting of the functions whose gradients are vector
charges. Here we present a necessary and sucient condition for the existence
of a bounded nonlinear extension operator BV () BV (Rn ). We nd necessary and sucient conditions for the validity of embedding theorems for
the space BV (), which are similar to those obtained for L11 () in Chap. 5.
In some integral inequalities we obtain the best constants. The results of
Sects. 9.5 and 9.6 on traces of functions in BV () make it possible to discuss boundary values of bad functions dened on bad domains. Along
with the results due to Burago and the author in Chap. 9 we present the
De GiorgiFederer theorem on conditions for the validity of the GaussGreen
formula.
Introduction
xxvii
Chapters 29 mainly concern functions with rst derivatives in Lp or in C .

This restriction is essential since the proofs use the truncation of functions
along their level surfaces. The next six chapters deal with functions that have
derivatives of any integer, and sometimes of fractional, order.
In Chap. 10 we collect (sometimes without proofs) various properties of
Bessel and Riesz potential spaces and of Besov spaces in Rn . In Chap. 10 we
also present a review of the results of the theory of (p, l)-capacities and of
nonlinear potentials.
In Chap. 11 we investigate necessary and sucient conditions for the validity of the trace inequality

uLq () CuSpl , u C0 Rn ,
(0.0.1)

where Lq () is the space with the norm ( |u|q d)1/q , is a measure, and
Spl is one of the spaces just mentioned. For q p, (0.0.1) is equivalent to the
isoperimetric inequality connecting the measure and the capacity generated
by the space Spl . This result is of the same type as the theorems in Chaps. 29.
It immediately follows from the capacitary inequality

cap Nt ; Spl tp1 dt CupS l ,
0
where Nt = {x : |u(x)| t}. Inequalities of this type, initially found by the

author for the spaces L1p () and
L2p (Rn ) [543], have proven to be useful in a
number of problems of function theory and were intensively studied.
For q > p 1 the criteria for the validity of (0.0.1), presented in Chap. 11
do not contain a capacity. In this case the measure of any ball is estimated
by a certain function of the radius.
Chapter 12 is devoted to pointwise interpolation inequalities for derivatives
of integer and fractional order.
Further, in Chap. 13 we introduce and study a certain kind of capacity. In
comparison with the capacities dened in Chap. 10, here the class of admissible
functions is restricted, they equal the unity in a neighborhood of a compactum.
(In the case of the capacities in Chap. 10, the admissible functions majorize
the unity on a compactum.) If the order l of the derivatives in the norm of the
space equals 1, then the two capacities coincide. For l = 1 they are equivalent,
which is proved in Sect. 13.3.
The capacity introduced in Chap. 13 is applied in subsequent chapters to
prove various embedding theorems. An auxiliary inequality between the Lq norm of a function on a cube and a certain Sobolev seminorm is studied in
detail in Chap. 14. This inequality is used to justify criteria for the embedding
of
Llp () into dierent function spaces in Chap. 15. By
Llp () we mean the
completion of the space C0 () with respect to the norm l uLp () . It is

known that this completion is not embedded, in general, into the distribution
space D . In Chap. 15 we present the necessary and sucient conditions for the
embeddings of
Llp () into D , Lq (, loc), and Lp (). For p = 2, these results
xxviii
Introduction
can be interpreted as necessary and sucient conditions for the solvability of

the Dirichlet problem for the polyharmonic equation in unbounded domains
provided the right-hand side is contained in D or in Lq (). In Chap. 16
we nd criteria for the boundedness and the compactness of the embedding
Llp (, )
operator of the space
Llp (, ) into Wqr (), where is a measure and
is the completion of C0 () with respect to the norm

1/p

|l u| dx +
|u| d
The topic of Chap. 17 is a necessary and sucient condition for density

of C0 () in a certain weighted Sobolev space which appears in applications.
Finally, Chap. 18 contains variations on the theme of Molchanovs discreteness
criterion for the spectrum of the Schr
odinger operator as well as two-sided
estimates for the rst DirichletLaplace eigenvalue.
Obviously, it is impossible to describe such a vast area as Sobolev spaces
in one book. The treatment of various aspects of this theory can be found
in the books by Sobolev [713]; R.A. Adams [23]; Nikolsky [639]; Besov, Ilin,
and Nikolsky [94]; Gelman and Mazya [305]; Goldshtein and Reshetnyak
[316]; Jonsson and Wallin [408]; Ziemer [813]; Triebel [758760]; D.R. Adams
and Hedberg [15]; Mazya and Poborchi [576]; Burenkov [155]; Hebey [361];
Haroske, Runst, and Schmeisser [354]; Hajlasz [342]; Salo-Coste [687]; Attouch, Buttazzo, and Michaille [54]; Tartar [744]; Haroske and Triebel [355];
Leoni [486]; Mazya and Shaposhnikova [588]; Mazya [565]; and A. Laptev
(Ed.) [479].
1
Basic Properties of Sobolev Spaces
The plan of this chapter is as follows. Sections 1.1 and 1.2 contain the prerequisites on Sobolev spaces and other function analytic facts to be used in the
book. In Sect. 1.3 a complete study of the one-dimensional Hardy inequality
with two weights is presented. The case of a weight of unrestricted sign on
the left-hand side is also included here, following Mazya and Verbitsky [593].
Section 1.4 contains theorems on necessary and sucient conditions for the
Lq integrability with respect to an arbitrary measure of functions in Wpl ().
These results are due to D.R. Adams, p > 1, [2, 3] and the author, p = 1,
[551]. Here, as in Sobolevs papers, it is assumed that the domain is good,
for instance, it possesses the cone property. In general, in requirements on a
domain in Chap. 1 we follow the all or nothing principle. However, this rule
is violated in Sect. 1.5 which concerns the class preserving extension of functions in Sobolev spaces. In particular, we consider an example of a domain for
which the extension operator exists and which is not a quasicircle.
In Sect. 1.6 an integral representation of functions in Wpl () that vanish
on along with all their derivatives up to order k 1, 2k l, is obtained.
This representation entails the embedding theorems of the Sobolev type for
any bounded domain . In the case 2k < l it is shown by example that some
requirements on are necessary. Section 1.7 is devoted to an approximation
of Sobolev functions by bounded ones. Here we reveal a dierence between
the cases l = 1 and l > 1. The chapter nishes with a discussion in Sect. 1.8
of the maximal subalgebra of Wpl () with respect to multiplication.
1.1 The Spaces Llp(), Vpl() and Wpl ()

1.1.1 Notation
Let be an open subset of n-dimensional Euclidean space Rn = {x}. Con stand for
nected open sets will be called domains. The notations and
the boundary and the closure of , respectively. Let C () denote the space
V. Mazya, Sobolev Spaces,
Grundlehren der mathematischen Wissenschaften 342,
c Springer-Verlag Berlin Heidelberg 2011
DOI 10.1007/978-3-642-15564-2 1,
1 Basic Properties of Sobolev Spaces
we mean the space of

of innitely dierentiable functions on ; by C ()
n
restrictions to of functions in C (R ).
In what follows D() or C0 () is the space of functions in C (Rn ) with
and C k () of functions
compact supports in . The classes C k (), C k (),
0
and
with continuous derivatives of order k and the classes C k, (), C k, (),
k,
C0 () of functions for which the derivatives of order k satisfy a Holder
condition with exponent (0, 1] are dened in an analogous way.
Let D () be the space of distributions dual to D() (cf. L. Schwartz [695],
Gelfand and Shilov [304]). Let Lp (), 1 p < , denote the space of
Lebesgue measurable functions, dened on , for which
1/p
|f |p dx
f Lp () =
< .
We use the notation L () for the space of essentially bounded Lebesgue

measurable functions, i.e., uniformly bounded up to a set of measure zero. As
a norm of f in L () one can take its essential supremum, i.e.,

f L () = inf c > 0 : |f (x)| c for almost all x .
By Lp (, loc) we mean the space of functions locally integrable with power
p in . The space Lp (, loc) can be naturally equipped with a countable
system of seminorms uLp (k ) , where {k }k1is a sequence of domains with
k k+1 , and k k = . Then Lp (, loc)
compact closures
k ,
becomes a complete metrizable space.
If = Rn we shall often omit in notations of spaces and norms. Integration without indication of limits extends over Rn . Further, let supp f be
the support of a function f and let dist(F, E) denote the distance between
the sets F and E. Let B(x, ) or B (x) denote the open ball with center x
and radius , B = B (0). We shall use the notation mn for n-dimensional
Lebesgue measure in Rn and vn for mn (B1 ).
Let c, c1 , c2 , . . . , denote positive constants that depend only on dimensionless parameters n, p, l, and the like. We call the quantities a and b equivalent and write a b
if c1 a b c2 a. If is a multi-index (1 , . . . , n ),
then, as usual, || = j j , ! = 1 !, . . . , n !, D = Dx11 , . . . , Dxnn , where
n
1
Dxi = /xi , x = x
1 , . . . , xn . The inequality means that i i
for i = 1, . . . , n. Finally, l = {D }, where || = l and = 1 .
1.1.2 Local Properties of Elements in the Space Llp ()
Let Llp () denote the space of distributions on with derivatives of order l
in the space Lp (). We equip Llp () with the seminorm
l uLp () =

p/2 1/p
D u(x)
2
.
||=l
1.1 The Spaces Llp (), Vpl () and Wpl ()
Theorem. Any element of Llp () is in Lp (, loc).

Proof. Let and g be bounded open subsets of Rn such that g .
Moreover, we assume that the sets and g are contained in g and along
with their neighborhoods. We introduce D() with = 1 on g, take an
arbitrary u Llp (), and set T = u. Further, let D be such that = 1
in a neighborhood of the origin and supp B .
It is well known that the fundamental solution of the polyharmonic operator l is

for 2l < n or for odd n 2l,
cn,l (1)l |x|2ln ,
(x) =
l1
2ln
cn,l (1) |x|
log |x|, for even n 2l.
Here the constant cn,l is chosen so that l = (x) holds.
It is easy to see that l ( ) = + with D(Rn ) and denoting
Diracs function. Therefore,
T + T =
l!
D ( ) D T,
!
||=l
where the star denotes convolution. We note that T C (Rn ). So, we

have to examine the expression D ( ) D T. Using the formula
D (u) =
!
D D u,
!( )!
we obtain
D T = D (u) = D u,
in g. Hence,
D ( ) D T = D ( ) D u,
in . To conclude the proof, we observe that the integral operator with a weak

singularity, applied to D u, is continuous in Lp ().
Corollary. Let u Llp (). Then all distributional derivatives D u with
|| = 0, 1, . . . , l 1 belong to the space Lp (, loc).
l||
The proof follows immediately from the inclusion D u Lp

the above theorem.
() and
Remark. By making use of the results in Sect. 1.4.5 we can rene the
theorem to obtain more information on local properties of elements in Llp ().
By the above theorem, if is connected, we can supply Llp () with the
norm
(1.1.1)
uLlp () = l uLp () + uLp () ,
where is an arbitrary bounded open nonempty set with
.
1.1.3 Absolute Continuity of Functions in L1p ()

First we mention some simple facts concerning the approximation of functions
in Lp () by smooth functions.
Let D, 0, supp B1 , and

(x) dx = 1.
With any u Lp () that vanishes on Rn \, we associate the family of
its mollications

xy
n
u(y) dy.
(M u) =
The function is called a mollier and is called a radius of mollication.

We formulate some almost obvious properties of a mollication:
1. M u C (Rn );
2. If u Lp (), then M u u in Lp () and M uLp (Rn ) uLp () ;
3. If is a bounded domain
, then for suciently small
D M u = M D u,
in . Hence, for u Llp (),
D M u D u
in Lp (), || l.
The properties of a mollication enable us to prove easily that

l uLp () = 0 is equivalent to asserting that u is a polynomial of a degree not higher than l 1.
We now discuss a well-known property of L1p (), p 1. A function dened
on is said to be absolutely continuous on the straight line l if this function
is absolutely continuous on any segment of l, contained in .
Theorem 1. Any function in L1p () (possibly modied on a set of zero
measure mn ) is absolutely continuous on almost all straight lines that are
parallel to the coordinate axes. The distributional gradient of a function in
L1p () coincides with the usual gradient almost everywhere.
In the proof of this assertion we use the following lemma.
Lemma. There is a sequence {k } of functions in D(0, 1) such that inclusion g L1 (0, 1) and equations
g(t)k (t) dt = 0
for all k = 1, 2, . . . , imply that g(t) = const a.e. on (0, 1).
(0, 1).
Proof. Let be any interval with rational endpoints such that
Let () denote the collection of mollications of the characteristic function
1
i
=
with the radii dist(, R \(0, 1))/2 , i = 1, 2, . . . . Clearly the union
()
is
a
countable
subset
of
D(0,
1),
hence
is
a
sequence
=
{
}
k
k=1 .
We observe that if f L1 (0, 1) and

f (t)k (t) dt = 0
0
for all k = 1, 2, . . . , then

f (t) dt = 0
e
for any interval e (0, 1) and hence for any measurable subset e of (0, 1).
Thus f = 0 a.e. on (0, 1).
Now k can be dened by

1
t
(s) (s)
k (t) =
k ( ) d ds,
0
where D(0, 1) and
(t) dt = 1.
0
Indeed, if g L1 (0, 1) and

g(t)k (t) dt = 0 for k = 1, 2, . . . ,
we have

0=

g(t) k (t) (t)
1

g(t)
=
0
Therefore

k (s) ds dt

g(s)(s) ds k (t) dt.
g(s)(s) ds
g(t) =
a.e. on (0, 1).
For the proof of Theorem 1.1.3/1 it suces to assume that = {x : 0 <

xi < 1, 1 i n}. Let x = (x1 , . . . , xn1 ). By Fubinis theorem
1
u
(x , t)
dt < for almost all x ,
0 t
where u/t is the distributional derivative. Therefore, the function

x v(x) =
0
xn
u
(x , t) dt
t
is absolutely continuous on the segment [0, 1] for almost all x and its
classical derivative coincides with u/xn for almost all xn (0, 1).
Let be an arbitrary function in D() and let {k } be the sequence from
the above lemma. After integration by parts we obtain
1
1
v
v(x , t)k (t) dt =
k (t) (x , t) dt, k = 1, 2, . . . .
t
0
0
Multiplying both sides of the last equality by (x ) and integrating over , we
obtain

v

v(x)k (xn )(x ) dx =
k (xn )(x )
dx.
x
n
By the denition of distributional derivative,

v

u(x)k (xn )(x ) dx =
k (xn )(x )
dx.
xn
Hence the left-hand sides of the two last identities are equal. Since D()
is arbitrary, we have for almost all x

u(x , xn ) v(x , xn ) k (xn ) dxn = 0,
k = 1, 2, . . . .
By the Lemma, for the same x the dierence u(x , xn ) v(x , xn )

does not depend on xn . In other words, for almost any xed x
xn
u
(x , t) dt + const,
u(x) =
t
0
which completes the proof.
The converse assertion is contained in the following theorem.

Theorem 2. If a function u dened on is absolutely continuous on
almost all straight lines that are parallel to coordinate axes and the rst classical derivatives of u belong to Lp (). Then these derivatives coincide with
the corresponding distributional derivatives, and hence u L1p ().
Proof. Let vj be the classical derivative of u with respect to xj and let
D(). After integration by parts we obtain

vj dx =
u dx,
x
j
which shows that vj is the distributional derivative of u with respect to xj .
1.1.4 Spaces Wpl () and Vpl ()

We introduce the spaces
Wpl () = Llp () Lp ()
and
Vpl () =
l

Lkp (),
k=0
equipped with the norms

uWpl () = l uLp () + uLp () ,
uVpl () =
k uLp () .
k=0
We present here two examples of domains which show that, in general, the
spaces Llp (), Wpl (), and Vpl () may be nonisomorphic if is not suciently regular.
In his paper of 1933 Nikod
ym [637] studied functions with a nite Dirichlet
integral. There he gave an example of a domain for which W21 () = L12 ().
Example 1. The domain considered by Nikod
ym is the union of the
rectangles (cf. Fig. 2)

Am = (x, y) : 21m 21m < x < 21m , 2/3 < y < 1 ,

Bm = (x, y) : 21m m < x < 21m , 1/3 y 2/3 ,

C = (x, y) : 0 < x < 1, 0 < y < 1/3 ,
where m (0, 2m1 ) and m = 1, 2, . . . .
Positive numbers m are chosen so that the series
Fig. 2.
Fig. 3.
2
m
m2 (Am ),
(1.1.2)
m=1
diverges. Let u be a continuous function on that is equal to m on Am , zero

on C, and linear on Bm . Since the series (1.1.2) diverges, u does not belong
to L2 (). On the other hand, the numbers m can be chosen to be so small
that the Dirichlet integral
2

u
dx dy,
Bm y
m=1
converges.
Example 2. The spaces W22 () and V22 () do not coincide for the domain
shown in Fig. 3. Let
on P,
0
m
2
u(x, y) = 4 (y 1)
on Sm (m = 1, 2, . . .),
m+1
(y 1) 1 on Pm (m = 1, 2, . . .).
2
We can easily check that

(2 u)2 dx dy = 22m ,
Sm
u2 dx dy = 25m ,
Sm
u2 dx dy 2m/2 ,
Pm
Sm
(u)2 dx dy 23m ,

(u)2 dx dy 2m/2 .
Pm
Therefore, uL2 () = whereas uW22 () < .

1.1.5 Approximation of Functions in Sobolev Spaces by Smooth
Functions in
Let 1 p < . The following two theorems show the possibility of approximating any function in Llp () and Wpl () by smooth functions on .
Theorem 1. The space Llp () C () is dense in Llp ().
Proof. Let {Bk }k1 be a locally nite covering of by open balls Bk
with radii rk , Bk , and let {k }k1 be a partition of unity subordinate to
this covering. Let u Llp () and let {k} be a sequence of positive numbers
which monotonically tends to zero so that the sequence of balls {(1 + k )Bk }
has the same properties as {Bk }. If Bk = B (x), then by denition we put
cBk = Bc (x).
Let wk denote themollication of uk = k u with radius k rk .
Clearly, w =
wk belongs to C (). We take (0, 1/2) and choose k to
satisfy
uk wk Llp () k .
On any bounded open set ,
, we have

u=
uk ,
where the sum contains a nite number of terms. Hence,

u wLlp ()
uk wk Llp () (1 )1 .
Therefore, w Llp () C () and
u wLlp () 2.
The theorem is proved.

The next theorem is proved similarly.
Theorem 2. The space Wpl () C () is dense in Wpl () and the space

C () is dense in Vpl ().
Vpl ()
Remark. It follows from the proof of Theorem 1 that the space Llp ()
is dense in Llp () C()
if has a compact closure. The
C () C()
l
l
same is true if Lp is replaced by Wp or by Vpl .
In fact, let k be such that
k
uk wk C()
.
10
We put
VN =
wk +
k=1
Then
uk .
k=N +1
N +1
sup w(x) VN (x)

uk wk C()
,
2
k=N +1
since w is the limit of a sequence in C().

On the other
and hence w C()
hand,

uk wk C()
u wC()

2 ,
k=1
1.1.6 Approximation of Functions in Sobolev Spaces by Functions
in C ()
We consider a domain R2 for which C () cannot be replaced by C ()

in Theorems 1.1.5/1 and 1.1.5/2. We introduce polar coordinates (, ) with
0 < 2. The boundary of the domain = {(, ) : 1 < < 2, 0 < < 2}
consists of the circles = 1, = 2, and the interval {(, ) : 1 < < 2, = 0}.
The function u = is integrable on along with all its derivatives, but it is
not absolutely continuous on segments of straight lines x = const > 0, which
intersect . According to Theorem 1.1.3/1 the function u does not belong
to Llp (1 ), where 1 is the annulus = {(, ) : 1 < < 2, 0 < 2}.
Hence, the derivatives of this function cannot be approximated in the mean
by functions in C ().
in Sobolev
A necessary and sucient condition for the density of C ()
spaces is unknown. The following two theorems contain simple sucient conditions.
Denition. A domain Rn is called starshaped with respect to a point
O if any ray with origin O has a unique common point with .
Theorem 1. Let 1 p < . If is a bounded domain, starshaped with
is dense in W l () and V l (), p [1, ). The
respect to a point, then C ()
p
p
l
same is true for the space Lp (), i.e., for any u Llp () there is a sequence
such that
{ui }i1 of functions in C ()

l (ui u)
ui u in Lp (, loc)
and
0.
Lp ()
Proof. Let u Wpl (). We may assume that is starshaped with respect
to the origin. We introduce the notation u (x) = u( x) for (0, 1). We can
easily see that u u Lp () 0 as 1.
11
From the denition of the distributional derivative it follows that D (u ) =

(D u) , || = l. Hence u Wpl ( 1 ) and

D (u u )
D u D (u )L () + D u D u L ()
Lp ()
p
p

1 l D u L () + D u D u L () .
l
The right-hand side tends to zero as 1. Therefore, u u in Wpl ().

1 , the sequence of mollications of u converges to u in
Since
l
Wp (). Now, using the diagonalization process, we can construct a sequence
that approximates u in W l (). Thus we proved the
of functions in C ()
p

density of C () in Wpl (). The spaces Llp () and Vpl () can be considered
in an analogous manner.
Theorem 2. Let 1 p < . Let be a domain with compact closure
of the class C. This means that every x has a neighborhood U such
that U has the representation xn < f (x1 , . . . , xn1 ) in some system of
is dense in
Cartesian coordinates with a continuous function f. Then C ()
l
l
l
Wp (), Vp (), and Lp ().
Proof. We limit consideration to the space Vpl (). By Theorem 1.1.5/2 we
may assume that u C () Vpl ().
Let {U } be a small covering of such that U has an explicit
representation in Cartesian coordinates and let {} be a smooth partition of
unity subordinate to this covering. It is sucient to construct the required
approximation for u.
We may specify by

= x = (x , xn ) : x G, 0 < xn < f (x ) ,
f > 0 on G. Also we may assume that u has
where G Rn1 and f C(G),
a compact support in {x : x G, xn = f (x )}.
Let denote any suciently small positive number. Obviously, u (x) =
It is also clear that for any multi-index ,
u(x , xn ) is smooth on .
0 || l, we have

D (u
u)
= D u D uL () 0
L ()
p
as +0. The result follows.
Remark. The domain , considered at the beginning of this section, for

is not dense in Sobolev spaces, has the property = .
We
which C ()
provides the density
might be tempted to suppose that the equality =
in Ll (). The following example shows that this conjecture is not
of C ()
p
true.
Example. We shall prove the existence of a bounded domain Rn such
is not dense in L1 ().
and L1p () C()
that =
p
12
We start with the case n = 2. Let K be a closed nowhere dense subset of

the segment [1, 1] and let {Bi } be a sequence of open disks constructed on
adjacent intervals of K taken as their diameters. Let B be the disk x2 +y 2 < 4
and let = B\Bi . We can choose K so that the linear measure of = {x
K : |x| < 1/2} is positive. Consider the characteristic function of the upper
halfplane y > 0 and a function C0 (1, 1) which is equal to unity on
(1/2, 1/2).
The function U , dened by
U (x, y) = (x)(x, y),
belongs to the space L1p (B) for all p 1. Suppose that uj U in L1p (),
L1 (). According to our
where {uj }j1 is a sequence of functions in C()
p
assumption, for almost all x and for all (0, 1/2),

uj (x, ) uj (x, ) =
Hence
uj (x, ) uj (x, )
dx
uj (x, y)
dy.
y
grad uj (x, y)
dx dy,
()
where () = (, ).
Since uj U in L11 (), the integrals
grad uj (x, y)
dx dy,
j 1,
()
are uniformly small. Therefore, for each > 0 there exists a 0 > 0 such that
for all (0, 0 )
uj (x, ) uj (x, )
dx < .
Applying Fubinis theorem, we obtain that the sequence in the left-hand side
converges to
U (x, ) U (x, )
dx = m1 ( )
as j for almost all small . Hence m1 ( ) which contradicts the

the required counterexample has been
positiveness of m1 ( ). Since = ,
constructed for n = 2.
In the case n > 2, let 2 denote the plane domain considered previously,
put = 2 (0, 1)n2 , and duplicate the above argument.
1.1.7 Transformation of Coordinates in Norms of Sobolev Spaces
Let H and G be domains in Rn and let
13

T : y x(y) = x1 (y), . . . , xn (y) ,
be a homeomorphic map of H onto G.
We say that T is a quasi-isometric map, if for any y0 H, x0 G,
lim sup
yy0
|x(y) x(y0 )|
L,
|y y0 |
lim sup
xx0
|y(x) y(x0 )|
L,
|x x0 |
(1.1.3)
and the Jacobian det x (y) preserves its sign in H.

We can check that the estimates (1.1.3) are equivalent to
x (y) L
y (x) L
a.e. on H,
a.e. on G,
where x , y are the Jacobi matrices of the mappings y x(y), x y(x)

and is the norm of the matrix. This immediately implies that the quasiisometric map satises the inequalities
Ln
det x (y)
Ln .
(1.1.4)
l 1, if the
By denition, the map T belongs to the class C l1,1 (H),
It is easy to show that if
functions y xi (y) belong to the class C l1,1 (H).
then T 1 is of the class
T is a quasi-isometric map of the class C l1,1 (H),
l1,1
C
(G).
l 1,
Theorem. Let T be a quasi-isometric map of the class C l1,1 (H),
l
which maps H onto G. Let u Vp (G) and v(y) = u(x(y)). Then v Vpl (H)
and for almost all y H the derivatives D v(y), || l exist and are expressed
by the classical formula

(1.1.5)
D v(y) =
(y) D u x(y) .
1||||
Here
(y) =
cs
n

Dsij xi (y),
i=1 j
and the summation is taken over all multi-indices s = (sij ) satisfying the
conditions

|sij | 1 = || ||.
si,j = , |sij | 1,
i,j
i,j
Moreover, the norms vVpl (H) and uVpl (G) are equivalent.
Proof. Let u C (G) Vpl (G). Then v is absolutely continuous on almost all straight lines that are parallel to coordinate axes. The rst partial
derivatives of v are expressed by the formula
14

n

v(y) xi (y) u
=
x(y) ,
ym
y
x
m
i
i=1
(1.1.6)
for almost all y. Since

vLp (H) cuLp (G) ,
it follows by Theorem 1.1.3/2 that v Vp1 (H). After the approximation of an
arbitrary u Vpl (G) by functions in C (G) Vp1 (G) (cf. Theorem 1.1.5/2)
the result follows in the case l = 1.
For l > 1 we use induction. Let (1.1.5) hold for || = l 1. Since D u
1
Vp (G), the functions y (D u)(x(y)) belong to the space Vpl (H). This and
0,1
the inclusion
(H) imply that each term on the right-hand side of
C
(1.1.4) with || = l 1 belongs to Vp1 (H). Applying (1.1.6) to (1.1.5) with
|| = l, we obtain
l vLp (H) cuVpl (G) .
The result follows.

1.1.8 Domains Starshaped with Respect to a Ball
Denition. is starshaped with respect to a ball contained in if is

starshaped with respect to each point of this ball.
Lemma. Let be a bounded domain starshaped with respect to a ball B
with radius and with center at the origin of spherical coordinates (r, ). If
has a representation r = r(), then r() satises a Lipschitz condition.
Proof. We show that for all x, y with
|x y | < 1,
the inequality
(1.1.7)
|x y| 2D2 1 |x y |
holds where D is the diameter of .

The inequality (1.1.7) means that the angle between the vectors x and
y is less than /3. We shall show that the straight line l, passing through the
points x, y, cannot intersect the ball B/2 .
In fact, if there exists a point z l B/2 , then z belongs to the segment
xy since is starshaped with respect to z. Consider the triangles Oxz, Oyz.
The inequalities |x| , |y| , |z| /2 imply |z| |y z|, |z| |x z|.
Hence Oxz /3, Oyz /3, and = Oxz Oyz /3, which
contradicts (1.1.7).
The distance from the origin O to the line l is |x||y||x y|1 sin which
is less than /2 since l B/2 = . Therefore
|x y| 21 |x||y| sin 41 D2 sin(/2) = 21 D2 | x y |,
and the result follows.
15
Remark. It is easy to see that the converse assertion also holds. Namely, if
is a bounded domain and has the representation r = r() in spherical
coordinates with r() satisfying a Lipschitz condition, then is starshaped
with respect to a ball with its center at the origin.
1.1.9 Domains of the Class C 0,1 and Domains Having the Cone
Property
Denition 1. We say that a bounded domain belongs to the class C 0,1 if
each point x has a neighborhood U such that the set U is represented by the inequality xn < f (x1 , . . . , xn1) in some Cartesian coordinate
system with function f satisfying a Lipschitz condition.
By Lemma 1.1.8 any bounded domain starshaped with respect to a ball
belongs to the class C 0,1 .
Denition 2. A domain possesses the cone property if each point of
is the vertex of a cone contained in along with its closure, the cone being
represented by the inequalities x21 + + x2n1 < b x2n , 0 < xn < a in some
Cartesian coordinate system, a, b = const.
Remark 1. It is easy to show that bounded domains of the class C 0,1 have
the cone property. The example of a ball with deleted center shows that the
converse assertion is not true.
Lemma 1. Let be a bounded domain having the cone property. Then
is the union of a nite number of domains starshaped with respect to a ball.
Since a domain having the cone property is a union of congruent cones
and hence it is a union of domains starshaped with respect to balls of a xed
radius, then Lemma 1 follows immediately from the next lemma.
Lemma 2. If a bounded domain is a union of an innite number of
domains G starshaped with respect to balls B G of a xed radius R > 0,
then for each r < R there exists a nite number of domains k (1 k N )
starshaped
with respect to balls of radius r, contained in k , and such that

=
.
k k
Proof.
Let G1 be a domain in the collection {G }. Consider the domain
1 = G , where the union is taken over all domains G for which the
distance between the centers of the balls B and B1 is R r. Obviously,
any of the balls B contain the ball C1 of radius r concentric with B1 . Since
any G is starshaped with respect to C1 , then 1 is starshaped with respect
to C1 .
We dene G2 to be any of the domains G such that G 1 = .
Repeating the preceding construction, we dene a domain 2 starshaped with
16
respect to the ball C2 of radius r with the center situated at a distance d >
R r from the center of the ball C1 . Analogously, we construct a domain 3
starshaped with respect to a ball C3 of radius r with the center situated at a
distance d > R r from the centers of the balls C1 and C2 , and so on.
Clearly this process will stop after a nite number of steps since the centers
of the balls C1 , C2 , . . . , are contained in a bounded domain and the distance
between centers is more than R r > 0. The result follows.
Remark 2. Domains of the class C 0,1 are sometimes called domains having the strong Lipschitz property, whereas Lipschitz domains are dened as
follows.
Denition 3. A bounded domain is called a Lipschitz domain if each
point of its boundary has a neighborhood U Rn such that a quasi-isometric
transformation maps U onto a cube.
Clearly, domains of the class C 0,1 are Lipschitz domains. The following
example shows that the converse is not true, i.e., a Lipschitz domain may not
have a strong Lipschitz property. Moreover, the Lipschitz domain considered
in the next example fails to have the cone property (cf. Remark 1).
Example. Let R2 be the union of the rectangles Pk = {x : |x1 2k | <
2
, 0 x2 < 2k2 }, k = 1, 2, . . . , and the square Q = {x : 0 < x1 <
1, 1 < x2 < 0}. Obviously, does not have the cone property. We shall
show that can be mapped onto the square Q by a quasi-isometric map.
We can easily check that the mapping T0 : x y = (y1 , y2 ), being the
identity on Q and dened on Pk by

y 2 = x2 ,
y1 = x1 2k 1 2k x2 + 2k ,
k2
is quasi-isometric. The image T0 is the union of the square T0 Q and the set
{y : 0 < y1 < 1, 0 y2 < f (y1 )} with f satisfying the Lipschitz condition
with the constant 4 and 0 f (y1 ) 1/8 (cf. Fig. 4).
Let be a piecewise linear function, = 1 for y > 0 and = 0 for y < 1.
The Lipschitz transformation T1 : y z, dened by
z1 = y1 ,
z2 = y2 f (y1 )(y2 ),
maps T0 onto the square {z : 0 < z1 < 1, 1 < z2 < 0}. The Jacobian of T1
is greater than 1/2; therefore T1 is a quasi-isometric mapping.
Thus, is mapped onto Q by the quasi-isometric mapping T0 T1 .
1.1.10 Sobolev Integral Representation
Theorem 1. Let be a bounded domain starshaped with respect to a ball B ,
B , and let u Llp (). Then for almost all x
17
Fig. 4.
u(x) =
x
||<l

y
u(y) dy
f (x; r, )
D u(y) dy,
rn1
(1.1.8)
||=l
where r = |y x|, = (y x)r1 , D(B1 ), f are innitely dierentiable

functions in x, r, such that
|f | c (D/)n1 ,
where c is a constant that is independent of and D is the diameter of .
Proof. It suces to put = 1. Let D(B1 ) and

(z) dz = 1.
B1
If x , z B1 , the line segment [z, x] is

First we assume that u C ().
contained in and so the following Taylors formula applies:
1
D u(z)
1

(1t)l1
u(x) =
(xz) +l
D u z+t(xz) (xz) dt.
!
!
0
||<l
||=l
Multiplying this equality by (z) and integrating over z B1 , we obtain
18
u(x) =
D u(z)
(x z) (z) dz
!
||<l B1
1 1
+l
dt
! 0
||=l

(1 t)l1 D u z + t(x z) (x z) (z) dz.
(1.1.9)
B1
Since

D u(z)(x z) (z) dz

||
u(z)Dz (x z) (z) dz,
= (1)
B1
|| < l,
B1
the former of the last two sums in (1.1.9) is a polynomial of degree l 1 in

Rn , which can be written in the form of the rst term on the right-hand side
of (1.1.8). We extend D u to be zero outside for || = l and perform the
change of variable y = z + t(x z) in the last integral over B1 in (1.1.9). Then
x z = (1 t)1 (x y),
dz = (1 t)n dy,
and the general term of the second sum in (1.1.9) is

1
dt
1
y tx
dy
D u(y)(x y)
.
! Rn
1
t
(1
t)n+1
0
The last expression can be written as

1
D u(y)(x y) K(x, y) dy,
! Rn
with
K(x, y) =

dt
y tx
n
=
r
(x + )n1 d,
1 t (1 t)n+1
r
where r = |x y| and = (y x)/|x y|. Here we have made the change of

variables
y tx
= x + ,
1t
that is, y x = (1 t) and r = (1 t).
Note that the function Rn \ {y} x K(x, y) is in C (Rn \ {y}) for any
xed y Rn . Moreover,
K(x, y) = 0 if the line segment [x, y] is not contained

in the cone Vx = zB1 [x, z]. Thus, formula (1.1.8) holds where
f (x, r, ) =
(1)l l
(x + )n1 d.
19
To check the estimate |f | c Dn1 , we observe that

D
(x + )
n1 d
f (x, r, )
l
! r
l n1
(x + )
d.
D
!
x+B1
The last integral is dominated by 2 sup{|(z)| : z B1 }, which gives the
required estimate.
Suppose u
The representation (1.1.8) has been proved for u C ().
l
Lp () and p [1, ). By Theorem 1.1.6/1, there is a sequence {ui }i1 such
that
ui u
ui C (),
in Lp (, loc),
l (ui u)Lp () 0.
Passing to the limit in (1.1.8) for ui and using the continuity of the integral
operator with a weak singularity in Lp (), we arrive at (1.1.8) in the general
case. This completes the proof of the theorem.

For = Rn we obtain a simpler integral representation of u D.
Theorem 2. If u D, then

dy

(1)l l
u(x) =
D u(y) nl ,
n
r
Rn !
(1.1.10)
||=l
where n is the (n 1)-dimensional measure of S n1 , i.e.,

n =
2n/2
,
(n/2)
(1.1.11)
and as in Theorem 1, r = |y x|, = (y x) r1 .

Proof. For xed x Rn and S n1 , we have

(1)l
l
u(x) =
r l1 l u(x + r) dr.
(l 1)! 0
r
Since
l!
l
(D u)(x + r),
u(x
+
r)
=
rl
!
||=l
it follows that
u(x) = (1) l
0
rl1

(D u)(x + r) dr.
!
||=l
Integration with respect to implies (1.1.10).
20
1.1.11 Generalized Poincar

e Inequality
The following assertion, based on Lemma 1.1.9/1 and Theorem 1.1.10/1, will
be used in Sect. 1.1.13.
Lemma. Let be a bounded domain having the cone property and let
be an arbitrary open set,
. Then for any u Llp (), p 1, there exists
a polynomial

(u, )x ,
(1.1.12)
(x) =
||l1
such that
l1

k (u )
L
p ()
Cl uLp () .
(1.1.13)
k=0
Here D() and C is a constant independent of u.

Proof. Clearly, we may assume that is a ball.
Let G be any subdomain of starshaped with respect to a ball referred to
in Lemma 1.1.9/1 and let B be the corresponding ball. We construct a nite
family of open balls {Bi }M
i=0 such that B0 = B, Bi Bi+1 = , BM = .
Since G is starshaped with respect to any ball contained in B0 B1 , then by
the integral representation (1.1.8) and by continuity of the integral operator
with the kernel |x y|lkn in Lp (G) we obtain

(1.1.14)
k uLp (G) C l uLp (G) + uLp (B0 B1 ) , 0 k < l.
Also, for i = 1, . . . , M 1,

uLp (Bi ) C l uLp (Bi ) + uLp (Bi Bi+1 ) .
Therefore,

k uLp (G) C l uLp () + uLp () .
Summing over all G we obtain

k uLp () C l uLp () + uLp () .
(1.1.15)
From the integral representation (1.1.8) for a function on it follows that

u Lp () Cl uLp () ,
where
(x) =

||<l

x
(y)u(y) dy,
D().
It remains to replace u by u in (1.1.15). The Lemma is proved.

The Lemma implies the following obvious corollary.
21
Corollary. The spaces Vpl (), Wpl (), and Llp () coincide for having
the cone property.
Remark. In this section we deliberately do not give a general formulation
of the lemma for functions on domains that have the cone property. Such a formulation follows immediately from the Lemma combined with Theorem 1.4.5,
which will appear later.
On the other hand, the class of domains, considered in the Lemma, is also
not maximal. The statements of the Lemma and of the Corollary are true
for any bounded domain which is a union of domains of the class C, dened
in Theorem 1.1.6/2. The proof is essentially the same, but we must use the
following simple property of domains of the class C instead of (1.1.14).
Let

= x : x21 + + x2n1 < 2 , 0 < xn < f (x1 , . . . , xn1 ) ,
with a continuous function f dened on the ball x21 + + x2n1 2 . Let G
denote the base of , i.e., the cylinder

x : x21 + + x2n1 < 2 , 0 < xn < min f (x1 , . . . , xn1 ) .
Then, for all u C 1 (),

uLp () C uLp () + uLp (G) .
The last inequality follows from the elementary inequality

a

b
a

p
f (t)
p dt c ap
f (t)
dt + a
f (t)
p dt ,
b 0
0
0
(1.1.16)
where f C 1 [0, a], 0 < b < a,k and c depends only on p.

The proof of (1.1.16) runs as follows. Let

1/p
a
f (t) dt
,
a = a
0
be the value of |f | at some point of (0, a). We have

a
|f (t)| dt a(p1)/p f Lp (0,a),
|a b |
0
and (1.1.16) follows.

Example. Considering the domain

= (x, y) R2 : |y| < exp(1/x), 0 < x < 1 ,
and the function u(x, y) = x2l exp(1/(p x)), we may easily check that, in
general, the space Lp () on the left-hand side of (1.1.13) cannot be replaced
by any of the spaces Lq (), q > p, for domains of the class C.
22
1.1.12 Completeness of Wpl () and Vpl ()

In the following theorem is an arbitrary open subset of Rn .
Theorem. The spaces Wpl () and Vpl () are complete.
Proof. Let {uk }k1 be a Cauchy sequence in Wpl (). Let uk u in Lp ()
and let D uk v , || = l in Lp (). For any D() we have

uD dx = lim
uk D dx
k

D uk dx = (1)l
v dx.
= (1)l lim
k
Thus, v = D u and the sequence {uk } converges to u Wpl (). The result
follows for the space Wpl (). The case of Vpl () can be considered in the same
way.
l () and Its Completeness

1.1.13 The Space L
p
Let be a domain.
Denition. L lp () is the factor space Llp ()/Pl1 , where Pk is the subspace of polynomials of degree not higher than k.
We equip L lp () with the norm l uLp () . The elements of L lp () are
classes u = {u + } where Pl1 , u Llp ().
Theorem 1. The space L lp () is complete.
Proof. Let {u k }k1 be a Cauchy sequence in L lp (). This means that for
any uk in the class u k and any multi-index , || = l, D uk T in Lp ().
We shall show that there exists a u Llp () such that D u = T .
, and let {j }j0 be a sequence of domains
Let B be an open ball B
with compact closures and smooth boundaries such that

0 ,
B
j j+1 ,
j = .
j
Let k be a polynomial specied by the set = B and by the function uk in

Lemma 1.1.11. Since uk is in the class u k , then
vk = uk k
is in the same class. By Lemma 1.1.11, {vk } is a Cauchy sequence in Lp (j )
for any j. We denote the limit function by u. Clearly, for any D() and
any multi-index with || = l,

u, D = lim vk , D = lim (1)l D uk , = (1)l (T , ).
k
23
The result follows.

The proof of Theorem 1 also contains the following assertion.
Theorem 2. Let {uk } be a sequence of functions in Llp () such that

k
l (uk u)
0
Lp ()
for some u Llp (). Then there exists a sequence of polynomials k Pl1
with uk k u in Llp (, loc).
l () and
1.1.14 Estimate of Intermediate Derivative and Spaces W
p
l
Lp ()
We start with an optimal estimate of the rst derivative by the L1 norms of
the function and its second derivative.
Lemma. Let < a < b < . For all f W12 (a, b) and for every
x [a, b]
b
b

f (x)

f (t)
dt +
f (t)
dt.
(1.1.17)
2
(b a) a
a
Both constants 4(a b)2 and 1 in front of the integrals on the right-hand
side are sharp.
Proof. By dilation, (1.1.17) is equivalent to the inequality
f (x)

f (t)
dt +
1
1
f (t)
dt.
(1.1.18)
Integrating by parts, one checks the identity

f (x) =
1
K (x, t)f (t) dt +

t
where
K (x, t) =
1
t
(sign t)f (t) dt,

1
(1 |y|) dy
(|y| 1) dy
for t < x,
for t > x.
Inequality (1.1.18) follows from

1 |y| dy = 1.
Putting f (t) = t in (1.1.18), we see that 4(b a)2 is the best constant in
(1.1.17). To show the sharpness of the constant 1, we choose f as a smooth
24
approximation of the function 1 (t 1 + )+ , where is an arbitrarily small

positive number.

Remark. The just proved lemma leads directly to the inequality

(1.1.19)
f pLp (a,b) c(p) (b a)f pLp (a,b) + (b a)1 f pLp (a,b) ,
where p [1, ]. In its turn, (1.1.19) implies

f pLp (R) c(p) f pLp (R) + 1 f pLp (R) ,
(1.1.20)
with an arbitrary > 0. Clearly, (1.1.20) is equivalent to the inequality

1/p 1/2
1/2
f Lp (R) 2c(p)
f Lp (R) f Lp (R) ,
(1.1.21)
which can be found in Hardy, Littlewood, and P

olya [351], Sect. 7.8 (see also
Lemma 1.8.1/2 in the present book).
Let, as before, denote an open subset of Rn . We conclude this section
l () and W
l (), as completions of C () in the
by introducing the spaces V
p
p
0
norms of Vpl () and Wpl (), 1 p < . In fact, these new spaces coincide,
which results from the one-dimensional inequality (1.1.21).
Another space
Llp (), to be used frequently in what follows, is dened as
the completion of C0 () in the norm l uLp () . Important properties of
Llp () will be studied in Chap. 15.

1.1.15 Duals of Sobolev Spaces
Theorem 1. Let 1 p < . Any linear functional on Llp () can be expressed
as

f (u) =
g (x)D u(x) dx,
(1.1.22)
||=l
where g Lp (), pp = p + p , and

2 1/2

f = inf
g

||=l
(1.1.23)
Lp ()
Here the inmum is taken over all collections {g }||=l , for which (1.1.22)
holds with any u Llp ().
Proof. Obviously, the right-hand side of (1.1.22) is a linear functional on
Llp () and

2 1/2

f
g
.

||=l
Lp ()
25
To express f (u) as (1.1.22), consider the space Lp () of vectors v = {v }||=l

with components in Lp (), equipped with the norm
1/2

v2

||=l
Lp ()
Since the space Llp () is complete, the range of the operator l : Llp ()
Lp () is a closed subspace of Lp (). For any vector v = l u we dene
(v) = f (u). Then
= f ,
and by the HahnBanach theorem has a norm-preserving extension to

Lp (). The proof is complete.
l () be the completion
As before, let Wpl () = Llp () Lp () and let W
p
l
of D() with respect to the norm in Wp ().
The following assertion is proved in the same manner as Theorem 1.
l ()) has the form
Theorem 2. Any linear functional on Wpl () (or on W
p
f (u) =

g (x)D u(x) + g(x)u(x) dx,
(1.1.24)
||=l
where g Lp (), g Lp (), and

1/2

2
2

g
+
g
f = inf

||=l
(1.1.25)
Lp ()
Here the inmum is taken over all collections of functions g , g Lp () for
l ()).
which (1.1.24) holds with any u Wpl () (or u W
p
The following simpler characterization of the space of linear functionals on
l () is a corollary of Theorem 2.
W
p
pl () can be identied with a genCorollary. Any linear functional on W

eralized function f D () given by

(1)l D g + g,
(1.1.26)
f=
||=l
where g and g belong to Lp (). The norm of this functional is equal to the
right-hand side of (1.1.25), where the inmum is taken over all collections of
functions g , g, entering the expression (1.1.26).
26
1.1.16 Equivalent Norms in Wpl ()

The following theorem describes a wide class of equivalent norms in Wpl ().
Theorem. Let be a bounded domain such that Llp () Lp () (for
example, has the cone property). Let F (u) be a continuous seminorm in
Wpl () such that F (l1 ) = 0 for any nonzero polynomial l1 of degree not
higher than l 1. Then the norm
l uLp () + F (u),
(1.1.27)
is equivalent to the norm in Wpl ().

Proof. Let I be the identity mapping of Wpl () into the space B()
obtained by the completion of Wpl () with respect to the norm (1.1.27). This
mapping is one-to-one, linear, and continuous. By Theorem 1.1.13/1 on the
completeness of
Llp () it follows that B() Llp (). Since Llp () = Wpl (),
then I maps Wpl () onto B(). By the Banach theorem (cf. Bourbaki [128],
Chap. 1, 3, Sect. 3), I is an isomorphism. The result follows.

Remark. The functional
F (u) =
f (u)
,
0||<l
where f are linear functionals in Wpl () such that

det f x = 0, ||, || l 1,
satises the conditions of the Theorem. For example, we can put

f (u) =
uD dx,
where D() and

(x) dx = 0.
Let P be a projector of Wpl () onto the polynomial subspace Pl1 , i.e.,

a linear continuous mapping of Wpl () onto Pl1 such that P2 = P. Then
we may take |P(u)| as F (u).
Since P(u P(u)) = 0, by the Theorem we have the equivalence of the
seminorms l uLp () and u P(u)Wpl () (cf. Lemma 1.1.11).
1.1.17 Extension of Functions in Vpl () onto Rn
In this section we discuss space-preserving extensions of functions in Vpl ()
onto the exterior of . We begin with the well-known procedure of niteorder reection.
27
We introduce the notation. Let x = (x , xn ), where x = (x1 , . . . , xn1 ), let

be an n-dimensional parallelepiped, P = (a/(l +1), a), P+ = (0, a),
P = P \P+ .
Theorem. For any integer l 0 there exists a linear mapping
C (P+ ) u u C l (P ),
such that u = u on P+ . It can be uniquely extended to a continuous mapping
Lkp (P+ ) Lkp (P ), p 1, for k = 0, 1, . . . , l.
The extension u u can be dened on the space C k,1 (P+ ), k =
0, 1, . . . , l 1. It is a continuous mapping into C k,1 (P ).
This mapping has the following property: if dist(supp u, F ) > 0 where F
is a compactum in P+ , then u = 0 in a neighborhood of F .
Proof. Let u C (P+ ). We set u = u in P+ and
u (x) =
l+1

cj u(x , jxn )
in P ,
j=1
where the coecients cj satisfy the system

l+1

(j)k cj = 1,
k = 0, . . . , l.
j=1
The determinant of this system (the Vandermonde determinant) does not

vanish.
Obviously, u C l (P ). It is also clear that
k uLp (P ) ck uLp (P+ ) .
Since C (P+ ) is dense in Lkp (P+ ) by Theorem 1.1.6/1, the mapping u u
admits a unique extension to a continuous mapping Lkp (P+ ) Lkp (P ). The
continuity of the mapping C k,l (P+ ) u u C k,l (P ) can be checked
directly.
Let dist(supp u, F ) > 0. We denote an arbitrary point of F by x and
introduce a small positive number such that (l + 1) < dist(supp u, F ).
Since u = 0 on {x P+ : |x | < (l + 1)xn }, we have u = 0 on {x P :
|x |2 + (l + 1)x2n < (l + 1)2 2 }. Thus, u = 0 if |x| < . This completes the
proof.

and with suciently
Let be a domain in Rn with compact closure
smooth boundary . Using the described extension procedure along with a
partition of unity and a local mapping of onto the halfspace, it is possible
to construct a linear continuous operator E : Vpl () Vpl (Rn ) such that
E u| = u for all u Vpl ().
28
If such an operator exists for a domain then, by denition, this domain

belongs to the class EVpl .
Thus, domains with smooth boundaries are contained in EVpl . The bounded
domains of the class C 0,1 turn out to have the same property. The last assertion is proved in Stein [724], 3, Ch.VI (cf. also Comments to the present
section).
In Sect. 1.5 we shall return to the problem of description of domains
in EVpl .
1.1.18 Removable Sets for Sobolev Functions
Here we describe a class of sets of removable singularities for elements in
Vpl (), Rn . The description will be given in terms of the (n 1)dimensional Hausdor measure. We begin with the denition of the sdimensional Hausdor measure.
Let E be a set in Rn . Consider various coverings of E by countable collections of balls of radii . For each s 0, let

ris ,
s () = vs inf
i
where ri is the radius of the ith ball, vs > 0, and the inmum is taken over
all such coverings. By the monotonicity of s , there exists the limit (nite or
innite)
Hs (E) = lim s ().
+0
This limit is called the s-dimensional Hausdor measure of E. If s is a pos(s)

itive integer, then vs = ms (B1 ), otherwise vs is any positive constant. For
s/2
example, one may put vs = / (1 + s/2) for any s 0. In the case where
s is a positive integer, s n, the Hausdor measure Hs agrees with the sdimensional area of an s-dimensional smooth manifold in Rn . In particular,
Hn (E) = mn (E) for Lebesgue measurable sets E Rn (see, e.g., Federer
[271, 3.2], Ziemer [812, 1.4.2]).
Theorem. Let u Vpl ( \ F ), where Rn is an open set and F
is a closed set in satisfying Hn1 (F ) = 0. Then u Vpl ().
Proof. If u Vpl ( \ F ), then u and D u, || l are dened almost
everywhere on and hence D u Lp (). It suces to verify that D u is
the generalized derivative of u on , i.e., for all D(),

uD dx = (1)||
D u dx.
(1.1.28)
Let
pi (x) = (x1 , . . . , xi1 , xi+1 , . . . , xn ),
1 i n,
29
denote the projection of a point x Rn on the coordinate hyperplane orthogonal to the xi axis. By assumptions, each set pi (F ) has (n 1)-dimensional
Lebesgue measure zero. Thus, almost every straight line that is parallel to the
xi axis is disjointed from F . According to Fubinis theorem, we have

u
u
dx =
dx
dxi , 1 i n,
x
x

i
i
(x )
where = pi (), x = pi (x) and (x ) is the intersection of with the line
x = const. Note that (x ) is in \ F for almost all x . An application
of Theorem 1.1.3/2 leads to

u

dx
dxi =
dx
u
dxi
u
dx.
x
x
x

i
i
i
(x )
(x )
Hence (1.1.28) follows for || = l = 1. The general case can be concluded by

induction on l. Indeed, let l 2 and let (1.1.28) hold for u Vpl1 ( \ F ) with
|| l 1. If u Vpl ( \ F ), || = l, and D = Di D for some 1 i n,
the left-hand side of (1.1.28) equals

u
D dx,
(1.1.29)
xi
since u Vp1 (\F ). By the induction hypothesis, expression (1.1.29) coincides
with the right-hand side of (1.1.28) (because u/xi Vpl1 ( \ F )). This
completes the proof.

Let us show that the condition Hn1 (F ) = 0 in the above theorem cannot
be replaced by the niteness of Hn1 (F ).
Example. Let = {x R2 : |x| < 2} and let F be the segment {x :
x2 = 0, |x1 | 1}. We introduce the set S = {x : |x1 | 1, x2 > 0} and
dene the function u on by

0
on \S,
u(x) =
exp((1 x21 )1 ) on S.
We have H1 (F ) = 2, u Vpl (\S) for any l, but u
/ Vpl ().
1.1.19 Comments to Sect. 1.1
The space Wpl () was introduced and studied in detail by Sobolev [711
713]. (Note that as early as in 1935 he also developed a theory of distributions in (C0l ) [710].) The denitions of the spaces Llp () and L lp () are
borrowed from the paper by Deny and Lions [234]. The proofs of Theorems 1.1.2, 1.1.5/1, 1.1.12, 1.1.13/1, and 1.1.13/2 follow the arguments of
30
this paper where similar results are obtained for l = 1. Theorem 1.1.5/1 was
also proved by Meyers and Serrin [599]. Concerning the contents of Sect. 1.1.3
we note that the mollication operator M was used by Leray in 1934 [487]
and independently by Sobolev in 1935 [709]. A detailed exposition of properties of molliers including those with a variable radius can be found in Chap. 2
of Burenkovs book [155]. For a history of the molliers see Naumann [624].
The property of absolute continuity on almost all straight lines parallel
to coordinate axes served as a foundation for the denition of spaces similar
ym [637], Morrey [612], and others. The
to L1p in papers by Levi [489], Nikod
example of a domain for which W22 () = V22 () (see Sect. 1.1.4) is due to the
author. The example considered at the beginning of Sect. 1.1.6 is borrowed
from the paper by Gagliardo [299]. Remark 1.1.6 and the subsequent example
are taken from the paper by Kolsrud [440]. Theorem 1.1.6/1 is given in the
textbook by Smirnov [705] and Theorem 1.1.6/2 was proved in the abovementioned paper by Gagliardo. The topic of Sect. 1.1.6 was also discussed by
Fraenkel [285] and Amick [46]. The following deep approximation result was
obtained by Hedberg (see [370]).
l () for
Theorem. Let be an arbitrary open set in Rn and let f W
p
some positive integer l and some p, 1 < p < . Then there exists a sequence
of functions { }1 , 0 1, such that supp is a compact subset of ,
l )(), and
f (L W
p
lim f f Wpl () = 0.
An ingenious approximation construction for functions in a two-dimensional bounded Jordan domain was proposed by Lewis in 1987 [492] (see also
is dense in W 1 () for 1 1. If a planar domain is not Jordan, it may
happen that for l > 1 bounded functions in the space Llp () are not dense in
Llp () (Mazya and Netrusov [572], see Sect. 1.7 in the sequel).
The problem of the approximation of Sobolev functions on planar domains
by C functions, which together with all derivatives are bounded on , was
treated by Smith, Stanoyevitch, and Stegenga in [707], where some interesting
counterexamples are given as well.
In [346] Hajlasz and Mal
y studied the approximation of mappings u :
Rm from the Sobolev space [Wp1 ()]m by a sequence {u }1 of more regular
mappings in the sense of convergence

f (x, u , u ) dx
f (x, u, u) dx
for a large class of nonlinear integrands.

During the last two decades a theory of variable exponent Sobolev spaces
W 1,p() () was developed, where the Lebesgue Lp -space is replaced by the
31
Lebesgue space with variable exponent p(x). This topic in not touched in
the present book, but we refer to Kovacik and R
akosnik [461] where such
spaces rst appeared and to the surveying papers by Diening, H
ast
o and
Nekvinda [236], Kokilashvili and Samko [439] and S. Samko [690]. In particular, the density of C0 (Rn ) in W 1,p() (Rn ) was proved by S. Samko [688]
under the so-called log-condition on p(x), standard for the variable exponent
analysis:
const
p(x) p(y)

| log |x y||
for small |x y|. This question is nontrivial because of impossibility to use
molliers directly. The Hardy type inequality in variable Lebesgue spaces are
studied by Rafeiro and Samko [669].
In connection with Sect. 1.1.7 see the books by Morrey [613], Reshetnyak [677], and the article by the author and Shaposhnikova [578], see also
Sect. 9.4 in the book [588].
The condition of being starshaped with respect to a ball and having the
cone property were introduced into the theory of Wpl spaces by Sobolev [711
713]. Lemma 1.1.9/2 was proved by Glushko [312]. The example given in
Sect. 1.1.9 is due to the author; another example of a Lipschitz domain that
does not belong to C 0,1 can be found in the book by Morrey [613]. Properties
of various classes of domains appearing in the theory of Sobolev spaces were
investigated by Fraenkel in [285]. Fraenkels paper [286] contains a thorough
study of the conditions on domains guaranteeing the embedding of the
in C (0,) ()
when > 0.
space C 1 ()
Integral representations (1.1.8) and (1.1.10) were obtained by Sobolev [712,
713] and used in his proof of embedding theorems. Various generalizations
of such representations are due to Ilin [396, 397], Smith [706] (see also the
book by Besov, Ilin, and Nikolsky [94]), and to Reshetnyak [675]. We follow
Burenkov [154] in the proof of Theorem 1.1.10/1.
The Poincare inequality for bounded domains that are the unions of domains of the class C was proved by Courant and Hilbert [216]. Properties
of functions in Llp () for a wider class of domains were studied by J.L. Lions [499].
Stanoyevitch showed [720] (see [721] for the proof) that the best constant
in the one-dimensional Poincare inequality
u u
Lp (1,1) CuLp (1,1)
is equal to 1 if p = 1 and
(p )1/p p1/p

p sin(/p)
=
,
(1/p) (1/p )
(p 1)1/p
if 1 < p < . Moreover, a unique extremal function exists if and only if
1 < p .
Theorem 1.1.16 on equivalent norms in Wpl () is due to Sobolev [713, 714].
32
The extension procedure described at the beginning of Sect. 1.1.17 was pro (see also Lichtenstein [494]). The
posed by Hestenes [378] for the space C k ()
same method was used by Nikolsky [638] and Babich [59] for Wpl (). The fact
that a space-preserving extension for functions in Wpl () (1 < p < ) onto
Rn is possible for domains of the class C 0,1 was discovered by Calderon [163].
His proof is based on the integral representation (1.1.8) along with the theorem on the continuity of the singular integral operator in Lp . A method that
is appropriate for p = 1 and p = was given by Stein [724]. The main part of
his proof was based on the extension of functions dened in a neighborhood
of a boundary point. Then, using a partition of unity, he constructed a global
extension. For the simple domain

= x = (x , xn ) : x Rn1 , xn > f (x ) ,
where f is a function on Rn1 satisfying a Lipschitz condition, the extension
of u is dened by

u (x , xn ) =
u x , xn + (x , xn ) () d, xn < f (x ).
1
Here is an innitely dierentiable function, equivalent to the distance to

. The function is dened and continuous on [1, ), decreases as
more rapidly than any power of 1 , and satises the conditions

() d = 1,
k () d = 0, k = 1, 2, . . . .
1
More information on extension operators acting on Sobolev spaces can be

found in Sect. 1.5.
Theorem 1.1.18 on removable sets for Sobolev functions is due to Vaisala
[771] (see also Reshetnyak [677, Chap. 1, 1.3]). Koskela showed that Wp1 removability of sets lying in a hyperplane depends on their thickness measured
in terms of a so-called p porosity [455].
1.2 Facts from Set Theory and Function Theory

In this section we collect some known facts from set theory and function theory
that will be used later.
1.2.1 Two Theorems on Coverings
The following assertion is a generalization of the classical Besicovitch covering
theorem, see Guzman [332] or DiBenedetto [235] for a proof.
Theorem 1. Let S be a set in Rn . With each point x S we associate
a ball Br(x) (x), r(x) > 0, and denote the collection of these balls by B. We
assume that
33
Fig. 5.
() the radii of balls in B are totally bounded,

() the sequence of radii of any disjoint sequence of balls in B tends to
zero.
Then we can choose a sequence of balls {Bm } in B such that:

1. S m Bm ;
2. There exists a number M , depending on the dimension of the space only,
such that every point of Rn belongs to at most M balls in {Bm };
3. The
m are disjoint;
balls (1/3)B

4. BB B m 4Bm .
Remark 1. Theorem 1 remains valid if balls are replaced by cubes with
edges parallel to coordinate planes. This result is contained in the paper by
Morse [615]. It also follows from [615] that balls and cubes can be replaced
by other bodies. The best value of M was studied by Sullivan [733] as well as
F
uredi and Loeb [297].
We anticipate another covering theorem by the next auxiliary assertion.
Lemma. Let g be an open subset of Rn with a smooth boundary and let
2mn (Br g) = mn (Br ). Then
s(Br g) cn r n1 ,
where cn is a positive constant which depends only on n, and s is the (n 1)dimensional area.
Proof. Let and be the characteristic functions of the sets g Br and
Br \g. For any vector z = 0 we introduce a projection mapping Pz onto the
(n1)-dimensional subspace orthogonal to z (see Fig. 5). By Fubinis theorem,
34

(1/4)vn2 r2n
= mn (g Br ) mn (Br \g) =
(x)(y) dx dy
Rn Rn

(x)(x + z) dz dx
=
n
n
R R

mn {x : x Br g, x + z Br \g} dz.
=
|z|2r
Since every interval connecting x g Br with (x + z) Br \g intersects

Br g, the last integral does not exceed

mn1 Pz (Br g) dz (2r)n+1 vn s(Br g).
2r
|z|2r
The lemma is proved.
Remark 2. The best value of cn equals the volume of the (n1)-dimensional

unit ball (cf. Lemma 5.2.1/1 below).
Theorem 2. Let g be a bounded open subset of Rn with smooth boundary.
There exists a covering of g by a sequence of balls with radii i , i = 1, 2, . . . ,
such that

n1
c s(g),
(1.2.1)
j
j
where c is a constant which depends only on n.

Proof. Each point x g is the center of a ball Br (x) for which
1
mn (Br (x) g)
= .
mn (Br (x))
2
(1.2.2)
(This ratio is a continuous function of r, which is equal to 1 for small values

of r and converges to zero as r .) Let us put B = {B3r(x) (x)}, where
Br(x) (x) satises (1.2.2). By Theorem 1 there exists a sequence of disjoint
balls Brj (xj ) such that

B3rj (xj ).
g
j=1
(Here we actually use a weaker variant of Theorem 1 (cf. Dunford and

J. Schwartz [244], III.12.1).)
The Lemma together with (1.2.2) implies
s(Brj (xj ) g) cn rjn1 .
Therefore,
s(g)
s(Brj (xj ) g) 31n cn
Thus, {B3rj (xj )} is the required covering.
(3rj )n1 .
35
1.2.2 Theorem on Level Sets of a Smooth Function

We recall the Vitali covering theorem (see Dunford and J. Schwartz [244],
III.12.2).
Let E R1 and let M be a collection of intervals. We say that M forms
a covering of E in the sense of Vitali if for each t E and any > 0 there
exists an interval i M such that t i, m1 (i) < .
Theorem 1. If E is covered by a collection M of intervals in the sense of
Vitali, then we can select a countable
or nite set of intervals {ik } such that

ik il = for k = l, m1 (E\ k ik ) = 0.
Consider a function f
x f (x) = t R1 .
The set
K1 = {x : f (x) = 0}
is called critical.
If E then f (E) is the image of E under the mapping f . If A R1
then f 1 (A) is the pre-image of A in . We shall briey denote f 1 (t) by Et .
Theorem 2. Let be an open set in Rn and f C (). Then

m1 f (K1 ) = 0.
Proof. It is sucient to assume that is a bounded set.

1. We introduce the notation

Kn = x : (f )(x) = 0, . . . , (n f )(x) = 0 .
First we show that

m1 f (Kn ) = 0.
For any > 0 and each x Kn we choose a number rx > 0 such that
B(x, rx ) and
osc f < rxn .
B(x,rx )
We x a point t f (Kn ) and consider any point

x(t) Et Kn .
Then we cover t by intervals (t , t + ) with
n
.
< rx(t)
(1.2.3)
36
The collection of all these intervals forms a covering of f (Kn ) in the sense
of Vitali. We choose a countable system of disjoint intervals i1 , i2 , . . . , which
covers f (Kn ) up to a set of linear measure zero.
Let
im = (tm m , tm + m ),
and xm Etm Kn . By (1.2.3), < rxnm . Hence

f 1 (im ) B xm , (m /)1/n ,
and therefore,
Since the intervals im

property. Thus,

m
mn f 1 (im ) vn .
are mutually disjointed, their pre-images have the same

mn f 1 (im ) mn (),
vn m=1
i.e., m1 [f (Kn )] cmn (), hence m1 [f (Kn )] = 0.

2. Now we use an induction on n. The theorem holds for n = 1. Assume
it holds for n 1.
Consider the set K1 \Kn . For any x K1 \Kn there exists a multi-index
, || < n, and an integer i n such that

D f (x) = 0,
D f (x) = 0.
(1.2.4)
xi
Let H be a set of points for which (1.2.4) holds. This set is obviously dened
by the pairs (, i). We show that

m1 f (H) = 0.
Without loss of generality we may assume that i = n. With the notation
g = D f , we have
g(x) = 0,
g
= 0 for x H.
xn
By the implicit function theorem, for any x0 H there exists a neighborhood

U such that

U x : g(x) = 0 x : xn = (X) ,
where X = (x1 , . . . , xn1 ) and is an innitely dierentiable function in some
domain G Rn1 . Since we can select a countable covering from any covering
of H, it is sucient to prove that

m1 f (H U ) = 0.
If x H U , then
37

def
f (x) = f X, (X) = h(X),
where X G. Let P denote the projection of H U onto the plane xn = 0.

Since h = 0 for X P , by the induction hypothesis we have

m1 h(P ) = 0.
Taking into account that h(P ) = f (H U ), we complete the proof.
From Theorem 2 and the implicit function theorem we immediately obtain

the following corollary.
Corollary. Let f C () (f D()), then for almost all t the sets
Et = {x : f (x) = t} are C -manifolds (C -compact manifolds).
1.2.3 Representation of the Lebesgue Integral as a Riemann
Integral along a Halfaxis
Lemma. Let (X, B, ) be a space with a (nonnegative) -nite measure and
let u : X R1 be a -measurable nonnegative function. Then

u(x)(dx) =
(Lt ) dt =
(Mt ) dt,
(1.2.5)
0
where

Lt = x X : u(x) > t ,

Mt = x X : u(x) t .
Proof. Let (a,b) denote the characteristic function of the interval (a, b) of
real axis. Writing
u(x) =
(0,u(x)) (t) dt,

0
and using Fubinis theorem on the product space X (0, ), we obtain

u(x)(dx) =
dt
(0,u(x)) (t)(dx) =
(Lt ) dt.
X
Thus the rst equality (1.2.5) is established. The second equality can be obtained in the same way. The lemma is proved.

Remark. We can easily derive a generalization of (1.2.5) for the integral

u(x)(dx),
X
where is a charge, u is not necessarily of a denite sign, and

u(x)
||(dx) < .
X
38
1.2.4 Formula for the Integral of Modulus of the Gradient

Here we establish the following assertion (we refer the reader to the beginning
of Sect. 1.1.18 for the denition of Hausdor measure).
Theorem. Let be a Borel measurable nonnegative function on and
let u C 0,1 (), where is an open subset of Rn . Then

(x)
u(x)
dx =
dt
(x) ds(x),
(1.2.6)
Et
where s is the (n 1)-dimensional Hausdor measure, Et = {x :

|u(x)| = t}.
We shall derive (1.2.6) in the following weaker formulation, which will be
used in this chapter.
If C(), 0, and u C (), then (1.2.6) holds.
(Here we may assume s to be the (n 1)-dimensional Lebesgue measure,
since by Corollary 1.2.2 the sets Et are smooth manifolds.)
Proof. Let w be an n-tuple vector function in D(). Using integration by
parts and applying Lemma 1.2.3, we obtain

wu dx =
u div w dx

=
dt
0
div w dx +
dt
ut
div w dx.
ut
Since u C (), the sets {x : u(x) = t} are innitely dierentiable manifolds

for almost all t. Therefore for almost all t > 0,

wu
ds,
div w dx =
w ds =
u>t
u=t
u=t |u|
where (x) is the normal to {x : u(x) = t} directed into the set {x : u(x) t}.
The integral

div w dx,
ut
must be treated analogously. Consequently,

wu dx =
dt
Et
Setting
w=
w u
dx.
|u|
u
,
+ )1/2
(|u|2
where D() and is a positive number, we obtain
(u)2
dx =
((u)2 + )1/2
dt
Et
39
wu
ds.
|u|
Passing to the limit as 0 and making use of Beppo Levis monotone

convergence theorem, we obtain (1.2.6) for all D().
Let C(), supp and let Mh be a mollication of with
radius h. Since supp Mh for small values of h, we have

(Mh )(u) dx =
dt
Mh ds.
(1.2.7)
Et
Obviously, there exists a constant C = C() such that

Mh ds C
ds,
Et
(1.2.8)
Et

where D(), = 1 on h supp Mh , 0. By (1.2.6), applied to
= , the integral on the right-hand side of (1.2.8) is an integrable function
on (0, +). Since Mh uniformly and since s(Et supp ) < for
almost all t, then also

h0
Mh ds
ds
Et
Et
for almost all t. Now, Lebesgues theorem ensures the possibility of passing to
the limit as h 0 in (1.2.7).
Further, we remove the restriction supp . Let C(), 0
and
let m be a sequence of nonnegative functions in D() such that
m supp m = , 0 m 1 and m (x) = 1 for x supp m1 . Then
supp(m ) and

m |u| dx =
dt
m ds.
Et
Since the sequence m does not decrease, we may pass to the limit as m
by Beppo Levis theorem (see Natanson [627]). This completes the proof.
In Sect. 1.2 we collected auxiliary material most of which will be used in this
chapter.
Theorem 1.2.1/1 is due to Besicovitch in the two-dimensional case for
disks [86]. Morse generalized this result to more general spaces and shapes
[616]. Theorem 1.2.1/2 is due to Gustin [331]. Here we presented a simple
proof of Theorem 1.2.1/2 given by Federer [270]. Theorem 1.2.2/2 was proved
by Morse [614] for functions in C n . Here we followed the proof presented in the
40
book by Landis [475] Chap. II, 2. Whitney showed in [795] that there exist
functions f C n1 for which Theorem 1.2.2/2 fails.
Lemma 1.2.3 is contained in the paper by Faddeev [266]. The equality (1.2.6) was established by Kronrod [466] in the two-dimensional case
for asymptotically dierentiable functions. Federer proved a generalization
of Theorem 1.2.4 for Lipschitz mappings Rn Rm , in [269]. This result,
frequently called the coarea formula, is the identity

(x) Jm f (x) dx =
(x) dHnm (x) dy,
Rm
f 1 (y)
where Rn is an open set, f : Rm is Lipschitz, is integrable

: R, Jm f is its m-dimensional Jacobian, so that |Jm f | is the square
root of the sum of the squares of the determinants of the m m minors
of the dierential of f , and 1 m < n. Mal
y, Swanson, and Ziemer [514]
generalized the co-area formula to the case f [Wp1 (, loc)]m with p > m > 1
or p m = 1. Renements and consequences of this result concerning the
space W11 can be found in the paper by Swanson [735].
In the form of an inequality with an appropriate denition of the modulus
of the gradient of a function, Theorem 1.2.4 may be extended to abstract
metric spaces, cf. Bobkov and Houdre [116, 117].
1.3 Some Inequalities for Functions of One Variable

1.3.1 Basic Facts on Hardy-type Inequalities
Most of this section is concerned with variants and extensions of the following
Hardy inequality (cf. Hardy, Littlewood, and Polya [351], Sect. 9.9).
If f (x) 0, then
p

p
p
r
p
x F (x) dx
xr xf (x) dx,
(1.3.1)
|r
1|
0
0
where p > 1, r = 1 and
f (t) dt
F (x) =
for r > 1,
F (x) =
f (t) dt
for r < 1.
In this section we make some remarks concerning (1.3.1) and related inequalities.
(i) First of all, (1.3.1) fails if r = 1 for each of the above denitions of F .
One can see it by choosing either f or 1 f as the characteristic function of
the interval [0, 1].
41
(ii) The constant factor in front of the integral on the right-hand side of
(1.3.1) is sharp. Let, for example, r > 1, and let

r
p
x |F | dx C
xr+p |f |p dx
(1.3.2)
0
hold. We take
f (x) =

0
x
for x > 1,
rp1
for x > 1,
where is a positive suciently small number. Then

x
0
for x < 1,
r1
f (t) dt =
F (x) =
1
(x p
1) for x > 1.
r1
0
We have
xr |F |p dx =
Since (1 x
1r
p +
p
r 1 p
p
1x
1r
p +
p dx
.
x1+p
1r
p +
)p = 1 + O(x
) for x > 1, the last integral is equal to

dx
+ O(1) = (p)1 + O(1).
1+p
x
1
The right-hand side of (1.3.2) is

xr+p |f |p dx =
0
x1p dx = (p)1 .
Hence (1.3.2) becomes

pp
(r 1 p)p
1
+ O(1)
p
which implies
C
C(p)1 ,
pp
.
(r 1)p
The case r < 1 is quite similar.

(iii) A multidimensional variant of Hardys inequality is

p
u(y)
p |y|s dy
u(y)
p |y|ps dy,
|s m|
Rm
Rm
(1.3.3)
where u C0 (Rm ) and s = m. In the case s > m we require u(0) = 0. This

inequality follows from the one-dimensional inequality (1.3.3) written as
p

p m+ps1
|v(r)|p
p
v (r)
r
dr
dr,
sm+1
r
|s m|
0
0
42
if one passes to spherical coordinates r = |x|, = x/r and remarks that

| |u|. One should use, of course, that
| u
r

dx =
ds
r m1 dr.
Rm
S m1
The constant in (1.3.3) is sharp, which can be shown by radial functions.

(iv) A more general Hardys inequality
p

u(z)
p |y|s dz
u(z)
p |y|ps dz,
|s m|
Rm+n
Rm+n
where u C0 (Rm+n ), z = (x, y), x Rn , y Rm , s = m, and additionally
u(x, 0) = 0 for all x Rn if s > m, is obtained by integration of (1.3.3), with
u(z) instead of u(y), over Rn .
(v) Finally, a few words about the critical value s = m excluded in (1.3.3).
The one-dimensional inequality holds
p

p
|u(t)|p
dt
|u (t)|p tp1 dt,
(1.3.4)
t(log t)p
p1
1
1
where u(1) = 0 and p > 1. In fact, by introducing the new variable x = log t,
we rewrite (1.3.4) as the Hardy inequality
p

|
u(x)|p
p
dx
|
u (x)|p dx
xp
p1
0
0
with u
(x) = u(ex ).
As a consequence of (1.3.4) we obtain

p
p
|u(z)|p dz
u(z)
p |y|pm dz
m (log |y|)p
|y|
p
1
n+m
n+m
R
R
for all u C0 (Rn+m ) such that u(z) = 0 for |y| 1.
1.3.2 Two-weight Extensions of Hardys Type Inequality in the
Case p q
Theorem 1. Let and be nonnegative Borel measures on (0, ) and let
be the absolutely continuous part of . The inequality
q

x
1/q

1/p
f (t) dt
d(x)
C
,
(1.3.5)
f (x) d(x)
holds for all Borel functions f and 1 p q if and only if

1/q
B := sup [r, )
r>0
r
0
d
dx
1/(p1)
(p1)/p
dx
< .
(1.3.6)
43
Moreover, if C is the best constant in (1.3.5), then

BCB
q
q1
(p1)/p
q 1/q .
(1.3.7)
If p = 1 or q = , then B = C.
In the case q = the condition (1.3.6) means that

B = sup r > 0 : ([r, )) > 0 < ,
and
d
dx
> 0 for almost all x [0, B].
We begin with the proof of the following less general theorem on absolutely
continuous measures and .
Theorem 2. Let 1 p q . In order that there exists a constant C,
independent of f , such that
q 1/q

1/p
x
v(x)f (x)
p dx
dx
w(x)
f
(t)
dt
C
,
(1.3.8)
it is necessary and sucient that

1/q r
1/p
w(x)
q dx
v(x)
p dx
< ,
B := sup
r>0
(1.3.9)
where p = p/(p 1). Moreover, if C is the best constant in (1.3.8) and B is

dened by (1.3.9), then (1.3.7) holds. If p = 1 or p = , then B = C.
Proof. The case 1 < p q < . Necessity. If f 0 and supp f [0, r],
then from (1.3.8) it follows that
1/q
w(x)
q dx
r
0
Let
1/p
v(x)f (x)
p dx
.
f (t) dt C
0
v(x)
p/(p1) dx < .
We set f (x) = |v(x)|p for x < r and f (x) = 0 for x > r. Then

1/q r
1/p
w(x)
q dx
v(x)
p dx
C.
If
(1.3.10)
v(x)
p dx = ,
then we arrive at the same result, replacing v(x) by v(x) + sgn v(x) with
> 0 in (1.3.8) and passing to the limit as 0.
44
Suciency. We put
v(t)
p dt
h(x) =
1/qp
.
By Holders inequality,
q p/q

x
w(x)
f (t) dt
dx
0
0
x

q/p
w(x)
q
f (t)h(t)v(t)
p dt
h(t)v(t)
p dt
q/p
p/q
dx
.
(1.3.11)
Now we prove that

x
r 1/r
(x)
f (y) dy dx
f (y)
0
1/r
(x) dx
dy,
(1.3.12)
provided that 0, f 0 and r 1. In fact, the left-hand side in (1.3.12)
is equal to

r 1/r
(x)1/r f (y)[y,) (x) dy dx
,
0
where [y,) is the characteristic function of the halfaxis [y, ). By Minkowskis inequality the last expression does not exceed
1/r

r
(x)1/r f (y)[y,) (x) dx
dy
0
f (y)
0
1/r
(x) dx
dy.
By (1.3.12), the right-hand side in (1.3.11) is majorized by

f (t)h(t)v(t)
p
w(x)
q
h(y)v(y)
p dy
q/p
p/q
dx
dt.
(1.3.13)
Using here the expression for h, we rewrite the integral in x as
1/q q/p
x
y

w(x)
q
v(y)
p
v(z)
p dz
dy
dx.
t
Since
v(y)
p
(1.3.14)

0
v(z)
p dz
1/q
dy = q
v(x)
p dy
1/q
,
45
(1.3.14) is equal to
q/p
w(x)
q
(q )
v(y)
p dy
q/(p q )
dx.
By the denition of B this expression is majorized by

B q/q (q )q/p
w(x)
q
=B
q1
q/p
(q )
q/p
q/p
B (q )
w(y)
q dy
1/q
dx
1/q
w(x)
q dx
q
t
1/p
v(x)
p dx
q
0
= B (q )
qh(t)q .
(1.3.15)
Therefore, (1.3.11) has the following majorant

f (t)v(t)h(t)
p B q (q )q/p qh(t)q p/q dt
0

q p/p p/q
v(t)f (t)
p dt.
q
= B (q )
0
Hence, (1.3.8) holds with the constant B(q )(p1)/p q 1/q .

Now we consider the limit cases.
If p = then q = and (1.3.8) follows from the obvious estimate
x
f (t) dt
ess sup
w(x)
0<x<
0
x dt
ess sup
w(x)
ess sup
v(t)f (t)
.
0<x<
0 |v(t)| 0<t<x
If p = 1, q < , then from (1.3.12) it follows that
q 1/q
x
w(x)
f (t) dt
dx
0
0

1/q

f (t)
w(x)
q dx
v(t)
dt
|v(t)|
0
t

v(t)f (t)
dt.
B
Let q = , p = 1. Then
46
x
ess sup
w(x)
f (t) dt
0<x<
0

x
v(t)f (t)
dt
ess sup
w(x)
ess sup
0<x<
0<t<x |w(t)| 0

v(t)f (t)
dt.
B
0
If p > 1, then
f (t) dt
0<x<
0
1/p x
1/p
x

v(t)
p dt
v(t)f (t)
p dt
ess sup
w(x)
ess sup
w(x)
0<x<
v(t)f (t)
p dt
1/p
.
This completes the proof of Theorem 2.
Proof of Theorem 1. Setting f = 0 on the support of the singular part of

the measure , we obtain that (1.3.5) is the equivalent to

x

0
q
1/p
1/q
x
f (x)
p d dx
f (t) dt
d(x)
C
.
dx
0
The estimate B C can be derived in the same way as in the proof of

Theorem 2, if |v(x)|p is replaced by d /dx and

w(x)
q dx
r
by ([r, )).
Now we establish the upper bound for C. We may assume f 0. Let
{gn } be a sequence of decreasing absolutely continuous functions on [0, )
satisfying

0 gn (x) gn+1 (x) [x, ) ,

lim gn (x) = [x, ) ,
n
for almost all x. We have

q
1/q
x
f (t) dt d(x)
=
0

[x, ) d
q 1/q
f (t) dt
0
47
By the monotone convergence theorem the right-hand side is equal to

x
q 1/q
sup
gn (x) d
f (t) dt
n

x
= sup
n
q
f (t) dt
0
gn (x)
1/q
dx
.
(1.3.16)
The denition of the constant B and the sequence {gn } imply

1/q r p /p 1/p

d
gn (x) dx
dx
B.
dx
r
0
From this and Theorem 2 we conclude that the right-hand side in (1.3.16) is
not greater than

1/p

p d
,
B(q )(p1)/p q 1/q
dx
f (x)
dx
0

Replacing x by x1 we derive the following assertion from Theorem 1.
Theorem 3. Let 1 p q . In order that there exist a constant C,

independent of f and such that
q
1/q

1/p

f (t) dt
d(x)
C
,
(1.3.17)
f (x) d(x)
it is necessary and sucient that the value

1/(p1) (p1)/p

1/q
d
B := sup (0, r)
dx
,
dx
r>0
r
be nite. The best constant in (1.3.17) satises the same inequalities as in
Theorem 1.
Analogously, by the change of variable
(0, ) x y = x x1 (, +),
from Theorem 1 we obtain the next assertion.
Theorem 4. Let 1 p q . In order that there exist a constant C,
independent of f and such that
q
+

1/q
+
1/p
f (x) d(x)
f (t) dt
d(x)
C
,
(1.3.18)

B :=
sup
r(,+)

1/q
(, r)

r
d
dx
1/(p1)
(p1)/p
dx
< .
Constants B and C are related in the same way as in Theorem 1.
48
1.3.3 Two-Weight Extensions of Hardys Inequality in the Case

p>q
Lemma. Let 1 q < p and let be a nonnegative Borel function on
(0, b), where b (0, ]. In order that there exist a constant C, independent
of and such that
b
0
t
q 1/q
b
1/p
(t)
( ) d
dt
C
,
(t) dt
0
(1.3.19)

b
p/(pq)
B :=
0
(pq)/pq
< .
t(q1)p/(pq) dt
( ) d
(1.3.20)
If C is the best constant in (1.3.19), then

(q1)/q
pq
p1

q
1/q
BC
p
p1
(q1)/q
q 1/q B
for q > 1 and B = C for q = 1.

Proof. Suciency. First consider the case q > 1. We may assume
(t) 0. Integrating by parts on the left-hand side of (1.3.19) and using
H
olders inequality with exponents p/(p q), p/(q 1), and p, we obtain
b
0
( ) d
0
=q
q
(t)
b
1/q

1/q
q1
( ) d (t)
0
1/q
dt
( ) d
0
1/p b
(t) dt
0
b
( ) d
0
t(q1)p/(pq)
0
( ) d
dt
p
1/q
(q1)/p
dt
p/(pq) (pq)/p 1/q

dt
. (1.3.21)
From (1.3.20) and Hardys inequality (1.3.2), it follows that (1.3.21) is majorized by

1/p
b
(q1)/q
p
B
q 1/q
(t)p dt
.
p1
0
Necessity. Consider, for example, the case b = . The proof is similar for
b < . If (1.3.19) holds for the weight with the constant C, then it holds
49
for the weight n = [0,N ] , where [0,N ] is the characteristic function of the
segment [0, N ], with the same constant. We put
1/(pq)
fN (x) =
x(q1)/(pq) ,
N (t) dt
x

BN =
p/(pq)
N ( ) d
(pq)/p
t
(q1)p/(pq)
dt
From (1.3.19) we have

q/(pq)
CBN
=C
1/p
fN (x) dx
p
q
N (t)
fN ( ) d
1/q
dt
(1.3.22)
Integrating by parts, we nd that the right-hand side in (1.3.22) is equal to

q
fN (t)
N ( ) d
q1
fN ( ) d
1/q
dt
(1.3.23)
Since
q1
fN ( ) d
0
x(q1)/(pq)
=
N ( ) d
q1
dx
(q1)/(pq)
1/(pq)
N ( ) d

t
(p1)(q1)/(pq)
p1
pq
1q
,
we see that (1.3.23) is not less than

(1q)/q
p/(pq)
1/q
p1
(q1)p/(pq)
q
N ( ) d
t
dt
pq
0
t

(1q)/q
p1
p/(pq)
= q 1/q
BN
.
pq
1/q
Therefore,
BN q
1/q
pq
p1
(1q)/q
C,
and the same estimate is valid for B.

In the case q = 1 the condition (1.3.6) becomes especially simple:
b
p
( ) d
B=
0
1/p
dt
< .
50
To prove that in this case C B we integrate by parts on the left-hand side

of (1.3.19) and apply Holders inequality with exponents p and p (cf. (1.3.21)).
Then the right-hand side of (1.3.19) has the upper bound
b
( ) d
0
1/p
p
(t)
p dt
dt
1/p
.
Thus, we proved that C B.

To derive the inequality B C we substitute
1/(p1)
fN (x) =
N (t) dt
into (1.3.23). This yields BN C and hence B C. The lemma is proved.

Theorem 1. Let 1 q < p . Then (1.3.8) holds if and only if
x
0
dy
|v(y)|p
q1
w(y)
q dy
p/(pq)
dx
|v(x)|p
(pq)/pq
< .
(1.3.24)
If C is the best constant in (1.3.8) and B stands for the left-hand side of
(1.3.24), then
pq
p1
(q1)/q

q
1/q
BC
p
p1
(q1)/q
q 1/q B
for 1 < q < p and B = C for q = 1, 1 < p .

Proof. We may assume that f 0, since the right-hand side in (1.3.8) does
not change and the left-hand side increases if f is replaced by |f |. We may as
well assume f (x) = 0 for suciently large values of x. Let us put
x
v(y)
p dy.
t(x) =
0
Then (1.3.8) becomes

p
w(t)

v(t)

(t)
dt
1/q
(t)
p dt
1/p
,
where w(t(x))
= w(x), v(t(x)) = v(x),

x

t(x) =
f (y) dy,
0

b=
v(y)
p dy.
Now, in the case 1 q < p < the result follows from the Lemma.
51
Let p = . Then
x
q1
w(y)
q dy dx
|v(x)|
0
0
x

x
q 1/q
dy
w(x)
q
= q 1/q
dx
.
0
0 |v(y)|
B=
Hence

0
w(x)
x
0
dy
|v(y)|
1/q
q 1/q
f (t) dt
dx
Bq 1/q ess sup |vf |.
0<x<
To prove the necessity we note that v does not vanish on the set of positive
measure and put f = 1/v. The theorem is proved.

The following more general assertion can be derived from Theorem 1 in
the same way as Theorem 1.3.2/1 was derived from Theorem 1.3.2/2.
Theorem 2. Let and be nonnegative Borel measures on (0, ) and
let be the absolutely continuous part of . Inequality (1.3.8) with 1 q <
p holds for all Borel functions f if and only if

B=
< .

[x, )
x
d
dy
q1 p/(pq)
p
dy
d
dx
p
(pq)/pq
dx
The best constant C in (1.3.17) is related with B in the same manner as

in Theorem 1.
The change of variable (0, ) x y = x x1 (, +) leads to
the following necessary and sucient condition for the validity of (1.3.18):
+

(, x]
d
dy
q1 p/(pq)
p
dy
d
dx
p
dx < ,
where 1 q < p .
1.3.4 Hardy-Type Inequalities with Indenite Weights
Here we are concerned with inequalities similar to those in Sect. 1.3 with
weights of unrestricted sign. We start with the estimate
u(x)v(x)Q(x) dx
constu Lp (R+ ) v Lp (R+ )

(1.3.25)
R+
for all u, v C0 (R+ ).
52
Let us assume that Q is a locally integrable real- or complex-valued function such that
b

lim
Q(x) dx =
Q(x) dx
(1.3.26)
b+
exists for every a > 0.

Theorem 1. Under the above assumptions on Q, let

Q(t) dt, x > 0.
(x) =
x
Let 1 0
It is not dicult to see that (1.3.27) is equivalent to the pair of conditions

p1
p 1
(x)
p dx < . (1.3.28)
(x) dx < ,
sup a
sup a
a>0
a>0
Proof. For u, v C0 (R+ ), let

Qu, v =
Q(x)u(x)v(x) dx.
0
We can extend Qu, v by continuity to the case where

x
x
f (t) dt,
v(x) =
g( ) d,
u(x) =
0
for f, g C0 (R+ ), by setting

Qu, v = lim
Q(x)u(x)v(x) dx.
a+
To show that the limit on the right-hand side exists, assume that both f and
g are supported in (, b) R+ . Then clearly,

lim
a+
Q(x)u(x)v(x) dx =
0
Q(x)
f (t) dt
g( ) d
Q(x) dx
b
f (t) dt
dx
g( ) d.
Observe that we have to be careful here: In what follows one cannot estimate the two terms on the right-hand side of the preceding equation separately
because this would lead to the restriction
sup b
b>0
53
Q(x) dx
< ,
which is not necessary for the boundedness of the bilinear form.

Using Fubinis theorem, we obtain
f (t)g( )
Q(x) dx dt d
max(t, )
g( ) d
f (t)g( )
f (t) dt
Q(x) dx
By denition
Qu, v =
Q(x) dx dt d.
max(t, )

max{t, } =
Q(x) dx.
max(t, )
Thus, (1.3.25) is equivalent to the inequality

f (t)g( ) max{t, } dt d
constf Lp (R+ ) gLp (R+ )
(1.3.29)
for compactly supported f, g.

Using the reverse Holder inequality, the preceding estimate can be rewritten in the equivalent form
p

max{t, } f (t) dt
d cf pLp (R+ ) .

(1.3.30)
Clearly,

max{t, } f (t) dt = ( )
f (t) dt +
0
f (t) (t) dt.
(1.3.31)
Suppose now that (1.3.27), or equivalently, both inequalities in (1.3.28)

hold. Then the estimate involving the rst term in (1.3.31) is established by
means of the weighted Hardy inequality
p

f (t) dt

( )
d Cf pLp (R+ ) ,
(1.3.32)
which holds if and only if the rst part of condition (1.3.28) is valid (see
Theorem 1.3.2/1).
The second term in (1.3.31) is estimated by using a similar weighted Hardy
inequality
54
f (t) (t) dt
d Cf pLp (R+ ) ,
which, by Theorem 1.3.2/3, is equivalent to the second part of condition

(1.3.28). This proves the if part of the theorem.
To prove the only if part, it suces to assume that f (x) in (1.3.30) is
supported on an interval [, a], a > 0, and restrict the domain of integration
in on the left-hand side of (1.3.30) to (a, +). Taking into account that
the second term in (1.3.31) vanishes, we get
p

a

max{t,
}
f
(t)
dt
p
a
( )
p d Cf p
f (t) dt
Lp (R+ ) .
Applying the reverse Holder inequality again, we obtain the rst part of
(1.3.28)

p1
( )
p d C.
a
a
Since (1.3.30) is symmetric, a dual estimate in the Lp norm yields the
second part of (1.3.28)

( )
p d C.
ap 1
a
Hence (1.3.27) holds.
Remark 1. Notice that a similar argument works with minor changes if the
integration (1.3.25) is performed against real- or complex-valued measure dQ
in the place of Q(x) dx. However, the general case where Q is a distribution
requires taking care of some technical problems which are considered in detail
by the author and Verbitsky in [592], Sect. 2.
Remark 2. For any p (1, ), a simple condition
sup a
(a)
< ,
(1.3.33)
a>0
is sucient, but generally not necessary for (1.3.25) to hold. However, for
nonnegative Q, condition (1.3.33) is equivalent to (1.3.27).
Theorem 1 is easily carried over to the two-weight setting.
Theorem 2. Let W1 , W2 0 be locally integrable weight functions on R+
such that, respectively,
a
a
1p
W1 (x)
dx < + and
W2 (x)1p dx < +
0
55
for every a > 0. Then the two-weight bilinear inequality
const
u(x)v(x)Q(x) dx
1/p
|u (x)| W1 (x) dx
1/p
v (x)
p W2 (x) dx
(1.3.34)
holds for all u, v C0 (R+ ) if and only if the following pair of conditions
hold:
a
p1
(x)
p W2 (x)1p dx 0
and
W2 (x)
sup
a>0
1p
p 1
dx
where (x) =
(x)
p W1 (x)1p dx < ,
(1.3.36)

x
Q(t) dt.
For functions dened on the interval (0, 1), Theorem 2 can be recast in a
similar way.
Theorem 3. The inequality
1

v (x)
constu (x)
u(x)v(x)Q(x)
dx
Lp (0,1)
L
p (0,1)
(1.3.37)
holds for all u, v C (0, 1) such that u(0) = 0, v(0) = 0 if and only if Q can
be represented in the form Q = , where
1
(x)
p dx 0
as a 0+ . The corresponding compactness criterion holds with the preceding

condition replaced by
1
(x)
p dx = 0.
lim sup ap 1
a0+
For functions with zero boundary values at both endpoints, one only has
to add similar conditions at a = 1.
We now state the analog of Theorem 1 on the whole line R for the Sobolev
space Wp1 (R) which consists of absolutely continuous functions u : R C
such that
56

uWp1 (R) =
1/p

u(x)
p +
u (x)
p dx
< .
Theorem 4. Let 1 < p < , and p = max(p, p ). The inequality
u(x)v(x)Q(x) dx
const uW 1 (R) vW 1 (R)
(1.3.39)
p
p
R
holds for all u, v C0 (R), if and only if Q can be represented in the form
Q = + 0 , where and 0 satisfy the following conditions:
a+1
a+1
0 (x)
dx < .
(x) dx < ,
sup
(1.3.40)
sup
a>0
a>0
The proofs of Theorems 3 and 4 are similar to the proof of Theorem 1.

The usual approach to inequality (1.3.25), in the case where Q is real
valued, is to represent it in the form Q = Q+ Q , where Q+ and Q
are, respectively, the positive and negative parts of Q, and then treat them
separately. However, this procedure ignores a possible cancellation between
Q+ and Q and diminishes the class of admissible potentials Q.
The following examples demonstrate the dierence between sharp results
which follow from Theorem 1, and the usual approach where Q+ and Q are
treated separately.
Example 1. Let
Q(x) =
Then
(x) =
x
sin x
,
x1+
> 0.

cos x
1
sin t
dt = 1+
+ O 2+
t1+
x
x
as x +.
As x 0+, clearly, (x) = O(1) for < 1, (x) = O(log x) for = 1, and
(x) = O(x1
) for > 1. From this it is easy to see that (1.3.27) is valid if and
only if 0 2, and hence by Theorem 1, L :
L1p (R+ ) L1
p (R+ ) is bounded
for 1 < p < . Moreover, the multiplication operator Q :
L1p (R+ ) L1
p (R+ )
is compact if and only if 0 < < 2.
Note that the same Theorem 1 applied separately to Q+ and Q gives a
satisfactory result only for 1 2.
In the next example, Q is a charge on R+ , and the condition imposed on
Q depends explicitly on p.
Example 2. Let
Q=

j=1
cj (j j+1 ),
57
where a is a unit point mass at x = a. Then clearly

(x) =
cj (j,j+1) (x).
j=1
It follows that (1.3.25) holds if and only if
sup np
n1
|cj |p < .
j=n
In particular, for 1 < r 2, let cj = j 1/r if j = 2m , and cj = 0 otherwise.

Then L :
L1p (R+ ) L1
p (R+ ) if and only if r p r/(r 1). Note that in
this example condition (1.3.27) fails for all r > 1.
1.3.5 Three Inequalities for Functions on (0, )
Lemma 1. If f is a nonnegative nonincreasing function on (0, ) and p 1,
then

p

p
f (x) d xp
f (x) dx .
(1.3.41)
0
Proof. Obviously,
p1
x

p1
f (x) dx p
f (t) dt
f (x) dx
p
xf (x)
0
0
0

p
=
f (x) dx .
0
The result follows.

Lemma 2. If f (x) 0, then

a+b
f (x) dx

c(a, b, , )

xa1 f (x)a dx

xb1+ f (x)b dx ,
(1.3.42)
where a > 1, b > 1, 0 < < a, 0 < < b.
Proof. Obviously,

f (x) dx =
0
x(a1)/a f (x)
+
0
dx
x(a1)/a (1
+ x)
dx
x(b1+)/b f (x) (b1+)/b
.
x
(1 + x1 )
58
By Holders inequality

x(a1)/a f (x)
dx
x(a1)/a (1 + x)
1/a
xa1 f (x)a dx
,
dx
M
x(b1+)/b f (x) (b1+)/b
x
(1 + x1 )
b1+
1/b
f (x) dx
,
b
where
L=
M =
0
Hence

f (x) dx L
0
dx
x(a1)/(a1) (1 + x)a/(a1)
(a1)/a
,
(b1)/b
dx
(b1+)/(b1)
x
(1 + x1 )b/(b1)
1/a

xa1 f (x)a dx
+M
1/b
xb1+ f (x)b dx
.
Replacing f (x) by f (z/), where > 0, and setting z = x, we obtain

1
1/a
f (z/) dz L
(a)/a
+ M (+b)/b
a1
f (z/) dz
1/b
z b1+ f (z/)b dz
Thus for all measurable nonnegative functions on (0, ) and for any > 0,

0

(z) dz L
/a
1/a
z
0
+ M /b
a1
(z) dz
1/b
z b+1 (z)b dz
Taking the minimum of the right-hand side over , we obtain (1.3.42).
Lemma 3. If f is a nonnegative nonincreasing function on (0, ) and

p 1, then

(p 1)p1
p
p1
sup
x
f
(x)
sup
x
f (t) dt.
(1.3.43)
pp
x>0
x>0
x
The characteristic function of the interval (0, 1) turns (1.3.43) into an equality.
Proof. Let c be an arbitrary positive number. Since f does not increase,
we have
p
c
p1
59
p

cp/(p1)
p
pp
p1
f
c
f (t) dt
c
p1
(p 1)p1
c
pp
sup xp1
f (t) dt.
p1
(p 1)
x>0
x
p
Setting x = p1
c, we arrive at (1.3.43).
If f is equal to unity for 0 < x < 1 and to zero for x 1, then

p1
f (t) dt = sup xp1 (1 x)
sup x
x
0x1
(p 1)p1
(p 1)p1
=
sup xp f (x).
p
p
pp
x>0
Remark. If f is an arbitrary nonnegative measurable function on (0, ),

then the inequality, opposite to (1.3.43), holds:

1
sup xp1
sup xp f (x),
f (t) dt
(1.3.44)
p
1
x
x
x
the equality being attained for f (x) = xp .
In fact,

dt
1
p1
p1
sup xp f (x).
x
f (t) dt x
sup xp f (x) =
p
t
p
1 x
x
x
x
1.3.6 Estimates for Dierentiable Nonnegative Functions of One

Variable
Let be a strictly increasing continuous function on [0, ) such that (0) = 0
and (x) as x . By 1 we mean the inverse of . We introduce
the functions
|v(x) v(y)|
T (v; x) = sup
(|x y|)
yR
and
T (v; x) = sup
>0
(1.3.45)
|v(x ) v(x)|
.
( )
Lemma. Let f be an absolutely continuous nonnegative function on R and

let
t
(t) =
1 (y) dy.
(1.3.46)
0
Then for almost all x suppf
60
f (x)
T (f ; x) 1

f (x)
,
T (f ; x)
(1.3.47)
where 1 is the inverse of and the sign + or is taken if f (x) 0 or

f (x) 0, respectively.
Proof. It suces to consider the case f (x) 0. Let x R be xed. For
any t 0 we have
t
f (x + t) = f (x) +
f (x + ) d
0
t
f (x + ) f (x)

= f (x) + f (x)t +
( ) d.
( )
0
Since f is nonnegative it follows that
0 f (x)
f (x)
t + T+ (f ; x)
( ) d.
(1.3.48)
The right-hand side attains its minimal value at

|f (x)|
.
t = 1
T+ (f ; x)
Therefore by (1.3.48)
0 f (x)
f (x)
1
= f (x) T+ (f ; x)
|f (x)|
T+ (f ; x)
1 (
|f (x)|
)
+
T (f ;x)
+ T+ (f ; x)
1

0

|f (x)|
( ) d
T+ (f ; x)
d( ),
which is equivalent to (1.3.47).
Theorem. Let f be an absolutely continuous nonnegative function on R.

Then for all x supp f

f (x)
f (x)
T (f ; x) 1
,
(1.3.49)
T (f ; x)
where 1 is the inverse of (1.3.46). This inequality with x = 0 becomes an
equality for the function
x
( ) d.
(1.3.50)
f (x) = (1) x +
0
Proof. The estimate (1.3.49) follows from (1.3.47) if we notice that

T (v; x) T (v; x) and that the function
1
t
1 (y) dy,
t > 0,
61
(1.3.51)
is increasing.
To show the sharpness of inequality (1.3.49) we rst notice that
T (f ; 0) = 1. Further, the left-hand side of (1.3.49) is |f (0)| = 1 and its
right-hand side is

1 f (0) = 1 (1) = 1.
Thus, the equality sign is attained in (1.3.49) with x = 0 for f dened by
(1.3.50). The proof is complete.
For the particular case (t) = t we immediately have:

Corollary. Let f be a dierentiable nonnegative function on R. Then for
all x R and > 0

+1

|f (x) f (y)|
+1
f (x)
f (x) sup
.
(1.3.52)
|x y|
yR
The inequality (1.3.52) with x = 0 becomes an equality for the function
f (x) =
x+1 +
x.
+1
(1.3.53)
Remark 1. We introduce the seminorm

|v(x) v(y)|
.
x,yR (|x y|)
v = sup
Since T (f ; x) f , it follows from the above Theorem that

f (x)
f 1 f (x) .
f
(1.3.54)
If is concave then this inequality with x = 0 becomes an equality for the

function (1.3.50). In fact, it suces to show that f = 1. In view of the
concavity of
|(x) (y)|
(x) (y)
= sup
1.
(|x
y|)
x>y (x y)
x,yR
f = sup
On the other hand, the last ratio equals 1 for y = 0.

For the particular case (x) = x , (0, 1], this gives a rougher variant
of (1.3.52)

+1

|f (x) f (y)|
+1
f (x)
f (x)
sup
,
(1.3.55)
|x y|
x,yR
62
where the constant factor is still sharp since (1.3.55) with x = 0 becomes an
equality for the function (1.3.53).
Remark 2. Taking = 1 in (1.3.55) we immediately arrive at the classical
inequality

2
f (x)
2f (x) sup |f |.
(1.3.56)
We can easily improve (1.3.56) using (1.3.52) and the right and left maximal
functions dened by
1 x+
M+ (x) = sup
(y)
dy,
(1.3.57)
>0 x
x
(y)
dy.
(1.3.58)
M (x) = sup
>0 x
Clearly,
1
f (x ) f (x)
(M f )(x),
> 0.
Hence by (1.3.52) with = 1 we have the estimate

2
f (x)
2f (x)(M f )(x),
(1.3.59)
where the sign + or is taken if f (x) 0 or f (x) 0, respectively. By

Corollary the constant 2 in (1.3.59) is the best.
Concerning Sect. 1.3.1 we mention the following partial generalization of
(1.3.1) with a sharp constant where the role of F is played by the Riemann
Liouville integrals of any positive order l:
x
1
Fl (x) =
(x t)l1 f (t) dt
(1.3.60)
(l) 0
and
Fl (x) =
1
(l)
(t x)l1 f (t) dt,
(1.3.61)
where f L1 (R+ , loc) and f 0.

Theorem. (Section 329 in Hardy, Littlewood, and Polya [351]) Let p > 1
and l > 0. If Fl is dened by (1.3.60), then

0
Flp
dx
<
xlp
(1 1p )
(l + 1 1p )
p
f p dx,
except for the case when f vanishes almost everywhere.
1.4 Embedding Theorems of Sobolev Type
63
If Fl is dened by (1.3.61), then

p

( p1 )
l p
x f dx,
Flp dx <
1
(l + p )
0
0
except for the case when f vanishes almost everywhere. In both inequalities
the constants are sharp.
There are a number of papers where particular cases of the theorems in
Sects. 1.3.2 and 1.3.3 are considered. The rst criterion of such a type was
obtained by Kac and Krein [409] who dealt with the inequality (1.3.8) for q =
p = 2. For p = q, Theorems 1.3.2/1 and 1.3.2/2 are due to Muckenhoupt [620].
A dierent proof of Theorem 1.3.2/2 in the case p = q = 2 can be found in
Bobkov and Gotze [115]. The generalizations for p = q presented in Sects. 1.3.2
and 1.3.3 were obtained by Rosin and the author (see Mazya [552]). The case
p < q was independently investigated by Kokilashvili [438]. A new proof for
the case 0 < q < p was given in Sinnamon and Stepanov [702] where the
case p = 1 is included. For the history of Hardys inequality (1.3.1) and its
extensions see the book by Kufner, Maligranda, and Persson [468].
Stepanov found necessary and sucient conditions on the weight functions
u and v subject to the inequality

(k)
f (x)u(x)
p dx C
f (x)v(x)
p dx,
0
where k 1 and f vanishes together with all its derivatives up to order k 1

at x = 0 or at innity, [725, 726].
The material of Sect. 1.3.4 is borrowed from the article [593] by Mazya
and Verbitsky.
Inequality (1.3.41) is proved in the paper by Hardy, Littlewood, and
P
olya [350] and (1.3.42) is presented in the book [351] by the same authors.
Levin [490] found the best constant in (1.3.42):

s
t
) ( 1st
) a+b
( 1st
1
,
c(a, b, , ) =
s+t
(as)s (bt)t ( + ) ( 1st
)
where s = /(a + b) and t = /(a + b). Lemma 1.3.5/3 was published in
the authors book [552].
The results in Sect. 1.3.6 were obtained by Mazya and Kufner [571]. Historical remarks on multiplicative inequalities for dierentiable functions are
made in Sect. 13.5 of the book by Mazya and Shaposhnikova [579].

This section deals with a generalization of the Sobolev embedding theorem.
The heart of this result will be obtained as a corollary of estimates, in which
64
the norms in the space of functions, integrable with power p with respect to
an arbitrary measure, are majorized by the norms in Sobolev spaces. First we
shall consider functions dened on Rn and then we shall proceed to the case
of a bounded domain.
1.4.1 D.R. Adams Theorem on Riesz Potentials
Let be a measure in Rn , i.e., a nonnegative countably additive set function,
dened on a Borel algebra of Rn . Let Lq (Rn , ) = Lq () denote the space
of functions on Rn , which are integrable with power q with respect to . We
put

1/q
q
uLq () =
|u| d
.
The space Lq (, ), where is a measure on an open set , is dened in
an analogous manner.
To prove the basic result of this section, we need the classical Marcinkiewicz
interpolation theorem, which is presented here without proof (cf. for example,
Steins book [724]).
Suppose p0 , p1 , q0 , and q1 are real numbers, 1 pi qi < , p0 < p1 ,
and q0 = q1 . Let be a measure in Rn and let T be an additive operator
dened on D, its values being -measurable functions.
The operator T is said to be of weak type (pi , qi ) if there exists a constant
Ai such that for any f D and > 0,
qi

x :
(T f )(x)
> 1 Ai f Lpi .
Theorem 1. Let T be an operator of the weak types (p0 , q0 ) and (p1 , q1 ).
If 0 < < 1 and
1
1
=
+ ,
p
p0
p1
1
1
=
+ ,
q
q0
q1
then, for all f D,

T f Lq () cA01 A1 f Lp ,
and hence, T can be extended onto Lp (Rn ) as a continuous operator: Lp
Lq (). Here c = c(p1 , p2 , q1 , q2 , ) is a constant independent of , T , and f .
Now, we proceed to the statement and proof of the basic theorem of this
section.
Theorem 2. Let l > 0, 1 0
where s = q( np l), is bounded, that is, (B(x, )) const q( p l) .

n
Proof. Suciency. We show that

t(Lt )1/q vn1/p
where
p =
pq
sup M (x)1/q f Lp ,
(n pl)(q p)

Lt = y : Il |f | (y) > t ,
p
,
p1
(1.4.1)
t > 0.
Let t be the restriction of to Lt and let r be a positive number, which

will be specied later. Clearly,

dt (y)
t(Lt ) = t
Il |f | (y) dt (y)
n
Rn
R
f (x)
|x y|ln dt (y) dx.

=
Rn
Rn
By Lemma 1.2.3, the interior integral with respect to y can be expressed in

the form

ln
t y : |x y|
d =
t B(x, ) d ln .
0
(The minus sign appears since l < n and therefore = corresponds to

= 0 and vice versa.) Thus,

f (x)
t B(x, ) dxln1 d
t(Lt ) (n l)
n
0 r R

= (n l)
( )ln1 d + (n l)
( )ln1 d = A1 + A2 ,
0
where r is an arbitrary positive number.

Using the obvious inequality

1/p
M (x)1/p s/p ,
t B(x, ) t B(x, )
we obtain
r
A1 (n 1) sup M (x)
1/p
f Lp
0
Rn
1/p

t B(x, ) dx
ln1+s/p d.
66
Since
Rn

t B(x, ) dx = vn n (Lt ),
we have
A1

p(n l) 1/p
vn sup M (x)1/p f Lp (Lt )1/p r l(ns)/p .
pl n + s
Similarly,

A2 (n 1)f Lp (Lt )
1/p
r

Rn
1/p
t B(x, ) dx
ln1 d

p(n l) 1/p
f Lp (Lt )rln/p .
=
v
n pl n
Hence,
t(Lt )
1/p

f Lp vn1/p (n l)p

sup M (x)1/p l(ns)/p (Lt )1/p ln/p
.
r
r
+
pl n s
n pl
The right-hand side attains a minimum value at

rs = (Lt )/ sup M (x),
and this value is equal to

p(n l)s
v 1/p f Lp sup M (x)1/q (Lt )1/p1/q .
(n pl)(p l n + s) n
Thus, (1.4.1) is proved.

Applying interpolation Theorem 1, we nd that the operator Il : Lp
Lq () is continuous and
Il f Lp () c sup M (x)1/q f Lp .
(1.4.2)
Il f Lq() Cf Lp .
(1.4.3)
Necessity. Let
Let f denote the characteristic function of the ball B(x, ). Then, for z
B(x, ),

ln
(Il f )(z) (2)
dy = vn 2ln l .
B(x,)
This and (1.4.3) imply

1/q

B(x, )
2nl vn1/p Cl+n/p .
From Theorem 2 and the integral representation (1.1.10) we obtain the

following criterion.
Corollary. Let 1 p l.

1. For all u D,
uLq () Cl uLp ,
where
67
(1.4.4)

C q c1 sup (ln/p)q B(x, ) .
x;
2. If (1.4.4) holds for all u D, then

C q c2 sup (ln/p)q B(x, ) .
x;
Proof. The rst statement follows from Theorem 2 and the integral representation (1.1.10). The second assertion can be justied by setting

(1.4.5)
u(y) = 1 (y x) ,
where D(B2 ) and = 1 on B1 , into (1.4.4).

1.4.2 Estimate for the Norm in Lq (Rn , ) by the Integral of the
Modulus of the Gradient
In the next theorem we meet for the rst time the phenomenon of equivalence
of integral and isoperimetric inequalities:
Theorem 1. 1. Let
sup
{g}
(g)1/q
< ,
s(g)
(1.4.6)
where q 1 and {g} is a collection of subsets of an open set , g , with

compact closures and bounded by C manifolds. Then for all u D()
uLq (,) CuL1 () ,
where
C sup
{g}
(1.4.7)
(g)1/q
.
s(g)
(1.4.8)
2. Suppose that for all u D() the inequality (1.4.7) holds. Then
C sup
{g}
(g)1/q
.
s(g)
(1.4.9)
Proof. 1. By Lemma 1.2.3

uLq (,) =
0

(Lt ) d tq
1/q
,
(1.4.10)
68
where Lt = {x : |u(x)| > t}. Since (Lt ) does not increase, (1.3.41) implies

(g)1/q
(Lt )1/q dt sup
s(Lt ) dt.
uLq (,)
{g} s(g)
0
0
Here we used Corollary 1.2.2, according to which almost all sets Lt are
bounded by smooth manifolds. By Theorem 1.2.4, the last integral coincides
with uL1 () .
2. Let g be an arbitrary set in {g} and let d(x) = dist(x, g), gt = {x :
d(x) < t}. Into (1.4.6), we substitute the function u
(x) = [d(x)], where
(d) is a nondecreasing C function on [0,1], equal to one for d = 0 and to
zero for d > , > 0. According to Theorem 1.2.4,

|u | dx =
(t)s(gt ) dt.
Since s(gt ) s(g) as t 0, we have

|u | dx s(g).
(1.4.11)
On the other hand,

u Lq (,) (g)1/q .
(1.4.12)
Combining (1.4.11) and (1.4.12) with (1.4.7), we obtain

(g)1/q Cs(g),

From Theorem 1 and the classical isoperimetric inequality
mn (g)(n1)/n n1 vn1/n s(g)
(1.4.13)
(cf. Lyusternik [507], Schmidt [693], Hadwiger [334], and others), it follows
that for all u D()
uLn/(n1) n1 vn1/n uL1 ,
(1.4.14)
with the best constant.

In the case n > p 1 we replace u by |u|p(n1)/(np) in (1.4.14) and then
estimate the right-hand side by Holders inequality. We have
p(n1)(np)
uLpn/(np)

p(n 1) 1/n
|u|n(p1)/(np) u
vn
L1
n(n p)
p(n 1) 1/n
n(p1)/(np)
v
uLpn/(np)
uLp .
n(n p) n
Consequently,
uLpn/(np)
69
p(n 1) 1/n
v
uLp .
n(n p) n
This along with
|lk u|
n1/2 |lk+1 u|,
k = 1, . . . , l 1,
yields
lk uLpn/(nkp)
p(n 1)n1/2 1/n

vn
lk+1 uLpn/(n(k1)p) , (1.4.15)
n kp
where kp < n. Putting k = 1, 2, . . . , l in (1.4.15) and then multiplying all

inequalities obtained, we arrive at the next corollary.
Corollary. If n > lp, p 1, then for all u D

uLpn/(nlp)
n1
l
1/n
n1/2 vn
(n/p l)
l uLp .
(n/p)
(1.4.16)
Thus we obtained the Sobolev inequality for (p > 1) and the Gagliardo
Nirenberg inequality (p = 1) with an explicit (but not the best possible for
p > 1, l 1 or for p 1, l > 1) constant. In the case l = 1 the best constant
is known (2.3.1).
The following extension of Theorem 1 is proved in the same way, and more
general facts of a similar nature will be studied in Sect. 2.1.
Theorem 1 . The best constant C in the inequality
1/q

|u| d
q
CuL1 () ,
where 0, q 1, C(), 0, and u is an arbitrary function in

C0 (), is equal to
(g)1/q
sup
.
{g} g (x) ds
Here {g} is the same as in Theorem 1.
The next theorem shows that in the case = Rn the condition (1.4.6)
can be replaced by the equivalent one

sup (1n)q B (x) < .
(1.4.17)
x;
Theorem 2. 1. If (1.4.17) holds, then (1.4.7) holds for all u D(Rn )

with q 1 and
70

C q cq sup (1n)q B (x) ,
(1.4.18)
x;
where c depends only on n.

2. If (1.4.7) holds for all u D(Rn ), then

C q (nvn )q sup (1n)q B (x) .
(1.4.19)
x;
Proof. Let {B(xj , j )} be the covering of g constructed in Theorem 1.2.1/2.

By the obvious inequality

1/q
aj
1/q
aj ,
where aj 0, q 1, we have
(g)

1/q q
B(xj , j )
B(xj , j )
j
sup
(1n)q

B(x, )
x;
q
n1
j

(g) cq sup (1n)q B(x, ) s(g),
x;
which along with Theorem 1 yields (1.4.7).

The inequality (1.4.19) is an obvious corollary of (1.4.9). The theorem is
proved.

1.4.3 Estimate for the Norm in Lq (Rn , ) by the Integral of the
Modulus of the lth Order Gradient
Lemma. Let be a measure on Rn , n > l, 1 q < (n l + 1)(n l)1 and
1 = 1 n1 (q 1)(n l). Further, let

(I1 )(x) =
|x y|1n d(y).
Rn
Then for all x Rn and > 0

l1n I1 L (B(x,)) c
sup

r (ln)q B(x, r) .
xRn ,r>0
Proof. Without loss of generality we may put x = 0. By Minkowskis

inequality,

|x|
|y|<2
d(y)
|x y|n1
|y|<2
Since (n 1) < n, we have

|x|
|x|
1/
dx
dx
|x y|(n1)
1/
d(y).
ln1
|x|
(1.4.20)
dx
cn (n1) .
|x y|(n1)
Hence the right-hand side in (1.4.20) does not exceed cn

Consequently,
71
|y|<2
d(y)
|x y|n1
n+1
(B(2)).
1/

dx
c(ln)q B(2) .
On the other hand,

|x|<
|y|2
d(y)
|x y|n1
1/

n/
dx
c
|y|2
d(y)
.
|y|n1
The last integral is equal to

(n 1)

B(r)\B(2) rn dr,
2
and therefore it is majorized by

cq(nl)n+1 sup r (ln)q B(r) .
0<r<
The result follows.
Theorem. Let be a measure on Rn and let l n, q 1. The inequality

uLq () Cl uL1 ,
u D,
(1.4.21)
holds if and only if

K =
sup
xRn ,>0

1/q
ln B(x, )
< .
(1.4.22)
Moreover, K is equivalent to the best constant C in (1.4.21).

Proof. The estimate C c K is obvious. We prove the opposite one. In
the case l = n it follows from the identity
xn
x1
nu
dx1 , . . . , dxn , u D.
(1.4.23)
u(x) =
x1 xn
72
Let l < n. For l = 1 the result follows by Theorem 1.4.2/2. First consider the
case l > 1, q > n/(n 1). By Corollary 1.4.1
uLq () c K l1 uLn/(n1) .
By (1.4.14), we obtain that the right-hand side does not exceed cK l uL1 .
Now let l > 1, q n/(n 1). We use induction on the number of derivatives. Suppose the assertion holds for derivatives of orders 2, . . . , l 1. By the
integral representation (1.1.10),

( x) |u()|q
d
c0 q |u||u|q1 I1 d.
|u| d(x) c0
| x|n
Hence

|u|q d(x) cuq1
Ln/(n1) |u|Il L ,
where 1 = 1(q 1)(nl)n1 . By (1.4.16) the rst norm on the right-hand

side is majorized by
c
sup
xRn ,>0
l1n Il L (B(x,)) l uL1 ,
which follows by the induction hypothesis. Since q n(n 1)1 then q <
(n l + 1)(n l)1 and we may use the Lemma. Thus, the suciency of the
condition (1.4.22) as well as the estimate C cK are proved. The necessity
of (1.4.22) and the estimate C cK follow by the insertion of the test
function (1.4.5) into (1.4.21). This completes the proof.
1.4.4 Corollaries of Previous Results
The following assertion combines and complements Corollary 1.4.1 and Theorem 1.4.3.
Theorem 1. Let either k < l, p(l k) < n, 1 p < q < or l k = n,
p = 1 q . The best constant in

(1.4.24)
k uLq () Cl uLp , u D Rn ,
is equivalent to
K = sup lknp

1/q
B(x; )
.
x;
Proof. The estimate C c K is proved in Corollary 1.4.1 and in Theorem 1.4.3. Inserting

xy
u(y) = (x1 y1 )k
,

where > 0 and D(B2 ), = 1 on B1 , into (1.4.24), we obtain the lower
bound for C.
73
The next assertion is the analog of Theorem 1 for the space Vpl .
Theorem 2. Let the conditions of Theorem 1 relating the values of p, q,
l, k, and n hold. The best constant in
k uLq () CuVpl ,
u D,
is equivalent to
K1 =
sup
lknp

1/q
.
B(x, )
(1.4.25)
x;(0,1)
Proof. First we derive the upper bound for C. Let the cubes Qj form the
coordinate net in Rn with step 1 and let 2Qj be concentric homothetic cubes
with edge length 2. By {j } we denote a partition of unity subordinate to the
covering {2Qj } and such that |m j | c(m) for all j. Here c(m) is a positive
number and m is an integer. Since the multiplicity of the covering {2Qj } is
nite and depends on n only, it follows that

k (j u)
q d.
k (j u)
d c
|k u|q d
j
Applying Theorem 1 to each summand of the last sum, we obtain

k (j u)
L
q ()
c, sup lknp
x;

1/q
l (j u) .
2Qj B(x, )
L
p
Consequently,
k uLp () c K1 uVpl ,
where K1 is the constant dened by (1.4.25).
The lower bound for C can be obtained in the same way as in Theorem 1.
1.4.5 Generalized Sobolev Theorem
Theorem. Let be a domain in Rn with compact closure and let it be the
union of a nite number of domains of the class EVpl . (In particular, according
to Sect. 1.1.9 and the Stein extension theorem, mentioned in Sect. 1.1.17, this
assumption holds if has the cone property.)
Further, let be a measure on satisfying

sup s B(x, ) < ,
(1.4.26)
xRn ,>0
where s > 0 (for example, if s is an integer, then can be s-dimensional

Lebesgue measure on Rs ).
74
Then, for any u C () Vpl (),

k

j uLq (,) CuVpl () ,
(1.4.27)
j=0
where C is a constant independent of u, and the parameters q, s, p, l, and k

satisfy the inequalities
(a) p > 1, 0 < n p(l k) < s n, q sp(n p(l k))1 ;
(b) p = 1, 0 < n l + k s n, q s(n l + k)1 ;
(c) p > 1, n = p(l k), s n, q is any positive number.
If either of the conditions holds:
(d) p > 1, n < p(l k);
(e) p = 1, n l k;
then
k

sup |j u| CuVpl () .
(1.4.28)
j=0
If belongs to the class EVpl (for example, is in C 0,1 ), then in case

(d) the Theorem can be rened as follows.
(f) If p 1, (l k 1)p < n < (l k)p and = l k n/p, then for all
u Vpl () C ()
sup
x,x+h, h
=0
|k u(x + h) k u(x)|
CuVpl () .
|h|
(1.4.29)
(g) If (l k 1)p = n, then inequality (1.4.29) holds for all 0 < < 1 and
u Vpl () C ().
Proof. First we note that in cases (c) and (g) the result follows from (e)
and (f), respectively, since Vpl1 () Vpl2 () for p1 > p2 .
It is sucient to prove (1.4.27) and (1.4.28) for domains of the class EVpl .
Since for such a domain there exists an extension operator Vpl () Vpl (Rn ),
we can limit ourselves to consideration of the case = Rn . To obtain (1.4.27)
in cases (a) and (b) we refer to Theorem 1.4.4/2.
Let (d) hold. It is sucient to prove (1.4.28) for functions in Vpl (Rn )
with supports in a ball. Then (1.4.28) results from the integral representation (1.1.10) and Holders inequality.
In case (e) the estimate (1.4.28) follows directly from (1.4.23).
Let (f ) hold. Clearly, it is sucient to assume that k = 0. Since EVpl ,
then, as before, we may put = Rn . Furthermore, one can assume without
loss of generality that |h| < 1/4 and that u(y) = 0 outside the ball B(x, 1).
By (1.1.10)

K (x y)D u(y) dy,
u(x) =
||=l
Rn
where |K (z)| c|z|ln and
K (z + h) K (z)
c|h||z|l1n
Therefore,
75
for |z| 3|h|.
u(x + h) u(x)

|l u(y)|
c
dy
nl
|xy|4|h| |x y|

|l u(y)|
dy.
+ c|h|
|x
y|nl+1
|xy|4|h|
(1.4.30)
It remains to apply H
olders inequality to both integrals on the right-hand
side. The theorem is proved.
Remark 1. All the relations between n, p, l, k, and in cases (d)(g) of
the Theorem are the best possible. This fact can be veried using examples
of functions xk1 log | log |x||, |x| .
Remark 2. From the Theorem it follows that Vpl () is continuously embedded into Vqk (), q = np(n p(l k))1 for n > p(l k), p 1, if is
bounded and has the cone property. In the case n = p(l k) the same holds
for any q < . Also note that for this critical dimension there is no embedding of Vpl () C k1,1 () into C k (). The corresponding counterexample is
provided by the function
e
u(x) = x1 sin log log
,
|x|
dened on the unit ball.
In the cases p(l k) > n and p = 1, l k n the space Vpl () is embedded
into C k ().
Remark 3. If C 0,1 , then under the conditions (f ) and (g) the space
with
is embedded into the space, obtained by the completion of C ()
respect to the norm
Vpl ()
j uL () +
j=0
where
(t) =
sup
x,y,x
=y
|k u(x) k u(y)|
,
(|x y|)
t
t(1 + | log t|)(p1)/p
for (0, 1),

for = 1,
p1
(if = 1, the right-hand side in (1.4.30) is majorized by c|h| | log |h|| p

l uLp with the aid of H
olders inequality). The exponent (p 1)/p is sharp
for = 1, which can be checked by using the trial function B1/4 (0) x
xk+1
( log |x|) with (0, (p 1)/p), l k n/p = 1.
1
76
From conditions (a), (b), and (c) it follows that for integer s the restriction
operator
(1.4.31)
C () Vpl () u u|Rs ,
can be uniquely extended to a linear operator Vpl () Vqk (Rs ).
Using Lemma 1.1.11 we may rewrite (1.4.27) as
k

j (u )
Lq (,)
Cl uLp () ,
j=0
where is the polynomial (1.1.12). This enables us to introduce a continuous

restriction operator L lp Vpk (Rs )/Pl1 to Rs for = ms . Analogously, we may establish that in cases (d) and (e) and (f) and (g) the space
k,
()/Pl1 ,
L lp () is continuously embedded into C k ()/P
l1 and into C
respectively.
In conclusion we note that the Theorem of the present subsection renes
Theorem 1.1.2 on local properties of functions in Llp (), where is an arbitrary open subset of Rn .
1.4.6 Compactness Theorems
The embedding and restriction operators mentioned in Remark 1.4.5/2, which
are continuous by Theorem 1.4.5, turn out to be compact for certain values of
p, l, q, n, and s. This result will be proved at the end of the present subsection.
Lemma. Any bounded subset of the space of restrictions of the functions
in Vpl (Rn ) to a bounded domain is relatively compact in Vpl1 ().
Proof. It suces to limit consideration to the case l = 1. Let f be a
summable nonnegative function on [0, a + ], where a > 0, > 0. Then
a t+
a+
dt
f ( ) d
f (t) dt.
(1.4.32)
0
In fact, the integral on the left-hand side is

a+

a
dt
f ( + t) d =
d
f (t) dt
0
a+
f (t) dt.
0
Now let u C0 (Rn ). Obviously, for all h Rn ,

p

u(x + h) u(x)
p dx
|u| dl dx,
x,h
where x,h = [x, x + h]. Hence

u(x + h) u(x)
p dx |h|p1
|u|p dl dx.
x,h
77
Applying (1.4.32) with = |h| to the last integral, we obtain

1/p
u(x + h) u(x)
p dx
|h|uLp (Rn ) .
It remains to note that by Rieszs theorem, a set of functions, dened on a

bounded open domain , is compact in Lp (), if it is bounded in Lp () and
u(x + h) u(x)
p dx 0,
uniformly as |h| 0, where h is an arbitrary vector in Rn . This completes

the proof.

Theorem 1. Let a bounded domain Rn be the union of a nite
number of domains in EVpl (for example, has the cone property owing to
Lemma 1.1.9/1). Let be a nonnegative measure in Rn with support in .
Further, let k < l, p(l k) < n, and either 1 p < q < or 1 = p < q < .
bounded in V l (), is relatively comThen any subset of the space C (),
p
pact in the metric
k

j uLq (,)
,
(1.4.33)
j=0
if and only if

lim sup q(lkn/p) B(x, ) = 0.
(1.4.34)
0 xRn
Proof. Suciency. We may assume from the very beginning that EVpl .
Then it suces to prove that any subset of the space C (Rn ) Wpl (Rn ),
bounded in Wpl (Rn ) = Vpl (Rn ), is relatively compact in the metric (1.4.33).
According to (1.4.34), given any > 0, there exists a number such that

q(1kn/p) sup B(x, ) <
x
by balls with diameter 1,

for . We construct a covering {Bi } of
the multiplicity of the covering being bounded by a constant which depends
on n. Let i be the restriction of to Bi and let {i } be a partition of unity
subordinate to the covering {Bi }. Using Theorem 1.4.4/1, we obtain

j (ui )
q di c sup q(lkn/p) i B(x, ) ui q l
V (Bi )
Bi j=0
;x

l

p(jl)
j=0
Summing over i, we arrive at
Bi
q/p
|j u|p dx
.
78

k
j=0
|j u|q d cl uqLp (Rn ) + C()uqV l1 (

p
Bi )
It remains to note that, by the Lemma, any bounded set in Vpl (Rn ) is compact

in Vpl1 ( i Bi ).
Necessity. Let us take the origin of Cartesian coordinates to be an arbitrary
point O Rn . Let denote a function in D(B2 ) which is equal to unity on
B , < 1, and such that |j | cj , j = 1, 2, . . . .
Note that has no point charges by (1.4.34) and the inequality p(lk) < n.
: uV l (Rn ) 1} in the
From the relative compactness of the set {u C ()
p
metric (1.4.27) it follows that given any > 0, any function of this set and
any point O we have

|k u|q d
B2
for some . Inserting the function

u(x) =
xk1 (x)
,
k
x1 Vpl (Rn )
into the last inequality, we obtain

(k!)q (B ) xk1 qV l (B2 ) cq(n/pl+k) .
p
The result follows.
Theorem 2. Let a bounded domain Rn be the union of a nite

number of domains in EVpl . Then for l > k 0, p 1 we have:
(a) If s is a positive integer and n > (l k)p, then the restriction operator (1.4.31) is compact as an operator, mapping Vpl () into Vqk ( Rs ) for
n (l k)p < s n and q < sp(n (l k)p)1 .
(b) If s is a positive integer and n = (l k)p, then the operator (1.4.31)
is compact as an operator, mapping Vpl () into Vqk ( Rs ) for any q 1,
s n.
(c) If n < (l k)p, then the embedding of Vpl () into the space C k ()
equipped with the norm
k

sup |j u|
j=0
is compact.
Proof. Since Vpl1 () Vpl2 () for p1 > p2 , then (b) follows from (a). In
turn, (a) is a corollary of Theorem 1.
To obtain (c) it suces to prove the compactness of the unit ball in Vpl (Rn )
with respect to the metric of the space C k (G), where (l k)p > n and G is
and > 0. By Sobolevs estimate (1.4.28)
any bounded domain. Let x G

k

79
sup |j u| CuVpl (B(x,1)) .
j=0 B(x,1)
Applying a dilation with coecient , we obtain

k

j=0
j sup |j u| c
B(x,)
in/p i uLp (B(x,)) .
i=0
Therefore, for j = 0, . . . , k
sup |j u| cljn/p l uLp (B(x,)) + C()uVpl1 (B(x,))
B(x,)
and thus
k

sup |j u| c lkn/p uVpl (Rn ) + C()uVpl1 (G ) ,
j=0 G
Since is an arbitrarily small number,

where G is the neighborhood of G.
it follows by the Lemma that the unit ball in Vpl (Rn ) is compact in Vpl1 (G ).
Thus (c) is proved.

1.4.7 Multiplicative Inequalities
Most of this subsection is dedicated to a necessary and sucient condition for
the validity of the inequality
1
.
k uLq () Cl uLp uL
p
(1.4.35)
Lemma. Let be a measure in Rn with support in B = {x : |x| < }

and such that

(1.4.36)
K = sup r s B(x, r) <
x;r
for some s [0, n]. Further, let p 1, let k and l be integers k < l, and let
s > n p(l k) if p > 1, s n l + k if p = 1.
) and for q satisfying the inequalities l k n/p +
Then, for all v C(B
s/q > 0, q p, we have

k vLq (,B ) cK 1/q s/qn/pk l l vLp (B ) + vLp (B ) . (1.4.37)
1 ) can be exProof. According to Sect. 1.1.17, any function w C (B
tended to a function w C0l (B2 ) satisfying the inequality
l wLp (B2 ) cwVpl (B1 ) .
80
Since Vpl (B1 ) = Wpl (B1 ) (see Corollary 1.1.11), the last estimate is equivalent
to

l wLp (B2 ) c l wLp (B1 ) + wLp (B1 ) .
Thus, applying a dilation, we obtain that the function v, mentioned in the
statement of the Lemma, admits an extension v C0l (B2 ) such that

(1.4.38)
l vLp (B2 ) c l vLp (B ) + 1 vLp (B ) .
Let (l k)p < n, p > 1 or l k n, p = 1. By Theorem 1.4.4/1 we obtain
k vLt (,B2 ) cK 1/t l vLp (B2 ) ,
(1.4.39)
where t = ps/(n p(l k)).

In the case (l k)p = n, p > 1, we let p1 denote a number in [1, p),
suciently close to p. We put t = p1 s/(np1 (lk)). Then, by Corollary 1.4.1,
k vLt (,B2 ) cK 1/t l vLp1 (B2 )
cK 1/t n/p1 n/p l vLp (B2 ) .
(1.4.40)
In the case (l k)p > n, p 1 we put t = . By Sobolevs theorem

k vLt (,B2 ) clkn/p l vLp (B2 ) .
(1.4.41)
Combining (1.4.39)(1.4.41), we obtain

k vLt (,B2 ) c K 1/t lkn/p+s/t l vLp (B2 ) .
(1.4.42)
By Holders inequality,

1/q1/t
k uLq (,B ) (B )
k vLt (,B )
K 1/q1/t s(1/q1/t) k vLt (,B ) ,
which along with (1.4.42) gives
k vLt (,B2 ) cK 1/q s/q+lkn/p l vLp (B2 ) .
Using (1.4.38), we complete the proof.
Theorem. 1. Let be a measure in Rn which satises the condition

(1.4.36) for some s [0, n]. Let p 1 and let k, l be integers, 0 k l 1;
s > n p(l k) if p > 1 and s n l + k if p = 1. Then, for all u D,
the estimate (1.4.35) holds, where C cK 1/q , n/p l + k < s/q, q p and
= (k s/q + n/p)/l.
2. If (1.4.35) is valid for all u D, then C c K 1/q .
81
Proof. According to the Lemma, for all x Rn and > 0,

k uLq (,B(x,))

c K 1/q s/qn/pk l l uLp (B(x,)) + uLp (B(x,)) . (1.4.43)
We x an arbitrary 0 > 0. If the rst term on the right-hand side of (1.4.43)

exceeds the second for = 0 , then we cover a point x supp by the ball
B(x, ). Otherwise we increase until the rst term becomes equal to the
second. Then the point x is covered by the ball B(x, ), where
1/l
1/l
= uLp (B(x,)) l uLp (B(x,)) .

In both cases
sq(n/pl+k)
l uqLp (B(x,))
k uqLq (,B(x,)) cK 0

q(1 )
+ l uq
Lp (B(x,)) uLp (B(x,)) .
(1.4.44)
According to Theorem 1.2.1/1, we can select a subcovering {B (i) }i1 of

nite multiplicity, depending only on n, from the covering {B(x, )} of supp .
Summing (1.4.44) over all balls B (i) and noting that

a
i bi
a+
i
/(+)

ai
bi ,
/(+)
b+
i
where ai , bi , , and are positive numbers + 1, we arrive at

k vqLq ()
q(n/pl+k)
cK(0

i
q/p
l upLp (B(i) )
l upLp (B(i) )
q/p
(1 )q/p
upLp (B(i) )
Since the multiplicity of the covering {B (i) } depends only on n, the right-hand
side is majorized by
sq(n/pl+k)
(1 )q
cK 0
.
l uqLq + l uLqp uLp
Passing to the limit as 0 0, we complete the proof of case 1.
To prove case 2 it is sucient to insert the function u (x) = (y1 x1 )k
(1 (x y)), where D(B2 ), = 1 on B1 , into (1.4.35). The result
follows.
82
Corollary 1. 1. Let be a measure in Rn such that

sup
rs B(x, r) < ,
K1 =
(1.4.45)
xRn ,r(0,1)
for some s [0, n]. Further let p 1, let k and l be integers 0 k l 1;

s > n p(l k) if p > 1 and s n l + k if p = 1. Then, for all u D,
1
k uLq () C1 uVpl uL
,
p
(1.4.46)
1/q
where C1 cK1 , n/p l + k < s/q, q p, and = (k s/p + n/p)/l.

1/q
2. If (1.4.46) is valid for all u D, then C1 c K1 .
Proof. Let {Q(i) } denote a sequence of closed cubes with edge length 1
which forms a coordinate grid in Rn . Let O (i) be the center of the cube Q(i) ,
O (0) = O, and let 2Q (i) be the concentric homothetic cube with edge length 2.
We put i (x) = (x O (i) ),where C0 (2Q(0) ), = 1 on Q (0) .
Applying the Theorem of the present subsection to the function ui and
to the measure e (e Q (i) ), we obtain

k (ui )p
Lq ()
p
p/q
p(1 )
cK1 l (ui )L u1 Lp
.
p
Summing over i and using the inequality

p/q
ai
p/q
ai ,
where ai 0, we arrive at (1.4.46).

The second assertion follows by insertion of the function u , dened at the
end of the proof of the Theorem, into (1.4.46).
The next assertion follows immediately from Corollary 1.
Corollary 2. Suppose there exists an extension operator which maps
Vpl () continuously into Vpl (Rn ) and Lp () into Lp (Rn ) (for instance,
is a bounded domain of the class C 0,1 ). Further, let be a measure in

satisfying (1.4.45), where s is a number subject to the same inequalities as in
Corollary 1. Then for all u C l ()
1
k uLq (,)
CuV l () uL () ,
p
p
(1.4.47)
where n/p l + k < s/q, q p 1 and = (k s/q + n/p)/l.

the estimate (1.4.47) holds, then the measure with
2. If for all u C l ()
satises (1.4.45).
support in
83

Theorem 1.4.1/2 is due to D.R. Adams [2, 3]. The proof given above is borrowed from the paper by D.R. Adams [3]. The following analog of Corollary 1.4.1 was obtained by Mazya and Preobrazhenski [577] and will be proved
in Sect. 11.9.
If 1 < p < q, lp = n, then the best constant C in
uLq () CuWpl ,
is equivalent to
u C0

p1
1/q
2 p
.
log
Br (x)
r
xRn ,r(0,1)
sup
For = ms , i.e., for the s-dimensional Lebesgue measure in Rs , inequality (1.4.4) was proved by Sobolev [712] in the case s = n and by Ilin [394] in
the case s < n. They used the integral representation (1.1.10) and the multidimensional generalization of the following HardyLittlwood theorem (cf.
Hardy, Littlewood, and P
olya [351]).
If 1 < p < q < and = 1 p1 + q 1 , then the operator |x| f with
f : R1 R1 maps Lp (R1 ) continuously into Lq (R1 ).
For one particular case, Lieb [496, 497] found an explicit expression for the
norm of the operator |x| f , (0, n), acting on functions of n variables.
His result can be written as the inequality

f (x)g(y)
n n |x y| dx dy
R
R

n
1 ((n )/2)
(n/2) n
2
f L 2n gL 2n , (1.4.48)
(n /2)
(n)
2n
2n
with the equality if and only if f and g are proportional to the function
(|x x0 |2 + a2 )(2n)/2 , where a R and x0 Rn .
Theorems 1.4.2/1 and 1.4.2/2 are due to the author [543, 548]. Inequality
u(x)
n/(n1) dx
(n1)/n

Cn
u(x)
dx
(1.4.49)
was proved independently by Gagliardo [299] and Nirenberg [641] using the
same method, without discussion of the best value of Cn .
The proof based on the classical isoperimetric inequality (1.4.13) in Rn ,
which gives the sharp constant (see (1.4.14)), was proposed simultaneously
and independently by Federer and Fleming [273] and by Mazya [527].
Briey, the proof by Gagliardo [299] and Nirenberg [640] runs as follows.
One notes that
84
u(x)
21
yi (x1 , . . . , xi1 , yi , xi+1 , . . . , xn )

dyi
R
for 1 i n and for all u C0 . This yields
u(x)
n/(n1) 2n/(1n)

1/(n1)
dyi
.
R yi
1in
Integrating successively with respect to x1 , x2 , and so on, and using the generalized Holder inequality

f1 fn1 d

fj Lpj ()
1jn1
with p1 = p2 = = pn1 = n 1 after every integration, we arrive at the

inequality
1/n

dx
.
(1.4.50)
uLn/(n1) 2
Rn xi
1in
(Note that (1.4.50) is equivalent to the isoperimetric inequality

n1
cos(, xi )
ds,
mn (g)
21
1in
which can be proved by a straightforward modication of the proof of Theorem 1.4.2/1.)

By the inequality between the geometric and arithmetic means, we obtain
the estimate

u
(1.4.51)
uLn/(n1) (2n)
xi
dx.
n
R
1in
Its optimality is checked by a sequence of mollications of the characteristic

function of the cube {x : 0 xi 1}.
Obviously, (1.4.51) implies (1.4.49) with Cn = (2n1/2 )1 , but this value of
Cn is not the best possible.
It is worth mentioning that Gagliardos paper [299] contains a more general
argument based on the same idea which leads to the embedding of Wpl () to
Lq (s ), where s is an s-dimensional surface situated in .

Gromov [325] gave a proof of (1.4.49) where the integration is taken over
a normed n-dimensional space X. The value Cn = n1 in Gromovs proof
is the best possible provided the unit ball in X has volume 1. This proof is
based on the so-called increasing triangular mappings which were apparently
introduced to Convex Geometry by Knothe [437], who used them to obtain
some generalizations of the geometric BrunnMinkowski inequality. Such a
mapping transports a given probability measure on the Euclidean space to
another one, and under mild regularity assumptions, it is dened in a unique
85
way. These mappings have a simple description in terms of conditional probabilities, and were apparently known in Probability Theory before Knothes
work.
In time it became clear that triangular mappings may be used to obtain
various geometric and analytic inequalities. Bourgain [137] applied them to
prove Khinchin-type (i.e., reverse H
older) inequalities for polynomials of a
bounded degree over high-dimensional convex bodies, with constants that are
dimension free.
There is a discussion of this method in Bobkov [111, 112], where triangular
mappings were used to study geometric inequalities of dilation type.
Using wavelet decompositions, weak estimates, and interpolation, Cohen,
DeVore, Petrushev, and Xu [208] for n = 2, and Cohen, Meyer and Oru [209]
for n 2, obtained the following improvement of (1.4.49):
uL
(n1)/n
n
n1
CuL1
1/n
uB ,
(1.4.52)
1n
. One equivalent norm in
where B is the distributional Besov space B,
1n
,
B,
f sup t(n1)/2 Pt f L ,
t>0
where Pt is the heat semigroup on Rn , was used by Ledoux [485] in his direct
semigroup argument leading to (1.4.52).
Another powerful method for proving Sobolev-type inequalities is based
upon symmetrization of functions (it will be demonstrated in Sects. 2.3.5
and 2.3.8) was developed in dierent directions during the last 40 years. In
particular, it led to generalizations and renements of those inequalities for
the so-called rearrangement invariant spaces: Klimov [426, 427, 430]; Mossino
[619]; Kolyada [443, 444], Talenti [742, 743]; Klimov and Panasenko [436];
Edmunds, Kerman, and Pick [253]; Bastero, M. Milman, and Ruiz [76]; M.
Milman and Pustylnik [607]; Cianchi [197]; Kerman and Pick [418, 419]; Martin and M. Milman [518522]; Martin, M. Milman, and Pustylnik [524]; Pick
[659, 660]; Cianchi, Kerman, and Pick [205]; and Cianchi and Pick [207], et
al.
Using symmetrization methods, Martin and M. Milman [517] showed that
1
) B,
,
for < 0, and f (W11 + W
1

||
f (s) c(n, ) |f | (s) 1+|| f B1+||
,
t
where h (t) = 1t 0 h (s) ds. This gives another approach to (1.4.52) and
other inequalities of a similar nature.
Although the constant in (1.4.14) is the best possible, it can be improved
by restricting the class of admissible functions in this inequality. For example,
since for any N -gon N R2 the isoperimetric inequality

2
s(N ) (4/N ) tan(/N )m2 (N )
86
is valid (see [714]) then duplicating the proof of Theorem 1.4.2/1 we obtain
the following assertion.
Let uN be a function on R2 with compact support, whose graph is a polygon
with N sides. Then

2

|uN |2 dx
|uN | dx .
(4/N ) tan(/N )
R2
R2
Lemma 1.4.3 is a special case of a result due to D.R. Adams [2]. Theorem 1.4.3 was proved by the author [551].
Theorem 1.4.5 for = ms is the classical Sobolev theorem (see Sobolev
[712, 713]) with supplements due to Ilin [394], Gagliardo [299], Nirenberg [640], and Morrey [612]. Here we stated this theorem in the form presented by Gagliardo [299].
The continuity of functions in Wp1 () for p > 2, n = 2, was proved by
Tonelli [754].
To Remark 1.4.5 we add that if n = p(l k), l > k, p > 1, the inequality
p/(p1)

|k u(x)|
exp c
dx c0
(1.4.53)
uVpl ()
holds with positive constants c and c0 , as shown for the rst time by Yudovich
in 1961 [809]. (See also Pohozhaev [662] and Trudinger [762]. Concerning the
best value of c0 in inequalities of type (1.4.53) see Comments to Chap. 11.)
The estimate (1.4.32) is contained in the paper by Morrey [612]. Lemma 1.4.6 is the classical lemma due to Rellich [672]. Theorem 1.4.6/2 was
proved by Kondrashov [447] for p > 1 and by Gagliardo [299] for p = 1.
In connection with the estimate (1.4.35) we note that multiplicative inequalities of the form
j uLq cl uLp u1
Lr
and their modications are well known (see Ilin [393] and Ehrling [257]).
Their general form is due to Gagliardo [300] and Nirenberg [640] (see also
Solonnikov [717]). The papers by Gagliardo [300] and Nirenberg [640] contain
the following theorem.
Theorem 1. Let be a bounded domain having the cone property and let

1/
u =
|u| dx
for > 0. Then

,
j uq c l up + ur u1
r
(1.4.54)
where p 1, 1/q = j/n + (1/p 1/n) + (1 )/r for all [j/l, 1] unless
1 < p < and l j n/p is a nonnegative integer when (1.4.54) holds for
[j/l, 1).
1.5 More on Extension of Functions in Sobolev Spaces
87
In the paper by Nirenberg [641] the stated result is supplemented by the

following assertion.
Theorem 2. Let < 0, s = [n/], = s + n/ and let
u = sup |s u|
where
for = 0,
us = [s u]
for > 0,
[f ] = sup |x y|
f (x) f (y)
.
x
=y
Further, let 1/r = /n, > 0. Then (1.4.54) holds for j < l and for
all [(j )/(l ), 1], except the case mentioned in Theorem 1.
The proof is reduced to derivation of the inequality

(i)
q
(l)
p
u
dx c
u
dx + [u]p [u]qp
for functions of the variable x on a unit interval I .

1.5.1 Survey of Results and Examples of Domains
In Sect. 1.1.17, we introduced the class EVpl of domains in Rn for which there
exists a linear continuous extension operator E : Vpl () Vpl (Rn ). There we
noted that the class EVpl contains Lipschitz graph domains.
Vodopyanov, Goldshtein, and Latfullin [779] proved that a simply connected plane domain belongs to the class EV21 if and only if its boundary is a
quasicircle, i.e., the image of a circle under a quasiconformal mapping of the
plane onto itself. By Ahlfors theorem [30] (see also Rickman [678]) the last
condition is equivalent to the inequality
|x z| c|x y|,
c = const,
(1.5.1)
where x, y are arbitrary points of and z is an arbitrary point on that

subarc of which joins x and y and has the smaller diameter.
We give an example of a quasicircle of innite length.
Example 1. Let Q be the square {(x1 , x2 ); 0 < xi < 1, i = 1, 2}. We divide
the sides of the square Q into three equal parts and construct the squares
Qi1 , i1 = 1, . . . , 4, Qi1 Q = , on the middle segments. Proceeding in the
same manner with each Qi1 , we obtain the squares Qi1 ,i2 , i2 = 1, . . . , 4 with
edge length 32 . Repeating the procedure, we construct a sequence of squares
88
Fig. 6.
{Qi1 ,i2 ,...,ik } (k = 1, 2, . . . ; ik = 1, . . . , 4), whose union with Q is denoted by

(see Fig. 6). Clearly,
m1 () = 4
3k1 (2/3)k = .
k=1
Let x, y . It suces to consider the case x Qi1 ,...,ik and y

Qj1 ,...,jm where i1 = j1 , . . . , il = jl , il+1 = jl+1 . Then |x y| c1 3l and
any point z in (1.5.1) satises the inequality |x y| c2 3l . Thus, is a
quasicircle.
A domain in R2 which is bounded by a quasicircle belongs to the class
EVpl for all p [1, ), l = 1, 2, . . . (cf. Goldshtein and Vodopyanov [320]
for l = 1, Jones [404] for l 1). In the just-mentioned paper by Jones, a
class of n-dimensional domains in EVpl is described. It is larger than C 0,1 and
coincides with the class of quasidisks for n = 2. Goldshtein [313] showed that
the simultaneous inclusion of a plane simply connected domain and the
domain R2 \ in EVpl implies that is a quasicircle.
Is the last property true under the single condition EVp1 for some
p = 2? In other words, are quasidisks the only plane simply connected domains
contained in EVpl , p = 2? This question is discussed in the present subsection.
We give two examples which speak in favor of an armative answer. The
rst shows that cusps directed onto the exterior of a domain do not allow
us to construct an extension operator.
Example 2. Let = {(x1 , x2 ) : 0 < x1 < 1, 0 < x2 < x
1 } where > 1.
Suppose EVpl . Then Vpl () Vql1 () for 1 p < 2, q = 2p/(2 p);
for p > 2.
V2l () Vql1 () for any q < and Vql () C l1,12/p ()
89
Let u(x) = x1l . If < ( + 1)/p, then u Vpl (). Under the additional
condition that is close to ( + 1)/p the function u does not belong to
Vql1 () (p < 2, q = 2p/(2 p) or p = 2, q is a large number) and does not
(p > 2). Thus,
/ EVpl .
belong to C l1,11/p ()
The following example excludes domains with inward cusps at the boundary from EVpl , p > 1. It shows, incidentally, that the union of two domains in
EVpl is not always in the same class.
Example 3. Let be the domain considered above. We shall prove that

/ EVpl . We introduce polar coordinates (r, ) with origin x = 0 so
R 2 \
that the ray = 0 is directed along the halfaxis x1 > 0, x2 = 0. We put
u(x) = rl ()(x). Here, satises the inequality < 2/p and is close to
2/p; C0 (R2 ), (x) = 1 for r < 1 and is a smooth function on (0, 2],
() = 1 for small values of > 0 and () = 0 for [, 2]. Let v Vpl (R2 )
Since for small positive values of x1
be an extension of u Vpl (R2 \).

l1 v
1
,
x1 , x
1 cx1
l1
x1
l1 v
(x1 , 0) = 0,
x1l1
it follows that
p

x1
lv
lv
dx1
dx
dx2
xl1 x
xl1 x
(p1)
0
0
x1
2
2
1
1
p

l1
v

dx1

xl1 x1 , x1
(p1)
0
x1
1

1p(1)(p1)
c
x1
dx1 =
0
if p > 1. The latter contradicts the inclusion v EVpl (). Thus R2 \

/ EVpl
2
1
1
for p > 1. Nevertheless, we shall show that R \ EV1 . Let u V1 (R2 \).
Suppose for a moment that u = 0 for x1 > 1/2.
We put u (x) = u(x1 , x2 ) and u+ (x) = u(x1 , 2x
1 x2 ) for x . It is
clear that

u 1
+ u+ V 1 () cuV11 (R2 \)
.
V ()
1
The function v, dened in R2 by

u(x),
v(x) =
+
u (x) + x2 x
1 (u (x) u (x)),
x
/ ,
x ,
is absolutely continuous on almost all straight lines parallel to coordinate axes.

Also,

+

u u L () .
vL1 () u L1 () + u+ L1 () + cx
1
1
90
Since
+

+
u x1 , x u(x1 , 0)
u (x) u (x)

u+ (x) u x1 , x
1
1
+
u(x1 , 0) u (x)
,
we have
+

x
u u L1 ()
1

x1
+
ut (x1 , t) dt
x 1

1

+ x1
x
1
L1 ()

u x1 , x
+ x
1 u(x1 , 0) L
1
1 ()
ut (x1 , t)
dt

u+ V 1 () + u V 1 () +
1
L1 ()
1
u x1 , x
1 u(x1 , 0) dx1 .
Clearly, the last integral does not exceed cuV11 (R2 \)

. We put E0 u = v. Thus
we have
E0 uV11 (R2 ) cuV11 (R2 \)
.
In the general case we introduce a truncating function C (R1 ) equal to
unity on (, 1/3] and to zero on [1/2, +). Further, let 1 = {1/3 <
V 1 (R2 ) is dened
x1 < 1}. The required extension operator E : V11 (R2 \)
1
by

E u = E0 (u) + E1 (1 )u ,
1 ) V11 (R2 ) is a linear continuous extension operator.
where E1 : V11 (R2 \
In general, for p [1, ), l = 1, 2, . . . , Poborchi and the author proved
[576, Chap. 5] the existence of a linear bounded extension operator mapping
into the space Vpl (R2 , ) with the weighted norm
Vpl (R2 \)

l

R2 s=0
1/p
|s u| dx
p
1 p < ,
where is a function which is equal to unity outside and coincides

(1)(lp1)
with x1
on . Moreover, if there exists an extension operator:
V l (R2 ; ) and the weight is nonnegative, depends only on
Vpl (R2 \)
p
x1 on and increases then
(1)(lp1)
(x) c x1
for x and for small enough x1 .
c = const,
91
1.5.2 Domains in EVp1 which Are Not Quasidisks

The examples in Sect. 1.5.1 suggest that the class of Jordan curves that bound
domains in EVpl consists of quasicircles only. However, we shall show that this
conjecture is false.
Theorem. There exists a domain R2 with compact closure and Jordan boundary such that :
() is not a quasicircle.
() is of nite length and Lipschitz in the neighborhood of all but one
of its points.
() belongs to EVp1 for p [1, 2).
belongs to EV 1 for p > 2.
() R2 \
p
(From the aforementioned theorem by Goldshtein [313] and from the conditions (), (), and () we obtain in addition that
/ EVp1 for p 2 and
2
1
R \
/ EVp for p [1, 2].)
Before we prove this theorem we recall a well-known inequality that will
be used later.
Lemma 1. Let be a sector dened in polar coordinates by the inequalities
0 < < and 0 < r < a. Let u Wp1 (), u|r=a = 0 for p > 2. Then

u
p

uLp () .
(1.5.2)
r
|2
p|
Lp ()
This estimate is an immediate corollary of the following particular case of
Hardys inequality:
a
a
pp
|u|p r 1p dr
|u |p r dr
p
|2
p|
0
0
(cf. Sect. 1.3).
Proof of Theorem. Figure 7 presents a domain satisfying the conditions
()(). The corresponding upper and lower teeth come close so rapidly
that (1.5.1) does not hold. So is a quasicircle. The teeth almost do not
change their form and decrease in geometric progression, so () holds.
Now we verify (). Let G be the dierence of the rectangle R = {1/3 <
x1 < 1, 0 < x2 < 1/3} and the union T of the sequence of isosceles right
triangles {tk }k0 (cf. Fig. 8). The hypotenuse of tk is the segment [2k1 , 2k ].
Lemma 2. There exists a linear continuous extension operator E1 :
Vp1 (G) Vp1 (R), 1 p < 2, such that E1 u = 0 almost everywhere on the
interval x2 = 0, 0 < x1 < 1.
Proof. Since the saw {x T, x2 > 0} is a curve of the class C 0,1
there exists a linear continuous extension operator Vp1 (G) Vp1 (R). Let v be
92
Fig. 7.
Fig. 8.
an extension of u Vp1 (G). We introduce a truncating function which is

equal to unity on G and to zero almost everywhere on the interval x2 = 0,
0 < x1 < 1. Namely, we put = k on the triangle tk , where

4
2
arctan x1 2xk1
for 2k1 < x1 < 3 2k2 ,
k = 4
x2
for 3 2k2 < x1 < 2k .
arctan 2k x1
Clearly,

1
|k | c d1
k + dk+1
(1.5.3)
on tk , where dk (x) is the distance from x to the point x2 = 0, x1 = 2k .

The required extension of u is v. To prove this we need to verify the
inequality

(v)
cvVp1 (R) .
(1.5.4)
L (T )
p
We have

(v)
vLp (T ) + vLp (T ) .
Lp (T )
By (1.5.3),
vpLp (T ) c

d1 v p + + d1 v p ,
k
k+1
Lp (t )
Lp (t )
k
k0
93
where t+
k and tk are the right and left halves of tk . Since 1 p < 2, then
1

d v
c vLp (t+ ) + 2k vLp (t+ ) .
k
Lp (t+ )
k
The same estimate holds for d1

k+1 vLp (t ) . Consequently,
k
vpLp (T )
p

v

c vpLp (T ) +
.
|x|
Lp (T )
Applying Lemma 1, we obtain

v

cvVp1 (R) .
|x|
Lp (R)
Thus, (1.5.4) as well as Lemma 1 are proved.
Clearly, the domain + = {x : x2 > 0} (cf. Fig. 8) can be mapped

onto the domain G in Lemma 2 by a quasi-isometric mapping. Therefore,
any function in Vp1 (+ ) has a norm-preserving extension onto the upperhalfplane that vanishes on the halfaxis Ox1 . Applying the same reasoning to
+ , we conclude that () holds.
= \
Now we verify (). Let S = {x : 0 < x1 < 1, 0 < x2 < x1 /3}. Let k ,
k = 0, 1, . . . , denote the components of the set S\T (cf. Fig. 8), and let
denote the union k k+1 of legs of the triangle k . Further, let Vp1 (T ) be
the space of functions u Vp1 (T ) that satisfy the following condition. The
limit values of u out of the triangles tk and tk+1 coincide in their common
vertex for k = 0, 1, . . . . We equip Vp1 (T ) with the norm of Vp1 (T ).
Clearly, () follows immediately from the next lemma.
Lemma 3. There exists a linear continuous extension operator E2 :
Vp1 (T ) Vp1 (S) with p (2, ).
Proof. Consider the rectangle Q = {(x, y) : 0 < x < a, 0 < y < b} with
vertices O, A = (a, 0), B = (0, b), and C = (a, b). Let the function w Vpl ,
p > 2, be dened on the triangle OAC and let w(0) = 0.
We shall show that there exists a linear extension operator w f Vpl (Q)
such that f (0, y) = 0 for y (0, b) and
f L (Q) wL (OAC) ,
f Lp (Q) cwLp (OAC) .
Here, c is a constant which depends only on a/b and p. Clearly, we may assume
that Q is a square. We construct an even extension of w across the diagonal
OC to the triangle OBC. By we denote a smooth function of the polar angle
such that () = 1 for < /4 and (/2) = 0. Since w(0) = 0, it follows by
Lemma 1 that
r1 wLp (Q) cwLp (Q) ,
94
where r is the distance to the point O. Hence f = w is the required extension.

Using the described procedure, we can construct an extension vk of u
Vp1 (T ) to the triangle k such that vk (2k , y) = u(2k , 0) and
vk L (k ) uL(tk tk+1 ) ,
vk Lp (k ) c, uLp (tk tk+1 ) ,
where k = 1, 2, . . . . For k = 0 we obtain an extension v0 of u to the triangle
0 satisfying similar inequalities, where tk tk+1 is replaced by t0 .
We dene an extension of u to S by v = u on T , v = vk on k . Clearly,

(1.5.5)
vLp (S) + vL (S) c uLp (T ) + uL (T ) .
From the integral representation (1.1.8) it follows that
u(y)
dy .
osc u c
tk
|x y|
tk
Consequently,
osc u c2k(12/p) uLp (tk ) .

tk
Thus the right-hand side in (1.5.5) is equivalent to the norm in Vp1 (T ). The
lemma is proved.

1.5.3 Extension with Zero Boundary Data
l (G) denote the
Let G and be bounded domains in Rn , EVpl . Let V
p
G then,
completion of D(G) with respect to the norm in Vpl (G). If
l
l
n
multiplying the operator E : Vp (G) Vp (R ) by a truncating function
D(G), = 1 on , we obviously obtain a linear continuous operator E :
l (G). If G and the boundaries G, have a nonempty
Vpl () V
p
intersection, then proving the existence of E becomes a nontrivial problem.
Making no attempt at a detailed study, we shall illustrate possibilities arising
here by an example borrowed from the paper by Havin and the author [568].
In that paper, the above-formulated problem appeared in connection with
certain problems of approximation in the mean by analytic functions.
Let and G be plane domains such that EVp1 and G, and let the
origin be the only common point of intersection of the disk BR = {ei : 0
(see Fig. 9). If R is suciently small, then BR (G\)
is
< R} with G\
the union of two disjoint domains 1 and 2 . We assume that the intersection
of any circle |z| = , (0, R), with each domain j (j = 1, 2) is a single
arc. Let this arc be given by the equation z = ei with (j (), j ()),
where j and j are functions satisfying a Lipschitz condition on [0, R], and let
eij () , eij () G. Further, let j () = j ()j (), lj () = j ().
95
Fig. 9.
Theorem. The following properties are equivalent.

1 (G).
(1) The function u L1p () can be extended to a function in V
p
(2)
|u(eij () )|p

d < .
[lj ()]p1
(1.5.6)
(Here u(eij () ) is the boundary value of u at the point eij () . This
boundary value exists almost everywhere on .)
Proof. Since EVp1 , then to prove that (2) implies (1) we may assume
that u has already been extended to a function in Vp1 (B), where B is a disk
containing G.
Let satisfy a Lipschitz condition on the exterior of |z| = R, = 0 on
R2 \G, = 1 on and (ei ) = 1 ( j ())/j () for ei j , j = 1, 2.
Clearly, for (j (), j ()),
i

u e u eij ()

j ()
j ()

u ei
d
j ()

(u) ei
d
j ()
(p1)/p

j ()
j ()

(u) ei
p d
1/p
. (1.5.7)
j ()
Thus

j

R
|u(ei )|p
|u(eij () )|p
p
d d c uLp (j ) +
d .
[lj ()]p
[lj ()]p1
0
96
1 (G).
We can easily deduce that the preceding inequality implies u V
p
p1 (G) then by (1.5.7)
If u V

0
|u(eij () )|p

d upLp (j ) .
[lj ()]p1
Since lj () 2, it follows that the condition (1.5.6) with p 2 cannot
be valid for all u Vp1 () and hence the operator E does not exist. The
same holds for 1 p < 2 provided lj () = O(1+ ), > 0. In fact, the
function u Vp1 (), dened near 0 by the equality u(ei ) = 1+2/p with
0 < < (p 1)/p does not satisfy (1.5.6).
Now let 1 p < 2 and lj () c , c > 0. Using an estimate similar
to (1.5.7), we arrive at

0

p

u eij ()
p d c up
+ 1 uLp (j ) ,
L
p1
p (j )

which together with Hardys inequality (1.5.2) shows that (1.5.6) is valid for
all u Vp1 (). Consequently, for p [1, 2) and lj () c the operator E
exists.

P. Jones [404] introduced a class of so-called (, ) domains and showed that
these domains belong to EVpl for p [1, ] and l = 1, 2, . . . . For (0, ),
(0, ], Rn is an (, ) domain if any points x, y with |x y| <
can be joined by a rectiable arc such that
() |x y|/,
dist(z, ) |x z||y z|/|x y|,
where () is the length of and z an arbitrary point. It should be

noted that Jones bounded extension operator Vpl () Vpl (Rn ) depends on
l (while that of Stein does not). Any domain in C 0,1 with compact closure
is an (, ) domain for some , and for n = 2 the class of simply connected
(, ) domains coincides with the class of quasidisks [404]. It is interesting
to observe that multidimensional domains with isolated inward cusps satisfy
Jones theorem and hence lie in EVpl for all p 1 and l = 1, 2, . . . (cf.
Example 1.5.1/3). Fain [267] and Shvartsman [698] extended Jones theorem
to anisotropic Sobolev spaces, and Chua [188] extended the result of Jones
to weighted Sobolev spaces. We also mention here the paper by Rogers [680],
where a bounded extension operator Vpl () Vpl (Rn ), independent of l, p
was constructed for an (, ) domain.
Romanov [681] showed that the role of the critical value 2 in Theorem 1.5.2
is related to the particular domain dealt with in this theorem. He constructed
/ EVp1
a planar fractal domain such that EVp1 for p [1, q) and
97
for p q. In [803], S. Yang constructed an example of a homeomorphism of

Rn such that the image of Rn+ is in EVp1 for all p > 1, but does not satisfy P.
Jones condition [404]. In particular, it is not a quasidisk if n = 2.
P. Shvartsman [700] described a class of EWpl domains Rn whenever
p > n. Suppose that there exist constants C and such that the following
condition is satised: for every x, y such that |x y| , there exists a
rectiable curve joining x to y such that

pn
1n
dist(z, ) p1 ds(z) C|x y| p1 .
(1.5.8)
Then is an EWql domain for every l 1 and every q p.

For l = 1 and q > p this result was proved by Koskela [454].
Buckley and Koskela [148] showed that if a nitely connected bounded domain R2 is a EWp1 domain for some p > 2, then there exists a constant
C > 0 such that for every x, y there exists a rectiable curve
satisfying inequality (1.5.8) (with n = 2). Combining this result with Shvartsmans theorem [700] one obtains the following fact. Let 2 < p < and let
be a nitely connected bounded planar domain. Then is an EWp1 domain
if and only if for every x, y there exists a rectiable curve joining
x to y such that

p2
1
dist(z, ) 1p ds(z) C|x y| p1 ,
(1.5.9)
with some C > 0.

Here we mention some results by Poborchi and the author [575] on the
extension of Sobolev functions from cusp domains. A typical domain with the
vertex of an outward cusp on the boundary is

= x = (y, z) Rn : z (0, 1), |y| < (z) , n 2,
where is an increasing Lipschitz continuous function on [0, 1] such that
(0) = limz0 (z) = 0. Cusp domains are generally not in EVpl , but it is
possible to extend the elements of Vpl () to some weighted Sobolev space on
Rn .
Let be a bounded nonnegative measurable function on Rn , which is
separated away from zero on the exterior of any ball centered at the origin.
l
(Rn ) we mean the weighted Sobolev space with norm
By Vp,
uVp,
l (Rn ) =
k uLp (Rn ) .
k=0
Clearly this weighted space is wider than Vpl (Rn ). The following assertion
gives precise conditions on the weight .
98
Let have an outward cusp as above. In order that there exist a linear
l
(Rn ) it is sucient and if (x)
continuous extension operator Vpl () Vp,
depends only on |x| and is nondecreasing in the vicinity of the origin, then it
is also necessary that

c((|x|)/|x|))min{l,(n1)/p} if lp = n 1,
(x)
(|x|) (1p)/p
l
c( (|x|)
if lp = n 1,
|x| ) | log( |x| )|
in a neighborhood of the origin, where c is a positive constant independent
of x.
Let p (1, ) and l = 1, 2, . . . . We now give sharp conditions on the
exponent q [1, p) such that there is a linear continuous extension operator
Vpl () Vql (Rn ) for the same . This extension operator exists if and only
if
1 n/(1)
dt
t
< ,
(t)
t
0
where

1/q 1/p = l( 1)/ (n 1) + 1 for lq < n 1
and
1/q 1/p = (n 1)( 1)/np
for lq > n 1.
In the case lq = n 1 the factor | log((t)/t)| should be included into the

integrand above with

= (1 1/q)/(1/p 1/q),
= (np q)/ q(n 1) .
Example. Power cusp. Let (z) = c z , > 1. A linear continuous extension operator Vpl () Vql (Rn ) for q p1 + l( 1)/ 1 + (n 1) .
2
lq = n 1 and either the same inequality as in 1 holds or

q 1 = p1 + l( 1)/ 1 + (n 1) with 2q 1 < 1 + p1 .
lq > n 1 and q 1 > p1 (1 + ( 1)(n 1)/n).

Various results on extensions of functions in Wpl (), with deterioration of
the class can be found in Burenkovs survey [156]. Finally we note that properties of extension domains for functions with bounded variation are discussed
in Chap. 9 of the present book.
1.6 Inequalities for Functions
99
1.6 Inequalities for Functions with Zero Incomplete

Cauchy Data
Consider the following question. Are any restrictions on necessary for
Sobolevs inequalities to be valid for functions in Wpl () that vanish at
along with their derivatives up to the order k1 for some k l? To be precise,
k ().
we mean functions in the space Vpl () W
p
Obviously, no requirements on are needed for k = l. In the following,
we show that the same is true if k satises the inequality 2k l. This result
cannot be rened, for in the case 2k < l some conditions must be imposed.
The validity of Sobolevs inequalities for l 2k follows from the integral
representation for dierentiable functions in an arbitrary bounded domain,
which is derived in Sect. 1.6.2. The necessity of requirements on in the
case l > 2k is shown by an example (cf. Sect. 1.6.4).
1.6.1 Integral Representation for Functions of One Independent
Variable
k (a, b).
Lemma. Let k and l be integers 1 < l 2k and let z W1l (a, b) V
1
Then
b
K (t, )z (l) ( ) d,
(1.6.1)
z (l1) (t) =
a
where
K (t, ) =
ab
2[l/2]1 ( 2ba
)
2[l/2]1 ( a+b2
)
ba
for t > ,
for t < ,
and 2i1 is a polynomial of degree 2i 1, dened by

2i1 (s) + 2i1 (s) = 1
and the boundary conditions
(i1)
2i1 (1) = = 2i1 (1) = 0.

Proof. Let l be even, l = 2q. Consider the boundary value problem
y (2q) (x) = f (x),
y (j) (1) = 0,
j = 0, 1, . . . , q 1,
(1.6.2)
on the interval [1, 1]. Let g(x, s) denote the Green function of this problem.
Clearly, g(x, s) = g(x, s) and g(x, s) is a polynomial of degree 2q 1 in x
and s for x > s and x < s. First, let x > s. The derivative 2q1 g(x, s)/x2q1
does not depend on x; it is a polynomial of degree 2q 1 in s and satises the
boundary conditions (1.6.2) at the point s = 1. We denote this polynomial
by 2q1 (s).
100
In the case x > s, i.e., x < s, we have

2q1
2q1
g(x,
s)
=
g(x, s) = 2q1 (s).
x2q1
x2q1
Hence, dierentiating the equality
1
y(x) =
g(x, s)y (2q) (s) ds,
1
we obtain
y (2q1) (x) =
1
2q1 (s)y (2q) (s) ds
2q1 (s)y (2q) (s) ds.

x
Passing to the variables t, via

x=
2t a b
,
ba
s=
2 a b
,
ba
and setting z(t) = y(x), z( ) = y(s), we arrive at

t
2 a b (2q)
(2q1)
z
z
(t) =
2q1
( ) d
ba
a

b
a + b 2 (2q)
z
2q1
( ) d.
ba
t
(1.6.3)
The lemma is proved for l = 2q. In the case l = 2q +1 we express (z )(2q1)

in terms of (z )(2q) by the formula just derived. This concludes the proof of
the lemma.

Remark. For i = 1, 2, 3, the polynomials 2i1 are
1
1
(s + 1),
3 (s) = (2 s)(s + 1)2 ,
2
4
1
2
(3s 9s + 8)(s + 1)3 .
5 (s) =
16
1 (s) =
1.6.2 Integral Representation for Functions of Several Variables

with Zero Incomplete Cauchy Data
The basic result of this section is contained in the following theorem.
k ().
Theorem. Let l be a positive integer, l 2k, and let u Ll1 () V
1
Then for almost all x

dy
K, (x, y)D u(y)
.
(1.6.4)
D u(x) =
|x
y|n1
{:||=l}
101
Here, is an arbitrary multi-index of order l 1 and K, is a measurable

function on such that |K, (x, y)| c, where c is a constant depending
only on n, l, k.
Proof. First suppose that u is innitely dierentiable on an open set ,
. Let L be an arbitrary ray, drawn from the point x; let be a unit

vector with origin at x and directed along L, y = x + , R1 . Further, let
(x, ) be the rst point of intersection of L with . We also introduce the
notation
a(x, ) = b(x, ).
b(x, ) =
(x, ) x
,
k () and l u L1 (), the function
Since u V
1

a(x, ), b(x, ) z( ) = u(x + ),
satises the conditions of Lemma 1.6.1 for almost all in the (n 1)dimensional unit sphere S n1 . So, from (1.6.3) it follows that
z

2 a(x, ) b(x, ) (l)
(0) =
2[l/2]1
z ( ) d
b(x, ) a(x, )
a(x,)

b(x,)
a(x, ) + b(x, ) 2 (l)
z ( ) d. (1.6.5)
2[l/2]1
b(x, ) a(x, )
0
(l1)
We note that
z (l1) (0) =
{:||=l1}
(l 1)!
D u(x).
!
Let be any multi-index of order l 1 and let {P ()} be the system of all
homogeneous polynomials of degree l 1 in the variables 1 , . . . , n such that

P () ds = .
S n1
Before proceeding to further transformations we note that the function

(x, ) |(x, )x| can be considered as the limit of a sequence of measurable
functions on S n1 if is approximated by an increasing nested sequence
of polyhedrons. Hence, a and b are measurable functions.
Let r = |yx|, i.e., = r, if > 0 and = r, if < 0. We multiply (1.6.5)
by P () and integrate over S n1 . Then
(l 1)!
D u(x)
!

=
P () ds
S n1
a(x,)
P () ds
S n1
2[l/2]1
0

0
b(x,)
a(x, ) + b(x, ) 2r
b(x, ) a(x, )
(1)l 2[l/2]1
2r + a(x, ) + b(x, )
b(x, ) a(x, )
l u(r, )
dr
rl
l u(r, )
dr.
r l
102
Replacing by in the second term and noting that a(x, ) = b(x, )

and P () = (1)l1 P (), we nd that the rst and the second terms are
equal, i.e., D u(x) is equal to

2!
P () ds
(l 1)! S n1

b(x,)
a(x, ) + b(x, ) 2r l u(r, )
2[l/2]1
dr
b(x, ) a(x, )
rl
0

a(x, ) + b(x, ) 2r
= 2l!
P ()2[l/2]1
b(x, ) a(x, )
(x)

{:||=l}
dy

D u(y) n1 ,
!
r
where (x) = {y : S n1 , r < b(x, )}. Let K, (x, ) denote the

function

b(x, ) b(x, ) 2r
2 l!
y
P ()2[l/2]1
!
b(x, ) + b(x, )
extended to \(x) by zero. Then (1.6.4) follows for x .
Now we remove the assumption that u is smooth on . Let u satisfy the
condition of the theorem. Clearly, u can be approximated in the seminorm
l uL() by the functions that are smooth on
and which coincide with
u near . This and the continuity of the integral operator with the kernel
|x y|1n K, (x, y), mapping L() into L(), imply (1.6.4) for almost all
x . Since is arbitrary, the theorem is proved.

1.6.3 Embedding Theorems for Functions with Zero Incomplete
Cauchy Data
Now we proceed to applications of Theorem 1.6.2.
Theorem 1. Let m, l, k be integers, 0 m < l 2k. Let p 1 and let
k () for n < p(l m). Then u C m (), m u L () and
u Wpl () V
1
m uL () cdlmn/p l uLp () ,
(1.6.6)
where d is the diameter of .

k () into V m () is compact.
The embedding operator of Wpl () V
103
Proof. It is sucient to assume d = 1. Iterating (1.6.4), we obtain the

following representation for D u, || = m < l,

Q (x, y)D u(y) dy,
(1.6.7)
D u(x) =
{:||=l}
where Q = O(rlmn +1), if n = lm and Q = O(log 2r 1 ), if n = lm.

Applying Holders inequality to (1.6.7), we obtain (1.6.6). Since any function in Vpl () can be approximated in the Vpl () norm by functions that
coincide with u near and are smooth in
, we see that m u is continuous in . Here, as in the proof of Theorem 1.6.2, is an arbitrary open set,
. Thus, (1.6.6) is derived.

We can construct a covering {B (i) } of Rn by balls with diameter , with
a multiplicity not exceeding some constant which depends only on n. Let
i = B (i) and let {i } be a smooth partition of unity subordinate to
{B (i) } such that j i = O( j ). By (1.6.6),

m uL () c maxm (i u)L ( )
c
i
lmn/p
l uLp () + c()uVplp () .
It remains to note that, by virtue of (1.6.7), any bounded subspace of Vpl ()

1k () is compact in Vpl1 (). The theorem is proved.

V
Further applications of the integral representation (1.6.4) are connected
with the results of Sect. 1.4.
Let m and l be integers, 0 m < l, p > 1 and let be a measure in
such that

B (x) Ks , K = const, 0 < s n,
(1.6.8)
for any ball B (x).
Let be an open set,
, n > p(l m) > n s and q =
ps(n p(l m))1 . Further, let Lq (, ) be the space of
! functions on which
are integrable with respect to and Lq (, , loc) = Lq (, ). By Theorem 1.4.5 there exists a unique linear mapping : Vplm () Lq (, , loc)
such that:
, then v = v on
;
(i) if v Vplm () and v is smooth on
(ii) the operator : Vplm () Lq (, ) is continuous for an arbitrary
set .
Theorem 2. Let m, l, k be integers, 0 m < l 2k, p > 1, s >
n p(l m) > 0 and let be a measure in , satisfying (1.6.8). Then, for
k (),
any u Llp () V
1

(m u)
c K 1/q l uLp () ,
(1.6.9)
Lq (,)
where q = ps(n p(l m))1 .
104
The proof follows immediately from the integral representation (1.6.7) and
Theorem 1.4.5.
Theorem 3. Let m, l, k, p, q, and s be the same as in Theorem 2 and let
be a measure in satisfying the condition

(1.6.10)
lim sup s B (x) = 0.
0 xRn
k () Lq (, ) is compact.
Then the operator m : Wpl () V
1
Proof. Given any > 0 there exists a number > 0 such that

s sup B (x) <
x
for . We shall use the notation i , i , introduced in the proof of Theorem 1. Clearly,

m (i u)
q d.
|m u| d c
This inequality and Theorem 2 imply

m uLq (,) cl uLp () + C()uVpl1 () .
k () is compact in
It remains to use the fact that a unit ball in Wpl () V
1
l1
Vp ().

Remark 1. From Corollary 11.8 and Theorem 11.9.1/4 it follows that Theorems 2 and 3 remain valid for n = p(lm) if the conditions (1.6.8) and (1.6.10)
are replaced by

1
B (x) K| log |q(1p)/p , 0 < < ,
2

lim sup | log |q(p1)/p B (x) = 0,
0 xRn
where q > p > 1.

Remark 2. Sometimes assertions similar to Theorems 1, 2, and 3 can be
rened via the replacement of l uLp () by ()l/2 uLp () , where is
the Laplace operator. Namely, for any bounded domain and any function
21 () such that u Lp (), 2p > n, we have
uV
uL () c(diam )2n/p uLp () .
This inequality results from an obvious estimate for the Green function of the
Dirichlet problem for the Laplace operator, which in turn follows from the
maximum principle.
105
Analogous estimates can be derived from pointwise estimates for the Green
function Gm (x, s) of the Dirichlet problem for the m-harmonic operator in an
n-dimensional domain (see Mazya [550, 558, 559]). For instance, for n =
5, 6, 7, m = 2 or n = 2m + 1, 2m + 2, m > 2, we have
Gm (x, s)
c|x s|2mn , c = c(m, n).
This along with Theorem 1.4.1/2 implies

uLq (,) CK 1/q m uLp () ,
2m (); n > 2mp, p > 1 and is a measure in satisfying (1.6.8).
where u V
1.6.4 Necessity of the Condition l 2k
Here we show that the condition l 2k cannot be weakened in the theorems
of the preceding section. We present an example of a domain Rn for
k (), l < 2k, is not embedded into L () for pl > n > 2pk
which Vpl () V
p
and is not embedded into Lpn/(npl) () for n > pl.
Consider the function

k
v (x) = 1 2 |x|2 ,
in the ball B (0). Since v vanishes on B (0) along with its derivatives up
to order k 1, then |m v | c k km for k m in an neighborhood of
B (0). It is also clear that |m v | = O( m ) in B (0) and that |m v | = 0
for m 2k + 1.
We denote by P and Q the lower and upper points at which the axis 0xn
intersects B (0), and construct the balls B (P ), B (Q), < /2. Let be a
smooth function on Rn vanishing on B1/2 (0) and equal to unity on Rn \B1 (0).
On B (0) we introduce the function

w(x) = v (x) 1 (x P ) 1 (x Q) ,
and estimate its derivatives. On the exterior of the balls B (P ), B (Q) we
have

|j w| = O j
|j w| = 0 for j > 2k,
for j 2k.
Also,

min{j,2k}
|j w| c
mj |m v |
m=0
on B (P ) B (Q). Hence
|j w| c k kj ,
on B (P ) B (Q). This implies
j = 0, 1, . . . , l,
106
Fig. 10.
j wpLp (B (0)) c pk p(kj)+n
for j > 2k.
Similarly, since n > pk, we obtain that

j wpLp (B (0)) c npj + pk p(kj)+n c1 npj
Therefore
for j 2k.

wpV l (B (0)) c n2pk + pk np(lk) .
p
We set = , where is a number satisfying the inequalities

n pk
pk
<<
,
n p(l k)
n p(l k)
> 1.
Then
wpV l (B (0)) c ,
p
where = (n p(l k)) pk > 0.

Consider the domain (Fig. 10) which is the union of balls Bi with radii
i and centers Oi , joined by cylindrical necks Ci of arbitrary height and with
cross-section diameter 1 = i . In each ball Bi we specify wi as described
previously and extend wi by zero to \Bi . Then we put
1.7 Density of Bounded Functions
u(x) =
hi wi (x),
x ,
107
(1.6.11)
i=1
where {hi } is the sequence of numbers such that
|hi |p i < .
(1.6.12)
i=1
This condition means that u Vpl (). The partial sums of the series (1.6.11)
k (). Since wi = 1 in the center of Bi ,
pk (), and so u V
are functions in V
p
we have
uL () sup |hi |.
i
k ()
Clearly, the series (1.6.12) can converge as hi . Therefore, Vpl () V
p
is not embedded into L ().
In the case n > pl we put |hi |p = ilpn . Then
uqLq () c
|hi |q in = c
i=1
1,
i=1
with q = pn/(n lp). On the other hand,

upV l ()
p
i ,
i=1
where = ( 1)(n p(l k)) > 0. So, if {i } is a decreasing geometric

k (), whereas u
progression, then u Vpl () V
/ Lq ().
p
The restrictions on under which the Sobolev theorems hold for the space
k (), 2k < l, will be considered in Sect. 7.6.6.
Vpl () V
p
The content of the present section is borrowed from the authors paper [553].
1.7 Density of Bounded Functions in Sobolev Spaces

1.7.1 Lemma on Approximation of Functions in L1p ()
Lemma. If v L1p () then the sequence of functions

min{v(x), m}
if v(x) 0,
v (m) (x) =
max{v(x), m} if v(x) 0,
(m = 1, 2, . . . , ) converges to v in L1p ().
108
The same is true for the sequence
v(x) m
v(m) (x) = 0
v(x) + m1
if v(x) m1 ,
if |v(x)| < m1 ,
if v(x) m1 .
Proof. Since functions in L1p () are absolutely continuous on almost all

lines parallel to coordinate axes (Theorem 1.1.3/1), we have, almost everywhere in ,
v (m) = (m) v,
where (m) is the characteristic function of the set {x : |v(x)| < m1 }. Therefore

p
(v v (m) )
p dx =
|v|p 1 (m) dx.
The convergence to zero of the last integral follows from the monotone convergence theorem.

The proof for the sequence v(m) is similar.
1.7.2 Functions with Bounded Gradients Are Not Always Dense
in L1p ()
Lemma 1.7.1 says that the set of bounded functions is dense in L1p () whenever p [1, ). The following example shows that this set cannot be generally
replaced by L1 ().
Example. Let p (2, ) and {ai }i1 , {i }i1 be two sequences of positive
numbers satisfying
a1 + 1 < 1,
and
ai+1 + i+1 < ai ,

i 1,
lim ai = 0,
ai1p i < .
(1.7.1)
i1
The planar domain is the union of the square 1 = (1, 0) (0, 1), the
triangle

2 = (x, y) R2 : x (0, 1), y (0, x) ,
and the passages

(x, y) : y (ai , ai + i ), 0 x y ,
i = 1, 2, . . .
(see Fig. 11). Let u be dened on by u(x, y) = i 1 if (x, y) i , i = 1, 2,

and u(x, y) = x/y if (x, y) \ (1 2 ).
Then u L1p () because of (1.7.1) (in fact u C() L () L1p ()).
We shall show that u cannot be approximated by functions in L1 () in the
metric of the space L1p ().
109
Fig. 11.
Let v L1 () be an arbitrary function. This function coincides with a

function in C() a.e. on . Moreover, by Sobolevs embedding theorem, there
is a constant K > 0, independent of u and v, such that
vL (1 ) + v 1L (2 ) Kv uL1p () .
(1.7.2)
Using the absolute continuity of v on almost all line segments [(0, y), (y, y)]
with

(ai , ai + i ),
(1.7.3)
y
i1
we obtain
Thus
v(y, y) v(0, y)
=
y
0
v
(x, y) dx
yvL () .
x
v(y, y) v(0, y)
1/2
for suciently small y satisfying (1.7.3). Hence the left-hand side of (1.7.2) is
not less than 1/2, and the quantity u vL1p () cannot be less than (2K)1 .
1.7.3 A Planar Bounded Domain for Which L21 () L () Is Not
Dense in L21 ()
According to Lemma 1.7.1, the subspace of bounded functions is dense in
L1p () for an arbitrary domain if p [1, ). It turns out that this property
cannot be generally extended to Sobolev spaces of higher orders. In this section
we give an example of a bounded domain R2 and a function f L22 ()
such that f does not belong to the closure of L21 () Lq () in the norm of
L21 () with arbitrary q > 0. In particular, this implies that L2p () Lq () is
not dense in L2p () for p 2.
First we establish an auxiliary assertion. Below we identify functions in
L21 with their continuous representatives.
110
Lemma. Let G be a planar subdomain of the disk BR , starshaped with

respect to the disk Br . Then the following estimate holds for all f L21 (G):

f (z1 ) f (z2 )
c 2 f L (G) + |z1 z2 |r12/q f L (G) ,
(1.7.4)
1
q
and the constant c depends only on q and the ratio R/r.
where z1 , z2 G,
Proof. It will suce to consider the case r = 1 and then use a similarity
transformation. In view of Lemma 1.1.11 there is a linear function such that
f L1 (G) c 2 f L1 (G) .
we obtain two estimates
Hence by the Sobolev embedding L21 (G) C(G)
(f )(z1 ) (f )(z2 )
c2 f L (G) ,
1

Lmin{1,q} (G) c 2 f L1 (G) + f Lq (G) .
Now (1.7.4) follows from the obvious inequality
(z1 ) (z2 )
c|z1 z2 |L
.
min{1,q} (G)
This concludes the proof of the lemma.
We turn to the required example. Let {i }i0 , {hi }i0 be two sequences
of positive numbers satisfying i < 2i2 and

(1.7.5)
hi exp (1 + i)2 /i2 , i 0,
lim i 2ib = 0 for all b > 0.
(1.7.6)
Next, let {i }i0 be the sequence of open isosceles right triangles with hypotenuses of length 21i , placed on the lines y = Hi , where

(hs s ), j = 0, 1, . . . .
Hj = 21j +
sj
We assume that all vertices of right angles lie on the axis Oy under the
hypotenuses. Let i denote the intersection of i with the halfplane y
Hi+1 + hi . Clearly
the distance between i and i+1 is hi . By we mean the
complement of i0 i to the rectangle {(x, y) R2 : |x| < 1, 0 < y < H0 },
(see Figs. 12 and 13).
Let C0 (1, 1) and (t) = 1 for |t| 1/2. We now dene f on as
follows. For any strip

i = (x, y) : Hi+1 < y < Hi , i 0,
f is given on i by

i+1
(|x| i )
log |x| i + hi 2 x
f (x, y) = sign x 1 +
i log hi
111
Fig. 12.
Fig. 13.
for (x, y)
/ i , |x| > i , and f (x, y) = x/i for the remaining points of i .
Clearly i supp(2 f ) is placed in the set

(x, y) i \ i : i |x| 2i1 .
Furthermore, the following estimate holds for (x, y) i
|2 f | c
max{2i (i + 1), (|x| i + hi )1 }

i | log hi |
(here and in the following in this section c is a positive constant independent

of i). Hence
2 f 2L2 () =
2 f 2L2 (i ) c
i0
which is dominated by c

i1
| log hi | + (i + 1)2
i0
i2 due to (1.7.5).
(i log hi )2
112
We now prove by contradiction that f does not belong to the closure of

L21 () Lq () in the norm of L21 (). Let

0 = \
i .
i1
Since 0 has the cone property, Sobolevs theorem says that the space L21 (0 )
is embedded into C(0 ) L (0 ), and
uL (0 ) KuL21 () ,
K = const > 0,
(1.7.7)
for all u L21 (). Suppose that there exists a function g L21 () Lq ()
subject to
f gL21 () (2K)1 .
+
Put A
i = (i , Hi+1 +hi ), i 0. Since f (Ai ) = 1, f (Ai ) = 1, the estimate
+

g A g A
1
i
is valid in view of (1.7.7) On the other hand, an application of the above

lemma gives

2 gL ( ) + 2i+1+2i/q i gL ( ) .

g A
c
g A+
1
i
i
q
i
i
i
i
This inequality in conjunction with (1.7.6) implies that

= 0,
lim
g A+
g A
i
i
i
and we arrive at the required contradiction.

1.7.4 Density of Bounded Functions in L2p () for Paraboloids in
Rn
We have seen in the preceding section that bounded domains with a nonsmooth boundary may fail to have the property that the set L2p () L ()
is dense in L2p (). It turns out that for unbounded domains this property may
fail even when is very simple with smooth boundary.
Let f be a function in C 1 ([0, )) such that f is positive on (0, ), |f (t)|
const and f (t) as t . If f (0) = f (0) = 0, we impose an additional
condition f (t) > 0 in the vicinity of t = 0. Let

= (x, t) Rn : 0 < t < , |x| < f (t)
(see Fig. 14). We call this domain an n-dimensional paraboloid.
In this section c denotes various positive constants depending only on p,
n, and f . The following assertion gives a necessary and sucient condition
for the space L2p () L () to be dense in L2p ().
113
Fig. 14.
Theorem. The space L2p () L () is dense in L2p (), 1 n.)

The proof of this theorem will be given at the end of the section. First we
establish three auxiliary assertions.
Lemma 1. Let u L2p () L (),

T = (x, t) : t = T ,

T = (x, t) : 0 < t < T, |x| < f (t) ,
and
w(T ) = f (T )
1n
T
u
dx.
t
Then for all T > 0
w(T )
c
f (t)
1n
p1
(p1)/p
1/p
|2 u| dx dt
p
dt
Proof. We may assume that

1n
f (t) p1 dt < .
Since
w(T ) =
||<1
we have for S > T
(1.7.9)
\T
u
f (t), t
t=T d,
t
(1.7.10)
114
w(S) w(T )
S
dt
=

ut f (t), t d
t
T
||<1
2

S

u
u

=
d.
f
(t),
t
f
(t),
t
+
f
dt
(t)
x
2
t
T
||<1 t
By Holders inequality
w(T ) w(S)
(p1)/p
S
1n
p1
f (t)
dt
c
T
dt
|x|<f (t)
1/p
(2 u)(x, t)
p dx
.
Thus, the limit d = limt w(t) exists and it suces to deduce d = 0 to prove
(1.7.9). Consider the function u
on (0, ) dened by

u
(t) = f (t)1n
u(x, t) dx =
u f (t), t d.
|x|<f (t)
||<1
Putting v(, t) = u(f (t), t), we nd

f (t)
d
u
= w(t) +
dt
f (t)

||<1
( v)(, t) d.
(1.7.11)
Furthermore,

||<1
v d = (1 n)

v d +
||<1
||=1
v ds ,
where ds is the area element on the sphere || = 1. Since v is bounded, the

second term on the right in (1.7.11) tends to zero as t . So we have
(t) = d and d = 0 as long as u
L (0, ). This completes the
limt u
proof of Lemma 1.

Lemma 2. Let (1.7.8) hold and let

p

F () =
| | + f p | |p f n1 dt.
(1.7.12)
Then inf F () = 0, where the inmum is taken over the set

: C (R1 ), (t) = 1 for t 1, (t) = 0 for large positive t . (1.7.13)
Proof. First let p > n. Since f (t) ct for large t, it will suce to prove
the equality

p

| | + tp | |p tn1 dt = 0.
inf
(1.7.14)
1
115
Consider the function C (R1 ), (t) = 1 for t 1, (t) = 0 for t > 2. By

setting (t) = (t/N ), N = 1, 2, . . . , we obtain (1.7.14).
Let p n. Since
n1

f | | c | | +
f p1
,
the functional (1.7.12) is majorized by

n1
p
n1 n1
p dt
f p1
+
f p1 f p1
c
n1 ,
1
f p1
(1.7.15)
on the set (1.7.13). After passing to the new variable = (t) given by
t
1n
f () p1 d,
=
1
the integral in (1.7.15) takes the form

p

G() =
| | + | |p d,
( ) = (t).
It remains to observe that G(N ) 0, where N ( ) = ( / (N )), N =

1, 2, . . . , and is the function introduced previously. This concludes the proof
of the lemma.

The proof of the preceding lemma enables us to state the following assertion.
Corollary. If (1.7.8) holds, then there exists a minimizing sequence
{N }N 1 for functional (1.7.12) such that N C (R1 ), N (t) = 1 for
t N, N (t) = 0 for large positive t.
Lemma 3. Let u L2p (). For any > 0 there exist a linear function
and a function v L2p () such that v(x, t) = (x, t) for large t and
u vL2p () < .
Proof. Let

GN = (x, t) : N < t < N + f (N ) .
By Lemma 1.1.11, there is a linear function N satisfying

(uN )
+f (N )1 uN Lp (GN ) c f (N )2 uLp (GN ) . (1.7.16)
L (G )
p
We put

N (t) = 1 + (t N )/f (N ) ,
vN = (u N )N + N ,
where is the function introduced in the proof of Lemma 2. Clearly vN = u

for t N and vN = N for t N + f (N ). Next,
116

c2 (u vN )Lp ()

(1 N )2 uLp () +
(u N )
|N |Lp ()

+ (u N )|2 N |
.
Lp ()
Since
|N | cf (N )1 ,
2 N cf (N )2 ,
and in view of (1.7.16), it follows that

2 (u vN )
c2 uLp (\N ) .
L ()
p
The last norm tends to zero as N , and we can set v = vN for suciently
large N . This establishes Lemma 3.
Proof of Theorem. Let (1.7.10) hold. We check that L2p () L () is not
dense in L2p (). Let u L2p () L () and u t in the norm of L2p ().
By Lemma 1
dx
c(n, p, f, T )2 u Lp () = o(1) as .

(1.7.17)
T t
1 ). Since 2 \
1 C 0,1 , the spaces L2 and V 2
We have u t in L2p (2 \
p
p
coincide for this domain. Thus, by Sobolevs theorem,

u
dx mn1 T as
T t
for almost all T (1, 2). However, this contradicts (1.7.17).
Suppose (1.7.8) is valid. It will be shown that an arbitrary function u
L2p () can be approximated by functions in L2p () L (). According to
Lemma 3, it is sucient to assume that u = w + , where is a linear function
and w(x, t) = 0 for large t. Since any domain T is of class C, Theorem 1.1.6/2
with
applies, and w can be approximated in L2p () by functions in C ()
bounded supports. It remains to approximate the functions t, x1 , . . . , xn1 .
Let {N } be the sequence from the Corollary preceding Lemma 3. We set
t
N (s) ds for t > N
vN (x, t) = N +
N
and vN (x, t) = t for t N . Clearly vN L2p () L () and

2 (t vN )p
N (t)
p f (t)n1 dt = o(1) as N .
=
c
Lp ()
0
To approximate xj , 1 j n 1, we put wN (x, t) = xj N (t). Then

p

2 (xj wN )p
|N | + f p |N |p f n1 dt,
c
Lp ()
0
and a reference to the Corollary completes the proof.
1.8 The Maximal Algebra in Wpl ()
117

Lemma 1.7.1 is due to Deny and Lions [234]. In connection with the contents
of Sect. 1.7.2, we need two denitions. Let Cb () denote the set of functions
in C () with bounded gradients of all orders. A bounded domain Rn
has the interior segment property if to every x there correspond a
number r > 0 and a nonzero vector y Rn such that z + ty provided
0 t 1 and z Br (x). (Clearly domains of class C have the interior
segment property.)
If R2 is a bounded domain that is either starshaped with respect to a
point or satises the interior segment condition, then Cb () is dense in Vpl ()
for p [1, ) and l = 1, 2, . . . . This theorem is due to Smith, Stanoyevitch,
and Stegenga [707]. In particular, it gives sucient conditions for the space
Vpl () L () to be dense in Vpl (), R2 . The just-mentioned result
fails for multidimensional domains.
Example 1.7.2, borrowed from Sect. 2.2 of the book by Mazya and
Poborchi [576], is analogous to Example 7.1 in the paper by Smith, Stanoyevitch, and Stegenga [707], which was constructed to show that the space
Cb () is not always dense in Wp1 () (see also OFarrell [643]).
The contents of Sects. 1.7.3 and 1.7.4 are taken from the paper by Mazya
and Netrusov [572].
1.8 Maximal Algebra in Wpl ()

1.8.1 Main Result
Let A be a subset of a Banach function space. The set A is called an algebra
with respect to multiplication if there is a constant c > 0 such that the
inclusions u A, v A imply uv A and uv cuv.
Note that the space Wpl = Wpl (Rn ) is not an algebra when lp n, p > 1 or
l < n, p = 1. Indeed, if Wpl were an algebra, the following inequalities would
occur:
N 1/N N 1/N
u
u W l c uWpl ,
L
p
where u is an arbitrary function in Wpl and N = 1, 2, . . . . By letting N ,

we can obtain
uL cuWpl .
Clearly, the last inequality is not true for the values p, l mentioned earlier.
Thus, Wpl is generally not an algebra with respect to pointwise multiplication.
However, it is possible to describe the maximal algebra contained in Wpl .
Namely, the following assertion holds.
Theorem. The subspace Wpl L with the norm Wpl + L is the
maximal algebra contained in Wpl . In particular, the space Wpl is an algebra if
lp > n, p > 1 or if l n, p = 1.
118
We need an auxiliary multiplicative inequality to prove this theorem.

Lemma 1. Let 1 p and l 2 be an integer. For any u Wpl L
and any j = 1, . . . , l 1, the estimate
1j/l
j/l
j uLlp/j cuL uW l
(1.8.1)
holds with c independent of u.

We shall deduce Lemma 1 with the aid of another auxiliary inequality.
Lemma 2. If u C0 (Rn ), p [1, ), 1 q , and 2r1 = p1 + q 1 ,
then
1/2
1/2
uLr c2 uLp uLq ,
where c is a positive constant depending only on n.
Proof. It is sucient to assume p > 1 and q < . Then the extreme cases
p = 1 or q = follow by letting p tend to 1 or q tend to .
We remark that the required inequality is a consequence of the onedimensional estimate
r
r

2p
2q

r
r
p
q
|u | dx
|u| dx
|u | dx c0
(1.8.2)
with 2r 1 = p1 + q 1 and c0 an absolute constant (the integration is taken
over R1 ). Once this estimate has been established, the result follows by integrating with respect to the other variables and by applying H
olders inequality.
Note that for any u C0 (R1 )
2c1 u Lr (i) |i|1+r
p1
u Lp (i) + |i|p
r1 1
uLq (i) ,
(1.8.3)
where c1 is an absolute positive constant, i an interval, and |i| its length. The
last estimate follows from (1.1.18) and the H
older inequality.
Inequality (1.8.2) is a consequence of the estimate
c1 u Lr () u Lp uLq ,
1/2
1/2
(1.8.4)
where is an arbitrary interval in R1 of nite length. The last is veried as

follows.
Fix a positive integer k and introduce the closed interval i of length ||/k
with the same left point as . Let us consider inequality (1.8.3) for this
interval i. If the rst summand on the right of the inequality is greater than
the second, we put i1 = i. In this case
1+rr/p

pr

||
|u |r dx
|u |p dx .
(1.8.5)
cr1
k
i1
R1
Suppose the rst term on the right of (1.8.3) is less than the second. We then
increase the interval i leaving the left end point xed until these two terms are
119
equal (clearly, the equality must take place for some i with |i| < because
1 + r1 p1 > 0). Let i1 denote the resulting interval. Then
|u |r dx
cr1
r
2p
2qr
|u |p dx
|u|q dx
.
i1
i1
(1.8.6)
i1
Putting the end point of i1 to be the initial point of the next interval, repeat
this process with the same k. We stop it when the closed nite intervals
i1 , i2 , . . . (each of length at least ||/k) form a covering of the interval . Note
that the covering {i1 , i2 , . . . } contains at most k elements, each is supporting
estimates (1.8.5) or (1.8.6) (with i1 replaced by is ). Adding these estimates
and applying Holders inequality, one arrives at

cr1
1+rr/p
rp
||
p
|u | dx
|u | dx k
k
r

2p
2qr
+
.
|u |p dx
|u|q dx
r
Now (1.8.4) follows by letting k because 1 + r r/p > 1. The proof of

Lemma 2 is complete.

Proof of Lemma 1. It is sucient to assume p < . First let u C0 (Rn ).
If aj = j upl/j , Lemma 2 implies a2j caj1 aj+1 for j = 1, . . . , l 1. By
1j/l j/l
al , and Lemma 1 is established for

induction on l, we obtain aj ca0
smooth functions u with compact support.
Turning to the general case u Wpl L , we rst assume that supp u is
bounded and consider a mollication uh of u with radius h. Since uh
cu and by Lemma 1 applied to uh we obtain
j/l
1j/l
j uh Llp/j c uh W l uL .

p
Now passage to the limit as h 0 gives the required inequality for u with
bounded support.
To conclude the proof of the lemma, we remove the assumption on the
boundedness of supp u. To this end we introduce a cuto function C0 (B2 )
such that 0 1 and |B1 = 1. Let k (x) = (x/k), k = 1, 2, . . . . An
application of Lemma 1 to the function uk yields
j/l
1j/l
j uLlp/j (Bk ) cuk W l uL .

p
It remains to pass to the limit as k to establish Lemma 1 in the general

case.

We now give a proof of the theorem stated previously.
Proof of Theorem. Let A be an algebra contained in Wpl .
120
As we have already seen at the beginning of the section A L Wpl .

To show that the space A = L Wpl is an algebra, we let u, v A be
arbitrary. Then

l (uv)
Lp
l

|k u| |lk v|
Lp
k=0
k uLlp/k lk vLlp/(lk)
k=0
by H
olders inequality. It follows from Lemma 1 that

l (uv)
Lp
(lk)/l
uL
k/l
k/l
(lk)/l
l uLp vL l vLp
k=0
and hence

l (uv) c uL l vLp + vL l uLp ,
Lp
(1.8.7)
which does not exceed cuA vA . Thus, A is an algebra and hence the space
Wpl L is the maximal algebra contained in Wpl .
If lp > n, p > 1 or l n, p = 1, then Wpl L by Sobolevs embedding
theorem, and the space Wpl is an algebra for these p and l. This concludes the
proof.

Let be a domain in Rn . We may ask whether the space Wpl () L ()
is an algebra with respect to pointwise multiplication. Clearly, for l = 1 the
answer is armative. Since Steins extension operator from a domain C 0,1
is continuous as an operator

Wpl () L () Wpl Rn L Rn ,
the above question has the armative answer for nite sums of domains in
C 0,1 . For example, can be a bounded domain having the cone property.
However, it turns out that the space Wpl () L () is generally not an
algebra.
1.8.2 The Space W22 () L () Is Not Always a Banach Algebra
Here we give an example of a bounded planar domain such that W22 ()
L () is not an algebra.
Let be the union of the rectangle P = {(x, y) : x (0, 2), y (0, 1)},
the squares Pk with edge length 2k and the passages Sk of height 2k and
of width 2k , k = 1, 2, . . . , > 1 (see Figs. 15 and 16). Dene u = 0 on P ,
Fig. 15.
121
Fig. 16.
u(x, y) = 23k/2 (y 1)2 on Sk , k = 1, 2, . . . , and u(x, y) = 2k/2 (2(y 1) 2k )

on Pk for k 1. Straightforward calculations show that
2
2 u
= 8.
2 u2L2 (Sk ) = 22+(2)k , |u| 3,
L2 (P )
k
Thus, if > 2, then u W22 () L (), but u2

/ W22 ().
A general form of the inequality obtained in Lemma 1.8.1/1 is due to Gagliardo
[300] and Nirenberg [640]. The proof of Lemma 1.8.1/2 follows the paper
by Nirenberg [640] where it was also shown that L Wpl is an algebra.
A counterexample in Sect. 1.8.2 is taken from the paper by Mazya and
Netrusov [572].
2
Inequalities for Functions Vanishing
at the Boundary
The present chapter deals with the necessary and sucient conditions for the
validity of certain estimates for the norm uLq (,) , where u D() and
is a measure in . Here we consider inequalities with the integral

p
(x, u) dx,
on the right-hand side. The function (x, ), dened for x and Rn , is

positive homogeneous of degree one in . The conditions are stated in terms
of isoperimetric (for p = 1 in Sect. 2.1) and isocapacitary (for p 1, in
Sects. 2.22.4) inequalities. For example, we give a complete answer to the
question of validity of the inequality

uLq (,) C
(x, u)
p
1/p
dx
,
both for q p 1 and 0 < q < p, p 1. In particular, in the rst case there
hold sharp inequalities for the best constant C
1/p C p(p 1)(1p)/p 1/p ,
where
= sup
(F )p/q
,
(p, )-cap(F, )
with the so-called (p, )-capacity of a compact subset of in the denominator. Actually, this is a special case of a more general assertion concerning
BirnbaumOrlicz spaces.
Among other denitive results we obtain criteria for the validity of multiplicative inequalities of the form

1
uLp (,) C (, u)Lp () uL
r (,)

DOI 10.1007/978-3-642-15564-2 2,
123
124
2 Inequalities for Functions Vanishing at the Boundary
as well as the necessary and sucient conditions for compactness of related

embedding operators.
In Sect. 2.5 we give applications of the results in Sect. 2.4 to the spectral theory of the multidimensional Schrodinger operator with a nonpositive
potential. Here the necessary and sucient conditions ensuring the positivity and semiboundedness of this operator, discreteness, and niteness of its
negative spectrum are obtained.
Certain properties of quadratic forms of the type

u u
aij (x)
dx
x
n
i xj
R
are studied in Sects. 2.6.1 and 2.6.2. Finally, Sects. 2.7 and 2.8 are devoted to
sharp constants in some multidimensional inequalities of the Hardy type.

(the Case p = 1)
2.1.1 Criterion Formulated in Terms of Arbitrary Admissible Sets
A bounded open set g Rn will be called admissible if g and g is a C
manifold. In Chaps. 57 this denition will be replaced by a broader one.
Let N (x) denote the unit normal to the boundary of the admissible set
g at a point x that is directed toward the interior of g. Let (x, ) be a
continuous function on Rn that is nonnegative and positive homogeneous
of the rst degree with respect to . We introduce the weighted area of g

x, N (x) ds(x).
(2.1.1)
(g) =
g
Let and be measures in and n = s(B1 ).

The following theorem contains a necessary and sucient condition for
the validity of the multiplicative inequality:

uLq (,) C (, u)L1 () u1
Lr (,)
(2.1.2)
for all u D(). This result will be proved using the same arguments as in
Theorem 1.4.2/1.
Theorem. 1. If for all admissible sets
(g)1/q (g) (g)(1)/r ,
(2.1.3)
where = const > 0, [0, 1], r, q > 0, + (1 )r1 q 1 , then (2.1.2)

holds for all u D() with C r (r + 1 )(1)/r .
2.1 Conditions for Validity of Integral Inequalities (the Case p = 1)
125
2. If (2.1.2) holds for all u D() with q > 0, [0, 1], then (2.1.3)
holds for all admissible sets g and C.
Proof. 1. First we note that by Theorem 1.2.4

u
|u| dx
(x, u) dx =
x,
|u|
{x:|u|>0}

u
dt
x,
(Lt ) dt.
=
ds =
|u|
0
0
Et
(2.1.4)
Here we used the fact that |u| =

0 on Et = {x : |u(x)| = t} for almost
all t and that for such t the sets Lt = {x : |u(x)| > t} are bounded by C
manifolds. By Lemma 1.2.3

1/q
uLq (,) =
(Lt ) d tq
.
0
Since (Lt ) is a nonincreasing function, then, applying (1.3.41), we obtain

uLq (,)

1/
(Lt )/q d t
,
0
1
where = r(r+1) , q. Using the fact that the sets Lt are admissible
for almost all t, from (2.1.3) we obtain

uLq (,) 1/
1/
(Lt ) (Lt )(1)/r t1 dt
.
Since + (1 )/r = 1, then by Holders inequality

uLq (,) 1/
(Lt ) dt
0
(1)/r
(Lt )tr1 dt
which by virtue of (2.1.4) and Lemma 1.2.3 is equivalent to (2.1.2).

2. Let g be any admissible subset of and let d(x) = dist(x, Rn \g),
gt = {x , d(x) > t}. Let denote a nondecreasing function, innitely
dierentiable on [0, ), equal to unity for d 2 and to zero for d , where
is a suciently small positive number. Then we substitute u (x) = [d(x)]
into (2.1.2).
By Theorem 1.2.4,

2

(x, u ) dx =
(t)
x, N (x) ds(x),
0
gt
where N (x) is the normal at x gt directed toward the interior of gt . Since

t0
x, N (x) ds(x)(g),
gt
126
we obtain
(x, u ) dx (g).
Let K be a compactum in g such that dist(K, g) > 2. Then u (x) = 1

on K and
u Lq (,) (K)1/q .
Using 0 u (x) 1 and supp u g, we see that
u Lr (,) (g)1/r .
Now from (2.1.2) we obtain
(g)1/q = sup (K)1/q C(g) (g)(1)/r .
Kg
The result follows.

2.1.2 Criterion Formulated in Terms of Balls for = Rn
In the case (x, ) = ||, = Rn , = mn it follows from (2.1.2) that for all
balls B (x)

1/q
B (x)
A(n1)+(1)n/r .
(2.1.5)
With minor modication in the proof of Theorem 1.4.2/2 we arrive at the
converse assertion.
Theorem. If (2.1.5) holds with [0, 1]; q, r > 0, + (1 )/r 1/q
for all balls B (x), then

(1)
uLq () C (, u)L1 uLr
(2.1.6)
holds for all u D(Rn ) with C cA.

Proof. As already shown in the proof of Theorem 1.2.1/2, for any bounded
open set g with a smooth boundary there exists a sequence {Bi (xi )}i1 of
disjoint balls with the properties
B3i (xi ),
() g
i1
()
()

2mn q Bi (xi ) = vn ni ,

n1
.
s(g) c
i
i1
From (2.1.5) it follows that

(g)
B3i (xi ) Aq
(3i )q[(n1)+(1)n/r] .
i1
i1
(2.1.7)
127
Since q + (1 )n/r 1, it follows from (2.1.7) that

(g) cAq
q (n1)+(1)n/r
q+(1)n/r
q+(1)n/r
i
i1
which by H
olders inequality does not exceed

q (1)q/r
n1
ni
.
cAq
i
i1
i1
To conclude the proof it remains to apply Theorem 2.1.1.

2.1.3 Inequality Involving the Norms in Lq (, ) and Lr (, )
(Case p = 1)
The next theorem is proved analogously to Theorem 2.1.1.
Theorem. 1. If for all admissible sets g
(g)1/q (g) + (g)1/r ,
where 0, 0, q 1 r, then

uLq (,) (x, u)L() + uLr (,)
(2.1.8)
(2.1.9)
holds for all u D().

2. If (2.1.9) holds for all u D(), then (2.1.8) holds for all admissible
sets g.
2.1.4 Case q (0, 1)
Here we deal with the inequality

uLq (,) C (, u)L1 ()
(2.1.10)
for u C0 (). As a particular case of (2.1.9), we obtain from Theorem 2.1.3

that (2.1.10) holds with q 1 if and only if for all admissible sets g
(g)1/q (g)
(2.1.11)
and is the best value of C.

We shall show that (2.1.10) can be completely characterized also for q
(0, 1). Let us start with the basic properties of the so-called nonincreasing
rearrangement of a function.
Let u be a function in measurable with respect to the measure . We
associate with u its nonincreasing rearrangement u on (0, ), which is introduced by
128

u (t) = inf s > 0 : (Ls ) t ,
(2.1.12)
where Ls = {x : |u(x)| > s}.

Clearly u is nonnegative and nonincreasing on (0, ). We also have
u (t) = 0 for t (). Furthermore, it follows from the denition of u
that

(2.1.13)
u (Ls ) s
and
(Lu (t) ) t,
(2.1.14)
the last because the function s (Ls ) is continuous from the right.
The nonincreasing rearrangement of a function has the following important
property.
Lemma 1. If q (0, ), then

u(x)q d =
u (t)
dt.
Proof. The required equality is a consequence of the formula

u(x)q d =
(Lt ) d tq
0
and the identity

in which
Ls

m1 Ls = (Ls ),
= {t > 0 :
u (t)
s (0, )
(2.1.15)
> s}. To check (2.1.15) we rst note that

m1 Ls = sup t > 0 : u (t) > s
(2.1.16)
by the monotonicity of u . Hence, (2.1.13) yields

m1 Ls (Ls ).
For the inverse inequality, let > 0 and t = m1 (Ls ) + . Then (2.1.16)
implies u (t) s and therefore

m1 Ls (Lu (t) ) t
by (2.1.14). Thus (Ls ) m1 (Ls ) and (2.1.15) follows.
Denition. Let C () denote the inmum (g) for all admissible sets
such that (g) , where (g) is the weighted area dened by (2.1.1).
Theorem. Let be a domain in Rn and 0 < q < 1.
(i) (Suciency) If

()

D :=
0
s1/q
C (s)
q
1q
ds
< ,
s
129
(2.1.17)
then (2.1.10) holds for all u C (). The constant C satises the inequality
C c1 (q)D(1q)/q .
(ii) (Necessity) If there is a constant C > 0 such that (2.1.10) holds for all
u C (), then (2.1.17) holds and C c2 (q)D (1q)/q .
Proof. (Suciency) Note that (2.1.17) implies () < and that C is a
positive function. By monotonicity of (Lt ), one obtains

|u|q d =

j=
2j+1

(Lt ) d tq
2j

j 2q(j+1) 2qj ,
j=
where j = (L2j ). We claim that the estimate

m

q

j 2q(j+1) 2qj cD 1q (, u)Lq ()
(2.1.18)
j=r
is true for any integers r, m, and r < m. Once (2.1.18) has been proved,
(2.1.17) follows by letting m and r in (2.1.18). Clearly, the sum
on the left in (2.1.18) is not greater than
m 2q(m+1) +
(j1 j )2jq .
(2.1.19)
j=1+r
Let Sr,m denote the sum over 1 + r j m. Holders inequality implies

Sr,m
q
2 C (j1 )
j
j=1+r
We have
m

(j1 j )1/(1q)
C (j1 )1/(1q)
j=1+r
1/(1q)
(j1 j )1/(1q) j1
1/(1q)
1q
.
(2.1.20)
Hence, by the monotonicity of C , the sum in curly braces is dominated by

m

j=1+r
j1

C (t)q/(q1) d t1/(1q) ,
which does not exceed D/(1 q). By (2.1.4) the sum in square brackets in
(2.1.20) is not greater than
130

j=
Thus
L2j1 \L2j
(x, u) dx.
q

(j1 j )2qj cD1q (, u)L
1 ()
j=1+r
To conclude the proof of (2.1.18), we show that the rst term in (2.1.19) is
also dominated by the right part of (2.1.18). Indeed, if m > 0, then

q
q/(1q) 1q
m /C (m )
m 2mq 2m C (m )
m
m
q/(1q) 1q
q

t
c(, u)L1 ()
dt
.
C
(t)
0
The suciency of (2.1.17) follows.
We turn to the necessity of (2.1.17). We shall use the following two auxiliary assertions.
Lemma 2. Let u C00,1 (). There exists a sequence {u }1 of functions
u D() such that

x, u (x) u(x) dx 0 as .
(2.1.21)
Proof. Let u = M 1 u, where M stands for a mollication with radius

.
. Let U be a neighborhood of supp u, U
Clearly, supp u is situated in U for all suciently large . Since
(0,1)
C( S n1 ) and u C0 (), it follows that

(u (x) u(x))
x, u (x) u(x) = x,
u (x) u(x) ,
|(u (x) u(x))|
if u (x) = u(x). Therefore, the left-hand side in (2.1.21) does not exceed

max u (x) u(x) dx 0 as .
S n1
U
The proof is complete.
Lemma 3. Let {v1 , . . . , vN } be a nite collection of functions in the space

C() L1p (), p [1, ). Then, for x , the function

x v(x) = max v1 (x), . . . , vN (x)
belongs to the same space and

N

(, v)
(, vi )
.
L1 ()
L1 ()
131
(2.1.22)
i=1
Proof. An induction argument reduces consideration to the case N = 2.

Here

v(x) = max v1 (x), v2 (x) .
The left-hand side in (2.1.22) is equal to

(x, v1 ) dx +
v1 v2
(x, v2 ) dx,
v1 <v2
which implies (2.1.22) for N = 2.
Continuation of the proof of Theorem. (Necessity) First we remark that

the claim implies () < and that C (t) > 0 for all t (0, ()]. Let j be
any integer satisfying 2j (). Then there exists a subset gj of such that

(gj ) 2j , and ( gj ) 2C 2j .
By the denition of C and by (2.1.4) there is a function uj C () subject
to uj 1 on gj , uj = 0 on and

(x, uj ) dx 4C 2j .
Let s be the integer for which 2s < () < 2s+1 . For any integer r < s, we
introduce the Lipschitz function
fr,s (x) = max j uj (x),
rjs
where
x ,

1/(1q)
j = 2j /C 2j
.
By Lemmas 2 and 3

(, fr,s )
L
1 ()

j (, uj )L
1 ()
j=r
and one obtains the following upper bound:

(, fr,s )
L1 ()

j C 2j .
(2.1.23)
j=r
We now derive a lower bound for the norm of fr,s in Lq (, ). Since

fr,s (x) j for x gj , r j s, and (gj ) 2j , the inequality
132

x : fr,s (x) > r < 2j
implies r j . Hence

for t 0, 2j , r j s,
fr,s
(t) j
where fr,s
is the nonincreasing rearrangement of fr,s . Then
(t)
fr,s
()
dt
s

j=r
2j
2j1
which implies
fr,s qLq (,)
fr,s
dt
jq 2j1 ,
j=r
jq 2j1 .
(2.1.24)
j=r
Next we note that by Lemma 2 if inequality (2.1.10) holds for all u C0 (),
then it holds for all Lipschitz u with compact supports in . In particular,

fr,s Lq (,) C (, fr,s )L () .
1
Now (2.1.23) and (2.1.24) in combination with the last inequality give
Cc
s
jq 2j )1/q
j=r
s
j
j=r j (2 )
=c
s

j=r
2j/(1q)
(C (2j ))q/(1q)
(1q)/q
.
By letting r and by the monotonicity of C , we obtain

1q
1q
q
q
1q
1q
q
q
s
()

2j
t
j
2
c
dt
.
Cc
C (t)
C (t)
0
j=

This completes the proof of the Theorem.
2.1.5 Inequality (2.1.10) Containing Particular Measures

We give two examples that illustrate applications of the inequality (2.1.10).
Example 1. Let = Rn , Rn1 = {x Rn , xn = 0}, (A) = mn1 (A
), where A is any Borel subset of Rn . Obviously,
R
n1
(g)
and hence
uL1 (Rn1 )
for all u D(Rn ).
1
s(g)
2
1
uL1 (Rn )
2
133
Example 2. Let A be any Borel subset of Rn with mn (A) < and let

(A) =
|x| dx,
A
where [0, 1]. Further, let Br be a ball centered at the origin, whose ndimensional measure equals mn (A). In other words,

r=
Obviously,

|x| dx
A
|x|
n
mn (A)
n
dx + r
1/n
.
|x| dx.
mn (Br \A)
ABr
Br
So

(n1)/n
(A)(n1)/(n) (n )(1n)/(n) n(n1)/n(n) nmn (A)
.
Let g be any admissible set in Rn . By virtue of the isoperimetric inequality
nmn (g)
(n1)/n
n1/n s(g),
we have
(g)(n1)/(n) (n )(1n)/(n) n(1)/(n) s(g).
This inequality becomes an equality if g is a ball. Therefore
sup
{g}
(g)(n1)/(n)
= (n )(1n)/(n) n(1)/(n)
s(g)
and for all u D(Rn )

Rn
(n1)/(n)

u(x)(n)/(n1) |x| dx
(n )(1n)/(n) n(1)/(n) uL1 (Rn )
(2.1.25)
with the best possible constant.

2.1.6 Power Weight Norm of the Gradient on the Right-Hand Side
In this subsection we denote by z = (x, y) and = (, ) points in Rn+m with
(d)
x, Rn , y, Rm , m, n > 0. Further, let Br (q) be the d-dimensional ball
with center q Rd .
134
Lemma 1. Let g be an open subset of Rn+m with compact closure and

smooth boundary g, which satises

1
(2.1.26)
|| d
|| d = ,
(n+m)
(n+m)
2
Br
(z)g
Br
(z)
where > m for m > 1 and 0 > 1 for m = 1. Then

|| ds() cr n+m1 r + |y| ,
(n+m)
Br
(2.1.27)
(z)g
where s is the (n + m 1)-dimensional area.

The proof is based on the next lemma.
Lemma 2. Let > m for m > 1 and 0 > 1 for m = 1. Then for
any v C (Br
(n+m)
(n+m)
) there exists a constant V such that

v() V || d cr
v()|| d.
(2.1.28)
(n+m)
Br
Br
(n+m)
= B and
Proof. It suces to derive (2.1.28) for r = 1. We put B1
(m)
(n)
B1 B1 = Q. Let R() denote the distance of a point Q from the
origin, i.e., R() = (1 + ||2 )1/2 for || = 1, || < 1 and R() = (1 + ||2 )1/2 for
|| = 1, || < 1. Taking into account that B is the quasi-isometric image of Q
under the mapping /R(), we may deduce (2.1.28) from the inequality

v() V || d c v()|| d,
(2.1.29)
Q
which will be established now. Since (m + )|| = div(|| ), then, after

integration by parts in the left-hand side of (2.1.29), we nd that it does not
exceed

+1

v() V ds() . (2.1.30)
|v|||
d +
d
(m + )
(n)
(m)
B1
B1
(n)
(m)
(m)
For the sake of brevity we put T = B1 (B1 \B1/2 ). Let m > 1. The
second summand in (2.1.30) is not greater than

c
|v| d + c
|v V | d.
T
By Lemma 1.1.11, the last assertion and (2.1.30) imply (2.1.29), where V
is the mean value of v in T . (Here it is essential that T is a domain for m > 1.)
(n)
If m = 1 then T has two components T+ = B1 (1/2, 1) and T =
(n)
B1 (1, 1/2). Using the same argument as in the case m > 1, we obtain

(n)
B1

v(, 1) V d c

v() d c
135

v()|| d,
Q
where V are the mean values of v in T . It remains to note that

1

v

v()|| d,

|V+ V | c
d
d c
(n)
1
B1
Q
provided 0. So for m = 1 we also have (2.1.29) with V replaced by V+ or
V . This concludes the proof of the lemma.

(n+m)
(z). In (2.1.28)
Proof of Lemma 1. For the sake of brevity let B = Br
we replace v by a mollication of the characteristic function of the set g
with radius . Then the left-hand side in (2.1.28) is bounded from below by
the sum

|| d + |V |
|| d,
|1 V |
e1
e0
where ei = { B : () = i}, i = 0, 1.
Let be a suciently small positive number. By (2.1.26)

1
|1 V | + |V |
|| d cr
|| () d
2
B
B
for suciently small values of . Consequently,

1

|| d cr lim sup
|| () d = cr
|| ds().
2 B
+0
B
Bg
It remains to note that

|| d crm+n r + |y| .
Remark 1. Lemma 1 fails for m = 1, > 0. In fact, let g = { Rn+1 :

> or 0 > > }, where = const > 0. Obviously, (2.1.26) holds for
this g. However,

(n+1)
Br
|| ds() c ,
which contradicts (2.1.27).

Theorem 1. Let be a measure in Rn+m , q 1, > m. The best
constant in

|y| |z u| dz, u C0 Rn+m ,
uLq (Rn+m ,) C
(2.1.31)
Rn+m
is equivalent to
136

1nm (n+m) 1/q
K = sup + |y|

(z)
.
B
(2.1.32)
z;
Proof. 1. First, let m > 1 or 0 > 1, m = 1. According to Theorem 2.1.3

[(g)]1/q
C = sup
,
|y| ds(z)
g
g
where g is an arbitrary subset of Rn+m with a compact closure and smooth
boundary. We show that for each g there exists a covering by a sequence of
(zi ), i = 1, 2, . . . , such that
balls Bn+m
i

n+m1
+
|y
|
c
|y| ds(z).

i
i
i
g
(n+m)
(z) for which (2.1.26) is valid.

Each point z g is the center of a ball Br
In fact, the ratio in the left-hand side of (2.1.26) is a continuous function in
r that equals unity for small values of r and converges to zero as r . By
(n+m)
Theorem 1.2.1 there exists a sequence of disjoint balls Bri
(zi ) such that
g
(n+m)
B3ri
(zi ).
i=1
According to Lemma 1,

(n+m)
B ri
(zi )g

|y| ds(z) crin+m1 ri + |yi | .
(n+m)
(zi )}i1 is the required covering.

Thus {B3ri
Obviously,

(n+m)

1/q q
(n+m)
B3ri
(zi )
(zi )
(g)
B3ri
i
cK q

i

n+m1
ri
ri + |yi |
q

cK
q
|y| ds(z)
Therefore C cK for m > 1 and for m = 1, 0 > 1.

2. Now let m = 1, > 0. We construct a covering of the set { : = 0}
by balls Bj with radii j , equal to the distance of Bj from the hyperplane
{ : = 0}. We assume that this covering has nite multiplicity. By {j } we
denote a partition of unity subordinate to {Bj } and such that |j | c/j
(see Stein [724], Chap. VI, 1.). Using the present theorem for the case = 0,
which has already been considered (or equivalently, using Theorem 1.4.2/2),
we arrive at
137

1/q

(j u)
j uLq (Rn+1 ,) c sup n j B(n+1) (z)
,
L1 (Rn+1 )
;z
where j is the restriction of to Bj . It is clear that

1/q
sup n j B(n+1) (z)
c
;z
sup
j ,zBj

1/q
n B(n+1) (z)
.
Therefore,
j uLq (Rn+1 ,)
c
sup
( + j )
j ,zBj
(n+1) 1/q
B
(z)
Summing over j and using (2.1.29), we obtain

uLq (Rn+1 ,) cK
|u||| d +
Rn+1

Rn+1
Rn+1

(j u)|| d.

|u|||1 d .
Since

Rn+1
|u|||

Rn+1
|u||| d
for > 0, then C cK for m = 1, > 0.

3. To prove the reverse estimate, in (2.1.31) we put U () = (1 ( z)),
(n+m)
(n+m)
where C0 (B2
), = 1 on B1
. Since

|| | u| d c1
|| d cn+m1 + |y| ,
(n+m)
B2
(n+m)
(z)
B2
(z)
the result follows.
Corollary. Let be a measure in Rn , q 1, > m. Then the best

constant in (2.1.31) is equivalent to
sup
xRn ,>0

1/q
1mn B(n) (x)
.
For the proof it suces to note that K, dened in (2.1.32), is equivalent

to the preceding supremum if supp Rn .
Remark 2. The part of the proof of Theorem 1 for the case m = 1, > 0
is also suitable for m > 1, > 1 m since for these values of and for all
u C0 (Rn+m ) we have

|u|||1 d ( + m 1)1
|u||| d.
(2.1.33)
Rn+m
Rn+m
This implies that the best constant C in (2.1.31) is equivalent to
138
K1 =
sup
zRn+m ;<|y|/2
1/q

|y| 1nm B(n+m) (z)
for m 1, > 1 m.
Since (2.1.33) is also valid for < 1 m with the coecient (1 m )1
if u vanishes near the subspace = 0, then following the arguments of the
second and third parts of the proof of Theorem 1 with obvious changes, we
arrive at the next theorem.
Theorem 2. Let be a measure in { Rn+m : = 0}, q 1, < 1m.
Then the best constant in (2.1.31), where u C0 ({ : = 0}), is equivalent
to K1 .
2.1.7 Inequalities of HardySobolev Type as Corollaries of
Theorem 2.1.6/1
Here we derive certain inequalities for weighted norms which often occur in
applications. Particular cases of them are the Hardy inequality
l
|x| u
cl uLp (Rn )
L (Rn )
p
and the Sobolev inequality

uLpn/(nlp) (Rn ) cl uLp (Rn ) ,
where lp < n and u D(Rn ). We retain the notation introduced in Sect.
2.1.5.
Corollary 1. Let
1 q (m + n)/(m + n 1),
Then

|y| u
L
n+m )
q (R
=1+

c|y| uL
m
q1
(m + n) > .
q
q
(2.1.34)
n+m )
1 (R
for u D(Rn+m ).
Proof. According to Theorem 2.1.6/1 it suces to establish the uniform
boundedness of the value

1/q

1nm
+ |y|

||q d
|z|<
with respect to and z. Obviously, it does not exceed

1nm+n/q

c + |y|
1/q

|y|<
||q d
139
This value is not greater than c|y| 1(m+n)(q1)/q for c|y| and
for > c|y|. The result follows.

c
+1(m+n)(q1)/q
In (2.1.34) let us replace q 1 , , and by 1 p1 + q 1 , + (1 p)1 q,

and((1p1 )q +1), respectively, and u by |u|s with s = (p1)qp1 +1. Then
applying Holders inequality with exponents p and p/(p 1) to its right-hand
side we obtain the following assertion.
Corollary 2. Let m + n > p 1, p q p(n + m)(n + m p)1 , and
= 1 + (n + m)(1/p 1/q) > m/q. Then

|y| u
c|y| uL (Rn+m )
(2.1.35)
L (Rn+m )
q
for all u D(Rn+m ).

For p = 2, = 1 m/2, n > 0, the substitution of u(z) = |y| v(z)
into (2.1.35) leads to the next corollary.
Corollary 3. Let m + n > 2, 2 < q 2(n + m)/(n + m 2), and
= 1 + (n + m)(21 q 1 ). Then

2
v2
(m 2)2
2
|y| v
c
(v) dz
dz
(2.1.36)
2
Lq (Rn+m )
4
Rn+m
Rn+m |y|
for all v D(Rn+m ), subject to the condition v(x, 0) = 0 in the case m = 1.
In particular, the exponent vanishes for q = 2(m + n)/(m + n 2) and
we obtain

(m 2)2
v2
2
dz
(v)2 dz, (2.1.37)
cvL 2(m+n) (Rn+m ) +
2
4
Rn+m |y|
Rn+m
m+n2
which is a renement of both the Sobolev and the Hardy inequalities, the
latter having the best constant.
To conclude this subsection we present a generalization of (2.1.35) for
derivatives of arbitrary integer order l.
Corollary 4. Let m + n > lp, 1 p q p(m + n lp)1 (m + n), and
= l + (m + n)(p1 q 1 ) > mq 1 . Then

y| u|
c|y| l uL (Rn+m )
(2.1.38)
L (Rn+m )
q
for u D(Rn+m ).
Proof. Let pj = p(n + m)(n + m p(l j))1 . Successively applying the
inequalities

|y| u
c|y| uL (Rn+m ) ,
Lq (Rn+m )
p1

|y| j u
c|y| j+1 uLp +1 (Rn+m ) , 1 j < l,
Lp (Rn+m )
j
140
which follow from (2.1.35), we arrive at (2.1.38).
Inequality (2.1.38) and its particular cases (2.1.34) and (2.1.35) obviously
fail for = l+nq1 (m+n)p1 . Nevertheless for this critical we can obtain
similar inequalities that are also invariant under similarity transformations in
Rn+m by changing the weight function on the left-hand side.
The results of Sects. 2.1.12.1.3 and 2.1.5 are borrowed from the authors
paper [543] (see also [552]).
Properties of the weighted area minimizing function C introduced in Definition 2.1.4 were studied under the assumption that (x, ) does not depend
on x and is convex. In particular, the sharp generalized isoperimetric inequality

N (x) ds(x) nn1/n mn (g)
(2.1.39)
g
holds for all admissible sets g Rn . Here n is the volume of the set {
Rn : () 1} with
(x, )Rn
() = sup
x=0 (x)
(see Busemann [158] and Burago, Zalgaller [151]). The surfaces minimizing
the integrals of the form

N (x) ds(x)
g
over all sets g with a xed volume, called Wul shapes, appeared in 1901 (see
Wul [798]). The Wul shape is called the crystal of the function , which in
its turn is called crystalline if its crystal is polyhedral (see, in particular, J.E.
Taylor [745] for a theory of crystalline integrands as well as the bibliography).
The sharp constant

C =
+1
+2
1

+1
+2
+2
2
(sin t) dt
,
0,
in the weighted isoperimetric inequality

2
1/2
+1
+2
N1 + |x|2 N22
m2 (g)
C
ds,
g
where (N1 , N2 ) = N was found by Monti and Morbidelli [611], which is

equivalent to the sharp integral inequality

uL +2 (R2 ) C
+1
R2
u
x
2

+ |x|
u
y
2 1/2
dx
2.2 (p, )-Capacity
141
for all u C0 (R2 ).

Theorem 2.1.4 is borrowed from Mazya [560], its proof being a modication of that in Mazya and Netrusov [572] relating to the case p > 1. For the
contents of Sect. 2.1.6 see Sect. 2.1.5 in the authors book [556].
Obviously, in Theorem 2.1.6/1 the role of |y| can be played by the distance to the m-dimensional Lipschitz manifold F supporting the measure .
Horiuchi [383] proved the suciency in Theorem 2.1.6/1 for an absolutely
continuous measure and for a more general class of sets F depending on
the behavior as 0 of the n-dimensional Lebesgue measure of the tubular
neighborhood of F , {z Rn+m : dist(z, F ) < }.
The contents of Sect. 2.1.7 were published in [556], Sect. 2.1.6, for the rst
time. Estimates similar to (2.1.38) are generally well known (except, probably,
for certain values of the parameters p, q, l, and ) but they were established
by other methods (see Ilin [395]). The multiplicative inequality

|x| u
Lr (Rn )

a
C |x| |u|L
p (R
n)
1a
|x| u
Lq (Rn )
was studied in detail by Caarelli, Kohn, and Nirenberg [162]. Lin [498] has
generalized their results to include derivatives of any order.
The inequality (2.1.36) was proved by Mazya [556], Sect. 2.1.6. Tertikas
and Tintarev [749] (see also Tintarev and Fieseler [753], Sect. 5.6, as well
as Benguria, Frank, and Loss [83]) studied the existence and nonexistence
of optimizers in (2.1.37) and found sharp constants in some cases. In one
particular instance of (2.1.36), the sharp value of c will be given in Sect. 2.7.1.
In [277], Filippas, Mazya, and Tertikas showed that for any convex domains
Rn the inequality

|u|2 dx
1
4

u2
dx c()
d2
2n
n2
n
|u| n2 dx
C0 ()
holds where u
and d = dist(x, ). See Comments to Sect. 2.7 for
other contributions to this area.
2.2 (p, )-Capacity

2.2.1 Denition and Properties of the (p, )-Capacity
Let e be a compactum in Rn and let be the function specied in Sect.
2.1.1. The number

p
(x, u) dx : u N(e, ) ,
inf
where p 1, is called the (p, )-capacity of e relative to and is denoted by

(p, )-cap(e, ). Here
142

N(e, ) = u D() : u 1 on e .
If = Rn , we omit in the notations (p, )-cap(e, ), N(e, ), and so
on.
In the case (x, ) = ||, we shall speak of the p-capacity of a compactum
e relative to and we shall use the notation capp (e, ).
We present several properties of the (p, )-capacity.
(i) For compact sets K , F , the inclusion K F implies
(p, )-cap(K, ) (p, )-cap(F, ).
This is an obvious consequence of the denition of capacity. From the same
denition it follows that the (p, )-capacity of F relative to does not increase
under extension of .
(ii) The equality

(p, )-cap(e, ) = inf
(x, u)
p

dx : u P(e, ) ,
(2.2.1)
where P(e, ) = {u : u D(), u = 1 in a neighborhood of e, 0 u 1 in

Rn } is valid.
Proof. Since N(e, ) P(e, ) it is sucient to estimate (p, )-cap(e, )
from below. Let (0, 1) and let f N(e, ) be such that

p
(x, f ) dx (p, )-cap(e, ) + .
Let {m (t)}m1 denote a sequence of functions in C (R1 ) satisfying the

conditions 0 m (t) 1 + m1 , m (t) = 0 in a neighborhood of (, 0]
and m (t) = 1 in a neighborhood of [1, ), 0 m (t) 1 for all t. Since
m (f (x)) P(e, ), then

p

p
p
dx.
inf
(x, u) dx : u P(e, )
m f (x)
x, f (x)
Passing to the limit as m , we obtain

p
inf
(x, u) dx : u P(e, )

p
(x, f ) dx (p, )-cap(e, ) + .
(iii) For any compactum e and > 0 there exists a neighborhood G

such that
(p, )-cap(K, ) (p, )-cap(e, ) +
for all compact sets K, e K G.
2.2 (p, )-Capacity
143
Proof. From (2.2.1) it follows that there exists a u P(e, ) such that

p
(x, u) dx (p, )-cap(e, ) + .
Let G denote a neighborhood of e in which u = 1. It remains to note that

p
(x, u) dx
(p, )-cap(K, )
for any compactum K, e K G.
The next property is proved analogously.

(iv) For any compactum e and any > 0 there exists an open set ,
, such that
(p, )-cap(e, ) (p, )-cap(e, ) + .
(v) The Choquet inequality
(p, )-cap(K F, ) + (p, )-cap(K F, )
(p, )-cap(K, ) + (p, )-cap(F, )
holds for any compact sets K, F .
Proof. Let u and v be arbitrary functions in P(K, ) and P(F, ), respectively. We put = max(u, v), = min(u, v). Obviously, and have
compact supports and satisfy the Lipschitz condition in , = 1 in the
neighborhood of K F and = 1 in a neighborhood of K F . Since the set
{x : u(x) = v(x)} is the union of open sets on which either u > v or u < v,
and since u(x) = v(x) almost everywhere on {x : u(x) = v(x)}, then

p

p
(x, ) dx +
(x, ) dx

p

p
(x, u) dx +
(x, v) dx.
=
Hence, having noted that mollications of the functions and belong

to P(K F, ) and P(K F, ), respectively, we obtain the required
inequality.

A function of compact sets that satises conditions (i), (iii), and (v) is
called a Choquet capacity.
Let E be an arbitrary subset of . The number (p, )-cap(E, ) =
sup{K} (p, )-cap(K, ), where {K} is a collection of compact sets contained
in E, is called the (p, ) capacity of E relative to . The number
inf (p, )-cap(G, ),
{G}
144
where {G} is the collection of all open subsets of containing E, is called the
outer capacity (p, )-cap(E, ) of E . A set E is called (p, ) capacitable
if
(p, )-cap(E, ) = (p, )-cap(E, ).
From these denitions it follows that any open subset of is (p, ) capacitable. If e is a compactum in , then by property (iii), given > 0, there
exists an open set G such that
(p, )-cap(G, ) (p, )-cap(e, ) + .
Consequently, all compact subsets of are (p, ) capacitable.
From the general theory of Choquet capacities it follows that analytic sets,
and in particular, Borel sets are (p, ) capacitable (see Choquet [186]).
2.2.2 Expression for the (p, )-Capacity Containing an Integral
over Level Surfaces
Lemma 1. For any compactum F the (p, )-capacity (for p > 1) can be
dened by

(p, )-cap(F, ) =
inf
uN(F,)
d
ds 1/(p1)
( E [(x, u)]p |u|
)
1p
, (2.2.2)
where Et = {x : |u(x)| = t}.

We introduce the following notation: is the set of nondecreasing functions
C (R1 ), which satisfy the conditions (t) = 0 for t 0, (t) = 1 for
t 1, supp (0, 1); 1 is the set of nondecreasing functions that are
absolutely continuous on R1 and satisfy the conditions (t) = 0 for t 0,
(t) = 1 for t 1, (t) is bounded.
To prove Lemma 1 we shall use the following auxiliary assertion.
Lemma 2. Let g be a nonnegative function that is integrable on [0, 1].
Then
1
1p
1
dt
( )p g dt =
.
(2.2.3)
inf
1/(p1)
0
0 g
Proof. First we note that by Holders inequality

dt
1=
0
p
1/p
( ) g dt
0
dt
g 1/(p1)
11/p
,
and hence the left-hand side of (2.2.3) is not smaller than the right.
Let 1 , (t) = (t) for t [ 1 , 1 1 ], supp [ 1 , 1 1 ],
= 1, 2, . . . . We set
2.2 (p, )-Capacity

1
(t) = (t)
145
Since the sequence converges to on (0,1) and is bounded, it follows by

Lebesgues theorem that
1
1
p g d
( )p g d.
0
Mollifying , we obtain the sequence { }, C (R1 ), supp (0, 1),

1
1
1
d = 1,
p g d
( )p g d.
0
Setting
(t) =
d,
0
we obtain a sequence of functions in such that

1
1
( )p g d
( )p g d.
0
Hence,

inf
Let
( )p g d = inf

M = t : g(t) ,
( )p g d.
(2.2.4)
0 (t) =
d,
0
where (t) = 0 on R1 \M and

1

(t) = g(t)1/(1p)
for t M .
g 1/(1p) d
M
Obviously, 0 1 , and
(0 )p g d =

1p
g 1/(1p) d
By (2.2.4) the left-hand side of (2.2.3) does not exceed

1p

g 1/(1p) d
We complete the proof by passing to the limit as 0.
Proof of Lemma 1. Let u N(F, ), . From the denition of capacity

and Theorem 1.2.4 we obtain
146

(p, )-cap(F, ) =
p
(u)(x, u)
( )p g dt,
where
dx =
g(t) =
Et
p ds
.
(x, u)
|u|
By Lemma 2

1p
(p, )-cap(F, )
(2.2.5)
1/(1p)
To prove the opposite inequality it is enough to note that

p
(x, u)

dx
g d
0
1p
1/(1p)
Recalling the property (2.2.1) of the (p, )-capacity, note that, in passing,
we have proved here also the following lemma.
Lemma 3. For any compactum F the (p, )-capacity (p > 1) can be
dened as

(p, )-cap(F, ) =
inf
uP(F,)
dt
ds 1/(p1)
( Et [(x, u)]p |u|
)
1p
2.2.3 Lower Estimates for the (p, )-Capacity

Lemma. For any u D() and almost all t 0,
(Lt )
p/(p1)
d
mn (Lt )
dt

Lt
p ds
(x, u)
|u|
1/(p1)
, (2.2.6)
where, as usual, Lt = {x : |u(x)| > t}.

Proof. By H
olders inequality, for almost all t and T , t < T ,

p/(p1)
Lt \LT
|u|p1 (x, u) dx

Lt \LT
|u|p dx
Using Theorem 1.2.4, we obtain

Lt \LT
(x, u)
p
1/(p1)
dx
2.2 (p, )-Capacity

p/(p1)
p1
(L ) d
147
Lt \LT

|u|p dx
(x, u)
p ds
|u|
1/(p1)
.
We divide both sides of the preceding inequality by (T t)p/(p1) and estimate

the rst factor on the right-hand side

1
T t
Tp
p/(p1)
p1
(L ) d

T

p ds 1/(p1)
mn (Lt \LT )
1
d
.
(x, u)
T t
T t t
|u|
L
Passing to the limit as T t, we obtain (2.2.6) for almost all t > 0. The
lemma is proved.

From Lemma 2.2.2/3 and from the Lemma of the present subsection we
immediately obtain the following corollary.
Corollary 1. The inequality
(p, )-cap(F, )
inf

uP(F,)
1
0
d
d
mn (L )
d
[(L )]p/(p1)
1p
(2.2.7)
holds.
Denition. In what follows we use the function C introduced in Denition 2.1.4 assuming = mn , that is, C stands for the inmum (g) for all
admissible sets such that mn (g) . Then from (2.2.7) we obtain the next
corollary, containing the so-called isocapacitary inequalities.

(p, )-cap(F, )
mn ()
mn (F )
d
[C ()]p/(p1)
1p
(2.2.8)
is valid.
By virtue of the classical isoperimetric inequality

(n1)/n
s(g) n(n1)/n n1/n mn (g)
,
in the case (x, ) = || we have
C () = n(n1)/n n1/n (n1)/n .
Therefore,
(2.2.9)
148

n p1
mn ()(pn)/n(p1)
capp (F, )
p 1
1p
mn (F )(pn)/n(p1)
(2.2.10)
p
np/n n(np)/n
for p = n and
capp (F, ) n
n1

1n
mn ()
n log
mn (F )
(2.2.11)
for p = n.
In particular, for n > p,

capp (F )
np/n n(np)/n
np
p1
p1
mn (F )(np)/n .
(2.2.12)
2.2.4 p-Capacity of a Ball

We show that the estimates (2.2.10) and (2.2.11) become equalities if and
F are concentric balls of radii R and r, R > r, i.e.,
p1

(pn)/(p1)
1p
|n p|
R
capp (F, ) = n
r(pn)/(p1)
(2.2.13)
p1
for n = p and

1n
R
capn (F, ) = n log
r
(2.2.14)
for n = p.
Let the centers of the balls and F coincide with the origin O of spherical
coordinates (, ), || = 1. Obviously,

R p
u n1

capp (F, )
inf
d
d

uN(F,) B1
r
R p

u n1

d inf
d.
uN(F,) r
B1
The inner integral attains its inmum at the function
(pn)/(p1) (pn)/(p1)
R(pn)/(p1) (pn)/(p1)
r
[r, R] v() = R
1
)
log(R1
log(rR
for p = n,
for p = n.
This implies the required lower estimates for the p-capacity. The substitution
of v() into the integral |u|p dx leads to (2.2.13) and (2.2.14).
In particular, the p-capacity of the n-dimensional ball Br relative to Rn is
p1 np
r
for n > p and to zero for n p. Since the p-capacity
equal to n ( np
p1 )
is a monotone set function, then for any compactum p-cap(F, Rn ) = 0, if
2.2 (p, )-Capacity
149
n p. In the case p n the capacity of a point relative to any open set ,

containing this point, equals zero. If p > n, then the p-capacity of the center
p1 np
R
. Therefore, in the last
of the ball BR relative to BR equals n ( pn
p1 )
case, the p-capacity of any compactum relative to any bounded open set that
contains this compactum is positive.
2.2.5 (p, )-Capacity for p = 1
Lemma. For any compactum F
(1, )-cap(F, ) = inf (g),
where the inmum is taken over all admissible sets g in containing F .
Proof. Let u N(F, ). Applying Theorem 1.2.4, we obtain
1

(x, u) dx =
(Lt ) dt inf (g).
gF
This implies the lower estimate for the capacity.

Let g be an admissible set containing F . The function u (x) = (d(x))
dened in the proof of the second part of Theorem 2.1.1 belongs to N(F, )
for suciently small > 0. So

(1, )-cap(F, )
(x, u ) dx.
In the proof of the second part of Theorem 2.1.1, it was shown that the preceding integral converges to (g), which yields the required upper estimate
for the capacity. The lemma is proved.

2.2.6 The Measure mn1 and 2-Capacity
(n1)
Lemma. If B
is an (n 1)-dimensional ball in Rn , n > 2, then

n n2

,
(2.2.15)
cap2 B(n1) , Rn =
cn
where c3 = 3 , c4 = 1, and cn = (n 4)!!/(n 3)!! for odd n 5 and

cn = 2 (n 4)!!/(n 3)!! for even n 6.
Proof. We introduce ellipsoidal coordinates in Rn : x1 = sinh cos 1 ,
xj = cosh sin 1 , . . . , sin j1 cos j , j = 2, . . . , n 1, xn = cosh sin 1 ,
. . . , sin n1 . A standard calculation leads to the formulas

dx = n cosh2 sin2 1 (cosh )n2 d d,
2

1
2
2 u
(u) =
+ ,
cosh2 sin2 1
150
where d is the surface element of the unit ball in Rn and the dots denote a
positive quadratic form of all rst derivatives of u except u/. The equation
(n1)
in the new coordinates is = 0. Therefore
of the ball B

2

(n1) n
u
n2
n2
inf
cap2 B
,R
(cosh )
d d,
||=1 {u} 0
where {u} is a set of smooth functions on [0, ) with compact supports. The
inmum on the right-hand side is equal to
1

d
= c1
n .
n2
(cosh
)
0
This value is attained at the function

1

d
d
,
v=
(cosh )n2 0 (cosh )n2
(n1)
and decreases suciently rapidly at innity.

which equals unity on B
Substituting v into the Dirichlet integral, we obtain
2

(n1) n
v
n n2
n2
cap2 B
, R n
(cosh )n2 d =

.
cn
0
This proves the lemma.
We now recall the denition of the symmetrization of a compact set K in

Rn relative to the (n s)-dimensional subspace Rns .
Let any point x Rn be denoted by (y, z), where y Rns , z Rs . The
image K of the compact set K under symmetrization relative to the subspace
z = 0 is dened by the following conditions:
1. The set K is symmetric relative to z = 0.
2. Any s-dimensional subspace, parallel to the subspace y = 0 and crossing
either K or K also intersects the other one and the Lebesgue measures of
both cross sections are equal.
3. The intersection of K with any s-dimensional subspace, which is parallel to the subspace y = 0, is a ball in Rs centered at the hyperplane z = 0.
Below we follow Polya and Szego [666] who established that the 2-capacity
does not increase under the symmetrization relative to Rn1 . Let be an
(n 1)-dimensional hyperplane and let Pr F be the projection of F onto .
We choose so that mn1 (Pr F ) attains its maximum value. We symmetrize
F relative to and obtain a compactum that is also symmetrized relative to a
straight line perpendicular to . So we obtain a body whose capacity does not
exceed 2-cap F and whose intersection with is an (n 1)-dimensional ball
with volume mn1 (Pr F ). Thus the (n 1)-dimensional ball has the largest
area of orthogonal projections onto an (n 1)-dimensional plane among all
compacta with xed 2-capacity.
2.2 (p, )-Capacity
151
This and the Lemma imply the isocapacitary inequality

(n2)/(n1)
mn1 F Rn1
(n2)/(n1)

cn
n1
cap2 F , Rn ,
n1
n
(2.2.16)
where cn is the constant dened in the Lemma.

The capacity generated by the integral

f (x, u, u) dx
was introduced by Choquet [186] where it served as an illustration of general

capacity theory. Here the presentation follows the authors paper [543].
Lemma 2.2.2/1 for p = 2, (x, ) = || is the so-called Dirichlet principle
with prescribed level surfaces veried in the book by P
olya and Szeg
o [666]
under rigid assumptions on level surfaces of the function u. As for the general case, their proof can be viewed as a convincing heuristic argument. The
same book also contains isocapacitary inequalities, which are special cases of
(2.2.10) and (2.2.11).
Lemma 2.2.3, leading to lower estimates for the capacity, was published
for (x, ) = || in 1969 by the author [538] and later by Talenti in [741],
p. 709.
Properties of symmetrization are studied in the books by Polya and
Szeg
o [666] and by Hadwiger [334] et al. See, for instance, the book by
Hayman [357] where the circular symmetrization and the symmetrization
with respect to a straight line in R2 are considered. Nevertheless, Haymans
proofs can be easily generalized to the n-dimensional case. Lemma 2.2.5 is a
straightforward generalization of a similar assertion due to Fleming [281] on
1-capacity.
In the early 1960s the p-capacity was used by the author to obtain the necessary and sucient conditions for the validity of continuity and compactness
properties of Sobolev-type embedding operators [527, 528, 530, 531].
Afterward various generalizations of the p-capacity proved to be useful in
the theory of function spaces and nonlinear elliptic equations. A Muckenhoupt
Ap -weighted capacity was studied in detail by Heinonen, Kilpelainen, and
Martio [375] and Nieminen [636] et al. A general capacity theory for monotone
operators

Wp1 u div a(x, Du) Wp1
was developed by Dal Maso and Skrypnik [220], whose results were extended
to pseudomonotone operators by Casado-Daz [174]. Biroli [104] studied the
p-capacity related to the norm
152
m

i=1
1/p

|Xi u|p dx +
|u|p dx
where Xi are vector elds subject to Hormanders condition: They and their
commutators up to some order span at every point all Rn . Properties of a
1
with the variable exponent p :
capacity generated by the Sobolev space Wp()
n
R (1, ) were investigated by Harjulehto, Hast
o, Koskenoja, and Varonen
[353]. Another relevant area of research is the p-capacity on metric spaces with
a measure (see, for instance, Kinnunen and Martio [425] and Goldshtein,
Troyanov [317]), in particular, on the Carnot group and Heisenberg group
(see Heinonen and Holopainen [374]).
A generalization of the inequality (2.2.8) was obtained by E. Milman [603]
in a more general framework of measure metric spaces for the case (x, ) =
||. Similarly to Sect. 2.3.8, if we introduce the p-capacity minimizing function
p (t) = inf capp (F, ),
where the inmum is taken over compacta F with mn (F ) t, then for
any p1 p0 1
( qq0 1)p1 /q0
1
1 q1 p /q
p1 (t)
(1 q0 ) 1 1

mn ()
t
ds
q
/q
q /p
1
0
(s t)
p0 (s) 1 0
p1 /q1
,
(2.2.17)
where qi = pi /(pi 1) denote the corresponding conjugate exponents. Clearly

(2.2.17) coincides with (2.2.8) when p0 = 1 by Lemma 2.2.5.
Under an appropriate curvature lower bound on the underlying space, it
was also shown in [603, 605] that (2.2.17) may, in fact, be reversed, to within
numeric constants. An application of this fact was given by E. Milman [604]
to the stability of the rst positive eigenvalue of the Neumann Laplacian on
convex domains, with respect to perturbation of the domain.

(the Case p > 1)
2.3.1 The (p, )-Capacitary Inequality
Let u D() and let g be the function dened by (2.2.5) with p > 1. Further,
let

def
T = sup t > 0 : (p, )-cap(Nt , ) > 0 > 0,
(2.3.1)
where Nt = {x : |u(x)| t}. From (2.3.1) it follows that
def
(t) =
[g( )]1/(p1)
(2.3.2)
2.3 Conditions for Validity of Integral Inequalities (the Case p > 1)
153
for 0 < t < T . In fact, let

2
v(x) = t2 u(x) .
Since v N(Nt , ), then from Lemma 2.2.2/1 and (2.3.1) we obtain

1

p ds 1/(1p)
(x, v)
d
|v|
0
{x:v(x)= }

1/(1p)
(p, )-cap(Nt , )
< ,
and it remains to note that

1
1

p ds 1/(p1)
d
=
d.
(x,
v)
1/(p1)
|v|
{x:v(x)= }
0 g( )
0
Since by Theorem 1.2.4

g( ) d =
p
(x, u) dx < ,
it follows that g(t) < for almost all t > 0 and the function (t) is strictly
monotonic. Consequently, on the interval [0, (T )) the function t(), which is
the inverse of (t), exists.
Lemma. Let u be a function in D() satisfying condition (2.3.1). Then
the function t() is absolutely continuous on any segment [0, (T )], where
(0, T ), and
(t)

p

p
(x, u) dx
t () d.
(2.3.3)
0
If T = max |u|, then we may write the equality sign in (2.3.3).

Proof. Let 0 = 0 < 1 < < m = (T ) be an arbitrary partition
of the segment [0, (T )]. By H
olders inequality,
t(k+1 )
[t(k+1 ) t(k )]p
[t(k+1 ) t(k )]p
= t( )
g( ) d,
(k+1 k )p1
t(k )
[ t(kk+1
g( )1/(1p) d ]p1
)
and consequently,
m1

k=0
m1
t(k+1 )
[t(k+1 ) t(k )]p
g( ) d
(k+1 k )p1
k=0 t(k )

T

p
(x, u) dx. (2.3.4)
g( ) d
=
0
The last inequality follows from Theorem 1.2.4. By (2.3.4) and F. Rieszs
theorem (see Natanson [627]), the function t() is absolutely continuous and
its derivative belongs to Lp (0, (T )). By Theorem 1.2.4,
154

p
(x, u) dx lim
+0
g( ) d.
(2.3.5)
Since t() is a monotonic absolutely continuous function, we can make the

change of variable = t() in the last integral. Then
(T )
g( ) d =
0
(T )
t ()
t ()g() d =
0
p
d,
which, along with (2.3.5), completes the proof.
Theorem. (Capacitary Inequality) Let u D(). Then for p 1,

p

pp
(x, u) dx.
(p, )-cap(Nt , ) d tp
(2.3.6)
p1
(p 1)
0
For p = 1 the coecient in front of the integral on the right-hand side

of (2.3.6) is equal to one. The constant pp (p 1)1p is optimal.
Proof. To prove (2.3.6) it is sucient to assume that the number T , dened
in (2.3.1), is positive. Since by Lemma 2.2.5
(1, )-cap(Nt , ) (Lt )
for almost all t > 0, we see that (2.3.6) follows from (2.1.4) for p = 1.
Consider the case p > 1. Let (t) be a function dened by (2.3.2) and let
t() be the inverse of (t). We make the change of variable

(p, )-cap(Nt , ) d tp =

0
(T )

(p, )-cap(Nt() , ) d t()p .
Setting v = t2 u2 , = t2 2 in (2.3.2), we obtain

1
(t) =
0
{x:v(x)=}
(x, v)
p ds
|v|
1/(1p)
d.
(2.3.7)
Since v N(Nt , ), then by Lemma 2.2.2/1 the right-hand side of (2.3.7)

does not exceed
1/(1p)

.
(p, )-cap(Nt() , )
Consequently,

(T )
0
d[t()]p
=p
p1
(T )
Applying the Holder inequality and the Hardy inequality
t()
p1
t () d.
(T )
0
we arrive at

[t()]p
d
p
p
p1

p
(T )
t ()
p
d,
155
(2.3.8)
pp
(p 1)p1
(T )
t ()
p
d,
which together with Lemma 2.3.1 yields (2.3.6).

To show that the constant factor in the right-hand side of (2.3.6) is sharp,
it suces to put (x, y) = |y| and u(x) = f (|x|). Then (2.2.13) and (2.3.6)
imply

p n1
|f (r)|p n1
|n p|p
pp
f (r) r
r
dr
dr,
p1
p
p1
(p 1)
r
(p 1)
0
0
which is a particular case of the sharp Hardy inequality (1.3.1).
Remark 1. The inequality

p
p
(x, u) dx,
(p, )-cap(Nt , ) d t C
(2.3.9)
with a cruder constant than in (2.3.6) can be proved more simply in the
following way. By the monotonicity of capacity, the integral in the left-hand
side does not exceed
+

def
= 2p 1
2pj (p, )-cap(N2j , ).
j=
Let C (R1 ), (t) = 1 for t 1, (t) = 0 for t 0, 0 (t) 1 + ,

and let

uj (x) = 21j u(x) 1 .
Since uj N(N2j , ), we have

2pj
2p1
j=

2p1
j=
N2j1 \N2j
N2j1 \N2j
(x, uj )
p
dx

p
p
21j |u| 1
(x, u) dx
p
(x, u) dx.
(1 + ) 2
p 2p1
Letting tend to zero, we obtain (2.3.9) with the constant C = 22p1 , which
completes the proof.
Remark 2. In fact, the inequality just obtained is equivalent to the following
one stronger than (2.3.9) (modulo the best constant)
156

(p, )-cap(N2t , Lt ) d tp c
(x, u)
p
dx.
(2.3.10)
Such conductor inequalities will be considered in Chap. 3.

2.3.2 Capacity Minimizing Function and Its Applications
Denition. Let p (t) denote
inf(p, )-cap(
g , ),
where the inmum is taken over all admissible sets g with
(g) t.
Note that by Lemma 2.2.5

1 (t) = C (t),
where C is the weighted area minimizing function introduced in Denition 2.1.4.
The following application of the capacity minimizing function p is immediately deduced from the capacitary inequality (2.3.6)

p

p
p (Nt ) d t C
(2.3.11)
(x, u) dx,
0
where C pp (p1)1p . Conversely, minimizing the integral in the right-hand

side over P(F, ), we see that (2.3.11) gives the isocapacitary inequality

p (F ) C(p, )-cap(F, ).
If for instance, = mn , then (2.2.8) leads to the estimate
mn ()

0
mn (Nt )
d
[C ()]p/(p1)
1p

d tp
pp
(p 1)p1
p
(x, u)
dx.
(2.3.12)
In particular, being set into (2.3.6) with p = n and (x, ) = ||, the isocapacitary inequality (2.2.11) implies

0
mn ()
log
mn (Nt )
1n

d tn
n
(n 1)n n
for all u C0 (), where C = n(n 1)1n n1 .

|u|n dx,
(2.3.13)
157
Clearly, the inequality (2.3.11) and its special cases (2.3.12) and (2.3.13)
can be written in terms of the nonincreasing rearrangement u of u introduced
by (2.1.12):
mn ()
p
u (s) dp (s) C
p
(x, u)
dx,
where C is the same as in (2.3.11). In particular, (2.3.13) takes the form

mn ()
n
umn (s)
ds
(log
mn () n
)
s
|u|n dx
(2.3.14)
for all u C0 ().

2.3.3 Estimate for a Norm in a BirnbaumOrlicz Space
We recall the denition of a BirnbaumOrlicz space (see Birnbaum and Orlicz
[103], Krasnoselski and Ruticki [463], and Rao and Ren [671]).
On the axis 0, and continuous

from the right for t 0, satisfying the conditions (0) = 0, as t .
Such functions M are sometimes called Young functions. Further, let

(s) = sup t : (t) s ,
be the right inverse of (t). The function
|u|
(s) ds
P (u) =
0
is called the complementary function to M (u).

Let LM (, ) denote the space of -measurable functions for which

P (v) d 1 < .
uLM (,) = sup uv d :
In particular, if M (u) = q1 |u|q , q > 1, then P (u) = (q )1 |u|q , q =

q(q 1)1 and

uLM (,) = (q )1/q uLq (,) .
The norm in LM (, ) of the characteristic function E of the set E is

1
1
,
E LM (,) = (E)P
(E)
158
where P 1 is the inverse of the restriction of P to [0, ).

In fact, if v = P 1 (1/(E))E , then

P (v) d = 1,
and the denition of the norm in LM (, ) implies

E v d = (E)P 1 1/(E) .
E LM (,)
On the other hand, by Jensens inequality,

1
E v d (E)P 1
P (v) d ,
(E) E
and if we assume that
P (v) d 1,
then the denition of the norm in LM (, ) yields

E LM (,) (E)P 1 1/(E) .
Although formally M (t) = |t| does not satisfy the denition of the Birnbaum
Orlicz space, all the subsequent results concerning LM (, ) include this case
provided we put P 1 (t) = 1. Then we have LM (, ) = L1 (, ).
Theorem. 1. If there exists a constant such that for any compactum
F

(2.3.15)
(F )P 1 1/(F ) (p, )-cap(F, )
with p 1, then for all u D(),
p
|u|
LM (,)

C
(x, u)
p
dx,
(2.3.16)
where C pp (p 1)1p .
2. If (2.3.16) is valid for any u D(), then (2.3.15) holds for all compacta F with C.
Proof. 1. From Lemma 1.2.3 and the denition of the norm in LM (, )
we obtain

p
p
|u|
= sup
v d d :
P (v) d 1
LM (,)
N
0

sup
N v d :
P (v) d 1 d p
0
p
N Lm (,) d .
=
0
Consequently,

p
|u|
LM (,)
159

(N )P 1 1/(N ) d p .
Using (2.3.15) and Theorem 2.3.1, we obtain

p

|u|
(p, )-cap(N , ) d p
LM (,)
0

p
pp
(x, u) dx.
p1
(p 1)
2. Let u be any function in N(F, ). By (2.3.16),

p
F LM (,) C
(x, u) dx.
Minimizing the right-hand side over the set N(F, ), we obtain (2.3.15). The
theorem is proved.

Remark 1. Obviously the isocapacitary inequality (2.3.15) can be written
in terms of the capacity minimizing function p as follows:

sP s1 p (s).
Remark 2. Let (x, y) be a function satisfying the conditions stated in
Sect. 2.1.1 and let the function (x, u, y) : R1 Rn R1 , satisfy:
(i) the Caratheodory conditions: i.e., is measurable in x for all x, y, and
continuous in x and y for almost all x.
(ii) The inequality

p
(x, u, y) (x, y)
holds.
(iii) For all u D()

p
p
lim inf
(x, u) dx.
(x, u, u) dx K
+
Then (2.3.16) in the Theorem can be replaced by the following more general
estimate:

p
|u|
C
(x, u, u) dx.
(2.3.17)
L (,)
M
As an illustration, note that Theorem 2.1.1 shows the equivalence of the inequality

uLq ()
1 + (u)2 dx,
where u D() and q 1, and the isoperimetric inequality (g)1/q (g).

Here, to prove the necessity of (2.3.15) for (2.3.17), we must set u = v,
where v N(F, ), in (2.3.17) and then pass to the limit as . An
analogous remark can be made regarding Theorems 2.1.1, 2.1.2, and others.
160
2.3.4 Sobolev Type Inequality as Corollary of Theorem 2.3.3

Theorem 2.3.3 contains the following assertion, which is of interest in itself.
Corollary. 1. If there exists a constant such that for any compactum
F
(F )p (p, )-cap(F, ),
(2.3.18)
where p 1, > 0, p 1, then for all u D()

p
(x, u) dx,
upLq (,) C
(2.3.19)
where q = 1 and C pp (p 1)1p .

2. If (2.3.19) holds for any u D() and if the constant C does not
depend on u, then (2.3.18) is valid for all compacta F with = q 1 and
C.
Remark. Obviously, the isocapacitary inequality (2.3.18) is equivalent to
the weak-type integral inequality

(2.3.20)
sup t(Lt )1/q C 1/p (, u)L ()
p
t>0
with Lt = {x : |u(x)| > t} and this, along with the Corollary, can be interpreted as the equivalence of the weak and the strong Sobolev-type estimates
(2.3.20) and (2.3.19).
2.3.5 Best Constant in the Sobolev Inequality (p > 1)
From the previous corollary and the isoperimetric inequality (2.2.12) we obtain
the Sobolev (p > 1)-Gagliardo (p = 1) inequality
uLpn/(np) CuLp ,
(2.3.21)
where n > p 1, u D(Rn ) and

C = p(n p)(1p)/p n1/n n(pn)/pn .
The value of the constant C in (2.3.21) is sharp only for p = 1 (cf. Theorem 1.4.2/1). To obtain the best constant one can proceed in the following
way.
By Lemma 2.3.1
(max |u|)
0
Putting
t ()
p

d =
Rn
|u|p dx.
p1
1/(p1)
n
(n
p)
r(np)/(1p) ,
161
t() = (r),
and assuming t() = const for (max |u|), we obtain

p n1
(r) r
dr =
|u|p dx.
n
Rn
Furthermore, by Lemma 1.2.3,

Rn
max |u|
|u|pn/(np) dx =

mn (Nt ) d tpn/(np) .
The denition of the function (t), Lemma 2.2.3, and the isoperimetric inequality (2.2.9) imply
(t)
Consequently,
and
n1/(1p)

(np)/n(1p)
p1 n
mn (Nt )
.
n p n
mn (Nt() ) n n1 rn

|u|
pn/(np)
Rn
Since
n
dx
n

r n d (r)pn/(np) .

(r)p r n1 dr < ,
0 as r . After integration by parts, we

it follows that (r)r
obtain

pn/(np) n1
(r)
|u|pn/(np) dx n
r
dr.
(np)/p
Rn
Thus,
sup
uD
uLpn/(np)
(
= n1/n sup
uLp
{}

0
[(r)]pn/(np) r n1 dr)(np)/pn

, (2.3.22)
( 0 | (r)|p r n1 dr)1/p
where {} is the set of all nonincreasing nonnegative functions on [0, ) such

that (r)r(np)/p 0 as r . Thus, we reduced the question of the best
constant in (2.3.21) to a one-dimensional variational problem that was solved
explicitly by Bliss [109] as early as 1930 by classical methods of the calculus
of variations. Paradoxically, the best constant in the Sobolev inequality had
been obtained earlier than the inequality itself appeared. The sharp upper
bound in (2.3.22) is attained at any function of the form
1n/p

,
(r) = a + brp/(p1)
a, b = const > 0,
162
and equals
n1/p
p1
np
(p1)/p
1/n

n n(p 1)
p1
B
,
.
p
p
p
Finally, the sharp constant in (2.3.21) is given by

C = 1/2 n1/p
p1
np
(p1)/p
(1 + n/2) (n)
(n/p) (1 + n n/p)
1/n
,
(2.3.23)
and the equality sign can be written in (2.3.21) if

1n/p

,
u(x) = a + b|x|p/(p1)
(2.3.24)
where a and b are positive constants (although u does not belong to D it can
be approximated by functions in D in the norm uLp (Rn ) ).
2.3.6 Multiplicative Inequality (the Case p 1)
The following theorem describes conditions for the equivalence of the generalized Sobolev-type inequality (2.3.19) and a multiplicative integral inequality.
We denote by the best constant in the isocapacitary inequality (2.3.18).
Theorem. 1. For any compactum F let the inequality (2.3.18) hold
with p 1, > 0. Further, let q be a positive number satisfying one of the
conditions (i) q q = 1 , for p 1, or (ii) q < q = 1 , for p > 1.
Then the inequality

uLq (,) C
(x, u)
p
(1)/p
dx
u
Lr (,)
(2.3.25)
holds for any u D(), where r (0, q), = r(q q)/q(q r), C
c (1)/p .
2. Let p 1, 0 < q < , r (0, q ] and for some q (0, q ] and any
u D() let the inequality (2.3.25) hold with = r(q q)/q(q r) and a
constant C independent of u.
Then (2.3.18) holds for all compacta F with = (q )1 and
p/(1)
cC
.
Proof. 1. Let p 1. By H
olders inequality,

|u|q d =
|u|q (qr)/(q r) |u|r(q q)/(q r) d
(qr)/(q r)
(q q)/(q r)
|u|q d
|u|r d
,

or equivalently,
163
uLq (,) u1

Lq (,) uLr (,) .
Estimating the rst factor by (2.3.19), we obtain (2.3.25) for p 1. Let

p > 1. By Lemma 1.2.3,

q
|u| d = q
(Nt )tq1 dt.
0
To the last integral we apply inequality (1.3.42), where x = tq , f (x) = (Nt ),

b = p(q )1 > 1, a > 1 is an arbitrary number, = a(q r)q1 , =
p(q q)/q q

(q q)/a(q r)
a
(Nt ) tar1 dt
(Nt )tq1 dt c
0

(Nt )
p/q
p1
q (qr)/p(q r)
dt
.
Since a > 1 and (Nt ) does not increase, we can apply (1.3.41) to the rst
factor in the following way:

a

a ar1
r1
(Nt ) t
dt c
(Nt )t
dt .
0
Thus,

uLq (,) c
(Nt )
p/q
(1)/p
tp1 dt
u
Lr (,) .
From condition (2.3.18) and Theorem 2.3.1 we obtain

p
p/q p1
(Nt )
(x, u) dx.
t
dt c
The proof of the rst part of the theorem is complete.

. We x a number > 0 and we
2. Let G be a bounded open set G
put
= sup
(F )p
(p, )-cap(F, G)
on the set of all compacta F in G satisfying the condition (p, )-cap(F, G) .

(If (p, )-cap(F, G) = 0 for any compactum F G, then the substitution of
an arbitrary u N(F, G) into (2.3.25) immediately leads to = 0.) Obviously,
1 (G)p < .
Let v be an arbitrary function in N(F, G) and let = max(pr1 , q r1 ).
We substitute the function u = v into (2.3.25). Then
164

1/q
(F )
cC
p
(x, v)
p(1)
(1)/p
dx

v
Lr (,)
(2.3.26)
Let (t) be the function dened in (2.3.2), where u is replaced by v. In our

case T = max v = 1. Clearly,

v
d =

q /p
(Nt ) (t)
Nt d(t ) =
r
d(tr )
,
[(t)]q /p
where Nt = {x G : v(x) t}. Since Nt F , we have by Lemma 2.2.2/3

(Nt )(t)q
Hence
/p

v
Since [(t)]
q /p
(Nt )
q /p
.
/p
q
[(p, )-cap(Nt , G)]
q /p

v
1
q /p r
d t .
(t)
is a nonincreasing function, from (1.3.41) we obtain

r
q/p
q /p
1
q (1p)/r p
(t)
d t
(1)(rq
)/p

r/p
d(tp )
[(t)]p1
r/p
.
Setting t = t() in the last integral and applying the inequality (2.3.8) and
Lemma 2.3.1, we obtain
(1)
0
d[t()]p
c
p1
(1)
p
t ()
d = c
(x, v)
p
dx.
Thus,
v Lr (,)

q /pr
c
(1)(rq )/rp

(x, v)
(q r)/rp
q /pr
(p, )-cap(F, G)
c
p
/p
dx

(x, v)
p
/p
dx
(2.3.27)
The last inequality follows from the estimate
(1)
p1

1
(p, )-cap(F, G)
(see Lemma 2.2.2/3). Since 0 v 1 and 1, from (2.3.26) and (2.3.27)

it follows that
165
(q r)/rp
q /pr
(F )1/q cC
(p, )-cap(F, G)

[1+(1)]/p

p
(x,
v)
dx
.
G
Minimizing

p
(x, v) dx
on the set P(F, G), we obtain

q /pr
(F )1/q cC
(p, )-cap(F, G)
= cC
(p, )-cap(F, G)
q /pr
Hence
(q q)/(q r)
(F )p/q cC qp/q
Consequently,
cC pq(q
r)/q (qr)
1/p+(q r)/pr
q /qp
(p, )-cap(F, G).
= cC p/(1) .
Since is majorized by a constant that depends neither on nor G, using the

property (iv) of the (p, )-capacity we obtain cC p/(1) . The theorem is
proved.

Remark. The theorem just proved shows, in particular, the equivalence of
the multiplicative inequality (2.3.25) and the Sobolev-type inequality (2.3.19).
2.3.7 Estimate for the Norm in Lq (, ) with q q > 0
and necessary if p > q 1.
Denition. Let S = {gj }
j= be any sequence of admissible subsets of
gi , gi+1 ), and
with gi gi+1 . We put i = (gi ), i = (p, )-cap(

p/q q/(pq) (pq)/q

i
= sup
.
i
{S} i=
(2.3.28)
(The terms of the form 0/0 are considered to be zeros.)

Theorem. (i) If < , then

upLq (,) C

p
(x, u) dx,
where u D() and p > q > 0, C c.
(2.3.29)
166
(ii) If there exists a constant C such that (2.3.29) holds for all u D()
with p > q 1, then cC.
Proof. (i) Let tj = 2j + j , j = 0, 1, 2, . . . , where j is a decreasing
sequence of positive numbers satisfying j 2j 0 as j . We assume
further that the sets Ltj are admissible. Obviously,

uqLq (,) =
tj1

(Lt ) d tq c
2qj (Ltj ).
tj
j=
j=
Let gj = Ltj . We rewrite the last sum as

c
p/q q/p

pj q/p
j
2 j
j
j=
and apply Holders inequality. Then

uqLq (,)
q/p
q/p
2
pj
j=
Let C (R1 ), (t) = 1 for t 1, (t) = 0 for t 0, 0 (t) 1 + ,

( > 0) and let

|u(x)| tj+1
.
uj (x) =
tj tj+1
Since uj N(
gj , gj+1 ), it follows that
2pj j c
j=
=c
(tj tj+1 )p
j=

j=
gj+1 \gj
gj+1 \gj
(x, uj )
u tj+1
tj tj+1
p
p
dx
(x, u)
p
dx.
Letting tend to zero, we obtain
pj

j c

p
(x, u) dx.
j=
(ii) We introduce the sequence

S = {gj }
j=
and put N +1 = 0 and
k =
1/(pq)
N

j
j=k
(2.3.30)
167
for k = N, N +1, . . . , 0, . . . , N 1, N . By uk we denote an arbitrary function

in P(
gk , gk+1 ) and dene the function
uk = (k k+1 )uk + k+1
u = N
on gN ,
on gk+1 \gk ,
u = 0 on \gN +1 .
Since u D(), it satises (2.3.29). Obviously,

uq d = v

(Lt ) d tq

N

k=N
k+1
N

q
(Lt ) d tq
k kq k+1
.
k=N
q
) implies
Therefore, (2.3.29) and the inequality (k k+1 )q (kq k+1
p/q
k (k k+1 )

N

k=N
gk+1 \gk
k=N
(x, uk )
p
=C
(k k+1 )
k=N
gk+1 \gk
dx
(x, uk )
p
dx.
gk , gk+1 ), it follows by minimizing the

Since uk is an auxiliary function in P(
last sum that

p/q
(k k+1 )
k=N
(k k+1 )p k .
k=N
Putting here
1/(pq) 1/(qp)
k
,
k k+1 = k
we arrive at the result

(pq)/q
N

p/q q/(pq)

C.
k k

k=N
2.3.8 Estimate for the Norm in Lq (, ) with q < p (Second

Necessary and Sucient Condition)
Lemma. Let g1 , g2 , and g3 be admissible subsets of such that g1 g2 ,
g2 g3 . We set
gi , gj ),
ij = (p, )-cap(
168
where i < j. Then

1/(p1)
12
1/(p1)
+ 23
1/(p1)
13
gk , gk+1 ),
Proof. Let be any positive number. We choose functions uk P(
k = 1, 2, so that
1/(p1)
k,k+1
1

Ek
(x, uk )
p ds
|uk |
where Ek = {x : uk (x) = }. We put u(x) =

u(x) = (u1 (x) + 1)/2 for x g2 . Then
1
1/(p1)
d + ,
1
2 u2 (x)
for x g3 \g2 and

p ds 1/(1p)
d
(x, u1 )
|u1 |
0
E1

1

p ds 1/(1p)
=
(x, u)
d,
|u|
1/2
E

1

p ds 1/(1p)
(x, u2 )
d
|u2 |
0
E2

1/2

p ds 1/(1p)
=
(x, u)
d,
|u|
0
E
where E = {x : u(x) = }. Therefore,

1/(1p)
12
1/(1p)
23
1

p ds
(x, u)
|u|
E
1/(1p)
d + 2.
Since u P(
g1 , g3 ), by Lemma 2.2.2/3 the right-hand side of the last inequal1/(1p)
ity does not exceed 13
+ 2. The lemma is proved.

Let p be the capacity minimizing function introduced in Denition 2.3.2.
It can be easily checked that condition (2.3.15) is equivalent to
p (t) tP 1 (1/t)
and condition (2.3.18) to
p (t) tp .
The theorem of the present subsection yields the following necessary and
sucient condition for the validity of (2.3.29) with p > q > 0:
()

K=
0
p ( )
q/(pq)
d < .
(2.3.31)
169
Theorem. Let p > q > 0, p > 1.

1. If K < , then (2.3.29) holds for all u D() with C cK (pq)/q .
2. If (2.3.29) holds with C < , then (2.3.31) is valid with K (pq)/q cC.
Proof. 1. By Theorem 2.3.7 it suces to prove the inequality
q/(pq)
()
p/q q/(pq)

j
d,
sup
j
pq 0
p ( )
{S} j=
(2.3.32)
where the notation of Sect. 2.3.7 is retained.

Let the integral in the right-hand side converge, let N be a positive integer,
gj , gN +1 ) for j N , N +1 = . By the Lemma,
and let j = (p, )-cap(
1/(1p)
1/(1p)
1/(1p)
j+1
j N.
Since q(p 1)/(p q) 1, then

|a b|q(p1)/(pq) aq(p1)/(pq) bq(p1)/(pq)
and hence
q/(pq)
q/(pq)
q/(pq)
j+1
This implies
def
N =
N p/q q/(pq)
N

j
p/(pq) q/(pq)
q/(pq)
j
j+1
j
j
j=N
j=N
p/(pq)
p/(pq) q/(pq)
p/(pq) q/(pq)
j
j
j1
+ N
N
.
j=N +1
gj , ) p (j ). Since the function p does not

It is clear that j (p, )-cap(
decrease then
N
q/(pq)
d( p/(pq) )
p/(pq)
.
N
p (N )
[p ( )]q/(pq)
0
Similarly,
p/(pq)
q/(pq)
p/(pq)
j
j1
p (j )
j1
Consequently,
N
p ( )
d( p/(pq) )
.
[p ( )]q/(pq)
q/(qp) p/(pq)
d
.
The result follows.

2. With a -measurable function f we connect its nonincreasing rearrangement

(2.3.33)
f (t) = inf s : x : f (x) s t .
170
By Lemma 2.1.4/1

()
f Lq (,) =
f (t)
1/q
dt
0 < q < .
(2.3.34)
We note that the inequality (2.3.29) implies that () < and that
p (t) > 0 for all t (0, ()]. Let l be any integer satisfying 2l (). We
introduce an admissible subset gl of such that

(gl ) 2l ,
(p, )-cap(gl , ) 2p 2l .
gl , ) satisfying
By ul we denote a function in P(

p
(x, ul ) dx 4p 2l .
(2.3.35)
Let s be the integer for which 2s () < 2s+1 . We dene the function in
C00,1 ()
fr,s (x) = sup l ul (x), x ,
rls
where r < s and the values l are dened by

l 1/(pq)
2
l =
.
p (2l )
By Lemma 2.1.4/2 and by Lemma 2.1.4/3 with p instead of we have
s

p
(x, fr,s ) dx
lp
l=r
(x, ul )
p
dx.
By (2.3.35) the right-hand side is majorized by

c

lp p 2l .
l=r
Now we obtain a lower estimate for the norm of fr,s in Lq (, ). Since

fr,s (x) l on the set gl , r l s, and (gl ) 2l , the inequality

x : fr,s (x) > < 2l
implies l . Hence
fr,s
(t) l

for t 0, 2l , r l s.
(2.3.36)
By (2.3.34) and (2.3.36)

fr,s qLq (,) =
()
(fr,s ) (t)
dt c
s
s

q

lq 2l .
(fr,s ) (2l ) 2l c
l=r
l=r
171
Therefore,
s
uLq (,)
( l=r lq 2l )1/q

c
B := sup
s
( [(x, u)]p dx)1/p
( l=r lp p (2l ))1/p
s
1/q1/p
2lp/(pq)
=c
.
p (2l )q/(pq)
l=r
With r we obtain
s

B
l=

2lp/(pq)
p (2l )q/(pq)
s

2s1
c
0
l=
1/q1/p
tp/(pq)
dt
q/(pq)
t
(p (t))
1/q1/p
.
Hence by monotonicity of p we obtain

()
Bc
0
tp/(pq)
dt
(p (t))q/(pq) t
1/q1/p
.

We give a sucient condition for inequality (2.3.29) with = mn formulated in terms of the weighted isoperimetric function C introduced in Denition 2.2.3.
Corollary. If p > q > 0, p > 1, and

0
mn () mn ()
t
d
(C ())p/(p1)
q(p1)
pq
t pq dt < ,
then (2.3.29) with = mn and any u D() holds.

Proof. The result follows directly from the last Theorem and Corollary 2.7.2.

2.3.9 Inequality with the Norms in Lq (, ) and Lr (, )
(the Case p 1)
The next theorem gives conditions for the validity of the inequality

p
(x, u) dx + upLr (,)
upLq (,) C
(2.3.37)
for all u D() with q p r, p > 1 (compare with Theorem 2.1.3).
172
Theorem. Inequality (2.3.37) holds if and only if

p/r
g , G ) + (G )
(g)p/q cC (p, )-cap(
(2.3.38)
for all admissible sets g and G with g G .

Proof. Suciency. By Lemma 1.2.3 and inequality (1.3.41),

q p/q
p
(Lt ) d t
uLq (,) =
0
(Lt )

p/q p
d t c
2pj (gj )p/q ,
j=
where gj = Ltj and {tj } is the sequence of levels dened in the proof of
part (i) of Theorem 2.3.7. We set j = (p, )-cap(
gj , gj+1 ) and using the
condition (2.3.38), we arrive at the inequality:

p
pj
pj
p/r
.
(2.3.39)
uLq (,) cC
2 j +
2 (gj )
j=
j=
We can estimate the rst sum on the right-hand side of this inequality by
means of (2.3.30). The second sum does not exceed

p/r p
p/r
c
(Lt )
d t c
(Lt ) d tr
= cupLr (,) .
0
Necessity. Let g and G be admissible and let g G . We substitute any

function u P(
g , G ) into (2.3.37). Then

p
(g)p/q C
(x, u) dx + (G )p/r .
Minimizing the rst term on the right of the set P(

g , G), we obtain (2.3.38).

Remark. Obviously, a sucient condition for the validity of (2.3.37) is the
inequality

(2.3.40)
(g)p/q C1 (p, )-cap(g, ) + (g)p/r ,
which is simpler than (2.3.38). In contrast to (2.3.37) it contains only one set g.
However, as the following example shows, the last condition is not necessary.
Let = R3 , q = p = r = 2, (x, y) = |y|, and let the measures and
be dened as follows:
(A) =
(A) =

k=0

k=0
s(A B2k ),
s(A B2k +1 ),
173
where A is any Borel subset of R3 and s is a two-dimensional Hausdor

measure. The condition (2.3.40) is not fullled for these measures and the
2k 1 , k = 2, 3, . . . , we have
2-capacity. Indeed, for the sets gk = B2k +1 \B
k+1
3
, (gk ) = 0, 2-cap(gk , R ) = 4(2k + 1).
(gk ) = 4
We shall show that (2.3.37) is true. Let u D(R3 ) and let (, ) be
spherical coordinates with center O. Obviously,
k 2
2
u 2 ,
2k +1
2k
Hence

k 2
d 2
4k
u 2 ,
B1
2

2
u
(, ) d + 2 u 2k + 1, .

B2k +1 \B2k
Summing over k, we obtain

u2 d c
R3
R3
u

2

dx+24k
B1

|u|2 dx +
2
k
d.
u 2 +1,

u2 d .
R3

2.3.10 Estimate with a Charge on the Left-Hand Side
The following assertion yields a condition close in a certain sense to being
necessary and sucient for the validity of the inequality

p
(x, u) dx, u D(),
|u|p d c
(2.3.41)
where is an arbitrary charge in , not a nonnegative measure as in Theorem 2.3.6. (Theorem 2.1.3 contains a stronger result for p = 1.)
Theorem. Let + and be the positive and negative parts of the
charge .
1. If for some (0, 1) and for all admissible sets g and G with g G we
have the inequality
g , G) + (1 ) (G),
(g) C (p, )-cap(
(2.3.42)
where C = const, then (2.3.41) is valid with C cC .

2. If for all u D() inequality (2.3.41) holds, then
g , G ) + (G )
+ (g) C(p, )-cap(
(2.3.43)
for all admissible sets g and G , g G .

Proof. Let = (1 )1/2p and gj = Lj , j = 0, 1, . . . . By Lemma 1.2.3,
174

upLp (,+ )

p
(Lt ) d t =

+ (Lt ) d tp
j=
j+1

+ (Lj ) (j+1)p jp .
j=

upLp (,+ )

(p, )-cap(Lj , Lj1 ) (j+1)p jp
j=
+ (1 )

(Lj1 ) (j+1)p jp .
(2.3.44)
j=
Using the same arguments as in the derivation of (2.3.30), we obtain that

the rst sum in (2.3.44) does not exceed

p
( p 1) p
(x, u) dx.
( 1)p
Since (Lt ) is a nondecreasing function, then

(j1)p
(j2)p
j1
(Lj1 )

(Lt ) d tp
j2
and hence

(Lj1 ) (j+1)p jp 2p

(Lt ) d tp .
j=
Thus
upLp (,+ )
( p 1) p
C
( 1)p
(x, u)
p
dx + 2p (1 )upLp (,+ ) .
It remains to note that 2p (1 ) = 1.

2. The proof of the second part of the theorem is the same as that of
necessity in Theorem 2.3.9. The theorem is proved.

2.3.11 Multiplicative Inequality with the Norms in Lq (, ) and
Lr (, ) (Case p 1)
The following assertion gives a necessary and sucient condition for the validity of the multiplicative inequality

p
p(1)
p
(2.3.45)
(x, u) dx uLr (,)
uLq (,) C
175
for p 1 (cf. Theorem 2.1.1).

Theorem. 1. Let g and G be any admissible sets such that g G . If a
constant exists such that

g , G ) (G )(1)p/r ,
(2.3.46)
(g)p/q (p, )-cap(
then (2.3.45) holds for all functions u D() with C c, 1/q (1 )/r +
/p, r, q > 0.
2. If (2.3.45) is true for all u D(), then (2.3.46) holds for all admissible
sets g and G such that g G . The constant in (2.3.46) satises C.
Proof. 1. By Lemma 1.2.3 and inequality (1.3.41),

1/

q 1/q
1/
/q 1
uLq (,) =
(L ) d
(L )
d
,
0
0
1
where = pr[p(1 ) + r] , q. Consequently,

p/

p
2j (gj )/q
uLq (,) c
j=

c
(p, )-cap(
gj , gj+1 )
/p
p/
(gj+1 )
(1)/r
j=
where gj = Ltj and {tj } is the sequence of levels dened in the proof of the
rst part of Theorem 2.3.7. Hence,

p
pj
2 (p, )-cap(
gj , gj+1 )
uLq (,) c
j=
(1)p/r
rj
(gj+1 )
(2.3.47)
j=
By (2.3.30),
2pj (p, )-cap(

gj , gj+1 ) c
p
(x, u)
dx.
j=
Obviously, the second sum in (2.3.47) does not exceed curLr (,) . Thus
(2.3.45) follows.
2. Let g and G be admissible sets with g G . We substitute any function
u P(
g , G ) into (2.3.45). Then

p
p/q
C
(x, u) dx (G )(1)p/r ,
(g)
which yields (2.3.46). The theorem is proved.
176
2.3.12 On Nash and Moser Multiplicative Inequalities

An important role in Nashs classical work on local regularity of solutions to
second-order parabolic equations in divergence form with measurable bounded
coecients [625] is played by the multiplicative inequality

1+2/n
Rn

C
u2 dx
Rn
|u|2 dx
Rn
4/n
|u| dx
u C0 .
(2.3.48)
Another inequality of a similar nature

1+2/n
2/n
uL2+4/n cuL2 uL2 ,
u C0 ,
(2.3.49)
was used by Moser in his proof of the Harnack inequality for solutions of
second-order elliptic equations with measurable bounded coecients in divergence form [617].
These two inequalities are contained as very particular cases in the
GagliardoNirenberg inequality for all u C0 (Rn )
1
j uLq cl u
Lp uLr ,
(2.3.50)
where 1 p, r , 0 j < l, j/l , 1, and

1
j
1
l
1
= +
+
.
q
n
p n
r
If 1 2, the Nash and Moser inequalities follow directly by the Holder
inequality from the Sobolev inequality
uL2n/(n2) cuL2 ,
u C0 .
(2.3.51)
We know by the second part of Theorem 2.3.6 that conversely, (2.3.48) and
(2.3.49) imply (2.3.51). The just-mentioned theorem does not contain (2.3.48)
and (2.3.49) for n = 2, which formally corresponds to the exceptional case
= 0. However, we show here that both (2.3.48) and (2.3.49) with n = 2,
and even the more general inequality

|u| dx c
q
Rn
Rn
qr

n
|u| dx
n
Rn
|u|r dx,
(2.3.52)
where q r > 0 can be deduced from Theorem 2.3.11. In fact, by this theorem,
(2.3.52) holds if and only if

qr
g , G) n mn (G),
mn (g) const capn (
(2.3.53)
177
where g and G are arbitrary bounded open sets with smooth boundaries,
g G, and capn is the n capacity of g with respect to G. By the isocapacitary
inequality (2.2.11),

1n
mn (G)
capn (
g , G) nn1 n log
.
mn (g)
Hence (2.3.53) is a consequence of the boundedness of the function
(qr)(n1)

n
1
(0, 1) x x log
,
x
which, in its turn, implies the multiplicative inequality (2.3.52).
The original proof of (2.3.52) (see Nirenberg [640], p. 129) is as follows.

First, one notes that it suces to obtain (2.3.52) for large q. Then (2.3.52) results by putting |u|q(1n)/n instead of u into (1.4.49) and using an appropriate
Holders inequality.
Extensions of Nashs inequality (2.3.48) to weighted inequalities with indenite weights on the left-hand side were obtained by Mazya and Verbitsky
[594] with simultaneously necessary and sucient conditions on the weights.
The basic results of Sects. 2.3.12.3.4 were obtained by the author in [531,
534] for p = 2, (x, ) = ||, M (u) = |u|, and in [543] for the general case.
Some of these results were repeated by Stredulinsky [729]. We shall return to
capacitary inequalities similar to (2.3.6) in Chaps. 3 and 11. The inequality
(2.3.14) can be found also in Brezis and Wainger [146] and Hansson [348].
Regarding the criterion in Sect. 2.3.3, see Comments to Sect. 2.4, where
other optimal embeddings of BirnbaumOrliczSobolev spaces into C and
BirnbaumOrlicz spaces are discussed.
Inequality (2.3.21) is (up to a constant) the Sobolev (p > 1)-Gagliardo
Nirenberg (p = 1) inequality. The best constant for the case p = 1 (see (1.4.14))
was found independently by Federer and Fleming [273] and by the author [527]
using the same method.
The best constant for p > 1, presented in Sect. 2.3.5 was obtained by
Aubin [55] and Talenti [740] (the case n = 3, p = 2 was considered earlier
by Rosen [682]), whose proofs are a combination of symmetrization and the
one-dimensional Bliss inequality [109] (see Sect. 2.3.5). The uniqueness of the
Bliss optimizer was proved by Gidas, Ni, and Nirenberg [307].
A dierent approach leading to the best constant in the Sobolev inequality,
which is based on the geometric BrunnMinkowskiLyusternik inequality, was
proposed in Bobkov and Ledoux [118].
178
The extremals exhibited in (2.3.24) of the Sobolev inequality (2.3.21) in

the whole of Rn , with sharp constant C, are the only onessee CorderoErausquin, Nazaret, and Villani [212] who used the mass transportation techniques referred to in Comments to Sect. 1.4. Strengthened, quantitative versions of this inequality are also available. They involve a remainder term
depending on the distance of the function u from the family of extremals. The
rst result in this connection was established by Bianchi and Egnell [96] for
p = 2. The case when p = 1 was considered in Cianchi [199] and sharpened
in Fusco, Maggi, and Pratelli [296] as far as the exponent in the remainder
term is concerned. The general case when 1 n are contained in
Cianchi [202].
In [811], Zhang proved an improvement of the inequality (1.4.14), called
the L1 ane Sobolev inequality,

n

n
n
u f L1 dsu n
f n
(2.3.54)
L n ,
2
n1
n1
n1
S
where u f is the partial derivative of f in direction u, dsu is the surface measure on S n1 and the constant factor on the right-hand side is sharp. Modications of (2.3.54) for the Lp -gradient norm with p > 1 and for the Lorentz
and BirnbaumOrlicz settings are due to Zhang [811]; Lutwak, D. Yang, and
Zhang [510]; Haberl and Schuster [333]; Werner and Ye [794]; and Cianchi,
Lutwak, D. Yang, and Zhang [206].
A Sobolev-type trace inequality, which attracted much attention, is the
following trace inequality:
f L p(n1) (Rn+ ) Kn,p f Lp (Rn+ ) ,
(2.3.55)
np
where n > p > 1. In the case p = 2, Beckner [78] and Escobar [259], using
dierent approaches, found the best value of Kn,2 . Xiao [799] generalized their
result to the inequality

f (x)2 x12 dx,
f L 2(n1) (Rn+ ) C(n, )
(2.3.56)
n
Rn
+
n12
showing that

C(n, ) =
214
(2(1 ))

((n 1 2)/2)
((n 1 + 2)/2)

(n 1)
((n 1)/2)
2
n1
An idea in [799] is that it suces to prove (2.3.56) for solutions of the

Euler equation

div xn12 u = 0 on Rn+ .
Then by the Fourier transform with respect to x = (x1 , . . . , xn1 ) the integral
on the right-hand side of (2.3.56) takes the form
2.4 Continuity and Compactness of Embedding Operators

2
const (x )/2 uL
2 (R
n1 )
179
and the reference to the Lieb formula (1.4.48) gives the above value of C(n, ).
More recently, Nazaret proved, by using the mass transportation method
mentioned in Comments to Sect. 1.4, that the only minimizer in (2.3.55) has
the form

pn
const (xn + )2 + |x|2 2(p1) ,
where = const > 0. Sharp Sobolev-type inequalities proved to be crucial in
the study of partial dierential equations and nowadays there is extensive literature dealing with them. To the works mentioned earlier we add the papers:
Gidas, Ni, and Nirenberg [307]; Lieb [496]; Lions [501]; Han [335]; Beckner [78,
79]; Adimurthi and Yadava [29]; Hebey and Vaugon [362, 363]; Hebey [359];
Druet and Hebey [243]; Lieb and Loss [497]; Del Pino and Dolbeault [231];
Bonder, Rossi, and Ferreira [125]; Biezuner [99]; Ghoussoub and Kang [306];
Demyanov and A. Nazarov [233]; Bonder and Saintier [126]; et al.
The study of minimizers in the theory of Sobolev spaces based on the socalled concentration compactness is one of the topics in the book by Tintarev
and Fieseler [753] where relevant historical information can be found as well.
The material of Sects. 2.3.62.3.11 is due to the author [543]. The suciency in Theorem 2.3.8 can be found in Mazya [543] and the necessity is due
to Mazya and Netrusov [572].
The equivalence of the Nash and Moser inequalities (2.3.48) and (2.3.49)
for n > 2 and Sobolevs inequality (2.3.51) is an obvious consequence of
Theorem 2.3.6, which was proved by the author [543] (see also [552], Satz 4.3).
This equivalence was rediscovered in the 1990s by Bakry, Coulhon, Ledoux,
and Salo-Coste [64] (see also Sect. 3.2 in the book by Salo-Coste [687]) and
by Delin [232]. The best constant in (2.3.48) was found by Carlen and Loss
[167]:
C = 2n1+2/n (1 + n/2)1+n/2 zn1 n2/n ,
where zn is the smallest positive root of the equation

(1 + n/2)J(n2)/2 (z) + zJ(n2)/2
(z) = 0.
The existence of the optimizer is proved in Tintarev and Fieseler [753], 10.3.
2.4 Continuity and Compactness of Embedding

1 () into BirnbaumOrlicz
Operators of
L1p() and W
p
Spaces
pl () be completions of D() with respect to the norms
Let
Llp () and W
l uLp () and l uLp () + uLp () .
Let be a measure in . By LM (, ) we denote the BirnbaumOrlicz
space generated by a convex function M , and by P we mean the complementary function of M (see Sect. 2.3.3).
180
The present section deals with some consequences of Theorem 2.3.3, containing the necessary and sucient conditions for boundedness and compact 1 () into the space
ness of embedding operators which map
L1p () and W
p
1/p
Lp,M (, ) with the norm |u|p LM (,) , where is a measure in . In the
case p = 2, M (t) = |t| these results will be used in Sect. 2.5 in the study of
the Dirichlet problem for the Schr
odinger operator.
2.4.1 Conditions for Boundedness of Embedding Operators
With each compactum F we associate the number

(F )P 1 (1/(F ))
for capp (F, ) > 0,
capp (F,)
p,M (F, ) =
0
for capp (F, ) = 0.
In the case p = 2, M (t) = |t|, we shall use the notation (F, ) instead of
p,M (F, ).
The following assertion is a particular case of Theorem 2.3.3.
Theorem 1. 1. Suppose that
p,M (F, )
for any compactum F . Then, for all u D(),

p
|u|
C
|u|p dx,
L (,)
M
(2.4.1)
where C pp (p 1)1p .
2. If (2.4.1) is valid for all u D(), then p,M (F, ) C for all compacta F .
Using this assertion we shall prove the following theorem.
p
|u|
LM (,)

|u|p + |u|p dx,
(2.4.2)
where p < n is valid for all u D() if and only if, for some > 0,

(2.4.3)
sup p,M (F, ) : F , diam(F ) < ,
where, as usual, F is a compact subset of .
Proof. Suciency. We construct a cubic grid in Rn with edge length c,
where c is a suciently small number depending only on n. With each cube
Qi of the grid we associate a concentric cube 2Qi with double the edge length
and with faces parallel to those of Qi . We denote an arbitrary function in
181
D() by u. Let i be an innitely dierentiable function in Rn that is equal

to unity in Qi , to zero outside 2Qi , and such that |i | c0 /.
By Theorem 1,

|ui |p
LM (,)

(F )P 1 (1/(F ))
(ui )p dx
c sup
: F 2Qi
capp (F, 2Qi )
2Qi

(ui )p dx.
c sup p,M (F, ) : F , diam(F )
2Qi
Summing over i and noting that

p

p
|u|

|u
|
i

LM (,)
LM (,)

|ui |p
,
LM (,)
i
we obtain the required inequality

p

|u|
c sup p,M (F, ) : F , diam(F )
LM (,)

|u|p + p |u|p dx.
(2.4.4)
Necessity. Let F be any compactum in and let diam(F ) < 1. We

include F inside two open concentric balls B and 2B with diameters and 2,
respectively. Then we substitute an arbitrary u P(F, 2B ) into (2.4.2).
Since u = 1 on F , then by (2.4.2)

|u|p dx +
|u|p dx .
F LM (,) C
2B
2B
Consequently,
(F )P

1/(F ) C 1 + c p

|u|p dx.
2B
Minimizing the last integral over the set P(F, 2B ) we obtain

(F )P 1 1/(F ) C 1 + c p capp (F, 2B ).
It remains to note that since p < n, it follows that
capp (F, 2B ) c capp (F, ),
(2.4.5)
where c depends only on n and p.

In fact, if u N(F, ) and D(2B), = 1 on B, || c, then
u N(F, 2B) and hence
182

(u)p dx
capp (F, 2B )

|u| dx +
|u| dx
c
2B

2B
p
p
c
|u| dx + uLpn/(np) () .
2B

This and the Sobolev theorem imply (2.4.5). The theorem is proved.
2.4.2 Criteria for Compactness

The following two theorems give the necessary and sucient conditions for
1 () into
the compactness of embedding operators that map
L1p () and W
p
Lp,M (, ).
Theorem 1. The conditions

lim sup p,M (F, ) : F , diam(F ) = 0
0
and

lim sup p,M (F, ) : F \B = 0
(2.4.6)
(2.4.7)
are necessary and sucient for any set of functions in D(), bounded in
L1p () (p < n), to be relatively compact in Lp,M (, ).

Theorem 2. The condition (2.4.6) and
lim sup{p,M (F, ) : F \B , diam(F ) 1} = 0
(2.4.8)
are necessary and sucient for any set of functions in D(), bounded in
1 () (p < n), to be relatively compact in Lp,M (, ).
W
p
To prove Theorems 1 and 2 we start with the following auxiliary assertion.
Lemma. Let () be the restriction of to the ball B . For an arbi 1 (), p < n, to be relatively compact
trary set, bounded in
L1p () or in W
p
()
in Lp,M (, ) for all > 0, it is necessary and sucient that

(2.4.9)
lim sup p,M (F, ) : F B , diam(F ) = 0,
0
for any > 0.

Proof. Suciency. Since capacity does not increase under the extension
of , we see that for any compactum F B ,
p,M (F, B ) p,M (F, ).
183
This along with (2.4.9) implies

lim sup p,M (F, B ) : F B , diam(F ) = 0
0
for all > 0. This equality, together with (2.4.4), where the role of is played
by B2 , yields

p
p
|u|
|u| dx + C1 ()
|u|p dx
L (,(2) )
M
B2
B2
for any > 0 and for all u D(B2 ). Replacing u by u, where is a

truncating function, equal to unity on B and to zero outside B2 , we obtain

p
p
|u|
|u| dx + C2 ()
|u|p dx.
(2.4.10)
L (,() )
M
B2
It remains to note that in the case p < n any set, bounded in

L1p () (and
1
a fortiori in Wp ()), is compact in Lp (B ) for any > 0. The suciency

of (2.4.8) is proved.
Necessity. Let F B be a compactum and let diam(F ) < 1. We
include F inside concentric balls B and 2B with radii and 2, respectively. By
u we denote an arbitrary function in P(F, 2B ). Since any set of functions
p1 (), is relatively compact in Lp,M (, () ), then for
in D(), bounded in W
all v D()

B |v|p
()
|v|p + |v|p dx,
LM (,() )
where B is the characteristic function of B and () 0 as 0. To prove

this inequality we must note that Theorem 2.4.1/2, applied to the measure
() , implies () (2B) 0 as 0. Since u equals zero outside 2B we
have

|u|p dx c p
|u|p dx.
Therefore,

(F )P 1 1/(F ) 1 + c p ()

|u|p dx.
2B
Minimizing the last integral over P(F, 2B ) and using (2.4.5), we arrive at

p,M (F, ) 1 + c p ().
The necessity of (2.4.9) follows. The lemma is proved.
Proof of Theorem 1. Suciency. Let C (Rn ), 0 1, || c1 ,

= 0 in a neighborhood of B/2 , = 1 outside B . It is clear that
184

p 1/p
1/p
1/p
|u|
(1 )p |u|p L (,) + p |u|p L (,)
LM (,)
M
M

p 1/p

p 1/p

|u| L (,() ) + |u| L (,) .
(2.4.11)
M
/2 , by (2.4.7)
By the rst part of Theorem 2.4.1/1, applied to the set \B
and the inequality
/2 ) p,M (F, ),
p,M (F, \B
given any , there exists a number > 0 such that

|u|p 1/p
(u)Lp () .
LM (,)
Since || c1 c|x|1 and
1
|x| u
L
we have
p ()
cuLp () ,

|u|p 1/p
cuLp () .
LM (,)
The last inequality along with (2.4.11) implies

1/p

p 1/p
|u|
|u|p L (,() ) + cuLp () .
LM (,)
M
(2.4.12)
Obviously, (2.4.6) implies (2.4.9). Therefore, the lemma guarantees that any
set of functions in D(), bounded in
L1p (), is compact in Lp,M (, () ). This
together with (2.4.12) completes the proof of the rst part of the theorem.
Necessity. Let F be a compactum in with diam(F ) < 1. Duplicating
the proof of necessity in the Lemma and replacing () there by , we arrive
at the inequality p,M (F, ) (1 + cp )() and hence at (2.4.6).
. Using the compactness in Lp,M (, ) of any set of
Now let F \B
functions in D(), which are bounded in
L1p (), we obtain

\B |u|p 1/p
uLp () ,

L (,)
M
where 0 as 0 and u is an arbitrary function in D(). In particular,

the last inequality holds for any u P(F, ) and therefore

(F )P 1 1/(F ) p upLp () .
Minimizing the right-hand side over the set P(F, ), we arrive at (2.4.7). The
theorem is proved.

Proof of Theorem 2. We shall use the same notation as in the proof of
Theorem 1.
/2 ,
Suciency. From (2.4.4), where = 1 and is replaced by \B
together with (2.4.8), it follows that given any > 0, there exists a > 0 such
that

|u|p 1/p
LM (,)

(u)L
p ()
185

+ uLp () .
This together with (2.4.11) yields

1/p

p 1/p
|u|
|u|p L (,() ) + cuWp1 () .
LM (,)
M
The remainder of the proof is the same as the proof of suciency in the
preceding theorem.
Necessity. The condition (2.4.6) can be derived in the same way as in the
proof of necessity in Theorem 1.
, > 8, diam(F ) 1. From the compactness in Lp,M (, )
Let F \B
1 (), it follows that
of any set of functions in D(), bounded in W
p

\B |u|p
upW 1 () ,
/2
L (,)
M
where 0 as and u is an arbitrary function in D(). We include

F inside concentric balls B and 2B with radii 1 and 2 and let u denote any
function in P(F, 2B ). Using the same argument as in the proof of necessity
in the Lemma we arrive at
p,M (F, ) (1 + c) ,
which is equivalent to (2.4.8). The theorem is proved.
Remark. Let us compare (2.4.6) and (2.4.9). Clearly, (2.4.9) results from
(2.4.6). The following example shows that the converse assertion is not valid.
Consider a sequence of unit balls B () ( = 1, 2, . . . ), with dist (B () , B () )
1 for = . Let = Rn and

p(x) dx,
(F ) =
F
where
p(x) =
2+
0
for x B () ,
()
for x
/
.
=1 B
Here is the distance of x from the center of B () .

We shall show that the measure satises the condition (2.4.9) with p = 2,
M (t) = t. First of all we note that for any compactum F B ()

r(F )
(F ) =
2+1/ dx
n3+1/ d d,
F
B1
where

r(F ) =
n
mn (F )
1/n
.
To estimate cap(F ), i.e., cap2 (F, Rn ), we apply the isoperimetric inequality (2.2.12)
186
n1 (n
Now
2)
(n2)/n
n
cap(F )
mn (F )
n

F, Rn

n2
= r(F )
.
r(F )1/
,
(n 2)(n 2 + 1/)

If F is the ball {x : }, we have

F, Rn =
1/
.
(n 2)(n 2 + 1/)
Consequently,

sup F, Rn : F Rn , diam(F ) 2
1/
= (n 2)2
(n 2)(n 2 + 1/)
lim
and (2.4.6) is not valid.

The material of this section is borrowed from Sect. 2.5 of the authors book
[552]. Sharp embeddings of BirnbaumOrliczSobolev spaces of order one into
the space L () will be considered in Chap. 7 of the present book (see also
Mazya [528, 545]).
An optimal Sobolev embedding theorem in BirnbaumOrlicz spaces was
established by Cianchi in [194], and in alternative equivalent form, in [195].
A basic version of this result states that if is an open set in Rn with nite
measure, M is any Young function, and Mn is the Young function given by

Mn (t) = M H 1 (t) for t > 0,
where
s
H(s) =
0
t
M (t)
1
n1
n1
n
dt
for s 0,
then there exists a constant C, depending on n, such that

uLMn () CuLM ()
for every weakly dierentiable function u vanishing, in the appropriate sense,
on . Moreover, the BirnbaumOrlicz space LMn () is optimal. Analogous
results for functions that need not vanish on , and for domains with
innite measure [194]. The case of higher-order derivatives was dealt with by
Cianchi in [200].
187
Some necessary and sucient conditions for embeddings of Sobolev-type

spaces into BirnbaumOrlicz spaces will be treated in Chap. 11 of this book.
Analog of certain results in the present section were obtained by Klimov
[431435] for the so-called ideal function spaces, for which the multiplication
by any function with |(x)| 1 a.e. is contractive.
A few words on the so-called logarithmic Sobolev inequalities. Let be a
measure in , () = 1, p 1 and let p be the capacity minimizing function
generated by (see Denition 2.3.2). The inequality

1 1
+ 1
d 4uLp () exp
log
log p (s) ds
(2.4.13)
exp
|u|
p 0
for all u
L1p () was proved in 1968 by Mazya and Havin [568]. It shows, in
particular, that
1
p (s) ds = +
0
implies
log+
0
1
d = +
|u|
for all u L1p (). This fact allows for certain applications of (2.4.13) to
complex function theory [568] (see also Sect. 14.3 of the present book for
another logarithmic inequality of a similar nature).
Inequality (2.4.13) is completely dierent from the logarithmic Sobolev
inequality obtained in 1978 by Weissler [793],

2
4
2
|u| log |u| dx
|u|2 dx,
exp
n Rn
en Rn
where uL2 (Rn ) = 1, which is equivalent (see Beckner and Pearson [81]) to
the well-known Gross inequality of 1975 [327],
!

u2 log u2
u2 d d C
|u|2 d,
(2.4.14)
Rn
where
Rn
Rn

d = (2)n/2 exp |x|2 /2 dx.
Various extensions, proofs, and applications of (2.4.14) were the subject of

many studies: R.A. Adams [24]; Stroock and Zegarlinski [730]; Holley and
Stroock [381]; Davies [222]; Zegarlinski [810]; Beckner [77, 80]; Gross [328];
Aida, Masuda, and Shigekawa [32]; Aida and Stroock [33]; Bakry [63]; Bakry,
Ledoux, and Qian [65]; Chen [185]; F.-Y. Wang [787790]; Bodineau and Helffer [121]; Bobkov and G
otze [114]; Ledoux [483485]; Yosida [808]; Guionnet
and Zegarlinski [330]; Xiao [799]; Lugiewicz and Zegarlinski [509]; Otto and
Rezniko [655]; Inglis and Papageorgiou [398]; Cianchi and Pick [207]; Martin
and M. Milman [521]; et al.
188
2.5 Structure of the Negative Spectrum of the

Multidimensional Schr
odinger Operator
In this section we show how the method and results of Sect. 2.4 can be applied
to the spectral theory of the Schr
odinger operator.
2.5.1 Preliminaries and Notation
We start with some denitions from the theory of quadratic forms in a Hilbert
space H. Let L be a dense linear subset of H and let S[u, u] be a quadratic
form dened on L . If there exists a constant such that for all u L
S[u, u] u2H ,
(2.5.1)
then the form S is called semibounded from below. The largest constant
in (2.5.1) is called the greatest lower bound of the form S and is denoted by
(S). If (S) > 0, then S is called positive denite. For such a form the set L
is a pre-Hilbert space with the inner product S[u, u]. If L is a Hilbert space
the form S is called closed. If any Cauchy sequence in the metric S[u, u]1/2
that converges to zero in H also converges to zero in the metric S[u, u]1/2 ,
then S is said to be closable. Completing L and extending S by continuity
onto the completion L, we obtain the closure S of the form S.
Now, suppose that the form S is only semibounded from below. We do not
assume (S) > 0. Then for any c > (S) the form
S[u, u] + c[u, u]
(2.5.2)
is positive denite. By denition, S is closable if the form (2.5.2) is closable

for some, and therefore for any, c > (S). The form S + cE cE is called the
closure S of S.
It is well known and can be easily checked that a semibounded closable
for which
form generates a unique self-adjoint operator S,
u) = S[u, u]
(Su,
for all u L .
Let be an open subset of Rn , n > 2, and let h be a positive number.

We shall consider the quadratic form

Sh [u, u] = h
|u|2 dx |u|2 d(x)
dened on D().
We shall study the operator Sh generated by the form Sh [u, u] under the
condition that the latter is closable. If the measure is absolutely continuous
with respect to the Lebesgue measure mn and the derivative p = d/dmn is
locally square integrable, then the operator Sh is the Friedrichs extension of
the Schr
odinger operator h p(x).
2.5 Structure of the Negative Spectrum of the Multidimensional
189
In this section, when speaking of capacity, we mean the 2-capacity and use
the notation cap.
Before we proceed to the study of the operator Sh we formulate two lemmas
on estimates for capacity that will be used later. For the proofs of these lemmas
see the end of the section.
Lemma 1. Let F be a compactum in Br . Then for R > r

1/2
2r
log Rer ) cap(F, ) for n = 3,
(1 + Rr
cap(F, Br )
r
2
for n > 3.
(1 + n3
Rr ) cap(F, )
R . Then for r < R
Lemma 2. Let F be a compactum in \B

r
r ) 1 + 1
cap(F, ).
cap(F, \B
n2Rr
All the facts concerning the operator Sh will be formulated in terms of the
function

(F )
for cap(F, ) > 0,
(F, ) = cap(F,)
0
for cap(F, ) = 0,
which is a particular case of the function p,M (F, ), introduced in Sect. 2.4,
for M (t) = |t|, p = 2.
2.5.2 Positivity of the Form S1 [u, u]
The following assertion is a particular case of Theorem 2.4.1/1.
Theorem. 1. If for any compactum F
(F, ) ,
(2.5.3)
then for all u D()

|u|2 (dx) C
|u|2 dx,
(2.5.4)
where C 4.
2. If (2.5.4) holds for all u D(), then for any compactum F
(F, ) C.
(2.5.5)
Corollary. If
sup (F, ) <
1
,
4
190
then the form S1 [u, u] is positive, closable in L2 (), and hence it generates a
self-adjoint positive operator S1 in L2 ().
Proof. The positiveness of S1 [u, u] follows from the Theorem. Moreover,
inequality (2.5.4) implies
"
#
|u|2 dx.
(2.5.6)
S1 [u, u] 1 4 sup (F, )
F
Let {u }1 , u D(), be a Cauchy sequence in the metric S1 [u, u]1/2

and let u converge to zero in L2 (). Then by (2.5.6), u converges to zero
in
L12 () and it is a Cauchy sequence in L2 (, ). Since

2
|u | d 4 sup (F, )
|u |2 dx,
then u 0 in L2 (, ). Thus, S1 [u , u ] 0 and therefore the form S1 [u, u]

is closable in L2 (). The corollary is proved.
We note that close necessary and sucient conditions for the validity of
the inequality

|u|2 d C
|u|2 dx, u D(),
where is an arbitrary charge in , are contained in Theorem 2.3.10 for

(x, y) = |y|, p = 2. The conditions in question coincide for 0. They
become the condition sup{(F, ) : F } < , which follows from the
Theorem.
2.5.3 Semiboundedness of the Schr
odinger Operator
Theorem. 1. If

1
lim sup (F, ) : F , diam(F ) < ,
0
4
(2.5.7)
then the form S1 [u, u] is semibounded from below and closable in L2 ().
2. If the form S1 [u, u] is semibounded from below in L2 (), then

(2.5.8)
lim sup (F, ) : F , diam(F ) 1.
0
Proof. 1. If is a suciently large integer, then there exists > 0 such

that

1 1 n
.
(2.5.9)
sup (F, ) : F , diam(F )
4 +2
191
We construct a cubic grid in Rn with edge length H = /( + 2) n.

(1)
(2)
We include each cube Qi of the grid inside concentric cubes Qi and Qi
(1)
(2)
with faces parallel to those of Qi . Let the edge lengths of Qi and Qi be
(2)
( + 1)H and ( + 2)H, respectively. Since diam(Qi ) = then for any
(2)
compactum F Qi

n

1 1
(2)
(F, )
.
(2.5.10)
F, Qi
4 +2
Let u denote an arbitrary function in D() and let denote an innitely
(1)
dierentiable function on Rn which is equal to unity in Qi and to zero
(2)
outside Qi . By (2.5.10) and Theorem 2.5.2 we have
n

1
(u)2 dx.
|u|2 d
(2)
(2)
+2
Qi
Qi
This implies

|u| d
2
(1)
Qi
1
+2
n

c1
2
2
|u| + 2 |u| dx.
(2)
H
Qi
(2)
Summing over i and noting that the multiplicity of the covering {Qi }
(2)
does not exceed ( + 2)n and that of {Qi } is not less than n , we obtain

2
|u|2 d 1 n
|u|2 + c 2 |u|2 dx.
(2.5.11)
Thus, the form S1 [u, u] is semibounded. Moreover, if K is a suciently

large constant, then

|u|2 dx
|u|2 dx, > 0.
S1 [u, u] + K
Further, using the same argument as in the proof of Corollary 2.5.2, we can
easily deduce that the form S1 [u, u] is closable in L2 ().
2. Let F be an arbitrary compactum in with diam(F ) < 1. We
enclose F in a ball B with radius and construct the concentric ball B with
radius 1/2 .
We denote an arbitrary function in P(F, B ) by u. In virtue of the
semiboundedness of the form S1 [u, u] there exists a constant K such that

u2 d
(u)2 dx + K
u2 dx.
B
B
B
Obviously, the right-hand side of this inequality does not exceed
192

1 + K1

(u)2 dx,
B
where is the rst eigenvalue of the Dirichlet problem for the Laplace operator
in the unit ball.
Minimizing the Dirichlet integral and taking into account that u = 1 on F ,
we obtain

(F ) 1 + K1 cap(F, B ).
By Lemma 2.5.1,

cap(F, B ) 1 + o(1) cap(F, ),
where o(1) 0 as 0. Hence

sup (F, ) : F , diam(F ) 1 + o(1).
It remains to pass to the limit as 0. The theorem is proved.
The two assertions stated in the following are obvious corollaries of the
Theorem. The second is a special case of Theorem 2.4.1/2.
Corollary 1. The condition

lim sup (F, ) : F , diam(F ) = 0
0
(2.5.12)
is necessary and sucient for the semiboundedness of the form Sh [u, u] in

L2 () for all h > 0.

2
|u|2 + |u|2 dx,
|u| d C
where u is an arbitrary function in D() and C is a constant independent

of u, is valid if and only if

sup (F, ) : F , diam(F ) <
(2.5.13)
for some > 0.
We shall give an example that illustrates an application of Theorem 2.5.2
and the theorem of the present subsection to the Schrodinger operator generated by a singular measure.
Example. Let M be a plane Borel subset of R3 . We dene the measure
(F ) = m2 (F M ) for any compactum F R3 . (In the sense of distribution
theory the potential p(x) is equal to the Dirac function concentrated on the
plane set M .) Then
193
m2 (F M )

m2 (F M )
.
F, R3 =
cap(F )
cap(F M )
Since
1/2

cap(F M ) 8 1/2 m2 (F M )
(cf. (2.2.16)), we have

1/2
1/2

m2 (F M )
.
F, R3
8
(2.5.14)
By Theorem 2.5.2 the form

S1 [u, u] =
R3

|u|2 dx
|u|2 m2 (dx)
M
is positive if m2 (M ) 4 . Using Corollary 1, from (2.5.14) we obtain that

the form Sh [u, u] is semibounded and closable in L2 (R3 ) for all h > 0 for any
plane set M .
2.5.4 Discreteness of the Negative Spectrum
Let be a xed positive number and let () be the restriction of a measure
to the ball B = {x : |x| < }. Further, let = () .
To exclude the inuence of singularities of the measure , which are located
at a nite distance, we shall assume that any subset of D(), bounded in
21 () (or in
W
L12 ()), is compact in L2 (() ). In Lemma 2.4.2 it is shown
that this condition is equivalent to

(2.5.15)
lim sup (F, ) : F B , diam(F ) = 0
0
for any > 0.

Now we formulate two well-known general assertions that will be used in
the following.
Lemma 1. (Friedrichs [292]). Let A[u, u] be a closed quadratic form in a
Hilbert space H with the domain D[A], (A) being its positive greatest lower
bound. Further, let B[u, u] be a real form, compact in D[A]. Then the form
A B is semibounded from below in H and closed in D[A], and its spectrum
is discrete to the left of (A).
Lemma 2. (Glazman [309]). For the negative spectrum of a self-adjoint
operator A to be innite it is necessary and sucient that there exists a linear
manifold of innite dimension on which (Au, u) < 0.
Now we proceed to the study of conditions for the spectrum of the
Schrodinger operator to be discrete.
Theorem. Let the condition (2.5.15) hold.
194
1. If

1
lim lim sup (F, ) : F \B , diam(F ) < ,

4
(2.5.16)
then the form S1 [u, u] is semibounded from below closable in L2 (), and the
negative spectrum of the operator S1 is discrete.
2. If the form S1 [u, u] is semibounded from below and closable in L2 (),
and the negative spectrum of the operator S1 is discrete, then

(2.5.17)
lim lim sup (F, ) : F \B , diam(F ) 1.

Proof. 1. We show that the form S1 [u, u] is semibounded from below and
closable in L2 (), and that for any positive the spectrum of the operator
S1 +2I is discrete to the left of . This will yield the rst part of the theorem.
By (2.5.16), there exists a suciently large integer such that

1 2 n
lim sup (F, ) : F \B , diam(F )

4 +2
for all > 0.
Given any , we can nd a suciently large number = () so that

1 1 n

.
sup (F, ) : F \B , diam(F )
4 +2
Hence

sup
() (F )
: F , diam(F )
cap(F, )

n
1 1
.
4 +2
If we replace () here by , then we obtain the condition (2.5.9), which was
used in the rst part of Theorem 2.5.3 for the proof of inequality (2.5.11). We
rewrite that inequality, replacing by () :

2
|u|2 + c 2 |u|2 dx.
(2.5.18)
|u|2 d() 1 n
Let denote an arbitrary positive number. We specify > 0 by the equality c 2 (1 n ) 2 = and nd corresponding to . Then

|u|2 d() 1 n
|u|2 dx +
|u|2 dx.
Hence the form

|u|2
A[u, u] =

|u|2 d() + 2
|u|2 dx,
majorizes
n
195

|u|2 dx +
|u|2 dx.
This means that the form A[u, u] has a positive lower bound and is closable in
u] denote the closure of the form A[u, u]. Clearly, the domain
L2 (). Let A[u,
u] coincides with W
21 ().
of the form A[u,
By (2.5.15) and Corollary 2.5.3/2, the form

|u|2 d()
B[u, u] =
1 (). Lemma 2.4.2 ensures the

is continuous in
and is closable in W
2
1
compactness of the form B[u, u] in W2 (). It remains to apply Lemma 1 to

u] and B[u,
u].
A[u,
2. Suppose that

lim sup (F, ) : F \B , diam > 1 + , > 0,
W21 ()
for some . Then there exists a sequence of compacta F with diam(F ) ,

which tends to innity and satises
(F ) > (1 + ) cap(F , ).
()
B
(2.5.19)
()
B
We include F in a ball
with radius . Let
denote a concentric ball
with a suciently large radius that will be specied later. Without loss of
()
generality, we may obviously assume that the balls B are disjoint.
By Lemma 2.5.1/1,

cap F , B() 1 + () cap(F , ),
where () 0 as . This and (2.5.19) imply

(F ) > K cap F , B() ,
(2.5.20)
where
1+
.
1 + ()
Let be chosen so that the constant K exceeds 1. By (2.5.20) there exists a
()
function u in P(F , B ) such that

u2 d > K
(u )2 dx.
K=
()
B
()
B
S1 [u , u ] < (K 1) 2
u2 dx,

where is the rst eigenvalue of the Dirichlet problem for the Laplace operator
in the unit ball.
Now, Lemma 2 implies that the spectrum of the operator S1 has a limit

point to the left of (K 1)2 . So we arrive at a contradiction.
Hence
196
h
2.5.5 Discreteness of the Negative Spectrum of the Operator S
for all h
The following assertion contains a necessary and sucient condition for the
discreteness of the negative spectrum of the operator Sh for all h > 0. We
note that although the measure in Theorem 2.5.4 is supposed to have no
strong singularities at a nite distance (condition (2.5.17)), the corresponding
criterion for the family of all operators {Sh }h>0 is obtained for an arbitrary
nonnegative measure.
Corollary. The conditions

lim sup (F, ) : F , diam(F ) = 0
0
and

lim sup (F, ) : F \B , diam(F ) 1 = 0
(2.5.21)
(2.5.22)
are necessary and sucient for the semiboundedness of the form Sh [u, u] in
L2 () and for the discreteness of the negative spectrum of the operator Sh for
all h > 0.
We also note that the semiboundedness of the form Sh [u, u] for all h > 0
implies that Sh [u, u] is closable in L2 () for all h > 0.
Proof. Suciency. We introduce the notation

l() = lim sup (F, ) : F \B , diam(F ) .

First we note that (2.5.21) implies (2.5.15). Therefore, according to Theorem 2.5.4, the condition l() 0, combined with (2.5.21), is sucient for the
semiboundedness of the form Sh [u, u] and for the discreteness of the negative
spectrum of the operator Sh for all h > 0.
To prove the suciency of the conditions l(1) = 0 and (2.5.21) we represent
N
an arbitrary compactum F with diam(F ) , > , as the union =1 F ,
where diam(F ) and N depends only on / and n. Since cap(F, ) is a
nondecreasing function of F , then
(F )
(F )
.
cap(F, ) =1 cap(F , )
N
This and the monotonicity of l() immediately imply l() l( ) N l(),

which proves the equivalence of the conditions l() 0 and l(1) = 0.
Necessity. If the form Sh [u, u] is semibounded for all h > 0, then by Corollary 2.5.3/1 the condition (2.5.21) holds together with (2.5.15). Under (2.5.15)
Theorem 2.5.4 implies the necessity of l() 0 which is equivalent to l(1) = 0.
The corollary is proved.

197
2.5.6 Finiteness of the Negative Spectrum

Theorem. Suppose that the condition (2.5.15) holds.
1. If

1
lim sup (F, ) : F \B < ,

4
(2.5.23)
then the form S1 [u, u] is semibounded from below and closable in L2 (), and
the negative spectrum of the operator S1 is nite.
2. If the form S1 [u, u] is semibounded from below and closable in L2 (),
and the negative spectrum of the operator S1 is nite, then

(2.5.24)
lim sup (F, ) : F \B 1.

Proof. 1. Since for any compactum F

)
(F )
(F \B )
(F B
,
cap(F, )
cap(F \B , ) cap(F B , )
(2.5.25)
conditions (2.5.15) and (2.5.23) imply

1
lim (F, ) : F , diam(F ) < .
4
According to the last inequality and Theorem 2.5.3, the form S1 [u, u] is
semibounded and closable in L2 (). From (2.5.11) it follows that the metric

|u|2 dx + S1 [u, u]
C
1 () for C large enough.

is equivalent to the metric of the space W
2
Turning to condition (2.5.23), we note that there exists a positive constant
such that
1

sup (F, ) : F , F \B0 <
4
for suciently large 0 . Hence

(0 ) (F )
1
: F < ,
sup
cap(F, )
4
and by Theorem 2.5.2 the form

(1 4)
|u|2 dx
|u|2 (0 ) (dx)
is positive. Therefore for any u D()
198
S1 [u, u] 4
|u| dx
|u|2 (0 ) (dx).
We estimate the right-hand side by inequality (2.4.10) with = 2, p = 2,

M (t) = |t|:

|u|2 dx K
|u|2 dx.
(2.5.26)
S1 [u, u] 2
B20
Passing to the closure of the form S1 [u, u], we obtain (2.5.26) for all u
21 ().
W
Since any set, bounded in
L12 (), is compact in the metric
1/2

2
|u| dx
B
for any > 0, the form

|u|2 dx K
B20
|u|2 dx,
is nonnegative up to a nite-dimensional manifold. Taking (2.5.26) into account, we may say the same for the form S1 [u, u]. Now the result follows from
Lemma 2.5.4/2.
2. Suppose

lim sup (F, ) : F \B > 1 + ,

where > 0.
Let {k }k1 denote an increasing sequence of positive numbers such that
k 1
0.
k+1
k
(2.5.27)
We construct the subsequence {k }1 , dened as follows: Let k1 = 1. We

, such that (F1 , ) > 1 + .
nd a compactum F1 , contained in \B
k1
Further we select k2 to be so large that F1 is contained in Bk2 . Let F2 denote
a compactum in \Bk2 +1 such that (F2 , ) > 1 + . If numbers k1 , . . . , k
and compacta F1 , . . . , F have already been chosen, then k+1 is dened by
the condition F Bk+1 . The set F+1 \Bk+1 +1 must be chosen to
satisfy the inequality
(F+1 , ) > 1 + .
Thus we obtained a sequence of compacta F with F in the spherical

layer Bk+1 \B
and subject to the condition
k +1
(F ) > (1 + ) cap(F , ).
(2.5.28)
. By Lemma 2.5.1/2, where

We introduce the notation R = Bk+1 +1 \B
k
r = k and R = k +1 ,

cap(F , R )
1 + (n 2)1
k
k +1 k
199

cap(F , Bk+1 +1 ),
which along with condition (2.5.27) implies

cap(F , R ) 1 + o(1) cap(F , Bk+1 +1 ).
(2.5.29)
From Lemma 2.5.1/1 with r = k +1 and R = k+1 +1 it follows that

cap(F , Bk+1 +1 ) 1 + o(1) cap(F , ).
According to (2.5.29),

cap(F , R ) 1 + o(1) cap(F , ).
Hence by (2.5.28), for suciently large ,
(F ) > (1 + ) cap(F , R ),
where is a positive constant.
Now we can nd a sequence of functions u P(F , R ) such that

u2 (dx) > (1 + )
(u )2 dx,
R
which yields the inequality S1 [u , u ] < 0. It remains to note that the supports
of the functions u are disjoint and therefore the last inequality holds for all
linear combinations of u . This and Lemma 2.5.4/2 imply that the negative
spectrum of the operator S1 is innite. The theorem is proved.

2.5.7 Inniteness and Finiteness of the Negative Spectrum of the
h for all h
Operator S
We shall nd criteria for the inniteness and for the niteness of the negative
spectrum of the operator Sh for all h. We underline that here, as in the proof
of the discreteness criterion in Corollary 2.5.5, we obtain the necessary and
sucient conditions without additional assumptions on the measure .
Corollary 1. Conditions (2.5.21) and

sup (F, ) : F =
(2.5.30)
are necessary and sucient for the semiboundedness of the form Sh [u, u] in
L2 () and for the inniteness of the spectrum of the operator Sh for all h > 0.
Proof. By Corollary 2.5.3/1, (2.5.21) is equivalent to the semiboundedness
of the form Sh [u, u] for all h > 0.
We must prove that the criterion
200

lim sup (F, ) : F \B = ,
(2.5.31)
which follows from Theorem 2.5.6, is equivalent to (2.5.30). Obviously, (2.5.30)

is a consequence of (2.5.31). Assume that the condition (2.5.30) is valid. Taking
into account (2.5.21), we obtain

sup (F, ) : F B <
for any . On the other hand, (2.5.30) implies

lim sup (F, ) : F B = .

We choose a sequence such that

sup (F, ) : F B+1 > 2 sup (F, ) : F B .
From this and inequality (2.5.25) we obtain

sup (F, ) : F R ,+1 sup (F, ) : F B ,
. Hence
where R, = B \B

sup (F, ) : F R ,+1 ,


Corollary 2. Conditions (2.5.21) and

lim sup (F, ) : F \B = 0

(2.5.32)
are necessary and sucient for the semiboundedness of Sh and for the niteness of the negative spectrum of Sh for all h > 0.
The necessity and suciency of conditions (2.5.21) and (2.5.32) immediately follow from Theorem 2.5.6.
2.5.8 Proofs of Lemmas 2.5.1/1 and 2.5.1/2
The following facts are well known (cf. Landkof [477]). For n 3 and for any
open set Rn there exists a unique Green function G(x, y) of the Dirichlet
problem for the Laplace operator.
Let be a nonnegative measure in . Let V denote the Green potential
of the measure , i.e.,

V (x) =
G(x, y)(dy).
201
Obviously, V is a harmonic function outside the support of the measure .

There exists a unique capacitary distribution of a compactum F with respect
to , i.e., a measure F , supported on F , such that V F (x) 1 in and
F (F ) = (n 2)1 n1 cap(F, ).
The potential V F is called the capacitary potential of F relative to . If F is
the closure of an open set with C -smooth boundary, then V F is a smooth
function in \F , equal to unity on F , and continuous in .
Proof of Lemma 2.5.1/1. Using the continuity of the capacity from the
right, we can easily reduce the proof for an arbitrary compactum to the consideration of a compactum F Br , which is the closure of an open set
with a C -smooth boundary.
Let V F denote the capacitary potential of F relative to and let
denote a continuous piecewise linear function, equal to unity on [0, r], and to
zero outside [0, R].
L12 ( BR ) by
The function u(x) = (|x|)V F (x) can be approximated in
functions in N(F, BR ). Hence

|u|2 dx.
(2.5.33)
cap(F, BR )
BR
We extend V F to be zero outside . It is readily checked that

V F 2 2 dx + A + B,
|u|2 dx =
BR
where
(2.5.34)
BR

F 2
1
s(dx),
A=
V
R r Br

2 R |x|
n1
dx.
V F
B=
(R r)2 BR \Br
|x|
Obviously,

V F 2 2 dx cap(F, ).
(2.5.35)
Now we note that

F
V s(dx) =
F (dy)
B

G(x, y)s(dx)
B

F (dy)
B
s(dx)
.
|x y|n2
The integral over B is a single-layer potential and it is equal to a constant

on B . Hence, for y B ,

s(dx)
= n .
|x
y|n2
B
202
Thus
V F s(dx) (n 2)1 cap(F, ).
(2.5.36)
B
The following inequality is a direct consequence of the maximum principle

V F (x)
rn2
|x|n2
Now, the bound for A is

1
A (R r)
V F s(dx)
for |x| r.
r
cap(F, ).
(n 2)(R r)
Br
(2.5.37)
We introduce spherical coordinates (, ) in the integral B. Then

n1
B=
(R r)2
Hence
R
n2

r
V F
(dx).
B
(n 1)r n2
Rr
(R ) d
V F (dx).
d
r
B
Using (2.5.36), we obtain

B
n 1 rn2
n2 Rr
2n d cap(F, ),
r
which along with (2.5.33)(2.5.35) and (2.5.37) gives the nal result.
Proof of Lemma 2.5.1/2. The general case can be easily reduced to the
R , which is the closure of an open
consideration of a compactum F \B
F
set with a smooth boundary. Let V
denote the capacitary potential of F
relative to , extended by zero outside .
The function
F
for x \BR ,
V (x)
u(x) = R(|x|r)
V F (x) for x (BR \Br ),
|x|(Rr)
0
for x Br ,
r ) by the functions in N(F, \B
r ). Therefore,
can be approximated in
L12 (\B

r )
(u)2 dx.
cap(F, \B
r
\B
This implies
r )
cap(F, \B

F 2
r
dx +
R(R r)

(V F )2 s(dx).
BR
(2.5.38)
Since V F 1 and
203
V F s(dx) (n 2)1 R cap(F, )

BR
(cf. (2.5.36)), it follows that

2
r
r
(n 2)1 cap(F, ),
V F s(dx)
R(R r) Br
Rr
which together with (2.5.38) completes the proof.

The presentation follows the authors paper [534] and the main results were
announced in Mazya [531]. A number of results on the spectrum of the
Schrodinger operator are presented in the monograph by Glazman [309] who
used the so-called splitting method. Birman [100, 101] established some important results in the perturbation theory of quadratic forms in Hilbert spaces.
In particular, he proved that the discreteness (the niteness) of the negative
spectrum of the operator Sh = h p(x) in Rn for p(x) 0 and for all
h > 0 is equivalent to the compactness of the embedding of W21 (Rn )(
L12 (Rn ))
into the space with the norm
1/2

|u| p(x) dx
2
Rn
Using such criteria, Birman derived the necessary or sucient conditions for
the discreteness, niteness, or inniteness of the negative spectrum of Sh for
all h > 0. The statement of these conditions makes no use of the capacity. The
results of Birmans paper [101] were developed in the authors paper [534] the
content of which is followed here.
The theorems of Sect. 2.5 turned out to be useful in the study of the asymptotic behavior of eigenvalues of the Dirichlet problem for the Schr
odinger operator. Rozenblum [683] considered the operator H = + q(x) in Rn with
q = q+ q , where q Ln/2,loc , n 3. We state one of his results. Let a
cubic grid be constructed in Rn with d as the edge length of each cube and
let F (d) be the union of those cubes Q of the grid that satises the condition

q (x) dx
E
: E 2Q > ,
sup
cap(E)
where 2Q is the concentric homothetic cube having edge length 2d, = (n)
is a large enough number.
Then, for > 0, the number N (, H) of eigenvalues of H that are less
than satises the inequality
204

N (, H) c1
n/2
c3 q(x)
F (c2 1/2 )
dx,
where c1 , c2 , and c3 are certain constants depending only on n.

In the case = Rn by Theorem 2.5.2, the inequality (2.5.4) with = Rn
holds if and only if
(F )
sup
< ,
cap(F
)
F
where F is an arbitrary compact set in Rn . For the same case, other criteria
for the validity of (2.5.4) are known. The following one is due to Kerman and
Sawyer [420] (see Theorem 11.5/1 of the present book):
For every open ball B in Rn ,

d(x) d(y)
c(B).
n2
B B |x y|
Another two criteria for (2.5.4) were obtained by Mazya and Verbitsky [591]:
(i) The pointwise inequality
I1 (I1 )2 (x) cI1 ()(x) < a.e.
holds, where I1 stands for the Riesz potential of order 1, i.e., I1 = |x|1n .
(ii) For every compact set F Rn ,

(I1 )2 dx c cap(F ).
F
One more condition necessary and sucient for (2.5.4) was found by Verbitsky [775]:
For every dyadic cube P in Rn ,
(Q) 2
|Q| c(P ),
|Q|11/n
QP
where the sum is taken over all dyadic cubes Q contained in P and c does not
depend on P .
We now state the main result of the paper [592] by the author and Verbitsky, characterizing arbitrary complex-valued distributions V subject to the
inequality

2

|u| V dx c
|u|2 dx for all u D.
(2.5.39)

Rn
Rn
This characterization reduces the case of distributional potentials V to that

of nonnegative absolutely continuous weights. (Cf. Sect. 1.3.4, where similar
statements are established for functions of one variable.)
2.6 Properties of Sobolev Spaces Generated by Quadratic Forms
205
Theorem. Let V D , n > 2. Then the inequality (2.5.39) holds, if and

only if there is a vector eld L2 (Rn , loc) such that V = div and

Rn

u(x)2 (x)2 dx C

Rn

u(x)2 dx
for all u D. The vector eld can be chosen in the form = 1 V .
2.6 Properties of Sobolev Spaces Generated by

Quadratic Forms with Variable Coecients
2.6.1 Degenerate Quadratic Form
In the preceding sections of the present chapter
we showed that rather gen
eral inequalities, containing the integral [(x, u)]p dx, are equivalent to
isocapacitary inequalities that relate (p, )-capacity and measures. Although
such criteria are of primary interest, we should note that their verication in
particular cases is often dicult. Even for rather simple quadratic forms
(x, )
2
aij (x)i j ,
i,j=1
the estimates for the corresponding capacities by measures are unknown.

Thus, the general necessary and sucient conditions obtained in the
present chapter cannot diminish the value of straightforward methods of investigation of integral inequalities without using capacity. In the present section
this will be illustrated, using as an example the quadratic form
(x, )
2

= |xn | + |x |2 n2 + | |2 ,
where x = (x1 , . . . , xn1 ), = (1 , . . . , n1 ).

By Corollary 2.3.4, the inequality

2
2

u(x , 0) dx c
(x, u) dx
Rn1
(2.6.1)
Rn
holds for all u D(Rn ) if and only if

mn1 {x g, xn = 0} c(2, )-cap(g)
for any admissible set g. A straightforward proof of the preceding isoperimetric
inequality is unknown to the author. Nevertheless, the estimate (2.6.1) is true
and will be proved in the sequel.
206
Theorem 1. Let
n1

2
(u/xi )2 .
(x, u) = |xn | + |x |2 (u/xn )2 +
i=1
Then (2.6.1) is valid for all u D(R ).

n
Proof. Let the integral in the right-hand side of (2.6.1) be denoted by Q(u).
For any (0, 1/2) we have

(|xn | + |x |2 )1/2 u
u(x , 0)2 dx 2
u
dx
(1)/2 |x | x
n
Rn1
Rn |xn |

1/2

1/2
1 2
2
2 Q(u)
|xn | |x | |u| dx
. (2.6.2)
Rn
To give a bound for the last integral we use the following well-known generalization of the HardyLittlewood inequality:

2

2
f (y) dy
dx
c
(2.6.3)
f (y) dy.

n1

2
|x |
Rn1
Rn1 |x y|
Rn1
(For the proof of this estimate see Lizorkin [505]. It can also be derived as a
corollary to Theorem 1.4.1/2.) Since the convolution with the kernel |x |+1n
corresponds to the multiplication by | | of the Fourier transform, (2.6.3)
can be written as

2
(x )/2 u dx ,
|u|2 |x |2 dx c
Rn1
Rn1
where (x )/2 is the fractional power of the Laplace operator. Now we nd
that the right-hand side in (2.6.2) does not exceed

1
/2 2
c Q(u) +
(x ) u dx .
|xn |
(2.6.4)
Rn
From the almost obvious estimate

g 2 t1 dt c
(g )2 t dt +
0
it follows that
| |2

Rn

g 2 dt ,

(Fx u)( , xn )2 |xn |1 dxn
2

u

c
Fx xn ( , xn ) |xn | dxn
1
R

2
2

(Fx u)( , xn ) dxn ,
+ | |
R1
207
where Fx is the Fourier transform in Rn1 . So the second integral in (2.6.4)
does not exceed

|xn |(u/xn )2 + (x u)2 dx.
c
Rn
The result follows.

The next assertion shows that Theorem 1 is exact in a certain sense.
Theorem 2. The space of restrictions to Rn1 = {x Rn : xn = 0} of

functions in the set {u D(Rn ) : Q(u) + u2L2 (Rn ) 1} is not relatively
(n1)
compact in L2 (B1
(n1)
), where B
= {x Rn1 : |x | < }.
) such that (y) = (y),

Proof. Let denote a function in C0 (B1
L2 (Rn1 ) = 1 and introduce the sequence {m }
m=1 dened by m (y) =
m(n1)/2 (my). Since this sequence is normalized and weakly convergent to
(n1)
(n1)
zero in L2 (B1
), it contains no subsequences converging in L2 (B1
).
n
Further, let {vm }
m=1 be the sequence of functions in R dened by

2
vm (x) = F1
x exp |xn | Fx m ,
(n1)
where Rn1 , = (||2 + 1)1/2 .

Consider the quadratic form
2

2 u
2
2
T (u) =
|xn | + |x |
+ |x u| + |u| dx.
xn
n
R
It is clear that
T (u) = (2)1n
Rn

2

F u 2

+ F u + 2 |F u|2 d dxn .
|xn |
t

t
Dierentiating the function T (vm ), we obtain from the last equality that
T (vm ) does not exceed

1 + 2 |xn | + 4 |xn |3 2 |F m |2 + 4 |F m |2
c
n
R

exp 22 |xn | d dxn .
Thus we obtain
T (vm ) c
R

n1

2 |F m |2 + |F m |2 d
n1
= c1

xi m 2W 1 (Rn1 )
2
m 2L2 (Rn1 )
const.
i=1
Let C0 (B2
), = 1 on B1
. It is clear that (vm )|Rn1 = m
and T (vm ) const. The sequence {vm /(T (vm ))1/2 }
m=1 is the required
counterexample. The theorem is proved.

(n1)
(n1)
208
2.6.2 Completion in the Metric of a Generalized Dirichlet Integral


S[u, u] =
Rn
aij (x)

u u
+ u2 dx,
xi xj
where aij (x)ni,j=1 is a uniformly positive denite matrix, whose elements

aij (x) are smooth real functions.
Let the completion of C00,1 with respect to the norm (S[u, u])1/2 be denoted
by H(S).
Further, we introduce the space H(S) obtained as the completion
with respect to this norm of the set of functions in C 0,1 with the nite integral
S[u, u].
If the elements of the matrix aij ni,j=1 are bounded functions, then
21 , H(S) = W21 and both spaces, obviously, coincide. It is also

H(S)
= W
known that H(S)

= H(S) if the functions aij do not grow too rapidly at
innity. Here we consider the problem of the coincidence of H(S)

and H(S)
in the general case.
Denition. Let E Rn . In the present subsection the set E is said to
have nite H(S) capacity if there exists a function u C 0,1 H(S) that is
equal to 1 on E.
Theorem 1. The spaces H(S)

and H(S) coincide if and only if, for an
arbitrary domain G with nite H(S) capacity, there exists a sequence of functions {m }m1 in C00,1 that converges in measure to unity on G and is such
that

m m
aij (x)
dx = 0.
(2.6.5)
lim
m G
xi xj
Before we proceed to the proof, we note that if G is a bounded domain then
the sequence {m }m1 always exists. We can put m = where C00,1 ,
= 1 on G.
Proof. Suciency. We show that any function u C 0,1 H(S) can be
approximated in H(S) by functions in H(S).

Without loss of generality we
may assume that u 0.
First, we note that if t > 0 then Lt = {x : u(x) > t} is a set of nite
H(S) capacity. In fact, the function v(x) = t1 min{u(x), t} equals unity on
Lt , satises a Lipschitz condition, and S[v, v] t2 S[u, u] < .
From the Lebesgue theorem it follows that the sequences min{u, m}, (u
m1 )+ , m = 1, 2, . . . , converge to u in H(S) (see Sect. 5.1.2). So we may
assume from the very beginning that u is bounded and vanishes on the exterior
of a bounded set G of nite H(S) capacity.
We denote the complements of the set G by Gj and then dene the sequence

u
(m)
(x) =
u(x)
for x
for x R \
jm
n
209
Gj ,
jm
Gj ,
j = 1, 2, . . . . It is clear that u(m) u in H(S) as m . Since each u(m)

vanishes on the exterior of a nite number of domains, we may assume without
loss of generality that G is a domain.
Let {m } be the sequence of functions specied for the domain G in the
statement of the theorem. Replacing {m } by the sequence {m }, dened by
|m | = min{2, |m |}, sgn m = sgn m , we obtain a bounded sequence with
and m u u in L2 . Moreover,
the same properties. Obviously, m u H(S)
aij
(u um )
(u um ) dx
x
x
n
i
j
R

u u
m m
2 (1 m )2 aij
dx + 2
u2 aij
dx.
(2.6.6)
x
x
xi xj
i
j
G
G
Since the sequence
(1 m )2

i,j
aij
u u
,
xi xj
converges to zero in G with respect to the measure mn and is majorized by

the integrable function

u u
9
aij
,
x
i xj
i,j
the rst integral on the right in (2.6.6) converges to zero. The convergence
to zero of the second integral follows from the boundedness of u and equality (2.6.5). Thus um u in H(S). The required approximation is constructed.
Necessity. Let G be an arbitrary domain in Rn with nite H(S) capacity.
Let u denote a function in C 0,1 H(S), which is equal to unity on G. Since
H(S) and H(S)

coincide u can be approximated in H(S) by the sequence
{m }m1 contained in C00,1 . Noting that u = 1 on G and m u in L2 (G),
we obtain that m 1 in G in measure. Furthermore,

m m
m
aij
dx =
aij
(u m )
(u m ) dx 0.
x
x
x
x
i
j
i
j
G
G
So the theorem is proved.
Although the above result is not very descriptive, it facilitates verication
of concrete conditions for coincidence or noncoincidence of H(S) and H(S).

We now present some of them.
Theorem 2. (cf. Mazya [536]). The spaces H(S) and H(S)

coincide provided n = 1 or n = 2.
210
Proof. Taking into account Theorem 1 and the discussion that follows its
statement, we arrive at the equality H(S) = H(S)

if we show that any domain
G with nite H(S) capacity is bounded. The case n = 1 is obvious. Let n = 2
and u C 0,1 H(S), u = 1 on G.
Let O and P denote arbitrary points in G and let the axis Ox2 be directed
from O to P . Then
2

|P |
|P |

2
u
dx2
+ u2 dx1 c1
max u(x1 , x2 ) dx2 .
S[u, u] c
x1
x1
0
0
R1
Taking into account that G is a domain and u = 1 on G we arrive at

2
max u(x1 , x2 ) 1.
x1
Therefore diam(G) cS[u, u], which completes the proof.
The following assertion shows that for n 3 the form S[u, u] must be subjected to certain conditions by necessity. The result is due to Uraltseva [769].
Our proof, though dierent, is based on the same idea.
Theorem 3. Let n > 2. Then there exists a form S[u, u] for which H(S) =
H(S).
Proof. 1. Consider the domain G = {x : 0 < xn < , |x | < f (xn )}
where x = (x1 , . . . , xn1 ) and f is a positive decreasing function in C [0, ),
f (0) < 1. For x
/ G we put aij (x) = ij .
For arbitrary functions aij on G, for any u C 0,1 , u = 1 on G, we have
S[u, u] = u2W 1 . This implies that G is a domain with nite H(S) capacity
2
if and only if cap(G) < (here, as before, cap is the Wiener capacity, i.e.,
2-cap). Clearly,
cap(G)

cap {x G : j xn j + 1}
j=0

cap {x : |x | f (j), j xn j + 1} .
j=0
This and the well-known estimates for the capacity of the cylinder (cf. Landkof [477] or Proposition 13.1.3/1 of the present book) yield
cap(G) c
cap(G) c

j=0
n3
f (j)

log f (j)1
for n > 3,
for n = 3.
j=0
Therefore G is a domain with nite H(S) capacity provided
f (t)
0
n3
dt <

log f (t)1 dt <
211
for n > 3,
for n = 3.
2. In the interior of G we dene the quadratic form aij (x)i j by

2
n1

n1

g(x
)
n
aij (x)i j = 2 +
(x) f (xn )
xi i + n ,
f (xn )
i=1
where C0 (G), 0 1, (x) = 1 on the set {x : 1 < xn < , |x | <
1
2 f (xn )}, and g is an arbitrary positive function on [0, ) satisfying the con
dition

1n
dt < .
g(t)
0
Using the change of variable xn = yn , xi = f (yn )yi , 1 i n 1, we map

G onto the cylinder {y : 0 < yn < , |y | < 1}. Obviously,

2
n1

n1

g(xn )

(x) f (xn )
xi
+
dx
xi
xn
G f (xn )
i=1

n1 2

dy,
g(yn )
yn
C
where C = {y : 1 < yn < , |y | < 12 }. Applying the Cauchy inequality to the
last integral we obtain

2

1n

g(t)
dt
aij
dx
yn dyn dy .
xi xj
1
1
G
|y |1/2
If C00,1 then the right-hand side exceeds

2

max (y , yn ) dy
|y |<1/2 1<yn <
2 dy,
C1
where C1 = {y C : yn < 2}. Passing to the variables x1 , . . . , xn on the right,

we arrive at

1n

dx
dt
aij
dx
2
,
g(t)
xi xj
[f (xn )]n1
1
G
G1
where G1 = {x : |x | < 12 f (xn ), 1 < xn < 2}. Thus, for any sequence {m }m1
of functions in C00,1 converging in measure to unity in G we have

aij m /xi m /xj dx > 0.
lim inf
m
To conclude the proof, it remains to make use of Theorem 1.
212
Theorem 3 has an interesting application to the problem of the selfadjointness of an elliptic operator in L2 (Rn ), n 3 (cf. Uraltseva [769]).
Let the operator

u
u S0 u =
aij (x)
+u
xi
xj
be dened on C0 . If aij ni,j=1 is the matrix constructed in Theorem 3, then
H(S) = H(S)
and hence there exists a function w H(S), which does not
vanish identically and is orthogonal to any v C0 in H(S), i.e.,

w v
+ wv dx =
wS0 v dx.
0=
aij
xi xj
Rn
Rn
Therefore the range of the closure S0 does not coincide with L2 . If S0 is selfadjoint then w Dom(S0 ) and S0 w = 0. This obviously implies w = 0. We
arrived at a contradiction, which means that S0 is not self-adjoint. Thus, the
condition of the uniform positive deniteness of the matrix aij ni,j=1 alone is
insucient for the self-adjointness of S0 .
The results of Sect. 2.6.1 are due to the author [556], Sect. 2.6. We note that
the proof of Theorem 2.6.1/2 implies nondiscreteness of the spectrum of the
Steklov problem

u
aij (x)
+ a(x)u = 0 in ,
xj
xj
i,j=1
n

aij cos(, xj )
i,j=1
u
= u
xi
on ,
under the condition that is characteristic at least at one point. Here

is a normal to and the matrix aij ni,j=1 is nonnegative a(x) > 0. The
coecients aij , a, and the surface are assumed to be smooth.
In conclusion, we note that the topic of Sect. 2.6.2 was also considered in
the paper by S. Laptev [482] who studied the form

S[u, u] =
(x)(u)2 + u2 dx,
Rn
where (x) const > 0. He presented an example of a function for which
H(S) = H(S)
and showed that H(S) and H(S)
coincide in each of the following three cases: (i) is a nondecreasing function in |x|, (ii) (x) = O(|x|2 +1),
and (iii) n = 3 and depends only on |x|.
2.7 Dilation Invariant Sharp Hardys Inequalities
213

2.7.1 Hardys Inequality with Sharp Sobolev Remainder Term
Here we nd the best value of C for a particular case of the inequality

1
|v|2
|v|2 dx
dx + Cxn v2Lq (Rn+ ) ,
(2.7.1)
2
n
n
4
x
R+
R+
n
which is equivalent to (2.1.36) with m = 1.
Theorem. For all u C (Rn+ ), u = 0 on Rn1 , the sharp inequality
1
|u| dx
4
Rn
+
Rn
+
|u|2
dx
x2n

n/(n+1) (n2 1)
x1/(n+1) u2
n
n
L 2(n+1) (Rn
2/(n+1)
+)
4( ( 2 + 1))
n1
(2.7.2)
holds.
Proof. We start with the Sobolev inequality

|w|2 dz Sn+1 w2L 2(n+1) (Rn+1 )
Rn+1
(2.7.3)
n1
with the best constant

Sn+1 =
(n+2)/(n+1) (n2 1)
+ 1))2/(n+1)
4n/(n+1) ( ( n2
(2.7.4)
(see (2.3.23)).
Let us introduce the cylindrical coordinates (r, , x ), where r 0,
[0, 2), and x Rn1 . Assuming that w does not depend on , we write
(2.7.3) in the form
2

w
2

2
r + |x w| r dr dx
Rn1 0

(n1)/(n+1)

(2)(n1)/(n+1) Sn+1
|w|2(n+1)/(n1) r dr dx
.

Rn1
Replacing r by xn , we obtain

|w| xn dx (2)
2
Rn
+
2/(n+1)
1/2

Sn+1
|w|
2(n+1)/(n1)
Rn
+
It remains to set here w = xn v and to use (2.7.4).
(n1)/(n+1)
xn dx
.

214
2.7.2 Two-Weight Hardys Inequalities

As usual, here and elsewhere Rn+ = {x = (x1 , . . . , xn ) Rn , xn > 0} and
C0 (Rn+ ) and C0 (Rn+ ) stand for the spaces of innitely dierentiable functions
with compact support in Rn+ and Rn+ , respectively.

p

|u(x)|p
p
(2p)
xnp1 u(x) dx
2
2
1/2
(xn1 + xn )
Rn
Rn
+
+
(2.7.5)
holds for all u C0 (Rn+ ).

Proof. We put 2 = x2n1 + x2n and denote the integrals on the left- and
right-hand sides by I and J , respectively. Integrating by parts, we obtain

u
1
p1
I = p
xn |u|
sgn u
dx +
x2n 3 |u|p dx.
xn
Rn
Rn
We denote two summands in the right-hand side by I1 and I2 . Clearly, by
Holders inequality we have |I1 | (p1)/p J 1/p . To obtain a bound for I2
we introduce cylindrical coordinates (z, , ) with z Rn2 , xn1 + ixn =
exp(i). Then

u
2
I2 = p
dz
sin d
|u|p1 sgn u d I (p1)/p J 1/p .

n2
R
0
0
Thus I 2pI (p1)/p J 1/p and (2.7.5) follows.
In this section we are concerned with generalizations of the inequality

|u|2
2
xn |u| dx
dx, u C0 (Rn+ ).
(2.7.6)
(x2n1 + x2n )1/2
Rn
Rn
+
+
1/2
By substituting u(x) = xn v(x) into (2.7.6), one arrives at the improved

Hardy inequality

1
|v|2 dx
|v|2 dx
|v|2 dx
(2.7.7)
2
2
4 Rn+ xn
xn (xn1 + x2n )1/2
Rn
Rn
+
+
for all v C0 (Rn+ ).
1/2
More generally, replacing u by xn v(x) in the next theorem, we nd a
condition on the function q that is necessary and sucient for the inequality

1
|v|2 dx
2
|v| dx
4 Rn+ x2n
Rn
+

|v|2 dx
xn
C
q
,
(2.7.8)
2
2
2
1/2
(xn1 + xn )
xn (xn1 + x2n )1/2
Rn
+
215
where v is an arbitrary function in C0 (Rn+ ). This condition implies, in particular, that the right-hand side of (2.7.7) can be replaced by

|v|2 dx
.
C
n
2
x2n (1 log (x2 x+x
2 )1/2 )
Rn
+
n
n1
Theorem 2. (i) Let q denote a locally integrable nonnegative function on

(0, 1). The best constant in the inequality

xn
|u|2
xn |u|2 dx C
q
dx, (2.7.9)
2
2
2
1/2
(xn1 + xn )
(xn1 + x2n )1/2
Rn
Rn
+
+
for all u C0 (Rn+ ), is given by
/2
0
:= inf
(|y ()|2 + 14 |y()|2 ) sin d

,
/2
|y()|2 q(sin ) d
0
(2.7.10)
where the inmum is taken over all smooth functions on [0, /2].
(ii) Inequalities (2.7.9) and (2.7.8) with a positive C hold if and only if
sup (1 log t)
t(0,1)
Moreover,
q( ) d < .
(2.7.11)
t(0,1)
1
sup (1 log t)
q( ) d
(2.7.12)
where a b means that c1 a b c2 a with absolute positive constants c1 and

c2 .
Proof. (i) Let U C0 (R2+ ), C0 (Rn2 ), x = (x1 , . . . , xn2 ), and let
N = const > 0. Putting

u(x) = N (2n)/2 N 1 x U (xn1 , xn )
into (2.7.9) and passing to the limit as N , we see that (2.7.9) is equivalent
to the inequality

x2 |Ux1 |2 + |Ux2 |2 dx1 dx2
R2+
x2
C
q
2
(x1 + x22 )1/2
R2+
|U |2 dx1 dx2
,
(x21 + x22 )1/2
(2.7.13)
where U C0 (R2+ ). Let (, ) be the polar coordinates of the point (x1 , x2 )

R2+ . Then (2.7.13) can be written as
216

0

|U |2 + 2 |U |2 sin d 2 d C
By the substitution
|U |2 q(sin ) d d.
0
U (, ) = 1/2 v(, )
the left-hand side becomes

d
1
|v |2 + |v |2 + |v|2 sin d
4
0
0

vv d sin d.
Re
0
(2.7.14)
Since v(0) = 0, the second term in (2.7.14) vanishes. Therefore, (2.7.13) can
be written in the form

1
d
|v |2 + |v |2 + |v|2 sin d
4
0
0

d
C
(2.7.15)
|v|2 q(sin ) d .
0
0
Now, the denition (2.7.10) of shows that (2.7.9) holds with C = .
To show the optimality of this value of C, put t = log and v(, ) =
w(t, ). Then (2.7.9) is equivalent to

1 2
2
2
|wt | + |w | + |w| sin d dt
4
R1 0

C
|w|2 q(sin ) d dt.
(2.7.16)
R1
Applying the Fourier transform w(t, ) w(s,

), we obtain

1
|w
|2 + |s|2 +
|w|
2 sin d ds
4
R1 0

C
|w|
2 q(sin ) d ds.
R1
Putting here
(2.7.17)
w(s,
) = 1/2 (s/)y(),
where C0 (R1 ), L2 (R1 ) = 1, and y is a function on C ([0, ]), and

passing to the limit as 0, we arrive at the estimate

2 1

y () + y()2 sin d C
y()2 q(sin ) d,
(2.7.18)
4
0
0
where can be changed for /2 by symmetry. This together with (2.7.10)
implies . The proof of (i) is complete.
217
(ii) Introducing the new variable = log cot 2 , we write (2.7.10) as

= inf
z
Since
|z()|2
4(cosh )2 ) d
.

1
|z()|2 q( cosh
) d
cosh
0

z(0)2 2
(|z ()|2 +
(2.7.19)

z ()2 + z()2 d
1
0
and

z()2
e2
d
(1 + e2 )2
0

2 d

2
e2

z() z(0) 2 + 2 z(0)
d
2
(1 + e2 )2
0
0
2

z () d + z(0)2
8
0
it follows from (2.7.19) that

inf 0
z
|z ()|2 d + |z(0)|2
d
1
|z()|2 q( cosh
) cosh
(2.7.20)
Setting z() = 1 and z() = min{ 1 , 1} for all positive and xed > 0
into the ratio of quadratic forms in (2.7.20), we deduce that

min
q
0
1
cosh
Hence,
d
cosh
1

, sup
>0

c
1
cosh
d
cosh
1
.
1
sup (1 log t)
t(0,1)
q( ) d
To obtain the converse estimate, note that

d
1
z()2 q
cosh cosh
0

2
d
d
1
1
z() z(0)2 q
+2
.
2z(0)
q
cosh cosh
cosh cosh
0
0
The second term in the right-hand side is dominated by

2
1
d
z () d
q
8 sup
cosh
cosh
>0
0
(see Sect. 1.3.2). Therefore,
218

d
1
cosh cosh
0

1
1
d
d
q
q
8 max
, sup
cosh cosh >0
cosh cosh
0

2

z () d + z(0)2 ,

z()2 q
which together with (2.7.20) leads to the lower estimate

min
q
0
1
cosh
Hence,
d
cosh
1

, sup
>0
t(0,1)
1
cosh
d
cosh
1
.
1
sup (1 log t)
q( ) d
The proof of (ii) is complete.
Since (2.7.11) holds for q(t) = t1 (1 log t)2 , Theorem 2(ii) leads to the
Corollary 1. There exists an absolute constant C > 0 such that the inequality
1
|v| dx
n
4
R+
Rn
+
|v|2 dx
C
x2n

Rn
+
|v|2 dx
n
2
x2n (1 log (x2 x+x
2 )1/2 )
n1
(2.7.21)
holds for all v C0 (Rn+ ). The best value of C is equal to

:= inf
0
[|y ()|2 + 14 |y()|2 ] sin d

,
|y()|2 (sin )1 (1 log sin )2 d
0
(2.7.22)
where the inmum is taken over all smooth functions on [0, /2]. By numerical
approximation, = 0.16, . . . .
A particular case of Theorem 2 corresponding to q = 1 is the following
assertion.
Corollary 2. The sharp value of in (2.7.6) and (2.7.7) is equal to

:= inf
[|y ()|2 + 14 |y()|2 ] sin d

,
|y()|2 d
0
(2.7.23)
where the inmum is taken over all smooth functions on [0, ]. By numerical
approximation, = 0.1564, . . ..
219
of the operator
Remark 1. Let us consider the Friedrichs extension L
sin

z,
L : z (sin )z +
4
(2.7.24)
dened on smooth functions on [0, ]. It is a simple exercise to show that the

is compactly embedded into L2 (0, ). Hence, the spectrum
energy space of L
of L is discrete and dened by (2.7.23) is the smallest eigenvalue of L.

Remark 2. The argument used in the proof of Theorem 2(i) with obvious
changes enables one to obtain the following more general fact. Let P and Q
be measurable nonnegative functions in Rn , positive homogeneous of degrees
2 and 2 2, respectively. The sharp value of C in

2
P (x)|u| dx C
Q(x)|u|2 dx, u C0 Rn ,
(2.7.25)
Rn
Rn
is equal to

:= inf
S n1
P ()(| Y |2 + ( 1 + n2 )2 |Y |2 ) ds

,
Q()|Y |2 ds
S n1
where the inmum is taken over all smooth functions on the unit sphere S n1 .
The material of this subsection is borrowed from Mazya and Shaposhnikova
[587]. In Sect. 2.7.1 we are concerned with the inequality (2.7.1) which is a
special case of (2.1.36). Another inequality of a similar nature, whose generalizations are dealt with in Sect. 2.7.2, is (2.7.7). It is equivalent to (2.7.6) and
was obtained in 1972 by the author, proving to be useful in the study of the
generic case of degeneration in the oblique derivative problem for second-order
elliptic dierential operators [541].
Without the second term on the right-hand sides of (2.7.1) and (2.7.7),
these inequalities reduce to the classical Hardy inequality with the sharp
constant 1/4. An interesting feature of (2.7.1) and (2.7.7) is their dilation
invariance. The value = 1/16 in (2.7.7) obtained in Mazya [541] is not
the best possible. Tidblom replaced it by 1/8 in [752]. As a corollary of Theorem 2.7.2/2, we nd an expression for the optimal value of (see Corollary 2.7.2/2).
Sharp constants in Hardy-type inequalities as well as variants, extensions,
and renements of (2.7.1) and (2.7.7), usually called Hardys inequalities with
remainder term, became a theme of many subsequent studies (Davies [225,
226]; Brezis and Marcus [142]; Brezis and V
azquez [145]; Matskewich and
Sobolevskii [526]; Sobolevskii [715]; Davies and Hinz [227]; Marcus, Mizel,
and Pinchover [516]; Laptev and Weidl [481]; Weidl [792]; Yafaev [801];
Brezis, Marcus, and Shafrir [143]; V
azquez and Zuazua [773]; Eilertsen [256];
220
Adimurthi [26]; Adimurthi, Chaudhuri, and Ramaswamy [27]; Filippas and

Tertikas [278]; M. Homan-Ostenhof, T. Homan-Ostenhof, and Laptev [380];
Barbatis, Filippas, and Tertikas [73, 74]; Balinsky [67]; Barbatis, Filippas,
and Tertikas [72]; Chaudhuri [178]; Z.-Q. Wang and Meijun [791]; Balinsky,
Laptev, and A. Sobolev [68]; Davila and Dupaigne [228]; Dolbeault, Esteban, Loss, and Vega [237]; Filippas, Mazya, and Tertikas [275277]; Gazzola,
Grunau, and Mitidieri [303]; Tidblom [751, 752]; Colin [211]; Edmunds and
Hurri-Syrjanen [252]; Galaktionov [301, 302]; Samko [689]; Yaotian and Zhihui [804]; Adimurthi, Grossi, and Santra [28]; Alvino, Ferone, and Trombetti
[42]; Barbatis [70, 71]; Brandolini, Chiacchio, and Trombetti [140]; Dou, Niu,
and Yuan [241]; Evans and Lewis [264]; Tertikas and Tintarev [749]; Tertikas
and Zographopoulos [750]; Benguria, Frank, and Loss [83]; Bosi, Dolbeault,
and Esteban [127]; Frank and Seiringer [289, 290]; Frank, Lieb, and Seiringer
[288]; A. Laptev and A. Sobolev [480]; Cianchi and Ferone [203]; Kombe and
Ozaydin
[446]; Filippas, Tertikas, and Tidblom [279]; Pinchover and Tintarev
[661]; Avkhadiev and Laptev [58] et al.).
2.8 Sharp HardyLeray Inequality for Axisymmetric

Divergence-Free Fields
2.8.1 Statement of Results
Let u denote a C0 (Rn ) vector eld in Rn . The following n-dimensional generalization of the one-dimensional Hardy inequality,

4
|u|2
dx
|u|2 dx
(2.8.1)
2
(n 2)2 Rn
Rn |x|
appears for n = 3 in the pioneering paper by Leray on the NavierStokes
equations [487]. The constant factor on the right-hand side is sharp. Since one
frequently deals with divergence-free elds in hydrodynamics, it is natural to
ask whether this restriction can improve the constant in (2.8.1).
We show in the present section that this is the case indeed if n > 2 and the
vector eld u is axisymmetric by proving that the aforementioned constant
can be replaced by the (smaller) optimal value

4
8
1
,
(2.8.2)
(n 2)2
(n + 2)2
which, in particular, evaluates to 68/25 in three dimensions. This result is a
special case of a more general one concerning a divergence-free improvement
of the multidimensional sharp Hardy inequality

4
|x|22 |u|2 dx
|x|2 |u|2 dx.
(2.8.3)
2
(2
+
n
2)
n
n
R
R
2.8 Sharp HardyLeray Inequality for Axisymmetric Divergence-Free Fields
221
Let be a point on the (n2)-dimensional unit sphere S n2 with spherical

coordinates {j }j=1,...,n3 and , where j (0, ) and [0, 2). A point
x Rn is represented as a triple (, , ), where > 0 and [0, ]. Correspondingly, we write u = (u , u , u ) with u = (un3 , . . . , u1 , u ).
The condition of axial symmetry means that u depends only on and .
For higher dimensions, our result is as follows.
Theorem 1. Let = 1 n/2, n > 2, and let u be an axisymmetric
divergence-free vector eld in C0 (Rn ). We assume that u(0) = 0 for <
1 n/2. Then

|x|22 |u|2 dx Cn,
|x|2 |u|2 dx
(2.8.4)
Rn
Rn
with the best value of Cn, given by

Cn,

2
4
1
,
=
(2 + n 2)2
n + 1 + ( n/2)2
for 1, and by
2

n
1
+1
Cn, =
2

+ min n 1, 2 + min x +
x0
4(n 1)( 1)
x + n 1 + ( n/2)2
(2.8.5)

(2.8.6)
for > 1.
The two minima in (2.8.6) can be calculated in closed form, but their
expressions for arbitrary dimensions turn out to be unwieldy and we omit
them.
However, the formula for C3, is simple.
Corollary. For n = 3 inequality (2.8.4) holds with the best constant

2+(3/2)2
4
for 1,
2 4+(3/2)2
(2.8.7)
C3, = (2+1)
4
for > 1.
8+(1+2)2
For n = 2, we obtain the sharp constant in (2.8.4) without axial symmetry
of the vector eld.
Theorem 2. Let = 0, n = 2, and let u be a divergence-free vector eld
in C0 (R2 ). We assume that u(0) = 0 for < 0. Then inequality (2.8.4) holds

2
for [ 3 1, 3 1],
2 1+(1)
2
3+(1)
(2.8.8)
C2, =
( 2 + 1)1
otherwise.
222
2.8.2 Proof of Theorem 1

In the spherical coordinates introduced previously, we have

n1

u + 1 (sin )2n
(sin )n2 u
n3

( sin sin n3 sin k+1 )1 (sin k )k
(sin k )k uk
+
k
div u = 1n
k=1
+ ( sin sin n3 sin 1 )1
u
.
(2.8.9)
Since the components u and uk , k = 1, . . . , n 3, depend only on and ,

(2.8.9) becomes

n1

(sin )n2 u (, )
u (, ) + 1 (sin )2n
n3

u (, )
.
(2.8.10)
+
k(sin n3 sin k+1 )1 cot k k
sin
div u = 1n
k=1
By the linear independence of the functions

(sin n3 sin k+1 )1 cot k ,
1,
k = 1, . . . , n 3,
the divergence-free condition is equivalent to the collection of n 2 identities

u
+ (n 1)u +
+ (n 2) cot u = 0,
(2.8.11)
uk = 0,
k = 1, . . . , n 3.
(2.8.12)
If the right-hand side of (2.8.4) diverges, there is nothing to prove. Otherwise, the matrix u is O(|x|m ), with m > n/2, as x 0. Since
u(0) = 0, we have u(x) = O(|x|m+1 ) ensuring the convergence of the integral
on the left-hand side of (2.8.4). We introduce the vector eld
v(x) = u(x)|x|1+n/2 .
(2.8.13)
The inequality (2.8.4) becomes

1
Cn,
n
+1
2
2

|v|2
|v|2
dx
dx.
n
n2
Rn |x|
Rn |x|
The condition div u = 0 is equivalent to

n2
+ v .
div v =
2
(2.8.14)
(2.8.15)
223
To simplify the exposition, we assume rst that v = 0. Now, (2.8.15) can

be written as

n
v
+
v + Dv = 0,
(2.8.16)
2
where
+ (n 2) cot .
(2.8.17)
Note that D is the adjoint of / with respect to the scalar product

f ()g()(sin )n2 d.
D :=
A straightforward, though lengthy calculation yields
Hence

2
2
+ 2
2
2
v

v
+ v2 + (n 1)v2 + (n 2)(cot )2 v2 + 2 v Dv v
.
(2.8.18)

2 |v|2 = 2
2
2
2
2
2

v
v
v
v
+
+ 2
+
2
S n1

+ v2 + (n 1)v2 + (n 2)(cot )2 v2 + 4v Dv ds.

|v|2 ds =
S n1
(2.8.19)
Changing the variable to t = log and applying the Fourier transform with
respect to t,
v(t, ) w(, ),
we derive

Rn
and
|v|2
dx =
|x|n2

R

l2 + n 1 |w |2 + l2 n + 3 |w |2
S n1

w 2 w 2
2
2
+

+
+ (n 2)(sin ) |w |

+ 4(w Dw ) ds d

Rn
From (2.8.15), we obtain
|v|2
dx =
|x|n
(2.8.20)

R
|w|2 ds d.
S n1
(2.8.21)
224
Dw
,
i + n/2
(2.8.22)
|Dw |2
2 + (n/2 )2
(2.8.23)
w =
which implies
|w |2 =
and
(w Dw ) =
(n/2 )|Dw |2
.
2 + (n/2 )2
(2.8.24)
Introducing this into (2.8.20), we arrive at the identity

2

|v|2
|Dw |2
dx
=
+
n
1
n2
2 + (n/2 )2
Rn |x|
S n1
0

w 2

+ (n 2)(sin )2 |w |2
+ 2 n + 3 |w |2 +

2

1

+ 2
Dw
+ (n/2 )2

|Dw |2
n
4
ds d.
2
2 + (n/2 )2
We simplify the right-hand side to obtain

|v|2
n 1 + 2 + 4
dx =
+ 1 |Dw |2
n2
2 + (n/2 )2
0
Rn |x|
S n1

2

1
Dw ds d.
+ 2 n + 3 |w |2 + 2

2
+ (n/2 )
(2.8.25)
On the other hand, by (2.8.21) and (2.8.22)

|v|2
|Dw |2
2
dx =
+ |w | ds d.
n2
2
2
0
Rn |x|
S n1 + (n/2 )
(2.8.26)
Dening the self-adjoint operator

T :=
or equivalently,
T = +
D,
n2
,
(sin )2
(2.8.27)
(2.8.28)
where is the part of the LaplaceBeltrami operator on S n1 , we write

(2.8.25) and (2.8.26) as

|v|2
dx =
Q(, w ) ds d
(2.8.29)
n2
Rn |x|
R S n1
and
Rn
|v|2
dx =
|x|n
225

R
q(, w ) ds d,
(2.8.30)
S n1
respectively, where Q and q are sesquilinear forms in w , dened by

n 1 + 2 + 4
Q(, w ) =
+
1
T w w
2 + (n/2 )2

1
+ 2 n + 3 |w |2 + 2
|T w |2
+ (n/2 )2
and
q(, w ) =
T w w
+ |w |2 .
2 + (n/2 )2
(2.8.31)
The eigenvalues of T are = ( + n 2), Z+ . Representing w as

an expansion in eigenfunctions of T , we nd

n1 Q(, w ) ds d
inf R S
w
q(, w ) ds d
R S n1
+4
( n1+
+ 1) + 2 n + 3 +
2 +(n/2)2
2
= inf inf
R N
2
2 +(n/2)2
(2.8.32)
2 +(n/2)2 +1
Thus our minimization problem reduces to nding

inf inf f (x, , ),
x0 N
where

f (x, , ) = x n + 3 + 1

16(1 )
.
4x + 4 + (n 2)2
(2.8.33)
(2.8.34)
Since 1, it is clear that f is increasing in x, so the value (2.8.33) is equal

to

16(1 )
inf f (0, , ) = inf 3 n + 1
.
(2.8.35)
N
N
4 + (n 2)2
We have
16(1 )(n 2)
f (0, , ) = 1
.
(4 + (n 2)2 )2
(2.8.36)
Noting that
4 + (n 2)2 4(n 1) + (n 2)2 4 n 1(n 2),

we see that
1
f (0, , ) 1
> 0.
(n 1)(n 2)
(2.8.37)
(2.8.38)
226
Thus the minimum of f (0, , ) is attained at 1 = n 1 and equals

2( 1 + n/2)2
16(1 )
=
3 n + (n 1) 1
. (2.8.39)
4(n 1) + (n 2)2
n 1 + ( n/2)2
This completes the proof for the case v = 0.
If we drop the assumption v = 0, then, to the integrand on the right-hand
side of (2.8.19), we should add the terms

2
2
v
v
2
+
+ (sin sin n3 sin 1 )2 v2 .
(2.8.40)
The expression in (2.8.40) equals

2
2 v ei .
As a result, the right-hand side of (2.8.29) is augmented by

R(, w ) ds d,
R
where
(2.8.41)
(2.8.42)
S n1

2
R(, w ) = 2 |w |2 + w ei
(2.8.43)
with = (, n3 , . . . , ). Hence,

inf
v
|v|2
dx
Rn |x|n2
|v|2
dx
Rn |x|n

n1 (Q(, w ) + R(, w )) ds d
= inf R S
.
w ,w
(q(, w ) + |w |2 ) ds d
R S n1
(2.8.44)
Using the fact that w and w are independent, the right-hand side is the
minimum of (2.8.32) and

n1 R(, w ) ds d
.
(2.8.45)
inf R S
w
|w |2 ds d
R S n1
Since w ei is orthogonal to one on S n1 , we have

w ei 2 ds (n 1)
|w |2 ds.
S n1
(2.8.46)
S n1
Hence the inmum in (2.8.45) is at most n 1, which exceeds the value in

(2.8.39). The result follows for 1.
For > 1 the proof is similar. Dierentiation of f in gives
1+
16( 1)((n 2)2 + 4x)

,
(4x + 4 + (n 2)2 )2
(2.8.47)
which is positive. Hence the role of the value (2.8.39) is played by the smallest
value of f (, n 1, ) on R+ . Therefore,

inf
v
|v|2
dx
Rn |x|n2
|v|2
dx
Rn |x|n

= 2 + min x +
x0

4(n 1)( 1)
.
x + n 1 + ( n/2)2
227
(2.8.48)

Proof of Corollary 1. We need to consider only > 1. It follows directly

from (2.8.6) that

2
3
1
+ 1 + 2,
C3, =
2
which gives the result.
Remark. Using (2.8.22), we see that a minimizing sequence {vk }k1 ,
which shows the sharpness of inequality (2.8.4) with the constant (2.8.5),
can be obtained by taking vk = (v,k , v,k , 0) with the Fourier transform
wk = (w,k , w,k , 0) chosen as follows:
w,k (, ) = hk () sin ,
w,k (, ) =
1n
hk () cos . (2.8.49)
i + n/2
The sequence {|hk |2 }k1 converges in distributions to the delta function at

= 0. The minimizing sequence that gives the value (2.8.7) of C3, is
w,k (, 0) = 0,
w,k (, ) = 0,
and
w,k (, ) = hk () sin ,
where {|hk |2 }k1 is as previously.

2.8.3 Proof of Theorem 2
The calculations are similar but simpler than those in the previous section.
We start with the substitution v(x) = u(x)|x|2 and write (2.8.14) in the form

2
1
2 |v| dx
= 2 + inf R 2 2 .
(2.8.50)
v
C2,
|v| |x| dx
R2
In polar coordinates and , with [0, 2), we have

2
|v |2 +|v |2 +2 v2 + v2 4v ( v ) dx. (2.8.51)
|v| dx =
R2
R2
Changing the variable to t = log and applying the Fourier transform

v(, ) w(, ), we obtain that the right-hand side is
2

2

+ 1 |w |2 + |w |2 + | w |2
R 0

+ | w |2 4( w )w d d.
(2.8.52)
The divergence-free condition for u becomes
228
w =
which yields

2
|v| dx =
R2
w
,
i + 1

2 + 4 3
+ 1 | w |2
2 + (1 )2
R 0

2

|2 w |2
2
+ 2
+ + 1 |w | d d.
+ (1 )2
2
Analogously,

|v|2 |x|2 dx =

|w |2 + |w |2 d d
R 0

2
| w |2
2
+
|w
|
=
d d.
2 + (1 )2
R 0
R2
(2.8.53)
(2.8.54)
2
(2.8.55)
Therefore, by (2.8.50)
1
= 2 + inf inf f (x, , ),
x0 N0
C2,
where

f (x, , ) = x + 1 + 1

4(1 )
.
x + + (1 )2
(2.8.56)
(2.8.57)
Let rst 1. Then f is increasing in x, which implies f (x, , )

f (0, , ). Since the derivative
4(1 )3
,
f (0, , ) = 1
( + (1 )2 )2
(2.8.58)
is positive for 2, we need to compare only the values f (0, 0, ), f (0, 1, ),

and f (0, 2, ). An elementary calculation
shows that both f (0, 0, ) and
f (0, 2, ) exceed f (0, 1, ) for (1 3, 1 + 3).
Let now > 1. We have
4( 1)(x + (1 )2 )
f (x, , ) = 1 +
>0
(x + + (1 2 ))2
(2.8.59)
and therefore f (x, , ) f (x, 0, ) = x + 1 1. The proof of Theorem 2 is

complete.

Remark. Minimizing sequences that give C2, in (2.8.8) can be chosen as
follows:
w,k (, ) = 0,
w,k (, ) = hk (),
for (1 3, 1 + 3), and
229
hk () sin( 0 )
,
w,k = hk () cos( 0 ),
i + 1
when (1 3, 1 + 3), for any constant 0 . Here {|hk |2 }k1 converges

in distributions to the delta function at 0.
w,k =
Corollary. Let = 0. Denote by a real-valued scalar function in

C0 (R2 ) and assume, in addition, that (0) = 0 if < 0. Then the sharp
inequality

2

2
2(1)
|| |x|
dx C2,
x1 x1 + 2x21 x2 + x22 x2 |x|2 dx (2.8.60)
R2
R2
holds with C2, given in (2.8.8).

Indeed, for n = 2, inequality (2.8.4) becomes (2.8.60) if is interpreted as
a stream function of the vector eld u, i.e., u = .
The results of this section are borrowed from the paper by Costin and Mazya
[214]. In [715], Sobolevskii stated that the sharp constant in Hardys inequality
for arbitrary solenoidal vector functions in a convex domain coincides with the
same constant in the classical one-dimensional case.
3
Conductor and Capacitary Inequalities
with Applications to Sobolev-Type
Embeddings
3.1 Introduction
Let be an open set in Rn and let and be locally nite, nonzero Borel
measures on . We also use the following notation: l is a positive integer,
1 p , q > 0, dx is an element of the Lebesgue measure mn on Rn , and
f is an arbitrary function in C0 (), i.e., an innitely dierentiable function
with compact support in . By Lt we mean the set {x : |f (x)| > t},
where t > 0. We shall use the equivalence relation a b to denote that the
ratio a/b admits upper and lower bounds by positive constants depending
only on n, l, p, and q.
This chapter is a continuation of the previous one. We shall discuss the
conductor inequality

capp (Lat , Lt ) d tp c(a, p)
| grad f |p dx,
(3.1.1)
0
where a = const > 1 and capp is the p-capacity of Lat with respect to Lt , or
in other words, the p conductivity of the conductor L\Lat . (This inequality
was only mentioned in Sect. 2.3.1.)
By monotonicity of capp the conductor inequality (3.1.1) implies the capacitary inequality

p
C(p)
capp (Lt , ) d t
| grad f |p dx,
(3.1.2)
0
which was also proved in Theorem 2.3.1 with the best constant
C(p) = pp (p 1)1p .
(3.1.3)
Note that the left-hand side in (3.1.2) can be zero for all f C0 ().
(This happens if and only if either p > n and = Rn , or p = n and the
complement of has zero n capacity.) At the same time, the left-hand side in
DOI 10.1007/978-3-642-15564-2 3,
231
232
3 Conductor and Capacitary Inequalities
(3.1.1) is always positive if f = 0. The layer cake texture of the left-hand side
in the conductor inequality (3.1.1) admits signicant corollaries that cannot be
directly deduced from the inequality (3.1.2). For instance, as a straightforward
consequence of (3.1.1) and the isocapacitary inequality (2.2.11), one deduces

d(tn )
c(a)
| grad f |n dx,
(3.1.4)
mn (Lt ) n1
(log mn (Lat ) )
0
where n > 1 and a > 1. Note that (3.1.4) is stronger than the well-known
inequality

d(tn )
c
| grad f |n dx,
(3.1.5)
mn () n1
(log m
)
0
(L
)
n
t
(see Mazya [543], Hansson [348], and Brezis and Wainger [146]) which is
informative only if the volume of is nite.
In the case p = n and p > 1, another straightforward consequence of (3.1.1)
with a similar avor is the following improvement of the classical Sobolev
inequality:

pn
pn
mn (Lt ) n(p1) mn (Lat ) n(p1) 1p d tp c(p, a)
| grad f |p dx.
(3.1.6)
An immediate application of the conductor inequality (3.1.1) that seems to
be unattainable with the help of the capacitary inequality (3.1.2) is a necessary
and sucient condition for the two-measure Sobolev-type inequality
1/q

1/p

(x, grad f )p dx +
|f |q d
C
|f |p d
,
(3.1.7)
where q p, , and are locally nite Radon measures on , and the

function Rn (x, y) (x, y) is positively homogeneous in y of degree
1 (see Theorem 2.3.9). The just-mentioned characterization is formulated in
terms of the conductivity generated by the integral

(x, grad f )p dx.
(3.1.8)
In the one-dimensional case and when (x, grad f ) = |f |, the conductivity

is calculated explicitly (see Lemma 2.2.2/2) and this characterization takes the
following simple form:
p/q

const 1p + d+ (x) ,
(3.1.9)
d (x)
where d (x) denotes any open interval (x d, x + d), such that d+ (x) .
In the present chapter, we deal mostly with applications of conductor inequalities to two-measure Sobolev-type embeddings that seem to be unattainable with the help of capacitary inequalities. In particular, we sometimes assume that n = 1 and we study inequalities of the type
3.2 Comparison of Inequalities (3.1.4) and (3.1.5)
1/q

|f | d
q

(l) p
f dx +
1/p
233
|f | d
p
(3.1.10)
and their analogs involving a fractional Sobolev norm.

We conclude the Introduction with a brief outline of the contents of the
chapter. In Sect. 3.2 we discuss the inequalities (3.1.4) and (3.1.5). A proof
of (3.1.1) is given in Sect. 3.3. In Sects. 3.4 and 3.5 we characterize the inequality (3.1.7) and give a criterion for its multiplicative analog. A necessary
and sucient condition for the compactness and two-sided estimates of the
essential norm of the embedding operator associated with (3.1.7) are obtained
in Sect. 3.6.
In Sect. 3.7 we give a necessary and sucient condition for the inequality
1/q

|f |q d

p
f (x) dx +
1/p
|f |p d
(3.1.11)
with 1 < p q < , restricted to nonnegative functions f C0 (). This is

the estimate
p/q

const 12p + d+ (x) ,
(3.1.12)
d (x)
valid for all intervals d+ (x) . A simple example shows that (3.1.12)
does not guarantee (3.1.11) for all f C0 (). We also give counterexamples
showing that the necessary condition for (3.1.10)
p/q

const 1lp + d+ (x)
d (x)
(3.1.13)
is not sucient if l 3.
Section 3.8 is dedicated to multidimensional (p, l)-conductivity inequalities
for fractional Sobolev Lp norms of order l in (0, 1) and (1, 2). The section is
concluded with the necessary and sucient conditions for two-measure multidimensional inequalities of type (3.1.7) involving fractional norms.
3.2 Comparison of Inequalities (3.1.4) and (3.1.5)

Inequalities (3.1.4) and (3.1.5) follow directly from (3.1.1) combined with the
isocapacitary inequality (2.2.11).
Let us compare the integrals on the left-hand sides of (3.1.4) and (3.1.5):

d(tn )
,
(3.2.1)
mn (Lt ) n1
(log m
)
0
n (Lat )
and
d(tn )
(log
mn () n1
mn (Lt ) )
(3.2.2)
234
where mn () < . Clearly, the rst of them exceeds the second. However,
the convergence of the second integral does not imply the convergence of the
rst. In fact, let Br = {x Rn : |x| < r}, = B2 and

5 |x| for 0 |x| < 1,
f (x) =
(3.2.3)
2 |x| for 1 |x| < 2.
B2t
Lt = B1
B5t
We have
for 0 t < 1,
for 1 t 4,
for 4 < t 5.
Let 1 < a < 4. Then both sets Lt and Lat for 1 < t < 4a1 coincide with
the ball B1 which makes (3.2.1) divergent whereas integral (3.2.2) is nite.
Furthermore, integral (3.2.1) is convergent for a 4.
Therefore, inequality (3.1.4) is strictly better than (3.1.5), even for domains of nite volume. We see also that the convergence of integral (3.2.1)
for a bounded function f may depend on the value of a.
The same argument shows that inequality (3.1.6) with any f C0 ()
and 1 < p < n, i.e.,

d(tp )
c
| grad f |p dx,
(3.2.4)
1
1
p1
(
np
np )
0
mn (Lat ) n(p1)
mn (Lt ) n(p1)
np
improves the Lorentz space L np
,p () inequality
np
mn (Lt ) n d tp c

| grad f |p dx,
which results from the capacitary inequality (2.3.6) and is stronger, in its
turn, than the Sobolev estimate for the norm f L pn .
np
In conclusion, we add that the convergence of the integral on the left-hand

side of (3.2.4) may depend on the choice of a, as shown by the function (3.2.3).
3.3 Conductor Inequality (3.1.1)
Let g and G denote arbitrary bounded open sets in Rn subject to g G, G

. We introduce the p conductivity of the conductor G\
g (in other terms, the
relative p-capacity of the set g with respect to G) as

grad (x)p dx : C0 (G), 0 1 on G
capp (
g , G) = inf

and = 1 on a neighborhood of g .
(3.3.1)
3.3 Conductor Inequality (3.1.1)
235
This inmum does not change if the class of admissible functions is enlarged
to

(3.3.2)
C () : 1 on g, 0 on \G
(see Sect. 2.2.1).
Now we derive a generalization of the conductor inequality (3.1.1).
Proposition. For all f C0 () and for an arbitrary a > 1 inequality
(3.1.1) holds with
p log a
.
c(a, p) =
(a 1)p
Proof. We show rst that the function t capp (Lat , Lt ) is measurable.
Let us introduce the open set S := {t > 0 : | grad f | > 0 on Lt } whose complement has zero one-dimensional Lebesgue measure by the Morse theorem.
Let t0 S. Given an arbitrary > 0, there exists a function C0 (Lt0 ),
= 1 on a neighborhood Lat0 , and such that
grad pLp capp (Lat0 , Lt0 ) + .
Since t0 S we deduce from (3.3.1) that for all suciently small > 0
grad pLp capp (La(t0 ) , Lt0 + ).
Therefore,
capp (La(t0 ) , Lt0 ) capp (Lat0 , Lt0 ) + ,
which means that the function t capp (Lat , Lt ) is upper semicontinuous
on S. The measurability of this function follows.
Let denote a locally integrable function on (0, ) such that there exist
the limits (0) and (). Then there holds the identity

dt
= (0) () log a.
(3.3.3)
(t) (at)
t
0
Setting here

| grad f |p dx,
(t) :=
Lt
we obtain

| grad f |p dx
1
log a

0

Lt \Lat
| grad f |p dx
dt
.
t
By (3.3.1) the right-hand side exceeds

(a 1)p
capp (Lat , Lt ) d tp ,
p log a 0
236
3.4 Applications of the Conductor Inequality (3.1.1)

The following lemma, which can be easily obtained from (3.1.1), is a particular
case of Theorem 2.3.9.
Lemma 1. Let 1 p q. The inequality

1/q

1/p

q
p
p
|f | d
C
| grad f | dx +
|f | d
(3.4.1)
holds for all f C0 () if and only if there exists a constant K > 0 such that
, the inequality
for all open bounded sets g and G, subject to g G, G

1/p
g , G) + (G)
(3.4.2)
(g)1/q K capp (
holds.
From this lemma, we shall deduce a sucient condition for (3.4.1) which
does not involve the p conductivity.
Corollary. Let n < p q. If for all bounded open sets g and G in Rn
, we have
such that g G, G

1/p
(g)1/q K dist(g, G)np + (G)
,
(3.4.3)
then (3.4.1) holds for all f C0 ().
Proof. Let be an arbitrary admissible function in (3.3.1). By Sobolevs
theorem 1.1.10/1, for all y g and z \G

p

p
1 (y) (z) c|y z|np grad (x) dx,
which implies

np
c capp (
g , G).
dist(g, G)
It remains to refer to Lemma 1.
Let us see how criterion (3.1.9) follows from Lemma 1.

Theorem 1. Let n = 1 and 1 p q < . The inequality (3.4.1) holds
for all f C0 () if and only if condition (3.1.9) is satised.
Proof. Let g0 = (a, b), G0 = (A, B), and A < a < b < B. It is an easy
exercise to show that
capp (
g0 , G0 ) = (a A)1p + (B b)1p .
(3.4.4)
(For the proof of a more general formula for a weighted p conductivity see
Lemma 2.2.2/2.) Hence, by setting g = d (x) and G = d+ (x) into (3.4.2),
we obtain
237

1/q

1/p
d (x)
K 2 1p + d+ (x)
,
which implies the necessity of (3.1.9). To prove the suciency we need to
obtain (3.4.2) for all admissible sets g and G. Let G be the union of nonoverlapping intervals Gi and let gi = Gi g. Denote by hi the smallest interval
containing gi and by i the minimal distance from hi to R\Gi . By denition
(3.3.1) in the one-dimensional case, we have
i , Gi )
capp (
gi , Gi ) = capp (h
and
g , G) =
capp (
capp (
gi , Gi ).
Hence, and by (3.4.4) applied to the intervals hi and Gi ,

1p
g , G)
i .
capp (
(3.4.5)

1/q
,
(gi )1/q (hi )1/q A i1p + (Gi )
where A is a positive constant independent of g and G. Since q p, we have

(g)p/q
(gi )p/q ,
i
which, together with (3.4.5), implies

1p

i
+ (Gi ) Ap capp (
g , G) + (G) .
(g)p/q Ap
i
The result follows from Lemma 1.
Remark 1. A possible modication of Theorem 3.4 concerns the Birnbaum

Orlicz space LM (), where M is an arbitrary convex function on (0, ),
M (+0) = 0. Let P denote the complementary convex function to M . One
can easily deduce from Theorems 2.3.3 (compare with Theorem 3.4) that the
condition

1

1/p

1
const 1p + d+ (x)
d (x) P
(d (x))
is necessary and sucient for the inequality

1
(L )P 1
|f |p dx +
|f |p d
d p c
(L )
0
as well as for the inequality
238

p
|u|
LM ()
|f |p dx +

|f |p d .
It is well known that the weight w in the integral

p
f (x) w(x) dx
can be removed by the change of the variable x

dx
.
=
w(x)1/(p1)
Therefore, Theorem 3.4 leads to a criterion for the three-weight inequality
1/q

|f | d
1/p
p
|f | d +
|f | d
p
where is a nonnegative measure. Note that the singular part of does not
inuence the validity of the last inequality (compare with Muckenhoupt [620]
and Sect. 1.3.2).
Remark 2. Let n = 1. With p (1, ) and the measure , we associate a
function R of an interval d (x) by the equality

R d (x) = sup : 1p > d+ (x)
(3.4.6)
with d+ (x) as everywhere. Clearly,
1p

1p

inf : 1p + d+ (x) 2R d (x)
,
R d (x)
(3.4.7)
which shows that criterion (3.1.9) can be written as

(p1)/p
1/q
R d (x)
d (x)
< .
sup
d (x)
Remark 3. According to Theorem 2.1.3, inequality (3.4.1) with p = 1, q 1

is equivalent to the inequality

(g)1/q C 2 + (g) ,
where g is an arbitrary interval and C is the same constant as in (3.4.1).
Similarly to (3.4.1) with n = 1, we can characterize the inequality

1/q
|f |q d

C
/p
|f |p dx
(1)/r
|f |r d
by using the following particular case of Theorem 2.3.11.
(3.4.8)
239
Lemma 2. Let n 1, p 1, and [0, 1]. If the inequality

1/q

|f | d

/p
| grad f | dx
(1)/r
|f | d
(3.4.9)
is valid for all f C0 () and some positive r and q, then there exists a
constant such that for all open bounded subsets g and G of such that
, there holds the inequality
g G, G
(g)p/q capp (
g , G) (G)(1)p/r .
(3.4.10)
If (3.4.10) holds for all g and G as above, then (3.4.9) is valid for all
functions f C0 () with 1/q (1 )/r + /p.
Arguing as in the proof of Theorem 3.4, we arrive at the following criterion
for (3.4.8).
Theorem 2. Let n = 1, p 1, and [0, 1]. If the inequality (3.4.8) holds
for all f C0 () and some positive r and q, then there exists a constant
> 0 such that

1/q
d (x)

(1)/r
d+ (x)
(p1)/p
(3.4.11)
for all x , d > 0 and > 0 such that d+ (x) . Conversely, if (3.4.11)
is true for some positive r and q such that 1/q (1 )/r + /p, then (3.4.8)
holds.
Note that for p = 1 condition (3.4.11) is simplied:
r/q(1)

const d (x) .
d (x)
For the particular case = , inequality (3.4.8) admits the following
simpler characterization that results from Theorem 2.3.6.
Theorem 3. 1. Let n = 1 and let for all x , d > 0 and > 0 such
that d+ (x) , there holds

(3.4.12)
d (x) const (1p)/p ,
where p 1 and > 0. Furthermore, let q be a positive number satisfying
one of the conditions (i) q 1 if p 1 or (ii) q < 1 if p > 1. Then
the inequality

1/q
|f |q d

C
/p
|f |p dx
(1)/r
|f |r d
(3.4.13)
with r (0, q) and = (q r)/(1 r)q is valid for any function f C0 ().
240
2. Conversely, let p 1, > 0 and r (0, 1 ]. Furthermore, let the

inequality (3.4.13) with = (qr)/(1r)q hold for any function f C0 ().
Then (3.4.12) holds for all x and d such that d+ (x) .
Remark 4. By Theorem 3 the multiplicative inequality (3.4.13) is equivalent to
1/q

1/p

|f |q d
C
|f |p dx
with p q.
The next assertion concerning an arbitrary charge (not a nonnegative
measure as elsewhere) follows directly from Theorem 2.3.8.
Theorem 4. Let + and denote the positive and negative parts of the
charge , respectively.
(i) If for a certain (0, 1) and for any x , d > 0, > 0, such that
d+ (x) , there holds

+ d (x) C 1p + (1 ) d+ (x) ,
where p > 1, then for all f C0 ()

p
|f | d C
|f |p dx.
(3.4.14)
(ii) If (3.4.14) is true for any x , d > 0, > 0 such that d+ (x) ,
then

(3.4.15)
+ d (x) C 1p + d+ (x) .
Example. We show that (3.4.15) is not sucient for (3.4.14). Let + and
be the Dirac measures concentrated at the points 0 and 1, respectively.
We introduce the sequence of piecewise linear functions {m }
m=1 on R by
m (x) = 0 for |x| > mp/(p1) ,
m (0) = 1,
Then

R
|m |p d =

p
1 + o(1)
m
m (1) = 1 m1 .

and
R
|m |p dx mp
as m ,
and therefore (3.4.14) fails. However, (3.4.15) holds with C = 1. To check this,
we need to consider only the case + (d (x)) = 1 and (d+ (x)) = 0, when
clearly 1 and 1p 1.
3.5 A p-Capacity Depending on and its Applications
241
3.5 p-Capacity Depending on and Its Applications to

a Conductor Inequality and Inequality (3.4.1)
Let n 1 and let K denote a compact subset of . We introduce a relative
p-capacity of K with respect to , depending on the measure , by

p
p
|| d ,
(3.5.1)
capp (K, , ) = inf grad Lp +
where the inmum is extended over all functions C0 () such that 1

on K. Arguing as in Sect. 2.3.1, one can show that the inmum in (3.5.1) will
be the same if the set of admissible functions is replaced by { C0 () :
= 1 on K, 0 1 on }.
Making small changes in the proof of Proposition 3.3, one arrives at the
inequality

capp (Lat , Lt , ) d tp c(p) grad f pLp +
|f |p d ,
0
C0 ().
By this inequality one can easily obtain

where a = const > 1 and f
the following condition, necessary and sucient for (3.4.1) with q p:
(g)p/q const capp (g, , )
(3.5.2)
for all bounded open sets g with g .

The next lemma shows directly that (3.5.2) is equivalent to (3.4.2).
Lemma. There holds the equivalence relation

capp (K, , ) inf capp (K, G) + (G) ,
(3.5.3)
where the inmum is taken over all bounded open sets G such that K G
.
and G
Proof. Let > 0, f C0 (), f = 1 on K, 0 f 1 on , and let

|f |p d.
capp (K, , ) + grad f pLp +
Then
capp (K, , ) +
2p(k+1)
k=0
k=0
L2k1 \L2k

grad 2k+1 f 1 p dx +

2pk capp (L2k , L2k1 ) + (L2k1 ) .
1
0

(Lt ) d tp
242
Since capp (L2k , L2k1 ) capp (K, L2k1 ), it follows that

capp (K, , ) + c inf capp (K, G) + (G) .
G
The estimate
capp (K, , ) capp (K, G) + (G)

is obvious. The result follows.

We introduce the capacity minimizing function
Sp (t) = inf capp (g, , ),
where the inmum is taken over all bounded open sets g, g , satisfying
(g) > t. By the last lemma,

Sp (t) inf capp (g, G) + (G)
g,G
,
with the inmum extended over open sets g and G such that g G, G
and (g) > t. Obviously, the condition (3.5.2) is equivalent to
sup
tp/q
< .
Sp (t)
Making trivial changes in the proof of Theorem 2.3.8, we arrive at the

condition, necessary and sucient for (3.4.1) with 0 < q < p, p 1:

0
tp/q
Sp (t)
q/(pq)
dt
< .
t
(3.5.4)
It follows from the proof of Theorem 3.4/1 that in the one-dimensional

case there holds the equivalence relation

Sp (t) inf : 1p + d+ (x)
with the inmum taken over all x, d, and such that d+ (x) and

(3.5.5)
d (x) > t.
By (3.4.7),

1p
,
Sp (t) inf R d (x)
where the inmum is taken over all intervals d (x), d (x) , satisfying
(3.5.5).
3.6 Compactness and Essential Norm
243
3.6 Compactness and Essential Norm

1 () as the closure of C () with respect to the norm
We dene the space W
p
0

f W
1 () =
p
p
f (x) dx +

f (x)p d
1/p
.
The condition (3.1.9) is a criterion of boundedness for the embedding operator

1 () Lq ()
Ip,q : W
p
for q p 1.
In this section we establish a compactness criterion for Ip,q with q p 1
and obtain sharp two-sided estimates for the essential norm of Ip,q . We recall
that the essential norm of a bounded linear operator A acting from X into Y ,
where X and Y are linear normed spaces, is dened by
ess A = inf A T ,
T
with the inmum taken over all compact operators T : X Y .

Theorem 1. If q p 1, then the operator Ip,q is compact if and only if
(d (x) \ [M, M ])1/q
= 0.
M x,,d { 1p + (d+ (x))}1/p
lim sup
(3.6.1)
Proof. Suciency. Let stand for the restriction of to the segment

[M, M ] and let M = M . We dene the embedding operators
1 () Lq (M ) and
IM : W
p

1 () Lq ( ),
IM
:W
p
M
as well as the embedding operators

iM : Lq (M ) Lq ()
We have
We prove that
and
iM : Lq (M ) Lq ().

Ip,q = iM IM + iM IM
.

IM
(3.6.2)
is compact. Consider the embedding operators

C
1 () C [M, M ] ,
:W
IM
p

iC
M : C [M, M ] Lq (M ),
where C([M, M ]) is the space of continuous functions with the usual norm.

C
C
= iC
Clearly, IM
M IM . Since IM is compact for any M > 0 by the Arzela
244

theorem, the operator IM
is compact as well. The condition IM 0 as
M is equivalent to (3.6.1) owing to Theorem 3.4/1.
1 ().
Necessity. Let Ip,q be compact and let B denote the unit ball in W
p
The set Ip,q B is relatively compact in Lq (). Therefore, for any > 0 there
exists a nite -net {fj }N
j=1 Ip,q B = B for the set Ip,q B. Given any fj ,
there exists a number Mj () such that

fj (x)q d(x) < q .
|x|>Mj ()
Let M () be equal to supj Mj (). Then for any f B and for some i {1, N }
we have
1/q

1/q

q
f (x)q dM () (x)

fj (x) dMj () (x)
f fj Lq () +
< 2.
Hence inequality (3.1.10) holds, where l = 1, M () and 2 instead of and C.
Now (3.6.1) follows from the necessity part in Theorem 3.4/1.

Theorem 2. Let q p 1 and
(d (x) \ [M, M ])1/q
.
M x,,d { 1p + (d+ (x))}1/p
E(, ) := lim sup
There exist positive constants c1 and c2 such that

c1 E(, ) ess Ip,q c2 E(, ).
(3.6.3)
Proof. We use the same notation as in the previous theorem. The upper bound in (3.6.3) is a consequence of the suciency part in the proof of
Theorem 3.4/1.
1 () Lq () and let be any positive
Let T be any compact operator: W
p
number. We choose T to satisfy
ess Ip,q Ip,q T .
There exists a positive M () such that for any f B

T f (x)q dM () (x) < q .
We introduce the truncation operator M : Lq () Lq (M ) by

0,
|x| < M,
(M f )(x) =
f (x), |x| M.
(3.6.4)
(3.6.5)
3.7 Inequality (3.1.10) with Integer l 2
245
Using (3.6.4) and (3.6.5), we obtain

ess Ip,q IM () M () T IM () 2.
By Theorem 3.6,
M () (d (x))1/q
c1 E(, ).
1p + (
1/p
x,,d {
d+ (x))}
IM () c sup
The result follows.

Let us deduce a characterization of inequality (3.1.11) for nonnegative functions.
Theorem. Let n = 1 and 1 0 such that

1/q

1/p
d (x)
K 12p + d+ (x)
,
(3.7.1)
for all x , d > 0 and > 0 satisfying d+ (x) .
Proof. To prove the necessity, we set a function f in (3.1.11), which is
subject to f C0 (), f = 1 on d (x), f = 0 outside d+ (x), and 0
f (x) 1 on . Then
d (x)
1/q

C
p

f (y) dy + d+ (x)
1/p
.
d+ (x)\d (x)
Clearly, f can be chosen on d+ (x)\d (x) so that the integral on the right
does not exceed c(p)r12p . The estimate (3.7.1) follows.
For the proof of suciency we need the following lemma.
Lemma. If f D(R), f 0, then f 1/2 L12p (R) and

p

2p 1
|f |2p
dt
|f |p dt,
p
p1
R f
R
(3.7.2)
where the integration is taken over the support of f . The constant factor on
the right-hand side is optimal.
Proof. Obviously, for > 0,

1
|f |2p
dt =
|f |2p2 f (f + )1p dt.
p
1p R
R (f + )
246
Integrating by parts, we obtain

2p 1
|f |2p
|f |2(p1)
dt
=
f dt
p
p 1 R (f + )p1
R (f + )

1/p
(p1)/p
2p 1
|f |2p
p
|f | dt
dt
.
p
p1
R
R (f + )
Therefore
|f |2p
dt
(f + )p
2p 1
p1
p
R
|f |p dt.
It remains to pass to the limit as +0.

Let f (x) = x+21/p (x), where > 0, C0 (R) and = 1 near the
origin. Putting the family {f } into (3.7.2) we see that the constant factor on
the right-hand side of (3.7.2) is sharp. The lemma is proved.

Let us turn to the proof of suciency of (3.7.1). Let f C0 () satisfy
supp f and f 0. Then u = f 1/2 satises

|u|
2q
1/2q

1/2
d
cK
2p
|u |
|u|
2p
dx +
1/2p
by Theorem 3.4/1. This inequality, along with (3.7.2) gives (3.1.11).
The proof of the following conductor inequality, involving nonnegative

functions, is based upon the smooth level truncation and the last lemma.
Proposition. Let n 1, f C0 (), f 0, a = const > 1, and p > 1.
Then

p
c(p,
a)
cap+
(L
,
L
)
d
t
| grad2 f |p dx,
(3.7.3)
at
t
p,2
0
where grad2 = { 2 /xi xj }ni,j=1 and

g , G)
cap+
p,2 (
= inf

grad2 (x)p dx : C (G), 1 0 on G,
0
G

= 1 in a neighborhood of g .
(3.7.4)
(Concerning the measurability of the function t cap+

p,2 (Lat , Lt ) see the
beginning of the proof of Proposition 3.3.)
Proof. Let H C 2 (R),

H(x) =
0 for x < ,
1 for x > 1 ,
247
where is an arbitrary number in (0, 1). By (3.7.4),

p

+
dx
grad2 H f (x) t
capp,2 (Lat , Lt )

(a 1)t

| grad f |2p
c(a)
p
dx.
p
+
|
grad
f
|
2
t
fp
Lt \Lat
Hence the left-hand side in (3.7.3) is dominated by

| grad f |2p
dt
p
pc(a)
+ | grad2 f | dx .
p
f
t
0
Lt \Lat
Owing to (3.3.3), this can be written as

| grad f |2p
p
+ | grad2 f | dx,
pc(a) log a
fp
which does not exceed the right-hand side of inequality (3.7.3) in view of
(3.7.2). The result follows.

Remark. A direct generalization of this proof for derivatives of higher than
second order is impossible since there is no analog of the inequality (3.7.2) for
higher derivatives. In fact, the example of a function R1 t u C0 (R1 ),
u 0, coinciding with t2 for |t| < 1, shows that the niteness of the norm
u(l) Lp (R1 ) does not imply the niteness of the integral

(j) pl/j
u (t) u(t)p(jl)/j dt
R1
for l > 2.
Nevertheless, in Sect. 11.2.1 it will be shown that the smooth truncation
can be used in the proof of a capacitary inequality for = Rn and any integer
l > 0 being applied not to an arbitrary nonnegative function but to a potential
with nonnegative density.
Example 1. Let us show that condition (3.7.1) is not sucient for (3.1.11)
with p = 1. Let be Diracs measure concentrated at x = 0 and let d(x) =
(1 + x2 )1 dx. Obviously, condition (3.7.1) holds. We construct a sequence
of nonnegative functions m C0 (R), m = 1, 2, . . . , dened by m (x) =
m (x m 1), where m is a smooth, nonnegative, even function on R,
vanishing for x m + 1 and such that m (x) = m + 1 x for 1 x m.
Then m (0) = 0,

q
|m | dx = const,
m
d ,
R
i.e., inequality (3.1.11) with p = 1 fails.
248
Example 2. We shall check that (3.7.1) does not suce for (3.1.11) to be
valid for all f C0 (R) if p 1. Let and be Diracs measures concentrated
at 0 and 1, respectively. Consider the function 0 C0 (R) such that 0 (x) =
x for x [1, 1]. We set m (x) = 0 (x/m). Then

R
1/q
|m |q d
= m1 ,

R
1/p
|m |p dx
= cm2+1/p ,
and (3.1.11) fails for p > 1. The case p = 1 was treated in Example 1.
Example 3. Now we consider the case of the derivative of order l 3 in
inequality (3.1.10) for all f C0 () such that f (x) 0 on . By the obvious
relation
b

(l) p
f (x) dx : f C [a, b], f (x) 0, f (a) = 0, f (b) = 1
inf
a
= cl,p (b a)1lp ,
we obtain the following necessary condition for (3.1.10)
sup

1/q
d (x)

inf
1pl

+ d+ (x)
1/p
< .
(3.7.5)
d+ (x)
x,d>0
We shall verify that this condition is not sucient for (3.1.10) when l 3 and
p 1.
Suppose rst that p > 1. Let and be Diracs measures concentrated
at 0 and 1, respectively. Then (3.7.5) holds. Let 0 be a nonnegative function
in C0 (R) such that 0 (x) = x2 for |x| 1. We put m (x) = 0 (x/m),
m = 1, 2, . . . . Then

(l) p
(x) dx = cm1pl ,
|m |q d = m2q ,
R
and inequality (3.1.10) fails.

Consider the remaining case l = 3, p = 1. Let be Diracs measure at O.
Then (3.7.1) has the form

1/q
sup 1 + x2 (x, )
< const.
x
For d(x) = (1 + |x|)2q1 the last condition holds.

We introduce the sequence {m (x)}m1 by
x
m (x) =
m (t) dt for |x| 2m + 2,
0
where m is the same as in Example 2. For |x| 2m + 2 we dene m so that

m 0 and
3.8 Two-Weight Inequalities Involving Fractional Sobolev Norms
249
(3)
(t) dt < .
sup
2m+2
We see that

R
(3)
dt =
m
2m+2
|m | dt +
(3)
dt < ,
2m+2
and inequality (3.1.10) with p = 1, l = 3 does not hold.
3.8 Two-Weight Inequalities Involving Fractional

Sobolev Norms
Consider the inequality

1/q
Rn
|f |q d

c f pp,l +
Rn
1/p
|f |p d
(3.8.1)
where f C0 (Rn ), p 1, 0 < l < 1, and

|f (x) f (y)|p
dx dy.
f pp,l =
|x y|n+pl
Rn Rn
As is well known (Uspenski [770]), any smooth extension of f onto Rn+1
admits the estimate

f pp,l c
|xn+1 |p(1l)1 | grad F |p dx dxn+1 ,
(3.8.2)
Rn+1
and there exists a linear extension operator f F C (Rn+1 ), where F

decays to 0 at innity and such that

|xn+1 |p(1l)1 | grad F |p dx dxn+1 cf pp,l .
(3.8.3)
Rn+1
The same argument as in Proposition 3.3 leads to a conductor inequality,

similar to (3.1.1), for the integral

|xn+1 |p(1l)1 | grad F |p dx dxn+1 ,
(3.8.4)
Rn+1
with the left-hand side involving the conductivity generated by (3.8.4) (compare with (3.3.1)).
Minimizing (3.8.4) over all extensions of f and using (3.8.2) and (3.8.3),
we arrive at the fractional conductor inequality

capp,l (Lat , Lt ) d tp c(l, p, a)f pp,l ,
(3.8.5)
0
250
where a > 1 and

g , G) = infpp,l ,
capp,l (
(3.8.6)
with the inmum taken over all C0 (G) subject to = 1 on g, = 0

outside G, and 1 0 on G. This inmum does not change if one requires
C0 (Rn ), 1 on g, and 0 outside G.
By (3.8.5) we obtain the following criterion for (3.8.1).
Theorem 1. Let 1 p q. Inequality (3.8.1) holds for all f C0 (Rn )
if and only if there exists a constant K such that for all open bounded sets g
and G subject to g G there holds

1/p
g , G) + (G)
.
(g)1/q K capp,l (
(3.8.7)
Remark. The last criterion can be simplied for p = 1, q 1, as follows:

dx dy
+
(g)
(g)1/q K
npl
g Rn \g |x y|
for all open bounded sets g. In fact, the necessity results by setting the characteristic function of g into (3.8.1). The suciency follows from

u1,l = 2
|u(y)|
dt
|u(x)||u(y)|

=2
0
Lt
Rn \Lt
|u(x)|
dx dy
|x y|n+l
dx dy
dt
|x y|n+l
combined with (3.8.7) where p = 1.

We turn to the inequality

1/q

p
|f | d
c grad f p,1+l +
1/p
Rn
|f | d
p
Rn
(3.8.8)
where f C0 (Rn ), f 0, and 0 < l < 1.

Lemma 1. Let f be a nonnegative function on R with absolutely continuous f . Suppose the weight function w belongs to the Muckenhoupt class Ap ,
p (1, ). Then
2
(f )

cf wLp (w dx) .
(3.8.9)
f
Lp (w dx)
Proof. Since M g 2Mg, where Mg is the centered maximal function
3.8 Two-Weight Inequalities Involving Fractional Sobolev Norms
(Mg)(x) = sup
>0
1
2
251
x+

g(y) dy
and M are one-sided maximal functions introduced in (1.3.57) and (1.3.58),

the result follows from (1.3.59) and the weighted norm inequality for the
maximal function due to Muckenhoupt [621].

Corollary 1. Let the conditions in Lemma 1 hold and let C 2 (R) be
such that

supti1 (i) < , i = 1, 2.
t>0
Then

(f )
Lp (w dx)
cf Lp (w dx) .
Lemma 2. Let F C (Rn+1 ) and F 0. Then there exists a positive

constant c = c(n, p, l) such that

| grad F |2p
|xn+1 |p(1l)1
dx dxn+1
Fp
Rn+1

c
|xn+1 |p(1l)1 | grad2 F |p dx dxn+1 .
(3.8.10)
Rn+1
Proof. The inequality (3.8.10) with (F/x1 , . . . , F/xn ) instead of

grad F on the left-hand side follows immediately from (3.7.2). To estimate the
integral involving only the derivative F/xn+1 we need the next inequality
for nonnegative functions of one variable:

p

|f (t)|2p
|t|p(1l)1
dt
c
|t|p(1l)1 f (t) dt,
(3.8.11)
p
f (t)
R
R
which is a particular case of Lemma 1 since the weight |t|p(1l)1 belongs to

the Muckenhoupt class Ap .
We state a direct corollary of Lemma 2.
Corollary 2. Let F be the same as in Lemma 2 and let h be a function
in C 1,1 (0, ) such that C := sup{t > 0 : |h (t)| + |t||h (t)| < }. Then

|xn+1 |1l1/p grad2 h(F )
cC |xn+1 |1l1/p grad2 F L (Rn+1 ) .
L (Rn+1 )
p
Let f C0 (Rn ), f 0. The standard extension operator with nonnegative radial kernel gives a nonnegative extension F C (Rn+1 ) of f satisfying

|xn+1 |1l1/p grad2 F
cf p,1+l .
Lp (Rn+1 )
252
Therefore, arguing as in the proof of Proposition 3.7 and using the last inequality and the trace inequality (3.8.2), we arrive at the conductor inequality

p
c(l, p, a)f pp,1+l ,
cap+
(3.8.12)
p,1+l (Lat , Lt ) d t
0
where

cap+
g , G) = inf pp,1+l : C0 (G), 1 0 on G,
p,1+l (

Using (3.8.12), we arrive at the following criterion.
Theorem 2. Let 1 p q. The inequality

1/q
|f | d
q
Rn

c f pp,1+l +
1/p
|f | d
p
Rn
holds for all nonnegative f C0 (Rn ) if and only if there exists a constant K
such that

1/p
g , G) + (G)
(g)1/q K cap+
p,1+l (
for all open bounded sets g and G subject to g G.
3.9 Comments to Chap. 3

The material of this chapter is mostly borrowed from the authors paper [562].
(For p = 2 inequality (3.1.2) with C(2) = 4 was used without explicit formulation already in [531, 534], and [543].)
Capacitary inequality (3.1.2) and its various extensions are of independent
interest and have numerous applications to the theory of Sobolev spaces, linear
and nonlinear partial dierential equations, calculus of variations, theories of
Dirichlet forms, and also in Probability Theory, especially in connection with
multidimensional concentration phenomena and the problems on the convergence of Markov semigroups, and so on (see Mazya [543, 551]; D.R. Adams [5];
Dahlberg [219]; Hansson [348]; Kolsrud [441, 442]; Netrusov [631]; Rao [670];
D.R. Adams and Pierre [18]; Kaimanovich [410]; Mazya and Netrusov [572];
Vondracek [783]; D.R. Adams and Hedberg [15]; Amghibech [44]; Aikawa [36];
Verbitsky [774, 775]; Hansson, Mazya, and Verbitsky [349]; Grigoryan [324];
Takeda [739]; Fitzsimmons [280]; Hajlasz [341]; Fukushima and Uemura [294,
295]; Adams and Xiao [20, 21]; Ben Amor [82]; M. Chen [183, 184]; Barthe,
Cattiaux, and Roberto [75]; Mazya and Verbitsky [594]; Xiao [800]; Cattiaux,
Gentil, and Guillin [177]; Phuc and Verbitsky [658]; Bobkov and Zegarlinski
[120]; et al.).
It is, perhaps, worth mentioning that the proof of (3.1.1) is so simple and
generic that it works in a much more general frame of analysis on manifolds
253
and metric spaces (see Grigoryan [324], Hajlasz [341], et al.). We shall discuss
such generalizations in the next chapter.
Inequality (3.1.10) and its applications were the subject of extensive work.
See, for example, books by Mazya [556]; Mynbaev and Otelbaev [622]; Opic
and Kufner [646]; Davies [224]; Kufner and Persson [469]; Kufner, Maligranda,
and Persson [468]; papers by Muckenhoupt [620]; Mazya [543]; Otelbaev [653];
Oinarov [644]; Davies [226]; Nasyrova and Stepanov [626]; Chua and Wheeden
[190]; Mazya and Verbitsky [593]; Stepanov and Ushakova [727]; Prokhorov
and Stepanov [667]; and references therein.
Inequality (3.1.1) was generalized to SobolevLorentz spaces by Costea
and Mazya [213] as follows:

capp,q (Lat , Lt ) d tp c(a, p, q)f pLp,q (,mn ,Rn ) when 1 q p
0
and

0

capp,q (Lat , Lt )q/p d tq c(a, p, q)f qLp,q (,mn ,Rn )
when p < q <
with capp,q standing for the capacity generated by the SobolevLorentz norm.
In [213] one can nd generalizations of other results in Sect. 3.1 to the context
of SobolevLorentz spaces.
4
Generalizations for Functions on Manifolds
and Topological Spaces
4.1 Introduction
The results and arguments in Chaps. 2 and 3 based upon isoperimetric and
isocapacitary inequalities for sets in the Euclidean space can be readily extended to much more general situtations; the present chapter only touches
such opportunities.
Applications to estimates of the principal eigenvalue of the Dirichlet
Laplacian on a Riemannian manifold are presented in Sects. 4.2 and 4.3.
In Sects. 4.44.6 we derive some conductor inequalities for functions dened
on a locally compact Hausdor space X . It is worth mentioning that, unlike
the Sobolev inequalities, the conductor inequalities do not depend on the dimension of X . Furthermore, with a lower estimate for the p-conductivity by
a certain measure on X , one can readily deduce the SobolevLorentz type
inequalities involving this measure.
In Sect. 4.4 we are interested in conductor inequalities for the Dirichlettype integral

Fp [f ],
(4.1.1)
X
where Fp is a measure-valued operator acting on a function f and satisfying

locality and contractivity conditions. A prototype of (4.1.1) is the functional

(x, grad f )p dx +
|f |p d,
(4.1.2)
with a function y (x, y), positively homogeneous of order 1.

In Theorem 4.4 we obtain the conductor inequality

p
dt
1
M t capp (Lat , Lt )
Fp [f ];
M
c(a, p)
t
0
X
(4.1.3)
here and elsewhere Lt = {x X : |f (x)| > t}, M is a positive convex function

on (0, ), M (+0) = 0, and M 1 stands for the inverse of M . By capp we
mean the p-capacity generated by the operator Fp .
DOI 10.1007/978-3-642-15564-2 4,
255
256
In Sect. 4.5 we derive the conductor inequality

capp, (Lat , Lt )

q/p q p/q
d t
c(a, p, q)f pp, ,
(4.1.4)
where q p 1,

f pp,
1/p
:=
X

f (x) f (y)p (dx dy)
(4.1.5)
and capp, is the relative p-capacity corresponding to the seminorm (4.1.5).

We apply (4.1.4) to obtain a necessary and sucient condition for a twomeasure Sobolev inequality involving f pp, .
In Sect. 4.6 we handle variants of the sharp capacitary inequality (2.3.6).
We show in Theorem 4.6/2 that a fairly general capacitary inequality is a
direct consequence of a one-dimensional inequality for functions with the rst
derivative in Lp (0, ). A corollary of this result is the following inequality
with the best constant, complementing (2.3.6):

1/q
capp (Lt , )q/p d tq
1/p
1/p1/q
pq
( qp
)
| grad f |p dx
q
( qp
) (p q1
qp )
(4.1.6)
where q > p 1. Combined with the isocapacitary inequality, (4.1.6) with

q = pn/(n p), n > p, immediately gives the classical Sobolev estimate

|f |
pn
np
1p/n
dx

c
| grad f |p dx
(4.1.7)
with the best constant (see Sect. 2.3.5).

Another example of the application of Theorem 4.6/2 is the inequality

(4.1.8)
exp c capp (Lt , )1/(1p) d exp ctp/(p1) < ,
sup
0
where c = const, the supremum is taken over all f C0 () subject to

grad f Lp () 1.
Inequality (4.1.8) with p = n is stronger than the sharp form of the Yudovich inequality [809] due to Moser [618], which immediately follows from
(4.1.8) and an isocapacitary inequality. Yudovichs inequality was rediscovered
by Pohozhaev [662] and Trudinger [764].
A capacitary improvement of the FaberKrahn isoperimetric property of
the rst eigenvalue of the Dirichlet problem for the Laplacian on a Riemannian
manifold is given in Sect. 4.7. Finally, in Sect. 4.8 the capacitary inequality
4.2 Integral Inequalities for Functions on Riemannian Manifolds
257
(4.1.6) is used to obtain a sharp constant in the Sobolev-type two-weighted

Ilins inequality

(nb)p
u(x) (npa) dx
|x|b
Rn
npa
nb

Cp,a,b

u(x)p dx
|x|a
Rn
p1
for all u C0 (Rn ).
4.2 Integral Inequalities for Functions on Riemannian

Manifolds
Until now, and mostly in the sequel, we restrict our attention to Sobolev
spaces in subdomains of Rn . However, the methods of proof of many of the
previous results do not use in a specic way the properties of the Euclidean
space. Let us turn, for instance, to Corollary 2.3.4.
The assumption Rn is not essential for its proof: The capacitary
inequality (2.3.6) is also valid for functions on a Riemannian manifold, while
the rest of the argument has a general character. Let be an open subset
of a Riemannian manifold Rn and let = mn be the n-dimensional measure
on subsets of Rn . Then by virtue of Theorem 2.3.3 we have the following
two-sided estimate:
[capp (F, )]1/p
{F }
[mn (F )]1/q
inf
inf
uC0
grad u Lp ()
u Lp ()
()
[capp (F, )]1/p

(p 1)(p1)/p
inf
,
p
{F }
[mn (F )]1/q
(4.2.1)
where {F } is the family of all compact subsets of . Generally speaking, we

preserve the notation introduced in this chapter for the Euclidean case.
Let C be a function on (0, mn ()) such that for any open set g with
compact closure g and the smooth boundary one has the isoperimetric
inequality

s(g) C mn (g) .
(4.2.2)
Then, in view of (2.2.8), the following isocapacitary inequality holds:
mn ()
1p
dv
,
capp (F, )
p/(p1)
mn (F ) [C (v)]
and consequently,
inf
uC0 ()
grad u pLp ()
u pLp ()
(p 1)p1 pp
.
m ()
dv
p1 )
sup0<t<mn () (t( t n
)
p/(p1)
[C (v)]
(4.2.3)
258
Choosing for C a power function we derive the estimate

inf
uC0 ()
grad u pLp ()
u pLp ()

p
s(g)
pp inf
.
{g} mn (g)
(4.2.4)
In the case p = 2 the left-hand side of the last inequality coincides with the
rst eigenvalue () of Dirichlets problem for the Laplace operator on Rn .
Estimates of eigenvalues for the Dirichlet and Neumann problem by geometric
characteristics of the manifold have been the subject of many investigations
(see Berger, Gauduchon, and Mazet [85]; Yau [805]; Chavel [179]; Donnelly
[238, 239]; Escobar [258]; Chung, Grigoryan, and Yau [191]; Nadirashvili [623];
et al.). A signicant role here was played by Cheegers inequality from 1970
[181]

2
1
s(g)
inf
,
(4.2.5)
()
4 {g} mn (g)
which coincides with (4.2.4) if p = 2. The constant 1/4 in (4.2.5) cannot be
improved (Buser [159]).1 The earlier proof of (4.2.4) shows that Cheegers
inequality is a consequence of the estimate
()
1
cap(F, )
inf
,
4 F mn (F )
(4.2.6)
which was established by the author in 1962. Inequality (4.2.3) for p = 2 can
be written as
mn ()

1
dv
4
sup
t
.
(4.2.7)
()
[C (v)]2
0<t<mn ()
t
Specializing the function C in (4.2.2), one can derive from it strengthenings
of known lower bounds for the eigenvalue ().
Let us now turn to an example. Let be a subdomain of a two-dimensional
Riemannian manifold R2 , whose Gauss curvature K does not exceed 2 on
, = const > 0. Let us show that

1/2
4
1
2 1 1 + 2 m2 ()/4
.
()
(4.2.8)
This is a sharpening of the inequality of McKean [595], C () 2 /4, which

follows from (4.2.5). Indeed, as is known, in the hypothesis K 2 inequality (4.2.2) is fullled with C (v) = (4v + 2 v 2 )1/2 (BuragoZalgaller [151]).
Plugging this function into (4.2.7), we nd
1
Cheegers paper from 1970 is actually about the rst eigenvalue of the Neumann
(and not the Dirichlet) Laplacian, but the Dirichlet case is only simpler, Busers
example shows that the constant 1/4 in Cheegers inequality cannot be improved in
both the Dirichlet and no-boundary cases.
4.2 Integral Inequalities for Functions on Riemannian Manifolds
259

1
A(t + 1)
A
4
4
2 max t log
= 2 1
,
()
0<t<A
(A + 1)t
A+1
where A = 2 m2 ()/4, (y) = ye1y , 0 y 1, 1 being the inverse
of . Now (4.2.8) follows from the obvious inequality 1 (s) 1 (1 s)1/2 .
Analogous considerations can be done also in connection with other integral inequalities for functions on a domain on a Riemannian manifold. Then
the powers of integrability do not depend only on the amount of irregularity
of the boundary, but also on the geometry of the manifold. As before, the
properties of the embedding operator are xed by the constants in the corresponding inequalities between measures and capacities, so the main diculty
consists of the verication of these inequalities.
Recently this theme has been given considerable attention thanks to the
applications to physics and geometry. Let us turn to some such results.
As the isoperimetric inequality (1.4.13) is also true for subsets of the
Lobachevski space H n of constant negative curvature, the estimate (1.4.14)
with an exact constant remains in force for functions on H n .
For subsets g of a smooth k-dimensional submanifold of Rn without boundary the following inequality holds:

k/(k1)
v(g) ck s(g) + Q(g)
,
(4.2.9)
where v and s are the k-dimensional volume and the (k 1)-dimensional area
and Q the absolute integral mean curvature (Michael and Simon [600]). As
a particular case of Theorem 2.1.3 we nd that the isoperimetric inequality
(4.2.9) is equivalent to the integral inequality
(k1)/k

|u|k/(k1) dv
Rn

ck

|uH| dv , (4.2.10)

| grad u| dv +
Rn
Rn
where H is the mean curvature. If Rn is a manifold with boundary Rn , then

one must add to the right-hand side of (4.2.10) the norm of the trace of u in
L1 (Rn ). The constant ck depends only on k.
Let us remark that an isoperimetric inequality of the type (4.2.9) and
the corresponding embedding theorem have likewise been proved for objects
generalizing k-dimensional surfaces in Rn such as currents and varifolds (cf.
Burago and Zalgaller [151], Sect. 7.4).
Inequality (4.2.9) (and therefore also its consequence (4.2.10)) can be extended to manifolds M smoothly embedded in an n-dimensional Riemannian
manifold Rn (Homann and Spruck [379]). Then some natural geometrical
restrictions arise, which are connected with the volume of M, the injectivity
radius of Rn , and its sectional curvature.
The question of exact constants in inequalities of Sobolev type for functions on a Riemannian manifold have been studied by Aubin. He proved, in
particular, that for any compact two-dimensional Riemannian manifold R2
there exists a constant A(p) such that for all u Wp1 (R2 ), 1 p 2, one has
260
u L2p/(2p) (R2 ) K(2, p) grad u Lp (R2 ) + A(p) u Lp (R2 ) ,
(4.2.11)
where K(n, p) is the best constant in (2.3.21), as given in (2.3.23). For manifolds of arbitrary dimension one does not have very complete results. There
the necessity follows naturally from attempts to solve special nonlinear elliptic
equations on Riemannian manifolds, for example, in connection with the wellknown Yamabe problem concerning the conformal equivalence of an arbitrary
Riemannian metric with a metric of constant scalar curvature. The issue is to
nd a number such that the equation
4
n1
u + Ru = u(n+2)/(n2)
n2
(4.2.12)
has positive solution on a Riemannian manifold (M, g) of dimension n 3.

Here R = R(x) is the scalar curvature dened by the metric g and the
constant scalar curvature of the conformal metric. Yamabes paper [802] devoted to this problem contains an error. It was found by Trudinger [763] who
corrected it in the case of nonnegative scalar curvature. The next step was
done by Aubin [56]. He stated his results in terms of the best constant of a
certain inequality of Sobolev type. The nonlinear spectral problem (4.2.12)
can be associated with a variational problem where one looks for the inmum
J of the functional

n2
2
2
| grad u| +
Ru ds/ u 2L2n/(n2) (M ) ,
4(n 1)
M
dened in the space W21 (M ). Aubin proved that J [K(n, 2)]2 and that
Yamabes problem is positively solved if this inequality is sharp.
Let us further remark that for the n-dimensional ball S of unit measure
Aubin proved the inequality

2
u 2L2n/(n2) (S) K(n, 2) grad u 2L2 (S) + u 2L2 (S) ,
with an exact constant for both terms in the right-hand side. In 1986 GilMedrano [308] gave a partial solution to Yamabes problem for compact ndimensional locally conformally at manifolds with positive scalar curvature
and innite fundamental group, proving the inequality J < [K(n, 2)]2 . The
complete solution of Yamabes problem was achieved by Schoen [694]. Escobar
[260] contributed by solving this problem for manifolds with a boundary. The
work on Yamabes problem was an important step in the study of nonlinear
partial dierential equations on Riemannian manifolds.
We mentioned only a few early results to give the avor of a very large
area developed during the last 40 years in geometric analysis. Numerous studies showed the usefulness of isoperimetric inequalities, and to a smaller degree of isocapacitary inequalities, in the theory of Sobolev spaces and differential equations on manifolds, graphs, and Markov chains. Without aiming at completeness, we mention the monographs by Aubin [57]; Varopoulos,
4.3 The First DirichletLaplace Eigenvalue and Isoperimetric Constant
261
Salo-Coste, and Coulhon [772]; Hebey [360]; Chavel [180]; and Salo-Coste
[687]. Considerable attention was paid to extensions for the so-called subRiemannian dierential geometry, in particular to the analysis on Carnot and
Heisenberg groups (see Capogna, Danielli, Pauls, and Tyson [166]).
4.3 The First DirichletLaplace Eigenvalue and

Isoperimetric Constant
In this section we show by a counterexample that the fundamental eigenvalue
of the Dirichlet Laplacian is not equivalent to an isoperimetric constant, called,
as a rule, Cheegers constant, in contrast with an isocapacitary constant2 (see
(4.2.1)). This equivalence, even uniform with respect to the dimension, holds
for domains whose boundaries have nonnegative mean curvature, as proved
recently by E. Milman (oral communication), but as we shall see, it fails even
in the class of domains starshaped with respect to a ball.
Let be a subdomain of an n-dimensional Riemannian manifold Rn and
let () be the rst eigenvalue of the Dirichlet problem for the Laplace operator in , or more generally, the upper lower bound of the spectrum of
this operator
u 2L2 ()
() = inf
.
(4.3.1)
uC0 () u 2
L2 ()
By Corollary 2.3.4, () admits the two-sided estimate
1
() () ()
4
with
() := inf
{F }
(4.3.2)
cap(F ; )
.
mn (F )
By Theorem 2.1.3, the set function

() =
inf
uC0 ()
u L1 ()
u L1 ()
(4.3.3)
admits the geometric representation

() = inf
{g}
s(g)
,
mn (g)
(4.3.4)
where g is an arbitrary open subset of Rn with compact closure g in and

smooth boundary g, and s is the (n 1)-dimensional Hausdor measure.
Obviously, for all u C0 (),
2
By the equivalence of the set functions a and b, dened on subsets of Rn , we mean
here the existence of positive constants c1 and c2 depending only on n and such that
c1 a b c2 a.
262

()
|(u2 )| dx
u L2 ()
.
2
2 dx
u L2 ()
u
Hence
()2 4(),
(4.3.5)
which shows, together with (4.3.2) and (4.3.3), that

()2 4 ().
(4.3.6)
Note that (4.3.5) is nothing but (4.2.6), and we arrived at it in a much simpler way than in Sect. 4.2. One can ask whether an upper bound for ()
formulated in terms of () exists.
By the following counterexample we show that the answer is negative for
domains in Rn starshaped with respect to balls.
Example. Let be a subdomain of the n-dimensional unit ball B, starshaped with respect to a concentric ball B = {x : |x| < }. Here we check
that the inequality opposite to (4.3.6),
()2 C (),
(4.3.7)
is impossible with C independent of . Moreover, we shall construct a sequence

of domains {N }n1 situated in B and such that:
(i) N is starshaped with respect to a ball B(0, N ), where N 0,
(ii) (N ) ,
(iii) (N ) c, where c depends only on n.
n1
we deLet N stand for a suciently large integer number. By {j }N

j=1
note a collection of points on the unit sphere S n1 placed uniformly in the
sense that the distance from every point j to the set of other points of the
collection lies between c1 N 1 and c2 N 1 , where c1 and c2 are positive constants, depending only on n. Consider a closed rotational cone Cj with the
axis Oj and the vertex at the distance c0 N 1 from O, where c0 is an absolute constant large enough. The opening of Cj will be independent of j and
1n
denoted by N . Let N = o(N n2 ). Clearly, the complement of Cj is visible
from a suciently small ball B(0, N ). Therefore, the domain

N := B\
Cj
j
is starshaped with respect to B(0, N ) (see Fig. 17).

We shall nd the limit of (N ) as N as well as a lower estimate
for (N ). Clearly, (N ) (B) = n. Furthermore, by (4.3.4),

s(B) + s( j (B Cj ))
s(N )

(N )
=
mn (N )
mn (B) mn ( j (B Cj ))
n1
n + c1 n1
N N
n2 n1 ,
n /n c2 N N
4.3 The First DirichletLaplace Eigenvalue and Isoperimetric Constant
263
Fig. 17. The boundary of the subdomain N of the unit ball is the union of many
thin conic surfaces and a spherical part.
and therefore,
lim (N ) = n.
To estimate (N ) from below, we construct a covering of B by the balls

Bk := B(xk , 4c0 N 1 ), whose multiplicity does not exceed a constant depending only on n. Let |xk | c0 N 1 . Theorem 14.1.2, to be proved later, implies

u2 dx
|u|2 dx
(4.3.8)
cN n cap(Bk \N )
Bk
Bk
for all u C0 (N ), and the result will stem from a proper lower bound for
cap(Bk \N ).
First, let us consider n = 3. Clearly, Bk \N contains a right rotational
cylinder Tk with height c0 N 1 and diameter of the base N N 1 . Now, by
Proposition 13.1.3/1 to appear in Chap. 13,
cap(Tk ) cN 1 | log N |1 .
This estimate in combination with (4.3.8) gives

u2 dx
|u|2 dx.
cN 2 | log N |1
Bk
(4.3.9)
Bk
Choosing N = exp(N ) and summing (4.3.9) over all balls Bk , we obtain

(N ) cN . Hence (N ) where as (N ) c. Thus, in particular,
there is no inequality
and equivalently,
s(g)
inf
{g} m3 (g)
2
C inf
{F }
cap(F ; )
,
m3 (F )
264
2
s(g)
inf
C()
{g} m3 (g)
with constant factors C independent of the radius .
For dimensions greater than three, the very end of the argument remains
intact, but the estimation of cap(Bk \N ) becomes a bit more complicated
and the choice of N will be dierent.
Let Bk stand for the ball concentric with Bk and dilated with coecient
. We introduce the set sk = {j : Cj 12 Bk = }. With every j in sk we
associate a right rotational cylinder Tj coaxial with the cone Cj and situated
in Cj 12 Bk . The height of Tj will be equal to c0 N 1 and the diameter of
the base equal to N |xk |. We dene a cuto function j , equal to 1 on the
N |xk |-neighborhood of Tj , zero outside the 2N |xk |-neighborhood of Tj , and
satisfying the estimate

j (x) c(x)1 ,
where (x) is the distance from x to the intersection of Tj with the axis of Cj .
By Pk we denote the equilibrium potential of Bk \N . We have

(Pk j )2 dx
cap(Cj Bk )
jsk
Rn
jsk
Rn
|Pk |2 dx +
Rn

Pk2 2 dx .
Changing the constant c, one can majorize the last integral by the previous
one due to Hardys inequality. Hence,

cap(Tj ).
(4.3.10)
cap(Bk \N ) c
jsk
By Proposition 13.1.3/1 to be proved in the sequel,

n3 1
N .
cap(Tj ) c N |xk |
Furthermore, it is visible that the number of integers in sk is between two
multiples of |xk |1n . Now, by (4.3.10)

n3 1
cap(Bk \N ) c|xk |1n N |xk |
N
and by (4.3.8)
cN
n1
|xk |2 n3
N

u dx
2
Bk
Bk
|u|2 dx.
(4.3.11)
Since |xk | 1, it follows by summation of (4.3.11) over k that

n1
.
(N ) cn3
N N
Putting, for instance,

N = N (1n)/(n5/2) ,
we see that (N ) , and the desired counterexample is constructed for
n > 3.
4.4 Conductor Inequalities for a Dirichlet-Type Integral
265

with a Locality Property
Let X denote a locally compact Hausdor space and let C(X ) stand for the
space of continuous real-valued functions given on X . By C0 (X ) we denote
the set of functions f C(X ) with compact supports in X .
We introduce an operator Fp dened on a subset dom(Fp ) of C(X ) and
taking values in the cone of nonnegative locally nite Borel measures on X .
We suppose that 1 dom(Fp ) and Fp is positively homogeneous of order
p 1, i.e., for every real , f dom(Fp ) implies f dom(Fp ) and
Fp [f ] = ||p Fp [f ].
(4.4.1)
It is also assumed that Fp is contractive, that is, (f ) dom(Fp ) and

Fp (f ) Fp [f ],
(4.4.2)
for all f dom(Fp ), where is an arbitrary real-valued Lipschitz function
on the line R such that | | 1 and (0) = 0. We suppose that the following
locality condition holds:

Fp [f ] =
Fp [c].
(4.4.3)
f (x) = c R on a compact set C =
C
An example of the measure satisfying conditions (4.4.1)(4.4.3) is given

by (4.1.2), where = 0 and
Rn (x, z) (x, z) R
(4.4.4)
is a continuous function, positively homogeneous of degree 1 with respect to

z. One can take the space of locally Lipschitz functions on as dom(Fp ).
Let g and G denote open sets in X such that the closure g is a compact
subset of G. We introduce the p-conductivity of the conductor G\g (in other
terms, the relative p-capacity of the set g with respect to G) as

capp (
g , G) = inf
Fp [] : dom(Fp ), 0 1 on G
X

(4.4.5)
Using the truncation

( )+
,1 ,
() = min
1
with (0, 1) and R, we see that the inmum in (4.4.5) does not change
if the class of admissible functions is enlarged to
266
dom(Fp ) C0 (X ) : 1 on g, 0 on X \G
(4.4.6)
(compare with Sect. 2.2.1).

Lemma. Let f dom(Fp ) C0 (X ), a = const > 1, Lt = {x X :
|f (x)| > t}. Then the function t capp (Lat , Lt ) is upper semicontinuous.
Proof. It follows from (4.4.3) that

Fp [f ] =
X
Fp [f ] < .
(4.4.7)
suppf
Let t0 > 0 and > 0. There exist open sets g and G such that
Lat g,
g G,
G Lt .
(4.4.8)
It follows from the denition of capp that for all compact sets C g
capp (C, G) capp (Lat0 , Lt0 ) +
(compare with Sect. 2.2.1). By (4.4.8),

max f (x) : x g < at0 and
We denote
and
(4.4.9)

min f (x) : x G > t0 .

1 = t0 a1 max f (x) : x g

2 = min f (x) : x G t0 .
Then
La(t0 ) g
and G Lt0 +
for every (0, min{1 , 2 }). Putting C = La(t0 ) in (4.4.9) and recalling
that capp decreases with enlargement of the conductor, we obtain
capp (La(t0 ) , Lt0 + ) capp (Lat0 , Lt0 ) + .
(4.4.10)
Using the monotonicity of capp again, we deduce from (4.4.10) that

capp (Lat , Lt ) capp (Lat0 , Lt0 ) +
for every t suciently close to t0 . In other words, the function t

capp (Lat , Lt ) is upper semicontinuous. The result follows.
We prove a general conductor inequality in the integral form for the functional (4.1.1).
Theorem. Let M denote an increasing convex (not necessarily strictly
convex) function given on [0, ), M (0) = 0. Then the inequality (4.1.3) holds
for all f dom(Fp ) C0 (X ) and for an arbitrary a > 1.
Proof. We have
267

capp (Lat , Lt )
Fp []
for every dom(Fp ) C0 (X ) satisfying

= 1 on Lat ,
= 0 on X \Lt ,
By the homogeneity of Fp and by (4.4.3),

t capp (Lat , Lt )
and
We set here
( x) =
where
0 1 on X .
Lt
Fp [t].
t (f (x))
,
(a 1)t

t () = min (|| t)+ , (a 1)t ,
R,
(4.4.11)
with + = (|| + )/2. By t = const on Lat and by (4.4.3) we have

1
Fp t (f ) .
tp capp (Lat , Lt )
p
(a 1) Lt \Lat
Since the mapping t () is contractive and since the function t
F [f ] has, at most, a countable set of discontinuities, it follows that
Lt p

1
p
F
[f
]
+
t
Fp [1],
(4.4.12)
tp capp (Lat , Lt )
p
(a 1)p Lt \Lat
Lat
for almost every t > 0. Hence,

dt
M tp capp (Lat , Lt )
t
0

1
dt
p
M
F
[f
]
+
t
F
[1]
p
p
p
(a
1)
t
0
Lt \Lat
Lat

2
dt
1
M
Fp [f ]
2 0
(a 1)p Lt \Lat
t

dt
1
M 2tp
Fp [1]
+
.
2 0
t
Lat
(4.4.13)
Let denote a locally integrable function on (0, ) such that there exist
the limits (0) and (). Then the identity

dt

(t) (at)
= (0) () log a,
(4.4.14)
t
0
holds. Setting here
268

(t) := M
1
(a 1)p

Lt

Fp [f ] ,
and using the convexity of M , we obtain

2
dt
M
Fp [f ]
p
(a 1) Lt \Lat
t
0

2
dt
2
F
[f
]
M
F
[f
]
M
p
p
p
p
(a 1) Lt
(a 1) Lat
t
0

2
= log a, M
Fp [f ] .
(4.4.15)
(a 1)p X
By convexity of M ,

dt
p
M 2t
Fp [1]
t
0
Lat

2
M 2tp
Fp [1] tp1
Lat
t
p1

M 2p

p1
La

Fp [1] d = ap 1

0
Using the truncation

() =
||
a
Fp [1] dt

Fp [1] d tp1
Fp [1] dt
La
Lat

Fp [1] d .
(4.4.16)

1
p1
= M 2p
p
La
0
Clearly,

Lat

p1
L \La
Fp [1] d.
for || > a,
for || a
together with (4.4.2), we deduce from (4.4.17) and (4.4.14) that

ap 1
d
p1
Fp [1] d
Fp [f ]
p
a
La
L \L
0
0
a
ap 1
Fp [f ].
= log a
ap
X
Combining this with (4.4.16), we arrive at

1
ap 1
dt
p
M 2p log a
M 2t
Fp [1]
Fp [f ] .
t
p
ap
Lat
X
0
Adding (4.4.15) and the last inequality, we deduce from (4.4.13) that
(4.4.17)
269

dt
t
0

2
ap 1
1
1
log aM
Fp [f ] + M 2p log a
Fp [f ] ,
2
(a 1)p X
2p
ap
X

Remark 1. Suppose that (4.4.3) is replaced by the following more restrictive

locality condition:

f (x) = const on a compact set C =
Fp [f ] = 0,
(4.4.18)
C
which holds, for example, if the measure in (4.1.2) is zero. Then the above
proof becomes simpler. In fact, we can replace (4.4.13) by

dt
1
M
F
[f
]
.
p
t
(a 1)p Lt \Lat
0
0
Estimating the right-hand side by (4.4.15) we obtain the inequality

p
dt
1
M t capp (Lat , Lt )
Fp [f ] .
(4.4.19)
log aM
t
(a 1)p X
0
The next assertion follows directly from (4.1.3) and (4.4.19) by setting
M () = q/p for 0.
Corollary 1. Let q p and let Fp satisfy the locality condition (4.4.3).
Then for all f dom(Fp ) C0 (X ) and for an arbitrary a > 1
capp (Lat , Lt )
q/p q
d t
1/q

C
1/p
Fp [f ]
(4.4.20)
If additionally Fp is subject to (4.4.18), then one can choose

C=
(q log a)1/q
.
a1
Remark 2. Let Fp satisfy (4.4.18). Then for every sequence {tk }

k= such
that 0 < tk < tk+1 ,
tk 0 as k and
tk as k ,
the following discrete conductor inequality holds:

k=
(tk+1 tk )p capp (Ltk+1 , Ltk ) Fp [f ].
(4.4.21)
270
Putting tk = ak , where a > 1, we see that
apk capp (Lak+1 , Lak ) (a 1)p Fp [f ].
(4.4.22)
k=
Using Lemma 4.4 and monotonicity properties of the conductivity, we

check that (4.4.22) is equivalent to (4.4.20) with q = p modulo the value of
the coecient c.
The capacitary inequality
q/p q
d t
1/q
capp (Lt , X )
1/p
X
Fp [f ]
(4.4.23)
results directly from (4.4.20).

An immediate consequence of (4.4.23) is the following criterion for the
Sobolev-type inequality:

f Lq () C
1/p
X
Fp [f ]
(4.4.24)
where is a locally nite Radon measure on X , q p, and f is an arbitrary

function in dom(Fp ) C0 (X ).
Corollary 2. Inequality (4.4.24) holds if and only if
sup
(g)p/q
< .
capp (g, X )
(4.4.25)
Proof. The necessity of (4.4.25) is obvious and its suciency follows from
the well-known and easily checked inequality

(Lt ) d tq
1/q

1/p
(Lt )p/q d tp
,
where q p 1 (see Hardy, Littlewood, and Polya [350]).
4.5 Conductor Inequality for a Dirichlet-Type Integral

without Locality Conditions
Here the notations X and Lip0 (X ) have the same meaning as in Sect. 4.4.
Let stand for the Cartesian product of sets and let denote a nonnegative
symmetric Radon measure on X 2 := X X , locally nite outside the diagonal
{(x, y) X 2 : x = y}. We shall derive a conductor inequality for the seminorm
(4.1.4) where f is an arbitrary function in C0 (X ) such that
4.5 Conductor Inequality for a Dirichlet-Type Integral
f p, < .
271
(4.5.1)
Clearly, the seminorm (f )p, is contractive, that is,

(f ) p, f p, ,
with the same as in (4.4.2).
Let, as before, g and G denote open sets in X such that g is a compact subset of G. Similarly to Sect. 4.4, we introduce the conductivity of the conductor
G\g, in other terms, the capacity of g with respect to G:
g , G) = inf{f pp, : C0 (X ), 0 1 on G
capp, (
and = 1 on a neighborhood of g}.
It is straightforward that this inmum does not change if the class of admissible functions is replaced with (4.4.6) (compare with the denition of
capp (g, G) in Sect. 4.4).
Theorem. For all f C0 (X ), for all q p 1, and for an arbitrary
a > 1 the conductor inequality (4.1.4) holds.
Proof. The measurability of the function t capp, (Lat , Lt ) is proved
word for word as in Lemma 4.4.
Clearly,

p
(a 1)p tp capp, (Lat , Lt ) t (f ) p,
(4.5.2)
with t dened by (4.4.11). Let Kt denote the conductor Lt \Lat . Since
S 2 (S T ) (T S) (S\T )2
for all sets S and T and since is symmetric, it follows that
(a 1)p tp capp, (Lat , Lt )

t f (x) t f (y) p (dx dy)
+
2
K X
(X \Kt )2
t

f (x) f (y)p (dx dy) + 2(a 1)p tp Lat (X \Lt ) .
2
Kt X
(4.5.3)
By Minkowskis inequality,

p/q

q/p q1
(a 1)p
capp, (Lat , Lt )
t
dt
A + B,
0
where

A=2
0

Kt X

q/p
dt
t
p/q
272
and

B = 2(a 1)
Lat (X \Lt )
q/p
p/q
t
q1
dt
Since q p, we have
Aa
0
where

(t) =
Lt X
(t) (at)
dt
t
p/q
,

p/q
.

A 2(log a)p/q f pp .
(4.5.4)
Let us estimate B. Clearly,

B = 2(a 1)p
q/p

X2
(x, Lat )(y, X \Lt ) (dx dy)
p/q
tq1 dt
where (, S) is the characteristic function of a set S. By Minkowskis inequality,

p/q

B 2(a 1)p
(x, Lat )(y, X \Lt )tq1 dt
(dx dy)
X2
0

2(a 1)p
f (x)q aq f (y)q p/q (dx dy).
=
+
p/q
p
q a
X2
Obviously, the inequality |f (x)| a|f (y)| implies

f (x)q aq f (y)q f (x)q
Hence

q
aq
f (x) |f (y)| + .
q
(a 1)
B q p f pp .
Adding this estimate with (4.5.4), we arrive at (4.1.4) with

c(a, p, q) =
1 + 2(q log a)p/q

.
(a 1)p q p/q

To show the usefulness of inequality (4.1.4), we give a criterion of a twoweight Sobolev-type inequality involving the seminorm f p, .
Corollary. Let q p 1, q r > 0, and let and be locally nite
nonnegative Radon measures on X . The inequality
4.6 Sharp Capacitary Inequalities and Their Applications

q/r
q
r
|f | d C f p, +
|f | d
q
273
(4.5.5)
holds for every f dom(Fp ) C0 (X ) if and only if all bounded open sets g
and G in X such that g G satisfy the inequality

(4.5.6)
(g) Q capp, (g, G)q/p + (G)q/r .
The best constants C and Q in (4.5.5) and (4.5.6) are related by Q C
c(p, q)Q.
Proof. The necessity of (4.5.6) and the estimate G C are obtained by
putting an arbitrary function f dom(Fp ) C0 (X ) subject to f = 1 on g,
f = 0 on X \G, 0 f 1, into (4.5.5).
The suciency of (4.5.6) results by the following argument:

|f |q d =
(Lt ) d tq
0
X

Q
capp, (Lat , Lt )q/p d tq +
(Lt )q/r d tq
0
q/r

q
q/p
r
|f | d
,
Q c(a, p, q) f p, +
X
where c(a, p, q) is the same constant as in (4.1.4). The proof is complete.

Remark. Using the obvious identity

|f | 1, =
(Lt X \Lt ) dt,
0
instead of the conductor inequality (4.1.4), we deduce with the same argument
that the inequality
1/q

|f | d
C f 1, +
|f | d ,
q
(4.5.7)
with q 1 holds if and only if, for all bounded open sets g,

(g)1/q C g (X \g) + (g) ,
with the same value of C as in (4.5.7).

Let denote an open set in Rn and let the function
Rn (x, z) (x, z) R
274
be a continuous function, positively homogeneous of degree 1 with respect to

z. Clearly, the measure

p
Fp [f ] := x, grad f (x) dx
satises (4.4.1), (4.4.2), and (4.4.18). Hence, (4.4.20) implies the inequality
q/p q
capp (Lt , X )
d t
1/q

C
1/p

x, grad f (x) p dx
(4.6.1)
where capp is the p-capacity corresponding to the integral (3.1.8), C = const >
0, and f is an arbitrary function in C0 (). The next assertion gives the sharp
value of C for q > p. In the case q = p the sharp value of C is given by (2.3.6)
and is obtained by the same argument.
Theorem 1. Inequality (4.6.1) with q > p holds with

C=
pq
)
( qp
1/p1/q
.
q
q1
( qp
) (p qp
)
(4.6.2)
This value of C is sharp if (x, y) = |y| and if either is a ball or = Rn .

Proof. Let
1/(1p)

x, N (x) p grad f (x)p1 ds(x)
(t) =
|f (x)|=
with ds standing for the surface element and N (x) denoting the normal vector
at x directed inward L . Further, let t() denote the inverse function of (t).
Then

x, grad f (x) p dx =
t ()p d
(4.6.3)
(see Lemma 2.3.1 for more details). By Blisss inequality [109]

t()q
1/q
1+q(p1)/p
p
q(p 1)
1/q
C

t ()p d
1/p
(4.6.4)
with C as in (4.6.2), and by (4.6.3) this is equivalent to
d(t()q )
q(p1)/p
1/q
C
1/p

x, grad f (x) p dx
To obtain (4.6.1) with C given by (4.6.2) it remains to note that

capp (Lt )
1
(t)p1
(4.6.5)
275
(see Lemma 2.2.2/1). The constant (4.6.2) is the best possible since (4.6.5)
becomes equality for radial functions.

Following Denition 2.1.4, we introduce the weighted perimeter minimizing
function on (0, ) by

x, N (x) ds(x),
C (m) := inf
(4.6.6)
g
where the inmum is extended over all bounded open sets g with smooth
boundaries subject to
mn (g) m.
According to Corollary 2.2.3/2, the following isocapacitary inequality holds:

capp (g, G)
mn (G)
mn (g)
dm
C (m)p
1p
.
(4.6.7)
Therefore, (4.6.1) leads to

Corollary 1. For, all f C0 (),
mn ()

0
mn (Lt )
dm
C (m)p
q/p

1/q
d tq
1/p

x, grad f (x) p dx
(4.6.8)
with q > p and C dened by (4.6.2). For p = 1 the last inequality should be
replaced by

q 1/q

x, grad f (x) dx
C mn (Lt ) d tq
(4.6.9)
with q 1.
In addition, this corollary, combined with the isoperimetric inequality
s(g) n1/n n1/n mn (g)1/n ,

immediately gives the following well-known sharp result.
Corollary 2. (See (1.4.14) for p = 1 and Sect. 2.3.5 for p > 1) Let
n > p 1 and q = pn(n p)1 . Then every f C0 (Rn ) satises the Sobolev
inequality (4.1.7) with the best constant
c=
1/2 1/2
p1
np
1/p
(n) (1 + n/2)
(n/p) (1 + n n/p)
1/n
.
276
The next assertion resulting from (4.6.3) and (4.6.5) shows that a quite
general capacitary inequality is a consequence of a certain inequality for functions of one variable.
Theorem 2. Let and be positive nondecreasing functions on (0, )
such that

(4.6.10)
1p d t() <
sup
0
with the supremum taken over all absolutely continuous functions [0, )
t() 0 subject to t(0) = 0 and

t ()p d 1.
(4.6.11)
0
Then
sup

capp (Lt , ) d(t) <
(4.6.12)
with the supremum extended over all f subject to

x, grad f (x) p dx 1.
(4.6.13)
The least upper bounds (4.6.10) and (4.6.12) coincide.

In fact, the above Theorem 1 is a particular case of Theorem 2 corresponding to the choice
(t) = tq and () = q/p .
The next result is another consequence of Theorem 2.
Theorem 3. For every c > 0

c
d exp ctp
< ,
exp
sup
1/(p1)
capp (Lt , )
0
(4.6.14)
where the supremum is taken over all f C0 () subject to (4.6.13) and

p = p/(p 1), p > 1.
Proof. It follows from a theorem by Jodeit [402] that

sup
exp t()p d <
(4.6.15)
with the supremum taken over all absolutely continuous functions [0, )
t() 0 subject to t(0) = 0 and (4.6.11). Hence, for every c > 0,

exp ct()p c d < .
sup
0
277
It remains to refer to Theorem 4.6/2 with

(t) = exp ctp

() = exp c 1/(1p) .
and
A direct consequence of Theorem 3 and the isocapacitary inequality

(2.3.17) is the following result by Moser.
Corollary 3. (Moser [618]) Let mn () < and let

{f } := f C0 () : grad f Ln () 1 .
Then

n

exp nn1/(n1) f (x) dx < .
sup
{f }
Proof. The rst inequality (2.2.11) can be written as

mn (g) mn (G) exp nn1/(n1) capn (g, G)1/(1n) .
1/(n1)
Hence, putting c = nn

and p = n in (4.6.14), we obtain

mn (Lt ) d exp nn1/(n1) tn < .

The result follows.
One needs no changes in proofs to see that the main results of this section,
Theorems 13, hold if is an open subset of a Riemannian manifold Rn and
grad f is the Riemannian gradient. As an application, we obtain another Moser
inequality [618].
Corollary 4. Let be a proper subdomain of the unit sphere S 2 and let
{f } be dened as in Corollary 3. Then

exp 4f 2 () ds < .
sup
{f }
Proof. By Theorem 3 with c = 4 we have the capacitary integral inequality

sup
{f }

exp
4
cap2 (Lt , )

d exp 4t2 < .
The classical isoperimetric inequality on S 2

s(g)2 m2 (g) 4 m2 (g)
(4.6.16)
278
(see Rado [668]), combined with (4.6.7) implies the isocapacitary inequality
1

m2 (G)(4 m2 (g))
g , G) 4 log
.
cap2 (
m2 (g)(4 m2 (G))
Setting here G = , g = Lt , and using (4.6.16), we complete the proof.

Remark. We can go even further, extending the previous results to the

measure-valued operator Fp [f ] in Sect. 4.4 subject to the condition
p

(4.6.17)
Fp (f ) = (f ) Fp [f ],
with the same as in (4.4.2). In fact, (4.6.17) implies

t ()p d,
Fp [f ] =
X
(4.6.18)
where t() is the inverse of the function

(t) =
1/(1p)

d Fp [f ](L )
d.
d
Identity (4.6.18) is the core of the proof of the results in the present section.
4.7 Capacitary Improvement of the FaberKrahn

Inequality
By the Faber [265]Krahn [462] inequality for the DirichletLaplace eigenvalues, the ball has the minimum eigenvalue among all domains in Rn with the
same volume.
Here we obtain a capacitary version of this inequality valid for any open
subset of an arbitrary Riemannian manifold. We state and prove the main
result of this section.
Theorem 1. Let R > 0, u C0 (), and Nt = {x : |u(x)| t}. If
n > 2, then
j(n2)/2
R
2

mn (BR )
u 2L2 () ,
cap(Nt ; )
cap(BR ) + cap(Nt ; )
n
n2

d t2
(4.7.1)
where cap is the 2-capacity and j is the rst positive root of the Bessel function J . If n = 2, then

4
2
d t2 u 2L2 () .
exp
(4.7.2)
j0
cap(Nt ; )
0
4.7 Capacitary Improvement of the FaberKrahn Inequality
279
Proof. Let w be an arbitrary absolutely continuous function on (0, R], such

that w(R) = 0. The inequality
j(n2)/2
R
2
w()2 n1 d
w ()2 n1 d,
(4.7.3)
where n > 2, is equivalent to the fact that the rst eigenvalue of the Dirichlet
2
Laplace operator in the unit ball B equals j(n2)/2
. Similarly, the inequality
j0
R
2

w()2 d
w ()2 d,
(4.7.4)
is associated with n = 2.
In the case n > 2, we introduce the new variables
=
2n R2n
,
(n 2)n

t() = w () ,
and write (4.7.3) in the form

2

t()2 d
n j(n2)/2 R1
((n 2)n + R2n )2(n1)/(n2)
0

t ()2 d.
(4.7.5)
Similarly, for n = 2, putting

= (2)1 log
R
,

t() = w () ,
we write (4.7.4) as

(2j0 )
t() exp(4) d
2
t ()2 d.
(4.7.6)
Note that the function t in (4.7.5) and (4.7.6) is subject to the boundary
condition t(0) = 0. We write (4.7.5) and (4.7.6) as
n
j(n2)/2
R
2
n

t ()2 d
dt()2
((n 2)n + R2n )n/(n2)
(4.7.7)
and
j02
exp(4) dt()2
t ()2 d.
Now, as in Sect. 2.3.1, we introduce the function
(4.7.8)
280
(t) =
0
d
|u| ds
|u|=
(4.7.9)
as well as its inverse t(), and replace the integral on the right-hand side
of (4.7.7) and (4.7.8) by u 2L2 () . It remains to note that

1
(t) cap(Nt ; )
(4.7.10)
by Lemma 2.2.2/1.
Let us use the area minimizing function of
(v) = inf s(g),
(4.7.11)
where the inmum is extended over all sets g with smooth boundaries and
compact closures g , subject to the inequality mn (g) v. This function
is a particular case of the function C from Denition 2.1.4, corresponding to
(x, ) = ||. The function appears in the isocapacitary inequality
1
mn ()
dv
cap(F ; )
2
mn (F ) (v)
(see Corollary 2.2.3/2). Therefore, (4.7.1), (4.7.2), and the identity
cap(BR ) = (n 2)n Rn2 ,
lead to the following Lorentz-type estimates.
Corollary 2. If n > 2 and R > 0, then, for all u C0 (),
j(n2)/2
R
2
mn ()
cap(BR )
mn (BR )
mn (Nt )
dv
+1
(v)2
n
2n
u 2L2 () .

d t2
(4.7.12)
If n = 2, then, for all u C0 (),

mn ()

dv
exp 4
j02
d t2 u 2L2 () .
2
0
mn (Nt ) (v)
Remark. Since
(v) n
n1
n
nn v
n1
n
(4.7.13)
(4.7.14)
by the classical isoperimetric inequality for R , the estimates (4.7.12) and

(4.7.13) imply the FaberKrahn property

j(n2)/2 2
()
R
n
for any n-dimensional Euclidean domain with mn () = n1 n Rn .
4.7 Capacitary Improvement of the FaberKrahn Inequality
281
Theorem 1 is a very special case of the following general assertion.

Theorem 2. Let M be a decreasing nonnegative function on [0, ) and
let q > 0 and p 1. Suppose that for all absolutely continuous functions
t() on [0, ), the inequality

1/q

t()q dM()

t ()p d
1/p
(4.7.15)
holds. Then, for all u C0 (),

1/(1p) q
d t
M capp (Nt ; )
1/q
u Lp () ,
(4.7.16)
where capp is the p-capacity dened by

|u|p dx : u C0 (), u 1 on F .
capp (F ; ) = inf
(4.7.17)
Proof. The role of the function given by (4.7.9) is played in the present
proof by
t
d

.
(4.7.18)
(t) =
p1 ds)1/(p1)
0 ( |u|= |u|
We write the left-hand side of (4.7.15) in the form

q
M() d t()
1/q
and use the monotonicity of M and the inequality

1/(1p)

capp (Nt() ; )
(4.7.19)
proved in Lemma 2.2.2/1. It remains to apply (4.7.15) and the identity

f ()p d =
|u|p dx
(4.7.20)
0

found in Lemma 2.3.1.
Using the area minimizing function dened by (4.7.11) and the estimate

capp (F ; )
mn ()
mn (F )
dv
(v)p/(p1)
1p
(see Corollary 2.2.3/2), we obtain the following from Theorem 2.
(4.7.21)
282
Corollary 2. Let , p, and q be the same as in Theorem 2 and let (4.7.15)

hold. Then

mn ()
q 1/q
dv
d t
M
u Lp ()
(4.7.22)
p/(p1)
0
mn (Nt ) (v)
for all u C0 ().
Clearly, (4.7.22) is a generalization of the estimates (4.7.12) and (4.7.13)
which were obtained for p = 2 with a particular choice of . Another obvious
remark is that (4.7.15), where M is dened on the interval 0 < t < mn () by
mn ()
M
t
dv
(v)p/(p1)

= p,q t
with a constant p,q depending on mn (), implies the inequality

1/q
p,q u Lq () u Lp () ,
(4.7.23)
for all u C0 ().
4.8 Two-Weight Sobolev Inequality with Sharp Constant

Let the measure b be dened by
dx
|x|b
b (K) =
K
(4.8.1)
for any compact set K in Rn . In this section we obtain the best constant in
the inequality

p dx 1/p

,
u L,q (b ) C
u(x)
|x|a
Rn
where the left-hand side is the quasinorm in the Lorentz space L,q (b ), i.e.,

u L,q (b ) =

q/ q 1/q

b x : u(x) t
d t
.
As a particular case of this result we obtain the best constant in the HardySobolev inequality

u(x)q dx
|x|b
Rn
1/q

C

u(x)p dx
|x|a
Rn
1/p
.
(4.8.2)
Let denote an open set in Rn and let p [1, ). By (p, a)-capacity of a

compact set K we mean the set function

capp,a (K, ) = inf
283

|u|p |x|a dx : u C0 (), u 1 on K .
In the case a = 0, = Rn we write simply capp (K).

The following inequality is a particular case of a more general one obtained
in Theorem 4.6/1, where is an open subset of an arbitrary Riemannian
manifold and |(x, u(x))| plays the role of |u(x)||x|a/p .
Let q p 1 and let be an open set in Rn . Then for an arbitrary
u C0 (),

1/p
p a

q/p q 1/q

u(x) |x| dx
capp,a (Lt , )
d t
Ap,q
,
(4.8.3)
where

Ap,q =
1/p1/q
pq
)
( qp
(4.8.4)
q
q1
( qp
) (p qp
)
for q > p and

Ap,p = p(p 1)(1p)/p .
(4.8.5)
The sharpness of this constant is checked by a sequence of radial functions

in C0 (). Moreover, there exists a radial optimizer vanishing at innity, if
= Rn .
Being combined with the isocapacitary inequality
(K) p, capp,a (K, ),
(4.8.6)
where is a Radon measure in , (4.8.3) implies the estimate

(Lt )
q/p q
d t
1/q
1/p

u(x)p |x|a dx

Ap,q 1/p
p,
(4.8.7)
for all u C0 ().

This estimate of u in the Lorentz space Lp/,q () becomes the estimate in
Lq () for = p/q
1/p

u(x)p |x|a dx
.

u Lq () Ap,q 1/p
p,
In the next assertion we nd the best value of p, in (4.8.6) for the

measure = b dened by (4.8.1).
Lemma. Let
1 p < n,
Then
0 a < n p,
and
a+pb
an
.
np
(4.8.8)
284

Rn
dx
|x|b
npa
nb
p1
npa
p1
bpa
n nb
(n b)
npa
nb
capp,a (K).
(4.8.9)
The value of the constant factor in front of the capacity is sharp and the
equality in (4.8.9) is attained at any ball centered at the origin.
Proof. Introducing spherical coordinates (r, ) with r > 0 and S n1 ,
we have
p2

2
u
1
2

r n1a dr ds . (4.8.10)
capp,a (K) = inf
r + r 2 | u|
u|K 1
n1
S
Let us put here r = 1/ , where

=
npa
,
np
and y = (, ). The mapping (r, ) (, ) will be denoted by . Then

(4.8.10) takes the form

capp,a (K) =
p1
inf
v
2
p2
u
+ ()2 | u|2
dy,

n
(4.8.11)
where the inmum is taken over all v = u 1 . Since 0 1 owing to

the conditions p < n, 0 < a < n p, and a 0, inequality (4.8.11) implies

capp,a (K) p1 inf
(4.8.12)
|u|p dy p1 capp (K) ,
v
Rn
which together with the isocapacitary property of capp (see (2.2.12)) leads to
the estimate

p1
p

np
np
np
nn n n mn (K) n .
(4.8.13)
capp (K)
p1
Clearly,
b (K) =
with
=n

(K)
dy
|y|
nb
b(n p) an
=
0.
npa
(4.8.14)
Furthermore, one can easily check that

1 n
n1 n n
n mn (K)
b (K)
nb
(4.8.15)
(see, for instance, Example 2.1.5/2). Combining (4.8.15) with (4.8.13), we nd
b (K)
npa
nb
p1
np
p1
bpa
n nb
(n b)
npa
nb

capp (K) ,
285
(4.8.16)

which together with (4.8.12) completes the proof of (4.8.9).

The main result of this section is as follows.
Theorem. Let conditions (4.8.8) hold and let q p. Then for all u
C0 (Rn )
(npa)q
b (Lt ) (nb)p d tq
q1

Cp,q,a,b

u(x)p dx
|x|a
Rn
p1
,
(4.8.17)
where

Cp,q,a,b =
pq
)
( qp
p1 1q
p1
npa
1 p1
q
q1
( qp
) (p qp
)
p+ab

(nb)p
( n2 )
.
npa
n
2 2 (n b) p+ab
(4.8.18)
The constant (4.8.18) is the best possible that can be shown by constructing a
radial optimizing sequence in C0 (Rn ).
Proof. Inequality (4.8.17) is obtained by the substitution of (4.8.4) and
(4.8.9) into (4.8.7). The sharpness of (4.8.18) follows from the sharpness of
the constant Ap,q in (4.8.3) and from the fact that the isocapacitary inequality
(4.8.9) becomes equality for balls.

The last theorem contains the best constant in the Ilin inequality (4.8.2)
as a particular case q = (nb)p/(npa). We formulate this as the following
assertion.
Corollary. Let conditions (4.8.8) hold. Then for all u C0 (Rn )
(nb)p

u(x) (npa) dx
|x|b
Rn
npa
nb

Cp,a,b

u(x)p dx
|x|a
Rn
p1
,
(4.8.19)
where
p+ab
1 p1
(nb)p
( n2 )
p1
=
npa
n
npa
2 2 (n b) p+ab
pb

p(nb)
)
( p(nb)
pb
.
(nb)(p1)
( nb
)
(1
+
)
pb
pb
Cp,a,b
This constant is the best possible, which can be shown by constructing a radial
optimizing sequence in C0 (Rn ).
286

Sections 4.2, 4.3. We follow the articles by Mazya [557] and [564]. In the
class of convex domains Rn , it was shown by B. Klartag and E. Milman (oral communication) that the FaberKrahn inequality admits a matching converse inequality for some linear image T () of (having the same
measure). Specically, there exists a numerical constant C > 0, such that if
mn () = n1 n Rn , then

2 j(n2)/2 2
.
T () C log(1 + n)
R
Sections 4.44.6. Here we follow the authors article [561]. Various aspects of the theory of Sobolev spaces of the rst order on metric spaces with
a measure were studied in numerous papers, to name only a few: Korevaar
and Schoen [452]; Biroli and Mosco [105]; Hajlasz [338]; Kinnunen and Martio
[425]; Cheeger [182]; Franchi, Hajlasz, and Koskela [287]; Kilpel
ainen, Kinnunen, and Martio [421]; Shanmugalingam [696]; Koskela [456]; Bjorn, MacManus, and Shanmugalingam [106]; Heinonen [373]; Goldshtein and Troyanov
[318]; Ambrosio and Tilli [43]; Ostrovskii [650]; Hajlasz, Koskela, and Tuominen [345]; and Kinnunen and Korte [423].
In connection with the capacitary improvement (4.6.14) of Mosers inequality we note that Hudson and Leckband [387] established the existence of
an extremal in Jodeits inequality (4.6.15).
Section 4.7. Theorem 4.7/1 was proved in Mazya [564].
Section 4.8. The material of this section is borrowed from the paper by
Mazya and Shaposhnikova [587].
Inequality (4.8.2) was rst proved by Ilin in 1961 ([395], Theorem 1.4)
without discussion of the value of C. For certain values of parameters the
best constant C was found in Chua and Wheeden [190] (p = 2), in Mazya
[543], Sect. 2 (p = 1, a = 0), in Glazer, Martin, Grosse, and Thirring [310]
(p = 2, n = 3, a = 0), Lieb [496] (p = 2, n 3, a = 0), in Chou and Chu [187]
(p = 2, q 2), and in A. Nazarov [628] (1 < p < n, a = 0), where dierent
methods were used.
5
Integrability of Functions in the Space L11 ()
Now we turn to embedding theorems for functions with unrestricted boundary

values. The present chapter contains conditions on that are necessary and
sucient for the embedding operator L11 () Lq () to be continuous or
compact. These criteria are intimately connected with relative isoperimetric
inequalities and isoperimetric functions. In Sect. 5.2 we consider the cases
q 1 and 0 < q < 1 separately.
In more detail, by a relative isoperimetric inequality we mean the inequality of the form
mn (G ) const s(i G ), > 0,
(5.0.1)
connecting the volume of admissible subsets G of with the area of the
surfaces i G = G . The sets G should satisfy an additional requirement
mn (G ) M , where M (0, mn ()). By Theorem 5.2.3, the above isoperimetric inequality with 1 holds if and only if the embedding operator
L11 () L1/ () is continuous. The sets satisfying (5.0.1) are said to
belong to the class J . They can be characterized by the inequality
lim inf M () > 0,
+0
where M is an isoperimetric (in other words, area minimizing function) of ,

i.e., the least upper bound of the numbers s(i G ) over all sets G subject to
mn (G ) M . The embedding L11 () L1/ () for > 1 proves to be
continuous if and only if belongs to the class H dened by the inequality
M

0
s
M (s)
1
1
ds
< .
s
To nd more visible sucient conditions on the domain ensuring Sobolev

embeddings, we use the so-called subareal mappings in Sect. 5.3. Those are
mappings of one domain onto another that do not increase s(i G ) up to a
constant factor.

DOI 10.1007/978-3-642-15564-2 5,
287
288
5 Integrability of Functions in the Space L11 ()
A particular irregular domain that was considered by Nikod

ym in 1933 as
a counterexample to the Poincare inequality is discussed in Sect. 5.4.
In Sect. 5.5 we show that the embedding L11 () Lq () is always compact
if 0 < q < 1 and it is compact for q 1 if and only if
lim 1/q M () = .
+0
It might be natural to expect that a weakening of the assumptions on

the domain in embedding theorems of the Sobolev type can be compensated
by supplementary assumptions on the boundary behavior of the function.
The question of the inuence of such assumptions on Sobolev-type inequal1
(, ) be the compleities is considered in Sect. 5.6. Let the space Wp,r

1
equipped with the norm ( |u|p dx)1/p +

tion of Lp () C () C()

( |u|r ds)1/r .
1
(, ) is continuously embedded
We show in Theorem 5.6.3 that W1,r
into Lq () with q 1, r q, if and only if

mn (G )1/q C s(i G ) + s(e G )1/r
for every admissible G , where e G = G . Here the sharp inequality
u
Ln/(n1) ()

[ (1 + n/2)]1/n
u
L() +
u
L() ,
n
1
(, ), can be found.
valid for an arbitrary bounded set and u W1,1
5.1 Preliminaries
5.1.1 Notation
In this chapter, as well as in Chaps. 6 and 7, we shall use the symbols introduced in Sect. 1.1.1 and the following notation.
A bounded open subset G of the set is called admissible if G is
a manifold of the class C (this term was understood in a more restrictive
sense in Chap. 2).
be the closure of the set E Rn and let E be the boundary of E.
Let E
Further, let clos E be the closure of E in and let i E be the inner part of
E with respect to , i.e., i E = E.
We put = B , u+ = max{u, 0}, u = u+ u, Et = {x : |u(x)| = t},
Lt = {x : |u(x)| > t}, Nt = {x : |u(x)| t}.
As before, we shall write
1/q
|u| dx
u
Lq () =
5.1 Preliminaries
289
This notation also will be used for q (0, 1) when the right-hand side is a
pseudonorm. (We recall that a linear space is called pseudonormed if there is
a functional
x
0, dened on its elements, which satises the conditions
1. if
x
= 0, then x = 0; 2.
x
= ||
x
, where R1 ; 3. if
xm
0,
ym
0, then
xm + ym
0.) Clearly, in the case 0 < q < 1 the functional

(u, v) =
|u v|q dx
satises the axioms of a metric.

Let C 0,1 () denote the space of functions that satisfy a Lipschitz condition
on any compact subset of .
If is a domain, we endow the space Llp (), p 1, l = 1, 2, . . . (cf.
Sect. 1.1.2) with the norm
l u
Lp () +
u
Lp () ,
where is an open nonempty set with compact closure
. From (1.1.13)
it follows that varying leads to an equivalent norm.
l
() = Llp () Lr () be the space equipped with the
Further, let Wp,r
norm for r 1 and with the psuedonorm for r (0, 1) as follows:
u
Wp,r
l () =
l u
L () +
u
L () .
p
r
l
In accordance with Sect. 1.1.4, Wp,p
() = Wpl (). By Theorems 1.1.5/1
l
and 1.1.5/2, the sets Lp () C () and Wpl () C () are dense in Llp ()

and Wpl (), respectively.
1
() and
5.1.2 Lemmas on Approximation of Functions in Wp,r
1
Lp ()
Lemma 1. The set of functions in L1p () C () L () (p 1) with

1
bounded supports is dense in Wp,r
() ( > r > 0).
1
Proof. Let v Wp,r
(). We use the sequences v(m) and v(m) introduced
1
(), the set of bounded
in Lemma 1.7.1. Since v (m) v and v(m) v in Wp,r
1
1
functions v Lp () with mn (supp v) < is dense in Wp,r
(). Suppose v
satises these conditions. We dene the sequence

vm (x) = m1 x v(x), m = 1, 2, . . . ,
where C0 (B2 ), = 1 on B1 . Obviously,

1/p

1
1 () c
v
L (\B ) + c m
vm v
Wp,r
v
L () mn (supp v)
p
m
+
v
Lr (\Bm ) 0,
290
as m . To approximate each vm by smooth functions it is sucient

to use a partition of unity and mollifying operators (cf. the proof of Theorem 1.1.5/1).

From Lemma 1 we obtain the following corollary.
Corollary. If is domain with nite volume, then the set of functions in
L1p () C () L () with bounded supports is dense in L1p ().
Lemma 2. Let G be an open subset of and let u C 0,1 L1p (), u = 0
outside G. Then there exists a sequence of functions in L1p () C () that
also vanish outside G which converges to u in L1p ().
Proof. Since u can be approximated in L1p () by the sequence u(m) dened
in Lemma 1.7.1, we may assume that u = 0 outside some open set g G with
clos g G.
We let {B (k) } denote a locally nite covering of g by open balls B (k) ,
B (k) G and then we repeat the proof of Theorem 1.1.5/1. The lemma is
proved.

Remark. If we assume that the function u referred to in the statement
we may also assume the functions of an
of Lemma 2 is continuous on ,
approximating sequence to have the same property (cf. Remark 1.1.5).
If, in addition to the condition of Lemma 2, u Lr () then the approxi1
mating sequence can be taken to be convergent in Wp,r
().
Both of these assertions are immediate corollaries of Lemma 2 and are
proved similarly.
5.2 Classes of Sets J, H and the Embedding

L11 () Lq ()
5.2.1 Classes J
Denition. A bounded domain belongs to the class J (
exists a constant M (0, mn ()) such that
def
A (M ) = sup
{G }
mn (G )
< ,
s(i G )
n1
n )
if there
(5.2.1)
where {G } is a collection of admissible subsets of with mn (G ) M and s

is the (n 1)-dimensional area.
The condition (5.2.1) gives a local characterization of the boundary of .
We briey comment on this property. If is a domain with suciently smooth
boundary then it can be easily seen that the (n 1)-dimensional area of the
surface G is bounded from below (up to a constant factor) by the area
5.2 Classes of Sets J , H and the Embedding L11 () Lq ()
291
of i G = G for any G of suciently small volume. So by the classical

isoperimetric inequality we have the relative isoperimetric inequality
mn (G )(n1)/n const s(i G )
and hence belongs to the class J(n1)/n . If has cusps directed into
, then it is intuitively clear that the last inequality still holds. If a cusp is
directed outward from the domain then there exists a sequence of sets G
for which
mn (G )(n1)/n
lim
= .
s(i G )
Along with this property the domain may satisfy (5.2.1) for some > (n
1)/n. The exponent characterizes the degree of sharpness of a cusp.
In what follows we shall see that it is relative (with respect to ) isoperimetric inequalities of the type (5.2.1) (and more complicated ones) that determine the order of integrability of functions in Sobolev spaces.
Lemma 1. Let be an open unit ball and let g be an open subset of
such that i g is a manifold of the class C 0,1 . Then
vn n/(1n)
s(i g)n/(n1) .
min mn (g), mn (\g)
v
2 n1
(5.2.2)
The constant in (5.2.2) is the best possible. More generally, the minimum
value of s(i g) over all sets g with mn (g) = const < vn /2, is attained at the
ball which is orthogonal to .
Proof. It is sucient to assume that 2mn (g) < vn . Applying the spherical
symmetrization of g with respect to a ray l that emanates from the center of
the ball , we obtain a set f symmetric with respect to l and such that
mn (f ) = mn (g),
s(i f ) s(i g).
(The spherical symmetrization with respect to a ray is dened similarly to the

symmetrization with respect to an (n s)-dimensional subspace introduced
in Sect. 2.3.4; we just need to replace the s-dimensional subspaces orthogonal
to Rns by (n 1)-dimensional spheres centered at the origin of the ray.)
Let b be a ball such that b = f and mn (b ) = mn (f ). Since

mn f (b\) = mn (b),
then by the isoperimetric property of the ball we have
s( b) s(i f ).
An elementary calculation shows that the minimum value of s( b)
over all balls with mn ( b) = const < 12 mn () is attained at the ball which
is orthogonal to .
292
It can be easily checked that the function

s( b )
,
[mn ( b )](n1)/n
where b is a ball with radius , orthogonal to and such that mn (b ) <
1

2 mn (), decreases. The proof is complete.
The next corollary follows immediately from Lemma 1.
Corollary 1. If is an n-dimensional ball, then J(n1)/n and

1
1
mn () = vn1
A(n1)/n
(vn /2)(n1)/n ,
2
where vs is the volume of the s-dimensional unit ball.
Corollary 2. A bounded domain starshaped with respect to a ball belongs
to the class J(n1)/n .
is the quasi-isometric image
Proof. According to Lemma 1.1.8, the set
of a ball. This along with Lemma 1 implies
mn (g)(n1)/n C(M )s(i g)
(5.2.3)
for any constant M (0, mn ()) and for all open sets g such that i g
is a manifold of the class C 0,1 and mn (g) M . Here C(M ) is a constant
independent of g. The corollary is proved.

Remark. The condition (5.2.1) does not hold for < (n 1)/n since in
this case
mn (B )
= c lim 1n+n = .
lim
0 s(B )
0
We give an example of a domain that does not belong to the class J for
any .
Example. Consider the domain depicted in Fig. 18. For the sequence of
subsets Qm (m = 2, 4, . . . , ), we have
s(i Qm ) = m2m ,
m2 (Qm ) = m4 ,
m2 (Qm )
m
= m2m4 ,
s(i Qm )
and hence
/ J .
Lemma 2. If a bounded open set is the union of a nite number of open
sets of the class J then it also belongs to J .
Proof. Let = 1 2 , k J (k = 1, 2), and let M = min{M1 , M2 }
where M1 and M2 are constants for 1 and 2 in the denition of the class
J . Let G denote an admissible subset of with mn (G ) M .
293
Fig. 18.
Since
mn (G ) mn (G 1 ) + mn (G 2 )
and s( G ) s(k G ), then
mn (k G )
mn (G )
c
.
s( G )
s(k G )
2
k=1
This along with Lemma 1.1.9/1 and Corollary 5.2.1/2 implies the next
corollary.
Corollary 3. A bounded domain having the cone property belongs to the
class J(n1)/n .
5.2.2 Technical Lemma
The following assertion will be used in Sect. 5.2.3.
Lemma. Let G be an admissible subset of such that s(i G ) < . Then
there exists a sequence of functions {wm }m1 with the properties:
wm C 0,1 (),
wm = 0 in \G ,
wm [0, 1] in ,
for any compactum e G there exists a natural number m(e), such that
wm (x) = 1 for
x e and m m(e),
5. limm sup |wm | dx = s(i G ).
1.
2.
3.
4.
294
Proof. Let be a bounded open set,

, s[(\) G ] < m1 . There
exists a small positive number = (m) such that for some locally nite (in )
covering of (\) G by open balls Bi with radii ri < the inequality

rin1 < 2m1 ,
(5.2.4)
i
holds. Obviously, we may assume that each ball Bi intersects (\) G .

We introduce
the notation
2Bi is the ball with radius 2 ri concentric with

Bi ; C1 = i Bi , C2 = i 2Bi ; (x) = dist(x, i G ).
Let g = {x : x G , (x) < } where = (m) is a small number, (0, ),
such that g is contained in C1 (this can be achieved since the covering
{Bi } is locally nite in ).
We construct a function v(x), which is equal to zero in \G , to 1 (x)
for x g , and to unity on the remaining portion of . This function is
discontinuous on the sets (\) G and g . We eliminate this defect by
using a truncating function (x) that is dened as follows.
Let i C 0,1 (Rn ), i = 1 outside 2Bi , = 0 in Bi , 0 i 1, |i | ri1
in 2Bi , and (x) = inf i {i (x)}.
Obviously, C 0,1 (), = 0 in C1 , = 1 outside C2 . Consider the
function wm = v, which equals zero in C1 and hence vanishes in the
neighborhood of the set of discontinuities of v. Clearly,

(v) dx
|wm | dx =
|| dx +
|v| dx.
\C1
\C1
We note that

|| dx
\C1
|i | dx
\C1
\C1

i
|i | dx c
2Bi
rin1 .
Here and henceforth in this lemma c is a constant that depends only on n.

The preceding inequalities along with (5.2.4) imply

|| dx cm1 .
\C1
Further, since by Theorem 1.2.4

|v| dx 1
\C1
|| dx = 1
s( ) d,
0
where = {x g : (x) = }, it follows for suciently small = (m)

that

|v| dx s(i G ) + cm1 .
\C1
Finally we have
295
|wm | dx s(i G ) + cm1 .
The function wm is equal to zero outside G and to unity outside an neighborhood of G . The lemma is proved.

5.2.3 Embedding L11 () Lq ()
Let G be an open subset of and let
() def
AG
= sup
[mn (G )]
,
s(i G )
(5.2.5)
where the supremum is taken over all admissible sets G with clos G G.
()
Lemma 1. 1. If AG < , 1, then for all functions u C 0,1 ()

equal to zero outside G,
L11 (),
u
Lq () C
u
L1 () ,
(5.2.6)
where q = 1 and C AG .
2. If for all functions u C 0,1 () L11 (), equal to zero outside G, in()
equality (5.2.6) holds, then C AG .
()
Proof. 1. By Lemma 5.1.2/2, it is sucient to prove (5.2.6) for functions

u C () L11 () that are equal to zero outside G.
Since

u
qLq () =
mn (Nt ) d tq ,
0
(1.3.5) implies
u
Lq ()
mn (Nt )1/q dt.
Now we note that, for almost all t > 0,

()
mn (Nt )1/q AG s(Et ).

From this inequality and Theorem 1.2.4 we obtain

()
()
u
Lq () AG
s(Et ) dt = AG
u
L1 () .
0
2. Let G be an admissible subset of G with clos G G. We insert the sequence {wm }m 1 specied in Lemma 5.2.2 into (5.2.6). For any compactum
e G we have

1/q
|wm |q dx
Cs(i G )
and hence mn (G )1/q Cs(i G ). The lemma is proved.
296
For any domain with nite volume we put

1
mn () .
A = A
2
Theorem. Let be a domain with mn () < .
1. If J , where [(n 1)/n, 1], then, for all u L11 (),
inf
u c
Lq () C
u
L1 () ,
cR1
(5.2.7)
where q = 1 and C A .
2. If for all u L11 () inequality (5.2.7) is true, then J with
= q 1 and 2(q1)/q C A for q > 1, C A for 0 < q 1.
Proof. 1. By Corollary 5.1.2, it is sucient to obtain (5.2.7) for functions
u L11 () L () C () with bounded supports. Let denote a number
such that

2mn x : u(x) mn (),

2mn x : u(x) > mn ().
According to Lemma 1,
1/q

(u )q+ dx
A
|u| dx
1/q

( u)q+ dx
A
|u| dx,
{x:u(x)> }
and
{x:u(x)< }
which completes the proof of the rst part of the theorem.

2. For all u L11 (), let inequality (5.2.7) hold and let G be any admissible
subset of with 2mn (G ) mn (). We insert the sequence {wm } specied
in Lemma 5.2.2 into (5.2.7). For any compactum e G we have

|1 c| dx +
q
inf
cR1
1/q
|c| dx
Cs(i G )
q
\G
and consequently,

1/q
min |1 c|q mn (G ) + |c|q mn (\G )
Cs(i G ).
c
The minimum value of the left-hand side with q > 1 is attained at

c=
mn (G )1/(q1)
.
1/(q1)
mn (G )
+ mn (\G )1/(q1)
Hence
[mn (G )mn (\G )]1/q

1/(q1)
[mn (G )
+ mn (\G )1/(q1) ](q1)/q
297
Cs(i G ).
Taking into account the condition 2mn (G ) mn (), we obtain

mn (G )1/q 2(q1)/q Cs(i G ).
The case 0 < q 1 is treated similarly. The theorem is proved.
Lemma 2. Let be a domain with mn () < . The space L1p () is

embedded into Lq () (p 1, q > 0) if and only if
inf
u c
Lq () C
u
Lp ()
cR1
(5.2.8)
for all u L1p ().

Proof. Necessity. Let Z be the subspace of functions equal to a constant
1 () be the factor space W 1 ()/Z , equipped with the
on and let W
p,q
p,q
norm
inf
u c
Lq () +
u
Lp () .
cZ
1 into L 1 (). This mapping is linear,

Let E denote the identity mapping of W
p,q
q
1
continuous, and one to one. Since Lq () Lq (), we see that E is surjective.
By the Banach theorem (cf. Bourbaki [128], I, 3, 3), E is an isomorphism and
hence (5.2.8) holds.
Suciency. Let (5.2.8) be true. We must show that E is surjective.
1
p,q
() is closed in L1p (). So it is sucient to
By (5.2.8), the image of W
1 () considered as a
take into account that, by Corollary 5.1.2, the space W
p,q
1
1
subspace of Lp () is dense in Lp (). The lemma is proved.

Theorem and Lemma 2 immediately imply the next corollary.
Corollary. If is a domain with mn () < then L11 () is embedded
into Lq (), q 1, if and only if
sup
mn (G )1/q
< ,
s(i G )
where the supremum is taken over all admissible subsets G of with mn (G )

1
2 mn ().
Remark. Since a planar domain bounded by a quasicircle belongs to
the class EV11 (cf. Sect. 1.4.8), the embedding L11 () L2 () holds. The last
assertion along with Lemma 5.2.1/2 and the just-formulated Corollary implies
that the union of a nite number of quasidisks belongs to the class J1/2 .
298
5.2.4 Area Minimizing Function M and Embedding of L11 () into

Lq ()
Denition. Let M (0, mn ()). By M () we denote the greatest lower
bound of the numbers s(i G ) considered over all admissible sets G that
satisfy the condition mn (G ) M .
Obviously, M () is nondecreasing in and nonincreasing in M .
We can give an equivalent denition of the class J in terms of the function M . Namely, J if and only if
lim inf M () > 0.
+0
(5.2.9)
Lemma. If M (0, mn ()) and is a domain with nite volume, then

M () > 0
for all (0, M ].
Proof. Let 0 < min{M, mn () M } and let be a domain with a

smooth boundary such that
and 2mn (\) < . If G is an admissible
subset of , M mn (G ) , then obviously,
2mn (G )
and
2mn (\G ) mn () M .
Since is a smooth surface, we have

inf
u c
L() C()
u
L()
cR1
for all u L11 (). Hence, according to the second part of Theorem 5.2.3,

min mn (G ), mn (\G ) C()s( G )
for any admissible subset G of . Thus,
C()s( G )
and therefore M () > 0 for small values of . Since M () is nondecreasing
in , we conclude that M () > 0 for all (0, M ]. The lemma is proved.
It can be easily seen that the condition of the connectedness of as well
as the condition mn () < are essential for the validity of the lemma.
From the lemma we immediately obtain the next corollary.
Corollary. If is a domain of the class J and mn () < , then the
value

sup mn (G ) /s(i G ) : G is an admissible subset of , mn (G ) M
is nite for arbitrary constant M (0, mn ()).
299
In what follows we shall write () instead of mn ()/2 (). The next criterion is proved exactly in the same way as Theorem 2.1.4.
Theorem. Let be a domain with mn () < .
1. If
1
mn ()/2 1
ds
s
B :=
< ,
(s)
s
0
(5.2.10)
.
then (5.2.7) holds for all u L11 () with q = 1/ and C cB1
2. If there is a constant C such that (5.2.7) holds with q < 1 for all
.
u L11 (), then (5.2.10) holds with = q 1 and C cB1
It can be easily veried by the previous Lemma that (5.2.10) holds simultaneously with the condition
M

0
s
M (s)
1
1
ds
<
s
for every M in (0, mn ()).

We shall say that domains satisfying (5.2.10) belong to the class H .
5.2.5 Example of a Domain in J1
We shall show that the union of of the squares Qm = {(x, y) : 2m1
x 3 2m2 , 0 < y < 2m2 } and the rectangles Rm = {(x, y) : 3 2m2 <
x < 2m , 0 < y < 1} (m = 0, 1, . . .) (Fig. 19) belongs to the class J1 and
does not belong to J for < 1.
Let T be the triangle {(x, y) : 0 < y < x/3, 0 < x < 1} contained in .
Further let v = u on \T and v = 3yx1 u on T . Clearly,
Fig. 19.
300
v
L()

v
u
u

+ c
,
r
y L()
y L()
L(T )
where r = (x2 + y 2 )1/2 . Consequently,
u
L()

u

u
L() + c
.
r
L(T )
Since by the obvious inequality

1
1

w(r) dr
w (r)r dr,
0
w(1) = 0,
L11 () is embedded into the space with the norm

r1 u
L(T ) , it follows
that L11 () L(). Applying Lemma 5.2.3/2 we obtain J1 . On
the other hand, the rectangles Gm = Rm {(x, y) : y > 2m1 } satisfy
lim m2 (Gm )/s(i Gm ) = 1. Therefore c1 () c2 for small values of .
Hence L11 () is embedded into L1 () and is not embedded into Lq () with
q > 1.
5.3 Subareal Mappings and the Classes J and H

In this section we introduce and study properties of subareal mappings
of a domain, i.e., the mappings that do not essentially enlarge the (n 1)dimensional area of surfaces. We shall use these mappings to verify the conditions for concrete domains to be in J and H .
5.3.1 Subareal Mappings
Consider a locally quasi-isometric mapping
x Rn .
Let A denote the image of an arbitrary set A under the mapping .
Let x be the matrix (i /xk )ni,k=1 and let det x be the Jacobian of . The
notations x and det x have a similar meaning.
Denition. The mapping is called subareal if there exists a constant k
such that
(5.3.1)
s(i G ) ks(i G )
for any admissible subset G of .
Lemma 1. The mapping is a subareal if and only if
|det x | k
x
,
for almost all x , where

is the norm of a matrix.
(5.3.2)
301
To prove this lemma we need the following assertion.

Lemma 2. Let u C () L11 () and let E be any measurable subset
of . In order that

|x u| dx k
| u| d
(5.3.3)
E
with a constant k independent of u and E, it is necessary and sucient for

the mapping to satisfy (5.3.2). Moreover, (5.3.2) follows from the validity
of (5.3.3) for any ball in .
Proof. The suciency results immediately from (5.3.2) along with

n

u

| u| d =
xi | det x | dx

x
i
E
E
i=1

=
|x u|(x ) | det x |1 dx,
E
where = |x u|1 x u and ( ) stands for the transposed matrix.

Necessity. We x a unit vector and consider the ball B (x0 ) . We
put u(x) = x. By virtue of (5.3.3)

(x ) | det x |1 dx.
dx k
| u| d = k
B (x0 )
B (x0 )
B (x0 )
Passing to the limit as 0 and using the fact that is arbitrary, we

obtain (5.3.3) for almost all x0 . The lemma is proved.

Proof of Lemma 1. Suciency. Let G be an arbitrary admissible subset of . We choose an arbitrary point y i G . Let Uy be a neighborhood of y so small that the set Uy G is represented by the inequality
xn f (x1 , . . . , xn1 ) in some Cartesian coordinate system with an innitely
dierentiable f .
Let be a xed small positive number. Let V denote the set of points
x Uy dened by the equation xn = + f (x1 , . . . , xn1 ) where [0, ]. We
put
for xn < f (x1 , . . . , xn1 ),

1
u (x) = 1 1 for x V ,
0
for xn f (x1 , . . . , xn1 ) + .
Obviously, u (x) is a Lipschitz function. By (5.3.2) and Lemma 2,

|x u | dx k
| u | d.
Uy
Uy
Using Theorem 1.2.4, we can rewrite the last inequality as
302
s(V ) d k 1
s(V ) d.
Since s(V ) is continuous, we have

s(V0 ) k lim inf 1
0
s(V ) d.
0
Taking into account the lower semicontinuity of the area, we obtain

s(V0 ) ks(V0 ).
The latter implies (5.3.1) since y is an arbitrary point.
Necessity. Let be a subareal mapping of . Consider an arbitrary
u C () L11 (). According to Theorem 1.2.2, the level sets of |u| are
smooth manifolds for almost all t > 0. Let B be an arbitrary ball in . By
Theorem 1.2.4,

|x u| dx =
s(B Et ) dt,
B
0
| u| d =
s(B Et ) dt.
B
By the denition of subareal mappings, s(Et ) ks(Et ). Consequently, (5.3.3)

holds. Now it remains to refer to Lemma 2. Lemma 1 is proved.

5.3.2 Estimate for the Function in Terms of Subareal Mappings
The following theorem yields lower bounds for the function , introduced at
the end of Sect. 5.2.4.
Theorem. Let be a domain with nite volume for which there exists a
subareal mapping onto a bounded domain starshaped with respect to a ball.
We put

() = inf
{G }
| det x | dx,
where the inmum is taken over all admissible subsets G of such that
mn (G )
1
mn ().
2
Then there exists a constant Q such that

Q() ()(n1)/n .
Proof. Let G be an admissible subset of with mn (G ) <
put
(5.3.4)
1
2 mn ().
We
303
M = sup mn (G ).
{G }
Since | det x | const > 0 on any compact subset of , it follows that

M < mn (). Taking into account the fact that is starshaped with respect
to a ball and the fact that G is a manifold of the class C 0,1 , by (5.2.2) we
obtain
(n1)/n

Cs (i G ) .
mn (G )
The latter along with (5.3.1) implies

mn (G )
(n1)/n
Cks(i G ).
Thus, if mn (G ) , then

(n1)/n
Ck().
()(n1)/n mn (G )

5.3.3 Estimates for the Function for Special Domains
We give some applications of Theorem 5.3.2.
Example 1. Consider the domain

=
xR :
n
n1

1/2
x2i
< f (xn ), 0 < xn < a ,
i=1
where f is a nonnegative convex function with f (a 0) < and f (0) = 0

(cf. Fig. 20).
We show that the function () specied for satises

n1

n1
k f ()
() f ()
,
(5.3.5)
where k (0, 1) and () is dened by

()
= vn1
f ( )
n1
d.
Proof. Consider the domain Gt = {x : 0 < xn < t}, where t is subject

to the condition
t

n1
1
f ( )
d mn ().
mn (Gt ) = vn1
2
0
Obviously,

n1
.
s( Gt ) = vn1 f (t)
304
Fig. 20.
Since by the denition of (), [mn (Gt )] s( Gt ), the right-hand side

The mapping x = (x1 , . . . , xn1 , f (xn )) maps onto the truncated
cone

n1

i2 < n2 , 0 < n < f (a) .
= :
i=1
The condition (5.3.2) is equivalent to the boundedness of f . Hence the mapping x is subareal.
It remains to give the lower bound for the function () dened in Theorem 5.3.2. Since the Jacobian of the mapping x is equal to f (xn ), we
must estimate the integral

f (xn ) dx,
(5.3.6)
G
> mn (G ) . Since f is nondecreasing, the

from below, provided
integral (5.3.6) attains its minimum value at the set G() . Therefore,
1
m ()
2 n

G
Thus
f (xn ) dx
()
0
f (c) dc
dx =
|x |f (c)
n
vn1
f () .
n

n
() n1 vn1 f () .
Applying Theorem 5.3.2 to the last inequality we obtain the left-hand side
of (5.3.5). The proof is complete.
305
Fig. 21.
For the domain

=

x:
n1

x2i
<
x2
n ,
0 < xn < a ,
1,
i=1
from (5.3.5) it immediately follows that

c1 () c2 ,
for small values of . Consequently, J
(n 1)
,
(n 1) + 1
(n1)
(n1)+1
(5.3.7)
and
/ J for <
(n1)
(n1)+1 .
Example 2. Consider the domain

n1 1/2

2
xi
< f (xn ) ,
= x : 0 < xn < ,
i=1
where f is the nonnegative convex function with f (+0) > and

f (+) = 0 (cf. Fig. 21). Suppose mn () < , i.e.,

n1
f (t)
dt < .
0
We can show that

n1

n1
() f ()
,
k f ()
(5.3.8)
306
where k (0, 1) and () is dened by

n1
dt.
= vn1
f (t)
()
The proof is the same as in the previous example. The role of the auxiliary
subareal mapping is played by

x = x1 , . . . , xn1 , f (xn )
onto the truncated cone

:
n1

i2
<
n2 ,
0 > n > f (0) .
i=1
By (5.3.8) and Theorem 5.2.4, H , > 1, i.e., L11 () is embedded

into Lq (), q = 1/, if and only if
f (t)(n1)(1+ 1 ) dt

< .
( t f ( )n1 d ) 1
(5.3.9)
In particular, for the domain

= x : x21 + + x2n1 < (1 + xn )2 , 0 < xn < ,

we have
c1 () c2 ,
for small , i.e., J
(n1)
(n1)1
(n 1) > 1,
(n 1)
,
(n 1) 1
and
/ J for <
(5.3.10)
(n1)
.
(n1)1
Example 3. Consider the plane spiral domain (cf. Fig. 22) dened in
polar coordinates by
1 1 () > > 1 2 (),
0 < < .
Here 0 < 2 ( + 2) < 1 () < 2 () < 1, 1 , 2 are functions satisfying

a uniform Lipschitz condition on [0, ) and such that 2 1 is convex on
[0, ). Further, we suppose that the area of is nite, i.e.,

2 () 1 () d < .
0
Applying Theorem 5.3.2 to the subareal mapping

1 = 1
of onto the triangle

1
1 () + 2 () ,
2
2 =

1
2 () 1 () ,
2
307
Fig. 22.
| 1 | < 2 ,
0 < 2 <

1
2 () 1 () ,
2
and using the same arguments as in Example 1, we obtain

c1 2 () 1 () () c2 2 () 1 () ,
(5.3.11)
where () is the function dened by

=
(5.3.12)
d d.
{ei :>()}

ei : 1(8+)1 > > 1(8+)1 c(8+) , 0 < < , (5.3.13)
where 0 < c < 2( 1), > 1, we have c1 () c2 with =

/( 1) for small . Thus, the domain (5.3.13) belongs to J/(1) .
A more complicated example of a domain in J , > 1, is considered in
the following section.
Remark. Incidentally, if x is a subareal mapping of onto a bounded
starshaped domain, then, obviously, L11 () is embedded into the space with
the norm

|u|n/(n1) | det x | dx
(n1)/n
In particular, for domains in Examples 1 and 2 we have

L11 () Ln/(n1) , f (xn ) dx .
308
This is an illustration of the natural idea that the limit exponent n/

(n 1) is also preserved for bad domains if we consider a weighed Lebesgue
measure, degenerating at bad boundary points, instead of mn .
In connection with this remark we note that, although the present chapter
deals with the problem of the integrability of functions in L11 () with respect
to Lebesgue measure, the proofs from Sect. 5.2.3 do not change essentially
after replacing Lq () by the space Lq (, ), where is an arbitrary measure
in (cf. Chap. 2).
5.4 Two-Sided Estimates for the Function for the

Domain in Nikod
yms Example
In this section we consider the domain specied in Example 1.1.4/1 provided
m = (2m1 ), where is a Lipschitz function on [0,1] such that (2t) (t).
We shall show that () ().
Lemma. If G is an admissible subset of , G C = , then

s(i G ) k m2 (G ) , k = const > 0.
(5.4.1)
Proof. Let Gm = (Am Bm ) G . Let N denote the smallest number for

which

s(i GN ) 2N 1 .
This and the obvious inequality m2 (Gm ) 2m1 imply that

2N ks(i G ).
Gm
m2
(5.4.2)
mN
Since

s(i Gm ) < 2m1
for all m < N , then Am Bm does not contain components of i G , connecting

the polygonal line abcd with the segment ef (cf. Fig. 23). So for m < N
2s(i Gm ) s(e Gm ),
where e A = A\i A. From this along with the isoperimetric inequality
s(Gm )2 4m2 (Gm )
we obtain
m2 (Gm )1/2 cs(i Gm ).
(5.4.3)
5.4 Two-Sided Estimates for the Function for the Domain
309
Fig. 23.
Summing (5.4.3) over m and using the inequality

1/2
am
a1/2
m ,
with positive am , we obtain

m2

1/2
Gm
cs(i G ).
(5.4.4)
m<N
Combining (5.4.2) with (5.4.4), we arrive at the required estimate.

Proposition. The function satises
k1 () () k2 (),
where k1 and k2 are positive constants.
Proof. From (5.4.1) and Theorem 1.2.4 it follows that

m2 (Nt ) dt
|u| dx
k
0
(5.4.5)
for all u C () that vanish on C. From Lemma 5.1.2/2 we obtain that this
inequality is valid for all u C 0,1 (), u = 0 on C.
Now let u be an arbitrary function in C 0,1 () and let be a continuous
piecewise linear function on [0,1] equal to zero on [0, 1/3] and to unity on
(2/3,1). We put
310
u(1) (x, y) = u(x, y)(y),

u(2) (x, y) = u(x, y) 1 (y) .
Since |u| |u(1) | + |u(2) |, we have
(1)
(2)
Nt Nt/2 Nt/2
and hence
(1)
(2)
m2 (Nt ) m2 Nt/2 + m2 Nt/2 .
Taking into account the condition C 0,1 [0, 1], we obtain

(2)

(1)
m2 (Nt ) m2 Nt/2 + m2 Nt/2 .
This and (5.4.5) applied to u(1) yield

(1)
(2)

u dx dy +
u dx dy .
m2 (Nt ) dt K
Here the right-hand side does not exceed

|u| dx dy +
|u| dx dy ,
K
(5.4.6)
where = {(x, y) : |y| < 2/3}.

We give a bound for the integral over in (5.4.6). Let (x, y) and let
(x, z) C. Obviously

u(x, y) u(x, z) + u(x, y) d
y,
where the integration is taken over a vertical segment, contained in and
passing through (x, 0). After integration over x, y, z we obtain

|u| dx dy c
|u| dx dy +
|u| dx dy .
Thus

m2 (Nt ) dt K
|u| dx dy +

|u| dx dy .
The rectangle C belongs to the class J1/2 . Therefore, if u = 0 on a subset

of with area not less than 12 m2 (), then

|u| dx dy K
|u| dx dy.
C
Consequently,
5.5 Compactness of the Embedding L11 () Lq () (q 1)

m2 (Nt ) dt K
311
|u| dx dy.
(5.4.7)
Inserting the sequence {wm }, constructed in Lemma 5.2.2, into (5.4.7), we obtain (5.4.1) for any admissible set with an area that does not exceed 12 mn ().
Thus, the lower bound for () is obtained.
To derive the estimate () c() it is sucient to note that the sequence
{Gm }m1 , where Gm is the interior of Am Bm , satises

s(i Gm ) = 2m1 ,

1 m1
2
2
m2 (Gm ) 2m1 .
3
3
The proposition is proved.

Obviously, the Nikod
ym domain belongs to J ( 1) if and only if
lim inf t (t) > 0.
t+0
(5.4.8)
Similarly, H , > 1, i.e., L11 () L1/ () if and only if

1
1 1
ds
s
< .
(s)
s
0
We shall return to the domain considered here in Sect. 6.5.
5.5 Compactness of the Embedding L11 () Lq ()

(q 1)
5.5.1 Class J
Denition. The set belongs to the class J , > (n 1)/n, if
lim sup
mn (G )
= 0,
s(i G )
(5.5.1)
where the supremum is taken over all admissible subsets G of such that
mn (G ) .
Remark 1. It is clear that (5.5.1) is equivalent to
lim M () = ,
(5.5.2)
where M is a xed number in (0, mn ()).

Remark 2. The exponent in the denition of the class J exceeds (n
1)/n since in the case = (n 1)/n we have

(n1)/n
mn (B )
= const s(B ).
312
5.5.2 Compactness Criterion

Theorem. Let mn () < and let be a domain. For the compactness of
the embedding L11 () Lq (), where n/(n 1) > q 1, it is necessary and
sucient that belong to J with = q 1 .
Proof. Suciency. Let u be an arbitrary function in L11 () L ()
C () with bounded support. By Corollary 5.1.2 the set of such functions is
dense in L11 (). Obviously,

q
q
q
|u| dx c
|u| dx + mn ()
N

q
q
c1
|u| + dx + mn ()
for 0. Let be such that

mn (N ) ,
mn (L ) ,
where is an arbitrary number in (0, M ] and M is the constant in the denition of J .

Using Lemma 5.2.3/1, we obtain

q

q
|u| + dx
|u| dx .
(5.5.3)
[M ()]q
Let be a bounded set with smooth boundary such that
2mn (\) < .
Since mn (N ) , we have 2mn ( N ) . Consequently,

|u|q dx 21 q .
Thus,
c
u
Lq ()
1/q
M ()

|u| dx +
mn ()
1/q
u
Lq () .
(5.5.4)
Let {uk }k1 be a sequence satisfying
uk
L() +
uk
L() 1.
Since the boundary of is smooth, the embedding operator L11 () Lq ()
is compact and we may suppose that {uk }k1 is a Cauchy sequence in Lq ().
By (5.5.4),
5.5 Compactness of the Embedding L11 () Lq () (q 1)
c
um ul
Lq ()
313

1/q
mn ()
1/q
+
2
um ul
Lq ()
M ()
and hence
c lim sup
um ul
Lq () 2
m,l
1/q
.
M ()
It remains to pass to the limit in the right-hand side as +0 and

take (5.5.2) into account.
Necessity. Let the embedding L11 () Lq () be compact. Then L11 ()
L1 () and elements of a unit ball in W11 () have absolutely equicontinuous
norms in Lq (). Hence, for all u L11 ()
1/q
G
|u|q dx
()

|u| + |u| dx,
(5.5.5)
where G is an arbitrary admissible subset of whose measure does not exceed

and () converges to zero as +0.
We insert the sequence {wm } constructed in Lemma 5.2.2 into (5.5.5).
Then, for any compactum K G ,

mn (K)1/q c() s(i G ) + mn (G ) ,
and hence
mn (G )1/q c1 ()s(i G ).
Remark. One can prove, in a similar way, that the compactness of the
embedding L11 () into Lq (), q < 1, holds if and only if H1/q , i.e., the
embedding L11 () Lq (), q > 1, is compact and continuous simultaneously.
The role of the inequality (5.5.3) in the proof of the statement should be
played by the estimate
|u|
dx c
+
M
q
M ()
q
1q
1q
q
|u| dx
which can be obtained by obvious changes in the proof of Theorem 2.3.8.

Example. The condition (5.5.1) for the domain

= x = (x , xn ), x = (x1 , . . . , xn1 ) : |x | < f (xn ), 0 < xn < a

(cf. Example 5.3.3/1) is equivalent to

x

n1

1n
f ( )
f (x)
d
= 0.
lim
x0
Since
(5.5.6)
314
x
f (t)
n1

n1
d f (x)
x,
we see that (5.5.6) holds if

(n1)(1)
lim x f (x)
=0
x0
(in particular, (5.5.6) always holds for = 1).

By the last Remark, the condition (5.3.11) is the equivalent to the compactness of the embedding L11 () into Lq (), q > 1, for the innite funnel in
Example 5.3.3/2.
1
5.6 Embedding W1,r
(, ) Lq ()
5.6.1 Class K,
and
Let r > 0, u C()
1/r
|u|r ds
,
u
Lr () =
where s is the (n1)-dimensional Hausdor measure. If r 1, then

u
Lr ()
is a norm and it is a pseudonorm for r (0, 1) (cf. Sect. 5.1.1).
1
Denition 1. We denote by Wp,r
(, ) the completion of the set of
1
functions in Lp () C () C() with respect to the norm (pseudonorm)
u
Lp () +
u
Lr () .
1
(, )
In this section we study the conditions for the embedding W1,r
0,1
Lq (). Contrary to Sobolevs theorem for domains of the class C , the norm
in Lr () does not always play the role of a weak perturbation for the
L1 ()-norm of the gradient in the inequality

(5.6.1)
u
Lq () c
u
L1 () +
u
Lr () .
In the case of a bad boundary the exponent q may depend on the order of
integrability of the function on the boundary.
1
In particular, we shall see that functions in W1,r
(, ) are integrable
with power q = n/(n 1) in , if r = 1 and is an arbitrary open set.
Denition 2. An open set belongs to the class K, , if there exists a
constant E such that

mn (g) E s(i g) + s(e g)
(5.6.2)
for any admissible set g (here e g = g ).
1
5.6 Embedding W1,r
(, ) Lq ()
315
5.6.2 Examples of Sets in K,

Example 1. An arbitrary open set belongs to the class K(n1)/n,1 since
the condition (5.6.2) for = (n 1)/n, = 1 is the classical isoperimetric
inequality

(n1)/n [ (1 + n/2)]1/n
s(g)
(5.6.3)
mn (g)
n
(cf. Sect. 9.1.5 and Remark 9.2.2).
Proposition. If K,1 , where (n 1)/n then K, , where
is an arbitrary number in [(n 1)/n, 1].
Proof. From (5.6.3) and

mn (g)
it follows that

cE s(g)
mn (g)
E s(g),
(5.6.4)
with = + (1 )(n 1)/n
for any [0, 1]. If s(e g) s(i g), then

mn (g)
2cE s(i g).
(5.6.5)
Otherwise, if s(e g) > s(i g), then (5.6.4) yields

mn (g)

E s(g) .
(5.6.6)
Combining (5.6.5) with (5.6.6), we complete the proof.

Example 2. We show that the plane domain

= (x, y) : 0 < x < , 0 < y < (1 + x)1 ,

where 0 < 1, belongs to K1/,1 and by the Proposition, K/, ,
where is an arbitrary number in [/2, 1].
Let g be an arbitrary admissible subset of . Obviously,
s(g) s(Pr g),
(5.6.7)
Ox
where s is the length and PrOx is the orthogonal projection onto the axis Ox.
Since 1, we have

s(PrOx g)

(1 + x)1 dx
(1 + x)1 dx = 1 s(Pr g) + 1 1
PrOx g
Consequently,
0

1
s(Pr g) .
Ox
Ox
316
1/

1
(1 + x)
dx
1/ s(g).
(5.6.8)
PrOx g
Taking into account that

m2 (g)
(1 + x)1 dx,
PrOx g
from (5.6.8) we obtain

m2 (g)1/ 1/ s(g),
(5.6.9)
which means that K1/,1 .

Example 3. We shall show that any set in J with nite volume belongs
to the class K, , where is an arbitrary positive number.
For any admissible subset g of satisfying the condition mn (g) M we
have
(5.6.10)
mn (g) A(M )s(i g).
Let mn (g) > M and
2n s(e g) < M (n1)/n ,
where n =
[ (1 + n/2)]1/n
.
n
(5.6.11)
By the isoperimetric inequality (5.6.3),

M (n1)/n mn (g)(n1)/n n s(i g) + s(e g)
and hence
M (n1)/n 2n s(i g).
This implies
mn (g) mn () 2n M (1n)/n s(i g).
(5.6.12)
If (5.6.11) is not valid, then

mn (g) mn () (2n ) s(e g) .
(5.6.13)
From (5.6.10), (5.6.12), and (5.6.13) it follows that belongs to K, .

1
(, ) Lq ()
5.6.3 Continuity of the Embedding Operator W1,r
Theorem. If K, , where 1, , then (5.6.1) holds for all

1
(, ) with q = 1/, r = 1/.
u W1,1/
Proof. Consider the case 1 > . Let u be an arbitrary function in
with bounded support. From Lemma 1.2.3 and Lemma 1.3.5/1
C ()C()
we obtain that
1
5.6 Embedding W1,r
(, ) Lq ()
u
L1/ ()
mn (Nt )
d t
1/
317

.
(5.6.14)
We introduce the set

At = t : s(Et ) s(Nt )
and represent the integral above in the form

+
,
At
CAt
where CAt is the complement of At with respect to the positive halfaxis. We

give a bound for the integral

I1 =
mn (Nt )/ d t1/ .
At
Since K, , then
mn (Nt ) 2E s(Nt )
for almost all t At and hence

1/
1/
1/
I1 (2E )
(2E )
s(Nt ) d t
At
|u|1/ ds.
(5.6.15)
Now we consider the integral

I2 =
mn (Nt )
/ 1/
.
d t
CAt
Obviously,
I2
mn (Nt )
(1)/

dt sup mn (N )
.
CAt
By (5.6.2),
for CAt , and hence

I2 2E
mn (N )
2E s(E )

(1)/
s(Et ) dt sup mn (N )
.
>0
The preceding inequality and Theorem 1.2.4 yield

I2 2E 1
u
L1 ()
u
L1/ () .
(1)/
From (5.6.14)(5.6.16) it follows that
(5.6.16)
318
u
L1/ () cE
u
L1 ()
u
1
L1/ () +
u
L1/ () .
Now consider the case 1. The condition K, implies

mn (N ) dt E
s(Nt ) dt .
s(Et ) dt +
0
(5.6.17)
Applying Lemma 1.2.3 and Lemma 1.3.5/1, we obtain

|u|1/ dx
s(Nt )
1/
dt
mn (Nt )
1/
dt

s(Nt ) d t1/ =

|u|1/ ds.
These estimates together with (5.6.17) lead to

u
L1/ () E
u
L1 () +
u
L1/ () .
(5.6.18)
This concludes the proof.
We give an example that shows that the Theorem is not true if K, ,

> , and r = 1/.
Example. In Example 5.6.2/2 we showed that the domain

= (x, y) : 0 < x < , 0 < y < (1 + x)21
belongs to the class K1/2, with (0, 1/2). Consider the sequence of functions
(x, y) um (x, y) = (1 + x)m , m > .
1
(, ) since it can be
Each of these functions obviously belongs to W1,1/
approximated in the norm of the space mentioned after multiplication by an
expanding sequence of truncating functions. We have
um
L2 () = 21/2 (m )1/2 ,
um
L() = m /(m 2 + 1),
um
L1/ () > (m ) .
For u = um and m + 0 the left-hand side of the inequality

u
L2 () C
u
L() +
u
L1/ ()
grows more rapidly than its right-hand side and hence (5.6.1) is not true for
q = 1 , r = 1 .
Remark. A review of the proof of the previous Theorem shows that the
theorem also remains valid for < , provided (5.6.1) in its statement is
replaced by

u
Lq () C
u
L() +
s(Nt )
1/r

dt .
319
(5.6.19)
Corollary. The inequality
u
Ln/(n1) ()

[ (1 + n/2)]1/n
u
L() +
u
L() ,
n
(5.6.20)
1
(, ). The constant
holds for an arbitrary bounded set and u W1,1
in (5.6.20) is the best possible.
The proof follows immediately from the isoperimetric inequality (5.6.14)

and (5.6.18). That the constant is exact was already remarked in Sect. 1.4.2,
where the inequality (1.4.14) was derived.

A substantial part of the results presented in this chapter were stated in the
authors paper [547].
Section 5.2. Lemma 5.2.1/1 was proved by Burago and the author [151].
Other results of this section (except Lemma 5.2.3/2) are due to the author.
The Poincare-type inequality (5.2.7) with q = n/(n 1) was deduced
from the isoperimetric inequality (5.2.1), where = (n 1)/n and M =
1
m (), simultaneously and independently from the author by Fleming and
2 n
Rishel [282].
For the further development of the idea concerning the relation between integral and geometrical inequalities see Miranda [608], Burago and Mazya [150]
(the integrability of traces on of functions in BV (), and so on; these results are presented in Chap. 9), Federer [271, 272] (embedding theorems for
currents), Klimov [428430, 433] (embedding theorems involving Birnbaum
Orlicz and rearrangement invariant spaces), Michael and Simon [600], Homan
and Spruck [379], Aubin [55] (the SobolevGagliardo inequalities for functions
on manifolds), Otsuki [654], Martin and M. Milman [523] (embedding theorems involving rearrangement invariant spaces). Lemma 5.2.3/2 was proved
by Deny and J.L. Lions [234].
The following optimal result was obtained by Cianchi in [193].
Theorem. Let n 2 and q (0, n/(n 1)]. Let be the unit ball in Rn .
Then
1/q
vn
u
L1 ()
(5.7.1)
u u
Lq ()
2vn1
for u L11 (), where u denotes the mean value of u over . The constant
1/q
vn (2vn1 )1 is sharp.
320
Although the constant vn (2vn1 )1 is sharp in (5.7.1), it is not achieved.

It is actually achieved when u equals the characteristic function of a half-ball
in a version of (5.7.1) for functions of bounded variation, where
u
L1 () is
replaced by the total variation of the distributional gradient of u (see Chap. 9).
The fact that among all subsets of the Euclidean ball of a given measure,
the ones minimizing the surface area in the interior of the ball are exactly
spherical caps perpendicular to the boundary, is due to Burago and Mazya
[150] (see also Almgren [40] and Bokowski and Sperner [124]). Cianchi noted
in [193] that for the unit ball in Rn symmetrization arguments as in the proof
of Lemma 5.2.1/1 and basic calculus give the formula

vn /2 (s) = B A1 (s) for s [0, vn /2],
1/q
where
A() = vn1
(sin t) dt + (tan )
/2

(cos t) dt
n
for [0, /2]
and

B() = (n 1)vn1 (tan )
n1
/2
(cos t)n2 dt
for [0, /2].
Isoperimetric inequalities in the form (5.2.2) with optimal constants for

special convex sets in the plane are discussed in Cianchi [192]. The case
when is an n-dimensional convex cone is considered in P.-L. Lions and
Pacella [502].
Except for the methods of obtaining Sobolev-type inequalities mentioned
in the Comments to Sect. 1.4, there is another powerful approach to reduce
many multidimensional geometric and analytic inequalities to specic problems in dimension one, called the localization technique. It was proposed in
1960 by Payne and Weinberger [656] and developed in 1990s by Lov
asz and
Simonovits [508], together with Kannan in [413]. Localization technique was
also used by Gromov and V. Milman [326] to recover the isoperimetric theorem
on the sphere.
The advantage of localization over triangular mappings is that it often
leads to optimal results. For example, localization easily reduces the Gaussian
isoperimetric inequality to dimension one, which was demonstrated in Bobkov
[110]. More generally, with this technique one may obtain sharp dilation-type
inequalities for convex domains and so-called convex probability measures
(cf. F. Nazarov, Sodin, and Volberg [629], Bobkov and F. Nazarov [119]). Another interesting line of applications is Sobolev-type inequalities (cf. Kannan,
Lovasz, and Simonovits [413], Bobkov [113]).
Here is a statement from Bobkov [113] that renes one of main results in
Kannan, Lovasz, and Simonovits [413]. Let Rn be a bounded convex
domain with, for deniteness, volume one. The best constant in the Poincaretype inequality
321

|u u | dx A
|u| dx
(5.7.2)
in the class of all smooth functions u satises

2 1/4

4
2
|x| dx
|x| dx
,
AC
(5.7.3)
where C is an absolute constant. In the

case of the Euclidean ball centered
at the origin and of radius of order n, as n , so that it has volume
equal to 1, we have that the right-hand side is of order 1. This follows also
from (5.7.1) with q = 1. The localization method was further developed in
Fradelizi and Guedon [283, 284]. We also mention that the value 21 diam
of the constant A in (5.7.2) for a convex domain, which is the best constant
formulated in terms of diameter, was found by Acosta and Duran [1].
Membership of two-dimensional domains with Holder continuous boundolder exponent, was estabary in the classes J , with depending on the H
lished by Cianchi in [192]. The case of n-dimensional domains with boundaries
enjoying more general regularity properties was considered by Labutin in [471].
As Buckley and Koskela showed in [147], bounded planar Jordan domains are
in the class J1/2 only if they are Jordan domains. By Bojarski [122] this
condition is also sucient.
Section 5.3. The contents of this section are borrowed from the authors paper [538], where subareal mappings that preserve the space L11 (),
were considered in particular. Since then various properties of homeomorphisms of Euclidean domains generating a bounded composition operator
of Sobolev spaces were studied. Without aiming at completeness we mention the works by Vodopyanov and Goldshtein [778]; Mazya and Shaposhnikova [578, Sect. 9.4], [588]; Holopainen and Rickman [382]; Ukhlov [767];
[768], Goldshtein and Gurov [314]; Goldshtein, Gurov, and Romanov [315];
Vodopyanov [777]; Hajlasz [340]; Kauhanen, Koskela, and Mal
y [417]; Troyanov and Vodopyanov [761]; Hencl, Koskela, and Mal
y [377]; and Hencl and
Koskela [376], Goldshtein and Ukhlov [319].
A dierent direction is taken by a theory of the so-called Sobolev mappings of Riemannian manifolds and metric spaces. This vast and diverse area,
important for applications to geometry, physics, elasticity etc. is outside the
scope of the present book.
Section 5.4. The estimates for the relative isoperimetric function are
presented in Mazya [552].
Section 5.5. The results of this section are borrowed from [547].
Section 5.6. The contents of this section are taken from the authors thesis [529]. The inequality (5.6.20) can be found in Mazya [538]. This inequality
proved to be useful in the theory of the Robin boundary value problem for
arbitrary domains developed by Daners [221].
6
Integrability of Functions in the Space L1p ()
Here we continue our study of integral inequalities for functions with unrestricted boundary values started in the previous chapter. For the embedding
operator L1p () Lq (), p 1, we nd necessary and sucient conditions
on ensuring the continuity of this operator (Sects. 6.26.4). To get criteria,
analogous to those obtained in Sect. 2.2, for the space L1p (), we introduce
classes of sets dened with the aid of the so-called p-conductivity, which plays
the same role as p-capacity in Chap. 2. Geometrical conditions formulated in
terms of the isoperimetric inequalities prove to be only sucient if p > 1.
We give some details. Let G be a bounded open subset of and let F be
a relatively closed subset of G. The dierence K = G\F is called a conductor
and its p-conductivity cp (K) is dened as inf f pLp () extended over Lipschitz functions f such that f 1 on F and f 0 on \G. The inmum of
cp (K) taken over all conductors K with mn (F ) t and mn (G) M , where
M (0, mn ()), is called the p-conductivity minimizing function M,p (t). We
say that belongs to the class Ip, if
M,p (t) const tp ,
with t (0, M ]. For p = 1 this class coincides with the class J introduced
in Chap. 5. By one of the results in the present chapter, Theorem 6.3.3, the
inclusion Ip, implies the inequality of the form
1

uLq () C uLp () + uLs ()
u
Lr () ,
where q 1/ if p < 1 and q < 1/ if p > 1. Conversely, the last
multiplicative inequality with q < 1/ ensures the inclusion Ip, (see
Theorem 6.3.3 for other assumptions about p, q, s, and ). By this result and
by Theorem 6.4.2 the inequality

uL1/ () C uLp () + uLs ()
holds with s < 1/ if and only if Ip, in the case p 1 and Hp,
in the case p > 1. The class Hp, is dened by the condition
DOI 10.1007/978-3-642-15564-2 6,
323
324
6 Integrability of Functions in the Space L1p ()

M

0
p
M,p ( )
1/(p1)
d
<
in Sect. 6.4.2.
As stated in Proposition 6.3.5/1, the p-conductivity minimizing function
admits the lower estimate in terms of the area minimizing function
M
1p
p/(1p)

M,p (t)
d
,
M ()
t
which shows, for instance, that J+(p1)/p Ip, . The failure of the converse
inclusion is illustrated in Sect. 6.5, where we nd equivalent lower and upper
estimates for the p-conductivity minimizing function for the Nikod
ym domain
already dealt with in Sect. 5.4.
Some generalizations of the criteria mentioned are discussed in Sect. 6.6
and then, in Sect. 6.7, the necessary and sucient conditions for integral
inequalities involving domains of innite volume are found. Next, in Sect. 6.8
we characterize the domains for which the embedding L1p () Lq () is
compact. Sucient conditions on for the boundedness and compactness of
the embedding operator Llp () Lq (), l 1, following from the previous
results are obtained in Sect. 6.9. We present applications to the Neumann and
other boundary value problems for strongly elliptic operators giving criteria of
their solvability and discreteness of spectrum in Sects. 6.10 and 6.11. Another
topic of Sect. 6.11 concerns inequalities containing integrals over the boundary,
which extend those in Sect. 5.6 to the case p > 1. Here we prove that the
rened Friedrichs-type inequality

uLq () C uLp () + uLr ()
holds with n > p, (np)r p(n1), r 1, and q = rn/(n1) for an arbitrary
open set with nite volume and show that the exponent q = rn/(n 1) is
critical provided no regularity assumptions on the domain are made.
6.1 Conductivity
6.1.1 Equivalence of Certain Denitions of Conductivity
Let be an open set in Rn . Let F and G denote bounded closed (in ) and
open subsets of , respectively, F G.
As we mentioned in the introduction to this chapter, the set K = G\F is
called a conductor . In what follows U (K) is the class of functions f C 0,1 ()
with f (x) 1 for x F and f (x) 0 for x \G. The value

p
|f | dx : f U (K)
cp (K) = inf
is called the p-conductivity of the conductor K.
6.1 Conductivity
325
Lemma 1. Let V (K) denote the class of functions {f C () :

f (x) = 1 for x F, f (x) = 0 for x \G}. Then

p
cp (K) = inf
|f | dx : f V (K) .
Proof. Since U (K) V (K), it suces to obtain only the lower bound
for cp (K). Let (0, 1) and let f U (K) be such that

|f |p dx cp (K) + .
We put = (1 + ) f and = min(+ , 1). Then

2

1/p
(1 + )2 .
Lp () cp (K) +
(6.1.1)
We introduce some notation: 1 = {x : = 1}, 2 = {x : = 0}, and

Q = {x : 1 > (x) > 0}. Since Q = {x : (1 + )1 > f > (1 + )2 }, we have
clos Q K.
We construct a locally nite (in ) covering of the set clos Q by open
balls B0,i , i = 1, 2, . . . . Let A0 denote the union of these balls. By the inclusion
clos Q K we can choose the covering to satisfy clos A0 K. Next we
(k = 1, 2) by open
construct locally nite (in ) coverings of the sets k \A0
balls Bk,i such that the closures in of the sets Ak = i Bk,i (k = 1, 2)
are disjoint with clos Q. The latter is possible since (k \A0 ) clos Q = .
Clearly, F A1 and (\G) A2 .
Let k,i D(Bk,i ), k = 0, 1, 2; i = 1, 2, . . . , and let
2

k,i = 1
in .
k=0 i=1
For any i = 1, 2, . . . , we introduce a function i D(B0,i ) such that

(0,i i )
i .
L ()
p
Next we put v0,i = i , v1,i = 1,i , v2,i = 0,

u=
2
vk,i .
(6.1.2)
k=0 i=1
The function u is innitely dierentiable in since each point of is contained only in a nite number of balls Bk,i and therefore (6.1.2) has a nite
number of summands. Obviously,

( )
Lp ()
2

(k,i vk,i )
k=0 i=1
Lp ()
326
Since = 1 on B1,i , v1,i = 1,i and = 0 on B2,i , v2,i = 0, it follows that

(0,i i )

( u)
Lp ()
Lp ()
/(1 ).
(6.1.3)
i=1
Using F A1 and F A0 = we have v0,i = i = 0, v2,i = 0, v1,i = 1,i ,

0,i = 0, 2,i = 0 on F . Therefore
u=
v1,i =
i=1
2

k,i = 1 on F.
k=0 i=1
Also, vk,i = 0 on \G and hence u = 0 on the same set. Thus, u V (K).

Finally, by (6.1.1) and (6.1.3),

1/p
+ (1 )1 ,
uLp () (1 + )2 cp (K) +

Let T (K) = {f V (K) : 0 f (x) 1on K}.
Henceforth, the following modication of Lemma 1 will be useful.
Lemma 2. The equality

|f | dx : f T (K)
p
cp (K) = inf
is valid.
Proof. Since T (K) V (K), it suces to obtain only the lower bound
for cp (K). Let > 0, C (, +), (t) = 1 for t 1, (t) = 0 for
t 0, 0 (t) 1 + . Further, let V (K). We introduce the function
f = () T (K). Obviously,

p
() ||p dx (1 + )p
|f |p dx =
||p dx
and consequently,

|f | dx : f T (K)
p
inf
cp (K).
The result follows.

6.1.2 Some Properties of Conductivity
We shall comment on some simple properties of p-conductivity.
6.1 Conductivity
327
Consider two conductors K = G\F and K = G \F contained in . We

say that K is a part of K(K K) if F F G G.
The denition of p-conductivity immediately implies the following proposition.
Proposition 1. If K K, then
cp (K) cp (K ).
Proposition 2. Given any > 0 and any conductor K with nite pconductivity, we can construct a conductor K K such that
cp (K ) cp (K) 0.
(6.1.4)
The conductor K can be chosen so that i F and i G are C -manifolds.

Proof. The right inequality follows from Proposition 1.
Let a function f U (K) satisfy

cp (K) + /2 >
|f |p dx
(6.1.5)
and let

2 = 1
+ 2cp (K)
2 + 2cp (K)
1/p
.
We may assume that the sets {x : f (x) = 1 } and {x : f (x) = }

are C -manifolds since otherwise can be replaced by an arbitrarily close
number having the aforementioned property. We construct the conductor K
as follows: K = G \F , where F = {x : f (x) 1 }, G = {x :
f (x) > }. Then (6.1.5) implies
p

cp (K) + /2
f (x)

>
1 2 dx.
(1 2)p
The function (1 2)1 (f ) is contained in U (K ). Hence

cp (K) + /2
cp (K ),
(1 2)p
which is equivalent to the left inequality in (6.1.4).
Proposition 3. Let K1 = G1 \F1 and K2 = G2 \F2 be any conductors

in . Then
(6.1.6)
cp (K ) + cp (K ) cp (K1 ) + cp (K2 ),
where K = (G1 G2 )\(F1 F2 ), K = (G1 G2 )\(F1 F2 ).
328
Proof. Let be an arbitrary positive number and let f1 , f2 be functions

in U (K1 ), U (K2 ), respectively, satisfying

|fi |p dx < cp (Ki ) + .
(6.1.7)
Ki
We introduce the functions

M = sup{f1 , f2 },
m = inf{f1 , f2 }.
It is clear that M and m satisfy the Lipschitz condition in and that M 1

on F1 F2 , M 0 on \(G1 G2 ), as well as m 1 on F1 F2 , m 0 on
\(G1 G2 ). Besides,

|M | dx +
|m| dx =
2

i=1
|fi |p dx.
Ki
This along with (6.1.7) and Lemma 6.1.1/1 implies (6.1.6).

6.1.3 Dirichlet Principle with Prescribed Level Surfaces and Its
Corollaries
The proofs of the following Lemmas 1 and 2 and Corollaries 1 and 2 do not differ from the proofs of similar assertions on the (p, )-capacity in Sects. 2.2.1
2.2.3.
Lemma 1. For any conductor K in with nite p-conductivity we have

cp (K) =
inf
f V (K)
f 1
Lp1 (E ) d
1p
,
(6.1.8)
where E = {x : f (x) = }.
Lemma 2. Let f be in C () L1p (). Then, for almost all t,

s(Et )
p/(p1)
f 1
Lp1 (Et )

d
mn (Lt ) ,
dt
(6.1.9)
where Lt = {x : f (x) > t}.

Corollary 1. For any conductor K in the inequality
1p
1

d
dt
cp (K) inf
mn (Lt )
: f V (K)
[s(Et )]p/(p1)
0 dt
(6.1.10)
holds.
Corollary 2. Let F be an open set closed in and let G, H be bounded
open subsets of such that
6.2 Multiplicative Inequality for Functions Which Vanish on a Subset of
F G,
329
clos G H.
The conductors
K (1) = G\F,
K (2) = H\ clos G,
K (3) = H\F,

(1) 1/(p1) (2) 1/(p1) (3) 1/(p1)
+ cp K
cp K
.
cp K
(6.1.11)
We present one more property of the p-conductivity, which is proved similarly to Theorem 2.3.1.
Lemma 3. Let u C 0,1 (), u = 0 on the exterior of an open bounded set
G and let Kt be the conductor G\Nt . Then, for p 1,

p
pp
cp (Kt ) d t
|u|p dx.
(6.1.12)
(p 1)p1
0
(For p = 1 the coecient in front of the second integral in (6.1.12) is equal to
unity.)
6.2 Multiplicative Inequality for Functions Which Vanish

on a Subset of
In this section we nd a necessary and sucient condition for the validity of
the inequality
(6.2.1)
uLq () Cu1
Lp () uLr ()
for all functions that vanish on some subset of .
Let G be an open bounded subset of . For p > 1 we put
(p,)
AG
= sup
{F }
[mn (F )]
,
[cp (G\F )]1/p
where {F } is the collection of closed (in ) subsets of G with cp (G\F ) > 0.

(1,)
Let AG
()
denote the value AG introduced in Sect. 5.2.3, i.e.,

(1,)
AG
= sup
{G }
[mn (G )]
,
s(i G )
where {G } is the collection of admissible subsets of G.

The following assertion is a generalization of Lemma 5.2.3/1 (the case
q p = 1).
330
Lemma. Let p 1 and let G be an open bounded subset of .

(p,)
1. If AG < and the numbers q, , p satisfy either one of the following
conditions:
(i) q q = 1
(ii) q < q = 1
for p 1,
for p > 1,
then for all functions u C 0,1 () which vanish on the exterior of G, the
inequality (6.2.1) holds with r (0, q), = r(q q)/q(q r), and C
(p,)
c(AG )1 .
2. Let q > 0, r (0, q ) and for some q (0, q ] and for any function
u C 0,1 () which vanishes on the exterior of G, let the inequality (6.2.1) be
valid with = r(q q)/q(q r) and with a constant C that is independent
of u.
(p,)
Then C c(AG )1 .
Proof. 1. Duplicating the proof of the rst part of Theorem 2.3.6 (for
= mn ), we arrive at

uLq () c
mn (Nt )
(1)/p
t
p1
dt
u
Lr () .
(6.2.2)
For p > 1 this implies

uLq ()
(p,) 1
c AG

p1
cp (Kt )t
0
(1)/p
dt
u
Lr () ,
where Kt is the conductor G\Nt . Now the result follows from Lemma 6.1.3/3.
In the case p = 1 inequality (6.2.1) results from (6.2.2) along with
Lemma 5.1.2/2 and the formula

|u| dx =
s(Et ) dt
(6.2.3)
0
(cf. Theorem 1.2.4).

2. Let p > 1. We x a small positive number > 0 and put
= sup
[mn (F )]p
cp (G\F )
on the set of all F G with cp (G\F ) . (The substitution of an arbitrary

function in T (G\F ) into (6.2.1) implies

1
p/q
p/r
mn (G)
C p mn (F )
,
cp (G\F )
which means that the collection of sets F , contained in G and satisfying

the inequality cp (G\F ) , is not empty.) Obviously, 1 mn (G)p .
6.3 Classes of Sets Ip,
331
Further, we must duplicate the proof of the second part of Theorem 2.3.6
with Lemma 2.2.2/1 replaced by Lemma 6.1.3/1 and with (p, )-cap(F, )
replaced by cp (G\F ). Then we arrive at cC p(1) , which, since is
(p,)
arbitrary, implies C c(AG )1 for p > 1.
Consider the case p = 1. Let G be any admissible subset of G. We insert
the sequence of functions {wm }m1 specied in Lemma 5.2.2 into (6.2.1).
Then
1
1/q
/r

C 1 mn (e)
,
s(i G )
mn (G )
which, since e is arbitrary, yields

s(i G ) C 1 mn (G ) .

1
Corollary. If p1 > p 1 and 1 p1
, then AG 1
1 = p
(p ,1 )
(p,)
cAG
1
p1 in place of u
Proof. We put |u|q1 /q with q1 > q, q11 p1
1 = q
in (6.2.1). Then

q /q
()
1
uL1q
q1
q
(p,) 1

AG
1
q /q
|u|(q1 q)/q uL () uLr 1 () , (6.2.4)
p
where r1 = rq1 /q. Applying the Holder inequality, we obtain

|u|p(q1 q)/q |u|p dx

|u|
(q1 q)pp1 /(p1 p)q
(p1 p)/p1
p/p1
|u|
p1
dx
(p p)q /p
uLq1 () 1 1 upLp () .
1
1
dx
The preceding inequality and (6.2.4) imply

(p,) 11
1
1
uLq1 () c AG
u1
Lp () uLr
1
() ,
where 1 = r1 (q1 q1 )/q1 (q1 r1 ), q1 = 11 . Using the second part of the

(p , )
(p,)
above Lemma we obtain that AG 1 1 cAG . The corollary is proved.

6.3.1 Denition and Simple Properties of Ip,
Denition 1. A domain belongs to the class Ip, (p 1, > 0) if there
exists a constant M (0, mn ()) such that
332

def
[mn (F )]
< ,
1/p
{K} [cp (K)]
Ap, (M ) = sup
(6.3.1)
where {K} is the collection of conductors K = G\F in with positive pconductivity and such that mn (G) M .
Proposition 6.1.2/2 implies that mn (F ) = 0 for any conductor K = G\F
with zero p-conductivity provided Ip, .
Remark. The class Ip, is empty for < (n p)/np, n > p since in this
r ) = const rnp and therefore
case cp (B2r \B
[mn (Br )]
n((np)/np)
as r 0.
r )]1/p = const r
[cp (B2r \B
Proposition 1. If a domain is the union of a nite number of domains
in Ip, , then is in the same class.
Proof. Let = 1 2 , i Ip, , i = 1, 2. Then there exist constants
M1 , M2 such that

1/p
mn (Fi ) A(i)
p, (Mi ) cp (Ki )
for any conductor Ki = Gi \Fi in i with mn (Gi ) Mi .
Put M = min{M1 , M2 } and let K denote a conductor G\F with
mn (G) M . Further let Gi = G i , Ki = Gi \Fi . If cp (K1 ) = 0, then
mn (F1 ) = 0 and hence mn (F ) = mn (F2 ). Therefore,

1/p

1/p
mn (F ) A(2)
A(2)
.
p, (M2 ) cp (K2 )
p, (M2 ) cp (K)
In the case cp (Ki ) > 0, i = 1, 2, we have

2

[mn (F2 )]
[mn (F1 )]
[mn (F )]
c
+
A(i)
c
p, (Mi ).
[cp (K)]1/p
[cp (K1 )]1/p
[cp (K2 )]1/p
i=1
Denition 2. We say that K is an admissible conductor if K =

G \ clos g, where G and g are admissible subsets of (cf. the denition
at the beginning of Sect. 5.1.1).
Proposition 2. We have
[mn (g)]
,
1/p
{K } [cp (K )]
Ap, (M ) = sup
(6.3.2)
where {K } is the collection of admissible conductors K = G \ clos g with

positive p-conductivity and with mn (G ) M . (Hence we may restrict ourselves to admissible conductors in the denition of the class Ip, .)
333
Proof. We prove the inequality

[mn (g)]
.
1/p
{K } [cp (K )]
Ap, (M ) sup
(6.3.3)
The reverse inequality is obvious. Let K be any conductor in Denition 1.

Given any > 0 we can nd an admissible conductor K = G \ clos g,
K K, such that cp (K) (1 )cp (K ) (cf. Proposition 6.1.2/2). It is
clear that mn (G ) M and
[mn (F )]
[mn (g)]
,
[cp (K)]1/p
(1 )1/p [cp (K )]1/p
which immediately implies (6.3.3).

6.3.2 Identity of the Classes I1, and J
Lemma. The classes I1, and J coincide and
A1, (M ) = sup
{G }
[mn (G )]
,
s(i G )
(6.3.4)
where {G } is the collection of admissible subsets of with mn (G ) M .

Proof. Let G be an admissible subset of with mn (G ) M and let
{wm }m1 be the sequence of functions specied in Lemma 5.2.2. The properties of {wm } imply s(i G ) c1 (G \e) for any compactum e contained in G .
If I1, , then [mn (e)] A1, (M ) s(i G ) and hence

mn (G ) = sup mn (e) A1, (M ) s(i G ).
eG
Suppose J . Let K be an arbitrary conductor G\F with mn (G) M .

By (6.2.3) for any f T (K) we obtain

s(i G )
mn (F ) . (6.3.5)
|f | dx inf s(i G ) : G G F inf
{G }
{G } [mn (G )]
By Lemma 6.1.1/2 we have

inf
|f | dx : f T (K) = c1 (K),
so (6.3.5) implies
A1, (M ) sup
{G }
The result follows.
[mn (G )]
.
s(i G )
334
6.3.3 Necessary and Sucient Condition for the Validity

1
()
of a Multiplicative Inequality for Functions in Wp,s
Lemmas 6.2 and 6.3.2 imply the following obvious assertion.
Corollary. 1. Let Ap, (M ) < for some M (0, mn ()).
Then for all u C 0,1 () with mn (supp u) M the inequality (6.2.1)
holds with r (0, q), = r(q q)/q(q r), q being the number specied in
Lemma 6.2 and C c[Ap, (M )]1 .
2. If (6.2.1) holds for all u C 0,1 () with mn (supp u) M < mn (),
then C c[Ap, (M )]1 .
The introduction of the classes Ip, is justied by the following theorem.
Theorem. 1. Let Ip, and let q be a positive number satisfying either
one of the conditions: (i) either q q = 1 for p q , or (ii) q < q for
p > q .
1
() we have
Then for any u Wp,s
1

uLq () C1 uLp () + C2 uLs ()
u
Lr () ,
(6.3.6)
where s < q , r < q, = r(q q)/q(q r), C2 = cM (sq )/sq , and

C1 cAp, (M ).
1
2. Let q > 0 and for some q (0, q ] and all u Wp,s
(), let (6.3.6) hold
with 0 < s < q , 0 < r < q , = r(q q)/q(q r).

Then Ip, with = 1/q . Moreover, if the constant M in the desq /(sq )
, where c is a small
nition of the class Ip, is specied by M = c C2
enough positive constant depending only on p, q , s, then C1 c Ap, (M ).
Proof. 1. By Lemma 6.1.1/1, it suces to obtain (6.3.6) for functions u
L1p () L () C () with bounded supports. Let T = inf{t : mn (Nt ) <
M }. Clearly, mn (Lt ) M mn (Nt ). Further we note that

q
|u| T dx. (6.3.7)
|u|q dx
|u|q dx + cT q mn (NT ) + c
NT
\NT
To get a bound for the rst summand on the right we rewrite it as follows:

|u|q dx =
|u|q (qr)/(q r) |u|r(q q)/(q r) dx,
\NT
\NT
and use the Holder inequality

|u| dx
|u|
\NT
\NT
(qr)/(q r)
dx
(q q)/(q r)
|u| dx
.
r
\NT
Since u < T on \NT , the right-hand side of the last estimate does not exceed
(q s)(qr)/(q r)
335
(qr)/(q r)
|u| dx

\NT

(q q)/(q r)
r
|u| dx
.
\NT
This implies

|u| dx M
q
\NT
(sq )(qr)/s(q r)
\NT

(q q)/(q r)
|u| dx

NT
(q s)(qr)/s(q r)
s
|u| dx

(qr)/(q r)
|u|s dx
\NT
M (sq
)(qr)/s(q r)
q (qr)/s(q r)
|u|s dx

(q q)/(q r)
r
|u| dx
.
(6.3.8)
Next we estimate the second summand on the right in (6.3.7). Since |u(x)| T
on NT and mn (NT ) M , we obtain
T mn (NT ) M
q
(sq )(qr)/s(q r)
Nt
q (qr)/s(q r)
|u| dx
s
NT

(q q)/(q r)
r
|u| dx
.
Combining this inequality with (6.3.8) and (6.3.7) we arrive at

+
1
u
uLq () c |u| T Lq () + cM (1)(sq )/sq uL
Lr () .
s ()
It remains to apply the rst part of the Corollary to (|u| T )+ .
2. Let M be an arbitrary constant satisfying
2c0 M 1/s1/q C2 < 1,

where c0 is a constant that depends only on s, p, q, q , r to be specied at
the end of the proof.
Let denote a small enough positive number; by K we mean the conductor
G\F in with mn (G) M , cp (K) . Further we introduce the function
= sup[mn (G)]p/q /cp (K) where the supremum is taken over the above set
of conductors. It is clear that < and
336
1/p
Ap, (M ) = lim .
(6.3.9)
+0
Since s < q , for any u T (K), we have

uLs ()
[mn (Nt )]1/q

c sup
uLp ()
1/p
0<t<1 [cp (G\Nt )]
r(q s)/s(q t)
uLt ()
q (st)/s(q t)
where t < s (cf. the proof of the rst part of Lemma 6.2).
Now we note that [mn (Nt )]p/q cp (G\Nt ) since cp (G\Nt )

cp (K) . Besides,
uLt () M 1/t1/s uLs () .
Consequently,
1/p
uLs () c M 1/s1/q uLp () .

Using the preceding estimate, from (6.3.6) we obtain

1/q
mn (F )

1
1/p
C1 + c M 1/s1/q C2
u1
Lp () uLr () .
Further, we must duplicate the proof of the second part of Theorem 2.3.6
with Lemma 2.2.2/3 replaced by Lemma 6.1.3/1 and with (p, )-cap(F, G)
replaced by cp (K). As a result we obtain

1/p
1/p
c C1 + M 1/s1/q C2 .
This inequality and the denition of the constant M imply
(M ) (2c0 C1 )p .
It remains to make use of (6.3.9). This completes the proof.

1
6.3.4 Criterion for the Embedding Wp,s
() Lq (), p q
An important particular case of Theorem 6.3.3 is the following criterion for

1
the embedding of Wp,s
() into Lq ().
1
()
Corollary. 1. If Ip, , q p and s < q , then for any u Wp,s
uLq () C1 uLp () + c2 uLs () ,

where C1 cAp,1/q (M ) and C2 = cM (sq
)/sq
(6.3.10)
2. Let for all u

the inequality (6.3.10) hold, where 0 < s < q .
Then Ip,1/q and if the constant M in the denition of the class Ip,1/q
1
Wp,s
()
q /(sq )
is dened by M cC2
cAp,1/q (M ).
with suciently small c = c(q , s), then C1
337
6.3.5 Function M,p and the Relationship of the Classes Ip,

and J
Denition. Let M,p (t) be the inmum of cp (K) taken over the collection of
all conductors K = G\F with mn (F ) t, mn (G) M .
By an argument similar to that in the proof of Proposition 6.3.1/2 we
can prove that we may restrict ourselves to admissible conductors in this
denition.
Inequality (6.1.10) immediately implies the following assertion on the connection of M,p with the isoperimetric function M introduced in Sect. 5.2.4.
Proposition 1. The inequality
M
1p

p/(1p)
M ()
M,p (t)
d
(6.3.11)
holds.
Obviously, for p 1
1/p

.
Ap, (M ) = sup t M,p (t)
(6.3.12)
0<tM
Hence Ip, if and only if

lim inf tp M,p (t) > 0.
(6.3.13)
t+0
We have noted already in Sect. 5.2.4 that J = I1, if and only if

lim inf t M (t) > 0.
t+0
Proposition 2. If J+(p1)/p , then Ip, and

Ap, (M )
p1
p
(p1)/p
A1,+(p1)/p (M ).
(6.3.14)
Proof. From (6.3.11) we obtain

p
M,p (t) A1,+(p1)/p (M )

>
The result follows.
p
p1
p1

(p+p1)/(p1) d
1p
A1,+(p1)/p (M )
tp .
Remark. By Corollary 6.2, the class Ip, is a part of the class Ip1 ,1 and
1
.
Ap1 ,1 (M ) c Ap, (M ) provided p1 > p 1 and 1 p1
1 =p
338
6.3.6 Estimates for the Conductivity Minimizing Function M,p

for Certain Domains
We consider some domains for which explicit two-sided estimates for M,p are
valid.
Example 1. Let be the domain {x : |x | < f (xn ), 0 < xn < a} considered
at the beginning of Sect. 5.3.3.
We show that
(M )
1p

(1n)/(p1)
f ( )
d
kp
()

M,p ()
(M )
(1n)/(p1)
f ( )
1p
d
(6.3.15)
()
where k is the constant in the inequality (6.3.5) that depends on M and the
function is specied by
()
= vn1
n1
f ( )
d.
Consider the conductor K,M = G(M ) \ clos G() , where G = {x :

0 < xn < }. Let the function u be dened by u(x) = 0 outside G(M ) ,
u(x) = 1 on G() , and
(M )
u(x) =
xn

d
[f ( )](n1)/(p1)
(M )
()
d
[f ( )](n1)/(p1)
on G(M ) \G() . Clearly, u is contained in U (K,M ). Hence, inserting u into

the integral

|u|p dx,
we obtain

cp (K,M )
(M )
(1n)/(p1)
f ( )
1p
d
(6.3.16)
()
Taking into account the denition of M,p (), we arrive at the right inequality (6.3.15).
Substituting the left inequality (5.3.5) into (6.3.11), we obtain the required
lower bound for M,p .
From (6.3.15) we obtain that Ip, if and only if

lim sup
x+0
x
f ( )
n1
p/(p1)
d
a
(1n)/(p1)
f ( )
d < .

1/2
< axn , 0 < xn < 1
() = x : x21 + + x2n1
339
(6.3.17)
with > (p 1)/(n 1), for small t we have

ctp M,p (t) c tp ,
where = [(n 1) + 1 p]/p[(n 1) + 1]. Thus the domain (6.3.17) is
contained in Ip, for p < 1.
Example 2. Let be the domain in Example 5.3.3/2. Using the estimates (5.3.8) for M and following the same discussion as in Example 1 we
obtain
()
1p

(1n)/(p1)
kp
f ( )
d
(M )

M,p ()
()
(1n)/(p1)
f ( )
1p
d
(6.3.18)
(M )
Consequently, Ip, if and only if
p/(p1) x

n1

(1n)/(p1)
d
d < .
f ( )
f ( )
lim sup
x+

x : x21 + + x2n1 < (1 + xn )2 , 0 < xn <
with (n 1) > 1 we have
ctp M,p (t) c tp
for small t,
provided p = [(n 1) 1 + p][(n 1) 1]1 . Here p > 1.

For the domain

x : x21 + + x2n1 < ecxn , 0 < xn < , c = const > 0,
by (6.3.18) we obtain
c M,p () c ,
and hence this domain is in Ip,1/p .

Example 3. Let be the spiral domain considered in Example 5.3.3/3.
Proposition 6.3.5/1 along with the lower bound (5.3.11) for yields the following lower bound for M,p (t):
(t)
1p

1/(p1)
(1)
M,p (t) CM
()
d
,
(6.3.19)
(M )
where (t) is specied by (5.3.12) and = 2 1 .
340
To derive a similar upper bound consider the conductor G(t) \ clos G(M ) ,
where G = {ei : 0 < < } and the function u dened by
u = 0 outside G(t) ,
u() =

1/(1p)
d
()
u = 1 on G(M ) ,
(t)
()
(M )
1/(1p)
1
(M ) .
on G(t) \G
(M )
It is clear that
cp G(t) \ clos G (M )
1 p
u () d d
G(t) \G(M )

(2)
CM
(t)
()
1/(1p)
1p
d
.
(M )
1p
(t)

1/(1p)
(1)
d
()
CM
(M )

M,p (t)
(2)
CM
(t)
()
1/(1p)
1p
d
(6.3.20)
(M )

i
e : 1(8+)1 > > 1(8+)1 c(8+) , 0 < < (6.3.21)
with 0 < c < 2( 1), > 1, we have ctp M,p (t) c tp for small t,
where = ( 1 + p)/p( 1). Thus (6.3.21) is a bounded domain in the
class Ip, for p > 1.
Example 4. We show that the adjoining nonintersecting cylinders

Gj = x : |xn j | < aj , x21 + + x2n1 < b2j

with j aj bn1
< (cf. Fig. 23) form a domain which is not in Ip,1/p if
j
lim supj aj = . Let

Fj = x Gj : |xn j | aj /2
and let C0 ([0, 1]), (t) = 1 for 0 t <
(|xn aj |/aj ) into the norm uLp (Gj ) . Then
1
2.
cp (Gj \Fj ) cap

j mn (Fj ).
Thus
lim inf 1 M,p () = 0.

0
We insert the function
1
6.4 Embedding Wp,s
() Lq () for q < p
341
Fig. 23.
1
6.4 Embedding Wp,s
() Lq () for q < p
6.4.1 Estimate for the Norm in Lq () with q < p for Functions

which Vanish on a Subset of
If Ip, with p 1, we have by Theorem 6.3.3 that the embedding
1
operator of Wp,s
() into Lq () with q = 1 is bounded. The following
example shows that inequality (6.3.6) with the limit exponent q = q = 1
may fail provided p > 1 (or equivalently, provided p > q ).
Example. Consider the plane domain

= (x1 , x2 ) : |x1 | < (1 + x2 ) , 0 < x2 <
with > 1. In Example 6.3.6/2 it was established that this domain is in the
class Ip, 1/p+1/(1) . We show that inequality (6.3.6) is not valid for
q = q = p( 1)/(p + 1).
The sequence of functions um (x1 , x2 ) = (1+x2 )m with m < 1+( 1)/p,
m = 1, 2, . . . , satises
342

uqLq () = 2 1 + ( 1)/p / 1 p(m 1) ,

um pLp () = 2m / 1 p(m 1) ,
um sLs () = 2/( 1 m s).
Since s < q < p, the left-hand side in (6.3.6), written for um , tends to innity
as m more rapidly than the right-hand side. Thus the inclusion Ip,
with p > q = 1 is necessary, but not sucient for the validity of (6.3.6)
with the limit exponent q = q . A necessary and sucient condition will be
given in Theorem 6.4.2.
(p)
Let G be a bounded open subset of and let G (t) denote the inmum of
cp (K) taken over the collection of all conductors K = G\F with mn (F ) t.
We put
1/(p1) 1/p
mn (G)
(p,)
=
d
,
(6.4.1)
BG
(p)
0
G ( )
where p > 1, p 1.
(p,)
Lemma. 1. Let BG
< . Then the inequality
uLq () CuLp ()
(6.4.2)
holds for all u C 0,1 () which vanish outside G, with q = 1 , p > 1 and
(p,)
C cBG .
2. If (6.4.2) is valid for all u C 0,1 () which vanish outside G, with some
(p,)
q [1, p), then C cBG .

Proof diers only by obvious details from that of Theorem 2.3.8.
1
() Lq () for
6.4.2 Class Hp, and the Embedding Wp,s
0<q 1. We put
1/(p1)
1/p
M
Bp, (M ) =
d
,
(6.4.3)
M,p ( )
0
where M is a constant in (0, mn ()).
Denition. The set belongs to the class Hp, provided Bp, (M ) <
for some M (0, mn ()).
The next corollary immediately follows from Lemma 6.4.1.
Corollary 1. 1. Let p > 1 and Bp, (M ) < . Then for all u C 0,1 ()
such that mn (supp u) M we have (6.4.2) with q = 1 and C cBp, (M ).
1
6.4 Embedding Wp,s
() Lq () for q < p
343
2. If (6.4.2) holds for all u C 0,1 () with mn (supp u) M then C

cBp, (M ).
Duplicating with obvious simplications the proof of Theorem 6.3.3, we
obtain the following theorem from Corollary 1.
Theorem. 1. Let p > 1 and Hp, . Then
uLq () C1 uLp () + C2 uLs ()
(6.4.4)
1
() with q = 1 , s < q ,
for all u Wp,s
C2 = cM (sq
)/sq
C1 cBp, (M ).
1
2. Let (6.4.4) hold for all u Wp,s
() with 1 q < p, s < q . Then
Hp, . Moreover, if M in the denition of Hp, is specied by M =
sq /(sq )
cC2
, where c is a small enough positive constant that depends only on
p, q , s, then C1 cBp, (M ).
Corollary 2. If p1 > p 1 and 1 p1

= p1 , then BG 1 1
1
for any G. (Consequently, Bp1 ,1 (M ) cBp, (M ) and Hp,
Hp1 ,1 .)
(p , )
(p,)
cBG
This assertion can be proved in the same way as Corollary 6.2.

6.4.3 Embedding L1p () Lq () for a Domain with Finite
Volume
We present a necessary and sucient condition for the validity of the Poincaretype inequality
(6.4.5)
inf1 u cLq () CuLp () ,
cR
provided is a domain with mn () < .

By Lemma 5.2.3/2 the embedding L1p () Lq () (p 1, q 1) and
inequality (6.4.5) are equivalent. The case mn () = will be considered
in 6.7.5.
Theorem 1. Let be a domain with nite volume.
1. If Ap, ( 12 mn ()) < for p 1 or Bp, ( 12 mn ()) < for p > 1,
then (6.4.5) holds for all u L1p () with q = 1 , C cAp, ( 12 mn ()) for
p 1 and C cBp, ( 12 mn ()) for p > 1.
2. If there exists a constant C such that (6.4.5) holds for all u L1p () with
q 1, then C cAp, ( 12 mn ()) for 1 = q p and C cBp, ( 12 mn ())

for 1 = q < p.
Proof. The proof of suciency follows the same argument as in the
proof of Theorem 5.2.3 except that we must apply Corollary 6.3.3 instead
of Lemma 5.2.3/1 for p 1 and Corollary 6.4.2/1 for p < 1.
344
Necessity. Let u be an arbitrary function in C 0,1 () L1p () with

mn (supp u)
1
mn ().
2
(6.4.6)
There exists a number c0 such that

u c0 Lq () = inf1 u cLq () .
cR
(6.4.7)
The latter and (6.4.5) imply

|u c0 |q dx + |c0 |q mn (\ supp u) C q uqLp () .

supp u
Therefore
Since
1
mn ()|c0 |q C q uqLp () .
2

1/q
uLq () |c0 | mn ()
+ u c0 Lq () ,
making use of (6.4.5) and (6.4.7), we obtain

uLq () cCuLp ()
for all u C 0,1 () satisfying (6.4.6). A reference to Corollaries 6.3.3
and 6.4.2/1 completes the proof.

Theorem 2. Let be a domain with nite volume. The space L1p () is
embedded into Lq () if and only if Ip,1/q for p q and Hp,1/q
for p > q > 0.
Proof. The necessity follows from Lemma 5.2.3/2 and Theorem 1. To
prove the suciency we must show that (6.4.5) is valid provided Ap, (M )
and Bp, (M ) are nite for some M (0, 12 mn ()). Let u L1p () C 0,1 ()
and T = inf{t : mn (Nt ) M }. We note that

q
q
q
|u| T + dx + T M .
|u| dx c
Using Corollaries 6.3.3 and 6.4.2/1, we obtain

uLq () CuLp () + cT M 1/q ,
where C is a constant independent of u. Let denote a bounded subdomain
of with smooth boundary such that mn (\) < M/2. Obviously,

1
|u|p dx
|u|p dx T p M.
2
NT
Therefore
1
6.4 Embedding Wp,s
() Lq () for q < p
uLq () CuLp () + cM 1/q
1/p
uLp () .
345
(6.4.8)
Since satises
inf u cLp () KuLp () ,
cR1
then (6.4.8) implies (6.4.5) for .

6.4.4 Sucient Condition for Belonging to Hp,
Here we give the following sucient condition for a set to belong to Hp, :
M

0
M ( )
p/(p1)
d < .
(6.4.9)
Proposition. If p > 1, 1, then

M
Bp, (M ) (p 1)(p1)/p
0
M ( )
p/(p1)
1/p
d
(6.4.10)
Proof. By (6.4.3) and the inequality (6.3.11) we have

p/(p1)
Bp, (M )

M
1/(p1)
0
p/(1p)
M ()
(p1)/(p1)
d
d.
To nd a bound for the right-hand side we apply (1.3.1) in the form
q

M
q M
M

q
q
r
f () d d
qr f ( ) d,
(6.4.11)
1r
0
0
where f ( ) 0, r < 1, q > 1. Putting r = (1 p)1 , q = (p 1)/(p 1)
in (6.4.11), we arrive at (6.4.10).

6.4.5 Necessary Conditions for Belonging to the Classes Ip, and
Hp,
Using denitions of Ip, and Hp, we obtain some necessary conditions for
to be contained in these classes.
and let s(t) be the area
Proposition 1. Let O be an arbitrary point in
of the intersection of with the sphere Bt centered at O.
If Ip, then

p/(p1)
s(t) dt
0
for suciently small and for r < .
dt
const
[s(t)]1/(p1)
346
Proof. If Ip, , then obviously

mn (r )

1/p
const cp ( \ clos r )
,
(6.4.12)
for small enough and r < . Let u = 1 in r , u = 0 outside and

u(x) =
|x|
dt
[s(t)]1/(p1)

r
dt
[s(t)]1/(p1)
1
x \r .
Inserting u into the denition of p-conductivity, we obtain
1p

dt
cp ( \ clos r )
,
1/(p1)
r [s(t)]
(6.4.13)

The proof of the following assertion is similar.
Proposition 2. If mn () < and Ip, , then

p/(p1)
s(t) dt
r
dt
const
[s(t)]1/(p1)
(6.4.14)
for large enough and for r > .

Corollary. If is an unbounded domain with mn () < and Ip, ,
then p 1.
Proof. By the Holder inequality for r > we have
r
r=

s(t)
1/p
dt
[s(t)]1/p

1/p
s(t) dt

dt
[s(t)]1/(p1)
Therefore, (6.4.14) implies

r
s(t) dt)1/p

r const
.
( r s(t) dt)
(
Let the sequence {j } be specied by

s(t) dt = 2j .
j
Then
j j1 const
If p < 1, then the series
j
j1

( j
s(t) dt)1/p
s(t) dt)

j
= const 21/p .
(p1)/p
.
1
6.4 Embedding Wp,s
() Lq () for q < p
347
(j j1 )
converges, which contradicts the assumption that is unbounded. The result

follows.

Since the condition Ip,1/q is necessary for the continuity of the embedding operator L1p () Lq (), it follows by the Corollary that the embedding
L1p () Lq () for an unbounded domain with nite volume implies p q.
Next we present a necessary condition for to belong to Hp, stated in
terms of the function s introduced in Proposition 1.
Proposition 3. If mn () < and Hp, , p > 1, then

c
dt
[s(t)]1/(p1)
p1
p1
p1
s(t) dt
s() d < ,
where c is a positive constant.

The result follows from Denitions 6.3.5 and 6.4.2 combined with (6.4.13).
6.4.6 Examples of Domains in Hp,
Example 1. By Denition 6.4.2 and the two-sided estimate (6.3.18) the domain
considered in Example 6.3.6/2 is contained in Hp, if and only if
x
ds
p1
n1
f (s) p1
f (s)
n1
1/(p1)

n1
ds
dx < .
f (x)
(6.4.15)
By Proposition 6.4.4 and the left estimate (6.3.18) we obtain the sucient
condition for Hp,

f (t)
n1
p/(p1)
dt

(n1)
f (x)
dx < ,
(6.4.16)
where
=
Assuming that for large x

( 1)p 1
.
p 1
n1

n1
ds cx f (x)
,
(6.4.17)
we see that (6.4.16) follows from the condition

n1
xp/p1 f (x)
dx < .
(6.4.18)
f (s)
348
On the other hand, the inequality

f (2x) cf (x)
for large x
(6.4.19)
and the monotonicity of f show that (6.4.15) implies (6.4.18). Hence, (6.4.18)
is equivalent to the inclusion Hp, under the additional requirements
(6.4.17) and (6.4.19).
Example 2. Consider the spiral in Examples 5.3.3/3 and 6.3.6/3. We
assume in addition that

(2) c(),
() d c()
for large . Then, using the same argument as in Example 1 we can show that
Hp, if and only if

()p/(p1) d < .
(6.4.20)
0
A more complicated example of a domain contained in Hp, will be given

in Sect. 6.5.
6.4.7 Other Descriptions of the Classes Ip, and Hp,
We show that the class of conductors K = G\F used in Denition 6.3.1/1 of
Ip, can be reduced.
Theorem 1. Let be bounded domain and let be a xed open set with
. The domain belongs to Ip, (p 1, p1 n1 ) if and only if

sup
{F }
[mn (F )]
< ,
[cp (G\F )]1/p
(6.4.21)
where G = \
and {F } is the collection of closed (in ) subsets of G.
Proof. Let S() denote the left-hand side in (6.4.21). The necessity
of (6.4.21) is obvious since

S() Ap, mn (\
) .
We prove the suciency. Let D be a domain with smooth boundary and
. Further, let be a smooth function, = 1 on
such that
D D
\D, = 0 on and 0 1.
For any u L1p () we have
mn

p p
p
x : |u| (x) t
d t S()
(1)
cp Kt d tp ,
1
6.4 Embedding Wp,s
() Lq () for q < p
349
(1)
where Kt is the conductor (\

)\{x : (|u|)(x) t}. The preceding
inequality and Lemma 6.1.3/3 imply

p p
mn x : |u| (x) t
d t
0

p
|u| p dx
c S()
+c
mn

x : 1 (x) u(x) t d tp .
(6.4.22)
and D has a smooth boundary, we have

Since supp(1 )|u| D

(2)

mn x : 1 (x) u(x) t
const cp Kt ,
(2)
where Kt is the conductor D\{x : (1 (x))|u(x)| t}. Hence from (6.4.22)

and Lemma 6.1.3/3 we obtain

p p

mn x : u(x) t
d t
0

p
p

|u| dx +
(1 )|u| dx
const

p
p
const
|u| dx +
|u| dx .
Let M =
1
2 mn (D)
and mn (supp u) M . Then

|u|p dx const
|u|p dx,
D
and thus

mn

p

x : u(x) t
d(tp ) const

|u|p dx.
(6.4.23)
Consider the conductor K = G \F in subject to the condition

mn (G ) M . We insert u T (K ) into (6.4.23). Then we arrive at

p
mn F
const cp K ,
which is equivalent to the inclusion Ip, .
Remark. The assertion we just proved remains valid for p = 1 provided

S() designates the supremum of [mn (g)] /s(i g) taken over all admissible
.
sets g contained in with clos g \
Following the same argument as in the proof of Theorem 1, we can establish
(p,)
that the inclusion Hp, is equivalent to the niteness of BG
where
G = \
(cf. Sect. 6.4.1).
350
We can give a dierent description of the class Ip, for p < 1 replacing
G by (x) = B (x), x . Namely, we introduce the function
[0, 1] ap, () = sup sup
x {F }
[mn (F )]
,
[cp ( (x)\F )]1/p
where {F } is the collection of closed (in ) subsets of (x).

Theorem 2. Let be a bounded domain and let p < 1. Then Ip,
if and only if ap, () < for some .
Proof. Necessity does not require proof. Let ap, () < . We construct a
by open balls with radius and take a partition of unity
nite covering of
{i } subordinate to it. By Lemma 6.2

u i Lq () cap, ()(ui )L () ,
p
where q = 1 . Summing over i, we conclude that Wp1 () Lq (). Consequently, by Theorem 6.3.3 (part 2)) Ip, .

6.4.8 Integral Inequalities for Domains with Power Cusps
Let us consider in more detail the domain

() = x = (x , xn ) : |x | < axn , 0 < xn < 1 ,
presented in Example 6.3.6/1. There we showed that () Ip, provided
n > p, > (p 1)/(n 1), and = [(n 1) + 1 p]/p[(n 1) + 1].
By Theorem 6.4.3/2, the latter is equivalent to the inclusion L1p ( () )
Lq ( () ) with q = p[(n 1) + 1]/[(n 1) + 1 p]. Although here the
exponent q is the best possible, it is natural to try to rene the stated result
by using spaces with weighted norms (cf. Remark 5.3.3).
Let

1/r

r r

u(x) xn dx
.
uLr (, () ) =
()
The coordinate transformation : x dened by i = xi , 1 i n 1,

n = xn , maps () onto (1) . Since (1) is a domain of the class C 0,1 , it
easily follows from Corollary 2.1.7/2 that

vLq (, (1) ) c vLp (, (1) ) + vL1 (()) ,
where n > p 1, p q pn/(n p), = 1 + n(p1 q 1 ) > n/q and
is nonempty domain,
() . Returning to the variable x, we obtain
uLq (+(1)/q, () )

c x uLp (+(1)/p, () ) + u/xn Lp ((1)(p1)/p, () )

+ uL1 () ,
1
6.4 Embedding Wp,s
() Lq () for q < p
351
where x = (/x1 , . . . , xn1 ). Putting = ( 1)(p 1)/p, we have

uLq (+(1)/q,() ) c x uLp (1, () )

+ u/xn Lp ( () ) + uL1 () . (6.4.24)
Choosing q to eliminate the weight on the left-hand side, we arrive at

uLq ( () ) c x uLp (1, () ) + u/xn Lp ( () ) + uL1 () ,
where, as before, q = p[(n 1) + 1]/[(n 1) + 1 p]. Since > 1, the
preceding result is better than the embedding L1p ( () ) Lq ( () ).
For > 1 we can take q to be the limit exponent pn/(n p) in Sobolevs
theorem. Then (6.4.24) becomes
uLpn/(np) ((1)(n1)/n, () )

c x uLp (1, () ) + u/xn Lp ( () ) + uL1 () ,
which, in particular, guarantees the inclusion L1p () Lpn/(np) (( 1)(n
1)/n, () ). We can readily check by the example of the function xn with
= 1 + [(n 1) + 1]/p ( is a small positive number) that the power
exponent of the weight is exact.
In conclusion we remark that one can obtain an integral representation
similar to (1.1.8) for the domain () .
By (1.1.8), for any v C 1 ( (1) ) L11 ( (1) ) we have

v() =
()v() d +
(1)
n

i=1
(1)
fi (, ) v
() d,
| |n1 i
where D( (1) ) and fi L ( (1) (1) ). Therefore the function u = v

has the integral representation

(y)u(y) dy
u(x) =
()
n1

i=1

+
()
()
(|x
(|x
y |2
Fi (x, y)yn1
u
(y) dy
+ (xn yn )2 )(n1)/2 yi
y |2
Fn (x, y)
u
(y) dy,
2
(n1)/2
yn
+ (xn yn ) )
(6.4.25)
where D( () ) and Fi L ( () () ). It is easily seen that we can

take u to be an arbitrary function in C 1 ( () ). Since by Theorem 1.1.6/1 the
() ) is dense in L1 ( () ) we conclude that (6.4.25) is valid for all
space C 1 (
1
1
()
u L1 ( ).
352

ym Example
In this section we consider the domain described in Example 1.1.4/1 with
m = (2m1 ) where is a nondecreasing function such that 2(t) < t. Here
we show in particular that the convergence of the integral
1
0
1/(p1)
t
(t)
dt
(6.5.1)
is necessary and sucient for to belong to Hp, , for p > 1.

Lemma. (The Inverse Minkowski Inequality) If gm are nonnegative measurable functions on [0, 1], then
1
1p 1
1p
dt
dt

.
(6.5.2)
1/(p1)
1/(p1)
0 [
0 [gm (t)]
m gm (t)]
m
Proof. Let be an absolutely continuousnondecreasing function on [0,1]
with (0) = 0 and (1) = 1. We put g(t) = m gm (t). Then
( )p g dt =

( )p gm dt.
By H
olders inequality

m
p
( ) gm dt
0
1
dt)p
0
1/(1p)
g
dt)p1
0 m
1

m
dt
1/(p1)
1p
.
gm
Finally, by Lemma 2.2.2/2,

inf
Hence (6.5.2) holds.
p
( ) g dt =
0
dt
g 1/(p1)
1p
.
Next we prove that Hp, provided (6.5.1) is nite. Consider an arbitrary nonnegative function u L1p () that is innitely dierentiable in
Am Bm for any m and that vanishes in the rectangle C.
(m)
= {(x, y) :
We x an arbitrary number m and note that each level set Et
u(x, y) = t} (Am Bm ) consists of a nite number of smooth homeomorphic
images of a circle and simple arcs with end points on (Am Bm )\{y = 2/3}
for almost all levels t (0, ) (cf. Corollary 1.2.2). Henceforth we shall always
consider only such levels.
If t satises
(m)
(6.5.3)
s Et
2m3 ,
then

ym Example
(m) 1

(m)
1 + 2m+1 2m1 2m1 < s Et
,
m2 Nt
3
353
(6.5.4)
(m)
where Nt
= {(x, y) : u t} (Am Bm ). Let Bm denote the collection
of all levels t for which (6.5.3) holds. We show that for t
/ Bm one of the
following three cases occurs:
(m)

s Et
2m1 ,
(6.5.5)
(m)
m1
m2 Nt
k 2
,
(6.5.6)
where k is a constant depending on integral (6.5.1);
(m)

s Et
2m1 ,
(m)

m 2 Lt
k 2m1 ,
(m)
where Lt
= (Am Bm )\Nt
(m)
(6.5.7)
(6.5.8)
(m)

s Et
< 2m1 ,
(m)
(m)
m2 Nt
ks Et
.
(m)
(6.5.9)
(6.5.10)
joining cd and ef since

We note that there are no components of Et
(m)
t
/ Bm (cf. Fig. 24). Besides, the set Et
(t > 0) is disjoint with the line
y = 23 because u = 0 in C.
(m)
(a) Suppose that (6.5.5) is valid. Let Et
denote the upper component of
(m)
(m)
Et
which joins the polygonal line abc with f e. The set Et
divides Am Bm
Fig. 24.
354
into components; the component containing de will be denoted by Dm . Since

(m)
is placed below the line y = 23 + 2m and hence
t
/ Bm it follows that Et

(m)

1
m2 Nt \Dm 22m1 + 2m1 ,
3

(m)

1
m2 Lt \Dm 22m1 + 2m1 .
3
Taking into account that increases and 2(t) < t, we obtain
2m1

0
t
(t)

dt 2
2m1
1
0
t dt
22(m+1)
22
(t)
(2m1 )
(6.5.11)
with = (p 1)1 . Consequently,

(m)

m2 Nt \Dm k1 2m1 ,
where
k1 =
1
+ 21
3
(m)

m2 Lt \Dm k1 2m1 ,
(6.5.12)
1
t\(t)
dt.
(m)
m denote one of
The set Et \Dm divides \Dm into components. Let D
(m)
them with the boundary containing Et
and the end points of the segment
(m)
(m)
de. Suppose Dm Lt . We estimate m2 (Nt

Dm ). First we note that
(m)
(m)
Nt
Dm is bounded by the components of Et
which (with the exception
(m)
of Et ) are either closed in Am Bm or join points of the polygonal lines

abcde or def . This implies

(m)

(m)
Dm 2s Et Dm .
s (Am Bm ) Nt
Using the isoperimetric inequality, we obtain
1/2

(m)
3 (m)
m2 Nt
Dm
s Et D m .
2
(6.5.13)
Since t
/ Bm , the latter and (6.5.11) yield
(m)

9 2(m+1)
9 22 m1
m2 Nt
2
2
Dm
4
4
1
0
t
(t)
dt,
which together with (6.5.12) leads to (6.5.6).

m
(b) Inequality (6.5.8) can be derived in the same way provided that D
(m)
Nt .
(m)
(c) Suppose that (6.5.9) holds. Then Et
does not contain components
which join abcd and ef . Following the same argument as in the derivation
of (6.5.13), we obtain

ym Example
355
(m) 1/2
3 (m)
m2 Nt
.
s Et
2
The last inequality and (6.5.9) imply (6.5.10).

Thus, one of the following cases is possible: either (6.5.10) is valid,
or (6.5.5) and (6.5.6), or (6.5.7) and (6.5.8). Let Bm denote the set of levels
t for which (6.5.10) is valid. Let Bm and B
m be the sets of levels satisfying (6.5.5), (6.5.6) or (6.5.7), (6.5.8), respectively.
Let m (t) be dened by
t
m (t) =
(m)
Et
We have
|u|
p1
m (t)
0
1/(1p)
ds
.
d (m)
d
m2 N
(m)
d
[s(E )]p/(p1)
(cf. Corollary 6.1.3/1). We express the integral on the right-hand side as the
sum

+
+
.
Bm
By (6.5.10),

Bm
1
0
B
m
B
m
d
p1
d (m)
m2 N
.
(m) p/(p1)
(m)
d
[m2 (N )]
[m2 (Nt )]1/(p1)

p/(1p)

1/(1p)

2m1
sup m2 N(m) k 2m1
.
B
m
B
m
We note that

d
d
m2 N(m) =
m2 L(m) ,
d
d
and use (6.5.7) and (6.5.8). Then

(m)
d
d
sup m2 Lt
m2 L(m)
(m) p/(p1)

dt

Bm
Bm
Bm
[s(E )]
m1 1/(1p)
.
k 2
Consequently,

1/(1p) (m) 1/(1p)
m (t) 2k 2m1
+ m2 Nt
.
Let l denote the smallest number for which

(l)

m2 Nt
(2c)1p 2l1 .
(6.5.14)
356
Then (6.5.14) implies

1/(1p)

l (t) 4k 2l1
.
(6.5.15)
Next, taking into account that

(m) 1 m
< 2 ,
m2 Nt
3

we obtain
3
m2
4

Nt
(m)

2l1 ,
m=l
which together with (6.5.15) yields

p1
3
(m)
4k/l (t)
m2
Nt
.
4
(6.5.16)
m=l
Since
(m)

< (2c)1p 2m1
m2 Nt
for m < l, by (6.5.14) we have

(m) 1/(1p)
m (t) 2 m2 Nt
,
for m < l. Consequently,
l1

l1

1p
(m)
2p1
m (t)
m2
Nt
,
m=1
and thus

m2
l1

m=1

Nt
(m)
p2
m=1
l1

1p
m (t)
,
(6.5.17)
m=1
because (t) < t/2. If

m2
l1

m=1
then by (6.5.16)

Nt
(m)
3
< m2
4

Nt
(m)
m=l

p1
3
m2 (Nt ) 4k/l (t)
.
8
Otherwise, if the reverse of (6.5.18) is valid, then by (6.5.17)

l1

1p

3
m2 (Nt ) 2p2
.
m (t)
8
m=1
(6.5.18)

ym Example
357
Thus, we always have

1p
3
m (t)
,
m2 (Nt ) k
8
m=1
(6.5.19)
where k is the maximum of (4k)p1 and 2p2 .

Let (t) be dened by
t
(t) =
E
Since by Lemma
|u|
p1
(t)
1p
1/(1p)
ds
d.

1p
m (t)
,
m1
(6.5.19) implies

1p
3
m2 (Nt ) k (t)
.
8
(6.5.20)
Let F be an arbitrary subset of G = \C closed in and let u be an

arbitrary function in V (K), K = G\F . By (6.5.20) we have

1p
3
m2 (F ) k (1)
,
8
which together with Lemma 6.1.3/1 yields

3
m2 (F ) k cp (K).
8
Consequently,
3
t
8
k G (t),
(p)
(6.5.21)
(p)
where G is the function introduced in Sect. 6.4.1.

Taking into account the convergence of the integral (6.5.1) as well as
Lemma 6.4.1, from (6.5.21) we obtain that there exists a constant Q such
that
(6.5.22)
uLq () QuLp ()
for all u C 0,1 () which vanish in C with q = 1 .
0,1
Now let u be an arbitrary function
in C () and let be continuous in ,
vanishing in C, equal to unity in m1 Am and linear in Bm (m = 1, 2, . . . ).
Then
(6.5.23)
uLq () uLq (\C) + uLq (\ m1 Am ) .
358

uLq (\C) Q1 uLp () + uLp ( m1 Bm ) .
The latter and (6.5.23) imply

uLq () Q2 uLp () + uLp (\m1 Am ) .
(6.5.24)
We estimate the second norm on the right in (6.5.24). Let (x, y) \m1 Am
and (x, z) C. Obviously,

u(x, y)p Q3 u(x, z)p + u(x, y)p d
y ,
where the integral is taken over the vertical segment contained in and
passing through the point (x, 0). Integrating in x, y and z, we obtain

p
p
p
|u|
dx
dy
Q
|u|
dx
dy
+
|u|
dx
dy
.
4

\
m1
Am
Therefore,
m1
Am

uLq () Q5 uLp () + uLp (C) .
Hence

inf1 u cLq () Q5 uLp () + inf1 u cLp (C) .
cR
cR
Using the Poincare inequality for the rectangle C, we obtain

inf u cLq () Q6 uLp () ,
cR1
which according to Theorem 6.4.3/2 is equivalent to Hp, with = q 1 .

Next we show that the convergence of the integral (6.5.1) is necessary for
to be contained in Hp, . Let
1
t/(t)
dt = ,
where = q/(p q), q = 1 . Then

+1
m
= ,
(m )
(6.5.25)
m1
where m = 2m1 . Consider a continuous function um in vanishing in C,

linear in Bm and equal to

(+1)/p
m /(m )
,
in Am , m 1. For vN =

1mN
um we have

ym Example

p/(+1)
vN
dx dy
359
N
N

1 p/(+1)
um
m =
+1
.
(m )
m
3 m=1
m=1
On the other hand,

|vN |p dx dy = 3p
N

N

upm (m ) = 3p
m=1

+1
.
(m )
m
m=1
If is contained in Hp, , then for all u C 0,1 () vanishing in C

upLp/(+1) () Q
|u|p dx dy,
(6.5.26)
where Q does not depend on u. From (6.5.26) we obtain

N

+1
m
[(
m )]
m=1
Hence
(+1)/
Q
N

+1
m
.
[(
m )]
m=1
N

+1
m
Q ,
[(
)]
m
m=1
which contradicts (6.5.25).
Remark 1. From (6.5.2) and Lemma 6.1.3/2 it follows that

3
m2 (Nt ) d tp Q
|u|p dx dy
8
0
(6.5.27)
for all u C (), u = 0 in C. Further let C 0,1 [0, 1] and (2t) const (t).
Then, following the same argument as in the proof of Proposition 5.4,
from (6.5.2) we obtain p (t) k(t). The reverse estimate follows by considering the sequence of conductors {Gm \Fm } where Gm is the interior of
Am Bm and Fm = clos Am . In fact, for a piecewise linear function um
which vanishes in C and is equal to 1 in Am we have

cp (Gm \Fm )
|um |p dx dy = 3p 2m1 k m2 (F ) .
Thus, for the Nikod

ym domain and for any p 1,
p (t) (t),
(6.5.28)
lim inf tp (t) > 0
(6.5.29)
and so Ip, if and only if

t+0
360
(since (t) t then p 1). Similarly, Hp, if and only if

1/(p1)
1
d < .
( )
0
Remark 2. The domain considered in the present section is interesting
in the following respect. While the conditions for the domains in the examples of Section 6.3.6 to belong to J+11/p and Ip, coincide (i.e., Proposition 6.3.5/2 is exact), the Nikod
ym domain is simultaneously contained in
Jp, and Ip, by virtue of (6.5.29). This means that if, for example, (t) = t ,
> 1, then the sucient conditions for the embedding L1p () Lq ()
(p > 1) being formulated in terms of the function , give an incorrect value
of the limit exponent q = p/(1 + p( 1)).
The actual maximal value of q , obtained here by the direct estimate of
p is equal to p/.
6.6 Some Generalizations

The spaces Ls (), Lr (), and Lq () can be replaced by the spaces Ls (, ),
Lr (, ), and Lq (, ) of functions that are integrable of order s, r, q, respectively, with respect to the measure in ; we just need to replace the
Lebesgue measure by in the corresponding necessary and sucient conditions. As an example, we pause for a moment to present a generalization of
Lemma 6.2.
Let G be an open bounded subset of . For p > 1 we put
(p,)
AG, = sup
{F }
[(F )]
,
[cp (G\F )]1/p
where {F } is the collection of subsets of G which are closed in . Further, let

(1,)
AG, = sup
{g}
[(g)]
,
s(i g)
where {g} is the collection of admissible subsets of G. Thus by denition,

(p,)
(p,)
AG,mn = AG .
Theorem. Let p 1 and let G be an open bounded subset of .
(p,)
1. If AG, < and the numbers q, , p are related by any of the following
conditions:
(i) q q = 1 for p 1, (ii) q < q = 1 for p > 1, then, for all
u C 0,1 () vanishing outside G, we have
1
u
uLq (,) CuL
Lr (,) ,
p ()
with r (0, q), = r(q q)/(q r)q, C c[AG, ]1 .

(p,)
(6.6.1)
361
2. Let q > 0, r (0, q ) and let (6.5.1) hold for some q (0, q ) and for
any u C 0,1 () vanishing outside G with = r(q q)/(q r)q and with
(p,)
a constant C independent of u. Then C c [AG, ]1 .
The proof does not dier from that of Lemma 6.2.

(p,)
Next we present a sucient condition for the niteness of AG, .

Proposition. We have
(p,) 1/ (1,) /
(p,)
,
AG, c AG
AG,
(6.6.2)
where = p(p 1 + p)1 .

Proof. Let K = G\F and let u be any function in V (K). We put
t
(t) =
E
0
(p,)
The denition of AG
1/(1p)
|u|p1 ds
d.
and (6.1.8) imply

1/(1p) (p,)
p/(p1)
(t) cp (G\Nt )
mn (Nt )
AG
,
(6.6.3)
(p,)
where t [0, 1). Thus, if AG

< then (t) is nite and hence absolutely
continuous on any segment [0, 1 ], > 0. Let t() be the inverse of (t).
Since the function (Nt() ) does not increase, by Lemma 1.3.5/3 we
have
(1)
p/(p1)
p/(p1)

c sup (p1)/p
d,
(Nt() )
/ (Nt() )
for t() [0, 1 ]. The right-hand side is equal to

c
sup
(p1)/p
(t)
t[0,1]
[(N )]p/(p1) d
.
uLp1 (E )
(The change of variable = (t) is possible since (t) is absolutely continuous

on [0, 1 ].) The latter, (6.6.3) and Lemma 6.1.3/2 imply
p/(p1)

/ (Nt() )
(p,) 1/
c AG
sup
1
t[0,1] mn (Nt )
p/(p1)
1
[(N )]
d
mn (N ) d.
s(E )
d
t
(6.6.4)
362
(N )
(1,)
AG, s(E )
for almost all (0, 1), from (6.6.4) we obtain

/
(t)
(Nt )
p/(p1)
(p,) 1/ (1,) p/(p1)

AG,
c AG
1
1
d
mn (N ) d
sup
d
t[0,1] mn (Nt ) t
(p,) 1/ (1,) p/(p1)
AG,
c AG
.
Passing to the limit as t 1 0 we arrive at

(F )
(1p)/p
(p,) 1/ (1,) /
AG,
(1)
c AG
.
Minimizing the right-hand side over V (K), we obtain

(F )
(p,) 1/ (1,) /

1/p
AG,
cp (G\F )
c AG
.
The result follows.
Now we give an example of the application of the preceding proposition in

a concrete situation.
Example. Let

= x : |x | < xn , 0 < xn < ,
where > 1, x = (x1 , . . . , xn1 ) and let G = {x : 0 < xn < 1}.

Here the role of the measure is played by the (n 1)-dimensional measure
(p,)
s on the set = {x : x1 = 0}. We show that AG,s is nite and
the value of is the best possible provided (n 1) + 1 > p 1 and =
((n 1) + 1 p)/((n 2) + 1).
First let p = 1. It was shown in Example 5.3.3/1 that the mapping
x = (x1 , . . . , xn1 , xn ),
of onto the cone = { : | | < n , 0 < n < } is subareal. Hence
s(i g) cs(i g)
(6.6.5)
for any admissible set g with clos g G. Since is a cone and is its
cross section by a hyperplane, by Theorem 1.4.5 we have
uL() c uL()
(6.6.6)
for all u L1p () that vanish outside G. Therefore,

c s(i g) s( g).
(6.6.7)
363
(The latter estimate results from the substitution of the sequence {wm } constructed in Lemma 5.2.2 into (6.6.6).)
It is clear that

x1
dx2 , . . . , dxn .
(6.6.8)
s( g) =
n
g
Since > 1, the inmum of the integral on the right in (6.6.8), taken over all
sets g with the xed measure s( g), is attained at {x : 0 < xn < a},
where a is the number dened by

1 (n2)+1
vn2 (n 2) + 1
a
= s( g).
Therefore,

(n1)/((n2)+1)
s( g) c s( g)
,
which together with (6.6.5) and (6.6.7) yields

s(i g) c s( g) ,
(1,)
where = (n 1)/((n 2) + 1). Thus, AG,s < . On the other hand,

since Ip, with = ((n 1) + 1 p)p/((n 1) + 1) (cf. (6.3.16)), it
(,)
(p,)
follows that AG
< . Now (6.6.2) implies AG,s < with

1
(n 1) + 1 p
(n 1) (n 1) + 1 p
.
p p1+
=
(n 2) + 1 p((n 1) + 1)
(n 1) + 1
(n 1) + 1 p
=
.
p((n 2) + 1)
This value of is the best possible, which may be veried using the sequence
Fm = {x : 0 < xn < m1 }, m = 1, 2, . . . . In fact,

cp (G\Fm ) c
1p
(1n)/(p1)
m1
cmp1(n1) ,
(cf. (6.3.16)) and s(Fm ) = c m1(n2) . Thus the estimate

s(Fm )

1/p
const cp (G\Fm )
implies ((n 1) + 1 p)/p((n 2) + 1).

In conclusion we consider briey some other generalizations of the previous
results.
Following Chap. 2 with a minor modication in the proofs, we can generalize the results of the present chapter and Chap. 5 to encompass functions
with the nite integral
364
p
(x, u) dx,
and even those satisfying the following more general condition:

(x, u, u) dx <
(cf. Remark 2.3.3/2).

Another possible generalization, which needs no essential changes in the
proofs, is the replacement of Lq () by a BirnbaumOrlicz space (cf. Theorem 2.3.3).
1
6.7 Inclusion Wp,r
() Lq () (r > q) for Domains with
Innite Volume
6.7.1 Classes J and Ip,

The classes J , Ip, introduced previously characterize local properties of
the domain. In the present section we are interested in the structure of domains
at innity.
Consider, for example, the unbounded plane domain

= (x, y) : 0 < x < , |y| < x1/2 .
We have

uL2 () C uL1 () + uLr () ,
(6.7.1)
where r is an arbitrary positive number that does not exceed 2.

This inequality can be proved in the following way. Let Qm,n be an arbitrary square of the integral coordinate grid. Each of the domains m,n =
Qm,n can be mapped onto Q0,0 by a quasi-isometric mapping so that
the Lipschitz constants of the mapping functions are uniformly bounded and
the Jacobian determinants are uniformly separated from zero. This and Theorem 1.4.5 imply the sequence of inequalities

u2 dx c u2L1 (m,n ) + u2Lr (m,n )
m,n
with constant c independent of m, n. Summing over m,n and using

1/
1/
ai
with 1, ai > 0 we arrive at (6.7.1).

i
ai
1
6.7 Inclusion Wp,r
() Lq () (r > q) for Domains with Innite Volume
365
The condition r 2 is essential for the validity of (6.7.1). In fact, for

u(x, y) = (x + 1) with 3r/2 < < 34 , 2 < r < 3, the right-hand side
in (6.7.1) is nite whereas u
/ L2 ().
In a sense the estimate (6.7.1) is unsatisfactory: The norm Lr () is not
weaker than that in L2 () (contrary to the case mn () < ). If we discuss
the rate of decrease of a function at innity then the niteness of the norm in
Lr () (r < 2) is a more restrictive condition than that of the norm in L2 ().
So we may pose the following question. Let mn () = . What is the
1
() for large r?
space Lq () containing Wp,r
To this end we introduce classes similar to J , Ip, .
Denition 1. The set is contained in the class J if there exists a

constant M > 0 such that
sup
{g}
[mn (g)]
< ,
s(i g)
(6.7.2)
where the supremum is taken over all admissible sets g with mn (g) M .
We note, for the time being without proof, that the domain inside the
parabola, mentioned at the beginning of this section, is in the class J1/3 and
that

(6.7.3)
uL3 () C uL1 () + uLr ()
for any r 3. The exponent 3 on the left-hand side of the above inequality
cannot be reduced.
Let F be a bounded subset of closed in and let R = BR . By
the p-capacity of F relative to we mean the limit of cp (R \F ) as R .
We denote it by p-cap (F ).
Denition 2. The domain is contained in the class Ip, , if there exists

a constant M > 0 such that
def
Ap, (M ) = sup
{F }
[mn (F )]
< .
[p-cap (F )]1/p
(6.7.4)
Here the supremum is taken over all F with mn (F ) M , p-cap (F ) > 0.
Similarly to Lemma 6.3.2, we can prove that the classes I1, and J
coincide and that
[mn (g)]
,
A1, (M ) = sup
{g} s(i g)
where {g} has the same meaning as in (6.7.2).
Proposition 1. The class Jp, is empty provided > 1/p 1/n.
Proof. Let Ip, . If is a large enough positive number such that

mn ( ) M and R > , then by (6.7.4)
366
mn ( )

1/p
K cp (R \ clos )
,
K = const.
Let u(x) = (|x|) where is a piecewise linear continuous function which

vanishes for t > R and is equal to unity for t < . Since u U (R \ clos ),
we have
cp (R \ clos ) (R )p mn (R \ ).
Consequently,

1/p

R K mn (R ) mn ( )
mn ( )
.
We dene a sequence of numbers {j }j1 by mn (j ) = 2j M . Then j+1
1
1
j K2j(p ) . Summing, we get j 1 + cK2j(p ) , which together
with the denition of the sequence {j } yields

p1
.
j 1 + cK mn (j )
Since j then < p1 . Further we have

p1
j 1 + cK vn nj
and hence p1 n1 . The result follows.
Let F be a bounded subset of closed in and let F () = inf s(i g)

with the inmum taken over all admissible subsets g of which contain F
and with mn (g) .
An immediate corollary of inequality (6.1.10) is the next proposition.

p-cap (F )
F ()
p/(p1)
1p
d
(6.7.5)
mn (F )
is valid.
Proposition 3. If J+(p1)/p , then Ip, and
Ap, (M )
p1
p
(p1)/p
A1,+(p1)/p (M ).
(6.7.6)
Proof. If mn (F ) M then by (6.7.5) we obtain

p
p-cap (F ) A1,+(p1)/p (M )

mn (F )
(+11/p)p/(p1) d
1p
,
1
6.7 Inclusion Wp,r
367
1
() Lq () (r > q)
The proof of the following theorem concerning the embedding of the space
1
() = L1p () Lr () into Lq () is carried out by an argument similar
Wp,r
to that used to prove Theorem 6.3.3. We give it for the readers convenience
since it diers in details.
1
Theorem. 1. If Ip, and r > q = 1 , then for any u Wp,r
()
uLq () C1 uLp () + C2 uLr () ,

where
C2 = M (rq)/rq ,
(6.7.7)
C1 p(p 1)(1p)/p Ap, (M ).
2. If (6.7.7) holds with r > q then Ip, with = q 1 . Moreover,
M (rq)/rq = 1 C2 , Ap, (M ) (1 )1 C1 , where is an arbitrary number

in (0, 1).
Proof. 1. By Lemma 5.1.2/2 it suces to prove (6.7.7) for functions in
C () with supports in R for some R < . We choose a number T such
that
mn (LT ) M mn (NT ).
It can be readily checked that

q
q
|u| dx =
|u| dx +
LT

mn (Nt \LT ) d tp .
(6.7.8)
The rst summand on the right is estimated by Holders inequality

LT
q/r
|u|q dx M 1q/r
|u|r dx
By (1.3.41), the second integral in (6.7.8) does not exceed

p/q p
d t
mn (Nt \LT )
T
q/p
.
Therefore,

uLq () M
1/q1/r
T
uLr () +

p/q p 1/p
d t
.
mn (Nt )
Since mn (Nt ) M for t (0, T ), it follows that

1/p

1/p
Ap, (M ) p-cap (Nt )

Ap, (M ) cp (R \Nt )
.
(6.7.9)
Applying (6.7.9) and Lemma 6.1.3/2, we obtain

mn (Nt )
1/q
368
uLq (M ) M
1/q1/r
1/q1/r

uLr () + Ap, (M )

p 1/p
cp (R \Nt ) d t
uLr () + Ap, (M )
p
uLp () .
(p 1)(p1)/p
2. We put M = (C2 1 )rq/(rq) and consider an arbitrary bounded F

closed in with mn (F ) M . Let u be any function in C 0,1 () that vanishes
outside some ball and is equal to unity on F . It is clear that
1

|u|r dx =
mn (Nt ) d tr .
0
Since mn (Nt ) does not increase and q < r, by (1.3.41) we have
1

q/r q r/q
mn (Nt )
|u|r dx
d t
.
0
Taking into account that mn (Nt ) mn (F ) M , we obtain

1
q r/q
1/r1/q
mn (Nt ) d t
= C21 uLq () .
uLr () M
0
Now, (6.7.7) and the preceding inequality imply that

uLq () C1 (1 )1 uLp () .
Having in mind that u = 1 on F and minimizing uLp () , we nally obtain

1/q
mn (F )
1/p

C1 (1 )1 p-cap (F )
.
The rst part of the Theorem and Proposition 6.7.1/3 imply the next
corollary.
Corollary. If J , 1, p 1, r > q, p(1 ) < 1, q = p/[1

1
().
p(1 )] then (6.7.7) is valid for any u Wp,r
6.7.3 Example of a Domain in the Class Ip,

Example. Consider the paraboloid

= x Rn : x21 + + x2n1 < ax2
n , 0 < xn < ,
(6.7.10)
where 1 > > 0 and a = const > 0 (Fig. 25).

First we show that
0 < lim sup
M
[mn (g)]
< ,
s(i g)
(6.7.11)
1
6.7 Inclusion Wp,r
369
Fig. 25.
where the supremum is taken over all admissible subsets g of the domain
with mn (g) M and = (n 1)/((n 1) + 1). Taking the symmetrization
of g with respect to the ray Oxn and repeating the proof of Lemma 5.2.1/1
we obtain that the ball B, orthogonal to , has the smallest area i g among
all sets g with the xed volume M . After a routine calculation we obtain
(1 )
[mn (B)]
=
1 .
M s(i B)
vn1
lim
So (6.7.11) follows.
Thus, J(n1)/((n1)+1) and by Corollary 6.7.2 and the second
part of Theorem 6.7.2 we have Ip, with p < (n 1) + 1 and =

p1 ((n 1) + 1)1 . We show that this value of is the largest possible.
In fact, let (X ) = {x : xn X }. It is clear that

(X /xn )(n1) p dx = cX (n1)+1p
p-cap (X )
\(X )
and hence
[mn ((X ))]
X
c X [(n1)+1]()
[p-cap ((X ))]1/p
for > p1 [(n 1) + 1]1 = .
Thus, for the domain (6.7.10) and for (n 1) > p 1 0, the inequality (6.7.7) holds with r > q = [(n 1) + 1]p/[(n 1) + 1 p]. This value of
q cannot be reduced. We note that it exceeds the limit exponent np/(n p)
in the Sobolev theorem for < 1.
370
(0)
6.7.4 Space L1p () and Its Embedding into Lq ()

(0)
Let L1p () denote the completion of the set of functions in C ()L1p () with
bounded supports with respect to the norm uLp () . (Here and elsewhere
in this section mn () = .)
(0)
According to Lemma 5.1.2/2, the space L1p () coincides with the comple1
tion of Wp,r
() (r is an arbitrary positive number) with respect to the norm
uLp () .
Removing the conditions mn (g) M , mn (F ) M in the denitions of
the classes J , Ip, (i.e., putting M = 0) we obtain the denitions of the
classes J (0), Ip, (0).
Remark 6.3.1 implies 1/p 1/n provided n p and Ip, . On the

other hand, according to Proposition 6.6, 1/p 1/n. Thus, only the class
Ip, 1/p1/n (0) is not empty. (For example, the domain (6.7.10) with = 1
and the space Rn are contained in this class.)
Taking the latter into account and mimicking the proof of Theorem 2.3.3
with minor modications, we arrive at the following theorem.
Theorem. The inequality
uLq () CuLp () ,
p1
(6.7.12)
(0)
is valid for all u L1p () if and only if n > p, q = pn/(n p) and
Ip,1/p1/n (0).
The best constant in (6.7.12) satises
Ap,1/p1/n (0) C p(p 1)(1p)/p Ap,1/p1/n (0).
Remark. We note that the inclusion Ip,1/p1/n (0), n > p, does not
imply the Poincare-type inequality
inf u cLpn/(np) () CuLp () ,
cR1
u L1p ().
(6.7.13)
In fact, consider the domain in Fig. 26, which is the union of the two cones
{x : |x | < xn + 1} and {x : |x | < 1 xn }, x = (x1 , . . . , xn1 ). Each of the
cones is in the class Ip,1/p1/n (0). Hence their union is in the same class
(cf. Proposition 6.3.1/1). However, the left-hand side in (6.7.13) is innite for
a smooth function that is odd in xn , vanishes for 0 < xn < 1 and is equal to
unity for xn > 2, and which, obviously, belongs to L1p ().
At the same time, by the Theorem, the inclusion Ip,1/p1/n (0) is

equivalent to the inequality
1
6.7 Inclusion Wp,r
371
Fig. 26.
uLpn/(np) () CuLp () ,
1
u Wp,pn/(np)
().
(6.7.14)
In particular, from this it follows that (6.7.13) implies Ip,1/p1/n (0).

The preceding Theorem shows as well that the norm in Lpn/(np) () cannot be replaced by the norm in Lq () with q = pn/(n p) in (6.7.13).
6.7.5 Poincar
e-Type Inequality for Domains with Innite Volume
The following assertion gives a description of domains for which (6.7.13) is
valid.
Here (and only here) we shall assume that the condition of boundedness
of the sets G and F is omitted in the denition of p-conductivity.
Theorem. Let mn () = . The inequality (6.7.13) holds for all u
L1p (), p 1, if and only if is a connected open set in Ip,1/p1/n (0) and
the niteness of the p-conductivity of the conductor
K = G\F in implies either mn (F ) < or mn (\G) < .
(6.7.15)
For p = 1 the condition (6.7.12) is equivalent to the following:
if G is an open subset of such that i G is a smooth
manifold and s(i G) < then either mn (G) < or mn (\G) < .
(6.7.16)
(We recall that Ip,11/n (0) = J11/n (0).)
Proof. The necessity of the inclusion Ip,1/p1/n (0) was noted at the
end of Remark 6.7.4. The necessity of the connectedness of is obvious.
372
We prove that (6.7.13) implies (6.7.15). Let v be an arbitrary function

in U (K), where K is the conductor G\F with cp (K) < . We put u =
max{0, min{v, 1}} in (6.7.13). Then
1/q
|c|q mn (\G) + |1 c|q mn (F )
CuLp ()
with q = pn/(n p). Therefore, either mn (\G) < or mn (F ) < .

Now let p = 1. We prove the necessity of (6.7.16). First we note that the
condition that G is bounded was not used in the proof of Lemma 5.2.2. Let G
be an open subset of such that i G is smooth manifold and s(i G) < .
We insert any function wm from the sequence constructed in Lemma 5.2.2
in (6.7.13) in place of u. Then, starting with some index m, for any compactum
e G, we have
|cm |n/(n1) mn (\G) + |1 cm |n/(n1) mn (e)
(n1)/n
Cwm L1 () ,
where cm = const. Since

lim sup wm L1 () = s(i G) < ,
m
we have either mn (\G) < , or cm = 0 and

(n1)/n
mn (e)
Cs(i G).
Consequently, (6.7.16) holds.

The suciency is proved in several steps.
Lemma 1. 1. Let Ip,q (0) with q = pn/(n p), n > p. Then

1p/n
mn (F )
const cp (K),
(6.7.17)
holds for all conductors K = G\F , where G is an open subset of and F is

a subset of with nite volume, closed in .
2. Let J11/n (0). Then

11/n
mn (G)
const s(i G)
(6.7.18)
holds for all open sets G such that i G is a smooth manifold and
mn (G) < .
Proof. 1. Since Ip,q (0), it follows that

1p/n
mn (H)
const cp (G\H)
for any bounded set H F that is closed in . Now (6.7.17) follows from
cp (G\H) cp (G\F ) and mn (F ) = supH mn (H).
1
6.7 Inclusion Wp,r
373
2. Let J11/n (0), mn (G) < , s(i G) < . We note that we did not
use the boundedness of G in the proof of Lemma 5.2.2 and insert the sequence
constructed in this lemma into (6.7.14). Passing to the limit as m , we
obtain
11/n

const s(i G)
mn (e)
for any compactum e G.
Lemma 2. 1. If Ip,q (0) with q = pn/(n p), n > p, and (6.7.15)

holds, then for any u C () L1p () there exists a unique number c such
that

(6.7.19)
mn x : u(x) c < for all > 0.
2. The same is true for p = 1 provided J11/n (0) and (6.7.16) holds.
Proof. 1. We introduce the sets At = {x : u(x) > t}, Bt = {x : u(x) t},
Ct = \At , Dt = \Bt and put

(6.7.20)
c = inf t : mn (At ) < .
Suppose that c = +. Then mn (At ) = for all t R1 and mn (Ct ) < for
all t by
(6.7.21)
(T t)p cp (At \BT ) upLp ()
and (6.7.12). According to (6.7.17), we have

mn (Ct )
1p/n
const cp (DT \Ct )
(6.7.22)
for all T > t. Since

(T t)p cp (DT \Ct ) upLp () ,
the right-hand side in (6.7.22) tends to zero as T +. Consequently,
mn (Ct ) = 0 for all t and c < +.
Now let c = . Then mn (At ) < for all t R1 . Applying (6.7.17) to
the conductor At \BT with T > t, we arrive at

mn (BT )
1p/n
const cp (At \BT ).
Since by (6.7.21) the right-hand side tends to zero as t , we see that

mn (BT ) = 0 for all T . Thus c > .
Now we prove (6.7.19). From denition (6.7.20) it follows that

(6.7.23)
mn {x : u c } <
for > 0. On other hand, (6.7.20) gives mn ({x : u c /2}) = .
Since the p-conductivity of the conductor Ac/2 \Bc is nite
(cf. (6.7.21)), by (6.7.15) we have
374

mn {x : u < c } < .
(6.7.24)
Inequalities (6.7.23) and (6.7.24) are equivalent to (6.7.19). The uniqueness

of the constant c is an obvious corollary of the condition (6.7.15) and the
niteness of the conductivity of any conductor At \BT , t > T . The rst part
of the Lemma is proved.
2. Now let p = 1. The identity
+

|u| dx =
s(At ) dt, u C () L11 ()
implies
s(At ) <
for almost all t
(6.7.25)
and
lim inf s(At ) = lim inf s(At ) = 0.
t
t+
(6.7.26)
Further, it suces to duplicate the argument in the proof of the rst part of
the Theorem using (6.7.25) in place of the niteness of the conductivity of the
conductor At \BT = DT \Ct , and (6.7.26) in place of the convergence to zero

of cp (At \BT ) as t + or T . The lemma is proved.
We complete the proof of suciency in the Theorem. Let u C ()
According to Lemma 6.1.1/1 we may in addition assume that u
L ().
We put
u(x) c if u(x) > c + ,

u (x) = 0
if |u(x) c| ,
u(x) c + if u(x) < c ,

L1p ().
where c is the constant specied in Lemma 2. From (6.7.19) and the boundedness of u it follows that u Lpn/(np) (). Hence u can be inserted
into (6.7.14). Passing to the limit as +0, we arrive at (6.7.13). The
theorem is proved.
6.8 Compactness of the Embedding L1p() Lq ()

In this section we obtain the necessary and sucient conditions for the sets
bounded in L1p () to be compact in Lq (). Here is a domain with nite
volume.
6.8.1 Class Ip,
As before, by Ap, (M ) we mean the constant in the denition of the class
Ip, .
6.8 Compactness of the Embedding L1p () Lq ()
375
Denition. The domain is contained in the class Ip, provided that

Ap, (M ) 0 as M 0.
The equality (6.3.4) implies that I1, if and only if
[mn (g)]
=0
M 0 {g:mn (g)M } s(i g)
lim
sup
(6.8.1)
(as before, here g designates an admissible subset of ).

The value of in the denition of Ip, exceeds 1/p 1/n since
) = const np
cp (B2 \B
and hence
[mn (B )]1/p1/n

)]1/p = const > 0.
[cp (B2 \B
6.8.2 Compactness Criteria

Theorem 1. Let mn () < . The embedding operator of L1p () into Lq (),
1 < p q < , is compact if and only if Ip, for p 1, where
1 = q .
Proof. Suciency. Let u be an arbitrary function in C () L1p ()
L () with bounded support. (By Corollary 5.1.2, the set of such functions
is dense in L1p ().) Let

T = inf t : mn (Nt ) M .
Obviously,

uLq () c |u| T + L
By Corollary 6.3.3 we have

|u| T
+ L
q ()
q ()

1/q
+ T mn ()
.
(M )uLp () ,
where (M ) = cAp, (M ).
Let M denote a bounded subdomain of with C 0,1 boundary and with
mn (\M ) < M/2. Since mn (NT ) M , it follows that mn (NT M )
M/2. Consequently,
uLr (M ) 21/r T M 1/r
and we arrive at

1/q
uLq () c(M )uLp () + cM 1/r mn ()
uLr (M ) .
By Corollary 5.1.2 the latter is valid for all u L1p ().
(6.8.2)
376
Since M is a domain with a smooth boundary and compact closure, the

embedding operator of L1p (M ) into Lr (M ) is compact. Let {um }m1 with
um L1p () = 1 be a Cauchy sequence in Lr (M ). Then (6.8.2) implies

1/q
um ul Lq () c(M ) + cM 1/r mn ()
um uLr (M ) .
(6.8.3)
Given any > 0 we can nd an M such that c (M ) < /2. Next we choose
a large enough number N so that the second term in (6.8.3) does not exceed
/2 for m, l > N . Then um ul Lq () < for m, l > N and hence {um }
is a Cauchy sequence in Lq ().
Necessity. Suppose the embedding operator of L1p () into Lq () is compact. Then the elements of the unit sphere in L1p () have absolutely equicontinuous norms in Lq (). Thus, for all u L1p (),

uLq (G) (M ) uLp () + uL1 () ,
where is a bounded subdomain of ,
, G is an arbitrary open subset
of with mn (G) M and (M ) tends to zero as M 0.
Suppose that the function u L1p () vanishes outside G. Then

1
uLp () .
uLq () (M ) 1 (M )M 11/q
It remains to use the second part of Corollary 6.3.3. The theorem is proved.
Theorem 2. Let mn () < . The embedding operator of L1p () into
Lq (), p > 1, 0 < q < p, is compact if and only if it is bounded, i.e.,
Hp, , where 1 = q .
Proof. Let Hp, . Then Bp, ( 12 mn ()) < and since M,p (t)
mn ()/2 (t) for t < M we have Bp, (M ) 0 as M 0. It remains to notice
that the proof of suciency in Theorem 1 together with Corollary 6.4.2/1

imply (6.8.3) with (M ) = cBp, (M ). The proof is complete.
6.8.3 Sucient Conditions for Compactness of the Embedding
L1p () Lq ()
The inequality
Ap1 ,1 (M ) cAp, (M )
1
implies the embedding
with p1 > p 1, 1 p1
1 =p
Ip, Ip1 ,1 .
In particular,
def
J+1p1 = I1,+1p1 Ip, .
The preceding leads to the following corollary.
6.8 Compactness of the Embedding L1p () Lq ()
377
Corollary 1. If J and p(1 ) < 1, 1 p q = p/[1 + p(1 )]

then the embedding operator of L1p () into Lq () is compact.
By Proposition 6.4.4 the requirement

0
M
M ( )
p/(p1)
d <
(6.8.4)
implies Bp, (M ) 0 as M 0. Therefore we have the next statement.

Corollary 2. If the integral (6.8.4) converges, then the embedding operator
of L1p () into Lq () (q = 1 , p > 1) is compact.
Corollaries 1 and 2 immediately imply the following rougher assertion.
Corollary 3. If J and p(1 ) 1, p 1, 1 q < p/[1 + p( 1)]
then the embedding operator of L1p () into Lq () is compact.
1
Clearly, the space L1p () can be replaced by Wp,r
() with r q in
Theorems 6.8.2/1 and 6.8.2/2 and Corollaries 13.
6.8.4 Compactness Theorem for an Arbitrary Domain with Finite

Volume
The positiveness of the function M (cf. Lemma 5.2.4) and the estimate
(6.3.11) imply M,p (t) > 0 for t > 0. Hence there exists a nondecreasing
positive continuous function on (0, mn ()] such that (t)/M,p (t) tends to
zero as t +0.
Theorem. Let be an arbitrary domain with nite volume. Then from
any bounded sequence in L1p () we can select a subsequence {um }m1 with

m,k
mn x : um (x) uk (x) t d tp 0
0
and therefore any bounded subset of L1p () is compact in an n-dimensional

Lebesgue measure.
Proof. Let u be a function in C () L1p () L () with bounded
support. Obviously,

p

mn (Nt ) d t
mn (Nt ) d tp + T p mn () .
(6.8.5)
0
Here T = inf{t : mn (Nt ) }, where is a suciently small positive number

that is independent of u. The right-hand side in (6.8.5) does not exceed

p

( )
sup
cp (LT \Nt ) d tp + c mn () p
|u| dx ,
0< M,p ( ) T
378
where is the domain specied in the proof of Theorem 6.8.2 with M

replaced by . By Lemma 6.1.3/3 we have

( )
upLp ()
mn (Nt ) d tp sup
(
)
0 M,p
0

+ c mn () p upL( ) .
(6.8.6)
By Corollary 5.1.2, the last inequality extends to encompass all functions in
L1p (). It remains to apply the arguments used at the end of the proof of
suciency in Theorem 6.8.2/1.
6.8.5 Examples of Domains in the class Ip,
Example 1. The estimates (6.3.15) imply that the domain

= x : (x21 + + x2n1 )1/2 < f (xn ), 0 < xn < a ,
in Example 6.3.6/1 is contained in Ip, if and only if

x
lim
x+0
n1
f (t)
p/(p1)
dt
a
(1n)/(p1)
f (t)
dt = 0.
By (5.3.5) and Corollary 6.8.3/1, a sucient condition for to belong to
Ip, is
+11/p
x

1n

n1
f (t)
dt
= 0.
(6.8.7)
lim f (x)
x+0
Since f does not decrease, we see that (6.8.7) holds for p = 1 and also for
p < 1 provided limx+0 x f (x) = 0, = (p + p 1)/(n 1)(p 1).
Example 2. The domain = {x : 0 < xn < , (x21 + + x2n1 )1/2 <
f (xn )} in Example 6.3.6/2 is contained in Ip, if and only if

lim
x+
+
n1
f ( )
p/(p1)
x
(1n)/(p1)
f ( )
d = 0
(6.8.8)
(cf. estimates (6.3.18)). By (5.3.8) and (6.3.11) the last equality holds if
lim
x+
1n
f (x)

n1
f ( )
+11/p
d
= 0.
(6.8.9)
In particular, Ip,1/p provided

f ( ) = e( ) ,
( ) +
as +.
In the case f ( ) = ec , the domain under consideration is contained in Ip,1/p

and does not belong to Ip,1/p .
6.9 Embedding Llp () Lq ()
379
Similarly, a necessary and sucient condition for the spiral in Examples 5.3.3/3 and 6.3.6/3 to be in Ip,1/p is that

() d
lim
p/(p1)
()
1/(1p)
d = 0.
A simpler sucient condition is that

+11/p

() d
= o ()
as +.
In particular, Ip,1/p if () = e() , () + as +, and

Ip,1/p \Ip,1/p if () = ec .
6.9 Embedding Llp() Lq ()

We present sucient conditions for the boundedness and compactness of the
embedding operator of Llp () into Lq (), which are simple corollaries of
Theorems 6.3.3 and 6.4.2.
Theorem 1. If Ip, , 1 1/l 1 then the
embedding operator of Llp () into Lq (), q = p/(1 l + pl) is bounded.
The proof is by induction on the number of derivatives l. In addition, we
must use the embeddings
Ip, Ip1 ,1 ,
Hp, Hp1 ,1 ,
1
(cf. Corollaries 6.2 and 6.4.2/2).
with p1 > p 1, 1 p1
1 =p
In particular, Theorem 1 guarantees the continuity of the embedding operator of Llp () (lp < n) into Lq () with the same q = pn/(n lp) as in
the Sobolev theorem for domains of the class Ip,1/p1/n .
Theorem 1 and Proposition 6.3.5/2 imply the following corollary.
Corollary 1. Let J , 1 1/n 1, lp(1 ) 1. Then
Llp () Lq () where q = p/(1 p l(1 )) for p l(1 ) < 1 and q is
arbitrary for pl(1 ) = 1. (Then the exponent q = pn/(n pl) corresponds
to = 1 1/n.)
Example 1. Since the domain

() = x : x21 + + x2n1 < x2
n , 0 < xn < 1 ,
1,
belongs to the class J with = (n1)/((n1)+1), (cf. Example 5.3.3/1),

we have Llp ( () ) Lq ( () ), where 1 + (n 1) > pl and
380

q = p (n 1) + 1 / 1 + (n 1) pl .
(6.9.1)
The example of the function u(x) = xn with = l + (1 + (n 1))/p

( is small positive number) shows that the exponent q cannot be increased.
Similarly to Theorem 1 we obtain the following theorem stating some
conditions for the compactness of the embedding Llp () Lq () for domains
with nite measure mn .
Theorem 2. If Ip, , 1 1/l 1,
then the embedding operator of Llp () into Lq (), q = p/(1 pl(1 )), is
compact.
This theorem and Corollary 6.8.3/1 imply the next statement.
Corollary 2. If mn () < and J where 1 1/n < 1,
lp(1 ) < 1, then the embedding operator of Llp () into Lq (), q < p/(1
pl(1 )), is compact.
From this we immediately obtain the following coarser sucient conditions
formulated in terms of J .
Corollary 3. If mn () < and J , 11/n 1, lp(1) 1,
then the embedding operator of L1p () into Lq (), q < p/(1 pl(1 )) is
compact.
Example 2. Consider the domain in Example 1. By Corollary 3 the embedding operator of Llp ( () ) into Lq ( () ) is compact if

q 0 , where
u (x) = l((n1)+1)/p (x/),
is bounded in Llp ( () ), but not compact in Lq ( () ) with q specied
by (6.9.1).
6.10 Applications to the Neumann Problem for Strongly

Elliptic Operators
Here we present some applications of the previous results to the study of the
solvability and the discreteness of the spectrum of the Neumann problem in
domains with irregular boundaries.
6.10 Applications to the Neumann Problem for Strongly Elliptic Operators
381
6.10.1 Second-Order Operators

Let be a domain with nite volume in Rn and let aij (i, j = 1, . . . , n) be real
measurable functions in , aij = aji . Suppose that there exists a constant
c 1 such that
c1 ||2 aij i j c||2
for almost all x and for all vectors = (1 , . . . , n ).
We dene the operator Aq , 1 q < of the Neumann problem for the
dierential operator

u
u
aij
xi
xj
1
(), Aq u Lq (), 1/q + 1/q = 1; 2.
by the following conditions: 1. u W2,q
l
for all v W2,q () the equality

u v
vAq u dx =
aij
dx
(6.10.1)
xi xj
is valid. The mapping u Aq u is closed. It is clear that the range R(Aq ) is

contained in the set Lq () 1 of functions in Lq () that are orthogonal to
unity in .
1
() the
Lemma. R(Aq ) = Lq () 1 if and only if for all v W2,q
generalized Poincare inequality
inf v cLq () kvL2 ()
cR1
(6.10.2)
holds.
1
() is dense in L12 ().
Proof. Suciency. By Lemma 5.1.2/2, the set W2,q
1
() then it also holds for all v L12 ().
Thus, if (6.10.2) holds for all v W2,q
Therefore, the functional v vf dx is bounded in L12 () and can be
expressed in the form

aij (v/xj )(u/xi ) dx, u L12 (),
(6.10.3)
for an arbitrary function f Lq (). Since by (6.10.2) L12 () Lq (), then
1
() and hence Aq u = f .
u W2,q
Necessity. Let f Lq () 1, v L12 () Lq (), vL2 () = 1. Since
R(Aq ) = Lq () 1, the functional

f v(f ) =
f v dx,
dened on Lq 1, can be expressed in the form (6.10.3). Therefore, |v(f )|

CuL2 () and the functionals v(f ) are bounded for each f Lq (). Thus,
382
1
they are totally bounded, i.e., (6.10.2) holds for all v W2,q
(). The lemma
is proved.
Theorems 6.3.3, 6.4.2 and the above lemma imply the following criterion
for the solvability of the problem Aq u = f for all f Lq () 1.
Theorem 1. R(Aq ) = Lq () 1 if and only if I2,1/q for q 2 and
H2,1/q for q 2.
Let a be a real function in L () such that a(x) const > 0 for almost
all x . Then the operator u Aq u + au has the same domain as Aq .
Consider the Neumann problem Aq u + au = f where f Lq (). If q 2
then
its solvability is a trivial consequence of the continuity of the functional

f
v
dx in the space W21 () with the inner product

aij (v/xj )(u/xi ) + avu dx.
If q < 2 then a necessary and sucient condition for solvability is

vLq () CvW21 ()
for all v W21 ().
The preceding theorem and Theorem 6.3.3 directly imply the next result.
Theorem 2. R(Aq + aI) = Lq () with q < 2 if and only if I2,1/q .
By Rellichs lemma the problem of requirements on for the discreteness
def
of the spectrum of the operator A = A2 is equivalent to the study of the compactness of the embedding W21 () L2 (). Therefore, from Theorem 6.8.2
we immediately obtain the following criterion.
Theorem 3. The spectrum of the operator A is discrete if and only if
I2,1/2 .
Sucient conditions for a set to be in I2,1/q , H2,1/q , I2,1/2 as well as
examples of domains belonging to these classes were presented in previous
sections of this chapter.
6.10.2 Neumann Problem for Operators of Arbitrary Order
In this subsection we limit consideration to strongly elliptic operators with
the range in L2 ().
Let be a bounded subdomain of Rn . Let i, j denote multi-indices of
order not higher than l, l 1, and let aij denote bounded complex-valued
measurable functions in .
Suppose that for all u Ll2 ()
Re
|i|=|j|=l
aij Di uD j u dx Cl u2L2 () ,
383
(6.10.4)
where D i = { |i| /xi11 , . . . , xinn }. Let the operator A of the Neumann problem
for the dierential operator

u (1)l
Di aij Dj u
|i|=|j|=l
be dened by the following conditions: 1. u W2l (), Au L2 (), 2. for all

v W2l ()

vAu dx =
aij D j uDi v dx.
|i|=|j|=l
It is clear that the range R(A) is contained in the orthogonal complement

L2 () Pl1 where Pl1 is the space of polynomials of degree not higher
than l 1.
If for all v Ll2 ()
inf
Pl1
v L2 () kl vL2 () ,
(6.10.5)
then R(A) = L2 () Pl1 (cf. Lions and Magenes [500], 9.1, Chap. 2). By
a simple argument using induction on the number of derivatives, we show
that (6.10.5) follows from the Poincare inequality
inf u cL2 () kvL2 () .
cR1
Thus we have the following assertion.

Theorem 1. If I2,1/2 then R(A) = L2 ()Pl1 ; i.e., the Neumann
problem Au = f is solvable for all f L2 () Pl1 .
We can pose the Neumann problem for the more general operator

u (1)l
D i aij D j u
|i|,|j|l
in a similar way. We dene the operator B of this problem by the conditions:

1. u V2l (), Bu L2 (); 2. for all v V2l ()

vBu dx =
aij Dj uD i v dx.
(6.10.6)
|i|,|j|l
Theorem 2. If I2,1/2 then R(B + I) = L2 () for suciently

large values of Re . Moreover, the operator (B + I)1 : L2 () V2l () is
compact.
384
To prove the theorem we need the following lemma.

Lemma. If I2,1/2 , then, for all u W2l () and for any > 0,
l1

k uL2 () l uL2 () + C()uL2 () .
(6.10.7)
k=0
Proof. By (6.8.2) we have

uL2 () c A2,1/2 (M )uL2 () + c
mn ()
M
1/2
uL2 (M ) ,
where M is a subdomain of with boundary of the class C 1 and such that

2mn (\M ) < M . Therefore

k uL2 ()
mn ()
c A2,1/2 (M )k+1 uL2 () + c
M
1/2
k uL2 (M )
for all k = 0, 1, . . . , l 1. Since the boundary of M is in C 1 , it follows that

k uL2 (M ) k+1 uL2 (M ) + C (0) ()uL2 (M )
for all > 0. Therefore
k uL2 () c A2,1/2 (M )k+1 uL2 () + C (1) (M )uL2 (M ) .
Applying this inequality with indices k, k + 1, . . . , l 1, we obtain

lk
l uL2 () + C (2) (M )uL2 (M ) .
k uL2 () c A2,1/2 (M )
It remains to note that A2,1/2 (M ) 0 as M 0. The lemma is proved.
We established, incidentally, that (6.10.7) is valid for some > 0 provided
I2,1/2 , i.e., A2,1/2 (M ) < .
Proof of Theorem 2. By (6.10.4) we have

Re

|i|,|j|l
aij D i uDj u dx Cl u2L2 () C1
l1

k u2L2 () .
k=0
Applying (6.10.7) with = C/2C1 we obtain

Re
aij Di uDj u + |u|2 dx
|i|,|j|l
1
C1 l u2L2 () + (Re C2 )u2L2 () .
2
385
Thus the coercive inequality

l

k u2L2 ()
const Re

k=0
aij D
uDj u
+ |u|
dx
|i|,|j|l
holds for Re > C2 . This implies (cf., for instance, Lions and Magenes [500],
Chap. 2, 9.1) the unique solvability of the equation Bu + u = f for all
f L2 (). The compactness of (B + I)1 results from Theorem 6.9/2. The
proof is complete.

6.10.3 Neumann Problem for a Special Domain
The monograph by Courant and Hilbert [216] contains the following example
of a domain for which the Poincare inequality is false.
Let be the union of the square

Q = (x, y) : 0 < x < 2, 1 < y < 1
and the sequence of symmetrically situated squares Qm , Qm , m = 1, 2, . . . ,
connected with Q by the necks Sm , Sm (Fig. 27). Let the side lengths of the
squares Qm , Qm as well as the heights of the necks be equal to m = 2m .
Let the widths of the necks be
m.
In this subsection we show that the Poincare inequality holds only for
3 and the Rellich lemma holds only for < 3. Hence the Neumann
problem considered in Sect. 6.10.1 is solvable in L12 () for any right-hand
side in L12 () 1 if and only if 3 and the spectrum of this problem is
discrete if and only if < 3.
We introduce the sequence of functions {um }m1 dened by um = 1
m
on Qm , um (x, y) = 2
m (y1) on Sm , um = 0 outside Sm Sm Qm Qm .
This sequence satises

um dx dy = 0,
(um )2 dx dy = 23
,
u2m dx dy > 2.
m
Thus the Poincare inequality is false for if > 3 and the embedding
operator of W21 () into L2 () is not compact if 3.
Consider a nonnegative function u equal to zero outside Qm Sm (m is a
xed natural number) and innitely dierentiable in Qm Sm . We introduce
the notation

Ht = (x, y) : u < t ,
Nt = (x, y) : u t .
Et = (x, y) : u = t ,
If t (0, ) satises
then
s(Et ) 2(+1)m/2 ,
(6.10.8)
386
Fig. 27.
m2 (Nt ) c2(3)m/2 s(Et ).
(6.10.9)
The set of levels t for which (6.10.8) is valid will be denoted by P. Duplicating
with minor modication the arguments presented in Sect. 4.5, we can show
that one of the following cases occurs for t CP:
s(Et ) 2m ,
m2 (Nt ) < c2(1+)m ;
(6.10.10)
(6.10.11)
(6.10.12)
s(Et ) 2m ,
m2 (Ht ) < c2(1+)m ;
m
,
s(Et ) < 2
(3)m/2
s(Et ).
m2 (Nt ) c2
(6.10.13)
(6.10.14)
(6.10.15)
This implies that the set of all levels t (up to a set of measure zero) can be
represented as the union R R R where R is the set of those t for
which (6.10.14) and (6.10.15) are valid, R and R are the set of levels for
which (6.10.10) and (6.10.11) and (6.10.12) and (6.10.13) hold, respectively.
The function

1
t (t) =
0
|u| ds
387
admits the estimate

d
d
m2 (N )
d
[s(E )]2
(t)
0
(cf. Lemma 6.1.3/2). We express the right-hand side as the sum

+
+
.
R
R
R
The inequality (6.10.15) implies

R
c2(3)m
t
m2 (N )
2 d
m2 (N ) d c2(3)m /m2 (Nt ).
d

2m sup m2 (N ) d 2(1)m .
R
R
Similarly, by (6.10.12) and (6.10.13),

2m sup m2 (N ) 2(1)m .
R
R
Consequently,
2(3)m
(1)m
(t) c 2
+
.
m2 (Nt )
(6.10.16)
Since m2 (Nt ) 22m , (6.10.16) yields

m2 (Nt )(t) c2(3)m .
Therefore,

u2 dx dy = 2
Qm Sm
(6.10.17)
tt m2 (Nt() ) d c2(3)m
tt
d
.
The integral on the right-hand side does not exceed

0
d
t2 2
1/2
(t )2
1/2
(t )2 d
(u)2 dx dy.
=2
Qm Sm
Hence

u dx dy c2
2
Qm Sm
(3)m

(u)2 dx dy.
Qm Sm
(6.10.18)
388
It is clear that this inequality is valid for any function u L12 () that is
equal to zero outside Qm Sm and that a similar estimate holds for negative
indices m.
Now let v be an arbitrary function in L12 () and let be a truncating
function equal to zero in Q, to unity in Qm and in Qm , and linear in Sm
and Sm (m = 1, 2, . . . ). Then

2
2
2
v dx dy
(v) dx dy +
v dx dy
Qk Sk
|k|N
Sk
v 2 dx dy,
(6.10.19)

where N =
|k|N (Qk Sk ) and N is an arbitrary positive integer.
From (6.10.18) we obtain

(v)2 dx dy
Qm Sm

c2(3)m
(v)2 dx dy + 22m
v 2 dx dy . (6.10.20)
Qm Sm
Sm
We estimate the second integral on the right. We can easily check that

m
m
|v| dx dy 2
|v| dx dy + 2
|v| dx.
Sm
Hence
Sm Q
Sm

v 2 dx dy
Sm
1
2
v 2 dx dy + 22m1
Sm

m
v2 dx.
+2

(v)2 dx dy
Sm
Sm Q
olders inequality,
Since s(Sm Q) = 2m , by H

v 2 dx dy c22m
(v)2 dx dy
Sm
Sm
2m
+ c2
Therefore (6.10.20) implies

(v)2 dx dy
Qm Sm

c2(3)m
Qm Sm
v2L2/(1) (Sm Q) .
(6.10.21)

(v)2 dx dy + v2L2/(1) (Sm Q) .
(6.10.22)
389
Using (6.10.21) and (6.10.22), from (6.10.19) we obtain

2
(3)N
v dx dy c2
(v)2 dx dy
Qk Sk
|k|N
v2L2/(1) (Sk Q)

v 2 dx dy.
+
N
Again, by Holders inequality,

v2L2/(1) (Sk Q)
|k|N
2(1)|k|/2 v2L4/(1) (Sk Q)
|k|N

2|k|(1)/(+1)
(+1)/2
|k|N

|k|N
|v|
4/(1)
Sk Q
(1)/2
dx
cv2L4/(1) (Q) .
Thus

v dx dy c2
2
(3)N
(v)2 dx dy + cv2L4/(1) (Q)
+ cv2L2 (N ) .
(6.10.23)
Since for any r (1, )

vLr (Q) c vL2 () + vL2 () ,
(6.10.23) yields
vL2 () c 2(3)N/2 vL2 () + cvL2 (N ) .
(6.10.24)
The boundary of N is of the class C 0,1 , and so the embedding operator of

W21 (N ) into L2 (N ) is compact. This and (6.10.24) imply the compactness
of the embedding operator of W21 () into L2 () for < 3. Finally, for 3
from (6.10.24) we get the Poincare inequality

inf 1 v L2 () c vL2 () + inf 1 v L2 (1 ) c vL2 () .
R
6.10.4 Counterexample to Inequality (6.10.7)

We shall show that the validity of inequality (6.10.7) for some > 0 does not
imply its validity for all > 0.
390
Let be the domain considered in Sect. 6.10.3. Suppose the width of the
necks Sm and Sm is equal to 23m . In Sect. 6.10.3 we showed that in this
case I2,1/2 . So (6.10.7) holds for some > 0.
Consider the sequence of functions {um }m1 dened by um = 0 outside
Qm Sm , um (x, y) = 4m (y 1)2 on Sm , um (x, y) = 2m+1 (y 1) 1 in Qm .
We can easily see that

2
m
,
|um | dx dy 4 1 + 4
|2 um |2 dx dy = 4,

u2m dx dy 4m .
By (6.10.7) with l = 2 we have

1/2

2 + k()2m ,
2 1 + 4m
m = 1, 2, . . . ,
which fails for some > 0.
6.11 Inequalities Containing Integrals over the Boundary

1
(, ) Lq ()
The content of the present subsection is related to that of Sect. 5.6 where
1
the case p = 1 is considered. The space Wp,r
(, ) and the class K, are
dened in Sect. 5.6.1.
Theorem 5.6.3 implies the following sucient condition for the continuity
1
(, ) Lq () for p > 1.
of the embedding Wp,r
Theorem 1. If K, with 1, p(1 ) < 1, then for all
1
(, )
u Wp,r
1 (,) ,
uLq () cuWp,r
(6.11.1)
where q = p/(1 p(1 )), r = p/(1 p(1 )).
with bounded supports
Proof. By Theorem 5.6.3 for all v C ()C()
we have

1/
1/
|v|
dx
C vL() +
|v|
ds
.
We put v = |u|q and see that by Holders inequality

|u|
Hence
(p1)/(1p(1))
|u| dx uLp ()
11/p
|u|
p/(1p(1))
dx
391

1/p1/+1
(p1)/p
uLq () C1 uLp ()
uLq () + uLr () ,
Since any set is contained in K11/n,1 , we obtain the following corollary.


uLpn/(np) () c uLp () + uLp(n1)/(np) ()
(6.11.2)
1
holds for all Wp,p(n1)/(np)
(, ) with p < n for an arbitrary open set .
Replacing u by |u|r and applying Holders inequality in (5.6.20) we arrive

at the next assertion.
Corollary 2. The following rened Friedrichs type inequality

uLq () C uLp () + uLr ()
(6.11.3)
holds for (n p)r p(n 1), r 1, q = rn/(n 1) for an arbitrary open set
with nite volume.
We show that the exponent q = rn/(n 1) on the left in (6.11.3) cannot
be improved provided is not subject to additional conditions.
Example. Let the domain be the union of the semiball B = {x :
xn < 0, |x| < 1}, the sequence of balls Bm (m = 1, 2, . . . , ) and thin pipes Cm
connecting Bm with B (Fig. 28). Let m be the radius of the ball Bm and
let hm be the height of Cm . Let um denote a piecewise linear function equal
to unity in Bm and to zero outside Bm Cm . Suppose that there exists a
constant Q such that

um Lq () Q um Lp () + um Lr ()
Fig. 28.
392
for all um . This implies the estimate

mn (Bm )
1/q

1/p
1/r
.
Q h1
+ s(Bm Cm )
m mn (Cm )
Since the rst term on the right can be made arbitrarily small by diminishing
n/q
(n1)/r
the width of Cm , it follows that m = O(m
). Hence q rn/(n 1).
To formulate a necessary and sucient condition for the validity of (6.11.1)
we need the following modication of the p-conductivity.
Let K be a conductor in . We put

f 1 on F, f 0 on \G .
|f |p dx : f C ()C(),
cp (K) = inf
(6.11.4)
Similarly to Lemma 6.1.1/2 we can prove that cp (K) can be expressed as
follows:

f = 1 on F, f = 0 on \G .
cp (K) = inf
|f |p dx : f C ()C(),
(6.11.5)
Theorem 2. A necessary and sucient condition for the validity of
(6.11.1) with p > 1, q p r is

1/q
mn (F )

1/p
1/r
const cp (K)
+ s(e G)
,
where K is any conductor G\F in .

This assertion can be proved similarly to Theorem 2.3.9.
The same argument as in the proof of Theorem 6.8.2/1 leads to the following criterion for compactness.
Theorem 3. A necessary and sucient condition for compactness of the
1
(, ) into Lq () with mn () < and q
embedding operator of Wp,r
p r is

[mn (F )]1/q
lim sup
: mn (G) M = 0.
(6.11.6)
M 0
[
cp (K)]1/p + [s(e G)]1/r
This and Corollary 2 imply the following assertion.
Corollary 3. Let (n p)r p(n 1), r 1, q < rn/(n 1). Then the
1
(, ) into Lq () is compact for arbitrary open
embedding operator of Wp,r
set with nite volume.
The example of the present subsection shows that there exist domains
1
with nite volume for which the embedding Wp,r
(, ) Lrn/(n1) () is
not compact.
393
(n1)
6.11.2 Classes Ip,
and J(n1)
1 () be the completion of the set of functions in C ()

Denition 1. Let W
p
C() with bounded supports with respect to the norm in Wp1 ().
(n1)
Denition 2. We say that is contained in the the class Ip,

(n p)/p(n 1), if there exists a constant R > 0 such that
def
Mp, (R) = sup
, p 1,
[s(e F )]
< .
[
cp (K)]1/p
Here e F = F and the supremum is taken over the set of all conductors
K = G\F in with G = BR (x), x .
(n1)
The restriction (np)/p(n1) is due to the fact that the class Ip,
contains only sets with boundary having the (n 1)-dimensional Hausdor
measure zero provided < (n p)/p(n 1).
Proposition 1. If < (n p)/p(n 1), then either Mp, (R) =
identically, or s() = 0.
(n1)
Proof. Let Ip, , s() > 0 and let be a small enough positive
number. We construct a covering of by open balls Brj (xj ) with rj <
and

rjn1 cs().
j
Then
rjn1 c

s Brj (xj ) ,
j
and hence, for at least one ball, we have

rjn1 cs Brj (xj ) .
(6.11.7)
Consider the conductor Kj = Gj \Fj where Gj = B2rj (xj ) and Fj =

Brj (xj ). Since cp (Kj ) p-cap(Brj (xj ), B2rj (xj )) = c rjnp (cf. Sect. 2.2.4), by
(n1)
(n1)p
and by estimate (6.11.7) we obtain rj
denition of the class Ip,

const rjnp . Noting that rj < and is small, we obtain (n 1)p n p.

(n1)
Denition 3. The set is contained in the class J

a constant M (0, mn ()) such that
R (M )
if there exists
[s(e g)]
= sup
: g is an admissible subset of with mn (g) M
s(i g)
def
is nite.
394
We introduce the function

Pp, (M ) = sup
[s(e F )]
,
[
cp (K)]1/p
where the supremum is taken over the collection of all conductors K = G\F
in with mn (G) M .
In the same way as we proved Proposition 6.6 we can prove the following
assertion.

1/
/
R (M )
Pp, (M ) c Ap, (M )
holds for = p/(p 1 + p), where Ap, is the function in the denition
of the class Ip, (cf. Sect. 6.3.1).
From this proposition it follows that
(n1)
.
J(n1) Ip, Ip,
(6.11.8)
(n1)
6.11.3 Examples of Domains in Ip,
and J(n1)
Example 1. Let x = (x1 , . . . , xn1 ) and let

= x : x Rn1 , < xn < |x | ,
0 < < n 2.
Let g also denote an arbitrary admissible subset of with s(i g) < 1. We

have

1/2
1 + 2 |x |2(+1)
dx ,
s(e g) =
Pr(e g)
where Pr is the orthogonal projection mapping onto the plane xn = 0. It is

clear that the supremum of the integral

1/2
1 + 2 |x |2(+1)
dx ,
E
taken over all subsets E of the plane xn = 0 with a xed (n 1)-dimensional

measure, is attained at the (n 1)-dimensional ball centered at the origin.
Therefore

1/2
r n2 1 + 2 r 2(+1)
dr,
s(e g) c
0
where
This implies
1
1/(n1)
= vn1
s Pr(e g)
.

(n2)/(n1)
.
s(e g) cs Pr(e g)
However, since
we have
395

s Pr(e g) s(i g),

(n1)/(n2)
s(e g)
cs(i g).
In Example 6.3.6/2 we showed that is contained in Ip, with =

(n1)
(n 1 + (p 1))/(n 1 )p. Using (6.11.8) we obtain that Ip,
with = (n 1 + (p 1))/p(n 2 ).
This value of is the best possible which is checked using the sequence
of conductors Km = Gm \Fm , where Gm = {x : 0 < xn < 2m1 } and
Fm = {x : 0 < xn < m1 } (cf. Example 6.6).
Example 2. A similar argument shows that the set

= x : |x | < xn , 0 < xn <
(n1)
with 1 is contained in J(n1)/((n2)+1) . Since by Example 6.3.6/1,

Ip, with = ((n 1) + 1 p)/p((n 1) + 1), we see that (6.11.8)
(n1)
implies Ip,
with = ((n 1) + 1 p)/p((n 2) + 1).
6.11.4 Estimates for the Norm in Lq ()
Theorem 1. Let s() < .
(n1)
with p 1, then for all functions u C () C()

1. If Ip,
with bounded supports the inequality
uLq () CuWp1 ()
(6.11.9)
holds with q = 1 and with constant C that is independent of u.

2. If for the same set of functions u inequality (6.11.9) holds with 1/q
(n1)
(n p)/p(n 1), then Ip,1/q .
Proof. 1. We construct a covering of by equal open balls BR (xi ),
xi , such that the multiplicity of the covering is nite and depends
only on n. Let {i } be a partition of unity subordinate to this covering with
|i | c R1 .
Duplicating with obvious modications the proof of Theorem 2.3.3, we
obtain
ui qLq () c

s(e F )
(ui )q
,
Lp ()
q/p
cp (G\F )]
F BR (xi ) [
sup
where G = BR (xi ). Summing over i, we arrive at

uLq () cMp,1/q (R) uLp () + R1 uLp () .
(6.11.10)
396
2. We show that Pp,1/q (M ) is bounded for some small M > 0.

Consider an arbitrary conductor K = G\F in with mn (G) M , where
M is a constant which will be chosen at the end of the proof.
We insert any function f of the class specied in (6.11.5) into (6.11.9).
Then

f Lq () C f Lp () + f Lp ()

C f Lp () + M 1/p(n1)/nq f Lqn/(n1) () . (6.11.11)
By Corollary 6.11.1/1

f Lqn/(n1) () C f Lp () + f Lq () .
(6.11.12)
The inequalities (6.11.11) and (6.11.12) imply

f Lq () C f Lp () + M 1/p(n1)/nq f Lq () .
If from the very beginning the constant M is chosen to be so small that
2CM 1/p(n1)/nq < 1,
then
f Lq () 2Cf Lp () .
Minimizing the right-hand side, we obtain

s(e F )
1/q

1/p
2C cp (K)
.
The proof of Theorem 1 implies the following assertion.

(n1)
Corollary. Let s() < and p 1. The class Ip,

by the condition Pp, (M ) is nite for some M > 0.
can be dened
Theorem 2. Let s() < , mn () < .

(n1)
with
1. If Ip, , p < 1, then for all functions u C () C()
bounded supports the inequality
1 (,)
uLq () CuWp,p
(6.11.13)
holds with q = 1 and with a constant C independent of u.

2. If for any u of the same class (6.11.13) holds with q > p then
(n1)
Ip, , = q 1 .
The rst part of the theorem follows from (6.11.3) and Theorem 1, the
proof of the second part is similar to that of the second part of Theorem 1.
Theorem 3. Let be a domain with s() < , mn () < and let
q p. The inequality
inf u cLq () CuLp ()
cR1
397
(6.11.14)
with bounded support and with

holds for any function u in C () C()
(n1)
q p if only if Ip,1/q .
Proof. Suciency. Let K = G\F , mn (G) M . By Theorem 6.11.1/2 and
Corollary 6.11.1/3 we have

mn (F ) const cp (K) + s(e G) .
This and the corollary of the present subsection imply
mn (F ) const cp (K).
In other words, is contained in Ip,1/p . By Theorem 6.4.3/2 and
Lemma 5.2.3/1, for all u L1p (), we have
inf u cLp () CuLp () .
cR1
It remains to refer to Theorem 1.

Necessity. Substituting any function f of the class specied in (6.11.5)
into (6.11.14), we obtain

cp (K)]q/p .
min1 |1 c|q s(e F ) + |c|q s(\e G) C q [
cR
This leads to the estimate

s(e F )s(\e G)
1/(q1)
{[s(e F )]
+ [s(\e G)]1/(q1) }q1
Therefore,

q/p
C q cp (K)
.

q/p
s(e F ) 2q1 C q cp (K)
provided 2s(e G) s(). It remains to take any ball of suciently small

radius with center at as G. The theorem is proved.

(n1)
and Compactness Theorems
6.11.5 Class Ip,
(n1)
if
Denition. The set is contained in the class Ip,
lim Mp, (R) = 0,
R0
where Mp, (R) is the same as in Proposition 6.11.2/1.

In the proof of Proposition 6.11.2/1 we showed that > (n p)/p(n 1)
(n1)
provided s() > 0 and the class Ip,
is not empty.
398
Example. Consider the domains

1 = x : |x | < 1, 1 < xn < |x | , 0 < < n 2,

2 = x : |x | < xn , 0 < xn < 1 , 1.
(n1)
In Examples 6.11.3/1 and 6.11.3/2 we actually showed that 1 Ip,1

(n1)
and 2 Ip,2 , where 1 = (n 1 + (p 1))/(n 2 )p and 2 =
/ Ip,1 . Consequently,
((n 1) + 1 p)/((n 2) + 1)p as well as that i
i Ip,i (i = 1, 2) if and only if i > i .

Theorem 1. Let s() < and mn () < . The set of functions
having bounded supports and contained in the unit ball
in C () C()
1
of the space Wp () is relatively compact in Lq (), q p, if and only if
(n1)
.
I
p,1/q
(n1)
Proof. Suciency. Let Ip,1/q , q p. If uWp1 () 1, by Theorem 6.11.4/1 we have
uLp () + uLp () const.

1
(, ) into Lp () is
By Corollary 6.11.1/3 the embedding operator of Wp,p
1
compact and the unit ball in Wp () is a compact subset of Lp ().
Given any positive number we can nd an R such that Mp,1/q (R) < .
Hence, by (6.11.10), for all u C () C() with bounded supports we
have
uLq () uLp () + C()uLp () .
Now the result follows by a standard argument.

Necessity. Let be the set of functions specied in the statement of the
theorem. Since the traces on of functions in form a compact subset of
Lq (), given any > 0, we can nd an R such that
1/q
|u|q ds

BR (x)
for all u and for all balls BR (x). Let u be an arbitrary function in
with support in BR (x). We have
C () C()

uLp (BR (x)) uLp () + uLp () .
Since by Corollary 6.11.1/2

uLp () C uLp () + uLp () ,
it follows that
1/q

|u| ds
C uLp () +

1/p
|u| ds
.
BR (x)
399
BR (x)
Thus, if is small enough, then
1/q
|u| ds
2 CuLp () .

BR (x)
Let K be the conductor (BR (x) )\F . Substituting any function f of

the class specied by the formula (6.11.5) into the latter inequality, we obtain

1/q

1/p
s(e F )
2C cp (K)
.
While proving Theorem 1, we also obtained the following result.

Theorem 2. Let s() < and mn () < . The set of functions in
having bounded supports and contained in the unit ball of the
C C()
1
space Wp,p
(, ) is relatively compact in Lq (), q p, if and only if
(n1)
.
I
p,1/q
6.11.6 Criteria of Solvability of Boundary Value Problems for

Second-Order Elliptic Equations
In Sect. 6.10.1 we established necessary and sucient conditions for the solvability of the Neumann problem with homogeneous boundary data for uniformly elliptic second-order equations in the energy space as well as criteria
for the discreteness of the spectrum of this problem. The theorems of the
present section enable us to obtain similar results for the problem

aij
+ au = f in ,
Lu
xi
xj
(6.11.15)
u
M u aij
cos(, xj ) + bu = on ,
xi
where is an outward normal to . Here a and b are real functions, a
L (), b L () and aij = aji .
In what follows we assume that s() < , mn () < and that either
both a and b are separated from zero and positive or they vanish identically.
We assume for the moment that f L1 () and L1 (). The exact
formulation of the problem is as follows.
We require a function in W21 (, ) such that

u v
aij
+ auv dx +
buv ds =
f v dx +
v ds, (6.11.16)
xj xi
W21 (, ) with bounded support.

where v is an arbitrary function in C()
400
This formulation is correct since by denition of the space W21 (, )

and (6.11.3) the integrals on the left in (6.11.16) converge.
The case b = 0, = 0 was studied in Sect. 6.10.1. The same argument as
in Sect. 6.10.1 together with Theorems 6.11.1/16.11.1/3, 6.11.4/16.11.4/3,
6.11.5/1, and 6.11.5/2 leads to the following result.
Theorem 1. 1. If a = 0, b = 0, f = 0, q = q/(q 1) 2, then the
problem (6.11.15) is solvable for all Lq (), orthogonal to unity on ,
(n1)
if and only if I2,1/q .
2. If inf a > 0, b = 0 and f = 0 then the problem (6.11.15) is solvable for
(n1)
all Lq (), q 2 if and only if I2,1/q .
3. If a = 0, inf b > 0 and f = 0 then the problem (6.11.15) is solvable for
all L2 () for an arbitrary . Under the same assumptions on a, b, and
f a necessary and sucient condition for the solvability of (6.11.15) for all
(n1)
Lq (), q < 2, is the inclusion I2,1/q .
1
In each of these cases the solution of (6.11.15) is contained in W2,q
(, ).
4. Let = 0 and inf b > 0. The problem (6.11.15) is solvable for all
f Lq (), q 2n/(n+1), for an arbitrary set . A necessary and sucient
condition for the solvability of this problem for all f Lq (), q < 2n/(n+1),
is the condition of Theorem 6.11.1/2 with p = r = 2.
Theorem 2. 1. Under the assumptions 13 of the previous theorem, a
necessary and sucient condition for the compactness of the inverse operator
1
(, ),
Lq () W2,2
q 2,
(n1)
of the problem (6.11.15) is I2,1/q .
2. If the assumption 4 of Theorem 1 is valid, the inverse operator
1
(, )
Lq () W2,2
of the problem (6.11.6) is compact for any set provided q > 2n/(n + 1).
A necessary and sucient condition for the compactness of this operator
for q 2n/(n + 1) is the condition of Theorem 6.11.1/3 with p = r = 2.
In the case q = 2, Theorem 2 yields necessary and sucient conditions for
the discreteness of the spectrum of the problems
Lu = 0
in ,
Lu = u in ,
M u = u on ,
M u = 0 on .
An extension of the results in the present subsection to the mixed boundary

problem
Lu = f
on \E,
Mu =
on \E,
is a simple exercise.
where E is a subset of ,
u = 0 on E,
401

Section 6.1. Conductivity (i.e., 2-conductivity) was studied by P
olya and
Szeg
o [666]. This notion was applied to embedding theorems by the author [528]. Here the presentation follows the authors paper [547].
Sections 6.26.4. The content of these sections except Sect. 6.4.5 and
6.4.7 is taken from the authors paper [547].
In the paper by Amick [45] the decomposition of the space [L2 ()]N into
two orthogonal subspaces of solenoidal vector elds and of gradients of functions in W21 () are studied. This decomposition plays an important role in the
mathematical theory of viscous uids (cf. Ladyzhenskaya [474]). According to
the Amick theorem [45], this decomposition is possible for a bounded domain
if and only if the spaces W21 () and L12 () coincide. By Theorem 6.4.3/2 of
the present book, the last property is equivalent to the inclusion I2,1/2 .
In connection with Corollary 6.3.4 we note that Buckley and Koskela [147]
proved that L1p () L2p/(2p) (), 2 > p 1, where is a bounded simply
connected domain, only if is a John domain. The suciency of the last
condition is due to Bojarski [122]. In other words, I2,2p/(2p) if and only
if is a John domain.
Section 6.5. The result of this section for p = 2 was obtained by the
author [537].
Section 6.6 was rst published in the authors book [552].
Section 6.7. Most of this section (6.7.16.7.4 except Proposition 6.7.1/1)
is borrowed from the authors paper [547]. Proposition 6.7.1/1 as well as the
content of Section 6.7.5 follow the authors book [552].
Sections 6.86.9 are part of the authors paper [547].
Section 6.10 is partly contained in the authors paper [537]. Sections 6.10.2 and 6.10.4 were published in the authors book [552]. The equivalence of the Poincare inequality and the solvability of the Neumann problem
is well known. The same pertains to the interconnection of conditions for the
discreteness of the spectrum and the theorems on compactness (cf. Deny and
Lions [234], Necas [630], Lions and Magenes [500], and others).
Section 6.11 is borrowed from the authors book [552]. Payne and Weinberger [656] showed that the optimal constant C in the Poincare inequality
(6.11.14) with p = q = 2 and for convex domains, formulated in terms of diameter is equal to 1 diam . A sharp form of Corollary 6.11.1/1 with optimal
constants multiplying the quantities uLp () and uLp(n1)/(np) () has
recently been established in Maggi and Villani [512]. The conductivity c2 was
applied to boundary value problems for the Laplacian in very general domains
by Arendt and Warma [51], Biegert and Warma [97, 98] et al. (In these papers
the set function c2 is called relative capacity.)
There are a number of papers, where for special classes of domains (without the cone property or unbounded), manifolds, and metric spaces, theorems on the continuity and the compactness of the embedding operator
402
Wpl () Lq (), as well as necessary conditions for these properties are

proved: J.-L. Lions [499]; Stampacchia [719]; Bjorup [108]; Campanato [165];
Globenko [311]; Andersson [48, 49]; Hurd [388]; R.A. Adams [22, 23]; Edmunds [247]; Fraenkel [285]; Reshetnyak [676]; Goldshtein and Reshetnyak
[316]; Jerison [401]; Hurri [389, 390]; Martio [525]; Egnell, Pacella, and Tricarico [254]; Evans and Harris [262]; Hurri-Syrj
anen [391, 392]; Chua [189];
Evans and Harris [263]; Smith and Stegenga [708]; Stanoyevitch [721]; Stanoyevitch and Stegenga [722, 723]; Buckley and Koskela [147]; Maheux and SaloCoste [513]; Koskela and Stanoyevitch [460]; Labutin [470, 471, 473]; Hajlasz
and Koskela [343]; Ross [684]; Kilpel
ainen and Mal
y [422]; Besov [9092]; Edmunds and Hurri-Syrjanen [251]; Hajlasz [341]; Burenkov and Davies [157];
Koskela, Onninen. and Tyson [459]; Koskela and Onninen [458]; R.A. Adams
and Fournier [25]; Coulhon and Koskela [215]; Harjulehto and H
ast
o [352];
Mazya and Poborchi [574576]; Bj
orn and Shanmugalingam [107]; Martin
and M. Milman [523]; et al.
Let, for example, be a cusp domain of the form

= x = (y, z) Rn : z (0, 1), |y| < (z) ,
n 2,

(0) = limz0 (z) = 0. Poborchi and the author showed [449] that for 1 <
p q < , l = 1, 2, . . . , the space Wpl () is continuously embedded in Lq ()
if and only if the quantities A0 , A1 are nite, where A = supz(0,1) A (z),

1q
(z t)q(l1)(1) (t)n1 dt
A (z) =
0

n1
(t) 1p (t z)
p(l1)
p1
p1
p
dt
This embedding is compact if and only if limz+0 A0 (z) = limz+0 A1 (z) = 0.

The continuity of the embedding operator:
W1l () Lq (),
1 q < ,
is equivalent to the inequality sup{A(z) : z (0, 1)} < with

1/q
(z t)(l1)q (t)n1 dt
A(z) = (z)1n
0
and its compactness to limz+0 A(z) = 0.

Furthermore, the inequality
1 (l1)p/(p1)
z
dz < ,
(n1)/(p1)
0 (z)
p (1, )
403
is necessary and sucient for the continuity of the embedding Wpl ()

C() L (), and this embedding is automatically compact. The space
W1l () is continuously embedded in C() L () if and only if

sup z l1 (z)1n : z (0, 1) < ,
and this embedding is compact if and only if
lim z l1 (z)1n = 0.
z+0
Recently a great deal of attention has been paid to the Sobolev-type embeddings for the so-called -John domains.
Let 1. A bounded domain Rn is -John, if there is a constant
C > 0 and a distinguished point x0 such that every x can be joined
to x0 by a rectiable arc along which

dist(y, ) C (x, y) ,
y ,
where |(x, y)| is the length of the portion of joining x to y. Clearly the
class of -John domains increases with and coincides with the class of John
domains for = 1.
Hajlasz and Koskela [344] found an exact exponent q for p = 1 and almost
exact q for p > 1 providing the continuity of the embedding L1p () Lq ()
for a -John domain. This embedding with exact exponent q was established
by Kilpel
ainen and Mal
y [422]. It was shown in the last work that for a -John
domain the following estimate holds:

1/q
|u t| (x) dx
q
inf
t(,+)

1/p
|u| (x) dx
,
C = const,
where
(x) = dist(x, ),
b 1 n,
a > n,
1 p q < ,
q(n p) np,

(n + a)q 1 (n + b 1) + 1 p1 1,
and u is an arbitrary function on , such that b/p u Lp ().
This result was generalized by Besov [91] to weighted Sobolev spaces of
any order. In particular, Besov [91] proved the continuity of the embedding
Llp () Lq (), where Rn is a -John domain,
1 < p < q < ,

l 1 + (n 1) /p + n/q 0.
The exponent q cannot be made greater (see Labutin [473]).
404
Embeddings of the anisotropic space Wpl () for special classes of domains are considered in the book by Besov, Ilin, and Nikolsky ([94], Sect. 12,
Chap. 3), and in the papers by Besov [88, 89] and Trushin [765, 766].
An interesting approach to the study of boundary value problems and
Sobolev spaces for domains with fractal boundaries was proposed by Davies
[223]. He showed that properties of functions due to the irregularities of the
boundary are the same as those due to the singularities of the coecients
of some Riemannian metrics on the unit ball. For instance, there exists a
Lipschitz equivalence between the Koch snowake domain with the Euclidean
metric and the unit disk D with the metric

2 2
dx1 + dx22 ,
ds2 = 1 |x|
where
dim(D) = (1 )1 =
(see Theorem 5.1 in [223]).
log 4
log 3
7
Continuity and Boundedness of Functions
in Sobolev Spaces
If a domain has the cone property, then by the Sobolev theorem any function
u in Wpl (), pl > n, coincides almost everywhere with a continuous function
in , and
uL () CuWpl () ,
where the constant C does not depend on u.
A simple example of the function u(x) = x1 , > 0, dened on the plane
domain = {x : 0 < x1 < 1, 0 < x2 < x1 }, > 1, shows that the cone
property is essential for the validity of Sobolevs theorem. We can naturally
expect that for sets with bad boundaries the embedding Wpl () L ()
C() is valid in some cases under stronger requirements on p and l.
In this chapter we study the classes of domains for which the embedding
operator of Wpl () into L ()C() is bounded or compact. Some theorems
we prove give necessary and sucient conditions ensuring the continuity of
these operators and related integral inequalities (see Sects. 7.17.3). These
criteria are formulated in terms of the p-conductivity.
To be more precise, let y and let (y) be the intersection of with
the ball B (y). For any conductor of the form (y)\{y} we introduce the
function

p () = inf cp (y)\{y}
y
and show that the embedding operator Wp1 () C()L () is continuous

if and only if p () is not identically zero.
Inequalities for functions with derivatives in BirnbaumOrlicz spaces in
nonsmooth domains are studied in Sect. 7.4. Compactness of the embedding
Wp1 () C() L () is fully characterized in Sect. 7.5 by the condition
lim p () = as +0. In that section we show by a counterexample that
one should not expect that for domains with bad boundaries the simultaneous
fulllment of both continuity and compactness of this embedding operator
always takes place.

DOI 10.1007/978-3-642-15564-2 7,
405
406
7 Continuity and Boundedness of Functions in Sobolev Spaces
More visible sucient conditions for continuity and compactness of the

above-mentioned embedding operator, given in terms of the isoperimetric
function (), can be found in Sects. 7.1 and 7.5.
The continuity and compactness criteria are generalized to Sobolev spaces
of an arbitrary integer order in Sect. 7.6 with the help of the so-called (p, l)conductivity.
7.1 The Embedding Wp1 () C() L()

7.1.1 Criteria for Continuity of Embedding Operators of Wp1 ()
and L1p () into C() L ()
Let y be an arbitrary point in the domain and let > 0.
Here and in the next two subsections we consider only conductors of the
form (y)\{y}. Further, we introduce the function

(7.1.1)
p () = inf cp (y)\{y} , p > n,
y
on (0, +), where cp is the p-conductivity introduced in Chap. 6. Obviously,

p does not increase and vanishes for > diam(). The condition p > n in the
denition of p is justied by the fact that the inmum on the right in (7.1.1)
equals zero for p n by (2.2.13).
Noting that the function u(x) = (1 1 |x y|)+ is contained in the class
U ( (y)\{y}) we obtain p () cnp .
Theorem 1. The embedding operator of Wp1 () into C() L () is
bounded if and only if p 0.
Proof. Suciency. Let u be any function in C () Wp1 () and let y be
a point in such that u(y) = 0. Let R denote a positive number for which
p (R) > 0 and let denote an arbitrary number in (0, R]. We put

v(x) = (x y)/ u(x)/u(y),
where C0 (B1 ), (0) = 1. Since v(y) = 1 and v(x) = 0 outside (y), we
have

|v|p dx,
cp (y)\{y}
and therefore

u(y)p cp p (y)\{y} c

|u|p dx + p
(y)
Thus the suciency of p 0 follows.

(y)

|u|p dx .
(7.1.2)
7.1 The Embedding Wp1 () C() L ()
407
Necessity. For all u Wp1 (), let

uL () CuWp1 () .
(7.1.3)
Inserting an arbitrary u T ( (y)\{y}) into (7.1.3) we obtain

1 C uLp () + vn1/p n/p .
If is small enough then
(2C)p

|u|p dx.
Minimizing the preceding integral over T ( (y)\{y}) we obtain

cp (y)\{y} (2C)p .

The next assertion similar to Theorem 1 holds for the space L1p ().
Theorem 2. Let mn () < . The embedding operator of L1p () into

C() L () is bounded if and only if p 0.
Proof. We only need to prove the suciency of p 0. By virtue of
Lemma 5.1.2/2 we need to derive the inequality
uL () CuL1p ()
(7.1.4)
for functions in L1p () L (). Let denote a bounded set with

.
The estimate (7.1.3) implies

1/p

uL () C uLp () + uL () mn (\)
.
Choosing to satisfy 2C(mn (\))1/p < 1, we arrive at (7.1.4). The theorem
is proved.

l () denote the completions of the spaces
lp () and W
Remark 1. Let L
p
l
W l () with respect to the

C () C() Lp () and C () C()
p
l
l
norms in Lp () and Wp ().
If we replace p () in Theorems 1 and 2 by

p () = inf cp (y)\{y} ,
where cp is the p-conductivity dened by (6.11.5), then we obtain analogous

l () and L
l ().
assertions for the spaces W
p
p
The following theorem contains two-sided estimates for the constants in
inequality (7.1.5).
408
Let p () denote the inmum of cp (K) taken over the set of conductors
K = G\F in with mn (G) .
Since cp (K) is a nondecreasing function of F , we may assume F to be a
point.
Theorem 3. 1. If p () = 0 for some < mn (), then for all u
L1p () Lq (), p > n, 0 < q < ,
uL () k1 uLp () + k2 uLq () ,
(7.1.5)
where k1 [p ()]1/p , k2 1/q .

2. If for any u L1p () Lq () inequality (7.1.5) holds, then p ()
(2k1 )p with = (2k2 )q .
Proof. 1. It suces to derive (7.1.5) for all functions in C () L1p ()
Lq (). We choose a positive number t such that

mn {x : u(x) > t} ,
mn x : u(x) t .
Let T > t and {x : |u(x)| T } = . By the denition of p-conductivity, for
the conductor

Kt,T = x : u(x) > t \ x : u(x) T
we have
(T t) cp (Kt,T )
p

|u|p dx =
Consequently,

|u|p dx.

(T t)p p ()
|u|p dx.
Hence
1/p
uLp () + 1/q uLq ({x:|u(x)|t})

T p ()

2. Suppose (7.1.5) is valid. We put = (2k2 )q and consider an arbitrary
conductor K = G\F with mn (G) . Let {um } be a sequence of functions
in T (K) such that
um pLp () cp (K).
Clearly,
1/q
1
k2 um Lq () k2 mn (G)
k2 1/q = .
2
Moreover, by (7.1.5) we have
1/p
.
1 2k1 um Lp (K) 2k1 cp (K)
Consequently, p () (2k1 )p . The theorem is proved.
Remark 2. Theorems 1 and 3 imply that the conditions p 0 and p 0

are equivalent.
409
7.1.2 Sucient Condition in Terms of the Isoperimetric Function

for the Embedding Wp1 () C() L ()
From Corollary 6.1.3/2 and the denition of the functions p and M it immediately follows that
1p

d
p ()
,
(7.1.6)
p/(p1)
0 [M ( )]
where M , which together with Theorem 7.1.1/3 yields the following
sucient condition for the embedding Wp1 () C() L ().
Theorem. If for some M < mn ()
M
0
d
< ,
[M ()]p/(p1)
(7.1.7)
then the embedding operator of Wp1 () into C() L () is bounded.

This implies the following obvious corollary.
Corollary. If J and p(1 ) > 1, then the embedding operator of
Wp1 () into C() L () is bounded.
Example. Consider the domain

= x : (x21 + + x2n1 )1/2 < f (xn ), 0 < xn < a ,
(7.1.8)
in Example 5.3.3/1. From (5.3.5) it follows that the convergence of the integral (7.1.7) is equivalent to the condition
a
d
< .
(7.1.9)
(n1)/(p1)
[f
(
)]
0
We show that p () 0, i.e., Wp1 () is not embedded in C() L ()
if (7.1.9) fails. Let
F = {0 < xn },
G = {x : 0 < xn < },
where > and K is the conductor G\F . We introduce the function u

U (K) which vanishes outside G, and is equal to unity on F and to

d
[f ()](n1)/(p1)
xn
1
d
[f ()](n1)/(p1)
on G\F . Obviously,

cp (K)
|u| dx = c
p
d
[f ()](n1)/(p1)
1p
.
(7.1.10)
410
Therefore, cp (K) 0 as 0 and hence p 0 provided (7.1.9) diverges.

If f ( ) = c , 1, then is contained in J with = (n 1)/((n
1)+1) (cf., for instance, Example 5.3.3/1) and consequently Wp1 () C()
L () for p > 1 + (n 1).
The condition (7.1.9) implies that the embedding operator of Wp1 () into
C() L () is bounded for all p > n if for small the function f is dened
as

f ( ) = h log(1/ ) ,
where h(s) 0 and s1 log h(s) 0 as s +. (Of course, this domain
does not have the cone property.)
If for any point P in we can construct a quasiconic body situated in
and specied in some coordinate system with the origin at P by (7.1.8)
where f is subject to (7.1.9), then, obviously, p () 0 and the embedding
operator of Wp1 () into C() L () is bounded.
7.1.3 Isoperimetric Function and a BrezisGallou
etWainger-Type
Inequality
Theorem. Let be a domain with mn () < and let
mn ()/2
0
d
< ,
[()]p
where p + p = pp and () denotes M () for M = mn ()/2. Furthermore,

let for some r (1, p)
mn ()/2
0
d
= .
[()]r
Then for any (0, mn ()/2) and u L1p ()

osc u c(p, r)

d
[()]p
mn ()/2
1/p
d
[()]r
uLp ()
1/r

uLr () .
Proof. Let the numbers T and t be chosen so that

mn x : u(x) > T mn ()/2 mn x : u(x) T ,

mn x : u(x) > t mn x : u(x) t .
Furthermore, let S := ess sup u. Then
(7.1.11)
411
(S t)p cp (Kt,S )
|u|p dx
and
(t T ) cr (KT,t )
r
|u|r dx.
Now it follows from Corollary 6.1.3/2, that

1p

d
cp (Kt,S )
p
0 [()]
and

mn ()/2
cr (KT,t )
This leads to the estimate

ess sup u T
d
[()]r
1r
.
1/p
d
uLp (uT )
p
0 [()]
1/r
mn ()/2
d
+
uLr (uT ) .
[()]r
An analogous estimate for T ess inf u is proved in the same way. Adding
both estimates we arrive at (7.1.11).

We specify this theorem for domains in the class J1/r .
Corollary. Let p > r > 1, mn () < and let J1/r , i.e.,
() C1/r .
Then

1/r

mn ()
(pr)/pr
osc u c0
uLp () + log
uLr () ,
(7.1.12)
where (0, mn ()/2) and c0 depends only on C, p, and r.

Remark. Minimizing the right-hand side of (7.1.12) in we see that

1/r
,
osc u c1 1 + log c2 uLp ()
(7.1.13)
provided
uLr () = 1.
For the cusp (7.1.8) with f (xn ) = xn , > 1, we have by (5.3.7) that
() (n1)/((n1)+1) ,
and we may take r = (n 1) + 1.
412
This example can be used to show that the exponent 1/r of the power of
the logarithm in (7.1.12) is sharp by setting
u(x) =
log(xn + )1
(log 1 )1/[1+(n1)]
with a small > 0 into (7.1.12).
7.2 Multiplicative Estimate for Modulus of a Function

in Wp1 ()
7.2.1 Conditions for Validity of a Multiplicative Inequality
It is well known that the estimate
n/p
1n/p
uL () CuW 1 () uLp ()

p
(7.2.1)
holds for p > n provided the domain has the cone property. This is a particular case of the general multiplicative GagliardoNirenberg inequalities (cf.
Sects. 1.4.7, 1.4.8). The following theorem contains a necessary and sucient
condition for validity of the estimate
1/(r+1)
r/(r+1)
uL () CuW 1 () uLp () ,

p
(7.2.2)
where r is a positive number.

Theorem. If for some r > 0
lim inf r p () > 0,
+0
(7.2.3)
then (7.2.2) holds for all u Wp1 ().

Conversely, if (7.2.2) holds for all u Wp1 (), then satises (7.2.3).
Proof. Suciency. By (7.2.3) there exists a constant M such that
r p () = const > 0
for (0, M ]. Therefore, by Theorem 7.1.1/3,
uL () r/p 1/p uLp () + 1/p uLp () .
The minimum value of the right-hand side over is attained for
p/(r+1)

= 1/p r 1 uLp () /uLp ()
and it is equal to
(7.2.4)
7.2 Multiplicative Estimate for Modulus of a Function in Wp1 ()
413
c 1/p(r+1) uLp () uLp () .

1/(r+1)
r/(r+1)
If M then (7.2.2) follows. If > M , then

1/p r1 uLp () M 1+1/r uLp () ,
and (7.2.4) implies
uL () c1 1/p(r+1) uLp () uLp () + cM 1/p uLp () .
1/(r+1)
r/(r+1)
The inequality (7.2.2) is proved.

Necessity. We put M = (2 C r+1 )p/(r+1) . We may assume that the constant C in (7.2.2) is so large that M < mn (). Consider an arbitrary conductor K = G\F with mn (G) M . From (7.2.2), for any u T (K), we
have

1/(r+1)
1 C1/p(r+1) uLp () + M 1/p
.
Therefore, 2C r+1 r/p uLp () 1. Minimizing the left-hand side over
T (K) we obtain
r+1 p r
p () 1.
2C
Theorems 1 and 7.1.2 imply the following sucient condition for the validity of inequality (7.2.2).
Corollary. If J with 1 > 1 1/n and p(1 ) > 1, then for
any u Wp1 () inequality (7.2.2) is valid with r = p(1 ) 1.
Proof. Since J , then M () CM for < M where CM is a
positive constant. This and (7.1.6) imply
p1 1p(1)
p
p () CM
1 p/(p 1)
.
The result follows.
Example. For f (xn ) = c xn , 1, the domain (7.1.8) is in the class
J(n1)/((n1)+1) and hence by the Corollary inequality (7.2.2) is valid for
p > 1 + (n 1) with r = (p 1 (n 1))/(1 + (n 1)). This exponent
is the best possible since (7.1.6) implies

p (n1)+1 c1

(n1)/(p1) d
1p
where is any suciently small positive number.
= c2 1+(n1)p ,
414
7.2.2 Multiplicative Inequality in the Limit Case r = (p n)/n

Inequality (7.2.2) becomes (7.2.1) for r = (p n)/n. In this particular case a
necessary and sucient condition can be expressed in terms of the function
p dened by (7.1.1).
Theorem. The inequality (7.2.1) holds for all u Wp1 () if and only if
lim inf pn p () > 0.
+0
(7.2.5)
Proof. Suciency. Let r be so small that pn p () > > 0 for < r. By
virtue of (7.1.2) we have
c 1/p uL () 1n/p uLp () + n/p uLp ()
(7.2.6)
for r with c > 0. The minimum of the right-hand side in (7.2.6) over
> 0 is attained at
= n(p n)1 uLp () /uLp ()
and is equal to
n/p
1n/p
cuLp () uLp () .
If r then (7.2.1) follows. If > r then
uLp () (p n)r n1 uLp ()
and (7.2.6) implies
c 1/p uL () uLp () uLp () + r n/p uLp () .
n/p
1n/p
Thus the suciency of the condition (7.2.5) is proved.

Necessity. We insert an arbitrary u T ( (y)\{y}) into (7.2.1). Since

upLp () cn ,
upW 1 () c cp (y)\{y} + n ,
p
by (7.2.1) we have

n/p n(pn)/p

.
C p c cp (y)\{y} + n
Consequently,
C p
/n

c pn p () + p .
It remains to pass to the lower limit as +0.
In the next proposition we give a sucient condition for (7.2.6) which

generalizes the cone property. Let y and let S (y) denote the sector
7.2 Multiplicative Estimate for Modulus of a Function in Wp1 ()
415
Fig. 29.
{x : |x y| < , |x y|1 (x y) (y)}, where (y) is a measurable subset

of the (n 1)-dimensional unit sphere.
Proposition. Suppose that there exist positive constants R and such
that any point y in the set can be placed at the vertex of the sector SR (y)
contained in , satisfying the condition s((y)) > . Then (7.2.5) holds.
Proof. Let 0 < < R and let (r, ) be spherical coordinates centered at y.
Obviously,

cp (y)\{y} inf
|u|p dx,
S (y)
where the inmum is taken over all functions u C 0,1 (S (y)) with u(y) = 1,
u(, ) = 0 for (y). It remains to note that

p
u n1
r
|u|p dx
d
dr

0 r
S (y)
(y)
p

1p

u

dr
d
r (1n)/(p1) dr
0 r
0
(y)

p1
pn
>
np .

p1
We shall consider a domain that does not satisfy the condition of Proposition and for which (7.2.5) is nevertheless true.
Example. Let be the domain in Fig. 29. Further, let m = 2m , Qm =
{x : m+1 < |x| < m }, y Qm . Let u denote a function in T ( (y)\{y})
such that

|u|p dx cp (y)\{y} + , > 0.
416
We note that

u() u()p c| |p2

|u|p dx,
(7.2.7)
Qj
for any points , Qj , j = 1, 2, . . . . (This estimate is invariant with respect

to a similarity transformation and so it suces to limit consideration to Q1 .
However, inequality (7.2.7) for Q1 is contained in Theorem 1.4.5, part (f).)
We begin with the case < m when Qm B (y) = . Let j = m, = y
and Qm B (y) in (7.2.7). Then

1 cp2 cp (y)\{y} + .
Next suppose m . For all Qj {x = (x1 , x2 ) : x2 < 0} we have

u() u(0)p c||p2
|u|p dx.
Using (7.2.7), we nd that the last inequality holds for all Qj . Consequently,

p
osc u cp2
|u|p dx.
2 (0)
Noting that 2 (0) (y), we nally obtain

p

1 = osc u cp2 cp (y)\{y} + .
(y)
7.3 Continuity Modulus of Functions in L1p()

The following assertion is an obvious corollary of the denition of p-conductivity.
Theorem. Let mn () < , be a nondecreasing continuous function on
[0, ) and let u be an arbitrary function in L1p (). In order that for almost
all x, y the inequality

u(x) u(y) |x y| uL ()
(7.3.1)
p
be valid it is necessary and sucient that

p
cp (K) dist(i F, i G)
(7.3.2)
for any conductor K = G\F .

Since the conductivity is a nonincreasing function of G and nondecreasing
function of F , the last condition is equivalent to
7.3 Continuity Modulus of Functions in L1p ()

p
cp \{x} \{y} |x y|
,
x, y .
417
(7.3.3)
We say that the class u = {u + const} is contained in the space C () if

sup
x,y
|u(x) u(y)|
< .
(|x y|)
Thus the embedding operator of L 1p () into C () is continuous if and

only if (7.3.3) holds.
Example. Further in this section we consider the domain already studied
in Examples 5.3.3/1, 5.5.2, 6.3.6/1, and 7.1.2. Here we show that for this
domain the embedding operator of L1p () into C () is continuous if and
only if

(t) k
0
11/p
[f ()](n1)/(p1)
k = const > 0.
(7.3.4)
Proof. Necessity. We insert the function u equal to unity for xn < , to

zero for xn > t and to
t
1
t
d
d
(n1)/(p1)
(n1)/(p1)
xn [f ()]
[f ()]
for xn t into (7.3.1). (Here > 0 and t (, a).) Then

1 vn1 (t)
p1
d
[f ()](n1)/(p1)
which becomes (7.3.4) as +0.

The suciency of (7.3.4) is a simple corollary of the inequality

u(x) u(0) k

xn
0
d
[f ()](n1)/(p1)
11/p
uLp () .
(7.3.5)
in L1 () for
(We note that Theorem 1.1.6/1 implies the density of C ()
p
the domain under consideration.)
To prove (7.3.5), we need the following lemma.
Lemma. Let

b = x = (x , xn ) : |x | < f (xn ), 0 < xn < b
b ) with u(0) = 0 and u(x) 1 for xn = b.
and let u be a function in C (
Then
1p
b

d
p
|u| dx k
.
(7.3.6)
(n1)/(p1)
b
0 [f ()]
418
Proof. It suces to establish (7.3.6) under the assumption that u = 0 for

xn < , u = 1 for xn > b where is a small positive number. Then

|u|p dx cp (K ),
b
where K = G \F , F = clos , G = b . To estimate cp (K ) from

below we make use of (7.1.6) and (5.3.5). These inequalities are applicable
despite the fact that the measure of the set G is large. In fact, extending f
to [b, 2b] we obtain the enlarged domain 2b such that 2mn (G ) mn (2b )
with no modication of the conductor K . We have
1p
b
d
,
(7.3.7)
cp (K ) k
[f ()](n1)/(p1)
which together with (7.3.6) completes the proof of the lemma.
Proof of the inequality (7.3.5). First we note that smooth functions in the
closure of the domain

1

gx = y : xn 2f (a) f (xn ) < yn < xn , xn < a

u(z) u(y) C|z y|1n/p uL (g ) ,
p
x
(7.3.8)
where z, y are arbitrary points in gx and C is a constant that is independent

of x. The latter is a corollary of the Sobolev theorem on the embedding of L1p
into C 1n/p for domains with smooth boundaries.
a ), u(0) = 0, u(x) = 1 at some x
a . By (7.3.8),
Let u C (

p
np

C f (xn )
max 1 u(y)
|u|p dx.
y
gx
Therefore,

|u|p dx Cxnp
,
n
a
provided 2 min u < 1 in gx . This and the obvious estimate

xn
0
d
[f ()](n1)/(p1)
p1

k
p1
xn
(1n)/(p1) d
= k1 xnpn
imply (7.3.5).
Next we assume that 2u(y) 1 for all y gx . Then the function 2u
satises the conditions of the Lemma with
1
b = xn 2f (a) f (xn ),
7.4 Boundedness of Functions with Derivatives in BirnbaumOrlicz Spaces
419
and thus

|u| dx k

[f ()](n1)/(p1)
xn
k
0
1p
d
[f ()](n1)/(p1)
1p
.
Inequality (7.3.5) follows.
7.4 Boundedness of Functions with Derivatives

in BirnbaumOrlicz Spaces
Most of the results of the previous sections in this chapter can be generalized
to the space of functions with nite integral

|u| dx,
(7.4.1)
where is a convex function. For this purpose we must introduce the conductivity generated by the integral (7.4.1).
Here we consider only a sucient condition for the boundedness of functions with nite integral (7.3.1) which is formulated in terms of the function .
We also state some corollaries of this condition.
Lemma. If u C (), then, for almost all t,

ds
d
= mn (Lt ),
|u|
dt
Et
(7.4.2)
where Et = {x : u(x) = t} and Lt = {x : u(x) > t}.

Proof. The equality (7.4.2) follows from the identity

ds
dx =
d
,
u>t
E |u|
t
which in turn results from Theorem 1.2.4.
Theorem. Let be a convex nonnegative function with (0) = 0 and let

be the complementary function of (cf. Sect. 2.3.3). If has nite volume
and

mn ()/2
(1/()) d < ,
(7.4.3)
then any function u C () with the nite integral (7.4.1) is bounded.
420
Proof. Let denote a number such that

2mn (N ) mn (),
2mn (L ) mn (),
where N = {x : u(x) }. We introduce the notation

m(t) = mn (Lt ),
h(t) = s(Et ).
By the inequality () + () with , > 0, for u(x) we have

h(t) m (t)
dt
m (t) h(t)

u(x)
u(x)
h(t)
1

m
m (t) dt.
(t)
dt
(t)
m
h(t)
u(x)
u(x) =
Using (7.4.2) together with Jensens inequality, we obtain

dt
h(t)
ds
1

m
(t)
=
|u|
ds

m (t)
|u| Et |u|
Et |u| Et

ds
.
|u|
|u|
Et
Consequently,
u(x)

+
|u| dx +
mn ()/2

1/() d.
A similar estimate is valid for ( u(x))+ . Therefore, u is bounded and

|u| dx + 2
osc u
mn ()/2

1/() d.
Corollary. If J , < 1, and

(t)t11/ dt < ,
1
then any function u C () with the nite integral (7.4.1) is bounded.

In particular, u L () if

|u|1/(1)
m

/(1)
logk+ |u|
r
m+1
log+ |u| dx < ,
(7.4.4)
k=1
where m 0, r > /(1 ) and logk+ is the k-times iterated log+ . (For m = 0
the expression in the rst parentheses in (7.4.4) is absent.)
7.4 Boundedness of Functions with Derivatives in BirnbaumOrlicz Spaces
421
Proof. The rst assertion follows from the theorem just proved and from
the denition of J .
To prove the second assertion we must use the fact that the convex function
1
m

m+1 r(1)/
logk t
t
(t) = t1/
log
k=1
is equivalent to the complementary function of

(1 )t
1/(1)
m

/(1)
k
log t
r
logm+1 t
k=1
for large t (see Krasnoselski and Ruticki [168]).
The Sobolev theorem on the embedding Wp1 () C() L () for

p > n can be rened for domains having the cone property on the basis of
the previous Corollary. Namely, if J11/n , then the continuity and the
boundedness of functions in result from the convergence of the integral
n1
m

m+1
n1+
k
n
log+ |u|
|u|
log+ |u|
dx, > 0.
k=1
We shall show that one cannot put r = /(1 ) in (7.4.4).

Example. Consider the domain in Examples 5.3.3/1 and 5.5.2. By
(5.3.5), the condition (7.4.3) is equivalent to
1
n1
1n
f ( )
f ( )
d < .
0
Let f ( ) = c , 1. As already noted, J with = (n

x1
1)/((n 1) + 1). The function u(x) = logm+3
+
n , m 0, is unbounded
in . On the other hand, for small xn > 0 we have
m+1
(n1)

k
(n1)+1
log+ |u|
|u|
k=1

1 2 1 1
log x1
log xn
cx(n1)1
n
n
m+1 1 1 m+2 1 (n1)1
log
xn
xn
.
log
Therefore,

|u|
1/(1)
m+1

k=1
/(1)
logk+
|u|
dx < .
422

Wp1 () C() L()
7.5.1 Compactness Criterion
Let p be the function dened by (7.1.1) and let be a domain with nite
volume.
Theorem. The condition
lim p () =
+0
(7.5.1)
is necessary and sucient for the compactness of the embedding operator of

Wp1 () into C() L ().
Proof. Suciency. From (7.1.2) for small > 0 it follows that
1
upL () c p () upLp () + C()upLp () ,
where C() < for each > 0. We x a small number > 0 and we denote
an open set such that
, 2C() mn (\ ) < 1 by . Then
1
upL () 2c p () upLp () + 2C()upLp ( ) .
Consider the unit ball in Wp1 () and select a sequence {um } in this ball
that converges in Lp ( ). Then
1
lim sup uk ul pL () 2c p () .
(7.5.2)
k,l
Taking into account that p () as 0 and passing to the subsequence

{umk }, we obtain a sequence convergent in L () C().
Necessity. Let the embedding operator of Wp1 () into C() L () be
compact and let

(7.5.3)
p () = inf cp (y)\{y} < A.
y
We construct a sequence k that converges to zero and a sequence of points

yk such that

(7.5.4)
cp k (yk )\{yk } < A, k = 1, 2, . . . .
Since

lim cp B (y)\{y} =
o
for p > n (cf. Sect. 2.2.4), the limit points of the sequence {yk } are located
on . From (7.5.4) it follows that there exists a sequence of functions uk
7.5 Compactness of the Embedding Wp1 () C() L ()
T (k (yk )\{yk }) with
423

|uk |p dx < A.
Since 0 uk 1, this sequence is bounded in Wp1 () and hence it is

compact in C() L (). Therefore, given any > 0 we can nd a number
N such that um uk L () < for all m, k N . In particular, |uN (yN )
uk (yN )| < for all k > N . On the other hand, since yN is not a limit point of
{yk } and uk (x) = 0 outside k , uN (yN ) = 1, we have |uN (yN ) uk (yN )| = 1
for suciently large k. Thus assumption (7.5.3) is false. The theorem is proved.

Remark. Replacing p () by p () dened in Remark 7.1.1/1 in the last
theorem, we obtain a necessary and sucient condition for the compact We actually proved in the theorem
p1 () C().
ness of the embedding W
that (7.5.1) is necessary and sucient for the compactness of the embedding
L1p () C() L ().
7.5.2 Sucient Condition for the Compactness in Terms of the
Isoperimetric Function
Theorem. If the integral (7.1.7) converges for some M , then the embedding
operator of Wp1 () into C() L () is compact.
Proof. The denition of the function p implies

cp (y)\{y} p mn (y)
for all y . This and (7.1.6) yield

cp (y)\{y}

mn ( (y))
0
d
[M ( )]p/(p1)
1p
.
Since mn ( (y)) vn n , from the denition of p we obtain

vn n
p ()
0
d
[M ( )]p/(p1)
1p
.
Now the required assertion follows from Theorem 7.5.1.

Example. In Example 7.1.2 we noted that for

= x : |x | < f (xn ), 0 < xn < a
condition (7.1.7) is equivalent to convergence of the integral (7.1.9). Therefore
for such the embedding operator of Wp1 () into C() L () is compact
if and only if (7.1.9) holds.
424
7.5.3 Domain for Which the Embedding Operator of Wp1 () into

C() L () is Bounded but not Compact
The embedding operator of Wp1 () into C() L () is simultaneously
bounded and compact for the domain in Example 7.5.2. By Theorems 1.4.5
and 1.4.6/2 this is also valid if has the cone property.
The situation can be dierent with domains having bad boundaries.
As an example we shall consider a domain for which the function p is not
identically zero and is bounded. Theorems 7.1.2 and 7.5.1 imply that for such
a domain the embedding operator of Wp1 () into C() L () is bounded
without being compact.
Example. First we show that the domain depicted in Fig. 30 satises
p () 0 for p > 2. Consider the conductor (y)\{y} where is small
enough and y . If y Q then

cp Q (y)\{y} c2p ,
where Q (y) = B (y) Q (cf. Proposition 7.2.2) and hence

cp (y)\{y} c2p .
(7.5.5)
Let y be in the rectangle Rm and let G = B (y) (Q Rm ). By the

denition of p-conductivity,

cp (y)\{y} cp G\{y} .
(7.5.6)
Take an arbitrary function u T (G\{y}). Let

Et = x : u(x) = t .
Nt = x : u(x) t ,
We only need to consider those levels t for which Et is a smooth curve. If

m2 (Nt ) 2pm , where p = p/(p 1), then
m2 (Nt Q) pm m2 (Nt Rm ).
Since Q J1/2 , in the case m2 (Nt ) 2pm we have
s(Et )
cm2 (Nt Q)
1
cm2 (Nt ).
2
(7.5.7)
Let m2 (Nt ) < 2pm . If the set Et contains a component connecting points
of the polygonal lines abc and def (Fig. 31), then we can easily see that
2s(Et ) s( Nt ) and by the isoperimetric inequality we have
1/2
2 1/2 m2 (Nt )
s( Nt ) + s(Et ) 3s(Et ).
Thus either
(7.5.8)
7.5 Compactness of the Embedding Wp1 () C() L ()
Fig. 30.
425
Fig. 31.
s(Et ) m
or
m2 (Nt )
1/2
c0 s(Et )
provided m2 (Nt ) < 2pm .

Next we proceed to the estimate of cp (G\{y}). By Corollary 6.1.3 we
obtain
1p
1

dt
d
.
(7.5.9)
cp G\{y} inf
m2 (Nt )
[s(Et )]p
0 dt
We express the integral on the right-hand side of (7.5.9) as the sum of integrals
over the sets A1 , A2 , A3 , where

A1 = t : m2 (Nt ) 2pm ,

A2 = t : s(Et ) m \A1 ,

1/2

c0 s(Et ) \A1 .
A3 = t : m2 (Nt )

A1
p/2 1

(p2)/2(p1)
c
dt
d
c1 m2 (G)
m2 (Nt )
/2
p
2
[m2 (Nt )]
0 dt
The integral over A2 admits the obvious estimate

d
m
p/(1p)
(N
)
dt
2,
2
t
m
A2
A2 dt
and the integral over A3 is estimated by (7.5.8) as follows:
426
A3
p
(p2)/2(p1)
d
dt
c2 m2 (G)
m2 (Nt )
.
/2
p
dt
[m2 (Nt )]
The last three inequalities and (7.5.9) lead to cp (G\{y}) const, which together with (7.5.6) and (7.5.5) yield the required result.
We show that the embedding operator of Wp1 () into C() L ()
is not compact. We dene the sequence of functions {um }m1 as follows:
1/(1p)
um (x1 , x2 ) = m
(x2 1) if (x1 , x2 ) Rm , um (x1 , x2 ) = 0 if (x1 , x2 ) Q.
These functions are uniformly bounded in Wp1 () since
um Lp () = 1,
um Lp () = c1/(p1)

.
m
However, um uk L () = 2 for m = k and the sequence {um } is not

compact in C() L ().
7.6 Generalizations to Sobolev Spaces of an Arbitrary

Integer Order
7.6.1 The (p, l)-Conductivity
Let G be an open subset of the set and let F be a subset of G that is closed
in . We dene the (p, l)-conductivity of the conductor G\F by
cp,l (G\F ) = inf l upLp () ,
(7.6.1)
where the inmum is taken over all functions u C () that are equal to
zero on \G and to unity on F .
Proposition 1. If pl > n, p > 1, or l n, p = 1, then
) Rnpl
cp,l (BR \B
(7.6.2)
for R > 2.

Proof. If R = 1, then (7.6.2) follows from the Sobolev inequality
uL (B1 ) cl uLp (B1 ) ,
u C0 (B1 ).
The general case can be reduced to R = 1 by a similarity transformation.

Proposition 2. If n = pl and p > 1, then
)
cp,l (BR \B
for R > 2.
1p

R
log

(7.6.3)
7.6 Generalizations to Sobolev Spaces of an Arbitrary Integer Order
427
We shall establish (7.6.3) in the proof of Proposition 13.1.2/2 in the following.

Proposition 3. If pl n, p > 1 or l < n, p = 1, then for R > 2,
) npl .
cp,l (BR \B
(7.6.4)
Proof. It is clear that

cp,l (BR \B
) cp,l (B2 \B ).
cp,l Rn \B
It suces to show that the rightmost and the leftmost of these functions are
equivalent to npl . A similarity transformation reduces the proof to the case
= 1 where the required assertion follows from the Sobolev inequality
uLpn/(npl) cl uLp ,
u C0 .
In the present section we shall consider only conductors of the form G\{y}
with y G. Propositions 2 and 3 along with the denition of the (p, l)conductivity imply that cp,l (G\{y}) is identically zero provided pl n, p > 1
or l < n, p = 1. By Proposition 1,

cp,l B (y)\{y} = cnpl , c = const > 0,
for pl > n, p > 1 or for l n, p = 1.
7.6.2 Embedding Llp () C() L ()
Theorem. Let be a domain in Rn . The embedding operator of Llp () into
C() L () is continuous if and only if

inf cp,l (\
)\{y} > 0
(7.6.5)
y\
for some open set with compact closure

.
Proof. Suciency. Let be a bounded open set with smooth boundary
and such that
, . Let denote a function in C () which is
equal to unity outside and to zero on . Further, let u be any function in
C () Llp ().
We x an arbitrary point y \ for which u(y) = 0 and put v(x) =
(x)u(x)/u(y), x . Since v(y) = 1 and v(x) = 0 outside the set G = \
,
we have

cp,l G\{y} l vpLp () .
Therefore,
428
l

u(y)p cp,l G\{y} c
l (u)p
L
p ()
cl upLp () + CupW l1 ( ) .

p
k=0
Thus

1

l upLp () + CupW l1 ( ) .
sup |u|p c inf cp,l G\{y}
\
yG
(7.6.6)
The estimate for |u| in follows from the Sobolev theorem on the embedding
of Wpl into C for domains with smooth boundaries.
Necessity. Let the inequality

(7.6.7)
sup |u| C l uLp () + uLp ()
hold for all u C () Llp (), where is a domain with compact closure
,
. Consider any conductor G\{y} where G = \
and y G. The
insertion of an arbitrary function u C () L1p (), equal to unity at y
and to zero outside G, into (7.6.7) yields
1 sup |u| cuLp () .
Minimizing the last norm, we obtain cp,l (G\{y}) C p .
7.6.3 Embedding Vpl () C() L ()

Now we present a direct extension of Theorem 7.1.2 to the space Vpl ().
Let y and let be a positive number. Consider the conductor
(y)\{y}. Further, let pl > n or l = n, p = 1 and
l

cp,l (y)\{y} = inf
k upLp () ,
(7.6.8)
k=1
where the inmum is taken over all innitely dierentiable functions in the
class Vpl () that are equal to zero in \B (y) and to one at y.
Theorem. The embedding operator of Vpl () into C() L () is continuous if and only if

inf cp,l (y)\{y} 0.
(7.6.9)
y
Proof. Suciency. Let u be an arbitrary function in C () Vpl () and

let y be such that u(y) = 0. Further, let

inf cp,l (y)\{y} > 0
y
429
for some and let C0 (B1 ). Consider the function

v(x) = (x y)/ u(x)/u(y).
Since v(y) = 1 and v(x) = 0 outside (y), we have
l

cp,l (y)\{y}
p(kl) k vpLp () .
k=1
Consequently,
l

u(y)p inf cp,l (y)\{y} c
p(kl) k upLp () .
y
k=0
Necessity. For all innitely dierentiable functions in Vpl (), let the inequality
l

sup |u| C
k uLp ()
(7.6.10)
k=0
hold. Consider any conductor (y)\{y} with y . We insert an arbitrary

function in the denition of cp,l ( (y)\{y}) into (7.6.10). Obviously,
1/p

uLp () vn n
sup |u|.
(y)
1/n
Therefore, if < vn
(2C)p/n , then
sup |u| 2C
(y)
and thus
1 2C
l

k uLp ( (y))
k=1
l

k uLp ( (y)) .
k=1
Minimizing the right-hand side over V ( (y)\{y}) we obtain

cp,l (y)\{y} c C p .
7.6.4 Compactness of the Embedding Llp () C() L ()

Now we present a necessary and sucient condition for the compactness of
the embedding Llp () C() L ().
430
Theorem. Let mn () < . The embedding operator of Llp () into

C() L () is compact if and only if

(7.6.11)
lim inf cp,l G \{y} =
yG

for some monotone sequence of bounded open sets { }1 such that
and . Here G = \
.
, . By (7.6.6)
Proof. Suciency. Let be an open set with
we have

1
sup |u|p c inf cp,l G \{y}
l upLp () + CupW l1 ( ) + sup |u|p .
yG
It remains to use the compactness of the embedding of Llp () into Wpl1 ( )

and into C( ) along with (7.6.11) (see the proof of suciency in Theorem 7.5.1).
Necessity. Suppose that

(7.6.12)
lim inf cp,l G \{y} < A = const
yG
for an increasing sequence of open sets { }1 . We endow Llp () with the

norm
uLlp () = l uLp () + uLp (1 ) .
By (7.6.12) there exist sequences {y } and {u }, u V (G \{y }), such
that l u Lp () < A1/p . Since the sequence {u }1 is compact in C()
L (), given > 0, there exists a number N such that sup |u u | <
for a , > N . Using u (y ) = 1, we obtain

u (y ) 1 < , , > N.
(7.6.13)
Further, since , the point y is contained in for a xed > N and
for all large enough . Therefore, u (y ) = 0 which contradicts (7.6.13). The
theorem is proved.

We note that we derived (7.6.11) in the proof of necessity
for
any
mono
tonic sequence of bounded open sets with
, = .
7.6.5 Sucient Conditions for the Continuity and the Compactness
of the Embedding Llp () C() L ()
We present a sucient condition for the boundedness and the compactness of
the embedding operator of Llp () into C()L () which generalizes (7.1.7).
Theorem 1. Let mn () < , p 1, l, and let l be a positive integer.
Also let J , where 1 > > (n 1)/n and pl(1 ) > 1. Then
uL () CuLlp () ,
431
(7.6.14)
and the embedding operator of Llp () into C() L () is compact.

Proof. By Theorem 7.1.2/1,

uL () C uLpl () + uLp () ,
(7.6.15)
where is an open set,

and C is a constant independent of u. Since
p(l 1)(1 ) < 1 and pl = p/[1 p(l 1)(1 )], we have by Corollary 6.9/1
that

n
n

u

u
u

C
+
xi
l1 xi
xi Lp ()
Lpl ()
Lp ()
i=1
i=1

C1 l uLp () + uLp () .
(7.6.16)
Combining (7.6.15) and (7.6.16) we arrive at (7.6.14).
The compactness of the embedding Llp () C() L () follows
from (7.6.16) and Theorem 7.5.2 in which p is replaced by p l and the condition (7.1.7) is replaced by J .
Theorem 2. If is a domain with nite volume contained in J with
1 > > (n 1)/n and p l(1 ) > 1, then for all u Wpl ()
1/(r+1)
r/(r+1)
uL () CuW l () uLp () ,

p
r = pl(1 ) 1.
For the proof it suces to use Corollary 7.2.1 with p replaced by p l and
then to apply (7.6.16).
Example. (Cusp) Let be the domain in Examples 5.3.3/1 and 7.1.2 with
f ( ) = c , 1. Then the condition J in Theorem 1 is equivalent
to pl > (n 1) + 1 and therefore the embedding operator of Llp () into
C() L () is compact provided p l > (n 1) + 1. If the inequality sign
is replaced here by equality then the operator fails to be bounded. In fact,
the function u(x) = log | log xn | is not in L () and belongs to Llp () for
p l = (n 1) + 1.
In Mazya and Poborchi [575] one can nd the following more general and
exhaustive result concerning the same cuspidal domain with Lipschitz and
increasing f on [0, 1], such that
f (0) = lim f (x) = 0.
x0
If p (1, ) and l = 1, 2, . . . , then the continuity of the embedding

operator Wpl () C() L () is equivalent to
1 (l1)p/(p1)
t
dt < .
(7.6.17)
(n1)/(p1)
0 f (t)
432
This embedding operator is compact. The space W1l () is embedded into

C() L () if and only if

l1
t
: t (0, 1) < ,
sup
f (t)n1
and the compactness of the same embedding operator is equivalent to
tl1
= 0.
t0 f (t)n1
lim
k (), l > 2k
7.6.6 Embedding Operators for the Space Wpl () W
p
pk () satises the
In Sect. 1.6 we showed that for l 2k the space Wpl () W
Sobolev-type theorems for arbitrary bounded domains. Here we consider the
case l > 2k where, according to Sect. 1.6.4, some additional requirements on
are necessary.
By Corollary 6.9/1 and Theorem 7.6.5/2, the inclusion J with
(n 1)/n < 1 implies the compactness of the embedding Wpl ()
Wqm () where q 1 = p1 (lm)(1) if 1 > p(lm)(1), q is an arbitrary
positive number if 1 = p(l m)(1 ) and q = if 1 < p(l m)(1 ).
In the following we show that in the case m 2k the preceding result is also
pk (). We present an example of a
the best possible for the space Wpl () W
k ()
domain J for which the continuity of the embedding Wpl () W
p
m
1
1
Wq () implies the inequality q p (l m)(1 ).
Let be the union of the semi-ball B = {x = (y, z) : z < 0, |x| < 2} and
the sequence of disjoint semi-ellipsoids

2 2
e+
z + i2 |y Oi |2 < 1 ,
i = x = (y, z) : z > 0, i
where 0 < y < 1, i = 2i1 , |O Oi | = 3 2i (see Fig. 32).
In e+
i we dene the function
k

wi (x) = 1 i2 z 2 i2 |y Oi |2 z/ ,
where is a smooth function on (0, +), (z) = 0 near z = 0 and = 1 on
the half-axis ( 12 , +). We assume that each wi , i = 1, 2, . . . , is extended to
\e+
i .
We can easily check that
is+(n1+)/q ,
s < 2k,
s wi Lq ()
2k(1)s+(n1+)/q , s 2k.
i
k () be continuously embedded into W m (), m
Let the space Wpl () W
p
q
2k. Then
433
Fig. 32.
m wi Lq () Cwi Wpl () ,

or equivalently,
(ml)+(1/p1/q)(n1+)
cC 1 .
Consequently, 1/q 1/p (l m)(1 ) for = (n 1)/(n 1 + ).

Next we show that J . We introduce the sets

e+ = = (, ) Rn : > 0, 2 + 2 < 1 ,

+ = e+ : < i1 .
Since the embedding operator of W11 (e+ ) into Ls (e+ ) with s = 1/ is continuous, for all v W11 (e+ ) we have

vL2 (e+ ) c vL1 (e+ ) + i1 vL1 ( ) .
Let u W11 (). Applying the last inequality to the function v() = u(i +
Oi , i ), we obtain

uL1 (e+ ) c uL1 (e+ ) + i1 uL1 ( + ) ,
i
where i+ = {x e+
i : z < i }. Consequently

1
i uL1 ( + ) + uL1 () + uL1 (B ) .
uLs () c
i
(7.6.18)
i=1
Let i denote the mirror image of i+ with respect to the plane z = 0. It is

clear that

i1
|u| dx
|u| dx + i1
|u| dx.
i+
i+ i
434

uLs () c uL1 () +
|u|

dx
.
|x|
Since the second term on the right does not exceed cuW11 (B ) we have
inf u cLs () cuL1 ()
cR1
and the inclusion J follows from Theorem 5.2.3.

Most of the results of this chapter were announced in Mazya [528] and published with proofs in [543].
The p-conductivity criterion of the embedding
Wp1 () C() L ()
(7.7.1)
in Theorem 7.1.1/1 was used by Labutin [473] to obtain the following necessary
and sucient conditions on functions G on R such that G u Wp1 () for
all u Wp1 ():
(i) G L (R) if the embedding (7.7.1) fails,
(ii) G L (R, loc) if the embedding (7.7.1) holds.
In the case r = n/(n 1), if the boundary is Lipschitz, the inequality
(7.1.13) is a particular case of the BrezisGallouetWainger inequality [141],
[146], which has important applications to the theory of nonlinear evolution
equations (see for instance, Chap. 13 in M.E. Taylor [747], Brezis and Gallouet
[141], et al.).
Buckley and Koskela [148] showed that the inequality (7.3.1) with (t) =
t12/p , p > 2, is equivalent to (1.5.9).
8
Localization Moduli of Sobolev Embeddings
for General Domains
In the present chapter we study noncompact embeddings

Ep,q () : L1p () Lq (),
where is a connected open set in Rn , n > 1, and 1 p q < . In
the opposite case p > q, the boundedness of Ep,q implies compactness (see
Remark 5.5.2 and Theorem 6.8.2/2), which makes this case of no interest for
us in this chapter. Here we dene new measures of noncompactness of Ep,q
and characterize their mutual relations as well as their relations with local
isoperimetric and isoconductivity constants. To describe our results we need
to introduce some notation that will be frequently used in this chapter.
We supply L1p () with the norm
uL1p () = uLp () + uLp () ,
where stands for the gradient and is a nonempty open set with compact
closure . It is standard that a change of leads to an equivalent norm
in L1p (). In this chapter we often omit in notations of spaces and norms
if it causes no ambiguity.
Among other things, we study the essential norm of the embedding operator Ep,q : L1p Lq , i.e., the number
essEp,q = inf Ep,q T L1p Lq ,
where the inmum is taken over all compact operators T : L1p Lq .
Another characteristic of Ep,q to be dealt with later in this chapter is
dened by
C(Ep,q ) = inf C,
where C is a positive constant such that there exist > 0 and K > 0 subjected
to the inequality
uLq () CuL1p () + KuL1 ( )
for all u L1p

DOI 10.1007/978-3-642-15564-2 8,
(8.0.1)
435
436
with
8 Localization Moduli of Sobolev Embeddings for General Domains

= x : dist(x, ) > .
Together with the norm Ep,q of the embedding operator Ep,q , its essential norm essEp,q and the number C(Ep,q ), we shall make use of two
numbers M1 (Ep,q ) and M2 (Ep,q ). The rst of them is dened by

uLq
M1 (Ep,q ) = lim sup
: u L1p , u = 0 on s .
s0
uL1p
The denition of M2 is as follows:

uLq
1
: u Lp , supp u B(x, ) ,
M2 (Ep,q ) = lim sup sup
0 x
uL1p
where B(x, ) is the open ball with radius centered at x.
These two characteristics of Ep,q dier in the ways of localization of the
functions involved and it seems appropriate to call M1 and M2 the localization
moduli of the embedding Ep,q .
In Sect. 8.1 we show that
essEp,q = C(Ep,q ) = M1 (Ep,q )
(8.0.2)
provided 1 p q < pn/(n p) if n > p and 1 q < for p n. We

also prove that the three quantities in (8.0.2) are equal to M2 (Ep,q ) under
the additional assumption p < q. The last fact fails if p = q as shown in an
example of a domain for which M2 (Ep,q ) = 0 and
Ep,p = essEp,p = M1 (Ep,p ) = C(Ep,p ) =
(see Sect. 8.2).
In Sect. 8.3, we assume that is a bounded C 1 domain and that q =
pn/(n p), and n > p 1. In this case we nd an explicit value for
C(Ep,q ) = M1 (Ep,q ) = M2 (Ep,q ).
The results obtained in Sect. 8.1 are readily extended in Sect. 8.4 to the
embedding of L1p () to the space Lq (, ), where is a Radon measure.
Next, we turn to domains with a power cusp on the boundary and nd
explicit formulas for the measures of noncompactness under consideration
and apply these results to the Neumann problem for a particular Schr
odinger
operator (Sects. 8.5 and 8.6).
In the nal section we show relations between our measures of noncompactness and local isoconductivity and isoperimetric constants. In particular,
we obtain
essE1,q = C(E1,q ) = M1 (E1,q ) = M2 (E1,q ) = lim
s0
sup
g\s
(mn (g))1/q
,
Hn1 ( g)
8.1 Localization Moduli and Their Properties
437
where 1 < q < n/(n 1), g is a relatively closed subset of such that g
is a smooth surface, and Hn1 is the (n 1)-dimensional area. This together
with the results from Sect. 8.5 yields explicit values of the local isoperimetric
constants for power cusps.

Let us discuss relations between the moduli. First of all, obviously,
M1 (Ep,q ) M2 (Ep,q ).
(8.1.1)
Lemma 1. Let be a domain in Rn with mn () < . Suppose that

1 p < and 1 q < . Then for any u L1p () the following estimate
holds:

(8.1.2)
uLq () M1 (Ep,q ) + uL1p () + C()uLmax{q,p} ( ) ,
where is an arbitrary positive number, C() is a positive function of , and
is an open set with smooth boundary and compact closure .
Proof. Let denote a smooth function on R+ , such that 0 1 and
(t) = 0 for t (0, 1) and (t) = 1 for t 2. By d(x) we denote the distance
from x to . Let us introduce the cuto function Hs (x) = (d(x)/s).
We write

uLq Hs uLq + (1 Hs )uL ,
q
and note that the second term does not exceed

vLq ()
1
sup
: v Lp (), supp v \ 2s (1 Hs )uL1 .
p
vL1p ()
Hence, for suciently small s = s()

uLq M1 (Ep,q ) + uL1p + uHs Lp + Hs uLq .
Since the supports of Hs and its derivatives are in s the result follows.
Corollary. Let be a domain in Rn with mn () < . Suppose that

1 p < and 1 q < . Then

M1 Ep,q () < ,
if and only if
Ep,q L1p ()Lq () < .
Proof. By Lemma 1, M1 (Ep,q ()) < implies Ep,q L1p ()Lq () < .
The converse is obvious.

Lemma 2. Let be a domain in Rn with mn () < .
438
(i) If 1 q < and 1 p < , then

M1 (Ep,q ) essEp,q .
(ii) Let 1 q < pn/(n p) if 1 p < n and 1 q < if n p < . Then
essEp,q M1 (Ep,q ).
Proof. (i) By F we denote an operator of nite rank acting from L1p ()
into Lq () and given by

Fu =
cj (u)j ,
1jN
where j Lq () and cj are continuous functionals on L1p (). Let > 0. We

choose the operator F to satisfy
+ essEp,q Ep,q F .
(8.1.3)
Without loss of generality, we can assume that the functions j have compact
supports in . Hence, there exists a positive s() such that F u = 0 on \s()
for all u L1p (). Let s (0, s()). By (8.1.3), for all u L1p () vanishing on
s ,
uLq (\s() )
uLq ()
=
.
+ essEp,q
uL1p ()
uL1p ()
The required lower estimate for essEp,q follows from the denition of
M1 (Ep,q ).
(ii) Let be an arbitrary positive number and let s > 0 be so small that

vLq ()
1
: v Lp (), v = 0 on s M1 (Ep,q ) + .
sup
vL1p ()
We introduce a domain with smooth boundary and compact closure ,
, such that mn ( \ ) = mn ( \ s ) with any chosen (0, 1).
By we denote the characteristic function of . By the compactness of the
embedding L1p () Lq () we have
essEp,q sup
uL1p
u uLq ()
uLq (\)
= sup
.
uL1p ()
uL1p uL1p ()
It is obvious that for any positive T

|u| min T, |u| + |u| T + ,
where f+ means the nonnegative part of f . This implies
(8.1.4)

uLq (\) min T, |u| Lq (\) + |u| T + Lq (\) .
439
(8.1.5)
We use the notation Lt = {x : |u(x)| > t} and choose T as

T = inf{t > 0 : mn Lt < mn ( \ s )}.
Then, for q < ,

min T, |u|
T mn ( \ )1/q
Lq (\)

1/q
= mn ( \ s )
T 1/q uLq () .
Hence,
min{T, |u|}Lq (\)

1/q Ep,q .
uL1p ()
uL1p ()
sup
Combining this inequality with (8.1.4) and (8.1.5), we arrive at

essEp,q
sup
uL1p ()
(|u| T )+ Lq (\)

+ 1/q Ep,q .
uL1p ()
Let be an arbitrary positive number and let H be the function introduced in the proof of Lemma 1. Then

|u| T
|u| T + (1 H )L ()
+ Lq ()
q

+ |u| T + H L () .
(8.1.6)
q
To estimate the rst term on the right-hand side, we take so small that

vLq ()
1
sup
: v Lp (); v = 0 on 2 M1 (Ep,q ) + .
vL1p ()
We normalize u by uL1p () = 1. Then

|u| T (1 H )
+
Lq ()

M1 (Ep,q ) + 1 + |u| T + H L
p ()

.
(8.1.7)
Combining (8.1.6) and (8.1.7), we see that

(u T )+
M1 (Ep,q ) +
Lq ()

+ C() uLq (LT ) + uLp (LT ) (8.1.8)
with a constant C() depending only on , , p, and q but not on s. Using
the compactness of the restriction operator L1p () Lmax{p,q} ( ) and the
equality mn LT mn ( \ s ), we conclude that the two norms on the righthand side of (8.1.8) tend to zero as s 0. Therefore
440

lim sup |u| T + Lq () M1 (Ep,q ) +
s0
and hence
essEp,q M1 (Ep,q ) + + 1/q Ep,q .
The proof is completed by using the arbitrariness of and .
Theorem 1. Let be a domain in Rn with mn () < . Suppose that

1 p q < pn/(n p) if 1 p < n and 1 q < if n p < . Then
essEp,q = C(Ep,q ) = M1 (Ep,q ).
Proof. In the case q < pn/(n p), n > p and q arbitrary when p n the
trace map from L1p () into Lmax {p,q} ( ) is compact. Then from Lions and
Magenes [500, Theorem 16.4] it follows that
uLmax(p,q) ( ) uL1p () + C()uL1 ( ) ,
where > 0 is an arbitrary small number. This together with (8.1.2) allows
us to obtain
C(Ep,q ) M1 (Ep,q ).
Let > 0 be an arbitrary positive number. Then

uLq () C(Ep,q ) + uL1p () + c()uL1 (s() )
and hence
M1 (Ep,q ) C(Ep,q ) + .
We shall see that, in dealing with M2 (Ep,q ), we must distinguish between

the cases p < q and p = q.
Theorem 2. Let be a domain in Rn with mn () < and 1 p <
q < . Suppose that q < pn/(n p) for p < n and 1 q < for p n.
Then
essEp,q = C(Ep,q ) = M1 (Ep,q ) = M2 (Ep,q ).
Proof. By (8.1.1) and Theorem 1 it is sucient to assume that

M2 (Ep,q ) < and to prove the inequality
M2 (Ep,q ) C(Ep,q ).
Fix > 0 and let {Bi }i1 be an open covering of \ /2 by balls of radius centered at . Also let the multiplicity of the covering be nite and
depend only on n. Obviously the collection {Bi }i1 supplemented by the set
/2 forms an open covering of which has a nite multiplicity as well. We
441
introduce a family of nonnegative functions {i }i0 , such that 0 C0 (/2 )

and i C0 (Bi ) for i 1, and

i (x)p = 1 on .
(8.1.9)
i0
The estimates to be obtained in what follows will be rst proved for an arbitrary function u L1p () L (). Since L1p () L () by Lemma 1.7.1,
these estimates remain valid for u L1p ().
Clearly,

p
p

|i u|p
|i u|
uLq () =
L
()
=
q/p
Lq/p ()
i0
i0
i upLq () .
(8.1.10)
i0
Given > 0 and suciently small , we have

vLq ()
1
sup sup
: v Lp (), v = 0 on \ Bi M2 (Ep,q ) + .
vL1p ()
i1
Therefore, the right-hand side in (8.1.10) does not exceed
p

i upL1 () + 0 upLq () .
M2 (Ep,q ) +
p
i1
Using the elementary inequality

(a + b)p (1 + )ap + c()bp
(8.1.11)
for positive a and b, we see that

p
(1 + )
upLq () M2 (Ep,q ) +
i upLp ()

+ C(, )upLp ()
i1
+ 0 upLq () .
(8.1.12)
We note further that by (8.1.9) the sum over i 1 in (8.1.12) does not exceed
upLp () . Since q 1 > p1 n1 and the support of 0 is separated from
, it follows that by the compactness of the restriction mapping L1p () into
Lq (/2 ) we have the estimate
0 uLq () 0 uL1p () + c(, )0 uL1 () ,
where is an arbitrary small number.
(8.1.13)
442
Let be a positive number independent of . Since q > p, we conclude

that

p1 q 1
uLq (\ ) + uLp ( )
(8.1.14)
uLp () mn ( \ )
and
uLp ( ) uL1p () + c(, )uL1 ( ) .
(8.1.15)
Choosing and suciently small and using (8.1.12)(8.1.15) we arrive

at the inequality

uLq () M2 (Ep,q ) + c uL1p () + c()uL1 ( )
with a constant c independent of and u. This completes the proof.
8.2 Counterexample for the Case p = q

In the next example, we show that the condition p < q in Theorem 8.1/2
cannot be removed. To be more precise, for any p (1, ) we construct a
planar domain for which
Ep,p = essEp,p = M1 (Ep,p ) = C(Ep,p ) =
and M2 (Ep,p ) = 0.
Example. (See Fig. 33) Let R2 be the union of rectangles
Ai = (0, 1/2] (a2i2 , a2i ), i > 0,
B0 = [3/2, 2) (0, a1 ),
Bi = [3/2, 2) (a2i1 , a2i+1 ), i > 0,
Ci = [1/2, 3/2] (ai1 , ai ), i > 0,
where
a0 = 0 and
ai =
np
for i > 0.
n=1
Clearly, m2 () < . For each integer j > 0 we dene the continuous function
fj (x) as a function that is zero on

An
Bn
Cn ,
nj
nj
n<2j
fj (x) = i and fj (x) = i + 1 on Aj+i and Bj+i , respectively, and linear on

C2j+i with |fj | = 1 for i 0. The graph of each function fj (x) has the
shape of a staircase with slope 1 on C2j+i , and landings on Aj+i and Bj+i for
8.2 Counterexample for the Case p = q
443
Fig. 33.
i > 0. By a simple computation we obtain fj Lp < and fj Lp = for
each j > 0 that implies M1 (Ep,p ) = . Furthermore, by Theorem 8.1/2 we
have C(Ep,p ) = essEp,p = .
It remains to show that M2 (Ep,p ) = 0. Let x and < 1/4. By Q(x, )
we denote the square (x , x + )2 . By the denition of one obtains that
Q(x, ) is a union of open disjoint sets {Ii }, where Ii is either a rectangle
or the union of three rectangles.
For f L1p () with supp f Q(x, ), we have

f (x)p dx (c)p f (x)p dx.
Ii
Ii
Summing over {Ii }, we arrive at

Ii
and then
sup
1/p

f (x)p dx
c
Ii
1/p

f (x)p dx
uLp
: u L1p (), supp u Q(x, )
uL1p
for every x . This implies M2 (Ep,p ) = 0.

c
444
8.3 Critical Sobolev Exponent

Here we show that all our measures of noncompactness can be found explicitly if has a continuous normal and q is the critical Sobolev exponent
pn/(n p).
Theorem. Let n > p 1 and let be a bounded C 1 domain. Then
1/n
pn ) = M (E
pn ) = M (E
pn ) = 2
c(p, n),
C(Ep, np
1
2
p, np
p, np
where
c(p, n) = 1/2 n1/p
p1
np
11/p
(n) (1 + n/2)
(n/p) (1 + n n/p)
1/n
is the best constant in the Sobolev inequality

uL
pn
np
(Rn )
cuLp (Rn ) .
(8.3.1)
Proof. Let be a radial function in C0 (Rn ), > 0 and (x) = (x/). By

O we denote an arbitrary point at and we put ,O (x) := (x O) into
the inequality (8.0.1). We use q to denote pn/(n p). Using the denition of
C(Ep,q ), we obtain
lim sup
0
,O Lq ()
C(Ep,q ).
,O Lp ()
We note the existence of the limit

Lq (Rn+ )
,O Lq ()
21/n1/p Lq (Rn )
=
= 1/p
.
0 ,O L ()
Lp (Rn+ )
2
Lp (Rn )
p
lim
To obtain the lower estimate for C(Ep,q ) one needs only to recall that the best
constant in (8.3.1) is attained for a radial function when p > 1 and obtained
as a limit by a sequence of radial functions for p = 1.
Let us turn to the upper estimate for C(Ep,q ). We construct a nite open
covering of by balls Bi of radius with a nite multiplicity depending only
on n. Let either Bi or the center of Bi be placed on . We introduce a
family of nonnegative functions {i } such that i C0 (Bi ) and

i (x)p = 1 on .
i
Obviously,

p
|
u|
upLq () =
i

Lq /p ()
i

p

|i u| L () =
i upLq () .
q /p
(8.3.2)
8.3 Critical Sobolev Exponent
If B i , then
445

i uLq () c(p, n)(i u)Lp () .
Let be an arbitrary positive number. Since is suciently small and the

domain is in the class C 1 , one can easily show that

i uLq () 21/n c(p, n) + (i u)L ()
p
for the ball Bi centered at a point at . By using (8.1.11), we see that

(i u)p
i upLp () + ui pLp ()
L ()
p
(1 + )
i upLp ()
+ C()
ui pLp () .
(8.3.3)
We note that the rst sum on the right-hand side of (8.3.3) is equal to
upLp () . Now, it follows from (8.3.2) that

p
upLq () 21/n c(p, n) +

p
p
(1 + )uLp () + C()
ui Lp () . (8.3.4)
i
Furthermore the second sum in the right-hand side of (8.3.4) can be majorized
by C(, )upLp () , and then

p

upLq () 21/n c(p, n) + (1 + )upLp () + C(, )upLp () .
Since the embedding Ep,p () is compact, it follows from Lions and Magenes
[500, Theorem 16.4] that

p
upLp () 21/n c(p, n) +
q

p
(1 + )upLp () + uL1p () + C(, , )uL1 ( ) .
Using (8.1.11) once more, we arrive at (8.0.1). Hence
C(Ep,q ) 21/n c(p, n),
which, in combination with the lower estimate for C(Ep,q ), shows that
C(Ep,q ) = 21/n c(p, n).
(8.3.5)
Putting an arbitrary u L1p () vanishing outside \ s into (8.0.1) and

taking the limit as s 0, we conclude that
446
C(Ep,q ) M1 (Ep,q ).
(8.3.6)
Now, let u L1p (), u = 0 outside the ball B(x0 , ) with x0 and
suciently small . One can easily construct an extension u
of u onto the
whole ball B(x0 , ) so that

uLq (B(x,))
uLq ()
21/n (1 )
.
uL1p ()

uL1p (B(x,))
Choosing u in a such way that its extension u
is an almost optimizing function
from the Sobolev inequality (8.3.1), we arrive at
uLq ()
21/n c(p, n)(1 2),
uL1p ()
and the denition of M2 yields
M2 (Ep,q ) 21/n c(p, n).
(8.3.7)
Combining (8.3.5)(8.3.7) and the inequality M2 M1 , we complete the

proof.
8.4 Generalization
The previous results hold in a more general situation when there is a compact
subset of that is responsible for the loss of compactness of Ep,q and the
norm in the target space involves a measure.
Let K be a compact subset of and let \ K be locally Lipschitz (i.e.,
each point of \ K has a neighborhood U Rn such that there exists a
quasi-isometric transformation which maps U onto a cube). Dene

sK = x : dist(x, K) > s .
It is obvious that for each s > 0, the embedding

L1p sK Lq sK
is compact
if and only if the embedding
L1p (s ) Lq (s )
is compact.
Let be a measure on . We dene the embedding operator

Ep,q () : L1p () Lq (, ),
where
(8.4.1)
8.5 Measures of Noncompactness for Power Cusp-Shaped Domains

Lq (, ) =
1/q
u : uLq, =
|u| d
q
447

< .
We say that the measure is admissible with respect to K if for every s > 0
the trace operator L1p (sK ) Lq (sK , ) is compact.
Let us note that for 1 < p < q < and p < n the admissibility of is
equivalent to

lim sup q(pn)/p B(x, ) = 0
0 x K
s
(see Theorem 1.4.6/1). In the case 1 < p = n < q < the admissibility is
equivalent to

lim sup | log |q(p1)/p B(x, ) = 0
0 x K
s
by Theorem 11.9.1/4, which will be proved in the sequel.

We introduce the following modied versions of the localization moduli
dealt with previously:

uLq,
1
K
: u Lp (), v = 0 on \ s
M1 Ep,q (), K = lim sup
s0
uL1p
and

uLq,
M2 Ep,q (), K = lim sup sup
: u L1p (), supp u B(x, ) .
0 xK
uL1p
In the proofs of Theorem 8.1/1 and Theorem 8.1/2, we replace s ,
Lq (), M1 (Ep,q ), and M2 (Ep,q ) by sK , Lq (, ), M1 (Ep,q (), K), and
M2 (Ep,q (), K), respectively. Then we use (8.4.1) and the denition of the
admissible measure to obtain the following theorem.
Theorem. Let K be a compact subset of such that \ K is locally
Lipschitz, and let be an admissible measure with respect to K.
(i) Let 1 p q < pn/(n p) for n > p and let 1 q < for p n. Then

essEp,q () = C Ep,q () = M1 Ep,q (), K .
(ii) Let 1 p < q < pn/(n p) for n > p and let 1 q < for p n. Then

essEp,q () = C Ep,q () = M1 Ep,q (), K = M2 Ep,q (), K .
8.5 Measures of Noncompactness for Power Cusp-Shaped

Domains
In this section we nd explicit values of the measures of noncompactness of
the embedding Ep,q for power cusp-shaped domains.
448
Let be a bounded Lipschitz domain in Rn1 . Consider the -cusp

= x = (x , xn ) Rn : x xn , xn (0, 1) ,
where > 1.
Theorem. Let be the -cusp with > 1. Let p [1, ) and R.
We introduce the measure d = x
n dx and set
q=
(i) Let
((n 1) + 1 )p
.
(n 1) + 1 p
p( 1)(n 1)
<<p
np
(8.5.1)
for 1 < p < n,
or
< p < (n 1) + 1
for n p.
Then

essEp,q () = C Ep,q () = M1 Ep,q () = M2 Ep,q ()
1 1
1q

pq (n 1) + 1 p p q 1
p
= mn1 () pq
p1
q(p 1)
qp
pq
pq
( qp )
.
q
q1
( qp ) (p qp )
(ii) Let 1 < p < 1 + (n 1) and = p. Then

essEp,p () = C Ep,p () = M1 Ep,p ()

1

= M2 Ep,p () = p (n 1) + 1 p
.
(iii) Let p = 1 and 1 < < 1. Then

essE1,q () = C E1,q () = M1 E1,q ()

1/q

1q
,
= M2 E1,q () = mn1 () q (n 1) + 1
where q is given by (8.5.1) with p = 1.
Proof. (i) By denition of M2 we have

pq

M2 Ep,q () mn1 () pq lim Kp,q (, , ),
0
where
Kp,q (, , ) := sup

|v(t)|q t(n1) dt)1/q
0
,
( 0 |v (t)|p t(n1) dt)1/p
449
the supremum being taken over Lipschitz functions on [0, ] vanishing at .

Making the substitution t = /, with > 0, we derive
Kp,q (, , ) = sup
|v( )|q (n1) d )1/q

0
,

( 0 |v ( )|p (n1) d )1/p
where the supremum is taken over all Lipschitz functions on [0, ] vanishing
at . This implies that Kp,q (, , ) is constant in and then

( 0 |v(t)|q t(n1) dt)1/q
Kp,q (, , ) = sup p (n1)
,
( 0 |v (t)| t
dt)1/p
where the supremum is extended over all Lipschitz functions with compact
support on [0, ). By substitution of t = c , where c = (1 p)((n 1) + 1
p)1 , we obtain
Kp,q (, , )

1
qp 1
( 0 |v( )|q ((n1)+1)c1 d ) q

(n 1) + 1 p pq
=
sup
,

1
p1
(
|v ( )|p d ) p
0
where the supremum is taken over all Lipschitz functions with compact support on [0, ). Since

q(p 1)
,
(n 1) + 1 c 1 = 1
p
it follows by Blisss inequality (4.6.4) that
Kp,q (, , )
qp
qp 1
1q
pq
pq
( qp
)
p
(n 1) + 1 p pq
=
,
q
q1
p1
q(p 1)
( qp ) (p qp )
which gives the required lower estimate for the essential norm of Ep,q ().
We now derive an upper bound for the essential norm of Ep,q (). Let us
introduce the mean value of u over the cross section {(x , xn ) : x xn } by

1
u(xn ) :=
(z)u xn z, xn dz.
(8.5.2)
mn1 Rn1
Let u L1p (). By the triangle inequality
uLq (,) uLq (,) + u uLq (,) .
(8.5.3)
450
We estimate the rst term on the right-hand side
uLq (,) =

x
n

u(xn )q dx x dxn
n
1/q
1

u(xn )q x(n1) dxn
n
= (mn1 )1/q
0
Kp,q (1, , )(mn1 )
1/q
1/q
1

(u) (xn )p x(n1) dxn
n
1/p
.
(8.5.4)
Since

1
(z)u zxn , xn dz
(u) (xn ) =
xn mn1 Rn1

1
(z) n u + xn1 (z, x )u zxn , xn dz,
=
mn1 Rn1

it follows that

(u) (xn ) (n u)(xn ) + (diam )xn1 (x u)(xn ).
Using the last inequality, we obtain
1/p

(u) (x)p x(n1) dx

0
1/p

(n u)(x)p x(n1) dx
1/p

(x u)(x)p x(n1) dx
.
+ (diam )1
(8.5.5)
Since

v(xn )p x(n1) dxn
n
0
1/p
(mn1 )
1/p
x
n
|v| dx dxn
p
1/p
,
(8.5.6)
we have for suciently small

(u) (xn )p x(n1) dxn
n
1/p
(mn1 )1/p uLp (0 ) ,
(8.5.7)
where 0 = {x : xn < }. Then, for every > 0, there exists a suciently

small > 0 such that

pq
uLq (0 ,) Kp,q (1, , )(mn1 ) pq + uLp (0 ) .
451
(8.5.8)
Let us estimate the second term on the right-hand side of (8.5.3). Consider
a sequence {zk }k0 given by
zk+1 + zk+1
= zk ,
z0 1,
k 0.
zk 1.
One can easily verify that zk 0, zk+1 zk1 1. Moreover zk+1
Choosing z0 to be suciently small, we obtain zk+1 /2 < zk < 2zk+1 , k 1.
Let us set

Ck = x = (x , xn ) Rn : xn (zk+1 , zk ), x xn , k 1.
It follows from the construction of Ck that

/q
u uLq (Ck ,) 2/q zk
u uLq (Ck ) .
(8.5.9)
We obtain by Sobolevs theorem that

n( 1q p1 )
u uLq (Ck ) czk

u uLp (Ck ) + zk (u u)L
p (Ck )

, (8.5.10)
where c depends on , n, p, q and is independent of k.

By the Poincare inequality we have

u(x , xn ) u(x , xn )p dx cxp
x u(x , xn )p dx
n
x
n
x
n
for almost all xn (0, 1). Hence it follows from (8.5.10) and the previous
inequality that
n
(1 n
p+q)
u uLq (Ck ) c zk

uLp (Ck ) + uLp (Ck ) .
(8.5.11)
We deduce from (8.5.7) that

uLp (Ck ) = n uLp (Ck ) uLp (Ck ) .
(8.5.12)
Combining (8.5.9), (8.5.11), and (8.5.12) implies

n
q +(1 n
p+q)
u uLq (Ck ,) czk
uLp (Ck ) .
Using (8.5.13) and the inequality
we conclude
1/q
aqk
1/p
apk
ak 0, q p,
(8.5.13)
452
1/q
u
uqLq (Ck ,)
k=l
1/p
n
[ +(1 n
p + q )]p
zk q
upLp (Ck )
k=l
Since q + (1 np + nq ) > 0, it follows that for every > 0 there exists > 0
such that
(8.5.14)
u uLq (0 ,) uLp (0 ) .
Combining (8.5.3), (8.5.8), and (8.5.14) gives the upper estimate for the essential norm of Ep,q ().
(ii) Let us recall the Hardy inequality
p

f (t) dt x(n1)p dx
0
x
p
p
f p (x)x(n1) dx,
(8.5.15)
(n 1) + 1 p
0
where [p/((n 1) + 1 p)]p is the best constant. Then replacing Blisss
inequality in (i) by Hardys inequality (8.5.15) with appropriate changes in
the proof of (i), we obtain (ii).
(iii) As in (ii) we replace Blisss inequality by Hardys inequality
f (t) dt
0
(n1)
1/q

dx
c
f (x)x(n1) dx,

1/q
,
c = (n 1) + 1
and repeat the proof of (i).
8.6 Finiteness of the Negative Spectrum of a Schr

odinger
Operator on -Cusp Domains
Let Rn be the -cusp domain with > 1. In this section we study the
Neumann problem for the Schr
odinger operator
N u
u = f in ,
x2n
u
= 0 on \ {0},
(8.6.1)
where = const > 0 and is a normal. The corresponding quadratic form is

given by

|u|2
|u|2 dx
dx.
(8.6.2)
A (u, u) =
2
xn
8.6 Finiteness of the negative spectrum
453
We need the following Hardy-type inequality for functions in .

Lemma. If u L12 () and u = 0 for xn = 1, then

2 2
|u| xn dx c
|u|2 dx.
(8.6.3)
Proof. Let X = (s, t), where s Rn1 , t (0, ) and

t = ( 1)1 x1
1 .
s = x
n x ,
n
The change of variable x X maps onto the half-cylinder {X : s ,
t > 0}. Put v(X) = u(x). Since
|x u| |X v|xn (t)
and the Jacobian |Dx/DX| is equal to c xn (t)n , the desired inequality (8.6.3)
takes the form

v(s, t)2 xn (t)2+n dt ds
0

c
|X v|2 xn (t)(n2) dt ds.
(8.6.4)
Setting here xn (t) = (1+( 1)t)1/(1) , we see that (8.6.4) is a consequence

of the one-dimensional Hardy-type inequality

(n2)
n2
2
1
|w| (1 + t)
dt c
(w )2 (1 + t) 1 dt
0
with w(0) = 0, which follows from Theorem 1.3.2/1.

x2
n
As in Sect. 8.5, we denote d =

dx. Since, by the lemma just proved,
the space L12 () is continuously embedded into L2 (, ) when > 1, A is
well dened on L12 ().
We start by showing the semiboundedness of A which guarantees the
existence of the Friedrichs extension of (8.6.1).
Theorem 1. A is semibounded if and only if [((n 1) 1)/2]2 .
Proof. Let u be the mean value of u over the cross section (see (8.5.2)),
then by the triangle inequality
uL2 (,) uL2 (,) + u uL2 (,) .
(8.6.5)
Combining (8.5.4), (8.5.5), and (8.5.6) (with p = q = 2 and = 2), we

obtain
454

uL2 (0 ,) essE2,2 () n uL2 (0 )

+ (diam )1 x uL2 (0 ) ,
(8.6.6)
where 0 = {x = (x , xn ) Rn : x and xn < }.

Let us estimate the second term on the right-hand side of (8.6.5)

u(x , xn ) u(xn )2 dx dxn .
u u2L2 (0 ,) =
x2n
x
0
n
Using the Poincare inequality on the cross line, we see that

2
x u(x , xn )2 x22
u u2L2 (0 ,) mn1 ()
dx dxn
n

22
x
n
2
mn1 () x u2L2 (0 ) .
Therefore, by (8.6.5) and (8.6.6), we obtain

uL2 (0 ,) essE2,2 ()uL2 (0 )
for suciently small > 0.
Then
u2L2 (,) u2L2 (0 ,) + u2L2 (\0 ,)
2

essE2,2 () u2L2 (0 ) + c()u2L2 (\0 )

2
2/ (n 1) 1 u2L2 (0 ) + c()u2L2 () , (8.6.7)
which gives the semiboundedness for [((n 1) 1)/2]2 .
Let > [((n 1) 1)/2]2 . We set d = [((n 1) 1)/2]2 > 0. It
follows from Theorem 8.5 that

|u|2 dx

2
x2n
1
0
lim sup
= 2/ (n 1) 1 .
:
u
L
(),
supp
u
2
0
|u| dx
0
Then there exists a sequence of functions {ui }
i=1 such that supp ui 1/i
and

|ui |2
d
dx
>
|ui |2 dx.
2 x2n
Hence,

|ui |2 dx

|ui |2
|ui |2
d
dx
<
dx
2
xn
2 x2n

i2 d
|ui |2 dx.
Therefore A is not semibounded when > [((n 1) 1)/2]2 .
8.6 Finiteness of the negative spectrum
455
The next theorem gives a condition for niteness of the negative spectrum
of A .
Theorem 2. If < [((n 1) 1)/2]2 , then the negative spectrum of A
is nite.
Proof. Let < [((n 1) 1)/2]2 and M be a linear innite-dimensional
manifold in L12 (). Take
<
1 [2/((n 1) 1)]2
.
(2 + 2)
(8.6.8)
It follows from (8.6.7) that

2
u2L2 (,) 2/ (n 1) 1 u2L2 (0 ) + c()u2L2 (\0 )
for suciently small > 0, where 0 = {x = (x , xn ) Rn : x and
xn < }.
Since the restriction L12 () L2 ( \ 0 ) is compact, there exists a nite
rank operator F : L12 () L2 ( \ 0 ), for which
u F uL2 (\0 )
c()
1/2
uL12 () .
Note that

2
u2L2 (,) 2/ (n 1) 1 u2L2 (0 )
2
1/2
+ c()
uL12 () + F uL2 (\0 ) .
c()
Let M M be dened by
M := {u : F u = 0 and u M }.
Then M is a linear innite-dimensional manifold in L12 () and for every
u M we have

2
u2L2 (,) 2/ (n 1) 1 u2L2 (0 ) + u2L12 ()

2

2/ (n 1) 1
+ 2 u2L2 () + 2u2L2 () . (8.6.9)
Combining (8.6.2), (8.6.8), and (8.6.9), we obtain for each u M M

|u|2
A (u, u) =
|u|2 dx
dx
2
xn

1

2
|u|2
|u|2
+ 2
dx
dx
2/ (n 1) 1
2
2
xn
xn

2
1
2 2/ (n 1) 1
+ 2
|u|2 dx > 0.
456
Therefore, there does not exist a linear manifold of innite dimension on which

A (u, u) < 0. This together with Lemma 2.5.4/2 completes the proof.
8.7 Relations of Measures of Noncompactness with Local

Isoconductivity and Isoperimetric Constants
Let E and F denote arbitrary relatively closed disjoint subsets of . We use
the p-conductivity of the conductor K = (\F )\E, that is,

cp (K) = inf upLp : u 1 on E and

u = 0 on F, u is locally Lipschitz in
(this denition is equivalent to the one given in Sect. 6.1.1) and we dene the
local isoconductivity constants

((E))1/q
S(p, q, , ) = lim sup
: E \ s
s0
(cp ((\s )\E))1/p
and
q, , ) = lim sup sup
S(p,
0 x

((E))1/q
: E B(x, ) .
(cp (( B(x, ))\E))1/p
Theorem 1. Let 1 p q < pn/(n p) if n > p and 1 q < if

p n. Then

(8.7.1)
S(p, q, , ) essEp,q () K(p, q)S(p, q, , ).
If additionally p < q, then

q, , ),
q, , ) essEp,q () K(p, q)S(p,
S(p,
pq
( qp
)
(qp)/pq
q
q1 )
( ( qp
) (p qp
)
K(p, q) = (p 1)(1p)/p ,
1,
where
(8.7.2)
when 1 < p < q;

when 1 < p = q;
when 1 = p q.
Proof. The left-hand side inequality in (8.7.1) follows immediately from

the denition of M1 .
The right-hand side inequality in (8.7.1) is a consequence of the conductor
inequality

1/q

1/p

q/p
q
p
d(t )
K(p, q)
|u| dx
,
cp (\F )\Nt
0
457
where u = 0 on F which is a relatively compact subset of (see Lemma

6.1.3/3). The inequality (8.7.2) follows by the same arguments when M1 is
replaced by M2 .

In the case p = 1 the p-capacity can be replaced by the (n1)-dimensional
area Hn1 . By g we denote relatively closed subsets of such that g
are smooth surfaces. We introduce the local isoperimetric constants

((g))1/q
: g \ s
T (q, , ) = lim sup
s0
Hn1 ( g)
and
T(q, , ) = lim sup sup
s0 x

((g))1/q
: g B(x, ) .
Hn1 ( g)
Theorem 2. If 1 < q < n/(n 1), then

T (q, , ) = T(q, , ) = essE1,q (),
and if q = 1, then

T (q, , ) = essE1,q ().
Proof. It follows from Lemma 5.2.3/1 that M1 (E1,q ) = T (q, ), which

together with Theorem 8.1/1 completes the proof.

In view of Theorem 8.1/1 and Theorem 8.1/2, the role of the essential norm
of Ep,q () can be played by M1 (Ep,q ()) and C(Ep,q ()) and additionally by
M2 (Ep,q ()) when p < q.
The following corollary, which is an immediate consequence of Theorem 8.5
(iii) and the previous theorem, gives explicit values of the local isoperimetric
constants for power cusps.
Corollary. Let Rn be the -cusp domain with > 1 and
(1 , 1). We introduce the measure d = x
n dx and set
q=
Then
(n 1) + 1
.
(n 1)

1q
1/q
T (q, , ) = mn1 () q (n 1) + 1
.

The material of this chapter is borrowed from the paper by Lang and Mazya
[478]. Lemma 8.6 is a particular case of Lemma 5.1.3/3 in Mazya and Poborchi
[576].
458
Various characteristics of noncompact embeddings such as essential norms,

limits of the approximation numbers, certain measures of noncompactness,
and the constants in the Poincare-type inequalities, were investigated by Amick [45], Edmunds and Evans [248], Evans and Harris [261] (see Edmunds
and Evans [249] and [250] for a detailed account of this development), and
Yerzakova [806, 807].
9
Space of Functions of Bounded Variation
In the 1960s the family of spaces of dierentiable functions was joined by

the space BV () of functions whose derivatives are measures in the open set
Rn . This space turned out to be useful in geometric measure theory, the
calculus of variations, and the theory of quasilinear partial dierential equations. In the present chapter we study the connection between the properties
of functions in BV () and geometric characteristics of the boundary of .
Sections 9.1 and 9.2 contain well-known basic facts of the theory of BV ().
In Sect. 9.3 we nd requirements on in terms of relative isoperimetric inequalities that are necessary and sucient for any function in BV () to admit
an extension to Rn in the class BV () and for the boundedness of the extension operator. Some explicit formulas for the norm of the optimal extension
operator are found in Sect. 9.4.
We dene the notion of a trace on the boundary for a function in BV ()
and give conditions for the integrability of the trace (Sects. 9.5 and 9.6). Some
results on the relation between the relative isoperimetric inequalities and the
integral inequalities (of the embedding theorem type) for BV () similar to
those in Chap. 5 are established in Sect. 9.5. We also consider conditions for
the validity of the GaussGreen formula for functions in BV () (Sect. 9.6).
9.1 Properties of the Set Perimeter and Functions

in BV ()
9.1.1 Denitions of the Space BV () and of the Relative
Perimeter
The space of functions u that are locally integrable in , whose gradients
u (in the sense of distribution theory) are charges in , is called the space
BV (). We denote the variation of the charge over the whole domain by
uBV () . The perimeter of a set E relative to is dened by

DOI 10.1007/978-3-642-15564-2 9,
459
460
9 Space of Functions of Bounded Variation
P (E ) = E BV () ,
(9.1.1)
where A is the characteristic function of the set A . (We put P (E ) =

provided E
/ BV ().)
We introduce the perimeter of E relative to the closed set C = Rn \.
Namely, PC (E ) = inf GC PG (E ), where G is an open set.
If PRn (E ) < , then obviously
P (E ) = var Rn E (),
PC (E ) = var Rn E (C),
P (E ) + PC (E ) = PRn (E ).
(9.1.2)
We also note that

P (E1 ) = P (E2 )
(9.1.3)
if E1 = E2 .
Henceforth, in the cases where it causes no ambiguity, we write u instead
of u, Rn u and P (E ) instead of PRn (E ).
9.1.2 Approximation of Functions in BV ()
The denition of the mollication operator Mh (cf. Sect. 1.1.3) immediately
implies the following lemma.
Lemma 1. If u BV () and G is an open subset of with [G]h ,
where [G]h is the h-neighborhood of G, then

Mh uL1 () var u [G]h .
(9.1.4)
Lemma 2. If ui BV () and ui u in L(, loc), then

uBV () lim inf ui BV () .
i
(9.1.5)
Proof. It suces to consider the case

lim inf ui BV () < .
i
For any vector-function with components in D() we have

ui dx =
ui div dx
u div dx.
Therefore,

u div dx sup || lim inf ui BV () ,

i
i.e., u is a charge and inequality (9.1.5) holds.
(9.1.6)
9.1 Properties of the Set Perimeter and Functions in BV ()
461
By denition, the sequence {k }k1 of nite charges converges to a charge

locally weakly in if, for any function C() with compact support,

k (dx) =
(dx).
lim
k
Lemma 3. If ui BV (), ui u in L1 (, loc), and supi uBV () < ,

then
loc. weak.
ui u.
Proof. By Lemma 2 we have u BV (). It remains to refer to (9.1.6) and
to use the density of D() in the space of continuous functions with compact
supports in .

Theorem. Let u BV (). Then there exists a sequence of functions

{um }m1 in C () such that um u in L1 (, loc) and

|um | dx = uBV () .
lim
m
If, in addition, u L1 () then um u in L1 ().
Proof. Let {i }i0

be a sequence of open sets with compact closures i
i+1 and such that i i = , 0 = . We choose i so that

i = 0.
var u
i
This can be done, for instance, in the following way. Let

{i } be an arbitrary
,
sequence of open sets with compact closures such that i i = ,
i

t
i j = for i = j, i C . Let i denote the t-neighborhood of i .
For small t the surface it is smooth. The set of t for which var u(it ) =
0 is nite or countable as the collection of jumps of a monotonic function.
Therefore, for any i we can nd an arbitrarily small number ti such that
var u(iti ) = 0. It remains to put i = iti .
We x a positive integer m and choose bounded open sets Di , Gi such
i (
i+1 \i ) and
that Gi D

i+1 \i ) < m1 .
var u Gi \(
i
Let {i } be a partition of unity subordinate to the covering {Di } of .

i [Di ]h and
We can nd small numbers hi > 0 so that G
i
u Mhi uL1 (Di ) max |i | m1 2i .
We put um =
i Mhi u. Then um u in L1 (, loc). Also,
(9.1.7)
462
um L1 ()

i Mhi uL1 () +
Mhi ui

(9.1.8)
L1 ()
Since supp i Di , by Lemma 1 the rst term does not exceed
Mhi uL1 (Di )

var u(Gi )
var u() + m1 .
i
By the equality i i = 0 and (9.1.7), for the second summand on the right
in (9.1.8) we have

M
u
=
(M
u
u)
hi
i
hi
i

i
L1 ()
L1 ()
Mhi u uL1 (Di ) max |i |

i
D
m1 2i = m1 .
i=1
Consequently,
um L1 () var u() + 2m1
and
lim sup um L1 () var u().
m
The result follows from the last inequality together with Lemma 2.
Corollary. If u BV (), then the functions u+ , u , and |u| are also

contained in BV () and
+

u
u
|u|
+
=
u
.
(9.1.9)
BV
()
BV ()
BV ()
BV ()
In fact, let {um } be the sequence of functions in C () introduced in the
+
Theorem. Then u+
m u , um u in L1 (, loc) and

+

u m
+u
(9.1.10)
m BV () = |um | BV () = um BV () .
BV ()
This and Lemma 9.1.2/1 imply
+

u
+u BV () uBV () .
BV ()
(9.1.11)
Consequently, u+ , u BV (). The reverse inequality of (9.1.11) is obvious.

Passing to the limit in (9.1.10), we arrive at the inequality in (9.1.9).
463
9.1.3 Approximation of Sets with Finite Perimeter

By denition, the sequence {Ei } of sets Ei converges to E if Ei
E in L1 (, loc).
Lemmas 9.1.2/2 and 9.1.2/3 imply the next assertions.
Lemma 1. If Ek E , then P (E ) lim inf k P (Ek ) and if
supk P (Ek ) < , then
loc. weak.
Ek E .
Lemma 2. Let u L1 () and let E be the characteristic function of the
set E . Further let E uL1 () . Then for any t [, 1 ], > 0,
the inequality
E Nt L1 () 1
holds where Nt = {x : u(x) t}.
Proof. Obviously,

E uL1 (E \Nt ) +uE L1 (Nt \E )

u(x) dx.
1u(x) dx+
E \Nt
Nt \E
Since u(x) < t for x E \Nt and u(x) t for x Nt \E , we obtain

(1 t)mn (E \Nt ) + tmn (Nt \E ) E Nt L1 () .
Theorem. For any measurable set E having nite measure mn there

exists a sequence of sets Ei for which Ei \ is a C -smooth submanifold of Rn (however, not compact, in general). Moreover, Ei E in L1 ()
and P (Ei ) P (E ).
Proof. If P (E ) = , then the result follows. Let P (E ) < . Let um denote
the sequence constructed in Theorem 9.1.2 for u = E . Since 0 E 1, the
denition of um implies 0 um 1. Therefore, by Theorem 1.2.4 we have
1
(m)
s Et
dt,
(9.1.12)
um L1 () =
0
where Et
= {x : um (x) = t}. The sets Et
are C -manifolds for almost all
t (0, 1) (see Corollary 1.2.2). In what follows we consider only such levels t.
Let > 0. We choose m = m() to be so large that
(m)
(m)
E um L1 () < .
Then by Lemma 1,
E N (m) L1 () 1/2 ,
t
(9.1.13)
464

(m)
where Nt
= {x : um (t) t} and t [1/2 , 1 1/2 ]. Furthermore, for any
m there exists a t = tm (1/2 , 1 1/2 ) such that
(m)
1 21/2 s Etm
(m)
dt.
s Et
(9.1.14)
Inequalities (9.1.13) and (9.1.14), together with the equality

P (E ) = lim
(m)
s Et
dt,
which follows from (9.1.12) and Theorem 9.1.2, imply N (m) E in L1 ()

tm
and
(m)
lim sup s Etm P (E ).

Remark. If E is a set with compact closure E , then the smooth

manifolds constructed in the preceding theorem are compact.
9.1.4 Compactness of the Family of Sets with Uniformly Bounded
Relative Perimeters
Theorem. The collection of sets E with uniformly bounded relative
perimeters P (E ) is compact.
Proof. By Theorem 9.1.2, for any E there exists a sequence um that
converges to E in L1 (, loc) and is such that
lim um L1 () = P (E ).
Lemma 1.4.6 implies that the family {um } is compact in L1 () where is an

arbitrary open set with compact closure
and with a smooth boundary.

Therefore, the family {E } is compact in L1 ().

9.1.5 Isoperimetric Inequality
Theorem. If E is a measurable subset of Rn and mn () < , then
mn (E )(n1)/n n1 vn1/n P (E ).
(9.1.15)
Proof. It suces to consider the case P (E ) < . By Theorem 9.1.3 there

exists a sequence of open sets Ei with C -smooth boundaries Ei such that
mn (Ei ) mn (E ) and s(Ei ) P (E ), where s is the (n 1)-dimensional
area. Inequality (9.1.15) is valid for the sets Ei (cf. Lyusternik [507] et al.).
Passing to the limit we arrive at (9.1.15).
465
9.1.6 Integral Formula for the Norm in BV ()

Lemma. If u1 and u2 are nonnegative functions in L1 () then

|u1 u2 | dx =
mn Lt1 \Lt2 Lt2 \Lt1 dt,
0
where Lti = {x : x , ui (x) > t}.

Proof. It is clear that

|u1 u2 | dx = (u1 u2 ) dx +
(u2 u1 ) dx = I1 + I2 ,
\A
where A = {x : u1 > u2 }. By Lemma 1.2.3,

mn Lt1 A dt
mn Lt2 A dt.
I1 =
0
We note that u1 (x) > u2 (x) if x Lt1 \Lt2 . Therefore, (Lt1 \Lt2 ) A =
Lt1 \Lt2 and, similarly, (Lt2 \Lt1 ) (\A) = Lt2 \Lt1 . Hence

I1 =
mn Lt1 \Lt2 dt.
0
In the same way we obtain

I2 =

mn Lt2 \Lt1 dt.
This completes the proof.
Theorem. For any function u that is locally integrable in we have

+
uBV () =
P (Lt ) dt,
(9.1.16)
where Lt = {x : u(x) > t}.

Proof. By Corollary 9.1.2 we may assume u 0. According to Lemma 1.2.3,
for any smooth vector-function with compact support in ,

u div dx =
dt
Lt div dx .
0
Therefore,

u div dx max ||
P (Lt ) dt,

466

is the lower Lebesgue integral. Hence
where

uBV ()
P (Lt ) dt.
(9.1.17)
If uBV () = then (9.1.16) follows. Let u BV (). Consider the

sequence {um } constructed in the proof of Theorem 9.1.2. Note that um 0.
i and
Let {i } be
a sequence of open sets i with compact closures
such that i i = . Since um u in L1 (, loc), by the Lemma we obtain

|um u| dx =
mn Ltm \Lt Lt \Ltm i dt 0,
0
where Ltm = {x : um (x) > t}. Therefore, for almost all t and for all i,

m
mn Ltm \Lt Lt \Ltm i 0.
The latter means that Ltm Lt for almost all t. Hence by Lemma 9.1.3/1
we have
P (Lt )
lim inf
m
P Ltm
dt lim inf

P Ltm dt,
where
is the upper Lebesgue integral. According to (1.2.6), the last inte-
gral is equal to um L1 () and thus
P (Lt ) dt lim inf um L1 () = uBV () ,

m

9.1.7 Embedding BV () Lq ()
The contents of this subsection are closely connected with Chap. 5. First we
note that by Theorem 9.1.2 the inequality
inf u cLq () CuL1 () ,
u L1 (),
inf u cLq () CuBV () ,
u BV ().
cR1
implies
cR1
Therefore, for the domain with nite volume, by Theorem 5.2.3 the last
inequality (for q 1) holds for u BV () if and only if J , = q 1 .
9.2 GaussGreen Formula for Lipschitz Functions
467
In the same way we can establish theorems similar to Theorem 5.5.2/1

and 6.3.3 for the space BV (). By Theorem 9.1.3 the denitions of the classes
J and J can be formulated in terms of the ratio

mn (E ) /P (E ),
where E is a measurable subset of . The area minimizing function M introduced in Sect. 5.2.4 can be dened as the inmum of the numbers P (E )
taken over the collection of measurable sets E with mn (E ) M .
Further we note that by Lemma 5.2.1/1 for the unit ball and for any
E , the inequality
1
n/(n1)

n/(1n)
P (E )
,
min mn (E ), mn (\E ) vn vn1
2
(9.1.18)
holds with the best possible constant.

9.2.1 Normal in the Sense of Federer and Reduced Boundary
For xed x, Rn , = 0 we put

A+ = y : (y x) > 0 ,

A = y : (y x) < 0 ,

A0 = y : (y x) = 0 .
Denition 1. The unit vector is called the (outward) normal in the
sense of Federer to the set E at the point x if

lim n mn E B (x) A+ = 0,
0

(9.2.1)
lim n mn CE B (x) A = 0.
0
Denition 2. The set of those points x E for which normals to E exist

is called the reduced boundary E of E .
9.2.2 GaussGreen Formula
The remainder of Sect. 9.2 contains the proof of the following assertion.
Theorem 1. If P (E ) < , then the set E is measurable with respect to
s and var E . Moreover, var E (E \ E ) = 0 and, for any B E ,

E (B) =
(x)s(dx),
var E (B) = s(B).
(9.2.2)
B
468
These formulas immediately imply the next assertion.

Theorem 2. (The GaussGreen formula). If P (E ) < and u is a Lipschitz function in Rn with compact support, then

u(x) dx =
u(x)(x)s(dx).
(9.2.3)
E
It suces to prove (9.2.3) for smooth functions. By the denition of E

we have

E u dx =
uE (dx),
Rn
Rn
which together with (9.2.2) implies the result.
Remark. In particular, from Theorem 1 it follows that P (E ) s(E ). The

case P (E ) < s(E ) is not excluded. Moreover, the perimeter can be nite
whereas s(E ) = . As an example, it suces to consider the plane disk B1
from which a sequence of segments of innite total length has been removed.
In this case E = B1 .
9.2.3 Several Auxiliary Assertions
Lemma 1. Let the vector charge concentrated on E Rn satisfy the condition |(E )| = var (E ). Then = a where a is a constant vector and is a
scalar nonnegative measure.
Proof. Since the charge is absolutely continuous with respect to = var
then the derivative d/d = f = (f1 , . . . , fn ) exists almost everywhere. The
equality |(E )| = var (E ) implies that |f | < 1 is impossible on a set of
positive measure . In fact, then we have |(E)| < var (E) for some E with
(E) > 0. Since |(E \E)| var (E \E) then

(E ) (E) + (E \E) < var (E) + var (E \E) = var (E ),
and we arrive at a contradiction. Since |f | 1 almost everywhere, it follows
that |f | = 1 almost everywhere and since |di /d| 1, we have

di
d
i (E ) =
d
by the absolute continuity of i with respect to . Therefore,
1/2

2 1/2

2 1/2

di
(E ) =
fi d
d
i (E )2
=
=
.
E d
i
i
i
The condition (E ) = |(E )| means that the Minkowski inequality

2 1/2
fi d

469
1/2
fi2
holds with the equality sign. Then the functions x fi (x) do not change sign
and are proportional with coecients independent of x for almost all x.

Lemma 2. If P (E ) < and the equality E = a is valid in the ball
B , where a is a constant vector and is a scalar measure, then, up to a set
of the measure mn zero, we have

B E = x B : (x y)a > 0 ,
where y is a point in B .
Proof. We may assume that a = (1, 0, . . . , 0). Mollifying E with radius ,
we obtain
M E 0,
x1
M E = 0
xi
(i = 2, . . . , m),
in the ball B . Hence the function M E does not depend on x2 , . . . , xn

and does not decrease in x1 in B . Therefore, the same is true for the limit

function E .
Lemma 3. If P (E ) < , then, for almost all > 0,

1
B E (B ) =
x ds(x).
E B
Proof. For all > 0 except, at most, a countable set we have
var E (B ) = 0. Suppose is not contained in that exceptional set. Let
(t) denote a piecewise linear continuous function on (0, ), equal to 1 for
t and vanishing for t > + , > 0.
Since

E (x) |x| dx =
|x| E (dx),
Rn
Rn
we have

1
x
dx = E (B )
E (x)
|x| E (dx). (9.2.4)
B+ \B
|x|
B+ \B
By virtue of var E (B ) = 0, the last integral converges to zero as
+0. The left-hand side of (9.2.4) has the limit equal to

1

xs(dx)
B
for almost all . The result follows.
470
Lemma 4. If P (E ) < , then P (E Br (x)) < for any x Rn and

for almost all r > 0. Moreover,

P E Br (x) = PBr (x) (E ) + s E Br (x) .
Proof. We assume that x is located at the origin. By Theorem 9.1.3, there
exists a sequence of polyhedra i such that i E and P (i ) P (E ).
Using the Fubini theorem, we obtain s(i Br ) s(E Br ) for almost
all r > 0. Then
lim sup P (i Br ) lim P (i ) + lim s(i Br ) = P (E ) + s(E Br )
i
and thus P (E Br ) < . By Lemma 9.1.3/1, there exists a sequence of

polyhedra {i } such that
i Br E Br ,
i E .
(9.2.5)
Let the number r satisfy the equality

lim sup var i (Br ) = 0
i
(which can fail only for a countable set of values r). Then (9.2.5) implies
Br i Br E .
(9.2.6)
By the convergence s(i Br ) s(E Br ) we nd that the set functions

i dened by

i (B) =
Bi ds,
Br
where is the outward normal to Br , weakly converge to , where

(B) =
BE ds.
Br
Obviously, i Br = Br i + i . Passing here to the limit and taking

into account (9.2.5), (9.2.6), and the weak convergence of i to , we arrive
at E Br = Br E + . Since the set function Br E is supported on Br
and supp Br , the result follows from the last identity.

9.2.4 Study of the Set N

Let N denote the set of points x E that satisfy the following conditions:
(a) var E (B (x)) > 0 for all > 0,
(b) the limit

= lim
0
471
E (B (x))
var E (B (x))
exists and || = 1.
We note that the vector (x) exists and that |(x)| = 1 almost everywhere
with respect to the measure var E . Moreover,

(x) var E (dx).
(9.2.7)
E (B) =
BN
Lemma. If P (E ) < and x N , then

lim inf n mn E B (x) > 0,
0

lim inf n mn CE B (x) > 0,
0
lim sup 1n PB (x) (E ) < .
(9.2.8)
(9.2.9)
(9.2.10)
0
Proof. By denition of ,

PB (x) (E ) = var E B (x) 2E B (x)
for suciently small . By Lemma 9.2.3/3 the right-hand side of this inequality
does not exceed 2s(E B (x)). Using Lemma 9.2.3/1, we have

P E B (x) = PB (x) (E ) + s E B (x) .
Hence

P E B (x) 3s E B (x) .
(9.2.11)
This estimate, together with the isoperimetric inequality (9.1.15), shows that
(n1)/n

d
c mn E B (x)
mn E B (x)
d
(9.2.12)
for suciently small . The property (a) of the set N and Lemma 9.2.3/4
imply P (E B (x)) > 0; therefore,

mn E B (x) > 0.
Since the function mn (E B (x)) is absolutely continuous, it follows
from (9.2.12) that c1 n mn (E B (x)) for almost all .
It is clear that the last inequality is actually true for all . Thus (9.2.8)
follows. Replacing E by CE in the previous argument we arrive at (9.2.9).
From (9.2.11) we have

P E B (x) cn1
472
for almost all , which together with Lemma 9.2.3/1 yields

PB (x) (E ) cn1

for all .
Theorem. If P (E ) < and x N , then the normal at x exists and

= . Moreover, for any > 0

(9.2.13)
lim 1n var E B (x) A0 = vn1 ,
0
where A0 = {y : (y x) = 0} and [ ] is the -neighborhood.

Proof. It suces to check that any sequence > 0 contains a subsequence such that equalities (9.2.1) and (9.2.13) are valid. Let E denote
the set obtained from E by the similarity transformation with center x
and coecient . We may assume that x is located at the origin. Clearly,
PB (E ) = n1 PB1 (1 E ). By the Lemma the relative perimeters PB1 (1 E )
are uniformly bounded. Consequently, by Theorem 9.1.4, there exists a sequence i > 0 such that the sequence of sets B1 1
i E converges to some set
D. Moreover, Lemma 9.1.3/1 yields
B1 1 E B1 D .
i
Thus, for all r (0, 1) except, at most, a countable set we have

i
B1 1 E (Br ) = 1 E (Br ) D (Br ).

i
(9.2.14)
By denition of the set N we obtain

lim
|1 E (Br )|
i
var 1 E (Br )
= lim
|E (Bi r )|
= 1.
var E (Bi r )
Comparing the latter equalities with (9.2.14) and taking into account the
semicontinuity of the variation under the weak convergence we obtain

D (Br ) = lim 1 (Br ) = lim var 1 (Br ) var D (Br ).
E
E
i
(9.2.15)
Hence, by virtue of Lemmas 9.2.3/1 and 9.2.3/2, we conclude that the set
D Br coincides with {y Br : y < b} up to the set of measure mn zero.
We show that b = 0. In fact, if b < 0 then

1
0 = mn (D B|b| ) = lim mn 1
i E B|b| = lim i mn (E B|b| ),
i
which contradicts (9.2.8). Similarly, b > 0 contradicts (9.2.9). From the convergence
Br 1
i E Br D = Br A ,
473
it follows that equalities of the form (9.2.1) are valid for the sequence {i r},
where {i } is a subsequence of any given sequence 0 and r is arbitrarily
close to unity. Hence (9.2.1) is true.
It remains to prove (9.2.13). We choose a subsequence i such that
var 1 E ,
i
where satises the inequality

(B) var D (B)
for any B B1 . Furthermore, (9.2.15) implies the existence of numbers r < 1,
arbitrarily close to unity, such that

(Br ) = D (Br ) var D (Br ).
Therefore, = var D . Now, for almost all > 0 and r (0, 1), we have

lim (i r)1n var E Bi r A0 i = lim var 1 E Br A0
i
i
i

= var D Br A0 = r n1 vn1 ,


9.2.5 Relations Between varE and s on E
Theorem 9.2.4 implies E N . Moreover, since var E (Rn \N ) = 0, it follows that var E (Rn \ E ) = 0 and thus the sets N and E are measurable
relative to var E .
Next we need the following well-known general assertion.
Lemma 1. Let be a measure in Rn and let, for all points x in the
-measurable set B, the following inequality hold:

lim sup 1n B (x) > 0,
0
where does not depend on x. Then s(B) c(n)(B).

Proof. For any > 0 there exists an open set G such that
(G\B) + (B\G) < .
By the denition of the Hausdor measure s, given > 0, there exists a > 0
such that
s(G) c1 (n)
n1
+ ,
(9.2.16)
i
for any covering of G by balls Bi with i < . For any x B G consider
the family of balls B (x) G, 3 < , such that
474

1n B (x) /2.
(9.2.17)
By Theorem 1.2.1/1, there

exists a sequence of mutually disjoint balls Bi (xi )
satisfying the condition i B3i (xi ) B G. Then

Bi (xi ) (G) (B) + .
By (9.2.16) and (9.2.17) we have
s(G) c1 (n)
(3i )n1 + c2 (n)
n1
+
i

c2 (n) 1
Bi (xi ) + c3 (n) 1 (B) + + .
Denition 9.2.1/1 and the Fubini theorem imply the following lemma.
Lemma 2. If P (E ) < and x E , then

lim inf 1n s E B (x) A+ = 0,
0
where the lower limit is taken over any subset of measure 1 in the interval
0 < < 1.
Lemma 3. If P (E ) < and x E , then

lim sup 1n var E B (x) vn1 .
0
Proof. Let Q be a subset of the interval 0 < < 1 on which the identity
in Lemma 9.2.3/1 is valid. By Lemmas 9.2.3/3 and 2 we have

lim sup 1n var E B (x)

lim sup 1n E B (x) lim sup n
x ds
0
Q
0
E B (x)

= lim n
x ds = vn1 .
0
A B (x)
The result follows.
Taking into account the equality var E ( E \N ) = 0, from Lemmas 1

and 3 we obtain the next assertion.
Lemma 4. If P (E ) < , then s( E \N ) = 0.
Now to prove Theorem 9.2.2/1 it suces to verify that the vector measures
ds and var E (dx) coincide on N .
Lemma 5. Let P (E ) < and let the set B N be measurable relative to
var E . Then s(B) var E (B).
475
Proof. The function

1 1n

var E B (x)
x f (x) = vn1
is lower semicontinuous. This is a consequence of the relation

B (x)\B (x1 )
as x1 x.
Therefore, f (x) is measurable. Let i 0. By Theorem 9.2.4, the sequence

fi converges to f (x) 1 on N . By the Egorov theorem, for any > 0
there exists a set N N such that var E (N ) < and the sequence fi
converges uniformly on N \N . Therefore there exists a > 0 such that

(9.2.18)
var E Br (x) (1 + )vn1 r n1 ,
for all x N \N and r (0, ).
By denition of the measure s, there exists a nite covering of N \N by
balls Bri (xi ) with ri < , such that

(1 + )s B (N \N ) vn1
rin1 .

var E (B) + var E B (N \N )

var E Bri (x) + (1 + )vn1
rin1
+
i
i

+ (1 + )2 s B (N \N ) + (1 + )2 s(B).
The result follows since is arbitrary.
The next assertion is a modication of the classical VitaliCaratheodory

covering theorem.
Lemma 6. Let be a nite measure in Rn concentrated on a set E .
Further, let M be a family of closed balls having the following property: For
each point x E there exists a (x) > 0 such that Br (x) M for all r < (x)
and

rk 0, where Br (x) is any ball in M and , are positive constants
independent of r and x. Then there exists, at most,
a countable family of

mutually disjoint balls B (i) M such that (E \ i B (i) ) = 0.
Proof. We x a number a > 1 and construct a sequence of balls B (i) M in
the following way. Suppose that B (1) , . . . , B (j1) have already been specied.
Then we choose B (j) to satisfy
B (j) B (i) = for i < j,
(j)

a B
sup Br (x) : Br (x) B (i) = , 1 i < j .
476
If the process breaks at some point j then E

proved.
j
i=1 B
(i)
and the lemma is

Suppose the sequence {B (i) } is innite. Let C (i) denote the closed ball
concentric to B (i) with radius Ri = Qri where ri is the radius of B (i) and the
constant Q (1, ) will be specied later. Note that from the very beginning
we could have constructed the sequence {B (i) } so that C (i) is contained in
M simultaneously with B (i) .
We show that

j1

(i)
(i)
B
C
(9.2.20)
E E
i=1
i=j
j1
for some Q and for all j. In fact, let x E \ i=1 B (i) . Then there exists
a ball B M centered at x such that B B (i) = for i < j. Note that
we have B B (p) = for some p j. Indeed, if B B (p) = for all p
then the constructed sequence {B (i) } satises (B) a(B (p) ). Since the
balls B (p) are mutually disjoint, the last inequality contradicts niteness of
the measure .
Let the number p be such that B B (i) = for i < p and B B (p) = .
Inequalities (B) a(B (p) ) and (9.2.19) imply the estimates

r k (B) B (p) rpk ,
where r is the radius of B. Since the balls B and B (p) are disjoint, the distance
between their centers satises

d r + rp rp 1 + (a/)1/k .
Let the constant Q be equal to 1 + (a/)1/k . Then d Rp and hence
x C (p) . The inclusion (9.2.20) follows.
It remains to note that

j1

Qk
(i)
B (i)
C (i)
Rik Qk
rik
B .
E\
i=j
i=1
i=j
i=j
i=j
Since the series

i=1
(B (i) ) converges, we have

E\
j1
(i)
0 as j .
i=1
Lemma 7. If P (E ) < and B is a subset of N , measurable with respect

to var E , then B is s-measurable and
s(B) var E (B).
(9.2.21)
9.3 Extension of Functions in BV () onto Rn
477
Proof. By Theorem 9.2.4, for any (0, 1) the measure = var E

satises the conditions of Lemma 6 with = vn1 (1 ), = vn1 (1 + ),
k = n 1.
By the denition of the Hausdor measure, given any , there exists a
> 0 such that
s(B) vn1
rin1 +
for any nite covering of B by balls Bri (xi ) with ri < .
Let {B (i) } be the sequence of closed balls in Lemma 6. We assume their
radii to be less than . We choose a nite subsequence {B (i) }qi=1 such that

q

(i)
B
< .
B\
i=1
As was shown in the proof of Lemma 1, there existsa nite collection of

disjoint open balls C (j) with radii j < such that ( j C (j) ) < and the

concentric balls 3C (j) with radii 3j form a covering of B\ iq B (i) . Thus

3C (j)
B (i) B.
j
Now we have
iq
n1
n1
+ ,
j +
ri
s(B) vn1 3
j
iq
where ri is the radius of B (i) . Hence

(i)

s(B) (1 + ) c
+ (1 + ) c + (B) + ,
B (j) +
C
j
iq
and (9.2.21) follows because is arbitrary.

Since inequality (9.2.21) is valid for all -measurable sets, this implies that
B is s-measurable.

Combining Lemmas 4, 5, and 7, we complete the proof of Theorem 9.2.2/1.

With any set E we associate the value
(E ) =
inf
B=E
PC (B).
Theorem. (a) If for any function u BV () there exists an extension

u
BV (Rn ) such that
478

u
CuBV () ,
BV Rn
(9.3.1)
where C is a constant independent of u, then

(E ) (C 1)P (E )
(9.3.2)
for any set E .

(b) Conversely, if for any E the inequality (9.3.2) holds with a constant C independent of E , then for any u BV () there exists an extension
u
BV (Rn ) for which (9.3.1) is true.
9.3.1 Proof of Necessity of (9.3.2)
The inequality (9.3.2) is trivial provided P (E ) = . Let P (E ) < . By
hypothesis there exists an extension E of the characteristic function E such
that

E n CP (E ).
BV R
This and formula (9.1.16) imply

CP (E )

E > t} dt
PRn {x :

E > t} dt.
PRn {x :
Since {x :
E (x) > t} = E for t (0, 1), by (9.1.2) and (9.1.3) we obtain
CP (E )
inf
B=E
inf
B=E
PRn (B)
P (B) +
inf
B=E
PC (B) P (E ) + (E ).
Hence (9.3.2) follows.

9.3.2 Three Lemmas on PC (E )
To prove the suciency of (9.3.2) we need the following three auxiliary assertions.
Lemma 1. If B , (B) < , and P (B) < , then there exists a
set E Rn such that E = B and
PC (E ) = (B).
(9.3.3)
Proof. Let {Ei } be a sequence of subsets of Rn such that Ei = B and

lim PC (Ei ) = (B).
(9.3.4)
479
By (9.3.4), supi PC (Ei ) < and since P (Ei ) = P (B) < , we have
supi PRn (Ei ) < . Hence from Theorem 9.1.4 it follows that there exists a
subsequence (for which we retain the notation {Ei }) that converges to some
set E . By Lemma 9.1.3/1,
P (E ) lim inf P (Ei ).
i
Taking into account that E = B as well as equalities (9.1.2) and (9.1.3),

we obtain
PC (E ) lim PC (Ei ) = (B).
(9.3.5)
i
Comparing (9.3.5) with the denition of (B), we arrive at (9.3.3).

Lemma 2. Let E1 , E2 be measurable subsets of Rn . Then
PC (E1 E2 ) + PC (E1 E2 ) PC (E1 ) + PC (E2 ).
(9.3.6)
Proof. Let G be an open set, G C. Then by (9.1.16)

PG (E1 ) + PG (E2 ) (E1 + E2 )BV (G)

=
PG {x : E1 + E2 > t} dt

PG {x : E1 + E2 > t} dt
=
0

PG {x : E1 + E2 > t} dt
= PG (E1 E2 ) + PG (E1 E2 ).
Consider the sequence of open sets Gi such that Gi+1 Gi and
Since
PC (Ek ) = lim PGi (Ek ), k = 1, 2,
(9.3.7)

i Gi
= C.
applying (9.3.7), we obtain (9.3.6).

Lemma 3. Let PC (Ek ) < , k = 1, 2. We put Bk = Ek . Then
PC (E1 E2 ) = PC (E1 ),
PC (E1 E2 ) = PC (E2 ),
provided B1 B2 and
PC (Ek ) = (Bk ),
k = 1, 2.
(9.3.8)
Proof. Since E1 E2 = B1 and (E1 E2 ) = B2 , by the denition

of we have
480
(B1 ) PC (E1 E2 ),
(B2 ) PC (E1 E2 ).
(9.3.9)
Using (9.3.8) we can rewrite (9.3.6) as

PC (E1 E2 ) + PC (E1 E2 ) (B1 ) + (B2 ),

which together with (9.3.9) proves the lemma.

9.3.3 Proof of Suciency of (9.3.2)
1 . Plan of Proof. Starting from Nt = {x : u(x) t} we construct the

family of sets Bt satisfying the conditions Bt = Nt , PC (Bt ) = (Nt ),
Bt B for t > .
We rst construct Bt for a countable set {ti }, which is everywhere dense
on (, ) (item 2 ) and then for all other t (item 3 ). Finally, in item 4 we
introduce the function u
(x) = sup{t : x Bt } and prove that u
(x) satises
the conditions of Theorem 9.3.
2 . Since u BV () then for almost all t we have P (Nt ) < by
formula (9.1.16). Therefore, we can choose a countable set {ti }, ti = tj for
i = j, which is everywhere dense on (, ) and satises P (Nti ) < .
From (9.3.2) it follows that (Nti ) < .
We construct a sequence of sets Bti , i = 1, 2, . . . , such that
(a) Bti = Nti ,
(b) PC (Bti ) = (Nti ),
(c) Bti Btj , ti > tj .
By Lemma 9.3.2/1 there exists a set Bt1 satisfying the conditions (a)(b).
Suppose the sets Bt1 , . . . , Btn have already been constructed so that the conditions (a)(c) are fullled for i, j = 1, . . . , n 1. By Lemma 9.3.2/1 there
exists a set B(n) satisfying (a) and (b). Let t be the largest of those numbers
ti , i = 1, . . . , n 1, for which ti < tn and let t be the smallest of those
numbers ti , i = 1, . . . , n 1, for which tn < ti . We put

Btn = B(n) Bt Bt .
It is clear that Bt Btn Bt . Hence Btn Bti for tn > ti and Btn Bti
for tn < ti , i = 1, . . . , n 1. Since
B(n) = Ntn ,
Bt = Nt Ntn ,
Bt = Nt Ntn ,
we have Btn = Ntn . Applying Lemma 9.3.2/3 to the sets B(n) , Bt , and
then to the sets B(n) Bt , Bt , we obtain
PC (Btn ) = (Ntn ).
Thus the collection of sets Bt1 , . . . , Btn satises the conditions (a)(c) for
i, j = 1, . . . , n.
481
3 . Let t
/ {ti }. From the set {ti } we select two monotone sequences {i },
{i } such that i < t < i and limi i = limi i = t.
(0)
(0)
By Lemma 9.3.2/1, there exists a set Bt such that Bt = Nt and

(0)
(k)
(0)
PC (Bt ) = (Nt ). Consider the sequence of sets Bt = Bt Bk , k =
(k)
(k+1)
(k)
Bt . Using Lemma 9.3.2/3,
1, 2, . . . . It is clear that Bt = Nt , Bt
(k)
have PC (Bt ) = (Nt ). We introduce the notation
t =
B
(k)
Bt .
k=1
(k)
Since Bt
t as k , we see that
B

t ) lim inf PC B(k) = (Nt ).
PC (B
t
k
t ). Thus
t = Nt . Therefore (Nt ) PC (B
On the other hand, B
PC (Bt ) = (Nt ).
(k)
t B , k = 1, 2, . . . . In the
Next consider the sequence of sets Ct = B
k
(k)
same way as when we considered the sets Bt we conclude that the set
Bt =
(k)
Ct ,
k=1
is measurable and satises the conditions

1.
Bt = Nt ,
2.
PC (Bt ) = (Nt ),
3. Bi Bt Bi ,
i = 1, 2, . . . . Now let t, be arbitrary numbers, t < . Then condition 3

implies that Bt B .
4 . Consider the function u
dened by u
(x) = sup{t : x Bt }. We put

(x) t ,
Ct = x : u
(x) > t .
At = x : u
Obviously, At Bt Ct . The sets At \Ct are mutually disjoint for dierent t
and hence mn (At \Ct ) = 0 for almost all t. Thus the sets At , Ct are measurable
for almost all t. Moreover,
PRn (At ) = PRn (Bt ) = PRn (Ct ).
We prove that u
is locally integrable. It is well known that the inequality
(n1)/n
mn (E )
C(R, )PBR (E )
(9.3.10)
is valid for the subset E of the ball BR such that mn (E ) < mn (BR ) .
(In particular, this follows from Lemma 5.2.1/1.) Let the closed ball B be
contained in and let BR B . Then (9.3.10) implies
482

mn (E ) C(R, ) PBR (E ) + mn (E B )

C(R, ) PRn (E ) + mn (E B )
for any set E BR . Putting E = Bt BR for t 0 and E = BR \Bt for t < 0
in the latter inequality and using PC (Bt ) = (Nt ) and estimate (9.3.2),
we obtain

mn (Bt BR ) C(R, ) CP (Nt ) + mn (Nt B ) , t 0,

mn (BR \Bt ) C(R, ) CP (Nt ) + mn (\Nt ) B , t < 0.
Taking into account that mn (Bt ) = mn (At ) for almost all t, from the latter
two inequalities we obtain

0
mn (At BR ) dt +
mn (BR \At ) dt
0

P (Nt ) dt +
mn (Nt B ) dt
C(R, ) C
0

+
mn (B \Nt ) dt ,
0
which is equivalent to

|
u| dx C(R, ) CuBV () +
BR

|u| dx ,
B
whence the local integrability of u

follows. Applying (9.1.16), (9.3.2), and
recalling that PRn (Ct ) = PRn (Bt ) for almost all t, we obtain

P (Bt ) + PC (Bt ) dt
PRn (Ct ) dt =

u
=
BV Rn

P (Nt ) + (Nt ) dt C
=
P (Nt ) dt = CuBV () ,
i.e., u
BV (R ) and (9.3.1) is valid.
n
9.3.4 Equivalent Statement of Theorem 9.3

,
Theorem 9.3 can be rephrased in terms of the extension operator A : u u
which associates with each u BV () its extension u
BV (Rn ).
We put

uBV (Rn )
A = sup
: u BV () ,
uBV ()
and denote by || the inmum of those numbers k for which (E ) kP (E )
for all E . Now we have the following theorem.
Theorem. The operator A exists and is bounded if and only if || < .
Moreover, A 1 + || for any extension operator A and there exists an
operator A with A = 1 + ||.
483
9.3.5 One More Extension Theorem

The condition (9.3.2) in Theorem 9.3 is of a global nature. For example,
nonconnected sets do not satisfy it.
This impediment may be removed if we make the requirements on the
extension operator less restrictive. Specically, the following theorem holds.
Theorem. Let be a bounded open set. For any function u BV () to
have an extension u
BV (Rn ) with

(9.3.11)

uBV (Rn ) K uBV () + uL1 () ,
where K is independent of u, it is necessary and sucient that there exists a
> 0 such that (E ) CP (E ) for any E with a diameter less than
, the constant C being independent of E .
Proof. Necessity. Let E and let E be the characteristic function of
E while
E is an extension of E satisfying (9.3.11). We have
1

E BV (Rn )
P {x :
E > t} dt.
K P (E ) + mn (E )
0
Since {x :
E > t} = E for t (0, 1), it follows that

K P (E ) + mn (E ) inf P (B).
B=E
By the inclusion E B, the latter estimate and the isoperimetric inequality (9.1.15) imply

(n1)/n
K P (E ) + mn (E ) nvn1/n mn (E )
.
(9.3.12)
We put = n/(2K). Then from (9.3.12) under the condition diam E < it
follows that mn (E ) P (E ). Therefore,
2KP (E )
inf
B=E
P (B) (E ).
Suciency. Consider the partition of unity i (x), i = 1, . . . , , such that
supp i ,
diam supp i < ,
i=1
and |gradi | d = const. Let u BV (). We put i = ui and

Nt = {x : |i | t}. Since for all t = 0 we have diam Nt < then
(Nt ) CP (Nt ). Therefore, following the same argument as in the proof
of suciency in Theorem 9.3 we obtain the function i BV (Rn ) such that
i = i in and

i BV (Rn ) (C + 1)i BV () (C + 1) uBV () + duL1 () .

We put u
=
i . It is clear that u
= u in and

uBV (Rn ) (C + 1) uBV () + duL1 () .

484
9.4 Exact Constants for Certain Convex Domains

By Theorem 9.3 the norm of the extension operator BV () BV (Rn ) is
expressed by the exact constant in the isoperimetric inequality (9.3.2). This
constant can be found in some particular cases. For plane convex domains it
has a simple geometrical interpretation (cf. Corollary 9.4.4/2). The constant
is also easily calculated if is an n-dimensional ball.
9.4.1 Lemmas on Approximations by Polyhedra
Lemma 1. Let be a bounded convex domain in Rn and let E ,
PRn (E ) < . Then there exists a sequence of polyhedra k such that k E
and
lim P (k ) = P (E ),
lim PC (k ) = PC (E ).
(9.4.1)
Proof. Let be the domain obtained from by the similarity transformation with coecient 1 + and with the center at a xed point of . We
denote the image of E under the same transformation by E . It is clear that
P (E ) = (1 + )n1 P (E ),
PC (E ) = (1 + )n1 PC (E ).
Hence we can easily obtain that

lim P (E ) = P (E ),
lim PC (E ) = PC (E ).
(9.4.2)
In fact, (1 + )n1 P (E ) P (E ) and consequently

P (E ) lim inf P (E ) P (E ).
0
The latter inequality is a corollary of Lemma 9.1.3/1. Since

P (E ) + PC (E ) = P (E ),
P (E ) + PC (E ) = P (E ),
the rst inequality (9.4.2) implies the second. For almost all we have
var E () = 0.
(9.4.3)
Let be subject to (9.4.3). By Theorem 9.1.3 there exists a sequence of polyhedra k, such that k, E , P (k, ) P (E ) as k . This and
Lemma 9.1.2/3 yield
weakly
var k, var E .
By (9.4.3) we have
485
lim sup var k, () var E () = 0,
and therefore,
lim P (k, ) = P (E ),
lim PC (k, ) = PC (E ).
(9.4.4)
We choose a sequence of numbers i that satisfy (9.4.3), such that i 0 as

i . Then (9.4.2) and (9.4.4) imply
lim lim P (k,i ) = P (E ),
i k
lim lim PC (k,i ) = PC (E ).
i k
This concludes the proof.

In the following we shall use the following elementary assertion.
Lemma 2. Let be a bounded convex domain in Rn and let be a nite

polyhedron. Then s( C) s( ).
Lemma 3. Let be a bounded convex domain in Rn and let E ,
P (E ) < . Then there exists a sequence of polyhedra k such that k E
and
lim P (k ) = P (E ),
lim PC (k ) = PC (E ).
Proof. By Lemma 1, there exists a sequence of polyhedra k , k E ,

satisfying (9.4.1). It is clear that P (k ) = P (k ). According to
Lemma 9.3.2/3 we have
PC (k ) PC (k ).
Therefore,
lim PC (k ) lim PC (k ) = PC (E ),
(9.4.5)
lim P (k ) = lim P (k ) = P (E ).
(9.4.6)
Since k E , we obtain
PRn (E ) lim PRn (k ).
k
(9.4.7)
From (9.4.6) and (9.4.7) we conclude that

PC (E ) lim PC (k ),
k
486
9.4.2 Property of PC
Lemma. Let P () < and suppose a normal to exists s-almost everywhere on . Then, for any set E ,
PC (E ) + PC (\E ) = s().
Proof. By the equality = E + \E we have
var (C) var E (C) + var \E (C) = PC (E ) + PC (\E ).
Since a normal to exists s-almost everywhere on , by Theorem 9.2.2/1
we obtain
var (C) = PRn () = s().
Consequently,
s() PC (E ) + PC (\E ).
We prove the reverse inequality. Let A , B denote the reduced boundaries of
the sets E and \E , respectively. The sets A and B are s-measurable (cf.
Theorem 9.2.2/1). We note that the sets A and B are disjoint.
In fact, suppose there exists a point x common to A and B . Then
the volume densities of E and \E at the point x are equal to 1/2. This is
impossible because x . Consequently,

s A + s B s().
It remains to use the equalities

s A = s A = PC (E ),

s B = s B = PC (\E )
(cf. Theorem 9.2.2/1). The lemma is proved.
9.4.3 Expression for the Set Function (E ) for a Convex Domain

Theorem. If is a bounded convex domain in Rn , then the equality

(E ) = min PC (E ), PC (\E )
(9.4.8)
holds for any set E with P (E ) < .
Proof. For the sake of deniteness, let
PC (E ) PC (\E ).
Let the set B be such that B = E , PC (B) = (E ). Assume for the
moment that mn (B) < .
487
By Lemma 9.4.1/1, we can nd a sequence of polyhedra k , k B such

that
lim P (k ) = P (B),
lim PC (k ) = PC (B).
(9.4.9)
Since mn (B) < , the polyhedra k are nite. By Lemma 9.4.1/2 we have
PC (k ) PC (k ). This and (9.4.9) yield
lim sup PC (k ) PC (B).
(9.4.10)
Using k B , we obtain
P (B ) lim inf P (k ),
k
which together with (9.4.9) implies

PC (E ) = PC (B ) lim inf PC (k ).
k
Hence from (9.4.10) it follows that PC (E ) PC (B) = (E ). Thus

PC (E ) = (E ).
Now let mn (B) = . Since PC (B) < , we have mn (CB) <
(cf. (9.1.18)). Further we note that
PC (CB) = PC (B) = (E ) = (\E ).
Hence, according to what we proved earlier, (\E ) = PC (\E ) and therefore, by (9.4.8),
(E ) = PC (\E ) PC (E ).
Since, obviously, (E ) PC (E ), we conclude that (E ) = PC (E ).
9.4.4 The Function || for a Convex Domain

Corollary 1. Let be a bounded convex domain. Then

|| = inf k : PC (E ) kP (E ) ,
where E is any subset of with PC (E ) 12 s().
The result follows immediately from Theorem 9.4.3 and Lemma 9.4.2.
Corollary 2. Let be a bounded convex domain in R2 . Then
|| =
1
s(),
2h
where h is the minimum length of those lines whose end points separate
into arcs of equal length.
488
Proof. We take an arbitrary > 0. Let E be a measurable subset of

such that
P (E ) > 0,
PC (E ) <
1
s(),
2
and
||
PC (E )
<
P (E )
(cf. Corollary 1). By Lemma 9.4.1/3, we can nd a polyhedron such that

||
PC ( )
< 2.
P ()
Let A and B be the set of points of the intersection of with some component of the boundary of . Points A and B can be chosen so that the
segment of the component of being considered, bounded by points A and
B, lies entirely in . The segment AB divides into two sets Q and Q . Let
PC (Q) PC (Q ). It is clear that
PC (Q)/AB PC ( )/P ()
and therefore
PC (Q)
< 2.
(9.4.11)
AB
If PC (Q) = PC (Q ) then (9.4.11) implies the required assertion by virtue
of the inequality AB h and the fact that is arbitrary.
Let PC (Q) < PC (Q ). We shift the segment AB parallel to itself to a
new position A B (A , B ) so that PC (Q1 ) = PC (Q1 ) where
Q1 and Q1 are the domains into which the segment A B divides .
Elementary calculations show that
||
PC (Q1 )/A B PC (Q)/AB,

which together with (9.4.11) proves the corollary.
Lemma. Let be the unit ball in Rn . Then

inf PC (E ) : E , P (E ) = p = const n
equals the area of the spherical part of the boundary of the spherical segment
whose base has area p.
Proof. Let E , P (E ) = p. By Lemma 9.4.1/3 there exists a sequence
of polyhedra k , k E , such that
P (k ) p,
PC (k ) PC (E ).
We perform the spherical symmetrization of k relative to some ray

l with origin at the center of . We obtain the set k , symmetric relative to
l, with a piecewise smooth boundary and such that
9.5 Rough Trace of Functions in BV () and Certain Integral Inequalities
P (k ) P (k ),
489
PC (k ) = PC (k ).
We denote the spherical segment whose spherical part of the boundary is

k by Qk . It is clear that
P (Qk ) P (k ),
PC (Qk ) = PC (k ).

Hence the result follows.
The next assertions can be obtained from the Lemma by simple calculations.
Corollary 3. (a) If is the unit ball in Rn then
|| = n /2vn1 .
(b) If is the unit ball in R3 then, for any E ,

4P (E ) PC (E ) 4 PC (E ) .
(c) If is the unit disk then, for any E ,

1
P (E ) 2 sin PC (E ) .
2
9.5 Rough Trace of Functions in BV () and Certain

Integral Inequalities
9.5.1 Denition of the Rough Trace and Its Properties
On the reduced boundary of we dene the rough trace u of a function
u BV (). We put

u (x) = sup t : P (Nt ) < , x Nt ,
where x and Nt = {x : u(x) t}. (The supremum of the empty
set is assumed to be .)
It is clear that if u has a limit value at a point x then
u (x) = lim u(y).
yx
Lemma 1. Let s() < . Then P (E ) < implies P (E ) < for any
E .
Proof. Since s() is nite, we can construct a sequence of polyhedra
k , k , such that s(k ) K < . Since P (E ) < , we have
490
P (E k ) K1 < . Moreover, E k E . Hence the result follows by

Lemma 9.1.3/1.

Corollary 1. If s() < and u BV (), then P (Nt ) < for almost
all t.
Lemma 2. Let s() < and u BV (). Then the rough trace u is
s-measurable on and

s x : u (x) t = s Nt
(9.5.1)
for almost all t.
In fact, instead of (9.5.1) we prove that, for all t R except for a countable
subset,

s x : u (x) t Nt = 0,

where A B = (A \ B) (B \ A ).
Proof. Denote Bt = {x : u (x) t}, Yt = Nt , and Xt = Bt \ Yt .
One can see that Yt Bt . Thus, it remains to prove that s(Xt ) = 0.
The sets Yt are measurable and the sets Xt are disjoint. It is clear that the
inclusions Yt0 Yt1 and Yt0 Xt0 Yt1 Xt1 hold for t0 < t1 , which implies
Yt0 Xt1 . Thus

Yt \ Yt1 Xt1 .
t<t1
On the other
hand, the sets (t<t1 Yt )\Yt1 are measurable and disjoint. There
fore s(( t<t1 Yt ) \ Yt1 ) = 0 for almost all t1 R. This implies that the sets
Xt (being subsets of measure zero sets) are measurable and of measure zero
for almost all t R. Thereby the sets Bt are measurable.

For a set E , denote by x (E ) its relative density at the point x; i.e,

mn (E B(x, ))

x (E ) = lim
.
0 mn (
B(x, ))
In the most interesting case x , we have

x (E ) = 2vn1 lim n mn E B(x, ) .
0
x (E ) and (E ) are dened similarly.

Upper and lower relative densities
x
Lemma 3. Let P () < , E , P (E ) < . Then the density of E
is equal either 0 or 1 for s-almost all x .

Proof. It is clear that x (E ) = 1 for x E and x (E ) = 0 for
x \ E . We put

1
Ck = x : x , < x (E ) < 1 , k = 2, 3, . . . .
k
491
It remains to prove that s(Ck ) = 0, k = 2, 3, . . . . Using the inclusion Ck

, for x Ck we deduce

x (E ) = 2 lim sup mn E
B (x) k 1 .
(9.5.2)
vn 0
By Lemma 1 we have P (E ) < . Therefore, (9.1.18) implies

n/(n1)
B (x) C var Rn E B (x)
mn E
.
Comparing this with (9.5.2), we obtain

lim sup 1n var E B (x)
0
vn
2kC
(n1)/n
.
(9.5.3)

Since Ck E = , we have var E (Ck ) = 0. Thus (9.5.3) along with
Lemma 9.2.5/1 yields s(Ck ) = 0 and the result follows.

Lemma 4. For any u BV () and almost all x

u (x) = (u) (x).
(9.5.4)
Proof. This result is equivalent to the fact that for almost all x

sup t : x Nt = inf t : x ( \ Nt ) .
This equality means that

sup t : Nt (x) = 1 = inf t : (\Nt ) (x) = 1 .
In its turn, this is equivalent to the equality

sup t : Nt (x) = 1 = inf t : Nt (x) = 0 .
Denote by L and R the left and the right terms of the last equality. It is easy to
see that Nt (x) is a nonincreasing function on t. So L R. Consider the set of
points x such that L(x) < R(x). It suces to prove that the s-measure of this
set is zero. Let {ti }
i=1 be a countable set dense in R such that P (Eti ) < .
If L(x) < R(x), then there exists ti such that L(x) < ti < R(x). Now our

statement follows from the equality s{x : 0 < Nt (x) < 1} = 0.

Corollary 2. For any u BV () and for almost all x
+ +

u
= u
,
u
= u
,
(9.5.5)
and as a result u = (u+ ) (u ) .

Proof. The rst equality is obvious. The second one follows from Lemma
4. Indeed, (u ) = ((u)+ ) = ((u) )+ = (u )+ = (u ) .
492
9.5.2 Integrability of the Rough Trace

Theorem. Suppose that P () < and that a normal to exists s-almost
everywhere on . In order for any u BV ()

u cs(dx) kuBV () ,
(9.5.6)
inf
c
where k is independent of u, it is necessary and sucient that the inequality

min PC (E ), PC (\E ) kP (E )
(9.5.7)
holds for any E .
Proof. Necessity. Let E , P (E ) < . By Lemma 9.5.1/1 we have
P (E ) < . Let E be the characteristic function of the set E . Then

E (x) cs(dx)
inf
c

= min |1 c|s E + |c|s \ E
c

= min PC (E ), PC (\E ) .
= min s E , s \ E
(The preceding equality is valid since, by hypothesis, s(\ ) = 0.)
On the other hand, E BV () = P (E ). Applying (9.5.6), we arrive
at (9.5.7).
Suciency. Let u BV (). It is clear that s( Nt ) is a nonincreasing
function of t. In fact, let x Nt and let < t. We have
1 = lim 2vn n mn ( B ) lim 2vn n mn (N B )
0
lim 2vn
0
0
mn (Nt B ) = 1,
i.e., x N .
Similarly, s(\ Nt ) is a nondecreasing function of t. From (9.1.16) we
obtain

P (Nt ) dt
min s Nt , s \ Nt dt.
kuBV () = k
Putting

t0 = sup t : P (Nt ) < , s Nt s \ Nt ,
we obtain

kuBV ()

s Nt dt +
t0

=
t0
=

=
t0
493

s \ Nt dt
t0

s x : u (x) t dt +
s x : u (x) t dt

+

u (x) t0 s(dx) +
u (x) t0 s(dx)

u (x) t0 s(dx).
Consequently,

kuBV () inf
c

u cs(dx),
From Corollary 9.4.4/1 we obtain that the best constant in (9.5.8) is equal
to || provided is a convex domain. In particular, for a plane convex domain
this constant coincides with the ratio of 12 s() to the length of the smallest
chord dividing into arcs of equal length (Corollary 9.4.4/2). According to
Corollary 9.4.4/3, the best constant in (9.5.6) equals n /2vn1 for the unit
ball.
9.5.3 Exact Constants in Certain Integral Estimates for the Rough
Trace
()
Let
Denition 1. Let A .
A denote the inmum of those k for which
[PC (E )] kP (E ) holds for all E that satisfy

mn (E A ) + s A E = 0.
(9.5.8)
Theorem. Let P () < and suppose a normal to exists s-almost
everywhere on . Then, for any function u BV () such that u(A ) = 0
and u (A ) = 0, the inequality

u s(dx) (1) uBV ()
(9.5.9)
A
(1)
holds. Moreover, the constant A is exact.

Proof. We have

u s(dx) =

s x : u t dt +

s x : u t dt.
By (9.1.16), the rst integral on the right is equal to
494

s Nt dt =
PC (Nt ) dt.
0
Note that mn (A Nt ) + s(A Nt ) = 0 for almost all t. Consequently, by

(1)
the denition of A ,

(1)
s x : u t dt
PC (Nt ) dt A
P (Nt ) dt.
0
Similarly we have

s x : u t dt
Finally,
(1)
PC (\Nt ) dt A
P (Nt ) dt.

u s(dx) (1) uBV () .
(1)
To see the sharpness of the constant A it suces to put u = E

into (9.5.9), where E is a set satisfying (9.5.8).

Denition 2. We introduce the function

[PC (E )]
(S) = sup
: E , P (E ) > 0, PC (E ) S .
P (E )
The preceding theorem implies the following obvious corollary.
Corollary. Suppose that P () < and that a normal to exists salmost everywhere on . Then for any u BV () with s({x : u (x) =

0}) S the inequality

u s(dx) 1 (S)uBV ()
holds. Moreover, the constant 1 (S) is exact.

From Lemma 9.4.4 it follows that for a ball the function 1 (S) coincides
with the ratio of S to the area of the base of the spherical segment whose
spherical part of the boundary has the area S.
In particular, 1 (S) = S/(2 sin(S/2)) for n = 2 and 1 (S) = 4/(4 S)
for n = 3.
Lemma. Let be a domain with P () < and suppose that a normal
to exists s-almost everywhere on . Then

def
(S) = inf P (E ) : E , PC (E ) S, PC (\E ) S > 0.
Proof. Let {Ei } be the minimizing sequence for (S). If
495

lim inf min mn Ei , mn (\Ei ) > 0,
i
then the result follows from Theorem 9.1.3 and Lemma 5.2.4. Let this lower
limit be equal to zero and, for the sake of deniteness, let mn Ei 0. Then
mn (\Ei ) mn (). By Lemma 9.1.3/1 we obtain
lim inf P (\Ei ) P () = s()
i
and by Lemma 9.4.2,

s() = PC (Ei ) + PC (\Ei ).
Moreover, we always have
P (\Ei ) = P (\Ei ) + PC (\Ei ).
Thus, for any > 0 and for large enough i, we obtain
P (\Ei ) PC (Ei ) S ,
i.e., inf P (\E ) S. The lemma is proved.
We can easily see that the function introduced in the preceding lemma
is connected with by the inequality
(S) = S /(S).
The same lemma immediately implies that (S) is nite for all S
(0, s()) provided is a domain with P () = s() < and (S) <
for some S < s().
Hence from Theorem 9.5.2 we conclude that (9.5.6) holds if and only if
i (S) < for some S (0, s()).
9.5.4 More on Integrability of the Rough Trace
Theorem. Let P () < and suppose that a normal to exists s-almost
everywhere on . For any function u BV () to satisfy the inequality

u
k uBV () + uL1 () ,
(9.5.10)
L()
where the constant k is independent of u, it is necessary and sucient that
there exist a > 0 such that for each measurable set E with diameter
less than the inequality
PC (E ) k1 P (E )
holds where k1 is a constant independent of E .
(9.5.11)
496
Proof. The necessity of (9.5.11) easily follows by the insertion of u =

E into (9.5.10) and then by application of the isoperimetric inequality. The
suciency results from Theorem 9.5.3 if we use a partition of unity (cf. the
proof of Theorem 9.2.2/2).

Remark. If each of the sets 1 , 2 satises the hypothesis of the preceding

theorem then their union has the same property.
The proof follows from formula (9.3.6).
9.5.5 Extension of a Function in BV () to C by a Constant
In the present subsection we assume that P () < and s(\ ) = 0.
We introduce the notation uc (x) = u(x) for x , uc (x) = c for x C
where c R1 .
Lemma. The equality

uc BV (Rn ) = uBV () + u cL1 ()
(9.5.12)
holds.
Proof. We have
uc BV (Rn ) =
0
=

P {x : |uc c| > t} dt

P {x : |u c| > t} dt

PC {x : |u c| > t} dt.
(9.5.13)
It is clear that

P {x : |u c| > t} dt = uBV () .
(9.5.14)
Further, since s(\ ) = 0, we see that

PC {x : |u c| > t} dt
0
0

s x : (u c) > t dt +
s x : (u c) < t dt

(u c) s(dx) =
|u c|s(dx),
=
0
=
which together with (9.5.13) and (9.5.14) implies (9.5.12).

() denote the subset BV () which contains functions with
Let B V
497
u0 BV (Rn ) = uBV () .

() if and only if u = 0 for the
Then from (9.5.12) it follows that u B V
class of domains under consideration. Thus the elements of the quotient-space
() are classes of functions that have the rough traces u .
BV ()/B V
Formula (9.5.12) and Theorem 9.5.2 imply the next assertion.
Corollary 1. If, for any E ,

min PC (E ), PC (\E ) kP (E ),
(9.5.15)
where k is a constant that is independent of E , then there exists a c such that

uc BV (Rn ) (k + 1)uBV () .
(9.5.16)
Conversely, if for each u BV () there exists a c such that (9.5.16) holds

with k independent of u, then (9.5.15) holds for any E .
Corollary 2. For the inequality

u0 BV (Rn ) k uBV () + uL1 () ,
(9.5.17)
where k is independent of u, to hold for any u BV (), it is necessary and

sucient that there exists a > 0 such that for any measurable set E with
diam E < the inequality PC (E ) k1 P (E ) holds where k1 is independent
of E .
Proof. The necessity follows immediately from (9.5.12) and the isoperimetric inequality. The suciency results from (9.5.12) and Theorem 9.5.4.

Inequality (9.3.6) implies the following corollary.

Corollary 3. If each of the sets 1 , 2 satises the hypothesis of Corollary 2, then their union has the same property.
In particular, this implies that any function in BV () can be extended by
zero to the whole space so that (9.5.17) is valid for domains that are nite
unions of domains in C 0,1 .
9.5.6 Multiplicative Estimates for the Rough Trace
Let P () < and suppose a normal to exists s-almost everywhere on .
by () , (S) we mean the functions introduced
Let A denote a subset of ;
A
in Denitions 9.5.3/1 and 9.5.3/2.
The following assertion supplements Theorem 9.5.3.
(1/q )
Theorem. 1. If A
such that
< , where q 1, then, for any u BV ()
498
u(x) = 0
u (x) = 0
for x A ,
the inequality
for x A ,

Cu1
BV () u L
Lq ()
(9.5.18)
(9.5.19)
t ()
holds where 0 < t < q < q ,

=
t(q q)
,
q(q t)
(9.5.20)
(1/q )
and C q(q t)/q (qt) cA

.
2. If for all u BV () satisfying (9.5.18) the inequality (9.5.19) holds
with specied by (9.5.20) and with q > q, q > t, then
(1/q )
C q(q
t)/q (qt)
Proof. 1. Following the same line of reasoning as in the proof of Theorem 9.5.3, we obtain

1/q

(1/q )
d A
uBV () ,
(9.5.21)
s( )
0
where = {x : |u (x)| }.
In Lemma 1.3.5/2 we put = tq , f () = s( ), b = 1/q , a (1, ),
= a(q t)/q, and = (q q)/q q. Then

s( )
q1
d c
a
s( )
(q q)/q(q t)
at1

s( )
1/q
(9.5.22)
q (qt)/(q t)
d
Since a > 1 and the function s( ) does not increase, Lemma 1.3.5/1 can be
applied to the rst factor. Then we have

a

a
s( ) at1 d c
s( ) t1 d
0
0

a
t
u (x) s(dx) .
=c
(9.5.23)
Combining (9.5.21)(9.5.23) we arrive at item 1 of the theorem.

2. The lower bound for the constant C results by insertion of E , where
E satises (9.5.8), into (9.5.19).

Let us consider two domains for which we can obtain exact conditions for
()
the boundedness of the function A .
499
Example 1. Let x = (x , xn ), x Rn1 and let

= x : 1 < xn < |x | , |x | < 1 , 0 < < n 2.
Further let A = {x : xn = 1, |x | < 1}. We show that A < for =
(n 1)/(n 2 ). (Using the sequence of sets Em = {x : xn > m},
m = 1, 2, . . . , we can prove that this is the best possible.)
Although is not convex we can apply to it the proof of Lemma 9.3.2/3.
Therefore, it suces to verify the estimate
()
(n1)/(n2)
P (E )
cP (E )
(9.5.24)
for any set E that is the intersection of a polyhedron with , E A =

. Since (9.5.24) has already been obtained in Example 6.11.3/1, the result
follows.
Example 2. Using the same argument and referring to Example 6.11.3/2,
we can show that for

= x : |x | < xn , 0 < xn < 1 , 1,
and for A = {x : |x | < 1, xn = 1} the value A is nite for = (n
1)/((n 2) + 1) and that is the best as possible.
()
The theorem of the present subsection implies that boundedness of the

value 1/q (S) is necessary and sucient for the validity of (9.5.19) for any
function u BV () with s({x : u (x) = 0}) S (cf. Corollary 9.5.3). Hence
we easily conclude that the boundedness of 1/q (S) for some S < P () is
necessary and sucient for the validity of the inequality

1

u
u
C1 uBV () + C2 u
,
(9.5.25)
Lq ()
Lr ()
Lt ()
where u is any function in BV (), r < q and q , q, t, are the same as in

the theorem of the present subsection (cf. Theorem 6.3.3).
9.5.7 Estimate for the Norm in Ln/(n1) () of a Function in
BV () with Integrable Rough Trace
To conclude this section we prove an assertion similar to Corollary 5.6.3.
Theorem. Suppose that P () < and that a normal to exists s-almost
everywhere on . Then for any u BV () the inequality

uLn/(n1) () nvn1/n uBV () + u L ()
(9.5.26)
1
1/n
is valid. Moreover, the constant nvn
is exact.
500
Proof. By (5.6.14) we have

(n1)/n
mn (Nt )
dt +
uLn/(n1) ()
mn (\Nt )
(n1)/n
dt,
(9.5.27)
where Nt = {x : u(x) t}. By the isoperimetric inequality (9.1.15) we obtain

(n1)/n

mn (Nt )
nvn1/n P (Nt ) = nvn1/n P (Nt ) + s Nt .
Since s( Nt ) = s({x : u t}) for almost all t (Lemma 9.5.1/2), we
see that

(n1)/n

mn (Nt )
dt nvn1/n
P (Nt ) dt +
s x : u t dt .
0
Taking into account that PC (Nt ) + PC (\Nt ) = P () by Lemma 9.4.2,

we similarly obtain that
0
(n1)/n

dt
mn (\Nt )
nvn1/n

P (Nt ) dt +

s x : u t dt .
Consequently,

(n1)/n
|u|n/(n1) dx

nvn1/n
P (Nt ) dt +
s x : u t dt
0

u s(dx) .
= nvn1/n uBV () +
The sharpness of the constant in (9.5.26) is a corollary of the fact that (9.5.26)
becomes an equality for u = B where B is a ball in .

Similarly to the preceding theorem, we can generalize Theorem 5.6.3 to

functions in BV ().
9.6 Traces of Functions in BV () on the Boundary and

GaussGreen Formula
9.6.1 Denition of the Trace
Let be an open set in Rn and let the function u be integrable in a neighborhood of a point x . The upper and the lower traces of the functions
u at the point x are the numbers
9.6 Traces of Functions in BV () on the Boundary and GaussGreen
501
1
u(y) dy,
0 mn (B (x) ) B (x)

1
u(y) dy,
u(x) = lim inf
0 mn (B (x) ) B (x)

u
(x) = lim sup
respectively.
(x) of the function
If u
(x) = u(x), this common value is called the trace u
u at the point x .
9.6.2 Coincidence of the Trace and the Rough Trace
Lemma. Let u BV (), u 0 and let

u (x)s(dx) < .
Then for any x the inequality

u(x) u (x)
(9.6.1)
holds.
Proof. By Theorem 9.5.7, the function u is integrable in and hence the
function u(x) is dened.
Inequality (9.6.1) is trivial provided u (x) = 0. Suppose 0 0 and choose t to satisfy 0 0 such that
1<
mn (Nt Br (x))
1
mn ( Br (x))
for 0 < r < r0 (x). Since

u(y) dy =
(9.6.2)

mn N Br (x) d,
Br (x)
(9.6.2) implies
1
mn (Br )

Br
t
1
mn (N Br ) d
mn (Br ) 0
mn (Nt Br )
(1 )t,
t
mn ( Br )
u(y) dy
which proves (9.6.1) for u (x) < .

In the case u (x) = the arguments are similar.
502

everywhere on . If u BV () and

u s(dx) < ,
then the trace u

of the function u exists s-almost everywhere on and
coincides with the rough trace u .
Proof. By Theorem 9.5.7, the function u is integrable in and consequently, the upper and lower traces u
and u are dened.
First, consider the case of a nonnegative function u. Then by Lemma,
u(x) u (x) for all x . Next we prove that the inequality u
(x) u (x)
holds for s-almost all x , if u 0. Suppose that

(x) > u (x) > 0.
s x : u
(x) >
Then there exists a c > 0 such that s(Q) > 0 where Q = {x : x , u
u (x) + c}. Recalling the denition of u
(x), for x Q we have

1
c + u (x) lim sup
mn Nt B (x) dt.
0 mn (B (x) ) 0
Since x , we see that
lim
0

2
mn B (x) = 1.
vn n
(9.6.3)
Therefore,

2
lim n
mn Nt B (x) dt
vn 0
0
(n1)/n

(n1)/n

2
1n
lim
dt.
mn Nt B (x)
0
vn
0
c + u (x)
(9.6.4)
The equality (9.6.3) implies

mn Nt B (x) min mn Nt B (x) , mn B (x)\Nt ,
where does not depend on t and 1 as 0. Applying the relative
isoperimetric inequality (9.1.18), we obtain

mn Nt B (x)
Noting that
n1
n
n1
n
vn
2
n1
n

1
vn1
var Nt B (x) .
(9.6.5)
503

var Nt (B ) = var Nt (B ) + s Nt ,
and integrating (9.6.5) with respect to t, we obtain

(n1)/n
mn Nt B (x)
dt
0

n1

vn n 1
n
vn1
var Nt B (x) dt
2
0

+
s Nt dt
n1
n
0
n1
n
=
vn
2
n1
n
1
vn1
var u B (x) +
u (y)s(dy) .
B (x)
(9.6.6)
Comparing (9.6.4) and (9.6.6) and taking into account that 1 as 0
we obtain

1
lim sup 1n var u B (x)
c + u (x) vn1
0

+ lim sup 1n
u (y)s(dy) .
(9.6.7)
B (x)
0
By (9.2.13),

lim 1n var B (x) = vn1
0
for s-almost all x . On the other hand, var (B ) = s( B ).

Therefore, for s-almost all x Q inequality (9.6.7) can be rewritten in the
form

1
c + u (x) vn1
lim sup 1n var u B (x)
0

1
u (y)s(dy). (9.6.8)
+ lim sup
0 s( B (x)) B (x)
The integral
u (y)s(dy)
I(E ) =
E
is absolutely continuous relative to the measure s(E ). Thus the derivative

1
dI
(x) = lim
u (y)s(dy) = u (x)
0 s( B (x)) B (x)
ds

exists for s-almost all x (see, for instance, Hahn and Rosenthal [336],
p. 290). Therefore, for s-almost all x Q the inequality (9.6.8) can be rewritten as
504

cvn1 lim sup 1n var u B (x) .
(9.6.9)
0
Since var u(Rn ) < and var u(Q) = 0, Lemma 9.2.5/1 and (9.6.9) imply
s(Q) = 0. The assertion is proved.
Now let u be an arbitrary function in BV (). Then the functions
u+ =

1
u + |u| ,
2
u =

1
|u| u ,
2
are also in BV (). By what we proved previously, the equalities

u
+ (x) = u+ (x) ,
u
(x) = u (x) ,
(9.6.10)
of the function
hold for s-almost everywhere on . Consequently, the trace u
u exists s-almost everywhere on . Moreover,
(x).
u
(x) = u
+ (x) u
(9.6.11)
Further, by (9.5.5) we have

u (x) = u+ (x) u (x) .
(9.6.12)
Comparing equalities (9.6.10)(9.6.12), we conclude that u

(x) = u (x).
9.6.3 Trace of the Characteristic Function

The hypothesis of Theorem 9.6.2 may be weakened for the characteristic function. Namely, the following lemma holds.
E
Lemma. Let P () < , E , and P (E ) < . Then the trace of
of the function E exists and coincides with (E ) for s-almost all x .
It is just a reformulation of Lemma 9.5.1/3.
9.6.4 Integrability of the Trace of a Function in BV ()
everywhere on . Then:
1. If for any measurable set E the inequality

(9.6.13)
where k is independent of E , is valid, then the trace u
exists for any u
BV (). Moreover,

|
u c|s(dx) kuBV () .
(9.6.14)
inf
c
505
2. If the inequality (9.6.14), with a constant k independent of u, holds for

any u BV () having a trace u
on then, for any measurable set E ,
the estimate (9.6.13) is true.
Proof. 1. By Theorem 9.5.2, the rough trace of u is integrable on .
Consequently, by Theorem 9.6.2, s-almost everywhere on there exists the
trace u
which coincides with u . Therefore (9.5.6) implies (9.6.14).
2. Let E be the measurable subset of with P (E ) < . By Lemma 9.6.3
the trace
E of the function E exists s-almost everywhere and equals E .
Thus, by inserting u = E into (9.6.14), we obtain

cs(dx) kP (E ),
inf
E
which is equivalent to (9.6.13) (compare with the proof of necessity in Theorem 9.5.2). The theorem is proved.

9.6.5 GaussGreen Formula for Functions in BV ()

Lemma. For any function u BV () and any measurable set B the
equality

Nt (B) dt
(9.6.15)
u(B) =
holds where Nt = {x : u(x) t}.

Proof. It suces to prove (9.6.15) for u 0. Let be an innitely dierentiable function with compact support in . Then

u(dx) =
u(dx) =
Nt (x) dt(dx).
By the Fubini theorem, the double integral equals

dt
Nt (x)(dx).
0
Moreover, we note that

u(dx) =
Nt (dx) =
dx
Nt dt
(9.6.16)
for almost all t. Here we may reverse the order of integration since the equality (9.1.16) implies the niteness of the integral

dt
|| var Nt (dx).
0
Thus (9.6.15) immediately results from (9.6.16).
506
Theorem. (The GaussGreen Formula). Let P () < and suppose that

a normal to exists s-almost everywhere on . Then for any function
u BV () whose rough trace is integrable on the boundary of , the equality

u (x)(x)s(dx)
v() =
holds where (x) is the normal to at the point x.

Proof. Since E (Rn ) = 0 for any set E with P (E ) < , by the Lemma
we have

Nt () dt =
Nt Nt dt.
u() =
Using s(\ ) = 0 and the coincidence of the normal to Nt with that to

on Nt , we obtain

Nt Nt =
(x)s(dx) = Nt .
Nt
Thus
u() =

Nt dt +
=
0
=
0
=
Nt dt

Nt dt

\ Nt dt
0

x : u t dt
x : u t dt

u (dx) =
u (x)(x)s(dx).


The example of the disk with a slit

z = ei : 0 < < 1, 0 < < 2
and of the function u(z) = shows that the condition s(\ ) = 0 cannot
be omitted under our denition of the trace.
The last theorem and Theorem 9.6.2 immediately imply the following
corollary.
Corollary. Let P () < , s(\ ) = 0. If for any E

where k is independent of E , then the trace u
(x) exists for any u BV ()
and the GaussGreen formula

u
(x)(x)s(dx)
u() =
holds.
507

This chapter was written together with Yu.D. Burago.
The rst two sections contain well-known facts from the theory of sets
with nite perimeter and from the theory of functions in BV . The foundation
of this theory was laid by Caccioppoli [160, 161] and De Giorgi [229, 230]. Its
further development is due to Krickeberg [464], Fleming [281], Fleming and
Rishel [282], and others. The results in Sects. 9.1.39.1.5 are due (up to the
presentation) to De Giorgi [229, 230]. Theorem 9.1.2 is a modication of a
result by Krickeberg [464]. Formula (9.1.16) in 9.1.6 was obtained by Fleming
and Rishel [282].
The results of Sect. 9.2 were established by De Giorgi [230] and supplemented by Federer [268].
At the present time the theory of sets with a nite perimeter can be considered as a part of the theory of integral currents (cf. Federer [271], part 4.5).
Sections 9.39.6 contain an expanded presentation of the paper by Burago
and Mazya [150].
Bokowski and Sperner [124] obtained the estimates for the functions (S)
and M for convex domains by the radii of inscribed and circumscribed balls.
For various facts concerning isoperimetric inequalities see the book by
Burago and Zalgaller [151] and the review paper by Osserman [648].
Soucek [718] studied the properties of functions whose derivatives of order
l are charges.
In connection with the contents of the present chapter see also the paper
by Volpert [781].
The following observation concerns the integrability of the trace of a function in BV ().
The requirement (9.6.13) has a global character: It is not satised, for
example, by any nonconnected set . This is essential because it is a question
of inequality (9.6.14). If we, however, put (9.6.14) in the following (equivalent,
provided is connected) form

|
u|s(dx) const uBV () ,
then the role of inequality (9.6.13) is taken by the local requirement

PC (E)
sup lim sup
: E B(x, ) < .
(9.7.1)
P (E)
x +0
The following result on continuation to the boundary of the domain was established by Anzellotti and Giaquinta [50]:
If condition (9.7.1) is fullled, then for any function L1 ( ) there
exists a function u W11 () whose trace on equals and we have the
inequality
uW11 () CL1 ( ) ,
508
where C depends on but not on .

Thus, for domains subject to condition (9.7.1) the spaces of traces of
of functions in BV () and W11 () coincide with L1 ( ).
Recently the basic results of Sects. 9.5 and 9.6 were extended by Burago
and Kosovsky [149]: They assumed that is a (n 1)-rectiable set instead
of our conditions that P () < and normals in the sense of Federer exist
a.e. on . It means that now the results of Sects. 9.5 and 9.6 are extended,
in particular, to the class of domains with cuts, important for applications.
Baldi and Montefalcone [66] generalized the extension criterion (9.3.2) to
a certain class of metric spaces.
A few words about the extension criterion (9.3.2). Roughly speaking,
(E) may be viewed as the area of soap lm placed on the exterior of and
suspended on that part of the boundary of E which belongs to . Therefore
(9.3.2) has a simple geometric meaning. Let l be an arbitrary closed contour
on the boundary of a three-dimensional body and let se and si be the areas
of the lms, suspended on l and positioned, respectively, outside and inside
. Condition (9.3.2) means that the ratio se /si is bounded irrespective of l.
If n = 2 then it is a question of the ratio between the distance between two
arbitrary points of measured outside and inside . This heuristic observation was rigorously justied by Koskela, Miranda, and Shanmugalingam [457]
for a bounded simply connected domain R2 . They showed that is a
BV-extension domain if and only if there exists a constant C > 0 such that
for all x, y R2 \ there is a rectiable curve R2 \ connecting x and y
with length
l() C|x y|.
In other words, is a BV -extension domain if and only if the complement of
is quasiconvex. A corollary of this result is the necessity of quasiconvexity
for to be a W11 -extension domain, and it is conjectured in [457] that the
same necessary condition holds for all Wp1 -extension Jordan domains in R2
with an arbitrary p [1, 2].
Fig. 34.
509
A necessary and sucient condition for the extension of functions in

Wp1 (), p = 2, to Rn without changing the class is not known. It is easy
to guess the following not very visual condition, which is a direct analog of
the BuragoMazya criterion (9.3.2) for the extension of functions in BV ().
abutting
For each conductor K (i) in there exists a conductor K (e) in Rn \
to it (cf. Fig. 34) such that

cp K (e) cp K (i) .
It cannot be excluded that this requirement is indeed sucient, but this has
not been proved.
10
Certain Function Spaces, Capacities,
and Potentials
Section 10.1 is of an auxiliary nature. Here we collect (mostly without proofs)

the results of function theory that are applied later or related to the facts
used in the sequel. First we discuss the theorems on spaces of functions having derivatives of arbitrary positive order (Sect. 10.1). The theory of these
spaces is essentially presented in monographs (cf. Stein [724]; Peetre [657];
Nikolsky [639]; Besov, Ilin, and Nikolsky [94]; Triebel [755, 756]; and Runst
and Sickel [685]) though in some cases the reader interested in the proofs will
have to refer to the original papers.
Section 10.2 is concerned with the Bourgain, Brezis, and Mironescu theorem on the asymptotic behavior of the norm of the Sobolev-type embedding
operator Wps Lpn/(nsp) as s 1 and s n/p. Their result is extended to
all values of s (0, 1) and is supplied with an elementary proof. The relation

|u(x) u(y)|p
dx dy = 2p1 n upLp (Rn )
lim
s0 Rn Rn |x y|n+sp
is proved.
In Sect. 10.3 we prove the multiplicative GagliardoNirenberg-type inequality

/p
p
1s
uWps u1
uWp/
s c(n)
L ,
p1
1
where 0 < < 1, 0 < s < 1, 1 < p < , and uWps is the seminorm in
the fractional Sobolev space Wps (Rn ) and show that the dependence of the
constant factor in the right-hand side on each of the parameters s, , and p
is precise in a sense.
In Sect. 10.4, dealing with properties of capacities and nonlinear potentials,
we restrict ourselves to the formulation of the results.

DOI 10.1007/978-3-642-15564-2 10,
511
512
10 Certain Function Spaces, Capacities, and Potentials
10.1 Spaces of Functions Dierentiable of Arbitrary

Positive Order
10.1.1 Spaces wpl , Wpl , blp , Bpl for l > 0
For p 1 and integer l > 0, let wpl denote the completion of the space D with
respect to the norm l uLp . For p 1 and noninteger l, we dene wpl as the
completion of D with respect to the norm

1/p
y up [l] |y|np{l} dy
.
(10.1.1)
wp
Here and elsewhere y u(x) = u(x + y) u(x), [l] and {l} are the integer and
fractional parts of l, respectively.
Replacing the norm (10.1.1) by the norm

ublp =
2y upLp |y|npl dy
1/p
,
0 < l 1,
(10.1.2)
in the previous denition, we obtain the space blp (here 2y u(x) = u(x + y)
2u(x) + u(x y)). For l > 1 we put ublp = ubpl1 .
Further, let Wpl and Bpl be the complements of D with respect to the norms
uwpl + uLp and ublp + uLp .
For fractional l the norms in wpl and blp , as well as the norms in Wpl and
Bpl , are equivalent. In fact, the identity

2 u(x + h) u(x) = u(x + 2h) 2u(x + h) + u(x)

+ u(x + 2h) u(x)
implies the estimates

2 2l Hl u Gl u 2 2l Hl u,
0 < l < 1,
where

(Hl u)(x) =

(Gl u)(x) =

(y u)(x) p |y|npl dy
1/p
2 p npl
u (x) |y|
dy
y
,
1/p
.
The utility of the dened spaces is mostly owing to the following trace
theorem.
Theorem 1. For p [1, ), l > 0, m = 1, 2, . . . , we have
1
ublp (Rn ) inf
|y|m{l}p m+[l] U
L (Rn+1 ) ,
p
{U }
(10.1.3)
10.1 Spaces of Functions Dierentiable of Arbitrary Positive Order
513
where U D(Rn+1 ) is an arbitrary extension of u D(Rn ) to the space

Rn+1 = {(x, y) : x Rn , y R1 }.
Proof. We restrict ourselves to the derivation of (10.1.3) for 0 < l 1,
m = 2.
(2)
Let U D(Rn+1 ), u = U |Rn . We put
t u(x, t) = u(x, 2t) 2u(x, t) +
u(x, 0). It can be easily seen that
t

d2
(2)
(10.1.4)
(t ) 2 u(x, t + ) + u(x, t ) d.
t u(x, t) =
d
0
We can also readily check that

(2) u(x, 0) = 2(2) u x, |h| + 2(2) u x, |h| (2) u x, 2|h|
t
h
h
|h|

(2)
(2) u x h, |h| .
+ u x + h, |h| +
(10.1.5)
|h|
|h|
We consider only the rst and the second summands since the others can be
estimated in a similar way. By (10.1.4) we obtain
2|h|
2
(2)

u , |h|
2
y
2 U (x, y)
dy.
|h|
Lp (Rn )
dy
0
Lp (Rn )
Therefore (for l (0, 1], p 1),
(2)

u , |h|
p n dh
|h|
Lp (R ) |h|n+pl
n
R
p
2
2
d U
d
c
y
2 (, y)
dy
1+pl
dy

0
0
Lp (Rn )
2

d U
c
y (2l)p1
dy.
dy2 (, y)
0
Lp (Rn )
(10.1.6)
Next we proceed to the second item in (10.1.5). We have

1

d2
(2)
h u x, |h| =
(1 ) 2 u x + h, |h| + u x h, |h| d.
d
0
By the Minkowski inequality
(2)

u , |h|
h
L
Hence
(2)

u , |h|
p n
h
Lp (R )
Rn
n
p (R )

c|h|2
2,x u , |h|
L
dh
c
|h|n+pl

0
p (R
n)
y (2l)p1
2,x U (, y)
L
p (R
n)
dy,
which together with (10.1.6) yields the upper bound for the norm ublp . We
proceed to the lower estimate.
514
Let u D(Rn ) and let be the extension operator to the space Rn+1 =
{X = (x, y) : x Rn , y R1 } dened by

x
n
(u)(X) = y
u() d,
(10.1.7)
y
Rn
where
C0 (B1 ),
(x) dx = 1,
(x) = (x).
Rn
Using the evenness of the function , for || = 2 we obtain

x

Dx u (X) = y n2
u() d
D
y
Rn

1
(2)
D (h/y)h u(x) dh,
= y n2
2
Rn
(2)
where, as before, h u(x) = u(x + h) 2u(x) + u(x h). Therefore for || = 2

we have

(2)

u(x) dh.
Dx u (X) cy n2
h
By
Since
2(u)(X) = y n
it follows that

(h/y)(2)
y u(x) dh + 2u(x),
2

d

cy n2
u
(X)
dy2

(2)

u(x) dh.
h
By
We can easily check that the same estimate is also true for | 2 /xi y|. So
the second derivatives of u are bounded and

p
y 1+p(2l)
(2 u)(, y)
Lp (Rn ) dy
0

(2)

p
u() dh
y 1+p(l+n)
dy.
(10.1.8)
By
Lp (Rn )
By the Minkowski inequality the right-hand side does not exceed

p

(2)
u()
y 1pln
dh
dy
c
h
L (Rn )
0
c
Rn
=c
Rn
By
(2)
u()
p
h
Lp (Rn )
(2)
u()
p
h
Lp (Rn )
y 1pln dy dh
|h|
|h|npl dh,
and the required lower estimate for the norm ublp follows.
515
Next we show that the function u can be approximated by the sequence

of extensions Uk D(Rn+1 ) of the function u in the metric
2lp1
|y|
2 U
n+1 .
Lp (R
Let k (X) = (X/k) where D(Rn+1 ), = 1 for |X| < k and k = 1, 2 . . . .

Since (D u)(X) = O((|X| + 1)n|| ) for 0 < || 2 then 2 (u
k u) = O((|X| + 1)n2 ). Furthermore,

supp(u k u) X Rn+1 : |X| > k .
Consequently,
2lp1
|y|
2 (u k u)
L
p (R
n+1 )

= O k ln+n/p = o(1)
as k . It remains to approximate each of the functions k u by a sequence

of mollications.

Similarly we can show the following analogous assertion for the space
Bpl (Rn ).
Theorem 2. For p [1, ), l > 0, m = 1, 2, . . . , we have

1
uBpl (Rn ) inf
|y|m{l}p m+[l] U
Lp (Rn+1 ) + U Lp (Rn+1 ) .
{U }
Theorems 1 and 2 have a long history. For p = 2, {l} = m 12 they were

established by Aronszajn [52], Babich and Slobodecki [60], Slobodecki [704],
and Freud and Kr
alik [291]. The particular case n = p = 2, m = l = 1 is
actually contained in the papers by Douglas [242], 1931, and Beurling [95],
1940. The generalization for p = 2 is due to Gagliardo [298] for l = 1
1/p. Theorems 1 and 2 were proved, in a form similar to that above, by
Uspenski [770] (cf. also Lizorkin [504]).
A theory of spaces with weighted norms is discussed in the books by
Triebel [756], Kufner [465], and in the survey by Besov, Ilin, Kudryavtsev,
Lizorkin, and Nikolsky [93].
In conclusion we state the following trace theorem for functions in
w11 (Rn+1 ), W11 (Rn+1 ) restricted to Rn .
Theorem 3. We have
uL1 (Rn ) inf U L1 (Rn+1 ) inf U W11 (Rn+1 ) ,
{U }
{U }
where u D(Rn ) and {U } is the collection of all extensions of u to Rn+1 ,

U D(Rn+1 ).
This assertion is proved in the paper by Gagliardo [298] (cf. also the book
by Besov, Ilin, and Nikolsky [94]). For a more recent development of trace
theorems see the Comments to the present chapter.
516
10.1.2 Riesz and Bessel Potential Spaces

With each function u D = D(Rn ), we associate its Fourier transform

u
() = F u() = (2)n/2 eix u(x) dx.
The same notation will be retained for the Fourier transform of the distribution u contained in the space D dual of D. We denote the convolution of
distributions by a star, .
The scales of fractional spaces dierent from the spaces introduced in
Sect. 10.1.1 are dened by means of the operators

l/2
( + 1)l/2 = F 1 1 + ||2
F,
()l/2 = F 1 ||l F,
where is the Laplace operator.

Namely, let hlp and Hpl (1 0) denote the completion of the
space D with respect to the norms
uHpl =
( + 1)l/2 u
L .
uhlp =
()l/2 u
L ,
p
The following assertion is the Mikhlin theorem on Fourier integral multipliers [601].
Theorem 1. Let the function dened on Rn \{0} have the derivatives
()/j1 , . . . , jk , where 0 k n and 1 j1 < j2 < < jk n.
Further let

k ()
k
M = const .
||
j1 , . . . , jk
k
Then for all u Lp
F F u
cM uL ,
p
L
p
1 < p < ,
where c is a constant that depends only on n and p.

Corollary 1. Let l = 1, 2, . . . , then there exist positive numbers c and C
that depend only on n, p, l, such that
c
()l/2 u
L l uLp C
()l/2 u
L
(10.1.9)
p
for all u D.
Proof. Let be a multi-index with || = l. Then
F 1 F u = F 1 ||l ||l F u.
The function ||l satises the hypothesis of Theorem 1, which leads to the
rightmost estimate in (10.1.9). On the other hand,
||l = ||2l ||l =
517

c ||l ,
||=l
where c = l!/!, so
F 1 ||l F u =

||=l
c F 1

F u.
||l
Again, applying Theorem 1, we obtain the leftmost estimate in (10.1.9).

The following corollary has a similar proof.
Corollary 2. Let l = 1, 2, . . . . There exist positive numbers c and C that
depend only on n, p, l such that
cuWpl
( + 1)l/2 u
L CuWpl
p
for all u D.
Thus, wpl = hlp and Wpl = Hpl provided p > 1 and l is an integer.
For the proof of the following theorem see the paper by Havin and
Mazya [567].
Theorem 2. Let pl < n, p > 1. Then u hlp if and only if
u = ()l/2 f c|x|ln f,
where f Lp .
A similar well-known assertion for the space Hpl is contained in the following theorem.
Theorem 3. The function u belongs in Hpl , p > 1, if and only if
u = ( + 1)l/2 f Gl f,
where f Lp ,

Gl (x) = c|x|(ln)/2 K(nl)/2 |x| > 0,
K is the modied Bessel function of the third kind.

For |x| 1 the estimates
ln
0 < l < n,
c|x| ,
Gl (x) c log(2/|x|), l = n,
c,
l > n,
hold. If |x| 1, then
Gl (x) c|x|(ln1)/2 e|x| .
518
The integral operators

I
l
|x|ln f,
f
l
f
Gl f,
are called the Riesz potential and the Bessel potential, respectively. Thus
Theorems 2 and 3 state that each element of the space hlp (pl < n) (Hpl ) is the
Riesz (Bessel) potential with density in Lp .
Next we formulate the theorem due to Strichartz [728] on equivalent norms
in the spaces hlp and Hpl .
Theorem 4. Let {l} > 0 and let

(D{l} v)(x) =
||<1

v(x + y) v(x) d
2
y
12{l}
1/2
dy
. (10.1.10)
Then
uhlp D{l} [l] uLp ,
(10.1.11)
uHpl D{l} [l] uLp + uLp .
(10.1.12)
Two-sided estimates for the Lp -norm of a function u on R1 by its harmonic extension to R1 (0, ) are due to Littlewood, Paley, Zygmund, and
Marcinkiewicz. Let the LittlewoodPaley function g(u) be dened by

g(u) (x) =

U (x, y) 2 y dy
1/2
,
where U (x, y) is the Poisson integral of u. The basic result here is the equivalence of the norms uLp (R1 ) and g(u)Lp (R1 ) , 1 < p < . In the book by
Stein [724] this equivalence is proved for Rn .
The next theorem by Shaposhnikova [697], similar to Theorems 10.1.1/1
and 10.1.1/2, contains a characterization of hlp (Rn ) in terms of extensions to
Rn+k .
Theorem 5. The norm of u D(Rn ) in hlp (Rn ), 0 < l < 1 is equivalent
to

p/2
inf
{U }
Rn
Rk
|y|
22lk
|U | dy
2
1/p
dx
where the inmum is taken over all extensions U D(Rn+k ) of u to Rn+k =

{(x, y) : x Rn , y Rk }.
Similarly,

uHpl inf
{U }
Rn

Rk

|y|22lk |U |2 + |U |2 dy
1/p
p/2
dx
519
10.1.3 Other Properties of the Introduced Function Spaces

The following two theorems, whose proofs can be found in the books by
Stein [724], Peetre [657], Nikolsky [639], and Triebel [756], are classical facts
of the theory of the spaces blp , Bpl , hlp , and Hpl .
blp
Theorem 1. If 2 p < then hlp blp , Hpl Bpl . If 1 0, then
ublp (Rn ) inf U hl+1/p
.
(Rn+1 )
p
{U }
Here and in (ii) the notation {U } has the same meaning as in Theorem 10.1.1/1.
(ii) If p [1, ), l > 0 then
ublp (Rn ) inf U bl+1/p
.
(Rn+1 )
p
{U }
In items (i) and (ii) one can replace b and h by B and H, respectively.
A far-reaching generalization of relations (i) and (ii) is given in the article
by Jonsson and Wallin [407] where functions in Bpl on the so-called d-sets F
are extended to Rn . The latter sets are dened by the relation

Hd F B(x, ) d ,
which is valid for all x F and < , where Hd is d-dimensional Hausdor
measure (cf. Sect. 1.2.4).
Henceforth we denote by B the ball with radius r, concentric with the
ball B of radius r. Similarly, with the cube Q with edge length d we associate
the concentric cube Q with sides parallel to those of Q and with edge length
d.
Next we state the theorem that is easily derived from (10.1.12) (cf.
Strichartz [728]) for the spaces Hpl , {l} > 0. For Wpl and Bpl it is a simple
corollary of the denitions of these spaces.
Theorem 3. Let {B (j) }j0 be a covering of Rn by unit balls that has a
nite multiplicity depending only on n. Further, let O (j) be the center of B (j) ,
O(0) = O, and let j (x) = (x O(j) ) where C0 (2B (0) ), = 1 on B (0) .
Then

1/p
p
uj sl
,
(10.1.13)
uSpl
p
j0
where Spl = Hpl , Wpl or Bpl and slp = hlp , wpl or blp , respectively.
We can easily check that for any v C0 (B1 ) the inequality
520
vLp cvslp ,
(10.1.14)
holds. Therefore the norm uj slp in (10.1.13) can be replaced by the equivalent norm uj Spl .
When dealing with the embedding of a Banach space X into another
Banach space Y (notation: X Y ) we always mean a continuous embedding.
Theorem 4. 1. If l = 1, 2, . . . , then bl1 w1l .
2. If p > 1, l > 0, n > (l )p and l n/p = n/ then hlp h .
3. If p 1, l > > 0, n > (l )p, and l n/p = n/ then blp b .
4. If p 1, l > 0, n > lp, and 1/ = 1/p l/n then blp L .
5. If l = 1, 2, . . . , n l and l n = n/ then w1l b .
Replacing the letters b, h, w by B, H, and W in items 15 we also obtain
true assertions.
6. If p > 1, l > 0, n = (l )p, then Hpl H for any (1, ). If
p > 1, lp > n, then Hpl L C.
7. If p > 1, l > > 0, n = (l )p, then Bpl B for any (1, ).
8. If p > 1, l > 0, n = lp, then Bpl L for any (1, ).
9. If p > 1, lp > n, then Bpl L C.
Various proofs of assertions 14 and 69 can be found in the monographs
mentioned at the beginning of the present subsection. The proof of assertion
5 is due to Solonnikov [717]. The embedding in 2 is an immediate corollary of the continuity of the operator ()(l)/2 : Lp L established by
Sobolev [712]. Item 1 follows from the inequalities
l uL1 (Rn ) c1 l+1 U L1 (Rn+1 ) c2 yl+2 U L1 (Rn+1 )
and from Theorem 10.1.1/1. (Here U D(Rn+1 ) is an arbitrary extension of
the function u D(Rn ).) The same theorem together with the inequality
1{}1/
|y|
[]+1 U
L (Rn+1 ) c
|y|{l} [l]+1 U
L (Rn+1 ) ,
1
which results from Corollary 2.1.7/4, leads to (3) for p = 1. The same result
for p > 1 follows from the embedding
n+1
n+1
hl+1/p
R
h+1/
R
p
(cf. item 2 above and part (i) of Theorem 2).

The corresponding assertions for the spaces Hpl and Bpl can be obtained
in a similar way. The embedding in item 6 easily follows from the denition
of the Bessel potential. The properties in items 7 and 9 result from item 6
l+1/p
applied to the space Hp
(Rn+1 ), and item 8 is a corollary of item 7.
10.2 On the Bourgain, Brezis, and Mironescu theorem
521
10.2 Bourgain, Brezis, and Mironescu Theorem

Concerning Limiting Embeddings of Fractional
Sobolev Spaces
10.2.1 Introduction
s (Rn ) as the completion
Let s (0, 1) and let p 1. We introduce the space W
p
n
of C0 (R ) in the norm

Rn
Rn
|u(x) u(y)|p
dx dy
|x y|n+sp
1/p
.
s,p
(Q) of functions dened on the cube Q = {x
We also need the space W
n
R : |xi | < 1/2, 1 i n}, which are orthogonal to 1 and have the nite
norm
1/p

|u(x) u(y)|p
dx dy
.
n+sp
Q Q |x y|
The main result of the recent article by Bourgain, Brezis, and Mironescu
[139] is the inequality
upLq (Q) c(n)
1s
upW s,p (Q) ,
(n sp)p1
(10.2.1)
s,p
where u W
(Q), 1/2 s < 1, sp < n, q = pn/(n sp), and c(n) depends
only on n.
s (Rn ), valid for the whole
Figuring out a similar estimate for functions in W
p
interval 0 < s < 1, one can anticipate the appearance of the factor s(1 s)
ps (Rn ) blows up both as s 1 and
in the right-hand side since the norm in W
s 0. The following theorem shows that this is really the case.
Theorem. Let n 1, p 1, 0 < s < 1, and sp < n. Then, for an
s (Rn ), there holds
arbitrary function u W
p
upLq (Rn ) c(n, p)
s(1 s)
upW
s (Rn ) ,
p
(n sp)p1
(10.2.2)
where q = pn/(n sp) and c(n, p) is a function of n and p.

From this theorem, one can derive inequality (10.2.1) for all s (0, 1)
with a constant c depending both on n and p (Corollary 10.2.2/1). In the
case s 1/2 considered in [139], one has 1 < p < 2n and therefore the
dependence of the constant c on p can be eliminated. Thus, we arrive at the
BourgainBrezisMironescu result and extend it to the values s < 1/2.
The proof given in [139] relies upon some advanced harmonic analysis
and is quite complicated. Our proof of (10.2.2) is straightforward and rather
simple. It is based upon an estimate of the best constant in a Hardy-type
522
s (Rn ), which is obtained in Theorem 10.2.2 and

inequality for the norm in W
p
is of independent interest.
In Theorem 10.2.4 we derive a formula for lims0 supW
s (Rn ) that complep
ments an analogous formula for lims1 (1 s)upW

s (Rn ) found in [138].
p
10.2.2 Hardy-Type Inequalities

Theorem. Let n 1, p 1, 0 < s < 1, and sp < n. Then, for an arbitrary
ps (Rn ),
function u W
Proof. Let

u(x) p dx c(n, p) s(1 s) up
s (Rn ) .
W
p
|x|sp
(n sp)p
Rn
(10.2.3)

(h) = n1 n(n + 1) 1 |h| + ,
where h Rn and the subscript plus stands for the nonnegative part of a
real-valued function. We introduce the standard extension of u onto Rn+1
=
+
{(x, z) : x Rn , z > 0}

(h)u(x + zh) dh.
U (x, z) :=
Rn
A routine majoration implies

U (x, z) n(n + 1)(n + 2)
zn

|h|<1

u(x + zh) u(x) dh.
Hence, and by H
olders inequality, one has

p
z 1+p(1s) U (x, z) dx dz
0
Rn

n
(n + 1)p (n + 2)p
z 1ps
n
0

u(x + zh) u(x) p dx dh dz.
|h|<1
(10.2.4)
Rn
Setting = zh and changing the order of integration, one can rewrite (10.2.4)
as

p
Rn
0

|u(x) u(y)|p
n(n + 1)p (n + 2)p
dx dy.
(10.2.5)
n+sp
n (sp + n)
Rn Rn |x y|
By Hardys inequality
z
p

|x|

z 1sp
( ) d dz sp
0
|x|
523
p

z 1+p(1s) (z) dz,
one has
|x|
|u(x)|p
|x|p
0

p
z
|x|
U
|U (x, )|

+
d
(x,
)
p(1 s)
z 1sp dz

|x|
0
0

|U (x, z)| p
p(1 s) |x| 1+p(1s) U

z
dz.
z (x, z) +
sp
|x|
0
|u(x)|p
= p(1 s)
|x|sp
z 1+p(1s) dz
Now, the integration over Rn and Minkowskis inequality imply

|u(x)|p
dx
sp
Rn |x|

p

1/p
p

p(1 s)
1+p(1s) U
dz dx
(x,
z)
z
+
A
, (10.2.6)
z

sp
Rn 0
where
|x|
A :=
Rn
Clearly,
1+p(1s)
A 2
p/2
dx
Rn
z 1+p(1s)
which does not exceed

p/2
1+p(1s)
2
(cos )
n
S+
|x|

p
1/p
p

.
U (x, z) dz dx
|U (x, z)|p
dz dx,
(x2 + z 2 )p/2
|U |p n1sp d d,
(10.2.7)
where = (x2 + z 2 )1/2 , cos = z/, d is an element of the surface area on

n
the unit sphere S n and S+
is the upper half of S n . Using Hardys inequality

p

U p n1+p(1s)
p
p n1sp

|U |
d
d,

n sp
0
0
one arrives at the estimate
1/2 p
p

2 p
Ap
z 1+p(1s) U (x, z) dx dz.
n sp
0
Rn
Combining this with (10.2.6), one obtains
524

Rn
p

|u(x)|p
21/2 p
p(1 s)
1
+
dx
|x|sp
sp
n sp

p

z 1+p(1s) U (x, z) dx dz,

0
Rn
which, along with (10.2.5), gives

|u(x)|p
(1 s) p(n + 2p)3p
dx
upW
s (Rn ) .
sp
p
p
p
|x|
(n
sp)
s
n
n
R
(10.2.8)
To justify (10.2.3) we need to improve (10.2.8) for small values of s. Clearly,

|u(x)|p
dy
n
u(x) p dx.
dx
=
sp
sp
n+sp
2 sp Rn |x|
Rn |xy|>2|x| |x y|
Since |x y| > 2|x| implies 2|y|/3 < |x y| < 2|y|, we obtain

1/p
1/p

|u(x)|p
|u(x) u(y)|p
n
dx
dx
dy
n+sp
2sp sp Rn |x|sp
Rn |xy|>|x| |x y|

1/p

3sp 1
|u(y)|p
+ n sp
dy
.
2 sp Rn |y|sp
Hence,
1/p

1/p

sp
1/p
|u(x)|p
n
1 3 1
dx
21/p uW
s (Rn ) .
sp
p
2sp sp
Rn |x|
Let be an arbitrary number in (0, 1). If s (4p)1 p , we conclude

2sp1 sp
|u(x)|p
dx
upW
(10.2.9)
s (Rn ) .
sp
p
n (1 )p
Rn |x|
Setting = 21 and comparing this inequality with (10.2.8), we arrive at
(10.2.3) with
c(n, p) = n1 (n + 2p)3p pp+2 2(n+1)(n+2) .
From Theorem 10.2.2, we shall deduce an inequality, analogous to (10.2.3),

for functions dened on the cube Q. Unlike (10.2.3), this inequality contains
no factor s on the right-hand side, which is not surprising because, for smooth
u, the norm uWs,p (Q) tends to a nite limit as s 0.
Corollary 1. Let n 1, p 1, 0 < s < 1, and sp < n. Then any function
s,p
(Q) satises
u W

u(x) p dx c(n, p) 1 s up s,p .
(10.2.10)
W (Q)
|x|sp
(n sp)p
Q
525
Proof. Let us preserve the notation u for the extension by a reection of

s,p
(Q) to the cube 3Q, where aQ stands for the cube obtained from Q
u W
by dilation with the coecient a. We choosea cuto function , equal to 1
n
on Q and vanishing outside 2Q, say, (x) = i=1 min{1, 2(1 xi )+ }. By the
previous Theorem, it is enough to prove that
1
upW
c(n, p)upW s,p (Q) .
s (Rn ) s
(10.2.11)
Clearly, the norm in the left-hand side is majorized by

1/p
|u(x) u(y)|p
p
dx
(y)
dy
n+sp
3Q 3Q |x y|

1/p
|(x) (y)|p
p
+
dx|u(y)|
dy
n+sp
3Q 3Q |x y|

1/p

dy
(u)(x) p dx
+ 2
.
n+sp
3Q Rn \3Q |x y|
The rst term does not exceed 6n/p uWs,p (Q) ; the second term is not greater
than
1/p

p

dx
1/2

2n
u(y) dy
np(1s)
3Q 3Q |x y|
1/p

n
2+n/p
n3
uLp (Q) ,
p(1 s)
and the third one is dominated by

2

2Q
|xy|>1/2
1/p n+1+p 1/p

2
dy
u(x) p dx
uLp (Q) .
n+sp
|x y|
sp
Adding these estimates, one obtains

uW0s,p (Rn ) 6n/p uW
s (Q)
p
2+n/p 1/p
+ n3

s1/p + (1 s)1/p uLp (Q) .
(10.2.12)
We preserve the notation u for the mirror extension of u onto Rn . Recalling

that u1 on Q, we have

u(x) p dx
u(x) u(y) p dx dy
Q
Q Q

p

dh u(x + h) u(x) dx.
(10.2.13)
2Q
526
Let U be the extension of u to Rn+1

+ . For any z > 0 and h 2Q
u( + h) u()
Lp (Q)
z
z
U
U
( + h, ) d
(, ) d
0
0
Lp (Q)
Lp (Q)
+ U ( + h, z) U (, z) Lp (Q)
z
2
d +
U ( + h, z) U (, z)
L (Q) .
(, )
p
0
Lp (3Q)
Hence
p1/p (1 s)1/p |h|1s

u( + h) u()
Lp (Q)
|h|
1/p
1+p(1s)
u( + h) u()
p
=
z
dz
Lp (Q)
0
+ ,
where

p
p = 2|h|
(10.2.14)
|h| z
and
U (, )
p
d
z 1ps dz
Lp (3Q)
U ( + h, z) U (, z)
p
z 1+p(1s) dz.
Lp (Q)
|h|
p =
0
Using Hardys inequality already referred to at the beginning of the proof of

the previous Theorem, we arrive at
p

p |h|
2|h|
z 1+p(1s) dz.
p
(,
z)
s
z
0
Lp (3Q)
The trivial inequality

U (x + h, z) U (x, z) p dx |h|p
Q

U (x, z) p dx
3nQ
implies
U (, z)
p
z 1+p(1s) dz.
Lp (3nQ)
|h|
p |h|p
0
We put the just-obtained estimates for and into (10.2.14) and deduce
u( + h) u()
p
Lp (Q)
p

|h|
U (, z)
p
+ 1 |h|ps
p(1 s)
z 1+p(1s) dz.
Lp (2nQ)
s
0
527
Noting that |h| n for h 2Q, we nd
|h|ps
u( + h) u()
Lp (Q)
p n

2
U (x, z) p dxz 1+p(1s) dz.
+1
p(1 s)
s
0
3nQ
(10.2.15)
Now, let U be the same extension of u onto Rn+1
as in the beginning of
+
the proof of Theorem 10.2.2. Repeating with obvious changes the standard
argument in the proof of the previous
Theorem, which leads to (10.2.5), we
conclude that the integral over (0, n)3nQ in the right-hand side of (10.2.15)
is majorized by
c0

u(x + z) u(x) p dx d z 1ps dz
0
||<1 3nQ

c0
nps
u(x + ) u(x) p dx d,
||
n + ps ||< n
3nQ
where
c0 =
n
(n + 1)p (n + 2)p .
n
Therefore,
|h|ps
u( + h) u()
L (Q)
p

1
s
|u(x) u(y)|p
c0 3n+2p n2n p
dx dy.
n+ps
s
Q Q |x y|
(10.2.16)
Let s > 1/2. It follows from (10.2.13) that

|u(x) u(y)|p
u(x) p dx n(n+ps)/2
dx dy.
n+ps
Q
Q Q |x y|
This inequality together with (10.2.16) shows that for all s (0, 1)
uLp (Q) (4n)4n (1 s)1/p uWs,p (Q) .
Combining this inequality with (10.2.12), we arrive at (10.2.11) and hence
complete the proof.

Corollary 2. Let 0 < s < 1 and p 1. Then there holds
sup |h|s
u( + h) u()
L (Q) c(n, p)(1 s)1/p uWs,p (Q) .
q
Proof. The result follows from the well-known embedding Bps (Q)
if s 1/2 and from (10.2.16) if s > 1/2.
s
(Q)
Bp,
528
10.2.3 Sobolev Embeddings

Proof of Theorem 10.2.1. It is well known that the fractional Sobolev norm of
order s (0, 1) is nonincreasing with respect to the symmetric rearrangement
of functions decaying to zero at innity (see Wik [796], Almgren and Lieb
[41], Theorem 9.2, and Cianchi [196]). Let v(|x|) denote the rearrangement of
|u(x)|. Then

n 1/q
n
q
v(r) d r
,
(10.2.17)
u||Lq (Rn ) =
n 0
where n is the area of the unit sphere B1 . Recalling that an arbitrary
nonnegative nonincreasing function f on the semi-axis (0, ) satises

f (t) d t
t
1
f ( ) d
f (t) dt =
f (t) dt
(see Hardy, Littlewood and Polya [350]), one nds that the right-hand side in
(10.2.17) does not exceed
n
n
1/q
v(r) d r
nsp
1/p
=
(n sp)1/p
s/n
n1/q n
p dx
v |x|
|x|sp
n
R
1/p
.
We now see that (10.2.2) results from inequality (10.2.3).
Corollary. Let n 1, p 1, 0 < s < 1, and sp < n. Then any function

s,p
(Q) satises
u W
upLq (Q) c(n, p)
1s
upW s,p (Q) .
(n sp)p1
Proof. Let be the same cuto function as in Corollary 10.2.2/1. The

result follows by combining inequality (10.2.11) with Theorem 10.2.1, where
u is replaced by u.
s (Rn ) as s 0
10.2.4 Asymptotics of the Norm in W
p
Theorem. For any function u

0<s<1
ps (Rn ), there exists the limit

W
1
lim supW
n upLp (Rn ) .
s (Rn ) = 2p
s0
Proof. Since can be chosen arbitrarily small, inequality (10.2.9) implies

1
lim inf supW
n upLp (Rn ) .
s (Rn ) 2p
s0
(10.2.18)
529
Let us majorize the upper limit. By (10.2.18), it suces to assume that u

Lp (Rn ). Clearly,

1/p

dy
u(x)Rn p dx
supW
2
s
.
s (Rn )
n+sp
p
Rn |y|2|x| |x y|
1/p p

dx dy
u(y) p
+ s
n+sp
Rn
|y|2|x| |x y|

|u(x) u(y)|p
+ 2s
dx dy.
n+sp
Rn |x|<|y|<2|x| |x y|
The rst term in braces does not exceed

1/p
1/p
1/p

n
dy
|u(x)|p
u(x) p dx
s
=
dx
,
n+sp
sp
p1/p
Rn |y||x| |x y|
Rn |x|
hence its lim sups0 is dominated by n p1/p uLp (Rn ) . The second term in
braces is not greater than

1/p

|u(y)|p
1/p
n+sp
2
s
dy
dx
n+sp
Rn |y|
|x|<|y|/2

1/p
1/p
|u(y)|p
s s
n
=2
dy
,
sp
p
Rn |y|
1/p
so it tends to zero as s 0.
We claim that

lim sup s
s0
Rn
|x|<|y|<2|x|
|u(x) u(y)|p
dx dy = 0.
|x y|n+sp
(10.2.19)
p (Rn ) for a certain (0, 1). Let N

By assumption of the Theorem, u W
be an arbitrary number greater than 1 and let s < . We have

|u(x) u(y)|p
2s
dx dy
n+sp
Rn |x|<|y|<2|x| |x y|

|u(x) u(y)|p
dx dy
2sN p( s)
|x y|n+ p
Rn |x|<|y|<2|x|
|xy|N

|u(x) u(y)|p
+ 2s
dx dy.
|x y|n+sp
Rn |x|<|y|<2|x|
|xy|>N
The rst term on the right-hand side tends to zero as s 0 and the second
one does not exceed

dy
u(x) p dx c(n, p)
u(x) p dx,
2p+1 s
n+sp
|x|>N/3 |xy|>N |x y|
|x|>N/3
which is arbitrarily small if N is suciently large. The proof is complete.
530
Remark. Since the proof of the Theorem just proved holds for vector-valued
functions, one can write

|u(x) u(y)|p
1
u(x) p dx, (10.2.20)
lim s
dx dy = 2p n
n+sp
s0
|x
y|
n
n
n
R
R
R
for any function u such that
u

ps Rn .
W
0<s<1
Formula (10.2.20) complements the following relation that was established in

Bourgain, Brezis, and Mironescu [138]:

|u(x) u(y)|p
lim(1 s)
dx dy
n+sp
s1
Rn Rn |x y|

u(x) p dx,
| cos |p d
(10.2.21)
=
B1
Rn
where is the angle deviation from the vertical.
10.3 On the Brezis and Mironescu Conjecture

Concerning a GagliardoNirenberg Inequality
for Fractional Sobolev Norms
10.3.1 Introduction
Let s (0, 1) and let 1 < p . We introduce the space Wps (Rn ) of functions
in Rn with the nite seminorm

1/p
|u(x) u(y)|p
s
dx dy
.
uWp =
n+sp
Rn Rn |x y|
The relation (10.2.21) motivated Brezis and Mironescu to conjecture the following GagliardoNirenberg-type inequality:
1/2
1/2
uW s/2 c(n, p)(1 s)1/2p uWps uL
(10.3.1)
2p
(see [144], Remark 5). In [144] one can also read, It would be of interest to
establish
1
0 < < 1,
(10.3.2)
uWp/
s cuW s u
L ,
p
with control of the constant c, in particular when s 1.
In the next subsection we prove that (10.3.2) holds with
c = c(n, p, )(1 s)/p ,
which, obviously, contains inequality (10.3.1) predicted by Brezis and Mironescu. Our proof is straightforward and rather elementary. In concluding Remarks 1 and 2 we show that the dependence of c on each of the parameters
s, , and p is sharp in a certain sense.
531
10.3.2 Main Theorem

Theorem. For all u Wps L there holds the inequality
uWp/
s
p
c(n)
p1

1s
1
/p
1
uWps uL
,
(10.3.3)
where 0 < s < 1, 1 < p < , and 0 < < 1.

Proof. Clearly,
/p 1
1
uL
s max 2
,2
uWps .
uWp/
(10.3.4)
Hence it suces to prove (10.3.3) only for s 1/2.

Let Br (x) = { Rn : | x| < r} and Br (0) = Br . We introduce the
mean value ux,y of u over the ball Bx,y := B|xy|/2 ((x + y)/2). Since

u(x) u(y) p/ 21+p/ |u(x) ux,y |p/ + ux,y u(y) p/ ,
it follows that

uWp/
s 2
where

D(x) =
Rn
We note that

|xy|>
p/
D(x)
/p
dx
,
(10.3.5)
Rn
|u(x) ux,y |p/

dy
|x y|n+ps
/p
.
|u(x) ux,y |p/

2p/ n
p/
uL ps .
dy
|x y|n+ps
ps
(10.3.6)
= {(x, z) : x Rn , z > 0}
Let U be an arbitrary extension of u onto Rn+1
+
such that U L1 (Rn+1
+ , loc). By U x,y (z) we denote the mean value of U (, z)
in Bx,y . Using the identity
|xy|
(1s)p
= p(1 s)
z 1+p(1s) dz,
|x y|
0
we nd
|u(x) ux,y |p/

dy
|x y|n+ps
|xy|<

= p1/ (1 s)1/
|xy|<
|xy|

p
z 1+p(1s) u(x) ux,y dz
1/
dy
|x y|n+ps+(1s)p/
31+p/ p1/ (1 s)1/ (J1 + J2 + J3 ),
(10.3.7)
532
where

J1 :=
|xy|
z
|xy|<
1/
|xy|
z
|xy|<
p
u(x) U (x, z) dz
dy
,
|x y|n+ps+(1s)p/

J2 :=
1+p(1s)
1+p(1s)
1/
|U x,y (z) ux,y | dz
p
dy
,
|x y|n+ps+(1s)p/
and

J3 :=
|xy|
z
|xy|<

p
U (x, z) U x,y (z) dz
1+p(1s)
1/
dy
.
|x y|n+ps+(1s)p/
Clearly,

J1
|xy|<
p 1/
z
U (x, t)

t dt dz
0
dy
|x
|xy|<
z 1+p(1s)
y|n+ps+(1s)p/
|xy|
|xy|
z 1ps
p 1/
z
U (x, t)

t dt dz
0
dy
.
|x y|nps(1)/
By Hardys inequality

a

1sp
z

0
z
0
p

p

(t) dt dz s
p

z 1+p(1s) (z) dz,
one has
J1 sp/

|xy|<
n
p/
s ps(1 )
|xy|
0
z
0

1/
U (x, z) p
dy
dz
z 1+p(1s)
z
|x y|nps(1)/
p

1+p(1s) U (x, z)

dz
1/
ps(1)/ .
(10.3.8)
533
Duplicating the same argument, we conclude that J2 does not exceed

sp/
dy
|xy|<
|x y|nps(1)/
|xy|
p 1/

.
dz
1+p(1s) U x,y (z)
(10.3.9)
Let M denote the n-dimensional HardyLittlewood maximal operator

1
f () d.
(Mf )(x) = sup
r>0 |Br | Br (x)
Using the obvious inequality

U x,y (z)
M U (x, z),

z
z
we nd from (10.3.9)
J2
n
p/
s ps(1 )

0
p 1/

U
z 1+p(1s) M
dz
ps(1)/ . (10.3.10)
z
To estimate J3 we use the Sobolev-type integral representation in the form

given in Akilov and Kantorovich [38], Chap. 10, Sect. 3
U (x, z) U x,y (z) =
n

k=1
Bx,y
bk (, x) U (, z)
d,
|x |n1 k
(10.3.11)
where bk (, x) are continuous functions for x = admitting the estimate

n

bk (, x) |x y| .
n|Bx,y |
Clearly, (10.3.11) implies the estimate

n 1/2

U (x, z) U x,y (z) 2 n
n

Br (x)
where r = |x y|. Integrating by parts we nd

| U (, z)|
d
n1
Br (x) |x |

U (, z) d + (n 1)
= r1n
Br (x)

n|x y| M|U | (x, z).
r
0
| U (, z)|
d,
|x |n1
ds
sn

U (, z) d
Bs (x)
534
Therefore,

J3
2n n3/2
n
p/
|xy|<
dy
|x y|nps(1)/

(2n n3/2 )p/
(p)/
ps(1 )
|xy|

p
z 1+p(1s) M|U | dz
1/
1+p(1s)
M|U |
p
1/
dz
ps(1)/ .
(10.3.12)
Here and in the sequel, for the sake of brevity, by M|U | we mean (M|U |)(x,
z). Putting estimates (10.3.8), (10.3.10), and (10.3.12) into (10.3.7), we arrive
at

|u(x) ux,y |p/
dy
|x y|n+ps
|xy|<

1/

p
(1 s)1/
1+p(1s)
c(n)
M|U | dz
z
ps(1)/ .
1
0
This estimate together with (10.3.6) implies that D(x) is majorized by

c(n) uL

+
1s
1
1/p
1+p(1s)
M|U |
p
1/p
dz
s(1)

.
Minimizing the right-hand side, we conclude that

/p
/p

p
1s
1
1+p(1s)
D(x) c(n)
M|U | dz
uL
z
.
1
0
Hence and by (10.3.5)
/p
1s
c(n)
u1
L
1

/p

p
z 1+p(1s) M|U | dz dx
.
s
uWp/
Rn
Since
MuLp
n8n p
uLp
n (p 1)
(see Iwaniec [399], Sect. 2.5), it follows that the norm uWp/
s does not exceed

c(n)
p
p1

1s
1

0

Rn

p
/p
u1
L .
(10.3.13)
Now we dene U by the formula

(h)u(x + zh) dh,
U (x, z) :=
535
(10.3.14)
Rn
where

(h) = n n(n + 1) 1 |h| + ,
with the subscript plus standing for the nonnegative part of a real-valued
function. It follows directly from (10.3.14) that

U (x, z) n(n + 1)(n + 2)
u(x + zh) u(x) dh.
zn
|h|<1
Hence and by H
olders inequality

p
n
0
R

n
(n + 1)p (n + 2)p
z 1ps
n
0
|h|<1 Rn

p
u(x + zh) u(x) dx dh dz.
(10.3.15)
We have
z 1ps

|h|<1

u(x + zh) u(x) p dh dz

z

u(x + ) u(x) p d
z 1psn
n1 d
0
0
B

1

1
ps1
u(x + ) u(x) p d.
= (ps + n)
d
Thus,

0
B1

p
n(n + 1)p (n + 2)p
upWps . (10.3.16)
n (ps + n)
Rn
Combining (10.3.16) with (10.3.13) we complete the proof.

Remark 1. Let

uWp1 =
1/p
Rn

u(x) p dx
As a particular case of a more general inequality, Brezis and Mironescu [144]

obtained the inequality
cuLp u1
uWp/
L ,
0 < < 1.
536
They commented on this in the following way: We do not know any elementary (i.e., without the Littlewood-Paley machinery) proof of (10.3.2) when
s = 1. Obviously, the earlier proof of (10.3.3), complemented by the reference to formula (10.2.21), provides an elementary proof of the inequality
1
c(n, p)uLp uL

.
(1 )/p uWp/
The factor (1 )/p controls the blow-up of the norm in Wp/

as 1.
Remark 2. Note that passing to the limit as p in both sides of
(10.3.3) one obtains inequality (10.3.2) with p = and with a certain nite
constant c. Let us consider the case p 1 when the constant factor in (10.3.3)
tends to innity. It follows from (10.3.3) that the best value of c(n, p, ) in the
inequality
1
/p
uWp/
s c(n, p, )(1 s)
uWps uL
,
(10.3.17)
admits the upper estimate

lim sup(p 1) c(n, p, ) c(n)(1 ) .
(10.3.18)
p1
Now we obtain the analogous lower estimate

lim inf (p 1) c(n, p, ) 1.
p1
(10.3.19)
s
In fact, the characteristic function of the ball B1 belongs to Wps and Wp/
if and only if sp < 1, and there holds
Wp/
s = W s .
p
Putting u = into (10.3.17), where s = p1 with an arbitrarily small

> 0, we obtain

/p
,
1 c(n, p, ) (p 1)/p
which implies (10.3.19). Thus, the growth O((p 1) ) of the constant in
(10.3.3) as p 1 is the best possible.
10.4 Some Facts from Nonlinear Potential Theory

10.4.1 Capacity cap(e, Spl ) and Its Properties
With each function space Spl = Hpl , Wpl , Bpl , hlp , wpl , and blp introduced in
Sect. 10.1, we associate a set function called the capacity. Namely, for any
compactum e Rn we put

cap e, Spl = inf upS l : u C0 , u 1 on e .
p
537
If E is an arbitrary subset of Rn , then we call the numbers

cap E, Spl = sup cap(e, Spl ) : e E, e is a compactum ,

cap E, Spl = inf cap G, Spl : G E, G is an open set ,
the inner and outer capacities of E, respectively.
Theorem 10.1.3/3 implies the relation

cap e B (i) , Spl ,
cap e, Spl
(10.4.1)
i0
where {B (i) } is the sequence of balls introduced in Theorem 10.1.3/3. Similar

quasi-additivity relations for capacities were studied by D.R. Adams [6] and
Aikawa [34, 35].
We state certain well-known properties of the capacity cap(, Spl ), where
l
Sp = Hpl or hlp , p > 1 (cf. Reshetnyak [674], Meyers [596], Mazya and
Havin [567]).
1. If the set e Rn is compact then for each > 0 there exists an open
set G Rn such that G e and

cap e , Spl < cap e, Spl + ,
where e is an arbitrary compact subset of G.
2. If the set e Rn is compact, then

cap e, Spl = cap e, Spl .
3. If E1 E2 Rn then

cap E1 , Spl cap E2 , Spl ,
4. If {Ek }
k=1

cap E1 , Spl cap E2 , Spl .

is a sequence of sets in Rn and E = k Ek , then

cap E, Spl
cap Ek , Spl .
k=1
It is well known that any analytic (in particular, any Borel) set E
Rn is measurable with respect to the capacity cap(, Spl ) (i.e., cap(E, Spl ) =
cap(E, Spl )) (cf. Meyers [596], Mazya and Havin [567]).
We introduce one more capacity

ck,p (E) = inf f pLp : f Lp , f 0 and

k |x y| f (y) dy 1 for all x E ,
538
where k is a positive decreasing continuous function on the half-axis (0, +)

(cf. Meyers [596]).
We list some connections between the two capacities.
(i) If Spl = Hpl or hlp and k is the Bessel or Riesz kernel, then

ck,p (E) = c cap E, Spl .
(ii) If diam E 1 and pl < n, then

cap E, Hpl cap E, hlp
(10.4.2)
(cf. Adams and Meyers [16]). Relations similar to (10.4.2) are also valid for
other pairs of spaces, e.g., Bpl , blp and Wpl , wpl .
(iii) If 1 0, 1 0, 1 < p, and q < . For E Rn the inequality
nlp
nmq

cap E, hm
c cap E, hlp
,
q
mq < lp < n,
is valid.
(vi) If, in addition, E B1 , then
nlp
nmq

cap E, Hqm
c cap E, Hpl
, mq < lp < n,
1p

c0
c cap E, Hpl , mq < lp = n,
log
m
cap(E, Hq )
p1
q1

c cap E, Hpl
, mq = lp = n, p q.
cap E, Hqm
Putting E = Br , we conclude that all power exponents here are exact.

Items (v) and (vi) are due to Adams and Hedberg [14].
10.4.2 Nonlinear Potentials
Nonlinear potentials were introduced in the article by Mazya and Havin [566]
and their theory has turned out to be a useful tool in various areas. It was
developed in the papers by Mazya and Havin [567], D.R. Adams and Meyers [17], Hedberg and Wol [372], D.R. Adams [9], et al. The reader interested
in a detailed exposition of the nonlinear potential theory can nd it in the
539
comprehensive monograph by D.R. Adams and Hedberg [15]. Here we collect

some facts from this theory.
Let p (1, ), n > pl. Each nonnegative measure given on the Borel
-algebra of the space Rn generates the function Up,l dened on Rn by

(Up,l )(x) =
or equivalently,
|x y|
1/(p1)
ln
Rn
|z y|
ln
Rn
Up,l = Il (Il )p 1 ,
d(z)
dy,
(10.4.3)
p + p = pp .
For p = 2, by changing the order of integration in (10.4.3) and taking into

account the composition formula

|y z|ln |y x|ln dy = const |z x|2ln ,
(cf. Landkof [477]), we obtain

(U2,l )(x) = c
d(z)
.
|z x|n2l
The function U2,l is the Riesz potential of order 2l (for l = 1 it is the Newton potential). Similarly, Up,l is called the nonlinear Riesz potential ((p, l)potential).
The nonlinear Bessel potential is dened as

Vp,l = Jl (Jl )p 1 .
The potentials Up,l and Vp,l satisfy the following rough maximum principle for nonlinear potentials.
Proposition 1. Let P be one of the potentials Up,l or Vp,l . Then there
exists a constant M that depends only on n, p, and l, such that

(P )(x) M sup (P )(x) : x supp .
This assertion was proved in the papers by Mazya and Havin [567] and
D.R. Adams and Meyers [16]. It is well known that we can take M = 1 for
p = 2, l 1 (cf. Landkof [477]). In general, this is impossible even for p = 2
(cf. Landkof [477]).
The next assertion contains basic properties of the so-called (p, l)-capacitary measure (cf. Meyers [596], Mazya and Havin [567]).
Proposition 2. Let E be a subset of Rn . If cap(E, hlp ) < , then there
exists a unique measure E with the following properties:
p/(p1)
1. Il E Lp/(p1) = cap(E, hlp );
540
2. (Up,l E )(x) 1 for (p, l)-quasi all x E (the notion (p, l)-quasi everywhere means everywhere except for a set of zero outer capacity cap(, hlp ));
3. supp E E;
= cap(E, hl );
4. E (E)
p
5. (Up,l E )(x) 1 for all x supp E .
The measure E is called the capacitary measure of E and Up,l E is called
the capacitary potential of E. The last proposition remains valid after the
replacement of hlp by Hpl , of Up,l by Vp,l , and of Il by Jl . Besides, we note
that the capacity cap(e, Spl ) (for Spl = hlp or Hpl ) can be dened as

cap e, Spl = sup (e) : supp e, (P )(x) 1 ,
(10.4.4)
xe
where P = Up,l or Vp,l (cf. Mazya and Havin [567]). Next we present some
pointwise estimates for (p, l)-potentials that are obvious in the linear case and
nontrivial in the nonlinear case.
Proposition 3. (Mazya and Havin [567] and D.R. Adams [4]).
(i) If 2 l/n 1 and () = supx (B(x, )), then

(Vp,l )(x) c
0
()
nlp
1/(p1)
d
.

ec0
(10.4.6)
The same estimates, without the factor ec0 , are valid for the potential
Up,l .
It is almost obvious that the following estimate, opposite to (10.4.5),

(Vp,l )(x) c
(B(x, ))
nlp
1/(p1)
ec0
d
,

(10.4.7)
holds for all p (1, ), l > 0, whereas (10.4.5) is not true for p 2 l/n.
Wol showed (cf. Hedberg and Wol [372]) that for pl < n the inequality

p/(p1)
Il Lp/(p1) c Wp,l d,
(10.4.8)
where

(Wp,l )(x) =
0
(B(x, ))
nlp
1/(p1)
d

is valid. The analogous inequality for Bessel potentials is
541

p/(p1)
Jl Lp/(p1) c
Sp,l d,
(10.4.9)
where pl n and

(Sp,l )(x) =
0
(B(x, ))
nlp
1/(p1)
ec0
d
.

(10.4.10)
The estimates converse to (10.4.8) and (10.4.9) follow immediately from

(10.4.7).
Remark. The failure of the inequality (10.4.5) for p 2l/n caused serious
diculties in attempts at a satisfactory generalization of the basic facts of the
classical potential theory to the nonlinear case.
Using inequalities (10.4.8) and (10.4.9) Hedberg (cf. the previously mentioned paper by Hedberg and Wol) managed to surmount this diculty by
virtue of an analog of the nonlinear potential theory in which the roles of
Up,l and Vp,l are played by certain nonlinear potentials that are equivalent
to Wp,l and Sp,l (see also D.R. Adams and Hedberg [15], Sects. 4.4 and 4.5).
Upper pointwise estimates similar to (10.4.5) are obtained for the case
1 < p 2l/n under the additional assumption that the potential is bounded.
Namely, the following proposition is true.
Proposition 4. (D.R. Adams and Meyers [16]).
(i) If 1 < p < 2 l/n and (Up,l )(x) K for all x Rn , then

(Up,l )(x) cK
0
(B(x, ))
nlp
(n1)/(nlp)
d
,

(10.4.11)
where = ((2 p)n l)/(n lp).

(ii) If p = 2 l/n and (Up,l )(x) K for all x Rn , then

(Up,l )(x) c
0
p 1

(B(x, ))
nlp
d
p1
. (10.4.12)
log cK
nlp

(B(x, ))

(The condition (Up,l )(x) K for all x Rn implies the estimate

B(x, ) e1 aK p1 nlp .)
10.4.3 Metric Properties of Capacity

The following relations are useful (cf. Meyers [596]).
If pl < n and 0 < < 1, then
cap(B , Hpl ) npl .
(10.4.13)
542
If pl < n and 0 < < , then

cap B , hlp = cnpl .
(10.4.14)
For pl = n, 0 < 1 we have

cap B , Hpl (log 2/)1p .
(10.4.15)
If p l > n, then cap({x}, Hpl ) > 0. Thus, only the empty set has zero capacity
if pl > n.
The following equivalence relations for the capacity of a parallelepiped
were obtained by D.R. Adams [7].
Proposition 1. Let 0 < a1 a2 an , a = (a1 , a2 , . . . , an ) and let
Q(a) = {x Rn : |xj | aj , j = 1, . . . , n}.
(i) If k 1 < lp < k, k = 1, . . . , n, then
n

cap Q(a), hlp aklp
aj .
k
j=k+1
(Here the product equals unity provided k = n.)

(ii) If lp = k, k = 1, 2, . . . , n 1, then

cap
Q(a), hlp

1p
n
ak+1
min log
,1
aj .
ak
j=k+1
Similar two-sided estimates hold for cap(Q(a), Hpl ).

If T is a quasi-isometric mapping of Rn onto itself, then cap(T E, Spl )
cap(E, Spl ), where Spl = Hpl or hlp . This is a simple corollary of (10.4.4).
Meyers [597] showed that cap(P E, Spl ) cap(E, Spl ) provided P is a projection Rn Rk , k < n, and Spl = Hpl or hlp .
For any set E Rn and for a nondecreasing positive function on [0, )
we dene the Hausdor -measure

(ri ),
H(E, ) = lim inf
+0 {B(i) }
where {B (i) } is any covering of the set E by open balls B (i) with radii ri < .
If (t) = tq , then d is called the dimension of the Hausdor measure. The ddimensional Hausdor measure Hd (E) is equal to vd H(E, td ) (cf. Sect. 1.2.4).
For d = n the measure Hn coincides with the n-dimensional Lebesgue measure
mn .
The following propositions contain noncoinciding, but in a certain sense,
exact necessary and sucient conditions for positiveness of the capacity formulated in terms of the Hausdor measures.
543
Proposition 2. Let 1 0.
(The last fact is a corollary of (10.4.6); cf. Mazya and Havin [567].)
Proposition 3. Let E be a Borel set in Rn .
1. If n > pl and Hnpl (E) < , then cap(E, Spl ) = 0, where Spl = hlp or Hpl .
2. If n = pl and H(E, ) < , where (r) = | log r|1p , then cap(E, Hpl ) = 0
(cf. Meyers [597], Mazya and Havin [567]).
Next we present one more sucient condition for the vanishing of
cap(E, Hpl ) (Mazya and Havin [567]).
Proposition 4. Let N be a measurable nonnegative function on [0, ).
Suppose, for any positive r, the set E can be covered by at most N (r) closed
balls whose radii do not exceed r.
If

1/(1p) (npl)/(1p)1
r
dr = ,
N (r)
0
then
cap(E, Hpl )
= 0.
Using Propositions 2 and 4, we can give a complete description of the

n-dimensional Cantor sets E with positive cap(E, Hpl ).
Let L = {lj }
j=1 be a decreasing sequence of positive numbers such that
2lj+1 < lj (j = 1, 2, . . . , ) and let 1 be a closed interval with length l1 . Let e1
denote a set contained in 1 , which equals the union of two closed intervals 2
and 3 with length l2 and which contains both ends of the interval 1 . We put
E1 = e1 e1 e1 . Next we repeat the procedure with the intervals 2

n-times
and 3 (here the role of l2 passes to l3 ) and thus obtain four closed intervals
with length l3 . Let their union be denoted by e2 ; E2 = e2 e2 e2 and

n-times
so on.
We put

Ej .
E(L ) =
j=1
Proposition 5. (Mazya and Havin [567]). The following properties are

equivalent:
544
(i) cap(E(L ), Hpl ) > 0;

(np l)/(1p)
2jn/(1p) lj
<
(ii)
for n > pl;
j1
2jn/(1p) log
j1
lj
<
lj+1
for n = pl.
Sharp results on metric properties of capacities generated by the Besov

l
spaces Bp,q
, 0 < p < , 0 < q , 0 < lp < n, were obtained by Netrusov
in [633].
10.4.4 Rened Functions
The function in Hpl is called rened or (p, l)-rened if there exists a sequence
of functions {m }m1 in D that converges to in Hpl and such that for each
> 0 there exists an open set with cap(, Hpl ) < and m uniformly
on Rn \.
Another (equivalent) denition is as follows: The function Hpl is called
rened if for each > 0 there exists an open set such that cap(, Hpl ) <
and the restriction of to Rn \ is continuous.
We list the basic properties of rened functions.
(i) If Hpl then there exists a rened function that coincides with
almost everywhere (with respect to n-dimensional Lebesgue measure) in Rn .
(ii) If 1 and 2 are rened functions that coincide almost everywhere
(with respect to n-dimensional Lebesgue measure), then 1 and 2 coincide
quasi-everywhere.
(iii) Each sequence of rened functions in Hpl that converges to a rened
function in Hpl contains a subsequence that converges to quasi-everywhere.
For the proofs of these assertions see the paper by Havin and the author [567], where references to the earlier literature are given.
For pl > n these properties become trivial since Hpl C.
The following result due to Bagby and Ziemer [62] shows that a function
in Hpl coincides with a function in C m (m l) outside some set that is small
with respect to the corresponding capacity.
Proposition. Let u Hpl , 1 0 there exists a function u C m and an open set such
that cap(, Hplm ) < and u(x) = u (x) for all x Rn \.
Swanson proved that every function in the Bessel potential space Hpl may
be approximated in capacity and norm by smooth functions in C m, , 0 <
m + < l [734].
In conclusion we add to the previously mentioned literature on nonlinear
potentials the lectures by D.R. Adams [8], which also contain a survey of some
545
other problems that we do not touch upon here, and also D.R. Adams very
interesting reminiscences [12].

Section 10.1. Besides the literature cited in Sect. 10.1, in which embedding, trace, and extension theorems are established for various fractionalorder and more general spaces, we also cite (with no intention of claiming completeness) the papers by Solonnikov [716]; Besov [87]; Golovkin and
Solonnikov [323]; Aronszajn, Mulla, and Szeptycki [53]; Taibleson [737, 738];
Golovkin [321, 322]; Volevic and Panejah [782]; Ilin [396, 397]; Burenkov [152,
153]; Kudryavtsev [467]; the rst chapter of the book by Hormander [384]; and
the book by Gelman and the author [305].
Here we mention some works dedicated to trace and extension theorems for
Besov and Sobolev spaces. The following result concerning the case 0 1/p, p 1 and for s > 1 n + n/p, p < 1. Moreover, this operator
s1/p
s
has a right inverse that is bounded from Bp,q (Rn1 ) to Bp,q
(Rn ) for every
s R.
The borderline case s = 1 n + n/p, 0 < p < 1, was studied by Johnsen

[403].
A. Jonsson [405] (see also A. Jonsson and H. Wallin [408]) obtained generalizations of Theorems 10.1.1/1 and 10.1.1/2 for traces on the so-called d-sets
in Rn , i.e., the sets that support a measure subject to (B(x, r)) rd .
s
Explicit formulas for linear continuous extension operators Bp,q
()
s
n
s
s
n
Bp,q (R ) and Fp,q () Fp,q (R ), where B and F stand for Besov and
TriebelLizorkin spaces, 0 < p < , 0 < q , and is a Lipschitz
domain, are given by Rychkov [686].
A description of the traces of the functions in Wp1 () to the boundary of
a cusp domain was given by Poborchi and the author ([576], Chap. 7). Let
be a domain in Rn and p [1, ). By T Wp1 () we mean the space of the
traces u| of the functions u Wp1 (). The norm in this space is dened by

f T Wp1 () = inf uWp1 () : u Wp1 (), u| = f .
11/p
According to Gagliardos theorem, T Wp1 () = Wp

() for p (1, )
and T W11 () = L1 () if is a bounded Lipschitz domain. When has
cusps on the boundary, Gagliardos theorem generally fails. Consider a typical
domain with an outward cusp

= x = (y, z) Rn : z (0, 1), |y| < (z) , n > 2,
546

(0) = limz0 (z) = 0.
It turns out that the boundary values of functions in Wp1 () and in
1
Wp (Rn \ ) can be characterized in terms of the niteness of the norm

f p, =

f (x) p q(x) dsx

f (x) f () p Q(x, ) dsx ds
1/p
,
where q and Q are nonnegative weight functions and dsx , ds the area elements
on .
Let p (1, ) and let f be a function on vanishing outside a small
neighborhood of the origin. Then f belongs to T Wp1 () if and only if f p, <
, where x = (y, z), = (, ), 0 q(x) const (z),

|x |2np if |z | < (z) + (), z, (0, 1),
Q(x, ) =
0
otherwise.
In addition, the norm f p, is equivalent to
A necessary and sucient condition for f
that f p, < with
1p
(z)
q(x) = ((z)| log((z)/z)|)1p
(z)2n
f T Wp1 () .
to belong to T Wp1 (Rn \ ) is
for 1 n 1,
and Q(x, ) = 0 only if z, (0, 1). For these pairs x, , Q is dened as

follows. If p < n 1, then
Q(x, ) = |x |2np .
In the cases p n 1 and |x | < (z) + () the weight Q is determined
by the same formula. Finally, if |x | (z) + (), then
Q(x, ) =
and

p
|x |
((z) + ())2(1p)
log 1 +
,
|x |
(z) + ()

2n
Q(x, ) = |x |np2 (z)()
,
p = n 1,
p > n 1.
The norm f p, with these weights q, Q is equivalent to f T Wp1 (Rn \) .

If p = n 1, some additional restrictions are imposed on (not excluding
power cusps). A function f dened on and vanishing outside a small
neighborhood of the origin is in T W11 () if and only if f 1, < with
0 q(x) const (z),

((z) + ())1n
Q(x, ) =
0
547
for z, (0, 1), |z | < (z) + (),

otherwise.
Furthermore, the norms f T W11 (Rn \) and f 1, are equivalent.

The characterization of the space T W11 (Rn \) is the same as for Lipschitz
domains

T W11 Rn \ = L1 ().
In all the cases mentioned previously there exists a bounded extension
operator

T Wp1 () Wp1 () and T Wp1 Rn \ Wp1 Rn \ .
This operator is linear for p > 1 and nonlinear for p = 1. One can easily
obtain from the previous results that the space T Wp1 (Rn \ ) is continuously
embedded into T Wp1 () for p [1, ), n > 2. Hence the space of the traces
on of the functions in Wp1 (Rn ) coincides with T Wp1 (Rn \ ).
Shvartsman [699] described the restrictions of the space Wp1 (Rn ), p > n,
to an arbitrary closed subset S Rn via certain doubling measures supported
on S. Observe that Dynkin [245, 246] conjectured that every compact set
S Rn carries a nontrivial doubling measure . He constructed a doubling
measure on every compact set S R that satises a certain porosity condition. Dynkins conjecture was subsequently proved by Volberg and Konyagin
[780]. Moreover, Volberg and Konyagin showed that every compact set in Rn
carries a nontrivial measure satisfying the following condition: There exists
a constant C = C > 0 such that, for all x S, 0 < r 1 and 1 k 1/r,
we have (B(x, kr)) C k n (B(x, r)). Using their argument, Luukkainen
and Saksman [511] extended this result to all closed subsets of Rn . Jonsson
[406] showed that it may also be assumed that (B(x, 1)) 1 for all x S.
Jonsson and Wallin [405, 408] obtained generalizations of Theorems 1 and
2 for the family of d-sets in Rn , 0 < n d < p, 1 n, i.e., the quantity

f Wp1 (Rn )|S := inf F Wp1 (Rn ) : F Wp1 (Rn ) and continuous, F |S = f
can be calculated as follows:
f Wp1 (Rn )|S f Lp () + sup

0<t1
+
S (x,y)<1
|xy|<t
|f (x) f (y)|p
d(x) d(y)
tpn (B(x, t))2
p1
|f (x) f (y)|p
d(x) d(y)
S (x, y)pn (B(x, S (x, y)))2
p1
.
548
Sections 10.2 and 10.3. For the historical background we refer to the
introductions to these sections. The material is borrowed from the papers by
Mazya and Shaposhnikova [584, 585]. Results related to (10.2.3) and (10.3.1)
were subsequently obtained by Kolyada and Lerner [445], M. Milman [606],
and Karadzhov, M. Milman, and Xiao [414].
Section 10.4. The references are given in the text. Here we mention
additionally the paper by Mazya and Havin [569], where several applications
of the (p, l)-capacity and nonlinear potentials to the theory of exceptional
sets are considered (the uniqueness and approximation of analytic functions,
convergence of Fourier series, and removal of singularities of analytic and
polyharmonic functions). Properties of capacities in BirnbaumOrlicz spaces
were studied by Aissaoui and Benkirane [37].
11
Capacitary and Trace Inequalities for Functions
in Rn with Derivatives of an Arbitrary Order
11.1 Description of Results

According to Corollary 2.3.4, for q p > 1, the inequality
uLq (Rn ,) AuLp (Rn ) ,
u C0 ,
(11.1.1)
follows from the isocapacitary inequality

p/q
(E)

pp (p 1)p1 Ap cap E, wp1 .
Here and henceforth E is an arbitrary Borel set in Rn and wp1 is the completion
of C0 with respect to the norm uLp .
On the other hand, if (11.1.1) is valid for any u C0 , then
p/q
(E)

Ap cap E, wp1
for all E Rn .
The present chapter contains similar results in which the role of wp1 is
played by the spaces Hpl , hlp , Wpl , wpl , Bpl , and blp .
Namely, let Spl be any one of these spaces. Then the best constant in

uL () AuS l , u C0 ,
(11.1.2)
q
p
where q p, is equivalent to the best constant in the isoperimetric inequality

p/q

B cap E, Spl .
(E)
(11.1.3)
The estimate A B immediately follows from the denition of capacity.
The reverse estimate is a deeper fact, its proof being based on the capacitary
inequality

cap Nt , Spl tp1 dt CupS l ,
(11.1.4)
0

DOI 10.1007/978-3-642-15564-2 11,
549
550
where u Spl , Nt = {x : |u(x)| t} and C is a constant independent of u. We

dealt with inequalities of a similar nature in the previous chapter, but only in
the case l 2. In Sect. 11.2 we present two proofs of inequality (11.1.4) with
l (0, ) and dierent elds of application and in Sect. 1.3 we discuss criteria
for the validity of embedding theorems formulated in terms of isocapacitary
inequalities. Section 1.4 is dedicated to a counterexample showing that the
capacitary inequality for the norm in L22 () may fail for some domain.
We might ask if it is possible to replace arbitrary sets E in (11.1.3) by
balls. From the D.R. Adams Theorem 1.4.1 it follows that this is so for the
Riesz potential space Spl = hlp , pl < n. The condition given by D.R. Adams is

(11.1.5)
B(x, ) Cs ,
where s = q(n/p l) and B(x, ) is any ball with center x and radius .
Thus, inequality (11.1.5) with q > p implies the isoperimetric inequality (11.1.3) for any set E.
In Sect. 11.7 we give a direct proof of more general assertions of this kind.
Namely, for any ball B(x, r), let

(11.1.6)
B(x, r) cap Br , hlp ,
where Br = B(0, r), is an increasing function subject to some additional
requirements and is a measure in Rn . Then for all Borel sets E Rn

(11.1.7)
(E) c c cap E, hlp .
By this theorem along with the equivalence of (11.1.2) and (11.1.3), we
show in Sect. 11.8 that inequalities similar to (11.1.6) are necessary and sufcient for the validity of estimates for traces of Riesz and Bessel potentials
in BirnbaumOrlicz spaces LM () and, in particular, in Lq (). Besides, this
gives a new proof of the aforementioned D.R. Adams theorem where no interpolation is used. Another corollary, of interest in its own right, claims that
the inequality
uLq () cuHpl ,
where q > p > 1, lp = n is fullled if and only if

B(x, r) c| log r|q/p
for all balls B(x, r) with radii r (0, 12 ).
Next we state some other results relating the conditions for (11.1.2).
(a) If Spl = Hpl , pl < n, q > p, then (11.1.2) is valid simultaneously
with (11.1.5), where 0 < < 1 (see Sect. 11.8).
(b) In the case pl > n a necessary and sucient condition for (11.1.2) with
Spl = Hpl is

sup B(x, 1) : x Rn <
(see Sect. 11.8).
11.1 Description of Results
551
(c) For q = p, condition (11.1.5) is not sucient for (11.1.2) (cf. Remark
11.8/2). So in this case, which is probably the most important for applications,
we have to deal with a less explicit isocapacitary condition (11.1.3). However,
there are ball and pointwise criteria for (11.1.2) with q = p due to Kerman
and Sawyer and Mazya and Verbitsky. They are treated in Sect. 1.5.
(d) In Sect. 11.6 we give the following necessary and sucient condition
for the validity of (11.1.2) provided that p > 1 and p > q > 0:
q/(pq)

t
dt < ,
(t)
0
where (t) = inf{cap(F, Spl ) : (F ) t} (cf. Theorem 11.6.1/1). A noncapacitary criterion for (11.1.2), where p > q, obtained by Cascante, Ortega, and
Verbitsky, is formulated in the same Sect. 1.6.
(e) In the case pl > n, p > q a necessary and sucient condition for (11.1.2)
can be written in the essentially simpler form

p/(pq)
< ,
Q(i)
i
where {Q(i) } is the sequence of closed cubes with edge length 1, which forms
the coordinate grid in Rn (cf. Sect. 11.6.2).
(f) We note also that in the case q = 1, p > 1, the inequality (11.1.2) with
Spl = hlp or Spl = Hpl is equivalent to the inclusion Il Lp or Jl Lp ,
respectively (here Il and Jl are the Riesz and the Bessel potentials) (cf. Sect.
11.6.2).
(g) In Sect. 11.10 we consider the case p = 1. For S1l = bl1 in addition to
Theorem 1.4.3 it is shown that (11.1.2) holds simultaneously with (11.1.5),
where p = 1, q 1. If S1l = B1l , we have to add the condition (0, 1)
in (11.1.5). We recall that according to Theorem 1.4.3 the same pertains to
the cases sl1 = w1l , S1l = W1l .
(h) Using the interpretation of blp and Bpl , p > 1 as the trace spaces of the
corresponding potential spaces, we obtain theorems on blp and Bpl from the
theorems concerning hlp and Hpl (cf. Remark 11.8/3).
Section 11.9 contains some applications of the results obtained in
Sects. 11.211.7. In particular, we present necessary and sucient conditions
for the compactness of the embedding operator of Hpl into Lq (). In Sect. 11.9
we also state some corollaries to previous theorems. They concern the negative
spectrum of the operator ()l p(x), p(x) 0, x Rn .
Integral inequalities with certain seminorms of a function, involving arbitrary measures, on the left-hand side and the L1 norm of the gradient on the
right-hand side are characterized in Sect. 11.10. The concluding Sect. 11.11 is
devoted to the study of the inequality
1/q

u(x) u(y)q (dx, dy)
CuLp () ,
with q p > 1.
552
11.2 Capacitary Inequality of an Arbitrary Order

11.2.1 A Proof Based on the Smooth Truncation of a Potential
We use the HardyLittlewood maximal operator M dened by

1
g() d.
(Mg)(x) = sup
m
B
r>0
n r B(x,r)
(11.2.1)
Lemma 1. (Hedberg [365]) Let f be a nonnegative function and let I be

the Riesz potential of order . For all and such that 0 < < < n and
almost all x Rn there holds the inequality

/
1/
,
(Mf )(x)
(I f )(x) c (I f )(x)
0 < < < n.
(11.2.2)
Proof. Let t be an arbitrary positive number to be chosen later. We make

use of the equality

t
f (y) dy
ds
n
= (n )
f (y) dy n+1 + t
f (y) dy,
n
s
0
Bt (x) |x y|
Bs (x)
Bt (x)
(11.2.3)
which is readily checked by changing the order of integration on the right-hand
side. Hence

f (y) dy
ct (Mf )(x).
(11.2.4)
n
|x
y|
Bt (x)
Clearly we have

Rn \Bt (x)
f (y) dy
t
|x y|n
t

Rn \Bt (x)
f (y) dy
|x y|n
(I f )(x).
(11.2.5)
Adding this inequality and (11.2.3), we obtain

(I f )(x) ct (Mf )(x) + t (I f )(x).
Minimization of the right-hand side in t completes the proof.
Lemma 2. Let l be an integer, 0 < l < n, Il f = |x|ln f , where f 0

and let F be a function in C l (0, ) such that

tk1 F (k) (t) Q, k = 0, . . . , l.
Then

l F (Il f ) cQ Mf + |l Il f |
almost everywhere in Rn .
553
Proof. We have

(k)
l F (u) c
F (u)
k=1
cQ
|j1 u| |jk u|
j1 ++jk =l
k=1 j1 ++jk
Since |u| Ils f , it follows that

l

l F (u) cQ |l Il f | +
|j u|
|j1 u|
1jk /l .
1j1 /l
k
u
u
=l
k=1 j1 ++jk =l

Ilj1 f Iljk f
,
(Il f )1j1 /l (Il f )1jk /l

where the sum
is taken over all collections of numbers j1 , . . . , jk less than l.
The result follows from (11.2.2).

Let w be a nonnegative function in Rn satisfying the Muckenhoupt condition

p1

1
1
p
p
w dx
w
dx
< ,
(11.2.6)
sup
mn Q Q
mn Q Q
Q
where the supremum is taken over all cubes Q. This condition ensures the
continuity of the operators M and l Il in the space of functions with the
nite norm wLp (cf. Muckenhoupt [621], Coifman and Feerman [210]).
Theorem. Let p > 1, l = 1, 2, . . . , lp < n. Inequality (11.1.4), where Spl is
the completion of C0 with respect to the norm wl uLp , holds.
Proof. Let u C0 (Rn ), u = Il f , v = Il |f |. We can easily see that v
C (Rn ) and v(x) = O(|x|ln ) as |x| . Since v(x) |u(x)|, putting tj =
2j (j = 0, 1, . . . ,), we have
l

0

cap Nt , Spl d tp c
2pj j ,
(11.2.7)
j=
where j = cap({x : v(x) tj }, Spl ).

We shall use an operator of smooth level truncation. Let [0, 1] be a
nondecreasing function, vanishing near 0, and equal to one in a neighborhood
of 1, and let us introduce the function f C ((0, )) given on [tj+1 , tj ] by

F (u) = tj+1 + (u tj+1 )/(tj tj+1 ) (tj tj+1 ).
Then we obtain by denition of the capacity that

p
2pj j cF (v)S l ,
p
which implies
554

+

p
2pj j cF (v)S l .
p
j=
By Lemma 2, the preceding norm is majorized by

c wM|f |Lp + wl Il |f |Lp .
(11.2.8)
Since the weight function w satises (11.2.6), we can use the boundedness of
the operators M and l Il in the weighted Lp to show that the sum (11.2.8)
does not exceed

cwf Lp = c1 w()l Il uL cwl uLp .
p

Corollary. Inequality (11.1.4) holds, where
Spl
blp ,
p > 1, l > 0.
Proof. Let U be an arbitrary extension of u C0 (Rn ) to the space Rn+1 =

{X = (xn , xn+1 ) : x Rn , xn+1 R1 }. According to Theorem 10.1.1/1,
ublp (Rn ) inf U
[l]+1
1[l]1/p ,
(Rn+1 ,x
)
Lp
{U }
n+1
where
Lkp (Rn+1 , w) is the completion of C0 (Rn+1 ) with respect to the norm
wk uLp (Rn+1 ) . Consequently,

n+1 1{l}1/p
, xn+1
cap e, blp Rn cap e, L[l]+1
R
.
p
1[l]1/p
We can easily check that the function X w(X ) = xn+1

es (11.2.6). Therefore, the last theorem yields

cap Nt , blp Rn d tp cU p [l]+1 n+1 1[l]1/p .
Lp
(R
,xn+1
satis-
We complete the proof by minimizing the right-hand side over all extensions

of u to Rn+1 .
11.2.2 A Proof Based on the Maximum Principle for Nonlinear
Potentials
Let K be the linear Bessel or Riesz potential of order l with density and let

K(K)p 1 be the nonlinear potential generated by K. Further, let M denote

the constant in the rough maximum principle for the potential K(K)p 1
(cf. Proposition 10.4.2/1).
Theorem. Inequality

0

cap Nt , Spl tp1 dt CupS l
(11.2.9)
555
holds, where Spl is either Hpl or hlp (pl < n). The best constant C in (11.2.9)
satises
C (p )p1 M if p 2,
(11.2.10)
C (p )p p1 Mp1 if p < 2.
(Note that for p = 2 both estimates give C 2M.)
Proof. For the sake of brevity, let c(t) = cap(Nt , Spl ). It suces to assume
u = Kf , f 0, f Lp . Let t denote the capacitary measure of Nt (cf.
Proposition 10.4.2/2).
We choose a nite number N of arbitrary values t1 , . . . , tN satisfying 0 <
t1 <, . . . , < tN and introduce the step function N on (0, ) by
for t (0, t1 ),
0
N = tj for tj t < tj+1 , j = 1, . . . , N 1,
tN for t tN .
Now the sum
f (tj )(tj+1 tj ),
1jN 1
where f is a function dened on (0, ), can be written as the Stieltjes integral

f (t) dN (t).
0
To obtain (11.2.9), it suces to prove the inequality

c(t)tp1 dN (t) CupS l ,
(11.2.11)
with a constant C subject to (11.2.10). The left-hand side in the last inequality
does not exceed

p2
dN (t) = f dx
Kt tp2 dN (t),
Kf dt t
0
which is majorized by

f Lp

tp2 Kt dN (t)
L p
Thus, to get the result it suces to obtain the estimate

0
p
tp2 Kt dN (t)
dx C p 1

0
First, we note that by the maximum principle
c(t)tp1 dN (t).
(11.2.12)
556
(K )p 1 Kt dx Mc(t)
(11.2.13)
for all (0, ). Next, we consider separately the cases p 2 and p < 2.
Let p 2. The left-hand side in (11.2.12) can be written as

p 1
Kt t
p2
p2 dN ( ) dx.
dN (t)
By virtue of the Holder inequality this expression is majorized by

p

2p
p1 (K )p dN ( ) dx

(K )
p 1
p 1
Kt t
p2
dN (t) dN ( ) dx
which by (11.2.13) does not exceed

p Mp 1

2p
Kt pLp tp1 dN (t)
p 1
c(t)tp1 dN (t)
Thus (11.2.11) follows for p 2.

Let p < 2. The left-hand side in (11.2.12) is equal to

Kt t
p2
p 1
dN (t)
p2
dN ( )
dx.
Hence, by Minkowskis inequality, it is majorized by

p
t
p 1
(K )
0
p1
Kt dx
p 1
p2
dN ( )
tp2 dN (t).
Estimating this value by (11.2.13), we obtain that it is majorized by

pM

c(t)
0
and (11.2.11) follows for p < 2.
p 1
p2
tp2 dN (t),
11.3 Conditions for the Validity of Embedding Theorems

in Terms of Isocapacitary Inequalities
We state the generalization of Theorem 2.3.3 to the case of Bessel and Riesz
potential spaces in Rn . We omit the proof since it duplicates that of Theorem 2.3.3.
11.3 Conditions for the Validity of Embedding Theorems in Terms
Theorem. The best constant in the inequality

p
|u|
AupS l ,
L ()
M
557
(11.3.1)
where Spl = hlp for pl < n or Spl = Hpl for pl n, p (1, ), is equivalent to

B = sup

(E)N 1 (1/(E))
n
l
:
E
R
>
0
.
,
cap
E,
S
p
cap(E, Spl )
Namely, B A pBC, where C is the constant in (11.1.4) (cf. Theorem 11.2.2).

This assertion immediately implies the following corollary.
Corollary. The best constant Cp,q in
uLq () Cp,q uSpl ,
(11.3.2)
where q p > 1 and Spl is one of the spaces in the preceding theorem, satises
Bp,q Cp,q Bp,q (pC)1/p .

Here
Bp,q = sup

(E)p/q
: E Rn , cap E, Spl > 0
l
cap(E, Sp )
and C is the constant in (11.1.4).

A theorem due to D.R. Adams [5] states that inequality (11.3.2) with
q = p > 1, lp < n, and Spl = hlp holds if and only if, for all compact sets
e Rn ,

Il e pLp const (e),
(11.3.3)
where e is the restriction of the measure to e.
This result follows from the preceding corollary and the next proposition.
Proposition 1. Let p (1, ), lp < n. Then we have the relation Q R
where
Il e pLp
(e)
,
,
R
=
sup
Q = sup
l
p1
e cap(e, hp )
e [(e)]
and the suprema are taken over all compacta e in Rn .
Proof. For any u C0 , u 1 on e, we obtain

(e) u(x) de (x) ()l/2 e L ()l/2 uL ,
p
p
which can be rewritten as
558
(e) cIl e Lp uhlp .

Taking the minimum of the right-hand side over all functions u, we obtain

1/p

.
(e) cR1/p (e)1/p cap e, hlp
On the other hand, by the Corollary,

|u|p d cQuphl .
p
Therefore,

p

u de cQ(e)p1 ()l/2 up ,

L p
which yields
Il e Lp cQ1/p (e)1/p .

Thus h cQ. The proof is complete.
In the same way, we can obtain the relation

p
sup
e
Jl e Lp
(e)
sup
,
l
p1
cap(e, Hp )
e [(e)]
where e is either an arbitrary compactum in Rn or a compactum with a

diameter not exceeding unity.
An important renement of Proposition 1 due to Kerman and Sawyer [420]
will be considered in Sect. 11.5.
To conclude the present section, we note that with the same arguments as
in the proof of Proposition 1 together with Theorem 11.2.1 we arrive at the
following proposition.
Proposition 2. Let p > 1, l = 1, 2, . . . , lp < n and let w be a nonnegative function that satises the Muckenhoupt condition (11.2.6). Then the best
constant Cp in
uLp () Cp wl uLp , u C0 ,
satises the relation
Cp sup
e
w1 Il e pLp
[(e)]p1
where e is an arbitrary compactum in Rn .
11.4 Counterexample to the Capacitary Inequality for

the Norm in L22 ()
Let be a domain in Rn and let denote a subdomain of with compact
closure in . In this section we use the norm
11.4 Counterexample to the Capacitary Inequality
559
uL1p () = uLp () + l uLp () .

Let F be a relatively closed subset of . By the capacity of F generated by
the space Llp () we mean the set function

cap F ; Llp () = inf upLl () : u C (), u|F 1 .
p
Let be a Borel measure on , p (1, ), and suppose that the trace

inequality
uLp (,) CuLlp () , C = const,
holds for all u C (). Then the isocapacitary inequality

(F ) D cap F ; Llp () , D = const,
(11.4.1)
easily follows from the denition of cap(F ; Llp ()) for any relatively closed
F with D = C p . One can readily show that (11.4.1) is also sucient for
the trace inequality to be valid for all u C () if the capacitary inequality

cap x : u(x) t ; Llp () d tp cupLl () ,
(11.4.2)
p
has been established for all u C (). Indeed, in view of Lemma 1.2.3 we
have

u(x)p d =
(Nt )d tp ,
0
where

Nt = x : |u(x)| t .
By applying the above isocapacitary inequality with respect to F = Nt and

using the (11.4.2), one obtains the trace inequality with C p = cD.
We now remark that the isocapacitary inequality really holds true for
l = 1, 2, p (1, ) and nice domains , say, C 0,1 . In this case it is
a consequence of Steins extension theorem and the validity of the capacitary
inequality for the norm in Wpl (Rn ), (see Sect. 10.2).
First, we note that if p [1, ), the capacitary inequality (11.4.2) holds
for the norm in L1p () without restrictions on .
Theorem. Let be a domain in Rn . If u C () L1p (), 1 p < ,
then

cap Nt ; L1p () d tp c(p)upL1 () .
p
Proof. Since cap(Nt ; L1p ()) is a nonincreasing function of t, the integral

on the left-hand side of the required inequality does not exceed

S = 2p 1
2pj cap N2j ; L1p () .
j=
560
Given any (0, 1), let be a function in C (R1 ) such that 0 1, 0

1 + , = 0 in a neighborhood of (, 0] and = 1 in a neighborhood
of [1, ). Putting

uj (x) = 21j u(x) 1 ,
we observe that uj C (), uj (x) = 1 for x N2j , supp uj N2j1 . Hence

S 2p1 2p 1 (S1 + S2 ),
where
S1 =

2
pj
j=
S2 =
N2j1 \N2j
|uj |p dx,
|uj |p dx.
pj
N2j1
j=
Clearly |uj | (1 + )21j |u|, and

|u|p dx = cupLp (),

S1 c
j=
N2j1 \N2j
with c = (1 + )p 2p .
To bound the sum S2 , we note that |uj | 1 and that the function (0, )
t mn ( Nt ) is nonincreasing. Therefore, the general term of the sum is
not greater than
j1

1 2

2p 1 2p
mn ( Nt ) d tp .
2j2
Thus
p
12
S2

mn ( Nt ) d tp =

u(x)p dx.
Here Lemma 1.2.3 with = mn | has been used at the last step. By letting
tend to zero, we arrive at the required capacitary inequality with c = 23p1 .

In this section we show that the capacitary inequality for the norm in
Llp () fails for l > 1 unless some restrictions are imposed on . We describe
a bounded domain R2 and a Borel measure on such that there is no
constant C > 0 for which the trace inequality
uL2 (,) CuL22 ()
holds for all u C () in spite of the fact that the inequality

(F ) const cap F ; L22 ()
(11.4.3)
is true for all sets F closed in . According to what was said previously,
the capacitary inequality in L22 () fails for the same domain.
561
Before we proceed to the construction of , we prove two auxiliary assertions. In this section c designates various absolute positive constants.
Lemma 1. Let

T = (x, y) R2 : |x| < y <
and let u W22 (T ), where 0 < 2 < 1. Then

uL2 (T ) c| log |1/2 2 uL2 (T ) + 2 uL2 (T ) .
Proof. It will suce to consider the case = 1. Here the required estimate
is a consequence of the inequality
vL2 (T ) c| log |1/2 vL2 (T1 ) ,
where v W21 (T1 ), v(x, y) = 0 for r = (x2 + y 2 )1/2 1. Passing to the polar
coordinates (x, y) = (r, ), we observe that

v(r, ) =
2

v (, ) d | log r|
v (, )2 d,
if r (0, 1), (/4, 3/4). Thus,

v2L2 (T )
3/4
d
/4
v (, )2 d,
r| log r| dr
0

Lemma 2. Let 0 < 2 < 1 and suppose that u W22 (T ) satises
2 uL2 (T ) 1,
Then
uL (T ) 1,
uL2 (T ) | log |1/2 .

u(x, y) c 1 + y| log |1/2
for all (x, y) T .

Proof. There exists a linear function (x, y) = ax + by + d such that the
generalized Poincare inequality

2 u L2 (T ) + 1 (u )L2 (T ) c2 uL2 (T )
is valid. Therefore, by the Sobolev embedding and Lemma 1

(u )
c| log |1/2 .
u L (T ) c,
L2 (T )
By the assumptions of the lemma, we obtain
(11.4.4)
562
Fig. 35.
Fig. 36.
L (T ) c,
L2 (T ) c| log |1/2 .
In other words,
|a| + |b| c| log |1/2 ,
|d| c.
By using (11.4.4), we arrive at the desired estimate thus concluding the proof.

We now turn to the counterexample. Let {i }i1 be a sequence of open
nonoverlapping intervals of length 21i lying in interval (0,4) of the real axis.
For deniteness, i are supposed to concentrate near the origin, i.e., for any
small > 0, interval (, 4) contains only a nite number of i . By {i }i1
we denote the sequence of open right isosceles triangles situated over the axis
Ox with hypotenuses i .
Let i be the triangle symmetric to i with respect to the line y = 2i i ,
i (0, 2i ) being specied in the following. Furthermore, let R = [1, 4]
[1, 0]. We dene to be the interior of the union of R and all the triangles
i and i (see Figs. 35 and 36).
To introduce the measure , we use the points

Aij = i , 2i i + 2j hi ,
563
j = 1, 2, . . . , 2i ; i = 1, 2, . . . ,
where i is the abscissa of the middle point of i ,

hi = | log2 i |1/2 ,
log2 | log2 i | = 2i+2 .
Let F be an arbitrary Borel subset of and F its characteristic function.

The required measure is dened by
2
(F ) =
22ji/2 F (Aij ).
i1 j=1
Let us verify (11.4.3). To be more specic, we introduce the norm in L22 ()

by
uL22 () = 2 uL2 () + uL2 ((1,4)(1,0)) .
We deduce (11.4.3) by using the estimate

cap {Aij }; L22 () c22j ,
j = 1, 2, . . . , 2i ,
(11.4.5)
which will be checked later. Let F be a relatively closed subset of and let
m(F ) be the minimum value of j such that there exists a point Aij in the
set F . Put m(F ) = if there are no points Aij in F . Then (11.4.5) and the
denition of imply

cap F ; L22 () c22m ,
(F ) c22m .
Hence satises (11.4.3).
Inequality (11.4.5) is a consequence of the estimate

u(Aij ) c2j ,
where u is an arbitrary function in L22 () normalized by uL22 () = 1. Let G

be the interior of the union R i=1 i . Clearly, G and G possesses the
cone property. By Lemma 1 and the Sobolev embedding L22 (G) L (G), we
have
uL (ti ) c,
uL2 (ti ) ci | log i |1/2 ,
where ti = i i . Next, Lemma 2 applied to the triangles i and i {(x, y) :
y < 2i i } yields

u(x, y) c 1 + y 2i + 2i | log i |1/2 ,
for all (x, y) i . Since the ordinate of the point Aij is 2i i + 2j hi ,
inequality |u(Aij )| c2j follows.
We now dene a function f such that
f L22 () C(),
f L2 (, ).
564
Put f (x, y) = 0 for y 0 and

f (x, y) = 2i/4 hi 2i y y + i 2i log2 2i y ,
on the set

i = (x, y) i : y 2i i ,
where C (0, ), (t) = 0 for t 1/4, (t) = 1 for t 1/2. The linear
extension of this function f to the set {(x, y) i , y > 2i i } will be also
denoted by f , i.e.,

f (x, y) = 2i/4 h1
y 2i + i
i
for (x, y) i , y > 2i i . Clearly supp 2 f is placed in the set i1 i .
Furthermore, the estimate

1
|2 f | c2i/4 hi max 2i i, 2i y
holds for (x, y) i . Therefore
2 f 2L2 () =
2 f 2L2 (i )
i1

2i/2 h2i i2 + | log2 i | c
2i/2 <
i1
i1
and f L22 (). At the same time

2
f 2L2 (,)
2i/22j f (Aij )2 =
i1 j=1
1 = .
i1
The result follows.
11.5 Ball and Pointwise Criteria

It was shown by Kerman and Sawyer [420] (see also Sawyer [691]) that it is
enough to assume in (11.3.3) with Spl = hlp that E = B, where B = B(x, r) is
a ball (or cube) in Rn .
Theorem 1. Let p > 1 and 0 < l < n. The trace inequality (11.3.3) holds
if and only if

p

(11.5.1)
Il (B d) dx C(B)
B
for every ball B.

The following criterion for (11.3.3) with Spl = hlp formulated in purely
pointwise form was obtained by Mazya and Verbitsky [591].
565
Theorem 2. Let 1 < p < and let 0 < l < n. Let be a nonnegative
Borel measure on Rn . Then the following statements are equivalent.
(i) The trace inequality (11.3.3) or, equivalently
Il f Lp (d) Cf Lp
(11.5.2)
holds for all f Lp (Rn ).

(ii) Il < a.e. and

Il (Il )p (x) CIl (x)
a.e.
(11.5.3)
(iii) The trace inequality (11.5.2) holds with replaced by the absolutely

continuous measure d = (Il )p dx, or, equivalently,

(11.5.4)
(E) =
(Il )p dx C cap E; hlp ,
E
where C is a constant which is independent of a compact set E.

Both Theorems 1 and 2 will be proved in the present section.
We shall need the following integration by parts lemma which is an
analog of the elementary inequality

p
p1
ak
p
ak
aj
,
k=1
k=1
j=k
where 1 p < and {ak } is a sequence of nonnegative numbers. A similar statement for more general integral operators is proved in Verbitsky and
Wheeden [776]. (See also Kalton and Verbitsky [412] where a discrete analog
with sharp constants is used.)
Lemma. Let 0 < l < n and 1 p < . Let be a positive Borel measure
on Rn and let f L1loc (d), f 0. Suppose that Il (f d)(x) < . Then

p
p1
Il (f d)(x) CIl f Il (f d)
d (x),
(11.5.5)
where C is a constant that depends only on , p, and n.
Proof. For convenience we ignore the normalization constant cn in the
denition of Il and denote by C constants that depend only on l, n, and p.
For a xed x Rn such that A = Il (f d)(x) < , we set

p1
B = Il f I (f d)
d (x).
Then (11.5.5) may be rewritten as Ap CB.
We rst consider the more dicult case 1 2|xz|}
f (z)
d(z)
|x z|nl
p1
d(y).
To estimate I, notice that in the inside integral we have

|x z|ln C|y z|ln .
Hence

IC
Rn
f (y)
|x y|nl

Rn
f (z)
d(z)
|y z|nl
p1
d(y) = CB.
We estimate II using Holders inequality with exponents 1/(p 1) and

1/(2 p), which gives II A2p III p1 , where by Fubinis theorem,

f (y)
f (z)
III =
d(z)
d(y)
nl
nl
Rn |x y|
{z:|yz|>2|xz|} |x z|

f (z)
f (y)
d(y) d(z).
=
nl
nl
Rn |x z|
{y:|yz|>2|xz|} |x y|
We now estimate the function in square brackets on the right-hand side, which
we denote by

f (y)
G(z) =
d(y).
|x
y|nl
{y:|yz|>2|xz|}
Obviously, G(z) A. On the other hand, G(z) CIl (f d)(z). Indeed, by
the triangle inequality,

y : |y z| > 2|x z| y : |y z| 2|x y|
and hence
G(z)

{y:|yz|2|xy|}
f (y)
d(y) CIl (f d)(z).
|x y|nl
Combining the preceding inequalities, we get

p1
G(z) CA2p Il (f d)(z)
.
From this it follows that III CA2p B, and thus
II A2p III p1 CAp(2p) B p1 .
567
Combining the estimates for I and II , we obtain

Ap CB + CAp(2p) B p1 ,
which yields Ap CB for 1 2 follows from the case
p = 2 considered previously. By Holders inequality with exponents p 1 and
(p 1)/(p 2) we have

2

A2 = Il (f d)(x) CIl f Il (f d) d
1
p1 p1
p1

p2
1
d
C Il f Il (f d)
Il (f d) p2 = CB p1 A p1 ,
which clearly implies Ap CB. The proof of the Lemma is complete.
Remark. It is easy to see that the above Lemma is true without the restriction Il (f d)(x) < , i.e., the right-hand side of (11.5.5) is innite if the
same is true for the left-hand side (see Verbitsky and Wheeden [776]).
We now show that the pointwise condition

Il (Il )p (x) cIl (x) < a.e.
(11.5.6)
implies both of the trace inequalities

Il f Lp (d) Cf Lp ,

f Lp Rn
(11.5.7)
Il f Lp (d) Cf Lp ,

f Lp Rn ,
(11.5.8)
and
where is dened by
d = (Il )p dx.
(11.5.9)
Proposition 1. Let 1 < p < . Let be a positive Borel measure on Rn

and let be dened by (11.5.9). Suppose that the pointwise condition (11.5.6)
holds. Then both (11.5.7) and (11.5.8) hold.
Proof. Without loss of generality we may assume that f 0, and that f
is uniformly bounded and compactly supported. By the above Lemma with
d = dx and Fubinis theorem,

Il f (Il f )p1 d = C
f (Il f )p1 (Il ) dx.
Il f pLp (d) C
Rn
Rn
Here C is the constant in (2.1) that depends only on p, l, and n. From this,
by Holders inequality, we get
p1
Il f pLp (d) Cf Lp Il f L
.
p (d)
The preceding inequality shows that (11.5.8)(11.5.7) for any .
(11.5.10)
568
Repeating the above argument with in place of , we obtain

p1
,
p (d1 )
(11.5.11)
where by (11.5.6)

p

d1 = (Il )p dx = Il (Il )p
dx cp d.
Here c is the constant in (11.5.5). Hence by (11.5.11) and the preceding estimate
p1
.
p (d)
(11.5.12)
Assuming that Il f Lp (d) < , we get

Il f Lp (d) Cf Lp ,
which proves (11.5.8). Now (11.5.10) combined with the preceding estimate
yields (11.5.7).
It remains to check that Il f Lp (d) < , which follows easily from
(11.5.6) and the assumption that f is bounded and compactly supported.
Indeed, by (11.5.6) Il < a.e., which implies that Il is locally integrable.
Since I f is obviously bounded, it follows that

(Il f )p (Il )p dx C
Il f (Il )p dx
Il f pLp (d) =
n
n
R
R

p
=C
f Il (Il ) dx C
f Il dx < .

Rn
Rn
Proposition 2. Let 1 < p < . Let be a positive Borel measure on Rn .
Then the pointwise condition (11.5.6) is equivalent to the KermanSawyer
condition

p
Il (B d) dx C(B)
(11.5.13)
B
for every ball B = B(x, r).

Proof. By Proposition 1 it follows that the pointwise condition (11.5.6)
implies the trace inequality (11.5.7). By duality, (11.5.7) is equivalent to the
inequality

Il (g d)
CgLp (d) , g Lp (d).
(11.5.14)
L
p
Letting g = B , B = B(x, r), we see that (11.5.13) holds.

To prove the converse, we shall show that (11.5.13) implies

(B(x, t)) dt
p
n
(Il ) dx Cr
B(x, r) =
tn
t
r
B(x,r)
(11.5.15)
569
for every ball B(x, r), which readily gives (11.5.6).

Note that, clearly,
Il (B(x,r) d) C
(B(x, r))
B(x,r) ,
rnl
(11.5.16)
and hence it follows from (11.5.13) that

B(x, r) Crnlp .
(11.5.17)
In particular, the right-hand side of (11.5.15) is nite and hence Il < a.e.
The proof of (11.5.15) is based on the decomposition d = d1 + d2 ,
where
d1 (x) = B(x,2r) d(x)
For y B(x, r),
d2 (x) = (1 B(x,2r) ) d(x).
and
Il 2 (y) CIl 2 (x),
where C depends only on l and n. Hence by (11.5.13) and the preceding

estimate,

p
p
B(x, r) =
(Il ) dy C
(Il 1 ) dy + C
(Il 2 )p dy
B(x,r)
B(x,r)

p
C B(x, 2r) + Crn Il 2 (x) .
B(x,r)
Clearly,

Il 2 (x)
p

p
(B(x, t)) dt
=C
tnl
t
2r
p 1

(B(x,
t))
(B(x, s)) ds
dt

= Cp
.
nl
nl
t
s
s
t
2r
t
Now by (11.5.16) it follows that (B(x, s)) Csnlp , and hence

p 1

(B(x, s)) ds
Ctl .
snl
s
t
From this we obtain

Il 2 (x)
p

C
2r
Combining these estimates, we get
(B(x, t)) dt
.
tn
t

(B(x, t)) dt
B(x, r) C B(x, 2r) + Cr n
tn
t
2r

(B(x, t)) dt
n
.
Cr
tn
t
r
This proves (11.5.15).
570
To obtain (11.5.6), we notice that by (11.5.15) and Fubinis theorem,

(B(x, r)) dr
Il (x) = C
r nl
r
0
t

(B(x,
t)) dt dr
(B(x, t))
l
l dr dt
=C
r
r
C
tn
t r
tn
r
t
0
r
0
0

(B(x, t)) dt
=C
= CIl (x).
tnl
t
0
Hence

Il (x) = Il (Il )p (x) CIl (x) <
a.e.
We now are in a position to complete the proof of Theorems 11.5/1 and

11.5/2.
By Proposition 1 it follows that (11.5.6)(11.5.7) and (11.5.8). As was
observed in the proof of Proposition 11.5/1, it follows from Lemma 11.5 that
(11.5.8)(11.5.7). Since (11.5.7)(11.5.13) and (11.5.13)(11.5.6) by Proposition 1, we have shown that the following conditions are equivalent:
(11.5.7) (11.5.8) (11.5.6) (11.5.13).
The proofs of Theorems 1 and 2 are complete.
11.6 Conditions for Embedding into Lq () for p > q > 0

In the present section we discuss the necessary and sucient conditions for
the validity of (11.3.2) for p > q > 0, p > 1.
11.6.1 Criterion in Terms of the Capacity Minimizing Function
Theorem 1. Let p (1, ), 0 < q < p, l > 0. The inequality
uLq (Rn ,) Cp,q uSpl ,
(11.6.1)
where Spl is the same as in Theorem 11.3, holds for all u C0 (Rn ) if and
only if
q
(Rn ) p/q pq
dt
t
< ,
Dp,q =
(t)
t
0
where

(t) = inf cap F, Spl : F is a compactum in Rn , (F ) t ,

t 0, Rn .
571
Proof. Suciency. We prove the inequality

q/(pq) (pq)/pq

(Rn )
p
t
1/p
dt
uSpl .
uLq () (4pC)
pq 0
(t)
(11.6.2)
Let S be any sequence of open sets {gj }+
such
that
g
g
.
j+1
j We put
j=
l
j = (gj ), j = cap(gj , Sp ) and

Dp,q
pq/(pq) (pq)/pq
+
(j j+1 )1/q
= sup
.
1/p
{S} j=
We show that the best constant in (11.6.1) satises

Cp,q (4pC)1/p Dp,q ,
(11.6.3)
where C is the constant in (11.1.4).

Let f C0 , f 0, u = Kf , gj = {x : (Kf )(x) > j }, where > 1.
Obviously,

q(j+1) j j+1 q/p
q(j+1) (gj ) (gj+1 ) =
j .
uqLq ()
q/p
j
j
j
By H
olders inequality the last sum does not exceed

q/p
q
p(j+1)
Dp,q
j
.
j
Next we note that
p(j+1)
p2p
j p
1
j
j1

cap Nt , Spl tp1 dt.
Putting p = 2 in this inequality, we obtain

p(j+1) j 4p
cap Nt , Spl tp1 dt.
j
Consequently,
uLq () (4pC)1/p Dp,q uSpl ,
(11.6.4)

The inequality (11.6.2) results immediately from (11.6.4) and the estimates
p/(pq)
(j j+1 )p/q q/(pq)

jp/(pq) j+1
q/(pq)
j
j
j
j
(Rn )
d(tp/(pq) )
.
((t))q/(pq)
0
572
The proof of suciency is complete.

Necessity. First we remark that (11.6.1) implies (t) > 0 for t > 0. Let j
be any integer satisfying 2j < (Rn ). Then there exists a compact set Fj Rn
such that

cap Fj , Spl 2 2j .
(Fj ) 2j ,
Let j be a function for which
Fj Kl j ,

j pLp (Rn ) 4 2j ,
j 0,
where Kl is either the Riesz or the Bessel kernel.

Suppose that s is the integer for which 2s < (Rn ) 2s+1 if (Rn ) is
nite and s is an arbitrary integer otherwise. Let r < s be another integer
and let

1/(pq)
r,s = max j j , j = 2j / 2j
.
rjs
Furthermore, we put
fr,s = Kl r,s .
Clearly,
fr,s pS l = r,s pLp
jp j pLp 4
j=r

jp 2j .
j=r
To obtain a lower bound for the norm of fr,s in Lq (Rn , ), we introduce

the nonincreasing rearrangement (fr,s ) of fr,s

(fr,s ) (t) = inf > 0 : |fr,s | > t , t 0, Rn .
Since fr,s |Fj j and (Fj ) 2j , the inequality

|fr,s | > < 2j
implies j . Hence
fr,s
(t) j

for t 0, 2j , r j s,
which gives

fr,s qLq (Rn ,) =
(Rn )
fr,s
(t)
q
dt
s
j=r
2j
2j1
fr,s
(t)
q
dt
jq 2j1 .
j=r
Next, we note that if (11.6.1) is valid for all u C0 (Rn ), then (11.6.1) is
valid for all u Spl . In particular, for u = fr,s one obtains
Cp,q
1/q1/p
s
s
2pj/(pq)
( j=r jq 2j )1/q
c s
=c
.
( j=r jp (2j ))1/p
((2j ))q/(pq)
j=r
573
Letting r and using the monotonicity of , we have

s
1/q1/p
q
q
pq
pq
2s
1/q1/p
2j
t
j
2
c
dt
Cp,q c
(2j )
(t)
0
j=
with c independent of s. Letting s , we arrive at the estimate Cp,q
1/q1/p
for (Rn ) = . In the case of nite (Rn ) we have
cDp,q
(Rn )

q/(pq)
t/(t)
dt
2s
2s

q/(pq)
2t/(2t)
dt c
2s1
2s
q/(pq)
t/(t)
dt,
2s1
1/q1/p
and again the inequality Cp,q cDp,q
is true. The proof is complete.
The proof of suciency implies the following assertion.

Corollary. Let the assumptions of the above theorem be fullled. If A is a
Borel subset of Rn , then
q

1/q
(A) p/q pq
(pq)/pq
dt
t
q
|u| d
c
uSpl
(t)
t
0
A
for all u C0 (Rn ) with c depending only on p, q.
Proof. Let A be the restriction of to the set A. Then it is readily seen
that (t) A (t). The result follows by applying the suciency statement
of the theorem with respect to the measure A on Rn .

The following criterion of the inequality
uLq () cuhlp ,
(11.6.5)
with q < p, whose statement involves no capacity, was established by Cascante, Ortega, and Verbitsky [175] for q > 1 and by Verbitsky [775] for q > 0
(see [176]).
This criterion is formulated in terms of the nonlinear potential
p 1

dr
(B(x, r))
,
(Wl,p )(x) =
nlp
r
r
0
which appeared in Mazya and Havin [567] and was used by Wol in [372] in
his proof of inequality (10.4.8).
Theorem 2. Let 0 < q n the necessary and sucient condition for the validity of (11.3.2),
where Spl = Hpl , Wpl or Bpl , can be written in the following simple way.
Theorem 1. If pl > n, p > q and Cp,q is the best constant in (11.3.2),
then

(i) p/(pq) (pq)/pq
Q
,
(11.6.6)
Cp,q
i0
where {Q } is the sequence of closed cubes with edge length 1 forming the
coordinate grid in Rn .
(i)
Proof. Let O(i) be the center of Q(i) , O (0) = O, and let 2Q(i) be the
concentric cube of Q(i) with edges parallel to those of Q(0) and having edge
length 2. We put i (x) = (x O(i) ), where C0 (2Q(0) ), = 1 on Q(0) .
We have

uq (i) Q(i)
uqLq ()
C(Q
i0

(i) p/(pq) 1q/p
Q
up
C(Q(i) )
i0
q/p
. (11.6.7)
i0
Next we note that for pl > n

uC(Q(i) ) ui Spl ,
(11.6.8)
where Spl = hlp , wpl or blp (cf. (10.1.14) and items 6 and 9 of Theorem 10.1.3/4).
Now (10.1.14), (11.6.7), and (11.6.8) imply the upper bound for Cp,q .
To obtain the lower bound for Cp,q it suces to insert the function
uN (x) =

Q(i) i (x),
N = 1, 2, . . . ,
i=0
into (11.3.2). This concludes the proof.
The constant Cp,q can be easily calculated for q = 1.

Theorem 2. Let Spl be either Hpl or hlp . Then
Cp,1 = KLp ,
where K is either the Riesz or the Bessel potential.
Proof. Let |u| Kf , f 0 and f Lp = uSpl . We have

|u| d f K dx f Lp KLp ,
which gives Cp,1 KLp . The reverse inequality follows by the substitution
of u = K(K)1/(p1) into (11.3.2) with q = 1.
11.7 Cartan-Type Theorem and Estimates for Capacities
575

In this section we establish the equivalence of inequalities of the type (11.1.6)
and (11.1.7). This follows from a theorem giving an estimate for the set where
the functions Wp,l and Sp,l , introduced in Sect. 10.4.2, majorize a given
value. Such estimates were rst obtained for harmonic functions by Cartan
[172] (cf. also Nevanlinna [635]). For linear Riesz potentials they are given in
Landkof [477].
The same scheme is used here for the nonlinear case.
Lemma. Let 1 r0 . Further let D be the set {x Rn : (P )(x) > Y []}, where
P = Wp,l for pl > n, P = Sp,l for pl = n, and
(r)
( nlp )p1 dr
for 1 < p < n/l,
r
Y [] = 0 r

((r))p 1 ebr r1 dr for p = n/l.
0
Then D can be the covered by a sequence of balls of radii rk r0 such that

(11.7.1)
(rk ) c Rn .
k
Proof. First, consider the case 1 0. Then

(Wp,l )(x) =
0
(B(x, r))
r nlp
p 1
dr

0
(r)
r nlp
p 1
dr
.
r
The latter means that x

/ D. This contradiction shows that given any x D
there exists an r = r(x) (0, r0 ) such that (r) < (B(x, r)) (Rn ).
Applying Theorem 1.2.1/1 we select a covering {B(xk , rk )}, k = 1, 2, . . . , of
D with nite multiplicity c = c(n) in the union of balls {B(xk , r(x))}, x D.
It is clear that

(rk ) <
B(xk , rk ) c Rn ,
k
and the result follows for 1 0, let

+
where
(t)t1 dt c (u),
(11.7.2)
576

(v) =
(v(v 1 ))p 1

v((v 1p ))p 1
for 1 Rn
Then the set G = {x Rn : (P )(x) > m} can be covered by a sequence of
balls {B(xk , rk )} with

cap Brk , Spl < c cm1p Rn .
(11.7.3)
k
Here Spl = hlp for lp < n and Spl = Hpl for lp = n.

Proof. Let = cap(B1 , hlp ) for n > lp. For n = lp we dene as

= min t : cap Br , Hpl t| log r|1p , r < e1 .
Further let Q = (Rn ). In the Lemma we put (r) = Q for r > r0 and

Q(rnlp )/(r0nlp )
if pl < n, r r0 ,
(r) =
Q(| log r|1p )/(| log r0 |1p ) if pl = n, r r0 .
Here and henceforth r0 is a number that will be specied later to satisfy the
inequality m > Y [] (the number Y [] was dened in the Lemma).
1. Let 1 < p < n/l. We have

Y [] =
0
r0
(r)
r nlp
p 1

dr
p 1 (nlp)/(1p)
+ Qp 1
r
.
r
n lp 0
We show that the integral on the right does not exceed

cQp 1 r0
(nlp)/(1p)
This is equivalent to the inequality

nlp 1p
r0
r0

0
(rnlp )
r nlp
p 1
dr
(nlp)/(1p)
cr0
.
r
Putting r pln = t, r0pln = t0 , we rewrite the latter as

1 p 1 1
p 1
t t
t dt c t0 t1
,
0
t0
which is fullled by virtue of (11.7.2). Thus

Y [] < cQp 1 r0
(nlp)/(1p)
577
and the inequality Y [] < m is satised provided we put

p1

r0nlp = cm1
Q.
We introduce the set D = {x Rn : (P )(x) > Y []}, which is open by the
lower semicontinuity of P . Since m > cY [] then G D. Let {B(xk , rk )} be
the sequence of balls constructed in the Lemma for the set D by the function
specied here. Inequality (11.7.1) can be rewritten as
nlp

rk
c cm1p Q .
k
Thus we obtain the covering of G by balls {B(xk , rk )} satisfying (11.7.3).

2. Let p = n/l and let r0 < 1/e. We have

r0

p 1 cr 1

Y [] =
(r)
e r dr + Qp 1
ecr r1 dr.
(11.7.4)
0
r0
The second integral is majorized by

1

ebr r 1 dr <
r 1 dr +
r0
r0

ecr dr 1 + c1 ec | log r0 |.
1
We show that the rst integral on the right in (11.7.4) does not exceed cQp 1
| log r0 |. In other words, we prove that

1p r0
p 1 dr

| log r|1p
| log r0 |1p
< c| log r0 |.
r
0
Putting | log r| = t, | log r0 | = t0 , we rewrite the preceding inequality as

1p p 1

p 1
t
dt ct0 t01p
,
t0
which holds by (11.7.2). Therefore there exists a constant c (1, ) such that
Y [] < cQp 1 | log r0 |.

Thus the inequality Y [] < m is satised provided we set
p1

| log r0 |1p = cm1
Q.
The completion of the proof follows the same line of reasoning as for p
(1, n/l).

Remark 1. The proof of the theorem shows that in the case pl = n we can
take the radii of the balls, covering G, to be less than 1/e.
578
Corollary 1. Let 1 0
where Spl = hlp for pl < n and Spl = Hpl for pl = n. Then there exists a
covering of K by balls B(xk , rk ) such that

(11.7.5)
cap Brk , Spl < c c cap K, Spl ,
k
where c is a constant that depends on n, p, l, and on the function . In the

case pl = n we may assume that rk e1 .
Proof. We limit consideration to the case pl < n. For pl = n the argument
is the same. We need nonlinear potentials Vp,l dened by (4.4.4) in the
.) These
book by D.R. Adams and Hedberg [15]. (They used the notation Vl,p
potentials are comparable with Wp,l dened in Sect. 10.4 (see (4.5.3) in [15]).
We put

C(K) = inf
Vp,l d : Vp,l 1 (p, l)-quasi everywhere on K .
By (10.4.8) the capacities C(K) and cap(K, hlp ) are equivalent. In the
paper by Hedberg and Wol [372] it was shown that the extremal measure
K for the previous variational problem exists and that C(K) = K (K).
We introduce the set G = {x Rn : Vp,l K (x) 1 }, where > 0.
Since Vp,l K (x) 1 for (p, l)-quasi-every x K, then E G E0 where
cap(E0 , hlp ) = 0.
By the Theorem there exists a covering of G by balls B(xj , rj ) for
which (11.7.3) is valid with m = 1 and (Rn ) = cap(K, hlp ). Since (t)/t is
integrable on [1, +), the function (r) = (cap(Br , hlp )) satises (10.4.16).
This and Proposition 10.4.3/2 imply that the set E0 has zero Hausdor measure. Therefore, E0 can be covered by balls B(yi , i ) so that

cap B i , hlp < .
i
The balls B(xj , rj ) and B(yi , i ) form the required covering.
Corollary 2. Let p (1, n/l] and let Spl = hlp for lp < n, Spl = Hpl for
lp = n. Further, let be the function dened just before the Theorem. If a
measure is such that

B(x, ) c cap B , Spl ,
(11.7.6)
then, for any Borel set E with the nite capacity cap(E, Spl ), the inequality

(E) c c cap E, Spl ,
(11.7.7)
where c is a constant that depends on n, p, l, and is valid.
11.8 Embedding Theorems for the Space Spl
579
Proof. It suces to derive (11.7.7) for a compactum E. According to Corollary 1 there exists a covering of E by balls B(xk , rk ) satisfying (11.7.5). Using
the additivity of as well as estimate (11.7.6), we obtain

(E)
B(xk , rk )
B(xk , rk )
k

c cap Brk , Spl < c c cap E, Spl .
The result follows.
Remark 2. According to (10.4.2) we have the equivalence cap(E, Hpl )

cap(E, hlp ) if diam E 1. Therefore under the additional requirement
diam E 1 we may also put Spl = Hpl in Corollary 2 for pl < n.
To prove this assertion we need to verify that the measure Rn A
1 (A) = (A E) satises (11.7.6).
Let diam E 1 and, for all r (0, 1), let

(11.7.8)
B(x, r) cap Br , Hpl .
For r < 1 we have

1 B(x, r) = B(x, r) E B(x, r)

cap Br , Hpl c cap Br , hlp .
In the case r 1

1 B(x, r) B(y, 1)
for any y E. Hence, using (11.7.8) and the monotonicity of the capacity, we
obtain

1 B(x, r) cap B1 , Hpl c cap Br , hlp .
Thus the measure 1 satises (11.7.6).
11.8 Embedding Theorems for the Space Spl (Conditions

in Terms of Balls, p > 1)
Theorem. Let M be a convex function and let N be the complementary function of M . Further let be the inverse function of t tN 1 (1/t) subject to
condition (11.7.2). Then:
() The best constant A in (11.3.1) with Spl = hlp , lp < n, is equivalent to

C1 = sup lpn B(x, ) N 1 1/ B(x, ) : x Rn , > 0 .
() The best constant A in (11.3.1) with Spl = Hpl is equivalent to
580

C2 = sup lpn B(x, ) N 1 1/ B(x, ) : x Rn , 0 < < 1
if pl < n and to

1
C3 = sup | log |p1 B(x, ) N 1 1/ B(x, ) : x Rn , 0 < <
2
if pl = n.
The proof immediately follows from Theorem 11.3 and the equivalence
B Cj , j = 1, 2, 3, obtained in Corollary 11.7/2 and Remark 11.7/2.
Remark 1. We can easily see that in the case pl > n the constant A
in (11.3.1) with Spl = Hpl is equivalent to

C4 = sup B(x, 1) N 1 1/ B(x, 1) : x Rn .
Indeed, let { (j) } be a partition of unity subordinate to a covering of Rn
by unit balls {B(j) } with nite multiplicity. From the denition of the norm
in LM () and the Sobolev theorem on embedding Hpl into L we obtain
p
|u|
LM ()

u (j) p
LM ()

, B (j)
c1
L
j

(j) p

u l c1 C4
u (j) p l .
()
H
H
p
The last sum does not exceed cupH l (cf. Theorem 10.1.3/3). Hence A c2 C4 .
p
The opposite estimate follows from (11.3.1) by the substitution of the function
C0 (B(x, 2)), = 1 on B(x, 1).
Now the D.R. Adams Theorem 1.4.1 follows from () of the previous
theorem where M (t) = tq/p , q > p.
Remark 2. We show that the condition (11.1.5) with s = n pl is not
sucient for (11.1.2) to hold in the case q = p. Let q = p, n > pl. We choose
a Borel set E with nite positive (n pl)-dimensional Hausdor measure.
We can take E to be closed and bounded (since any Borel set of positive
Hausdor measure contains a bounded subset having the same property). By
the Frostman theorem (see Carleson [168], Theorem 1, Ch. 2) there exists a
measure = 0 with support in E such that

(11.8.1)
B(x, ) cnpl ,
where c is a constant that is independent of x and . By Proposition 10.4.3/3,
cap(E, Hpl ) = 0. On the other hand, from (11.1.2) it follows that (E)
A cap(E, Hpl ) and hence (E) = 0. This contradiction shows that (11.1.2)
fails although (11.8.1) holds.
11.8 Embedding Theorems for the Space Spl
581
Setting M (t) = tq/p in the previous theorem, we obtain the following result
for the case lp = n.
Corollary 1. If lp = n, q > p > 1 then the exact constant A in
uLq () AuHpl
(11.8.2)
is equivalent to

p/q
1
.
: x Rn , 0 < <
C5 = sup | log |p1 B(x, )
2
From the theorem of the present section we easily obtain the following

assertion relating the case lp = n, M (t) = exp(tp 1 ) 1 and measure of
power type.
p
|u|
L

exp(tp 1 1)
AupS l
holds if and only if for 0 < 1

B(x, ) c ,
with a certain > 0.
Proof. Since N (t) = (log t)p1 (1 + o(1)) as t we have 1 (t) =

tN (1/t) = (log t)1p (1 + o(1)). Hence, log (t) = tp 1 (1 + o(1)). Obviously, satises the condition (11.7.2). Now it remains to use cap(B , Hpl )
| log |1p with (0, 12 ) and to apply the Theorem. The proof is complete.
Remark 3. Since Bpl (Rn ) is the space of traces on Rn of functions in

l+1/p
Hp
(Rn+1 ), the Theorem and Corollary 1 still hold if the space Hpl (Rn ) is
replaced by Bpl (Rn ).
Remark 4. We can obtain assertions similar to the Theorem and Corollaries 1 and 2 by replacing u by k u in the left-hand sides of inequalities (11.3.1)
and (11.8.2). For example, the generalization of Corollary 1 runs as follows.
If (l k)p = n, q > p > 1 then the best constant in
k uLq () AuHpl
(11.8.3)
is equivalent to C5 .
The estimate A cC5 needs no additional arguments. To prove the reverse
inequality we place the origin at an arbitrary point of the space and put

log |x|
k
,
u(x) = x1
log
582
where (0, 12 ) and C (R1 ), (t) = 1 for t > 1, (t) = 0 for t <
1
into (11.8.3). It is clear that supp u B 1/2 Using standard but rather
2
cumbersome estimates we can show that, for |x| < 12 ,

1
(Dl u)(x) c|x|n/p log |x| ,
where Dl u is the Strichartz function (10.1.10) for {l} > 0 and Dl u = |l u|
for {l} = 0. Hence from Theorem 10.1.2/4 we have
upH l c| log |1p .
p
On the other hand,

k uqLq () k!(B ).
Consequently, A cC5 .
11.9 Applications
11.9.1 Compactness Criteria
The theorems proved in the present chapter imply necessary and sucient
conditions for compactness of embedding operators of the spaces Hpl , hlp ,
Wpl , wpl , Bpl , and blp into Lq (). The proof of these results follows by standard arguments (compare with Theorems 2.4.2/1 and 2.4.2/2), so we restrict
ourselves to the next four statements. The rst two theorems are based on
Corollary 11.3.
Theorem 1. Let p > 1, pl < n and let slp be any one of the spaces hlp , wpl ,
Any set of functions in C0 , bounded in slp , is relatively compact in Lq ()
if and only if

(e)
n
: e R , diam e = 0,
lim sup
(11.9.1)
0
cap(e, slp )

(e)
n
\B
= 0,
(11.9.2)
:
e
R
lim sup

cap(e, slp )
blp .
where B = {x : |x| < }.

Theorem 2. Let p > 1, pl n, and let Spl be any of the spaces Hpl , Wpl ,
and Bpl . A set of functions in C0 , bounded in Spl , is relatively compact in
Lq () if and only if condition (11.9.1) holds and

(e)
n
: e R \B , diam e 1 = 0.
(11.9.3)
lim sup

cap(e, Spl )
11.9 Applications
583
Theorems 3 and 4 below follow from Theorem 11.8 and Corollary 11.8/1,
respectively.
Theorem 3. Let p 1, l > 0, pl < n. Further let 1 q < if p = 1 and
p < q < if p > 1. Then the set {u C0 : uWpl 1} is relatively compact
in Lq () if and only if
(i)
(ii)

1/q
ln/p B(x, )
= 0,
lim
sup
lim

1/q
sup ln/p B(x, )
= 0.
+0 x; (0,)
|x| (0,1)
Theorem 4. Let p > 1, l > 0, pl = n and q > p. Then the set {u C0 :

uWpl 1} is relatively compact in Lq () if and only if
(i)
(ii)
lim
sup
0 x; (0,)

1/q
| log |11/p B(x, )
= 0,

1/q
lim sup | log |11/p B(x, )
= 0.
|x| 2 <1
11.9.2 Equivalence of Continuity and Compactness of the

Embedding Hpl Lq () for p > q
Here we prove the following assertion.
Theorem. Let 1 0. Suppose that is a nite
Borel measure on Rn . Then the compactness of the trace operator Hpl Lq ()
is equivalent to its continuity and to the inequality Dp,q < , where Dp,q is
dened in Theorem 11.6.1/1.
This theorem will be obtained as a consequence of the compactness properties of the convolution maps which are studied in the present subsection.
We consider integral convolution operators of the form f K f , where

K(x y)f (y) dy
(K f )(x) =
Rn
and K is the kernel of convolution. In the following we abbreviate Lp (Rn ) =

p , Lp (Rn ) = Lp . It will be shown that under some conditions on K and
the continuity of the convolution map Lp Lq () implies its compactness.
Lemma 1. Let p (1, ) and K Lp , p = p/(p 1). Suppose that
is a nite Borel measure on Rn having bounded support. Then the operator
Lp f K f Lq () is compact for any q (0, ).
Proof. An application of Holders inequality yields
584
u Kp f p ,
u = K f.
Furthermore

u(x + h) u(x) K( + h) K() f p
p
for any x, h Rn . Hence it follows that the elements of the set {K f :
f p 1} are uniformly bounded and equicontinuous on Rn . Since the support of is bounded, the required result is a consequence of the ArzelaAscoli
compactness theorem.

We continue the analysis of compactness properties of convolution operators with the following lemma.
Lemma 2. Let K be a nonnegative function in L1 whose support has
a nite volume. Suppose is a nite Borel measure on Rn with bounded
support. If p (1, ), 0 < q < p, then the continuity of convolution operator
Lp f K f Lq () implies its compactness.
Proof. Given > 0, put K (x) = K(x) if K(x) 1 and K (x) = 0
otherwise. Let K () = K K . According to Lemma 1, the operator f K f
is compact as an operator Lp Lq ().
Suppose the map f K f Lq () is not compact. Then the norm of
the operator
f K f = K () f Lq ()
is bounded below uniformly with respect to (0, ). Thus, there is a
constant c0 > 0 such that for any > 0 there exists a function f () Lp
satisfying
()

()
f 1,
K f ()
c0 .
p
L ()
q
Since the operator K : Lp Lq () is continuous, we can assume without

loss of generality that f () is a nonnegative continuous function on Rn with
compact support.
Put 1 = 1, f1 = f (1 ) . Since lim+0 K () 1 = 0 and

sup K () f1 (x) K () 1 f1 ,
xRn
it follows that K () f1 Lq () 0 as 0. Hence there is a number

2 (0, 1 ) for which

( )

K 1 K (2 ) f1
c0 /2.
Lq ()
Letting f2 = f (2 ) and arguing as before, one can select 3 (0, 2 ) subject
to
( )

K 2 K (3 ) f2
c0 /2.
Lq ()
11.9 Applications
585
Continuing this process, we construct a decreasing sequence {i }i1 of positive

numbers and a sequence {fi }i1 C0 (Rn ) such that fi p 1, fi 0 and
hi Lq () c0 /2,
i = 1, 2, . . . ,
where hi = (K (i ) K (i+1 ) ) fi 0. For m = 1, 2, . . . dene

gm (x) = m1/p max fi (x).
1im
Then

1/p
m

p
fi

1

1/p
gm p m
i=1
and also

m

hi

1/p
K gm Lq () m
i=1
Since
q
hi
.
Lq ()
m

min 1, mq1
hqi ,
i=1
we have
K gm Lq () m
q > 0,
i=1
1/p
m
1/q

q
11/q
min 1, m
hi Lq ()
i=1
and hence

c0
min m1/q1/p , m11/p .
2
However, the expression on the right is unbounded as m , and this
contradicts the continuity of the operator f K f Lq (). Lemma 2 is
proved.

K gm Lq ()
Proof of Theorem. Let Bk be the ball of radius k centered at the origin,

k = 1, 2, . . . . Let k denote the restriction of to Bk . Clearly, for any l > 0
the Bessel kernel can be expressed as the sum of two nonnegative kernels, one
of which has compact support and belongs to L1 while the other is in Lr for
all r > 0. According to the previous lemmas any subset of C0 (Rn ), bounded
in Hpl , is relatively compact in Lq (k ). Furthermore, by Corollary 11.6.1
q/(pq) 1/q1/p
1/q
(Rn \Bk )

t
q
|u| d
c(p, q)
dt
uHpl
(t)
0
Rn \Bk
for all u C0 (Rn ). Thus for the same u we have
uLq () k uHpl + constuLq (k ) ,
k = 1, 2, . . . ,
where k is a sequence of positive numbers converging to zero. By using the

last estimate and a diagonal method, one can select from any sequence in

C0 (Rn ), bounded in Hpl , a subsequence convergent in Lq ().
586
11.9.3 Applications to the Theory of Elliptic Operators

Corollary 11.3 has immediate applications to the spectral theory of elliptic
operators. Let
Sh = h()l p(x), x Rn ,
where p 0 and h is a positive number. The results in Sect. 2.5 concerning the
Schrodinger operator h p(x) have natural analogs for the operators Sh .
Duplicating with minor modications the proofs of Theorems 2.5.32.5.6 and
other assertions in Sect. 2.5, we can derive conditions for the semiboundedness
of the operator Sh as well as conditions for the discreteness, niteness, or inniteness of the negative part of its spectrum. Here we present the statements
of two typical theorems of this kind.
Theorem 1. Let n > 2l and let M denote the constant in the rough
maximum principle for the Riesz potential of order 2l.
1. If

p(x) dx
e
: diam e < (4M)1 ,
lim sup
0
cap(e, hl2 )
then the operator S1 is semibounded.
2. If the operator S1 is semibounded, then

p(x) dx
:
diam
e
1.
lim sup e
0
cap(e, hl2 )
Theorem 2. Let n > 2l. The conditions

p(x) dx
lim sup e
:
diam
e
= 0,
0
cap(e, hl2 )

p(x) dx
n
e
lim sup
: e R \B , diam e 1 = 0,

cap(e, hl2 )
are necessary and sucient for the semiboundedness of the operator Sh and
for the discreteness of its negative spectrum for all h > 0.
11.9.4 Criteria for the RellichKato Inequality
By the basic RellichKato theorem [673], [415], the self-adjointness of + V
in L2 (Rn ) is guaranteed by the inequality
V uL2 (Rn ) auL2 (Rn ) + buL2 (Rn ) ,
(11.9.4)
where a < 1 and u is an arbitrary function in C0 (Rn ).

Let n 3. The Sobolev embedding W22 (Rn ) L (Rn ) and an appropriate
choice of the test function in (11.9.4) show that (11.9.4) holds if and only if
there is a suciently small constant c(n) such that
11.9 Applications

sup
xRn

V (y)2 dy c(n).
587
(11.9.5)
B1 (x)
Here and elsewhere Br (x) = {y Rn : |y x| < r}. This class of potentials is

known as Stummels class Sn , which is dened also for higher dimensions by

2
4n

|x y|
lim sup
V (y) dy = 0 for n 5,
r0
Br (x)

lim sup
r0
log |x y|

V (y)2 dy

=0
for n = 4
Br (x)
(see Stummel [731]). Although the condition V Sn is sucient for (11.9.4)

for every n, it does not seem quite natural for n 4. As a matter of fact,
it excludes the simple potential V (x) = c|x|2 obviously satisfying (11.9.4) if
the factor c is small enough.
If n 5, a characterization of (11.9.4) (modulo best constants) results
directly from a necessary and sucient condition for the inequality

V uL2 (Rn ) CuL2 (Rn ) , u C0 Rn ,
contained in Theorem 11.9.3/1: The Rellich-Kato condition (11.9.4) holds in
the case n 4 if and only if there is a suciently small c(n) subject to

|V (y)|2 dy
F
sup
c(n)
(11.9.6)
2
{e:diam e1} cap(e, h2 )
(the values of c(n) in the suciency and necessity parts are dierent). For
2m = n one should add u2L2 (Rn ) to the last integral.
The condition

|V (y)|2 dy
F
0 as 0
(11.9.7)
sup
2
{e:diam e} cap(e, h2 )
is necessary and sucient for (11.9.4) to hold with an arbitrary a and b = b(a).
An obvious necessary condition for (11.9.4) is

supr1,xRn r 4n Br (x) |V (y)|2 dy c(n)
for n 5,

(11.9.8)
2 1
2
supr1,xRn (log r )
|V (y)| dy c(n) for n = 4,
Br (x)
where c(n) is suciently small. Standard lower estimates of cap2 by mn combined with the criterion (11.9.6) give the sucient condition

sup{e:diam e1} (mn e)(4n)/n e |V (y)|2 dy c(n) for n 5,

(11.9.9)
sup{e:diam e1} (log mvnne )1 e |V (y)|2 dy c(n)
for n = 4.
Though sharp and looking similar, (11.9.8) and (11.9.9) are not equivalent.
588
Finally, we observe that the inequality

V uL2 (Rn ) CuL2 (Rn ) u1
L2 (Rn )
holds for a certain (0, 1) and every u C0 (Rn ) if and only if for all
r (0, 1)

V (y)2 dy Cr n4
sup
x
Br (x)
(see Theorem 1.4.7).
11.10 Embedding Theorems for p = 1

11.10.1 Integrability with Respect to a Measure
The aim of the present subsection is to prove the following theorem which
complements Theorem 1.4.3.
Theorem 1. Let k be a nonnegative integer, 0 < l k n, 1 q < .
Then the best constant A in
k uLq (Rn ,) Aubl1
is equivalent to
(11.10.1)

1/q
.
K = sup lkn B(x, )
x, >0
Proof. () We show that A cK. We put u() = (x1 1 )k (1 (x )),

where C0 (B2 ), = 1 on B1 , into (11.10.1). Since

k uqLq (Rn ,) k! B(x, ) ,
(11.10.2)
ubl1 = cnl+k ,
(11.10.3)
then A cK.
() We prove that A cK. Let q > 1. By Theorem 11.8 and Remark
11.8/2 we have
k uLq (Rn ,) c sup
x,
(B(x, ))1/q
k(ln+n/t)+n/t
ublnn/t ,
t
where t is a number suciently close to unity, t > 1. It remains to apply items

(iii) and (iv) of Theorem 10.1.3/4.
Next we show that A cK for q = 1. It suces to consider the case k = 0.
Let l (0, 1). According to Corollary 2.1.6,

l

y| |U (z) dz,
uL1 (Rn ,) cK
Rn+1
589
where U C0 (Rn+1 ) is an arbitrary extension of a function u to Rn+1 .

Taking into account the relation

ubl1 inf
|y|l U (z) dz
U
Rn+1
contained in Theorem 10.1.1/1, we arrive at A cK.

If l = 1, then by Theorem 1.4.3
uL1 (Rn ,) cK2 U (z)L(Rn+1 ) .
Minimizing the right-hand side over all U we conclude that A cK for the
space b11 .
Suppose the estimate A cK is established for l (N 2, N 1), where
N is an integer N 2. We prove it for all l (N 1, N ]. We have

( x) u()

d d(x) c |u|I1 dx,
|u| d(x) = c
| x|n
where I1 = |x|1n . By the induction hypothesis, the latter integral does
not exceed

c sup rl1n
I1 () d ubl1 .
x,r
B(x,r)
By Lemma 1.4.3 with q = 1 the last supremum is majorized by cK. The

theorem is proved.

Remark 1. We substitute the function u dened by u(x) = (x/) where
C0 (Rn ), > 0, into (11.10.1). Let . Then (11.10.1) is not fullled
for l k > n provided = 0.
For l k = n, q < inequality (11.10.1) holds if and only if (Rn ) < .
Theorem 2. Let 0 < k < l, l k n, 1 q < . The best constant C0
in
k uLq (Rn ,) C0 uB1l
is equivalent to
K0 =
(11.10.4)

1/q
lkn B(x, )
.
sup
x; (0,1)
Proof. The estimate C0 cK0 follows in the same way as C cK in

Theorem 1. To prove the reverse inequality we use the partition of unity
{j }j1 subordinate to the covering of Rn by open balls with centers at the
nodes of a suciently ne coordinate grid and apply Theorem 1 to the norm
k (j u)Lq (Rn ,j ) where j is the restriction of to the support of j . Then

k (j u)q dj
|k u|q d c
j
cK0q
j uqbll cK0q
1
q
j ubl1
590
q

(Here we made use of the inequality
ai ( ai )q , where ai 0, q 1.)
Now a reference to Theorem 10.1.3/3 completes the proof.

Remark 2. For l k n the best constant in (11.10.4) is equivalent to one

of the following values:

1/q
sup B(x, 1)
xRn
if q 1,
1

(i) (1q)1 (1q)q
Q
if 0 < q < 1,
i0
where {Q(i) } is the same sequence as in Theorem 11.6.2/1. The proof is contained in Remark 11.8/1 and Theorem 11.6.2/1.
11.10.2 Criterion for an Upper Estimate of a Dierence Seminorm
(the Case p = 1)
Let us consider the seminorm

1/q

u(x) u(y)q (dx, dy)
,
uq, =
(11.10.5)
where is an open subset of a Riemannian manifold and is a nonnegative

measure on , locally nite outside the diagonal {(x, y) : x = y}. By
denition, the product 0 equals zero.
In this section, rst, we characterize both and subject to the inequality
uq, CuL1 () ,
(11.10.6)
where q 1 and u is an arbitrary function in C (). We show that (11.10.6)

is equivalent to a somewhat unusual relative isoperimetric inequality.
Theorem. Inequality (11.10.6) holds for all u C () with q 1 if and
only if for any open subset g of , such that g is smooth, the relative
isoperimateric inequality
1/q
(g, \
g ) + (\
g , g)
Cs( g)
(11.10.7)
holds with the same value of C as in (11.10.6).

Proof. Suciency. Denote by u+ and u the positive and negative parts
of u, so that u = u+ u . We notice that
uq, u+ q, + u q,
and

|u| dx =

|u+ | dx +
(11.10.8)
|u | dx.
(11.10.9)
591
First, we obtain (11.10.6) separately for u = u+ and u = u . Let a > b and

let t (a, b) = 1 if a > t > b and t (a, b) = 0 otherwise.
Clearly,
u(y) q
1/q

(dx, dy)
dt

u(x)

1/q

q

u(x),
u(y)
+
u(y),
u(x)
dt
=
(dx,
dy)
.
t
t

uq, =
By Minkowskis inequality,

uq,

q
t u(x), u(y) + t u(y), u(x) (dx, dy)

t u(x), u(y) + t u(y), u(x) (dx, dy)
=
=
(Lt , \Nt ) + (\Nt , Lt )
1/q
1/q
dt
1/q
dt
dt,
where Lt = {x : u(x) > t} and Nt = {x : u(x) t}.

By (11.10.7) and the co-area formula (1.2.6), the last integral does not
exceed

s x : u(x) = t dt = C u(x) dx.
C
0
Therefore,

u q, C

u (x) dx,
and the reference to (11.10.8) and (11.10.9) completes the proof of suciency.
Necessity. Let {wm } be the sequence of locally Lipschitz functions in
constructed in Lemma 5.2.2 with the following properties:
1. wm (x) = 0 in \g,
2. wm (x) [0, 1] in ,
3. for any compactum K g there exists an integer N (e) such that wm (x) =
1 for x K and m N (e),
4. the limit relation holds

lim sup wm (x) dx = s( g).
m
By Theorem 1.1.5/1, the inequality (11.10.6) holds for all locally Lipschitz
functions. Therefore,
(11.10.10)
wm q, Cwm L1 () ,
and by property 4,
lim supwm q, Cs( g).
m
(11.10.11)
592
Further,

wm qq,

wm (x)q (dx, dy)
=
xg
y\g
wm (y)q (dx, dy)
+
x\g

wm (x) wm (y)q (dx, dy),
+
g
which implies
yg
wm qq,
wm (y)q (\
g , dy).
wm (x) (dx, \
g) +
g
This, along with property 3, leads to

lim inf wm qq, (g, \
g ) + (\g, g).
m
Combining this relation with (11.10.10) and (11.10.11), we arrive at (11.10.7).

Corollary 1. (One-Dimensional Case). Let
= (, ),
The inequality

where < .

u(x) u(y)q (dx, dy)
1/q

u (x) dx
(11.10.12)
with q 1 holds for all u C () if and only if

+ (\I,
I) 1/q 2C
(I, \I)
(11.10.13)
for all open intervals I such that I , and

+ (\I,
I) 1/q C
(I, \I)
(11.10.14)
for all open intervals I such that I contains one of the end points of .
Proof. Necessity follows directly from (11.10.7) by setting g = I. Let us
check the suciency of (11.10.13). Represent an arbitrary open set g of
as the union on nonoverlapping open intervals Ik . Then by (11.10.13) and
(11.10.14)

1/q
(g, \
g ) + (\
g , g)

1/q
(Ik , \
=
g ) + (\
g , Ik )
k

k
g ) + (\
g , Ik )
(Ik , \
1/q
s( Ik ),
593
which is the same as (11.10.7). The result follows from the Theorem.
Remark 1. Suppose that the class of admissible functions in the Theorem

is diminished by the requirement u = 0 in a neighborhood of a closed subset
Then the same proof leads to the same criterion (11.10.7) with the
F of .
only dierence that the admissible sets g should be at a positive distance
from F . For the example F = , i.e., for the inequality (11.10.6) with any
u C0 (), the necessary and sucient condition (11.10.7) becomes the
isoperimetric inequality
1/q
(g, \
g ) + (\
g , g)
Cs(g)
(11.10.15)
for all open sets g with a smooth boundary and compact closure g . If,
in particular, in Corollary 1, the criterion of the validity of (11.10.12) for all
u C0 () is the inequality (11.10.13) for every interval I, I . In the
case u = 0 near one of the end points = (, ), one should require both
(11.10.13) and (11.10.14), but the intervals I should be at a positive distance
from that end point.
Needless to say, the condition (11.10.7) is simplied as follows for a symmetric measure , i.e., under the assumption (E, F ) = (F , E):
(g, \
g )1/q 21/q Cs( g),
for the same open sets g as in the Theorem.
Remark 2. The integration domain in (11.10.5) excludes inequalities
for integrals taken over . This can be easily avoided assuming additionally

and that u C()
C ().
that is dened on compact subsets of
Then, with the same proof, one obtains the corresponding criterion, similar
to (11.10.7):

g ) + (\
g , g) 1/q Cs( g).
(
g , \
As an application, consider the inequality

u(x) u(y)s(dx)s(dy) C
|u| dx,
(11.10.16)
which holds if and only if

s( g)s(\g) 21 Cs( g),
for the same sets g as in the Theorem.
By Corollary 9.4.4/3,
(i) If is the unit ball in R3 , then
4s( g) s( g)s(\g),
and
(11.10.17)
594
(ii) If is the unit disk on the plane, then

1
s( g) 2 sin s( g) .
2
Moreover, the last two inequalities are sharp. Hence, the inequality (11.10.16)
holds with the best constant C = 8 if = B. In case (ii),
s( g) 21 min
sin
s( g)s(\g).
( )
Since the last minimum equals 1 , it follows that the best value of C in the
inequality (11.10.16) for the unit disk is 4.
Example. We deal with functions in Rn and prove the inequality
1/q

|u(x) u(y)|q
dx
dy
C
|u| dx,
(11.10.18)
n+q
Rn Rn |x y|
Rn
where u C0 (Rn ), n > 1, 0 < < 1, and q = n/(n 1 + ).
Let us introduce the set function

g i(g) :=
g
Rn \g
dx dy
n+q .
|x y|
By the Theorem we only need to prove the isoperimetric inequality

n1
i(g) nq c(, n)s(g),
(11.10.19)
for q = n/(n 1 + ). Let be the Laplace operator in Rn . If u = r , we

may write

1
u = n1 rn1 ur r = ( + n 2)r2 .
r
Setting = 2 n q, we arrive at
y |x y|2nq = (n 2 + q)|x y|nq .
Using (1.4.13) and Example 2.1.5/2, we obtain

1
i(g) =
y |x y|2nq dy dx
q(n 2 + q) g Rn \g

1
|x y|2nq dy dx
=
q(n 2 + q) g g y

1
|x y|n1+q dx dsy
q g g
1 1q
1q
1q
1q
nvn n
(nvn )1 n1
(mn g) n s(g)
s(g)1+ n1 .
q(1 q)
q(1 q)
Since
1 q =
595
(n 1)(1 )
,
n1+
inequality (11.10.19) follows.

Remark 3. Inequality (11.10.18) can be interpreted as the embedding

q Rn ,
L11 Rn W
where
L11 (Rn ) is the completion of the space C0 (Rn ) in the norm uL1 (Rn )
(Rn ) is the completion of C (Rn ) in the fractional Sobolev norm
and W
q
0

Rn
Rn
|u(x) u(y)|q
dx dy
|x y|n+q
1/q
.
We can simplify the criteria (11.10.7) for = Rn , replacing arbitrary sets

g by arbitrary balls B(x, ) similarly to Theorem 1.4.2/2, where the norm

uLq () =
1/q
Rn
|u|q d
is treated in place of uq, . Unfortunately, the best constant in the suciency

part will be lost.
Corollary 2. (i) If q 1 and

sup (1n)q B(x, ), Rn \B(x, ) + Rn \B(x, ), B(x, ) < ,
xRn ,>0
(11.10.20)
then the inequality

1/q

Rn
Rn

u(x) u(y)q (dx, dy)
CuL1 (Rn )
(11.10.21)
holds for all u C (Rn ) and

C q cq sup (1n)q B(x, ), Rn \B(x, ) + Rn \B(x, ), B(x, ) ,
xRn ,>0
(11.10.22)
where c depends only on n.
(ii) If (11.10.21) holds for all u C (Rn ), then

C q nq sup (1n)q B(x, ), Rn \B(x, )
xRn ,>0

+ Rn \B(x, ), B(x, ) .
596
Proof. Let g be an arbitrary open set in Rn with smooth boundary and

let {B(xj , j )} be a covering of g subject to
n1
cs(g),
(11.10.23)
j
j
where c depends only on n (see Theorem 1.2.2/2). Then

1/q q
g, Rn \g
B(xj , j ), Rn \g
B(xj , j ), Rn \g
j
q

1/q q
n1
n
q
B(xj , j ), R \B(xj , j )
(cB)
j
,
j
where B is the value of the supremum in (11.10.20). This and (11.10.23) imply

q
g, B(xj , j ) cBs(g) .
Similarly

q
Rn \g, g cBs(g) ,
and the result follows from the Theorem.

The assertion (ii) stems from (11.10.7) by setting g = B(x, ).
11.10.3 Embedding into a Riesz Potential Space

It is possible to obtain a similar criterion for the embedding
(),
L11 R
q
(11.10.24)
where 0 < < 1, q 1 and Rq () is the completion of the space C0 in the

norm

u(y)

uRq () =
dy
, q 1,

n+1
Rn |x y|
Lq ()
where is a measure in Rn .
In the case q > 1 and = mesn this norm is equivalent to the norm

()/2 u
Lq
in the space of Riesz potentials of order with densities in Lq . We shall see
that embedding (11.10.24) is equivalent to the isoperimetric inequality of a
new type.
Theorem. Let q 1 and 0 < < 1. The inequality

uRq () C
|u| dx,
Rn
(11.10.25)
597
where is a continuous nonnegative function, holds for all u C0 if and only

if, for any bounded open g Rn with smooth boundary g, the isoperimetric
inequality holds:

y dsy

C
(x) dsx .
(11.10.26)

n+1
g |x y|
g
Lq ()
Proof. Necessity of (11.10.26) follows by the substitution of a mollication of
a characteristic function of g into (11.10.26).
Let us prove the suciency. By the co-area formula,

y dsy

dt
.
uRq () =

n+1
Et |x y|
Lq ()
Hence it follows from Minkowskis inequality and (11.10.26) that

y dsy

dt
uRq ()

n+1
Et |x y|
Lq ()

(x) dsx dt = C
|u| dx.
C
Rn
Et
The result follows.
Remark. The set of the inequalities of type (11.10.27) is not void. Let us
show, for example, that for
q=
n
,
n1
n 2, 0 < < 1,
there holds the isoperimetric inequality

y dsy
c(n, )s(g).

n+1
g |x y|
Lq
(11.10.27)
In fact, since q 2, we have the well-known inequality

uhq cubq ;
the norm in b
q on the right-hand side does not exceed cuL1 and it remains
to refer to Theorem 11.10.3.
11.11 Criteria for an Upper Estimate of a Dierence

Seminorm (the Case p > 1)
11.11.1 Case q > p
Now we deal with the inequality
598
uq, CuLp () ,
(11.11.1)
where q > p > 1, and show that it is equivalent to a certain isocapacitary

inequality.
The capacity to appear in the present context is dened as follows. Let F1
and F2 be nonoverlapping subsets of , closed in . The p-capacity of the
pair (F1 , F2 ) with respect to is given by

u(x)p dx,
capp (F1 , F2 ; ) = inf
{u}
where {u} is the set of all u C (), such that u 1 on F1 and u 0

on F2 . Obviously, this capacity does not change if F1 and F2 change places.
In fact, capp (F1 , F2 ; ) is nothing but the p-conductivity of the conductor
K = (\F1 )\F2 (see Sect. 6.1.1). We use the new notation only in this section
to emphasize the symmetric dependence of the following geometric criteria on
two sets F1 and F2 .
Furthermore, if F is a closed set in Rn and F G, where G is an open
, then capp (F, \G; ) coincides with the p-capacity
set, such that G
capp (F ; G) dened in Sect. 2.2.1.
Theorem. Inequality (11.11.1) with p (1, q) holds for all u C () if
and only if for any pair (F1 , F2 ) of nonoverlapping sets, closed in ,
(F1 , F2 )p/q B capp (F1 , F2 ; ),
(11.11.2)
where c1 C B c2 C with positive c1 and c2 depending only on p and q.

In the suciency part we may assume that F1 and F2 are sets with smooth
Fi .
In the proof of this theorem, we use the inequality

f ()q 1q/p d
1/q
R+
cf Lp (R+ )
(11.11.3)
(see (4.6.7)) and the inequality

R+
R+
|f () f ()|q
d d
| |2+q/p
1/q
cf Lp (R+ ) ,
(11.11.4)
where q > p > 1, p = p/(p 1) and f is an arbitrary absolutely continuous

+.
function on R
A short argument leading to (11.11.4) is as follows. Clearly, (11.11.4) results from the same inequality with R in place of R+ , which follows, in its
turn, from the estimate
f Bq1(qp)/pq (R) cf Wp1 (R)
(11.11.5)
599
by dilation with a coecient and the limit passage as 0+ . (The standard

notations B and W for Besov and Sobolev spaces with nonhomogeneous norms
are used in (11.11.5).) To obtain (11.11.5), we recall the well-known Sobolevtype inequality
hLp (R) chB(qp)/pq (R) ,
q
(see Theorem 4 , Sect. 5.1 of Stein [724]) and put h = ( + 1)1/2 f , which
shows that
cf B 1+(qp)/pq (R) .
(11.11.6)
f W 1
(R)
q
By duality, (11.11.6) is equivalent to (11.11.5).

With (11.11.4) at hand, we return to the Theorem.
Proof. Suciency. Arguing as at the beginning of the proof of Theorem
11.10.2, we see that it suces to prove (11.11.1) for a nonnegative u. By the
denition of the Lebesgue integral

u d =
(N ) d =
(L ) d,
R+
R+
where is a measure, and therefore

P (u) d =
R+
(N ) dP ( ),
(11.11.7)
where P is a nondecreasing function on R+ . Putting here u = 1/v and Q( ) =

P ( 1 ), we deduce

Q(u) du =
(\L ) dQ( ),
(11.11.8)
R+
where Q is nonincreasing. We obtain

u(x) u(y)q (dx, dy)

q
u(x) u(y) + (dx, dy)
=

q
+
u(y) u(x) + (dx, dy)

q

=
u(x) u(y) + (dx, dy) + (dy, dx) .
By (11.11.7) and (11.11.8), the last double integral is equal to

q1

t u(y) + (N , dy) + (dy, N ) d
q
R+

= q(q 1)
( )q2
(N , \L ) + (\L , N ) d d.
+
R+
R+
600
Now, (11.11.2) implies

uqq, 2q(q 1)B

R+
R+
( )q2
+ cap(N , \L ; ) d d,
and using the function t(), the inverse of (4.7.18), we arrive at the
inequality
uqq, 2q(q 1)Bq/p

q2
q/p
t() t()
cap(Nt() , \Lt() ; )
R+
t ()t () d d.
By Lemma 2.2.2/1, for >
cap(Nt() , \Lt() ; ) ( )1p ,
and therefore,
uqq, 2q(q 1)Bq/p

q2
( )q/p t() t()

t ()t () d d. (11.11.9)
R+
Integrating by parts twice on the right-hand side of (11.11.9), we obtain

q
(t() t())q
q
+
1
d d
uqq, 2B q/p

2+q/p
p
p
R+ 0 ( )

q/p
q1
+
t() t () d
R+
= B q/p
+
q
p

1
p

|t() t()|q
d d
2+q/p
R+ R+ | |

t()q 1q/p d .
q
+1
p
R+
Hence, we deduce from (11.11.3) and (11.11.4) that

uq, cB1/p t Lp (R+ ) ,
(11.11.10)
where c depends only on p and q. It remains to refer to (4.6.6).

Necessity. Let F1 and F2 be subsets of , closed in . We take an arbitrary
function u C (), such that u 1 on F1 and u 0 on F2 , and put it into
(11.11.1). Then

(F1 , F2 ; )
p/q
F1
F2
1/q

u(x) u(y)q (dx, dy)

C
|u|p dx.
601
It remains to minimize the right-hand side to obtain

(F1 , F2 ; )p/q C capp (F1 , F2 ; ).
The result follows.
A direct consequence of the Theorem and the isocapacitary inequality for

capp (F ; G) (see (2.2.10) and (2.2.11)) is the following sucient condition for
(11.11.1) formulated in terms of the n-dimensional Lebesgue measure:
q(1n)/n

mn (G)
(F, \G) c log
mn (F )
if p = n,
(11.11.11)
and
1p

(F, \G) cmn (G)(pn)/n(p1) mn (F )(pn)/n(p1) ,
(11.11.12)
if p = n. Choosing two concentric balls situated in as the sets F1 and \F2

in (11.11.2) and using the explicit formulas for the p-capacity of spherical condensers (see Sect. 2.2.4) we see that the inequalities (11.11.11) and (11.11.12),
with concentric balls F and G placed in , are necessary for (11.11.1).
In the one-dimensional case the Theorem can be written in a much simpler
form.
Corollary. Let
< .
= (, ),
The inequality

1/q

u(x) u(y)q (dx, dy)

C
1/p
p
u (x) dx
(11.11.13)
holds for every u C () if and only if, for all pairs of intervals I and J of
the three types
I = [x d, x + d]
and
J = (x d r, x + d + r),
I = (, x]
and
J = (, x + r),
(11.11.14)
I = [x, ]
and
J = (x r, ),
(11.11.15)
where d and r are positive and J , we have

1/q
B,
r(p1)/p (I, \J)
(11.11.16)
where B does not depend on I and J.

Proof. The necessity of (11.11.16) follows directly from that in the Theorem
and the inequality
602
capp (I, \J; ) 2r 1p

(see Lemma 2.2.2/2).
Let us prove the suciency. By G1 we mean an open subset of such that
1 \F2 . Connected components of \F2 will be denoted by
F1 G1 and G
Jk . Let Jk contain the closed convex hull Ik of those connected components
of G1 which are situated in Jk .
Then

p/q
(F1 , F2 )p/q (G1 , F2 )p/q

(Ik , \Jk )
(Ik , \Jk )p/q ,

k
and since by (11.11.16)
1p

,
(Ik , \Jk )p/q B p dist{Ik , R\Jk }
we obtain
(F1 , F2 )p/q B p
1p
dist{Ik , R\Jk }
(11.11.17)
Consider an arbitrary function u C (), such that u = 1 on G1 and

u = 0 on F2 . Clearly, u = 0 on Jk . We have

p
p
|u | dx
|u | dx
|
uk |p dx,
(11.11.18)
Jk
Jk
where u
k = u on Jk \Ik , u
k = 1 on Ik , and u
k = 0 on Jk . Hence

1p
|u |p dx
.
dist{Ik , R\Jk }
Comparing this estimate with (11.11.17), we arrive at

|u |p dx (F1 , F2 )p/q ,
and minimizing the integral in the left-hand side over all functions u, we obtain
(11.11.2).

Remark. It is straightforward but somewhat cumbersome to obtain a
more general criterion by replacing the seminorm on the right-hand side of
(11.11.13) with

1/p
p
u (x) (dx)
,
(11.11.19)
where is a measure in . In fact, one can replace by its absolutely continuous part (d /dx) dx and further, roughly speaking, the criterion will follow
from the Corollary by the change of variable x , where
1/(1p)

dx.
d = d /dx
Restricting myself to this hint, the author leaves the details to the interested
reader.
603
11.11.2 Capacitary Sucient Condition in the Case q = p

In the marginal case q = p the condition (11.11.2) in Theorem 11.11.1, being
necessary, is not generally sucient. In fact, let n = 1, = R, and
(dx, dy) =
dx dy
.
|x y|p+1
Then as shown in the proof of Corollary 11.10.2/2, (11.11.2) is equivalent to

(11.11.16), and (11.11.16) holds since

d
dt
(I, R\J) =
|t
|p+1
|tx|<d
| x|>d+r

d
=
dt
cr1p ,
p+1
|t
|
|t|<d
| |>d+r
and the same estimate holds for I and J dened by (11.11.14) and (11.11.15).
On the other hand, (11.10.15) fails because

|u(x) u(y)|p
dx dy =
|x y|p+1
R R
for every nonconstant function u.
In the next theorem we give a sucient condition for (11.11.1) with q =
p > 1 formulated in terms of an isocapacitary inequality.
Theorem. Given p (1, ) and a positive, vanishing at innity, nonincreasing absolutely continuous function on R+ , such that

1/(1p) d p1 d
()
()
< .
S := sup
>0
0
Suppose that

1p
(F1 , F2 ) capp (F1 , F2 ; )
(11.11.20)
for all nonoverlapping sets F1 and F2 closed in . Assume also that

p1
()
d < .
(11.11.21)
K :=
0
Then
up, 2
1/p

p
S
(p 1)p1
1/pp
K1/p uLp ()
(11.11.22)
for all u C ().

Proof. We assume that u Lp () and the integral in (11.11.22) involving
derivatives of is convergent. Arguing as in the proof of Theorem 11.11.1 and
using (11.11.20) instead of (11.11.2), we obtain
604
upp, 2p(p 1)
0

p2
( ) t() t()
t () dt () d.
(11.11.23)
Owing to (11.11.21), we can integrate by parts in the inner integral in
(11.11.23) and obtain

p
( ) t() t() p1 d t () d
up, 2p
0
= 2p
0

( ) t() t() p1 t () d d.
By H
olders inequality

upp, 2p
A()1/p B 1/p d,
(11.11.24)
where
A=
0
and

B=
p
| ( )|
t() t() d

( )( )p1 t ()p d.
Using Theorem 1.3.2/1 concerning a two-weight Hardy inequality, we obtain

pp
SB,
(p 1)p1
which together with (11.11.24) gives

upp,
(1p)/p
2p (p 1)
p
1/p

0

( )( )p1 t ()p d d.
Changing the order of integration, we arrive at

1/pp 1/p
K t Lp (R+ ) .
up, 21/p p (p 1)1p S
It remains to apply (4.6.6).
Remark 1. If the requirement
u=0
on a neighborhood of a closed subset E of
is added in the formulations of Theorems 11.11.1 and the theorem just proved,
the same proofs give conditions for the validity of (11.11.1), similar to (11.11.2)
and (11.11.16). The only new feature is the restriction
(\F2 )
is at a positive distance from E.
605
In the important particular case E = , which corresponds to zero Dirichlet

data on , the conditions (11.11.2) and (11.11.20) become
and
(F, \G)p/q B capp (F ; G)
(11.11.25)

1p
(F, \G) capp (F ; G)
,
(11.11.26)
respectively, where F is closed and G is open, G F , and the closure of G is

compact and situated in . The capacity capp (F ; G) is dened with = G
in Sect. 2.3.1.
Using lower estimates for the p-capacity in terms of area minimizing functions, one obtains sucient conditions from (11.11.2), (11.11.16), (11.11.25),
and (11.11.26) formulated in terms of the area minimizing function C , introduced in Denition 2.1.4, with (, v) = v. For example, by (11.11.25) and
(11.11.26), inequalities (11.11.2) and (11.11.20) hold for all u C0 () if,
respectively,
mn (\G)

dv
(F, \G)p/q B
C (v)p/(p1)
mn (F )
and

(F, \G)
mn (\G)
mn (F )

dv
,
C (v)p/(p1)
where F and G are the same as in (11.11.25) and (11.11.26).

By obvious modications of the proof of suciency in Corollary 11.10.2,
one deduces the following assertion from the Theorem in this subsection.
Corollary. (One-Dimensional Case) With the notation used in Corollary 11.11.1, suppose that
(I, \J) (r).
Then there exists a positive constant c depending only on p and such that
up, cS 1/pp K 1/p u Lp ()

for all u C ().
Remark 2. Let us show that the condition K < , which appeared in
the Theorem of this subsection, is sharp. Suppose that there exists a positive
constant C independent of u and such that

u(t) u( )p (t ) dt d C u (t)p dt,
(11.11.27)
R
where is a convex function in C (R). We take an arbitrary N > 0 and put

u(t) = min{|t|, N } into (11.11.27). Then
606
N/2
0
(t )p (t ) dt d 2CN,
and setting here t = + s, we obtain

N/2 N
N/2

1
sp1 (s) ds p
sp1 (s) ds d 2CN.
pN
2
0
0
0
Hence K 4p1 C.
Remark 3. It seems appropriate, in conclusion, to say a few words about
the lower estimate for the dierence seminorm up, , similar to the classical
Sobolev inequality
1/q

q
|u| (dx)
Cup, ,
(11.11.28)
where is a subdomain of a Riemannian manifold, and are measures in

and , respectively, and u is an arbitrary function in C0 (). Suppose
that q p 1. Then a condition, necessary and sucient for (11.11.28), is
the isocapacitary inequality
sup
{F }
(F )p/q
< ,
capp, (F ; )
(11.11.29)
where F is an arbitrary compact set in and the capacity is dened by

capp, (F ; ) = inf upp, : u C0 (), u 1 on F .
The necessity of (11.11.28) is obvious and the suciency results directly
from the inequality

capp, (Nt ; ) d tp c(p)upp,
(11.11.30)
0
(see (4.1.5)).
Although providing a universal characterization of (11.11.28), the condition (11.11.29) does not seem satisfactory when dealing with concrete measures and domains. This is related even to the one-dimensional case (cf. Problem 2 in Kufner, Maligranda, and Persson [468]). As an example of a more
visible criterion, consider the measure on Rn Rn given by
(dx, dy) = |x y|np dx dy,
(11.11.31)
with 0 < < 1 and p < n. This measure generates a seminorm in the
n
homogeneous Besov space b
p (R ). With this particular choice of , we have
by Theorem 11.10.1/1 and Remark 11.8/3 that (11.11.28) holds with q > p > 1
and q p = 1 if and only if
(B(x, ))p/q
< .
np
xRn ,>0
sup
(11.11.32)
The inequality (11.11.32) is the same as (11.11.29) with F = B(x, ) and is

dened in (11.11.31).
607

Section 11.2. Inequalities similar to (11.1.4) were proved in the authors
paper [543], where (11.1.4) (and even a stronger estimate in which the role of
the capacity of the set Qt is played by the capacity of the condenser Qt \Q2t )
were derived for l = 1 and l = 2. In the more dicult case l = 2, the
proof was based on the procedure of smooth truncation of the potential
near equipotential surfaces (see Proposition 3.7). By combining this procedure
with the Hedberg inequality (11.2.2). D.R. Adams [5] established (11.1.4)
for the Sobolev space Wpl for any integer l. The proof of D.R. Adams is
presented in Sect. 11.2.1. The same tools together with Theorem 10.1.1/1 on
traces of functions in the weighted Sobolev space were used by the author to
derive (11.1.4) for functions in Wpl for all p > 1, l > 0. This implies the validity
of (11.1.4) for the Bessel potential space Hpl for all fractional l > 0 but only
for p 2. The latter restriction was removed by Dahlberg [219] whose proof
is also based on smooth truncation and on subtle estimates for potentials
with nonnegative density. Hansson [347, 348] found a new proof of (11.1.4) for
spaces of potentials that uses no truncation. Hanssons approach is suitable for
a wide class of potentials with general kernels. In Sect. 11.2.2 we presented the
authors proof (cf. [554]) of inequality (11.1.4) based on Hanssons idea [347],
but apparently simpler. In [20] D.R. Adams and Xiao proved the inequalities

cap Nt , blp,q dtp Cupbl , 1 < q p < ,

0
p,q

q/p q
cap Nt , blp,q
dt Cuqbl ,
p,q
1 < p q < ,
where cap is the capacity associated with the space blp,q (Rn ). The case q = p
is due to Mazya [548] and the case 1 p q < was considered by Wu
[797].
Section 11.3. The equivalence of (11.1.1) and an isocapacitary inequality
was discovered by the author in 1962 for the particular case p = q = 2, l = 1
(cf. Mazya [531, 534]). Results of this kind were later obtained in the papers
by Mazya [543, 548], D.R. Adams [5], Mazya and Preobrazhenski [577], and
others.
Section 11.4. The counterexample to the isocapacitary inequality in the
norm L22 () is taken from the paper by Mazya and Netrusov [572].
Section 11.5. We follow the survey-article [775] by Verbitsky.
Section 11.6. The capacitary characterization of the embedding in the
case q > p > 0 in Sect. 11.6.1 is due to the author [556] and the author and
Netrusov [572].
Sections 11.711.8. The presentation mostly follows the paper by the
author and Preobrazhenski [577]. In comparison with this paper the requirements on the function are made less restrictive for 1 < p < 2 l/n by
virtue of the results due to Hedberg and Wol [372], which appeared later.
608
Remark 11.8/2 proved by D.R. Adams [5] and Corollary 11.8/2 was established earlier by D.R. Adams [2] by a dierent method.
Inequalities related to Corollary 11.8/2, with sharp constants are available.
In fact, Cianchi [201] obtained the following far-reaching result concerning the
Yudovich inequality (see the Comments to Sect. 1.4).
Theorem. Let be an open bounded subset of Rn , n 2. Let be a
positive Borel measure on such that

sup B(x, ) : x Rn , 0 < < 1 <
(11.12.1)
for some (0, n]. Then there exists a constant C = C(, ) such that

exp
n1
nvnn n |u(x)|
uLn ()
1
n
n1
for every u D(). The constant nvnn

1
n
n1
n
d(x) C
(11.12.2)
in (11.12.2) is sharp, i.e.,
n1
(11.12.2) fails if nvn n is replaced by any larger constant, whenever there

exist x0 and 1 , C1 > 0 such that

(11.12.3)
B(x, ) C1 , 0 < 1 .
The case when is the Lebesgue measure in this theorem was established
by Moser [618] and extended to the space of Bessel potentials and to higher
derivatives of integer order by D.R. Adams [11]. The existence of an optimizer
was proved in the planar case by Carleson and Chang [169]. The general case of
an arbitrary can be found in Cianchi [201], where inequalities for functions
not necessarily vanishing on the boundary are taken into account (see also
Cianchi [198]). In particular, trace inequalities on of Mosers type are
obtained in [201]. Note that a general optimal capacitary inequality implying
Mosers result can be found in Sect. 4.6 (see also Mazya [561]). An extension
of the Yudovich inequality to the case when the derivatives belong to the
Lorentz spaces Ln,q , where 1 < q n is due to Hudson and Leckband [386].
Capacitary inequalities were used by Schechter in the study of estimates
of the form

1
1
m
M0
M0 (u) d(x) CM
M (1 ) u d(x)
for u C (Rn ), where M0 , M are convex functions and , are measures
[692].
Section 11.9. Here various applications of the results obtained in Chap. 11
are given. Section 11.9.2 is borrowed from Mazaya and Poborchi [576],
Sect. 8.6.
Theorems of the present chapter were applied to the problem of the description of classes of multipliers in various spaces of dierentiable functions
609
(see the book by Mazya and Shaposhnikova [588]). Here we restrict ourselves
to mentioning just a few results.
By multipliers acting from a function space S1 into another one S2 we
mean functions such that multiplication by them sends S1 into S2 . Thus, with
the two spaces S1 and S2 we associate a third one, the space of multipliers
M (S1 S2 ). If S1 = S2 = S we will employ the notation M (S).
The norm of the operator of multiplication by the multiplier is taken as
the norm of in M (S1 S2 ).
We list some equivalent normings of multipliers in pairs of Sobolev spaces.
Let m > l, mp < n, and further, either q > p > 1 or q p = 1. Then
M (W m (Rn )W l Rn )
p
q
mn/p
sup
gradl Lq (Br (x)) + L1 (B1 (x)) . (11.12.4)
r
xRn ;0<r<1
For the formulation of an analogous result for q = p > 1, we require the

capacity cl,p , as introduced in Sect. 10.4.1.
If m > l and p > 1, then
M (Wpm (Rn )Wpl (Rn ))
sup
{eRn ,diam e1}
gradl Lp (e)

+ sup L1 (B1 (x)) . (11.12.5)
[cm,p (e)]1/p
xRn
For m = l the second term to the right in relations (11.12.4) and (11.12.5)
has to be replaced by L (Rn ) .
If mp = n then the rst term to the right in (11.12.4) takes the form
sup
xRn ;0<r<1
| log 2/r|(p1)/p gradl Lq (Br (x)) ,
and if mp > n the space M (Wpm (Rn ) Wql (Rn )) coincides with the space
Wql (Rn , unif) of functions with the nite norm
sup Wql (B1 (x)) .
xRn
Multipliers have a number of useful properties. For instance, the traces of mull1/p
()
tipliers are multipliers of traces. More exactly, each function in M Wp
l
can be continued to as a function in M Wp () (we assume that the boundary
of is smooth). The converse statement, that traces of functions in M Wpl ()
l1/p
(), is trivial.
belong to M Wp
Sections 11.10 and 11.11. The results of these sections are due to the
author [551, 560, 564].
12
Pointwise Interpolation Inequalities
for Derivatives and Potentials
In the present chapter we consider various pointwise interpolation inequalities

for derivatives of integer and fractional order and apply them to some weighted
integral inequalities. In Sect. 12.1 we prove, among others, the estimate
1z/

z/
MIz f (x) c MI f (x)
Mf (x)
(12.0.1)
for all x Rn and 0 < z < < n. Here Iz is the Riesz potential and
M is the HardyLittlewood maximal operator. As an application we easily
derive weighted multiplicative inequalities of the GagliardoNirenberg type
for complex powers of and 1 .
Section 12.2 is mostly dedicated to the proof of the following sharp inequality and its corollaries

M u(x)
u(x) n(n + 1)D (u; x)
,
(12.0.2)
nD (u; x)
where
D (v; x) = sup
r>0

1

v(x)
v(y)
ds
y ,

(r)
Br (x)
M is the maximal operator dened by

M f (x) = sup
r>0

yx
f (y) dy ,
Br (x) |y x|
(12.0.3)
(12.0.4)
and is a certain strictly increasing function generated by the continuity

modulus . The barred integral stands for the mean value. A simple corollary
of (12.0.2) is the inequality
2 8
u (x) M u (x) M u (x)
3
(12.0.5)
with the best constant.

DOI 10.1007/978-3-642-15564-2 12,
611
612
12 Pointwise Interpolation Inequalities for Derivatives and Potentials
One of the results in Sect. 12.3 is the estimate

p (1)/p
/q

M|u|q (x)
,
(Dp, u)(x) c Mu u(x) (x)
(12.0.6)
where 0 < < 1, p [1, ), q max{1, pn/(n + p)}, and

(Dp, u)(x) =
Rn
|u(y) u(x)|p
dy
|y x|n+p
1/p
.
Section 12.3 also contains simple proofs of weighted norm multiplicative inequalities of the GagliardoNirenberg type for integer and fractional derivatives based on pointwise interpolation inequalities.
We conclude this chapter showing in Sect. 12.4 that (12.0.6) is useful in the
question of continuity of the composition operator u f (u) in a fractional
Sobolev space.
12.1 Pointwise Interpolation Inequalities for Riesz and

Bessel Potentials
12.1.1 Estimate for the Maximal Operator of a Convolution
Lemma. Let k be a nonnegative, nonincreasing function in L1 (0, ) and let
g L1 (Rn , loc). Assume that Mg(x) < for some x Rn . Then the integral

(Kg)(x) =
k |y x| g(y) dy,
Rn
is absolutely convergent and

MKg(x) 2n+1
Rn

k || d Mg(x)
Let x = 0 and r (0, ). We have

(Kg)(y) dy 2n g(z)
k |y| dy dz
Br
B
Br (z)
2r

+
k |z y| g(z) dz dy.
Proof.

Br
The rst term on the right in (12.1.1) is majorized by

k |y| dy g(z) dz.
2n
Rn
(12.1.1)
Rn \B2r
(12.1.2)
B2r
Since in the second term |z y| |z| r |z|/2, we have k(|z y|) k(|z|/2)
and it follows that this term does not exceed
12.1 Pointwise Interpolation Inequalities for Riesz and Bessel Potentials
613

k |z|/2 g(z) dz
Rn

n

tn dk(t) Mg(0)
=
g() d dk(t) 2 vn
0
0
B
2t

k || d Mg(0).
= 2n
Rn
This together with (12.1.2) completes the proof.

12.1.2 Pointwise Interpolation Inequality for Riesz Potentials
Let z be a complex number with z < n. The Riesz potential Iz is dened by

f (y)
Iz f (x) = c
dy,
nz
Rn |y x|
where the constant c is chosen in such a way that Iz = ()z/2 , i.e.,
1
Iz f (x) = Fx
||z Fx f,
with F denoting the Fourier transform in Rn .

Theorem. Let f L1,loc (Rn ) and let 0 < z < < n. Then

z/
1z/
.
Mf (x)
MIz f (x) c MI f (x)
(12.1.3)
Proof. We introduce a function in the Schwartz space S such that F = 1

in a neighborhood of the origin, and the functions

z
1
(12.1.4)
P (x) = c1 Fx
|| F () ,
z

1
Q(x) = c2 Fx
||
1 F () .
(12.1.5)
It is then evident that

Iz f (0) = P I f (0) + Q f (0).
Let m be a positive integer such that
0 < m + z 1.
Since
m!
P (x) = c
(y)x |x y|2mn+z dy,
! x Rn
||=m
where x is the gradient, we have
(12.1.6)
614

P (x) c |x| + 1 n+z .
Therefore, P L1 (Rn ) and by Lemma 12.1.1,
M(P I f )(0) c MI f (0).
(12.1.7)
We observe that the function ||z (1 F ()) is smooth, so for |y| 1

and for suciently large N

Q(y) c(N )|y|N .
For |y| < 1

Q(y) c|y|n+z + Iz (y).
Since the second term on the right is bounded, the last two estimates imply
Q L1 (Rn ). By Lemma 12.1.1,
M(Q f )(0) c Mf (0).
Combining this with (12.1.7) and (12.1.6), we nd

MIz f (0) c MI f (0) + Mf (0) .
The dilation y y/t with an arbitrary t > 0 implies

MIz f (0) c t(z) MI f (0) + tz Mf (0) ,
and it remains to minimize the right-hand side in t.
12.1.3 Estimates for |Jw |

Let z be a complex number. The Bessel potential Jz is dened by Jz =
( + 1)z/2 , i.e.,

z/2
1
1 + ||2
Jz f (x) = Fx
Fx f.
Another formula for Jz is
Jz f (x) = c
where

Rn
Gz (x y)f (y) dy,
(12.1.8)

Gz (x) = c|x|(zn)/2 K(nz)/2 |x| .
K is the modied Bessel function of the third kind. We formulate some

estimates of the kernel Gz used in the following (see Aronszajn [53]).
For |x| 1 one has

for z = n + k, k even,
k Gz (x) c log(2/|x|)
(12.1.9)
znk
+ 1) for other values of z.
c(|x|
Further, for |x| > 1

k Gz (x) c|x|(zn1)/2 e|x| .
615
(12.1.10)
As in Sect. 12.1.2, by we denote a function in the Schwartz space S such

that F = 1 in B1 . We put
(x) = (mn B )1 (x/).
Lemma. The following inequalities are valid:

( + 1)w/2 (x)
for |x| > 1 and arbitrary L > 0,

cL |x|
nw
c((|x| + )
+ 1) for |x| 1 and w = n,
c| log(2|x| + )|
for |x| 1 and w = n.
(In the trivial case of even positive w sharper estimates hold.)
Proof. For |x| > 1 and arbitrarily large L we have

|x|2L ( + 1)w/2 (x)
1

w/2
( )L 2 + 1
Fx
= Fx

2L

k, 2 + 1 w/2 |2Lk, Fx | d.
c
(12.1.11)
Rn k=0
By using Fx = (F )() we nd
|2Lk Fx |
cL 2Lk
,
(|| + 1)N
where N is an arbitrarily large positive number. Therefore, the integral on

the right in (12.1.11) is majorized by
c
2L

2Lk
k=0
Rn \B1/
|| + 1
wk
N
|| + 1
d

(z)2L
N
|| + 1
|| + 1
d c2L(z)n + c.
+c
B1/
Hence

( + 1)w/2 (x) c|x|2L
for |x| > 1.
For |x| <

( + 1)w/2 (x)

L
c
||w ||
d + w
Rn \B1/

d
B1/
(12.1.12)
cnw . (12.1.13)
616
It remains to estimate the left-hand side of (12.1.13) for < |x| < 1. We
start with w 0. The case w = 0 is trivial. If w = 0 we write

w
( + 1) (x) =
Gw (x y) (y) dy.
(12.1.14)
Rn
We divide the integration domain into B2 (x) and Rn \B2 (x). By (12.1.10) the
integral over Rn \B2 (x) is dominated by
c
Rn \B2 (x)
|x y|
|y|
L

dy cLn
Rn \B1
dy
.
|y|N +L
The estimates of the integral over B2 (x) are also straightforward. Let, for
example, w > n. By (12.1.9) the integral in question is majorized by

|x y|
B3
nw
cLn
|y| +
dy
B3 \B2|x|
|y|n+w+L

+ cLn |x|nw
B|x|/2
c|x|
nw
L
dy
+c
Ln
|x|L

B2|x| \B|x|/2
|x y|nw dy
dy
(|y| + )L
Setting s = [w] > 0 and t = {w} we proceed by induction in s. Let the

required estimates be proved for [w] < s. We use the identity
(+1)
s+t
2
= ()
s+1
(+1)
ts
2 1
s

(1)
sk
k=0

ts
s+1
(+1)k1+ 2 ,
k
which can be veried directly. By induction hypothesis the functions

( + 1)k1+
ts
+iw
2
have the majorant c|x|nst for < |x| < 1. Hence we are left with estimating the function
()s+1 ( + 1) 2 1+iw (x)
(s + 1)!
=c
y (y)x G2+2sw (x y) dy.
!
Rn
ts
(12.1.15)
||=s+1
We introduce a cuto function C0 (B4 \B1/4 ), = 1 on B2 \B1/2 .

Suppose n + t > 1. The remaining case n = 1, t = 0 will be dealt with
separately. Integrating by parts, we have

617

y
1
y (y)x G2+2sw (x y) dy
|x|
Rn

s+1

y

2s+2k G2+2sw (x y) dy.

(y) k,y 1
c

|x|
Rn
k=0
(12.1.16)
By (12.1.9) and (12.1.10) for k s + 1

2s+2k G2+2sw (x y) c|x y|nst+k .
(12.1.17)
(Note that n s t + k is negative since n + t > 1.) Hence the sum on the
right-hand side of (12.1.16) is majorized by

(y)|x y|nst dy
c
B|x|/2 (Rn \B2|x| )
s+1

|x|

+
B4|x| \B2|x|
k=1
B|x|/2 \B|x|/4

nst+k
(y)|x y|
dy .
(12.1.18)
Clearly,

(y)|x y|nst dy c|x|nst
B|x|/2
and

Rn \B2|x|

(y)|x y|nst dy cL
|x|
L
|x|nst .
Similarly, the integrals over B4|x| \B2|x| and B|x|/2 \B|x|/4 in (12.1.18) are estimated by
L
|x|nst+k .
cL
|x|
Thus

y

y (y)x G2+2sw (x y) dy c|x|nst . (12.1.19)
n 1 |x|
R
To estimate

y
y (y)x G2+2sw (x y) dy
|x|
Rn

(12.1.20)
we use (12.1.17) for k = s + 1 and obtain that for any suciently large K > 0
(12.1.20) does not exceed
618
K
cK
|x y|1tn dy
|x|
B4|x| \B|x|/4
Kns1
cK
|x|tns .
|x|
ns1
Combining this with (12.1.19), we nd

()s+1 ( + 1) ts
2 1+iw (x) c |x|nst
|x|
L
for < |x| < 1. Thus, for n + t > 1 the result follows.
Let n = 1, t = 0. As before, we have to estimate the function (12.1.15),
that is, the integral

ys+1 (y)xs+1 G2+s (x y) dy
R1

y
ys (y)xs+2 G2+s (x y) dy
=
1
|x|
R1

y
ys (y)xs+2 G2+s (x y) dy.
+
(12.1.21)
|x|
1
R
The rst integral on the right does not exceed

s

2s+2k

k

y

(y) y 1
c
G2+s (x y) dy.
x

|x|
R1
k=0
In view of (12.1.17) the argument previously used to estimate the right-hand

side of (12.1.16) gives the majorant c|x|1s for the last sum.
We pass to the second integral on the right in (12.1.21). Restricting ourselves to x > 0 without loss of generality, we can rewrite it as

4x
y s
s
(y) (1)x (x) xs+2 Gs+2 (x y) dy
x y
x/4
3x/4
s
ts+2 Gs+2 (t) dt.
(12.1.22)
+ (1)x (x)
3x
It is enough to assume that s is odd. We have

ts+2 Gs+2 (t) = const t1 + O(1)
as t 0,
because s+2 ( 2 + 1)(s+2)/2 is asymptotically equal to sgn as || .

Therefore, the second term in (12.1.22) does not exceed cL x1s (/x)L1s
and the rst one is dominated by

c
619
|( xy )ys (y) (1)xs (x)|

dy
|x y|
x/4

L

y s
.
cx sup y
y (y) cL x1s
x
x
(x/4,4x)
4x
Thus the second integral on the right-hand side of (12.1.21) has the majorant cL |x|1s (/x)L . Hence the function (12.1.21) is dominated by c|x|1s ,

12.1.4 Estimates for |Jw ( )|
Lemma. (i) For an arbitrarily large L there exists a constant cL such that
for |x| >

( + 1)w/2 ( )(x) cL
|x|
L
nw .
Here and elsewhere stands for the Dirac function.

(ii) There exists a constant c such that for |x| <
wn
for w n, w = n,
c|x|

( + 1)w/2 ( )(x) c log(2|x|1 ) for w = n,
wn
for w < n.
c
Proof. (i) We need to estimate the absolute value of
w/2
2

1
+1
1 (F )() .
(x) = Fx
Let |x| > and let N be a suciently large, positive integer. We have

N 2

+ 1 w/2 1 (F )() d
|x|2N (x)
Rn
2N

k, 1 (F )()
Rn k=0
d
( 2
1)(2N kw)/2
Since F = 1 in a neighborhood of the origin,

|x|2N (x) c
2N

Rn k=0
( 2
k d
= c2N nw ,
+ )2N kw
which gives the result.

(ii) Now let |x| < . In the case w n, w = n the assertion follows
from (12.1.9) and Lemma 12.1.3. Setting = and X = x/ we obtain
620

1
(x) = wn FX
In the case w < n

(x) cwn
2 + 2
2 + 1
w/2
w/2

1 (F )() .
d = cwn .
Rn
In the remaining case w = n we notice that for supp(1 F )

2 + 2
w/2

w/2

n1
= 2 + 1
+ O || + 1
,
uniformly with respect to (0, 1). Consequently,

1
(x) = wn FX
and therefore
2 + 1

w/2
1 (F )() + O(1)

(x) Gw (X) + Jw (X).
The second term on the right is bounded uniformly with respect to (0, 1)
and the rst term is O(log(2|X|1 )) if w = n, and is bounded if w = n,

w = 0. Hence (x) admits the required estimates for |x| < .
12.1.5 Pointwise Interpolation Inequality for Bessel Potentials
Theorem. Let f L1 (Rn , loc) and let 0 < z < . Then
1z/

z/
Mf (x)
MJz f (x) c MJ f (x)
(12.1.23)
for all x Rn .
Proof. We introduce the functions
P = ( + 1)(z)/2 ,
Q = Jz ( ),
(12.1.24)
(12.1.25)
Jz f = P J f + Q f.
(12.1.26)
where (0, 1). Clearly,
We claim that

(P J f )(y) dy c(z) MJ f (0)
(12.1.27)
Br
for all r > 0 and (0, 1). In fact,

d
d
P |t| dt c
c(z) .
+
L
n+(z)
Rn
Rn (|| + 1)
Rn (|| + )
Hence (12.1.27) follows from Lemma 12.1.1.

By Lemma 12.1.1,

zn

(/|x|)n+1
Q (x) c
c|x|zn

Therefore
Rn
and by Lemma 12.1.1

621
for |x| > ,

for |x| < .

Q (x) dx cz

(Q f )(y) dy cz Mf (0).
(12.1.28)
Br
Combining this with (12.1.27), we arrive at

MJz f (0) c (z) MJ f (0) + z Mf (0)
for all (0, 1). If
MJ f (0) <
(12.1.29)
z
Mf (0),
( z)
then the minimum of the right-hand side of (12.1.29) on (0, 1] is attained at

=
and is equal to
( z) MJ f (0)
z Mf (0)
1/
(0, 1)
1z/
z/

Mf (0)
,
c MJ f (0)
which gives (12.1.23). If

MJ f (0)
z
Mf (0),
( z)
we have by (12.1.29)

MJz f (0) c MJ f (0) + Mf (0) c1 MJ f (0).
Since by Lemma 12.1.1
MJ f (0) cMf (0),
we arrive at (12.1.23). The proof is complete.
Remark 1. Inequality (12.1.3) can be deduced from (12.1.23) by dilation

x x/ and by passage to the limit as 0. However, we included an
independent proof of (12.1.3) because it is simpler than that of (12.1.23).
622
Remark 2. If z = 0 and f 0, one has

MIz f (x) cIz f (x)
and
MJz f (x) cJz f (x)
(see, in particular, Lemma 12.1.1). This means that for real z, , and nonnegative f inequalities (12.1.3) and (12.1.23) are equivalent to Hedbergs inequalities (cf. [365] and [15], Sect. 3.1), which contains no operator M in front of
the potentials.
Obvious changes in the proofs of Theorems 12.1.2 and 12.1.5 give the
following generalization of (12.1.3) and (12.1.23).
Proposition. Let f L1loc (Rn ).
(i) If k < < n then for all x Rn
k/

1k/
(Mf )(x)
(Mk I f )(x) c (MI f )(x)
.
(12.1.30)
(ii) If k < then for all x Rn

k/

1k/
(Mf )(x)
(Mk J f )(x) c (MJ f )(x)
.
(12.1.31)
12.1.6 Pointwise Estimates Involving Mk u and l u

Here we obtain two multiplicative inequalities involving integer powers of the
Laplacian.
Corollary. Let k and l be positive integers. Then for all x Rn

1 2lk
k
(Mk u)(x) (Mu)(x)
Ml u (x) 2l ,
(Mk u)(x)
k

1 2l+1
k
Ml u (x) 2l+1 ,
Mu (x)
k < 2l,
(12.1.32)
k < 2l+1. (12.1.33)
Proof. Setting = l and u = Jl f in (12.1.31), we arrive at

1 kl
k
(Mk u)(x) (Mu)(x)
M( + 1)l/2 u (x) l .
Now (12.1.32) follows by the dilation x x and passage to the limit as
0.
By (12.1.32) with k replaced by k 1 we have

1 k1

2l
k1
u
u
Ml u (x) 2l , (12.1.34)
(x)
M
(x)
Mk1
xi
xi
where i = 1, 2, . . . , n. Using (12.1.32) once more with k = l = 1 we obtain

1/2
1/2
u
M
(x) c (Mu)(x)
(Mu)(x)
.
xi
623
(12.1.35)
To estimate Mu we write u = v and note that by (12.1.32)

1

1 2l1
(Ml v)(x) 2l ,
(M div v)(x) c (Mv)(x)
which implies

1 2l1
1
Ml u (x) 2l .
(Mu)(x) c (Mu)(x)
Combining this inequality with (12.1.35) we nd
1

2l
(Mu)(x) c (Mu)(x) 2l+1 Ml u (x) 2l+1 ,
which together with (12.3.13) completes the proof of (12.1.33).
12.1.7 Application: Weighted Norm Interpolation Inequalities for

Potentials
Let w be a nonnegative measurable function on Rn . By Lp (w dx) we mean
the space of measurable functions f with the nite norm

1/p
p

.
(12.1.36)
f Lp (w dx) =
f (x) w(x) dx
Rn
Proposition. Let 1 < q < , 1 < p < , and let

z 1 z 1
1
= 1
+
.
s
p q
Further, let the weight function w belong to the Muckenhoupt class Amin{q,p} .
Then
z/
1z/
Iz f Ls (w dx) cI f Lq (w dx) f Lp (w dx) ,

z/
1z/
Jz f Ls (w dx) cJ f Lq (w dx) f Lp (w dx) ,
0 < z < < n,
(12.1.37)
0 < z < .
(12.1.38)
Proof. By Theorem 12.1.2

1/s

s(1z/)
sz/
MI f (x)
Mf (x)
w(x) dx
,
Iz f Ls (w dx) c
Rn
which by H
olders inequality is majorized by
z/
1z/
cMI f Lq (w dx) Mf Lp (w dx) .

It remains to refer to the Muckenhoupt theorem on the boundednesss of the
operator M in L (w dx) for 1 < < . Inequality (12.1.37) is proved.
Duplicating the same arguments and using Theorem 12.1.5 one arrives at
(12.1.38).
624
12.2 Sharp Pointwise Inequalities for u

12.2.1 The Case of Nonnegative Functions
Let v be a function in Rn and let the function T (f ; x) be dened by (1.3.45),
where R is replaced by Rn . We start with the following almost obvious multidimensional corollary of the one-dimensional estimate (1.3.47).
Theorem. Let f be a dierentiable nonnegative function on Rn . Then for
all x supp f

f (x)
f (x) T (f ; x) 1
,
(12.2.1)
T (f ; x)
where 1 is the inverse of (1.3.46). If is concave, then this inequality with
x = 0 becomes an equality for
xn
f (x) = (1) xn +
( ) d.
0
Proof. Let h(t) = f (x + tex ), where ex is a unit vector parallel to the

gradient direction of f at the point x and t R. Theorem 1.3.6 applied to h
yields

h(t)
h (t) T (h ; x) 1
.
T (h ; x)
The estimate (12.2.1) follows if we notice that the function (1.3.51) is increasing.
The sharpness of inequality (12.2.1) is proved in the same way as in Theorem 1.3.6.

Remark. The last theorem obviously implies the following n-dimensional
generalization of (1.3.52):

|f (x) f (y)|
f (x)+1 + 1
f (x) sup
,
(12.2.2)
n
|x y|
yR
where > 0. The equality sign in (12.2.2) with x = 0 is attained at
f (x) =
+
x+1
n
xn .
+1
12.2.2 Functions with Unrestricted Sign. Main Result

Let M be the maximal operator dened by (12.0.4) where u is a locally
integrable function on Rn , n 1, Br (x) is the ball {y Rn : |x y| < r},
625
and the bar stands for the mean value of the integral. Clearly, M u(x) does
not exceed the sharp maximal function of Feerman and Stein [274]:

M u(x) = sup
r>0
Br (x)

u(y)
Br (x)

u(z) dz dy.
We write Br instead of Br (0) and use the notation |B1 | = mn (B1 ) and
|S n1 | = s(B1 ). We introduce the mean value of the vector-valued function v : Rn Rn over the sphere Br (x) as follows:

Av(x; r) =
v(y) dsy ,
(12.2.3)
Br (x)
and set
D (v; x) = sup
r>0
|v(x) Av(x; r)|

.
(r)
(12.2.4)
In particular, for a function v of one variable we have

|2v(x) v(x + r) v(x r)|
.
2(r)
D (v; x) = sup
r>0
In what follows, we assume that is a continuous nondecreasing function

on [0, ) such that (0) = 0 and () = .
The objective of this section is the next result.
Theorem. Let the function

(1 n + n) n1 (t) d
(t) :=
(12.2.5)
be strictly increasing on [0, ) and let 1 be the inverse function for .

Further let
t
1 ( ) d.
(t) =
0
Then for any u C (R )

1

u(x) n(n + 1)D (u; x) 1

M u(x)
,
nD (u; x)
(12.2.6)
where 1 is the inverse function for .

Let C 1 (0, ). Suppose the function t (t) is nondecreasing on (0, )

and that, for n > 1, the function t (t) is nondecreasing on (0, ). Let R be
a unique root of the equation
626
n(n + 1)(t) = 1.
(12.2.7)
Inequality (12.2.6) with x = 0 becomes an equality for the function

1
xn (1 n 0 n1 (|x|) d) for 0 |x| R,
0 for |x| (n + 1)R,

1
u(x) = nxn
n1
( (n+1)R|x|
) d
|x| ((n + 1)R |x|) 0 ((n + 1) n)

n
for R < |x| < (n + 1)R.

(12.2.8)
12.2.3 Proof of Inequality (12.2.6)

It suces to prove (12.2.6) for x = 0. We have

u(0) u(y) 1 |y| dy
B1

n1
1
y

=
u(y) dy.
u(0) S
n(n + 1)
|y|
B1
Hence,
|B1 |
u(0) =
n+1
(12.2.9)
y
dy
|y|

1 n1

+ S n1
r
(1 r) u(0) Au(0; r) dr. (12.2.10)
u(y)
B1
After the scaling y y/t, r r/t, equality (12.2.10) becomes

1
y
|B1 |
tu(0) = n
u(y) dy
n+1
t Bt
|y|

n1 1 t n1

r
(t r) u(0) Au(0; r) dr. (12.2.11)
+ S
tn 0
This implies

u(0) n + 1 M u(0) + n(n + 1) D (u; 0)
t
tn+1
r n1 (t r)(r) dr,
0
which can be written as

t

( ) d. (12.2.12)
0 tu(0) + (n + 1)M u(0) + n(n + 1)D (u; 0)
0
Since is strictly increasing, it follows that the right-hand side in (12.2.12)

attains its minimum value at

|u(0)|
1
.
(12.2.13)
t =
n(n + 1)D (u; 0)
627
Thus, by (12.2.5) one has

0 (n + 1)M u(0) u(0) 1
|u(0)|
n(n + 1)D (u; 0)
|u(0)|
1 ( n(n+1)D
)
(u;0)
+ n(n + 1) D (u; 0)
( ) d
|u(0)|
1 ( n(n+1)D
)
(u;0)
= (n + 1)M u(0) n(n + 1)D (u; 0)
x d(x).
Therefore,
M u(0) nD (u; 0)
|u(0)|
n(n+1)D (u;0)
1 ( ) d,

12.2.4 Proof of Sharpness
Let us prove rst that (12.2.7) has a unique root. Note that (0) = 0 by
(12.2.5). Since t (t) is nondecreasing, one has () = . It remains to
show that (t) > 0 for t > 0. To this end, we only need to check that
t (t)|t=0 = 0. Since the function (12.2.5) can be written as
(t) =
n
tn+1
n ( ) d
0
n1
tn
n1 ( ) d,
(12.2.14)
we see that

t
1 t n
( ) d n
( ) d
n
t 0
0
t

n+1 t n
n
n1 ( ) d n
( ) d.
= n1
t
t
0
0
(12.2.15)
Hence,
1
tn
n ( ) d (n 1)
0
( ) d = t(t)
(12.2.16)
and thus
t (t) = (t)
n
tn+1
n ( ) d n(t).
(12.2.17)
The last relation, combined with (0) = (0) = 0, shows that t (t) vanishes
at t = 0.
628
Let us now prove that u dened by (12.2.8) is in C 1 (Rn ). We claim that u is

continuous on the sphere |x| = R together with its rst partial derivatives. We
use spherical coordinates to write xn = r cos . Denote the function u(x)/ cos
by u1 (r) for 0 |x| R and by u2 (r) for R < |x| < (n + 1)R, i.e.,
r
n
tn1 (t) dt
(12.2.18)
u1 (r) = r n1
r
0
and

2
u2 (r) = n
n+1
nr
t (t) dt
where
n1
r
n1

(t) dt ,
(12.2.19)

r = n1 (n + 1)R r .
(12.2.20)
Clearly,

u2 (R) = n2
n+1
Rn
tn (t) dt
0
n
Rn1

tn1 (t) dt .
(12.2.21)
By (12.2.14) we can write

u2 (R) = n (n + 1)R (R)
1
Rn1

tn1 (t) dt .
It follows from (12.2.7) and (12.2.18) that the right-hand side is equal to
u1 (R). Let us show that
(12.2.22)
u2 (R) = u1 (R).
u1 (r) = 1 +
n(n 1)
rn
tn1 (t) dt n(r).
(12.2.23)
By (12.2.19)
u2 (r) = n2
n+1
n+1
r
tn (t) dt
n1
nr
tn1 (t) dt

(r )
.
n
(12.2.24)
Therefore,
u2 (R)

2
=n
n+1
Rn+1
n1
t (t) dt
Rn
t
0
n1

(R)
,
(t) dt
n
which by (12.2.14) can be rewritten as

u2 (R) = n(n + 1)(R) +
n(n 1)
Rn
tn1 (t) dt n(R).

0
629
Using (12.2.7) and (12.2.23) we arrive at (12.2.22). It remains to note that by

(12.2.19) and (12.2.24)

u2 (n + 1)R = 0.
u2 (n + 1)R = 0,
Hence, u is in C 1 (Rn ).
Our next goal is to show that
M u(0) =
R
n
n+1
(t) dt.
(12.2.25)

y
1
u(y) dy
r Mr u :=

|Br | Br |y|
(12.2.26)
Let us nd the maxima of the function
on [0, R] and [R, (n+1)R] separately. Recall that, for 0 |x| R, the function
u can be written as cos u1 (r), where u1 is dened by (12.2.18). It is clear
that the function (12.2.26) is equal to
r
/2
2|S n2 |
n
d
(cos )2 (sin )n2 d
|B1 |rn
0
0

r
/2
n
tn1 (t) dt d
(cos )2 (sin )n2 d .
Since

2S n2
/2

(cos )2 (sin )n2 d = n1 S n1 ,
(12.2.27)
it follows that

r
n r n1
t
(t) dt d
n
n+1 r 0 0

r
r
1
r
1
n1 ( ) d n
n ( ) d ,
=
n n1
n+1
r
r 0
0
Mr u =
and by (12.2.16) we arrive at

r
Mr u =
n
n+1
(t) dt.
(12.2.28)
As was proved above, (t) > 0 for t > 0. Therefore,

max Mr u = MR u.
(12.2.29)
max
(12.2.30)
0rR
We prove that
Rr(n+1)R
Mr u = MR u.
630
By (12.2.26) one has

Mr u =

R
1
R
n
|R
(t)
dt
|B
n
1
|B1 |rn
n+1
0

r
n2 /2

+ 2S
(cos )2 (sin )n2 d
u2 ()n1 d .
0
In view of (12.2.27) we obtain

r

R
R
n
(t) dt +
u2 ()n1 d .
Mr u = r n Rn
n+1
0
R
Hence and by (12.2.28), to prove (12.2.30) we need to show that the function

r
R

R
n
(t) dt
u2 ()n1 d (12.2.31)
G(r) := r n Rn
n+1
0
R
is nonnegative on the interval [R, (n + 1)R]. Clearly,

R
1
R

n1
n
(t) dt u2 (r) .
G (r) = nr
n+1
n
0
Note that by (12.2.19) and (12.2.14)
r

r
1
(t) dt r (r ) = n
t (t) dt.
u2 (r) = n
n
0
0
Integrating by parts and using (12.2.7), we nd
R

n1
n
t (t) dt n
G (r) = nr
0
t (t) dt
(12.2.32)

0.
Hence and by G(R) = 0, one has G(r) 0 for R r (n + 1)R. Thus,

(12.2.25) holds.
Let us now justify the relation
sup
r>0
|u(0) Au(0; r)|

= 1,
(r)
(12.2.33)
where A is dened by (12.2.3). Let 0 r R. It follows from (12.2.8) that

r

n2 r n1
n
u
=1 n
tn1 (t) dt (cos )2 n(r) n
t
(t) dt ,
xn
r 0
r 0
(12.2.34)
which together with (12.2.27), implies
A
u
(0; r) = 1 (r).
xn
631
Hence,
u
u
(0) A
(0; r) = (r).
xn
xn
Combining this fact with the formulas
u
(0) = 0,
xk
we obtain
u
(0; r) = 0,
xk
k = 1, . . . , n 1,
|u(0) Au(0; r)|

= 1 for 0 r R.
(r)
(12.2.35)
We now claim that

(r )
Note that
u
u
(0) A
(0; r) (r)
xn
xn
for R r (n + 1)R.
(12.2.36)

1
u2 (r)
u
= u2 (r) (cos )2
u2 (r) ,
xn
r
r
where u2 is given by (12.2.19). In view of (12.2.27) one has

A
1
u
n 1 u2 (r)
+ u2 (r).
(0; r) =
xn
n
r
n
(12.2.37)
By (12.2.19) and (12.2.24),

r
n(n 1) n + 1 r n
n
u
n1
(0; r) =
t (t) dt n1
t
(t) dt
A
xn
r
nr
r
0
0

r
r
n1
n+1
n
n1
t (t) dt
t
(t) dt (r ).
+n
nr
n+1
r
0
0
(12.2.38)
Next observe that

r
n + 1 r n
n
t
(t)
dt
tn1 (t) dt
nr
n1
r
0
0
n1
r
t
t
1
=
t (t) dt 0,
r
r
0
because is nondecreasing. This, together with (12.2.38) and the inequality
r r , yields

r

n 1 r n1
n
u
(0; r) n(n+1) n+1
tn (t) dt n
t
(t) dt (r ).
A
xn
r
r
0
0
By (12.2.14), this inequality can be written as
632
u
(0; r) n(n + 1)(r ) (r ).
xn
(12.2.39)
Since 1 is strictly increasing and r < R, it follows that

A
u
(0; r) n(n + 1)(R) (r ).
xn
We now use the identity

u
(0)
xn
n(n + 1)(R) = 1 =
to obtain
A
u
u
(0; r)
(0) (r ),
xn
xn
which implies the left inequality in (12.2.36).

To prove the right inequality in (12.2.36), we note that, by (12.2.15), relation (12.2.38) can be rewritten as

r
r
n(n 1) 1
u
n
(0; r) =
t
(t)
dt
n
(t)
dt
A
xn
r
n
0
r r 0
n
+ n+1
tn (t) dt + n(r ) (r ).
(12.2.40)
r
0
Let
B(r) := r(r) r + rA
u
(0; r).
xn
(12.2.41)
The right inequality in (12.2.36) is equivalent to the inequality B(r) 0 for

all values of r in [R, (n + 1)R]. By (12.2.40) we have

r
r
1
n
t (t) dt n
(t) dt
B(r) := r(r) r + n(n 1) n
r 0
0

r
n
tn (t) dt + n(r ) (r ) .
(12.2.42)
+ r n+1
r
0
Using the relations R = R and n(n + 1)(R) = 1, we obtain
n2
B(R) = n
R
t (t) dt n (n 1)
n
(t) dt
n
R.
n+1
(12.2.43)
We note that relation (12.2.43), together with (12.2.16) for t = R, gives

B(R) = 0.
The next step is to show that B (r) 0 for r [R, (n + 1)R]. Combining
(12.2.37) and (12.2.32), we see that

u
n(n 1) r
A
(0; r) =
t (t) dt r (r ),
xn
r
0
633
which together with (12.2.41), gives

B(r) = r(r) r + n(n 1)
t (t) dt rr (r ).
Clearly,

r
.
t (t)
B (r) = r(r) 1 nr (r ) +
n
t=r
(12.2.44)
For n = 1, by (12.2.17) and (12.2.14), one has

2 t
1 t 2
t (t) = (t) 2
( ) d = 2
( ) d.
t 0
t 0
Since t (t) is nondecreasing, it follows that

t
t2
2 ( ) d 0.
0
Thus, t (t) is also nondecreasing for n = 1. Hence, and by assumption of the

theorem, both the functions t (t) and t (t) are nondecreasing for n 1.
Therefore, the last term on the right-hand side in (12.2.44) is nonnegative and
r (r ) R (R) for r R. Thus,
B (r) (R) + R (R) 1 nR (R)
(12.2.45)
for r [R, (n + 1)R]. Owing to relation (12.2.17) for t = R, the last inequality
can be rewritten as
R
n2
1
B (r) R (R) (n 1)(R) + n+1
tn (t) dt
. (12.2.46)
R
n
+
1
0
By (12.2.14) for t = R, relation (12.2.46) gives
B (r) R (R) (n 1)(R) +
n(n 1)
Rn
tn1 (t) dt.

0
Integrating by parts on the right-hand side, we obtain

R

n1 R n
1

B (r) R (R)
t (t) dt n
tn t (t) dt.
n
R
R 0
0
Since the function t (t) is nondecreasing, it follows that the right-hand side
is nonnegative. This implies the right inequality (12.2.36), and together with
(12.2.35), leads to (12.2.33).
Finally, we must show that inequality (12.2.6) with x = 0 becomes an
equality for u given by (12.2.12). It follows from (12.2.7) and (12.2.25) that
634
1
n(n+1)
n
0
1 ( ) d
=n

t d(t) = n R(R)
R
n
n+1
(t) dt
0
(t) dt = M u(0).
By (12.2.33), the right-hand side of (12.2.6) is equal to

1
1
= 1.
n(n + 1)
n(n + 1)
12.2.5 Particular Case (r) = r , > 0
Setting (r) = r with > 0 in Theorem 12.2.2 and using the notation
D (u; x) = sup
r>0
|u(x) Au(x; r)|

,
r
we obtain the following corollary to Theorem 12.2.2.

Corollary 1. Let u C 1 (Rn ), and let > 0. Then inequality
1

+1

u(x) C1 M u(x) +1 sup |u(x) Au(x; r)|
r
r>0
(12.2.47)
holds with the best constant

C1 = (n + 1)
1

+1
n
+1
.
(n + )(n + + 1)
Inequality (12.2.47) with x = 0 becomes
n
n+
|x| )
xn (1
1
((n+1)R|x|)+1 xn
u(x) = n (n+)(n++1)
|x|
0
where

R=
(12.2.48)
an equality for the function

for 0 |x| R,
for R < |x| < (n + 1)R,
for |x| (n + 1)R,
(n + )(n + + 1)
( + 1)n(n + 1)
1
.
denote the modCorollary 2. (Local version of Corollary 12.2.5.) Let M

ied maximal operator given by

M u(x) = sup
0<r<1
Br (x)
and let
635

yx
u(y) dy
|y x|
(u; x) = sup |u(x) Au(x; r)| .

D
r
0<r<1
Then, for any > 0, the inequality

1
1

(u; x) +1 + C2 M
u(x) +1 M
u(x) +1 (12.2.49)
u(x) C1 D
holds with the best constants C1 dened by (12.2.48) and C2 = n + 1.
Proof. It suces to set x = 0. It follows from (12.2.11) that

n
(u; 0)t .
u(0)t1 +
u(0) (n + 1) M
D
(n + )(n + + 1)
The right-hand side attains its minimum value either at

t=
u(0)
(n + )(n + + 1) M
(u; 0)
n
D
1
+1
<1
or at t = 1. Thus we arrive at the following alternatives: either

u(0)
M
and
n
(u; 0)
D
(n + )(n + + 1)

1
u(0) +1 D
(u; 0) +1 ,
u(0) C1 M
(12.2.50)
with C1 dened by (12.2.48), or

u(0)
M
and
n
(u; 0)
D
(n + )(n + + 1)

1
1

+1
+1

C1
u(0) +1 .
M
D (u; 0)
+ C2 M u(0)
+1
(12.2.51)
Inequalities (12.2.50) and (12.2.51) imply (12.2.49).
To show that the constant C1 is sharp, we make the dilation x x,
0 < < 1, in (12.2.49). Then

u(0)

1

1

u(x) C1 D (u; x) +1 + C2 M u(x) +1 M u(x) +1 .
Passing to the limit as 0, we arrive at (12.2.49) with the best constant C1 .
636
To prove that the constant C2 is sharp, we set u(x) = xn . Then (12.2.49)

becomes

u(x).
u(x) (n + 1)M
(12.2.52)
Clearly,

x
x2n

xn dx = sup
dx
0<r<1
Br |x|
Br |x|

r
2n|S n2 | /2
2
n2
= sup
(cos
)
(sin
)
d
n d,
n1 |r n
0<r<1 |S
0
0

M u(0) = sup
0<r<1

u(0) = (n + 1)1 .
M
Thus inequality (12.2.52) with x = 0 becomes an equality. This completes the
proof of the corollary.

12.2.6 One-Dimensional Case
For n = 1 the operator (12.2.4) becomes
D (u ; x) = sup
r>0
and M is dened by
M u(x) = sup
r>0
|2u (x) u (x + r) u (x r)|

2(r)

1 x+r
.
sign(y
x)u(y)
dy

2r xr
(12.2.53)
The next corollary immediately follows from Theorem 12.2.2.

Corollary 1. Let u C 1 (R). Then the inequality

M u
u (x) 2D (u ; x) 1
D (u ; x)
(12.2.54)
holds, where 1 is the inverse function for

t
1 ( ) d,
(t) =
0
with
standing for the inverse function for

1
(t) =
(t) d.
0
Suppose t (t) is nondecreasing on (0, ). Then inequality (12.2.54) with

x = 0 becomes an equality for the odd function u given on the semi-axis x 0
by the formula
1
d)
x(1 0 (x)
1
u(x) = (2R x) 0 (2 1)((2R x)) d
637
for 0 x R,
for R < x < 2R,
for x 2R,
where R is a unique root of the equation 2(t) = 1.

Set
T (u ; x) = sup
yR
|u (x) u (y)|

(|x y|)
and note that D (u ; x) T (u ; x). Moreover, if u is odd, then D (u ; 0) =
T (u ; 0). Therefore, Corollary 12.2.6 implies the following assertion.
Corollary 2. Let u C 1 (R). Then

u (x) 2 T (u ; x) 1

M u
.
T (u ; x)
(12.2.55)
Inequality (12.2.55) becomes an equality for the same function as in Corollary

12.2.6. As in Corollary 12.2.6, we here assume that r (r) is nondecreasing
on (0, ).
In the special case (t) = t , > 0, Corollaries 12.2.6 and 12.2.6 can be
stated as follows.
Corollary 3. Let u C 1 (R), and let > 0. The inequality
1

+1

u (x) C1 M u(x) +1 sup |2u (x) u (x + r) u (x r)|
r
r>0
(12.2.56)

2( + 1) +1
.
(12.2.57)
C1 =
1
( + 2)
Inequality (12.2.56) becomes an equality for the odd function u whose values for x 0 are given by
+1
1/
for 0 x ( +2
,
( + 1)x x
2 )
+2 1/
+2 1/
+1
1/
u(x) = +2 (2( 2 )
x)
for ( 2 )
< x < 2( +2
,
2 )
+2 1/
0
for x 2( 2 ) .
(12.2.58)
Note that the sharp estimate (12.0.5) is a particular case of (12.2.56) for
= 1. It implies a rougher, but nevertheless sharp estimate
2

u (x) 8 M u (x)u L
638
with the equality sign for x = 0 provided by the function (12.2.58) with = 1.
Taking into account that the second dierence at 0 for an odd function
(12.2.58) is twice as big as the corresponding rst dierence, we arrive at the
Corollary 4. Let u C 1 (R), and let > 0. Then

+1

|u (y) u (x)|
2+1 + 1
u (x)
M u(x) sup
,
+2
|y x|
yR
(12.2.59)
which becomes an equality for the same function as in Corollary 12.2.6.
12.3 Pointwise Interpolation Inequalities Involving

Fractional Derivatives
Let m be positive and a noninteger with [m] and {m} denoting its integer and
fractional parts. We introduce the operator Dp,m

(Dp,m u)(x) =
Rn

[m] u(x) [m] u(y)p |x y|np{m} dy
1/p
.
Sometimes, we call Dp,m u the fractional derivative of u of order m. In

Sect. 12.3.1 we derive pointwise interpolation inequalities with Dp,m u in their
right-hand sides.
12.3.1 Inequalities with Fractional Derivatives on the Right-Hand
Sides
Theorem. (i) Let k and l be integer, and let m be noninteger, 0 l k < m.
Then

mk
kl
k u(x) c (Ml u)(x) ml (Dp,m u)(x) ml
(12.3.1)
for almost all x Rn .
(ii) Let k and m be integer, and let l be noninteger, 0 < l < k m. Then

mk
kl
k u(x) c (Dp,l u)(x) ml (Mm u)(x) ml
(12.3.2)

(iii) Let k be integer, and let l and m be noninteger, 0 < l < k < m. Then

mk
kl
k u(x) c (Dp,l u)(x) ml (Dp,m u)(x) ml
(12.3.3)

Proof. (i) It suces to prove inequality (12.3.1) for l = 0 and x = 0. Let
be a function in the ball B1 with Lipschitz derivatives of order m 2, which
vanishes on B1 together with all these derivatives. Also let
12.3 Pointwise Interpolation Inequalities Involving Fractional Derivatives
639

(y) dy = 1.
B1
Let t be an arbitrary positive number to be chosen later. We shall use the

Sobolev integral representation

(y)
v(y)(y/t) dy + (1)[m]k [m] k
tn
v(0) =
!
Bt
||<[m]k

y
dy
y

(12.3.4)
v(y)
n1 d n
!
|y|
|y|
|y|/t
Bt
||=[m]k
(see Sect. 1.1.10).

By setting here v = u with an arbitrary multi-index of order k and
integrating by parts in the rst integral, we arrive at the identity

1 +
k n
y (y/t) dy
u(0) = (1) t
u(y)
!
Bt
||<[m]k

[m]k
(1)
[m] k
+
||=[m]k
Bt
y +
u(y)
!

y
dy
n1 d n .

|y|
|y|
|y|/t
(12.3.5)
Hence, for k < [m] we have

k u(0) c tk Mu(0) + t[m]k [m] u(0)

|[m] u(y) [m] u(0)|
+
dy
.
|y|n[m]+k
Bt
Holders inequality implies

|[m] u(y) [m] u(0)|
dy ctmk (Dp,m u)(0).
n[m]+k
|y|
Bt
(12.3.6)
(12.3.7)
Let be an arbitrary multi-index of order [m]. The identity

y
y
n
n
u(y) dy + t
u(0) (y) dy
u(0) = t
t
t
Bt
Bt
implies

y
u(y)([m] )
dy
t
Bt
1/q

(y)q |y|( np +{m})q dy
(Dp,m u)(0),
+ t{m}

[m] u(0) tn[m]
Bt
(12.3.8)
640
where p1 + q 1 = 1. Combining (12.3.7)(12.3.8) we arrive at

k u(0) c tk (Mu)(0) + tmk (Dp,m u)(0) .
Minimizing the right-hand side in t we complete the proof of (i).
(ii) It is sucient to take l (0, 1) and x = 0. Since the function
+ (y (y)) in (12.3.5) is orthogonal to 1 in L2 (B1 ), it follows from (12.3.5)
that

|m u(y)|
k u(0) c tnk
u(y) u(0) dy +
dy
.
(12.3.9)
nm+k
Bt
Bt |y|
If m k n, the second integral does not exceed

mkn
m u(y) dy.
t
Bt
In the case m k < n the same integral is estimated by (11.2.3)

n
dy
mk
n
m u(y)
m u(y) dy.
t
sup
nm+k
|y|
m
k
t
Bt
B
(12.3.10)
Hence, and by H
olders inequality, applied to the rst integral in (12.3.9) we
have

k u(0) c tk Dp,l u(0) + tmk Mm (x) .
The result follows.

(iii) By (12.3.9) with m replaced by [m]

k u(0) c tlk Dp,l u(0) + t[m]k [m] u(0)

|[m] u(y) [m] u(0)|
dy
.
+
|y|n[m]+k
Bt
By (12.3.7) the third term on the right does not exceed
tmk Dp,m u(0).
Now we note that (12.3.8) implies

[m] u(0) c tl[m] Dp,l u(0) + t{m} Dp,m u(0) .
Hence

k u(0) c tk Dp,l u(0) + tmk Dp,m u(0) .
The result follows.
12.3 Pointwise Interpolation Inequalities Involving Fractional Derivatives
641
12.3.2 Inequality with a Fractional Derivative Operator on the

Left-Hand Side
Theorem. Let 0 < < 1 and let p [1, ). Then
p (1)/p

/q
,
M|u|q (x)
(12.3.11)
where q max{1, pn/(n + p)}.

Proof. Let rst p n/(n 1). Then q = pn/(n + p). Let t be an arbitrary
positive number to be chosen later. By Sobolevs embedding theorem

1/p
|u(y) u(x)|p
dy
n+p
Bt (x) |y x|

q 1/q

u(y) u(x)

c1
y |y x|+n/p dy
Bt (x)
1/q

|u(y) u(x)|q
1
+ c2 t
dy
.
(+n/p)q
Bt (x) |y x|
The right-hand side of this inequality does not exceed

c1
Bt (x)
|u(y)|q dy
|y x|(+n/p)q
1/q

+ c2
Bt (x)
|u(y) u(x)|q
dy
|y x|(+1+n/p)q
1/q
,
which by Hardys inequality is dominated by

c
Bt (x)
|u(y)|q dy
|y x|(+n/p)q
1/q
.
Estimating this by (11.2.3) we arrive at the inequality

Bt (x)
|u(y) u(x)|p
dy
|y x|n+p
1/p

1/q
ct1 M|u|q (x)
.
(12.3.12)
Now let p [1, n/(n 1)) and let (n 1 + , n). By Holders inequality
with exponents n/(n 1)p and n/(n (n 1)p) we have

|u(y) u(x)|p
dy
n+p
Bt (x) |y x|

n
(n1)p

n
|u(y) u(x)| n1
dy
|y x|
Bt (x)

1 (n1)p
n
dy
.
(n+p())n
Bt (x) |y x| n(n1)p
642
The right-hand side is equal to

ct
p(n+1)
Bt (x)
|u(y) u(x)|
|y x|
n
n1
(n1)p
n
dy
which by the GagliardoNirenberg inequality does not exceed

p

|u(y) u(x)|
y u(y) u(x) dy + t1
ctp(n+1)
dy

|y x|
|y x|
Bt (x)
Bt (x)
p

|u(y)|
|u(y) u(x)|
p(n+1)
ct
dy +
dy .
+1
Bt (x) |y x|
Bt (x) |y x|
By Hardys inequality the second integral on the right is dominated by the
rst one. Hence and by (11.2.3) we obtain
p

|u(y) u(x)|p
|u(y)|
p(n+1)
dy
ct
dy
n+p
Bt (x) |y x|
Bt (x) |y x|

p
ctp(1) M|u| (x) .
(12.3.13)
Unifying this with (12.3.12), we nd

p/q
|u(y) u(x)|p
dy ctp(1) M|u|q (x)
,
n+p
Bt (x) |y x|
(12.3.14)
where q = max{1, pn/(n + p)}. For any nonnegative f and any a > 0 there
holds the inequality

f (y) dy
cta (Mf )(x),
(12.3.15)
|y
x|n+a
n
R \Bt (x)
which follows from the readily checked identity

f (y) dy
n+a
Rn \Bt (x) |x y|

ds
= (n + a)
f (y) dy n+a+1 tan
f (y) dy.
s
t
Bs (x)
Bt (x)
(see Hedberg [365]). By (12.3.15)

p

|u(y) u(x)|p
dy ctp Mu u(x) (x).
n+p
Rn \Bt (x) |y x|
(12.3.16)
(12.3.17)
Summing up (12.3.14) and (12.3.17) we arrive at

p

p

p/q

(Dp, u)(x) c tp(1) M|u|q (x)
+ tp Mu u(x) (x) .
Minimizing the right-hand side in t > 0 we complete the proof.
12.4 Application of (12.3.11) to Composition Operator
643
12.3.3 Application: Weighted GagliardoNirenberg-Type

Inequalities for Derivatives
Proposition. Let 1 < q < , 1 < p < , and let

k 1
k 1
1
=
+ 1
.
s
mp
m q
(i) Let k be an integer and let l and m be noninteger, 0 < l < k < m. Then
for any nonnegative measurable w
mk
kl
k uLs (w dx) cDp,l uLml

Dp,m uLml
.
q (w dx)
p (w dx)
(12.3.18)
(ii) Let k and l be integers and let m be noninteger, 0 l k < m. Further

let w Aq . Then
mk
kl
k uLs (w dx) cl uLml

Dp,m uLml
.
q (w dx)
p (w dx)
(12.3.19)
(iii) Let k and m be integers and let l be noninteger, 0 < l < k m. Further
let w Ap . Then
mk
kl
m uLml
.
k uLs (w dx) cDp,l uLml
q (w dx)
p (w dx)
(12.3.20)
Proof. We use the same argument as in the proof of Proposition 12.1.7.

Inequality (12.3.18) follows from (12.3.3) and H
olders inequality. To show
(12.3.19) we notice that by (12.3.1) its right-hand side does not exceed

c
Rn
1/s

s mk
s kl
ml
Dp,m u(x) ml w(x) dx
Ml u(x)
which by H
olders inequality is majorized by
mk
kl
cMl uLml
Dp,m uLml
,
q (w dx)
p (w dx)
and it remains to refer to the Muckenhoupt theorem on the boundedness of
the operator M in L (w dx)), 1 < < . Inequality (12.3.20) is proved in
the same way.

in Fractional Sobolev Spaces
12.4.1 Introduction
In this section we use the space Wps of distributions with the nite norm
644
uWps = [s] uWp{s} ,

where {s} > 0 and

vWp{s} =

Rn
Rn
|v(x) v(y)|p
dx dy
|x y|n+{s}p
1/p
.
The application mentioned in the title of this section is as follows.

Theorem. Let p 1 and let s be noninteger, 1 < s < . For every complex-valued function f dened on R and such that f (0) = 0 and
f , . . . , f ([s]+1) L (R) there holds

f (u)

f (l)
uWps (Rn ) + usLps (Rn ) ,
L (R)
[s]+1
Wps (Rn )
(12.4.1)
l=1
where c is a constant independent of f and u.

If additionally, f ([s]+1) C(R) then the map

n

1
Wps Rn Wsp
R u f (u) Wps Rn
(12.4.2)
is continuous.
We formulate a particular case of the inequality (12.3.11), which is used
in the proof of this theorem.
Corollary. Suppose (0, 1), p 1, and u Wp1 (Rn , loc). Then for
almost all x Rn
p (1)/p

/p
,
(12.4.3)
M|u|p (x)
where M is the HardyLittlewood maximal operator.
Another ingredient in the proof of the theorem is the GagliardoNirenbergtype inequality
sk
k1
s1
k uLps/k (Rn ) cuLs1
n Dp,s uL (Rn ) ,
ps (R )
p
(12.4.4)
where 1 k < s, s is noninteger, and p 1.

A short argument leading to (12.4.4) is by the pointwise inequality

sk
k1
k u(x) c M|u| (x) s1 (Dp,s u)(x) s1 .
(12.4.5)
In fact, one uses (12.4.5) to majorize the Lps/k -norm of |k u| and applies
Holders inequality with exponents
(s 1)k/(s k)
and
(s 1)k/(k 1)s
together with the boundedness of M in Lps (Rn ). Inequality (12.4.5) was

proved in Sect. 12.3.1.
645
12.4.2 Proof of Inequality (12.4.1)

Since f (0) = 0, we have

f (u)
f L (R) uLp (Rn ) .
Lp (Rn )
(12.4.6)
By the Leibnitz rule,

Dp,s f (u)
Lp (Rn )
[s]

l=1

l

i u
Dp,{s} f (l) (u)

|+...+| |=[s]

|1
i=1
l
|i |1
Lp (Rn )
(12.4.7)
We continue by putting
v = f (l) (u)
and
w=
l

i u
i=1
in the obvious inequality

Dp,{s} (vw)
vDp,{s} wLp (Rn ) + wDp,{s} vLp (Rn )
Lp (Rn )
(12.4.8)
and arrive at

Dp,s f (u)
L
n
p (R )
[s]

l=1

l

i
(l)
u Dp,{s} f (u)

|+...+| |=[s]

|1
l
|i |1
i=1

l

(l)

i

D
+ f
u

p,{s}
L (R)

i=1
i=1

.
(12.4.9)
Lp (Rn )
l

i
(l)
Il := u Dp,{s} f (u)

We set
Lp (Rn )
(12.4.10)
Lp (Rn )
By inequality (12.4.3) with f (l) (u) in place of u,

1{s}

p {s}/p
Dp,{s} f (l) (u)(x) cf (l) L (R) Mf (l) (u) (x)
1{s}
{s}

{s}/p
cf (l) L (R) f (l+1) L (R) M|u|p (x)
.
Hence, using H
olders inequality with exponents s/{s}, s/|i |, i = 1, . . . , l, we
nd

1{s}
{s}
Il cf (l) L (R) f (l+1) L (R)
l

i u
i=1
ps
|i |
(Rn )

M|u|p {s}/pn .
Ls (R )
(12.4.11)
646
Since M is bounded in Ls (Rn ), the last factor on the right is majorized by

l
{s}
cu
n . By (12.4.4) the product
i=1 in (12.4.11) does not exceed
Lps (R )
c
l

s|i |
|i |1
s1
s1
uLps
(Rn ) Dp,s uLp (Rn )
i=1
sl[s]
[s]l
s1
= cuLs1
n Dp,s uL (Rn ) .
ps (R )
p
(12.4.12)
Hence and by (12.4.11)

s(l1+{s})
[s]l
1{s}
{s}

s1
Il cf (l) L (R) f (l+1) L (R) uLps s1
D
u
n
p,s
(R )
Lp (Rn ) .

l

Jl := Dp,{s}
i u

Let
(12.4.13)
i=1
(12.4.14)
Lp (Rn )
Clearly,
J1 Dp,s uLp (Rn ) .
Now let l > 1. By (12.4.8),

l

i
1
Jl u Dp,{s} u

i=2
Lp (Rn )
(12.4.15)

l

1

i
+ u Dp,{s}
u

i=2
Lp (Rn )
(12.4.16)
Applying rst the inequality (12.4.3) and then Holders inequality with exponents

s/|i |, 2 i l, s/|1 | 1 {s} , and s/ 1 + |1 | {s},
we nd

l

i u Dp,{s} 1 u

i=2
Lp (Rn )

l

p
(1{s})/p
p
{s}/p

c i u M 1 u
M |1 u

i=2
l

i=2

i u
L
Lp (Rn )
ps
|i |
(Rn )
p 1{s}
p {s}
1 p
M 1 u p

L s (Rn ) M u
L
|1 |
s
1+|1 |
(Rn ) .
Hence, noting that s > 1 + |1 | and using the boundedness of M in Lq (Rn )

with q > 1, we obtain that the rst term on the right-hand side of (12.4.16)
is dominated by
l

i u
L
i=2
ps
|i |
(Rn )
1{s}
1 u
L
ps
|1 |
(Rn )
{s}
1 u
L ps
1+|1 |
(Rn )
647
(12.4.17)
By (12.4.4) this does not exceed

s(l1)
sl
s1
cuLs1
n Dp,s uL (Rn ) .
ps (R )
p
(12.4.18)
We estimate the second term in the right-hand side of (12.4.16). By

(12.4.3),

l

1

i
u
u Dp,{s}

n
i=2
Lp (R )

p (1{s})/p l
l
p {s}/p

1

i
i
c u M u
u
M

i=2
i=2
which by H
olders inequality with exponents

s/|1 |, s/ [s] |1 | 1 {s} , and
Lp (Rn )

s/ 1 + [s] |1 | {s},
is majorized by

c 1 uL

(1{s})/p
l

p

M i u
ps (Rn )

|1 |
i=2
s
[s]|1 |
Using the Lq -boundedness of M with

q = s/ [s] |1 | and
(Rn )
p {s}/p

l

i u
M

i=2
s
1+[s]|1 |
(Rn )

q = s/ 1 + [s] |1 | ,
we conclude that

l

1

i
u
u Dp,{s}

i=2
Lp (Rn )

1{s}
l

c 1 uL ps (Rn ) i u

|1 |
i=2
ps
[s]|1 |
(Rn )

{s}
l

i u

i=2
ps
1+[s]|1 |
(Rn )
(12.4.19)
By H
olders inequality with exponents ([s] |1 |)/|i |, 2 i l, the second
norm on the right-hand side of (12.4.19) does not exceed
c
l

i u
L
i=2
ps
|i |
(Rn )
(12.4.20)
648
Again by H
olders inequality, now with exponents

1 + [s] |1 | /|j |, 2 j l, j = i, and 1 + [s] |1 | / 1 + |i | ,
the third norm on the right-hand side of (12.4.19) is dominated by
l
l

j u
L
i=2
j=2
j=i
ps
|j |
(Rn )

i u
ps
1+|i |
(Rn )
(12.4.21)
Combining (12.4.19)(12.4.21) with (12.4.4) we nd that the left-hand side

of (12.4.19) does not exceed (12.4.18). Thus, (12.4.18) is a majorant for the
second term in (12.4.16). Hence and by (12.4.15),
s(l1)
sl
s1
Jl cuLs1
n Dp,s uL (Rn ) ,
ps (R )
p
1 l [s].
(12.4.22)
Inserting (12.4.13) and (12.4.22) into (12.4.9) we arrive at

Dp,s f (u)
L

f (l)
[s]+1
p
(Rn )
L (R)

Dp,s uLp (Rn ) + usLps (Rn ) .
l=1
Hence and by (12.4.6) the proof of (12.4.1) is complete.

12.4.3 Continuity of the Map (12.4.2)
Let u u in Wps (Rn ) L1sp (Rn ). Since
(l)

(l+1)
f (u ) f (l) (u)

f
c
u uLp (Rn )
n
Lp (R )
L (R)
for l = 0, . . . , [s], we have
f (l) (u ) f (l) (u)

in Lp Rn .
(12.4.23)
We shall prove that

Dp,s f (u ) f (u)
L
p (R
n)
0.
(12.4.24)
Let be a multi-index of order || = [s]. By the Leibnitz rule,

c(l, 1 , . . . , l )
f (u) f (u ) =
[s]
l=1
(l)
(u)
l

i=1
uf
i
(l)
(u )
l

i=1

i
u ,
649
where the second sum is taken over all l-tuples of multi-indices {1 , . . . , l }

such that 1 + + l = and |i | 1. We rewrite the dierence
f (l) (u)
l

i u f (l) (u )
i=1
l

i u
(12.4.25)
i=1
using the identity

l

ai
i=0
l

bi =
i=0
l

b0 , . . . , bk1 (ak bk )ak+1 , . . . , al ,
(12.4.26)
k=0
where the products of either bi or ai are missing if k = 0 or k = l, respectively.

Setting
a0 = f (l) (u),
ai = i u,
b0 = f (l) (u ),
1 i l,
b i = i u ,
in (12.4.26), we nd that (12.4.25) is equal to

f (l) (u) f (l) (u )
l

i u + f (l) (u )
i=1
l k1

l

i u k (u u )
k=1 i=1
i u.
i=k+1
Consequently,

Dp,{s} [s] f (u) f (u )
Lp (Rn )

[s]
l

(l)

c
i u
Dp,{s} f (u) f (l) (u )

| |+...+| |=[s]
l=1
i=1
l
|i |1
Lp (Rn )

l
k1
l

(l)
i
k
i
+
u (u u )
u
Dp,{s} f (u )

k=1
i=1
i=k+1

.
Lp (Rn )
(12.4.27)
By (12.4.22), (12.4.23) and the boundedness of derivatives of f we can
apply the Lebesgue dominated convergence theorem to conclude that

l

(l)
i u
0.
(12.4.28)
f (u) f (l) (u ) Dp,{s}

n
i=1
Lp (R )
Using (12.4.3) with u replaced by f (l) (u) f (l) (u ) and employing Holders
inequality with exponents s/|i |, 1 i l, and s/{s}, we obtain
650

l

(l)

i
(l)
u Dp,{s} f (u) f (u )

i=1
Lp (Rn )
l

p 1{s}

p
c
i u Mf (l) (u) f (l) (u )

i=1

(l)
p {s}
(l)
p

M f (u) f (u )

n
Lp (R )
1{s}
cf (l) L (R)
l

i u
i=1
ps
|i |
(Rn )
(l)
{s}
M f (u) f (l) (u ) p p
Ls (Rn ) .
The boundedness of M in Ls (Rn ) implies that the left-hand side of the last
inequality is dominated by
l

1{s}

i u
cf (l) L (R)
L
i=1
ps
|i |
(Rn )
(l)

f (u) f (l) (u ) {s} n .
Lps (R )
(12.4.29)
l
By (12.4.4), the product i=1 has the majorant (12.4.12). Obviously,
(l)

f (u) f (l) (u )
Lps (Rn )

(l)

(u u )L (Rn )
c f
L (R)
ps
(l+1)

(l+1)

+ f
(u) f
(u ) uLps (Rn ) .
Hence and by Lebesgues dominated convergence theorem
l

(l)

i
(l)
0.
u Dp,{s} f (u) f (u )

n
i=1
Lp (R )
This together with (12.4.28) implies that the rst term in brackets in the
right-hand side of (12.4.27) tends to zero.
We now show that

k1

l

(l)
i
k
j
u (u u)
u
0 (12.4.30)
f (u )Dp,{s}

n
i=1
j=k+1
Lp (R )
for any k = 1, . . . , l. Here the products

k1

i=1
i u
and
l

j u
j=k+1
are missing for k = 1 and k = l, respectively. Clearly, (12.4.30) holds for l = 1.

Let l > 1. By (12.4.8), the left-hand side in (12.4.30) is majorized by
651

k1
l

i
j
k
u
u Dp,{s} (u u)

(R)

i=1
j=k+1
Lp (Rn )

k1

l

. (12.4.31)
+ k (u u) Dp,{s}
i u
j u

n
(l)
f
L
i=1
j=k+1
Lp (R )
Applying inequality (12.4.3), we nd that the rst term in brackets in (12.4.31)

does not exceed

l
k1

p 1{s}

i
p
c
u
j u M k (u u)

i=1
j=k+1

p {s}

p
M k (u u)
.
(12.4.32)

n
Lp (R )
Holders inequality with exponents

s
,
|i |
s
,
|j |
1 i k 1,
s
,
|k |(1 {s})
k + 1 j l,
s
,
(1 + |k |){s}
as well as the boundedness of M in Lq (Rn ) with q = s/|k | and q = s/(1 +

|k |) yield that (12.4.32) is dominated by
c
k1

i u
i=1
ps
|i |
(Rn )
l

j u
j=k+1
ps
|j |
(Rn )

1{s}
{s}
k (u u)L ps (Rn ) k (u u)L ps
|k |
1+|k |
(Rn )
(12.4.33)
By (12.4.4) applied to each factor we see that (12.4.33) and therefore the rst
term in brackets in (12.4.31) tends to zero.
Making use of inequality (12.4.3) once more, we obtain that the second
term in brackets in (12.4.31) is majorized by

p 1{s}
k1
p
l

k

i
j
c (u u) M
u
u

i=1
j=k+1
k1

p {s}
p
l

i
j
M
u
u

i=1
j=k+1
.
Lp
(Rn )
Applying Holders inequality with exponents

s
,
|k |
s
,
([s] |k |)(1 {s})
s
,
(1 + [s] |k |){s}
(12.4.34)
652
and using the Lq -boundedness of M for q = s/([s] |k |) and q = s/(1 + [s]

|k |), we nd that (12.4.34) does not exceed

1{s} {s}
cN1
N2 k (u u)L ps (Rn ) ,
(12.4.35)
|k |
k1

l

i
j
N1 :=
u
u

where
i=1
j=k+1
ps
[s]|k |

k1
l

i u
j u
N2 :=

and
i=1
By (12.4.4),
j=k+1

k (u u)
L
ps
|k |
(Rn )
(Rn )
ps
1+[s]|k |
(Rn )
0.
It remains to show the boundedness of N1 and N2 . Using H

olders inequality
with exponents
[s] |k |
,
|i |
1 i k 1,
we nd
N1 c
k1

i u
i=1
[s] |k |
,
|j |
and
ps
|i |
l

j u
L
(Rn )
j=k+1
k + 1 j l,
ps
|j |
(Rn )
which is bounded owing to (12.4.4). Again by Holders inequality, now with

exponents
1 + [s] |k |
,
|i |
1 + [s] |k |
,
1 + |r |
1 i k 1, i = r,
and
1 + [s] |k |
, k + 1 j l, j = r,
|j |
we nd that N2 is majorized by
k1 k1

i u
r u
c
n
L
r=1
i=1
i=r
(R )
ps
|i |
l
k1

i u
r=k+1 i=1
l

j u
L
j=k+1
j=r
ps
|j |
ps
|i |
ps
1+|r |
1 + [s] |k |
,
1 + |r |
l

j u
L
(Rn )
j=k+1
ps
|j |
(Rn )
(Rn )

r u
(Rn )
L

ps
1+|r |
(Rn )
By (12.4.4) every norm on the right is bounded. The proof is complete.
653

First pointwise multiplicative inequalities for Riesz potentials are due to Hedberg: Lemma 11.2.1/1 was proved in [365] (see also [15], Sect. 3.1). A version
of inequalities (11.2.4) and (11.2.5) appeared already in [364], Lemma 2.
Bojarski and Hajlasz noticed in [123], 1993, that Hedbergs inequality
(11.2.4) combined with the classical Sobolev representation theorem gives the
pointwise estimate

f (x) f (y) M|f | (x) + M|f | (y) |x y|
(12.5.1)
as well as its generalizations to gradients of higher order. Inequality (12.5.1)
inuenced the development of analysis on arbitrary metric measure spaces
(see, for example, Hajlasz [337, 338, 342], Hajlasz and Koskela [344]).
Hedbergs Lemma 11.2.1/1 was generalized in Mircea and Szeptycki [609]
to potentials with kernels k satisfying the homogeneity condition k(tx) =
tn k(x) for all t > 0 and x Rn \ {0}. The result is the following. Let
Iu(x) = k u(x) and let M be the maximal operator dened by

1
u(x y) dy,
M u(x) = sup
t>0 vol(tX) tX
where X = {x Rn \ {0} : |K(x)| 1} {0}. Then

Iu(x) A M u(x) 1(1)p u(1)p
Lp
for any u Lp and almost all x Rn with the best constant

1
p
A=
vol(X) .
(1 )p 1 (1 )p
Section 12.1. Lemma in Sect. 12.1.1 is a renement of Steins estimate
k u ckL1 Mu
(see [724] Theorem 2, Chap. 3, Sect. 2.29). It was proved by Mazya and Shaposhnikova in [581]. Interpolation inequalities for Riesz and Bessel potentials
of complex order (Theorems 12.1.2 and 12.1.5) were obtained in [581].
Section 12.2. Theorem 12.2.1 is proved in the paper by Kufner and
Mazya [571]. Other results in this section were obtained in Mazya and Shaposhnikova [586] where, together with D (v, x), the nonlinear operator

1
v(y) dy
D (v; x) = sup
v(x)

(r)
r>0
Br (x)
was considered and a sharp analog of inequality (12.0.2) involving D was
obtained. In the case (r) = r , > 0, one has the following result.
654
Proposition. Let u C 1 (Rn ), and let > 0. Then the inequality

u(x) C2 M u(x) +1 sup r u(x)

r>0
Br (x)
1
+1

u(y) dy
(12.5.2)
1

+1
+1
C2 = (n + 1)
.
n++1
(12.5.3)
Inequality (12.5.2) with x = 0 becomes an equality for the odd function given
for x 0 by the formula
|x| )
for 0 |x| R,
xn (1
n
+1 xn
u(x) = n++1 ((n + 1)R |x|)
for R < |x| < (n + 1)R,
|x|
0
for |x| (n + 1)R,
where

R=
n++1
( + 1)(n + 1)
1
.
Note that for n = 1 estimate (12.5.2) takes the form

u (x) C2 M u(x) +1 sup u (x)
r>0
1
u(x+r)+u(xr) +1

2r
r
(12.5.4)
(compare with (12.2.56)) with the best constant

C2 =
1

+1
2( + 1)
.
+2
(12.5.5)
We note that the operator

(T u)(x) = sup
yRn
|u(y) u(x)|
|y x|
has been extensively studied (see, e.g., Carton-Lebrun and Heinig [173]).
Section 12.3. The paper by Kalamajska [411] is dedicated to some integral representation formulas for dierentiable functions and pointwise interpolation inequalities on bounded domains with the operator M both on the
right- and left-hand sides. Kalamajska proved, in particular, that if

u(x) dx = 0, where a Rn and r > 0,
lim Rk
R
BrR (aR)
then, for any polynomial P of degree less than j,
655

jk
k
Mk u(x) M u(x) P (x) j Mj u(x) j .
Interpolation inequalities for k u and Dp,m u collected in Theorem 12.3.1 are
borrowed from Mazya and Shaposhnikova [580]. A slightly weaker version of
Theorem 12.3.2 can be found in [583].
In [582] the interpolation inequality

k
k
k u(x) c (Mu)(x) 1 r (Sr u)(x) r ,
0 < k < r, {r} > 0,
involving the Strichartz function

2

[r] u(x + h) [r] u(x) dh
(Sr u)(x) =
0
By
dy
y 1+2{r}+2n
12
,
was proved and used to obtain a description of the maximal algebra embedded
in the space of multipliers between Bessel potential spaces.
Section 12.4. Inequality (12.4.1), with Bessel potential space Hps (Rn ),
p (1, ), and s > 0, instead of Wps (Rn ), was obtained by D.R. Adams and
Frazier [13].
For p > 1, a theorem on the boundedness and continuity of the operator
(12.4.2) was proved by Brezis and Mironescu [144]. As they say, their proof is
quite involved being based on a microscopic improvement of the Gagliardo
Nirenberg inequality, in the TriebelLizorkin scale, namely Wps L Fq,

for
every and on an important estimate on products of functions in the Triebel
Lizorkin spaces, due to Runst and Sickel [685]. To this Brezis and Mironescu
added: It would be interesting to nd a more elementary argument which
s
scale.
avoids this excursion into Fp,q
In the present section, which follows the paper by Mazya and Shaposhnikova [583], such an elementary argument, which includes p = 1, is given.
The problem of the description of functions of one real variable acting by
composition on spaces on the Sobolev type, dealt with here, was the subject
of numerous studies: Dahlberg [218]; Marcus and Mizel [515]; Bourdaud [129
132]; Bourdaud and Meyer [135]; D.R. Adams and Frazier [13]; Bourdaud and
M. Kateb [134]; Oswald [651]; Bourdaud and D. Kateb [133]; Runst and Sickel
[685]; Labutin [472]; Bourdaud, Moussai, and Sickel [136]; et al.
13
A Variant of Capacity
In Sect. 10.4.1 we introduced the capacity cap(e, Spl ) of a compactum e Rn

for any one of the spaces Spl = Hpl , hlp , Bpl , and so on. In other words, we
considered the set function

cap e, Spl = inf upS l : u N(e) ,
p
where N(e) = {u D : u 1 on e}.

Replacing N(e) by

M(e, ) = u C0 () : u = 1 in a neighborhood of e ,
we obtain another set function that appears to be useful in various applications. This new set function will also be called the capacity. It will be denoted
by Cap(e, Spl ()).
Having in mind further applications we restrict ourselves to the capacities
Cap generated by the spaces
Llp () and Wpl , l = 1, 2, . . . . Various estimates
for these capacities are collected in Sect. 13.1. Section 13.2 deals briey with
the so-called (p, l)-polar sets in Rn and in Sect. 13.3 we show the equivalence
of Cap(e,
Llp ) and cap(e,
Llp ) for p > 1.
Finally, in Sect. 13.4 an example of application of Cap to the problem of
removable singularities of polyharmonic functions is given.
13.1 Capacity Cap

13.1.1 Simple Properties of Cap(e,
Llp ())
The inner and outer capacities of an arbitrary subset of the set are dened
by

Cap E,
Llp () ,
Llp () = sup Cap e,
eE

DOI 10.1007/978-3-642-15564-2 13,
657
658
13 A Variant of Capacity

Llp () ,
Cap E,
Llp () = inf Cap G,
GE
respectively. Here e is an arbitrary compact subset of E and G is an arbitrary

open subset of containing E. If the inner and outer capacities coincide,
their common value is called the capacity of the set E and is denoted by
Cap(E,
Llp ()). Henceforth when speaking of the (p, l)-capacity we shall mean
this set function.
The denition of the capacity Cap(e,
Llp ()) immediately implies the following two properties.
Monotonicity. If e1 e2 and 1 2 , then

Llp (1 ) Cap e2 ,
Llp (2 ) .
Cap e1 ,
Right continuity. For each > 0 there exists a neighborhood of the
compactum e with
such that for an arbitrary compactum e with
e e we have

Llp () Cap e,
Llp () + .
Cap e ,
The following three propositions establish the simplest connections between the capacities Cap(, S1 ) and Cap(, S2 ).
Proposition 1. Let and be open sets in Rn with D = diam <
and
. Then for any compactum e , located at the distance e from
, the inequality
lp

D
Cap e,
Llp () c
Cap e,
Llp ()
e
is valid.
Proof. Let u M(e, ) and let C0 (), = 1 in a neighborhood of e
and |i (x)| ci
e , i = 1, 2, . . . (cf. Stein [724], 2, Chap. 6). Then
l

p
p

Cap e,
Llp () l (u)Lp () c
e(jl)p j uLp () .
j=0
It remains to make use of the Friedrichs inequality

j uLp () c Dlj l uLp () .


The following proposition can be proved in a similar way.
Proposition 2. Let e , D = diam , e = dist{e, }.Then

c1 min 1, lp
e

Cap(e, Wpl )
c2 max 1, Dlp ,
l
Cap(e, L ())
p
where
Wpl
Wpl (Rn ).
13.1 Capacity Cap
659
n
An immediate corollary of the embedding
Llp (Rn )
Lm
q (R ) for appropriate values of p, l, q, and m is the following.
Proposition 3. Let n > lp, p > 1 or n l, p = 1. If e Br , then

(13.1.1)
Llp ,
Cap e,
Llp (B2r ) c Cap e,
where c = c(n, p, l).
r , B2r ). We have
Proof. Let u M(e), M(B
l
l

l (u)
c

u
c
j uLqj c2 l uLp ,
j
1
Lp (B2r )
Lp (B2r )
j=0
j=0
where qj = pn(n (l j)p)1 .
Next we present two lower bounds for Cap(e,

Llp ()).
Proposition 4. Let e Rs , s n. Then

Cap e,
Llp () c (ms e)p/q ,
where q = ps/(n lp) for p > 1, s > n lp > 0, or s n l 0 for p = 1.
The constant c depends only on n, p, l, and q.
For the proof it suces to apply the inequality
uLq (Rs ) c l uLp (Rn )
to an arbitrary u M(e, ).
Proposition 5. If is an open set, e is a compactum in and

de = min max |x y|,
xe y
then

l
,
Cap e,
Llp () c dnp
e
(13.1.2)
where p l > n, p > 1 or l n, p = 1.

Proof. For any u M(e, ) and a certain x e we have

|l u|p dy, x e.
1 =
u(x)
c dpln
e
|yx|<de
ln/p
Therefore, 1 c de
l uLp () . The result follows.
Corollary. Let p l > n, p > 1 or l n, p = 1 and let x0 be a point in B .

Then

Llp (B2 ) nlp .
(13.1.3)
Cap x0 ,
Proof. The lower estimate for the capacity follows from Proposition 5
and the upper estimate results from substituting the function u(x) = ((x

x0 )1 ), where C0 (B1 ), into the norm l uLp (B2 ) .
660
13.1.2 Capacity of a Continuum

Proposition 1. Let n > lp > n 1, p 1 and let e be a continuum with
diameter d. Then

(13.1.4)
Cap e,
Llp dnlp .
d with radius d and we denote the conProof. We include e in the ball B
centric ball with radius 2d by B2d . Using the monotonicity of the capacity, we
obtain

d ,
Llp = c dnlp .
Cap e,
Llp Cap B
Let O and P be points in e with |O P | = d. Let the axis Oxn be directed
from O to P . We introduce the notation
x = (x , xn ),
(n1)
B2d
x = (x1 , . . . , xn1 ),
e(t) = e {x : xn = t},

n1
1
l = l /x
1 , . . . , xn1 ,
(t) = B2d {x : xn = t},
1 + + n1 = l.
For any u M(e, B2d ) we have

B2d
|l u|p dx
dt

0
d
(n1)
B2d
(t)
|l u|p dx

(n1)
Cap e(t),
Llp B2d (t) dt.
d , and p l > n 1, it follows that

Since e(t) = , e(t) B

(n1)
Cap e(t),
Llp B2d (t) c dn1lp .
Minimizing l upLp (B2d ) over the set M(e, B2d ), we obtain

Cap e,
Llp (B2d ) c dnlp .

To complete the proof it remains to use estimate (11.1.1).
Proposition 2. If n = lp, p > 1, then for any continuum e with diameter

d, 2d < D, the equivalence

Cap e,
Llp (BD )
1p

D
log
d
(13.1.5)
holds. Here BD is the open ball with radius D and with center O e.
Proof. First we derive the upper bound for the capacity. Let the function
v be dened on BD \Bd as follows:
13.1 Capacity Cap
D
v(x) = log
d
1
log
661
D
.
|x|
Let denote a function in C [0, 1] equal to zero near t = 0, to unity near

t = 1 and such that 0 (t) 1. Further let u(x) = [v(x)] for x BD \Bd ,
u(x) = 1 in Bd and u(x) = 0 outside BD . It is clear that u M(Bd , BD ).
Also, we can easily see that
1

l u(x)
c log D
|x|l
d
on BD \Bd . This implies

d ,
Cap B
Llp (BD )
l u(x)
p dx
BD
D
c log
d
p
lp
BD \Bd
|x|
D
dx = c log
d
1p
.
We proceed to the lower bound for the capacity. Let P and Q be points in
e with |P Q| = d. By (r, ) we denote the spherical coordinates of a point in
the coordinate system with origin Q, r > 0, B1 (Q). Let u be a function
in M(e, B2D (Q)) such that

|l u|p dx ,
B2D (Q)
where

Llp (BD )
= Cap e,
Llp B2D (Q) Cap e,
and is a small positive number. We introduce the function

U (r) = u(r, )Lp (B1 (Q)) .
Since u = 1 at least one point of the sphere {x : |x Q| = r}, where r < d
and p l > n 1, we have

1 U (r)
cu(r, ) U (r) l
.
W (B (Q))
p
Hence
1 2d1
d
d/2
U (r) dr
c
dj1 j uLp (Bd (Q)\Bd/2 (Q)) .
(13.1.6)
j=1
Using (l 1)p < n, we obtain

(j1)p
p
r
|j u| dx c
B2D (Q)
B2D (Q)
|l u|p dx,
1 j < l.
(13.1.7)
662
Therefore, the right-hand side in (13.1.6) does not exceed

c l uLp (B2D (Q)) c0 ( ).
If (2c0 )1 , then the required estimate follows. Let < (2 c0 )1 . Hence

U (r) dr > d/4,

d/2
and U (r0 ) >

that
1
2
for some r0 (d/2, d). Using (13.1.7) once more, we conclude

2D
|ru|p
d
B1 (Q)
dr
.
r
By the Holder inequality we have
p
1p
2D

rU (r)
p dr c log 2D
U
(r)
dr
r
d
r0
r0

1p
1p
2D
D
= c log
U (r0 )p 2p c log
.
d
d

2D
The result follows.

13.1.3 Capacity of a Bounded Cylinder
Proposition 1. Let C,d be the cylinder

x : (x , xn ) : |x | , |xn | d/2 ,
where 0 < 2 < d and Q2d = {x : |xi | < d}. Then
np l1

for n 1 > pl,
d
Cap C,d ,
Llp (Q2d )
d(log d )1p for n 1 = pl.
Proof. Let u M(C,d , Q2d ). Obviously,

|l u|p dx
Q2d
d/2
d/2
dxn
(n1)
|l u|p dx ,
(13.1.8)
Q2d
n1
1
where l = { l /x
= {x :
1 , . . . , xn1 }, 1 + + n1 = l, and Q2d
|xi | < d, i = 1, . . . , n 1}. The inner integral on the right in (13.1.8) exceeds
(n1) l (n1)
,
,
Lp B
Cap B
(n1)
where B
is the (n1)-dimensional ball {x : |x | < } and = 2(n1)1/2 d.
Hence from Propositions 13.1.2/1 and 13.1.2/2 it follows that the integral
under consideration majorizes c np l1 for n 1 > p l and c (log d/)1p for
(n1)
13.2 On (p, l)-Polar Sets
663
n1 = p l. Minimizing the left-hand side of (13.1.8) over the set M(C,d , Q2d ),
we obtain the required lower estimate for the capacity.
We proceed to the upper bound. Let x u(x ) be a smooth function
(n1)
with support in the ball Bd
that is equal to unity in a neighborhood
(n1)
of the ball B
. Further, we introduce the function d (x) = (x/d), where
1 , Q2 ). Since the function x d (x)u(x ) is in the class M(C,d , Q2d ),
M(Q
we have

l (ud )
p dx
Cap C,d ,
Llp (Q2d )
Q2d
cd
dpk

(n1)
Bd
k=0
c1 d
(n1)
|lk u|p dx
|l u|p dx .
Bd
Minimizing the latter integral and using Propositions 13.1.2/1 and 13.1.2/2,
we arrive at the required estimate.

13.1.4 Sets of Zero Capacity Cap(, Wpl )

The denition of the capacity Cap(,
Llp ()) and Proposition 13.1.1/2 imply
l
that Cap(e, Wp ) = 0 if and only if there exists a bounded open set containing e such that Cap(e,
Llp ()) = 0. The choice of is not essential by
Proposition 13.1.1/2.
From Corollary 13.1.1, for lp > n, p > 1 and for l n, p = 1, we obtain
that the equality Cap(e, Wpl ) = 0 is valid only if e = .
Proposition 13.1.1/3 shows that in any one of the cases n > lp, p > 1 or
Llp ) = 0 are equivalent.
n l, p = 1 the equalities Cap(e, Wpl ) = 0 and Cap(e,
Corollary 13.1.1 and Propositions 13.1.2/1 and 13.1.2/2 imply that no similar
property is true for n lp, p > 1. To be precise, Cap(e,
Llp ) = 0 for any
compactum e provided n lp, p > 1.
13.2 On (p, l)-Polar Sets

Let Wpl
denote the space of linear continuous functionals T : u (u, T ) on

l
Wp .
The set E Rn is called a (p, l)-polar set if zero is the only element in
l
Wp with support in E.
Theorem 1. The space D() is dense in Wpl if and only if C is a
(p, l)-polar set.
664
Proof. 1. Suppose D() is not dense in Wpl . Then there exists a nonzero
functional T Wpl
, equal to zero on D(), i.e., with support in C. (Here
we make use of the following corollary of the HahnBanach theorem. Let M
be a linear set in the Banach space B and let x0 be an element of B situated
at a positive distance from M . Then there exists a nonzero functional T B
such that (x, T ) = 0 for all x M .) Consequently, C is not a (p, l)-polar
set.
2. Suppose D() is dense in Wpl . For any functional T Wpl
with support
in C we have
(u, T ) = 0
for all u D(). Therefore, the last equality is valid for all u Wpl and so
T = 0. Thus, C is a (p, l)-polar set.

Theorem 2. The set E is (p, l)-polar if and only if Cap(E, Wpl ) = 0.
Proof. 1. Let Cap(E, Wpl ) = 0 and T Wpl
, supp T E. Without loss of
generality we may assume that supp T is a compactum (otherwise we could
take T with D instead of T ). We take an arbitrary D and a sequence
{um }m1 of functions in D which equal unity in a neighborhood of supp T
and tend to zero in Wpl . Since (1 um ) = 0 in a neighborhood of T , we have
(, T ) = (um , T ). The right-hand side converges to zero as m ; hence
(, T ) = 0 for all D. Since D is dense in Wpl , we conclude that T = 0.
2. Let E be a (p, l)-polar set. Then any compactum K in E is also a (p, l)polar set, and D(Rn \K) is dense in Wpl . Let v M(K). By the density of
D(Rn \K) in Wpl , there exists a sequence vm D(Rn \K) that converges to
v in Wpl . Every function vm v equals unity near K, has compact support,
and vm vWpl 0 as m . Therefore, Cap(K, Wpl ) = 0. The proof is
complete.

Taking into account the just proved assertion we can give an equivalent
formulation of Theorem 1.
Theorem 3. The space D() is dense in Wpl if and only if

Cap C, Wpl = 0.
13.3 Equivalence of Two Capacities

We compare the capacities Cap(e,
L1p ) and cap(e,
L1p ), p 1. Obviously, the
rst capacity majorizes the second. One can easily check that the converse
inequality also holds. In fact, for the function v = min{(1 )1 u, 1} there
Llp for an arbitrary number
exists a sequence in M(e) that converges to v in
(0, 1) and a function u N(e). Therefore,
13.3 Equivalence of Two Capacities

Cap e,
L1p
|v |p dx (1 )p
665

|u|p dx
L1p ). Thus the capacities Cap(e,

L1p ) and cap(e,
L1p )
and so Cap(e,
L1p ) cap(e,
coincide.
Since the truncation along the level surfaces does not keep functions in the
spaces
Llp and Wpl for l > 1, the previous argument is not applicable to the
proof of equivalence of the capacities Cap and cap, generated by these spaces.
Nevertheless, in the present section we show that the equivalence occurs for
p > 1.
Theorem 1. Let p > 1, n > p l, l = 1, 2, . . . . Then

cap e,
Llp Cap e,
Llp c cap e,
Llp
(13.3.1)
for any compactum e.

Proof. The left inequality results from the inclusion M(e) N(e). We
proceed to the upper bound for Cap. Let G denote a bounded open set such
that G e and

Ll cap e,
Llp + .
cap G,
p
(cf. Sect. 10.4.2). By
Let U be the (p, l)-capacitary potential of the set G
Proposition 10.4.2/2, the inequality U 1 is valid (p, l)-quasi-everywhere in
and therefore, quasi-everywhere in some neighborhood of the compactum e.
G,
We apply the mollication with radius m1 , m = 1, 2, . . . , to U and multiply
the result by the truncating function m , m (x) = (x/m) where C0 ,
(0) = 1. Thus we obtain a sequence of functions {Um }m1 in C0 such that
0 Um c in Rn , Um 1 in a neighborhood of e and

(13.3.2)
Llp + .
lim l Um pLp cap e,
m
(The inequality Um c follows from Proposition 10.4.2/1.)

We introduce the function
l+1

w = 1 (1 Um )+
,
which obviously has compact support and is in C l . Also,

l wLp l (1 Um )l+1 L .
p
Applying (1.8.7), we obtain

l wLp c l Um Lp .
It remains to make use of inequality (13.3.2).
The following theorem has a similar proof.
666
Theorem 2. Let p > 1, l = 1, 2, . . . . Then

cap e, Wpl Cap e, Wpl c cap e, Wpl
(13.3.3)
for any compactum e.

Corollary 1. Let p > 1, l = 1, 2, . . . , and let e be a closed subset of the
d = {x : 2|xi | d}. Then
cube Q

cap e,
Llp (Q2d ) Cap e,
Llp (Q2d ) c cap e,
Llp (Q2d ) .
(13.3.4)
Proof. It suces to derive (13.3.4) for d = 1. The left inequality is trivial;
the right one follows from Proposition 13.1.1/2 and Theorem 1.

Remark. The proofs of Theorems 1 and 2 and of Corollary 1 do not change

provided we replace the class M(e, ) in the denition of the capacity Cap
by the smaller one

P(e, ) = u C0 () : u = 1 in a neighborhood of e, 0 u 1 .
d.
Corollary 2. Let p > 1, l = 1, 2, . . . , and let e1 , e2 be compacta in Q
Then
2

l
Cap ei ,
Llp (Q2d ) ,
(13.3.5)
Cap e1 e2 , Lp (Q2d ) c
i=1
where c is a constant that depends only on n, p, and l.

Proof. Let ui P(ei , Q2d ), i = 1, 2, and let

l ui pLp c cap ei ,
Llp (Q2d ) +
(13.3.6)
(cf. preceding Remark). The function u = u1 + u2 is contained in C0 (Q2d )

and satises the inequality u 1 on e1 e2 . Hence from Corollary 1 we obtain

Cap e1 e2 ,
Llp (Q2d ) c cap e1 e2 ,
Llp (Q2d )
c l upLp
p1
l ui pLp ,
i=1
which together with (13.3.6) leads to (13.3.5).
13.4 Removable Singularities of l-Harmonic Functions

in Lm
2
Let F be a compactum in the open unit ball B. Here we show that the
capacity Cap enables one to characterize removable singularity sets for lharmonic functions in Lm
2 (B). For other results of such a kind see Mazya and
Havin [569].
13.4 Removable Singularities of l-Harmonic Functions in Lm

2
667
Proposition. Let l and m be integers, l m 0. Every function u

(B)
satisfying the equation
Lm
2
l u = 0
on B\F
is a solution of the same equation on B if and only if

Cap F,
L2lm
(B) = 0.
2
(13.4.1)
Proof. Suciency. Let (13.4.1) hold. For every v D(B\F )

l u v dx = 0.
B
Hence

{:||=m}
m!
!
Since the norm

D u D lm v dx = 0.
(13.4.2)
lm

v
Lm (B)
(13.4.3)
is equivalent to the norm in

L2lm
(B), it follows from Theorems 13.2/1 and
2
13.2/2 that D(B\F ) is dense in D(B) in the sense of the norm (13.4.3).
Therefore, using the inclusion u Lm
2 (B), we obtain (13.4.2) for all v D(B),
which is equivalent to l u = 0 on B.
(B) providing the inmum
Necessity. Let u denote the function in
L2lm
2
of the functional

2
(2l m)!
D dx
!
B
{:||=2lm}
dened on the set of in M(F, B). For every v D(B\F )

(2l m)!
D u D v dx = 0.
!
B
(13.4.4)
{:||=2lm}
Hence
2lm u = 0 on B\F,
which can be written as

l lm u = 0 on B\F.
(13.4.5)
The function lm u belongs to Lm

2 (B\F ). Furthermore, by (13.4.5) it is lharmonic on B\F and we deduce from the necessity assumption that

l lm u = 0 on B,
(B),
i.e., u is (2lm)-harmonic on B. Since u was chosen as a function in
L2lm
2
we see that u = 0 on B. Hence (13.4.1) holds.

668

Section 13.1. The polyharmonic capacity Cap(e,
Ll2 ()) was introduced by
the author [533]. The content of this section follows, to a large extent, the
authors paper [544].
Section 13.2. The denition of a (2, l)-polar set is borrowed from the
paper by Hormander and Lions [385]. The results of this section are due to
Littman [503]. For Theorem 13.2/2 see the earlier paper by Grushin [329].
Llp )
Section 13.3. The equivalence of the capacities Cap(e,
Llp ) and cap(e,
was established by the author [539, 542] for integer l. For fractional l the
equivalence of the corresponding capacities is proved by D.R. Adams and
Polking [19].
Section 13.4. This application of the polyharmonic capacity to description of removable singularities of polyharmonic functions was not published.
14
Integral Inequality for Functions on a Cube
Let Qd be an n-dimensional cube with edge length d and with sides parallel
to coordinate axes. Let p 1 and k, l be integers, 0 k l. We denote a
function in Wpl (Qd ), p 1, by u.
The inequality
uLq (Qd ) C
l
dj1 j uLp (Qd )
(14.0.1)
j=k+1
with q in the same interval as in the Sobolev embedding theorem often turns
out to be useful. This inequality occurs repeatedly in the following chapters.
Obviously, (14.0.1) is not valid for all u Wpl (Qd ), but it holds provided u is
subject to additional requirements.
In the present chapter we establish two-sided estimates for the best constant C in (14.0.1). In Sects. 14.1 and 14.2 we mainly consider the case of u
d and k = 0. The existence of C is equivvanishing near a compactum e Q
alent to the positiveness of the (p, l)-capacity of e. For a (p, l)-negligible set
e, upper and lower bounds for C are stated in terms of this capacity. If q p
and e is (p, l)-essential, i.e., its capacity is comparable with the capacity of
the cube, then C is estimated by the so-called (p, l)-inner diameter.
In Sect. 14.3 the function u is a priori contained in an arbitrary linear
subset C of the space Wpl (Qd ). There we present a generalization of the basic
theorem in Sect. 14.1 and give applications for particular classes C. In this
connection we have to introduce some functions of the class C that play a role
similar to that of the (p, l)-capacity.
In conclusion we note that the statements as well as the proofs of results in
the present chapter remain valid after replacing the cube Qd by an arbitrary
bounded Lipschitz domain with diameter d.

DOI 10.1007/978-3-642-15564-2 14,
669
670
14 Integral Inequality for Functions on a Cube
14.1 Connection Between the Best Constant and

Capacity (Case k = 1)
14.1.1 Denition of a (p, l)-Negligible Set
d . In either of the cases
Denition. Let e be a compact subset of the cube Q
n pl, p > 1 or n > l, p = 1 we say that e is a (p, l)-negligible subset of Qd if

(14.1.1)
Cap e,
Llp (Q2d ) dnpl ,
where is a suciently small constant that depends only on n, p, and l. For
the purposes of the present chapter we can take to be an arbitrary positive
number satisfying the inequality
4pn .
(14.1.2)
d . For
If (14.1.1) fails, then, by denition, e is a (p, l)-essential subset of Q
n < pl, p > l or for n l, p = 1 only the empty set is called (p, l)-negligible.
d will be denoted
The collection of all (p, l)-negligible subsets of the cube Q
by N (Qd ).
14.1.2 Main Theorem
Let u
Qd denote the mean value of u on the cube Qd , i.e.,

1
u
Qd = mn (Qd )
u dx.
Qd
In what follows, by c, c1 , and c2 we mean positive constants depending only

on n, p, l, k, and q.
We introduce the seminorm
u
p,l,Qd
l
djl j uLp (Qd ) .
j=1
d.
Theorem. Let e be a closed subset of the cube Q

1. For all u C (Qd ) with dist(supp u, e) > 0 the inequality
uLq (Qd ) C u
p,l,Qd ,
(14.1.3)
where q [1, pn(n pl)1 ] for n > pl, p 1, and q [1, ) for n = pl, p > 1,
holds. Moreover, the constant C admits the estimate

C p c1 dnp/q Cap e,
Llp (Q2d ) .
(14.1.4)
d ) with dist(supp u, e) > 0, let
2. For functions u C (Q
uLq (Qd/2 ) C u
p,l,Qd ,
(14.1.5)
where e N (Qd ) and q satises the same conditions as in item 1. Then

C p c2 dnp/q Cap e,
Llp (Q2d ) .
(14.1.6)
14.1 Connection Between the Best Constant and Capacity (Case k = 1)
671
For the proof of this theorem we need the two-sided estimate for Cap.
1 . There exists a constant c such that
Lemma. Let e be a compactum in Q

1 ), dist(supp u, e) > 0
c1 Cap e,
Llp (Q2 ) inf 1 upV l (Q1 ) : u C (Q
p

(14.1.7)
c Cap e,
Llp (Q2 ) .
Proof. To obtain the left estimate we need the following well-known assertion (cf. Theorem 1.1.17).
There exists a linear continuous mapping
d ) C k1,1 (Q
2d ),
A : C k1,1 (Q
k = 1, 2, . . . ,
d , (ii) if dist(supp v, e) > 0, then dist(supp Av, e) >

such that (i) Av = v on Q
0, and (iii)
i AvLp (Q2d ) ci vLp (Qd ) ,
i = 0, 1, . . . , l, 1 p .
(14.1.8)
Let v = A(1 u). Let denote a function in D(Q2 ) that is equal to unity
in a neighborhood of the cube Q1 . Then

l (v)
p dx cvp l
.
(14.1.9)
Cap(e, Q2 ) c
V (Q2 )
p
Q2
Now the left estimate in (14.1.7) follows from (14.1.8) and (14.1.9).
Next we derive the rightmost estimate in (14.1.7). Let w M(e, Q2 ). Then
wpV l (Q1 ) c
l
k wpLp (Q2 ) cl wpLp (Q2 ) .
k=0
Minimizing the last norm over the set M(e, Q2 ) we obtain

Llp (Q2 ) .
wpV l (Q1 ) c Cap e,
p
We complete the proof of the Lemma by minimizing the left-hand side.
Proof of the theorem. It suces to consider only the case d = 1 and then
use a similarity transformation.
1. Let N = uLp (Q1 ) . Since dist(supp u, e) > 0, it follows by the Lemma
that

p

p

Cap e,
Llp (Q2 ) c1 N 1 uV l (Q1 ) = cN p u pp,l,Q1 + c1 N 1 uLp (Q1 ) ,
p
i.e.,

Llp (Q2 ) c u
N p Cap e,
p
p,l,Q1
+ cN upLp (Q1 ) .
(14.1.10)
672
Without loss of generality, we may assume that u

Q1 0. Then
uQ1 Lp (Q1 ) u u
Q1 Lp (Q1 ) .
|N u
Q1 | = uLp (Q1 )
Consequently,
N uLp (Q1 ) N u
Q1 Lp (Q1 ) + u u
Q1 Lp (Q1 )
2u u
Q1 Lp (Q1 ) .
(14.1.11)
By (14.1.10), (14.1.11), and the Poincare inequality

u u
Q1 Lp (Q1 ) cuLp (Q1 )
we obtain

Cap e,
Llp (Q2 ) upLp (Q1 ) c u
p
p,l,Q1 .
From the Sobolev embedding theorem and the preceding inequality we conclude that

1
Llp (Q2 )
u pp,l,Q1 .
upLq (Q1 ) c u pp,l,Q1 + upLp (Q1 ) c 1 + Cap e,
Thus the rst item of the theorem follows.
2. For pl > n, p > 1 or for l n, p = 1 the assertion is trivial. Consider
the other values of p and l. Let M(e, Q2 ) be such that

Llp (Q2 ) + ,
(14.1.12)
l pLp (Q2 ) Cap e,
and let u = 1 . Applying the inequality
j Lp (Q2 ) cl Lp (Q2 ) ,
j = 1, . . . , l 1,
we obtain
u
p,l,Q1
p,l,Q1
cl Lp (Q2 ) .
Hence from (14.1.12) it follows that

1/p
Llp (Q2 ) +
.
uLp (Q1/2 ) cC Cap e,
By Holders inequality we have

1/p
Q1/2 cC Cap e,
Llp (Q2 )
.
1 Q1/2 = u
(14.1.13)
It remains to show that the mean value of on Q1/2 is small. Noting that

1
1
|w| dt
1
1
|t||w | dt
1
1
|w | dt
for any function w C 1 [1, 1] satisfying w(1) = w(1) = 0, we obtain
14.1 Connection Between the Best Constant and Capacity (Case k = 1)

2
dx
x1
2
dx
Q2
Q2
Q2 x1

l
(p1)n/p

l Lp (Q2 ) . (14.1.14)
l
dx 2
Q2 x1

dx
Q1/2
Therefore,
673
|| dx

1/p
Llp (Q2 ) +
.
Q1/2 2(2p1)n/p Cap e,
This and (14.1.1) and (14.1.2) imply

Q1/2 2n/p ,
which together with (14.1.13) completes the proof of the second part of the
theorem.

14.1.3 Variant of Theorem 14.1.2 and Its Corollaries
In the following theorem, which will be used in Chap. 16, we prove an assertion
similar to the rst part of Theorem 14.1.2 and relating it to a wider class of
functions. The parameters p, l, and q are the same as in Theorem 14.1.2.
d and let be a number in the
Theorem. Let e be a closed subset of Q
interval (0, 1). Then for all functions in the set

d) : u
u C (Q
Qd 0, u(x) dn/p uLp (Qd ) for all x e ,
inequality (14.1.3) is valid and

Llp (Q2d ) .
C p c(1 )p dnp/q cap e,
Proof. Duplicating the proof of Lemma 14.1.2, we obtain

1 ), u 0 on e
c1 cap e,
Llp (Q2 ) inf 1 upV l (Q1 ) : u C (Q
p

c cap e,
Llp (Q2 ) .
Further, we note that the inequality 1 N 1 u 1 on e implies

p

(1 )p cap e,
Llp (Q2 ) c1 N 1 uV l (Q2 )
p
and we follow the argument of the proof of the rst part of Theorem 14.1.2.
d . Then the inequality
Corollary 1. Let e be a closed subset of Q

upLq (Qd ) c dpn+np/q upLp (Qd )
dnp/q
l upLp (Qd )
+
cap(e,
Ll (Q2d ))
p
d ) that vanish on e.
is valid for all functions u C (Q

(14.1.15)
674
Proof. It suces to put d = 1. Let

P (u) =

x
(y)u(y) dy
Q1
0<l
be the polynomial in the generalized Poincar

e inequality for the cube Q1 (see
Lemma 1.1.11). Further let S(u) = P (u) Q1 0 (y)u(y) dy. Since all functions
are orthogonal to unity for || > 0, then
S(u)
cuL (Q ) .
(14.1.16)
p
1
It suces to obtain (14.1.15) under the assumption
uLp (Q1 ) uLp (Q1 ) ,
where = (n, p, l) is a small constant. Then the function v = u S(u)
satises the inequality
v(x)
cvL (Q )
p
1
on e. We can assume that vQ1 0. Applying the theorem in the present
section to the function v, we arrive at

c
j vpLp (Q1 ) .
cap(e,
Llp (Q2 ))
l
vpLq (Q1 )
j=1
Hence from Lemma 1.1.11 we obtain

u S(u)p
Lq (Q1 )

c
j u P (u) p
Lp (Q1 )
l
cap(e, L (Q2 ))
l
j=1
l upLp (Q1 ) .
cap(e, Llp (Q2 ))

Now a reference to (14.1.16) completes the proof.

Corollary 14.1.3 implies the following assertion.
d . The inequality
Corollary 2. Let e be a closed subset of Q

upLq (Qd ) c dpkn+np/q k upLp (Qd )
+
dnp/q
l upLp (Qd )
cap(e,
Llk+1
(Q
))
p
2d

(14.1.17)
d ) vanishing on e along with all derivatives

holds for all functions u C (Q
up to order k, k < l.
Proof. It suces to derive (14.1.17) for q = p and d = 1. By Corollary 1,
14.2 Connection Between Best Constant and the (p, l)-Inner Diameter

j upLp (Q1 ) c j+1 upLp (Q1 ) +
1
l upLp (Q1 )
cap(e,
Llj
(Q
))
p
1
675
for j = 0, 1, . . . , k 1. Therefore

k1

upLp (Q1 ) c k upLp (Q1 ) +
j=0
Since
1
cap(e, Llj
p (Q1 ))

l upLp (Q1 ) .

lk+1
cap e, Llj
(Q1 )
p (Q1 ) cap e, Lp
for j = 0, . . . , k 1, the result follows.
Remark 1. By Corollary 13.3/1, for p > 1 we can replace cap by Cap in

the statements of Theorem and Corollaries 1 and 2.
Remark 2. From Proposition 13.1.1/3 it follows that we can replace cap(e,
Llp (Q2d )) by cap(e,

Llp (Rn )) in the statements of the Theorem and Corollary 1
for n > lp. A similar remark applies to Corollary 2 in the case n > p(lk +1).
Remark 3. Proposition 10.4.2/2 and the properties of (p, l)-rened functions (see 10.4.4) imply that in the denition of the capacity cap(E, hlp ) of a
Borel set we can minimize the norm uhlp over all (p, l)-rened functions in
hlp satisfying the inequality u(x) 1 for (p, l)-quasi-every x E. Therefore, in
the theorem of the present subsection, we can deal with (p, l)-rened functions
in Vpl (Qd ) for which the inequality
u(x) dn/p uLp (Qd )
d . Similarly, in Corolis valid (p, l)-quasi everywhere on the Borel set E Q
lary 1 we can consider a Borel set E Qd and (p, l)-rened functions in

Vpl (Qd ) equal to zero quasi-everywhere on E. The class of functions in Corollary 2 can also be enlarged if we consider the class Ck (E) of rened functions
u Vpl (Qd ) such that D u(x) = 0 for (p, l ||)-quasi all x E and for all
multi-indices of order || k.
14.2 Connection Between Best Constant and the

(p, l)-Inner Diameter (Case k = 1)
14.2.1 Set Function lp,q (G)
Denition. With any open set G Qd we associate the number
lp,q (G) = inf
p
p,l,Qd
,
p
uLp (Qd )
676
d ) vanishing
where p 1 and the inmum is taken over all functions u C (Q
in a neighborhood of Qd \G.
d , then
By Theorem 14.1.2, if Qd \G is a (p, l)-negligible subset of Q

Llp (Q2d ) .
lp,q (G) dnp/q Cap Qd \G,
This relation fails without the condition of smallness on Cap(Qd \G,
Llp (Q2d )).
In fact, if G is small, then the value lp,q (G) becomes large (for instance, we
can easily check that lp,q (G) nplnp/q provided G is a cube with small
edge length ) whereas

cap Qd \G,
Llp (Q2d ) cdnpl .
In the present section we give a description of the set function lp,q (G) for
q p 1 in certain new terms connected with the (p, l)-capacity under the
/ N (Qd ).
condition that Qd \G
14.2.2 Denition of the (p, l)-Inner Diameter
We x the cube Qd and we denote by Q an arbitrary cube in Qd with edge
length and with sides parallel to those of Qd .
Denition. Let G be an open subset of Qd . The supremum of for which
the set {Q : Q \G N (Q )} is not empty will be called the (p, l)-inner
(cubic) diameter of G relative to Qd and denoted by Dp,l (G, Qd ). In the case
Qd = Rn we shall use the notation Dp,l (G) and call it the (p, l)-inner (cubic)
diameter of G. Obviously, Dp,l (G, Qd ) = d provided Qd \G is a (p, l)-negligible
d.
subset of Q
Let n < pl, p > 1 or n = l, p = 1. By denition, for such p and l, all
the sets except the empty set are (p, l)-essential. Therefore, for any open set
G Qd , the (p, l)-inner (cubic) diameter Dp,l (G, Qd ) coincides with the inner
(cubic) diameter D(G), i.e., with the supremum of edge lengths of cubes Q
inscribed in G.
14.2.3 Estimates for the Best Constant in (14.1.3) by the
(p, l)-Inner Diameter
The following theorem contains two-sided estimates for lp,q (G) for q p 1.
Theorem 1. Let G be an open subset of Qd such that Qd \G is a (p, l)d.
essential subset of Q
d ) vanishing in a neighborhood of Qd \G
1. For all functions u C (Q
inequality (14.1.3) is valid with q p 1 and
ln(p1 q 1 )

.
C c1 Dp,l (G, Qd )
(14.2.1)
14.2 Connection Between Best Constant and the (p, l)-Inner Diameter
677
d ) that vanish in a neighborhood of the set

2. If for all functions u C (Q
Qd \G inequality (14.1.3) holds, then

ln(p1 q1 )
C c2 Dp,l (G, Qd )
.
(14.2.2)
Proof. 1. Assume for the moment that Dp,l (G, Qd ) < d. We denote an
arbitrary number in (Dp,l (G, Qd ), d] by . The denition of the (p, l)-inner
, the set e = Q \G is a (p, l)-essential
diameter implies that, for any cube Q
subset, i.e.,

Llp (Q2 ) > npl .
(14.2.3)
cap Q e,
(Here and in what follows Qc is the open cube with edge length c whose
center coincides with that of Q and whose sides are parallel to the sides of
Q .) In the case Dp,l (G, Qd ) = d we put = d. Then (14.2.3) is also valid
= Q
d.
since, by hypothesis, e is a (p, l)-essential subset of the cube Q
According to the rst part of Theorem 14.1.2 and inequality (14.2.3), we
have
upLq (Q )
c np/q
u
cap(Q e,
Llp (Q2 ))
c lpn(1p/q) u
p
p,l,Q
p
p,l,Q .
(14.2.4)
We construct a covering of Qd by cubes Q whose multiplicity does not exceed

some number that depends only on n. Next we sum (14.2.4) over all cubes of
the covering. Then
upLp (Qd ) c lp
l
p(jl) j upLp (Qd ) .
(14.2.5)
j=1
Using a well-known multiplicative inequality, we obtain

l
j/l

1j/l
j vLp (Qd ) cvLp (Qd )
dil i vLp (Qd )
(14.2.6)
i=0
(cf. Lemma 1.4.7). Putting v = u u

Qd in (14.2.6) and applying the Poincare
inequality
u u
Qd Lp (Qd ) cduLp (Qd ) ,
we obtain
1j/l
j uLp (Qd ) cuLp (Qd ) u
j/l
p,l,Qd .
Hence from (14.2.5) with q = p we obtain

1c
l

j=1
u p,l,Qd
uLp (Qd )
l
pj/l
.
678
Therefore,
uLp (Qd ) c l u
p,l,Qd ,
(14.2.7)
and the rst part of the theorem follows for q = p. Let q > p. Summing the
inequality

upLp (Q ) c lpn(1p/q) l upLp (Qd ) + pl upLp (Qd )

over
all cubes of the covering {Q } and making use of the inequality ( ai )
ai , where ai > 0, 0 < < 1, we conclude that

upLp (Qd ) c lpn(1p/q) l upLp (Qd ) + l upLp (Qd ) .
It remains to apply inequality (14.2.7).
be a cube having a (p, l)-negligible
2. Let 0 < < Dp,l (G, Qd ) and let Q
intersection with Qd \G. Let denote a function in C (Q ) that vanishes
near Q , is equal to unity on the cube Q/2 , and satises |j | c j ,
) that vanishes near
j = 1, 2, . . . . If v is an arbitrary function in C (Q
Qd \G then the function u = v extended by zero on the exterior of Q
satises (14.1.5) by the hypothesis of the theorem. Therefore
vLp (Q/2 ) C
l

j (v)
j=1
cC
Lp (Q )
cC
j
l
kj k vLp (Q )
j=1 k=0
p,l,Q

+ vLp (Q ) .
l
This and the estimate

1
1
vLp (Q ) c vLp (Q ) + c n(p q ) vLq (Q/2 )
yield
vLp (Q/2 ) c C
p,l,Q
+ l+n(p
1
q 1 )

vLq (Q/2 ) .
(14.2.8)
We may assume that 2c C l+n(p q ) < 1 since the reverse inequality is
the required inequality (14.2.3). Then by (14.2.8)
vLq (Q/2 ) 2c C v
p,l,Q ,
and (14.2.2) follows from the second part of Theorem 14.1.2 applied to the

cube Q . The proof is complete.
In each of the cases pl > n, p > 1 and l = n, p = 1, Theorem 1 can be
stated in terms of the inner diameter D(G). To be precise, we formulate the
following statement.
Theorem 2. Let G be an arbitrary open subset of Qd , G = Qd and let the
numbers n, p, and l satisfy either of the conditions pl > n, p > 1 or l = n,
p = 1. Further, let C be the best constant in (14.1.3) with q [p, ). Then
1
C D(G)ln(p
q 1 )
(14.2.9)
14.3 Estimates for the Best Constant C in the General Case
679
14.3 Estimates for the Best Constant C in the General

Case
Let C denote an arbitrary linear subset of the space Wpl (Qd ). Our goal is the
study of the inequality
l
uLq (Qd ) C
djl j uLp (Qd ) ,
(14.3.1)
j=k+1
where u C and q, p, and l are the same as in Theorem 14.1.2. The norm on
the right in (14.3.1) can be replaced by an equivalent one retaining only the
summands corresponding to j = l and j = k + 1.
14.3.1 Necessary and Sucient Condition for Validity of the Basic
Inequality
be the closure of C in the metric of the space Vpl (Qd ) and let Pk be the
Let C
set of polynomials of degree k l 1, normalized by

||p dx = 1.
(14.3.2)
dn
Qd
= .
Theorem. The inequality (14.3.1) holds if and only if Pk C
Proof. The necessity of this condition is obvious. We will prove the suciency.
= then in C
we can introduce the norm
If Pk C
u
l
dj1 j uLp (Qd ) ,
j=k+1
into
which makes it a Banach space. Let I be the identity mapping from C
l
Lp (Qd ). Since C Vp (Qd ) Lq (Qd ), we see that I is dened on C. We will

show that it is closed. Let |um |C 0 and um uLq (Qd ) 0 as m .
Then there exists a sequence of polynomials {m }m1 of a degree not higher
than k such that um m 0 in Lq (Qd ). Consequently, u = lim m in the
= then u = 0. Thus I is closed. Now from
space Vpl (Qd ) and since Pk C
the Banach theorem it follows that E is continuous, that is, (14.3.1) holds.

Example. Consider the class Cr (E) (r = 0, . . . , l 1, E is a Borel subset
d ) of (p, l j)-rened functions u Vpl (Qd ), p > 1, such that j u = 0
of Q
(p, l j)-quasi-everywhere on E, j = 0, . . . , r.
Since any sequence of (p, l)-rened functions that converges in Vpl (Qd ) contains a subsequence that converges (p, l)-quasi-everywhere (cf. Sect. 10.4.4),
it follows that Cr (E) is a closed subset of Vpl (Qd ).
680
Thus, by the Theorem, inequality (14.3.1) is valid for all u Cr (E) if and
only if Pk does not contain a polynomial such that j = 0 (p, l j)quasi-everywhere on E, j = 0, . . . , r.
14.3.2 Polynomial Capacities of Function Classes
Let be a polynomial in Pk and let

cap C, ,
Llp (Q2d ) = inf

|l u|p dx,
Q2d
where the inmum is taken over all functions u

Llp (Q2d ) such that the
restriction of u to Qd is contained in a linear subset C of the space

Vpl (Qd ).
With C we associate l capacities

Llp (Q2d ) =
inf
cap C, ,
Llp (Q2d ) ,
CAPk C,
{:Pk }
k = 0, . . . , l 1. In other words,

CAPk C,
Llp (Q2d ) = inf
{,u}

|l u|p dx,
(14.3.3)
Q2d
where the inmum is taken over all pairs {, u} such that Pk , u|Q d C
and u
Llp (Q2d ).
Llp (Q2d )) does not increase as k increases. The
It is clear that CAPk (C,
following inequality holds:

(14.3.4)
CAPk C,
Llp (Q2d ) cdnpl .
1 , Q2 ) and let d (x) = (x/d). Since 1 Pk and the
In fact, let M(Q
restriction of the function 1d to Qd equals zero, the pair {1, d } is admissible
for the problem (14.3.3). This implies (14.3.4).
We introduce the norm
uVpl (Qd ) =
l
djl j uLp (Qd ) .
j=0
The next assertion is similar to Lemma 14.1.2.

Lemma. The capacity CAPk (C,
Llp (Qd )) is equivalent to the following capacity of the class C:
(14.3.5)
inf upV l (Qd ) ,
p
where the inmum is taken over all pairs {, u} such that Pk , u

Vpl (Qd ), u C.
681
Proof. We have

Llp (Q2d )
CAPk C,

l d ( Au)
p dx,
Q2d
where A is the extension operator in Lemma 14.1.2. Obviously, the right-hand

side does not exceed c AupV l (Qd ) . From (14.1.8) it follows that A = .
p
Therefore, using (14.1.8) once more, we obtain

Llp (Q2d ) c upV l (Qd ) .
CAPk C,
p
Minimizing the right-hand side, we arrive at the required upper bound for
CAPk .
We now prove the lower estimate. Since ( u)|Qd can be extended to a
function in
Llp (Q2d ), the classes of admissible functions in the denitions of
both capacities under consideration are simultaneously empty or nonempty.
Llp (Q2d ), ( v)|Q d C. Then the capacity (14.3.5) does
Let Pk , v
not exceed
vVpl (Qd ) cl vLp (Q2d ) .

14.3.3 Estimates for the Best Constant C in the Basic Inequality
From Theorem 14.3.1 it follows that (14.3.1) holds if and only if

Llp (Q2d ) > 0.
CAPk C,
The next theorem yields two-sided estimates for the best constant C in (14.3.1)
expressed in terms of the capacity CAPk (C,
Llp (Q2d )).
Theorem. 1. If CAPk (C,
Llp (Q2d )) > 0 then, for all u C, the inequality (14.3.1) holds with

1/p
Llp (Q2d )
.
C cdn/q CAPk C,
(14.3.6)
2. If (14.3.1) holds for all u C and if

Llp (Q2d ) c0 dnpl ,
CAPk C,
where c0 is a small enough constant that depends only on n, p, l, and k, then

1/p
C cdn/q CAPk C,
Llp (Q2d )
.
(14.3.7)
Proof. 1. Let u C be normalized by

uLp (Qd ) = dn/p ,
(14.3.8)
682
and let be any polynomial in Pk . According to Lemma 14.3.2 we have

k

1/p
Llp (Q2d )
c
dil i ( u)L
CAPk C,
p (Qd )
i=0
+c
l
dil i uLp (Qd ) .
(14.3.9)
i=k+1
Hence from the inequality

i vLp (Qd ) cdli l vLp (Qd ) + cdi vLp (Qd ) ,
we obtain that the rst sum in (14.3.9) does not exceed
cdl uLp (Qd ) + cl uLp (Qd ) .
Therefore

1/p
CAPk C,
Llp (Q2d )
cdl uLp (Qd ) + c
l
dil i uLp (Qd ) .
(14.3.10)
i=k+1
For each u Vpl (Qd ) there exists a polynomial of degree less than k + 1
such that
(14.3.11)
uLp (Qd ) c dk+1 k+1 uLp (Qd ) .
First suppose that
k+1 uLp (Qd ) > (2c )1 dn/pk1 .
Then, by (14.3.4) we have

1/p
Llp (Q2d )
CAPk C,
cdkl+1 k+1 uLp (Qd ) .
Now let
(14.3.12)
k+1 uLp (Qd ) (2c )1 dn/pk1 .

uLp (Qd ) 21 dn/p = 21 uLp (Qd )
and consequently
21 dn/p Lp (Qd ) 3 21 dn/p .
We put
= dn/p 1
Lp (Qd ) .
(14.3.13)
683
Then (14.3.13) implies

uLp (Qd ) 2 dn/p Lp (Qd ) uL
p (Qd )
Obviously, the right-hand side does not exceed
2 uLp (Qd ) + 2uLp (Qd )

dn/p Lp (Qd ) 1
= 2 uL (Q ) + 2
L (Q ) uL (Q )
4 uL
p
p (Qd )

uLp (Qd ) 4c dk+1 k+1 uLp (Qd ) ,
which together with (14.3.10) and (14.3.12) implies the estimate

l

1/p

c
dil i uLp (Qd )
Llp (Q2d )
CAPk C,
(14.3.14)
i=k+1
for all u C, normalized by (14.3.8).

From the Sobolev embedding theorem and (14.3.4) we obtain
1
q 1 )
k+1 uLp (Qd ) + cdn(q p ) uLp (Qd )

1/p
1
Llp (Q2d )
cdkl+nq +1 CAPk C,
k+1 uLp (Qd )
uLp (Qd ) cdk+1+n(p

+ cdn(q
p1 )
uLp (Qd ) ,
which together with (14.3.14) yields (14.3.1) with the constant C satisfying (14.3.6).
Llp (Q2d ), (
2. Let be an arbitrary positive number, Pk ,
)|Qd C and let

Llp (Q2d ) + dnlp .
|l |p dx CAPk C,
Q2d
d is contained in C, by the hypothesis of

Since the restriction of to Q
the theorem we have
Lq (Qd ) C
l
djl j Lp (Qd ) .
j=k+1
The right-hand side does not exceed

1/p

Llp (Q2d ) + dnlp
cCl Lp (Qd ) cC CAPk C,
because
Llp (Q2d ). Similarly,
(14.3.15)
684
Lq (Qd ) cdl+n(q
p1 )
l Lp (Q2d ) c(c0 + )1/p dn/p .
Thus
Lq (Qd ) Lq (Qd ) c(c0 + )1/p dn/q
and by (14.3.15)

1/p
. (14.3.16)
Lq (Qd ) c(c0 +)1/p dn/q +cC CAPk C,
Llp (Q2d ) +dnlp
Now we note that
1
Lp (Qd ) cdn(p
q 1 )
Lq (Qd ) .
The preceding estimate follows from the Holder inequality for p q. In the
case p > q it results as follows:

1
1
Lp (Qd ) c dk+1 k+1 Lp (Qd ) + dn(p q ) Lq (Qd )
= cdn(p
q 1 )
Lq (Qd ) .
Since Pk , we have Lp (Qd ) = dn/p . Therefore Lq (Qd ) cdn/q .

Using the smallness of the constant c0 , we arrive at (14.3.7). The theorem is
proved.

14.3.4 Class C0 (e) and Capacity Capk (e,
Llp (Q2d ))
The rest of the section deals with the class

d ) : dist(supp u, e) > 0 ,
C0 (e) = u C (Q
(14.3.17)
d.
where e is compact subset of the cube Q
We introduce the following set function:

Llp (Q2d ) = inf inf
Capk e,
(14.3.18)
Pk {f }
|l f |p dx,
Q2d
where p 1 and {f } is a collection of functions in

Llp (Q2d ) such that f =
in a neighborhood of e where Pk .
Since P0 = {1}, we have

Cap0 e,
Llp (Q2d ) = Cap e,
Llp (Q2d ) .
Llp (Q2d )) and CAPk (C0 (e),
Llp (Q2d )) are
We show that the capacities Capk (e,
equivalent.
Lemma. The following inequalities are valid:

CAPk C0 (e),
Llp (Q2d )
Llp (Q2d ) Capk e,

Ll (Q2d ) .
c CAPk C0 (e),
p
685
Proof. The upper estimate for the CAPk inequality is an obvious corollary
of the denitions of the two capacities.
We shall prove the lower estimate. Let Pk , u C0 (e), let A be the
extension operator in Lemma 14.1.2, and let d be the function used in the
proof of inequality (14.3.4). From property (ii) of the operator A it follows
that d ( Au) is contained in the class {f } introduced in the denition of
Capk (e,
Llp (Q2d )). Therefore
p

Llp (Q2d ) d ( Au)Ll (Q
Capk e,
p
2d )
c AupV l (Q2d ) .
p
Taking into account the equality A = and the estimate (14.1.8) for the
function v = u we complete the proof by reference to Lemma 14.3.2.
From Theorem 14.3.1, applied to the class C0 (e), and from the preceding lemma there immediately follows an assertion that coincides with Theorem 14.1.2 for k = 0.
Corollary. 1. If Capk (e,
Llp (Q2d )) > 0, then, for all u C0 (e), the inequality (14.3.1) holds and

1/p
Llp (Q2d )
.
C cdn/p Capk e,
2. If (14.3.1) holds for all u C0 (e) and if

Capk e,
where c0 is small enough constant that depends only on n, p, and l, then

1/p
Llp (Q2d )
.
C cdn/q Capk e,
14.3.5 Lower Bound for Capk

We derive a lower bound for the capacity Capk (e,
Llp (Q2d )) by the capacity
lk
Cap(e,
Lp (Q2d )).
Proposition. The following inequality holds:

Lplk (Q2d ) .
Llp (Q2d ) cdkp Cap e,
Capk e,
Proof. It suces to consider the case d = 1. From the inequality
l vLp (Q2 ) clk vLp (Q2 ) ,
we obtain
v
Llp (Q2 ),
(14.3.19)
686

Cap e,
Llp (Q2 ) c Cap e,
Lplk (Q2 ) .
(14.3.20)
Llp (Q2d ) such that f = in a

Let Pk and let f be a function in
neighborhood of e. Obviously, the dierence /xi f /xi is contained in
C0 (e) for all i = 1, . . . , n. For some i, let
/xi Lp (Q2 ) ,
(14.3.21)
where is a positive number (that depends only on k, l, or n) which will be

specied later. Then
p

f
p

Lpl1 (Q2 ) .
l f Lp (Q2 ) l1
p Capk1 e,
(14.3.22)

xi Lp (Q2 )
If for all i = 1, . . . , n the inequality (14.3.21) fails, the condition Lp (Q1 ) = 1
implies
(x)
1
c, x Q2 .
We can take = (2c)1 . Since = f on e, we have |f (x)| 12 , x e, and
hence

Llp (Q2 ) .
l f pLp (Q2 ) 2p Cap e,
The preceding result and (14.3.22) yield

Llp (Q2 ) c min Capk1 e,
Lpl1 (Q2 ) , Cap e,
Llp (Q2 ) .
Capk e,
Applying (14.3.20), we complete the proof.
We present an example of a set for which

L22 (Q2 ) = 0 and Cap e,
Cap1 e,
L22 (Q2 ) > 0.
Example. Let n = 3, p = 2, l = 2, and let e be the center of the cube
Q1 = {x : |xi | < 12 }. Since by the Sobolev embedding theorem

u(e)
2 c
|2 u|2 dx
Q2
for all u D(Q2 ), it follows that Cap(e,

L22 (Q2 )) c1 > 0.
L22 (Q2 )) = 0. Let = 2 3x. Obviously, P1 .

We show that Cap1 (e,
, Q2 ) such that |j | cj . The function
Let denote a function in M(Q
coincides with in a neighborhood of e. Hence

2
2 ( )
2 dx.
Cap1 e, L2 (Q2 )
Q2
L22 (Q2 )) = 0.
On the other hand, the last integral is O(). Thus Cap1 (e,
687
Remark. In connection with the previous example consider the quadratic

forms
2
2

3
3
2u
u
+
S1 (u, u) =
dx,
xi
Q1 i,j=1 xi xj
i,j=1
2
3

2u
dx,
S2 (u, u) =
Q1 i,j=1 xi xj

1 ) that vanish near the center of the cube

dened on functions u C (Q
Q1 . The forms generate the operators 2 and 2 with the Neumann
boundary data on Q1 and with the complementary condition u = 0 at the
point e. Corollary 14.3.4 and the above example imply that the rst operator
is positive denite and that the second is not.
In general, for p = 2, the basic results of the present section can be reformulated as necessary and sucient conditions for positive deniteness and as
two-sided estimates for the rst eigenvalue of the elliptic operator generated
by the quadratic form S(u, u). This form is given on a linear subset C of the
space V2l (Q1 ) and satises the coerciveness condition
c1
l
l
j u2L2 (Q1 ) S(u, u) c2
j=k+1
j u2L2 (Q1 )
j=k+1
for all u C.
14.3.6 Estimates for the Best Constant in the Case of Small
(p, l)-Inner Diameter
Here we show that the best constant in (14.3.1) (for q p 1 and C = C0 (e))
d \e provided this
is equivalent to some power of the (p, l)-inner diameter of Q
diameter is small.
Lemma. Let G be an open subset of the cube Qd such that
Dp,l (G, Qd ) c0 d,
(14.3.23)
where c0 is a small enough constant that depends only on n, p, and l. Then, for
d ) vanishing in a neighborhood of Q
d \G, the inequality
all functions u C (Q

lj
j uLp (Qd ) c Dp,l (G, Qd )
l uLp (Qd )
(14.3.24)
holds, where j = 0, 1, . . . , l 1.
Proof. It suces to assume that d = 1 and l > 1. We put D = Dp,l (G, Q1 ).
Since < 1, it follows that Qd \G
/ N (Q1 ). Therefore, by Theorems 14.2.3/1
and 14.2.3/2, we have
688
uLp (Q1 ) cDl
l
j uLp (Q1 ) .
j=1

j uLp (Q1 ) c l uLp (Q1 ) + uLp (Q1 ) ,
we obtain that

uLp (Q1 ) cD l l uLp (Q1 ) + uLp (Q1 ) .
Thus, (14.3.24) follows for j = 0.
To obtain the estimate for j uLp (Q1 ) with j 1 we can use the inequality

j/l
(lj)/l
uLp (Q1 ) .
j uLp (Q1 ) c l uLp (Q1 ) + uLp (Q1 )
Theorem. Let q be the same number as in Theorem 14.1.2 and let con d ) that vanish in a
dition (14.3.23) hold. Then for all functions u C (Q
neighborhood of Qd \G the inequality

uLq (Qd ) C
l
dj1 j uLp (Qd ) ,
(14.3.25)
j=k+1
where k = 0, 1, . . . , l 1 holds. The best constant in (14.3.25) satises the

inequalities
ln(p1 q1 )
ln(p1 q1 )

C c2 Dp,l (G, Qd )
. (14.3.26)
c1 Dp,l (G, Qd )
(In the case n < pl, p > 1 or n = l, p = 1 the value Dp,l (G, Qd ) can be
replaced by inner diameter D(G, Qd ) in (14.3.25) and (14.3.26).)
Proof. The right estimate in (14.3.26) follows from (14.2.1) and the previous lemma, and the left estimate is contained in the second part of Theorem 14.2.3/1 and in Theorem 14.2.3/2.

Remark. The smallness of the (p, l)-inner diameter is crucial for the validity
of the Theorem. In fact, let G be the cube Q1 R3 with center excluded.
Then d = 1, D(G) = 12 , whereas, according to Remark 14.3.5, the inequality
uL2 (Q1 ) C2 uL2 (Q1 )
(14.3.27)
is not true. (This can be seen directly by insertion of the function u(x) =
x1 (x/), where > 0, = 0 on B1 (0), = 1 outside B2 (0) into (14.3.27).)
689
14.3.7 A Logarithmic Sobolev Inequality with Application to the

Uniqueness Theorem for Analytic Functions in the Class
L1p (U )
From the inequality (14.1.3) we can deduce an estimate for the integral of the
logarithm of the modulus of a function in L1p that characterizes the smallness
of the set of zeros of this function that suces for this integral to be nite
(see Mazya and Havin [569]).
Let E be Borel set in Rn1 = {x = (x , xn ) Rn : xn = 0} and let {B}
be a collection of n-dimensional open balls with centers in Rn1 . We denote
the concentric ball with doubled radius by 2B. Let

s(B) = mn1 Rn1 B ,

1 (2B) ,
c(E B) = cap E B, W
2

s(B).
S=
{B}
We denote the number of dierent balls B that contain a point x by .

Lemma. Let be a (p, l)-rened function in the class L1p ( B) that vanishes on E G. Then

1
log
(x )
(x ) dx
S (BRn1 )
1
s(B)
s(B) log
pS
c(E B)
{B}

1
c
||p dx ,
(14.3.28)
+ log
p
S B
where c = c(n, p), p 1.
Proof. If (X , , ) is a measure space with nite and f is a nonnegative
function dened on X and measurable with respect to the -algebra , then
for any p > 0

1/p

1
1
p
exp
log f d
f d
.
(14.3.29)
(X ) X
(X ) X
Therefore
1
s(B)

1
1
p

log || dx log
|| dx ,
p
s(B) BRn1
BRn1
where the ball B is arbitrary. Now we note that, for p 1,

1
||p dx
s(B) BRn1

1
1
c
||p dx +
||p dx
p1 (2B)) B
mn (B) B
cap(B, W
690
and use (14.1.3) and Remark 14.1.3/3. We have

1
c
p
|| dx
||p dx.
mn (B) B
c(E B) B

c
1
p

|| dx
||p dx
s(B) BRn1
c(E B) B
and consequently,

log || dx

(BRn1 )

=
log || dx
Then
{B}
BRn1

S(B)
c
1
1
p
s(B) log
s(B) log
+
|| dx .
p
c(E B)
p
s(B) B
{B}
{B}
Applying (14.3.29) to the last sum we arrive at the required estimate.
By virtue of (14.3.28) we can prove the uniqueness theorem for analytic

functions of the class Llp in the unit disk U (see the paper by Havin and the
author [569], where this problem is considered in detail and where a bibliography is given).
Let A be the set of all functions analytic in the disk U and let X A. We
say that a set E, contained in the interval (0, 2), is the uniqueness set for
X if each function f X with limr10 f (rei ) = 0 for any E vanishes
identically on U .
Let E be a Borel set in the interval (0, 2) and let {} be the set of
pairwise, disjoint open intervals (0, 2). Denote the length of by l(),
the disk with diameter by B, and the concentric disk with doubled radius
1 (2B)).
by 2B. We also use the notation c(E ) = cap(E , W
p
Let E satisfy the condition

l()
=
(14.3.30)
l() log
c(E )
and let an analytic function f L1p (U ) satisfy limr10 f (rei ) = 0 for any
E.

log lim
f (rei )
d =
0
r10
(see Havin and Mazya [569]) which together with the well-known uniqueness
theorem for analytic functions of the Hardy class H 1 shows that f (z) = 0 for
all z U . Thus E is the uniqueness set for L1p (U ).
Since c(E ) 1 for p > 2, then in this case we can omit c(E )
in (14.3.30). For p < 2 we can replace cap(E , Wp1 (R2 )) by cap(E ).
691

Section 14.1. Theorem 14.1.2 is due to the author [533] for p = 2 (see
also Mazya [544] for the case p 1). Similar results were rediscovered by
Donoghue [240], Polking [665], and R.A. Adams [23].
Inequality (14.1.17) and the particular case (14.1.15) of it, which renes
the rst part of Theorem 14.1.2, was established by Hedberg [368] for p > 1
who used a dierent method for which the restriction p = 1 is important. In
10.1.3 these results follow from Theorem 14.1.3, which was implicitly proved
in the authors paper [544]. We note that our proof is also valid for the case
p = 1.
Inequality (14.1.17) plays an important role in the aforementioned paper
by Hedberg where the well-known problem of spectral synthesis in Sobolev
spaces is solved (see [369] for the history of this problem). The basic result of
Hedberg runs as follows.
Let u Wpl (Rn ), p > 1, and let m be a positive integer. Let K be a
closed subset of Rn and D u|K = 0 for all with 0 || l 1. Then
l (Rn \K), i.e., there exist functions um C (Rn \K) such that
uW
p
0
lim u um Wpl (Rn ) = 0.
An obvious corollary to this theorem is the following uniqueness theorem

for the Dirichlet problem (see Hedberg [368]).
Let Rn be an open bounded set in Rn and let u be a solution of
l u = 0 in in the space W2m (Rn ) satisfying D u| = 0, 0 || l 1.
Then u = 0 in .
The existence of the spectral synthesis in the Sobolev space Wpl (Rn ) was
proved by a dierent method by Netrusov [634] who also solved the problem
for a wide class of function spaces including Besov and Lizorkin-Triebel spaces
[632]. The above-mentioned results by Netrusov with complete proofs can be
found in the books by D.R. Adams and Hedberg [15] and by Hedberg and
Netrusov [371].
pl () was given by Swanson
The following characterization of the space W
and Ziemer [736].
l ().
Theorem. Let l be a positive integer, let 1 < p < , and let f W
p
l () if and only if
If Rn is an arbitrary open set, then f W
p

k f (y)
dy = 0
lim
r0 r n B(x,r)
for (l k)-quasi-every x Rn \ and for all integer k, 0 k l 1.
Section 14.2. A set function similar to l2,2 was introduced and applied
to the investigation of the uniqueness conditions for the solution of the rst
boundary value problem by Kondratiev [448, 449]. In these papers it is called
692
n
the capacity Cl,d
. The connection of lp,q with the (p, l)-inner diameter was
studied by the author [546].
Section 14.3. The results of this section (except Proposition 14.3.5 and
those of Sect. 10.3.6) are borrowed from the authors paper [544].
Proposition 14.3.5 was published in the authors book [555]. This proposition together with Corollary 14.3.4 shows that the constant C in Corollary 14.3.4 satises the inequality
1/p

,
Llk
C cdk+n/q Cap e,
p (Q2d )
which was obtained earlier by a direct method in the paper by Hedberg [367]
(compare with the stronger inequality (14.1.17) that was discussed in the
comments to Sect. 14.1).
In connection with the content of the present chapter we mention the paper
by Meyers [598] in which the inequality
u LuWpk () Ck+1 uWplk1 () ,
u Wpl (),
where L is a projection mapping Wpl () Pk and is a Lipschitz domain,

is studied.
A certain family of polynomial (p, l)-capacities similar to cap(C0 (e), ,
Llp (Q2d )) was used by Bagby [61] in the study of approximation in Lp by solutions of elliptic equations. Nystrom obtained a lower estimate for CAPl1 by
the Bessel (p, l)-capacity assuming that the set e preserves Markovs inequality
[642].
We conclude this chapter by mentioning the problem of characterization
of sets of uniqueness for Sobolev spaces, by considering the particular case
of Wpl , p > 1. By the set of uniqueness we mean a set E Rn enjoying the
following property. From the condition u Wpl , u(x) = 0 for all x Rn \E,
outside a suitable subset of zero (p, l)-capacity follows that u = 0.
A description of the set of uniqueness in Wpl in the Theorem below was
1/2
given by Hedberg [366]. The rst result of this kind, for the space W2
the circle, was obtained in 1950 by Ahlfors and Beurling [31].
on
Theorem. Let E be a Borel subset of Rn and let cl,p stand for the capacity
generated by the space Wpl . The following conditions are equivalent:
(i) E is the uniqueness set for Wpl ;
(ii) cl,p (G\E) = cl,p (G) for each open set G;
(iii) for almost all x
lim sup n cp,l (B
\E) > 0.
If lp > n then E is a uniqueness set for Wpl if and only if it does not have
inner points.
15
Embedding of the Space
Llp () into Other
Function Spaces
15.1 Preliminaries
If n > pl, p > 1, or n l, p = 1, then by Sobolevs theorem the mapping
Llp () u u Lq (), where q = pn/(n pl), is continuous. We can

easily show that this operator is one to one. In fact, let zero be the image
of u
Llp () in Lq () and let a sequence {um }m1 of functions in D()
converge to u in Llp (). Then for all multi-indices with || = l and for all
D()

l
lim
D um dx = lim (1)
um D dx = 0.
m
Since the sequence D um converges in Lp (), it tends to zero.

The above considerations show that each element of
Llp () (for n > p l,
p > 1 or n l, p = 1) can be identied with a function in Lq () and the
identity mapping
Llp () u u Lq ()
is one to one, linear, and continuous (i.e., it is a topological embedding).
If n lp, p > 1 or n < l, p = 1 and if
uLq () Cl uLp ()
for some q > 0 with a constant C independent of u D(), then we arrive
at the same conclusion. However, for these values of n, l, and p the above inequality for the norms does not hold for all domains and moreover, in general,
Llp () is not necessarily embedded into the space of distributions D ().

The theorems of the present chapter contain necessary and sucient conditions for the topological embedding of
Llp () into D (), Lq (, loc), and
Lq () (Sects. 15.115.5). In Sect. 15.6 we obtain criteria for the compactness
of the embedding
Llp () Lq (). In Sect. 15.7 results of the previous sections are applied to the study of the rst boundary value problem for linear
elliptic equations of order 2 l, and in Sect. 15.8 we consider applications to the
Dirichlet and Neumann problems for quasilinear elliptic equations.
DOI 10.1007/978-3-642-15564-2 15,
693
694

Llp () into Other Function Spaces
15.2 Embedding
Llp() D ()
The aim of the present section is to prove the following assertion.
Theorem. The space
Llp () (1 p < ) is topologically embedded into
D () if and only if any one of the following conditions holds:
1. n > pl, p > 1 or n l, p = 1;
2. C is not (p, n/p)-polar if n = p l, p > 1;
3. C is not empty if n < p l and n/p is not integer ;
4. C is not (p, n/p)-polar or it is not contained in an (n1)-dimensional
hyperplane if n < pl and n/p is integer.

Taking into account Theorem 13.2/2, we may read the set of zero inner
(p, n/p)-capacity for the (p, n/p)-polar set in the last theorem.
15.2.1 Auxiliary Assertions
Lemma 1. The space
Llp () is topologically embedded into D () if and only
if, for any D(), the functional

u dx (u, ) D ()
(15.2.1)
D() u
is continuous with respect to the norm l uLp () .

Proof. 1. If
Llp D (), then any sequence um D () that is a Cauchy
sequence in the norm l uLp () converges in D (). Consequently, the functional (u, ) is continuous.
2. Let the functional (u, ) be continuous in the norm l uLp () . Then
the mapping (15.2.1) can be continuously extended to
Llp (). It remains to
show that the resulting mapping is one to one. Let um be a Cauchy sequence
in
Llp () that converges to zero in D (). Obviously, for all D() and
all multi-indices with || = l we have

l
D um dx = (1)
um D dx 0.
Since the sequence D um converges in Lp (), it tends to zero.
The just-proved lemma can be reformulated as follows.

Corollary. The space
Llp () is embedded into D () if and only if

(u, ) Kl uL (Rn )
(15.2.2)
p
for any D() and for all u D().
Lemma 2. The space
Llp is embedded into D () if and only if for any
D() there exists a distribution T D () supported in C such that
15.2 Embedding
Llp () D ()

(u, T ) Kl uL (Rn ) ,
p
u D(Rn ).
695
(15.2.3)
Proof. Necessity. Let

Llp () D (). Then by the previous corollary
inequality (15.2.2) is valid for all u D(). The space D() can be identied
with the subspace D(Rn ) by zero extension to C. By the HahnBanach
theorem, functional (15.2.1) can be extended to a functional u (u, s) on
D(Rn ) that is continuous with respect to the norm l uLp (Rn ) , i.e., to a
functional satisfying

(u, s) Kl uL (Rn ) ,
(15.2.4)
p
where K does not depend on u. Since s = on , the support of the distribution T = s is contained in C. The estimates (15.2.3) and (15.2.4) are
equivalent.
Suciency. Suppose (15.2.3) is valid. Then the functional D(Rn ) u
(u, s) with s = T is continuous with respect to the norm l uLp (Rn )
on D(Rn ) and hence on D() where it coincides with (15.2.1). It remains to
refer to the Corollary.

Lemma 3. Let n p l for p > 1 or n < l for p = 1. If for any D()
n
there exists a distribution T Wpl
(R ) with compact support in C such
that

(15.2.5)
x , T = 0
for all multi-indices with || k = [l n/p], then the space
Llp () is

embedded into D ().
Proof. Let B be the ball {x : |x| < } containing the supports of and T .
We show that (15.2.3) is valid for all u D(Rn ). Indeed, for any polynomial
P of a degree not higher than k we have

(u, T ) = (u P, T ) = (u P ), T ,
where D(B2 ), = 1 in B . Let K = T W l
n . Then
(R )
p

(u, T ) K (u P )
Wpl (Rn )

C l uLp (Rn ) + u P Lp (B2 ) .
Applying the inequality

inf u P Lp (B2 ) Ck+1 uLp (B2 ) ,
Pk
where Pk is the set of all polynomials of a degree not higher than k, as well
as the inequalities
k+1 uLp (B2 ) Ck+1 uLq (B2 ) cl uLp (Rn ) ,
where q = pn[n p(l k 1)]1 , we arrive at (15.2.3). The lemma is proved.
696

15.2.2 Case = Rn
Lemma. Let n lp for p > 1 or n < l for p = 1 and let be a multi-index
of order || 1 n/p. Then there exists a sequence of functions u D(Rn )
such that u x in D (Rn ) and u 0 in
Llp (Rn ).
Proof. Let be an innitely dierentiable function on (0, ) that is equal
to unity in a neighborhood of [0, 1] and to zero in a neighborhood of [2, )
and let || < l n/p. On the one hand, clearly, the sequence

u (x) = x 1 |x|
converges to x in D (). On the other hand,
l u Lp (Rn ) const n/p+||l 0.

Let || = l n/p. We put
v (x) =
1
2
log
log
|x|
for x B 2 \B = {x : 2 > |x| } ( > 2). By (t) we denote a function

in C [0, 1] equal to zero near t = 0 and to unity near t = 1. Further, let
w (x) = [v (x)] for x B 2 \B , w (x) = 1 in B and w (x) = 0 outside
B 2 . On the one hand, it is clear that u (x) = x w (x) converges to x in
D (Rn ). On the other hand, since

j w (x) const(log )1 |x|j
for j > 0, we have
l u Lp (Rn ) const(log )1
= const(log )

B 2 \B
1/p1
1/p
|x|(||l)p dx
0.

Theorem. The space
Llp (Rn ) is embedded into D (Rn ) if and only if either
n > lp, p > 1, or n l, p = 1.
The necessity follows immediately from the Lemma and the suciency
results from the estimate
uLpn/(nlp) () c(n, l, p)l uLp ()
and the proof of the one-to-one correspondence presented in Sect. 15.1.
Corollary. The space

Llp () is embedded into D () for any open set
if and only if n > lp, p > 1 or n l, p = 1.
15.2 Embedding
Llp () D ()
697
Proof. The necessity was proved in the preceding theorem. Extending any
u D() by zero to C, we obtain the embedding D() D(Rn ) and hence
the embedding D (Rn ) D (). Therefore,
Llp (Rn ) D (Rn ) D ().

Llp ()
Here and henceforth means a topological embedding.
15.2.3 Case n = p l, p > 1

Lemma. If a closed set E is not a (p, l)-polar set, then there exists a distrin
bution T Wpl
(R ) with compact support in E such that (1, T ) = 1.
n
Proof. Since E is not (p, l)-polar, there exists a distribution S Wpl
(R ),
S = 0 with support in E. Let the function T = S satisfy (, S) = 1.
Obviously, the distribution T = S satises the requirements of the lemma.
Theorem. Let n = lp, p > 1. In order that

Llp () D () it is necessary
and sucient that C is not a (p, n/p)-polar set.
Proof. Necessity. From Lemma 15.2.2 it follows that, if the function
D() with (1, ) = 0 is arbitrary, then the inequality

(u, ) Cl uL (Rn )
p
cannot be valid for all u D(Rn ). Since
Llp () D (), there exists a

n
distribution T D (R ) such that supp T C and

(u, T ) cl uL (Rn )
p
for all u D(Rn ) (see Lemma 15.2.1/2). Obviously, T = 0. Moreover,

(u, T ) cl uL (Rn ) + (u, ) cuW l (Rn ) ,
p
p
n
i.e., T Wpl
(R ). Thus C is not a (p, n/p)-polar set.
Suciency. Since C is not a (p, n/p)-polar set, there exists a distribun

tion T0 Wpl
(R ) with compact support in C such that (1, T0 ) = 1. Let
D(). We put T = (, 1)T0 . Since (1, T ) = 0, then, according to
Lemma 15.2.1/3 (where k = 0),
Llp () D (). The theorem is proved.
15.2.4 Case n < p l and Noninteger n/p
Theorem. If n < pl and if n/p is noninteger then the condition C = is
necessary and sucient for the embedding
Llp () D().
698

Proof. Necessity. If C = , then by Theorem 15.2.2 the space

Llp () is

not embedded into D ().
Suciency. Let C = . We may assume 0 C. We put
T =
(1)
||k
(, x )
D (x),
!
where is the Dirac delta-function, D is the usual gradient, k = [l n/p],

n
and D(). Since k < l n/p, we have D Wpl
(R ) for || k and
n
T Wpl
(R ). Besides, obviously, (x , T ) = 0 for || k. It remains to
make use of Lemma 15.2.1/3. The theorem is proved.

15.2.5 Case n < pl, 1 < p < , and Integer n/p
Theorem. Let n/p be an integer, n < pl, 1 < p < . The space
Llp () is

embedded into D () if and only if C is not a (p, n/p)-polar set and is not
contained in a (n 1)-dimensional hyperplane.
Proof. We put k = l n/p.
Suciency. (a) Suppose the C is not a (p, n/p)-polar set. Then by
(Rn ) with compact supLemma 15.2.3 there exists a distribution T0 Wpkl

port in C such that (1, T0 ) = 1. We put

T =
a (1)|| D T0 .
||k
n
Obviously, T is a distribution in Wpl
(R ) with compact support in C.
It remains to show that given any D() we can nd numbers a so
that (15.2.5) holds for all multi-indices with || k. In other words, the
linear algebraic system

a D T0 , x = , x , || k,
||k
must be solvable. Since (D T0 , 1) = 0 for || > 0, this system can be rewritten

in the form

!
T0 , x = , x .
(15.2.6)
a
( )!
The matrix of system (15.2.6) is triangular with nonzero elements on the main
diagonal (it consists of the numbers !(T0 , 1) = !). Thus (15.2.6) is solvable.
(b) Suppose C is not in an (n 1)-dimensional hyperplane. Suppose also
that the points 0, a1 , . . . , an are situated in C and are anely independent.
Further, let D(). We introduce the distribution
15.2 Embedding
Llp () D ()
T = P0 (i)(x) +
699
Pj (i)(x aj ),
j=1
where P0 , P1 , . . . , Pn are homogeneous polynomials of degrees not higher than

n
k 1. Since p(l k + 1) > n, we have T Wpl
(R ).
We choose polynomials P0 , P1 , . . . , Pn so that (15.2.5) holds. Let Q()
and R() be the sums of terms of a degree not higher than k in the Taylor
expansions of the Fourier transforms T() and (),

respectively. We denote
the homogeneous part of R() of degree k by r(). It is clear that
T() = P0 () +

Pj () exp i aj , ,
j=1
Q() =
Pj () i
j=0
aj , Pj ().
j=1
Since the forms aj , are independent, we can choose P1 , . . . , Pn so that

i
aj , Pj () = r().
j=1
We dene the polynomial P0 by P0 = (P1 + + Pn ) + R r (obviously, the

degree of P0 is less than k). Thus Q() = R(), which is equivalent to (15.2.5).
Necessity. Let
Llp () D () and let C Rn1 (for deniteness we
n1
= {x : x1 = 0}). We show that C is not a (p, n/p)-polar set. By
put R
Lemma 15.2.1/2 for any D() there exists a distribution T with support
in C such that (15.2.3) holds. Since C Rn1 , it follows that
j
l
l

1
T =
(x1 ) Sj (x2 , . . . , xn ) =
Tj ,
i x1
j=0
j=0
where Sj D (Rn1 ), supp Sj C (cf. L. Schwartz [695]). We show that

n
Tj = 0 and Tj Wpl
(R ). We have
T() =
1j Sj (2 , . . . , n ) =
j=0
Tj ().
j=0
Let 0 , . . . , l be distinct numbers in (0, 1). According to the Lagrange inter

l
polation formula there exist constants 0 , . . . , l such that ak = j=0 j P (j )
for any polynomial P (t) = a0 + a1 t + + al tl . In particular, if P (t) = tk ,
then
l

1=
j jk .
(15.2.7)
j=0
700

For the polynomial

T(t1 , 2 , . . . , n ) =
1j Sj (2 , . . . , n )tj ,
j=0
we have
1k Sk (2 , . . . , n )
j T (j 1 , 2 , . . . , n ).
j=0
Consequently,

Tk =
1
i x1
k
(x1 ) Sk (x2 , . . . , xn )

l

j
x1
=
T
, x2 , . . . , xn .
j
j=0 j
We dene the function
k =

l

x1
j
, x2 , . . . , xn .
j
j=0 j
Inequality (15.2.3) for T implies

(u, k Tk ) Kl uL (Rn )
p
(15.2.8)
(15.2.9)
for all u D(Rn ). We may assume that the function D(Rn ) satises
(, xk1 ) = 1. This along with (15.2.7) and (15.2.8) yields (k , xk1 ) = 1. Suppose
that Tk = 0. Then for all u D(Rn )

(u, k ) Kl uL (Rn ) .
p
Hence by this and Lemma 15.2.2 we obtain (k , xk1 ) = 0. Thus Tk = 0, or
n
equivalently, Sk = 0. Since D(Rn ), (15.2.9) implies Tk Wpl
(R ).
(Rn ). It is well known (see, for
Next we show that (x1 ) Sk Wpkl

instance, Stein [724], 4, Chap. 6) that for any collection of functions gj
Wplj (Rn ) (j = 0, 1, . . . , l 1) there exists a function Wpl (Rn ) such that
j /xj = gj for x Rn1 and
Wpl (Rn ) K
l1

gj Wplj (Rn ) .
j=0
Let gj = 0 for j = k and gk = u, where u is an arbitrary function in

Wplk (Rn ). Then
15.3 Embedding
Llp () Lq (, loc)
u, (x1 ) Sk =

,
1
i x1
k
701

(x1 ) Sk
= (, Tk ).
n
Since Tk Wpl
(R ), applying the above assertion, we get

u, (x1 ) Sk KW l (Rn ) Ku lk n .
Wp (R )
p
n/p
Thus (x1 )Sk Wp

is proved.
(Rn ) and C is not a (p, n/p)-polar set. The theorem

Thus, the proof of Theorem 15.2 is complete.
15.3 Embedding
Llp() Lq (, loc)
The following assertion shows that Theorem 15.2 contains the necessary and
sucient conditions for the embedding
Llp () Lq (, loc).
Theorem. Let n p l for p > 1 or n < l for p = 1. If the space
Llp ()
is embedded into D (), then it is also embedded into Lq (, loc) for n = p l,
where q is any positive number, and into C() for n < pl.
Proof. Let G be a domain in Rn with compact closure and smooth boundary G = . First we show that there exists a family of functions
D(G ), || = l 1, such that the matrix ( , x ) is not degenerate. Let be any function in D(G ) with (, 1) = 0 and let = D .
Obviously, for > we have

, x = (1)|| , D x = 0,
and the matrix ( , x ) is triangular. Since the main diagonal terms are
(1)|| !(, 1) = 0, the determinant is not zero and the existence of the
functions follows.
Since
Ll () D (), by Corollary 15.2.1 we have
p

( , u) Kl uL
p ()
(15.3.1)
where K is a constant independent of u. By the Sobolev embedding theorem

uLq (G) K l uLp (G) +
( , u)
||=l1
for all u D(), which together with (15.3.1) yields

uLq (G) Kl uLp () .

702
In the same way as in Sect. 15.1 we can prove that the mapping
Llp () u

u Lq (, loc) is one to one. The theorem is proved.
Next we present two corollaries to Theorem 14.1.2 that complement the
theorem of this section.
Corollary 1. Let n = p l, p > 1, and let Qd be a cube for which
d \,
Llp (Q2d )) > 0.
Cap(Q
Then, for all u D(),
upLq (Qd ) Cl upLp () ,
(15.3.2)
d \,
where C c dnp/q [Cap(Q
Llp (Q2d ))]1 .
Proof. By Theorem 14.1.2,
1

d \,
Llp (Q2d )
u
upLq (Qd ) c dnp/q Cap Q
p
p,l,Q2d .
(15.3.3)
Since
j uLp (Q2d ) c d lj j uLqj () c d lj l uLq ()
for j 1, qj = pn[n p(l j)]1 , we have
u
p,l,Q2d
c l uLp () ,
which together with (15.3.3) yields (15.3.2).

n/p
d \,
Lp (Q2d )) >
Corollary 2. 1. If p l > n, n/p is an integer, and Cap(Q
0, then, for all u D(),
max |u|p Cl upLp () ,
d
Q
(15.3.4)
d \,
Lp (Q2d ))]1 .
where C c dlpn [Cap(Q
d \ = , then (15.3.4) is valid
2. If p l > n, n/p is not an integer, and Q
for all u D() with C c dlpn .
n/p
Proof. 1. Let kp = n, q > n. Using the Sobolev theorem, we obtain

max |u| c d lk1 max |lk1 u| c d lkn/q lk uLq (Qd ) .
d
Q
d
Q
(15.3.5)
By (15.3.2), with l replaced by k and u replaced by lk u, we have

1/p
d \,
lk uLp (Qd ) c dn/p Cap Q
Lkp (Q2d )
l uLp () .
This and (15.3.5) imply (15.3.4).
15.4 Embedding
Llp () Lq () (the Case p q)
703
2. Let j be an integer such that l p1 n < j < l p1 n + 1. Since

p(l j) < n, then
j uLq () Cl uLp () ,
where q = pn[n (l j)p]1 . The condition p(l j + 1) > n is equivalent to
q > n. Therefore
max |j1 u| cd1q
d
Q
j uLq (Qd ) cdp
+lj+1
l uLp () .
Hence, from this and

max |u| c dj1 max |j1 u|
d
Q
d
Q
we obtain (15.3.4). The corollary is proved.
15.4 Embedding
Llp() Lq () (the Case p q)
In this section we nd the necessary and sucient conditions for the validity
of the inequality
(15.4.1)
uLq () Cl uLp () ,
where u is an arbitrary function in D(). The results we present here can be
deduced (although only for p > 1) from the more general theorems proved
in Section 16.2 where the space
Llp (, ) is studied. However, the separate
exposition seems reasonable because of the importance of this particular case,
the possibility of including p = 1 and the simpler statements.
15.4.1 A Condition in Terms of the (p, l)-Inner Diameter
If we put d = in the proofs of Theorems 14.2.3/1 and 14.2.3/2, then we
arrive at the following theorem.
Theorem. Let q satisfy any one of the conditions:
(i) q [p, np(n p l)1 ] if p 1, n > pl;
(ii) q [p, ) if p > 1, n = pl;
(iii) q [p, ] if pl > n, p > 1 or l = n, p = 1.
Then:
The inequality (15.4.1) holds if and only if
() Dp,l () < for n > pl, p 1 or n = pl, p > 1;
() D() < for n < pl, p > 1 or n = l, p = 1.
The best constant C in (15.4.1) satises

1
1
[Dp,l ()]ln(p q ) in case (),
C
1
1
in case ().
[D()]ln(p q )
(15.4.2)

704
15.4.2 A Condition in Terms of Capacity

By the Sobolev embedding theorem, the inequality (15.4.1) is valid for any
set provided q = pn(n pl)l , n > pl, or q = , p = 1, l = n. Therefore it
remains to consider only the cases q < pn(n pl)1 for n = pl and q for
n < pl, p 1.
We present a necessary and sucient condition for the validity of (15.4.1)
stated in dierent terms and resulting from Theorem 14.1.2/1.
Theorem 1. Let q be the same as in Theorem 15.4.1/1. The inequality (15.4.1) is valid if and only if

d \,
Llp (Q2d ) > 0
(15.4.3)
inf Cap Q
for some d > 0. Here the inmum is taken over all cubes Qd with edge length
d and with sides parallel to coordinate axes.
Proof. Suciency. We construct the cubic grid with edge length d. Suppose
d \,
Llp (Q2d )) > 0
dlpn Cap(Q
for any cube of the grid. By Theorem 14.1.2, we have
upLp (Qd ) c 1
dpk k upLp (Qd ) .
k=1
Summing over all cubes of the grid, we obtain

upLp () c 1
dpk k upLp () .
k=1
To estimate the right-hand side we use the inequality

k/l
1k/l
k uLp () cl uLp () uLp () .
(15.4.4)
Then
upLp () c
l

dpl l/k l upLp ()
k/l
p(1k/l)
uLp () .
(15.4.5)
k=1
Since

d \,
Llp (Q2d ) cdnpl ,
Cap Q
it follows that c and hence the right-hand side in (15.4.5) does not exceed
21 upLp () + cdpl l l upLp () .
Thus
15.4 Embedding
Llp () Lq () (the Case p q)
upLp () cdpl l l upLp () .
705
(15.4.6)
We now prove (15.4.2) for q > p. By Theorem 14.1.2

upLp (Qd )
cd
np/q

d
|j u|p dx.
pjn
Qd
j=1
Summing over all cubes Qd and using the inequality

ai
ai , ai 0, 0 < 1,
i
we obtain
upLq () c 1 dnp/q
l

j=1

|j u|p dx.
dpjn
Now (15.4.4) yields

upLq () c 1 dnp/q
pj/l
p(1j/l)
dpjn l uLp () uLp () .
j=1
Applying (15.4.6), we nally obtain

upLp () c l dnp/q+pln l upLq () .
(15.4.7)
Necessity. Let Qd be an arbitrary cube with edge length d and let u be an

d ) such that dist(Q
d \, supp u) > 0. Replacing u
arbitrary function in C (Q
to be a function in (15.4.1), where D(Qd ), = 1 on Qd/2 , |j | c dj ,
we obtain
uLp (Qd/2 ) Cl (u)Lp (Qd ) .
Hence
uLq (Qd/2 ) c C
djl j uLp (Qd ) .
j=0
Applying the well-known inequality

uLp (Qd ) cduLp (Qd ) + cuLp (Qd/2 )
and the H
older inequality we obtain
uLq (Qd/2 ) c C
djl j uLp (Qd ) + c Cdl+n(qp)/pq uLq (Qd/2 ) .
j=1
Thus for d, so large that 2c C dl+n(qp)/pq < 1, we have

706
uLq (Qd/2 ) c C
dj1 j uLp (Qd ) .
j=1
Since
uLq (Qd ) cdn(pq)/pq+1 uLp (Qd ) + cuLq (Qd/2 ) ,
we obtain
l

djl j uLp (Qd ) .
uLq (Qd ) c C + dl+n(pq)/pq
j=1
By the second part of Theorem 14.1.2, we have either

d \,
Cap Q
Llp (Q2d ) dnpl ,
where satises inequality (14.1.2), or

p
d \,
Llp (Q2d ) cdnp/q C + dl+n(pq)/pq
.
Cap Q

Theorem 1 can be rephrased as follows.
Theorem 2. One of the following conditions is necessary and sucient

for the validity of (15.4.1),
1. For some d > 0,

d \,
Llp Rn > 0 if n > pl.
inf Cap Q
Qd
2. For some d > 0,

d \,
inf Cap Q
Llp (Q2d ) > 0
Qd
if n = pl, p > 1.
3. The domain does not contain arbitrarily large cubes if (i) n < pl,
p > 1, (ii) n l, p = 1, (iii) n = p l and C is connected.
Proof. Part 1 follows from Theorem 1 and Proposition 13.1.1/3, part 2 is
contained in Theorem 1. For n < pl the condition

d \,
Llp (Q2d ) cdnpl
Cap Q
d \ = , which proves part 3 for n = pl, p > 1.
is equivalent to Q
Let n = pl, p > 1 and let C be connected. If contains arbitrarily
large cubes, then obviously, (15.4.3) does not hold. Suppose that cubes of an
arbitrary size cannot be placed in and that the number d0 is so large that
any cube Qd with d > d0 has a nonempty intersection with . Then in Rn \
there exists a continuum that contains points in Qd and in Rn \Q2d . It remains
to apply Proposition 13.1.2/1. This concludes the proof.
15.5 Embedding
Llp () Lq () (the Case p > q 1)
707
We present an example of an unbounded domain for which the hypotheses

of the Theorem are valid.
(i)
Example. In each cube Q1 , i 1, of the coordinate grid with edge length

1 we select a closed subset ei lying in an s-dimensional plane s > n pl 0.
Let inf ms ei const > 0.

We denote the complement of the set i ei by . By Proposition 13.1.1/4
we have

(i)
Llp (Q2 ) const > 0
Cap ei ,
(i)
for any cube Q2 . Hence (15.4.1) holds for .
15.5 Embedding
Llp() Lq () (the Case p > q 1)
15.5.1 Denitions and Lemmas
We continue to study inequality (15.4.1). Here we obtain a necessary and
sucient condition for q [1, p). Contrary to the case q p considered
previously, this condition does not depend on q. It means that up to a small
error the set is the union of cubes Q (i) with a nite multiplicity of the
intersection and with edge lengths {di }i1 satisfying
n+lpq/(pq)
di
< .
(15.5.1)
i=1
The smallness is described in terms of the capacity

l
Capl1 e, Lp (Q2d ) = inf inf

|l u|p dx
Pl1 {u}
Q2d
introduced in 14.3.4. We recall that Pl1 is the set of polynomials of a degree

not higher than l 1 normalized by the equality

dn
||p dx = 1,
Qd
and {u} is the set of functions in

Llp (Q2d ), equal to polynomials Pl1 in
d.
a neighborhood of a compactum e Q
We shall use the following assertion that is a particular case (k = l 1) of
Corollary 14.3.4.
d with
Lemma 1. 1. Let e be a compact subset of the cube Q

Llp (Q2d ) > 0.
Capl1 e,
Then
708

uLq (Qd ) Al uLp (Qd )

for any function u C (Qd ) that vanishes in a neighborhood of e. Here
1 q pn(n pl)1 for n > pl; 1 q < for n = pl, 1 q for
n < pl, and
1/p

.
A cdn/q Capl1 e,
Llp (Q2d )
d ) that vanishes in a neigh2. If (15.4.1) holds for any function u C (Q
borhood of the compactum e Qd and if

Capl1 e,
where c0 is a small enough constant, then

1/p
Llp (Q2d )
.
A cdn/p Capl1 e,
Denition 1. Let be a suciently small constant depending only on n,
d is said to be (p, l, l 1)-negligible if
p, l. A compact subset e of a cube Q

(15.5.2)
Llp (Q2d ) < dnpl .
Capl1 e,
Otherwise e is called (p, l, l 1)-essential.
d is denoted
The collection of (p, l, l 1)-negligible subsets of the cube Q
by Nl1 (Qd ).
Lemma 2. Let 1 q p and let (15.4.1) hold for any u C0 (). Then
there exists a constant c that depends only on n, p, q, and l, and is such that
Qd \ is a (p, l, l 1)-essential subset of Qd for any cube Qd with edge length
d satisfying
d cC pq/(n(pq)+lpq) .
d ) with dist(supp u, Q
d \) > 0. Let
Proof. Consider a function u C (Q
C0 (Q1 ), = 1 on Q1/2 and let d = (x/d).

The insertion of the function ud into (15.4.1) yields
uLq (Qd/2 ) ud Lq (Qd ) Cl (ud )Lp (Qd )

cC l uLp (Qd ) + dl uLp (Qd ) .
uLp (Qd ) cdl l uLp (Qd ) + cdn(qp)/pq uLq (Qd/2 )
(see, for instance, Lemma 1.1.11) we obtain

uLq (Qd/2 ) c0 C l uLp (Qd ) + dn(qp)/pql uLq (Qd/2 ) .
(15.5.3)
15.5 Embedding
709
Consequently,
uLq (Qd/2 ) 2c0 Cl uLp (Qd )
for 2c0 Cdn(qp)/pql < 1. On the other hand, (15.5.3) and the Holder inequality imply
uLp (Qd ) dn(pq)/pq uLp (Qd )

c dn(pq)/pq+l l uLp (Qd ) + uLq (Qd/2 ) .
Therefore

uLq (Qd ) c1 dn(pq)/pq+l + C l uLp (Qd )
for 2c0 Cdn(pq)/pql < 1. Suppose, in addition, that dn(pq)/pq+l > C. Then
uLq (Qd ) 2c1 dn(pq)/pq+l l uLp (Qd ) .
(15.5.4)
d \
If Q
/ Nl1 (Qd ), we have nothing to prove. Otherwise, by part 2 of
Lemma 1, inequality (15.5.4) implies

1/p
d \,
Llp (Q2d )
2c1 dn(pq)/pq+l ,
cdn/q Capl1 Q
or equivalently,
Capl1

d \,
Q
Llp (Q2d )
c
2c1
p
dnpl .
d \
We may always assume that (c/(2 c1 ))p . Therefore Q
/ Nl1 (Qd ).

Denition 2. The cube QD = QD (x) with center x is called critical
if

d \ Nl1 (Qd ) .
D = sup d : Q
Lemma 2 implies the following assertion.
Corollary. Let 1 q < p and let (15.4.1) hold for any u C0 (). Then
for any x there exists a critical cube QD (x).
In what follows Q(i) is an open cube with edges parallel to the coordinate
axes and with edge lengths di , i = 1, 2, . . . . Further, let cQ (i) be a concentric
cube with edge length cdi and with sides parallel to those of the cube Q (i) .
Let denote a positive constant that depends only on n.
Denition 3. A covering {Q (i) }i1 of a set is in the class Cl,p,q if:
1. P (i) \ Nl1 (P (i) ), where P (i) = Q (i) ;
2. P (i) P (j) = for i = j;
3. the multiplicity of the covering {Q (i) } does not exceed a constant that
depends only on n;
4. Q (i) \
/ Nl1 (Q (i) );
5. the series (15.5.1) converges.
710

15.5.2 Basic Result

Theorem. Let 1 q < p. The inequality (15.4.1) holds for all u
Llp () if
and only if there exists a covering of in Cl,p,q .
Proof. Necessity. Let x , Qd = Qd (x) and let D be the edge length of
the critical cube centered at x. We put

d \,
Llp (Q2d ) .
g(d) = dpln Capl1 Q
Let d denote a number in the interval [D, 2D] such that g(d) where is
the constant in (15.5.2). Further, let M be the collection of cubes {Qd }x .
We show that series (15.5.1) converges for any sequence {Q(i) } of disjoint
cubes in M . By Lemma 15.5.1/1, given an arbitrary number i > 0, there
exists a function vi C (Q (i) ) with dist(supp vi , Q (i) \) > 0 such that

(i)

l
(i) \,
|l vi |p dx cdn
Cap
Q
L
|vi |p dx.
2Q
+
i
p
l1
i
Q(i)
Q(i)
We assume that i = dpl

. Then, by (15.5.2),
i

pl
p
|l vi | dx cdi
Q(i)
Q(i)
|vi |p dx.
(15.5.5)
Estimating the right-hand side by the inequality

n(pq)/pq
vi Lp (Q(i) ) c dli l vi Lp (Q(i) ) + cdi
vi Lp ( 12 Q(i) )
(15.5.6)
(see Lemma 1.1.11) and using the smallness of the constant , we arrive at
the estimate

n(pq)/qpl
|l vi |p dx cdi
vi pL ( 1 Q(i) ) .
(15.5.7)
q 2
Q(i)
Let i D(Q (i) ), i = 1 in 12 Q (i) , |k i | cdk

i , k = 1, 2, . . . . We introduce
the function ui = i vi . It is clear that
l ui Lp (Q(i) ) c
dkl k vi Lp (Q(i) )
k=0

c l vi Lp (Q(i) ) + dl vi Lp (Q(i) ) .
Applying (15.5.6), we obtain
l ui Lp (Q(i) ) cl vi Lp (Q(i) ) + cdn(qp)/pql vi Lp ( 12 Q(i) ) .
n(qp)/pql
l ui Lp (Q(i) ) cdi
ui Lq (Q(i) ) .
(15.5.8)
15.5 Embedding
711
By the hypothesis of the theorem, (15.4.1) holds for any u D(). We

normalize ui by
n/qpl/(qp)
ui Lq (Q(i) ) = di
(15.5.9)
N
and put u = i=1 ui into (15.4.1). Then

p/q
ui qLq (Q(i) )

p/q
|u|q dx
Cp
i=1
N

i=1
Q(i)
|l ui |p dx.
By (15.5.8) we have

p/q
ui qLq (Q(i) )
cC p
i=1
n(qp)/qpl
di
ui pLq (Q(i) ) ,
i=1
which together with (15.5.9) yields

(pq)/q
np lq/(qp)
di
cC p .
i=1
Thus the series (15.5.1) converges.

According to Theorem 1.2.1, there exists a sequence of cubes {Q (i) }i1
M that forms a covering of of nite multiplicity with Q (i) Q (j) = ,
i = j. The convergence of series (15.5.1) was proved previously (the arguments
should be applied to the sequence of mutually disjoint cubes Q(i) ). Therefore
{Q (i) }i1 is a covering in the class Cl,p,q .
Suciency. Let u C0 () and let {Q (i) }i1 be a covering of in the
class Cl,p,q . Obviously,

q/p q/p
q
|u| dx
i i
|u|q dx,
pn/q
where i = di
ity, we obtain
|u|q dx
Q(i)
i1
Capl1 (Q (i) \,
Llp (2Q (i) )). Applying the Holder inequal
q/(pq)
(pq)/q
i1
p/q q/p

i
i1
Q(i)
|u|q dx
By Lemma 15.5.1/1,
p/q

c i
Q(i)
|u|q dx
Q(i)
|l u|p dx.
712

This implies

|u|q dx c
(pq)/p
q/p
|l u|p dx
,
q/(pq)
i1
/ N (Q (i) ), we conclude that

and since Q (i) \
uLq () c
(pq)/pq
n+lpq/(pq)
di
l uLp () .
(15.5.10)
i1
This completes the proof.
In the proof of necessity we incidentally obtained the following necessary

condition for the validity (15.4.1).
Proposition. Let {Q (i) }i1 be a sequence of disjoint cubes in . Then
the divergence of the series (15.5.1) is necessary for the validity of (15.4.1)
with q < p.
15.5.3 Embedding
Llp () Lq () for an Innite Funnel
Example. Consider the domain

= x = (x , xn ) : x = (x1 , . . . , xn1 ), xn > 0, |x | < (xn ) ,
where is a bounded decreasing function.
We shall show that (15.4.1) with p > q 1 holds if and only if

(15.5.11)
(t) dt < ,
0
where = n 1 + lp q/(p q).

Proof. Let {ai } and {bi } be two number sequences dened as follows:
a0 = 0;
b0 = 0;
ai+1 ai = 2(ai ), i 1,
2
bi+1 bi =
(bi ), i 1.
n1
Clearly ai , bi 0 as i , and the dierences ai+1 ai , bi+1 bi decrease.

Dene two sequences of cubes as follows:

(i)
Qext = ai < xn < ai+1 , 2|x | < ai+1 ai , 1 v n 1 ,

(i)
Qint = bi < xn < bi+1 , 2|x | < bi+1 bi , 1 n 1
(i)
(i)
(see Fig. 37). The cubes Qext cover . All (n 1)-dimensional faces of Qext
except for two of them are contained in Rn \ and
15.5 Embedding
713
Fig. 37.
(i)
(i)
Cap Qext \,
L1p 2Qext c(bi+1 bi )np .
(i)
This along with Proposition 14.3.5 implies that Qext \ is a (p, l, l1)-essential
(i)
subset of Qext .
We suppose that the integral (15.5.11) converges and show the convergence
of the series (15.5.1). In fact,
(ai+1 ai )+1
i=0
(ai+1 ai ) (ai ai1 ) + a+1

1
i=1

+1
(ai ) (ai ai1 ) + (0)
.
i1
Since does not increase, we have

(ai ) (ai ai1 )
ai
(t)
dt.
ai1
Hence

+1
(ai+1 ai )+1 (0)
+

2 (t) dt,
i=0
and the sucient condition in Theorem 15.5.2 follows.

We prove the necessity of (15.5.11). Suppose that

(t) dt = ,
0
and let the series (15.5.1) converge for any sequence of disjoint cubes in .
By monotonicity of we have
714

(bi bi1 )
+1
i=1
(bi bi1 ) (bi+1 bi )
i=1

(bi ) (bi+1 bi )
i=1
i=1
bi+1
(t)
dt.
bi
(i)
Consequently the series (15.5.1) diverges for the sequence of cubes Qint . Thus
we arrived at a contradiction. It remains to apply Proposition 15.5.2.

Llp() Lq ()
In this section we obtain the necessary and sucient conditions for the compactness of the embedding operator of
Llp () into Lq () with p, q 1.
15.6.1 Case p q
Theorem. The set

F = u D() : l uLp () 1 ,
is relatively compact in Lq () if and only if one of the following conditions
holds:
1. For any d > 0

d \,
infn
Cap Q
Llp () > kdnpl
(15.6.1)
lim
Qd R \B
if n > p l. Here k is a positive constant that is independent of d.

2. For any d > 0

d \,
inf
Cap Q
Ll (Q2d ) > k
lim
Qd Rn \B
(15.6.2)
if n = pl.
3. The set does not contain an innite sequence of disjoint cubes if
pl > n or pl = n and the set Rn \ is connected.
Proof. Suciency. First we note that by Propositions 13.1.1/3, 13.1.2/1,
and Corollary 13.1.1 the conditions of the theorem are equivalent to

d \,
lim
infn
Cap Q
Llp (Q2d ) > kdnpl
(15.6.3)
Qd R \B
(cf. the proof of Theorem 15.4.2/1). This and the rst part of Theorem 15.4.2/1
imply (15.4.1) for all u D() and hence the boundedness of F in Wpl ().
Since any bounded subset of Wpl () is compact in Lq (\B ), it suces
to prove the inequality

Llp () Lq ()
uLq (\B ) uWpl ()
715
(15.6.4)
with arbitrary positive and suciently large .

Let C (Rn ), = 0 in B1/2 , = 1 outside B1 , and (x) = (x/).
We denote by d a small number that depends on and which will be specied
later. By (15.6.3), there exists a suciently large radius (d) such that

d \,
dpln Cap Q
Llp (Q2d ) > kdnp l
for > (d) and for all cubes Qd Rn \B/4 . Hence, by (15.4.1) with
replaced by \B/2 , we have
u Lq (\B/2 ) ckl dlpn+np/q l (u )Lp (\B/2 ) .
We could choose d beforehand to satisfy
c k l dlpn+np/q < .
Then
uLq (\B ) c
jl j uLp () cuWpl () ,
j=0
which completes the proof of the rst part of the theorem.

Necessity. Let be any positive number. Suppose the set F is relatively
compact in Lq (). Then there exists a number = () so large that
uLq () l uLp ()
). Let Qd denote any cube with edge length d situated
for all u D(\B
outside the ball B . In the proof of the second part of Theorem 15.4.2/1 it
was shown that either

d \,
Cap Q
Llp (Q2d ) dnpl ,
where is a constant satisfying (14.1.2), or

p

d \,
Llp (Q2d ) cdnp/q + dl+n(pq)/pq
.
Cap Q
15.6.2 Case p > q

The following assertion shows that the embedding operator of
Llp () into
Lq () is compact and continuous simultaneously for p > q 1.
Theorem. The set

F = u D() : l uLp () 1
716

is relatively compact in Lq (), 1 q < p if and only if there exists a covering

of in the class Cl,p,q .
Proof. The necessity follows immediately from Theorem 15.5.2. We prove
the suciency. Let {Q (i) }i1 be a covering of in the class Cl,p,q , let di be
the edge length of Q (i) , let be a positive number, and let N be an integer
so large that
n+lpq/(pq)
di
< pq/(pq) .
(15.6.5)
iN +1
We denote the radius of a ball B = {x : |x| < } such that B/4 contains the
cubes Q (1) , . . . , Q(N ) by . By (15.5.10) we obtain
(pq)/pq

n+lpq/(pq)
l (u )
di
,
u Lq (\B/2 ) c
L (\B
)
p
/2
iN +1
where is the same function as in the proof of Theorem 15.6.1. This

and (15.6.5) immediately imply that
uLq (\B ) l uLp () + c()uWpl1 ((B \B/2 )) .
Now the result follows from the compactness of the embedding of
Llp ()
l1
Lq () into Wp ( (B \B/2 )).
15.7 Application to the Dirichlet Problem for a Strongly

Elliptic Operator
Let l be a positive integer and let i, j be multi-indices of orders |i|, |j| l.
Let aij be bounded measurable functions in such that aij = aji for any
pair (i, j). Suppose

aij (x)i j
|j |2 , = const > 0,
(15.7.1)
|i|=|j|=l
|j|=l
for all complex numbers i , |i| = l, and for almost all x .

We dene the quadratic form

aij Di T Dj T dx
A(T, T ) =
|i|=|j|=l
on the space Ll2 (). Obviously, the seminorms A(T, T )1/2 and l T L2 ()
are equivalent.
In the following we apply the results of the previous sections to the study
of the Dirichlet problem for the operator

Au = (1)l
Dj aij Di u .
|i|=|j|=l
15.7 Application to the Dirichlet Problem for a Strongly Elliptic Operator
717
15.7.1 Dirichlet Problem with Nonhomogeneous Boundary Data

Lemma. Let
Ll2 () be a subspace of D (). Then any function T Ll2 ()
can be expressed in the form
T = u + h,
(15.7.2)
where u
Ll2 (), h Ll2 () and Ah = 0 (in the sense of distributions).
Proof. We equip
Ll2 () with the norm [A(u, u)]1/2 . Let T = ui + hi (i =
1, 2) be two decompositions of the form (15.7.2). Since A(h1 h2 ) = 0 and
Ll2 (), it follows that A(u1 u2 , h1 h2 ) = 0. Consequently, A(u1
(u1 u2 )
u2 , u1 u2 ) = 0 and u1 = u2 . The uniqueness of the representation (15.7.2)
is proved.
The space Ll2 () becomes the Hilbert space provided we equip it with any
of the inner products
AN (T, G) = A(T, G) + N 1 (T, G)L2 () ,
N = 1, 2, . . . ,
where is a nonempty open bounded set,

. Let uN denote the projection
of the function T Ll2 () onto
Ll2 () in the space Ll2 () with the norm
[AN (G, G)]1/2 (by hypothesis,
Ll2 () is a subspace of Ll2 ()). Then, for any
l
L2 (),
AN (T uN , ) = 0.
(15.7.3)
In Sect. 15.3 we noted that the embedding
Ll2 () D () implies the
l
embedding L2 () L2 (, loc), an, hence, the estimate

uN 2L2 () CA(uN , uN ),
where C is a constant that is independent of uN . This and the obvious inequality A(uN , uN ) AN (T, T ) show that the sequence uN converges weakly
in
Ll2 () and in L2 () to some u
Ll2 (). Passing to the limit in (15.7.3)
we obtain that h = T u satises A(h, ) = 0, where is any function in

Ll2 (). The lemma is proved.
The representation (15.7.2) enables one to nd the solution of the equation
Ah = 0 which has the same boundary values as T along with its derivatives of
order up to l 1, i.e., to solve the Dirichlet problem for the equation Ah = 0.
Therefore, the conditions for the embedding
Ll2 () D () in Theorem 15.2
imply criteria for the solvability of the Dirichlet problem formulated in terms
of the (2, l)-capacity. Namely, we have the following statement.
Theorem. For any function T Ll2 () to be represented in the form
(15.7.2) it is necessary and sucient that any one of the following conditions
be valid: 1. n > 2l; 2. C = for odd n, n < 2l; 3. C is a set of the
positive (2, n/2)-capacity for n = 2l; and 4. C is not contained in an (n1)dimensional hyperplane or is a set of positive (2, n/2)-capacity for even n <
2l.
718

15.7.2 Dirichlet Problem with Homogeneous Boundary Data

The results of Sects. 15.4 and 15.5 give conditions for the unique solvability
in
Ll2 () of the rst boundary value problem for the equation Au = f with
f Lr ().
We rst formulate the problem. Let f be a given function in Lq (), q =
q(q 1)1 , 1 < q . We require a distribution T
Ll2 () that satises
AT = f .
The following fact is well known.
Lemma. The above Dirichlet problem is solvable for any f Lq () if
and only if
(15.7.4)
uLq () Cl uL2 ()
for all u D().
Proof. Suciency. Since

(f, u) f L () uL () Cf L () l uL ()
q
2
q
q
for all u D(), the functional (f, u) dened on the linear set D(), which
is dense in Ll2 (), is bounded in
Ll2 (). Hence, by the Riesz theorem, there
l
exists T
L2 () such that
(f, u) = A(T, u)
for all u D(). This is equivalent to AT = f .
Necessity. Any u D() with uLl2 () = 1 generates the functional
(v, f ), dened on Lq (). Since to any f Lq () there corresponds a solution
T f of the Dirichlet problem,

(v, f ) l T f L () .
2
Consequently, the functionals (v, f ) are bounded for any f Lq (). Hence
the norms of (v, f ) are totally bounded, which is equivalent to (15.7.4). The
lemma is proved.

Inequality (15.7.4) cannot be valid for all u with the same constant if
either q 2n(n 2 l)1 , n > 2 l, or if q = , n = 2 l. On the other hand,
for q = 2n(n 2 l)1 , n > 2 l, (15.7.4) holds for an arbitrary domain . The
other cases were studied in Sects. 15.4 and 15.5. For q 2, Theorem 15.4.2/1
together with the preceding lemma leads to the following statement.
Theorem 1. The Dirichlet problem in question is solvable in
Ll2 () for

all f Lq () (2 q > 2n/(n + 2l), n 2l and for 2 q 1, n < 2l) if
and only if one of the following conditions is valid:
1. There exists a constant d > 0 such that

d \,
Ll2 > 0 for n > 2l.
inf Cap Q
Qd
15.7 Application to the Dirichlet Problem for a Strongly Elliptic Operator
2. There exists a constant d > 0 such that

d \,
Ll2 (Q2d ) > 0
inf Cap Q
Qd
719
for n = 2l.
3. The domain does not contain arbitrarily large cubes if n < 2 l or if

n = 2 l and Rn \ is connected.
For q < 2, Theorem 15.5.2 and the Lemma imply the following theorem.
Theorem 2. The Dirichlet problem for the equation AT = f is solvable
in
Ll2 () for all f Lq (), 1 < q < 2, if and only if there exists a covering
of the set belonging to the class Cl,2,q and having a nite multiplicity.
15.7.3 Discreteness of the Spectrum of the Dirichlet Problem
The quadratic form A(u, u) generates a selfadjoint operator A in L2 (). By
the well-known Rellich theorem, a necessary and sucient condition for the
discreteness of the spectrum of this operator is the compactness of the embedding
Ll2 () L2 (). So Theorem 15.6.1 implies the following criterion
for the discreteness of the spectrum of A stated in terms of the (2, l)-capacity.
Theorem. The spectrum of the operator A is discrete if and only if one
of the following conditions is valid:
1. For any constant d > 0

d \, Ll > kdn2l
infn
Cap Q
lim
2
Qd R \B
if n > 2l.
Here and in what follows k is a positive number which does not exceed d.
2. For any d > 0

d \,
infn
Cap Q
Ll2 (Q2d ) > k
lim
Qd R \B
if n = 2l.
3. The domain does not contain an innite sequence of disjoint congruent cubes if n < 2l, or if n = 2l and Rn \ is connected.
15.7.4 Dirichlet Problem for a Nonselfadjoint Operator

B(u, u) =
aij (x)D i uDj u dx,
|i|,|j|l
where i, j are n-dimensional multi-indices.
720

We state the Dirichlet problem with homogeneous boundary data for the
operator

(1)|j| Dj (aij (x)Di u)
Bu =
|i|,|j|l
l (). We require
in the following way. Let f be a continuous functional on W
2
l
an element in W2 () such that

B(u, ) = (, f ),
(15.7.5)
2l () is arbitrary.
where the function W
The next assertion is a particular case of a well-known theorem of Hilbert
space theory (see, for instance, Lions and Magenes [500], Chap. 2, 9.1).
Lemma. If
2W l () C|B(, )|
2
l (), then the Dirichlet problem (15.7.5) is uniquely solvable.

for all W
2
Let denote a positive constant such that
Re
aij i j
|i|,|j|l,
|i|+|j|<2l
|i |
1/2
|i|l
1/2
|j |
|j|<l
for all complex numbers i , |i| l, and introduce the set function

Ll2 () .
= inf l u2L2 () : uL2 () = 1, u
Theorem. Let
D2,l () < c0 /
D() < c0 /
if n 2l,
if n < 2l,
where D2,l () is the (2, l) inner diameter of , D() is the inner diameter
of , is the constant in (15.7.1), and c0 is a constant that depends only on
n, l. Then the Dirichlet problem (15.7.5) is uniquely solvable.
Proof. Obviously,
A(u, u) Re B(u, u)
l

k u2L2 ()
1/2 l1

k=0
1/2
k u2L2 ()
k=0
Using the inequality

k/l
1k/l
k uL2 () cl uL2 () uL2 () ,

we obtain
u
Ll2 (),
15.8 Applications to the Theory of Quasilinear Elliptic Equations
A(u, u) Re B(u, u)
721
c 2l
l u2L2 () + 2l1 u2L2 () .
2
Consequently,

c 2l
2
Re B(u, u) uL2 () + 2l1 u2L2 () .
4
4
It remains to note that by Theorem 15.4.1

D2,l ()2l

D()2l
(15.7.6)
if n 2l,
if n < 2l.

Thus the Dirichlet problem (15.7.5) is uniquely solvable for domains with
small (2, l) inner diameter for n 2l or with small inner diameter for n < 2l.
15.8 Applications to the Theory of Quasilinear Elliptic

Equations
In the present section we apply capacitary criteria for integral inequalities
obtained in the previous and present chapters to the Dirichlet and Neumann
problems for quasilinear elliptic equations of the type treated by Leray and
Lions in [488]. We show in Sect. 15.8.1 that the inequality (15.4.1) is necessary
and sucient for the solvability of the Dirichlet problem with zero boundary
data. It is assumed that the right-hand side of the equation is integrable with
some power. Combined with the results of Sect. 15.4, this criterion provides
explicit conditions for the solvability of the boundary value problem given in
terms of the (p, l)-capacity. The corresponding result for linear elliptic equations was obtained in Sect. 15.7.
In Sect. 15.8.3 we prove the uniqueness theorem for the bounded solution
of the Dirichlet problem vanishing outside certain compact subsets of the
boundary with zero (p, l)-capacity. A similar result concerning the Neumann
problem for quasilinear second-order equations is derived in Sect. 15.8.4.
Let be an unbounded open subset of Rn . Here we use the notations
up = uLp () and Qd = {x : 2|xi | < d, i = 1, . . . , n}. When the domain
of the integration is not indicated, it is assumed to be . In this section we
suppose that 1 < p < .
15.8.1 Solvability of the Dirichlet Problem for Quasilinear
Equations in Unbounded Domains
Consider the equation
722


Au (1)l D a (x, l u) = f (x),
x ,
(15.8.1)
where f is integrable in , is a multi-index of order l, and D =

n
1
l /x
1 , . . . , xn .
Suppose that the functions a are continuous in all variables, but x for
almost all x and measurable with respect to x for all other variables.
Besides, we assume that for any vector v = {v }

a (x, v) |v|p1
(15.8.2)
a (x, v)v |v|p ,
with some p > 1. Further, we impose the monotonicity condition

a (x, v) a (x, w) (v w ) > 0
(15.8.3)
for w = v.
We show that the Dirichlet problem for (15.8.1) is solvable for all f
Lq () if and only if
(15.8.4)
vq Cl vp
with q 1 holds for all v C0 ().
We say that u
Llp () L(, loc) is a solution of the Dirichlet problem
Au = f
in ,
u=0
if
on ,
(15.8.5)
f dx
(15.8.6)

a (x, l u)D dx =
for all C0 ().

Lemma 1. Suppose that (15.8.4) for some q 1 and for all v C0 ().
Then, for any f Lq () there exists a unique solution of (15.8.5).
be an expanding sequence of bounded open sets such
Proof. Let {k }
k k+1 and
that
k k = . We introduce a sequence {uk } of solutions
of the problem
Auk = 0 in k ,
uk = 0 on k .
(15.8.7)
Such solutions exist by the LerayLions theorem (see [488]). Since uk
Llp (k ), we have

(15.8.8)
a (x, l uk )D uk dx = f uk dx
with uk extended by zero outside k . Hence
l uk pp1 Cf q ,
uk qp1 C p f q .
(15.8.9)
723
Let {vk } be a subsequence of {uk } weakly converging both in

Llp () and
in Lq (). If u is the weak limit of {vk }, it follows that

(15.8.10)
lim
a (x, l u) D vk D u dx = 0.
k
By (15.8.8),

a (x, l vk )D vk dx
f u dx.
(15.8.11)
Let wm C0 (), wm u in Llp (). Since supp wm k for a xed m and

suciently large k, we obtain

a (x, l vk )D u dx = a (x, l vk )D (u wm ) dx + f m dx.


lim sup a (x, l vk )D u dx f u dx
k

wm ) p + f (u wm )dx

f q cC l (u wm )p + u wm q .

c lim sup l vk pp1 l (u

k
Thus

a (x, l vk )D u dx
f u dx,
which together with (15.8.10) and (15.8.11) gives

Jk =
a (x, l vk ) a (x, l u) D vk D u dx 0.
Next we take a subsequence {wk } of the sequence {vk } such that

lim a x, l wk (x) a x, l u(x) D wk (x) D u(x) = 0 (15.8.12)
for almost all x . Let x be a point where (15.8.12) holds, let be the
limit of l wk (x), and = l u(x). We show that | | < . In fact,

a (x, l wk ) a (x, l u) D wk D u

a (x, l wk )D wk c |l wk |p1 + |l wk | + 1 .
Under the assumption = we nd

a x, l wk (x) D wk (x) ,
which contradicts (15.8.12). Since a (x, y) is continuous in y, we have [a (x,
) a (x, )]( ) = 0. Therefore, = , that is,
724

l wk (x) l u(x),

a x, l wk (x) a x, l u(x) ,
almost everywhere. Next we use the following simple property (see Leray and
Lions [488]): If gk , g Lp (), gk p c and gk g almost everywhere in
, it follows that gk g weakly in Lp (). This implies that a (x, l wk )
a (x, l u) weakly in Lp (). Hence

a (x, l u)D w dx = lim

a (x, l wk )D w dx
k
for any w C0 (). Taking into account that supp w k for large enough
k, we see that

a (x, l wk )D w dx = f w dx.
Thus, u is a solution of the equation Au = f .
Let u1 ,u2 be two solutions of the problem (15.8.5). Since (u1 u2 )
Llp (),
we have

a (x, l ui )D (u1 u2 ) dx = f (u1 u2 ) dx
with i = 1, 2 and therefore,

a (x, l u1 ) a (x, l u2 ) D u1 D u2 dx = 0.
This and (15.8.3) imply that u1 = u2 almost everywhere in . The Lemma is
proved.
Now we prove a converse assertion.
Lemma 2. If (15.8.5) is solvable for any f Lq (), then (15.8.4) holds
for all v C0 ().1
Proof. Let v C0 () with v
Ll () = 1. The functional
p

v(f ) =
f v dx,
dened on Lq (), can be written in the form

v(f ) = a (x, l u)D v dx,
where u
Llp () L(, loc). Hence

v(f ) c l upp1 ,
and the functionals v(f ) are bounded for each f Lp (). Therefore, the
norms of v(f ) are uniformly bounded and (15.8.4) holds.
1
Note that Lemma 2 holds without the condition (15.8.3).
725
Combining Lemmas 1 and 2 with Theorem 14.1.2, we arrive at the following assertion.
Theorem. Let q [p, pn/(n pl)] for n pl, and q [p, ] for n > pl.
The problem (15.8.5) is solvable for any f Lq () if and only if satises
one of the conditions:
1. For certain d > 0 and n > pl

d C,
Llp > 0
inf Cap Q
Qd
with the inmum here and below taken over all cubes Qd with edge length d.
2. For certain d > 0 and n = pl

d C,
inf Cap Q
Llp (Q2d ) > 0.
Qd
3. The set does not contain large cubes if n < pl and the complement of
is connected if n = pl.
15.8.2 A Weighted Multiplicative Inequality
The goal of this subsection is to prove the following auxiliary fact.
Proposition. Let be an open subset of Rn and let u and z be functions
in Llp () L () such that the product uz has a compact support. Then
m
m

m
l l
z m u pl cu1
z l upl + cu zm
pl ,
(15.8.13)
where m = 1, 2, . . . , l 1, and c is a constant depending only on m, l, and p.

The proof is based on the next two lemmas.
Lemma 1. Let a0 , a1 , . . . , al , A be nonnegative numbers satisfying the
inequality
1/2 1/2

ak C1 ak1 ak + ak1 A
(15.8.14)
for k = 1, 2, . . . , l. Then
and therefore,
k
1/(k+1) k+1

ak C2 a0
ak+1 + a0 Ak
(15.8.15)
lk k

ak C3 a0 k all + a0 Ak .
(15.8.16)
Here C1 , C2 , and C3 are constants depending on l.

Proof. The result follows by induction.
Lemma 2. For all functions v and z given on R and such that v and
z are absolutely continuous, and the support of vz is compact, the following
inequality holds:

726

1/2

v |z| q C v |z|+1 q + v|z|1
+1
q
+1

z q ,
(15.8.17)
where p is the norm in Lp (R), 1, q > + 1, and C is a constant

depending only on and q.
Proof. It may be assumed that
+1
v |z|
q
+1
We introduce the notation

q
b0 = |v| 1 |z|q dx,

q
b2 = |v | +1 |z|q dx,
< .
b1 =

b=
|v | |z|q dx,
|z |q dx.
(The integration is carried out over R.)

We rst examine the case q < 2. Let =
positive number. By the Holder inequality
|v| +
and let be an arbitrary
|v |
|z|q dx
(|v| + )
q

1 2
q

1
q
|v| +
|z| dx
b1 =
q(2q)
2(1) ,
|v |2 |z|q
2q
q
2
dx
.
(15.8.18)
(|v| + ) 1
Integrating by parts, we nd that the last integral is not greater than

1 2q

1 2q
1
1
|v| +
c
|v| +
|v ||z|q dx + c
|v ||z ||z|q1 dx.
We denote these integrals by i1 and i2 and estimate them with the aid of the
Holder inequality:

|v |
i1

i2
+1
q1

+1
q
q
q

1
1
0
q
|z| dx
|z| dx
b2 q b0 q ,
|v| +
q
q
q1
q
q

1
q
|z| dx
|v | |z| dx
|v| +

q
+1
q1
1 +1 1
0
b1q b0 q b q ,
|z | dx
q
and these estimates combined with (15.8.18) imply the estimate (15.8.17).
Let q 2. After integration by parts in b1 we obtain

q
q
b1 c |vv ||v | 2 |z|q dx + c |v||v | 1 |z ||z|q1 dx.
727
q
q
We apply the triple Holder inequality with exponents p1 = q2
, p2 = 1
,
q
and p3 = +1 to the rst integral on the right-hand side. We estimate the
q
,
second integral by means of the same inequality with exponents p1 = q
q
p2 = 1
, and p3 = q. Then
1 2 1 +1
1 1 1
b1 c b1 q b0 q b2 q + b1 q bo q b q .
This proves the lemma.

Now we are in a position to justify (15.8.13).
Proof of Proposition. Let q > m + 1. By Lemma 2 we have
m

1
1
z m u q cz m+1 m+1 u 2 q z m1 m1 u 2 q
m
m+1
m1

+ cz m1 m1 u q zq .
(15.8.19)
m1
Hence, by Lemma 1 we arrive at (15.8.13).
15.8.3 Uniqueness of a Solution to the Dirichlet Problem with an

Exceptional Set for Equations of Arbitrary Order
Let be a bounded open set in Rn with smooth boundary and let e be
a compact subset of . We say that a function u given on belongs to the
e = , u Llp (, loc).
space Llp (, e, loc) if, for any open set with
l
l
Let Lp (, e, loc) be the set of functions u in Lp (, e, loc) such that for any
open set with
e = , u is a Wpl ()-limit of a sequence of functions
in Wpl () vanishing near .
Lemma 1. Let M(e, Rn ) and let u
Llp (, e, loc) L (). Then
m

m
1 m
c(1 )m m u pl u l (1 )l l upl + u l pl
(15.8.20)
with m = 1, . . . , l 1. Here and elsewhere in this subsection by c we mean

positive constants depending only on p, m, and l.
Proof. The result follows from (15.8.13) combined with the Gagliardo
Nirenberg inequality (1.8.1), where j = m and u = .

We say that u is a bounded solution of the Dirichlet problem for (15.8.1)
with an exceptional compact set e if u
Llp (, e, loc) L () and, for
all
Llp () L () vanishing in a neighborhood of e,

(15.8.21)
a (x, l u)D dx = f dx, f L1 ().

728
Lemma 2. Let P(e, Rn ), where P is the class of functions introduced

in Remark 13.3. Also let u be a solution of the Dirichlet problem with an
exceptional set e such that u
Llp (, e, loc) L (). Then

1/p
c(1 )l l up u l p + u1/p
f 1 ,
(15.8.22)
where c is a positive constant independent of u and .

Proof. Let z = 1 . Setting = uz lp in (15.8.21), we nd

z lp a (x, l u)D u dx

!
= a (x, l u)
D uD z lp dx
( )!!
>0

+ f z lp u dx,

c

z lp |l u|p dx
|l u|p1

|lk u|k z lp dx + u f 1 .
k=1
Applying the Holder inequality on the right-hand side, we obtain

l
k

l

z lk |lk u|
z ki
|mj | + A,
c z l u p

i
i=1
k=1
j=1
mj =k
where Ap = u f 1 . The rst term on the right-hand side does not exceed
l

lk

z lk u
c
k=1
k

pl
lk
i=1
i
j=1
mj
lp
mj
mj =k
By the GagliardoNirenberg inequality (1.8.1) with j replaced by mj and

u by we nd
l

k

l pl z lk lk u
c z l l u p
k=1
pl
lk
+ A.
Using (15.8.20) with m = l k, we arrive at

l

k
k

1 k

l l
l pl u
cz l lk up
z l u p l
k=1
+ u l p + A
729

z l l up + c u l p + A
for all > 0. The estimate (15.8.22) is proved.
Lemma 3. If e is a compact (p, l)-polar subset of and is a polynomial of degree at most l 1 from
Llp (, e, loc) L (), then 0.
Qd . Since Cap(e,
Llp (Q2d )) = 0, it follows that cap(e,
Proof. Let
= 0. Let > 0 and let g be an open set such that e g Q2d
and Cap(g,
L1p (Q2d )) < . We cover \g by a nite number of open cubes qi
with qi e = . By the monotonicity and semiadditivity of the (p, 1)-capacity
we have

cap qi ,
L1p (Q2d )
L1p (Q2d ))

cap ()\g,
L1p (Q2d )

L1p (Q2d ) cap ,
L1p (Q2d )
cap ,
L1p (Q2d ) cap g,
with a suciently small > 0. Therefore, there exists an open cube Q such
that

,
e = , cap Q
L1p (Q2d ) > 0.
Q
) vanishing in a neighborhood
Let {um } be a sequence of functions in C (Q
\ and converging to in W l (Q ). Clearly,
of Q
p
l um Lp (Q ) = l um Lp (Q ) 0
as m . By Corollary 14.3.4 and Theorem 14.1.2 we have um 1 in
Lp (Q ). Hence = 0 on Qd and thus = 0 everywhere.

Theorem. The Dirichlet problem for the equation (15.8.1) with an exceptional (p, l)-polar set has at most one bounded solution.
Qd and let u and v be two bounded solutions of the
Proof. Let
Dirichlet problem. Further, let z = 1 with P(e, Q2d ) (see Remark
13.3). Then

a (x, l u) a (x, l v) D z lp (u v) dx = 0
and hence

z lp a (x, l u) a (x, l v) D (u v) dx

=
a (x, l u) a (x, l v)
def
J =

>0

!
D (u v)D z lp dx.
( )!!
730

This implies the inequality

z l(p1) a (x, l u) + z l(p1) a (x, l v)
J c
p
p
k

l pl z lk lk (u v)
k=1
pl
lk
By Lemma 1 we obtain
l
k
k

p1
1 k

l l
z l (u v)p l
l pl u v
J c z l l upp1 + z l l v p
+ u
1 k
v l p l .
k=1
Combining the condition u, v L () with Lemma 2, we see that the right1/l

hand side of the last inequality does not exceed F (l p )l p , where F
is a continuous function on [0, +). Let {k } be a sequence of functions in
P(e, Q2d ) with l k Lp (Q2d ) 0. Then

lim (1 k )lp a (x, l u) a (x, l v) D (u v) dx = 0.
k
Since k 0 in measure, it follows that

a (x, l u) a (x, l v) D (u v) dx = 0
and hence l (u v) = 0. Thus we nd that u v is a polynomial of a degree
at most l 1 and therefore, by Lemma 3, u = v. The theorem is proved.
15.8.4 Uniqueness of a Solution to the Neumann Problem for
Quasilinear Second-Order Equation
Let be an open bounded set in Rn . Consider the second-order equation
(15.8.1), that is, the equation

ai (x, u) = f,
xi
f L(),
(15.8.23)
with coecients satisfying (15.8.2) and (15.8.3).

Let e be a closed subset of . We say that a function u L1p (, e, loc) is a
bounded solution of the Neumann problem for (15.8.23) with the exceptional
set e if u L () and for all L1p ()L () vanishing in a neighborhood
of e,

dx = f dx.
(15.8.24)
ai (x, u)
xi
731
Lemma. Let z C (Rn ), z 0, and let z = 0 in a neighborhood of

e . For any bounded solution of the Neumann problem for (15.8.23) with
the exceptional set e the inequality

(15.8.25)
z p |u|p dx c osc uzp f 1 + c(osc u)p zpp
holds.
Proof. We put
= z p (u C),
C = const,
in (15.8.24). Then

u
z p ai (x, u)
dx
xi
u C zp f 1 + pu C

z
z p1 ai (x, u)
dx.
xi
The result follows by (15.8.2) and Holders inequality.

Corollary. Let e be a closed (p, 1)-polar subset of and let u be a
bounded solution of the Neumann problem for (15.8.23) with the exceptional
set e. Then u L1p ().
Qd and let {k } be a sequence of functions in P(e, Q2d )
Proof. Let
with k Lp (Q2d ) 0. By (15.8.25) with z = 1 k we have

lim sup (1 k )p |u|p dx c osc uf 1 .
k
Since k 0 in measure, we conclude that

|u|p dx c osc uf 1 .
(15.8.26)
The corollary is proved.

Theorem 1. The dierence of any two bounded solutions of the Neumann
problem for (15.8.23) with an exceptional (p, 1)-polar set is a constant.
Proof. Let u and v be two solutions. For any z in C (Rn ), z 0, vanishing
in a neighborhood of e, we have

(u v)z p dx = 0.
ai (x, u) ai (x, v)
xi
Hence

z p ai (x, u) ai (x, v)
(u v) dx
xi
732


pu v ai (x, u) ai (x, v) z p1 p zp

pu v zupp1 + zvpp1 zp .
(15.8.27)
Qd and let {zk } be a sequence of functions in P(e, Q2d ) with

Let
zk Lp (Qd ) 0. Passing to the limit in (15.8.27) and using (15.8.26), we
arrive at

(u v) dx = 0.
ai (x, u) ai (x, v)
xi
Thus u v = const.
1 (, e) we denote
Denition. Let e be a compact subset of . By W
p
1
1
the completion in Wp () of the set of functions in Wp () vanishing in a
neighborhood of e.
The next theorem also concerns (15.8.23). Under an additional assumption
on we prove the necessity of the (p, 1)-polarity of e.
Theorem 2. Let be the image of a cube under a bi-Lipschitz mapping.
If the Neumann problem for (15.8.23) with an exceptional set e has the only
bounded solution u const, then e is a (p, 1)-polar set.
Qd . Suppose that cap(e,
Proof. Let
L1p (Qd )) > 0. We introduce two
compact sets e1 and e2 with

L1p (Qd ) > 0 (i = 1, 2).
e1 e2 = , ei e, cap ei ,
(The existence of such sets easily follows from the semi-additivity of the (p, 1)capacity.) Let be a neighborhood of e with the properties
Qd and
e2
= and let h P(e, ). By the LerayLions theorem (see [488]), there
p1 (, e1 e2 ), such that
exists u Wp1 () with u h W
ai (x, u)
dx = 0
(15.8.28)
xi
p1 (, e1 e2 ). Since the functions in Wp1 () are absolutely confor all W
tinuous on almost all straight lines parallel to the coordinate axes, it can be
1 (, e1 e2 ). Plugging
easily shown that both (u 1)+ and u belong to W
p
them into (15.8.28) as and using (15.8.2), we nd that 0 u 1 almost
everywhere in . Together with (15.8.28) this implies that u is a bounded
solution of the Neumann problem with the exceptional set e1 e2 e.
p1 (, e2 ) and
It remains to prove that u1 is a constant. Since u W
1
Lp (Qd )) > 0 it follows from Theorem 14.1.2 that
cap(e2 ,
up Cup .
Hence there exists a constant C1 independent of u such that
uWp1 () C1 up .
733
The assumption that is a Lipschitz image of a cube implies the existence

1 (Qd , e1 ) and
of an extension v of u onto Qd with the properties v h W
p
vLp (Qd ) C2 uWp1 () .
Therefore,

0 < cap e1 ,
L1p (Qd ) vpLp (Qd ) (C1 C2 )p upp
and u = const. The theorem is proved.

Sections 15.1, 15.2. In [234] Deny and Lions studied the orthogonal projection method with respect to the Dirichlet problem for the Laplace operator.
In particular, they gave the following description of the sets satisfying
L12 () D (). This embedding occurs for n 3 for arbitrary set , for
n = 2 if the 2-capacity of C is positive and for n = 1 if C = .
For integers l the problem of the embedding
Ll2 () D () was solved
by Hormander and Lions [385]. Their result was formulated in terms of (2, l)polarity. In the authors paper [535] it was noted that the conditions of the
Hormander and Lions theorem can be restated in terms of the l-harmonic
capacity.
Our proof of Theorem 15.2 follows the paper by Khvoles and the author [570], where the next result is obtained.
Let p > 1, l > 0 and let
hlp () be the completion of D() with respect to
the norm ()l/2 uLp (Rn ) (in particular,
hlp () =
Llp () for integer l).
Theorem. The space
hlp () is embedded into D () if and only if one of
the following conditions is valid:
1. n > pl; 2. cap(C, Hpl ) > 0, if n = pl; 3. C = if n 0 or C does not lie in a

(n 1)-dimensional hyperplane, if n < pl, l n/p is noninteger.
Sections 15.315.6. The results of these sections are due to the author [544, 546, 549]. A proof of Theorem 15.4.2/1, based upon polynomial
capacities in Chap. 14, was proposed by Wannebo [784].
A topic close to the theme of this chapter, but not treated here, is the
question of the validity of the Hardy-type inequality

p
dx
u(x)p
cp u(x) dx for all u C0 (), (15.9.1)
p
[dist(x,
)]
and its generalizations, under minimal restrictions on . Let, for brevity,

d = dist(x, ). Corollary 2.3.4 shows that (15.9.1) holds if and only if for all
compact subsets F of
734

dx
const capp (F, ).
dp
(15.9.2)
Moreover, the best constant cp in (15.9.1) satises

p
d dx
(p 1)p1
F
cp .
sup
pp
cap
F
p (F, )
Since (15.9.1) results from the inequality

dx
dx
|u| p c0
|u| p1 ,
d
d
by changing u for |v|p1 v, the following sucient condition for (15.9.1) is

a consequence of Theorem 2.1.3. For all bounded open sets g with smooth
boundaries and closures g , the weighted isoperimetric inequality holds:

dx
ds
c0
,
p
p1
g d
g d
and the best constant in (15.9.1) is subject to the inequality
p
d dx
g
p
.
cp p sup
d1p ds
g
g
A dierent approach to the characterization of (15.9.1) for general domains
was initiated by Ancona [47] and developed by Lewis [493]. They proved
(15.9.1) for the so-called uniformly p-thick domains, which means that for
every x C and for all > 0

)\, B(x, 2) const capp B(x,
), B(x, 2) .
capp B(x,
Due to Ancona (n = 2) and Lewis (n 2), the uniform thickness condition is
equivalent to (15.9.1) in the critical case p = n.
A comprehensive survey of further development of this area, together with
new results, can be found in Kinnunen and Korte [424]. We only give a list of
publications related to the topic: Mikkonen [602]; Wannebo [784786]; Hajlasz
[339]; Sugawa [732]; Bj
orn, MacManus, and Shanmugalingam [106]; Calude
and Pavlov [164]; Korte and Shanmugalingam [453]; Laptev and A. Sobolev
[480]; et al. To this we add that in [785], Wannebo studied the generalized
Hardy inequality

m

|k u|p dt(mk)p dx c |m u|p dt dx,
k=0
where u C0 () and t is less than some small positive number. He obtained

a sucient condition on formulated in terms of the polynomial capacity
Capm1 .
735
For Hardy inequalities of type (15.9.1) of fractional order (0, 1) we

refer to Rafeiro and Samko [669]. The authors show that such an inequality
holds under the additional assumption

inf mn B(x, r) \ Crn
x
if and only if a certain property of (related to a property of to be a

pointwise multiplier in the space of the Riesz potentials of order ) holds.
Sections 15.715.8. The presented scheme of applications of integral
inequalities to boundary value problems in variational form is well known
(see, for instance, Deny and Lions [234], Lions and Magenes [500], Leray and
Lions [488]). The results in these sections are due to the author.
The criterion for the discreteness of the spectrum of the Dirichlet problem for the Laplace operator (Theorem 15.7.3 for l = 1) was found by
Molchanov [610]. It was improved by the author and Shubin in [590] (see
Chapter 18 of the present book). For any integer l, Theorem 15.7.3 was proved
by the author [533]. Similar results for more general operators are presented
in Sect. 16.5.
The unique solvability of the Dirichlet problem for the operator B (see
Sect. 15.7.4) in domains with small (2, l) inner diameter (which is understood in a sense dierent from ours) was shown by Kondratiev [448, 449]
(see the Comments to Sect. 14.2). Among other applications of the results of
Chap. 14 to the theory of elliptic equations we list the following: theorems
of the PhragmènLindel
of type for elliptic equations of arbitrary order (see
Landis [476]), estimates for eigenvalues of the operator of the Dirichlet problem in an unbounded domain (see Rozenblum [683] and Otelbaev [652, 653]),
necessary and sucient conditions for the Wiener regularity of a boundary
point with respect to higher-order elliptic equations (see Mazya [559]).
Applications of the results of Chaps. 14 and 15 to the quasilinear elliptic
equations given in Sect. 15.8 are due to the author [540].
The weighted multiplicative inequality (15.8.13) was proved in Mazya
[542].
16
Embedding
Llp (, ) Wrm ()
In this chapter we denote by

Llp (, ) the completion of D() with respect
to the metric
l uLp () + uLp (,) ,
where p > 1, is an open set in Rn and is a measure in .
For instance, in Sect. 16.2 we show that the space D() equipped with
the norm
Llp (Rn , ) is continuously embedded into Lp (Rn ) if and only if
inf (Qd \e) const > 0
{e}
(16.0.1)
for any cube Qd with suciently large edge length d. Here {e} is the collection
of all subsets of the cube Qd with small enough capacity cap(e,
Llp (Q2d )). The
corresponding embedding operator is compact if and only if condition (16.0.1)
holds and inf {e} (Qd \e) tends to innity as the cube Qd tends to innity
(Sect. 16.3).
In Sect. 16.4 we study the closability of certain embedding operators. One
of the theorems in this section asserts that the identity operator dened on
Llp () is closable if and only if the meaD() and acting from Lp (, ) into
sure is absolutely continuous with respect to the (p, l)-capacity. In Sect. 16.5
the previously proved criteria are reformulated for p = 2 as necessary and sufcient conditions for the positive deniteness and for the discreteness of the
spectrum of the selfadjoint elliptic operator generated by the quadratic form

a (x)D uD u dx +
|u|2 d, u D().
||=||=l
16.1 Auxiliary Assertions

d ) we have
Lemma 1. For any u C (Q

DOI 10.1007/978-3-642-15564-2 16,
737
738
16 Embedding
Llp (, ) Wrm ()
upLp (Qd ) c1 dpl u
p
p,l,Qd
cdn
upLp (Qd ,) ,
inf (Qd \e)
(16.1.1)
d
where is a measure in Qd , the inmum is taken over all compacta e Q
l
npl
with cap(e, Lp (Q2d )) d

( is an arbitrary constant), and the seminorm
p,l,Qd is that introduced in 14.1.2.
Proof. We assume that the average value u
Qd of the function u in Qd is
nonnegative and put

d : u(x) .
2 = dn/p uLp (Qd ) ,
e = x Q
Obviously,
uLp (Qd ) u Lp (Qd ) + dn/p
and hence
uLp (Qd ) 2u Lp (Qd ) .
(16.1.2)
First consider the case cap(e ,

Llp (Q2d )) > dnpl . If u
Qd , then applying Theorem 14.1.3 to the function u and using (16.1.2) we deduce the
estimate
upLp (Qd ) c1 dpl u pp,l,Qd .
(16.1.3)
Further, if u
Qd < , then by virtue of the inequality
u u
Qd Lp (Qd ) cduLp (Qd ) ,
we obtain

Qd dn/p 2cduLp (Qd ) .
uLp (Qd ) 2 uLp (Qd ) u
(Here we used the fact that u
Qd 0.) So, for cap(e ,
Llp (Q2d )) > dnpl the
l
estimate (16.1.3) is valid. In the case cap(e , Lp (Q2d )) dnpl we have

2p dn
2 p dn
upLp (Qd ) = 2p dn p
upLp (Qd ,) .
|u|p d
(Qd \e ) Qd \e
inf (Qd \e)
This and (16.1.3) imply (16.1.1).

d with
Lemma 2. Let E be a compact subset of Q

cap E,
Llp (Q2d ) < dnpl ,
(16.1.4)
where is a suciently small positive constant which depends only on n, p,

and l. Then
inf
uC0 (Qd \e)
u
Ll (Rn ,)
p
uLp (Rn )

c dl + dn/p (Qd \e)1/p .
(16.1.5)

Llp (, ) Wrm ()
739
Proof. Clearly, it suces to consider the case d = 1. By Remark 13.3 there

exists a P(e, Q2 ) such that 0 1 and
l Lp (Rn ) c0 1/p .
(16.1.6)
Let be an arbitrary function in C0 (Q1 ) that is equal to unity on Q1/2 and

satises 0 1. For the function u = (1 ), which is obviously in
C0 (Q1 \E), we obtain

|u|p d (Q1 \E),
(16.1.7)
Q1

l uLp (Rn ) c l Lp (Rn ) + l ()Lp (Rn )
c1 + c2 l Lp (Rn ) c1 + c0 c2 1/p .
(16.1.8)
We obtain the following lower bound for the norm of u in Lp :

uLp (Rn ) Lp (Rn ) Lp (Rn )
2n/p l Lp (Rn )
sup
uC0 (Q2 )
uLp (Rn )
.
l uLp (Rn )
This and (16.1.6) along with the smallness of imply uLp (Rn ) 21n/p .
Combining this estimate with (16.1.7) and (16.1.8) we arrive at (16.1.5). The
lemma is proved.

From the proof of Lemma 2 it follows that can be subjected to the
inequality
l uLp (Rn )
inf
,
(16.1.9)
2np cp
p
0
uC0 (Q2 ) uL (Rn )
p
where c0 is the constant in (16.1.6).
In what follows in this chapter, the subsets of Qd that satisfy the inequality

Cap e,
Llp (Q2d ) dnpl ,
where n pl, = c1
, c is the constant in (13.3.5), will be called (p, l)negligible (cf. Denition 14.1.1). As before for pl > n, by denition, the only
(p, l)-negligible set is the empty one.
Similar to Chap. 14, the collection of all (p, l)-negligible closed subsets of
d will be denoted by N (Qd ).
the cube Q
Llp(, ) Wrm()
Let be an arbitrary open set Rn and let be a measure in . We denote by
d whose intersections with Rn \ are (p, l)-negligible.
F the set of all cubes Q
740
16 Embedding
Llp (, ) Wrm ()
We introduce the number

D = Dp,l (, ) = sup
d F
Q
d : dnpl
inf
eN (Qd )

d \e) .
(Q
(16.2.1)
Obviously, D is a nondecreasing function of the set . For = 0 the number D

coincides with the (p, l) inner diameter of the set , introduced in Sect. 14.2.
Theorem 1. Let 0 m l, p r < , l m > n/p n/r. Then
(a) The inequality
uWrm (Rn ) Cu
Ll (,)
p
(16.2.2)
is valid for all u C0 () if and only if there exist positive constants d and
k such that
d \E) k
(16.2.3)
(Q
d in F and for all compacta E in N (Qd ).
for all cubes Q
(b) The best constant in (16.2.2) satises the estimates

c1 C Dln(1/p1/r) max D m , 1 cC.
(16.2.4)
Proof. We begin with the right inequality in (16.2.4) and with the necessity
of condition (16.2.3). From the denition of D it follows that for any > 0
d F with
there exists a cube Q
dnpl
inf
eN (Qd )
d \e)
(Q
and with D d D if D < , d > 1 if D = .

d \)
Let e N (Qd ). According to Corollary 13.3/2 the set E = e (Q
satises (16.1.4). By Lemma 16.1/2 we can nd a function u C0 (Qd \E)

such that

pl
d \E) 1/p uL (Q )
+ dn (Q
u
p
d
Ll (,) c1 d
p
c2 dl uLp (Qd ) .
(16.2.5)
Making use of the obvious inequalities

uLp (Qd ) c3 min dn/pn/r uLr (Qd ) , dn/pn/r+m m uLr (Rn )

c4 dn/pn/r min 1, dm uWrm (Rn ) ,
from (16.2.5) we obtain

l+n/pn/r
min 1, dm uWrm (Rn ) .
u
Ll (,) c5 d
p
Since is arbitrarily small, the right inequality in (16.2.4) is proved.

Llp (, ) Wrm ()
741
If (16.2.2) holds then the right-hand side of (16.2.4) together with l > n/p
n/r imply D < . In this case, for d = 2D and k = dnpl we have (16.2.3).
Now we shall prove the suciency of (16.2.3) and the left inequality
d } where d is chosen to satin (16.2.4). Cover Rn by the cubic grid {Q
isfy (16.2.3). If the cube Qd has a (p, l)-essential intersection with Rn \ then,
by Theorem 14.1.2,
upLp (Qd ) cdpl u pp,l,Qd .
d \
/ N (Qd ) then by Lemma 16.1/1
On the other hand, if Q
upLp (Qd ) cdpl u
p
p,l,Qd
+ ck1 dn upLp (Qd ,) .
(16.2.6)
Summing over all cubes of the grid, we obtain

upLp () c
l
dpj j upLp () + ck1 dn upLp (,) .
(16.2.7)
j=1
Applying (15.4.4) and H

olders inequality from (16.2.7) we obtain
upLp () cdpl l upLp () + ck1 dn upLp (,) .
(16.2.8)
Since the embedding operator of Wpl into Wrm is continuous, the suciency
Let
1/pl

.
R = max d, dn k1
Then by (16.2.8) we have
uLp cRl u
Ll (,) ,
p
which implies the following estimate for the best constant C in (16.2.2):
C c sup
uD
m uLr + uLr
.
l uLp + Rl uLp
Replace u(x/R) by u. Then

Rm+n/r m uLr + Rn/r uLr
Rn/pl (l uLp + uLp )
uD

uWrm
cRln/p+n/r max Rm , 1 sup
.
uD uWpl
C c sup
(16.2.9)
We may assume that D < . The denition of D implies the validity

of (16.2.3) for d = 2D, k = dnpl . Inserting = 2D into (16.2.9) we arrive at the left inequality in (16.2.4). The theorem is proved.
742
16 Embedding
Llp (, ) Wrm ()
Since for pl > n the only negligible set is the empty set, Theorem 1 can,
in this case, be restated in the following equivalent formulation without the
notion of capacity.
Theorem 2. Let pl > n, 0 m l, p r < , and l m > n/p n/r.
Then:
(a) Inequality (16.2.2) holds for all u D() if and only if the estimate
d .
(Qd ) > k is valid for some d > 0, k > 0 for all cubes Qd with Q
(b) The best constant C in (16.2.2) satises (16.2.4) with

d) .
(16.2.10)
D = Dp,l (, ) = sup d : dnpl (Q
d
Q
Part (a) of Theorem 1 can also be simplied for pl = n when Rn \ is

connected.
Theorem 3. Let pl = n and let Rn \ be connected. Inequality (16.2.2) is
valid for all u D() if and only if there exist constants d > 0, k > 0 such
d and for all (p, l)-negligible
that (16.2.3) holds for all cubes Qd with Q
d.
compacta F Q
Proof. We need only establish the suciency since the necessity is contained in Theorem 1.
d be a cube of the coordinate grid having a nonempty intersection
Let Q
n
with R \. Then Q2d contains a continuum in Rn \ with a length not less
than d. So according to Proposition 13.1.2/2 we have

2d \,
Llp (Q4d ) c.
cap Q
This and Theorem 14.1.2 imply
upLp (Qd ) cdpl u
p
p,l,Qd ,
u D().
(16.2.11)
d that intersects Rn \, together

The latter estimate, applied to each cube Q
d leads to (16.2.7). The further arguwith inequality (16.2.6) for cubes Q
ments are just the same as in the proof of the suciency of condition (16.2.3)
in Theorem 1. The theorem is proved.
Llp(, ) Wrm()
16.3.1 Essential Norm of the Embedding Operator
Let E be the identity mapping of the space C0 () considered as an operator
from
Llp (, ) into Wrm (Rn ).
With E we associate its essential norm, i.e., the value

Llp (, ) Wrm ()
=
p,l,m = inf E T ,
743
(16.3.1)
{T }
where {T } is the set of all compact operators
Llp (, ) Wrm (Rn ).

Theorem. Let 0 m < l, p r < , l m > n/p n/r. Then:
(a)
< if and only if D = Dp,l (, ) < ;
(b) there exists a constant c > 1 such that

c1
Dln/p+n/r max Dm , 1 c
with
(16.3.2)
N ),
D = Dp,l (, ) = lim Dp,l (, \Q
(16.3.3)
where QN is a cube with center 0 and edge length N .

Proof. Part (a) follows from Theorem 16.2/1. We prove the left-hand
side of (16.3.2). Let TN be the operator of multiplication by (N 1 x),
N = 1, 2, . . . , where C0 (Q2 ) = 1 on Q1 . By Theorem 16.2/1, for
any u C0 (Q2 ) we have
l
c1 u
Ll (,) l uLp + D uLp + uLp (,) cu
Ll (,) . (16.3.4)
p
Hence if N D then
TN
Ll (,)
Ll (,) c.
p
(16.3.5)
From (16.3.4) and the well-known compactness theorem it follows that the
Llp (, ) Wrm is compact. Applying Theorem 16.2/1 to the
mapping TN :
set \QN we obtain for any u C0 () that
(E TN )
ln/p+n/r
Wrm
cDN
m
max DN
, 1
(E TN )u
Ll (,) ,
p
N ). Hence, using (16.3.5) and passing to the limit as

where DN = Dp,l (, \Q
N , we arrive at the left inequality in (16.3.2).
We prove the right inequality in (16.3.2). We may assume that D = 0.
Obviously D1 D2 D. Using the same arguments as in the proof of

the right inequality (16.3.4), for any large enough number N we construct a
N ) with the diameter of the support not exceeding
function uN C0 (\Q
2D and such that
l
uN
Ll (,) cD ,
p
Obviously,
uN Lp = 1.
(16.3.6)
744
16 Embedding
Llp (, ) Wrm ()
u
uN Lp = D n/p GD
N Lp
u
u
c1 Dn/p min{GD
n Lr , m GD
N Lr },
(16.3.7)
where Ga is the operator dened by (Ga u)(x) = u(ax). We choose a sequence

{Ni }i1 so that the distance between the supports of uNi and uNj (i = j) is

more than 2 nD. From (16.3.6) and (16.3.7) we obtain

(u
(u
1 c2 Dn/p min
GD
Nj uNi ) L , m GD
Ni uNj ) L
r
r

c3 Dn/pn/r min Dm , 1 uNi uNj Wrm .

(16.3.8)
We denote an arbitrary compact operator by T :
Llp (, ) Wrm . Passing, if necessary, to a subsequence we may assume that the sequence {T uNi }
converges in Wrm . Further,
(E T )(uNi uNj )
m uNi uNj W m
T (uNi uNj )
m ,
r
W
W
r
which along with (16.3.8) shows that

c5 lim sup
(E T )(uNi uNj )
W m Dln/p+n/r max Dm , 1 .
Ni ,Nj

ln/p+n/r
max Dm , 1 .
c6 E T
Ll (,)W m D
p

16.3.2 Criteria for Compactness
From Theorem 16.3.1 we immediately obtain the following criterion.
Theorem 1. Let 0 m l, p r < , and l m > n(p1 r1 ). Then

the set

F = u D() : u
Ll (,) 1
p
is relatively compact in Wrm if and only if D = Dp,l (, ) < and

N ) = 0.
lim Dp,l (, \Q
The preceding theorem admits the equivalent formulation.

Theorem 2. Let 0 m l, p r < , l m > n(p1 r1 ). Then the
set F is relatively compact in Wrm if and only if :
(a) there exist positive constants d0 and k such that inequality (16.2.3) is
d F and for any compactum E N (Qd );
valid for any cube Qd0 with Q
0
0

Llp (, ) Wrm ()
745
(b) we have
inf
eN (Qd )
d \e)
(Q
(16.3.9)
d F , tends to innity, where d is an arbitrary positive

as the cube Qd , Q
number.
Proof. Necessity. Inequality (16.2.3) follows from Theorem 16.2/1. If
(16.3.9) is not valid then there exists a sequence of cubes {Q (i) }i1 , with
a limit point at innity and satisfying the conditions (i) the set {di } of the
edge lengths of the cubes Q (i) is bounded, (ii) the inequality
inf
eN
(Q(i) )

Q (i) \e < c0
holds. This and Lemma 16.1/2 imply that we can nd a sequence of functions
{ui }i1 in D() with
diam supp ui c,
ui
Ll (,) c,
dist(supp ui , supp uj ) c > 0, i = j,

ui Wrm c1 ,
ui Lp c2 > 0.
Suciency. Let conditions (16.2.3) and (16.3.9) be valid. By Theorem 16.2/1 and (16.2.3) the value Dp,l (, ) is nite. Now (16.3.9) and (16.2.8)
yield
sup
uWrm /u
lim
Ll (,) = 0.
N uC (\Q
N )
N imply
This and the right inequality (16.2.4) applied to \Q
N ) = 0.
lim Dp,l (, \Q
Hence, F is relatively compact in Wrm by virtue of Theorem 1. The proof is

complete.

Clearly, in the case pl > n Theorem 2 can be stated as follows.
Theorem 3. Let pl > n, 0 m < l, p r < , and l m > n/p
n/r. The set F is relatively compact in Wrm if and only if the condition of
Theorem 16.2/1 holds and if
(Qd )
(16.3.10)
d tends to innity.
as the cube Qd with Q
In the case pl = n under the hypothesis of the connectedness of Rn \ the
statement of Theorem 2 can be simplied in the following way.
Theorem 4. Let pl = n, 0 m < l, p r < , m < n/r, and suppose
the complement of is connected. Then F is relatively compact in Wrm if and
746
16 Embedding
Llp (, ) Wrm ()
only if the condition of Theorem 16.2/3 holds and (16.3.9) is valid as the cube
d , tends to innity.
Qd , Q
Proof. Taking into account Theorem 2 we can limit ourselves to proving suciency. According to the GagliardoNirenberg theorem (see Theorem 1.4.7),
k uLr cl uLp u1
Lp ,
where 0 k l and = l1 (n/p n/r + k). Hence, it suces to obtain the
relative compactness of F in Lp . By Theorem 16.2/3, F is bounded in Wpl (Rn )
and consequently it is relatively compact in Lp (B ) for any
(0, ). It
remains to show that for any we can nd a
=
() such that
uLp (Rn \B ) .
(16.3.11)
d that
Put dl = in (16.1.1) and choose
=
() so that for any cube Q
intersects Rn \B we have
inf
eN (Qd )
d \e) p dn .
(Q
Then for such cubes

uLp (Qd ) c u
p,l,Qd

+ uLp (Qd ,) .
(16.3.12)
d that intersect both Rn \ and Rn \B the estimate

For all Q
uLp (Qd ) c u
p,l,Qd
(16.3.13)
holds by (16.2.11). We take the pth power of (16.3.12) and of (16.3.13) and
then sum them over all cubes of the coordinate grid with edge length d which
have a nonempty intersection with Rn \B . This implies (16.3.11). The theorem is proved.
16.4 Closability of Embedding Operators

Let be a measure in Rn . We call it (p, l) absolutely continuous if the equality
cap(B, Wpl ) = 0, where B is a Borel set, implies (B) = 0.
For instance, by Proposition 10.4.3/2 the Hausdor -measure is (p, l)
absolutely continuous provided the integral (10.4.16) converges.
Let E be the identity operator dened on D(), which maps Lp () into
Llp (, ).
Theorem 1. Let n pl, p > 1. The operator E is closable if and only if
the measure is (p, l) absolutely continuous.
Proof. Suciency. Suppose a sequence of functions {uk }k1 in D() conLlp (, ). Further,
verges to zero in Lp () and that it is a Cauchy sequence in
16.4 Closability of Embedding Operators
747
let D uk v in Lp () for any multi-index with || = l. Then for all

D()

v dx = lim
uk D dx = 0
(1)||
k
l (). The
and v = 0 almost everywhere in . Consequently, uk 0 in W
p
sequence {uk } contains a subsequence {wk } that converges to zero (p, l)-quasieverywhere (see 10.4.4). The (p, l)-absolute continuity of the measure implies
wk 0 -almost everywhere. Since {uk } is a Cauchy sequence in Lp (, )
then uk 0 in the same space.
Necessity. Suppose there exists a Borel set B with cap(B, Wpl ) = 0
and (B) > 0. Let F denote a compact subset of B satisfying 2(F ) > (B).
Further let {k }k0 be a sequence of open sets with the properties 1. F
k , Wpl ) 0, and 3. (k ) (F ). We introduce the
k+1 k , 2. cap(
capacitary measure k and the capacitary Bessel potential wk = Vp,l k of the
set
k (see Proposition 10.4.2/2).
We show that wk 0 in the space C(e) where e is any compactum that
does not intersect F . Let = dist(e, F ) and let k be a number satisfying
2 dist(e,
k ) > . We set

gk (y) =
1/(p1)
Gl (y z) dk (z)
(see the denition of the Bessel potential in 10.1.2). If 4|x y| < then
4|z y| for all z
k and consequently
1/(p1)

1/(p1)

gk (y) c() k (
k )
= c() cap
k , Wpl
.
Therefore,
1/(p1)

Gl (x y)gk (y) dy c1 () cap
k , Wpl
.
4|xy|
On the other hand,

1/(p1)

Gl (x y)gk (y) dy c2 ()gk Lp = c2 () cap
k , Wpl
.
4|xy|>
These estimates and the equality

wk (x) = Gl (x y)gk (y) dy
imply wk 0 in C(e).
The arguments used in the proof of Theorem 13.3/1 show that the functions vk = 1 [(1 wk )+ ]l satisfy
vk Wpl cwk Wpl .
(16.4.1)
748
16 Embedding
Llp (, ) Wrm ()
Moreover, it is clear that vk = 1 in a neighborhood of F , 0 vk 1 in Rn

and that the sequence {vk } converges uniformly to zero on any compactum
that does not intersect F .
Let uk denote a mollication of vk with small enough radius. Using the
equality lim cap(
k , Wpl ) = 0 and the estimate (16.3.9) we obtain uk Wpl 0.
Let be any function in P(F, ) with support S. Clearly, uk = 1 in a
neighborhood of F and uk Wpl 0. We show that uk is a Cauchy sequence
in Lp (, ). Let be an arbitrary positive number and let M be so large that
(M \F ) < . We choose an integer N to satisfy

uk (x)p /(S), x S\M
for k > N . Hence, for any k, l > N , we obtain

(uk ul )p d c
|uk ul |p d +
S\M
M \F

|uk ul |p d
< c.
Therefore uk is a Cauchy sequence in Lp (, ). Since uk = 1 on F then

uk pLp (,) (F ) > 0.
The existence of such a sequence contradicts the closability of the operator E .

In the case pl > n the operator E is always closable, which trivially follows
from the Sobolev theorem on the embedding Wpl () C().
pl ()
Theorem 2. The identity operator dened on D(), which maps W
into Lp (, ), is closable if and only if the measure is (p, l) absolutely continuous.
Proof. Suciency. Let {uk }k1 , uk D() be a Cauchy sequence in
l (). Then it converges to zero in
Lp (, ) that converges to zero in W
p
Lp (, ) by the previous theorem.
Necessity. Suppose there exists a Borel set B with cap(B, Wpl ) = 0 and
(B) > 0. Under this condition, in Theorem 1 we constructed a sequence of
functions uk D(), which is a Cauchy sequence in Lp (, ) that converges
pl () and does not converge to zero in Lp (, ). The result follows.
to zero in W

Theorem 3. 1. The identity operator dened on D(), which maps
Llp ()
into Lp (, ) (lp < n), is closable if and only if the measure is (p, l) absolutely continuous.
2. The same holds for n = pl provided Rn \ is a set of positive (p, l)capacity.
Proof. 1. Suciency. Let a sequence of functions uk in D() be a Cauchy
Llp (). Since lp < n, we have
sequence in Lp (, ) that converges to zero in
16.5 Application: Positive Deniteness and Discreteness of the Spectrum
749
the embedding
Llp () Lq (), q = pn(npl)1 and hence {uk } converges to
zero in n-dimensional Lebesgue measure. Further, using the same approach as
in the proof of suciency in Theorem 1, from {uk } we can select a subsequence
{wk } that converges to zero almost everywhere.
Necessity follows from Theorem 2.
2. The case n = pl is considered in the same way, making use of the
embedding
Llp () Lq (, loc) (see Sect. 15.3).
16.5 Application: Positive Deniteness and Discreteness

of the Spectrum of a Strongly Elliptic Operator
We dene two quadratic forms
A(u, u) =
||=||=l
a (x)D uD u
dx,

|u|2 d,
B(u, u) = A(u, u) +
where u D(), a are measurable functions and is a measure in .

Suppose
1 l u2L2 () A(u, u) 2 l u2L2 ()
for all u D().
By denition the form B is closable in L2 () if any Cauchy sequence um
D() that converges to zero in L2 () in the norm (B(u, u))1/2 converges to
zero in the norm (B(u, u))1/2 .
This condition implies the existence of the unique selfadjoint operator B
in L2 () with
(u, Bv) = B(u, v) for all u, v D().
Clearly, the closability of the form B is necessary for the existence of B.
By Theorem 16.4/1, B is closable in L2 () if and only if the measure is
(p, l) absolutely continuous. We shall assume that has this property.
The next assertion is a particular case of Theorem 16.2/1.
Theorem 1. 1. The operator B is positive denite if and only if
(Qd \F ) k
(16.5.1)
d having (2, l)-negligible intersection with Rn \ and for all

for all cubes Q
d with certain d > 0, k > 0.
(2, l)-negligible compacta F Q
2. The lower bound of the spectrum of the operator B satises
c1 1 D2l c2 2 D 2l ,
(16.5.2)
where c1 and c2 are constants that depend only on n and l and D is dened
by (16.2.1) with p = 2.
750
16 Embedding
Llp (, ) Wrm ()
Corollary 1. 1. For 2l > n a necessary and sucient condition for the

positive deniteness of B is the inequality (Qd ) k, where d and k are
d is any cube in .
certain positive constants and Q
2. The lower bound of the spectrum of B satises (16.5.2) with D dened
by (16.2.10) for p = 2.
Corollary 2. If 2l = n and the set Rn \ is connected, then a necessary and sucient condition for the positive deniteness of B is the valid d and for all (2, n/2)-negligible compacta
ity of (16.5.1) for all cubes Q
F Qd with certain d > 0 and k > 0.

This corollary is a special case of Theorem 16.2/3.
Theorem 2. The lower bound of points of condensation of the spectrum
of B satises
(16.5.3)
c1 1 D2l c2 2 D2l ,
where D is dened by (16.3.3) with p = 2.

Proof. Let
be the number dened by (16.3.1) for p = 2. By Theorem 16.3.1 it suces to prove that =
2 . From Theorems 16.3.1 and 1
it follows that = 0 if and only if
= . So we may suppose = 0 and
= .
We introduce the family {E } of orthogonal projective operators that form
a resolution of the identity generated by the selfadjoint operator B in L2 ().
Then
2
(16.5.4)
1 B(u, u)
(E E )u
L () .
2
Since the projective operator E is nite-dimensional for < , then (16.5.4)

and the denition of
imply
2 .
For any , 0 < < , in L2 () there exists an orthogonal and normalized
innite sequence of functions {i }i1 such that (Bi , j ) = 0 for i = j and
(Bi , j ) + .
Let T be an arbitrary compact operator mapping
Ll2 (, ) into L2 (). We
have that
(E T )(i j )L2 ()
i j
Ll (,)
2
i j L2 () T (i j )L2 ()
(B(i j ), i j )1/2
=
(i 2L2 () + j 2L2 () )1/2 T (i j )L2 ()
[(Bi , i ) + (Bj , j )]1/2
( + )1/2 21/2 ( )1/2

T (i j )
L2 () .
The upper limit of the right-hand side as i, j equals ( + )1/2 . So by
the arbitrariness of and the denition of
we obtain
2 1 . The theorem
is proved.
751
The preceding result implies the next theorem.

Theorem 3. The spectrum of the operator B is discrete if and only if
D = 0.
This theorem is equivalent to the following assertion.
Theorem 4. The spectrum of the operator B is discrete if and only if
inf
F N (Qd )
d \F )
(Q
(16.5.5)
d (d is an arbitrary positive number), having a (2, l)-negligible

as the cube Q
intersection with Rn \, tends to innity.
For 2l > n this criterion can be reformulated as follows.
Theorem 5. Let 2l > n. The operator B has a discrete spectrum if and
only if
d)
(Q
d , where d > 0, tends to innity.
as the cube Q
Theorem 16.3.2/4 implies the following result.
Theorem 6. Let 2l = n and let Rn \ be connected. The operator B has
d , where
a discrete spectrum if and only if (16.5.5) holds as the cube Q
d > 0, tends to innity.

In 1934, K. Friedrichs [292] proved that the spectrum of the Schrodinger
operator + V in L2 (Rn ) with a locally integrable potential V is discrete
provided V (x) + as |x| (see also Berezin and Shubin [84]). On the
other hand, if we assume that V is semibounded in the following, then the
discreteness of spectrum easily implies that for every d > 0

V (x) dx + as Qd ,
(16.6.1)
Qd
where Qd is an open cube with the edge length d and with the edges parallel
to coordinate axes, Qd means that the cube Qd goes to innity (with
xed d). This was rst noticed by A.M. Molchanov in 1953 (see [610]) who
also showed that this condition is, in fact, necessary and sucient in the case
n = 1, but not sucient for n 2. Moreover, in the same paper Molchanov
discovered a modication of condition (16.6.1) that is fully equivalent to the
discreteness of the spectrum in the case n 2. In other words, he found a
criterion for the compactness of the embedding
L12 (, ) into L2 (), where
= V dx. Molchanovs test states that for every d > 0
752
16 Embedding
Llp (, ) Wrm ()
inf
F
Qd \F
V (x) dx +
as Qd ,
(16.6.2)
d which
where the inmum is taken over all compact subsets F of the closure Q
are called negligible. The negligibility of F in the sense of Molchanov means
that
cap(F ) cap(Qd ),
(16.6.3)
where cap is the harmonic capacity and > 0 is a suciently small constant.
More precisely, Molchanov proved that we can take = cn where for n 3

1
.
cn = (4n)4n cap(Q1 )
(16.6.4)
The notion of negligibility will be studied in detail in Chap. 18 following the

article by Mazya and Shubin [590].
The results of the present chapter develop the result by Molchanov [610]
for the case of higher derivatives. Such criteria were obtained for the space
Ll2 (, ) with 2l > n (when there is no need for a capacity) by Birman and
Pavlov [102]. In the case q p > 1, l = 1, 2, . . . , the necessary and sucient
conditions for the boundedness and compactness of the embedding operator of
Llp (, ) into Lq () were established by the author [544]. Two-sided estimates

for the norm and for the essential norm of this operator are due to the author
and Otelbaev [573], where the space with the norm
()l/2 u
+ uLp (,)
L ()
p
is considered (l is any positive number). The embedding theorems for more

general spaces with weighted norms were obtained by Ole nik and Pavlov [645],
Lizorkin and Otelbaev [506], and Otelbaev [653]. The function

inf
Qd (x)\e ,
Dp,l (x) = sup d : dnpl
eN (Qd (x))
(compare with (16.2.1)) was used by Otelbaev [652] in the derivation of bounds
for the Kolmogorov diameter of the unit ball in
Llp (, ) measured in Lp ().
We present his result.
The Kolmogorov diameter of a set M in a Banach space B is dened to
be the number

inf {Lk } supf M inf gLk f gB , k = 1, 2, . . . ,
dk (M, B) =
k = 0,
supf M f B ,
where {Lk } is the set of subspaces of B with dim Lk k.
Theorem. (Otelbaev [652].) Let M be the unit ball in the space
Llp (, )
and let N () be the number of Kolmogorov diameters of M in Lp () which
exceed 1 , > 0. Then
753

c1 N (c) n/l mn x : Dp,l (x) 1/l cN c1 ,
where c does not depend on , , and .
By denition, the embedding operator of
Llp (, ) into Lp () is in the
class l if

< .
dk M, Lp ()
k=0
The just-stated Otelbaev theorem shows that the embedding operator under consideration is in the class l if and only if l > n and

ln
Dp,l (x)
dx < .
An immediate application of these results are two-sided estimates for the

eigenvalues of the Dirichlet problem for the operator ()l/2 + and the
conditions for the nuclearity of the resolvent of this operator.
In conclusion we mention the article [39] by Alekseev and Oleinik, where
two-sided estimates for Kolmogorov diameters of the unit ball of
Llp () in the
space Lp (, ) for pl > n and arbitrary measure are found.
17
Approximation in Weighted Sobolev Spaces
17.1 Main Results and Applications

For p 1 we dene Lm,p (Rn ) as the set of distributions u on Rn such that

um,p =
1/p
m

p
k u(x) dx
< .
k=1
Here k u denotes the vector (D u)||=k ,

Lm,p (Rn ) is the completion of
n
m,p
C0 (R ) with respect to the L
seminorm. Now let Rn be an open
set and let be a nontrivial positive Radon measure on . We will study the
space Hm,p (), dened as the completion of Lp () Lm,p (Rn ) C0 () with
respect to the norm
um,p; = uLp () + um,p .
m,p (). Note that if p < n
The closure of C0 () in Hm,p () is denoted H
then by Sobolevs inequality the elements in

Lm,p can be identied with funcnp
tions in Lp , where p = np
. If a domain in the notations of a space is
m,p () are naturally
not indicated it is assumed to be Rn . The elements in H
m,p
m,p
Lp (Rn , loc).
identied with elements in
L . Note also that L
The theorem to be proved in this chapter is the following. (See Sect. 17.2
for denitions of the capacities Bm,p and H1nm used in the following.)
Theorem. Let be a nontrivial positive Radon measure concentrated on
Rn and let C denote Bm,p for p > 1 and H1nm for p = 1. Then
m,p () = H m,p () if either p n or p < n and C( c ) = .
(i) H
m,p () = H m,p () if
Suppose now that p < n and C( c ) < . Then H
and only if
(ii) () = , when either m n, p = 1 or mp > n, p > 1;
DOI 10.1007/978-3-642-15564-2 17,
755
756
17 Approximation in Weighted Sobolev Spaces
(iii) (F c ) = for every closed set F Rn satisfying C(F ) < , when

n
or m < n, p = 1.
either 1 n or 1 = p < n, m n the condition
C( c ) < is just a complicated way of saying that c should be bounded.
Among other results of this chapter there is the following criterion obtained
in Sect. 17.3. The identity operator dened on C0 () and mapping L1 ()
into the completion of C0 () with respect to the norm

|m u(x)| dx + |u| d
is closable if and only if is absolutely continuous with respect to the Hausdor
capacity H1nm . Similar facts for p > 1 were established in Sect. 16.4.
To show the usefulness of the above Theorem we give its applications to
the problem of equality of the minimal and maximal Dirichlet Schr
odinger
forms and to the question of selfadjointness of a dierential operator.

Q[u, u] =
| u|2 dx +
|u|2 (dx).
(17.1.1)
The closure of Q dened on the set

u C (Rn ) : supp u , Q[u, u] < ,
which may contain functions with noncompact support, will be denoted by
Qmax . Another quadratic form Qmin is introduced as the closure of Q dened
on C0 ().
The following necessary and sucient condition for this equality results
directly from the above theorem, where cap stands for the 2-capacity with
respect to Rn .
Proposition 1. (i) If either n 2 or n > 2 and cap(Rn \) = , then
Qmax = Qmin .
(ii) Suppose that n > 2 and cap(Rn \) < . Then Qmax = Qmin if and
only if
(\F ) =
for every closed set F with cap(F ) < .
(17.1.2)
Note that the equality Qmax = Qmin is equivalent to (17.1.2) in the particular case = Rn .
The next example, which illustrates a possible shape of a set with innite
capacity, was treated at the end of the proof of Theorem 2.6.2/3.
17.2 Capacities
757
Example. Consider the domain complementing the innite funnel

x = (x , xn ) : xn 0, |x | f (xn ) ,
where f is a continuous and decreasing function on [0, ) subject to f (t)
cf (2t). It was shown in the proof of Theorem 2.6.2/3 that cap(Rn \) <
if and only if the function f (t)n2 for n > 3 and the function (log(t/f (t)))1
for n = 3 are integrable on (1, ).
Now we turn to another application of the Theorem concerning the question of selfadjointness.
Proposition 2. Let L1 (Rn , loc), n 3, be a positive function, locally
bounded away from zero, and let the operator A = 1 be dened on
C0 (Rn ).
A necessary condition for A to be essentially selfadjoint in L2 () is that

(x) dx =
Fc
whenever F is a closed set such that cap(F ) < .

Proof. Suppose does not satisfy the condition in the theorem. Then
by Theorem 17.1 and the Hahn-Banach theorem there is a nonzero function
u H1,2 (Rn ) such that

u(x)v(x)(x) dx + u(x) v(x) dx = 0
where A is the closure of A. Thus
1,2 (Rn ) D(A),
for all v H
L () = 0
(u, v + Av)
2
and so u D((I + A) ) = D(A ) and u + A u = 0. Now suppose A = A .
and hence
Then u D(A)

u(x)2 (x) dx + u(x)2 dx = 0,
which implies that u = 0. This contradiction shows that A is not selfadjoint.
17.2 Capacities
We will denote dierent constants, not depending on the essential functions
or variables considered, by A. The ball with radius r and centered at x will be
denoted by B(x, r). If x = 0 we will write only B(r). The annulus B(R)\B(r)
is denoted by A(R, r).
758
We start by listing the convolution kernels needed in the following:

Bessel kernels G , Riesz kernels I , and the truncated Bessel kernels G;1
dened by
G;1 (x) = (x)G (x),
where C0 (B(1))+ is an arbitrary but xed function such that = 1 on
B( 12 ). For 1 we dene
K = I1 G1 .
It is easy to see, analogously to the cases with Riesz or Bessel potentials,
that, for 1 1 we dene

B,p (E) = inf f pp : f L+
p , G f 1 on E ,

R,p (E) = inf f pp : f L+
p , I f 1 on E ,

C,p (E) = inf f pp : f L+
p , K f 1 on E ,
and

B,p;1 (E) = inf f pp : f L+
p , G;1 f 1 on E .
It was proved by D.R. Adams [6] that B,p and R,p are nite simultanen
ously, although not comparable, for 1 < p <
. It is not hard to see from the
proof that C,p and B,p are nite at the same time for 1 < p < n. The set
functions B,p and B,p;1 are comparable. See D.R. Adams and Hedberg [15]
for the proof of this fact. We need the following lemma, which, for technical
convenience, is the reason for introducing the truncated Bessel kernel.
Lemma 1. Let p > 1. If F Rn is closed then

G;1 f 1 on F .
B,p;1 (F ) = inf f pp : f L+
p, C
Proof. We may assume B,p;1 (F ) nite. Let A0 = B(1) and Aj =
B(j + 1)\B(j) for j 1. It was proved by D.R. Adams [6] that
B,p (F Aj ) AB,p (F ).
j=1
Hence
Ba,p;1 (F Aj ) < .
j=1
Now choose fj Lp+ such that
(17.2.1)
17.2 Capacities
fj pp 2j + B,p;1 (F Aj )
and
759
G;1 fj 1
on a neighborhood of F Aj . This can be done since G;1 fj is lower semicontinuous. It is seen from the denition of B,p;1 that we may assume that
supp fj Aj where we denote E = {x : dist(x, E) 1}. By mollication we
obtain functions gj C0 (Aj )+ such that

gj pp A 2j + B,p;1 (F Aj )

and G;1 gj 1 on F Aj . Consequently, if we set g = j=M gj for some
M to be specied later we get G;1 g 1 on F \B(M ). Also, since the sum
dening g is uniformly locally nite, it follows that G;1 g C and
gpp A
gj pp A
j=M
2j + B,p;1 (F Aj ) .
(17.2.2)
j=M
Now let > 0. By (17.2.1) and (17.2.2) we get gp < if M is large enough.
By the same argument with lower semicontinuity and mollication we can
nd a function h C0 (Rn )+ such that

1/p
hp B,p;1 F B(M )
+
and
G;1 h 1 on F B(M ).
Setting f = g + h we obtain
f p B,p;1 (F )1/p + 2 and
C G;1 f 1 on F.
Since was arbitrary, the proof is complete.

For p = 1 the appropriate capacities are Hausdor capacities dened as
follows.
Denition 3. Let 0 d < n. Then, for subsets E of Rn , we dene
Hd (E) = inf
rid ,
i=1
where the inmum is taken over all countable coverings

with ri . For d < 0 we dene Hd (E) = H0 (E).

i=1
B(xi , ri ) E,
The following lemma is immediate except for (iii) which is a variant of the
well-known Frostman lemma [293].
Lemma 2. Let 0 < d < n. Then:
d
d
d
d
(i) H
(E) H1d (E) AH
(E) + A(H
(E))n/d . In particular H
and
d
H1 are nite at the same time.

d
d
(ii)
j=0 H1 (EAj ) AH1 (E), where Aj is as in the proof of Lemma 17.2.
760
(iii) H1d (E) is comparable to sup{(E) : (B(x, r)) rd , r 1, x Rn },

where is a positive measure, for Borel sets E.
Lemma 3. Let m be an integer, 0 < m < n. Then for closed sets F Rn ,
is comparable to

inf 1 + m 1 : C , 0 1, = 1 on a neighborhood of F .
H1nm (F )
Proof. Suppose C , = 1 on a neighborhood of F and 1 +

m 1 < . If is a positive measure supported by F then by Theorem
1.4.2/2 we have
(B(x, r))
(F ) d A sup
1 + m 1 .
nm
r
x:0<r1
Taking the supremum over with
(B(x, r))
1,
r nm
x;0<r1
sup
we get by Lemma 2(iii) that H1nm (F ) A(1 + m 1 ), which proves

one direction of the lemma.
To prove the other direction suppose rst that F is compact. Cover F by
balls B(xi , ri ), ri 1, 1 i s, such that
s
rinm H1nm (F ) + ,
i=1
where > 0. By Lemma 3.1 of Harvey and Polking [356] there are functions
||
i C0 (B(xi , 2ri )), 1 i s, such that |D i | A ri
and such that
s
= i=1 i satises = 1 on a neighborhood of F . We get
1 + m 1
s

i=1

i (x) + m i (x) dx
B(xi ,2ri )
s
rin + rinm AH1nm (F ) + A.
i=1
Since is arbitrary we are done in the case where F is compact.

For the general case we introduce a partition of unity 1 = j=0 j , where
0 j 1, j = 1 on a neighborhood of A2j , supp j A2j and |k j | A
for 1 k m. Here Aj and Aj are as in the proof of Lemma 1. Now choose
functions j corresponding to the sets F A2j according to the construction
for compact sets in a way that
j 1 + m j 1 AH1nm (F A2j ) + 2j ,
17.3 Applications of Lemma 17.2/3
where > 0. Letting =

Lemma 2(ii),
1 + m 1
m

A
A
j=0 k=0

A2j
j=0 j j
761
we obtain, using Leibnizs rule and

k (x) dx

nm
j (x) + m j (x) dx
H1
(F A2j ) + 2j
j=0
2 + A
j=0
H1nm (F Aj ) 2 + AH1nm (F ).
j=0
Since is arbitrary the lemma follows.

We record here some generalizations, depending on Lemma 17.2/3, to the case
p = 1 of some results in Chap. 16. As before let be a positive Radon measure
concentrated on Rn and let W , X, and Y be the completions of C0 ()
with respect to the norms

uW =
u(x) dx + m u(x) dx,

uX = u(x) d + m u(x) dx,
and
uY =

m u(x) dx,
respectively. Then we have the following theorems.

Theorem 1. The identity operator dened on C0 () and mapping L1 ()
into X is closable if and only if is absolutely continuous with respect to
H1nm .
Theorem 2. The identity operator dened on C0 () and mapping W
into L1 () is closable if and only if is absolutely continuous with respect to
H1nm .
Theorem 3. Let m n. Then the identity operator dened on C0 ()
and mapping Y into L1 () is closable if and only if is absolutely continuous
with respect to H1nm .
Remark. Note that for m n and for m = n, respectively, the above
condition on absolute continuity is vacuous so the operators are always clos-
762
able. In the proof of Theorem 1 below quasi-everywhere can be read as

everywhere in this case.
For the proofs we will need two lemmas, where a function u is called H1d quasicontinuous if it is dened H1d -quasi-everywhere and if for every > 0
there is an open set G such that .u|Gc is continuous and H1d (G) < .
Lemma 1. Suppose that uj C0 () and that uj+1 uj W < 4j . Then
uj converges H1nm -quasi-everywhere to an H1nm -quasicontinuous function.
Proof. Let be a positive Radon measure such that (B(x, r)) rnm
for all x Rn and all r 1. Then by Theorem 1.4.3 we have

|uj+1 uj | d A 4j .
By monotone convergence, u
(x) = limj uj (x) exists a.e. with respect to .
Now, let

F = x:u
(x) is dened
and

u(x) uj (x)| 2j .
Ej = x F : |
Then, by part (iii) of Lemma 17.2/2, H1nm (F c ) = 0. Also,

(Ej ) A2j |
u uj | d A2j ,

so H1nm (Ej ) A2j . Let Fk = jk Ej . Then H1nm (Fk ) A2k . Thus,
given > 0, we may choose k and an open set Gk with H1nm (Gk ) < such
that Fk F c Gk . Since uj u
uniformly on Gck , the lemma is proved.
Lemma 2. Suppose that u is H1d -quasicontinuous and that E = {x :
u(x) = 0} is a Borel set with mn (E) = 0. Then H1d (E) = 0.
Proof. Suppose that H1d (E) = c > 0. There is an open set G with H1d (G) <
such that u|Gc is continuous. We do not specify here because the choice
of it depends on a certain constant, appearing later in the proof. However,
is a xed positive number less than c.
Let K E\G be a compact set such that H1d (K) > c and set Kj =
{x : dist(x, K) 1j }. By Lemma 17.2/2, part (iii), we can choose measures
j supported by Kj such that
j (B(x, r))
A
rd
x;0<r1
sup
and j (Kj ) = H1d (Kj ).

Dene j (x) = j n (jx), where C0 (B(1)) is a function such that
0 A and = 1. Set j = j j . Then we have

j j (y) =
Thus, for r
763

1
Aj nd .
j (y t) dj (t) Aj n j B y,
j
1
j
j (B(x, r))
A(jr)nd A.
rd
For
1
j
r 1 we have

j (B(x, r))
1
(y)
j (y t) dj (t) dy
B(x,r)
rd
rd

1
1
= d
B(x,r) j (t) dj (t) d
B(x,r+ 1j ) (t) dj (t)
r
r
j (B(x, r + 1j ))
A
A.
(r + 1j )d
Thus we obtain
j (B(x, r))
A.
rd
x;0<r1
sup
Also,

n
j (R ) =

=
j (x y) dj (y) dx
j (x y) dx dj (y)
= j (Kj ) = H1d (Kj ) H1d (K) c .

Let Kj = supp j . Then, since mn (E) = 0, we get

H1d Kj \E A1 j E c = A1 j Rn A1 (c ).
Now we are in the position to specify : Take any positive satisfying A1 (c
) > . Then we obtain

H1d Kj \E > H1d (G).

Hence there are points xj Kj E c Gc . We may assume that xj converges to
some point x0 . Since xj Kj there are points yj K such that |xj yj | < 2j .
Then yj x0 so, in particular, x0 K. By the continuity of u on Gc we
obtain that 0 = u(xj ) u(x0 ) = 0. From this contradiction we conclude that
H1d (E) = 0 and the lemma is proved.

Proof of Theorem 1. We start with the suciency part. Suppose {uj }
C0 () is a Cauchy sequence in X, converging to zero in L1 (). Then D uj
converges in L1 () for || = m and since obviously D uj 0 as distributions
we get D uj 0 in L1 (). Hence, passing to a subsequence, we may assume
764
that uj converges H1nm -quasi-everywhere by Lemma 1. Also, by Lemma 1

we get that, H1nm -quasi-everywhere, uj 0. Thus uj 0 a.e. with respect
to the measure , and since {uj } is a Cauchy sequence in L1 (), we obtain
uj 0 in X.
For the necessity part suppose F is a compact set satisfying
H1nm (F ) = 0 and (F ) > 0. Let Gj F be shrinking open sets such
j ) 0 and (Gj \F ) 0 as j . Then, by Lemma 17.2/3,
that H1nm (G
we can nd j C0 () such that j = 1 on Gj and
j 1 + m j 1 0.
Moreover, by construction, j 0 uniformly outside every neighborhood of
F and there is a compact set K such that supp j K for all j. Now let
> 0 and choose j so that (Gj \F ) < /4. Then, for i and k large enough,

|i k | d
|i k | d + 2(Gj \F ) < + = .
2
2
K\Gj
Hence {j } is a Cauchy sequence in X, converging to zero in L1 (). However,
since
j L1 () (F ) > 0,
we cannot have j 0 in X. This proves the necessity part.
The proofs of Theorems 2 and 3 follow the same lines as the proofs of the
analogous facts in Sect. 16.4, making use of the proof of Theorem 1.
Using Lemma 17.2/3 we can also obtain the necessary and sucient conditions for continuity and compactness of the embedding of X into the Sobolev
space W k,q (Rn ). To state these theorems we need rst some denitions. Let
m < n. Then a set F B(x, r) is called (m, 1)-negligible if
nm
(F ) rnm ,
H
where is a suciently small constant, depending only on m and n. For

m n only the empty set is called (m, 1)-negligible.
Let F () be the family of all balls B(x, r) such that B(x, r)\ is (m, 1)negligible. Then we dene

D1,m (, ) = sup r; B(x, r) F (), inf B(x, r)\F rnm ,

where the inmum is taken over all (m, 1)-negligible closed sets F B(x, r).
We then have the following two theorems.
Theorem 4. Let 0 k m, 1 q < and m k > n(1 1/q). Then
uq + k uq AuX
for all u C0 () if and only if there are positive constants r and a such that
17.4 Proof of Theorem 17.1
765
B(x, r)\F a
for all balls B(x, r) F () and all (m, 1)-negligible sets F B(x, r). The
best constant A is comparable to

Dmn(11/q) max D k , 1 ,
where D = D1,m (, ).
Theorem 5. Let 0 k m, 1 q < , and m k > n(1 1/q). Then
X is compactly embedded into Wqk (Rn ) if and only if D1,m (, ) < and

lim D1,m , \B(R) = 0.
The proofs of these statements are the same as those of the corresponding
theorems for p > 1 in Sects. 16.216.3, relying now on Lemma 17.2/3.

We will need some basic results on the function spaces Lm,p and
Lm,p . We
state rst a lemma of the Hardy type (see Sect. 1.3.1).
Lemma 1.
(i) If 1 p < n and u
L1,p (Rn ) then

|u(x)|p
u(x)p dx.
dx
A
p
|x|
(ii) If p > n and u L1,p (Rn ) then

|u(x)|p
u(x)p dx + u(x)p dx .
dx
A
|x|p
B(1)c
B(1)
(iii) If p = n and u L1,p (Rn ) then

|u(x)|p
u(x)p dx + u(x)p dx .
dx
A
p
B(2)c (|x| log |x|)
B(2)
Lemma 2. Let 1 p n. Then for each u Lm,p (Rn ) there is a unique
constant c such that u c
Lm,p (Rn ).
We turn now to the proof of Theorem 17.1, divided into four cases starting
with the main one.
n
or p = 1, m < n. We start by proving the suciency
The case 1 p m
part. Suppose that satises the condition in the theorem. We will show then
766
that C
Hm,p
Lm,p . Let u C
Hm,p and suppose that c = 0, where
uc
Lm,p . We can assume that c > 0. Let

c

F = x : u(x) c
2
and suppose rst that p > 1. Then u c can be written u c = Km f where

f Lp and we get

2p
c
p f pp < .
Cm,p (F ) Cm,p
x : Km |f |
2
c
If p = 1 then in the same way as in the proof of Lemma 17.2/3

nm
H (F ) A m u(x) c dx < .
On the other hand, since |u|
c
2
on F c , we have
p
2
(F c )
|u|p d < .
c
This contradicts the condition on , hence c = 0. It follows that u

Lm,p .
c
c
Note that since F we must have c = 0 if C( ) = , without using any
condition on .
Now let R C0 (B(2R)) satisfy 0 R 1, R = 1 on B(R) and
|k R | ARk for k m. Then for R 1 we get
u uR pm,p

k u(x)p l (1 R )(x)p dx
1l+km
m
k=1
B(R)c
m

k=1

k u(x)p dx + A

1l+km
l1

k u(x)p dx + ARp
B(R)c
Rlp

k u(x)p dx
A(2R,R)

u(x)p dx.
A(2R,R)
Thus R u u in Lm,p as R by Lemma 1(i). Also

|R u u|p d
|u|p d 0
B(R)c
m,p (), i.e., C

as R since u Hm,p (). It follows that u H
m,p () and hence H m,p () = H

m,p ().
Hm,p () H
m,p ()
We now turn to the necessity part. Suppose that Hm,p () = H
c
and that F is a closed set such that Bm,p (F ) < and (F ) < where now
767
p > 1. Assume also that Bm,p ( c ) < . Then there is an open set G c
< . By Lemma 17.2/1 we can nd f L+ such that
such that Bm,p (G)
p
Now let T be a smooth function

Gm;1 f C and Gm;1 f 1 on F G.
on R+ such that T (t) = 1 if t 1 and

suptk1 T (k) (t) < for 0 k m.
t>0
Then by Lemma 17.2/1, which works also for the truncated kernel, there is a
function g Lp such that T (Gm;1 f ) = Gm;1 g and gp Af p . We
set u = 1 Gm;1 g. Then u C

Lm,p , but by Lemma 17.4, u
/
Lm,p
since Gm;1 g
Lm,p . Moreover

|u|p d F c <
|u|p d =
Fc
m,p (). This contradiction shows

by assumption on F , so u Hm,p ()\H
that must satisfy the condition in the theorem.

For p = 1 we use instead u = 1 , where is the function constructed
Since
in Lemma 17.2/3, satisfying = 1 on a neighborhood of F G.

(x) dx + m (x) dx < ,
Lm,1 . Hence u
/
it follows easily, using the previous multiplier R , that
m,1
(). But
H
|u| d F c < so u Hm,1 ,

and we have again obtained a contradiction.
The case 1 n or 1 = p < n, m n. For the suciency we
again decompose u = v + c where v
Lm,p . Using Sobolevs inequality

1/p

1/p

v(y)p dy
m v(y)p dy
+A
sup |v| A
B(x,1)
B(x,1)
B(x,1)
for every x Rn , we see that v(y) 0 uniformly as |x| . Thus, if c = 0,

c
we can nd R such that |v| |c|
2 on B(R) . Then, if is not nite,
p

|c|
p
B(R)c = ,
|u| d
2
and we have a contradiction. Thus c = 0 and we can proceed as in the rst case,
again using Lemma 1(i). Note that if c is unbounded it follows immediately
that c = 0, without any condition on since v = c on c .
For the necessity we observe that if c is bounded then there is u C

such that u(x) = 1 for large |x|. Then u Lm,p \
Lm,p , since u
/ Lp . If is
m,p
m,p
nite we get u H ()\H () and we are done.
768
The case p > n is proved as previously, now invoking Lemma 1(ii).

The case p = n is proved by Lemma 1(iii), this time using the multiplier

1
R2
R (x) =
log
,
log R
|x|
where is a smooth function satisfying (t) = 0 for t
t 34 . This completes the proof.
1
4
and (t) = 1 for

Remark. The same question of the density of test functions can be asked
about the more general norm
u = uLp () +
m
k up ,
k=l
where m l 2. In the case lp n approximation is always possible. This

is proved in the same way as in the present chapter, using only somewhat
dierent Hardy inequalities. In the general case it is easy to give an implicit
necessary and sucient condition. Namely, with obvious notation, Hl,m,p =
l,m,p if and only if
H

l,m,p
u H , P Pl1 ,
|u P |p d < P = 0.
However, it is not clear whether this condition can be stated in a more transparent way in the spirit of Theorem 17.1, for example, in terms of the polynomial capacities dealt with in Sect. 14.3.

In this chapter we follow the article by Carlsson and the author [171].
Section 17.1. The question as to when the equality Qmax = Qmin holds
was raised by Kato [416]. Simon studied a similar question concerning the
n
magnetic Schr
odinger operator in R [701]. Proposition 17.1/2 improves the
condition (x) dx = , necessary for the essential selfadjointness of the
operator A, which was obtained by Eidus [255].
Section 17.2. Lemma 17.2/3 was partly proved by D.R. Adams [10].
Section 17.3. Lemmas 17.3/1 and 17.3/2 were proved in Carlsson [170].
18
Spectrum of the Schr
odinger Operator
and the Dirichlet Laplacian
Consider the Schrodinger operator + V in L2 (Rn ) with a potential V ,

locally integrable and semibounded below. As we mentioned in Sect. 16.6,
Molchanovs criterion (16.6.2) involves the so-called negligible sets F , that is,
sets of suciently small harmonic capacity.
In Sects. 18.218.3 we show that the constant cn given by (16.6.4) can be
replaced by an arbitrary constant , 0 < < 1. We even establish a stronger
result allowing negligibility conditions with depending on d and completely
describe all admissible functions . More precisely, in the necessary condition
for the discreteness of spectrum we allow arbitrary functions : (0, +)
(0, 1). If (d) = O(d2 ) in the negligibility condition (16.6.3), then it fails to
be sucient, i.e., it may happen that it is satised but the spectrum is not
discrete (Sect. 18.4). However, we show that in the sucient condition we
can admit arbitrary functions with values in (0, 1), dened for d > 0 in a
neighborhood of d = 0 and satisfying
lim sup d2 (d) = +.
(18.0.1)
d0
All such relations involving functions : (0, +) (0, 1), are necessary
and sucient for the discreteness of spectrum. Therefore two conditions with
dierent functions are equivalent, which is far from being obvious a priori.
This equivalence means the following striking eect: If (16.6.2) holds for very
small sets F , then it also holds for sets F which almost ll the corresponding
cubes.
Another important question is whether the operator + V with V 0
is strictly positive, i.e., the spectrum is separated from zero. Unlike the discreteness of spectrum conditions, it is the large values of d that are relevant
here. The following necessary and sucient condition for the strict positivity
was obtained in Sect. 16.5: There exist positive constants d and such that
for all cubes Qd

inf
V (x) dx ,
F
Qd \F

DOI 10.1007/978-3-642-15564-2 18,
769
770

odinger Operator and the Dirichlet Laplacian
d that are negligible

where the inmum is taken over all compact sets F Q
in the sense of Molchanov. We prove in Sect. 18.5 that here again an arbitrary
constant (0, 1) in the negligibility condition is admissible.
The above-mentioned results are proved in a more general context. The
family of cubes Qd is replaced by a family of arbitrary bodies homothetic to a
standard bounded domain which is starshaped with respect to a ball. Instead
of locally integrable potentials V 0 we consider positive measures. We also
include operators in arbitrary open subsets of Rn with the Dirichlet boundary
conditions.
The goal of the last Sect. 18.7 is to obtain explicit lower and upper estimates of the rst eigenvalue of the Laplacian with Dirichlet data formulated
in terms of a capacitary inner radius.
18.1 Main Results on the Schr

odinger Operator
Let V be a positive Radon measure in an open set Rn , n 2. We will
consider the Schrodinger operator, which is formally given by an expression
+ V. It is dened in L2 () by the quadratic form

|u|2 dx +
|u|2 V(dx), u C0 ().
(18.1.1)
hV (u, u) =
For the associated operator to be well dened we need a closed quadratic

form. The form above is closable in L2 () if and only if V is (2, 1) absolutely
continuous (see Sect. 16.4). In the present chapter we will always assume that
this condition is satised. The operator, associated with the closure of the
form (18.1.1), will be denoted HV .
In particular, we can consider an absolutely continuous measure V which
has a density V 0, V L1 (Rn , loc), with respect to the Lebesgue measure
dx. Such a measure will be absolutely continuous with respect to the capacity
as well.
Instead of the cubes Qd a more general family of test bodies will be used.
Let us start with a standard open set G Rn . We assume that G satises the
following conditions:
(a) G is bounded and starshaped with respect to an open ball B (0) of radius
> 0, with the center at 0 Rn ;
(b) diam(G) = 1.
The rst condition means that G is starshaped with respect to every point
of B (0). It implies that G can be presented in the form

G = x : x = r, || = 1, 0 r < r() ,
(18.1.2)
where r() (0, +) is a Lipschitz function on the unit sphere
S n1 Rn (see Lemma 1.1.8). Condition (b) is imposed for the convenience
of formulations.
18.1 Main Results on the Schr

odinger Operator
771
For any positive d > 0 denote by Gd (0) the body {x | d1 x G} which is

homothetic to G with coecient d and with the center of homothety at 0. We
will denote by Gd a body that is obtained from Gd (0) by a parallel translation:
Gd (y) = y + Gd (0) where y is an arbitrary vector in Rn . The notation Gd
means that the distance from Gd to 0 goes to innity.
Denition. Let (0, 1). The negligibility class N (Gd ; ) consists of
all compact sets F Gd satisfying the following conditions:
d
d \ F G
G
(18.1.3)
cap(F ) cap(Gd ).
(18.1.4)
and
The notation cap will be introduced in Sect. 18.2.

Now we formulate our main result about the discreteness of spectrum.
Theorem 1. (i) (Necessity) Let the spectrum of HV be discrete. Then for
every function : (0, +) (0, 1) and every d > 0
inf
F N(d) (Gd ,)
V(Gd \ F ) +
as Gd .
(18.1.5)
(ii) (Suciency) Let a function d (d) (0, 1) be dened for d > 0 in a

neighborhood of 0, and satisfy (18.0.1). Assume that there exists d0 > 0 such
that (18.1.5) holds for every d (0, d0 ). Then the spectrum of HV in L2 ()
is discrete.
Let us make some comments about this theorem.
Remark 1. It suces for the discreteness of spectrum of HV that the condition (18.1.5) holds only for a sequence of ds, i.e., d {d1 , d2 , . . . , }, dk 0
and d2
k (dk ) + as k +.
Remark 2. As we will see in the proof, in the suciency part the condition
(18.1.5) can be replaced by a weaker requirement: There exist c > 0 and
d0 > 0 such that for every d (0, d0 ) there exists R > 0 such that
dn
inf
F N(d) (Gd ,)
V(Gd \ F ) c d2 (d)
(18.1.6)
whenever Gd ( \BR (0)) = (i.e., for distant bodies Gd having nonempty intersection with ). Moreover, it suces that the condition (18.1.6) is satised
for a sequence d = dk satisfying the condition formulated in Remark 1.
Note that unlike (18.1.5), the condition (18.1.6) does not require that the
left-hand side goes to + as Gd . What is actually required is that the
left-hand side has a certain lower bound, depending on d for arbitrarily small
d > 0 and distant test bodies Gd . Nevertheless, the conditions (18.1.5) and
772

(18.1.6) are equivalent because each of them is equivalent to the discreteness

of spectrum.
Remark 3. If we take = const (0, 1), then Theorem 1 gives Molchanovs
result, but with the constant = cn replaced by an arbitrary constant
(0, 1).
Remark 4. For any two functions 1 , 2 : (0, +) (0, 1) satisfying the
requirement (18.0.1), the conditions (18.1.5) are equivalent, and so are the
conditions (18.1.6), because any of these conditions is equivalent to the discreteness of spectrum.
It follows that the conditions (18.1.5) for dierent constants (0, 1) are
equivalent. In the particular case, when the measure V is absolutely continuous
with respect to the Lebesgue measure, we see that the conditions (16.6.2) with
dierent constants (0, 1) are equivalent.
Remark 5. The previous results are new even for the operator H0 =
in L2 () (but for an arbitrary open set Rn with the Dirichlet boundary
conditions on ). In this case the discreteness of spectrum is completely
determined by the geometry of . Namely, for the discreteness of spectrum of
H0 in L2 () it is necessary and sucient that there exists d0 > 0 such that
for every d (0, d0 )
lim inf cap(Gd \ ) (d) cap(Gd ),
Gd
(18.1.7)
where d (d) (0, 1) is a function, which is dened in a neighborhood

of 0 and satises (18.0.1). The conditions (18.1.7) with dierent functions ,
satisfying the previous conditions, are equivalent. This is a nontrivial property
of capacity. It is necessary for the discreteness of spectrum that (18.1.7) holds
for every function : (0, +) (0, 1) and every d > 0, but this condition
may not be sucient if does not satisfy (18.0.1) (see Theorem 2 below).
The following result demonstrates that the condition (18.0.1) is precise.
Theorem 2. Assume that (d) = O(d2 ) as d 0. Then there exists
an open set Rn and d0 > 0 such that for every d (0, d0 ) the condition (18.1.7) is satised, but the spectrum of in L2 () with the Dirichlet
boundary conditions is not discrete.
Now we will state our positivity result. We will say that the operator HV
is strictly positive if its spectrum does not contain 0. Equivalently, we can say
that the spectrum is separated from 0. Since HV is dened by the quadratic
form (18.1.1), the strict positivity is equivalent to the existence of > 0 such
that
(18.1.8)
hV (u, u) u 2L2 () , u C0 ().
Theorem 3. (i) (Necessity) Suppose that HV is strictly positive, so that
(18.1.8) is satised with a constant > 0. Let us take an arbitrary (0, 1).
18.2 Discreteness of Spectrum: Necessity
773
Then there exist d0 > 0 and > 0 such that

dn
inf
F N (Gd ,)
V(Gd \ F )
(18.1.9)
for every d > d0 and every Gd .

(ii) (Suciency) Assume that there exist d > 0, > 0, and (0, 1),
such that (18.1.9) is satised for every Gd . Then the operator HV is strictly
positive.
Instead of all bodies Gd it is sucient to take only the ones from a nite
multiplicity covering (or tiling) of Rn .
Remark 6. Considering the Dirichlet Laplacian H0 = in L2 () we see
from Theorem 3 that for any choice of a constant (0, 1) and a standard
body G, the strict positivity of H0 is equivalent to the following condition:

d > 0, such that cap Gd (Rn \ ) cap(Gd ) for all Gd . (18.1.10)
In particular, it follows that for two dierent s these conditions are equivalent.

In this section we prove the necessity part (i) of Theorem 18.1/1. We will start
by recalling some denitions and introducing the necessary notations.
If F is a compact subset in an open set D Rn , then the harmonic
capacity of F with respect to D is dened as

2
0,1

(18.2.1)
cap(F, D) = inf
u(x) dx : u C0 (D), u|F = 1 ,
Rn
which is nothing but 2-cap(F, ).
We will write Bd for a ball Bd (z) with unspecied center z. Let us use
the notation cap(F ) for cap(F, Rn ) if F Rn , n 3, and for cap(F, B2d ) if
F Bd R2 , where the disks Bd and B2d have the same center. The choice
of these disks will usually be clear from the context, otherwise we will specify
them explicitly.
Note that the inmum does not change if we restrict ourselves to the
Lipschitz functions u such that 0 u 1 everywhere (see Sect. 2.2.1).
We will also need the potential theoretic denition of the harmonic capacity cap(F ) for a compact set F Bd . For n 3 it is as follows:

n
E(x y) d(y) 1 on R \ F ,
(18.2.2)
cap(F ) = sup (F ) :
F
where the supremum is taken over all positive nite Radon measures on F
and E = En is the standard fundamental solution of in Rn , i.e.,
774

E(x) =
1
|x|2n ,
(n 2)n
(18.2.3)
with n being the area of the unit sphere S n1 Rn . If n = 2, then

cap(F ) = sup (F ) :
G(x, y) d(y) 1 on B2d \ F ,
(18.2.4)
F
where G is the Green function of the Dirichlet problem for in B2d , i.e.,
G( y) = ( y),
y B2d ,
G(, y)|B2d = 0 for all y B2d . The maximizing measure in (18.2.2) or in

(18.2.4) exists and is unique. We will denote it F and call it the equilibrium
measure. Note that

(18.2.5)
cap(F ) = F (F ) = F Rn = F , 1.
The corresponding potential will be denoted PF , so

PF (x) =
E(x y) dF (y), x Rn \ F, n 3,
F

PF (x) =
G(x, y) dF (y),
x B2d \ F, n = 2.
We will call PF the equilibrium potential or capacitary potential. We will extend it to F by setting PF (x) = 1 for all x F . It follows from the maximum
principle that 0 PF 1 everywhere in Rn if n 3 (and in B2d if n = 2).
In the case when F is the closure of an open subset with a smooth boundary, u = PF is the unique minimizer for the Dirichlet integral in (18.2.1) where
we should take D = Rn if n 3 and D = B2d if n = 2. In particular,

|PF |2 dx = cap(F ),
(18.2.6)
where the integration is taken over Rn (or Rn \ F ) if n 3 and over B2d (or
B2d \ F ) if n = 2.
The following lemma provides an auxiliary estimate which is needed for
the proof.
Lemma 1. Assume that G has a C boundary, and P is the equilibrium
potential of Gd . Then

d ),
|P |2 ds nL1 d1 cap(G
(18.2.7)
Gd
where the gradient P in the left-hand side is taken along the exterior of Gd ,
ds is the (n 1)-dimensional volume element on Gd . The positive constants
775
, L are geometric characteristics of the standard body G (they depend on the

choice of G only, but not on d): was introduced at the beginning of Sect. 18.1,
and
1
L = inf r (x)
,
(18.2.8)
xG
(x), (x) is the unit normal vector to G at x which is

where r (x) =
directed to the exterior of G.

x
|x|
Proof. It suces to consider Gd = Gd (0). For simplicity we will write G

instead of Gd (0) in this proof, until the size becomes relevant.
Also
We will rst consider the case n 3. Note that P = 0 on Rn \ G.
so, in fact, |P | = |P/|. Using the Green formula, we obtain
P = 1 on G,

P
x
P
P
0=
dx =
P dx
r
|x|
Rn \G
Rn \G

x
P x
P dx
P ds
P
=
|x|
|x|
Rn \G
G

P
P P
xi P
=
dx
ds
x
x
|x|
x
r
n
j
j
i
G
i,j R \G

P ij P
xi xj P P
dx +
dx
=
x
|x|
x
|x|3 xi xj
n
n
j
i
i,j R \G
i,j R \G

2P
xi P
P P
dx
ds
xj xi xj
r
n |x|
G
i,j R \G

2

1
1 P
|P |2 dx +
=

dx
Rn \G |x|
Rn \G |x| r

xi
1
2
|P | dx
|P |2 r ds.
2 i Rn \G |x| xi
G
we see that it equals
Integrating by parts in the last integral over Rn \G,

xi
1
xi
1
2
|P | dx +
|P |2 i ds
2 i Rn \G xi |x|
2 i G |x|

1
1
n1
|P |2 r ds,
|P |2 dx +
=
2
|x|
2
n
R \G
G
where i is the ith component of . Returning to the calculation above, we
obtain

2

n3
1
1 P
2
0=
|P | dx +

dx
2
Rn \G |x|
Rn \G |x| r

1
|P |2 r ds.
(18.2.9)
2 G
776

It follows that

|P | r ds (n 1)
2
Rn \G
1
|P |2 dx.
|x|
Recalling that G = Gd (0), we observe that |x|1 (d)1 . Now using (18.2.6),
we obtain the desired estimate (18.2.7) for n 3 (with n 1 instead of n).
Let us consider the case n = 2. Then, by denition, the equilibrium potential P for G = Gd (0) is dened in the ball B2d (0). It satises P = 0
in B2d (0) \ G and the boundary conditions P |G = 1, P |B2d (0) = 0. Let us
rst modify the previous calculations by taking the integrals over B (0) \ G
where d < < 2d. We will get additional boundary terms
(instead of Rn \G),
with the integration over B (0). Instead of (18.2.9) we will obtain

2

1
1
1 P
2
0=
|P | dx +

dx
2 B (0)\G |x|
B (0)\G |x| r

P 2
1
1
|P |2 ds.
2
|P |2 r ds +
2 G
2 B (0)
r
Therefore,

|P |2 r ds
G
B (0)\G

P 2
2

2
|P | ds
r
B (0)

|P |2 dx +
|P |2 ds.
1
|P |2 dx +
|x|
B2d (0)\G
B (0)
Now let us integrate both sides with respect to over the interval [d, 2d] and
divide the result by d (i.e., take the average over all ). Then the left-hand
side and the rst term in the right-hand side do not change, while the last
term becomes d1 times the volume integral with respect to the Lebesgue
measure over B2d (0) \ Bd (0). By (18.2.6) the right-hand side can be estimated
by (1 + )(d)1 cap(Gd ). Since 0 < 1, we get the estimate (18.2.7) for
n = 2.

Proof of Theorem 18.1/1, part (i). (a) We use the same notations as previously. Let us x d > 0, take Gd = Gd (z), and assume that G has a C
boundary. Let us take a compact set F Rn with the following properties:
(i) F is the closure of an open set with a C boundary;
(ii) Gd \ F B3d/2 (z);
(iii) cap(F ) cap(Gd ) with 0 < < 1.
Let us recall that the notation Gd \ F means that Gd \ is contained in the
interior of F . This implies that V(Gd \ F ) < +. The inclusion F B3d/2 (z)
and the inequality (iii) hold, in particular, for compact sets F which are small
neighborhoods (with smooth boundaries) of negligible compact subsets of Gd ,
and it is exactly such F s that we have in mind.
777
We shall refer to the sets F satisfying conditions (i)(iii) above as regular

ones.
Let P and PF denote the equilibrium potentials of Gd and F , respectively.
The equilibrium measure Gd has its support in Gd and has density P/
with respect to the (n 1)-dimensional Riemannian measure ds on Gd . So
for n 3 we have

P
(x) dsx , y Rn ;
E(x y)
P (y) =
G
d
P
(x) dsx = cap(Gd );
Gd
P (y) = 1
for all y Gd ,
0 P (y) 1 for all y Rn .
(If n = 2, then the same holds only with y B2d and with the fundamental
solution E replaced by the Green function G.) It follows that

P
P
ds =
(x) dsx dF (y) F (F ) = cap(F ).
PF
E(x y)
Gd
F Gd
Therefore,

cap(Gd ) cap(F )
(1 PF )
Gd
P
ds,
and, using Lemma 1, we obtain

cap(Gd ) cap(F )
2
2
P
ds
(1 PF )
Gd
1 PF 2L2 (Gd ) P 2L2 (Gd )

nL(d)1 cap(Gd ) 1 PF 2L2 (Gd ) ,
(18.2.10)
where L is dened by (18.2.8).

(b) Our next goal will be to estimate the norm 1PF L2 (Gd ) in (18.2.10)
by the norm of the same function in L2 (Gd ). We will use the polar coordinates
(r, ) as in (18.1.2), so in particular Gd is presented as the set {r()|
S n1 }, where r : S n1 (0, +) is a Lipschitz function (C as long as we
assume the boundary G to be C ). Assuming that v C 0,1 (Gd ), we can
write

r()n1
|v|2 ds =
|v|2
d
r
Gd
S n1

v r(), 2 r()n1 d,
(18.2.11)
L
S n1
where d is the standard (n 1)-dimensional volume element on S n1 .

Using the inequality
778


f ()2 2
we obtain

v r(), 2

2r()

f (t)2 dt + 2
2r()
[(1 )r()]n1
+

f (t)2 dt,

f C 0,1 [0, ] , > 0,

v (, )2 d +
r()
(1)r()
2
r()

v(, )2 d
r()
(1)r()

v (, )2 n1 d
r()
(1)r()
2
r()[(1 )r()]n1

v(, )2 n1 d.
r()
(1)r()
It follows that the integral in the right-hand side of (18.2.11) is estimated by

2r() d
(1 )n1
S n1
+
S n1

v (, )2 n1 d
r()
(1)r()

2 d
v(, )2 n1 d.
n1
(1 )
r()
Taking 1/2, we can majorize this by

2n
2n d
|v|2 dx +
|v|2 dx,
d
Gd
Gd
where (0, 1] is the constant from the description of G in Sect. 18.1. Recalling (18.2.11), we see that the resulting estimate has the form

2n L
|v|2 ds 2n Ld
|v|2 dx +
|v|2 dx.
d Gd
Gd
Gd
Now, taking v = 1 PF , we obtain

2n L
(1 PF )2 ds 2n Ld cap(F ) +
(1 PF )2 dx.
d
Gd
Gd
Using this estimate in (18.2.10), we obtain

cap(Gd ) cap(F )
1
2
n2 L cap(G d ) cap(F ) +
n
1
d2

(1 PF ) dx .
2
Gd
(18.2.12)
(c) Now let us consider G, which is starshaped with respect to a ball, but
does not necessarily have a C boundary. In this case we can approximate
the function r() (see Sect. 18.1) from above by a decreasing sequence of
779
C functions rk () (e.g., we can apply a standard mollifying procedure to

r() + 1/k), so that for the corresponding bodies G (k) the constants Lk are
uniformly bounded. It is clear that in this case we will also have k , and
(k)
cap(Gd ) cap(Gd ). So we can pass to the limit in (18.2.12) as k + and
conclude that it holds for an arbitrary G (which is starshaped with respect to
a ball). But for the moment we still retain the regularity condition on F .
(d) Let us dene

L = u : u C0 (), hV (u, u) + u 2L2 () 1 ,
(18.2.13)
where hV is dened by (18.1.1). By a standard functional analysis argument

the spectrum of HV is discrete if and only if L is precompact in L2 (), which,
in turn, holds if and only if for every > 0 there exists R > 0 such that

|u|2 dx for every u L.
(18.2.14)
\BR (0)
Equivalently, we can write that

|u|2 dx
|u|2 dx +
|u|2 V(dx) ,
\BR (0)
(18.2.15)
for every u C0 ().

Therefore, it follows from the discreteness of spectrum of HV that for every
> 0 there exists R > 0 such that for every Gd with Gd (Rn \ BR (0)) =
and every u C0 (Gd )

2
2
2
|u| dx
|u| dx +
|u| V(dx) .
(18.2.16)
Gd
Gd
Gd
In other words, = (Gd ) 0 as Gd for the best constant in (18.2.16).

(Note that (Gd )1 is the bottom of the Dirichlet spectrum of HV in Gd .)
Since 1 PF = 0 on F (hence in a neighborhood of Gd \ ), we can take
u = (1 PF ), where (0, 1) to be chosen later, C0 (Gd ) is a cuto
function satisfying 0 1, = 1 on G(1)d , and | | Cd1 with
C = C(G). Then, using integration by parts and the equation PF = 0 on
G \ F , we obtain

2
| |2 (1 PF )2 2 (1 PF )PF
|u| dx =
Gd
Gd

+ 2 |PF |2 dx

2
2
2
2
=
| | (1 PF ) dx C (d)
(1 PF )2 dx.
Gd
Therefore, from (18.2.16)
Gd
780


Gd

|u|2 dx C 2 (d)2
(1 PF )2 dx + V(Gd \ F ) ,
Gd
hence

(1 PF ) dx C (d)
2
G(1)d
(1 PF ) dx + V(G d \ F ) .
2
Gd
Now, applying the obvious estimate

(1 PF )2 dx
(1 PF )2 dx + mn (Gd \ G(1)d )
Gd
G(1)d
G(1)d
(1 PF )2 dx + C1 dn
with C1 = C1 (G), we see that

2
2
2
2
(1 PF ) dx C (d)
(1 PF ) dx + V(G d \ F ) + C1 dn ,
Gd
Gd
hence

Gd
(1 PF )2 dx 2V(Gd \ F ) + 2C1 dn ,
provided
C 2 (d)2 1/2.
Returning to (18.2.12) and using (18.2.17) we obtain

2

cap(F )
1 n
2
C2 + d
(1 PF ) dx
1
cap(Gd )
Gd

C2 + 2C1 1 + 21 dn V(Gd \ F ) ,
(18.2.17)
(18.2.18)
(18.2.19)
where C2 = C2 (G). Without loss of generality we will assume that C2 1/2.

Recalling that cap(F ) cap(Gd ), we can replace the ratio cap(F )/ cap(Gd )
in the left-hand side by . Now let us choose
=
(1 )2
,
4C2
(1 )2
(1 )4
=
.
8C1
32C1 C2
(18.2.20)
Then 1/2 and for every xed (0, 1) and d > 0 the condition (18.2.18)
will be satised for distant bodies Gd because = (Gd ) 0 as Gd .
(More precisely, there exists R = R(, d) > 0, such that (18.2.18) holds for
every Gd such that Gd (Rn \ BR (0)) = .)
If and are chosen according to (18.2.20), then (18.2.19) becomes
dn V(Gd \ F ) (16C2 )1 (1 )4 ,
(18.2.21)
which holds for distant bodies Gd if (0, 1) and d > 0 are arbitrarily xed.
18.3 Discreteness of Spectrum: Suciency
781
(e) Up to this moment we worked with regular sets F see conditions

(i)(iii) in part (a) of this proof. Now we can get rid of the regularity requirements (i) and (ii), retaining (iii). So let us assume that F is a compact set,
Gd \ F Gd and cap(F ) cap(Gd ) with (0, 1). Let us construct a
sequence of compact sets Fk F , k = 1, 2, . . . , such that every Fk is regular,
F1 F2 . . . ,
and
Fk = F.
k=1
We have then cap(Fk ) cap(F ) as k + due to the well-known continuity

property of the capacity (see, e.g., Sect. 2.2.1). According to the previous steps
of this proof, the inequality (18.2.21) holds for distant Gd s if we replace F
by Fk and by k = cap(Fk )/ cap(Gd ). Since the measure V is positive, the
resulting inequality will still hold if we replace V(Gd \Fk ) by V(Gd \F ). Taking
limit as k +, we obtain that (18.2.21) holds with = cap(F )/ cap(Gd )
instead of . Since , (18.2.21) immediately follows for arbitrary compact
F such that Gd \ F Gd and cap(F ) cap(Gd ) with (0, 1).
(f) Let us x G and take the inmum over all negligible F s (i.e., compact
sets F , such that Gd \ F Gd and cap(F ) cap(Gd )) in the right-hand
side of (18.2.21). We then get for distant Gd s
dn
inf
F N (Gd ,)
V(Gd \ F ) (16C2 )1 (1 )4 .
(18.2.22)
Now let us recall that the discreteness of spectrum is equivalent to the condition = (Gd ) 0 as Gd (with any xed d > 0). If this is the case,
then it is clear from (18.2.22) that for every xed (0, 1) and d > 0, the
left-hand side of (18.2.22) tends to + as Gd . This concludes the proof
of part (i) of Theorem 18.1/1.

18.3 Discreteness of Spectrum: Suciency

In this section we will establish the suciency part of Theorem 18.1/1.
Let us recall the Poincare inequality

u(x)2 dx,
u u
2L2 (Gd ) A(G)d2
Gd
u C 0,1 (Gd ),
where Gd R was described in Sect. 18.1,

1
u
=
u(x) dx
|Gd | Gd
n
is the mean value of u on Gd , |Gd | is the Lebesgue volume of Gd , and A(G) > 0
is independent of d.
782

The following Lemma slightly generalizes (to an arbitrary body G) a particular case of the rst part of Theorem 14.1.2 with essentially the same proof.
Lemma 1. There exists C(G) > 0 such that the following inequality holds
for every function u C 0,1 (Gd ) which vanishes on a compact set F Gd (but
is not identically 0 on Gd ):

C(G) Gd |u(x)|2 dx

.
(18.3.1)
cap(F )
|Gd |1 Gd |u(x)|2 dx
The next lemma is an obvious adaptation of Lemma 16.1 to general test
bodies Gd (instead of cubes Qd ). In its proof Lemma 1 should be used.
Lemma 2. Let V be a positive Radon measure in . There exists C2 (G) > 0
such that for every (0, 1) and u C 0,1 (Gd ) with u = 0 in a neighborhood
of Gd \ ,

C2 (G)d2
C2 (G)dn
|u|2 dx
|u|2 dx +
|u|2 V(dx), (18.3.2)
V (Gd , ) Gd
Gd
Gd
where
V (Gd , ) =
inf
F N (Gd ,)
V(Gd \ F ).
(18.3.3)
(Here the negligibility class N (Gd , ) was introduced in Denition 18.1.)

Now we move to the proof of the suciency part in Theorem 18.1/1.
We start with the following proposition, which gives a general (albeit complicated) sucient condition for the discreteness of spectrum.
Proposition 1. Given an operator HV , let us assume that the following condition is satised: there exists 0 > 0 such that for every (0, 0 ) we can nd
d = d() > 0 and R = R() > 0, so that if Gd satises Gd ( \ BR (0)) = ,
then there exists = (Gd , ) (0, 1) such that
d2 1
and
dn V (Gd , ) 1 .
(18.3.4)
Then the spectrum of HV is discrete.

Proof. Recall that the discreteness of spectrum is equivalent to the following condition: For every > 0 there exists R > 0 such that (18.2.15) holds
for every u C0 (). This will be true if we establish that for every > 0
there exist R > 0 and d > 0 such that

2
2
2
|u| dx
|u| dx +
|u| V(dx)
(18.3.5)
Gd
Gd
Gd
for all Gd such that Gd ( \ BR (0)) = and for all u C (Gd ), such
that u = 0 in a neighborhood of Gd \ . Indeed, assume that (18.3.5) is true.
18.4 A Suciency Example
783
Let us take a covering of Rn by bodies Gd so that it has a nite multiplicity

m = m(G) (i.e., at most m bodies Gd can have a nonempty intersection).
Then, taking u C0 () and summing up the estimates (18.3.5) over all
bodies Gd with Gd ( \ BR (0)) = , we obtain (18.2.15) (hence (18.2.14))
with m instead of .
Now Lemma 18.3/2 and the assumptions (18.3.4) immediately imply
(18.3.5) (with replaced by C2 (G)).

Instead of requiring that the conditions of Proposition 1 are satised for
all (0, 0 ), it suces to require it for a monotone sequence k +0. We
can also assume that d(k ) 0 as k +. Then, passing to a subsequence,
we can assume that the sequence {d(k )} is strictly decreasing. Keeping this
in mind, we can replace the dependence d = d() by the inverse dependence
= g(d), so that g(d) > 0 and g(d) 0 as d +0 (and here we can
also restrict ourselves to a sequence dk +0). This leads to the following,
essentially equivalent, but more convenient reformulation of Proposition 1.
Proposition 2. Given an operator HV , assume that the following condition is satised: There exists d0 > 0 such that for every d (0, d0 ) we can
nd R = R(d) > 0 and = (d) (0, 1), so that if Gd ( \ BR (0)) = ,
then
(18.3.6)
d2 g(d)1 and dn V (Gd , ) g(d)1 ,
where g(d) > 0 and g(d) 0 as d +0. Then the spectrum of HV is discrete.
Proof of Theorem 18.1/1, part (ii). Instead of (ii) in Theorem 18.1/1 it
suces to prove the (stronger) statement formulated in Remark 18.1/4. So
suppose that d0 > 0, c > 0, d (0, d0 ), R = R(d) > 0, (d) (0, 1),
satisfying (18.0.1), such that (18.1.6) holds for all Gd with Gd (\BR (0)) = .
Since the left-hand side of (18.1.6) is exactly dn V(d) (Gd , ), we see that
(18.1.6) can be rewritten in the form
dn V (Gd , ) cd2 (d),
hence we can apply Proposition 18.3 with g(d) = c1 d2 (d)1 to conclude
that the spectrum of HV is discrete.


In this section we prove Theorem 18.1/2. We shall construct a domain
Rn , such that the condition (18.1.7) is satised with (d) = Cd2 (with an
arbitrarily large C > 0), and yet the spectrum of in L2 () (with the
Dirichlet boundary condition) is not discrete. This will prove Theorem 18.1/2
showing that the condition (18.0.1) is precise. We assume that n 3.
Let us use the following notations:
784

L(j) is the spherical layer {x Rn : log j |x| log(j + 1)}. Its width is
log(j + 1) log j which is < j 1 for all j and equivalent to j 1 for large j.
(j)
{Qk }k1 is a collection of closed cubes that form a tiling of Rn and have
edge length (n) j 1 , where (n) is a suciently small constant depending
on n (to be adjusted later).
(j)
(j)
xk is the center of Qk .
(j)
(j)
{Bk }k1 is the collection of closed balls centered at xk with radii j
given by

n
= C (n)/j ,
n (n 2) n2
j
where n is the area of the unit sphere S n1 Rn and C is an arbitrary
constant. The last equality can be written as
(j)
(j)
cap Bk = C mn Qk .
(18.4.1)
(j)
Among the balls Bk we select a subcollection which consists of the balls

(j)
with the additional property Bk L(j) . We shall refer to these balls as
(j) . By an abuse of notation
selected ones and denote selected balls by B
k
we do not introduce a special letter for the subscripts of the selected balls.
(j) the corresponding cubes Q(j) , so that
We also denote by Q
k
k
(j) .
(j) = Q(j) B
Q
k
k
k

(j) L(j) .
B
k

is the complement of j1 (j) .
Br (P ) is the closed ball with radius r 1 centered at a point P . We shall
make a more precise choice of r later.
(j) =
k1
Proposition 1. The spectrum of in (with the Dirichlet boundary

condition) is not discrete.
Proof. Let j 7 and P L(j) , i.e.,
log j |P | log(j + 1).

Note that the ball Br (P ) is a subset of the spherical layer lsm L(s) if and
only if
log m |P | r and |P | + r log(l + 1).
Therefore, if
log m log j r
and
then Br (P )
log(j + 1) + r log(l + 1),

lsm
(s)
. The last two inequalities can be written as
m j er
j + 1 (l + 1)er .
and
785
(18.4.2)
If we take, for example,

m = [j/3] and
l = 3j,
then, by the inequality j 7, we deduce that

L(s) .
Br (P )
(18.4.3)
[j/3]s3j
Using (18.4.2), the denition of and subadditivity of capacity, we obtain

(s)
cap Br (P ) \ = cap Br (P )

[j/3]s3j k1
s1

(s)
cap Br (P ) B
k

(s)
.
mn Q
k
[j/3]s3j {k:Br (P )Q
(s)
=}
k
(s) ,
We see that the multiplicity of the covering of Br (P ) by the cubes Q
k
participating in the last sum, is at most 2, provided (n) is chosen suciently
small. Hence,

(18.4.4)
cap Br (P ) \ c(n)Cr n .
On the other hand, we know that the discreteness of spectrum guarantees
that for every r > 0

lim inf cap Br (P ) \ (n) r n2 ,
|P |
where (n) is a constant depending only on n (cf. Remark 18.1/7). For suciently small r > 0 this contradicts (18.4.4).

Proposition 2. The domain satises

lim inf cap Br (P ) \ (n) C r n ,
|P |
(18.4.5)
where (n) > 0 depends only on n.

(s)
(s) (extended by zero to
Proof. Let k be the capacitary measure on B
k
(s) ), and let 1 (n) denote a suciently small constant to be chosen
Rn \ B
k
later. We introduce the measure
(s)
k ,
= 1 (n)
k,s
786

where the summation here and in the following is taken over k, s which cor (s) . Taking P L(j) , let us show that
respond to the selected balls B
k

E(x y) d(y) 1 on Rn ,
(18.4.6)
Br/2 (P )
where E(x) is given by (18.2.3). It suces to verify (18.4.6) for x Br (P )

because for x Rn \ Br (P ) this will follow from the maximum principle.
Obviously, the potential in (18.4.6) does not exceed

(s)
1 (n)
E(x y) dk (y).
(s)
B
k
(s)
Br/2 (P )
=}
{s,k:B
k

and
, the rst sum being extended
We divide this sum into two parts
(s)
over all points xk with the distance j 1 from x. Recalling that x Br (P )
and using (18.4.3), we see that the number of such points does not exceed a
certain constant c1 (n). We dene the constant 1 (n) by

1
.
1 (n) = 2c1 (n)
(s)
Since k is the capacitary measure, we have

1 (n) c1 (n) = 1/2.
Furthermore, by (18.4.1)

c2 (n)

= c2 (n) C
(s)

c3 (n) C
We can assume that
(s)
)
cap(B
k

(s)
mn Q
k
(s)
|x xk |n2
|x xk |n2
dy
< c4 (n) C r2 .
|x
y|n2
Br (P )
r (2c4 (n)C)1/2 ,

1/2. Therefore (18.4.6) holds.
which implies
It follows that for large |P | (i.e., for P with |P | R = R(r) > 0), or
equivalently, for large j, we have

(s) (s)
cap Br (P ) \
1 (n)k B
k
(s)
Br/2 (P )}
{s,k:B
k
= 1 (n)
(s)
cap B
k
(s)
Br/2 (P )}
{s,k:B
k
= 1 (n) C
(s)
Br/2 (P )}
{s,k:B
k
(s)
mn Qk (n) C rn .
18.6 Structure of the Essential Spectrum of HV
This ends the proof of Proposition 2, hence of Theorem 18.1/2.
787

Remark. Slightly modifying the previous construction, we provide an example of an operator H = + V (x) with V C (Rn ), n 3, V 0, such
that the corresponding measure V dx satises (18.1.5) with (d) = Cd2 and
an arbitrarily large C > 0, but the spectrum of H in L2 (Rn ) is not discrete. So
the condition (18.0.1) is precise even in the case of the Schrodinger operators
with C potentials.
18.5 Positivity of HV
In this section we prove Theorem 18.1/3.
Proof of Theorem 18.1/3 (Necessity). Let us assume that the operator HV
is strictly positive. This implies that the estimate (18.2.16) holds with some
> 0 for every Gd (with an arbitrary d > 0) and every u C0 (Gd ). But
then we can use the arguments of Sect. 18.2 which lead to (18.2.22), provided
(18.2.18) is satised. It will be satised if d is chosen suciently large.

Proof of Theorem 18.1/3 (Suciency). Let us assume that there exist
d > 0, > 0 and (0, 1) such that for every Gd the estimate (18.1.9) holds.
Then by Lemma 18.3/2, for every Gd and every u C (Gd ), such that u = 0
in a neighborhood of Gd \ , we have

C2 (G)d2
C2 (G)dn
|u|2 dx
|u|2 dx +
|u|2 V(dx).
Gd
Gd
Gd
Let us take a covering of Rn of nite multiplicity N by bodies Gd . It follows
that for every u C0 ()

1 dn2
2
2
2
2
,
|u| dx N C2 (G)d max
|u| dx +
|u| V(dx) ,

which proves positivity of HV .
18.6 Structure of the Essential Spectrum of HV

Lemma. If the spectrum of HV is not purely discrete, the essential spectrum
of HV extends to innity. Moreover, if 0 belongs to the essential spectrum of
HV , then this spectrum coincides with [0, ).
Proof. Let be the bottom of the essential spectrum. Then there exists a
sequence of real-valued functions { }1 in C0 () subject to the conditions
L2 () = 1,
0 weakly in L2 (),
(18.6.1)
788


(HV )
We set
L2 ()
0.
(18.6.2)

n

xk ,
u = exp i( )1/2
k=1
where > . We see that u satises (18.6.1) and that

2
n

2
2

(HV )u
(HV )
=
+
4(
/x

k
L2 ()
L2 ()

k=1
L2 ()
Since the right-hand side does not exceed

(HV ) 2
+ 4n( )Q[ , ],
L2 ()
we have

2

(HV )u 2
(HV ) L () + 4n( )(HV ) L ()
L2 ()
2
2
+ 4n( ).
By (18.6.2)

lim sup(HV )u L
2 ()
(),
where () = 2(n ())1/2 . It follows that any segment [(), +()]

contains points of the essential spectrum. If, in particular, = 0 then every
positive belongs to the essential spectrum.

In concert with this lemma the pairs (, V) can be divided into three
nonoverlapping classes.
(i) The rst class includes (, V) such that the spectrum of HV is
discrete.
(ii) The pair (, V) belongs to the second class if the essential spectrum
of HV is unbounded and strictly positive.
(iii) Finally, (, V) is of the third class if the essential spectrum of HV
coincides with [0, ).
By Theorem 18.1, condition (i) is equivalent to (18.1.5) and (ii) holds if
and only if (18.1.9) is valid. Finally, (iii) is equivalent to the failure of (18.1.7)
by Theorem 18.1/3 and Lemma. In other words, (iii) holds if and only if
lim inf
inf
Gd F N(d) (Gd ,)
for every d > 0.
d \ F ) = 0
V(G
18.7 Two-Sided Estimates of the First Eigenvalue of the Dirichlet Laplacian
789
18.7 Two-Sided Estimates of the First Eigenvalue of the

Dirichlet Laplacian
18.7.1 Main Result
Let us consider an open set Rn , n > 2 and denote the bottom of the
spectrum of its minus Dirichlet Laplacian ()Dir by (), i.e.,

|u|2 dx

() = inf
.
(18.7.1)
uC0 ()
|u|2 dx
Obviously, implies () ( ). In particular, if Br is an open

ball of radius r, such that Br , then () (Br ) = Cn r 2 where
Cn = (B1 ). It follows that for the interior radius of , which is dened as
r = sup{r : Br },
we have
2
.
() Cn r
However, this estimate is not good for unbounded domains or domains with
complicated boundaries. For instance, a similar estimate for () from below
does not hold for the complement of any Cartesian coordinate grid in Rn
when () = 0 and r < .
The way to improve this estimate is to relax (as in Sect. 15.4) the requirement for Br to be completely inside by allowing some part of Br , which
has a small harmonic capacity, to stick out of . Namely, let us take an arr negligible (or more precisely,
bitrary (0, 1) and call a compact set F B
-negligible) if
r ).
(18.7.2)
cap(F ) cap(B
(Here cap(F ) denotes the Wiener (harmonic) capacity of F , cap(F ;
L12 (Rn )).)
Now write
r \ is -negligible}.
r, = sup{r : Br , B
This is the interior capacitary radius.
Theorem. Let us x (0, 1). Then there exist c = c(, n) > 0 and
C = C(, n) > 0, such that for every open set Rn
2
2
cr,
() Cr,
.
(18.7.3)
Explicit values of constants c = c(, n) and C = C(, n) are provided in

(18.7.22) and (18.7.38), respectively.
Let us formulate some corollaries of this theorem.
Corollary 1. () > 0 if and only if r, < .
790

This corollary gives a necessary and sucient condition of strict positivity

of the operator ()Dir in .
Since the condition () > 0 does not contain , we immediately obtain
the following corollary.
Corollary 2. Conditions r, < , taken for dierent s, are equivalent.
Writing F = Rn \ (which can be an arbitrary closed subset in Rn ), we
obtain from the previous corollary (comparing = 0.01 and = 0.99):
Corollary 3. Let F be a closed subset in Rn , which has the following
property: There exists r > 0 such that
r ) 0.01 cap(B
r )
cap(F B
for all Br . Then there exists r1 > 0 such that
r )
r ) 0.99 cap(B
cap(F B
1
1
for all Br1 .
The inequalities (18.7.3) for suciently small > 0 were established in
Chaps. 14 and 15. The Theorem provides a substantial improvement, in particular, allowing Corollaries 2 and 3 and providing explicit values of the constants.
18.7.2 Lower Bound
In this section we will establish the lower bound for () from Theorem
18.7.1, which is an easier part of this theorem. The key part of the lower
bound proof is presented in the following lemma, which is a particular case of
a more general Theorem 14.1.2, part 1, though without an explicit constant,
which we provide to specify explicit constants in Theorem 18.7.1.
Lemma 1. The following inequality holds for every complex-valued funcr ) which vanishes on a compact set F B
r (but is not
tion u C 0,1 (B
identically 0 on Br ):

Cn Br |u(x)|2 dx
cap(F ) n
,
(18.7.4)
r
|u(x)|2 dx
Br

where
Cn = 4n
2
1 2
n
(18.7.5)
Beginning of Proof. A. Clearly, it is sucient to consider the ball Br cenr ). By scaling we see that it
tered at 0, and real-valued functions u C 0,1 (B
791
suces to consider the case r = 1. (The corresponding estimate for an arbitrary r > 0 follows from the one with r = 1 with the same constant Cn .) So
we need to prove the estimate

Cn
2
|u| dx
|u|2 dx,
(18.7.6)
cap(F ) B1
B1
1 , u C 0,1 (B
1 ), and u|F = 0.
where F is a compact subset of B
n
0,1
Consider the function U C (R )

1 |u(x)|,
if |x| 1,
U (x) =
|x|2n (1 |u(|x|2 x)|), if |x| 1,
i.e., U extends 1 |u| to {x : |x| 1} as the Kelvin transform of 1 |u|.
Clearly, U |F = 1, |U | = |u| almost everywhere in B1 , U (x) = O(|x|2n )
and |U (x)| = O(|x|1n ) as |x| . It follows that U can serve as a test
function in (18.2.1), i.e.,

|U |2 dx.
(18.7.7)
cap(F )
Rn
Using the harmonicity of |x|2n and the GreenStokes formula, we obtain by
a straightforward calculation

2

2
2
1 u() d, (18.7.8)
|U | dx = 2
|u| dx + (n 2)
n
R
B1
B1
where d means the area element on B1 .
B. For a function v on B1 dene its average

1
v d =
v d.
v =
n B1
B1
To continue the proof of Lemma 1, we will need the following elementary
Poincare-type trace inequality.
Lemma 2. For any v C 0,1 (B1 ),

2
|v v| d
B1
|v|2 dx.
(18.7.9)
B1
Proof of Lemma 2. It suces to prove it for real-valued functions v. Let

us expand v in spherical functions. Let
{Yk,l | l = 0, 1, . . . , nk , k = 0, 1, . . . , }
792

be an orthonormal basis in L2 (B1 ) which consists of eigenfunctions of the

(negative) LaplaceBeltrami operator on B1 , so that the eigenfunctions
Yk,l = Yk,l () with a xed k have the same eigenvalue k(k+n2) (which has
multiplicity nk +1). Note that the zero eigenvalue (corresponding to k = 0) has
1/2
multiplicity 1 and Y0,0 = const = n
for the corresponding eigenfunction.
Writing x = r, where r = |x|, = x/|x|, we can present v in the form

vk,l (r)Yk,l ().
(18.7.10)
v(x) = v(r, ) =
k,l
Then

v(x)2 dx =
B1
k,l
and

vk,l (r)2 r n1 dr
(18.7.11)

v()2 d =
vk,l (1)2 .
B1
It follows that
1
(18.7.12)
k,l

v() v2 d =
B1

vk,l (1)2 .
(18.7.13)
{k,l:k1}

2
v
|v| = + r2 | v|2 ,
r
2
where means the gradient along the unit sphere with variable and xed r,
we also get

1
2 k(k + n 2)

vk,l
vk,l (r)2 rn1 dr.
|v|2 dx =
(r) +
r2
B1
0
k,l
(18.7.14)
Comparing (18.7.13) and (18.7.14), and taking into account that k(k + n 2)
increases with k, we see that it suces to establish that the inequality

1
2 n 1

g (r) +
g(r)2 rn1 dr
g(1)2
r2
0
holds for any real-valued function g C 0,1 ([0, 1]). To this end write
1
1
n2 2
n2

r
2r
g dr =
g g + (n 2)rn3 g 2 dr
g(1)2 =
0
0

1
1
n1 2

n 1 2 n1
g 2 +
r
r
g + (n 1)rn3 g 2 dr =
g
dr,
r2
0
0
which proves Lemma 2.

793
Proof of Lemma 1 (continuation). C. Let us normalize u by requiring |u| = 1,

i.e., average of |u| over B1 equals 1. Then by Lemma 2

2
1 |u| d
|u|2 dx.
B1
B1
Combining this with (18.7.7) and (18.7.8), we obtain

|u|2 dx.
cap(F ) n
B1
Removing the restriction |u| = 1, we can conclude that for any u C 0,1 (B1 )
2
|u| d
B1
n
cap(F )

|u|2 dx.
(18.7.15)
B1
Note that for any real-valued function v C 0,1 (B1 )

|v v|2 d =
|v|2 d v2 ,
B1
B1
hence, using (18.7.9), we get

2
2
|v| d = v +
B1
1
|v v| d v +
n
B1
2

|v|2 dx.
Applying this to v = |u| and using (18.7.15), we obtain

n n
|u|2 d 1 +
|u|2 dx.
cap(F
)
B1
B1
B1
(18.7.16)
D. Note that our goal is an estimate that is similar to (18.7.16), but with
the integral over B1 in the left-hand side replaced by the integral over B1 .
To this end we again use the expansion (18.7.10) of v = |u| over spherical
functions, and the identities (18.7.11), (18.7.12), and (18.7.14). Let us take a
real-valued function g C 0,1 ([0, 1]) and write
g 2 (r)rn1 dr.
Q=
0
Integrating by parts, we obtain

Q=
2
n
1
0
gg r n dr +
1 2
g (1).
n
Using the inequality 2ab a2 + 1 b2 , where a, b R, > 0, and taking into

account that r 1, we obtain
794

1

1
1
g 2 (r) + g 2 (r) rn1 dr + g 2 (1)
n
0
1
1
1
g 2 (r)rn1 dr + g2 (1),
= Q+
n
n 0
n
1
n
hence for any (0, n)

1
Q
(n )
g 2 (r)rn1 dr +
1
g 2 (1).
n
Taking = n/2, we obtain

4
Q 2
n
g 2 (r)rn1 dr +
2 2
g (1).
n
(18.7.17)
Now we can argue as in the proof of Lemma 2, expanding v = |u| over spherical
harmonics Yk,l . Then the desired inequality follows from the inequalities for
the coecients vk,l = vk,l (r), with the strongest one corresponding to the
case k = 0 (unlike k = 1 in Lemma 2). Then using the inequality (18.7.17)
for g = v0,0 we obtain

4
2
|u|2 dx 2
|u|2 dx +
|u|2 d.
(18.7.18)
n B1
n B1
B1
Using (18.7.16), we deduce from (18.7.18)

2
4
nn
|u|2 dx
+
|u|2 dx.
1
+
n2
n
cap(F )
B1
B1
(18.7.19)
Taking into account the inequality

1 ) = (n 2)n ,
cap(F ) cap(B
we can estimate the constant in front of the integral in the right-hand side of
(18.7.19) as follows:

2
4
nn
4n
2
+
1+
1 2 ,
n2
n
cap(F )
cap(F )
n
which ends the proof of Lemma 1.

The lower bound in (18.7.3) is given by the following lemma.

Lemma 3. There exists c = c(, n) > 0 such that for all open sets Rn
2
() c r,
.
(18.7.20)
795
r has
Proof. Let us x (0, 1) and choose any r > r, . Then any ball B
n
a nonnegligible intersection with R \ , i.e.,
r ).
r \ ) cap(B
cap(B
r \ , it follows from Lemma 1 that for
Since any u C0 () vanishes on B
any such u

Cn
Cn
2
2
|u| dx n
|u| dx n
|u|2 dx.
r
cap
(
B
\
)
r
cap
(
B
)
r
r
Br
Br
Br
1 )r n2 , we obtain
r ) = cap(B
Taking into account that cap(B

Cn r 2
2
2
|u| dx
1 ) B |u| dx.
cap(B
r
B
r
r = B
r(k) , k = 1, 2, . . . , so that
Now let us choose a covering of Rn by balls B
the multiplicity of this covering is at most N = N (n). For example, we can
make
N (n) n log n + n log(log n) + 5n, n 2,
(18.7.21)
which holds also for the smallest multiplicity of coverings of Rn by translations
of any convex body (see Theorem 3.2 in Rogers [679]).
Then summing up the previous estimates over all balls in this covering,
we see that

Cn r2
2
2
|u| dx
|u| dx
|u|2 dx
1 )
cap(
B
r(k)
r(k)
B
B
Rn
k
k

Cn N r 2
2
1 ) Rn |u| dx.
cap(B
Recalling (18.7.1), we see that
() cr2 ,
with
c = c(, n) =
1 )
cap(B
n2 (n 2)
=
.
Cn N
4(n2 2)N
Taking the limit as r r, , we obtain (18.7.20) with the same c.
(18.7.22)

18.7.3 Upper Bound

We divide the proof of the upper bound into parts.
1. According to (18.7.1), to get an upper bound for () it is enough to
take any test function u C0 () and write
796


|u|2 dx
()
.
|u|2 dx
(18.7.23)
For simplicity of notations we shall write instead of () everywhere in

this section. The inequality (18.7.23) can be written as follows:

|u|2 dx 1
|u|2 dx.
(18.7.24)
By approximation, it suces to take u C00,1 () or even u

L12 ().
In particular, choosing a ball Br , we can take
u C00,1 ( Br ) = C00,1 () C00,1 (Br ).
(18.7.25)
3r/2 , such that F is the closure of an open set

Let us take a compact set F B
with a smooth boundary. (In this section we will call such sets regular subsets
3r/2 .) Denote by PF its capacitary potential (see Sect. 18.2). Regularity
of B
of F implies that PF C 0,1 (Rn ). By denition PF = 1 on F , so 1 PF = 0
on F . Let us also assume that
r \ ,
Int F B
where Int F means the set of all interior points of F . Then 1 PF = 0 in a
r \ . Therefore, multiplying 1 PF by a cuto function
neighborhood of B
C0 (Br ), we will get a function u = (1 PF ), satisfying the requirement
(18.7.25), hence t to be a test function in (18.7.23).
In the future we will also assume that the cuto function C0 (Br ) has
the properties
0 1 on Br ,
|| 2(r)1
= 1 on B(1)r ,
on Br ,
where 0 < < 1 and the balls Br and B(1)r are supposed to have the same
center. Using integration by parts and the equation PF = 0 on Br \ F , we
obtain for the test function u = (1 PF )

||2 (1 PF )2 2 (1 PF )PF + 2 |PF |2 dx
|u|2 dx =
Br
Br

=
||2 (1 PF )2 dx 4(r)2
(1 PF )2 dx.
Br
Br
Therefore, from (18.7.24) we obtain

2
1
2
|u| dx 4(r)
Br
(1 PF )2 dx.
Br
Since 0 PF 1, the last integral in the right-hand side is estimated by

mn (Br ) = n1 n r n .
Therefore,
797
|u|2 dx 4n1 n 1 2 r n2 .
Br
Restricting the integral in the left-hand side to B(1)r , we obtain

(1 PF )2 dx 4n1 n 1 2 rn2 .
(18.7.26)
B(1)r
2. Now we need to provide an appropriate lower bound for the left-hand

side of (18.7.26). To this end we rst restrict the integration to the spherical
layer
Sr1 ,r2 = Br2 \ Br1 ,
where 0 < r1 < r2 < r. In the sequel we shall take
r1 = (1 2)r,
r2 = (1 )r,
(18.7.27)
where 0 < < 1/2, though it is convenient to write some formulas in a greater
generality. Let us denote the volume of the layer Sr1 ,r2 by |Sr1 ,r2 |, i.e.,

|Sr1 ,r2 | = mn Sr1 ,r2 = n1 n r2n r1n .
We also need the notation

f (x) dx =
Sr1 ,r2
1
|Sr1 ,r2 |

f (x) dx
Sr1 ,r2
for the average of a positive function f over Sr1 ,r2 . In particular, restricting
the integration in (18.7.26) to Sr1 ,r2 (with r1 , r2 as in (18.7.27)) and dividing
by |Sr1 ,r2 |, we obtain

41 2 rn2
(1 PF )2 dx
.
r2n r1n
Sr1 ,r2
Hence, by the CauchySchwarz inequality,

2
2

41 2 r n2
1
PF dx =
(1 PF ) dx
. (18.7.28)
r2n r1n
Sr1 ,r2
Sr1 ,r2
To simplify the right-hand side, let us estimate (r2n r1n )1 from above. We
see that

r2n r1n = (r2 r1 ) r2n1 + r2n2 r1 + + r1n1

nrr1n1 = nrn (1 2)n1 nrn 1 2(n 1) .
Now note that
1
1 + 4(n 1),
1 2(n 1)
798

provided
0<
1
.
4(n 1)
Under this condition it follows that

1
n1 1 r n 1 + 4(n 1) ,
n
n
r2 r1
and (18.7.28) takes the form

2

1
PF dx 4n1 3 1 + 4(n 1) 1 r2 .
(18.7.29)
(18.7.30)
(18.7.31)
Sr1 ,r2
3. For simplicity of notation and without loss of generality we may assume

that the ball Br is centered at 0 Rn (and so are smaller balls and spherical
layers).
To provide a lower bound for the left-hand side of (18.7.31), we will give
an upper bound for the average of PF . According to the denition of PF and
notations from Sect. 18.7.1, we can write

PF dx =
E(x y) dF (y) dx
Sr1 ,r2
Sr
1 2

,r

E(x y) dx dF (y).
(18.7.32)
Sr1 ,r2
The inner integral on the right-hand side can be explicitly calculated as the
potential of a uniformly charged spherical layer with total charge 1. The result
of this calculation is |Sr1 ,r2 |1 Vr1 ,r2 (y), where
2 2
r2 r1
,
if |y| r1 ,
2(n2)
r22
r1n
|y|2
(18.7.33)
Vr1 ,r2 (y) = 2n + 2(n2) n(n2)|y|n2 , if r1 |y| r2 ,
n
n
r2 r1
if |y| r2 .
n(n2)|y|n2 ,
The function y Vr1 ,r2 (y) belongs to C 1 (Rn ) and is spherically symmetric; it
tends to 0 as |y| ; it is harmonic in Rn \ Sr1 ,r2 and satises the equation
Vr1 ,r2 = 1 in Sr1 ,r2 . These properties uniquely dene the function Vr1 ,r2 .
Dierentiating it with respect to |y|, we easily see that it is decreasing with
respect to |y|, hence its maximum is at y = 0 (hence given by the rst row in
(18.7.33)). So we obtain, using (18.7.30),

n(r22 r12 )
E(x y) dx |Sr1 ,r2 |1 Vr1 ,r2 (0) =

2(n 2)n (r2n r1n )
Sr1 ,r2
nr2 (1 )
nr(r1 + r2 )
2(n 2)n (r2n r1n )

(n 2)n (r2n r1n )
1 + (4n 5)
(1 )[1 + 4(n 1)]
(n 2)n r n2
(n 2)n r n2
Finally, using the equality cap(Br ) = (n 2)n rn2 , we obtain

1 + (4n 5)
E(x y) dx
r ) ,
cap(B
Sr1 ,r2
799
(18.7.34)
provided r1 , r2 chosen as in (18.7.27) and (18.7.29) is satised.

4. Using (18.7.34) in (18.7.32) and taking into account (18.2.5), we obtain

1 + (4n 5)
PF (x) dx
dF (y)
r )
cap(B
Sr1 ,r2
F

cap(F )

= 1 + (4n 5)
1 + (4n 5) , (18.7.35)
cap(Br )
provided F is -negligible. (i.e., satises (18.7.2)). Taking into account
(18.7.29), we can set

1
1
,
(18.7.36)
,
= min
4(n 1) 2(4n 5)
so that (18.7.29) is satised, and besides,

1+
1
1 + (4n 5)
=1
,
2
2
so that (18.7.35) becomes
PF (x) dx 1
Sr1 ,r2
1
.
2
Taking this into account in (18.7.31) and using (18.7.29), we obtain

(1 )2
4n1 3 1 + 4(n 1) 1 r 2 8n1 3 1 r2 ,
4
hence
32(1 )2 3 r 2 .
(18.7.37)
We are now ready to prove Theorem 18.7.1.

The lower bound for was established in Lemma 18.7.2/3. We obtained
the estimate (18.7.37) under the condition that there exist (0, 1), a ball
3r/2 (here the balls Br and B3r/2 have
Br , and a regular compact set F B
r \ .
the same center), such that F is -negligible and its interior includes B
(The estimate then holds with = (, n) given by (18.7.36).) It follows, in
r \ is -negligible.
particular, that B
r \ is -negligible, then we can approximate it by regular
Conversely, if B
r \ , Int Fk Fk+1 , and
compact sets Fk , k = 1, 2, . . . , such that Int Fk B
Br \ is the intersection of all Fk s. Then
800

r \ )
lim cap(Fk ) = cap(B
by the continuity property of the capacity (see Sect. 2.2.1). In this case, for
any > 0 the sets Fk will be ( + )-negligible for suciently large k. It
follows that the estimate (18.7.37) will hold if we only know that there exists
r \ is -negligible. Then the estimate still holds if we
a ball Br such that B
replace r by the least upper bound of the radii of such balls, which is exactly
the interior capacitary radius r, . This proves the upper bound in (18.7.3)
with
(18.7.38)
C(, n) = 32(1 )2 3 ,
where is dened by (18.7.36).

18.7.4 Comments to Chap. 18

Sections 18.118.5. The material is borrowed from the article by Mazya
and Shubin [590], which improves the results of Sect. 18.4 for the particular
case p = 2, l = 1.
Section 18.2. A survey of the necessary and sucient conditions ensuring
various spectral properties of the Schr
odinger operator can be found in [563].
M.E. Taylor gave an alternative formulation of the discreteness of spectrum
criterion formulated in terms of the so-called scattering length [748].
The necessary and sucient conditions for the discreteness and strict positivity of magnetic Schrodinger operators with a positive scalar potential
2
n

+
ia
+V
j
xj
j=1
are obtained by Kondratiev, Mazya, and Shubin in [450] and [451].
Section 18.6. Lemma 18.6 is due to Glazman [309].
Section 18.7. We follow the article [589] by Mazya and Shubin which
improves the results of Sect. 15.4 for the particular case p = 2, l = 1. A variant
of Lemma 18.7.2/1 was obtained by Mazya [532] as early as in 1963.
Lieb [495] used geometric arguments to establish a lower bound for ()
which is similar to (18.7.20), but with capacity replaced by the Lebesgue
measure. Such lower bounds can be also deduced from Theorem 18.7.1 if we
use isoperimetric inequalities between the capacity and Lebesgue measure

n/(n2)
,
mn F An cap(F )
(18.7.39)
with the equality for balls (see, e.g., [666] or Sects. 2.2.3 and 2.2.4 in [556]),
so

n/(n2)
An = (mn B1 ) cap(B1 )
= n1 (n 2)n/(n2) n2/(n2) .
Namely, let us write for any (0, 1)
801

(m )
r,n = sup r : Br , mn (Br \ ) mn Br .
Then (18.7.39) implies that
(m )
r,n r,
provided
= n/(n2) .
Therefore, we obtain for every (0, 1)

(m ) 2
,
() c(, n) r,n
where = (n2)/n .
Here c(, n) is given by (18.7.22). This is exactly Liebs inequality (1.2) in

[495], though with a dierent constant.
There are numerous results that give lower bounds for (). We will
mention only a few. The FaberKrahn inequality ([265, 462, 666]) gives a
lower bound of () in terms of the area of R2 . Under miscellaneous
topological and geometric restrictions on , the interior radius was shown
to provide a lower bound (hence a two-sided estimate) for () in the case
n = 2 by Hayman [358], Osserman [647649], M. E. Taylor [746], Croke [217],
Ba
nuelos and Carroll [69], and also in the case n 3 ([358, 649]). The following
two-sided estimate for () was established by the author in [531, 534] (see
Sects. 2.4 and 4.3 of this book),
cap(F, )
cap(F, )
1
inf
() inf
,
4 F mn (F )
mn (F )
where the inmum is taken over all compact sets F .
References
1. Acosta, G., Duran, R.: An optimal Poincare inequality in L1 for convex

domains, Proc. Am. Math. Soc. 132 (2004), no. 1, 195202.
2. Adams, D. R.: Traces of potentials arising from translation invariant
operators, Ann. Sc. Norm. Super. Pisa 25 (1971), 203217.
3. Adams, D. R.: A trace inequality for generalized potentials, Stud. Math.
48 (1973), 99105.
4. Adams, D. R.: Traces of potentials II, Indiana Univ. Math. J. 22 (1973),
907918.
5. Adams, D. R.: On the existence of capacitary strong type estimates in
Rn , Ark. Mat. 14 (1976), 125140.
6. Adams, D. R.: Quasi-additivity and sets of nite Lp -capacity, Pac. J.
Math. 79 (1978), 283291.
7. Adams, D. R.: Sets and functions of nite Lp -capacity, Indiana Univ.
Math. J. 27 (1978), 611627.
8. Adams, D. R.: Lectures on Lp -potential theory, Preprint no. 2 (1981),
174, Dept. of. Math., Univ. of Umea.
9. Adams, D. R.: Weighted nonlinear potential theory, Trans. Am. Math.

Soc. 297 (1986), no. 1, 7394.
10. Adams, D. R.: A note on the Choquet integrals with respect to Hausdor
capacity, Lect. Notes Math. 1302 (1988), 115124.
11. Adams, D. R.: A sharp inequality of J. Moser for higher order derivatives, Ann. Math. 128 (1988), 385398.
12. Adams, D. R.: My love aair with the Sobolev inequality, Sobolev Spaces
in Mathematics I, Sobolev Type Inequalities, Springer, Berlin, 2008, 1
24.
13. Adams, D. R., Frazier, M.: Composition operators on potential spaces,
Proc. Am. Math. Soc. 114 (1992), 155165.
14. Adams, D. R., Hedberg, L. I.: Inclusion relations among ne topologies
in non-linear potential theory, Indiana Univ. Math. J. 33 (1984), no. 1,
117126.

DOI 10.1007/978-3-642-15564-2,
803
804
References
15. Adams, D. R., Hedberg, L. I.: Function Spaces and Potential Theory,
Springer, Berlin, 1996.
16. Adams, D. R., Meyers, N. G.: Thinness and Wiener criteria for nonlinear potentials, Indiana Univ. Math. J. 22 (1972), 139158.
17. Adams, D. R., Meyers, N. G.: Bessel potentials. Inclusion relations
among classes of exceptional sets, Indiana Univ. Math. J. 22 (1973),
873905.
18. Adams, D. R., Pierre, M.: Capacitary strong type estimates in semilinear
problems, Ann. Inst. Fourier (Grenoble) 41 (1991), 117135.
19. Adams, D. R., Polking, J. C.: The equivalence of two denitions of
capacity, Proc. Am. Math. Soc. 37 (1973), 529534.
20. Adams, D. R., Xiao, J.: Strong type estimates for homogeneous Besov
capacities, Math. Ann. 325 (2003), no. 4, 695709.
21. Adams, D. R., Xiao, J.: Nonlinear potential analysis on Morrey spaces
and their capacities, Indiana Univ. Math. J. 53 (2004), no. 6, 16291663.
22. Adams, R. A.: Some integral inequalities with applications to the imbedding of Sobolev spaces dened over irregular domains, Trans. Am. Math.
Soc. 178 (1973), 401429.
23. Adams, R. A.: Sobolev Spaces, Academic Press, New York, 1975.
24. Adams, R. A.: General logarithmic Sobolev inequalities and Orlicz
imbeddings, J. Funct. Anal. 34 (1979), no. 2, 292303.
25. Adams, R. A., Fournier, J. J. F.: Sobolev Spaces, Academic Press, New
York, 2003.
26. Adimurthi: HardySobolev inequalities in H 1 () and its applications,
Commun. Contemp. Math. 4 (2002), no. 3, 409434.
27. Adimurthi, Chaudhuri, N., Ramaswamy, M.: An improved Hardy
Sobolev inequality and its applications, Proc. Am. Math. Soc. 130
(2002), 489505.
28. Adimurthi, Grossi, M., Santra, S.: Optimal HardyRellich inequalities,
maximum principle and related eigenvalue problems, J. Funct. Anal.
240 (2006), no. 1, 3683.
29. Adimurthi, Yadava, S. L.: Some remarks on Sobolev type inequalities,
Calc. Var. Partial Dier. Equ. 2 (1994), 427442.
30. Ahlfors, L.: Lectures on Quasiconformal Mappings, Van Nostrand,
Toronto, 1966.
31. Ahlfors, L., Beurling, A.: Conformal invariants and function-theoretic
null-sets, Acta Math. 83 (1950), 101129.
32. Aida, A., Masuda, T., Shigekawa, L.: Logarithmic Sobolev inequalities
and exponential integrability, J. Funct. Anal. 126 (1994), 83101.
33. Aida, S., Stroock, D.: Moment estimates derived from Poincare and
logarithmic Sobolev inequalities, Math. Res. Lett. 1 (1994), 7586.
34. Aikawa, H.: Quasiadditivity of Riesz capacity, Math. Scand. 69 (1991),
1530.
35. Aikawa, H.: Quasiadditivity of capacity and minimal thinness, Ann.
Acad. Sci. Fenn., Ser. A 18 (1993), 6575.
References
805
36. Aikawa, H.: Capacity and Hausdor content of certain enlarged sets,
Mem. Fac. Sci. Eng. Shimane Univ., Ser. B, Math. Sci. 30 (1997), 121.
37. Aissaoui, N., Benkirane, A.: Capacites dans les espaces dOrlicz, Ann.
Sci. Math. Que. 18 (1994), no. 1, 123.
38. Akilov, G. P., Kantorovich, L. V.: Functional Analysis, Pergamon, New
York, 1982.
39. Alekseev, A. B., Oleinik, V. L.: Estimates of the diameters of the unit
ball of
Llp in Lp (), Probl. Math. Phys. 7 (1974), 37.
40. Almgren, F. J.: Existence and regularity almost everywhere of solutions
to elliptic variational problems with constraints, Mem. Am. Math. Soc.
4 (1976), no. 165, 1199.
41. Almgren, F. J., Lieb, E. H.: Symmetric decreasing rearrangement is
sometimes continuous, J. Am. Math. Soc. 4 (1989), 683773.
42. Alvino, A., Ferone, V., Trombetti, G.: On the best constant in a Hardy
Sobolev inequality, Appl. Anal. 85 (2006), no. 13, 171180.
43. Ambrosio, L., Tilli, P.: Topics on Analysis in Metric Spaces, Oxford Lecture Series in Mathematics and Its Applications 25, Oxford University
Press, Oxford, 2004.
44. Amghibech, S.: Inegalites isoperimetriques pour les graphes, Potential
Anal. 6 (1997), 355367.
45. Amick, C. J.: Some remarks on Rellichs theorem and the Poincare inequality, J. Lond. Math. Soc., II. Ser. 18 (1978), 8193.
46. Amick, C. J.: Approximation by smooth functions in Sobolev spaces,
Bull. Lond. Math. Soc. 11 (1979), no. 1, 3740.
47. Ancona, A.: On strong barriers and an inequality by Hardy for domains
in RN , Lond. Math. Soc. (2) 34 (1986), 274290.
48. Andersson, R.: Unbounded Sobole regions, Math. Scand. 13 (1963),
7589.
49. Andersson, R.: The type set of a generalized Sobolev operator, Medd.
Lunds Univ. Mat. Semin. 19 (1972), 1101.
50. Anzellotti, G., Giaquinta, M.: BV functions and traces (in Italian. English summary), Rend. Semin. Mat. Univ. Padova 60 (1979), no. 1978,
121.
51. Arendt, W., Warma, M.: The Laplacian with Robin boundary conditions
on arbitrary domains, Potential Anal. 19 (2003), 341363.
52. Aronszajn, N.: On coercive integro-dierential quadratic forms, Conference on Partial Dierential Equations, Univ. of Kansas, Report no. 14
(1954), 94106.
53. Aronszajn, N., Mulla, P., Szeptycki, P.: On spaces of potentials connected with Lp -classes, Ann. Inst. Fourier 13 (1963), 211306.
54. Attouch, H., Buttazzo, G., Michaille, G.: Variational Analysis in Sobolev
and BV Spaces, Applications to PDEs and Optimization, MPS/SIAM
Series on Optimization, 6, SIAM, Philadelphia, 2006.
55. Aubin, T.: Problèmes isoperimetriques et espaces de Sobolev, C. R.
Acad. Sci., Paris 280 (1975), 279281.
806
References
56. Aubin, T.: Equations

dierentielles non lineaires et problème de Yamabe
concemant la courbure scalaire, J. Math. Pures Appl. 55 (1976), 269
296.
57. Aubin, T.: Nonlinear Analysis on Manifolds: MongeAmpere Equations,
58. Avkhadiev, F., Laptev, A.: Hardy inequalities for nonconvex domains,
Around the Research of Vladimir Mazya I: Functions Spaces, International Mathematical Series 11, Springer, Berlin, 2010, 112.
59. Babich, V. M.: On the extension of functions, Usp. Mat. Nauk 8 (1953),
111113.
60. Babich, V. M., Slobodecki, L. N.: On the boundedness of the Dirichlet
integral, Dokl. Akad. Nauk SSSR 106 (1956), 604607 (in Russian).
61. Bagby, T.: Approximation in the mean by solutions of elliptic equations,
Trans. Am. Math. Soc. 281 (1984), 761784.
62. Bagby, T., Ziemer, W. P.: Pointwise dierentiability and absolute continuity, Trans. Am. Math. Soc. 191 (1974), 129148.
63. Bakry, D.: On Sobolev and logarithmic Sobolev inequalities for Markov
semigroups, Trends in Stochastic Analysis, World Scientic, New York,
1997, 4357.
64. Bakry, D., Coulhon, T., Ledoux, M., Salo-Coste, L.: Sobolev inequalities in disguise, Indiana Univ. Math. J. 44 (1995), no. 4, 10331074.
65. Bakry, D., Ledoux, M., Qian, Z.: Logarithmic Sobolev inequalities,
Poincare inequalities and heat kernel bounds, ESAIM Probab. Stat. 1
(1995/1997), 391407.
66. Baldi, A., Montefalcone, F.: A note on the extension of BV functions in
metric measure spaces, J. Math. Anal. Appl. 340 (2008), no. 1, 197208.
67. Balinsky, A.: Hardy type inequalities for AharonovBohm potentials
with multiple singularities, Math. Res. Lett. 10 (2003), 18.
68. Balinsky, A., Laptev, A., Sobolev, A. V.: Generalized Hardy inequality
for the magnetic Dirichlet forms, J. Stat. Phys., nos. 114 (2004), 507
521.
69. Ba
nuelos, R., Carroll, T.: An improvement of the Osserman constant
for the bass note of a drum, Stochastic Anal. (Ithaca, NY, 1993), Proc.
Symp. Pure Math. 57, Am. Math. Soc., Providence, 1995, 310.
70. Barbatis, G.: Improved Rellich inequalities for the polyharmonic operator, Indiana Univ. Math. J. 55 (2006), no. 4, 14011422.
71. Barbatis, G.: Best constants for higher-order Rellich inequalities in
Lp (), Math. Z. 255 (2007), 877896.
72. Barbatis, G., Filippas, S., Tertikas, A.: Rened geometric Lp Hardy
inequalities, Commun. Contemp. Math. 5 (2003), no. 6, 869883.
73. Barbatis, G., Filippas, S., Tertikas, A.: Series expansion for Lp Hardy
inequalities, Indiana Univ. Math. J. 52 (2003), no. 1, 171190.
74. Barbatis, G., Filippas, S., Tertikas, A.: A unied approach to improved
Lp Hardy inequalities with best constants, Trans. Am. Math. Soc. 356
(2004), no. 6, 21692196.
References
807
75. Barthe, F., Cattiaux, P., Roberto, C.: Concentration for independent
random variables with heavy tails, AMRX Appl. Math. Res. Express
2005, no. 2, 3960.
76. Bastero, J., Milman, M., Ruiz, F.: A note on L(, q) spaces and Sobolev
embeddings, Indiana Univ. Math. J. 52 (2003), 12151230.
77. Beckner, W.: Sobolev inequalities, the Poisson semigroup and analysis
on the sphere S n , Proc. Natl. Acad. Sci. USA 89 (1992), 48164819.
78. Beckner, W.: Sharp Sobolev inequalities on the sphere and the Moser
Trudinger inequality, Ann. Math. (2) 138 (1993), no. 1, 213242.
79. Beckner, W.: Asymptotic estimates for GagliardoNirenberg embedding
constants, Potential Anal. 17 (2002), 253266.
80. Beckner, W.: Sharp inequalities and geometric manifolds, J. Fourier
Anal. Appl. 3 (1997), 825836.
81. Beckner, W., Pearson, M.: On sharp Sobolev embedding and the logarithmic Sobolev inequality, Bull. Lond. Math. Soc. 30 (1998), 8084.
82. Ben Amor, A.: On the equivalence between trace and capacitary inequalities for the abstract contractive space of Bessel potentials, Osaka
J. Math. 42 (2005), no. 1, 1126.
83. Benguria, R. D., Frank, R. L., Loss, M.: The sharp constant in the
HardySobolevMazya inequality in the three dimensional upper halfspace, Math. Res. Lett. 15 (2008), no. 4, 613622.
84. Berezin, F. A., Shubin, M. A.: The Schr
odinger Equation, Kluwer Academic, Dordrecht, 1991.
85. Berger, M., Gauduchon, P., Mazet, E.: Le spectre dune variete Riemannienne, Lect. Notes Math. 194, Springer, Berlin, 1971.
86. Besicovitch, A. S.: A general form of the covering principle and relative
dierentiation of additive functions, Proc. Camb. Philos. Soc. 41 (1945),
103110; II. 42 (1946), 110.
87. Besov, O. V.: Investigation of a family of function spaces in connection
with imbedding and extension theorems, Tr. Mat. Inst. Steklova, Akad.
Nauk SSSR 60 (1961), 4281 (in Russian). English translation: Am.
Math. Soc. Transl. 40 (1964), 85126.
88. Besov, O. V.: Embeddings of an anisotropic Sobolev space for a domain
with a exible horn condition, Proc. Steklov Inst. Math. 4, 1989, 113.
Studies in the theory of dierentiable functions of several variables and
its applications, XII.
89. Besov, O. V.: Embeddings of SobolevLiouville and LizorkinTriebel
spaces in a domain, Dokl. Akad. Nauk 331 (1993), no. 5, 538540 (in
Russian). English translation: Russ. Acad. Sci. Dokl. Math. 4 (1994),
no. 1, 130133.
90. Besov, O. V.: Compactness of embeddings of weighted Sobolev spaces
on a domain with an irregular boundary, Dokl. Math. 63 (2001), no. 1,
95100.
91. Besov, O. V.: Sobolevs embedding theorem for a domain with an irregular boundary, Sb. Math. 192 (2001), no. 34, 323346.
808
References
92. Besov, O. V.: Function spaces of LizorkinTriebel type on an irregular

domain, Proc. Steklov Inst. Math. 260 (2008), 2536.
93. Besov, O. V., Ilin, V. P., Kudryavtsev, L. D., Lizorkin, P. I., Nikolsky,
S. M.: Imbedding theory for classes of dierentiable functions of several
variables, Partial Dierential Equations, Nauka, Moscow, 1970, 3863
(in Russian).
94. Besov, O. V., Ilin, V. P., Nikolsky, S. M.: Integral Representations of
Functions and Imbedding Theorems, Nauka, Moscow, 1975 (in Russian).
English translation: V. H. Winston and Sons, Washington, D.C., vol. I,
1978; vol. II. 1979.
95. Beurling, A.: Ensembles exceptionnels, Acta Math. 72 (1939), 113.
96. Bianchi, G., Egnell, H.: A note on the Sobolev inequality, J. Funct. Anal.
100 (1991), no. 1, 1824.
97. Biegert, M., Warma, M.: Regularity in capacity and the Dirichlet Laplacian, Potential Anal. 25 (2006), no. 3, 289305.
98. Biegert, M., Warma, M.: Removable singularities for a Sobolev space,
J. Math. Anal. Appl. 313 (2006), 4963.
99. Biezuner, R. J.: Best constants in Sobolev trace inequalities, Nonlinear
Anal. 54 (2003), 575589.
Perturbation of quadratic forms and the spectrum of
100. Birman, M. S.:
singular boundary value problems, Dokl. Akad. Nauk SSSR 125 (1959),
471474 (in Russian).
On the spectrum of boundary value problems, Mat. Sb.
101. Birman, M. S.:
55 (1961), 125174 (in Russian).
Pavlov, B. S.: On the complete continuity of certain
102. Birman, M. S.,
imbedding operators, Vestn. Leningr. Univ., Mat. Mekh. Astron. No. 1
(1961), 6174 (in Russian).
103. Birnbaum, Z. W., Orlicz, W.: Uber

die Verallgemeinerung des Begries
der zueinander konjugierten Potenzen, Stud. Math. 3 (1931), 167.
104. Biroli, M.: Schr
odinger type and relaxed Dirichlet problems for the
subelliptic p-Laplacian, Potential Anal. 15 (2001), 116.
105. Biroli, M., Mosco, V.: Sobolev inequalities on homogeneous spaces, Potential Anal. 4 (1995), 311324.
106. Bj
orn, J., MacManus, P., Shanmugalingam, N.: Fat sets and pointwise
boundary estimates for p-harmonic functions in mertic spaces, J. Anal.
Math. 85 (2001), 339369.
107. Bj
orn, J., Shanmugalingam, N.: Poincare inequalities, uniform domains
and extension properties for NewtonSobolev functions in metric spaces,
J. Math. Anal. Appl. 332 (2007), no. 1, 190208.
108. Bj
orup, K.: On inequalities of Poincares type, Math. Scand. 8 (1960),
157160.
109. Bliss, G. A.: An integral inequality, J. Lond. Math. Soc. 5 (1930), 4046.
110. Bobkov, S. G.: Localization proof of the BakryLedoux isoperimetric
inequality and some applications, Theory Probab. Appl. 47 (2002), no. 2,
308314.
References
809
111. Bobkov, S. G.: Large deviations via transference plans, Adv. Math. Res.
2 (2003), 151175.
112. Bobkov, S. G.: Large deviations and isoperimetry over convex probability measures with heavy tails, Electron. J. Probab. 12 (2007), 1072
1100.
113. Bobkov, S. G.: On isoperimetric constants for log-concave probability
distributions, Geom. Aspects Funct. Anal., Lect. Notes Math. 1910,
Springer, Berlin, 2007, 8188.
114. Bobkov, S. G., Gotze, F.: Exponential integrability and transportation
cost related to logarithmic Sobolev inequalities, J. Funct. Anal. 163
(1999), 128.
115. Bobkov, S. G., G
otze, F.: Hardy-type inequalities via Riccati and SturmLiouville equations, Sobolev spaces in mathematics I, Sobolev Type Inequalities, Int. Math. Ser. 8, Springer, Berlin, 2009, 6986.
116. Bobkov, S. G., Houdre, C.: Isoperimetric constants for product probability measures, Ann. Probab. 25 (1997), no. 1, 184205.
117. Bobkov, S. G., Houdre, C.: Some connections between isoperimetric and
Sobolev-type inequalities, Mem. Am. Math. Soc. 129 (1997), no. 616,
viii+111.
118. Bobkov, S. G., Ledoux, M.: From BrunnMinkowski to sharp Sobolev
inequalities, Ann. Mat. Pura Appl. (4) 187 (2008), no. 3, 369384.
119. Bobkov, S. G., Nazarov, F. L.: Sharp dilation-type inequalities with
xed parameter of convexity, J. Math. Sci. 152 (2008), no. 6, 826839.
120. Bobkov, S. G., Zegarlinski, B.: Distributions with slow tails and ergodicity of Markov semigroups, innite dimensions, Around the Research
of Vladimir Mazya I, Function Spaces, Int. Math. Series 11, Springer,
Berlin, 2010, 1379.
121. Bodineau, T., Heler, B.: The log-Sobolev inequalities for unbounded
spin systems, J. Funct. Anal. 166 (1999), 168178.
122. Bojarski, B.: Remarks on Sobolev Imbedding Inequalities, Lect. Notes
Math., 1351, Springer, Berlin, 1989, 5268.
123. Bojarski, B., Hajlasz, P.: Pointwise inequalities for Sobolev functions,
Stud. Math. 106 (1993), 7792.
124. Bokowski, J., Sperner, E.: Zerlegung konvexer Korper durch minimale
Trenn
achen, J. Reine Angew. Math. 311/312 (1979), 80100.
125. Bonder, J. F., Rossi, J. D., Ferreira, R.: Uniform bounds for the best
Sobolev trace constant, Adv. Nonlinear Stud. 3 (2003), 181192.
126. Bonder, J. F., Saintier, N.: Estimates for the Sobolev trace constant with
critical exponent and applications, Ann. Mat. Pura Appl. 187 (2008),
no. 4, 683704.
127. Bosi, R., Dolbeualt, J., Esteban, M. J.: Estimates for the optimal constants in multipolar Hardy inequalities for Schrodinger and Dirac operators, Commun. Pure Appl. Anal. 7 (2008), 533562.
128. Bourbaki, N.: Espaces Vectoriels Topologiques, Hermann, Paris, 1953.
810
References
129. Bourdaud, G.: Le calcul fonctionnel dans les espaces de Sobolev, Invent.
Math. 104 (1991), 435466.
130. Bourdaud, G.: Fonctions qui opèrent sur les espaces de Besov et de
Triebel, Ann. I.H.P.AN 10 (1993), 413422.
131. Bourdaud, G.: The functional calculus in Sobolev spaces, Function
Spaces, Dierential Operators and Nonlinear Analysis (Friedrichroda,
1992), Teubner-Texte Math. 133, Teubner, Stuttgart, 1993, 127142.
132. Bourdaud, G.: Une propriete de composition dans lespace H s , C. R.
Acad. Sci. Paris, Ser. I 340 (2005), 221224.
133. Bourdaud, G., Kateb, D.: Fonctions qui opèrent sur les espaces de Besov,
Math. Ann. 303 (1995), no. 4, 653675.
134. Bourdaud, G., Kateb, M. E. D.: Calcul fonctionnel dans lespace de
Sobolev fractionnaire, Math. Z. 210 (1992), no. 4, 607613.
135. Bourdaud, G., Meyer, Y.: Fonctions qui opèrent sur les espaces de
Sobolev, J. Funct. Anal. 97 (1991), 351360.
136. Bourdaud, G., Moussai, M., Sickel, W.: An optimal symbolic calculus
on Besov algebras, Ann. I.H.P.AN 23 (2006), 949956.
137. Bourgain, J.: On the distribution of polynomials on high dimensional
convex sets, Lect. Notes Math. 1469 (1991), 127137.
138. Bourgain, J., Brezis, H., Mironescu, P.: Another look at Sobolev spaces,
Optimal Control and Partial Dierential Equations Menaldi, J. L., Rofman, E., Sulem, A. (Eds.), IOS Press, Amsterdam, 2001, 439455.
139. Bourgain, J., Brezis, H., Mironescu, P.: Limiting embedding theorems
for W s,p when s 1 and applications, J. Anal. Math. 87 (2002), 77101.
140. Brandolini, B., Chiacchio, F., Trombetti, C.: Hardy inequality and
Gaussian measure, Commun. Pure Appl. Math. 6 (2007), no. 2, 411
428.
141. Brezis, H., Gallouet, T.: Nonlinear Schr
odinger evolution equations,
Nonlinear Anal. 4 (1980), 677681.
142. Brezis, H., Marcus, M.: Hardys inequality revisited, Ann. Sc. Norm.
Super. Pisa 25 (1997), 217237.
143. Brezis, H., Marcus, M., Shafrir, I.: Extremal functions for Hardys inequality with weight, J. Funct. Anal. 171 (2000), 177191.
144. Brezis, H., Mironescu, P.: GagliardoNirenberg, composition and products in fractional Sobolev spaces, J. Evol. Equ. 1 (2001), no. 4, 387404.
145. Brezis, H., Vazquez, J. L.: Blow-up solutions of some nonlinear elliptic
problems, Rev. Mat. Univ. Complut. Madr. 10 (1997), 443469.
146. Brezis, H., Wainger, S.: A note on limiting cases of Sobolev embeddings
and convolution inequalities, Commun. Partial Dier. Equ. 5 (1980),
773789.
147. Buckley, S., Koskela, P.: SobolevPoincare implies John, Math. Res.
Lett. 2 (1995), 577593.
148. Buckley, S., Koskela, P.: Criteria for imbeddings of SobolevPoincare
type, Int. Math. Res. Not. 18 (1996), 881902.
149. Burago, Yu., Kosovsky, N.: Boundary trace for BV functions in regions
with irregular boundary, Analysis, Partial Dierential Equations and
References
150.
151.
152.
153.
154.
155.
156.
157.
158.
159.
160.
161.
162.
163.
164.
165.
811
Applications, The Vladimir Mazya Anniversary Volume, Operator Theory, Advances and Applications Birkh
auser, Basel, 2009, 114.
Burago, Yu. D., Mazya, V. G.: Some questions of potential theory and
function theory for domains with non-regular boundaries, Zap. Nauchn.
Semin. Leningr. Otd. Mat. Inst. Steklova 3 (1967), 1152 (in Russian).
English translation: Semin. Math. Steklov Math. Inst. Leningr. 3 (1969),
168. Consultants Bureau, New York.
Burago, Yu. D., Zalgaller, V. A.: Geometric Inequalities, Springer,
Berlin, 1988.
Burenkov, V. I.: On imbedding theorems for the domain Rk = {i h <
xki i < i h; 0 < h < 1}, Mat. Sb. 75 (1968), 496501 (in Russian).
English translation: Math. USSR Sb. 4 (1968) 457462.
Burenkov, V. I.: On the additivity of the spaces Wpr and Bpr and imbedding theorems for regions of a general kind, Tr. Mat. Inst. Steklova
105 (1969), 3045 (in Russian). English translation: Proc. Steklov Inst.
Math. 105 (1969), 3553.
Burenkov, V. I.: The Sobolev integral representation and the Taylor
formula, Tr. Mat. Inst. Steklova 131 (1974), 3338 (in Russian). English
translation: Proc. Steklov Inst. Math. 131 (1975).
Burenkov, V. I.: Sobolev Spaces on Domains, Teubner-Texte zur Mathematik, Band 137, 1998.
Burenkov, V. I.: Extension theorems for Sobolev spaces, The Mazya
Anniversary Collection, Vol. 1 Operator Theory: Advances and Applications 109, Birkhauser, Basel, 1999, 187200.
Burenkov, V. I., Davies, E. B.: Spectral stability of the Neumann Laplacian, J. Dier. Equ. 186 (2002), 485508.
Busemann, H.: The isoperimetric problem for Minkowski area, Am. J.
Math. 71 (1949), 743762.
Buser, P.: On Cheegers 1 h2 /4 in Geometry of the Laplace operator,
Proc. Symp. Pure Math. 36 (1980), 2977.
Caccioppoli, R.: Misure e integrazione sugli insiemi dimensionalmente
orientati, Rend. R. Accad. Naz. Lincei 12 (1952), 311.
Caccioppoli, R.: Misure e integrazione sugli insiemi dimensionalmente
orientati, Rend. R. Accad. Naz. Lincei 12 (1952), 137146.
Caarelli, L., Kohn, R., Nirenberg, L.: First order interpolation inequalities with weights, Compos. Math. 53 (1984), 259275.
Calder
on, A. P.: Lebesgue spaces of dierentiable functions and distributions, Proc. Symp. Pure Math. IV (1961), 3349.
Calude, C. S., Pavlov, B.: The Poincare-Hardy inequality on the complement of a Cantor set, Interpolation Theory, Systems Theory and
Related Topics (Tel Aviv/Rehovot, 1999), Operator Theory Adv. Appl.
134, Birkhauser, Basel, 2002, 187208.
Campanato, S.: Il teoreme di immersione di Sobolev per una classe di
aperti non dotati della proprietà di cono, Ric. Mat. 11 (1962), 103122.
812
References
166. Capogna, L., Danielli, D., Pauls, S. D., Tyson, J. T.: An Introduction to
the Heisenberg Group and the Sub-Riemannian Isoperimetric Problem,
Progress in Mathematics 259, Birkhauser, Basel, 2007.
167. Carlen, E. A., Loss, M.: Sharp constant in Nashs inequality, Int. Math.
Res. Not. 7 (1993), 213215.
168. Carleson, L.: Selected Problems on Exceptional Sets, Van Nostrand,
Toronto, 1967.
169. Carleson, L., Chang, S.-Y. A.: On a sharp inequality concerning the
Dirichlet integral, Am. J. Math. 107 (1985), 10151033.
170. Carlsson, A.: Inequalities of PoincareWirtinger type, Licentiate thesis
no. 232, Linkoping, 1990.
171. Carlsson, A., Mazya, V.: On approximation in weighted Sobolev spaces
and self-adjointness, Math. Scand. 74 (1994), 111124.
172. Cartan, H.: Sur les systèmes de fonctions holomorphes à varietes
lineaires et leurs applications, Ann. Sci. Ecole Norm. Super. 3 (1928),
255346.
173. Carton-Lebrun, C., Heinig, H. P.: Weighted norm inequalities involving
gradients, Proc. R. Soc. Edinb. 112A (1989), 331341.
174. Casado-Daz, J.: The capacity for pseudomonotone operators, Potential
Anal. 14 (2001), 7391.
175. Cascante, C., Ortega, J. M., Verbitsky, I.: Trace inequalities of Sobolev
type in the upper triangle case, Proc. Lond. Math. Soc. (3) 80 (2000),
391414.
176. Cascante, C., Ortega, J. M., Verbitsky, I. E.: On Lp Lq trace inequalities, J. Lond. Math. Soc. 74 (2006), 497511.
177. Cattiaux, P., Gentil, I., Guillin, A.: Weak logarithmic Sobolev inequalities and entropic convergence, Probab. Theory Relat. Fields 139 (2007),
no. 34, 563603.
178. Chaudhuri, N.: Bounds for the best constant in an improved Hardy
Sobolev inequality, Z. Anal. Anwend. 22 (2003), no. 4, 757765.
179. Chavel, I.: Eigenvalues in Riemannian Geometry, Academic Press, New
York, 1984.
180. Chavel, I.: Isoperimetric Inequalities: Dierential Geometric and Analytic Perspectives, Cambridge University Press, Cambridge, 2001.
181. Cheeger, J.: A lower bound for the smallest eigenvalue of the Laplacian,
Problems in Analysis, Gunning, R. (Ed.), Princeton University Press,
Princeton, 1970, 195199.
182. Cheeger, J.: Dierentiability of Lipschitz functions on metric spaces,
Geom. Funct. Anal. 9 (1999), 428517.
183. Chen, M. F.: Eigenvalues, Inequalities and Ergodic Theory, Probability
and Its Applications, Springer, Berlin, 2005.
184. Chen, M. F.: Capacitary criteria for Poincare-type inequalities, Potential
Anal. 23 (2005), 303332.
References
813
185. Chen, M. F., Wang, F.-Y.: Estimates of logarithmic Sobolev constant:

an improvement of BakryEmery criterion, J. Funct. Anal. 144 (1997),
287300.
186. Choquet, G.: Theory of capacities, Ann. Inst. Fourier 5 (1955), 131395.
187. Chou, K. S., Chu, Ch. W.: On the best constant for weighted Sobolev
Hardy inequality, J. Lond. Math. Soc. (2) 48 (1993), no. 1, 137151.
188. Chua, S.-K.: Extension theorems on weighted Sobolev spaces, Indiana
Univ. Math. J. 41 (1992), 10271076.
189. Chua, S.-K.: Weighted Sobolev inequalities on domains satisfying the
chain condition, Proc. Am. Math. Soc. 117 (1993), no. 2, 449457.
190. Chua, S.-K., Wheeden, R. L.: Sharp conditions for weighted 1dimensional Poincare inequalities, Indiana Univ. Math. J. 49 (2000),
143175.
191. Chung, F., Grigoryan, A., Yau, S.-Y.: Higher eigenvalues and isoperimetric inequalities on Riemannian manifolds and graphs, Commun.
Anal. Geom. 8 (2000), 9691026.
192. Cianchi, A.: On relative isoperimetric inequalities in the plane, Boll.
Unione Mat. Ital. B 3 (1989), no. 7, 289325.
193. Cianchi, A.: A sharp form of Poincare type inequalities on balls and
spheres, Z. Angew. Math. Phys. 40 (1989), 558569.
194. Cianchi, A.: A sharp embedding theorem for OrliczSobolev spaces, Indiana Univ. Math. J. 45 (1996), 3965.
195. Cianchi, A.: Boundedness of solutions to variational problems under general growth conditions, Commun. Partial Dier. Equ. 22 (1997), 1629
1646.
196. Cianchi, A.: Rearrangements of functions in Besov spaces, Math. Nachr.
230 (2001), 1935.
197. Cianchi, A.: Symmetrization and second-order Sobolev inequalities,
Ann. Mat. 183 (2004), 4577.
198. Cianchi, A.: Moser-Trudinger inequalities without boundary conditions
and isoperimetric problems, Indiana Univ. Math. J. 54 (2005), 669705.
199. Cianchi, A.: A quantitative Sobolev inequality in BV, J. Funct. Anal.
237 (2006), no. 2, 466481.
200. Cianchi, A.: Higher-order Sobolev and Poincare inequalities in Orlicz
spaces, Forum Math. 18 (2006), 745767.
201. Cianchi, A.: Moser-Trudinger trace inequalities, Adv. Math. 217 (2008),
20052044.
202. Cianchi, A.: Sharp SobolevMorrey inequalities and the distance from
extremals, Trans. Am. Math. Soc. 360 (2008), 43354347.
203. Cianchi, A., Ferone, A.: Best remainder norms in SobolevHardy inequalities, Indiana Univ. Math. J. 58 (2009), 10511096.
204. Cianchi, A., Fusco, N., Maggi, F., Pratelli, A.: The sharp Sobolev inequality in quantitative form, J. Eur. Math. Soc. 11 (2009), no. 5, 1105
1139.
205. Cianchi, A., Kerman, R., Pick, L.: Boundary trace inequalities and rearrangements, J. Anal. Math. 105 (2008), 241265.
814
References
206. Cianchi, A., Lutwak, E., Yang, D., Zhang, G.: Ane MoserTrudinger
and MorreySobolev inequalities, Calc. Var. Partial Dier. Equ., 36
(2009), no. 3, 419436.
207. Cianchi, A., Pick, L.: Optimal Gaussian Sobolev embeddings, J. Funct.
Anal. 256 (2009), no. 11, 35883642.
208. Cohen, A., DeVore, R., Petrushev, P., Xu, H.: Non linear approximation
and the space BV (R2 ), Am. J. Math. 121 (1999), 587628.
209. Cohen, A., Meyer, Y., Oru, F.: Improved Sobolev inequalities, Semin.
X-EDP Ecole
Polytech. XVI (1998), 116.
210. Coifman, R., Feerman, C.: Weighted norm inequalities for maximal
functions and singular integrals, Stud. Math. 5 (1974), 241250.
211. Colin, F.: Lemmes de decomposition appliques à des inegalities de type
HardySobolev, Potential Anal. 23 (2005), 181206.
212. Cordero-Erausquin, D., Nazaret, B., Villani, C.: A mass-transportation
approach to sharp Sobolev and GagliardoNirenberg inequalities, Adv.
Math. 182 (2004), no. 2, 307332.
213. Costea, S., Mazya, V.: Conductor inequalities and criteria for Sobolev
Lorentz two weight inequalities, Sobolev Spaces in Mathematics II, Int.
Math. Ser. (N. Y.) 9, Springer, New York, 2009, 103121.
214. Costin, O., Mazya, V.: Sharp Hardy-Leray inequality for axisymmetric
divergence-free elds, Calc. Var. Partial Dier. Equ. 32 (2008), no. 4,
523532.
215. Coulhon, T., Koskela, P.: Geometric interpretations of LP -Poincare inequalities on graphs with polynomial volume growth, Milan J. Math. 72
(2004), 209248.
216. Courant, R., Hilbert, D.: Methoden der mathematischen Physik, 2,
217. Croke, C. B.: The rst eigenvalue of the Laplacian for plane domains,
Proc. Am. Math. Soc. 81 (1981), 304305.
218. Dahlberg, B. E. J.: A note on Sobolev spaces. Harmonic analysis in
Euclidean spaces, Proc. Sympos. Pure Math. Williams Coll., Part 1
(Williamstown, MA, 1978), Proc. Sympos. Pure Math. XXXV, Am.
Math. Soc., Providence, 1979, 183185.
219. Dahlberg, B. E. J.: Regularity properties of Riesz potentials, Indiana
Univ. Math. J. 28 (1979), 257268.
220. Dal Maso, G., Skrypnik, I. V.: Capacity theory for monotone operators,
Potential Anal. 7 (1997), 765803.
221. Daners, D.: Robin boundary value problems on arbitrary domains,
Trans. Am. Math. Soc. 352 (2000), 42074236.
222. Davies, E. B.: Heat Kernels and Spectral Theory, Cambridge University
Press, Cambridge, 1989.
223. Davies, E. B.: Two-dimensional Riemannian manifolds with fractal
boundaries, J. Lond. Math. Soc. (2) 49 (1994), no. 2, 343356.
224. Davies, E. B.: Spectral Theory and Dierential Operators, University
Press, Cambridge, 1995.
References
815
225. Davies, E. B.: The Hardy constant, Q. J. Math. (2) 46 (1995), 417431.
226. Davies, E. B.: A review of Hardy inequalities, The Mazya Anniversary
Collection, vol. 2, Operator Theory, Advances and Applications 110,
Birkhauser, Basel, 1999.
227. Davies, E. B., Hinz, A. M.: Explicit constants for Rellich inequalities in
Lp (), Math. Z. 227 (1998), no. 3, 511523.
228. D
avila, J., Dupaigne, L.: Hardy-type inequalities, J. Eur. Math. Soc. 6
(2004), no. 3, 335365.
229. De Giorgi, E.: Su una teoria generale della misure (r l)-dimensionale
in uno spazio ad r dimensioni, Ann. Mat. Pura Appl. (4) 36 (1954),
191213.
230. De Giorgi, E.: Nuovi teoremi relative alle misure (r 1)-dimensionale
in spazio ad r dimensioni, Ric. Mat. 4 (1955), 95113.
231. Del Pino, M., Dolbeault, J.: Best constants for GagliardoNirenberg inequalities and applications to nonlinear diusions, J. Math. Pures Appl.
81 (2002), no. 9, 847875.
232. Delin, H.: A proof of the equivalence between Nash and Sobolev inequalities, Bull. Sci. Math. 120 (1996), no. 4, 405411.
233. Demyanov, A. V., Nazarov, A. I.: On the existence of an extremal
function in the Sobolev embedding theorems with critical exponent,
Algebra Anal. 17 (2005), no. 5, 108142.
234. Deny, J., Lions, J.-L.: Les escapes du type de Beppo Levi, Ann. Inst.
Fourier 5 (1953/1954), 305370.
235. DiBenedetto, E.: Real Analysis, Birkh
auser, Basel, 2002.
236. Diening, L., H
ast
o, P., Nekvinda, A.: Open problems in variable exponent Lebesgue and Sobolev spaces, Function Spaces, Dierential Operators and Nonlinear Analysis, Proceedings of the Conference held in
Milovy, Bohemian-Moravian Uplands, May 28June 2, 2004, Math. Inst.
Acad. Sci. Czech Republick, Praha, 2005, 3858.
237. Dolbeault, J., Esteban, M. J., Loss, M., Vega, L.: An analytical proof
of Hrady-like inequalities related to the Dirac operator, J. Func. Anal.
216 (2004), no. 1, 121.
238. Donnelly, H.: Eigenvalue estimates for certain noncompact manifolds,
Mich. Math. J. 31 (1984), 349357.
239. Donnelly, H.: Bounds for eigenfunctions of the Laplacian on compact
Riemannian manifolds, J. Funct. Anal. 187 (2001), 247261.
240. Donoghue, W. F. Jr.: A coerciveness inequality, Ann. Sc. Norm. Super.
Pisa, Cl. Sci. 20 (1966), 589593.
241. Dou, J., Niu, P., Yuan, Z.: A Hardy inequality with remainder terms
in the Heisenberg group and weighted eigenvalue problem, J. Inequal.
Appl. 2007 (2007), 32585.
242. Douglas, J.: Solution of the problem of Plateau, Trans. Am. Math. Soc.
33 (1931), 263321.
816
References
243. Druet, O., Hebey, E.: The AB program in geometric analysis: sharp
Sobolev inequalities and related problems, Mem. Am. Math. Soc. 160
(2002), no. 761, 198.
244. Dunford, N., Schwartz, J.: Linear Operators I, Interscience, New York,
1958.
245. Dynkin, E. M.: Homogeneous measures on subsets of Rn , Lect. Notes
in Math. 1043, Springer, Berlin, 1983, 698700.
246. Dynkin, E. M.: Free interpolation by functions with derivatives in H 1 ,
Zap. Nauchn. Sem. Leningr. Otdel. Mat. Inst. Steklov (LOMI) 126
(1983), 7787. English transl. in J. Sov. Math. 27 (1984), 24752481.
247. Edmunds, D. E.: Embeddings of Sobolev spaces, nonlinear analysis,
function spaces and applications, Proc. of a Spring School Held in Horni
Bradlo, 1978 Teubner, Leipzig, 1979, 3858.
248. Edmunds, D. E., Evans, W. D.: Spectral theory and embeddings of
Sobolev spaces, Q. J. Math. Oxf. Ser. (2) 30 (1979), no. 120, 431453.
249. Edmunds, D. E., Evans, W. D.: Spectral Theory and Dierential Operators, Clarendon/Oxford University Press, New York, 1987.
250. Edmunds, D. E., Evans, W. D.: Hardy Operators, Function Spaces and
Embeddings. Springer Monographs in Mathematics, Springer, Berlin,
2004.
251. Edmunds, D. E., Hurri-Syrj
anen, R.: Necessary conditions for Poincare
domains, Papers on Analysis: A volume dedicated to Olli Martio on
the occasion of his 60th birthday, Report. Univ. Jyvaskyla 83 (2001),
6372.
252. Edmunds, D. E., Hurri-Syrjanen, R.: Weighted Hardy inequalities,
J. Math. Anal. Appl. 310 (2005), 424435.
253. Edmunds, D. E., Kerman, R., Pick, L.: Optimal imbeddings involving
rearrangement invariant quasinorms, J. Funct. Anal. 170 (2000), 307
355.
254. Egnell, H., Pacella, F., Tricarico, M.: Some remarks on Sobolev inequalities, Nonlinear Anal.Theory Methods Appl. 13 (1989), no. 6, 671681.
255. Eidus, D.: The perturbed Laplace operator in a weighted L2 -space,
J. Funct. Anal. 100 (1991), 400410.
256. Eilertsen, S.: On weighted fractional integral inequalities, J. Funct. Anal.
185 (2001), no. 1, 342366.
257. Ehrling, G.: On a type of eigenvalue problem for certain elliptic dierential operators, Math. Scand. 2 (1954), 267285.
258. Escobar, J. F.: On the spectrum of the Laplacian on complete Riemannian manifolds, Commun. Partial Dier. Equ. 11 (1986), 6385.
259. Escobar, J. F.: Sharp constant in a Sobolev trace inequality, Indiana
Univ. Math. J. 37 (1988), 687698.
260. Escobar, J. F.: The Yamabe problem on manifolds with boundary, J. Differ. Geom. 35 (1992), 2184.
261. Evans, W. D., Harris, D. J.: Sobolev embeddings for generalized ridged
domains, Proc. Lond. Math. Soc. (3) 54 (1987), no. 1, 141175.
References
817
262. Evans, W. D., Harris, D. J.: On the approximation numbers of Sobolev

embeddings for irregular domains, Q. J. Math. Oxf. 40 (1989), no. 2,
1342.
263. Evans, W. D., Harris, D. J.: Fractals, trees and the Neumann Laplacian,
Math. Ann. 296 (1993), 493527.
264. Evans, W. E., Lewis, R. T.: Hardy and Rellich inequalities with remainders, J. Math. Inequal. 1 (2007), no. 4, 473490.
265. Faber, C.: Beweis, dass unter allen homogenen Membranen von gleicher
Flache und gleicher Spannung die kreisf
ormige den tiefsten Grundton
gibt, Sitzungsber. Bayer. Acad. Wiss. Math. Phys., Munich, 1923, 169
172.
266. Faddeev, D. K.: On representations of summable functions by singular
integrals at Lebesgue points, Mat. Sb. 1 (1936), 352368 (in Russian).
267. Fain, B. L.: On extension of functions in anisotropic Sobolev spaces,
Proc. Steklov Inst. Math. 170 (1984), 248272 (in Russian).
268. Federer, H.: A note on the GuassGreen theorem, Proc. Am. Math. Soc.
9 (1958), 447451.
269. Federer, H.: Curvature measures, Trans. Am. Math. Soc. 93 (1959),
418491.
270. Federer, H.: The area of nonparametric surfaces, Proc. Am. Math. Soc.
11 (1960), 436439.
271. Federer, H.: Geometric Measure Theory, Springer, Berlin, 1969.
272. Federer, H.: A minimizing property of extremal submanifolds, Arch.
Ration. Mech. Anal. 9 (1975), 207217.
273. Federer, H., Fleming, W. H.: Normal and integral currents, Ann. Math.
72 (1960), 458520.
274. Feerman, C., Stein, E.: H p -spaces of several variables, Acta Math. 129
(1972), 137193.
275. Filippas, S., Mazya, V., Tertikas, A.: Sharp HardySobolev inequalities,
C. R. Acad. Sci. Paris 339 (2004), no. 7, 483486.
276. Filippas, S., Mazya, V., Tertikas, A.: On a question of Brezis and Marcus, Calc. Var. Partial Dier. Equ. 25 (2006), no. 4, 491501.
277. Filippas, S., Mazya, V., Tertikas, A.: Critical HardySobolev inequalities, J. Math. Pures Appl. 87 (2007), 3756.
278. Filippas, S., Tertikas, A.: Optimizing improved Hardy inequalities,
J. Funct. Anal. 192 (2002), 186233.
279. Filippas, S., Tertikas, A., Tidblom, J.: On the structure of Hardy
SobolevMazya inequalities, J. Eur. Math. Soc. 11 (2009), 11651185.
280. Fitzsimmons, P. J.: Hardys inequality for Dirichlet forms, J. Math.
Anal. Appl. 250 (2000), 548560.
281. Fleming, W. H.: Functions whose partial derivatives are measures, Ill.
J. Math. 4 (1960), 452478.
282. Fleming, W. H., Rishel, R. W.: An integral formula for total gradient
variation, Arch. Math. 11 (1960), 218222.
818
References
283. Fradelizi, M., Guedon, O.: The extreme points of subsets of s-concave
probabilities and a geometric localization theorem, Discrete Comput.
Geom. 31 (2004), no. 2, 327335.
284. Fradelizi, M., Guedon, O.: A generalized localization theorem and geometric inequalities for convex bodies, Adv. Math. 204 (2006), no. 2,
509529.
285. Fraenkel, L. E.: On regularity of the boundary in the theory of Sobolev
spaces, Proc. Lond. Math. Soc. 39 (1979), 385427.
in C (0,) (),
J. Lond.
286. Fraenkel, L. E.: On the embedding of C 1 ()
Math. Soc. (2) 26 (1982), 290298.
287. Franchi, B., Hajlasz, P., Koskela, P.: Denitions of Sobolev classes on
metric spaces, Ann. Inst. Fourier 49 (1999), 19031924.
288. Frank, R. L., Lieb, E. H., Seiringer, R.: HardyLiebThirring inequalities for fractional Schr
odinger operators, J. Am. Math. Soc. 21 (2008),
no. 4, 925950.
289. Frank, R. L., Seiringer, R.: Non-linear ground state representations and
sharp Hardy inequalities, J. Funct. Anal. 255 (2008), no. 12, 34073430.
290. Frank, R. L., Seiringer, R.: Sharp fractional Hardy inequalities in halfspaces, Around the Research of Vladimir Mazya I: Functions Spaces,
International Mathematical Series 11, Springer, Berlin, 2010, 161167.
291. Freud, G., Kr

alik, D.: Uber
die Anwendbarkeit des Dirichletschen
Prinzips f
ur den Kreis, Acta Math. Hung. 7 (1956), 411418.
292. Friedrichs, K.: Die Rand- und Eigenwertprobleme aus der Theorie der
elastischen Platten (Anwendung der direkten Methoden der Variationsrechnung), Math. Ann. 98 (1928), 205247.
293. Frostman, O.: Potentiel dequilibre et capacite des ensembles avec
quelques applications à la theorie des fonctions, Medd. Lunds Univ.
Mat. Sem. 3 (1935), 1118.
294. Fukushima, M.: Uemura. T.: Capacitary bounds of measures and ultracontractivity of time changed processes, J. Math. Pures Appl. 82 (2003),
no. 5, 553572.
295. Fukushima, M., Uemura, T.: On Sobolev and capacitary inequalities for
contractive Besov spaces over d-sets, Potential Anal. 18 (2003), no. 1,
5977.
296. Fusco, N., Maggi, F., Pratelli, A.: The sharp quantitative Sobolev inequality for functions of bounded variation, J. Funct. Anal. 244 (2007),
no. 1, 315341.
297. F
uredi, Z., Loeb, P. A.: On the best constant for the Besikovitch covering
theorem, Proc. Am. Math. Soc. 121 (1994), 4.
298. Gagliardo, E.: Caratterizzazioni delle tracce sulla frontiera relative ad
alcune classi di funzioni in piu variabili, Rend. Semin. Mat. Univ. Padova
27 (1957), 284305.
299. Gagliardo, E.: Proprieta di alcune classi di funzioni in piu variabili, Ric.
Mat. 7 (1958), 102137.
References
819
300. Gagliardo, E.: Ulteriori proprieta di alcune classi di funzioni in piu variabili, Ric. Mat. 8 (1959), 2451.
301. Galaktionov, V.: On extensions of Hardys inequalities, Commun. Contemp. Math. 7 (2005), 97120.
302. Galaktionov, V.: On extensions of higher-order Hardys inequalities, Differ. Integral Equ. 19 (2006), 327344.
303. Gazzola, F., Grunau, H.-Ch., Mitidieri, E.: Hardy inequalities with optimal constants and remainder terms, Trans. Am. Math. Soc. 356 (2004),
21492168.
304. Gelfand, I. M., Shilov, G. E.: Spaces of Fundamental and Generalized
Functions, Nauka, Moscow, 1958 (in Russian). English edition: Academic Press, New York, 1968.
305. Gelman, I. W., Mazya, V. G.: Abschatzungen f
ur Dierentialoperatoren im Halbraum, Akademie Verlag, Berlin, 1981.
306. Ghoussoub, N., Kang, X. S.: HardySobolev critical elliptic equations
with boundary singularities, Ann. Inst. H. Poincare Anal. Non Lineaire
21 (2004), 767793.
307. Gidas, B., Ni, W.-M., Nirenberg, L.: Symmetry and related properties
via the maximum principle, Commun. Math. Phys. 68 (1979), 209243.
308. Gil-Medrano, O.: On the Yamabe problem concerning the compact locally conformally at manifolds, J. Funct. Anal. 66 (1986), 4253.
309. Glazman, I. M.: Direct Methods of Qualitative Spectral Analysis of
Singular Dierential Operators, Davey, New York, 1966.
310. Glazer, V., Martin, A., Grosse, H., Thirring, W.: A family of optimal
conditions for the absence of bound states in a potential, Studies in
Mathematical Physics, Lieb, E. H., Simon, B., Wightman, A. S. (Eds.),
Princeton University Press, Princeton, 1976, 169194.
311. Globenko, I. G.: Some questions of imbedding theory for domains with
singularities on the boundary, Mat. Sb. 57 (1962), 201224 (in Russian).
312. Glushko, V. P.: On regions which are star-like with respect to a sphere,
Dokl. Akad. Nauk SSSR 144 (1962), 12151216 (in Russian). English
translation: Sov. Math. Dokl. 3 (1962), 878879.
313. Goldshtein, V. M.: Extension of functions with rst generalized derivatives from planer domains, Dokl. Akad. Nauk SSSR 257 (1981), 268271
(in Russian). English translation: Sov. Math. Dokl. 23 (1981), 255258.
314. Goldshtein, V. M., Gurov, L.: Applications of change of variable operators for exact embedding theorems, Integral Equ. Oper. Theory 19
(1994), no. 1, 124.
315. Goldshtein, V. M., Gurov, L., Romanov, A.: Homeomorphisms that
induce monomorphisms of Sobolev spaces, Isr. J. Math. 91 (1995), 31
60.
316. Goldshtein, V. M., Reshetnyak, Yu. G.: Introduction to the Theory of
Functions with Generalized Derivatives, and Quasiconformal Mappings,
Nauka, Moscow, 1983; English translation: Quasiconformal Mappings
and Sobolev Spaces, Kluwer, Dordrecht, 1990.
820
References
317. Goldshtein, V. M., Troyanov, M.: Axiomatic theory of Sobolev spaces,

Expo. Math. 19 (2001), 289336.
318. Goldshtein, V. M., Troyanov, M.: Capacities in metric spaces, Integral
Equ. Oper. Theory 44 (2002), 212242.
319. Goldshtein, V. M., Ukhlov, A.: Sobolev homeomorphisms and composition operators, Around the Research of Vladimir Mazya I: Function
Spaces, International Mathematical Series 11, Springer, Berlin, 2010,
207220.
320. Goldshtein, V. M., Vodopyanov, S. K.: Prolongment des fonctions de
classe Llp et applications quasi conformes, C. R. Acad. Sci., Paris 290
(1980), 453456.
321. Golovkin, K. K.: Imbedding theorems for fractional spaces, Tr. Mat.
Inst. Steklova 70 (1964), 3846 (in Russian). English translation: Am.
Math. Soc. Transl. 91 (1970), 5767.
322. Golovkin, K. K.: Parametric-normed spaces and normed massives, Tr.
Mat. Inst. Steklova 106 (1969), 1135 (in Russian). English translation:
Proc. Steklov Inst. Math. 106 (1969), 1121.
323. Golovkin, K. K., Solonnikov, V. A.: Imbedding theorems for fractional
spaces, Dokl. Akad. Nauk SSSR 143 (1962), 767770 (in Russian). English translation: Sov. Math. Dokl. 3 (1962), 468471.
324. Grigoryan, A.: Isoperimetric inequalities and capacities on Riemannian
manifolds, The Mazya Anniversary Collection, vol. 1, Operator Theory,
Advances and Applications 110, Birkhauser, Basel, 1999, 139153.
325. Gromov, M.: Isoperimetric inequalities in Riemannian manifolds,
Asymptotic Theory of Finite-Dimensional Normed Spaces, Milman, V.
D., Schechtman, G. (Eds.), Lect. Notes Math. 1200, Springer, Berlin,
1986. With an appendix by M. Gromov.
326. Gromov, M., Milman, V. D.: Generalization of the spherical isoperimetric inequality to uniformly convex Banach spaces, Compos. Math. 62
(1987), 263282.
327. Gross, L.: Logarithmic Sobolev inequalities, Am. J. Math. 97 (1975),
no. 4, 10611083.
328. Gross, L.: Logarithmic Sobolev inequalities and contractivity properties
of semigroups, Lect. Notes Math. Springer, Berlin, 1993, 5488.
329. Grushin, V. V.: A problem for the entire space for a certain class of
partial dierential equations, Dokl. Akad. Nauk SSSR 146 (1962), 1251
1254 (in Russian). English translation: Sov. Math. Dokl. 3 (1962), 1467
1470.
330. Guionnet, A., Zegarlinski, B.: Lectures on logarithmic Sobolev inequalities, Seminaire de Probabilites, XXXVI Lect. Notes Math. 1801,
331. Gustin, W.: Boxing inequalities, J. Math. Mech. 9 (1960), 229239.
332. Guzman, M.: Dierentiation of Integrals in Rn . Lect. Notes Math. 481
References
821
333. Haberl, C., Schuster, F.: Asymmetric ane Lp Sobolev inequalities,

J. Funct. Anal. 257 (2009), 641658.
334. Hadwiger, H.: Vorlesungen u
ber Inhalt, Ober
ache und Isoperimetrie,
335. Han, Z.-C.: Asymptotic approach to singular solutions for nonlinear
elliptic equations involving critical Sobolev exponenet, Ann. Inst. H.
Poincare Anal. Non Lineaire 8 (1991), 159174.
336. Hahn, H., Rosenthal, A.: Set Functions, The University of New Mexico
Press, Albuquerque, 1948.
337. Hajlasz, P.: Geometric approach to Sobolev spaces and badly degenerated elliptic equations, Nonlinear Analysis and Applications, (Warsaw,
1994), GAKUTO Internat. Ser. Math. Sci. Appl. 7, Gakk
otosho, Tokyo,
1996, 141168.
338. Hajlasz, P.: Sobolev spaces on an arbitrary metric space, Potential Anal.
5 (1996), no. 4, 403415.
339. Hajlasz, P.: Pointwise Hardy inequalities, Proc. Am. Math. Soc. 127
(1999), no. 2, 417423.
340. Hajlasz, P.: Sobolev mappings, co-area formula and related topics, Proc.
on Analysis and Geometrym, Novosibirsk 2000, 227254.
341. Hajlasz, P.: Sobolev inequalities, truncation method, and John domains,
Rep. Univ. Jyvaskyl
a Dep. Math. Stat., 83, Jyv
askyl
a, 2001.
342. Hajlasz, P.: Sobolev spaces on metric-measure spaces, Heat Kernels and
Analysis on Manifolds, Graphs, and Metric Spaces (Paris, 2002), Contemp. Math. 338, Am. Math. Soc., Providence, 2003, 173218.
343. Hajlasz, P., Koskela, P.: Isoperimetric inequalities and imbedding theorems in irregular domains, J. Lond. Math. Soc., Ser. 2 58 (1998), 425
450.
344. Hajlasz, P., Koskela, P.: Sobolev met Poincare, Mem. Am. Math. Soc.
145 (2000), no. 688, x+101.
345. Hajlasz, P., Koskela, P., Tuominen, H.: Measure density and extendability of Sobolev functions, Rev. Mat. Iberoam. 24 (2008), no. 2, 645669.
346. Hajlasz, P., Mal
y, J.: Approximation in Sobolev spaces of nonlinear
expressions involving the gradient, Ark. Mat. 40 (2002), 245274.
347. Hansson, K.: On a maximal imbedding theorem of Sobolev type and
spectra of Schr
odinger operators. Link
oping Studies in Science and Technology. Dissertation. Link
oping, 1978.
348. Hansson, K.: Imbedding theorems of Sobolev type in potential theory,
Math. Scand. 45 (1979), 77102.
349. Hansson, K., Mazya, V., Verbitsky, I. E.: Criteria of solvability for
multi-dimensional Riccatis equation, Ark. Mat. 37 (1999), no. 1, 87
120.
350. Hardy, G. H., Littlewood, J. E., P
olya, G.: Some simple inequalities
satised by convex functions, Messenger Math. 58 (1929), 145152.
351. Hardy, G. H., Littlewood, J. E., P
olya, G.: Inequalities, Cambridge University Press, Cambridge, 1934.
822
References
352. Harjulehto, P., H

ast
o, P.: A capacity approach to the Poincare inequality
and Sobolev imbeddings in variable exponent Sobolev spaces, Rev. Mat.
Complut. 17 (2004), no. 1, 129146.
353. Harjulehto, P., H
ast
o, P., Koskenoja, M., Varonen, S.: Sobolev capacity
on the space W 1,p() (Rn ), J. Funct. Spaces Appl. 1 (2003), no. 1, 1733.
354. Haroske, D., Runst, T., Schmeisser, H.-J.: Function spaces, dierential
operators and nonlinear analysis, The Hans Triebel Anniversary Volume, Proceedings of the 5th International Conference Held in Teistungen, 2001.
355. Haroske, D. D., Triebel, H.: Distributions, Sobolev Spaces and Elliptic
Equations, EMS Textbooks in Mathematics, European Mathematical
Society, Z
urich, 2008.
356. Harvey, R., Polking, J.: Removable singularities of solutions of linear
partial dierential equations, Acta Math. 125 (1970), 3956.
357. Hayman, W. K.: Mulivalent Functions, Cambridge University Press,
Cambridge, 1958.
358. Hayman, W. K.: Some bounds for principal frequency, Appl. Anal. 7
(1977/1978), 247254.
359. Hebey, E.: Optimal Sobolev inequalities on complete Riemannian manifolds with Ricci curvature bounded below and positive injectivity radius,
Am. J. Math. 118 (1996), 291300.
360. Hebey, E.: Sobolev Spaces on Riemannian Manifolds, Lect. Notes Math.
1635, Springer, Berlin, 1996.
361. Hebey, E.: Nonlinear Analysis on Manifolds: Sobolev Spaces and Inequalities, Courant Lecture Notes in Mathematics 5, AMS, Providence,
1999.
362. Hebey, E., Vaugon, M.: The best constant problem in the Sobolev embedding theorem for complete Riemannian manifolds, Duke Math. J. 79
(1995), 235279.
363. Hebey, E., Vaugon, M.: Meilleures constantes dans le theorème
dinclusion de Sobolev, Ann. Inst. H. Poincare Anal. Non Lineaire 13
(1996), 5793.
364. Hedberg, L. I.: Weighted mean approximations in Caratheodory regions,
Math. Scand. 23 (1968), 113122.
365. Hedberg, L. I.: On certain convolution inequalities, Proc. Am. Math.
Soc. 36 (1972), 505510.
366. Hedberg, L. I.: Approximation in the mean by solutions of elliptic equations, Duke Math. J. 40 (1973), 916.
367. Hedberg, L. I.: Two approximation problems in function spaces, Ark.
Mat. 16 (1978), 5181.
368. Hedberg, L. I.: Spectral synthesis in Sobolev spaces and uniqueness of
solutions of the Dirichlet problem, Acta Math. 147 (1981), 237264.
369. Hedberg, L. I.: Spectral synthesis in Sobolev spaces, Linear and Complex
Analysis Problem Book, Lect. Notes Math. 1043, Springer, Berlin, 1984,
435437.
References
823
370. Hedberg, L. I.: Approximation in Sobolev spaces and nonlinear potential

theory, Proc. Symp. Pure Math., Part 1 45 (1986), 473480.
371. Hedberg, L. I., Netrusov, Y.: An axiomatic approach to function spaces,
spectral synthesis, and Luzin approximation, Mem. Am. Math. Soc. 188
(2007), no. 882, 197.
372. Hedberg, L. I., Wol, T. H.: Thin sets in nonlinear potential theory,
Ann. Inst. Fourier 33 (1983), 4.
373. Heinonen, J.: Lectures on Analysis on Metric Spaces, Universitext,
Springer, New York, 2001.
374. Heinonen, J., Holopainen, I.: Quasiregular maps on Carnot groups,
J. Geom. Anal. 7 (1997), no. 1, 109148.
375. Heinonen, J., Kilpelainen, T., Martio, O.: Nonlinear Potential Theory of
Degenerate Elliptic Equations, Oxford University Press, Oxford, 1993.
376. Hencl, S., Koskela, P.: Regularity of the inverse of a planar Sobolev
homeomorphism, Arch. Ration. Mech. Anal. 180 (2006), no. 1, 7595.
377. Hencl, S., Koskela, P., Mal
y, Y.: Regularity of the inverse of a Sobolev
homeomorphism in space, Proc. R. Soc. Edinb., Sect. A 136 (2006),
no. 6, 12671285.
378. Hestenes, M. R.: Extension of the range of dierentiable function, Duke
Math. J. 8 (1941), 183192.
379. Homann, D., Spruck, J.: Sobolev and isoperimetric inequalities for Riemannian submanifolds I, Comment. Pure Appl. Math. 27 (1974), 715
727; II. Commun. Pure Appl. Math. 28 (1975) 765766.
380. Homann-Ostenhof, M., Homann-Ostenhof, T., Laptev, A.: A geometrical version of Hardys inequality, J. Funct. Anal. 189 (2002), 539548.
381. Holley, R., Stroock, D. W.: Logarithmic Sobolev inequalities and
stochastic Ising models, J. Stat. Phys. 46 (1987), 11591194.
382. Holopainen, I., Rickman, S.: Quasiregular mappings, Heisenberg group
and Picards theorem, Proceedings of the Fourth FinnishPolish Summer School in Complex Analysis at Jyv
askyla, 1992, 2535.
383. Horiuchi, T.: The imbedding theorems for weighted Sobolev spaces,
J. Math. Kyoto Univ. 29 (1989), no. 3, 365403.
384. H
ormander, L.: Linear Partial Dierential Operators, Springer, Berlin,
1963.
385. H
ormander, L., Lions, J.-L.: Sur la completion par rapport a une
integrale de Dirichlet, Math. Scand. 4 (1956), 259270.
386. Hudson, S., Leckband, M.: A sharp exponential inequality for Lorentz
Sobolev spaces on bounded domains, Proc. Am. Math. Soc. 127 (1999),
no. 7, 20292033.
387. Hudson, S., Leckband, M.: Extremals for a MoserJodeit exponential
inequality, Pac. J. Math. 206 (2002), no. 1, 113128.
388. Hurd, A. E.: Boundary regularity in the Sobolev imbedding theorems,
Can. J. Math. 18 (1966), 350356.
389. Hurri, R.: Poincare domains in Rn , Ann. Acad. Sci. Fenn. Ser. A. I.
Math. Diss. 71 (1988), 142.
824
References
390. Hurri, R.: The weighted Poincare inequalities, Math. Scand. 67 (1990),
no. 1, 145160.
391. Hurri-Syrj
anen, R.: Unbounded Poincare domains, Ann. Acad. Sci.
Fenn. Ser. A. I. Math. 17 (1992), 409423.
392. Hurri-Syrjanen, R.: An improved Poincare inequality, Proc. Am. Math.
Soc. 120 (1994), no. 1, 213222.
393. Ilin, V. P.: Some inequalities in function spaces and their application
to the investigation of the convergence of variational processes. Dissertation, Leningrad, LGU, 1951 (in Russian).
394. Ilin, V. P.: On a imbedding theorem for a limiting exponent, Dokl.
Akad. Nauk SSSR 96 (1954), 905908 (in Russian).
395. Ilin, V. P.: Some integral inequalities and their application to the theory
of dierentiable functions of many variables, Mat. Sb. 54 (1961), 331
380 (in Russian).
396. Ilin, V. P.: Integral representations of dierentiable functions and their
application to questions of the extension of functions of the class Wpl (G),
Sib. Mat. Zh. 8 (1967), 573586 (in Russian). English translation: Sib.
Math. J. 8 (1967), 421432.
397. Ilin, V. P.: Integral representations of functions of the class LlP (G) and
imbedding theorems, Zap. Nauchn. Semin. Leningr. Otd. Math. Inst.
Steklova 19 (1970), 95155 (in Russian).
398. Inglis, J., Papageorgiou, I.: Logarithmic Sobolev inequalities for innite dimensional H
ormander type generators on the Heisenberg group,
Potential Anal. 31 (2009), 79102.
399. Iwaniec, T.: Nonlinear Dierential Forms, University Printing House,
Jyv
askyla, 1998.
400. Jawerth, B.: Some observations on Besov and LizorkiTriebel spaces,
Math. Scand. 40 (1977), 94104.
401. Jerison, D.: The Poincare inequality for vector elds satisfying
Hormanders condition, Duke Math. J. 53 (1986), 503523.
402. Jodeit, M.: An inequality for the indenite integral of a function in Lq ,
Stud. Math. 44 (1972), 545554.
403. Johnsen, J.: Traces of Besov spaces revisited, Z. Anal. Anwend. 19
(2000), no. 3, 763779.
404. Jones, P. W.: Quasiconformal mappings and extendability of functions
in Sobolev spaces, Acta Math. 147 (1981), 7188.
405. Jonsson, A.: The trace of potentials on general sets, Ark. Mat. 17 (1979),
118.
406. Jonsson, A.: Besov spaces on closed subsets of Rn , Trans. Am. Math.
Soc. 341 (1994).
407. Jonsson, A., Wallin, H.: A Whitney extension theorem in Lp and Besov
spaces, Ann. Inst. Fourier 28 (1978), 139192.
408. Jonsson, A., Wallin, H.: Functions Spaces on Subsets of Rn , Mathematical Reports 2, Harwood Academic, London, 1984.
References
825
409. Kac, I. S., Krein, M. G.: Criteria for discreteness of the spectrum of a
singular string, Izv. Vyss. Uchebn. Zaved. 2 (1958), 136153 (in Russian).
410. Kaimanovich, V.: Dirichlet norms, capacities and generalized isoperimetric inequalities for Markov operators, Potential Anal. 1 (1992), no. 1,
6182.
411. Kalamajska, A.: Pointwise interpolative inequalities and Nirenberg type
estimates in weighted Sobolev spaces, Stud. Math. 108 (1994), no. 3,
275290.
412. Kalton, N. J., Verbitsky, I. E.: Nonlinear equations and weighted norm
inequalities, Trans. Am. Math. Soc. 351 (1999), 9.
413. Kannan, R., Lov
asz, L., Simonovits, M.: Isoperimetric problems for convex bodies and a localization lemma, Discrete Comput. Geom. 13 (1995),
541559.
414. Karadzhov, G. E., Milman, M., Xiao, J.: Limits of higher order Besov
spaces and sharp reiteration theorems, J. Funct. Anal. 221 (2005), no.
2, 323339.
415. Kato, T.: Fundamental properties of Hamiltonian operators of
Schrodinger type, Trans. Am. Math. Soc. 70 (1951), 195211.
416. Kato, T.: Remarks on Schrodinger operators with vector potentials, Integral Equ. Oper. Theory 1 (1978), 103113.
417. Kauhanen, J., Koskela, P., Mal
y, J.: Mappings of nite distortion: condition N, Mich. Math. J. 49 (2001), 169181.
418. Kerman, R., Pick, L.: Optimal Sobolev imbeddings, Forum Math. 18
(2006), no. 4, 535570.
419. Kerman, R., Pick, L.: Compactness of Sobolev imbeddings involving
rearrangement-invariant norms, Stud. Math. 186 (2008), no. 2, 127160.
420. Kerman, R., Sawyer, E.: The trace inequality and eigenvalue estimates
for Schr
odinger operators, Ann. Inst. Fourier (Grenoble) 36 (1986), 207
228.
421. Kilpel
ainen, T., Kinnunen, J., Martio, O.: Sobolev spaces with zero
boundary values on metric spaces, Potential Anal. 12 (2000), 233247.
422. Kilpel
ainen, T., Mal
y, J.: Sobolev inequalities on sets with irregular
boundaries, Z. Anal. Anwend. 19 (2000), no. 2, 369380.
423. Kinnunen, J., Korte, R.: Characterizations of Sobolev inequalities on
metric spaces, J. Math. Anal. Appl. 344 (2008), no. 2, 10931104.
424. Kinnunen, J., Korte, R.: Characterizations for the Hardy inequality,
Around the Research of Vladimir Mazya I: Functions Spaces, International Mathematical Series 1113, Springer, Berlin, 2010.
425. Kinnunen, J., Martio, O.: The Sobolev capacity on metric spaces, Ann.
Acad. Sci. Fenn. Math. 21 (1996), no. 2, 367382.
426. Klimov, V. S.: Embedding theorems for symmetric spaces, Mat. Sb. (N.
S.) 79 (1969), no. 121, 171178 (in Russian).
427. Klimov, V. S.: The rearrangements of dierentiable functions, Mat. Zametki 9 (1971), 629638 (in Russian).
826
References
428. Klimov, V. S.: Imbedding theorems for Orlicz spaces and their applications to boundary value problems, Sib. Mat. Zh. 13 (1972), 334348 (in
Russian). English translation: Sib. Math. J. 13 (1972) 231240.
429. Klimov, V. S.: Isoperimetric inequalities and imbedding theorems, Dokl.
Akad. Nauk SSSR 217 (1974), 272275 (in Russian). English translation:
Sov. Math. Dokl. 15 (1974), 10471051.
430. Klimov, V. S.: Imbedding theorems and geometric inequalities, Izv.
Math. USSR Izv. 10 (1976), 615638.
431. Klimov, V. S.: To embedding theorems of anisotropic function classes,
Mat. Sb. 127 (1985), no. 2, 198208.
432. Klimov, V. S.: Capacities of sets and embedding theorems for ideal
spaces, Dokl. Akad. Nauk 341 (1995), no. 5, 588589 (in Russian).
433. Klimov, V. S.: Functional inequalities and generalized capacities, Mat.
Sb. 187 (1996), no. 1, 4154 (in Russian). English translation: Sb. Math.
187 (1996), no. 1, 3952.
434. Klimov, V. S.: On the symmetrization of anisotropic integral functionals,
Izv. Vyssh. Uchebn. Zaved. Mat. 8 (1999), no. 8, 2632 (in Russian).
English translation: Russ. Math. (Iz. VUZ) 43 (1999), no. 8, 2329.
435. Klimov, V. S.: Generalized multiplicative inequalities for ideal spaces,
Sib. Mat. Zh. 45 (2004), no. 1, 134149 (in Russian). English translation:
Sib. Math. J. 45(1) (2004), 112124.
436. Klimov, V. S., Panasenko, E. S.: Geometric properties of ideal spaces
and the capacities of sets, Sib. Math. J. 40 (1999), no. 3, 488499.
437. Knothe, H.: Contributions to the theory of convex bodies, Mich. Math.
J. 4 (1957), 3952.
438. Kokilashvili, V. M.: On Hardy inequalities in weighted spaces, Soobshch.
Gruz. Akad. Nauk SSR 96 (1979), 3740 (in Russian).
439. Kokilashvili, V. M. Samko, S.: Weighted boundedness of the maximal,
singular and potential operators in variable exponent spaces, Analytic
Methods of Analysis and Dierential Equations, A. A. Kilbas, S. V. Rogosin (Eds.), Cambridge Scientic Publisher, Cambridge, 2008, 139
164.
440. Kolsrud, T.: Approximation by smooth functions in Sobolev spaces.
A counterexample, Bull. Lond. Math. Soc. 13 (1981), 167169.
441. Kolsrud, T.: Condenser capacities and removable sets in W 1,p , Ann.
Acad. Sci. Fenn., Ser. A I Math. 8 (1983), no. 2, 343348.
442. Kolsrud, T.: Capacitary integrals in Dirichlet spaces, Math. Scand. 55
(1984), 95120.
443. Kolyada, V. I.: Estimates or reaarangements and embedding theorems,
Mat. Sb. 136 (1988), 323 (in Russian). English translation: Math. USSR
Sb. 64 (1989), 121.
444. Kolyada, V. I.: Rearrangements of functions and embedding of
anisotropic spaces of Sobolev type, East J. Approx. 4 (1998), 111199.
References
827
445. Kolyada, V. I., Lerner, A. K.: On limiting embeddings of Besov spaces,

Stud. Math. 171 (2005), no. 1, 113.
446. Kombe, I., Ozaydin,

M.: Improved Hardy and Rellich inequalities on
Riemannian manifolds, Trans. Am. Math. Soc. 361 (2009), 61916203.
447. Kondrashov, V. I.: On some properties of functions form the space Lp ,
Dokl. Akad. Nauk SSSR 48 (1945), 563566 (in Russian).
448. Kondratiev, V. A.: On the solvability of the rst boundary value problem
for elliptic equations, Dokl. Akad. Nauk SSSR 136 (1961), 771774 (in
Russian). English translation: Sov. Math. Dokl. 2 (1961) 127130.
449. Kondratiev, V. A.: Boundary problems for elliptic equations in domains
with conical or angular points, Tr. Mosk. Mat. Obs. 16 (1967), 209292
(in Russian). English translation: Trans. Mosc. Math. Soc. 16 (1967)
227313.
450. Kondratiev, V. A., Mazya, V., Shubin, M.: Discreteness of spectrum
and strict positivity criteria for magnetic Schr
odinger operators, Commun. Partial Dier. Equ. 29 (2004), 489521.
451. Kondratiev, V., Mazya, V., Shubin, M.: Gauge optimization and spectral properties of magnetic Schr
odinger operators, Commun. Partial Differ. Equ. 34 (2009), no. 1012, 11271146.
452. Korevaar, N., Schoen, R.: Sobolev spaces and harmonic maps for metric
space targets, Commun. Anal. Geom. 1 (1993), no. 4, 561659.
453. Korte, R., Shanmugalingam, N.: Equivalence and self-improvement of
p-fatness and Hardys inequality, and association with uniform perfectness, Math. Z. 264 (2010), no. 1, 99110.
454. Koskela, P.: Extensions and imbeddings, J. Funct. Anal. 159 (1998),
369383.
455. Koskela, P.: Removable sets for Sobolev spaces, Ark. Mat. 37 (1999),
291304.
456. Koskela, P.: Sobolev spaces and quasiconformal mappings on metric
spaces, European Congress of Mathematics, vol. I (Barcelona, 2000),
Progr. Math. 201, Birkhauser, Basel, 2001, 457467.
457. Koskela, P., Miranda, Jr. M., Shanmugalingam, N.: Geometric properties of planar BV-extension domains, Around the Research of Vladimir
Mazya I: Function Spaces, International Math. Ser. 11, Springer, Berlin,
2010, 255272.
458. Koskela, P., Onninen, J.: Sharp inequalities via truncation, J. Math.
Anal. Appl. 278 (2003), no. 2, 324334.
459. Koskela, P., Onninen, J., Tyson, J. T.: Quasihyperbolic boundary conditions and Poincare domains, Math. Ann. 323 (2002), no. 4, 811830.
460. Koskela, P., Stanoyevitch, A.: Poincare inequalities and Steiner symmetrization, Ill. J. Math. 40 (1996), no. 3, 365389.
461. Kov
acik, O., R
akosnik, J.: On spaces Lp(x) and W k,p(x) , Czechoslov.
Math. J. 41 (1991), no. 116, 592618.
462. Krahn, E.: Uber

eine von Rayleigh formulierte Minimaleigenschaft des
Kreises, Math. Ann. 94 (1925), 97100.
828
References
463. Krasnoselski, M. A., Ruticki, Ya. B.: Convex functions and Orlicz
spaces, Nauka, Moscow, 1961. English edition: Noordho, Groningen,
1961.
464. Krickeberg, K.: Distributionen, Funktionen beschr
ankter Variation und
Lebesguescher Inhalt nichtparametrischer Fl
achen, Ann. Mat. Pura
Appl., IV Ser. 44 (1957), 105134.
465. Kronrod, A. S.: On functions of two variables, Usp. Mat. Nauk 5 (1950),
24134 (in Russian).
466. Kudryavtsev, L. D.: Direct and inverse imbedding theorems. Applications to the solution of elliptic equations by the variational method, Tr.
Am. Math. Soc. Transl. 42 (1974).
467. Kufner, A.: Weighted Sobolev Spaces, Teubner, Leipzig, 1980.
468. Kufner, A., Maligranda, L., Persson, L. E.: The Hardy Inequality. About
Its History and Some Related Results, Vydavatelsk
y Servis, Plze
n, 2007.
469. Kufner, A., Persson, L.-E.: Weighted Inequalities of Hardy Type, World
Scientic, Singapore, 2003.
470. Labutin, D. A.: Integral representations of functions and embeddings
of Sobolev spaces on cuspidal domains, Mat. Zametki 61 (1997), no. 2,
201219 (in Russian). English translation: Math. Notes 61 (1997), 164
179.
471. Labutin, D. A.: Embedding of Sobolev spaces on H
older domains, Proc.
Steklov Inst. Math. 227 (1999), 163172 (in Russian). English translation: Tr. Mat. Inst. 227 (1999), 170179.
472. Labutin, D. A.: Superposition operator in Sobolev spaces on domains,
Proc. Am. Math. Soc. 128 (2000), no. 11, 33993403.
473. Labutin, D. A.: The unimprovability of Sobolev inequalities for a class of
irregular domains, Tr. Mat. Inst. Steklova 232 (2001), 218222 (in Russian). English translation: Proc. Steklov Inst. Math. 232 (2001), no. 1,
211215.
474. Ladyzhenskaya, O. A.: Mathematical Questions of the Dynamics of a
Viscous Incompressible Fluid, Nauka, Moscow, 1970 (in Russian).
475. Landis, E. M.: Second Order Equations of Elliptic and Parabolic Types,
Nauka, Moscow, 1971 (in Russian).
476. Landis, E. M.: On the behaviour of solutions of higher-order elliptic
equations in unbounded domains, Tr. Mosk. Mat. Obs. 31 (1974), 35
58 (in Russian). English translation: Trans. Mosc. Math. Soc. 31 (1974)
3054.
477. Landkof, N. S.: Foundations of Modern Potential Theory, Nauka,
Moscow, 1966 (in Russian). English edition: Springer, Berlin, 1972.
478. Lang, J., Mazya, V.: Essential norms and localization moduli of Sobolev
embeddings for general domains, J. Lond. Math. Soc. 78 (2008), no. 2,
373391.
479. Laptev, A. (Ed.): Around the Research of Vladimir Mazya I, International Math. Series 11, Springer, Berlin, 2010.
References
829
480. Laptev, A., Sobolev, A. V.: Hardy inequalities for simply connected
planar domains, Am. Math. Soc. Transl., Ser. 2, Adv. Math. Sci., 225,
(2008), 133140.
481. Laptev, A., Weidl, T.: Hardy inequalities for magnetic Dirichlet forms,
Mathematical Results in Quantum Mechanics (Prague, 1998), Oper.
Theory Adv. Appl. 108, Birkh
auser, Basel, 1999, 299305.
482. Laptev, S. A.: Closure in the metric of a generalized Dirichlet integral,
Dier. Uravn. 7 (1971), 727736. (in Russian). English translation: Differ. Equ. 7 (1971) 557564.
483. Ledoux, M.: Concentration of measures and logarithmic Sobolev inequalities, Lect. Notes Math. 1709 (1999), 120216.
484. Ledoux, M.: Logarithmic Sobolev inequalities for unbounded spin systems revisited, Seminaire de Probabilites, XXXV, Lect. Notes Math.
1755, Springer, New York, 2001, 167194.
485. Ledoux, M.: On improved Sobolev imbedding theorems, Math. Res.
Lett. 10 (2003), no. 56, 659669.
486. Leoni, G.: A First Course in Sobolev Spaces, Graduate Studies in Mathematics 105. American Mathematical Society, Providence, 2009.
487. Leray, J.: Sur le mouvement visqueux emplissant lespace, Acta Math.
63 (1934), 193248.
488. Leray, J., Lions, J.-L.: Quelques resultats de Vishik sur les problemes
elliptiques non lineaires par les methodes de MintyBrowder, Bull. Soc.
Math. Fr. 93 (1965), 97107.
489. Levi, B.: Sul prinzipio di Dirichlet, Rend. Palermo 22 (1906), 293359.
490. Levin, V. I.: Exact constants in inequalities of the Carleson type, Dokl.
491. Lewis, J. L.: Approximations of Sobolev functions and related topics,
Complex analysis, I (College Park, MD, 198586), Lect. Notes Math.
1275, Springer, Berlin, 1987, 223234.
492. Lewis, J. L.: Approximation of Sobolev functions in Jordan domains,
Ark. Mat. 25 (1987), 255264.
493. Lewis, J. L.: Uniformly fat sets, Trans. Am. Math. Soc. 308 (1988),
177196.
494. Lichtenstein, L.: Eine elementare bemerkung zur analysis, Math. Z. 30
(1929), 794795.
495. Lieb, E. H.: On the lowest eigenvalue of the Laplacian for the intersection
of two domains, Invent. Math. 74 (1983), 441448.
496. Lieb, E. H.: Sharp constants in the HardyLittlewoodSobolev and related inequalities, Ann. Math. 118, (1983), 349374.
497. Lieb, E., Loss, M.: Analysis, Graduate Studies in Mathematics 14, AMS,
Provindence, 2001.
498. Lin, C. S.: Interpolation inequalities with weighted norms, Commun.
Partial Dier. Equ. 11 (1986), 15151538.
499. Lions, J.-L.: Ouverts m-reguliers, Rev. Uni
on Mat. Argent. 17 (1955),
103116.
830
References
500. Lions, J.-L., Magenes, B.: Problemes aux limites non homogenes et
applications, Dunod, Paris, 1968. English edition: Non-Homogeneous
Limit Problems and Applications, Springer, Berlin, 1972.
501. Lions, P.-L.: The concentration-compactness principle in the calculus of
variations. The limit case, II, Rev. Mat. Iberoam. 1 (1985), 145121.
502. Lions, P. L., Pacella, F.: Isoperimetric inequalities for convex cones,
Proc. Am. Math. Soc. 109 (1990), 477485.
503. Littman, W.: Polar sets and removable singularities of partial dierential
equations, Arch. Math. 7 (1967), 19.
504. Lizorkin, P. I.: Boundary properties of functions from weight classes,
Dokl. Akad. Nauk SSSR 132 (1960), 514517 (in Russian). English translation: Sov. Math. Dokl. 1 (1960) 589593.
505. Lizorkin, P. I.: Estimates of integrals of potential type in norms with
dierence relations, Theory of Cubature Formulas and Applications of
Functional Analysis to Some Problems of Mathematical Physics, Novosibirsk, 1975, 94109 (in Russian).
506. Lizorkin, P. I., Otelbaev, M.: Imbedding theorems and compactness for
spaces of Sobolev type with weights, I, Mat. Sb. 108 (1979), 358377
II. Mat. Sb. 112 (1980) 5685 (in Russian). English translation: Math.
USSR Sb. 36 (1980) 331349; 40 (1981) 5177.
507. Lyusternik, L. A.: Brunn-Minkowski inequality for arbitrary sets, Dokl.
508. Lovasz, L., Simonovits, M.: Random walks in a convex body and an
improved volume algorithm, Random Struct. Algorithms 4 (1993), no. 4,
359412.
509. Lugiewicz, P., Zegarlinski, B.: Coercive inequalities for H
ormander type
generators in innite dimensions, J. Funct. Anal. 247 (2007), no. 2, 438
476.
510. Lutwak, E., Yang, D., Zhang, G.: Sharp ane Lp Sobolev inequalities,
J. Dier. Geom. 62 (2002), 1738.
511. Luukkainen, J., Saksman, E.: Every complete doubling metric space
carries a doubling measure, Proc. Am. Math. Soc. 126 (1998), 531534.
512. Maggi, F., Villani, C.: Balls have the worst best Sobolev inequalities,
J. Geom. Anal. 15 (2005), no. 1, 83121.
513. Maheux, P., Salo-Coste, L.: Analyse sur les boules dun operateur souselliptique, Math. Ann. 303 (1995), 713740.
514. Mal
y, J., Swanson, D., Ziemer, W. P.: The coarea formula for Sobolev
mappings, Trans. Am. Math. Soc. 355 (2003), 477492.
515. Marcus, M., Mizel, V. J.: Every superposition operator mapping one
Sobolev space into another is continuous, J. Funct. Anal. 33 (1979),
217229.
516. Marcus, M., Mizel, V. J., Pinchover, Y.: ON the best constant for
Hardys inquality in Rn , Trans. Am. Math. Soc. 350 (1998), 32373255.
517. Martin, J., Milman, M.: Sharp GagliardoNirenberg inequalities via
symmetrization, Math. Res. Lett. 14 (2006), no. 1, 4962.
References
831
518. Martin, J., Milman, M.: Symmetrization inequalities and Sobolev embeddings, Proc. Am. Math. Soc. 134 (2006), 23352347.
519. Martin, J., Milman, M.: Higher order symmetrization inequalities and
applications, J. Math. Anal. Appl. 330 (2007), 91113.
520. Martin, J., Milman, M.: Self improving SobolevPoincare inequalities
truncation and symmetrization, Potential Anal. 29 (2008), no. 4, 391
408.
521. Martin, J., Milman, M.: Isoperimetry and symmetrization for logarithmic Sobolev inequalities, J. Funct. Anal. 256 (2009), no. 1, 149178.
522. Martin, J., Milman, M.: Pointwise symmetrization inequalities for
Sobolev functions and applications, arXiv:0908.1751 (2009).
523. Martin, J., Milman, M.: Isoperimetric Hardy type and Poincare inequalities on metric spaces, Around the Research of V. Mazya I: Functions
Spaces, Springer, Berlin, 2010, 285298.
524. Martin, J., Milman, M., Pustylnik, E.: Symmetrization and selfimproving Sobolev via truncation, truncation and symmetrization,
J. Funct. Anal. 252 (2007), 677695.
525. Martio, O.: John domains, bilipschitz balls, and Poincare inequality,
Rev. Roum. Math. Pures Appl. 33 (1988), 107112.
526. Matskewich, T., Sobolevskii, P. E.: The best possible constant in a generalized Hardys inequality for convex domains in Rn , Nonlinear Anal.
28 (1997), 16011610.
527. Mazya, V. G.: Classes of regions and imbedding theorems for function spaces, Dokl. Akad. Nauk SSSR 133 (1960), 527530 (in Russian).
English translation: Sov. Math. Dokl. 1 (1960), 882885.
528. Mazya, V. G.: The p-conductivity and theorems on imbedding certain
function spaces into a C-space, Dokl. Akad. Nauk SSSR 140 (1961),
299302 (in Russian). English translation: Sov. Math. Dokl. 2 (1961),
12001203.
529. Mazya, V. G.: Classes of sets and imbedding theorems for function
spaces, Dissertation, MGU, Moscow, 1962 (in Russian).
530. Mazya, V. G.: On solvability of the Neumann problem, Dokl. Akad.
Nauk SSSR 147 (1962), 294296 (in Russian). English translation: Sov.
Math. Dokl. 3 (1962).
531. Mazya, V. G.: The negative spectrum of the n-dimensional Schr
odinger
operator, Dokl. Akad. Nauk SSSR 144 (1962), 721722 (in Russian).
English translation: Sov. Math. Dokl. 3 (1962), 808810.
532. Mazya, V. G.: On the boundary regularity of solutions of elliptic equations and of a conformal mapping, Dokl. Akad. Nauk SSSR 152 (1963),
533. Mazya, V. G.: The Dirichlet problem for elliptic equations of arbitrary
order in unbounded regions, Dokl. Akad. Nauk SSSR 150 (1963), 1221
1224 (in Russian). English translation: Sov. Math. Dokl. 4 (1963), 860
863.
832
References
534. Mazya, V. G.: On the theory of the multidimensional Schrodinger operator, Izv. Akad. Nauk SSSR 28 (1964), 11451172 (in Russian).
535. Mazya, V. G.: Polyharmonic capacity in the theory of the rst boundary
value problem, Sib. Mat. Zh. 6 (1965), 127148 (in Russian).
536. Mazya, V. G.: On closure in the metric of the generalized Dirichlet integral, Zap. Nauchn. Semin. Leningr. Otd. Mat. Inst. Steklova 5 (1967),
537. Mazya, V. G.: Neumanns problem in domains with nonregular boundaries, Sib. Mat. Zh. 9 (1968), 13221350 (in Russian). English translation: Sib. Math. J. 9 (1968), 9901012.
538. Mazya, V. G.: On weak solutions of the Dirichlet and Neumann problems, Tr. Mosk. Mat. Obs. 20 (1969), 137172 (in Russian). English
translation: Trans. Mosc. Math. Soc. 20 (1969), 135172.
539. Mazya, V. G.: Classes of sets and measures connected with imbedding
theorems, Imbedding Theorems and Their Applications (Baku, 1966),
Nauka, Moscow, 1970, 142159 (in Russian).
540. Mazya, V. G.: Applications of certain integral inequalities to the theory of quasi-linear elliptic equations, Comment. Math. Univ. Carol. 13
(1972), 535552 (in Russian).
541. Mazya, V. G.: On a degenerating problem with directional derivative,
Mat. Sb. 87 (1972), 417454 (in Russian). English translation: Math.
USSR Sb. 16 (1972), 429469.
542. Mazya, V. G.: Removable singularities of bounded solutions of quasilinear elliptic equations of any order, Zap. Nauchn. Semin. Leningr. Otd.
J. Sov. Math. 3 (1975), 480492.
543. Mazya, V. G.: Certain integral inequalities for functions of many variables, Problems in Mathematical Analysis 3 LGU, Leningrad, 1972, 33
68 (in Russian). English translation: J. Sov. Math. 1 (1973), 205234.
544. Mazya, V. G.: On (p, l)-capacity, imbedding theorems and the spectrum
of a selfadjoint elliptic operator, Izv. Akad. Nauk SSSR, Ser. Mat. 37
(1973), 356385 (in Russian). English translation: Math. USSR Izv., 7
(1973), 357387.
545. Mazya, V. G.: On the continuity and boundedness of functions from
Sobolev spaces, Problems in Mathematical Analysis, 4 LGU, Leningrad,
1973, 4677 (in Russian).
546. Mazya, V. G.: On the connection between two kinds of capacity, Vestn.
Leningr. Univ., Mat. Mekh. Astron. 7 (1974), 3340 (in Russian). English translation: Vestn. Leningr. Univ. Math. 7 (1974), 135145.
547. Mazya, V. G.: On the summability of functions in S. L. Sobolev spaces,
Problems in Mathematical Analysis, 5. LGU, Leningrad 1975, 6698 (in
Russian).
548. Mazya, V. G.: Capacity-estimates for fractional norms Zap. Nauchn.
Semin. Leningr. Otd. Mat. Inst. Steklova 70 (1977), 161168 (in Russian). English translation: J. Sov. Math. 23 (1983), 19972003.
References
833
549. Mazya, V. G.: On an integral inequality, Semin. Inst. Prikl. Mat. Dokl.
Tbil. Univ., 1213 (1978), 3336 (in Russian).
550. Mazya, V. G.: Behaviour of solutions to the Dirichlet problem for the
biharmonic operator at a boundary point, Lect. Notes Math. 703 (1979),
250262.
551. Mazya, V. G.: Summability with respect to an arbitrary measure of
functions in S. L. Sobolev-L. N. Slobodecki spaces, Zap. Nauchn. Semin.
Leningr. Otd. Mat. Inst. Steklova 92 (1979), 192202 (in Russian).
552. Mazya, V. G.: Einbettungss
atze f
ur Sobolewsche Raume, Teil 1, Teubner, Leipzig, 1979; Teil 2, 1980.
553. Mazya, V. G.: Integral representation of functions satisfying homogeneous boundary conditions and its applications, Izv. Vyssh. Uchebn.
Zaved., Mat. 2 (1980), 3444 (in Russian). English translation: Sov.
Math. 24 (1980), 3544.
554. Mazya, V. G.: On an imbedding theorem and multipliers in pairs of S.
L. Sobolev spaces, Tr. Tbil. Mat. Inst. Razmadze 66 (1980), 5960 (in
Russian).
555. Mazya, V. G.: Zur Theorie Sobolewscher R
aume, Teubner, Leipzig,
1981.
556. Mazya, V. G.: Sobolev Spaces, Springer, Berlin, 1985.
557. Mazya, V. G.: Classes of domains, measures and capacities in the theory
of dierentiable functions, Analysis III, Spaces of Dierentiable Functions, Encyclopaedia of Math. Sciences 26, Springer, Berlin, 1991, 141
211.
558. Mazya, V. G.: On the Wiener-type regularity of a boundary point for
the polyharmonic operator, Appl. Anal. 71 (1999), 149165.
559. Mazya, V. G.: The Wiener test for higher order elliptic equations, Duke
Math. J., 115 (2002), no. 3, 479572.
560. Mazya, V. G.: Lectures on isoperimetric and isocapacitary inequalities
in the theory of Sobolev spaces, Contemp. Math. 338 (2003), 307340.
561. Mazya, V. G.: Conductor and capacitary inequalities for functions on
topological spaces and their applications to Sobolev type imbeddings,
J. Funct. Anal. 224 (2005), no. 2, 408430.
562. Mazya, V. G.: Conductor inequalities and criteria for Sobolev type twoweight imbeddings, J. Comput. Appl. Math., 194 (2006) no. 1, 94114.
563. Mazya, V. G.: Analytic Criteria in the Qualitative Spectral Analysis
of the Schr
odinger Operator, Part 1, Proc. of Symposia in Pure Mathematics 76, Am. Math. Soc., Providence, 2007.
564. Mazya, V. G.: Integral and isocapacitary inequalities, Linear and Complex Analysis, Amer. Math. Soc. Transl. Ser. 2 226, Am. Math. Soc.,
Providence, 2009, 85107.
565. Mazya, V. G. (Ed.): Sobolev Spaces in Mathematics I, Sobolev Type
Inequalities, International Mathematical Series 8, Springer/Tamara
Rozhkovskaya Publisher, New York/Novosibirsk, 2009.
834
References
566. Mazya, V. G., Havin, V. P.: A nonlinear analogue of the Newton potential and metric properties of the (p, l)-capacity, Dokl. Akad. Nauk
SSSR 194 (1970), 770773 (in Russian). English translation: Sov. Math.
Dokl. 11 (1970), 12941298.
567. Mazya, V. G., Havin, V. P.: Nonlinear potential theory, Usp. Mat. Nauk
27 (1972), 67138 (in Russian). English translation: Russ. Math. Surv.
27 (1972), 71148.
568. Mazya, V. G., Havin, V. P.: On approximation in the mean by analytic
functions, Vestn. Leningr. Univ. 23 (1968), 6274 (in Russian). English
translation: Vestn. Leningr. Univ. Math. 1 (1974), 231245.
569. Mazya, V. G., Havin, V. P.: Use of (p, l)-capacity in problems of the
theory of exceptional sets, Mat. Sb. 90 (1973), 558591 (in Russian).
English translation: Math. USSR Sb., 19 (1973), 547580.
570. Mazya, V. G., Khvoles, A. A.: On imbedding the space Llp () into the
space of generalized functions, Tr. Tbil. Mat. Inst. Razmadze 66 (1981),
7083 (in Russian).
571. Mazya, V. G., Kufner, A.: Variations on the theme of the inequality
(f )2 2f sup|f |, Manuscr. Math. 56 (1986), 89104.
572. Mazya, V. G., Netrusov, Yu.: Some counterexamples for the theory of
Sobolev spaces on bad domains, Potential Anal. 4 (1995), 4765.
573. Mazya, V. G., Otelbaev, M.: Imbedding theorems and the spectrum of
a psuedodierential operator, Sib. Mat. Zh. 18 (1977), 10731087 (in
Russian). English translation: Sib. Math. J. 18 (1977), 758769.
574. Mazya, V. G., Poborchi, S.: Embedding and Extension Theorems for
Functions on Non-Lipschitz Domains, St. Petersburg University Publishers, St. Petersburg, 2006.
575. Mazya, V. G., Poborchi, S.: Imbedding theorems for Sobolev spaces on
domains with peak and in Holder domains, St. Petersburg Math. J. 18
(2007), no. 4, 583605.
576. Mazya, V. G., Poborchi, S. V.: Dierentiable Functions on Bad Domains, World Scientic, Singapore, 1997.
577. Mazya, V. G., Preobrazhenski, S. P.: On estimates of (p)(l)-capacity
and traces of potentials, Wiss. Inf. Tech. Hochs. Karl-Marx-Stadt, Sekt.
Math., 28 (1981), 138 (in Russian).
578. Mazya, V. G., Shaposhnikova, T. O.: Change of variable as an operator
in a pair of S. L. Sobolev spaces, Vestn. Leningr. Univ. 1 (1982), 43
48 (in Russian). English translation: Vestn. Leningr. Univ. Math. 15
(1983), 5358.
579. Mazya, V. G., Shaposhnikova, T.: Jacques Hadamard, a Universal
Mathematician, AMS/LMS, Providence, 1998.
580. Mazya, V. G., Shaposhnikova, T.: On pointwise interpolation inequalities for derivatives, Math. Bohem. 124 (1999), no. 23, 131148.
581. Mazya, V. G., Shaposhnikova, T.: Pointwise interpolation inequalities
for Riesz and Bessel potentials, Analytical and Computational Methods
References
582.
583.
584.
585.
586.
587.
588.
589.
590.
591.
592.
593.
594.
595.
835
in Scattering and Applied Mathematics, Chapman and Hall, London,

2000, 217229.
Mazya, V. G., Shaposhnikova, T.: Maximal Banach algebra of multipliers between Bessel potential spaces, Problems and Methods in Mathematical Physics, The Siegfried Prossdorf Memorial Volume, J. Elschner,
I. Gohberg, B. Silbermann (Eds.), Operator Theory: Advances and Application 121, Birkhauser, Basel, 2001, 352365.
Mazya, V. G., Shaposhnikova, T.: An elementary proof of the Brezis and
Mironescu theorem on the composition operator in fractional Sobolev
spaces, J. Evol. Equ. 2 (2002), 113125.
Mazya, V. G., Shaposhnikova, T.: On the Bourgain, Brezis, and
Mironescu theorem concerning limiting embeddings of fractional
Sobolev spaces, J. Funct. Anal. 195 (2002), no. 2, 230238. Erratum:
J. Funct. Anal. 201 (2003), no. 1, 298300.
Mazya, V. G., Shaposhnikova, T.: On the Brezis and Mironescu conjecture concerning a GagliardoNirenberg inequality for fractional Sobolev
norms, J. Math. Pures Appl. (9) 81 (2002), no. 9, 877884.
Mazya, V. G., Shaposhnikova, T.: Sharp pointwise interpolation inequalities for derivatives, Funct. Anal. Appl. 36 (2002), no. 1, 3048.
Mazya, V. G., Shaposhnikova, T.: A collection of sharp dilation invariant integral inequalities for dierentiable functions. Sobolev spaces in
mathematics, I: Sobolev type inequalities, IMS 8 (2008), 223.
Mazya, V. G., Shaposhnikova, T.: Theory of Sobolev Multipliers with
Applications to Dierential and Integral Operators, Grundlehren der
Mathematischen Wissenschaften 337. Springer, Berlin, 2009.
Mazya, V. G., Shubin, M. A.: Can one see the fundamental frequency
of a drum? Lett. Math. Phys. 74 (2005), no. 2, 135150.
Mazya, V. G., Shubin, M. A.: Discreteness of spectrum and positivity
criteria for Schr
odinger operators, Ann. Math. 162 (2005), 919942.
Mazya, V. G., Verbitsky, I.: Capacitary estimates for fractional integrals, with applications to partial dierential equations and Sobolev
multipliers, Ark. Mat. 33 (1995), 81115.
Mazya, V. G., Verbitsky, I.: The Schr
odinger operator on the energy
space: boundedness and compactness criteria, Acta Math. 188 (2002),
263302.
Mazya, V. G., Verbitsky, I.: Boundedness and compactness criteria for
the one-dimensional Schrodinger operator, Function Spaces, Interpolation Theory and Related Topics (Lund, 2000), de Gruyter, Berlin, 2002,
369382.
Mazya, V. G., Verbitsky, I. E.: Innitesimal form boundedness and
Trudingers subordination for the Schr
odinger operator, Invent. Math.
162 (2005), 81136.
McKean, H. P.: An upper bound to the spectrum of on a manifold of
negative curvature, J. Dier. Geom. 4 (1970), 359366.
836
References
596. Meyers, N. G.: A theory of capacities for potentials of functions in

Lebesgue classes, Math. Scand. 26 (1970), 255292.
597. Meyers, N. G.: Continuity of Bessel potentials, Isr. J. Math. 11 (1972),
271283.
598. Meyers, N. G.: Integral inequalities of Poincare and Wirtinger type,
Arch. Ration. Mech. Anal. 68 (1978), 113120.
599. Meyers, N. G., Serrin, J.: H = W , Proc. Natl. Acad. Sci. USA 51 (1964),
10551056.
600. Michael, J. H., Simon, L. M.: Sobolev and mean-value inequalities on
generalized submanifolds of Rn , Commun. Pure Appl. Math. 26 (1973),
362379.
601. Mikhlin, S. G.: Multidimensional Singular Integrals and Integral Equations, Nauka, Moscow, 1962 (in Russian).
602. Mikkonen, P.: On the Wol potential and quasilinear elliptic equations
involving measures, Ann. Acad. Sci. Fenn. Math. Diss. 104 (1996), 171.
603. Milman, E.: On the role of convexity in functional and isoperimetric
inequalities, Proc. Lond. Math. Soc. 99 (2009), no. 3, 3266.
604. Milman, E.: On the role of convexity in isoperimetry, spectral-gap and
concentration, Invent. Math. 177 (2009), no. 1, 143.
605. Milman, E.: A converse to the Mazya inequality for capacities under
curvature lower bound, Around the Research of Vladimir Mazya I:
Function Spaces, International Mathematical Series 11, 2010, 321348.
606. Milman, M.: Notes on limits of Sobolev spaces and the continuity of
interpolation scales, Trans. Am. Math. Soc. (2005), no. 9, 34253442.
607. Milman, M., Pustylnik, E.: On sharp higher order Sobolev embeddings,
Commun. Contemp. Math. 6 (2004), 495511.
608. Miranda, M.: Disuguaglianze di Sobolev sulle ipersuperci minimali,
Rend. Semin. Mat. Univ. Padova 38 (1967), 6979.
609. Mircea, M., Szeptycki, P.: Sharp inequalities for convolution operators
with homogeneous kernels and applications, Indiana Univ. Math. J. 46
(1997), 3.
610. Molchanov, A. M.: On conditions for discreteness of the spectrum of
selfadjoint dierential equations of the second order, Tr. Mosk. Mat.
Obs. 2 (1953), 169200 (in Russian).
611. Monti, R., Morbidelli, D.: Isoperimetric inequality in the Grushin plane,
J. Geom. Anal. 14 (2004), no. 2, 355368.
612. Morrey, C. B.: Functions of several variables and absolute continuity II,
Duke Math. J. 6 (1940), 187215.
613. Morrey, C. B.: Multiple Integrals in the Calculus of Variations, Springer,
Berlin, 1966.
614. Morse, A. P.: The behaviour of a function on its critical set, Ann. Math.
40 (1939), 6270.
615. Morse, A. P.: A theory of covering and dierentiation, Trans. Am. Math.
Soc. 55 (1944), 205235.
616. Morse, A. P.: Perfect blankets, Trans. Am. Math. Soc. 61 (1947), 418
442.
References
837
617. Moser, J.: On Harnacks theorem for elliptic dierential equations, Commun. Pure Appl. Math. 14 (1961), 577591.
618. Moser, J.: A sharp form of an inequality by N. Trudinger, Indiana Univ.
Math. J. 20 (1971), 10771092.
619. Mossino, J.: Inegalites Isoperimetriques et Applications en Physique,
Trabaux en Cours, Hermann, Paris, 1984.
620. Muckenhoupt, B.: Hardys inequality with weights, Stud. Math. 44
(1972), 3138.
621. Muckenhoupt, B.: Weighted norm inequalities for the Hardy maximal
function, Trans. Am. Math. Soc. 165 (1972), 207226.
622. Mynbaev, K. T., Otelbaev, M. O.: Weighted Functional Spaces and
Dierential Operator Spectrum, Nauka, Moscow, 1988.
623. Nadirashvili, N.: Isoperimetric inequality for the second eigenvalue of a
sphere, J. Dier. Geom. 61 (2002), 335340.
624. Naumann, J.: Remarks on the prehistory of Sobolev spaces, Preprint
series: Inst f
ur Math., Humboldt-Universit
at zu Berlin (2002), 134.
625. Nash, J.: Continuity of solutions of parabolic and elliptic equations, Am.
J. Math. 80 (1958), 931954.
626. Nasyrova, M., Stepanov, V.: On maximal overdetermined Hardys inequality of second order on a nite interval, Math. Bohem. 124 (1999),
293302.
627. Natanson, I. P.: Theory of Functions of a Real Variable, Nauka, Moscow,
1974 (in Russian). English edition: Ungar, New York, 1955. German
edition: Akademie-Verlag, Berlin, 1981.
628. Nazarov, A. I.: HardySobolev inequalities in a cone, J. Math. Sci. 132
(2006), no. 4, 419427.
629. Nazarov, F., Sodin, M., Volberg, A.: The geometric KLS lemma,
dimension-free estimates for the distribution of values of polynomials,
and distribution of zeroes of random analytic functions, Algebra Anal.
14 (2002), no. 2, 214234 (in Russian). Translation in: St. Petersburg
Math. J., 14(2) (2003), 351366.
630. Necas, J.: Les Methodes Directes en Theorie des Equations Elliptiques,
Academia, Prague, 1967.
631. Netrusov, Yu. V.: Sets of singularities of functions in spaces of Besov
and LizorkinTriebel type, Tr. Math. Inst. Steklov 187 (1990), 185203.
632. Netrusov, Yu. V.: Spectral synthesis in spaces of smooth functions, Dokl.
Ross. Akad. Nauk 325 (1992), 923.
633. Netrusov, Yu. V.: Estimates of capacities associated with Besov spaces,
Zap. Nauchn. Sem. St. Peterburg. Otdel. Mat. Inst. Steklov 201 (1992),
Issled. Linein. Oper. Teor. Funktskii 20, 124156, 191; English translation: J. Math. Sci. 78 (1996), no. 2, 199217.
634. Netrusov, Yu. V.: Spectral synthesis in the Sobolev space generated by
an integral metric, J. Math. Sci. 85 (1997), no. 2, 18141826.
838
References
635. Nevanlinna, R.: Eindeutige Analytische Funktionen, Springer, Berlin,

1936. Russian edition: Single-Valued Analytic Functions. GTTI,
Moscow, 1941.
636. Nieminen, E.: Hausdor measures, capacities and Sobove spaces with
weights, Ann. Acad. Sci. Fenn. 81 (1991), 1.
637. Nikod
ym, O.: Sur une classe de fonctions considerees dans le probleme
de Dirichlet, Fundam. Math. 21 (1933), 129150.
638. Nikolsky, S. M.: Properties of certain classes of functions of several variables on dierentiable manifolds, Mat. Sb. 33 (1953), 261326 (in Russian).
639. Nikolsky, S. M.: Approximation of Functions of Several Variables and
Imbedding Theorems, Nauka, Moscow, 1977 (in Russian).
640. Nirenberg, L.: On elliptic partial dierential equations (lecture II), Ann.
Sc. Norm. Super. Pisa, Cl. Sci. 13 (1959), 115162.
641. Nirenberg, L.: An extended interpolation inequality, Ann. Sc. Norm.
Super. Pisa, Sci. Fis. Mat. 20 (1966), 733737.
642. Nystr
om, K.: An estimate of a polynomial capacity, Potential Anal. 9
(1998), 217227.
643. OFarrell, A.: An example on Sobolev space approximation, Bull. Lond.
Math. Soc. 29 (1997), no. 4, 470474.
644. Oinarov, R.: On weighted norm inequalities with three weights, J. Lond.
Math. Soc. 48 (1993), 103116.
645. Oleinik, V. L., Pavlov, B. S.: On criteria for boundedness and complete
continuity of certain imbedding operators, Probl. Math. Phys. 4 (1970),
646. Opic, B., Kufner, A.: Hardy-type Inequalities. Pitman Research Notes
in Math. 129. Longman, Harlow, 1990.
647. Osserman, R.: A note on Haymans theorem on the bass note of a drum,
Comment. Math. Helv. 52 (1977), 545555.
648. Osserman, R.: The isoperimetric inequality, Bull. Am. Math. Soc. 84
(1978), 11821238.
649. Osserman, R.: Bonnesen-style isoperimetric inequalities, Am. Math.
Mon. 86 (1979), 129.
650. Ostrovskii, M. I.: Sobolev spaces on graphs, Quaest. Math. 28 (2005),
no. 4, 501523.
651. Oswald, P.: On the boundedness of the mapping f |f | in Besov spaces,
Comment. Math. Univ. Carol. 33 (1992), no. 1, 5766.
652. Otelbaev, M.: Two-sided estimates for diameters and applications, Dokl.
Sov. Math. Dokl. 17 (1976), 16551659.
653. Otelbaev, M.: Imbedding theorems for weighted spaces and their applications in the study of the spectrum of the Schr
odinger operator, Tr.
Mat. Inst. Steklova 150 (1979), 265305 (in Russian). English translation: Proc. Steklov Inst. Math. 150 (1981), 281321.
References
839
654. Otsuki, T.: A remark on the Sobolev inequality for Riemannian submanifolds, Proc. Jpn. Acad. 51 (1975), 785789.
655. Otto, F., Rezniko, M. G.: A new criterion for the logarithmic Sobolev
inequality and two applications, J. Funct. Anal. 243 (2007), 121157.
656. Payne, L. E., Weinberger, H. F.: An optimal Poincare inequality for
convex domains, Arch. Ration. Mech. Anal. 5 (1960), 286292.
657. Peetre, J.: New Thoughts on Besov Spaces, Duke Univ. Math. Series I.
Duke University Press, Durham, 1976.
658. Phuc, N. C., Verbitsky, I. E.: Quasilinear and Hessian equations of Lane
Emden type, Ann. Math., 168 (2008), 859914.
659. Pick, L.: Optimality and interpolation (English summary), Interpolation theory and Applications, Comtemp. Math. 445, Am. Math. Soc.,
Providence, 2007, 253264.
660. Pick, L.: Optimality of functions spaces in Sobolev embeddings, Sobolev
Spaces in Mathematics I: Sobolev Type Inequalities, V. Mazya (Ed.),
661. Pinchover, Y., Tintarev, K.: On the HardySobolevMazya inequality
and its generalizations, Sobolev Spaces in Mathematics I, International
Mathematical Series 8, Springer, Berlin, 2009, 281297.
662. Pohozhaev, S. I.: On the imbedding Sobolev theorem for pl = n, Doklady Conference, Section Math. Moscow Power Inst. (1965), 158170 (in
Russian).
663. Poincare, H.: Sur les equations de la physique mathematique, Rend.
Circ. Mat. Palermo 8 (1894), 57156.
664. Poincare, H.: La methode de Neumann et le problème de Dirichlet, Acta
Math. 20 (1897), 59142.
665. Polking, J. C.: Approximation in Lp by solutions of elliptic partial differential equations, Am. J. Math. 94 (1972), 12311244.
666. P
olya, G., Szeg
o, G.: Isoperimetric Inequalities in Mathematical Physics,
Ann. Math. Stud. 27. Princeton University Press, Princeton, 1951.
667. Prokhorov, D. V., Stepanov, V. D.: Inequalities with measures of
Sobolev embedding theorems type on open sets of the real axis, Sib.
Mat. Zh. 43 (2002), no. 4, 864878. English translation: Sib. Math. J.
43 (2002), no. 4, 694707.
668. Rado, T.: The isoperimetric inequality on the sphere, Am. J. Math. 57
(1935), no. 4, 765770.
669. Rafeiro, H., Samko, S.: Hardy type inequality in variable Lebesgue
spaces, Ann. Acal. Sci. Fenn. Math. 34 (2009), 279289. Corrigendum
in Ann. Acad. Sci. Fenn. Math. 35, 2010.
670. Rao, M.: Capacitary inequalities for energy, Isr. J. Math. 61 (1988),
no. 1, 179191.
671. Rao, M. M., Ren, Z. D.: Theory of Orlicz Spaces, Pure and Applied
Mathematics 146. Marcel Dekker, New York, 1991.
672. Rellich, F.: Ein Satz u
ber mittlere Konvergenz, Math. Nachr. 31 (1930),
3035.
840
References
673. Rellich, F.: Storungstheorie der Spektralzerlegung IV, Math. Ann. 113
(1937), 600619; 113 (1937), 677685; 116 (1939), 555570; 117 (1940),
356382; 118 (1942), 462484.
674. Reshetnyak, Ju. G.: On the concept of capacity in the theory of functions
with generalized derivatives, Sib. Mat. Zh. 10 (1969), 11091138 (in
Russian). English translation: Sib. Math. J. 10 (1969), 818842.
675. Reshetnyak, Yu. G.: Some integral representations of dierentiable functions, Sib. Mat. Zh. 12 (1971), 420432 (in Russian). English translation:
Sib. Math. J. 12 (1971), 299307.
676. Reshetnyak, Yu. G.: Integral representations of dierentiable functions
in domains with nonsmooth boundary, Sib. Mat. Zh. 21 (1980), no. 6,
108116 (in Russian). English translation: Sib. Math. J. 21, 833839
(1981).
677. Reshetnyak, Yu. G.: Spatial Transformations with Bounded Distortion,
Nauka, Novosibirsk, 1982 (in Russian).
678. Rickman, S.: Characterizations of quasiconformal arcs, Ann. Acad. Sci.
Fenn. Ser. AI-395 (1966), 730.
679. Rogers, C. A.: Packing and Covering, Cambridge University Press, Cambridge, 1964.
680. Rogers, L. G.: Degree independent Sobolev extension on locally uniform
domains, Preprint, Cornell University, Ithaca, 2005, 151.
681. Romanov, A. S.: On extension of functions of Sobolev spaces, Sib. Mat.
Zh. 34 (1993), no. 4, 149152.
682. Rosen, G.: Minimum value for C in the Sobolev inequality 3
Cgrad 3 , SIAM J. Appl. Math. 21 (1971), 3033.
683. Rozenblum, G. V.: On the estimates of the spectrum of the Schrodinger
operator, Probl. Math. Anal. Leningr. 5 (1975), 152165 (in Russian).
684. Ross, M.: A RellichKondrachov theorem for spiky domains, Indiana
Univ. Math. J. 47 (1998), no. 44, 14971509.
685. Runst, T., Sickel, W.: Sobolev Spaces of Fractional Order, Nemytskij
Operators, and Nonlinear Partial Dierential Equations, de Gruyter,
Berlin, 1996.
686. Rychkov, V.: On restrictions and extensions of the Besov and Triebel
Lizorkin spaces with respect to Lipschitz domains, J. Lond. Math. Soc.
(2) 60 (1999), no. 1, 237257.
687. Salo-Coste, L.: Aspects of Sobolev-Type Inequalities, LMS Lecture
Note Series 289. Cambridge University Press, Cambridge, 2002.
688. Samko, S.: Density of C0 (Rn ) in the generalized Sobolev spaces
W m,p(x) (Rn ), Dokl. Math. 60 (1999), 382385.
689. Samko, S.: Best Constant in the weighted Hardy inequality: the spatial
and spherical version, Fract. Calc. Appl. Anal. 8 (2005), no. 1, 3952.
690. Samko, S.: On a progress in the theory of Lebesgue spaces with variable
exponent: maximal and singular operators, Integral Transforms Spec.
Funct. 16 (2005), nos. 56, 461482.
References
841
691. Sawyer, E. T.: Two weight norm inequalities for certain maximal and
integral operators, Trans. Am. Math. Soc. 308 (1988), 533545.
692. Schechter, M.: Potential estimates in Orlicz spaces, Pac. J. Math. 133
(1998), no. 2, 381399.
693. Schmidt, E.: Uber

das isoperimetrische Problem im Raum von n Dimension, Math. Z. 44 (1939), 689788.
694. Schoen, R.: Conformal deformation of a Riemannian metric to constant
scalar curvature, J. Dier. Geom. 20 (1984), 479495.
695. Schwartz, L.: Theorie des Distributions, Hermann, Paris, 1973.
696. Shanmugalingam, N.: Newtonian spaces: an extension of Sobolev spaces
to metric spaces, Rev. Mat. Iberoam. 16 (2000), no. 2, 243279.
697. Shaposhnikova, T. O.: Equivalent norms in spaces of functions with
fractional or functional smoothness, Sib. Mat. Zh. 21 (1980), 184196.
(in Russian). English translation: Sib. Math. J. 21 (1980), 450460.
698. Shvartsman, P. A.: Extension theorems preserving locally polynomial
approximations, Preprint, Deposited at VINITI 6457-86, Yaroslavl University, 1986 (in Russian).
699. Shvartsman, P. A.: Sobolev Wp1 -spaces on closed subsets of Rn , Adv.
Math. 220 (2009), 18421922.
700. Shvartsman, P. A.: On Sobolev extension domains in Rn , J. Funct. Anal.
258 (2010), 22052245.
701. Simon, B.: Maximal and minimal Schr
odinger forms, J. Oper. Theory 1
(1979), no. 1, 3747.
702. Sinnamon, G., Stepanov, V. D.: The weighted Hardy inequality: new
proofs and the case p = 1, J. Lond. Math. Soc. (2) 54 (1996), no. 1,
89101.
703. Sj
odin, T.: Capacities of compact sets in linear subspaces of Rn , Pac. J.
Math. 78 (1978), 261266.
704. Slobodecki, L. N.: Generalized S. L. Sobolev spaces and their application to boundary value problems for partial dierential equations, Uch.
Zap. Leningr. Pedagog. Inst. Gercena 197 (1958), 54112 (in Russian).
705. Smirnov, V. I.: Course in Higher Mathematics, Vol. 5, Nauka, Moscow,
1959 (in Russian). German edition: Lehrgang der hoheren Mathematik,
Deutscher Verlag der Wissenschaften, Berlin, 1960.
706. Smith, K. T.: Inequalities for formally positive integro-dierential forms,
Bull. Am. Math. Soc. 67 (1961), 368370.
707. Smith, W., Stanoyevitch, A., Stegenga, D. A.: Smooth approximation of
Sobolev functions on planar domains, J. Lond. Math. Soc. (2) 49 (1994),
309330.
708. Smith, W., Stegenga, D.: Holder domains and Poincare domains, Trans.
Am. Math. Soc. 319 (1990), 67100 (in Russian).
709. Sobolev, S. L.: A general theory of diraction of waves on Riemannian
surfaces, Tr. Fiz.-Mat. Inst. Steklova 9 (1935), 39105 (in Russian).
710. Sobolev, S. L.: Le probleme de Cauchy dans lespace des fonctionelles,
Dokl. Akad. Nauk SSSR 3(8) (1935), no. 7(67), 291294.
842
References
711. Sobolev, S. L.: On some estimates relating to families of functions having

derivatives that are square integrable, Dokl. Akad. Nauk SSSR 1 (1936),
712. Sobolev, S. L.: On theorem in functional analysis, Sb. Math. 4 (1938),
471497 (in Russian). English translation: Am. Math. Soc. Trans. 34
(1963), no.. 2, 3968.
713. Sobolev, S. L.: Applications of Functional Analysis in Mathematical
Physics, Izd. LGU im. A. A. Zdanova,

Leningrad, 1950 (in Russian).
English translation: Am. Math. Soc. Trans. 7 (1963).
714. Sobolev, S. L.: Introduction to the Theory of Cubature Formulas,
Nauka, Moscow, 1974 (in Russian).
715. Sobolevskii, P. E.: Hardys inequality for the Stokes problem, Nonlinear
Anal. 30 (1997), no. 1, 129145.
716. Solonnikov, V. A.: On certain properties of Wpl spaces of fractional order, Dokl. Akad. Nauk SSSR 134 (1960), 282285 (in Russian). English
translation: Sov. Math. Dokl. 1 (1960), 10711074.
717. Solonnikov, V. A.: Inequalities for functions of the classes Wpm (Rn ), Zap.
Nauchn. Semin. Leningr. Otd. Mat. Inst. Steklova 27 (1972), 194210
(in Russian). English translation: J. Sov. Math. 3 (1975), 549564.
718. Soucek, J.: Spaces of functions on the domain whose k-th derivatives
Cas.
are measures dened on ,

Pest. Mat. 97 (1972), 1046.
719. Stampacchia, G.: Problemi al contorno ellittici, con dati discontinui,
dotati di soluzioni holderaine, Ann. Mat. Pura Appl. IV. Ser. 51 (1958),
138.
720. Stanoyevitch, A.: Geometry of Poincare domains, Doctoral Dissertation,
Univ. of Michigan, 1990.
721. Stanoyevitch, A.: Products of Poincare Domains, Proc. Am. Math. Soc.
117 (1993), no. 1, 7987.
722. Stanoyevitch, A., Stegenga, D. A.: The geometry of Poincare disks,
Complex Var. Theory Appl. 24 (1994), 249266.
723. Stanoyevitch, A., Stegenga, D. A.: Equivalence of analytic and Sobolev
Poincare inequalities for planar domains, Pac. J. Math. 178 (1997), no.
2, 363375.
724. Stein, E. M.: Singular Integrals and Dierentiability Properties of Functions, Princeton University Press, Princeton, 1970.
725. Stepanov, V. D.: Weighted inequalities of Hardy type for higher-order
derivatives and their applications, Dokl. Akad. Nauk SSSR 302 (1988),
no. 5, 10591062 (in Russian). English translation: Sov. Math. Dokl.
38(2) (1989), 389393.
726. Stepanov, V. D.: Weighted inequalities of Hardy type for Riemann
Liouville fractional integrals, Sib. Mat. Zh. 31 (1990), no. 3, 186197,
219 (in Russian). English translation: Sib. Math. J. 31 (1990), no. 3,
513522 (1991).
727. Stepanov, V. D., Ushakova, E. P.: On integral operators with variable
limits of integration, Tr. Math. Inst. Steklov 232 (2001), 298317.
References
843
728. Strichartz, R. S.: Multipliers on fractional Sobolev spaces, J. Math.

Mech. 16 (1967), 10311060.
729. Stredulinsky, E. W.: Weighted Inequalities and Degenerate Elliptic Partial Dierential Equations, Lect. Notes Math. 1074. Springer, Berlin,
1984.
730. Stroock, D., Zegarlinski, B.: The logarithmic Sobolev inequality for continuous spin systems on a lattice, J. Funct. Anal. 104 (1985), 299313.
731. Stummel, F.: Singul
are elliptische Dierentialoperatoren in Hilbertschen
Raumen, Math. Ann. 132 (1956), 150178.
732. Sugawa, T.: Uniformly perfect sets: analytic and geometric aspects,
S
ugaku Expo. 16 (2003), no. 2, 225242. Translation of S
ugaku 53
(2001), no. 4, 387402.
733. Sullivan, J. M.: Sphere packing give an explicit bound for the Besikovitch
covering theorem, J. Geom. Anal. 4 (1994), no. 2, 219231.
734. Swanson, D.: Pointwise inequalities and approximation in fractional
Sobolev spaces, Stud. Math. 149 (2002), 147174.
735. Swanson, D.: Area, coarea, and approximation in W 1,1 , Ark. Mat. 45
(2007), no. 2, 381399.
736. Swanson, D., Ziemer, W. P.: Sobolev functions whose inner trace at the
boundary is zero, Ark. Mat. 37 (1999), 373380.
737. Taibleson, M. H.: Lipschitz classes of functions and distributions in En ,
Bull. Am. Math. Soc. 69 (1963), 487493.
738. Taibleson, M. H.: On the theory of Lipschitz spaces of distributions on
Euclidean n-space. I Principal properties, J. Math. Mech. 13 (1964),
407479.
739. Takeda, M.: Lp -independence of the spectral radius of symmetric
Markov semigroups, Can. Math. Soc. Conf. Proc. 29 (2000), 613623.
740. Talenti, G.: Best constant in Sobolev inequality, Ann. Mat. Pura Appl.
IV. Ser. 110 (1976), 353372.
741. Talenti, G.: Elliptic equations and rearrangements, Ann. Sc. Norm. Super. Pisa, Cl. Sci. 3 (1976), no. 4, 697718.
742. Talenti, G.: An inequality between u and |grad u| (English summary).
General Inequalities, 6 (Oberwolfach, 1990), Int. Ser. Numer. Math. 103,
Birkhauser, Basel, 1992, 175182.
743. Talenti, G.: Inequalities in rearrangement-invariant function spaces,
Nonlinear Analysis, Function Spaces and Applications 5, Prometheus,
Prague, 1995, 177230.
744. Tartar, L.: An Introduction to Sobolev Spaces and Interpolation Spaces,
Lecture Notes of the Unione Matematica Italiana 3. Springer/UMI,
Berlin/Bologna, 2007.
745. Taylor, J. E.: Crystalline variational problems, Bull. Am. Math. Soc. 84
(1978), no. 4, 568588.
746. Taylor, M. E.: Estimate on the fundamental frequency of a drum, Duke
Math. J. 46 (1979), 447453.
844
References
747. Taylor, M. E.: Partial Dierntial Equations III: Nonlinear Equations,

748. Taylor, M. E.: Scattering length and the spectrum of + V , Can.
Math. Bull. 49 (2006), no. 1, 144151.
749. Tertikas, A., Tintarev, K.: On existence of minimizers for the Hardy
SobolevMazya inequality, Ann. Mat. Pura Appl. 186 (2007), no. 4,
645662.
750. Tertikas, A., Zographopoulos, N. B.: Best constants in the Hardy
Rellich inequalities and related improvements, Adv. Math. 209 (2007),
no. 2, 407459.
751. Tidblom, J.: A geometrical version of Hardys inequality for W01,p , Proc.
Am. Math. Soc. 138 (2004), no. 8, 22652271.
752. Tidblom, J.: A Hardy inequality in the half-space, J. Funct. Anal. 221
(2005), 482495.
753. Tintarev, K., Fieseler, K.-H.: Concentration Compactness. FunctionalAnalytic Grounds and Applications, Imperial College Press, London,
2007.
754. Tonelli, L.: Lestremo assoluto degli integrali doppi, Ann. Sc. Norm.
Super. Pisa 2 (1933), 89130.
755. Triebel, H.: Spaces of BesovHardySobolev type, Teubner, Leipzig,
1978.
756. Triebel, H.: Interpolation Theory, Function Spaces, Dierential Operators, VEB Deutscher Verlag der Wissenschaften, Berlin, 1978.
757. Triebel, H.: Theory of Function Spaces, Akademishe Verlagsgesellschaft,
Leipzig, 1983.
758. Triebel, H.: Theory of Functions Spaces II, Monographs in Mathematics
84, Birkh
auser, Basel, 1992.
759. Triebel, H.: The Structure of Functions, Monographs in Mathematics
97, Birkh
auser, Basel, 2001.
760. Triebel, H.: Theory of Functions Spaces III, Monographs in Mathematics
100, Birkhauser, Basel, 2006.
761. Troyanov, M., Vodopyanov, S.: Liouville type theorems for mappings
with bounded (co)-distortion, Ann. Inst. Fourier (Grenoble) 52 (2002),
no. 6 17531784.
762. Trudinger, N. S.: On imbeddings into Orlicz spaces and some applications, J. Math. Mech. 17 (1967), 473483.
763. Trudinger, N. S.: Remarks concerning the conformal deformation of Riemannian structures on compact manifolds, Ann. Sc. Norm. Super. Pisa
22 (1968), 265274.
764. Trudinger, N. S.: An imbedding theorem for H 0 (G, ) spaces, Stud.
Math. 50 (1974), 1730.
765. Trushin, B. V.: Sobolev embedding theorems for a class of anisotropic
irregular domains, Proc. Steklov Inst. Math. 260 (2008), 287.
766. Trushin, B. V.: Sobolev embedding theorems for a class of irregular
anisotropic domains, Dokl. Math. 77 (2008), no. 1, 6467.
References
845
767. Ukhlov, A. D.: Mappings that generate embeddings of Sobolev spaces,

Sib. Math. J. 34 (1993), no. 1, 165171.
768. Ukhlov, A. D.: Dierential and geometrical properties of Sobolev mappings, Math. Notes 75 (2004), no. 2, 291294.
769. Uraltseva, N. N.: On the non-selfadjointness in L2 (Rn ) of an elliptic operator with rapidly growing coecients, Zap. Nauchn. Semin. Leningr.
Otd. Mat. Inst. Steklova 14 (1969), 288294 (in Russian).
770. Uspenski, S. V.: Imbedding theorems for classes with weights, Tr. Mat.
Inst. Steklova 60 (1961), 282303 (in Russian). English translation: Am.
Math. Soc. Transl. 87 (1970), 121145.
771. V
aisal
a, J.: Removable sets for quasiconformal mappings, J. Math.
Mech. 19 (1969), no. 1, 4951.
772. Varopoulos, N., Salo-Coste, L., Coulhon, T.: Analysis and Geometry
on Groups, Cambridge University Press, Cambridge, 1993.
773. V
azquez, J. L., Zuazua, E.: The Hardy inequality and the asymptotic behaviour of the heat equation with an inverse-square potential, J. Funct.
Anal. 173 (2000), 103153.
774. Verbitsky, I. E.: Superlinear equations, potential theory and weighted
norm inequalities, Proceedings of the Spring School VI, Prague, May
31June 6, 1998, 233269.
775. Verbitsky, I. E.: Nonlinear potentials and trace inequalities, The Mazya
Anniversary Collection, vol. 2, Operator Theory, Advances and Applications 110, Birkhauser, Basel, 1999, 323343.
776. Verbitsky, I. E., Wheeden, R. L.: Weighted norm inequalities for integral
operators, Trans. Am. Math. Soc. 350 (1998), 33713391.
777. Vodopyanov, S. K.: Topological and geometrical properties of mappings
with summable Jacobian in Sobolev classes I, Sib. Math. J. 41 (2000),
no. 1, 1939.
778. Vodopyanov, S. K., Goldshtein, V. M.: Structure isomorphisms of
spaces Wn1 and quasiconformal mappings, Sib. Math. J. 16 (1975), 224
246.
779. Vodopyanov, S. K., Goldshtein, V. M., Latfullin, T. G.: Criteria for
extension of functions of the class L12 from unbounded plane domains,
Sib. Mat. Zh. 20 (1979), 416419 (in Russian). English translation: Sib.
Math. J. 20 (1979), 298301.
780. Volberg, A. L., Konyagin, S. V.: On measures with doubling condition,
Math. USSR, Izv. 30 (1988), 629638.
781. Volpert, A. I.: The spaces BV and quasi-linear equations, Mat. Sb. 73
(1967), 255302 (in Russian). English translation: Math. USSR Sb. 2
(1967), 225267.
782. Volevic, L. R., Paneyakh, B. P.: Certain spaces of generalized functions
and embedding theorems, Usp. Mat. Nauk 20 (1965), 374 (in Russian).
English translation: Russ. Math. Surv. 20 (1965), 173.
783. Vondracek, Z.: An estimate for the L2 -norm of a quasi continuous function with respect to smooth measure, Arch. Math. 67 (1996), 408414.
846
References
784. Wannebo, A.: Hardy inequalities, Proc. Am. Math. Soc. 109 (1990), no.
1, 8595.
785. Wannebo, A.: Equivalent norms for the Sobolev spaces W0m,p (), Ark.
Mat. 32 (1994), 245254.
786. Wannebo, A.: Hardy inequalities and imbeddings in domains generalizing C 0 domains, Proc. Am. Math. Soc. 122 (1994), no. 4, 11811190.
787. Wang, F.-Y.: Logarithmic Sobolev inequalities on noncompact Riemannian manifolds, Probab. Theory Relat. Fields 109 (1997), 417424.
788. Wang, F.-Y.: On estimation of logarithmic Sobolev constant and gradient estimates of heat semigroups, Probab. Theory Relat. Fields 108
(1997), 87101.
789. Wang, F.-Y.: Logarithmic Sobolev inequalities conditions and counterexamples, J. Oper. Theory 46 (2001), 183197.
790. Wang, F.-Y.: A generalization of Poincare and log-Sobolev inequalities,
Potential Anal. 22 (2005), 115.
791. Wang, Z.-Q., Meijun, Z.: Hardy inequalities with boundary terms, Electron. J. Dier. Equ. 43 (2003), 18.
792. Weidl, T.: A remark on Hardy type inequalities for critical Schrodinger
operators with magnetic elds, Operator Theory, Advances and Applications 110, Birkhauser, Basel, 1999, 345352.
793. Weissler, F. B.: Logarithmic Sobolev inequalities for the heat-diusion
semi-group, Trans. Am. Math. Soc. 237 (1978), 255269.
794. Werner, E., Ye, D.: New Lp ane isoperimetric inequalities, Adv. Math.
218 (2008), 712780.
795. Whitney, H.: A function not constant on a connected set of critical
points, Duke Math. J. 1 (1935), 514517.
796. Wik, I.: Symmetric rearrangement of functions and sets in Rn , Report
no. 1, Department of Mathematics, University of Ume
a (1977), 136.
797. Wu, Z.: Strong type estimates and Carleson measures for Lipschitz
spaces, Proc. Am. Math. Soc. 127 (1999), no. 11, 32433249.
798. Wul, G.: Zur Frage Geschwindigkeit des Wachsthums und der
Au
osung der Krystallachen, Z. Krystallogr. Mineral. 34 (1901), 449
530.
799. Xiao, J.: A sharp Sobolev trace inequality for the fractional-order derivatives, Bull. Sci. Math. 130 (2006), no. 1, 8796.
800. Xiao, J.: Homogeneous endpoint Besov space embeddings by Hausdor
capacity and heat equation, Adv. Math. 207 (2006), no. 2, 828846.
801. Yafaev, D.: Sharp constants in the Hardy-Rellich inequalities, J. Funct.
Anal. 168 (1999), 121144.
802. Yamabe, H.: On a deformation of Riemmanian structures on compact
manifolds, Osaka Math. J. 12 (1960), 2137.
803. Yang, S.: A Sobolev extension domain that is not uniform, Manuscr.
Math. 120 (2006), no. 2, 241251.
804. Yaotian, S., Zhihui, C.: General Hardy inequalities with optimal constants and remainder terms, J. Inequal. Appl. 3 (2005), 207219.
References
847
805. Yau, S. T.: Isoperimetric constants and the rst eigenvalue of a compact
manifold, Ann. Sci. Ecole Norm. Super. 8 (1975), 487507.
806. Yerzakova, N. A.: The measure of noncompactness of Sobolev embeddings, Integral Equ. Oper. Theory 19 (1994), no. 3, 349359.
807. Yerzakova, N. A.: On measures of non-compactness and applications to
embeddings, Nonlinear Anal. 30 (1997), no. 1, 535540.
808. Yosida, N.: Application of log-Sobolev inequality to the stochastic dynamics of unbounded spin systems on the lattice, J. Funct. Anal. 173
(2000), 74102.
809. Yudovich, V. I.: Some estimates connected with integral operators and
with solutions of elliptic equations, Dokl. Akad. Nauk SSSR 138 (1961),
805808. English translation: Sov. Math. Dokl. 2 (1961), 746749.
810. Zegarlinski, B.: On log-Sobolev inequalities for innite lattice systems,
Lett. Math. Phys. 20 (1990), no. 3, 173182.
811. Zhang, G.: The ane Sobolev inequality, J. Dier. Geom. 53 (1999),
183202.
812. Ziemer, W. P.: Weakly Dierentiable Functions, Springer, Berlin, 1969.
813. Ziemer, W. P.: Weakly Dierentiable Functions. Sobolev Spaces and
Functions of Bounded Variation, Graduate Texts in Mathematics, 10.
Springer, New York, 1989.
List of Symbols
Function Spaces
D() = C (),
C (), C (),
0
k
k
C (), C0 (), C k, (),
D (), D = D(Rn ),
C k, (),
Lp (), Lp (, loc) . . . . . . . . . . . 1.1.1
Llp () . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.2
Wpl (), Vpl () . . . . . . . . . . . . . . . . . 1.1.4
L lp (), Pk . . . . . . . . . . . . . . . . . . . . 1.1.13
l () . . . . . . . . . . . . . . . . . 1.1.14
Llp (), V
p
l
() . . . . . . . . . . . . . . . . . . . . . . . . 1.1.17
W
p
(X, B, ) . . . . . . . . . . . . . . . . . . . . . . . 1.2.3
Lq (Rn , ) = Lq (), Lq (, ) . . . . 1.4.1
N(e, ), P(e, ) . . . . . . . . . . . . . . . 2.2.1
X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1
1
() . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1
Wp,r
1
Wp,r
(, ) . . . . . . . . . . . . . . . . . . . 5.6.1
U (K), V (K), T (K). . . . . . . . .6.1.1
BV (). . . . . . . . . . . . . . . . . . . . . . . . .9.1.1
wpl , Wpl , blp , Bpl . . . . . . . . . . . . . . . . 10.1.1
hlp , Hpl . . . . . . . . . . . . . . . . . . . . . . . . 10.1.2
s (Rn ) . . . . . . . . . . . . . . . . . . . . . . . 10.2.1
W
p
s (Rn ) . . . . . . . . . . . . . . . . . . . . . . . 10.2.1
W
p
Spl . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4.1
M (S1 S2 ), M (S) . . . . . . . . . . . . 11.12
Wps . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.4.1
M(e, ) . . . . . . . . . . . . . . . . . . . . . . . 13.1.1
Wpl
. . . . . . . . . . . . . . . . . . . . . . . . . . 13.1.4
C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14.3
Pk , Cr (E) . . . . . . . . . . . . . . . . . . . . . 14.3.1
C0 (e) . . . . . . . . . . . . . . . . . . . . . . . . . 14.3.4
Llp (, ) . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Lm,p (Rn ) . . . . . . . . . . . . . . . . . . . . . . . 17.1
Lm,p (Rn ) . . . . . . . . . . . . . . . . . . . . . . . 17.1

(0)
m,p
1
Lp () . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7.4 H () . . . . . . . . . . . . . . . . . . . . . . . . 17.1
m,p () . . . . . . . . . . . . . . . . . . . . . . . . 17.1
1 (), W
1 () . . . . . . . . . . . . . . . . 6.11.1 H
L
p
p
Subsets of Rn
, supp f , B(x,
) = B (x), B . 1.1.1 E . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2.1
Lt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 [ ] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2.4
d (x) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1
Qd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Et , Lt . . . . . . . . . . . . . . . . . . . .2.1.1; 5.1.1
E, E, clos E, i E, , Nt . . . 5.1.1 F (). . . . . . . . . . . . . . . . . . . . . . . . . . .17.3

DOI 10.1007/978-3-642-15564-2,
849
850
List of Symbols
Classes of Sets in Rn
C 0,1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.9
EVpl . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.17
J . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1
J . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.1
K, . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.6.1
Ip, . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1
Hp, . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.2
J , Ip, . . . . . . . . . . . . . . . . . . . . . . 6.7.1
J (0), Ip, (0) . . . . . . . . . . . . . . . . 6.7.4

Ip, , Hp, . . . . . . . . . . . . . . . . . . . . . 6.8.1
(n1)
(n1)
Ip, , J
. . . . . . . . . . . . . . 6.11.2
(n1)
Ip,
. . . . . . . . . . . . . . . . . . . . . . . 6.11.5
N (Qd ) . . . . . . . . . . . . . . . . . . . . . . . 14.1.1
Cl,p,q . . . . . . . . . . . . . . . . . . . . . . . . . .15.5.1
Set Functions
mn , dist(E, F ) . . . . . . . . . . . . . . . . . 1.1.1
Hs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.18
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1.4.1
s, (g) . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1
(p, )-cap(e, ), p-cap(e, ) . . . . 2.2.1
p,M (F, ), (F, ) . . . . . . . . . . . . 2.4.1
capp (K, , ) . . . . . . . . . . . . . . . . . . . . 3.5
cap+
g , G) . . . . . . . . . . . . . . . . . . . . . . 3.7
p,l (
() . . . . . . . . . . . . . . . . . . . . . . . . 4.2, 4.3
() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3
() . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3
()
AG , A() . . . . . . . . . . . . . . . . . . . . . . 5.2.3
cp (K) . . . . . . . . . . . . . . . . . . . . . . 6.1.1;8.7
(p,)
AG . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2
(p,)
BG . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.1
cp (K) . . . . . . . . . . . . . . . . . . . . . . . . . 6.11.1
cp,l (G\F ) . . . . . . . . . . . . . . . . . . . . . . 7.6.1
cp,l ( (y)\{y}). . . . . . . . . . . . . . . . .7.6.3
essEp,q . . . . . . . . . . . . . . . . . . . . . . . . . . 8
C(Ep,q ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
M1 (Ep,q ) . . . . . . . . . . . . . . . . . . . . . . . . . . 8
M2 (Ep,q ) . . . . . . . . . . . . . . . . . . . . . . . . . . 8
S(p, q, , ) . . . . . . . . . . . . . . . . . . . . . . 8.7
q, , ) . . . . . . . . . . . . . . . . . . . . . . 8.7
S(p,
T (q, , ), T(q, , ) . . . . . . . . . . . . . 8.7
P (E ), P (E ), PC . . . . . . . . . . . . . 9.1.1
(E ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.3
||, A . . . . . . . . . . . . . . . . . . . . . . . . 9.3.4
()
A . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.3
cap(e, Spl ), cap(E, Spl ),
cap(E, Spl ) . . . . . . . . . . . . . . . . . 10.4.1
E . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4.2
H(E, ) . . . . . . . . . . . . . . . . . . . . . . . 10.4.3
C(K) . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.7
Cap(e,
Llp ()) . . . . . . . . . . . . . . . . . 13.1.1
l
p,q (G) . . . . . . . . . . . . . . . . . . . . . . . 14.2.1
Dp,l (G, Qd ), Dp,l (G), D(G) . . . 14.2.2
Capk (e,
Llp (Q2d )) . . . . . . . . . . . . . . 14.3.4
Dp,l (, ) . . . . . . . . . . . . . . . . . . . . . . . 16.2
=
p,l,m , D = Dp,l (, ) . . . . . 16.3.1
C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.1
Bm,p . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.2
H1nm . . . . . . . . . . . . . . . . . . . . . . . . . . 17.2
D1,m (, ) . . . . . . . . . . . . . . . . . . . . . 17.3
V (Gd , ) . . . . . . . . . . . . . . . . . . . . . . . 18.3
Functionals
F () . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.16 Qmin , Qmax . . . . . . . . . . . . . . . . . . . . . 17.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.6 hV (u, u). . . . . . . . . . . . . . . . . . . . . . . . .18.1
Fp [f ] . . . . . . . . . . . . . . . . . . . . . . . . . 4.1;4.4
List of Symbols
851
Operators
D , l , = 1 . . . . . . . . . . . . . . . 1.1.1
M . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.3
M+ , M . . . . . . . . . . . . . . . . . . . . . . 1.3.6
Il . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1;10.1.2
E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.3
Sh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.1
M . . . . . . . . . . . . . . . . . . . . . . . . 3.8; 10.3.2
Ep,q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.6
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.1.1
Fu = u
, Jl , D{l} . . . . . . . . . . . . . . 10.1.2
Up,l , Vp,l , Wp,l , Sp,l . . . . . . 10.4.2
T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2.1
I2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Dp, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Jz . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.1.4
Gz . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.1.4
M . . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.2
Constants
c, c1 , c2 , . . . , vn . . . . . . . . . . . . . . . . 1.1.1 j . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.7
n . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.10 Cp,q,a,b . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8
p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.15 M . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4.2
c(a, b, , ) . . . . . . . . . . . . . . . . . . . . . 1.3.7 Dp,q . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.6
Functions
T (v; x), T (v; x) . . . . . . . . . . . . . . 1.3.6
Fl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.7
M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1
(x, ) . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1
N (x) . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1
C (
) . . . . . . . . . . . . . . . . . . . . . . . 2.1.4;4.2
p (t) . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.2
R(d (x)). . . . . . . . . . . . . . . . . . . . . . . . .3.4
Sp (t) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5
u , u+ . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1
A (M ) . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1
M , . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.4
Ap, (M ) . . . . . . . . . . . . . . . . . . . . . . . 6.3.1
M,p . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.5
Bp, (M ) . . . . . . . . . . . . . . . . . . . . . . . 6.4.2
Ap, (M ), A1, (M ) . . . . . . . . . . . . . 6.7.1

p . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.11.1
p , p . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1
u (x) . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.1
(S) . . . . . . . . . . . . . . . . . . . . . . . . . . 9.5.3
u
(x), u(x) . . . . . . . . . . . . . . . . . . . . . . 9.6.1
D (v; x). . . . . . . . . . . . . . . . . . . . . . . . . . 12
(t) . . . . . . . . . . . . . . . . . . . . . . . . . . 12.2.2
(t). . . . . . . . . . . . . . . . . . . . . . . . . . .12.2.2
d . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14.3.2
PF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .18.2
Other Symbols
Llp (Q2d )),
= (1 , . . . , n ), ||, !, a b . 1.1.1 Cap(C, ,
Y [] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.7
Llp (Q2d )) . . . . . . . . . . 14.3.2
Capk (C,
u
Qd , u p,l,Qd . . . . . . . . . . . . . . . . . 14.1.2
Subject Index
(p, )-capacity, 141

(p, l)-inner diameter, 675, 676
(p, l)-negligible set, 670, 739
(p, l)-polar set, 663
(p, l)-quasi everywhere, 540
(p, l)-rened function, 544
-John domains, 415
d-sets, 519
p-capacity of a ball, 148
p-conductivity, 336
Absolute continuity, 4
Admissible subset, 124, 288
Ahlfors theorem, 87
Approximation by polyhedra, 484
Approximation of functions, 9, 10,
107, 289
Approximation of sets with nite
perimeter, 463
Area minimizing function, 140
Area minimizing function , 299
Area minimizing function M , 298
ps (Rn )
Asymptotics of the norm in W
as s 0, 528
Besicovitch covering theorem, 32
Bessel potential, 518
Best constant in the Sobolev inequality, 160
BirnbaumOrlicz space, 157
Boundedness of embedding operators,
180
Boundedness of functions with derivatives in BirnbaumOrlicz spaces,

419
Bourgain, Brezis and Mironescu theorem, 511
Brezis and Mironescu Conjecture, 530
Cantor sets E with positive
cap(E, Hpl ), 543
Capacitary FaberKrahn inequality,
278
Capacitary inequality, 154, 231, 549
Capacitary inequality with the best
constant, 256, 274
Capacitary YudovichMoser type inequality, 256, 276
Capacity, 536, 657
Capacity cap(e, Spl ), 536
Capacity Cap(e,
Llp ()), 657
Capacity Capk (e,
Llp (Q2d )), 684
Capacity minimizing function, 156
Capacity of a cylinder, 662
Capacity of a parallelepiped, 542
Carnot group, 152
Cartan type theorem, 575
Case of small (p, l)-inner diameter,
687
Cheegers inequality, 258
Choquet capacity, 143
Class C0 (e), 684
Class Hp, , 354
Class Ip, , 343

DOI 10.1007/978-3-642-15564-2,
853
854
Class
Class
Class
Class
Class
Subject Index
K, , 314
(n1)
Ip, , 409
EVpl , 87
T (K), 338
V (K), 336
p,
Classes J and I , 376

(n1)
(n1)
, 405
Classes Ip, , J
Classes J , 290
Classes of sets, 28, 290, 300, 311, 343,
354, 376, 377, 386
Classical isoperimetric inequality, 280
Closability, 188, 746
Closability of embedding operators,
746, 761
Coarea formula, 40
Compactness of the embedding
Llp () Lq (), 714

L11 () Lq () (q 1), 311
L1p () Lq (), 386
Llp () C() L (), 429
Wp1 () C() L (), 422
Compactness of the embedding operator
Llp (, ) Wrm (), 742
Compactness theorem for an arbitrary domain with nite volume,
389
Compactness theorems, 76, 182, 387,
389, 714, 744
Complementary function, 157
Completeness of Wpl () and Vpl (),
22
Composition operator in fractional
Sobolev spaces, 643
Conditions for embedding into Lq ()
for p > q > 0, 570
Conductivity, 336
Conductor, 336
Conductor inequality, 231
Cone property, xxiii, 15
Continuity modulus of functions in
L1p (), 416
Continuity of the embedding operator
Llp (, ) Wrm (), 739
Contractivity condition, 255

Counterexample to the capacitary inequality for the norm in L22 (),
558
Coverings, 32
Criteria of solvability of to boundary
value problems for second order
elliptic equations, 411
Criterion of positivity of the
Schr
odinger operator HV , 787
D.R. Adams theorem, 64
Degenerate quadratic form, 205
Density of bounded functions in
L2p (), 112
Density of bounded functions in
Sobolev spaces, 107
Description of the traces of the functions in Wp1 (), 545
Dirichlet principle with prescribed
level surfaces, 340
Dirichlet problem, 716, 718
Dirichlet problem for a strongly elliptic operator, 716
Discreteness of the negative spectrum, 193
Discreteness of the negative spectrum
of the operator Sh , 196
Discreteness of the spectrum, 193,
196, 394, 749
Discreteness of the spectrum of the
Dirichlet problem, 719
Domain of the class C, 11
Domain of the class C 0,1 , 15
Domain starshaped with respect to a
point, 10
Domains in EVp1 which are not quasidisks, 91
Domains of the class C 0,1 , 15
Domains with power cusps, 362
Dual of a Sobolev space, 24
Embedding
Llp () D (), 694
Embedding
Llp () Lq () (the case
p q), 703
Embedding
Llp () Lq () (the case
p > q 1), 707
Embedding
Llp () Lq () for an
innite funnel, 712
Subject Index
Embedding
Llp () Lq (, loc), 701
Embedding Vpl () C() L (),
428
Embedding Wp1 () C()L (),
406
Embedding into a Riesz potential
space, 596
Embedding operators for the space
pk (), l > 2k, 432
Wpl () W
Embedding theorems for p = 1, 588
Embedding theorems for the space
Spl , 579
Embedding theorems of the Sobolev
type, 63
Equivalence of continuity and compactness of the embedding Hpl
Lq () for p > q, 583
Equivalence of two capacities, 664
Equivalent norms in Wpl (), 26
Essential norm of the embedding operator Ep,q : L1p Lq , 435
Estimate for the norm in Lq (Rn , )
by the integral of the modulus
of the l-th order gradient, 70
Estimate for the norm in Lq (Rn , )
by the integral of the modulus
of the gradient, 67
Estimates for capacities, 575
Estimates for nonnegative functions
of one variable, 59
Expression for the (p, )-capacity
containing an integral over level
surfaces, 144
extension domains for functions with
bounded variation, 98
Extension of functions, 26, 87
Extension of functions in BV (), 477
Extension of functions with zero
boundary data, 94
Extension of Sobolev functions from
cusp domains, 97
Extension operators, 477, 514
Extension operators acting on
Sobolev spaces, 32
FaberKrahn inequality, 280
Finiteness of the negative spectrum,
197
855
First DirichletLaplace eigenvalue,

261
Formula for the integral of modulus
of the gradient, 38
Fourier transform, 516
Friedrichs lemma, 193
Friedrichs inequality, xxiii
Friedrichs type inequality, 403
Frostmans lemma, 759
Function p (), 406
Function M,p , 349
Function p (), 407
Function || for a convex domain, 487
Functions on Riemannian manifolds,
257
Functions with bounded gradients,
108
Fundamental solution of the polyharmonic operator, 3
GagliardoNirenberg inequality, 69,
83, 86
GaussGreen formula, 467, 500, 505,
506
Generalized Dirichlet integral, 208
Generalized Poincare inequality, 20
Generalized Sobolev theorem, 73
Glazmans lemma, 193
Gustins covering theorem, 34
H
ormander and Lions theorem, 733
HahnBanach theorem, 25
Hardy inequalities of fractional order,
735
Hardys inequality, 40, 48
Hardys inequality with sharp
Sobolev remainder term, 213
HardyLittlewood maximal operator,
611
HardyLeray inequality, 220
HardyLittlewood theorem, 83
HardySobolev type inequalities, 138
Hardy-type inequalities with indenite weights, 51
Hausdor -measure, 542
Hausdor capacities, 759
Hausdor measure, 28, 38
Heisenberg group, 152
856
Subject Index
Inequality of McKean, 258

Inniteness and niteness of the negative spectrum, 199
Inner and outer capacities, 537, 657
Integral formula for the norm in
BV (), 465
Integral representation for functions
of several variables with zero incomplete Cauchy data, 100
Integral representations, 16, 99
Interior capacitary radius, 789
Interior segment property, 117
Isocapacitary inequality, 147
Isoperimetric inequality, 133, 464
Kerman and Sawyer criterion, 204
Koch snowake domain, 416
Kolmogorov diameter, 752
Lebesgue integral as a Riemann integral, 37
Liebs theorem on Riesz potentials, 83
Lipschitz domain, 16
LittlewoodPaley function, 518
Local isoconductivity constants, 456
Locality condition, 255
Localization moduli, 437
Logarithmic Sobolev inequalities, 187
Lorentz space, 234
M. Rieszs theorem, 77
Magnetic Schr
odinger operators, 800
Marcinkiewicz interpolation theorem,
64
Maximal algebra in Wpl (Rn ), 117
Maximal algebra in Wpl (), 117
Mazya and Verbitsky criterion, 204,
564
Mikhlin theorem on Fourier integral
multipliers, 516
Molchanovs criterion, 751, 771
Mollications, 4
Moser multiplicative inequality, 176
Muckenhoupt condition, 553
Multiplicative GagliardoNirenberg
inequality, 511, 530
Multiplicative inequality, 79, 124,
162, 341
Nash multiplicative inequality, 176
Negative spectrum of the multidimensional Schr

odinger operator, 188
Negligibility class N (Gd ; ), 771
Neumann problem, 392
Neumann problem for operators of arbitrary order, 394
Neumann problem for strongly elliptic operators, 392
Nikod
yms example, 308, 364
No embedding of Vpl () C k1,1 ()
into C k (), 75
Nonincreasing rearrangement, 127
Nonlinear potential theory, 536
Nonlinear potentials, 539
Normal in the sense of Federer, 467
One-dimensional Poincare inequality,
31
Operator Fp , 255
Perimeter, 459
Poincare inequality for domains with
innite volume, 383
Pointwise estimates involving Mk u
and l u, 622
Pointwise interpolation inequalities
involving fractional derivatives, 638
Pointwise interpolation inequality for
Bessel potentials, 620
Pointwise interpolation inequality for
Riesz potentials, 613
Positive deniteness and discreteness
of the spectrum of a strongly elliptic operator, 749
Positivity of the form S1 [u, u], 189
Problem of spectral synthesis in
Sobolev spaces, 691
Pseudonormed space, 289
Quasi-isometric map, 13
Quasidisk, 87
Reduced boundary, 467
Rened functions, 544
Relation of H(S) and H(S),

209
Relations of capacities, 538
Relative p-capacity, 234
Subject Index
Relative p-capacity depending on the
measure, 241
Relative isoperimetric inequality, 291
Relative perimeter, 459
Rellich lemma, xxiv
RellichKato theorem, 586
Removable sets for Sobolev functions,
28
Riesz and Bessel potential spaces, 516
Riesz potential, 64, 611
Right and left maximal functions, 62
Rough maximum principle for nonlinear potentials, 539
Rough trace, 489, 499
Schr
odinger operator on -cusp domains, 452
Self-adjointness of a dierential operator, 756
Semiboundedness of the Schr
odinger
operator, 190
Set function lp,q (G), 675
Set function (E ) for a convex domain, 486
Set of nite capacity, 208
Set of uniqueness in Wpl , 692
Sets of zero capacity Cap(, Wpl ), 663
Sharp pointwise inequalities for u,
624
Smooth level truncation, 246
Sobolev integral representation, 16
Sobolev mappings, 321
Sobolev spaces, 30
Sobolev type inequality, 160
Sobolevs inequality, 69
SobolevLorentz spaces, 253
Solvability of the Dirichlet problem
for quasilinear equations in unbounded domains, 721
Space
Llp (, ), 1, 737
Space Hpl , 517
Space L21 ()L () is not necessarily dense in L21 (), 109
1
(), 379
Space Wp,r
Space W22 () L () is not always
a Banach algebra, 120
pl () and
Spaces W
Llp (), 23
Spaces wpl , Wpl , blp , Bpl for l > 0, 512
Spaces Wpl () and Vpl (), 7
857
Starshaped domain, 10, 14

Starshaped with respect to a ball, 14
Stein extension theorem, 73
Steins extension operator, 32
Steklov problem, 212
Strichartz theorem, 518
Subareal mappings, 300
Symmetrization, 150, 291
Theorem on critical sets of a smooth
function, 35
Trace, 500
Trace theorem, 512
Transformation of coordinates in
norms of Sobolev spaces, 12
Two measure Sobolev-type inequality, 232
Two-sided estimates of the rst eigenvalue of the Dirichlet Laplacian,
789
Two weight Hardys inequalities, 214
Two-weight Hardys inequality, 42
Two-weight inequalities involving
fractional Sobolev norms, 249
Two-weighted Sobolev inequality
with sharp constant, 282
Uniqueness of a solution to the
Dirichlet problem with an exceptional set for equations of arbitrary order, 727
Uniqueness of a solution to the Neumann problem for quasilinear
second order equations, 730
Uniqueness theorem for analytic functions in the class L1p (U ), 689
Upper and lower traces, 500
Upper estimate of a dierence seminorm (the case p = 1), 590
Upper estimate of a dierence seminorm (the case p > 1), 597
Variable exponent, 30
Verbitsky criterion, 204
Vitali covering theorem, 35
858
Subject Index
Weighted area of g, 124

Weighted norm inequalities, 612
Weighted norm interpolation inequalities for potentials, 623
Wul shapes, 140

Yamabe problem, 260
Young functions, 157
Author Index
Acosta, 321, 803

Adams, D. R., xxviii, 1, 64, 83, 86,
252, 537542, 544, 550, 557, 578,
580, 607, 608, 655, 668, 691, 758,
768, 803, 804
Adams, R. A., xxviii, 187, 402, 691,
804
Adimurthi, 179, 220, 804
Ahlfors, 87, 692, 804
Aida, 187, 804
Aikawa, 252, 537, 804
Aissaoui, 548, 805
Akilov, 533, 805
Alekseev, 753, 805
Almgren, 320, 528, 805
Alvino, 220, 805
Ambrosio, 286, 805
Amghibech, 252, 805
Amick, 30, 401, 458, 805
Ancona, 734, 805
Andersson, 402, 805
Anzellotti, 507, 805
Arendt, 401, 805
Aronszajn, 515, 545, 614, 805
Attouch, xxviii, 805
Aubin, 177, 260, 319, 805, 806
Avkhadiev, 220, 806
Babich, 32, 515, 806
Bagby, 544, 692, 806
Bakry, 179, 187, 806
Baldi, 508, 806
Balinsky, 220, 806
Ba
nuelos, 801, 806
Barbatis, 220, 806
Barthe, 252, 806
Bastero, 85, 807
Beckner, 178, 179, 187, 807
Ben Amor, 252, 807
Benguria, 141, 220, 807
Benkirane, 548, 805
Berezin, 751, 807
Berger, 258, 807
Besicovitch, 39, 807
Besov, xxviii, 31, 402404, 511, 515,
545, 807, 808
Beurling, 515, 692, 804, 808
Bianchi, 178, 808
Biegert, 401, 808
Biezuner, 179, 808
Birman, 203, 752, 808
Birnbaum, 157, 808
Biroli, 151, 286, 808
Bj
orn, 286, 402, 734, 808
Bj
orup, 402, 808
Bliss, 161, 177, 274, 449, 808
Bobkov, 40, 63, 85, 177, 187, 252, 320,
808, 809
Bodineau, 187, 809
Bojarski, 321, 401, 653, 809
Bokowski, 320, 507, 809
Bonder, 179, 809
Bosi, 220, 809
Bourbaki, 26, 297, 809
Bourdaud, 655, 809, 810

DOI 10.1007/978-3-642-15564-2,
859
860
Author Index
Bourgain, 85, 521, 530, 810

Brandolini, 220, 810
Brezis, 177, 219, 232, 434, 521, 530,
535, 655, 810
Buckley, 97, 321, 401, 402, 434, 810
Burago, 140, 258, 259, 319, 320, 507,
508, 810, 811
Burenkov, xxviii, 30, 31, 98, 402, 545,
811
Busemann, 140, 811
Buser, 258, 811
Buttazzo, xxviii, 805
Caccioppoli, 507, 811
Caarelli, 141, 811
Calder
on, 32, 811
Calude, 734, 811
Campanato, 402, 811
Capogna, 261, 811
Carlen, 179, 812
Carleson, 580, 608, 812
Carlsson, 768, 812
Carroll, 801, 806
Cartan, 575, 812
Carton-Lebrun, 654, 812
Casado-Daz, 151, 812
Cascante, 551, 573, 812
Cattiaux, 252, 806, 812
Chang, 608, 812
Chaudhuri, 220, 804, 812
Chavel, 258, 261, 812
Cheeger, 286, 812
Chen, 187, 252, 812
Chiacchio, 220, 810
Choquet, 143, 144, 151, 813
Chou, 286, 813
Chu, 286, 813
Chua, 96, 253, 286, 402, 813
Chung, 258, 813
Cianchi, 85, 178, 186, 187, 220, 319
321, 528, 608, 813, 814
Cohen, 85, 814
Coifman, 553, 814
Colin, 220, 814
Cordero-Erausquin, 178, 814
Costea, 253, 814
Costin, 229, 814
Coulhon, 179, 261, 402, 806, 814, 845
Courant, xxiv, 31, 385, 814

Croke, 801, 814
D
avila, 220, 815
Dahlberg, 252, 607, 655, 814
Dal Maso, 151, 814
Daners, 321, 814
Danielli, 261, 811
Davies, 187, 219, 253, 402, 404, 811,
814, 815
De Giorgi, 507, 815
Del Pino, 179, 815
Delin, 179, 815
Demyanov, 179, 815
Deny, 29, 117, 319, 401, 733, 735, 815
DeVore, 85, 814
DiBenedetto, 32, 815
Diening, 31, 815
Dolbeault, 179, 220, 809, 815
Donnelly, 258, 815
Donoghue, 691, 815
Dou, 220, 815
Douglas, 515, 815
Druet, 179, 815
Dunford, 34, 35, 816
Dupaigne, 220, 815
Dur
an, 321, 803
Dynkin, 547, 816
Edmunds, 85, 220, 402, 458, 816
Egnell, 178, 402, 808, 816
Ehrling, 86, 816
Eidus, 768, 816
Eilertsen, 219, 816
Escobar, 178, 258, 260, 816
Esteban, 220, 809, 815
Evans, 220, 402, 458, 816, 817
Faber, 278, 801, 817
Faddeev, 40, 817
Fain, 96, 817
Federer, 28, 39, 40, 83, 177, 319, 467,
507, 817
Feerman, 553, 625, 814
Ferone, 220, 805, 813
Ferreira, 179, 809
Fieseler, 141, 179, 844
Filippas, 141, 220, 806, 817
Fitzsimmons, 252, 817
Author Index
Fleming, 83, 151, 177, 319, 507, 817
Fournier, 402, 804
Fradelizi, 321, 817, 818
Fraenkel, 30, 31, 402, 818
Franchi, 286, 818
Frank, 141, 220, 807, 818
Frazier, 655, 803
Freud, 515, 818
Friedrichs, xxiii, 193, 751, 818
Frostman, 759, 818
Fukushima, 252, 818
F
uredi, 33, 818
Fusco, 178, 813, 818
Gagliardo, xxiii, 30, 69, 83, 84, 86,
121, 176, 818
Galaktionov, 220, 819
Gallouet, 434, 810
Gauduchon, 258, 807
Gazzola, 220, 819
Gelfand, 2, 819
Gelman, xxviii, 545, 819
Gentil, 252, 812
Ghoussoub, 179, 819
Giaquinta, 507, 805
Gidas, 177, 179, 819
Gil-Medrano, 260, 819
Glazer, 286, 819
Glazman, 193, 203, 800, 819
Globenko, 402, 819
Glushko, 31, 819
Goldshtein, xxviii, 87, 88, 91, 152,
286, 321, 402, 819, 820, 845
Golovkin, 545, 820
G
otze, 63, 187, 809
Grigoryan, 252, 253, 258, 813, 820
Gromov, 84, 320, 820
Gross, 187, 820
Grosse, 286, 819
Grossi, 220, 804
Grunau, 220, 819
Grushin, 668, 820
Guedon, 321, 817, 818
Guillin, 252, 812
Guionnet, 187, 820
Gurov, 321, 819
Gustin, 39, 820
Guzman, 32, 820
861
Haberl, 178, 820

Hadwiger, 68, 151, 821
Hahn, 503, 821
Hajlasz, xxviii, 30, 252, 253, 286, 321,
402, 403, 653, 734, 809, 818, 821
Han, 179, 821
Hansson, 177, 232, 252, 607, 821
Hardy, 24, 40, 48, 62, 63, 83, 91, 96,
270, 528, 821
Harjulehto, 152, 402, 821, 822
Haroske, xxviii, 822
Harris, 402, 458, 816
Harvey, 760, 822
H
ast
o, 31, 152, 402, 815, 821, 822
Havin, 94, 187, 517, 537540, 543,
544, 548, 573, 666, 689, 690, 822,
834
Hayman, 151, 801
Hebey, xxviii, 179, 261, 815, 822
Hedberg, xxviii, 30, 252, 538541,
552, 578, 607, 622, 642, 653, 691,
692, 758, 803, 822, 823
Heinig, 654, 812
Heinonen, 151, 152, 286, 823
Heler, 187, 809
Hencl, 321, 823
Hestenes, 32, 823
Hilbert, xxiv, 31, 385, 814
Hinz, 219, 815
Homan, 259, 319, 823
Homann-Ostenhof, M., 220, 823
Homann-Ostenhof, T., 220, 823
Holley, 187, 823
Holopainen, 152, 321, 823
Horiuchi, 141, 823
H
ormander, 545, 668, 733, 823
Houdre, 40, 809
Hudson, 286, 608, 823
Hurd, 402, 823
Hurri, 402, 823
Hurri-Syrj
anen, 220, 402, 816, 824
Ilin, xxiii, xxviii, 31, 83, 86, 141, 286,
404, 511, 515, 545, 808, 824
Inglis, 187, 824
Iwaniec, 534, 824
Jawerth, 545, 824
Jerison, 402, 824
862
Author Index
Jodeit, 276, 286, 824

Johnsen, 545, 824
Jones, 88, 96, 97, 824
Jonsson, xxviii, 519, 545, 547, 824
Kac, 63, 824
Kaimanovich, 252, 825
Kalamajska, 654, 825
Kalton, 565, 825
Kang, 179, 819
Kannan, 320, 825
Kantorovich, 533, 805
Karadzhov, 548, 825
Kateb, D., 655, 810
Kateb, M., 655, 810
Kato, 586, 768, 825
Kauhanen, 321, 825
Kerman, 85, 204, 551, 558, 564, 813,
816, 825
Khvoles, 733, 834
Kilpel
ainen, 151, 286, 402, 403, 823,
825
Kinnunen, 152, 286, 734, 825
Klimov, 85, 187, 319, 825, 826
Knothe, 84, 826
Kohn, 141, 811
Kokilashvili, 31, 63, 826
Kolsrud, 30, 252, 826
Kolyada, 85, 548, 826
Kombe, 220, 827
Kondrashov, xxiii, 86, 827
Kondratiev, 691, 735, 800, 827
Konyagin, 547, 845
Korevaar, 286, 827
Korte, 286, 734, 825, 827
Koskela, 32, 97, 286, 321, 401403,
434, 508, 653, 810, 814, 818, 821,
823, 825, 827
Koskenoja, 152, 822
Kosovsky, 508, 810
Kov
acik, 31, 827
Krahn, 278, 801, 827
Kr
alik, 515, 818
Krasnoselski, 157, 421, 827
Krein, 63, 824
Krickeberg, 507, 828
Kronrod, 40, 828
Kudryavtsev, 515, 545, 808, 828
Kufner, 63, 253, 515, 606, 653, 828,

834, 838
Labutin, 321, 402, 403, 434, 655, 828
Ladyzhenskaya, 401, 828
Landis, 40, 735, 828
Landkof, 200, 210, 539, 575, 828
Lang, 457, 828
Laptev, A., xxviii, 219, 220, 734, 806,
823, 828
Laptev, S., 212, 829
Latfullin, 87, 845
Leckband, 286, 608, 823
Ledoux, 85, 177, 179, 187, 806, 809,
829
Leoni, xxviii, 829
Leray, 30, 721, 722, 724, 732, 735, 829
Lerner, 548, 826
Levi, 30, 829
Levin, 63, 829
Lewis, 30, 220, 734, 829
Lichtenstein, 32, 829
Lieb, 83, 179, 220, 286, 528, 800, 801,
805, 818, 829
Lin, 141, 829
Lions, J.-L., 29, 31, 117, 319, 383, 385,
401, 402, 440, 445, 668, 720722,
724, 732, 733, 735, 815, 829, 855
Lions, P.-L., 179, 320, 830
Littlewood, 24, 40, 62, 63, 83, 270,
518, 528, 821
Littman, 668, 830
Lizorkin, 206, 515, 691, 752, 808, 830
Loeb, 33, 818
Loss, 141, 179, 220, 807, 812, 815, 829
Lov
asz, 320, 825, 830
Lugiewicz, 187, 830
Lutwak, 178, 813, 830
Luukkainen, 547, 830
Lyusternik, 68, 464, 830
MacManus, 286, 734, 808
Magenes, 383, 385, 401, 440, 445, 720,
735, 830
Maggi, 178, 401, 813, 818, 830
Maheux, 402, 830
Maligranda, 63, 253, 606, 828
Mal
y, 30, 40, 321, 403, 821, 823, 825,
830
Author Index
Marcinkiewicz, 64, 518
Marcus, 219, 655, 810, 830
Martin, 85, 187, 286, 319, 402, 819,
830, 831
Martio, 151, 152, 286, 402, 823, 825,
831
Masuda, 187, 804
Matskewich, 219, 831
Mazya, xxviii, 1, 30, 31, 54, 63, 83,
86, 90, 94, 105, 107, 117, 121,
140, 141, 151, 177, 179, 186, 187,
203, 204, 209, 212, 219, 220, 229,
232, 252, 253, 286, 319321, 401,
402, 431, 434, 457, 507, 517, 537
540, 543545, 548, 551, 564, 573,
607609, 653, 655, 666, 668, 689
692, 733, 735, 752, 768, 800, 811,
812, 814, 817, 819, 821, 827, 828,
831835
Mazet, 258, 807
McKean, 258, 835
Meijun, 220, 846
Meyer, 85, 655, 810, 814
Meyers, 30, 537539, 541543, 692,
804, 835, 836
Michael, 259, 319, 836
Michaille, xxviii, 805
Mikhlin, 516, 836
Mikkonen, 734, 836
Milman, E., 152, 836
Milman, M., 85, 187, 319, 402, 548,
807, 825, 830, 831
Milman, V., 320, 820
Miranda, 319, 508, 827, 836
Mircea, 653, 836
Mironescu, 521, 530, 535, 655, 810
Mitidieri, 220, 819
Mizel, 219, 655, 830
Molchanov, 735, 751, 752, 836
Montefalcone, 508, 806
Monti, 140, 836
Morbidelli, 140, 836
Morrey, 30, 31, 86, 836
Morse, 33, 39, 836
Mosco, 286, 808
Moser, 176, 256, 277, 286, 608, 837
Mossino, 85, 837
Moussai, 655, 810
863
Muckenhoupt, 63, 238, 251, 253, 553,

837
Mulla, 545, 805
Mynbaev, 253, 837
Nadirashvili, 258, 837
Nash, 176, 837
Nasyrova, 253, 837
Natanson, 39, 153, 837
Naumann, 30, 837
Nazaret, 178, 814
Nazarov, A., 179, 286, 815, 837
Nazarov, F., 320, 809, 837
Necas, 401, 837
Nekvinda, 31, 815
Netrusov, 30, 117, 121, 141, 179, 252,
544, 607, 691, 823, 834, 837
Nevanlinna, 575, 837
Ni, 177, 179, 819
Nieminen, 151, 838
Nikod
ym, xxiv, 7, 30, 838
Nikolsky, xxviii, 31, 32, 404, 511, 515,
519, 808, 838
Nirenberg, xxiii, 69, 83, 86, 87, 121,
141, 177, 179, 811, 819, 838
Niu, 220, 815
Nystr
om, 692, 838
OFarrell, 117, 838
Oinarov, 253, 838
Oleinik, 752, 753, 805, 838
Onninen, 402, 827
Opic, 253, 838
Orlicz, 157, 808
Ortega, 551, 573, 812
Oru, 85, 814
Osserman, 507, 801, 838
Ostrovskii, 286, 838
Oswald, 655, 838
Otelbaev, 253, 735, 752, 753, 830, 834,
837, 838
Otsuki, 319, 838
Otto, 187, 839
Ozaydin,
220, 827
Pacella, 320, 402, 816, 830
Paley, 518
Panasenko, 85, 826
Paneyakh, 545, 845
Papageorgiou, 187, 824
864
Author Index
Pauls, 261, 811

Pavlov, 734, 752, 808, 811, 838
Payne, 320, 401, 839
Pearson, 187, 807
Peetre, 511, 519, 839
Persson, 63, 253, 606, 828
Petrushev, 85, 814
Phuc, 252, 839
Pick, 85, 187, 813, 814, 816, 825, 839
Pierre, 252, 804
Pinchover, 219, 220, 830, 839
Poborchi, xxviii, 90, 97, 117, 402, 431,
457, 545, 608, 834
Pohozhaev, 86, 256, 839
Poincare, xxiv, 839
Polking, 668, 691, 760, 804, 822, 839
P
olya, 24, 40, 62, 63, 83, 150, 151, 270,
401, 528, 821, 839
Pratelli, 178, 813, 818
Preobrazhenski, 83, 607, 834
Prokhorov, 253, 839
Pustylnik, 85, 831, 836
Qian, 187, 806
Rado, 278, 839
Rafeiro, 31, 735, 839
R
akosnik, 31, 827
Ramaswamy, 220, 804
Rao, 157, 252, 839
Rellich, 86, 586, 839
Ren, 157, 839
Reshetnyak, xxviii, 31, 32, 402, 537,
819, 840
Rezniko, 187, 839
Rickman, 87, 321, 823, 840
Rishel, 319, 507, 817
Roberto, 252, 806
Rogers, 96, 795, 840
Romanov, 96, 321, 819, 840
Rosen, 177, 840
Rosenthal, 503, 821
Rosin, 63
Ross, 402, 840
Rossi, 179, 809
Rozenblum, 203, 735, 840
Ruiz, 85, 807
Runst, xxviii, 655, 822, 840
Ruticki, 157, 421, 828

Rychkov, 545, 840
Saintier, 179, 809
Saksman, 547, 830
Salo-Coste, xxviii, 179, 261, 402,
806, 830, 840, 845
Samko, 31, 220, 735, 826, 839, 840
Santra, 220, 804
Sawyer, 204, 551, 558, 564, 825, 840
Schechter, 608, 841
Schmeisser, xxviii, 822
Schmidt, 68, 841
Schoen, 260, 286, 827, 841
Schuster, 178, 820
Schwartz, J., 34, 35, 816
Schwartz, L., 2, 699, 841
Seiringer, 220, 818
Serrin, 30, 836
Shafrir, 219, 810
Shanmugalingam, 286, 402, 508, 734,
808, 827, 841
Shaposhnikova, xxviii, 31, 63, 219,
286, 321, 518, 548, 609, 653, 655,
834, 835, 841
Shigekawa, 187, 804
Shilov, 2, 819
Shubin, 735, 751, 752, 800, 807, 827,
835
Shvartsman, 96, 97, 547, 841
Sickel, 655, 810, 840
Simon, 259, 319, 841
Simonovits, 320, 825, 830
Sinnamon, 63, 841
Sj
odin, 538, 841
Skrypnik, 151, 814
Slobodecki, 515, 806, 841
Smirnov, 30, 841
Smith, 30, 31, 117, 402, 841
Sobolev, A., 220, 734, 806, 828
Sobolev, S. L., xxiii, xxviii, 1, 2931,
83, 86, 520, 841, 842
Sobolevskii, 219, 229, 831, 842
Sodin, 320, 837
Solonnikov, 86, 520, 545, 820, 842
Soucek, 507, 842
Sperner, 320, 507, 809
Spruck, 259, 319, 823
Stampacchia, 402, 842
Author Index
Stanoyevitch, 30, 31, 117, 402, 841,
842
Stegenga, 30, 117, 402, 841, 842
Stein, 28, 32, 64, 136, 511, 518, 519,
599, 625, 658, 700, 817, 842
Stepanov, 63, 253, 837, 839, 841, 842
Stredulinsky, 177, 843
Strichartz, 518, 519, 842
Stroock, 187, 804, 823, 843
Stummel, 587, 843
Sugawa, 734, 843
Sullivan, 33, 843
Swanson, 40, 544, 691, 830, 843
Szeg
o, 150, 151, 401, 839
Szeptycki, 545, 653, 805, 836
Taibleson, 545, 843
Takeda, 252, 843
Talenti, 85, 151, 177, 843
Tartar, xxviii, 843
Taylor, J. E., 140, 843
Taylor, M. E., 434, 800, 801, 843, 844
Tertikas, 141, 220, 806, 817, 844
Thirring, 286, 819
Tidblom, 219, 220, 817, 844
Tilli, 286, 805
Tintarev, 141, 179, 220, 839, 844
Tonelli, 86, 844
Tricarico, 402, 816
Triebel, xxviii, 511, 515, 519, 545, 691,
822, 844
Trombetti, 220, 805, 810
Troyanov, 152, 286, 321, 819, 820, 844
Trudinger, 86, 256, 260, 844
Trushin, 404, 844
Tuominen, 286, 821
Tyson, 261, 402, 811, 827
Uemura, 252, 818
Ukhlov, 321, 820, 844, 845
Uraltseva, 210, 212, 845
Ushakova, 253, 842
Uspenski, 249, 515, 845
V
ais
al
a, 32, 845
Varonen, 152, 822
Varopoulos, 260, 845
Vaugon, 179, 822
V
azquez, 219, 810, 845
865
Vega, 220, 815

Verbitsky, 1, 54, 63, 177, 204, 252,
253, 551, 564, 565, 567, 573, 607,
812, 821, 825, 835, 839, 845
Villani, 178, 401, 814, 830
Vodopyanov, 87, 88, 321, 820, 844,
845
Volberg, 320, 547, 837, 845
Volevic, 545, 845
Volpert, 507, 845
Vondracek, 252, 845
Wainger, 177, 232, 434, 810
Wallin, xxviii, 519, 545, 547, 824
Wang, F.-Y., 187, 812, 846
Wang, Z.-Q., 220, 846
Wannebo, 733, 734, 845, 846
Warma, 401, 805, 808
Weidl, 219, 829, 846
Weinberger, 320, 401, 839
Weissler, 187, 846
Werner, 178, 846
Wheeden, 253, 286, 565, 567, 813, 845
Whitney, 40, 846
Wik, 528, 846
Wol, 538, 540, 541, 573, 578, 607,
823
Wu, 607, 846
Wul, 140, 846
Xiao, 178, 187, 252, 548, 607, 804,
825, 846
Xu, 85, 814
Yadava, 179, 804
Yafaev, 219, 846
Yamabe, 260, 846
Yang, D., 178, 813, 830
Yang, S., 97, 846
Yaotin, 220, 846
Yau, 258, 813, 846
Ye, 178, 846
Yerzakova, 458, 847
Yosida, 187, 847
Yuan, 220, 815
Yudovich, 86, 256, 608, 847
Zalgaller, 140, 258, 259, 507, 811
Zegarlinski, 187, 252, 809, 820, 830,
843, 847
866
Author Index
Zhang, 178, 813, 830, 847

Zhihui, 220, 846
Ziemer, xxviii, 28, 40, 544, 691, 830,
843, 847
Zographopoulos, 220, 844

Zuazua, 219, 845
Zygmund, 518

2997-Sobolev Spaces - With Applications To Elliptic Partial Differential Equations (2nd)

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

2997-Sobolev Spaces - With Applications To Elliptic Partial Differential Equations (2nd)

Загружено:

Авторское право:

Доступные форматы

Grundlehren der

For further volumes:

Professor Vladimir Mazya

Sobolev spaces, i.e., the classes of functions with derivatives in Lp , occupy

or subsection we rst indicate the number of this section or subsection. For

Basic Properties of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . .

1.2.3 Representation of the Lebesgue Integral as a Riemann

1.7.4 Density of Bounded Functions in L2p () for

Inequalities for Functions Vanishing at the Boundary . . . . . . 123

2.3.9 Inequality with the Norms in Lq (, ) and Lr (, )

1 () into BirnbaumOrlicz Spaces . . . . . . . . . . . . . 179

Conductor and Capacitary Inequalities with Applications

Generalizations for Functions on Manifolds and

Integrability of Functions in the Space L11 () . . . . . . . . . . . . . . 287

5.4 Two-Sided Estimates for the Function for the Domain in

Integrability of Functions in the Space L1p () . . . . . . . . . . . . . . 323

0 < q < p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 342

6.5 More on the Nikod

6.7.1 Classes J and Ip, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364

6.7.3 Example of a Domain in the Class Ip, . . . . . . . . . . . . . . 368

6.7.4 Space L1p () and Its Embedding into Lq () . . . . . . . . . . 370

Continuity and Boundedness of Functions in Sobolev

7.1.3 Isoperimetric Function and a BrezisGallouet

Localization Moduli of Sobolev Embeddings for General

Space of Functions of Bounded Variation . . . . . . . . . . . . . . . . . . 459

9.6.2 Coincidence of the Trace and the Rough Trace . . . . . . . . 501

11.7 Cartan-Type Theorem and Estimates for Capacities . . . . . . . . . 575

12.3.2 Inequality with a Fractional Derivative Operator on

13 A Variant of Capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 657

14.3.7 A Logarithmic Sobolev Inequality with Application

Comments to Chap. 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691

15 Embedding of the Space

Llp (, ) Wrm () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739

Llp (, ) Wrm () . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 742

In [711713] Sobolev proved general integral inequalities for dierentiable

A sketch of a fairly rich prehistory of Sobolev spaces can be found in Naumann

of Chap. 2 are applied to the spectral theory of the Schrodinger operator in

Chapters 29 mainly concern functions with rst derivatives in Lp or in C .

where Nt = {x : |u(x)| t}. Inequalities of this type, initially found by the

completion of the space C0 () with respect to the norm l u Lp () . It is

can be interpreted as necessary and sucient conditions for the solvability of

is the completion of C0 () with respect to the norm

The topic of Chap. 17 is a necessary and sucient condition for density

1.1 The Spaces Llp(), Vpl() and Wpl ()

1 Basic Properties of Sobolev Spaces

we mean the space of

We use the notation L () for the space of essentially bounded Lebesgue

1.1 The Spaces Llp (), Vpl () and Wpl ()

Theorem. Any element of Llp () is in Lp (, loc).

where the star denotes convolution. We note that T C (Rn ). So, we

The proof follows immediately from the inclusion D u Lp

1 Basic Properties of Sobolev Spaces

1.1.3 Absolute Continuity of Functions in L1p ()

The function is called a mollier and is called a radius of mollication.

The properties of a mollication enable us to prove easily that

for all k = 1, 2, . . . , imply that g(t) = const a.e. on (0, 1).

1.1 The Spaces Llp (), Vpl () and Wpl ()

 with the radii dist(, R \(0, 1))/2 , i = 1, 2, . . . . Clearly the union

We observe that if f L1 (0, 1) and

for all k = 1, 2, . . . , then

where D(0, 1) and

Indeed, if g L1 (0, 1) and

g(t)k (t) dt = 0 for k = 1, 2, . . . ,

completion of the space C0 () with respect to the norm l uLp () . It is

with the radii dist(, R \(0, 1))/2 , i = 1, 2, . . . . Clearly the union

Therefore, uL2 () = whereas uW22 () < .